diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_0/index_tensor([8800, 8800])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_0/index_tensor([8800, 8800])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..19c80c6515a25f09b565b071505fd8761ea291e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_0/index_tensor([8800, 8800])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle shot captures the pinnacle of athleticism, the camera smoothly tracking a male gymnast executing a powerful lunge. Starting with an explosive push of his hands on the leather-covered pommel horse, his lean, muscular body soars into the air. He wears a tight, sleeveless royal blue and white leotard and crisp white trousers, his short, dark hair damp with sweat. Under the bright, even glare of the overhead stage lights, shadows are minimal, and his body forms a perfectly straight line. Then, gracefully and decisively, he twists through the air, legs firmly together, toes pointed, completing a full rotation with astonishing precision. His face is a mask of intense concentration, jaw set, eyes fixed on his landing. The soft-focus blur of the vast arena and the audience in the distance in the background isolates the gymnast from the deep blue of the thick landing mat below, creating an intense and awe-inspiring atmosphere that celebrates a moment of perfect form and controlled strength.", + "seed": 13689, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_0/index_8800_seed_13689_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1598755544051528, + "vlm_logprob": -0.1598755544051528, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.1598755544051528, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4374346435070038 + ], + "opposite_logprob_tokens": [ + -1.5624346733093262 + ], + "logprob_reward": -0.4374346435070038, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.4374346435070038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12971512973308563, + 0.04071006178855896, + 0.06208197399973869, + 0.5076497793197632, + 0.03646896779537201, + 0.07666707783937454, + 0.0486614964902401, + 0.019899588078260422, + 0.05160047486424446, + 0.0265454463660717 + ], + "credit_map": [ + 0.015344020910561085, + 0.00957554206252098, + 0.021599993109703064, + 0.02129051461815834, + 0.012014942243695259, + 0.0160902738571167, + 0.014445368200540543, + 0.01935446821153164, + 0.0045325965620577335, + 0.003942732699215412, + 0.004285864997655153, + 0.004927706904709339, + 0.005395346321165562, + 0.004168007522821426, + 0.008360002189874649, + 0.005097805056720972, + 0.0066748312674462795, + 0.004878966137766838, + 0.010158615186810493, + 0.014616429805755615, + 0.005538411904126406, + 0.006317140068858862, + 0.007750217337161303, + 0.006147366017103195, + 0.05450037866830826, + 0.18677020072937012, + 0.1021738201379776, + 0.02877173386514187, + 0.06817170977592468, + 0.02650168351829052, + 0.023114560171961784, + 0.01764562726020813, + 0.004496536683291197, + 0.003218985628336668, + 0.005131806246936321, + 0.0047403061762452126, + 0.004236969631165266, + 0.003416030202060938, + 0.0037188518326729536, + 0.007509482093155384, + 0.010795476846396923, + 0.00845966674387455, + 0.006897950079292059, + 0.006980223581194878, + 0.015845144167542458, + 0.01382414996623993, + 0.005425982642918825, + 0.00843848567456007, + 0.005992602091282606, + 0.009839937090873718, + 0.0051348102279007435, + 0.004423466976732016, + 0.008130313828587532, + 0.006452053785324097, + 0.004975678864866495, + 0.0037126357201486826, + 0.0023228980135172606, + 0.0035323072224855423, + 0.00225512171164155, + 0.002004518173635006, + 0.0032767080701887608, + 0.002774616703391075, + 0.0023306270595639944, + 0.0014027933357283473, + 0.00674057612195611, + 0.006394732277840376, + 0.005941844545304775, + 0.0026797426398843527, + 0.003052615327760577, + 0.015272644348442554, + 0.0028173443861305714, + 0.008700975216925144, + 0.0034839645959436893, + 0.004577277693897486, + 0.004357228055596352, + 0.0016660343389958143, + 0.0030087572522461414, + 0.0032731664832681417, + 0.0028139294590801, + 0.003365088952705264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.4374346435070038 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06211763247847557 + ], + "opposite_logprob_tokens": [ + -2.812117576599121 + ], + "logprob_reward": -0.06211763247847557, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06211763247847557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06713367998600006, + 0.044037919491529465, + 0.028320694342255592, + 0.6832326650619507, + 0.0201032105833292, + 0.04374798387289047, + 0.05501937493681908, + 0.013795440085232258, + 0.026544757187366486, + 0.018064258620142937 + ], + "credit_map": [ + 0.007715817075222731, + 0.005281044635921717, + 0.01340776402503252, + 0.010979064740240574, + 0.00463530607521534, + 0.007871679030358791, + 0.006456397939473391, + 0.010786608792841434, + 0.004010220058262348, + 0.004416997544467449, + 0.004326672293245792, + 0.004848361946642399, + 0.007654077839106321, + 0.006723188329488039, + 0.006383958272635937, + 0.005674444604665041, + 0.0034002983011305332, + 0.002566634677350521, + 0.004479427821934223, + 0.0064046368934214115, + 0.002206826815381646, + 0.0023891525343060493, + 0.003892172360792756, + 0.002981546800583601, + 0.07316321134567261, + 0.2535938620567322, + 0.13353988528251648, + 0.03817212954163551, + 0.09882735460996628, + 0.03456806018948555, + 0.0304722897708416, + 0.020895887166261673, + 0.0020038061775267124, + 0.0024314026813954115, + 0.0038196523673832417, + 0.004573068581521511, + 0.0015680871438235044, + 0.0020486728753894567, + 0.0016344834584742785, + 0.0020240377634763718, + 0.005519421771168709, + 0.004953915718942881, + 0.0033633511047810316, + 0.0040552192367613316, + 0.010440126992762089, + 0.0057040611281991005, + 0.003708980977535248, + 0.006002906244248152, + 0.005578764248639345, + 0.009583592414855957, + 0.004651454742997885, + 0.006535555236041546, + 0.009860923513770103, + 0.006434927228838205, + 0.007813592441380024, + 0.004560564178973436, + 0.00147810741327703, + 0.0028006478678435087, + 0.0016566469566896558, + 0.001376708853058517, + 0.002249842043966055, + 0.0018926648190245032, + 0.0012990464456379414, + 0.0010417758021503687, + 0.0028039596509188414, + 0.0035476461052894592, + 0.0052516576834023, + 0.0021771222818642855, + 0.0025578155182302, + 0.0030066275503486395, + 0.002869051182642579, + 0.004330876749008894, + 0.0022398338187485933, + 0.0015792511403560638, + 0.0037263664416968822, + 0.0012151036644354463, + 0.00211282423697412, + 0.0019128852291032672, + 0.0011568459449335933, + 0.004121147096157074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.06211763247847557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306282848119736 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306282848119736, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306282848119736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051750048995018005, + 0.018223216757178307, + 0.04353884607553482, + 0.7391274571418762, + 0.02690533548593521, + 0.022482670843601227, + 0.057058185338974, + 0.007517004385590553, + 0.019812338054180145, + 0.013584805652499199 + ], + "credit_map": [ + 0.00408914964646101, + 0.00402116822078824, + 0.008744271472096443, + 0.011225338093936443, + 0.0055288467556238174, + 0.005120603367686272, + 0.004671648144721985, + 0.008349028415977955, + 0.0017505738651379943, + 0.0015037690754979849, + 0.0018930104561150074, + 0.0020250212401151657, + 0.0024673352017998695, + 0.0020392811857163906, + 0.0037911124527454376, + 0.0027531147934496403, + 0.004106862470507622, + 0.003098404034972191, + 0.006800118833780289, + 0.00968173611909151, + 0.0035560044925659895, + 0.004641205072402954, + 0.00605250708758831, + 0.005602010991424322, + 0.07771537452936172, + 0.2784131169319153, + 0.13776564598083496, + 0.0383070707321167, + 0.116383396089077, + 0.03673909977078438, + 0.03164651617407799, + 0.022157300263643265, + 0.003396166255697608, + 0.0028538815677165985, + 0.0035970958415418863, + 0.005065409932285547, + 0.002513263840228319, + 0.0029804129153490067, + 0.002577360486611724, + 0.0039217486046254635, + 0.0034109894186258316, + 0.0032984670251607895, + 0.001986840972676873, + 0.002224231604486704, + 0.0044689844362437725, + 0.0027382206171751022, + 0.0018547289073467255, + 0.002500209491699934, + 0.005025292746722698, + 0.009434374049305916, + 0.004371499177068472, + 0.007663254160434008, + 0.009400026872754097, + 0.006752315443009138, + 0.009782843291759491, + 0.0046285842545330524, + 0.0008171806111931801, + 0.001249520806595683, + 0.0009511216776445508, + 0.0007471534190699458, + 0.0012248640414327383, + 0.0008339163032360375, + 0.0009933013934642076, + 0.0006999471224844456, + 0.0025704996660351753, + 0.0028157723136246204, + 0.0026702696923166513, + 0.0013527919072657824, + 0.0017120426055043936, + 0.005628618877381086, + 0.001654015271924436, + 0.001408328884281218, + 0.0019113970920443535, + 0.0022472802083939314, + 0.0018849842017516494, + 0.0012205980019643903, + 0.0014882853720337152, + 0.0017506902804598212, + 0.0011377615155652165, + 0.0019438096787780523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.04306282848119736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252124160528183 + ], + "opposite_logprob_tokens": [ + -1.5021241903305054 + ], + "logprob_reward": -0.252124160528183, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252124160528183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2669181227684021, + 0.14470726251602173, + 0.09453669935464859, + 0.12573213875293732, + 0.08155129849910736, + 0.0817115530371666, + 0.06821921467781067, + 0.03133625537157059, + 0.06710466742515564, + 0.038182806223630905 + ], + "credit_map": [ + 0.025403397157788277, + 0.01968216709792614, + 0.03799101337790489, + 0.0578913576900959, + 0.024065474048256874, + 0.031322624534368515, + 0.0277106910943985, + 0.04285140708088875, + 0.012370344251394272, + 0.01060049794614315, + 0.015332421287894249, + 0.016870049759745598, + 0.01836853288114071, + 0.01727791130542755, + 0.033237166702747345, + 0.020650329068303108, + 0.012752016074955463, + 0.010052932426333427, + 0.012782602570950985, + 0.013057885691523552, + 0.008618605323135853, + 0.010315285995602608, + 0.012895356863737106, + 0.014062016271054745, + 0.011273841373622417, + 0.04034655913710594, + 0.014524556696414948, + 0.006806389894336462, + 0.028069401159882545, + 0.009177970699965954, + 0.00709187425673008, + 0.008441534824669361, + 0.009205090813338757, + 0.009226628579199314, + 0.010254565626382828, + 0.018394343554973602, + 0.007930845953524113, + 0.008995258249342442, + 0.00786418467760086, + 0.009680375456809998, + 0.013008917681872845, + 0.009046497754752636, + 0.0064963167533278465, + 0.009283933788537979, + 0.011170567944645882, + 0.018497945740818977, + 0.006234585307538509, + 0.007972783409059048, + 0.00854028295725584, + 0.011836584657430649, + 0.00908772461116314, + 0.006062189117074013, + 0.00851343385875225, + 0.009322774596512318, + 0.006207792088389397, + 0.008648427203297615, + 0.003473383840173483, + 0.0038271048106253147, + 0.003632055129855871, + 0.003344953525811434, + 0.00455909688025713, + 0.004227759782224894, + 0.0050355298444628716, + 0.0032363696955144405, + 0.01067871879786253, + 0.009942689910531044, + 0.007519163656979799, + 0.00682145357131958, + 0.006373696960508823, + 0.010094825178384781, + 0.007689329329878092, + 0.00798479188233614, + 0.008391570299863815, + 0.0040774173103272915, + 0.003978535067290068, + 0.0030763058457523584, + 0.0037523226346820593, + 0.004940785001963377, + 0.003326976438984275, + 0.0066388919949531555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.252124160528183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638507030904293 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638507030904293, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638507030904293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07207083702087402, + 0.06728117167949677, + 0.08815024793148041, + 0.32085680961608887, + 0.024243509396910667, + 0.18475688993930817, + 0.06431607156991959, + 0.045937955379486084, + 0.04878984019160271, + 0.08359664678573608 + ], + "credit_map": [ + 0.008495436981320381, + 0.004989003296941519, + 0.014150665141642094, + 0.009414290077984333, + 0.0078032249584794044, + 0.01081555150449276, + 0.006022733170539141, + 0.010379931889474392, + 0.008932639844715595, + 0.005901703145354986, + 0.008853782899677753, + 0.009824290871620178, + 0.007377154193818569, + 0.005874416325241327, + 0.010240145958960056, + 0.010277030989527702, + 0.008974304422736168, + 0.00694690877571702, + 0.013171598315238953, + 0.01906520500779152, + 0.008935452438890934, + 0.009486031718552113, + 0.010973862372338772, + 0.010596883483231068, + 0.034955430775880814, + 0.11690687388181686, + 0.0686502531170845, + 0.0194429699331522, + 0.03911060839891434, + 0.016457803547382355, + 0.014973212964832783, + 0.010359654203057289, + 0.0021716798655688763, + 0.0023264114279299974, + 0.0033866255544126034, + 0.0052566854283213615, + 0.0030565711203962564, + 0.002419081050902605, + 0.0031111971475183964, + 0.0025152568705379963, + 0.03419212996959686, + 0.01857224851846695, + 0.01457192562520504, + 0.012157783843576908, + 0.030525460839271545, + 0.04754941910505295, + 0.006885398179292679, + 0.020302530378103256, + 0.007166016381233931, + 0.011937370523810387, + 0.005285957362502813, + 0.007204053923487663, + 0.010700960643589497, + 0.00788635853677988, + 0.008474727161228657, + 0.005660624708980322, + 0.009919176809489727, + 0.006498800124973059, + 0.008058322593569756, + 0.003078901208937168, + 0.005468405783176422, + 0.006415869574993849, + 0.002396609866991639, + 0.004101870581507683, + 0.0054842326790094376, + 0.006077272817492485, + 0.007975972257554531, + 0.0038618098478764296, + 0.004575909581035376, + 0.0073631526902318, + 0.0046796598471701145, + 0.008771829307079315, + 0.015578887425363064, + 0.00792273972183466, + 0.02439774014055729, + 0.0035704616457223892, + 0.006148813758045435, + 0.009960610419511795, + 0.0030515268445014954, + 0.012965871952474117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.004638507030904293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot captures the pinnacle of athleticism, the camera smoothly tracking a male gymnast executing a powerful lunge. Starting with an explosive push of his hands on the leather-covered pommel horse, his lean, muscular body soars into the air. He wears a tight, sleeveless royal blue and white leotard and crisp white trousers, his short, dark hair damp with sweat. Under the bright, even glare of the overhead stage lights, shadows are minimal, and his body forms a perfectly straight line. Then, gracefully and decisively, he twists through the air, legs firmly together, toes pointed, completing a full rotation with astonishing precision. His face is a mask of intense concentration, jaw set, eyes fixed on his landing. The soft-focus blur of the vast arena and the audience in the distance in the background isolates the gymnast from the deep blue of the thick landing mat below, creating an intense and awe-inspiring atmosphere that celebrates a moment of perfect form and controlled strength.", + "seed": 13690, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_0/index_8800_seed_13690_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14768786625936628, + "vlm_logprob": -0.14768786625936628, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.14768786625936628, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22647272050380707 + ], + "opposite_logprob_tokens": [ + -1.9764727354049683 + ], + "logprob_reward": -0.22647272050380707, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.22647272050380707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1373075246810913, + 0.0769338309764862, + 0.08275022357702255, + 0.13261626660823822, + 0.11818492412567139, + 0.13824303448200226, + 0.11384443193674088, + 0.034461572766304016, + 0.1253572702407837, + 0.040300942957401276 + ], + "credit_map": [ + 0.02059374563395977, + 0.017247682437300682, + 0.021679548546671867, + 0.018405964598059654, + 0.015456545166671276, + 0.014485541731119156, + 0.01434677466750145, + 0.015091713517904282, + 0.008337232284247875, + 0.010123022831976414, + 0.011274897493422031, + 0.009042833000421524, + 0.008344119414687157, + 0.007701328955590725, + 0.010316325351595879, + 0.01179406139999628, + 0.00834618415683508, + 0.008166259154677391, + 0.009407002478837967, + 0.013163807801902294, + 0.010005471296608448, + 0.010540966875851154, + 0.01025751419365406, + 0.012863012962043285, + 0.011748092249035835, + 0.01294056884944439, + 0.014400867745280266, + 0.025254853069782257, + 0.02722819522023201, + 0.015463368035852909, + 0.014626238495111465, + 0.010954071767628193, + 0.019333407282829285, + 0.016504697501659393, + 0.014135845005512238, + 0.014915093779563904, + 0.014522284269332886, + 0.017511574551463127, + 0.013351998291909695, + 0.007910014130175114, + 0.017718179151415825, + 0.014409697614610195, + 0.017863458022475243, + 0.014598360285162926, + 0.017387740314006805, + 0.018282456323504448, + 0.012235124595463276, + 0.025748012587428093, + 0.020692111924290657, + 0.016675224527716637, + 0.013875567354261875, + 0.00895706657320261, + 0.010346819646656513, + 0.026144783943891525, + 0.005340288393199444, + 0.011812563985586166, + 0.004236807115375996, + 0.005366211291402578, + 0.004634390119463205, + 0.002593747805804014, + 0.005072271451354027, + 0.006002708803862333, + 0.003578389994800091, + 0.002977042691782117, + 0.015242607332766056, + 0.018460718914866447, + 0.0299172755330801, + 0.005628456361591816, + 0.006387705449014902, + 0.007926030084490776, + 0.009665198624134064, + 0.03212926909327507, + 0.007705563213676214, + 0.004526330623775721, + 0.005220391787588596, + 0.0034974501468241215, + 0.005202448461204767, + 0.005731250625103712, + 0.003681373782455921, + 0.004736131522804499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.22647272050380707 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209690868854523 + ], + "opposite_logprob_tokens": [ + -2.8120968341827393 + ], + "logprob_reward": -0.06209690868854523, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209690868854523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1970396637916565, + 0.041719693690538406, + 0.0914061963558197, + 0.11352844536304474, + 0.15386375784873962, + 0.18702614307403564, + 0.04781731218099594, + 0.06045372411608696, + 0.07463613897562027, + 0.03250890597701073 + ], + "credit_map": [ + 0.019333014264702797, + 0.021225139498710632, + 0.03238741680979729, + 0.024331901222467422, + 0.024628443643450737, + 0.019958896562457085, + 0.023597022518515587, + 0.03157782182097435, + 0.004941860213875771, + 0.004728719592094421, + 0.006956390105187893, + 0.006263489834964275, + 0.005711001344025135, + 0.0043944926001131535, + 0.003923769574612379, + 0.004799970425665379, + 0.008041751570999622, + 0.012935430742800236, + 0.010498715564608574, + 0.015447007492184639, + 0.011069927364587784, + 0.010922490619122982, + 0.009587525390088558, + 0.012903343886137009, + 0.010037938132882118, + 0.017798766493797302, + 0.00992971658706665, + 0.020098257809877396, + 0.015555400401353836, + 0.014063450507819653, + 0.015630273148417473, + 0.010414643213152885, + 0.04282194375991821, + 0.01613583043217659, + 0.010576212778687477, + 0.011942708864808083, + 0.011903216131031513, + 0.04139251634478569, + 0.010739177465438843, + 0.008352170698344707, + 0.02472163923084736, + 0.033301446586847305, + 0.013938832096755505, + 0.020387478172779083, + 0.0300289373844862, + 0.03248385712504387, + 0.01644967310130596, + 0.015714270994067192, + 0.007818680256605148, + 0.006163924001157284, + 0.005259191617369652, + 0.004589863587170839, + 0.0061692469753324986, + 0.010100307874381542, + 0.0032106684520840645, + 0.004505431745201349, + 0.008875194005668163, + 0.00833350233733654, + 0.006262127310037613, + 0.00989766139537096, + 0.008339556865394115, + 0.007572619244456291, + 0.005204834043979645, + 0.0059682270511984825, + 0.006828329060226679, + 0.011387764476239681, + 0.018557202070951462, + 0.0044119954109191895, + 0.00742505444213748, + 0.00706520676612854, + 0.0067100864835083485, + 0.012250497937202454, + 0.005865838378667831, + 0.004066270776093006, + 0.0026353260036557913, + 0.0025241789408028126, + 0.004597880877554417, + 0.0057091498747467995, + 0.0037947159726172686, + 0.00331554445438087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.06209690868854523 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205290004611015 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.014205290004611015, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205290004611015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11214440315961838, + 0.08134939521551132, + 0.06854350864887238, + 0.30338454246520996, + 0.11444714665412903, + 0.09905393421649933, + 0.03060690127313137, + 0.032949090003967285, + 0.09291376918554306, + 0.0646074041724205 + ], + "credit_map": [ + 0.010981891304254532, + 0.012452792376279831, + 0.021476944908499718, + 0.012653395533561707, + 0.012474249117076397, + 0.01302928663790226, + 0.014169791713356972, + 0.014906042255461216, + 0.005772273521870375, + 0.009920720010995865, + 0.011984475888311863, + 0.009861175902187824, + 0.009922035038471222, + 0.006717041600495577, + 0.010478164069354534, + 0.01669350080192089, + 0.0055633531883358955, + 0.014115002006292343, + 0.007155410014092922, + 0.011216516606509686, + 0.0062449052929878235, + 0.006248727440834045, + 0.009332611225545406, + 0.008666973561048508, + 0.02832268923521042, + 0.032351646572351456, + 0.051165539771318436, + 0.03544312343001366, + 0.0466945581138134, + 0.03155971318483353, + 0.04326169565320015, + 0.034585535526275635, + 0.036317046731710434, + 0.01125284843146801, + 0.006610960699617863, + 0.008293922059237957, + 0.007722157519310713, + 0.032566893845796585, + 0.006149062421172857, + 0.005534247029572725, + 0.022721771150827408, + 0.007976899854838848, + 0.011302868835628033, + 0.013302634470164776, + 0.011286665685474873, + 0.011933967471122742, + 0.007941585965454578, + 0.012587533332407475, + 0.0055130948312580585, + 0.003539395984262228, + 0.003153510857373476, + 0.0027517564594745636, + 0.005682562477886677, + 0.004699391778558493, + 0.0016646543517708778, + 0.003602531971409917, + 0.00431418139487505, + 0.004844842478632927, + 0.0034387768246233463, + 0.0028046383522450924, + 0.004496176261454821, + 0.004534543491899967, + 0.003885426791384816, + 0.004630500450730324, + 0.008804487995803356, + 0.00789766013622284, + 0.018735062330961227, + 0.00832277350127697, + 0.011853748001158237, + 0.011379474774003029, + 0.007930034771561623, + 0.017990518361330032, + 0.014848161488771439, + 0.007323578465729952, + 0.004964052699506283, + 0.0047060404904186726, + 0.007877356372773647, + 0.012119270861148834, + 0.006644668523222208, + 0.006124267354607582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.014205290004611015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289277493953705 + ], + "opposite_logprob_tokens": [ + -1.0539277791976929 + ], + "logprob_reward": -0.4289277493953705, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289277493953705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30088257789611816, + 0.052404507994651794, + 0.10808622092008591, + 0.07773945480585098, + 0.09087029844522476, + 0.1663144826889038, + 0.04047587886452675, + 0.025543944910168648, + 0.09531989693641663, + 0.042362749576568604 + ], + "credit_map": [ + 0.04388846084475517, + 0.028437305241823196, + 0.0564696379005909, + 0.022860508412122726, + 0.038548726588487625, + 0.0333402119576931, + 0.03506466746330261, + 0.04227305203676224, + 0.0059928689152002335, + 0.0043442570604383945, + 0.00804070197045803, + 0.009808712638914585, + 0.00612353440374136, + 0.0055239237844944, + 0.006273928098380566, + 0.0062965876422822475, + 0.01032619085162878, + 0.01451568678021431, + 0.012387862429022789, + 0.01589732989668846, + 0.011726133525371552, + 0.012958803214132786, + 0.0129702715203166, + 0.017303945496678352, + 0.006342429202049971, + 0.012141885235905647, + 0.007166295312345028, + 0.015444395132362843, + 0.008775360882282257, + 0.011163009330630302, + 0.009694971144199371, + 0.007011100649833679, + 0.02406477928161621, + 0.01104036532342434, + 0.0061896429397165775, + 0.0070175426080822945, + 0.006637096870690584, + 0.02422439493238926, + 0.006293353158980608, + 0.005403124261647463, + 0.03207436203956604, + 0.016194801777601242, + 0.02194046974182129, + 0.017436007037758827, + 0.02407078631222248, + 0.019742323085665703, + 0.015117297880351543, + 0.0197384525090456, + 0.00722717447206378, + 0.0040699029341340065, + 0.006871524732559919, + 0.004356264136731625, + 0.00592120923101902, + 0.0050774854607880116, + 0.0021379159297794104, + 0.004814400803297758, + 0.0031515012960880995, + 0.003460894338786602, + 0.0027884719893336296, + 0.0031340382993221283, + 0.00422549806535244, + 0.0036953752860426903, + 0.0019393800757825375, + 0.0031487836968153715, + 0.006860882043838501, + 0.010700730606913567, + 0.030240759253501892, + 0.005403489805757999, + 0.007193433586508036, + 0.005734141916036606, + 0.007014723028987646, + 0.022171737626194954, + 0.007530158385634422, + 0.008234847337007523, + 0.0041632927022874355, + 0.0024571926333010197, + 0.004672219045460224, + 0.005473749246448278, + 0.0038680618163198233, + 0.0059632305055856705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.4289277493953705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736662704497576 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736662704497576, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736662704497576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12218225747346878, + 0.28088414669036865, + 0.19485177099704742, + 0.06938628107309341, + 0.06283323466777802, + 0.09046941995620728, + 0.03505122289061546, + 0.04913352057337761, + 0.056185923516750336, + 0.03902213275432587 + ], + "credit_map": [ + 0.02512473613023758, + 0.01426231861114502, + 0.0155175244435668, + 0.01239049807190895, + 0.01372726820409298, + 0.014906245283782482, + 0.014603559859097004, + 0.011650120839476585, + 0.030040651559829712, + 0.033122383058071136, + 0.05432598665356636, + 0.042305637151002884, + 0.030053989961743355, + 0.023180143907666206, + 0.036059923470020294, + 0.03179545700550079, + 0.01148802787065506, + 0.01338417548686266, + 0.015781912952661514, + 0.023692142218351364, + 0.029674366116523743, + 0.026473773643374443, + 0.040669359266757965, + 0.03368803858757019, + 0.009397385641932487, + 0.00572046497836709, + 0.006059172563254833, + 0.008179672993719578, + 0.015048366971313953, + 0.012761558406054974, + 0.005943710450083017, + 0.006275954190641642, + 0.011830339208245277, + 0.006959886290132999, + 0.006525911390781403, + 0.005445215385407209, + 0.006833448074758053, + 0.0114975580945611, + 0.006139669101685286, + 0.007601213175803423, + 0.012897583656013012, + 0.009049749933183193, + 0.008633513934910297, + 0.009827806614339352, + 0.014827387407422066, + 0.015745466575026512, + 0.007428746670484543, + 0.012059169821441174, + 0.004735379945486784, + 0.004417561460286379, + 0.004375021904706955, + 0.004491129890084267, + 0.004007842391729355, + 0.0066778468899428844, + 0.0029729902744293213, + 0.0033734538592398167, + 0.004859075881540775, + 0.005777646787464619, + 0.008444982580840588, + 0.007791685406118631, + 0.004846577532589436, + 0.008752183057367802, + 0.00403875345364213, + 0.004622623790055513, + 0.0067863320000469685, + 0.014746188186109066, + 0.011192982085049152, + 0.0026983513962477446, + 0.004756424110382795, + 0.007678875233978033, + 0.0036785511765629053, + 0.0046482267789542675, + 0.005365655291825533, + 0.006207847967743874, + 0.005097958259284496, + 0.002544112503528595, + 0.005312427878379822, + 0.004460089839994907, + 0.002827900694683194, + 0.0072061424143612385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.006736662704497576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot captures the pinnacle of athleticism, the camera smoothly tracking a male gymnast executing a powerful lunge. Starting with an explosive push of his hands on the leather-covered pommel horse, his lean, muscular body soars into the air. He wears a tight, sleeveless royal blue and white leotard and crisp white trousers, his short, dark hair damp with sweat. Under the bright, even glare of the overhead stage lights, shadows are minimal, and his body forms a perfectly straight line. Then, gracefully and decisively, he twists through the air, legs firmly together, toes pointed, completing a full rotation with astonishing precision. His face is a mask of intense concentration, jaw set, eyes fixed on his landing. The soft-focus blur of the vast arena and the audience in the distance in the background isolates the gymnast from the deep blue of the thick landing mat below, creating an intense and awe-inspiring atmosphere that celebrates a moment of perfect form and controlled strength.", + "seed": 13691, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_0/index_8800_seed_13691_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10325236395001411, + "vlm_logprob": -0.10325236395001411, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.10325236395001411, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2078801691532135 + ], + "opposite_logprob_tokens": [ + -2.4578802585601807 + ], + "logprob_reward": -0.2078801691532135, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.2078801691532135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42184117436408997, + 0.069553904235363, + 0.06698235124349594, + 0.06226428598165512, + 0.2029774785041809, + 0.03987165912985802, + 0.03723987564444542, + 0.0317414253950119, + 0.04222358390688896, + 0.025304235517978668 + ], + "credit_map": [ + 0.035209961235523224, + 0.030240410938858986, + 0.08509210497140884, + 0.07645473629236221, + 0.029166871681809425, + 0.021652694791555405, + 0.0713212713599205, + 0.07270313799381256, + 0.0070378719829022884, + 0.004285756032913923, + 0.005724980961531401, + 0.007390283979475498, + 0.0045545208267867565, + 0.00798931997269392, + 0.015543480403721333, + 0.017027689144015312, + 0.0050000292249023914, + 0.0052033099345862865, + 0.01740645244717598, + 0.006835128180682659, + 0.008616509847342968, + 0.0058966707438230515, + 0.008727160282433033, + 0.009297095239162445, + 0.0071593052707612514, + 0.00944477692246437, + 0.010170437395572662, + 0.008013852871954441, + 0.005440298467874527, + 0.0037777090910822153, + 0.0064623975194990635, + 0.011795504949986935, + 0.01393507793545723, + 0.02238447777926922, + 0.03087279200553894, + 0.03327730670571327, + 0.027636293321847916, + 0.0175369530916214, + 0.026776691898703575, + 0.030557885766029358, + 0.005084612872451544, + 0.0049410611391067505, + 0.00732682179659605, + 0.005857414100319147, + 0.004707720130681992, + 0.002521586138755083, + 0.0037567324470728636, + 0.005675711669027805, + 0.0043808226473629475, + 0.0045244102366268635, + 0.005173820070922375, + 0.002900040941312909, + 0.004018078558146954, + 0.005124777089804411, + 0.004527685232460499, + 0.0065902466885745525, + 0.0036948705092072487, + 0.004927651025354862, + 0.0037753519136458635, + 0.0037102550268173218, + 0.003176224883645773, + 0.0034542777575552464, + 0.00453062541782856, + 0.004472165834158659, + 0.005793066695332527, + 0.00420967023819685, + 0.003203665604814887, + 0.002559519372880459, + 0.009490680880844593, + 0.007889976724982262, + 0.004584952723234892, + 0.0044920495711266994, + 0.0039788829162716866, + 0.0029520487878471613, + 0.0036549873184412718, + 0.003550595371052623, + 0.0025137350894510746, + 0.003001484787091613, + 0.0022545468527823687, + 0.0033979539293795824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.2078801691532135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209870055317879 + ], + "opposite_logprob_tokens": [ + -2.812098741531372 + ], + "logprob_reward": -0.06209870055317879, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209870055317879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13928833603858948, + 0.09868012368679047, + 0.12761302292346954, + 0.04546085745096207, + 0.293600469827652, + 0.05300898477435112, + 0.11581765115261078, + 0.036786146461963654, + 0.053843118250370026, + 0.035901300609111786 + ], + "credit_map": [ + 0.014571929350495338, + 0.009002619422972202, + 0.028901925310492516, + 0.025921661406755447, + 0.00837193988263607, + 0.007782964967191219, + 0.02212199755012989, + 0.022613300010561943, + 0.004793610889464617, + 0.008675715886056423, + 0.007754352409392595, + 0.01031348668038845, + 0.012754907831549644, + 0.0078012957237660885, + 0.02186136692762375, + 0.024725377559661865, + 0.00963139720261097, + 0.011498657986521721, + 0.02746487781405449, + 0.013388033956289291, + 0.016899816691875458, + 0.016738994047045708, + 0.015423737466335297, + 0.01656750962138176, + 0.004594217520207167, + 0.00539835449308157, + 0.005614595487713814, + 0.006418207194656134, + 0.003795845666900277, + 0.002991368994116783, + 0.007884391583502293, + 0.008763878606259823, + 0.02398323081433773, + 0.038216449320316315, + 0.040155649185180664, + 0.0445018894970417, + 0.039780713617801666, + 0.024302735924720764, + 0.039338674396276474, + 0.043321143835783005, + 0.005104224197566509, + 0.00645548477768898, + 0.009774550795555115, + 0.006067086011171341, + 0.004397035576403141, + 0.004694937728345394, + 0.006115484982728958, + 0.010400179773569107, + 0.008836011402308941, + 0.01750396378338337, + 0.01013717520982027, + 0.009596562013030052, + 0.01925482228398323, + 0.010839157737791538, + 0.022882863879203796, + 0.016767092049121857, + 0.0034407067578285933, + 0.005546513944864273, + 0.0041119009256362915, + 0.003143316600471735, + 0.007285690400749445, + 0.005072514060884714, + 0.003991801757365465, + 0.004193700384348631, + 0.006689179223030806, + 0.006652436684817076, + 0.0031622012611478567, + 0.0033422044944018126, + 0.01416253112256527, + 0.008299234323203564, + 0.005508097819983959, + 0.006027232389897108, + 0.0041518001817166805, + 0.003098847344517708, + 0.008600163273513317, + 0.004327384755015373, + 0.0030389109160751104, + 0.004085834603756666, + 0.002571963705122471, + 0.0060263932682573795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.06209870055317879 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607481949031353 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.01607481949031353, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607481949031353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19122476875782013, + 0.12166018038988113, + 0.07026028633117676, + 0.06491310894489288, + 0.07539279758930206, + 0.06902697682380676, + 0.0642215758562088, + 0.0698920488357544, + 0.19680197536945343, + 0.07660628110170364 + ], + "credit_map": [ + 0.025287356227636337, + 0.016047634184360504, + 0.02988019958138466, + 0.03653779625892639, + 0.016252921894192696, + 0.01494714803993702, + 0.02659638039767742, + 0.02567533403635025, + 0.012911384925246239, + 0.008542180992662907, + 0.013129904866218567, + 0.009640214964747429, + 0.00881285686045885, + 0.011326439678668976, + 0.028882687911391258, + 0.028414510190486908, + 0.005491278599947691, + 0.00760889146476984, + 0.016785556450486183, + 0.006505067925900221, + 0.008686409331858158, + 0.007468918804079294, + 0.006820867769420147, + 0.010893297381699085, + 0.0067856102250516415, + 0.009987195022404194, + 0.009154186584055424, + 0.010651379823684692, + 0.005100930575281382, + 0.0047074067406356335, + 0.008567542769014835, + 0.009958853013813496, + 0.007623001467436552, + 0.007506371010094881, + 0.009631614200770855, + 0.008246642537415028, + 0.015071706846356392, + 0.0046986243687570095, + 0.012243100441992283, + 0.010371735319495201, + 0.008399143815040588, + 0.009396345354616642, + 0.011133085936307907, + 0.007803917862474918, + 0.009525055065751076, + 0.004947890527546406, + 0.009105358272790909, + 0.008716181851923466, + 0.006012071389704943, + 0.006883335765451193, + 0.008437425829470158, + 0.0065077985636889935, + 0.008925397880375385, + 0.006061052903532982, + 0.009439654648303986, + 0.011954836547374725, + 0.006549003068357706, + 0.01015603356063366, + 0.007158635184168816, + 0.006628780625760555, + 0.009159675799310207, + 0.013835299760103226, + 0.007903432473540306, + 0.008501187898218632, + 0.02295462228357792, + 0.025551648810505867, + 0.014759954065084457, + 0.01563652977347374, + 0.03976172208786011, + 0.02454165183007717, + 0.024121562018990517, + 0.0294742863625288, + 0.008102050051093102, + 0.008492307737469673, + 0.015896020457148552, + 0.010752083733677864, + 0.0069198003038764, + 0.008247335441410542, + 0.0061880755238235, + 0.012008605524897575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.01607481949031353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557342052459717 + ], + "opposite_logprob_tokens": [ + -1.6005734205245972 + ], + "logprob_reward": -0.22557342052459717, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557342052459717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17678533494472504, + 0.08320004492998123, + 0.10452073812484741, + 0.04469120129942894, + 0.3431437015533447, + 0.04699404165148735, + 0.06573887169361115, + 0.03493678569793701, + 0.04528165981173515, + 0.054707691073417664 + ], + "credit_map": [ + 0.024502648040652275, + 0.015391546301543713, + 0.019246134907007217, + 0.028478149324655533, + 0.011388675309717655, + 0.016102122142910957, + 0.03323272243142128, + 0.02844332344830036, + 0.00996146909892559, + 0.00653871288523078, + 0.011591307818889618, + 0.011819858103990555, + 0.006119584199041128, + 0.006713511422276497, + 0.014012309722602367, + 0.0164432842284441, + 0.00779346376657486, + 0.010108774527907372, + 0.014925730414688587, + 0.011833409778773785, + 0.01310492679476738, + 0.013820548541843891, + 0.016257405281066895, + 0.016676470637321472, + 0.005662961862981319, + 0.005321426782757044, + 0.006526736076921225, + 0.006548186298459768, + 0.004109967965632677, + 0.002959759207442403, + 0.0061437287367880344, + 0.007418431341648102, + 0.026343828067183495, + 0.030205704271793365, + 0.04156412184238434, + 0.045395951718091965, + 0.06005208194255829, + 0.02480800449848175, + 0.06685178726911545, + 0.04792221263051033, + 0.006457225885242224, + 0.004537200089544058, + 0.012406102381646633, + 0.004424032289534807, + 0.004213795531541109, + 0.0027052746154367924, + 0.00385404285043478, + 0.008396366611123085, + 0.005399099085479975, + 0.00855166558176279, + 0.008081294596195221, + 0.005497407633811235, + 0.011151927523314953, + 0.005753334145992994, + 0.012003140524029732, + 0.009300998412072659, + 0.003076222725212574, + 0.003827691078186035, + 0.004328865557909012, + 0.004897475242614746, + 0.0037286977749317884, + 0.0050222366116940975, + 0.005828866735100746, + 0.004226725082844496, + 0.006188575178384781, + 0.006630711257457733, + 0.0042116823606193066, + 0.004625846166163683, + 0.007918818853795528, + 0.006128829438239336, + 0.005014726892113686, + 0.004562467336654663, + 0.005259030498564243, + 0.007207478862255812, + 0.011661517433822155, + 0.009335671551525593, + 0.00335964048281312, + 0.005135367624461651, + 0.004995808005332947, + 0.007753173820674419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.22557342052459717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2689671516418457, + 0.0700593814253807, + 0.14326803386211395, + 0.021203307434916496, + 0.08037152141332626, + 0.1243550106883049, + 0.14006583392620087, + 0.06808747351169586, + 0.05491601675748825, + 0.028706200420856476 + ], + "credit_map": [ + 0.038329727947711945, + 0.01798836700618267, + 0.03998878225684166, + 0.047393057495355606, + 0.016737673431634903, + 0.023028714582324028, + 0.04992884770035744, + 0.035571977496147156, + 0.005940854549407959, + 0.006086504086852074, + 0.01047773752361536, + 0.007055334281176329, + 0.010590939782559872, + 0.00932446401566267, + 0.009950499050319195, + 0.010633053258061409, + 0.014528438448905945, + 0.015503993257880211, + 0.018564166501164436, + 0.010360397398471832, + 0.015541690401732922, + 0.022606566548347473, + 0.017824511975049973, + 0.028338270261883736, + 0.0030396084766834974, + 0.003552248701453209, + 0.0031212831381708384, + 0.002980532357469201, + 0.001971879042685032, + 0.0016570501029491425, + 0.002533213933929801, + 0.0023474907502532005, + 0.010281859897077084, + 0.007071764208376408, + 0.012898306362330914, + 0.012921428307890892, + 0.00992600992321968, + 0.006628076080232859, + 0.010390941984951496, + 0.010253134183585644, + 0.016392726451158524, + 0.007622231263667345, + 0.03296138346195221, + 0.014952977187931538, + 0.009724142961204052, + 0.009089798666536808, + 0.007753188721835613, + 0.025858569890260696, + 0.015595652163028717, + 0.012319075874984264, + 0.014954001642763615, + 0.01139678806066513, + 0.022799775004386902, + 0.015726467594504356, + 0.022450832650065422, + 0.02482323721051216, + 0.005801620427519083, + 0.007033730857074261, + 0.004041597247123718, + 0.004590146709233522, + 0.017715997993946075, + 0.013331233523786068, + 0.007653306238353252, + 0.00791984237730503, + 0.00794031098484993, + 0.005392453633248806, + 0.0033883319701999426, + 0.0033276553731411695, + 0.012275751680135727, + 0.00999924261122942, + 0.00607980927452445, + 0.00651246402412653, + 0.002988799475133419, + 0.003309488296508789, + 0.004050562623888254, + 0.0035530629102140665, + 0.00321368221193552, + 0.006074708420783281, + 0.0022877769079059362, + 0.0032281181775033474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.004634710028767586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot captures the pinnacle of athleticism, the camera smoothly tracking a male gymnast executing a powerful lunge. Starting with an explosive push of his hands on the leather-covered pommel horse, his lean, muscular body soars into the air. He wears a tight, sleeveless royal blue and white leotard and crisp white trousers, his short, dark hair damp with sweat. Under the bright, even glare of the overhead stage lights, shadows are minimal, and his body forms a perfectly straight line. Then, gracefully and decisively, he twists through the air, legs firmly together, toes pointed, completing a full rotation with astonishing precision. His face is a mask of intense concentration, jaw set, eyes fixed on his landing. The soft-focus blur of the vast arena and the audience in the distance in the background isolates the gymnast from the deep blue of the thick landing mat below, creating an intense and awe-inspiring atmosphere that celebrates a moment of perfect form and controlled strength.", + "seed": 13692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_0/index_8800_seed_13692_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12478288542479277, + "vlm_logprob": -0.12478288542479277, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.12478288542479277, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3301297128200531 + ], + "opposite_logprob_tokens": [ + -1.8301297426223755 + ], + "logprob_reward": -0.3301297128200531, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.3301297128200531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23761841654777527, + 0.04108739644289017, + 0.24217656254768372, + 0.05598228797316551, + 0.0913294106721878, + 0.1198607087135315, + 0.07801856845617294, + 0.02915581874549389, + 0.032928865402936935, + 0.07184194028377533 + ], + "credit_map": [ + 0.015047930181026459, + 0.03459544852375984, + 0.02773510292172432, + 0.04025300219655037, + 0.02105652168393135, + 0.014296742156147957, + 0.0441056452691555, + 0.04052801430225372, + 0.004320033360272646, + 0.0034818851854652166, + 0.009232976473867893, + 0.00496528297662735, + 0.003330229315906763, + 0.004058190155774355, + 0.00460412772372365, + 0.007094672415405512, + 0.02011999487876892, + 0.004987918306142092, + 0.08833761513233185, + 0.013159633614122868, + 0.0096448864787817, + 0.013770519755780697, + 0.01037327479571104, + 0.08178271353244781, + 0.0047920322977006435, + 0.00430098595097661, + 0.009727993048727512, + 0.008915457874536514, + 0.005121180322021246, + 0.0063031637109816074, + 0.009033998474478722, + 0.007787473499774933, + 0.0044989376328885555, + 0.009443535469472408, + 0.01943216659128666, + 0.014839147217571735, + 0.008534413762390614, + 0.005144117865711451, + 0.015179328620433807, + 0.014257757924497128, + 0.017164945602416992, + 0.013152570463716984, + 0.014209614135324955, + 0.018647635355591774, + 0.014309336431324482, + 0.013296796008944511, + 0.013202337548136711, + 0.01587746851146221, + 0.009400127455592155, + 0.0070914956741034985, + 0.012363490648567677, + 0.01212785579264164, + 0.012109579518437386, + 0.009669958613812923, + 0.0067830574698746204, + 0.00847300048917532, + 0.0031362702138721943, + 0.003993649035692215, + 0.00402257451787591, + 0.003799279686063528, + 0.0021407052408903837, + 0.0037741733249276876, + 0.001964099705219269, + 0.006325067486613989, + 0.004462139680981636, + 0.0031955703161656857, + 0.0051013510674238205, + 0.004432888701558113, + 0.004039594437927008, + 0.004004627000540495, + 0.002967985114082694, + 0.004724711179733276, + 0.008939714170992374, + 0.009816962294280529, + 0.00907967984676361, + 0.006627482362091541, + 0.011005825363099575, + 0.009520436637103558, + 0.00789210107177496, + 0.008959741331636906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.3301297128200531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038174986839294434 + ], + "opposite_logprob_tokens": [ + -3.288175106048584 + ], + "logprob_reward": -0.038174986839294434, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038174986839294434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23997712135314941, + 0.15103907883167267, + 0.20100755989551544, + 0.06745772808790207, + 0.06736192852258682, + 0.03976986184716225, + 0.07002338021993637, + 0.04360703006386757, + 0.053353823721408844, + 0.06640247255563736 + ], + "credit_map": [ + 0.01689974032342434, + 0.023156819865107536, + 0.03437050059437752, + 0.02734402008354664, + 0.035017386078834534, + 0.020779002457857132, + 0.04551130533218384, + 0.03689834475517273, + 0.008605262264609337, + 0.012003603391349316, + 0.048934418708086014, + 0.014106049202382565, + 0.00850970484316349, + 0.0072323656640946865, + 0.014515506103634834, + 0.03713217377662659, + 0.01513548195362091, + 0.005805093329399824, + 0.06579993665218353, + 0.014438771642744541, + 0.012258618138730526, + 0.010189476422965527, + 0.016010314226150513, + 0.061369866132736206, + 0.008258973248302937, + 0.004259067587554455, + 0.012617289088666439, + 0.008316301740705967, + 0.004980253521353006, + 0.00937420129776001, + 0.008959952741861343, + 0.010691683739423752, + 0.004231337457895279, + 0.007916403003036976, + 0.008595962077379227, + 0.011090945452451706, + 0.011612210422754288, + 0.005262959748506546, + 0.010827981866896152, + 0.007824127562344074, + 0.004223864991217852, + 0.005400638561695814, + 0.005473456811159849, + 0.005199749022722244, + 0.004187304992228746, + 0.0050180223770439625, + 0.00435579102486372, + 0.005911031737923622, + 0.007095388602465391, + 0.005847536027431488, + 0.012625275179743767, + 0.009968946687877178, + 0.007604903541505337, + 0.011910667642951012, + 0.00572575768455863, + 0.009244902990758419, + 0.005170310381799936, + 0.0071982876397669315, + 0.005905782803893089, + 0.006977396551519632, + 0.003362464951351285, + 0.005480294581502676, + 0.0036133804824203253, + 0.005899115931242704, + 0.0043241605162620544, + 0.0048324065282940865, + 0.009343327954411507, + 0.010569052770733833, + 0.004744643811136484, + 0.006077838595956564, + 0.005060620605945587, + 0.008401773869991302, + 0.008679616264998913, + 0.01027410663664341, + 0.007549447473138571, + 0.006063682027161121, + 0.014010093174874783, + 0.005367042496800423, + 0.006547965109348297, + 0.007910517044365406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.038174986839294434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232976246625185 + ], + "opposite_logprob_tokens": [ + -3.7732975482940674 + ], + "logprob_reward": -0.0232976246625185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232976246625185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18650506436824799, + 0.1663852334022522, + 0.05660224333405495, + 0.2689393162727356, + 0.027469558641314507, + 0.05892463028430939, + 0.03741128742694855, + 0.038378484547138214, + 0.05023430287837982, + 0.10914988815784454 + ], + "credit_map": [ + 0.010480882599949837, + 0.02150905504822731, + 0.025678083300590515, + 0.020796600729227066, + 0.029516585171222687, + 0.015460950322449207, + 0.031686943024396896, + 0.03137596696615219, + 0.008409570902585983, + 0.011272438801825047, + 0.05777730420231819, + 0.014806736260652542, + 0.007893465459346771, + 0.007153393235057592, + 0.014993862248957157, + 0.04407846927642822, + 0.00672941142693162, + 0.0036200829781591892, + 0.008971297182142735, + 0.008227082900702953, + 0.007478645537048578, + 0.005165141075849533, + 0.009266889654099941, + 0.007143689319491386, + 0.03336235135793686, + 0.015670225024223328, + 0.048855021595954895, + 0.02673986926674843, + 0.02501525729894638, + 0.047536302357912064, + 0.03817731514573097, + 0.03358296677470207, + 0.0029040013905614614, + 0.0024011426139622927, + 0.0036711948923766613, + 0.004561529960483313, + 0.003688220167532563, + 0.0032252671662718058, + 0.003638314548879862, + 0.0033798874355852604, + 0.007694799918681383, + 0.00966006238013506, + 0.006736794952303171, + 0.008920816704630852, + 0.007041024975478649, + 0.006300157401710749, + 0.0055418056435883045, + 0.0070291669107973576, + 0.005259912461042404, + 0.0035575609654188156, + 0.006277704145759344, + 0.00440216576680541, + 0.004842123948037624, + 0.005667784716933966, + 0.003022783901542425, + 0.004381250124424696, + 0.002970605855807662, + 0.0030215252190828323, + 0.00764979375526309, + 0.005310860462486744, + 0.0021693757735192776, + 0.0037189358845353127, + 0.0036716293543577194, + 0.009865756146609783, + 0.004572507459670305, + 0.005440851673483849, + 0.007269489578902721, + 0.009393706917762756, + 0.005886150058358908, + 0.004986621905118227, + 0.00545916985720396, + 0.007225804030895233, + 0.012250362895429134, + 0.017870435491204262, + 0.012207526713609695, + 0.009745247662067413, + 0.024356437847018242, + 0.009362732991576195, + 0.010496747680008411, + 0.012860397808253765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0232976246625185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557437419891357 + ], + "opposite_logprob_tokens": [ + -1.6005743741989136 + ], + "logprob_reward": -0.22557437419891357, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557437419891357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44076794385910034, + 0.04648004472255707, + 0.11200603097677231, + 0.06267998367547989, + 0.16884569823741913, + 0.021834122017025948, + 0.02236427739262581, + 0.020652417093515396, + 0.05889839306473732, + 0.04547104239463806 + ], + "credit_map": [ + 0.02120049297809601, + 0.039982523769140244, + 0.05086173117160797, + 0.06420162320137024, + 0.048635706305503845, + 0.026271695271134377, + 0.12007419764995575, + 0.06953997910022736, + 0.0026344216894358397, + 0.0048300656490027905, + 0.012628168798983097, + 0.005636568181216717, + 0.0034522206988185644, + 0.002643187064677477, + 0.004576400388032198, + 0.01007900945842266, + 0.009356440044939518, + 0.0033289724960923195, + 0.037044357508420944, + 0.011476668529212475, + 0.004429738502949476, + 0.004974364768713713, + 0.008517548441886902, + 0.03287794813513756, + 0.006467678118497133, + 0.0041111004538834095, + 0.00887280609458685, + 0.010027419775724411, + 0.008330711163580418, + 0.00942578911781311, + 0.00808601826429367, + 0.007358464878052473, + 0.009096580557525158, + 0.024799935519695282, + 0.027784958481788635, + 0.027394752949476242, + 0.027583302929997444, + 0.010009905323386192, + 0.022795479744672775, + 0.01938079297542572, + 0.0018318917136639357, + 0.004152285400778055, + 0.0028249132446944714, + 0.0024417899549007416, + 0.0029269689694046974, + 0.0028275654185563326, + 0.0020651028025895357, + 0.00276360590942204, + 0.0032738111913204193, + 0.0022021832410246134, + 0.003323927754536271, + 0.00279223988763988, + 0.0029582621064037085, + 0.0033913799561560154, + 0.0014672820689156651, + 0.002955191070213914, + 0.0017014776822179556, + 0.0029044023249298334, + 0.0034430266823619604, + 0.0022233850322663784, + 0.002704818034544587, + 0.0018326363060623407, + 0.002414645394310355, + 0.0034280254039913416, + 0.006323360372334719, + 0.00484229251742363, + 0.008614412508904934, + 0.0091320825740695, + 0.006846938282251358, + 0.009453818202018738, + 0.005316810216754675, + 0.008368675597012043, + 0.004244574345648289, + 0.006416352931410074, + 0.00596995372325182, + 0.004426018334925175, + 0.006980072241276503, + 0.006006914656609297, + 0.004451872315257788, + 0.0069752829149365425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.22557437419891357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle looking up at the gymnast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly track the male gymnast as he executes a powerful lunge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gymnast wearing a tight, sleeveless royal blue and white leotard with crisp white trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gymnast perform a full rotation in the air with legs firmly together and toes pointed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a vast arena and audience while the gymnast is isolated against a deep blue landing mat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737728603184223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737728603184223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14765408635139465, + 0.060408495366573334, + 0.10432365536689758, + 0.06213926896452904, + 0.044382866472005844, + 0.06278485804796219, + 0.03735903650522232, + 0.1525566130876541, + 0.11281893402338028, + 0.2155722677707672 + ], + "credit_map": [ + 0.007688685320317745, + 0.01716773584485054, + 0.019547225907444954, + 0.016815772280097008, + 0.021924277767539024, + 0.012043887749314308, + 0.032734498381614685, + 0.019732002168893814, + 0.004948332905769348, + 0.005737306084483862, + 0.013533394783735275, + 0.007562700193375349, + 0.005790232215076685, + 0.004553760401904583, + 0.005508920177817345, + 0.0127738481387496, + 0.011137370020151138, + 0.0031108229886740446, + 0.03146801143884659, + 0.00631082383915782, + 0.00608287425711751, + 0.009030140936374664, + 0.00696145836263895, + 0.030222157016396523, + 0.007392243482172489, + 0.0036131541710346937, + 0.012787123210728168, + 0.008050884120166302, + 0.004966569133102894, + 0.009180731140077114, + 0.00793484691530466, + 0.008213719353079796, + 0.002768364967778325, + 0.004503593780100346, + 0.005458908621221781, + 0.007177148945629597, + 0.007198024075478315, + 0.003930617123842239, + 0.008094978518784046, + 0.005251231137663126, + 0.005920123308897018, + 0.010226169601082802, + 0.006785515230149031, + 0.004360266495496035, + 0.009216072969138622, + 0.012276220135390759, + 0.004885583650320768, + 0.009114908054471016, + 0.0047812205739319324, + 0.00405515544116497, + 0.005088448990136385, + 0.0033280218485742807, + 0.004900122527033091, + 0.007496015168726444, + 0.0028226200956851244, + 0.0048874313943088055, + 0.013525150716304779, + 0.01113390177488327, + 0.03400993347167969, + 0.012896076776087284, + 0.007841566577553749, + 0.01537979394197464, + 0.011619457975029945, + 0.046150725334882736, + 0.012007235549390316, + 0.011324552819132805, + 0.01575734280049801, + 0.018862349912524223, + 0.01357933972030878, + 0.013239037245512009, + 0.010030874982476234, + 0.018018199130892754, + 0.024962814524769783, + 0.029225103557109833, + 0.030148280784487724, + 0.02156663127243519, + 0.03144975006580353, + 0.025443773716688156, + 0.022612640634179115, + 0.03016328066587448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.006737728603184223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_tensor([2695, 2695])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_tensor([2695, 2695])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..087f6d0444b1c8566c0799164492a8fe8e582a08 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_tensor([2695, 2695])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, shot from a slightly upward angle and panning slowly to the right, captures a solitary dancer, lost in the ecstasy of a twirl in the center of a grand yet empty Baroque ballroom. Her long, dark brown curls are elegantly tied into a loose bun, a few strands of hair freed from their restraints as she spins, gently brushing against her serene face, eyes closed and a hint of a smile playing on her lips. She wears a deep, royal blue, sleeveless silk gown, the tight bodice emphasizing her slender waist. The massive, multi-tiered skirt, lifted high by centrifugal force as she steadily and effortlessly spun, pivoting on the toe of her left foot, unfolds like a fully blossoming blue flower, almost parallel to the ground, its pleats drawing graceful, undulating curves in the air. A massive crystal chandelier atop the soaring dome casts a warm, radiant golden light, illuminating the mirror-like polished marble floor, which reflects the dancers' twirling figures and the blurred motion of their skirts, as well as the intricate gold reliefs on the walls. Light filters through the swirling skirts, casting ever-shifting, interwoven patterns of light and shadow on the floor. The entire scene, bathed in warm hues, creates an atmosphere of profound tranquility, dreamy elegance, and a touch of solitary elegance.", + "seed": 29085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_2695_seed_29085_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.034479815792292354, + "vlm_logprob": -0.034479815792292354, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.034479815792292354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006964574102312326 + ], + "opposite_logprob_tokens": [ + -5.256964683532715 + ], + "logprob_reward": -0.006964574102312326, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006964574102312326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18231621384620667, + 0.15696656703948975, + 0.058486342430114746, + 0.018780410289764404, + 0.06995252519845963, + 0.0356208011507988, + 0.092221699655056, + 0.32356196641921997, + 0.023343056440353394, + 0.03875043988227844 + ], + "credit_map": [ + 0.021762050688266754, + 0.021943863481283188, + 0.019912535324692726, + 0.025304727256298065, + 0.03417688608169556, + 0.018310649320483208, + 0.017269140109419823, + 0.023636344820261, + 0.03697708994150162, + 0.01725330390036106, + 0.011260727420449257, + 0.011129901744425297, + 0.009572339244186878, + 0.042095690965652466, + 0.0165882371366024, + 0.012089257128536701, + 0.006629159674048424, + 0.006966718938201666, + 0.011512289755046368, + 0.00934712402522564, + 0.004801909439265728, + 0.01103490311652422, + 0.004269617609679699, + 0.003924617078155279, + 0.0028571179136633873, + 0.0034921381156891584, + 0.0025625235866755247, + 0.0026180590502917767, + 0.0024956357665359974, + 0.001525603118352592, + 0.001731649972498417, + 0.0014976818347349763, + 0.00956716388463974, + 0.014405196532607079, + 0.006549433339387178, + 0.0063900877721607685, + 0.008643019944429398, + 0.008414563722908497, + 0.006135813891887665, + 0.009847241453826427, + 0.005328380968421698, + 0.0063823070377111435, + 0.004290401935577393, + 0.002741133561357856, + 0.003998592495918274, + 0.004397067707031965, + 0.002923582913354039, + 0.005559330340474844, + 0.01827784813940525, + 0.018758531659841537, + 0.009852888062596321, + 0.006722026504576206, + 0.007553292904049158, + 0.015252957120537758, + 0.008129452355206013, + 0.007674707565456629, + 0.0666981041431427, + 0.08424083143472672, + 0.03799736127257347, + 0.023858435451984406, + 0.02026359550654888, + 0.0500338189303875, + 0.015259979292750359, + 0.025209825485944748, + 0.0033061548601835966, + 0.003954487387090921, + 0.0026460180524736643, + 0.002704332349821925, + 0.0026909855660051107, + 0.0031345561146736145, + 0.002826013369485736, + 0.002080506645143032, + 0.0033519440330564976, + 0.006597225088626146, + 0.0039007309824228287, + 0.0036206424701958895, + 0.008165178820490837, + 0.005704216193407774, + 0.004832979757338762, + 0.002577515784651041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.006964574102312326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636165712028742 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636165712028742, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636165712028742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1315351128578186, + 0.12673836946487427, + 0.09223674237728119, + 0.09881121665239334, + 0.13776618242263794, + 0.10635657608509064, + 0.05787579342722893, + 0.07781775295734406, + 0.04039209708571434, + 0.1304701417684555 + ], + "credit_map": [ + 0.024456705898046494, + 0.01868424378335476, + 0.011478697881102562, + 0.012966515496373177, + 0.01952986605465412, + 0.020916074514389038, + 0.011624096892774105, + 0.01187891885638237, + 0.011233154684305191, + 0.011327487416565418, + 0.012370032258331776, + 0.01693994738161564, + 0.015514658764004707, + 0.015583142638206482, + 0.02858233079314232, + 0.015187613666057587, + 0.009564823471009731, + 0.012225599959492683, + 0.00822308100759983, + 0.015964137390255928, + 0.01595611311495304, + 0.009864745661616325, + 0.010972595773637295, + 0.009465646930038929, + 0.014968043193221092, + 0.018349958583712578, + 0.010419469326734543, + 0.011795945465564728, + 0.011602395214140415, + 0.009359764866530895, + 0.010398403741419315, + 0.01191724557429552, + 0.017457084730267525, + 0.019466618075966835, + 0.02680015005171299, + 0.01651415228843689, + 0.014812626875936985, + 0.014239785261452198, + 0.014099893160164356, + 0.01437588781118393, + 0.020711973309516907, + 0.016724010929465294, + 0.011452960781753063, + 0.010232113301753998, + 0.012034857645630836, + 0.01206185482442379, + 0.009591645561158657, + 0.013547160662710667, + 0.008228067308664322, + 0.009347335435450077, + 0.008218243718147278, + 0.005708970595151186, + 0.006669315043836832, + 0.00617265235632658, + 0.007825849577784538, + 0.005705360323190689, + 0.011166546493768692, + 0.01476964633911848, + 0.01096353679895401, + 0.0065084705129265785, + 0.009219431318342686, + 0.011886351741850376, + 0.006784919649362564, + 0.006518852896988392, + 0.007987621240317822, + 0.007293555419892073, + 0.004217176232486963, + 0.004797968082129955, + 0.004839922767132521, + 0.005227550398558378, + 0.003264488186687231, + 0.0027638175524771214, + 0.02387065626680851, + 0.02887498401105404, + 0.012910325080156326, + 0.010757840238511562, + 0.017421355471014977, + 0.01373963337391615, + 0.012936183251440525, + 0.009959163144230843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.007636165712028742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735597271472216 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735597271472216, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735597271472216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13007616996765137, + 0.20702247321605682, + 0.03800113871693611, + 0.08112066239118576, + 0.08205118775367737, + 0.13950581848621368, + 0.09701221436262131, + 0.06224193051457405, + 0.03567294403910637, + 0.12729546427726746 + ], + "credit_map": [ + 0.023608068004250526, + 0.017468081787228584, + 0.01849348284304142, + 0.015683334320783615, + 0.01879746839404106, + 0.013293772004544735, + 0.009802283719182014, + 0.012929676100611687, + 0.05258236080408096, + 0.017824534326791763, + 0.018741238862276077, + 0.018456902354955673, + 0.015125352889299393, + 0.060286253690719604, + 0.010755259543657303, + 0.013250580057501793, + 0.004702150821685791, + 0.006307908799499273, + 0.005120518151670694, + 0.004650596063584089, + 0.0051740133203566074, + 0.004268286284059286, + 0.00372916366904974, + 0.004048500210046768, + 0.011807630769908428, + 0.01781771145761013, + 0.009978462010622025, + 0.009789949283003807, + 0.010145510546863079, + 0.007881443947553635, + 0.007593385875225067, + 0.006106569431722164, + 0.009985696524381638, + 0.014441853389143944, + 0.007644173223525286, + 0.008217249996960163, + 0.011153433471918106, + 0.010827183723449707, + 0.01039588451385498, + 0.009385707788169384, + 0.027487199753522873, + 0.022338874638080597, + 0.013148536905646324, + 0.011493293568491936, + 0.01299103070050478, + 0.014006881974637508, + 0.010557075031101704, + 0.027482930570840836, + 0.011628213338553905, + 0.012616686522960663, + 0.011100572533905506, + 0.011967519298195839, + 0.01653273217380047, + 0.009618749842047691, + 0.014195462688803673, + 0.009352277033030987, + 0.008882293477654457, + 0.008037087507545948, + 0.0069468277506530285, + 0.007547326385974884, + 0.007484695874154568, + 0.007085421122610569, + 0.007559080142527819, + 0.008699200116097927, + 0.006385136861354113, + 0.0067233191803097725, + 0.003996009472757578, + 0.00424245884642005, + 0.004175127483904362, + 0.004472666885703802, + 0.0028979587368667126, + 0.0027802647091448307, + 0.024865703657269478, + 0.032038554549217224, + 0.01158063393086195, + 0.011524248868227005, + 0.009972265921533108, + 0.015120524913072586, + 0.01269481610506773, + 0.009498720988631248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.006735597271472216 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205534189939499 + ], + "opposite_logprob_tokens": [ + -2.8120553493499756 + ], + "logprob_reward": -0.06205534189939499, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205534189939499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22431178390979767, + 0.1090320348739624, + 0.0568319633603096, + 0.06783337891101837, + 0.07643114775419235, + 0.07142417877912521, + 0.09375999867916107, + 0.061450593173503876, + 0.07363761216402054, + 0.1652873307466507 + ], + "credit_map": [ + 0.038202088326215744, + 0.023094981908798218, + 0.021709099411964417, + 0.01765843667089939, + 0.050060562789440155, + 0.02461719699203968, + 0.02553025260567665, + 0.023439140990376472, + 0.02214271016418934, + 0.011142787523567677, + 0.008074830286204815, + 0.013254450634121895, + 0.009547139517962933, + 0.024685177952051163, + 0.009294276125729084, + 0.01089065708220005, + 0.007656360976397991, + 0.011733009479939938, + 0.0066327257081866264, + 0.007628818042576313, + 0.006807852070778608, + 0.007715034764260054, + 0.004263598471879959, + 0.004394558724015951, + 0.008374902419745922, + 0.010456698015332222, + 0.011952054686844349, + 0.008404617197811604, + 0.009944615885615349, + 0.005634963046759367, + 0.00685159582644701, + 0.006213925778865814, + 0.012937819585204124, + 0.01123037002980709, + 0.007601364981383085, + 0.008043019101023674, + 0.009612941183149815, + 0.007006446365267038, + 0.009558847174048424, + 0.010440335609018803, + 0.015785647556185722, + 0.010427580215036869, + 0.009277763776481152, + 0.00542563758790493, + 0.006612259428948164, + 0.007125832140445709, + 0.005496959667652845, + 0.011272496543824673, + 0.012998366728425026, + 0.017181748524308205, + 0.009241546504199505, + 0.01192873902618885, + 0.011791856028139591, + 0.009289385750889778, + 0.012779762037098408, + 0.008548591285943985, + 0.00980888120830059, + 0.010252249427139759, + 0.007456645835191011, + 0.006706714164465666, + 0.007802355568856001, + 0.0082887914031744, + 0.006731875240802765, + 0.004403077065944672, + 0.013654901646077633, + 0.012968073599040508, + 0.007355256471782923, + 0.007489056792110205, + 0.009821295738220215, + 0.00919856783002615, + 0.007709662429988384, + 0.005440786015242338, + 0.019823385402560234, + 0.019431836903095245, + 0.019020039588212967, + 0.02067880891263485, + 0.02387906238436699, + 0.01754075102508068, + 0.027529770508408546, + 0.017383670434355736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.06205534189939499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900739997625351 + ], + "opposite_logprob_tokens": [ + -2.4640073776245117 + ], + "logprob_reward": -0.08900739997625351, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900739997625351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18597449362277985, + 0.21832510828971863, + 0.06285691261291504, + 0.0688055008649826, + 0.04295293986797333, + 0.08078642189502716, + 0.04043048620223999, + 0.08633919060230255, + 0.08397059887647629, + 0.12955836951732635 + ], + "credit_map": [ + 0.03688553348183632, + 0.02429376170039177, + 0.022646885365247726, + 0.020140819251537323, + 0.023720156401395798, + 0.027583645656704903, + 0.015609432011842728, + 0.015094266273081303, + 0.054421316832304, + 0.022627225145697594, + 0.015410025604069233, + 0.02294032648205757, + 0.014299072325229645, + 0.06080808863043785, + 0.011340701952576637, + 0.016478346660733223, + 0.008199715986847878, + 0.009601451456546783, + 0.006944073364138603, + 0.010657901875674725, + 0.00757569819688797, + 0.008553599938750267, + 0.005290338769555092, + 0.00603412976488471, + 0.008734067901968956, + 0.008621186949312687, + 0.010740342549979687, + 0.00788906216621399, + 0.009740411303937435, + 0.00504931528121233, + 0.006503428798168898, + 0.011527687311172485, + 0.0060465033166110516, + 0.006705342326313257, + 0.0040456079877913, + 0.003603025572374463, + 0.0027965856716036797, + 0.005954009015113115, + 0.004297593608498573, + 0.009504268877208233, + 0.01103529054671526, + 0.009277037344872952, + 0.01242435909807682, + 0.008713630959391594, + 0.011891256086528301, + 0.008376352488994598, + 0.006454606540501118, + 0.012613880448043346, + 0.006927466951310635, + 0.005994822829961777, + 0.005098729394376278, + 0.0035087796859443188, + 0.007256030570715666, + 0.004515041131526232, + 0.0036768505815416574, + 0.0034527694806456566, + 0.016530951485037804, + 0.01792626827955246, + 0.009253081865608692, + 0.007739679887890816, + 0.009459836408495903, + 0.010755598545074463, + 0.00782011728733778, + 0.006853654980659485, + 0.016228364780545235, + 0.017816271632909775, + 0.011816113255918026, + 0.007786889094859362, + 0.006067538168281317, + 0.012722565792500973, + 0.005343923810869455, + 0.006188924890011549, + 0.01903766207396984, + 0.01536630094051361, + 0.01770983636379242, + 0.012774269096553326, + 0.02235010638833046, + 0.013719616457819939, + 0.01812577247619629, + 0.01047480572015047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.08900739997625351 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a slightly upward angle and panning slowly to the right, captures a solitary dancer, lost in the ecstasy of a twirl in the center of a grand yet empty Baroque ballroom. Her long, dark brown curls are elegantly tied into a loose bun, a few strands of hair freed from their restraints as she spins, gently brushing against her serene face, eyes closed and a hint of a smile playing on her lips. She wears a deep, royal blue, sleeveless silk gown, the tight bodice emphasizing her slender waist. The massive, multi-tiered skirt, lifted high by centrifugal force as she steadily and effortlessly spun, pivoting on the toe of her left foot, unfolds like a fully blossoming blue flower, almost parallel to the ground, its pleats drawing graceful, undulating curves in the air. A massive crystal chandelier atop the soaring dome casts a warm, radiant golden light, illuminating the mirror-like polished marble floor, which reflects the dancers' twirling figures and the blurred motion of their skirts, as well as the intricate gold reliefs on the walls. Light filters through the swirling skirts, casting ever-shifting, interwoven patterns of light and shadow on the floor. The entire scene, bathed in warm hues, creates an atmosphere of profound tranquility, dreamy elegance, and a touch of solitary elegance.", + "seed": 29086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_2695_seed_29086_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03518621418625116, + "vlm_logprob": -0.03518621418625116, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03518621418625116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004155097529292107 + ], + "opposite_logprob_tokens": [ + -5.879155158996582 + ], + "logprob_reward": -0.004155097529292107, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004155097529292107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13492439687252045, + 0.08101075142621994, + 0.07792255282402039, + 0.06031493470072746, + 0.08615615218877792, + 0.13422583043575287, + 0.15934982895851135, + 0.07197093218564987, + 0.14914338290691376, + 0.0449812188744545 + ], + "credit_map": [ + 0.014726720750331879, + 0.017633873969316483, + 0.02586081437766552, + 0.01049555093050003, + 0.013483935035765171, + 0.02864345908164978, + 0.008399363607168198, + 0.015680674463510513, + 0.009358235634863377, + 0.010341654531657696, + 0.008278464898467064, + 0.011763780377805233, + 0.014106935821473598, + 0.009364794939756393, + 0.009187748655676842, + 0.008609134703874588, + 0.016349568963050842, + 0.010030229575932026, + 0.01082320511341095, + 0.005782437976449728, + 0.008476502262055874, + 0.011282422579824924, + 0.008580244146287441, + 0.0065979440696537495, + 0.011603491380810738, + 0.009790786541998386, + 0.008681776002049446, + 0.006169530097395182, + 0.007201282773166895, + 0.00671049952507019, + 0.005391258746385574, + 0.004766309168189764, + 0.013477914035320282, + 0.006834147032350302, + 0.012368828058242798, + 0.005499690305441618, + 0.007714062929153442, + 0.02916608937084675, + 0.004114174749702215, + 0.00698124710470438, + 0.019938893616199493, + 0.011775266379117966, + 0.017426209524273872, + 0.01838604174554348, + 0.016385238617658615, + 0.0098760686814785, + 0.025160126388072968, + 0.015277979895472527, + 0.024902360513806343, + 0.023280009627342224, + 0.02234126627445221, + 0.01486071664839983, + 0.010375789366662502, + 0.018416400998830795, + 0.010813827626407146, + 0.034359466284513474, + 0.009697472676634789, + 0.008120749145746231, + 0.010013199411332607, + 0.011375131085515022, + 0.010288351215422153, + 0.005353996530175209, + 0.010376670397818089, + 0.006745362654328346, + 0.02811930887401104, + 0.01751631870865822, + 0.027184369042515755, + 0.015113078989088535, + 0.013278783299028873, + 0.0170885119587183, + 0.013609019108116627, + 0.017233995720744133, + 0.006707639433443546, + 0.006101810373365879, + 0.006723443511873484, + 0.003398555563762784, + 0.005583321675658226, + 0.007439762819558382, + 0.003661250462755561, + 0.00536543782800436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.004155097529292107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645604364573956 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645604364573956, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645604364573956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2869393825531006, + 0.1105039194226265, + 0.05807845667004585, + 0.04178997129201889, + 0.09086007624864578, + 0.0806611105799675, + 0.19570717215538025, + 0.06296062469482422, + 0.045071277767419815, + 0.027427993714809418 + ], + "credit_map": [ + 0.025121591985225677, + 0.01856895722448826, + 0.089835524559021, + 0.015707483515143394, + 0.01368054747581482, + 0.09377939254045486, + 0.014319535344839096, + 0.015926342457532883, + 0.014041129499673843, + 0.013081761077046394, + 0.014348846860229969, + 0.015058438293635845, + 0.016040470451116562, + 0.017084574326872826, + 0.012403829023241997, + 0.008444870822131634, + 0.009807359427213669, + 0.009288541972637177, + 0.005989325698465109, + 0.005828553345054388, + 0.008103062398731709, + 0.007977794855833054, + 0.0066538783721625805, + 0.004429941065609455, + 0.00678988778963685, + 0.004892630502581596, + 0.005477086175233126, + 0.005672124680131674, + 0.0048089115880429745, + 0.004409246612340212, + 0.005702011287212372, + 0.004038074053823948, + 0.02346954494714737, + 0.008069204166531563, + 0.009923445992171764, + 0.005574481096118689, + 0.00797309074550867, + 0.025857506319880486, + 0.0038119920063763857, + 0.006180811673402786, + 0.01334543526172638, + 0.006870969198644161, + 0.01129046082496643, + 0.010858899913728237, + 0.008381509222090244, + 0.008993389084935188, + 0.013264927081763744, + 0.007655516266822815, + 0.026691485196352005, + 0.018164878711104393, + 0.02839515544474125, + 0.028544599190354347, + 0.018311550840735435, + 0.018773114308714867, + 0.026156773790717125, + 0.030669627711176872, + 0.009140890091657639, + 0.007340149488300085, + 0.010328313335776329, + 0.007944997400045395, + 0.007161058951169252, + 0.006417488679289818, + 0.0073349494487047195, + 0.007292776834219694, + 0.008260656148195267, + 0.006940476596355438, + 0.007652406580746174, + 0.0042420350946486, + 0.005424275994300842, + 0.0032643054146319628, + 0.00524717103689909, + 0.004039952997118235, + 0.004724529106169939, + 0.0037406927440315485, + 0.003278092248365283, + 0.002599804662168026, + 0.0034671956673264503, + 0.003106477903202176, + 0.0032230308279395103, + 0.003288168925791979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.008645604364573956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096094518899918 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096094518899918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096094518899918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30513811111450195, + 0.044053852558135986, + 0.07567889988422394, + 0.04025208204984665, + 0.19159041345119476, + 0.06432900577783585, + 0.11394153535366058, + 0.06070694699883461, + 0.08259790390729904, + 0.02171129360795021 + ], + "credit_map": [ + 0.02125929296016693, + 0.021602975204586983, + 0.0994870737195015, + 0.010943771339952946, + 0.01564008742570877, + 0.11504137516021729, + 0.007280605845153332, + 0.01388292945921421, + 0.005217745900154114, + 0.006753832567483187, + 0.005258663557469845, + 0.006021406501531601, + 0.005727240815758705, + 0.005352814681828022, + 0.005576658062636852, + 0.004145491402596235, + 0.014802811667323112, + 0.00949047040194273, + 0.008409393019974232, + 0.008953744545578957, + 0.008036350831389427, + 0.009408953599631786, + 0.00988807063549757, + 0.006689111702144146, + 0.005849280394613743, + 0.0058485716581344604, + 0.003920731600373983, + 0.005641734693199396, + 0.004330470692366362, + 0.003999096807092428, + 0.005880807060748339, + 0.004781386349350214, + 0.05596093833446503, + 0.02109476551413536, + 0.025459973141551018, + 0.010175425559282303, + 0.012061774730682373, + 0.044394608587026596, + 0.008844230324029922, + 0.013598677702248096, + 0.009804253466427326, + 0.005780658684670925, + 0.007293943781405687, + 0.009400063194334507, + 0.008056960068643093, + 0.00792929157614708, + 0.010041790083050728, + 0.006022050976753235, + 0.014784569852054119, + 0.009439148008823395, + 0.016315234825015068, + 0.017037615180015564, + 0.011052513495087624, + 0.009954795241355896, + 0.015450319275259972, + 0.019907338544726372, + 0.010524001903831959, + 0.005644542630761862, + 0.011859509162604809, + 0.006355386693030596, + 0.005274650640785694, + 0.007618006318807602, + 0.005253517534583807, + 0.008177331648766994, + 0.009047429077327251, + 0.00994717888534069, + 0.008850506506860256, + 0.010384578257799149, + 0.007906753569841385, + 0.01143457181751728, + 0.01192499604076147, + 0.013101890683174133, + 0.002729743020609021, + 0.0027206214144825935, + 0.0022612009197473526, + 0.003554572584107518, + 0.002795619424432516, + 0.0025599936489015818, + 0.002615304896607995, + 0.0024742353707551956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.004096094518899918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002948969602585 + ], + "opposite_logprob_tokens": [ + -2.6950294971466064 + ], + "logprob_reward": -0.07002948969602585, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002948969602585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25300437211990356, + 0.07125046104192734, + 0.04759359359741211, + 0.09067213535308838, + 0.10707662254571915, + 0.1906309723854065, + 0.07324538379907608, + 0.051591020077466965, + 0.08693922311067581, + 0.027996202930808067 + ], + "credit_map": [ + 0.01719125173985958, + 0.018738694489002228, + 0.07942423224449158, + 0.00872671976685524, + 0.013925984501838684, + 0.09393493831157684, + 0.007336478214710951, + 0.013726086355745792, + 0.011597747914493084, + 0.00794819463044405, + 0.012446925975382328, + 0.00861434731632471, + 0.008044370450079441, + 0.010572204366326332, + 0.0065352958627045155, + 0.005491374526172876, + 0.009598463773727417, + 0.007477411534637213, + 0.0056955525651574135, + 0.0032775639556348324, + 0.005207863636314869, + 0.006313676480203867, + 0.004686621483415365, + 0.005336441099643707, + 0.018500566482543945, + 0.011439730413258076, + 0.010515467263758183, + 0.01115338783711195, + 0.01052960753440857, + 0.00937536358833313, + 0.010828133672475815, + 0.00832988042384386, + 0.023795871064066887, + 0.008098001591861248, + 0.009565313346683979, + 0.009015384130179882, + 0.009512058459222317, + 0.03166999667882919, + 0.007923533208668232, + 0.007496467791497707, + 0.03288383409380913, + 0.020108239725232124, + 0.02526603825390339, + 0.025935400277376175, + 0.018026018515229225, + 0.015291356481611729, + 0.033900681883096695, + 0.01921939104795456, + 0.010556506924331188, + 0.008298481814563274, + 0.014171022921800613, + 0.008007657714188099, + 0.0053434669971466064, + 0.007494637742638588, + 0.007683437783271074, + 0.01169017143547535, + 0.00840453989803791, + 0.005436758976429701, + 0.007347176317125559, + 0.005945321638137102, + 0.0051469034515321255, + 0.00475788488984108, + 0.004884009715169668, + 0.009668421000242233, + 0.010196826420724392, + 0.009451672434806824, + 0.009919004514813423, + 0.008597683161497116, + 0.007413411978632212, + 0.012564367614686489, + 0.015677548944950104, + 0.013118707574903965, + 0.00411082012578845, + 0.003666896838694811, + 0.004724650178104639, + 0.0027439785189926624, + 0.0035499127116054296, + 0.003456509904935956, + 0.002791836392134428, + 0.0029515980277210474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.07002948969602585 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900478482246399 + ], + "opposite_logprob_tokens": [ + -2.4640047550201416 + ], + "logprob_reward": -0.08900478482246399, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900478482246399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2675253450870514, + 0.06272391229867935, + 0.051309190690517426, + 0.06871431320905685, + 0.07088644057512283, + 0.1749611347913742, + 0.07541844993829727, + 0.09646499902009964, + 0.10910762846469879, + 0.022888628765940666 + ], + "credit_map": [ + 0.023594794794917107, + 0.019276736304163933, + 0.08957675844430923, + 0.008742030709981918, + 0.013378949835896492, + 0.09315329790115356, + 0.0065602874383330345, + 0.013242472894489765, + 0.013101317919790745, + 0.010334868915379047, + 0.006324816029518843, + 0.0069891358725726604, + 0.0043507167138159275, + 0.010449630208313465, + 0.0036323415115475655, + 0.007541083265095949, + 0.010054812766611576, + 0.0068526240065693855, + 0.006263568066060543, + 0.005086833145469427, + 0.005050936713814735, + 0.0062564280815422535, + 0.005909701809287071, + 0.00583428330719471, + 0.011198767460882664, + 0.01010225061327219, + 0.009432241320610046, + 0.006650542374700308, + 0.010107018984854221, + 0.008348959498107433, + 0.006168598309159279, + 0.006705926265567541, + 0.012333405204117298, + 0.0074718003161251545, + 0.008929057978093624, + 0.006252393126487732, + 0.010376191698014736, + 0.013471740297973156, + 0.005507726687937975, + 0.006544115487486124, + 0.032134123146533966, + 0.017037516459822655, + 0.02349097840487957, + 0.023662878200411797, + 0.019695941358804703, + 0.015310974791646004, + 0.030253244563937187, + 0.013375455513596535, + 0.011998044326901436, + 0.012345902621746063, + 0.016829654574394226, + 0.007091067265719175, + 0.004441589117050171, + 0.008293421939015388, + 0.005743729416280985, + 0.008675030432641506, + 0.016231929883360863, + 0.00903225876390934, + 0.01696711778640747, + 0.012943236157298088, + 0.009162399917840958, + 0.009754640981554985, + 0.011828732676804066, + 0.01054468285292387, + 0.017459120601415634, + 0.013018480502068996, + 0.011796453967690468, + 0.012931399047374725, + 0.010698267258703709, + 0.01250289287418127, + 0.014884352684020996, + 0.01581665687263012, + 0.0033592400141060352, + 0.0034988143015652895, + 0.003075558692216873, + 0.00232460699044168, + 0.003006408456712961, + 0.0028491823468357325, + 0.001858808333054185, + 0.0029160098638385534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.08900478482246399 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a slightly upward angle and panning slowly to the right, captures a solitary dancer, lost in the ecstasy of a twirl in the center of a grand yet empty Baroque ballroom. Her long, dark brown curls are elegantly tied into a loose bun, a few strands of hair freed from their restraints as she spins, gently brushing against her serene face, eyes closed and a hint of a smile playing on her lips. She wears a deep, royal blue, sleeveless silk gown, the tight bodice emphasizing her slender waist. The massive, multi-tiered skirt, lifted high by centrifugal force as she steadily and effortlessly spun, pivoting on the toe of her left foot, unfolds like a fully blossoming blue flower, almost parallel to the ground, its pleats drawing graceful, undulating curves in the air. A massive crystal chandelier atop the soaring dome casts a warm, radiant golden light, illuminating the mirror-like polished marble floor, which reflects the dancers' twirling figures and the blurred motion of their skirts, as well as the intricate gold reliefs on the walls. Light filters through the swirling skirts, casting ever-shifting, interwoven patterns of light and shadow on the floor. The entire scene, bathed in warm hues, creates an atmosphere of profound tranquility, dreamy elegance, and a touch of solitary elegance.", + "seed": 29087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_2695_seed_29087_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031974362954497335, + "vlm_logprob": -0.031974362954497335, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.031974362954497335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004135153256356716 + ], + "opposite_logprob_tokens": [ + -6.1291351318359375 + ], + "logprob_reward": -0.004135153256356716, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004135153256356716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2306092381477356, + 0.05886188894510269, + 0.029903700575232506, + 0.11392409354448318, + 0.06718113273382187, + 0.08874697983264923, + 0.11752691119909286, + 0.0706057921051979, + 0.13518759608268738, + 0.08745260536670685 + ], + "credit_map": [ + 0.04754041135311127, + 0.027708416804671288, + 0.027210239320993423, + 0.022307349368929863, + 0.02651861682534218, + 0.02833852730691433, + 0.025478096678853035, + 0.025507597252726555, + 0.011686187237501144, + 0.006560604553669691, + 0.005668071564286947, + 0.0039062276482582092, + 0.009437531232833862, + 0.011775537393987179, + 0.004298449028283358, + 0.005529280751943588, + 0.0047286017797887325, + 0.0038026210386306047, + 0.003207378787919879, + 0.0030800073873251677, + 0.004375170450657606, + 0.005639044567942619, + 0.00269644265063107, + 0.0023744336795061827, + 0.014586465433239937, + 0.01303641777485609, + 0.009906645864248276, + 0.008745124563574791, + 0.020530227571725845, + 0.012583848088979721, + 0.01248369924724102, + 0.02205166406929493, + 0.011786749586462975, + 0.007573888637125492, + 0.0056572663597762585, + 0.005722689907997847, + 0.007423946633934975, + 0.007443277630954981, + 0.007982232607901096, + 0.013591080904006958, + 0.011725804768502712, + 0.009289778769016266, + 0.013054472394287586, + 0.010347986593842506, + 0.011621108278632164, + 0.009672760032117367, + 0.012114938348531723, + 0.010920130647718906, + 0.017405502498149872, + 0.016729822382330894, + 0.012884090654551983, + 0.010675160214304924, + 0.010631900280714035, + 0.02633858658373356, + 0.009388890117406845, + 0.01347296591848135, + 0.01276931818574667, + 0.010863919742405415, + 0.012611107900738716, + 0.006431881804019213, + 0.007384800352156162, + 0.010176163166761398, + 0.00543171213939786, + 0.004936893004924059, + 0.014106083661317825, + 0.01728425920009613, + 0.013619823381304741, + 0.01209248322993517, + 0.023939255625009537, + 0.01255557406693697, + 0.01719810627400875, + 0.02439201809465885, + 0.02320818230509758, + 0.01477004773914814, + 0.00915167760103941, + 0.00969737395644188, + 0.009524911642074585, + 0.01035647839307785, + 0.005787970963865519, + 0.0049559506587684155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.004135153256356716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21022216975688934, + 0.10716909915208817, + 0.04751840978860855, + 0.059807609766721725, + 0.03256605565547943, + 0.04787541180849075, + 0.13188482820987701, + 0.05649244785308838, + 0.26418638229370117, + 0.04227765277028084 + ], + "credit_map": [ + 0.02827131561934948, + 0.02516479790210724, + 0.021367976441979408, + 0.0208933986723423, + 0.03445304557681084, + 0.030158312991261482, + 0.024259645491838455, + 0.02565365470945835, + 0.012628097087144852, + 0.011260087601840496, + 0.008047417737543583, + 0.013679313473403454, + 0.016730081290006638, + 0.008797290734946728, + 0.018381444737315178, + 0.017645353451371193, + 0.005199596285820007, + 0.005599054973572493, + 0.0035266370978206396, + 0.00636327126994729, + 0.010640552267432213, + 0.004660178441554308, + 0.006056339479982853, + 0.005472773220390081, + 0.0074378917925059795, + 0.007314289454370737, + 0.004724000580608845, + 0.004079881589859724, + 0.013177726417779922, + 0.007769616320729256, + 0.004708003252744675, + 0.010596195235848427, + 0.00463115656748414, + 0.003409049706533551, + 0.002528879791498184, + 0.0036419108510017395, + 0.002609697636216879, + 0.004234997555613518, + 0.005097600165754557, + 0.006412759888917208, + 0.005606553517282009, + 0.005344112869352102, + 0.004704468883574009, + 0.004129206761717796, + 0.01026997622102499, + 0.005870571359992027, + 0.005965820048004389, + 0.005984701216220856, + 0.013988737016916275, + 0.019017472863197327, + 0.010576416738331318, + 0.013456962071359158, + 0.019135814160108566, + 0.025101924315094948, + 0.016338950023055077, + 0.014268537983298302, + 0.0067809391766786575, + 0.007103721611201763, + 0.008174585178494453, + 0.007591972127556801, + 0.006228409241884947, + 0.007320879027247429, + 0.00767563097178936, + 0.005616303998976946, + 0.02817920222878456, + 0.0324765108525753, + 0.024501163512468338, + 0.02722197212278843, + 0.044589217752218246, + 0.02609928883612156, + 0.033453766256570816, + 0.04766524210572243, + 0.009241866879165173, + 0.005884217564016581, + 0.004613666795194149, + 0.005304336082190275, + 0.004326586611568928, + 0.004626156762242317, + 0.0049268752336502075, + 0.0033539424184709787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18978339433670044, + 0.04159518703818321, + 0.025331854820251465, + 0.03054036945104599, + 0.045430686324834824, + 0.246242493391037, + 0.0712781548500061, + 0.1043350026011467, + 0.17849935591220856, + 0.06696347147226334 + ], + "credit_map": [ + 0.0268438458442688, + 0.023830119520425797, + 0.021534401923418045, + 0.019587010145187378, + 0.023533256724476814, + 0.025389373302459717, + 0.021933015435934067, + 0.027132362127304077, + 0.007465100847184658, + 0.004784869495779276, + 0.003349184524267912, + 0.003911588340997696, + 0.0048748645931482315, + 0.004810013342648745, + 0.007083272095769644, + 0.005316291935741901, + 0.004818716552108526, + 0.0025293149519711733, + 0.002253755694255233, + 0.0014482991537079215, + 0.004482828080654144, + 0.003664323128759861, + 0.002809179713949561, + 0.00332543789409101, + 0.003929402679204941, + 0.0033869449980556965, + 0.0027821471448987722, + 0.0029708719812333584, + 0.002794543979689479, + 0.002297365805134177, + 0.005204368382692337, + 0.007174723315984011, + 0.0066173141822218895, + 0.004496965557336807, + 0.004317087586969137, + 0.00427528191357851, + 0.006036998238414526, + 0.005098872818052769, + 0.007528328336775303, + 0.007059839088469744, + 0.028513764962553978, + 0.026852689683437347, + 0.028607897460460663, + 0.024417847394943237, + 0.0440095029771328, + 0.024504929780960083, + 0.03172402083873749, + 0.037611834704875946, + 0.01336290780454874, + 0.009028361178934574, + 0.009042328223586082, + 0.006371446419507265, + 0.008055407553911209, + 0.01065276749432087, + 0.006320040673017502, + 0.008444888517260551, + 0.010661862790584564, + 0.012447445653378963, + 0.007144198752939701, + 0.013997890055179596, + 0.013534781523048878, + 0.0115735474973917, + 0.021303175017237663, + 0.013672091998159885, + 0.021529408171772957, + 0.02261454612016678, + 0.018940221518278122, + 0.016735227778553963, + 0.02661730907857418, + 0.019584449008107185, + 0.02082398347556591, + 0.03165421634912491, + 0.01006192434579134, + 0.013165194541215897, + 0.008353703655302525, + 0.006492342334240675, + 0.006641789339482784, + 0.008810230530798435, + 0.007550750859081745, + 0.005887535400688648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.004636727273464203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002481818199158 + ], + "opposite_logprob_tokens": [ + -2.6950247287750244 + ], + "logprob_reward": -0.07002481818199158, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002481818199158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21436946094036102, + 0.06938084214925766, + 0.09200534224510193, + 0.05999123677611351, + 0.05321042984724045, + 0.08890826255083084, + 0.09961020201444626, + 0.09440059959888458, + 0.044877175241708755, + 0.18324647843837738 + ], + "credit_map": [ + 0.028953738510608673, + 0.026168983429670334, + 0.03072063997387886, + 0.027908992022275925, + 0.03194897621870041, + 0.03184136375784874, + 0.015236752107739449, + 0.02159002050757408, + 0.012580703012645245, + 0.007797032129019499, + 0.006381000857800245, + 0.007457860745489597, + 0.009164717048406601, + 0.009007981047034264, + 0.007523435167968273, + 0.009468112140893936, + 0.011625266633927822, + 0.009763444773852825, + 0.006445032544434071, + 0.012143936939537525, + 0.02053019031882286, + 0.009429165162146091, + 0.010807513259351254, + 0.011260785162448883, + 0.009508898481726646, + 0.006833627820014954, + 0.006523836404085159, + 0.004087368026375771, + 0.009385445155203342, + 0.007654153276234865, + 0.004599049221724272, + 0.011398856528103352, + 0.011039300821721554, + 0.006352108437567949, + 0.005804797168821096, + 0.003930887673050165, + 0.006523235235363245, + 0.007511825766414404, + 0.005194852594286203, + 0.006853422150015831, + 0.00918604526668787, + 0.010540485382080078, + 0.005375867709517479, + 0.01100278738886118, + 0.0083198556676507, + 0.011380448937416077, + 0.01662726327776909, + 0.016475511714816093, + 0.012869243510067463, + 0.013165290467441082, + 0.011220529675483704, + 0.01274868380278349, + 0.008685595355927944, + 0.018142392858862877, + 0.010227915830910206, + 0.012550555169582367, + 0.010012513026595116, + 0.01069978903979063, + 0.00782359391450882, + 0.011289672926068306, + 0.015608012676239014, + 0.00918525829911232, + 0.017243970185518265, + 0.012537786737084389, + 0.00754130445420742, + 0.006913518067449331, + 0.004934280179440975, + 0.005512393079698086, + 0.006375886034220457, + 0.006450669839978218, + 0.0036047608591616154, + 0.0035443631932139397, + 0.049495190382003784, + 0.030399369075894356, + 0.019203539937734604, + 0.020303552970290184, + 0.01597144827246666, + 0.020741868764162064, + 0.015713844448328018, + 0.011417659930884838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.07002481818199158 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999247521162033 + ], + "opposite_logprob_tokens": [ + -2.6949925422668457 + ], + "logprob_reward": -0.06999247521162033, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999247521162033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16889113187789917, + 0.11468972265720367, + 0.06791193783283234, + 0.051492661237716675, + 0.058364830911159515, + 0.11105524003505707, + 0.07482639700174332, + 0.06074841693043709, + 0.11854149401187897, + 0.1734781712293625 + ], + "credit_map": [ + 0.03589800000190735, + 0.02252732776105404, + 0.019808053970336914, + 0.01501497346907854, + 0.01861386187374592, + 0.0234880018979311, + 0.016494307667016983, + 0.01704660803079605, + 0.020274372771382332, + 0.011612570844590664, + 0.007493240293115377, + 0.014496242627501488, + 0.015289919450879097, + 0.010510310530662537, + 0.01676691137254238, + 0.01824616640806198, + 0.01613638922572136, + 0.008327432908117771, + 0.008479859679937363, + 0.008084415458142757, + 0.006996451411396265, + 0.008583281189203262, + 0.0050369626842439175, + 0.006267146673053503, + 0.006816153414547443, + 0.006193677894771099, + 0.0038485461845993996, + 0.0034925369545817375, + 0.01139158383011818, + 0.006962766405194998, + 0.004124821629375219, + 0.008662574924528599, + 0.012887965887784958, + 0.007321282755583525, + 0.008680921979248524, + 0.004367102403193712, + 0.004316564649343491, + 0.0073149967938661575, + 0.005860811565071344, + 0.00761518394574523, + 0.014299442991614342, + 0.011287684552371502, + 0.01372367050498724, + 0.0085859140381217, + 0.018296901136636734, + 0.01169531513005495, + 0.014803942292928696, + 0.018362367525696754, + 0.011447129771113396, + 0.010600075125694275, + 0.008239608258008957, + 0.006365044042468071, + 0.009151044301688671, + 0.013635415583848953, + 0.007921097800135612, + 0.007466979790478945, + 0.009185456670820713, + 0.008071226999163628, + 0.008183054625988007, + 0.009278587996959686, + 0.005409980658441782, + 0.008910652250051498, + 0.006344859022647142, + 0.005364598706364632, + 0.028563261032104492, + 0.017311112955212593, + 0.017254533246159554, + 0.010159796103835106, + 0.013181529007852077, + 0.017196739092469215, + 0.008132305927574635, + 0.006742215249687433, + 0.03158113732933998, + 0.0482378713786602, + 0.02066735550761223, + 0.012833540327847004, + 0.011099150404334068, + 0.03033309057354927, + 0.00936499610543251, + 0.009361031465232372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.06999247521162033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a slightly upward angle and panning slowly to the right, captures a solitary dancer, lost in the ecstasy of a twirl in the center of a grand yet empty Baroque ballroom. Her long, dark brown curls are elegantly tied into a loose bun, a few strands of hair freed from their restraints as she spins, gently brushing against her serene face, eyes closed and a hint of a smile playing on her lips. She wears a deep, royal blue, sleeveless silk gown, the tight bodice emphasizing her slender waist. The massive, multi-tiered skirt, lifted high by centrifugal force as she steadily and effortlessly spun, pivoting on the toe of her left foot, unfolds like a fully blossoming blue flower, almost parallel to the ground, its pleats drawing graceful, undulating curves in the air. A massive crystal chandelier atop the soaring dome casts a warm, radiant golden light, illuminating the mirror-like polished marble floor, which reflects the dancers' twirling figures and the blurred motion of their skirts, as well as the intricate gold reliefs on the walls. Light filters through the swirling skirts, casting ever-shifting, interwoven patterns of light and shadow on the floor. The entire scene, bathed in warm hues, creates an atmosphere of profound tranquility, dreamy elegance, and a touch of solitary elegance.", + "seed": 29088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_1/index_2695_seed_29088_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0340023448690772, + "vlm_logprob": -0.0340023448690772, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.0340023448690772, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004686443600803614 + ], + "opposite_logprob_tokens": [ + -6.00468635559082 + ], + "logprob_reward": -0.004686443600803614, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004686443600803614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058081354945898056, + 0.026451271027326584, + 0.019081957638263702, + 0.08869417756795883, + 0.04820583760738373, + 0.18895460665225983, + 0.4394589364528656, + 0.04982941225171089, + 0.019559282809495926, + 0.06168316304683685 + ], + "credit_map": [ + 0.011338994838297367, + 0.006335684098303318, + 0.00731122400611639, + 0.0058652255684137344, + 0.006693627685308456, + 0.0077228303998708725, + 0.006373871583491564, + 0.0064398981630802155, + 0.00323466956615448, + 0.002269479213282466, + 0.004006274044513702, + 0.0038952534087002277, + 0.0020966199226677418, + 0.004532348830252886, + 0.003427385352551937, + 0.0029892416205257177, + 0.002362958388403058, + 0.0019223071867600083, + 0.0024232855066657066, + 0.0026300547178834677, + 0.0024339864030480385, + 0.0018771402537822723, + 0.0031493904534727335, + 0.0022828367073088884, + 0.011013858020305634, + 0.01857811026275158, + 0.008822411298751831, + 0.006646140478551388, + 0.007998388260602951, + 0.021914707496762276, + 0.0052131484262645245, + 0.008507415652275085, + 0.013622559607028961, + 0.004187570419162512, + 0.004680895712226629, + 0.002436045790091157, + 0.002302284352481365, + 0.015272551216185093, + 0.0021496694535017014, + 0.003554263850674033, + 0.04485512524843216, + 0.030297202989459038, + 0.05222443863749504, + 0.010694220662117004, + 0.009220940992236137, + 0.02090553753077984, + 0.011999179609119892, + 0.008757954463362694, + 0.05339110642671585, + 0.09880118072032928, + 0.042909685522317886, + 0.03796904534101486, + 0.04028460383415222, + 0.1040051057934761, + 0.026602458208799362, + 0.03549572080373764, + 0.005445083137601614, + 0.017936958000063896, + 0.0035626254975795746, + 0.0028930753469467163, + 0.005014370661228895, + 0.008648555725812912, + 0.004058242775499821, + 0.0022704992443323135, + 0.0024431145284324884, + 0.002017129911109805, + 0.004295827355235815, + 0.002540529938414693, + 0.0026252716779708862, + 0.0024190449621528387, + 0.00181450170930475, + 0.0014038636581972241, + 0.013258333317935467, + 0.010852526873350143, + 0.005421423818916082, + 0.0054170796647667885, + 0.005283092148602009, + 0.005905379541218281, + 0.01130704302340746, + 0.004238280933350325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.004686443600803614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084526777267456 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084526777267456, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084526777267456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06643425673246384, + 0.03031977079808712, + 0.01576252467930317, + 0.033231593668460846, + 0.03781772032380104, + 0.06172259896993637, + 0.6183664798736572, + 0.04404419660568237, + 0.06130700185894966, + 0.03099389560520649 + ], + "credit_map": [ + 0.0072271511889994144, + 0.007477685809135437, + 0.011541360057890415, + 0.00602070614695549, + 0.008662360720336437, + 0.006251852959394455, + 0.01052479911595583, + 0.008728339336812496, + 0.0021307722199708223, + 0.0023627751506865025, + 0.003878525225445628, + 0.0049718646332621574, + 0.004926995374262333, + 0.002144408645108342, + 0.006380264647305012, + 0.003524161409586668, + 0.001970588928088546, + 0.0029757244046777487, + 0.0019501999486237764, + 0.0017240300076082349, + 0.0015411503845825791, + 0.0019547482952475548, + 0.001933194464072585, + 0.001712885219603777, + 0.0038956962525844574, + 0.004119894001632929, + 0.005049080587923527, + 0.004447268787771463, + 0.004069729242473841, + 0.004292136523872614, + 0.00384173053316772, + 0.0035160519182682037, + 0.009455258958041668, + 0.0030332908499985933, + 0.0036269277334213257, + 0.003086433047428727, + 0.00310019520111382, + 0.009716455824673176, + 0.002999167423695326, + 0.002799989888444543, + 0.015407193452119827, + 0.00986445602029562, + 0.014980902895331383, + 0.0039879316464066505, + 0.004168956074863672, + 0.005246885120868683, + 0.00475316122174263, + 0.003313106717541814, + 0.07660600543022156, + 0.13756109774112701, + 0.05980280041694641, + 0.05242261290550232, + 0.056904297322034836, + 0.14556381106376648, + 0.03943348675966263, + 0.0500723235309124, + 0.004444863647222519, + 0.01278967410326004, + 0.002978160046041012, + 0.0034702925477176905, + 0.003948707599192858, + 0.004538580309599638, + 0.007615376263856888, + 0.004258536268025637, + 0.0057855164632201195, + 0.020334947854280472, + 0.012620460242033005, + 0.005117577034980059, + 0.004448567051440477, + 0.006006932817399502, + 0.003851707559078932, + 0.003141285153105855, + 0.003943809773772955, + 0.0044682505540549755, + 0.0035919274669140577, + 0.003256372408941388, + 0.0041422974318265915, + 0.004063992295414209, + 0.004171470645815134, + 0.003355772700160742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.011084526777267456 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254980642348528 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254980642348528, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254980642348528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03064371645450592, + 0.02050194703042507, + 0.019451988860964775, + 0.12969034910202026, + 0.05370419844985008, + 0.07588110864162445, + 0.5515161156654358, + 0.026585444808006287, + 0.05997595563530922, + 0.032049186527729034 + ], + "credit_map": [ + 0.004281637258827686, + 0.0024954481050372124, + 0.004198566544800997, + 0.0024032003711909056, + 0.00411802576854825, + 0.003980925306677818, + 0.005465524736791849, + 0.003700385568663478, + 0.001843354431912303, + 0.0017737076850607991, + 0.0018697374034672976, + 0.004301426466554403, + 0.0027634822763502598, + 0.0014957425883039832, + 0.0040185111574828625, + 0.0024359833914786577, + 0.0027854046784341335, + 0.0030910144560039043, + 0.0019913334399461746, + 0.0021084127947688103, + 0.0015186702366918325, + 0.0025637377984821796, + 0.002968194894492626, + 0.002425219863653183, + 0.018295859917998314, + 0.020973382517695427, + 0.013735052198171616, + 0.013563951477408409, + 0.012531864456832409, + 0.028151707723736763, + 0.00998689141124487, + 0.012451637536287308, + 0.01674935594201088, + 0.00256103347055614, + 0.004927507136017084, + 0.0026622030418366194, + 0.0022077751345932484, + 0.01844203658401966, + 0.0024229385890066624, + 0.003731345757842064, + 0.02115052565932274, + 0.010146057233214378, + 0.022323627024888992, + 0.004107379820197821, + 0.003185061737895012, + 0.00677205016836524, + 0.00458223232999444, + 0.003614165121689439, + 0.06684430688619614, + 0.1219698116183281, + 0.05275890231132507, + 0.04792187735438347, + 0.05064360052347183, + 0.13011400401592255, + 0.03672286123037338, + 0.044540710747241974, + 0.0022178925573825836, + 0.0070135705173015594, + 0.0019036374287679791, + 0.0027395428624004126, + 0.0031785336323082447, + 0.0038828628603368998, + 0.003004420083016157, + 0.002644983120262623, + 0.00497419061139226, + 0.017331019043922424, + 0.018151894211769104, + 0.0034132669679820538, + 0.0033807451836764812, + 0.005635649897158146, + 0.003321486059576273, + 0.003767698537558317, + 0.0045721568167209625, + 0.005719385575503111, + 0.0029654211830347776, + 0.0027607667725533247, + 0.0029774534050375223, + 0.004266731441020966, + 0.004482673015445471, + 0.004304594825953245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.005254980642348528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700293779373169 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.0700293779373169, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700293779373169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018661553040146828, + 0.006236739456653595, + 0.007358239032328129, + 0.07032101601362228, + 0.03343474492430687, + 0.021843818947672844, + 0.7569651007652283, + 0.01653333380818367, + 0.05115809291601181, + 0.01748736947774887 + ], + "credit_map": [ + 0.002330006565898657, + 0.0022007967345416546, + 0.0027796830981969833, + 0.0013861358165740967, + 0.0026457770727574825, + 0.002199511043727398, + 0.0030470441561192274, + 0.0020725992508232594, + 0.0006795912049710751, + 0.0006536603905260563, + 0.0008178857387974858, + 0.0007063930970616639, + 0.0007728468044660985, + 0.0011296645971015096, + 0.0008364153327420354, + 0.000640282582025975, + 0.0008315751329064369, + 0.0008515460649505258, + 0.0010152600007131696, + 0.0009176884195767343, + 0.0007721952279098332, + 0.0010631964541971684, + 0.0011240891180932522, + 0.0007826882065273821, + 0.009520469233393669, + 0.014112592674791813, + 0.006295519880950451, + 0.005039194133132696, + 0.006680626887828112, + 0.01805613562464714, + 0.004182526376098394, + 0.006433946080505848, + 0.01106974482536316, + 0.0015741651877760887, + 0.003163462271913886, + 0.0011120201088488102, + 0.001516283256933093, + 0.011479299515485764, + 0.0014015573542565107, + 0.002118209842592478, + 0.005395529791712761, + 0.0022410890087485313, + 0.007293350994586945, + 0.001137329963967204, + 0.0010576288914307952, + 0.0020386932883411646, + 0.0016906779492273927, + 0.0009895183611661196, + 0.0906214714050293, + 0.17115239799022675, + 0.07313304394483566, + 0.06563813984394073, + 0.0695180594921112, + 0.17964531481266022, + 0.04592369124293327, + 0.06133295223116875, + 0.001413771533407271, + 0.001917605404742062, + 0.0014560091076418757, + 0.0021664348896592855, + 0.002801529597491026, + 0.0014817917253822088, + 0.003377980552613735, + 0.0019182114629074931, + 0.004183677490800619, + 0.017034918069839478, + 0.014483439736068249, + 0.0032852324657142162, + 0.002117760479450226, + 0.0049313209019601345, + 0.0025240015238523483, + 0.0025977434124797583, + 0.004764953628182411, + 0.0029862283263355494, + 0.001837324001826346, + 0.0015090181259438396, + 0.0012507810024544597, + 0.001641537994146347, + 0.0012147678062319756, + 0.002282757544890046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.0700293779373169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, empty Baroque ballroom with gold reliefs on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a female dancer wearing a deep royal blue sleeveless silk gown with a tight bodice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's skirt a massive multi-tiered blue skirt that is lifted high and unfolds parallel to the ground while she spins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dancer spinning on the toe of her left foot with her eyes closed and a hint of a smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a wide-angle shot from a slightly upward angle that pans slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895639538764954 + ], + "opposite_logprob_tokens": [ + -2.578956365585327 + ], + "logprob_reward": -0.07895639538764954, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895639538764954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0712331086397171, + 0.031366124749183655, + 0.030569395050406456, + 0.20177023112773895, + 0.0795195922255516, + 0.10606382042169571, + 0.33566373586654663, + 0.024820882827043533, + 0.0682460144162178, + 0.050747066736221313 + ], + "credit_map": [ + 0.01514696329832077, + 0.011329811066389084, + 0.008729236200451851, + 0.004716854076832533, + 0.006061609834432602, + 0.011090192943811417, + 0.00519056199118495, + 0.008967877365648746, + 0.004578025080263615, + 0.0033263631630688906, + 0.004128977190703154, + 0.002553720725700259, + 0.0029800981283187866, + 0.00754174217581749, + 0.002391419606283307, + 0.003865778213366866, + 0.004486510064452887, + 0.00315709225833416, + 0.0035371738485991955, + 0.004828477744013071, + 0.0031549304258078337, + 0.0024918841663748026, + 0.004859637934714556, + 0.004053687211126089, + 0.02468203194439411, + 0.03963390365242958, + 0.01932086981832981, + 0.015750188380479813, + 0.01944715529680252, + 0.04919307306408882, + 0.01365465484559536, + 0.02008836157619953, + 0.023328682407736778, + 0.0036897705867886543, + 0.006587136536836624, + 0.005045758560299873, + 0.004136206582188606, + 0.02524598501622677, + 0.004998440388590097, + 0.006487609818577766, + 0.032440491020679474, + 0.012913254089653492, + 0.026003170758485794, + 0.007768738083541393, + 0.005847640335559845, + 0.007403761148452759, + 0.008306042291224003, + 0.005380723159760237, + 0.06738273799419403, + 0.04958343878388405, + 0.0602177232503891, + 0.027251174673438072, + 0.032679542899131775, + 0.03495190665125847, + 0.0379478894174099, + 0.0256492979824543, + 0.0020814232993870974, + 0.003172115655615926, + 0.0025549493730068207, + 0.0025265482254326344, + 0.0036161320749670267, + 0.0038832747377455235, + 0.0029109157621860504, + 0.004075523931533098, + 0.006418307311832905, + 0.015294925309717655, + 0.01639379747211933, + 0.006379401311278343, + 0.005733748897910118, + 0.006385107524693012, + 0.0053240275010466576, + 0.006316700484603643, + 0.009497478604316711, + 0.006054141093045473, + 0.0051426962018013, + 0.005423457361757755, + 0.005719659850001335, + 0.005538675934076309, + 0.004846332129091024, + 0.00852462649345398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.07895639538764954 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_tensor([6936, 6936])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_tensor([6936, 6936])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a504570b29c9ff454ca13ed8aaba0fcc682fe440 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_tensor([6936, 6936])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing, empty shot captures a corner of a bare room. A warm, slanting beam of sunset light filters through an unrecognized window, casting a precise shadow across the dusty wooden floor, illuminating a pair of aluminum medical crutches resting quietly against a mottled corner. Their gray rubber grips and armpit pads appear worn, and the gleaming, cold metal brackets bear slight scratches, silently telling a forgotten story. As the camera moves forward ever more slowly, an incredible process unfolds: the cold metal surface of the crutches begins to petrify inch by inch, its original luster gradually fading, replaced by a rough, grayish-white texture. Microscopic cracks creep from the contact points like a spiderweb, strikingly clear in the warm, golden light. Finally, when the camera lingers in close-up, the pair of crutches has completely transformed into a silent sculpture, covered with deep and poignant cracks, as if solidifying a heavy memory in an eternal form. The whole picture, with its sharp contrast of warm light and cold texture, creates a sad atmosphere of forgetfulness, silence and the solidification of time.", + "seed": 24576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_6936_seed_24576_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29092521239072083, + "vlm_logprob": -0.29092521239072083, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.29092521239072083, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4086955785751343 + ], + "opposite_logprob_tokens": [ + -0.2836955487728119 + ], + "logprob_reward": -1.4086955785751343, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4086955785751343, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.196616992354393, + 0.1153893768787384, + 0.12565571069717407, + 0.08028731495141983, + 0.05908160284161568, + 0.04988017678260803, + 0.12336020916700363, + 0.13392913341522217, + 0.056675054132938385, + 0.05912443995475769 + ], + "credit_map": [ + 0.02441151812672615, + 0.022898871451616287, + 0.016194041818380356, + 0.021869588643312454, + 0.026786021888256073, + 0.038310568779706955, + 0.024189408868551254, + 0.021956956014037132, + 0.012197869829833508, + 0.012541605159640312, + 0.009563405066728592, + 0.012235199101269245, + 0.019800422713160515, + 0.025939764454960823, + 0.012844840995967388, + 0.010266263037919998, + 0.0260036401450634, + 0.017421457916498184, + 0.009436553344130516, + 0.0070886993780732155, + 0.012012061662971973, + 0.017366763204336166, + 0.00913909636437893, + 0.027187447994947433, + 0.011681979522109032, + 0.009829066693782806, + 0.009348399937152863, + 0.006055811885744333, + 0.010660543106496334, + 0.011847077868878841, + 0.010553005151450634, + 0.010311434976756573, + 0.009420620277523994, + 0.008025966584682465, + 0.009809393435716629, + 0.004053782671689987, + 0.005833141505718231, + 0.01138503197580576, + 0.005019632633775473, + 0.005534031894057989, + 0.010975290089845657, + 0.0064533911645412445, + 0.00450193602591753, + 0.0034385465551167727, + 0.0032362695783376694, + 0.00917388778179884, + 0.006857596337795258, + 0.005243259482085705, + 0.019090188667178154, + 0.01089485827833414, + 0.0087751355022192, + 0.017606418579816818, + 0.012927093543112278, + 0.01784382201731205, + 0.021404754370450974, + 0.014817927032709122, + 0.020235639065504074, + 0.011201148852705956, + 0.009528305381536484, + 0.013725523836910725, + 0.010609449818730354, + 0.030833149328827858, + 0.018812552094459534, + 0.018983373418450356, + 0.007244572974741459, + 0.006475085858255625, + 0.003861866658553481, + 0.009446450509130955, + 0.0068115415051579475, + 0.007581572979688644, + 0.00688142841681838, + 0.008372536860406399, + 0.0077823493629693985, + 0.006116344127804041, + 0.003884586039930582, + 0.010890983045101166, + 0.004776710644364357, + 0.007512866519391537, + 0.010585763491690159, + 0.0075748395174741745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -1.4086955785751343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029833149164915085 + ], + "opposite_logprob_tokens": [ + -3.5298330783843994 + ], + "logprob_reward": -0.029833149164915085, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029833149164915085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3453940451145172, + 0.039774324744939804, + 0.16547149419784546, + 0.048087552189826965, + 0.04756039008498192, + 0.05901039391756058, + 0.0714876726269722, + 0.0983816534280777, + 0.0934784784913063, + 0.03135392442345619 + ], + "credit_map": [ + 0.04179468750953674, + 0.041940733790397644, + 0.025687113404273987, + 0.032882727682590485, + 0.04658007249236107, + 0.0783575028181076, + 0.04186852648854256, + 0.03628268465399742, + 0.0037272614426910877, + 0.0042542642913758755, + 0.003334728069603443, + 0.004478943534195423, + 0.005989202298223972, + 0.007988969795405865, + 0.005954785738140345, + 0.004046172369271517, + 0.043717160820961, + 0.055224187672138214, + 0.009397188201546669, + 0.008058511652052402, + 0.015332542359828949, + 0.012524006888270378, + 0.009788897819817066, + 0.011429001577198505, + 0.006795709021389484, + 0.006126314867287874, + 0.004693519324064255, + 0.003810569643974304, + 0.006519841030240059, + 0.00673044053837657, + 0.00748981861397624, + 0.005921341944485903, + 0.005984166637063026, + 0.005533824674785137, + 0.007252620533108711, + 0.0054659657180309296, + 0.005563986487686634, + 0.005024062469601631, + 0.005575660616159439, + 0.007160102482885122, + 0.014700917527079582, + 0.008055703714489937, + 0.005004770588129759, + 0.004548412747681141, + 0.0034548561088740826, + 0.010476801544427872, + 0.006436934694647789, + 0.006332001648843288, + 0.010341041721403599, + 0.006849939934909344, + 0.010413942858576775, + 0.011143704876303673, + 0.006162772886455059, + 0.007936229929327965, + 0.009240805171430111, + 0.009399229660630226, + 0.015135981142520905, + 0.01203057263046503, + 0.00764029286801815, + 0.010523194447159767, + 0.013066863641142845, + 0.01663677766919136, + 0.013848762027919292, + 0.009499212726950645, + 0.012784883379936218, + 0.011820245534181595, + 0.006656934041529894, + 0.017985045909881592, + 0.008576810359954834, + 0.012445596978068352, + 0.011569592170417309, + 0.011639367789030075, + 0.0035926238633692265, + 0.0032380856573581696, + 0.002322198823094368, + 0.006843786220997572, + 0.002394626382738352, + 0.0031070788390934467, + 0.006286760326474905, + 0.00356876477599144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.029833149164915085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264822859317064 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264822859317064, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264822859317064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04764861240983009, + 0.06553798168897629, + 0.06942138075828552, + 0.028346078470349312, + 0.08333928883075714, + 0.12311204522848129, + 0.0618237666785717, + 0.08952441066503525, + 0.0839192271232605, + 0.34732720255851746 + ], + "credit_map": [ + 0.006687450222671032, + 0.0060703265480697155, + 0.0037104864604771137, + 0.006581770721822977, + 0.006576309911906719, + 0.0052945371717214584, + 0.0073645757511258125, + 0.005363157019019127, + 0.006813609041273594, + 0.00791384931653738, + 0.005512894131243229, + 0.009317376650869846, + 0.01047825813293457, + 0.011417663656175137, + 0.00847578328102827, + 0.005608550738543272, + 0.007528640329837799, + 0.01582001894712448, + 0.0076692900620400906, + 0.004704039078205824, + 0.004748029168695211, + 0.006398804020136595, + 0.004872267134487629, + 0.01768028922379017, + 0.0043235826306045055, + 0.0037412047386169434, + 0.0032725012861192226, + 0.002194733126088977, + 0.0035743152257055044, + 0.004313402809202671, + 0.0032911545131355524, + 0.00363518251106143, + 0.010794484987854958, + 0.012157445773482323, + 0.009211953729391098, + 0.008178861811757088, + 0.011018713936209679, + 0.009281226433813572, + 0.008061401546001434, + 0.014635197818279266, + 0.01986025832593441, + 0.012955045327544212, + 0.011911630630493164, + 0.015172086656093597, + 0.011730298399925232, + 0.011172102764248848, + 0.023597942665219307, + 0.01671268418431282, + 0.009330116212368011, + 0.007444233167916536, + 0.004810075741261244, + 0.009266044944524765, + 0.006040614098310471, + 0.01206403411924839, + 0.0072480239905416965, + 0.005620625335723162, + 0.011257250793278217, + 0.012190517038106918, + 0.006554309278726578, + 0.011614405550062656, + 0.010385175235569477, + 0.010902758687734604, + 0.01434459537267685, + 0.01227540522813797, + 0.011021975427865982, + 0.008583529852330685, + 0.007017519325017929, + 0.014621594920754433, + 0.0077627915889024734, + 0.011034570634365082, + 0.015298496000468731, + 0.008578754961490631, + 0.04369417950510979, + 0.03442702442407608, + 0.02659098245203495, + 0.0699949711561203, + 0.024697313085198402, + 0.03796908259391785, + 0.06600500643253326, + 0.04394863545894623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.005264822859317064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739741191267967 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739741191267967, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739741191267967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05848564952611923, + 0.05616327375173569, + 0.06094659864902496, + 0.02687545120716095, + 0.04892290383577347, + 0.06752894073724747, + 0.03459906950592995, + 0.11705058813095093, + 0.05932939797639847, + 0.4700981080532074 + ], + "credit_map": [ + 0.008155051618814468, + 0.00771839777007699, + 0.005869108717888594, + 0.00645531015470624, + 0.0078821936622262, + 0.008630937896668911, + 0.007706863805651665, + 0.006067783106118441, + 0.006494029425084591, + 0.006694409996271133, + 0.004874407779425383, + 0.005448478274047375, + 0.0097840316593647, + 0.01158001460134983, + 0.006360089872032404, + 0.004927807953208685, + 0.012582216411828995, + 0.018060093745589256, + 0.004828061442822218, + 0.003419035580009222, + 0.004712425637990236, + 0.005439389497041702, + 0.00427224300801754, + 0.0076331328600645065, + 0.003600668627768755, + 0.0034608852583914995, + 0.0038239415735006332, + 0.0024842708371579647, + 0.0022073029540479183, + 0.0053784968331456184, + 0.002859466476365924, + 0.003060418413951993, + 0.0064342147670686245, + 0.007351808249950409, + 0.004912602715194225, + 0.003313040593639016, + 0.006003785878419876, + 0.012836119160056114, + 0.004660638980567455, + 0.0034106923267245293, + 0.013625514693558216, + 0.01078068371862173, + 0.004607770591974258, + 0.006120836827903986, + 0.005557615775614977, + 0.011021674610674381, + 0.009586703032255173, + 0.006228141952306032, + 0.0053761801682412624, + 0.002619390841573477, + 0.0035635235253721476, + 0.003680919064208865, + 0.003401330905035138, + 0.004864375572651625, + 0.005731899756938219, + 0.005361449904739857, + 0.01867164485156536, + 0.01944383606314659, + 0.007404693402349949, + 0.012319136410951614, + 0.01316219661384821, + 0.020432308316230774, + 0.011250886134803295, + 0.014365882612764835, + 0.008257164619863033, + 0.006684924941509962, + 0.005208007991313934, + 0.0084363529458642, + 0.006345145870000124, + 0.009042786434292793, + 0.009450234472751617, + 0.005904782097786665, + 0.06023000180721283, + 0.04655000567436218, + 0.03547796607017517, + 0.09397667646408081, + 0.03311358764767647, + 0.05011223629117012, + 0.09192199259996414, + 0.05871564894914627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.006739741191267967 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092770162969828 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092770162969828, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092770162969828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05281691998243332, + 0.07789577543735504, + 0.0851694867014885, + 0.07737012952566147, + 0.17169061303138733, + 0.08864864706993103, + 0.22491975128650665, + 0.05938855931162834, + 0.07782171666622162, + 0.084278404712677 + ], + "credit_map": [ + 0.008161363191902637, + 0.00663950527086854, + 0.004580607172101736, + 0.007015982177108526, + 0.007574682589620352, + 0.005911948624998331, + 0.0075498609803617, + 0.005382972303777933, + 0.008303554728627205, + 0.008612630888819695, + 0.00595129793509841, + 0.0070634749718010426, + 0.013741549104452133, + 0.018318163231015205, + 0.008678876794874668, + 0.007226223591715097, + 0.011567250825464725, + 0.014184372499585152, + 0.012205484323203564, + 0.006117003504186869, + 0.009632866829633713, + 0.012481942772865295, + 0.009496305137872696, + 0.009484262205660343, + 0.009926196187734604, + 0.0139747504144907, + 0.007075471803545952, + 0.004599949810653925, + 0.00959368608891964, + 0.015640396624803543, + 0.006058545783162117, + 0.010501130484044552, + 0.022039249539375305, + 0.027626775205135345, + 0.017710821703076363, + 0.018550336360931396, + 0.01939089596271515, + 0.021260561421513557, + 0.014771671034395695, + 0.03034030646085739, + 0.0142460186034441, + 0.010343837551772594, + 0.009619983844459057, + 0.010903002694249153, + 0.008379427716135979, + 0.00850171409547329, + 0.01623259298503399, + 0.010422071442008018, + 0.030383091419935226, + 0.02285441756248474, + 0.022691842168569565, + 0.02453836239874363, + 0.02321748062968254, + 0.044972874224185944, + 0.031695302575826645, + 0.024566374719142914, + 0.00925393495708704, + 0.006547562777996063, + 0.004528244491666555, + 0.007247547153383493, + 0.004728671163320541, + 0.0137621546164155, + 0.0061703454703092575, + 0.007150100078433752, + 0.01004827581346035, + 0.009196119382977486, + 0.005035272799432278, + 0.017002331092953682, + 0.007050287909805775, + 0.00894507672637701, + 0.010621835477650166, + 0.009922521188855171, + 0.011040767654776573, + 0.008639847859740257, + 0.006155628710985184, + 0.01635804958641529, + 0.00603480963036418, + 0.008827653713524342, + 0.016462119296193123, + 0.010759527795016766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.004092770162969828 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing, empty shot captures a corner of a bare room. A warm, slanting beam of sunset light filters through an unrecognized window, casting a precise shadow across the dusty wooden floor, illuminating a pair of aluminum medical crutches resting quietly against a mottled corner. Their gray rubber grips and armpit pads appear worn, and the gleaming, cold metal brackets bear slight scratches, silently telling a forgotten story. As the camera moves forward ever more slowly, an incredible process unfolds: the cold metal surface of the crutches begins to petrify inch by inch, its original luster gradually fading, replaced by a rough, grayish-white texture. Microscopic cracks creep from the contact points like a spiderweb, strikingly clear in the warm, golden light. Finally, when the camera lingers in close-up, the pair of crutches has completely transformed into a silent sculpture, covered with deep and poignant cracks, as if solidifying a heavy memory in an eternal form. The whole picture, with its sharp contrast of warm light and cold texture, creates a sad atmosphere of forgetfulness, silence and the solidification of time.", + "seed": 24577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_6936_seed_24577_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.037257101759314536, + "vlm_logprob": -0.037257101759314536, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.037257101759314536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.071683868765831 + ], + "opposite_logprob_tokens": [ + -2.821683883666992 + ], + "logprob_reward": -0.071683868765831, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.071683868765831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45302751660346985, + 0.0765562653541565, + 0.06695324182510376, + 0.01499757170677185, + 0.05541950464248657, + 0.03571076691150665, + 0.045517269521951675, + 0.05788937211036682, + 0.09723643213510513, + 0.09669207781553268 + ], + "credit_map": [ + 0.08060546219348907, + 0.028387190774083138, + 0.07398722320795059, + 0.04384386166930199, + 0.02433018758893013, + 0.06824872642755508, + 0.02690502442419529, + 0.10671987384557724, + 0.009215809404850006, + 0.00871796440333128, + 0.010513068176805973, + 0.010820579715073109, + 0.005540706217288971, + 0.012610548175871372, + 0.007342521101236343, + 0.011795071884989738, + 0.012611630372703075, + 0.004096314311027527, + 0.009946336969733238, + 0.0059495409950613976, + 0.0035981438122689724, + 0.010689789429306984, + 0.004840973298996687, + 0.015220504254102707, + 0.0015898160636425018, + 0.0013318967539817095, + 0.002590000629425049, + 0.0013315572869032621, + 0.0012044681934639812, + 0.002460529562085867, + 0.0012485802872106433, + 0.0032407224643975496, + 0.004820514470338821, + 0.009555405005812645, + 0.005757702514529228, + 0.006723188795149326, + 0.005209457129240036, + 0.009110779501497746, + 0.004738686606287956, + 0.009503770619630814, + 0.0032380938064306974, + 0.006417963188141584, + 0.002729250816628337, + 0.002712295623496175, + 0.003871292108669877, + 0.005909864790737629, + 0.00293363886885345, + 0.007898366078734398, + 0.008068198338150978, + 0.004146311432123184, + 0.004833762999624014, + 0.003472006879746914, + 0.005554631352424622, + 0.0057937828823924065, + 0.004406406544148922, + 0.009242168627679348, + 0.009090905077755451, + 0.0039338404312729836, + 0.004956747870892286, + 0.0031602545641362667, + 0.006440095137804747, + 0.011563240550458431, + 0.0037516928277909756, + 0.01499259378761053, + 0.01591399870812893, + 0.009173817001283169, + 0.013606640510261059, + 0.007640485651791096, + 0.013765810988843441, + 0.012463221326470375, + 0.01028989627957344, + 0.014382564462721348, + 0.014378449879586697, + 0.019529789686203003, + 0.010340259410440922, + 0.005289280787110329, + 0.013375972397625446, + 0.012272405438125134, + 0.009122436866164207, + 0.012383487075567245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.071683868765831 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009819891303777695 + ], + "opposite_logprob_tokens": [ + -4.634819984436035 + ], + "logprob_reward": -0.009819891303777695, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009819891303777695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13763980567455292, + 0.10725697129964828, + 0.07198845595121384, + 0.026062162593007088, + 0.055801697075366974, + 0.04641621187329292, + 0.03203440457582474, + 0.18640337884426117, + 0.13093647360801697, + 0.20546042919158936 + ], + "credit_map": [ + 0.021873963996767998, + 0.010861380957067013, + 0.021917320787906647, + 0.0222568791359663, + 0.007682436145842075, + 0.01570136472582817, + 0.012885056436061859, + 0.024461403489112854, + 0.009785530157387257, + 0.008635829202830791, + 0.015699928626418114, + 0.019902406260371208, + 0.007585098501294851, + 0.015881715342402458, + 0.011642013676464558, + 0.018124449998140335, + 0.00759940454736352, + 0.005761391017585993, + 0.01092919334769249, + 0.006845427677035332, + 0.005529686808586121, + 0.011175092309713364, + 0.006951924879103899, + 0.017196334898471832, + 0.00253435131162405, + 0.0021888581104576588, + 0.003969710320234299, + 0.002604406327009201, + 0.0018299956573173404, + 0.004632050171494484, + 0.002152307191863656, + 0.006150483153760433, + 0.006953204516321421, + 0.005530534777790308, + 0.006648138165473938, + 0.004349351394921541, + 0.0028842852916568518, + 0.013676386326551437, + 0.0024239742197096348, + 0.013335819356143475, + 0.008039595559239388, + 0.007330710534006357, + 0.003830173285678029, + 0.0034240451641380787, + 0.004003445152193308, + 0.007504866924136877, + 0.004180604591965675, + 0.00810276996344328, + 0.004518427886068821, + 0.004540007095783949, + 0.0021746791899204254, + 0.0027960774023085833, + 0.005741787143051624, + 0.0037779139820486307, + 0.003814808325842023, + 0.004670706111937761, + 0.024150315672159195, + 0.018560808151960373, + 0.015061341226100922, + 0.014619166031479836, + 0.01960596814751625, + 0.03231046721339226, + 0.02908233553171158, + 0.033012986183166504, + 0.01895688846707344, + 0.01452895998954773, + 0.018527396023273468, + 0.01063027884811163, + 0.019873682409524918, + 0.016068438068032265, + 0.013099394738674164, + 0.01925143413245678, + 0.02620210312306881, + 0.042730849236249924, + 0.020584644749760628, + 0.012938021682202816, + 0.028938792645931244, + 0.022841546684503555, + 0.02969779260456562, + 0.02152668498456478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.009819891303777695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905647695064545 + ], + "opposite_logprob_tokens": [ + -2.4640564918518066 + ], + "logprob_reward": -0.08905647695064545, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905647695064545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15087291598320007, + 0.06218951195478439, + 0.02817612886428833, + 0.04597526043653488, + 0.0334448479115963, + 0.026566537097096443, + 0.09748324006795883, + 0.14627741277217865, + 0.14818574488162994, + 0.2608284056186676 + ], + "credit_map": [ + 0.02956751175224781, + 0.009334285743534565, + 0.02292962372303009, + 0.018627578392624855, + 0.008543817326426506, + 0.019890060648322105, + 0.01083212997764349, + 0.031147904694080353, + 0.005216354504227638, + 0.006139328703284264, + 0.00927379634231329, + 0.013023646548390388, + 0.004676147364079952, + 0.008956865407526493, + 0.0056138597428798676, + 0.009289512410759926, + 0.003189947223290801, + 0.00253369752317667, + 0.003436013590544462, + 0.0032231248915195465, + 0.002377151045948267, + 0.004700563848018646, + 0.002816763473674655, + 0.0058988663367927074, + 0.003725717542693019, + 0.0028333053924143314, + 0.006185299716889858, + 0.004679222125560045, + 0.0022796106059104204, + 0.00875464640557766, + 0.002638702280819416, + 0.014878759160637856, + 0.002724254038184881, + 0.006266768090426922, + 0.003514682175591588, + 0.004201128613203764, + 0.003614061279222369, + 0.004752510692924261, + 0.002928204368799925, + 0.005443238187581301, + 0.0031773510854691267, + 0.0023564735893160105, + 0.0023913842160254717, + 0.0025279384572058916, + 0.0023276968859136105, + 0.00542157469317317, + 0.0021505835466086864, + 0.00621353555470705, + 0.0071935285814106464, + 0.016477923840284348, + 0.007164087612181902, + 0.009192166849970818, + 0.01873394288122654, + 0.011317958123981953, + 0.013048558495938778, + 0.014355073682963848, + 0.014906954020261765, + 0.015089825727045536, + 0.01476058829575777, + 0.016614872962236404, + 0.020701328292489052, + 0.022538725286722183, + 0.021592499688267708, + 0.02007262036204338, + 0.0196217130869627, + 0.014921985566616058, + 0.022121289744973183, + 0.01153455302119255, + 0.020625922828912735, + 0.019676558673381805, + 0.016668172553181648, + 0.02301555871963501, + 0.03243229538202286, + 0.053448304533958435, + 0.02648610807955265, + 0.01610563136637211, + 0.03626204654574394, + 0.031569529324769974, + 0.03598896786570549, + 0.028535539284348488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.08905647695064545 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087002232670784 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.011087002232670784, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087002232670784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24227303266525269, + 0.0284480769187212, + 0.014575048349797726, + 0.01151752658188343, + 0.020426597446203232, + 0.017473552376031876, + 0.037649933248758316, + 0.15549588203430176, + 0.07217695564031601, + 0.3999634385108948 + ], + "credit_map": [ + 0.04857214167714119, + 0.014920963905751705, + 0.03868516907095909, + 0.024121839553117752, + 0.013387939892709255, + 0.038395706564188004, + 0.015420163050293922, + 0.048769108951091766, + 0.0027980851009488106, + 0.0026798914186656475, + 0.004105271305888891, + 0.005679509602487087, + 0.0021735853515565395, + 0.004148837178945541, + 0.003160133957862854, + 0.003702759277075529, + 0.001683936920017004, + 0.0013910538982599974, + 0.0017728251405060291, + 0.0016336182598024607, + 0.0013589588925242424, + 0.0022350165527313948, + 0.0014811625005677342, + 0.0030184760689735413, + 0.0008742267382331192, + 0.000824737420771271, + 0.001504896441474557, + 0.0011466466821730137, + 0.0006985287764109671, + 0.0021122083999216557, + 0.0007807016954757273, + 0.003575579496100545, + 0.0016367794014513493, + 0.0019426578655838966, + 0.002392577938735485, + 0.0017321029445156455, + 0.002380063058808446, + 0.004899337887763977, + 0.0012960623716935515, + 0.004147014115005732, + 0.002062720013782382, + 0.0016728841001167893, + 0.001363809104077518, + 0.001191400457173586, + 0.0014798275660723448, + 0.0043213688768446445, + 0.0011647676583379507, + 0.0042167725041508675, + 0.004433873575180769, + 0.004919493570923805, + 0.0030594163108617067, + 0.0032777595333755016, + 0.006106238346546888, + 0.0043619731441140175, + 0.005142584443092346, + 0.006348594091832638, + 0.02223113924264908, + 0.013360263779759407, + 0.013162628747522831, + 0.011204770766198635, + 0.01789373718202114, + 0.027333542704582214, + 0.0217987559735775, + 0.028511041775345802, + 0.010680004954338074, + 0.008029255084693432, + 0.007095192559063435, + 0.005778543185442686, + 0.01123686321079731, + 0.009416110813617706, + 0.006566439755260944, + 0.013374540954828262, + 0.051854655146598816, + 0.07756075263023376, + 0.037407614290714264, + 0.02261088415980339, + 0.05328383296728134, + 0.050838109105825424, + 0.04830153286457062, + 0.05810604616999626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.011087002232670784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638269543647766 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638269543647766, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638269543647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1896563470363617, + 0.06524929404258728, + 0.03359024599194527, + 0.023752273991703987, + 0.02594091184437275, + 0.025649605318903923, + 0.06407900899648666, + 0.1235911026597023, + 0.17276452481746674, + 0.2757267951965332 + ], + "credit_map": [ + 0.03486480936408043, + 0.014030689373612404, + 0.029888072982430458, + 0.021496664732694626, + 0.010310777463018894, + 0.028401140123605728, + 0.010639769956469536, + 0.04002441465854645, + 0.004542156588286161, + 0.00621189596131444, + 0.009335910901427269, + 0.011318453587591648, + 0.006096833385527134, + 0.010102993808686733, + 0.008281172253191471, + 0.009359870105981827, + 0.0030698601622134447, + 0.003915075212717056, + 0.003960983827710152, + 0.00478091137483716, + 0.0031167929992079735, + 0.004519594833254814, + 0.004951214883476496, + 0.00527581013739109, + 0.0017562336288392544, + 0.003682450158521533, + 0.0028940618503838778, + 0.0032328013330698013, + 0.0022089816629886627, + 0.00344750564545393, + 0.0016186352586373687, + 0.004911602474749088, + 0.0027044399175792933, + 0.0037756238598376513, + 0.0031961887143552303, + 0.0031724630389362574, + 0.002831161255016923, + 0.003603659337386489, + 0.0026779042091220617, + 0.003979469183832407, + 0.0026696587447077036, + 0.005870998837053776, + 0.0020898920483887196, + 0.002062019193544984, + 0.003784107742831111, + 0.0040776077657938, + 0.0020638802088797092, + 0.003031437750905752, + 0.008365445770323277, + 0.008229088969528675, + 0.005530992988497019, + 0.008341602049767971, + 0.01072740275412798, + 0.007447478361427784, + 0.009537965059280396, + 0.005899024195969105, + 0.013630508445203304, + 0.014810782857239246, + 0.014920425601303577, + 0.013161612674593925, + 0.015955282375216484, + 0.01746293157339096, + 0.014963771216571331, + 0.018685780465602875, + 0.025761792436242104, + 0.019677618518471718, + 0.023256391286849976, + 0.013657600618898869, + 0.0206135343760252, + 0.023940404877066612, + 0.016896072775125504, + 0.02896108850836754, + 0.0350915901362896, + 0.05786800757050514, + 0.027721436694264412, + 0.017314789816737175, + 0.03843340277671814, + 0.031010083854198456, + 0.03658369556069374, + 0.03170372545719147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004638269543647766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing, empty shot captures a corner of a bare room. A warm, slanting beam of sunset light filters through an unrecognized window, casting a precise shadow across the dusty wooden floor, illuminating a pair of aluminum medical crutches resting quietly against a mottled corner. Their gray rubber grips and armpit pads appear worn, and the gleaming, cold metal brackets bear slight scratches, silently telling a forgotten story. As the camera moves forward ever more slowly, an incredible process unfolds: the cold metal surface of the crutches begins to petrify inch by inch, its original luster gradually fading, replaced by a rough, grayish-white texture. Microscopic cracks creep from the contact points like a spiderweb, strikingly clear in the warm, golden light. Finally, when the camera lingers in close-up, the pair of crutches has completely transformed into a silent sculpture, covered with deep and poignant cracks, as if solidifying a heavy memory in an eternal form. The whole picture, with its sharp contrast of warm light and cold texture, creates a sad atmosphere of forgetfulness, silence and the solidification of time.", + "seed": 24578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_6936_seed_24578_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2622671049553901, + "vlm_logprob": -0.2622671049553901, + "vlm_yes_no_margin": 3.375000011920929, + "vlm_token_credit": -0.2622671049553901, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2263214588165283 + ], + "opposite_logprob_tokens": [ + -0.3513215184211731 + ], + "logprob_reward": -1.2263214588165283, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2263214588165283, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.040088851004838943, + 0.0300284530967474, + 0.038057226687669754, + 0.5017012357711792, + 0.07409942150115967, + 0.10153219848871231, + 0.04101637005805969, + 0.05150556564331055, + 0.08978273719549179, + 0.032187946140766144 + ], + "credit_map": [ + 0.005950918421149254, + 0.0041560223326087, + 0.005257313139736652, + 0.004618092905730009, + 0.002819397486746311, + 0.005366954021155834, + 0.003995685838162899, + 0.007924468256533146, + 0.003335080575197935, + 0.0035756018478423357, + 0.0036461791023612022, + 0.004895421676337719, + 0.0027317346539348364, + 0.004339579027146101, + 0.0032893079333007336, + 0.0042155468836426735, + 0.006348239723592997, + 0.0033027827739715576, + 0.005193494260311127, + 0.0037910686805844307, + 0.0038892661686986685, + 0.005814325530081987, + 0.004538212437182665, + 0.005179836880415678, + 0.1095808893442154, + 0.014741598628461361, + 0.17455202341079712, + 0.02616744674742222, + 0.01572711020708084, + 0.0409126840531826, + 0.022683536633849144, + 0.09733590483665466, + 0.0074135921895504, + 0.013440522365272045, + 0.006447598338127136, + 0.005564525723457336, + 0.01601000688970089, + 0.009422524832189083, + 0.007772018201649189, + 0.00802863109856844, + 0.010118677280843258, + 0.00865514948964119, + 0.012428049929440022, + 0.007856369949877262, + 0.019163621589541435, + 0.013172128237783909, + 0.017605317756533623, + 0.012532883323729038, + 0.0030820344109088182, + 0.0033406438305974007, + 0.005507758818566799, + 0.006647121161222458, + 0.009053720161318779, + 0.004448197782039642, + 0.005624603480100632, + 0.0033122890163213015, + 0.0037939732428640127, + 0.005005079787224531, + 0.010150689631700516, + 0.00488328980281949, + 0.006880846805870533, + 0.005196410696953535, + 0.006395511794835329, + 0.009199763648211956, + 0.008895962499082088, + 0.009822643361985683, + 0.014403310604393482, + 0.008660746738314629, + 0.016985954716801643, + 0.0071855392307043076, + 0.014531604945659637, + 0.009296978823840618, + 0.003513421630486846, + 0.004950577858835459, + 0.004711736459285021, + 0.003377453191205859, + 0.005159583408385515, + 0.003046578262001276, + 0.004833817947655916, + 0.00259477854706347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -1.2263214588165283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210923194885254 + ], + "opposite_logprob_tokens": [ + -2.8121092319488525 + ], + "logprob_reward": -0.06210923194885254, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210923194885254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1222287267446518, + 0.11134959757328033, + 0.06345710158348083, + 0.3102567791938782, + 0.08558446913957596, + 0.09690982848405838, + 0.051024891436100006, + 0.03608095273375511, + 0.06982627511024475, + 0.05328144133090973 + ], + "credit_map": [ + 0.015228920616209507, + 0.01654459536075592, + 0.016239354386925697, + 0.016745135188102722, + 0.010919558815658092, + 0.017079917713999748, + 0.012133299373090267, + 0.017337942495942116, + 0.014121541753411293, + 0.010857023298740387, + 0.0155123770236969, + 0.013076475821435452, + 0.007503564469516277, + 0.02201908268034458, + 0.01049338560551405, + 0.01776614971458912, + 0.011475890874862671, + 0.008189301937818527, + 0.008254839107394218, + 0.009129677899181843, + 0.0057106828317046165, + 0.0059594810009002686, + 0.006612891331315041, + 0.008124343119561672, + 0.07314399629831314, + 0.00945187732577324, + 0.1103387326002121, + 0.017333056777715683, + 0.008183754980564117, + 0.022864650934934616, + 0.01291151437908411, + 0.05602918937802315, + 0.010212923400104046, + 0.018891630694270134, + 0.009110881015658379, + 0.006068396847695112, + 0.017254196107387543, + 0.01046923827379942, + 0.0055923848412930965, + 0.00798481609672308, + 0.00946570374071598, + 0.008455020375549793, + 0.012121402658522129, + 0.012034227140247822, + 0.01720057800412178, + 0.013602962717413902, + 0.013724042102694511, + 0.01030589360743761, + 0.0061934455297887325, + 0.004586735740303993, + 0.006812296807765961, + 0.006664750631898642, + 0.0103943245485425, + 0.005000205244868994, + 0.007269260007888079, + 0.004103875253349543, + 0.0031693659257143736, + 0.004077264107763767, + 0.005788142792880535, + 0.003900486044585705, + 0.0066033098846673965, + 0.0030604235362261534, + 0.004023769870400429, + 0.005458191502839327, + 0.007397322915494442, + 0.007687471341341734, + 0.010714716278016567, + 0.00862470455467701, + 0.010741312056779861, + 0.008019465021789074, + 0.010011312551796436, + 0.00662996293976903, + 0.008230868726968765, + 0.006156462244689465, + 0.0074097514152526855, + 0.005172653589397669, + 0.007513265125453472, + 0.006422746926546097, + 0.005968809127807617, + 0.006406884640455246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06210923194885254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751700770109892 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751700770109892, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751700770109892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0600305013358593, + 0.04239148274064064, + 0.04783468320965767, + 0.34342530369758606, + 0.10673429071903229, + 0.13284792006015778, + 0.04800741747021675, + 0.053943321108818054, + 0.08981329202651978, + 0.0749717578291893 + ], + "credit_map": [ + 0.008294936269521713, + 0.009030984714627266, + 0.006542574148625135, + 0.006535657215863466, + 0.0064048245549201965, + 0.006525756325572729, + 0.008500217460095882, + 0.00819555576890707, + 0.004873694386333227, + 0.004188185092061758, + 0.0042160130105912685, + 0.004318247549235821, + 0.0035066523123532534, + 0.009012545458972454, + 0.004212530795484781, + 0.008063620887696743, + 0.006005463656038046, + 0.0060008978471159935, + 0.004237798508256674, + 0.009282082319259644, + 0.00488448329269886, + 0.0035660015419125557, + 0.0056321872398257256, + 0.00822577252984047, + 0.06727204471826553, + 0.010277306661009789, + 0.11452390998601913, + 0.021016903221607208, + 0.011253497563302517, + 0.029552990570664406, + 0.017865411937236786, + 0.07166324555873871, + 0.011061268858611584, + 0.02383064292371273, + 0.010658294893801212, + 0.0070872982032597065, + 0.0141189219430089, + 0.016029315069317818, + 0.011271046474575996, + 0.012677510268986225, + 0.010952264070510864, + 0.01032232865691185, + 0.015352779068052769, + 0.012982426211237907, + 0.024128250777721405, + 0.029053447768092155, + 0.013915153220295906, + 0.016141271218657494, + 0.0035500165540724993, + 0.0056897480972111225, + 0.008360340259969234, + 0.0050396667793393135, + 0.013002302497625351, + 0.004286124370992184, + 0.004974941723048687, + 0.003104281146079302, + 0.004451871383935213, + 0.005781400948762894, + 0.011117521673440933, + 0.004250604193657637, + 0.006910673342645168, + 0.005563726648688316, + 0.0059249806217849255, + 0.009942544624209404, + 0.007459939923137426, + 0.009377680718898773, + 0.015662293881177902, + 0.008367732167243958, + 0.01529756560921669, + 0.0074785430915653706, + 0.017017817124724388, + 0.009151719510555267, + 0.006711688358336687, + 0.0060290927067399025, + 0.01188577339053154, + 0.007847927510738373, + 0.010375289246439934, + 0.007657775189727545, + 0.010741224512457848, + 0.013722989708185196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.006751700770109892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036190745886415243 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036190745886415243, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036190745886415243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03906327486038208, + 0.024757377803325653, + 0.04195653274655342, + 0.4337080717086792, + 0.1878765970468521, + 0.05180004611611366, + 0.06232168897986412, + 0.05506015196442604, + 0.06457055360078812, + 0.03888566046953201 + ], + "credit_map": [ + 0.005126434378325939, + 0.0031253681518137455, + 0.0038757543079555035, + 0.006634885910898447, + 0.0023509878665208817, + 0.006107979919761419, + 0.004315649159252644, + 0.007526211440563202, + 0.002803520066663623, + 0.002832360565662384, + 0.0025166599079966545, + 0.0034335863310843706, + 0.0019843238405883312, + 0.004094661213457584, + 0.0030282095540314913, + 0.004064055159687996, + 0.008107699453830719, + 0.002503669122233987, + 0.006167296785861254, + 0.006795031949877739, + 0.002942189807072282, + 0.005663636606186628, + 0.003803215455263853, + 0.005973792169243097, + 0.10848083347082138, + 0.013242031447589397, + 0.15455996990203857, + 0.02332129515707493, + 0.011731400154531002, + 0.030333593487739563, + 0.01759585738182068, + 0.07444310188293457, + 0.01785552315413952, + 0.04020034521818161, + 0.017795026302337646, + 0.00967559963464737, + 0.04407157003879547, + 0.027630390599370003, + 0.01304000336676836, + 0.017608141526579857, + 0.004363314714282751, + 0.0044552977196872234, + 0.0052177635952830315, + 0.0038160935509949923, + 0.008356975391507149, + 0.013368678279221058, + 0.0055679818615317345, + 0.006653941236436367, + 0.0045898607932031155, + 0.005168246570974588, + 0.010923969559371471, + 0.008367019705474377, + 0.017853526398539543, + 0.0045237005688250065, + 0.006815781816840172, + 0.004079586826264858, + 0.0047537945210933685, + 0.005239536520093679, + 0.011474896222352982, + 0.004260025452822447, + 0.006970020476728678, + 0.004977068863809109, + 0.006973575334995985, + 0.010411235503852367, + 0.006203507538884878, + 0.007102055009454489, + 0.009559527039527893, + 0.01056786347180605, + 0.009247181005775928, + 0.004754611290991306, + 0.010124007239937782, + 0.0070117986761033535, + 0.0028367850463837385, + 0.004687457345426083, + 0.005481678526848555, + 0.004246218595653772, + 0.006924821995198727, + 0.004095633514225483, + 0.0060083866119384766, + 0.004604677204042673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.0036190745886415243 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534058652818203 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012534058652818203, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534058652818203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05688964203000069, + 0.023678036406636238, + 0.05527620390057564, + 0.2925766408443451, + 0.06389965116977692, + 0.08500351756811142, + 0.10149554908275604, + 0.046513330191373825, + 0.13409040868282318, + 0.14057697355747223 + ], + "credit_map": [ + 0.0072737568989396095, + 0.008828538469970226, + 0.0053783743642270565, + 0.007256339304149151, + 0.004616434685885906, + 0.007204727735370398, + 0.005947056692093611, + 0.010384412482380867, + 0.002582427579909563, + 0.0020518575329333544, + 0.0028398383874446154, + 0.0019925301894545555, + 0.0018453258089721203, + 0.005307648330926895, + 0.0027860235422849655, + 0.004272386431694031, + 0.012647629715502262, + 0.0032171986531466246, + 0.01005561463534832, + 0.006863797083497047, + 0.0042862012051045895, + 0.006626633461564779, + 0.0042343540117144585, + 0.007344773970544338, + 0.07531439512968063, + 0.009392398409545422, + 0.10356809943914413, + 0.014296215027570724, + 0.011103675700724125, + 0.02054079994559288, + 0.011122397147119045, + 0.04723867028951645, + 0.007587565574795008, + 0.010363566689193249, + 0.00788787193596363, + 0.004933541174978018, + 0.007968271151185036, + 0.009682601317763329, + 0.007756604813039303, + 0.007719631306827068, + 0.00692997220903635, + 0.005877960007637739, + 0.00826592929661274, + 0.006837678141891956, + 0.0135743273422122, + 0.022604545578360558, + 0.00884514395147562, + 0.01206796895712614, + 0.008380153216421604, + 0.008930470794439316, + 0.016200195997953415, + 0.011656549759209156, + 0.02822914533317089, + 0.00847933441400528, + 0.012235937640070915, + 0.007383761927485466, + 0.005512280389666557, + 0.004458679351955652, + 0.008827181532979012, + 0.004423655103892088, + 0.005372828803956509, + 0.0051213097758591175, + 0.003847630461677909, + 0.008949765004217625, + 0.00949557963758707, + 0.012452209368348122, + 0.021961243823170662, + 0.016275394707918167, + 0.02287282422184944, + 0.009959603659808636, + 0.029578614979982376, + 0.011494931764900684, + 0.012152699753642082, + 0.010084503330290318, + 0.025268767029047012, + 0.01887199841439724, + 0.017762543633580208, + 0.009794640354812145, + 0.019515635445713997, + 0.027126183733344078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.012534058652818203 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing, empty shot captures a corner of a bare room. A warm, slanting beam of sunset light filters through an unrecognized window, casting a precise shadow across the dusty wooden floor, illuminating a pair of aluminum medical crutches resting quietly against a mottled corner. Their gray rubber grips and armpit pads appear worn, and the gleaming, cold metal brackets bear slight scratches, silently telling a forgotten story. As the camera moves forward ever more slowly, an incredible process unfolds: the cold metal surface of the crutches begins to petrify inch by inch, its original luster gradually fading, replaced by a rough, grayish-white texture. Microscopic cracks creep from the contact points like a spiderweb, strikingly clear in the warm, golden light. Finally, when the camera lingers in close-up, the pair of crutches has completely transformed into a silent sculpture, covered with deep and poignant cracks, as if solidifying a heavy memory in an eternal form. The whole picture, with its sharp contrast of warm light and cold texture, creates a sad atmosphere of forgetfulness, silence and the solidification of time.", + "seed": 24579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_10/index_6936_seed_24579_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1730561071773991, + "vlm_logprob": -0.1730561071773991, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.1730561071773991, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.831224799156189 + ], + "opposite_logprob_tokens": [ + -0.581224799156189 + ], + "logprob_reward": -0.831224799156189, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.831224799156189, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17379792034626007, + 0.021656159311532974, + 0.057172276079654694, + 0.019112873822450638, + 0.02346617728471756, + 0.31502506136894226, + 0.1263398677110672, + 0.1707400679588318, + 0.05314076319336891, + 0.0395488440990448 + ], + "credit_map": [ + 0.030048979446291924, + 0.015483258292078972, + 0.02710334025323391, + 0.018855862319469452, + 0.011752673424780369, + 0.011769709177315235, + 0.026458293199539185, + 0.032325804233551025, + 0.004047133028507233, + 0.0023192018270492554, + 0.0037068426609039307, + 0.0029875212348997593, + 0.0019694757647812366, + 0.0014915155479684472, + 0.0021928322967141867, + 0.002941635437309742, + 0.004363971762359142, + 0.003934667911380529, + 0.006317295599728823, + 0.005189272575080395, + 0.0036033436190336943, + 0.00729813938960433, + 0.010640375316143036, + 0.01582520455121994, + 0.002443572971969843, + 0.0027270568534731865, + 0.003202501218765974, + 0.003444394562393427, + 0.0020686842035502195, + 0.0011114338412880898, + 0.002356437034904957, + 0.0017587925540283322, + 0.003118757624179125, + 0.0029969527386128902, + 0.0039901877753436565, + 0.003232849296182394, + 0.0034499880857765675, + 0.0020030320156365633, + 0.002560289576649666, + 0.002114120637997985, + 0.033248040825128555, + 0.1257733255624771, + 0.03139496594667435, + 0.031133992597460747, + 0.0298057422041893, + 0.01664884388446808, + 0.025229472666978836, + 0.02179068885743618, + 0.014072608202695847, + 0.01556384563446045, + 0.0189799964427948, + 0.02681124582886696, + 0.013632956892251968, + 0.009243039414286613, + 0.013964042998850346, + 0.014072127640247345, + 0.01538650132715702, + 0.01374759990721941, + 0.05098586156964302, + 0.04202898591756821, + 0.013062477111816406, + 0.008436916396021843, + 0.015748636797070503, + 0.011343104764819145, + 0.005165994167327881, + 0.005979303270578384, + 0.011320460587739944, + 0.010795225389301777, + 0.004069071263074875, + 0.003748389193788171, + 0.007966944947838783, + 0.004095376003533602, + 0.0028633365873247385, + 0.00865462701767683, + 0.003666137345135212, + 0.005709260702133179, + 0.004803488031029701, + 0.0036060207057744265, + 0.0064910100772976875, + 0.003754962468519807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.831224799156189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020610801875591278 + ], + "opposite_logprob_tokens": [ + -3.895610809326172 + ], + "logprob_reward": -0.020610801875591278, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020610801875591278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1481594741344452, + 0.062239136546850204, + 0.07688073068857193, + 0.04409680888056755, + 0.35687634348869324, + 0.07630601525306702, + 0.052007947117090225, + 0.07373329997062683, + 0.05667072534561157, + 0.05302952602505684 + ], + "credit_map": [ + 0.01999698393046856, + 0.01666710153222084, + 0.021555321291089058, + 0.022682225331664085, + 0.011832119897007942, + 0.010827096179127693, + 0.020556792616844177, + 0.024041837081313133, + 0.004507385194301605, + 0.005206437315791845, + 0.008662635460495949, + 0.005453831981867552, + 0.003029686864465475, + 0.003160889260470867, + 0.011164958588778973, + 0.021053310483694077, + 0.006437223870307207, + 0.007439008913934231, + 0.008902314119040966, + 0.01144173089414835, + 0.008125828579068184, + 0.007431321777403355, + 0.015651198104023933, + 0.011452103964984417, + 0.005513949319720268, + 0.004451121669262648, + 0.005726383533328772, + 0.006114589050412178, + 0.004734090529382229, + 0.0053783804178237915, + 0.00682047801092267, + 0.005357817281037569, + 0.04223906621336937, + 0.046177878975868225, + 0.042148884385824203, + 0.03827338293194771, + 0.07099815458059311, + 0.04449602961540222, + 0.03871840238571167, + 0.033824533224105835, + 0.007018322125077248, + 0.021107740700244904, + 0.011588740162551403, + 0.006255779881030321, + 0.007065309211611748, + 0.006417520344257355, + 0.0097190598025918, + 0.007133543491363525, + 0.006796093191951513, + 0.006564790848642588, + 0.010069235228002071, + 0.007803276646882296, + 0.005923292133957148, + 0.004487321712076664, + 0.004903995431959629, + 0.005459940060973167, + 0.007962757721543312, + 0.00789093691855669, + 0.020845353603363037, + 0.01215214654803276, + 0.008295439183712006, + 0.006152222864329815, + 0.006438462529331446, + 0.003995982464402914, + 0.005546161904931068, + 0.0045500388368964195, + 0.00842051487416029, + 0.009858781471848488, + 0.008397193625569344, + 0.0055122594349086285, + 0.009841522201895714, + 0.0045442525297403336, + 0.004411470610648394, + 0.010259617120027542, + 0.005445072427392006, + 0.008879413828253746, + 0.005850919988006353, + 0.0038667728658765554, + 0.010584247298538685, + 0.00373200885951519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.020610801875591278 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003629170823842287 + ], + "opposite_logprob_tokens": [ + -5.628629207611084 + ], + "logprob_reward": -0.003629170823842287, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003629170823842287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1614028364419937, + 0.15497735142707825, + 0.06950388848781586, + 0.054178398102521896, + 0.04624989256262779, + 0.3123054504394531, + 0.041476231068372726, + 0.049160946160554886, + 0.0689973309636116, + 0.041747696697711945 + ], + "credit_map": [ + 0.02773907594382763, + 0.019163833931088448, + 0.02582034096121788, + 0.02238817699253559, + 0.01281861960887909, + 0.011204227805137634, + 0.01836121268570423, + 0.023907355964183807, + 0.014974230900406837, + 0.018143191933631897, + 0.023336755111813545, + 0.02174580842256546, + 0.008632595650851727, + 0.007361162919551134, + 0.022037317976355553, + 0.038746293634176254, + 0.007345291785895824, + 0.00739121763035655, + 0.007402045652270317, + 0.008642373606562614, + 0.006182510871440172, + 0.0074620116502046585, + 0.012000106275081635, + 0.013078337535262108, + 0.007075897417962551, + 0.0074925050139427185, + 0.010692622512578964, + 0.010199270211160183, + 0.003975408151745796, + 0.002290119417011738, + 0.006750172935426235, + 0.005702408961951733, + 0.006453266367316246, + 0.0045532300136983395, + 0.008373650722205639, + 0.004448402673006058, + 0.007117105647921562, + 0.004379833582788706, + 0.00628959946334362, + 0.004634808283299208, + 0.03155836462974548, + 0.11958326399326324, + 0.043363526463508606, + 0.02506856434047222, + 0.029082486405968666, + 0.017423829063773155, + 0.026961112394928932, + 0.019264336675405502, + 0.0038140746764838696, + 0.005021861754357815, + 0.006575551349669695, + 0.006916968151926994, + 0.004629603587090969, + 0.00397919025272131, + 0.004532522056251764, + 0.006006461102515459, + 0.0038743363693356514, + 0.0050589051097631454, + 0.013965834863483906, + 0.010247671976685524, + 0.003431385848671198, + 0.003161756554618478, + 0.005532484967261553, + 0.0038885781541466713, + 0.006528103724122047, + 0.006726960651576519, + 0.008876354433596134, + 0.01317930780351162, + 0.007107348181307316, + 0.009523537941277027, + 0.00959024392068386, + 0.007465489208698273, + 0.004666138906031847, + 0.005771637428551912, + 0.004205562174320221, + 0.00620019156485796, + 0.005467326380312443, + 0.003866208251565695, + 0.007347505073994398, + 0.004223132040351629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.003629170823842287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011822147062048316 + ], + "opposite_logprob_tokens": [ + -6.7511820793151855 + ], + "logprob_reward": -0.0011822147062048316, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011822147062048316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10616900026798248, + 0.041837092489004135, + 0.057591378688812256, + 0.04464399442076683, + 0.03845749422907829, + 0.3483273386955261, + 0.12336619198322296, + 0.08006718754768372, + 0.10522446036338806, + 0.05431581288576126 + ], + "credit_map": [ + 0.023464035242795944, + 0.012341439723968506, + 0.01613861694931984, + 0.009615002200007439, + 0.006483591627329588, + 0.008012352511286736, + 0.011807327158749104, + 0.01830664835870266, + 0.0043260385282337666, + 0.004342252854257822, + 0.006089163478463888, + 0.003889101091772318, + 0.002101167570799589, + 0.0028487006202340126, + 0.006880671251565218, + 0.011360003612935543, + 0.006455133203417063, + 0.008765839971601963, + 0.008116752840578556, + 0.006667117588222027, + 0.0048341997899115086, + 0.007959883660078049, + 0.007106950506567955, + 0.007685509044677019, + 0.005836631637066603, + 0.006927827838808298, + 0.007873207330703735, + 0.00787150114774704, + 0.0037342493887990713, + 0.002727205865085125, + 0.004157663322985172, + 0.005515714641660452, + 0.005494318436831236, + 0.004256904125213623, + 0.0070234364829957485, + 0.0050542657263576984, + 0.004473915323615074, + 0.0037497668527066708, + 0.004431203473359346, + 0.0039736866019666195, + 0.04282306507229805, + 0.13537192344665527, + 0.020471584051847458, + 0.036434803158044815, + 0.03833405300974846, + 0.02113652229309082, + 0.03443954512476921, + 0.019315898418426514, + 0.013372761197388172, + 0.01596766524016857, + 0.019445234909653664, + 0.02320941723883152, + 0.009847458451986313, + 0.012572597712278366, + 0.01327089499682188, + 0.015680180862545967, + 0.008499491959810257, + 0.01062669325619936, + 0.015531332232058048, + 0.014275187626481056, + 0.008190516382455826, + 0.007846799679100513, + 0.009128492325544357, + 0.005968678742647171, + 0.009144386276602745, + 0.008576156571507454, + 0.026034966111183167, + 0.02603420987725258, + 0.007848121225833893, + 0.006065915338695049, + 0.014012581668794155, + 0.007508128881454468, + 0.006681147497147322, + 0.007141842041164637, + 0.005477385129779577, + 0.008076172322034836, + 0.007674454711377621, + 0.005012118257582188, + 0.007969733327627182, + 0.0062829600647091866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0011822147062048316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty corner of a bare room with a dusty wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warm, slanting beam of sunset light visible filtering through a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of aluminum medical crutches with gray rubber grips and armpit pads visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crutch metal surface gradually transform from gleaming cold metal to a rough, grayish-white petrified texture with cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move forward slowly and end in a close-up shot of the transformed crutches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633549325168133 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633549325168133, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633549325168133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048185158520936966, + 0.04028550535440445, + 0.15884099900722504, + 0.027828695252537727, + 0.02812756970524788, + 0.31825166940689087, + 0.1916002333164215, + 0.06627383828163147, + 0.07779423892498016, + 0.04281212389469147 + ], + "credit_map": [ + 0.005933536682277918, + 0.01064350176602602, + 0.008511279709637165, + 0.00601274473592639, + 0.0030506066977977753, + 0.0026674065738916397, + 0.0037664039991796017, + 0.007599672768265009, + 0.004092648159712553, + 0.0036119227297604084, + 0.005358456633985043, + 0.004301166627556086, + 0.0023319250904023647, + 0.0025949114933609962, + 0.006563081871718168, + 0.01143138948827982, + 0.012395823374390602, + 0.010728014633059502, + 0.016532858833670616, + 0.020844554528594017, + 0.009276771917939186, + 0.011374414898455143, + 0.03542577475309372, + 0.04226277023553848, + 0.003313259454444051, + 0.003636902431026101, + 0.004347397014498711, + 0.005518678110092878, + 0.0030401868280023336, + 0.0025310395285487175, + 0.002840435830876231, + 0.0026007932610809803, + 0.003316919319331646, + 0.0033711267169564962, + 0.004050017800182104, + 0.0034453850239515305, + 0.003470429452136159, + 0.0021228333935141563, + 0.0048694550059735775, + 0.0034814015962183475, + 0.0326155349612236, + 0.13718578219413757, + 0.0408870093524456, + 0.029715223237872124, + 0.028872303664684296, + 0.01651163399219513, + 0.02034643664956093, + 0.012117737904191017, + 0.019387759268283844, + 0.02427526004612446, + 0.03228786587715149, + 0.040788374841213226, + 0.025275159627199173, + 0.012129693292081356, + 0.01846720091998577, + 0.018988898023962975, + 0.0070947278290987015, + 0.006281007546931505, + 0.016291862353682518, + 0.015010041184723377, + 0.0062210774049162865, + 0.004826697055250406, + 0.0060124085284769535, + 0.004536005202680826, + 0.009873162023723125, + 0.00817611999809742, + 0.014454931020736694, + 0.014062339439988136, + 0.00789659284055233, + 0.00917532853782177, + 0.008700569160282612, + 0.005455187987536192, + 0.003695051185786724, + 0.009205132722854614, + 0.0042573390528559685, + 0.007038758136332035, + 0.005949296988546848, + 0.0033081138972193003, + 0.006763963494449854, + 0.002594468416646123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.008633549325168133 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_11/index_tensor([9452, 9452])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_11/index_tensor([9452, 9452])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..968df3ec37e96efbc5f66803b51b2aff99870609 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_11/index_tensor([9452, 9452])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static shot, shot from a horizontal angle, yet accompanied by constant, high-frequency micro-vibrations due to the train's breathtaking speed, expertly and directly captures the interior of a high-speed maglev train. The interior showcases a minimalist, futuristic design, with ergonomically designed, streamlined seats constructed of a light gray, high-tech composite material, neatly arranged. Bright sunlight streams through the large picture windows, blending with the soft, even cool white light emitted by LED strips embedded in the ceiling, illuminating the nearly invisible dust particles in the air. The rural landscape outside the windows is reduced to a blur of rapidly passing dynamic streaks, with only swift flashes of green and brown. The immense power of the train generates a deep resonance throughout the train, transforming into a barely perceptible vibration that reflects fine, flowing light off the metal handrails and window glass, creating a unique atmosphere of cutting-edge technology, extreme speed, and cool efficiency.", + "seed": 7773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_11/index_9452_seed_7773_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08317736107856036, + "vlm_logprob": -0.08317736107856036, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.08317736107856036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15897585451602936 + ], + "opposite_logprob_tokens": [ + -2.283975839614868 + ], + "logprob_reward": -0.15897585451602936, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.15897585451602936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2434467375278473, + 0.06638462841510773, + 0.07053949683904648, + 0.035923026502132416, + 0.03517420217394829, + 0.11339160799980164, + 0.03346547484397888, + 0.08300874382257462, + 0.20928609371185303, + 0.10938003659248352 + ], + "credit_map": [ + 0.01911163330078125, + 0.01231993455439806, + 0.020738255232572556, + 0.030454101040959358, + 0.015237411484122276, + 0.04775078594684601, + 0.07576394081115723, + 0.022070636972784996, + 0.00419958820566535, + 0.007099242880940437, + 0.008172380737960339, + 0.008619108237326145, + 0.0056277187541127205, + 0.008125805296003819, + 0.01496679987758398, + 0.00957398023456335, + 0.004345719236880541, + 0.0051157549023628235, + 0.009595349431037903, + 0.020663168281316757, + 0.0059960270300507545, + 0.007231866009533405, + 0.008884463459253311, + 0.008707139641046524, + 0.0029208853375166655, + 0.003111274214461446, + 0.006710807327181101, + 0.005431766621768475, + 0.002075913129374385, + 0.003382681868970394, + 0.006044187117367983, + 0.006245506927371025, + 0.0033692128490656614, + 0.0063763633370399475, + 0.0030239252373576164, + 0.006344819441437721, + 0.0023129405453801155, + 0.00198484119027853, + 0.007043893914669752, + 0.0047182017005980015, + 0.007939840666949749, + 0.013119738548994064, + 0.008084588684141636, + 0.017565827816724777, + 0.02313939481973648, + 0.015039972960948944, + 0.019443880766630173, + 0.00905835721641779, + 0.0030912114307284355, + 0.0020361943170428276, + 0.005750157404690981, + 0.007231652271002531, + 0.0027836174704134464, + 0.0026857962366193533, + 0.003981507383286953, + 0.0059053367003798485, + 0.005109755787998438, + 0.008189191110432148, + 0.007503196131438017, + 0.009235204197466373, + 0.011678089387714863, + 0.006076069548726082, + 0.015352919697761536, + 0.019864309579133987, + 0.01110632624477148, + 0.006141184829175472, + 0.008461261168122292, + 0.009565901011228561, + 0.007887083105742931, + 0.007367353420704603, + 0.0892292857170105, + 0.06952769309282303, + 0.009333123452961445, + 0.006971851456910372, + 0.010705835185945034, + 0.014508080668747425, + 0.00765142822638154, + 0.010973784141242504, + 0.034152086824178696, + 0.015083834528923035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.15897585451602936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043119918555021286 + ], + "opposite_logprob_tokens": [ + -3.1681199073791504 + ], + "logprob_reward": -0.043119918555021286, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043119918555021286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.254629522562027, + 0.03075447678565979, + 0.04771824926137924, + 0.06902345269918442, + 0.05408203974366188, + 0.08936865627765656, + 0.027170928195118904, + 0.05122091993689537, + 0.29113519191741943, + 0.08489659428596497 + ], + "credit_map": [ + 0.012173122726380825, + 0.013500477187335491, + 0.016222819685935974, + 0.03362920135259628, + 0.02944406308233738, + 0.046253275126218796, + 0.07810082286596298, + 0.02530573308467865, + 0.002667340449988842, + 0.0032403937075287104, + 0.00406407518312335, + 0.003760985331609845, + 0.0036131301894783974, + 0.0028976767789572477, + 0.0059722671285271645, + 0.00453860592097044, + 0.003117663785815239, + 0.003872315865010023, + 0.0058984821662306786, + 0.014508204534649849, + 0.0027204661164432764, + 0.00511562405154109, + 0.006269706413149834, + 0.0062157888896763325, + 0.004969998262822628, + 0.005907359998673201, + 0.006523624062538147, + 0.015129609033465385, + 0.005478228442370892, + 0.005435822997242212, + 0.01690809056162834, + 0.008670718409121037, + 0.003939296584576368, + 0.0042259874753654, + 0.004621697124093771, + 0.01350440364331007, + 0.003212468232959509, + 0.0023910196032375097, + 0.01670045033097267, + 0.00548672117292881, + 0.006673344876617193, + 0.012363242916762829, + 0.006976582575589418, + 0.010105618275702, + 0.02484632097184658, + 0.007954584434628487, + 0.012667923234403133, + 0.007781036663800478, + 0.0022072407882660627, + 0.001527664135210216, + 0.0045415377244353294, + 0.00598379410803318, + 0.0018337239744141698, + 0.002042659791186452, + 0.0035044190008193254, + 0.005529890768229961, + 0.002922327723354101, + 0.0050511350855231285, + 0.005559000186622143, + 0.004923341795802116, + 0.005973299033939838, + 0.003000749507918954, + 0.010906266048550606, + 0.012884797528386116, + 0.01578356698155403, + 0.007662895135581493, + 0.010977924801409245, + 0.012635140679776669, + 0.011198612861335278, + 0.00789880845695734, + 0.12899018824100494, + 0.09598806500434875, + 0.006665547378361225, + 0.00739953713491559, + 0.00800320878624916, + 0.009267610497772694, + 0.012511853128671646, + 0.008527626283466816, + 0.022906752303242683, + 0.009614454582333565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.043119918555021286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110340960323811 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110340960323811, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110340960323811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13720303773880005, + 0.3070775866508484, + 0.10104850679636002, + 0.03696481138467789, + 0.12639902532100677, + 0.05122370645403862, + 0.019944924861192703, + 0.03648807480931282, + 0.031254347413778305, + 0.15239593386650085 + ], + "credit_map": [ + 0.012773176655173302, + 0.007269520778208971, + 0.010723046958446503, + 0.01690215803682804, + 0.012315783649682999, + 0.02424410916864872, + 0.03766128048300743, + 0.015313982963562012, + 0.023898493498563766, + 0.04479958117008209, + 0.03221071884036064, + 0.04047564044594765, + 0.05188433825969696, + 0.025317799299955368, + 0.05838824808597565, + 0.03010280802845955, + 0.0060546123422682285, + 0.007420276757329702, + 0.008974438533186913, + 0.019841743633151054, + 0.018252940848469734, + 0.013162901625037193, + 0.011636200360953808, + 0.015705404803156853, + 0.002952749375253916, + 0.0033690426498651505, + 0.004442234989255667, + 0.006118100136518478, + 0.0037925622891634703, + 0.003413355676457286, + 0.00883396714925766, + 0.004042803309857845, + 0.013054451905190945, + 0.02210872620344162, + 0.013367168605327606, + 0.02010699361562729, + 0.012604297138750553, + 0.00664121937006712, + 0.018947744742035866, + 0.01956843212246895, + 0.0037997066974639893, + 0.004576484207063913, + 0.0054574026726186275, + 0.013097556307911873, + 0.009081568568944931, + 0.0037827747873961926, + 0.006926253437995911, + 0.004501963034272194, + 0.0017196544213220477, + 0.0018715516198426485, + 0.003386810654774308, + 0.003167626913636923, + 0.0018648148979991674, + 0.001384215080179274, + 0.0023220358416438103, + 0.004228217527270317, + 0.0025906378868967295, + 0.003649171208962798, + 0.00406982097774744, + 0.0043645864352583885, + 0.004774461500346661, + 0.0023479473311454058, + 0.007497135549783707, + 0.007194317877292633, + 0.0037605674006044865, + 0.004180471878498793, + 0.004457384813576937, + 0.0025396074634045362, + 0.0036149248480796814, + 0.0031453650444746017, + 0.0051815747283399105, + 0.004374453332275152, + 0.014834584668278694, + 0.009028278291225433, + 0.01369544118642807, + 0.024367989972233772, + 0.01101821567863226, + 0.015103651210665703, + 0.0442962720990181, + 0.0200515054166317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.004110340960323811 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794618487358093 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794618487358093, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794618487358093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13146445155143738, + 0.25175315141677856, + 0.0947311520576477, + 0.033062465488910675, + 0.0334569476544857, + 0.0461665578186512, + 0.02397391013801098, + 0.012937759980559349, + 0.347563236951828, + 0.024890311062335968 + ], + "credit_map": [ + 0.010323568247258663, + 0.008021574467420578, + 0.013155265711247921, + 0.014311996288597584, + 0.0070725386030972, + 0.01988573558628559, + 0.036756254732608795, + 0.021937530487775803, + 0.01944609358906746, + 0.023766154423356056, + 0.025577252730727196, + 0.03303183615207672, + 0.019691601395606995, + 0.03390773385763168, + 0.057263582944869995, + 0.03906891494989395, + 0.00666813924908638, + 0.007766369264572859, + 0.012184835970401764, + 0.01827043667435646, + 0.016358299180865288, + 0.007807902991771698, + 0.012580729089677334, + 0.013094439171254635, + 0.002493640873581171, + 0.0027541134040802717, + 0.004350139293819666, + 0.005384945776313543, + 0.002745966427028179, + 0.0032738815061748028, + 0.007766281720250845, + 0.004293498117476702, + 0.003618113696575165, + 0.005815358366817236, + 0.004820208065211773, + 0.005956197157502174, + 0.0031484588980674744, + 0.0016179136000573635, + 0.005088150035589933, + 0.003392549231648445, + 0.002499251626431942, + 0.0056321038864552975, + 0.0036573426332324743, + 0.007669327314943075, + 0.013246427290141582, + 0.0038876174949109554, + 0.006351496558636427, + 0.0032229896169155836, + 0.0015920582227408886, + 0.0017302149208262563, + 0.004546889569610357, + 0.004151926375925541, + 0.002489856444299221, + 0.0016797618009150028, + 0.002852061064913869, + 0.004931141622364521, + 0.0012738211080431938, + 0.001242210972122848, + 0.002185400342568755, + 0.0022613792680203915, + 0.0011201053857803345, + 0.0007204879075288773, + 0.00281896092928946, + 0.001315394532866776, + 0.019762899726629257, + 0.008954761549830437, + 0.013141357339918613, + 0.015701774507761, + 0.0135661531239748, + 0.009205047972500324, + 0.15679116547107697, + 0.11044006049633026, + 0.0023098543751984835, + 0.0015520913293585181, + 0.002092845505103469, + 0.004128618631511927, + 0.002763147931545973, + 0.0021705699618905783, + 0.006421762052923441, + 0.003451421158388257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.029794618487358093 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798860728740692 + ], + "opposite_logprob_tokens": [ + -1.8048861026763916 + ], + "logprob_reward": -0.1798860728740692, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798860728740692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09494510293006897, + 0.015493053011596203, + 0.01723618432879448, + 0.01200924627482891, + 0.015196261927485466, + 0.024058161303400993, + 0.00742859672755003, + 0.01957639865577221, + 0.7722010016441345, + 0.021856006234884262 + ], + "credit_map": [ + 0.005875930190086365, + 0.0066415537148714066, + 0.008439083583652973, + 0.010641909204423428, + 0.014972374774515629, + 0.0098428251221776, + 0.020970938727259636, + 0.01756049133837223, + 0.0011776095489040017, + 0.001061552087776363, + 0.002222688402980566, + 0.0016751751536503434, + 0.002008941723033786, + 0.0021775106433779, + 0.003136146580800414, + 0.002033428754657507, + 0.0011734538711607456, + 0.001024672412313521, + 0.0023916850332170725, + 0.00429163221269846, + 0.0011664838530123234, + 0.001540356664918363, + 0.0024427168536931276, + 0.0032051855232566595, + 0.0009176282910630107, + 0.0012204806553199887, + 0.00163921935018152, + 0.0018033151281997561, + 0.0010414059506729245, + 0.001058113295584917, + 0.002201997209340334, + 0.0021270865108817816, + 0.0016305774915963411, + 0.0017417363123968244, + 0.0026818623300641775, + 0.0026738785672932863, + 0.0010817323345690966, + 0.0009217889164574444, + 0.00254584732465446, + 0.001918840454891324, + 0.0018926794873550534, + 0.0013418117305263877, + 0.0016129163559526205, + 0.002245373325422406, + 0.0012469306821003556, + 0.005466626491397619, + 0.008411153219640255, + 0.0018406723393127322, + 0.0005181838641874492, + 0.000693461683113128, + 0.0011177388951182365, + 0.001455987337976694, + 0.0009834475349634886, + 0.0005914143403060734, + 0.001087266718968749, + 0.0009810971096158028, + 0.0010152148315683007, + 0.0017872436437755823, + 0.0018576643196865916, + 0.0022942633368074894, + 0.002606133697554469, + 0.001298696850426495, + 0.0038038622587919235, + 0.004913320764899254, + 0.042467404156923294, + 0.020708395168185234, + 0.028541943058371544, + 0.03409182280302048, + 0.030112385749816895, + 0.019802801311016083, + 0.34517690539360046, + 0.2512993812561035, + 0.0022969318088144064, + 0.0013900203630328178, + 0.002097135176882148, + 0.004036827944219112, + 0.00190881232265383, + 0.0017684795893728733, + 0.00566650228574872, + 0.0026912973262369633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.1798860728740692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot from a horizontal angle, yet accompanied by constant, high-frequency micro-vibrations due to the train's breathtaking speed, expertly and directly captures the interior of a high-speed maglev train. The interior showcases a minimalist, futuristic design, with ergonomically designed, streamlined seats constructed of a light gray, high-tech composite material, neatly arranged. Bright sunlight streams through the large picture windows, blending with the soft, even cool white light emitted by LED strips embedded in the ceiling, illuminating the nearly invisible dust particles in the air. The rural landscape outside the windows is reduced to a blur of rapidly passing dynamic streaks, with only swift flashes of green and brown. The immense power of the train generates a deep resonance throughout the train, transforming into a barely perceptible vibration that reflects fine, flowing light off the metal handrails and window glass, creating a unique atmosphere of cutting-edge technology, extreme speed, and cool efficiency.", + "seed": 7774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_11/index_9452_seed_7774_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.224356946721673, + "vlm_logprob": -0.224356946721673, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.224356946721673, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.49346858263015747 + ], + "opposite_logprob_tokens": [ + -0.9934685826301575 + ], + "logprob_reward": -0.49346858263015747, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.49346858263015747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2240176498889923, + 0.07466863095760345, + 0.0860993042588234, + 0.11043310910463333, + 0.10720590502023697, + 0.1519133299589157, + 0.08725874125957489, + 0.059486258774995804, + 0.05954934284090996, + 0.0393676832318306 + ], + "credit_map": [ + 0.019581960514187813, + 0.027133801952004433, + 0.032798074185848236, + 0.01862289570271969, + 0.02380402386188507, + 0.031996577978134155, + 0.03115607611835003, + 0.03892423212528229, + 0.009685853496193886, + 0.008976061828434467, + 0.011038396507501602, + 0.006275119259953499, + 0.010134373791515827, + 0.00813826359808445, + 0.00886830035597086, + 0.01155226118862629, + 0.012708174996078014, + 0.01257838774472475, + 0.009668209590017796, + 0.006407448556274176, + 0.010364037938416004, + 0.016503117978572845, + 0.0070689222775399685, + 0.010801001451909542, + 0.014450989663600922, + 0.0162531565874815, + 0.01763979345560074, + 0.010024919174611568, + 0.01618344895541668, + 0.01230061799287796, + 0.013033879920840263, + 0.010546297766268253, + 0.008622894063591957, + 0.01486954465508461, + 0.018147893249988556, + 0.0195806585252285, + 0.01149227749556303, + 0.016245581209659576, + 0.008179466240108013, + 0.010067594237625599, + 0.01233394630253315, + 0.022008506581187248, + 0.024922935292124748, + 0.017682675272226334, + 0.014691338874399662, + 0.019370514899492264, + 0.01643042452633381, + 0.024472983554005623, + 0.008373218588531017, + 0.01123877614736557, + 0.011053104884922504, + 0.010324827395379543, + 0.008406855165958405, + 0.010274664498865604, + 0.01519924495369196, + 0.012388047762215137, + 0.007196150254458189, + 0.00824003480374813, + 0.007427229080349207, + 0.007844250649213791, + 0.00836166925728321, + 0.006121874321252108, + 0.005703039932996035, + 0.008592008613049984, + 0.006676855962723494, + 0.007453951518982649, + 0.010856924578547478, + 0.005147870630025864, + 0.005902907811105251, + 0.009027771651744843, + 0.007369783706963062, + 0.007113280240446329, + 0.004040118306875229, + 0.00784389115869999, + 0.005232747178524733, + 0.003572791814804077, + 0.004525004420429468, + 0.005179442930966616, + 0.004395024385303259, + 0.004578664433211088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.49346858263015747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14305713772773743 + ], + "opposite_logprob_tokens": [ + -2.018057107925415 + ], + "logprob_reward": -0.14305713772773743, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14305713772773743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.331755131483078, + 0.05818721652030945, + 0.10557924956083298, + 0.1315138190984726, + 0.08868815749883652, + 0.02240275777876377, + 0.05179080739617348, + 0.07423731684684753, + 0.09822700172662735, + 0.037618573755025864 + ], + "credit_map": [ + 0.032879024744033813, + 0.05024627596139908, + 0.06692212074995041, + 0.02701549232006073, + 0.03787049278616905, + 0.03566405922174454, + 0.040350157767534256, + 0.04080748185515404, + 0.006134450901299715, + 0.00839904136955738, + 0.006951606832444668, + 0.004310953430831432, + 0.011225746013224125, + 0.006758348550647497, + 0.00683360593393445, + 0.007573462091386318, + 0.019160401076078415, + 0.012905721552670002, + 0.011672060936689377, + 0.006211840081959963, + 0.014554179273545742, + 0.020349252969026566, + 0.008335402235388756, + 0.012390385381877422, + 0.012875143438577652, + 0.02207759954035282, + 0.015574682503938675, + 0.012340129353106022, + 0.021806467324495316, + 0.017554128542542458, + 0.015532690100371838, + 0.013752974569797516, + 0.006648636423051357, + 0.012257931753993034, + 0.012955420650541782, + 0.012009412050247192, + 0.012279605492949486, + 0.013432716019451618, + 0.009233187884092331, + 0.009871245361864567, + 0.0025998372584581375, + 0.003131608245894313, + 0.002797120250761509, + 0.002341846702620387, + 0.002351515693590045, + 0.002830860670655966, + 0.0023064794950187206, + 0.004043488763272762, + 0.003227145643904805, + 0.007176779210567474, + 0.003921994939446449, + 0.004342997446656227, + 0.008179398253560066, + 0.008199741132557392, + 0.008704082109034061, + 0.008038668893277645, + 0.008873946964740753, + 0.0096439179033041, + 0.009023047983646393, + 0.012720488011837006, + 0.009115934371948242, + 0.007233559153974056, + 0.009971069172024727, + 0.007655342109501362, + 0.009002361446619034, + 0.013044631108641624, + 0.015237060375511646, + 0.0069497027434408665, + 0.012166105210781097, + 0.017204612493515015, + 0.009667567908763885, + 0.014954952523112297, + 0.0023087195586413145, + 0.009236697107553482, + 0.0032251253724098206, + 0.004616851918399334, + 0.010651953518390656, + 0.003182795597240329, + 0.0018361062975600362, + 0.00256032170727849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.14305713772773743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038126446306705475 + ], + "opposite_logprob_tokens": [ + -3.2881264686584473 + ], + "logprob_reward": -0.038126446306705475, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038126446306705475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29182541370391846, + 0.06927162408828735, + 0.07804462313652039, + 0.08050297200679779, + 0.0897015780210495, + 0.03885721415281296, + 0.15315943956375122, + 0.10334373265504837, + 0.037722934037446976, + 0.05757051333785057 + ], + "credit_map": [ + 0.02257642336189747, + 0.04613982513546944, + 0.026173090562224388, + 0.022475240752100945, + 0.048163510859012604, + 0.03784681484103203, + 0.05154600366950035, + 0.036904506385326385, + 0.008073784410953522, + 0.009203411638736725, + 0.008433928713202477, + 0.007140226662158966, + 0.009474614635109901, + 0.007487314287573099, + 0.009491419419646263, + 0.00996692106127739, + 0.006621996872127056, + 0.010660937055945396, + 0.00896829180419445, + 0.004833472892642021, + 0.008931899443268776, + 0.017096949741244316, + 0.007776593789458275, + 0.013154483400285244, + 0.008833217434585094, + 0.01210994366556406, + 0.0076057882979512215, + 0.008968180045485497, + 0.010778969153761864, + 0.011267184279859066, + 0.011942497454583645, + 0.008997189812362194, + 0.010113393887877464, + 0.012346065603196621, + 0.01855549030005932, + 0.009761259891092777, + 0.010465763509273529, + 0.01102522388100624, + 0.008255282416939735, + 0.009179101325571537, + 0.004471341148018837, + 0.005916341673582792, + 0.0067130629904568195, + 0.003744869027286768, + 0.00362375914119184, + 0.004750385414808989, + 0.0033763207029551268, + 0.006261132191866636, + 0.011132139712572098, + 0.02224106527864933, + 0.016969799995422363, + 0.017344404011964798, + 0.022443046793341637, + 0.01936846226453781, + 0.026380037888884544, + 0.017280492931604385, + 0.009907009080052376, + 0.014230425469577312, + 0.011312116868793964, + 0.013529150746762753, + 0.012772767804563046, + 0.012445429340004921, + 0.01034170389175415, + 0.018805133178830147, + 0.004133254289627075, + 0.004659030120819807, + 0.00811090413480997, + 0.002973079914227128, + 0.0034204174298793077, + 0.00561892194673419, + 0.004174848552793264, + 0.004632477182894945, + 0.004684843588620424, + 0.0158065278083086, + 0.004688694607466459, + 0.004071615636348724, + 0.014909223653376102, + 0.005016123875975609, + 0.003840754274278879, + 0.004552734084427357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.038126446306705475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819760538637638 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.01819760538637638, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819760538637638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2481139600276947, + 0.08903798460960388, + 0.03720194846391678, + 0.04664655402302742, + 0.06209680810570717, + 0.07398276776075363, + 0.06681395322084427, + 0.09273559600114822, + 0.08402016758918762, + 0.1993502527475357 + ], + "credit_map": [ + 0.024780290201306343, + 0.039354465901851654, + 0.032484956085681915, + 0.01935739815235138, + 0.04228072240948677, + 0.030921680852770805, + 0.03532906994223595, + 0.023605380207300186, + 0.009912529028952122, + 0.012156492099165916, + 0.012954465113580227, + 0.008791723288595676, + 0.014940683729946613, + 0.009114598855376244, + 0.010802178643643856, + 0.010365323163568974, + 0.0035143583081662655, + 0.0049671269953250885, + 0.003387612756341696, + 0.002627010690048337, + 0.005099360831081867, + 0.007315696217119694, + 0.004095997661352158, + 0.006194781977683306, + 0.006813284009695053, + 0.006518030073493719, + 0.0062272208742797375, + 0.004528983496129513, + 0.00754267955198884, + 0.007255933713167906, + 0.004240610636770725, + 0.003519815858453512, + 0.005888141226023436, + 0.009390927851200104, + 0.008976462297141552, + 0.012327229604125023, + 0.0064032115042209625, + 0.007480272091925144, + 0.006845674477517605, + 0.0047848885878920555, + 0.0077415271662175655, + 0.010034951381385326, + 0.008877790533006191, + 0.010682412423193455, + 0.007579977624118328, + 0.009577224031090736, + 0.008517964743077755, + 0.010970914736390114, + 0.006108179688453674, + 0.009303810074925423, + 0.00646924925968051, + 0.006430173292756081, + 0.009154293686151505, + 0.011354717426002026, + 0.008953244425356388, + 0.0090402876958251, + 0.008865797892212868, + 0.01231762208044529, + 0.011162170208990574, + 0.011779886670410633, + 0.01343159656971693, + 0.011705005541443825, + 0.008981711231172085, + 0.01449180394411087, + 0.00953658577054739, + 0.010643672198057175, + 0.01976543664932251, + 0.006722644902765751, + 0.006709533277899027, + 0.010599007830023766, + 0.009802760556340218, + 0.010240525007247925, + 0.01920427568256855, + 0.03451039269566536, + 0.03139722719788551, + 0.0188747588545084, + 0.02724933996796608, + 0.025374753400683403, + 0.02148301526904106, + 0.021256491541862488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.01819760538637638 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289349615573883 + ], + "opposite_logprob_tokens": [ + -1.053934931755066 + ], + "logprob_reward": -0.4289349615573883, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289349615573883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.305154025554657, + 0.09363257139921188, + 0.10150144249200821, + 0.11941701918840408, + 0.10467007756233215, + 0.09371112287044525, + 0.03201083838939667, + 0.0496666319668293, + 0.04645547643303871, + 0.05378081277012825 + ], + "credit_map": [ + 0.03511698544025421, + 0.05062997713685036, + 0.04648298770189285, + 0.023604588583111763, + 0.035536523908376694, + 0.03916555643081665, + 0.03411287069320679, + 0.04050455614924431, + 0.018111201003193855, + 0.012299048714339733, + 0.01074899174273014, + 0.012678203172981739, + 0.01095549575984478, + 0.008877175860106945, + 0.010302801616489887, + 0.009659646078944206, + 0.010870884172618389, + 0.014604896306991577, + 0.010416978970170021, + 0.00999236386269331, + 0.015672320500016212, + 0.012646329589188099, + 0.012296983040869236, + 0.015000689774751663, + 0.015936188399791718, + 0.016933681443333626, + 0.015323479659855366, + 0.010824183002114296, + 0.015445197932422161, + 0.013856527395546436, + 0.01043306477367878, + 0.020664701238274574, + 0.01400539930909872, + 0.017392830923199654, + 0.01537454966455698, + 0.015155099332332611, + 0.011295698583126068, + 0.013828332535922527, + 0.007391860242933035, + 0.010226305574178696, + 0.01292470470070839, + 0.014382289722561836, + 0.020014779642224312, + 0.008559434674680233, + 0.007804690860211849, + 0.009681451134383678, + 0.0065937889739871025, + 0.013749982230365276, + 0.003824150189757347, + 0.005298319738358259, + 0.004017158877104521, + 0.0033363024704158306, + 0.0038221003487706184, + 0.004416844341903925, + 0.004153123591095209, + 0.003142837667837739, + 0.007863156497478485, + 0.00765698729082942, + 0.0077663687989115715, + 0.004513683263212442, + 0.007214442826807499, + 0.005746198818087578, + 0.0036370076704770327, + 0.005268788896501064, + 0.004732470493763685, + 0.00573794636875391, + 0.007757184095680714, + 0.004323259927332401, + 0.005444651935249567, + 0.008109292015433311, + 0.004041032865643501, + 0.006309636868536472, + 0.004130109213292599, + 0.007566566579043865, + 0.007648017723113298, + 0.01004551537334919, + 0.007143711671233177, + 0.007616425398737192, + 0.00454288674518466, + 0.005087584722787142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.4289349615573883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot from a horizontal angle, yet accompanied by constant, high-frequency micro-vibrations due to the train's breathtaking speed, expertly and directly captures the interior of a high-speed maglev train. The interior showcases a minimalist, futuristic design, with ergonomically designed, streamlined seats constructed of a light gray, high-tech composite material, neatly arranged. Bright sunlight streams through the large picture windows, blending with the soft, even cool white light emitted by LED strips embedded in the ceiling, illuminating the nearly invisible dust particles in the air. The rural landscape outside the windows is reduced to a blur of rapidly passing dynamic streaks, with only swift flashes of green and brown. The immense power of the train generates a deep resonance throughout the train, transforming into a barely perceptible vibration that reflects fine, flowing light off the metal handrails and window glass, creating a unique atmosphere of cutting-edge technology, extreme speed, and cool efficiency.", + "seed": 7775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_11/index_9452_seed_7775_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13916279133409262, + "vlm_logprob": -0.13916279133409262, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.13916279133409262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19169919192790985 + ], + "opposite_logprob_tokens": [ + -1.9416991472244263 + ], + "logprob_reward": -0.19169919192790985, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.19169919192790985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09578771144151688, + 0.10797806084156036, + 0.10076738148927689, + 0.052135810256004333, + 0.1648973971605301, + 0.07038198411464691, + 0.0459732860326767, + 0.09966054558753967, + 0.16731643676757812, + 0.09510141611099243 + ], + "credit_map": [ + 0.009428087621927261, + 0.012356877326965332, + 0.014153199270367622, + 0.01079978235065937, + 0.011524437926709652, + 0.00910230539739132, + 0.013452078215777874, + 0.014970943331718445, + 0.012679981999099255, + 0.01268642209470272, + 0.013791093602776527, + 0.018053652718663216, + 0.011019197292625904, + 0.014538435265421867, + 0.01155935414135456, + 0.013649926520884037, + 0.007895716466009617, + 0.00974019430577755, + 0.012616544961929321, + 0.010593126527965069, + 0.01309636328369379, + 0.02066316269338131, + 0.013038290664553642, + 0.01312397513538599, + 0.0038672504015266895, + 0.0052508143708109856, + 0.00857518706470728, + 0.007422445807605982, + 0.008090152405202389, + 0.006754028145223856, + 0.005968726240098476, + 0.00620720349252224, + 0.010008727200329304, + 0.029460901394486427, + 0.01446008775383234, + 0.028798148036003113, + 0.021138004958629608, + 0.029817242175340652, + 0.016067812219262123, + 0.015146475285291672, + 0.00553131615743041, + 0.008442179299890995, + 0.008896848186850548, + 0.008215303532779217, + 0.008131062611937523, + 0.009168687276542187, + 0.013057255186140537, + 0.008939331397414207, + 0.00444365106523037, + 0.0069763557985424995, + 0.007103321608155966, + 0.004650901537388563, + 0.004164597485214472, + 0.0057725682854652405, + 0.006822105962783098, + 0.006039783824235201, + 0.0077467928640544415, + 0.02251683920621872, + 0.0053391712717711926, + 0.007130924146622419, + 0.009138788096606731, + 0.006415525916963816, + 0.014459121972322464, + 0.02691338211297989, + 0.020793816074728966, + 0.02810778096318245, + 0.022943852469325066, + 0.021055862307548523, + 0.02482045255601406, + 0.016588617116212845, + 0.013871234841644764, + 0.019134825095534325, + 0.029961464926600456, + 0.007970107719302177, + 0.008112375624477863, + 0.007760053966194391, + 0.004981264006346464, + 0.00483902869746089, + 0.02465076930820942, + 0.006826351396739483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.19169919192790985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11310653388500214 + ], + "opposite_logprob_tokens": [ + -2.2381064891815186 + ], + "logprob_reward": -0.11310653388500214, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11310653388500214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10329655557870865, + 0.09136784821748734, + 0.21327316761016846, + 0.06256897002458572, + 0.038207925856113434, + 0.11535503715276718, + 0.053585976362228394, + 0.0505971722304821, + 0.10343506932258606, + 0.16831226646900177 + ], + "credit_map": [ + 0.009068858809769154, + 0.016497274860739708, + 0.009522692300379276, + 0.011637059971690178, + 0.01826392486691475, + 0.015903539955615997, + 0.011209185235202312, + 0.011194020509719849, + 0.012549230828881264, + 0.006573401857167482, + 0.01198845449835062, + 0.014270112849771976, + 0.008380181156098843, + 0.012138099409639835, + 0.014281305484473705, + 0.011187056079506874, + 0.019082404673099518, + 0.022187279537320137, + 0.0257980078458786, + 0.016359861940145493, + 0.03022572211921215, + 0.04812159016728401, + 0.024783095344901085, + 0.02671521157026291, + 0.003928810358047485, + 0.014185859821736813, + 0.0051730358973145485, + 0.006320948712527752, + 0.011534668505191803, + 0.010759435594081879, + 0.004416514653712511, + 0.006249701138585806, + 0.002869503339752555, + 0.008676117286086082, + 0.004177617374807596, + 0.004263150505721569, + 0.005737301893532276, + 0.005134584382176399, + 0.0026243433821946383, + 0.004725304897874594, + 0.006765511352568865, + 0.013367497362196445, + 0.008812863379716873, + 0.012772849760949612, + 0.027297940105199814, + 0.0227364469319582, + 0.01184266246855259, + 0.011759258806705475, + 0.004781212192028761, + 0.007118626032024622, + 0.007166970521211624, + 0.006713556125760078, + 0.007095315959304571, + 0.00703761400654912, + 0.00825747475028038, + 0.005415207706391811, + 0.003210279159247875, + 0.008415447548031807, + 0.003528417320922017, + 0.0035928182769566774, + 0.003895118599757552, + 0.004281626548618078, + 0.009323464706540108, + 0.014349999837577343, + 0.007970527745783329, + 0.01702910289168358, + 0.016886351630091667, + 0.015087582170963287, + 0.014568361453711987, + 0.010856111533939838, + 0.009840787388384342, + 0.011196247301995754, + 0.017186900600790977, + 0.016614122316241264, + 0.01839446648955345, + 0.018707716837525368, + 0.01502798218280077, + 0.020159129053354263, + 0.0332341231405735, + 0.028987837955355644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.11310653388500214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337226502597332 + ], + "opposite_logprob_tokens": [ + -3.4087226390838623 + ], + "logprob_reward": -0.0337226502597332, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337226502597332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06322525441646576, + 0.0543242022395134, + 0.3707968294620514, + 0.03879542648792267, + 0.05001882091164589, + 0.09851377457380295, + 0.022505443543195724, + 0.029337203130126, + 0.17535138130187988, + 0.0971316546201706 + ], + "credit_map": [ + 0.006406985688954592, + 0.006766739767044783, + 0.010810842737555504, + 0.01056166086345911, + 0.005773451644927263, + 0.006304136477410793, + 0.007806353736668825, + 0.008795072324573994, + 0.00486468430608511, + 0.008094269782304764, + 0.007982898503541946, + 0.006072643678635359, + 0.00631297891959548, + 0.009414463303983212, + 0.004966472275555134, + 0.006615788675844669, + 0.027077559381723404, + 0.029487842693924904, + 0.06742838025093079, + 0.030768614262342453, + 0.04804141819477081, + 0.06384716928005219, + 0.05073332414031029, + 0.053412504494190216, + 0.0034145573154091835, + 0.004326063673943281, + 0.006371479947119951, + 0.005067202728241682, + 0.006207428406924009, + 0.005324017256498337, + 0.004221613984555006, + 0.003863063408061862, + 0.003130743745714426, + 0.010991654358804226, + 0.005306561943143606, + 0.004590123891830444, + 0.008583908900618553, + 0.006036878097802401, + 0.006132860668003559, + 0.005246082786470652, + 0.006100823171436787, + 0.011691998690366745, + 0.008603318594396114, + 0.011071893386542797, + 0.021651076152920723, + 0.018823498860001564, + 0.009768697433173656, + 0.01080246176570654, + 0.0017633967800065875, + 0.00443057157099247, + 0.0026488578878343105, + 0.0021948497742414474, + 0.0032721974421292543, + 0.003052155254408717, + 0.002905392786487937, + 0.0022380196023732424, + 0.0031003206968307495, + 0.004882833454757929, + 0.002949782647192478, + 0.004496871959418058, + 0.0040766336023807526, + 0.002816841471940279, + 0.0038827366661280394, + 0.003131178207695484, + 0.012368050403892994, + 0.04040243849158287, + 0.019801991060376167, + 0.02512529492378235, + 0.023660458624362946, + 0.020780257880687714, + 0.016074853017926216, + 0.017138024792075157, + 0.03436420485377312, + 0.008870959281921387, + 0.006702450569719076, + 0.006347845308482647, + 0.0063690603710711, + 0.004158892203122377, + 0.0233309306204319, + 0.006987301167100668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.0337226502597332 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644304238259792 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644304238259792, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644304238259792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21396920084953308, + 0.05560401827096939, + 0.14211112260818481, + 0.06353161484003067, + 0.05457240715622902, + 0.10940878093242645, + 0.04762570932507515, + 0.057264190167188644, + 0.08729610592126846, + 0.1686168909072876 + ], + "credit_map": [ + 0.014888562262058258, + 0.016740186139941216, + 0.02310088649392128, + 0.03273481875658035, + 0.03554303199052811, + 0.03337418660521507, + 0.028296222910284996, + 0.0292913056910038, + 0.00690826028585434, + 0.005375158507376909, + 0.007346096448600292, + 0.006726707797497511, + 0.004608050920069218, + 0.007742601912468672, + 0.007448619231581688, + 0.009448518045246601, + 0.014943554066121578, + 0.014636173844337463, + 0.017924794927239418, + 0.011462576687335968, + 0.02035941369831562, + 0.028637228533625603, + 0.018776461482048035, + 0.015370918437838554, + 0.004497984889894724, + 0.010091662406921387, + 0.007670570630580187, + 0.00781679805368185, + 0.011224357411265373, + 0.009276344440877438, + 0.005798052065074444, + 0.007155849132686853, + 0.00339462049305439, + 0.008072775788605213, + 0.0046933903358876705, + 0.011178051121532917, + 0.005587302148342133, + 0.007425423711538315, + 0.008729828521609306, + 0.005491010844707489, + 0.009200380183756351, + 0.016014661639928818, + 0.009608260355889797, + 0.011901137419044971, + 0.018723424524068832, + 0.018877144902944565, + 0.011010395362973213, + 0.014073383063077927, + 0.004219344817101955, + 0.006731635890901089, + 0.004915900994092226, + 0.004378085024654865, + 0.007264153566211462, + 0.005535709206014872, + 0.007265785243362188, + 0.0073150936514139175, + 0.002938488032668829, + 0.008702417835593224, + 0.002917894162237644, + 0.003408589633181691, + 0.003308940911665559, + 0.0026356512680649757, + 0.01298690028488636, + 0.020365307107567787, + 0.007496302016079426, + 0.011470474302768707, + 0.015968821942806244, + 0.011130956001579762, + 0.013284381479024887, + 0.010600745677947998, + 0.007283911574631929, + 0.010060501284897327, + 0.061721980571746826, + 0.010791179724037647, + 0.01018805243074894, + 0.012059864588081837, + 0.0069710733368992805, + 0.005736387334764004, + 0.049461979418992996, + 0.011686380952596664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.008644304238259792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486412763595581 + ], + "opposite_logprob_tokens": [ + -1.223641276359558 + ], + "logprob_reward": -0.3486412763595581, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486412763595581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07226091623306274, + 0.07832658290863037, + 0.20775656402111053, + 0.050842177122831345, + 0.1020495742559433, + 0.12931452691555023, + 0.11028274148702621, + 0.05746869370341301, + 0.053876928985118866, + 0.1378212571144104 + ], + "credit_map": [ + 0.007189679425209761, + 0.008793340064585209, + 0.012276522815227509, + 0.01144731231033802, + 0.007864722050726414, + 0.006426235195249319, + 0.010237285867333412, + 0.008025824092328548, + 0.010263578034937382, + 0.010222317650914192, + 0.011123073287308216, + 0.010113691911101341, + 0.007672141306102276, + 0.009967487305402756, + 0.009075879119336605, + 0.009888417087495327, + 0.031978026032447815, + 0.025519756600260735, + 0.016264913603663445, + 0.0171956866979599, + 0.03164132311940193, + 0.04469885304570198, + 0.02060142159461975, + 0.019856588914990425, + 0.004254224244505167, + 0.008028870448470116, + 0.008052400313317776, + 0.006897986400872469, + 0.006715416442602873, + 0.0057077836245298386, + 0.006664011627435684, + 0.004521488212049007, + 0.00567657733336091, + 0.01832125522196293, + 0.008163793943822384, + 0.017604902386665344, + 0.011301564052700996, + 0.020692123100161552, + 0.009771179407835007, + 0.010518195107579231, + 0.013876699842512608, + 0.017763376235961914, + 0.014034133404493332, + 0.015806108713150024, + 0.019834650680422783, + 0.017853036522865295, + 0.014922227710485458, + 0.015224297530949116, + 0.01545930840075016, + 0.018314188346266747, + 0.013983849436044693, + 0.011471936479210854, + 0.01262086071074009, + 0.014622665010392666, + 0.013412493281066418, + 0.010397445410490036, + 0.004017633851617575, + 0.008494134061038494, + 0.002784131560474634, + 0.0035194188822060823, + 0.004708521533757448, + 0.0028212431352585554, + 0.011990050785243511, + 0.019133562222123146, + 0.004669428803026676, + 0.007123230490833521, + 0.006344439461827278, + 0.009673015214502811, + 0.006194133311510086, + 0.006447731051594019, + 0.007408728823065758, + 0.00601622462272644, + 0.021628979593515396, + 0.014338470064103603, + 0.016026169061660767, + 0.013423596508800983, + 0.011817567981779575, + 0.019213279709219933, + 0.01960628665983677, + 0.021766917780041695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.3486412763595581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot from a horizontal angle, yet accompanied by constant, high-frequency micro-vibrations due to the train's breathtaking speed, expertly and directly captures the interior of a high-speed maglev train. The interior showcases a minimalist, futuristic design, with ergonomically designed, streamlined seats constructed of a light gray, high-tech composite material, neatly arranged. Bright sunlight streams through the large picture windows, blending with the soft, even cool white light emitted by LED strips embedded in the ceiling, illuminating the nearly invisible dust particles in the air. The rural landscape outside the windows is reduced to a blur of rapidly passing dynamic streaks, with only swift flashes of green and brown. The immense power of the train generates a deep resonance throughout the train, transforming into a barely perceptible vibration that reflects fine, flowing light off the metal handrails and window glass, creating a unique atmosphere of cutting-edge technology, extreme speed, and cool efficiency.", + "seed": 7776, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_11/index_9452_seed_7776_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06865769196301699, + "vlm_logprob": -0.06865769196301699, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.06865769196301699, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047842808067798615 + ], + "opposite_logprob_tokens": [ + -3.2978427410125732 + ], + "logprob_reward": -0.047842808067798615, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.047842808067798615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3008544445037842, + 0.05502947047352791, + 0.04107128083705902, + 0.24274928867816925, + 0.036283258348703384, + 0.03919309750199318, + 0.056774288415908813, + 0.06096627563238144, + 0.1347295045852661, + 0.0323491208255291 + ], + "credit_map": [ + 0.026566816493868828, + 0.03159574419260025, + 0.020425502210855484, + 0.04753676801919937, + 0.04211263731122017, + 0.030396759510040283, + 0.05554782226681709, + 0.046672407537698746, + 0.00662354938685894, + 0.006232037674635649, + 0.003755975980311632, + 0.007280109450221062, + 0.006596093997359276, + 0.006121673621237278, + 0.00910037662833929, + 0.009319650009274483, + 0.003826016327366233, + 0.006848297081887722, + 0.002763625467196107, + 0.007617248222231865, + 0.006348320748656988, + 0.005248053930699825, + 0.005016551818698645, + 0.0034031711984425783, + 0.006886609829962254, + 0.09873487055301666, + 0.003481154562905431, + 0.003480939194560051, + 0.10320325940847397, + 0.01174358930438757, + 0.008664509281516075, + 0.006554371677339077, + 0.004084879532456398, + 0.006400588899850845, + 0.0031149794813245535, + 0.003681443864479661, + 0.006029009353369474, + 0.004187604412436485, + 0.0048531037755310535, + 0.003931649960577488, + 0.004134987946599722, + 0.005768261384218931, + 0.005784006789326668, + 0.005169464275240898, + 0.006319636013358831, + 0.005040759686380625, + 0.003683098591864109, + 0.0032928837463259697, + 0.0049546267837285995, + 0.011738431639969349, + 0.0036385119892656803, + 0.004946407862007618, + 0.0129008824005723, + 0.007042282726615667, + 0.005474782548844814, + 0.0060783629305660725, + 0.003504893509671092, + 0.013142568990588188, + 0.002684307750314474, + 0.004464076366275549, + 0.014352891594171524, + 0.0059118131175637245, + 0.008540242910385132, + 0.008365479297935963, + 0.004491183906793594, + 0.05284196510910988, + 0.0032897633500397205, + 0.004670195747166872, + 0.05476699769496918, + 0.006804255302995443, + 0.003745847847312689, + 0.004119303543120623, + 0.002657938050106168, + 0.006437892559915781, + 0.0018231043359264731, + 0.002613226417452097, + 0.007896379567682743, + 0.0033114636316895485, + 0.003992226906120777, + 0.00361688993871212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.047842808067798615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012578792870044708 + ], + "opposite_logprob_tokens": [ + -4.387578964233398 + ], + "logprob_reward": -0.012578792870044708, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012578792870044708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15706922113895416, + 0.06355662643909454, + 0.036751341074705124, + 0.3563525080680847, + 0.041680652648210526, + 0.02557629719376564, + 0.07278095185756683, + 0.1466194987297058, + 0.05681576952338219, + 0.04279714450240135 + ], + "credit_map": [ + 0.012080186977982521, + 0.028048651292920113, + 0.010271680541336536, + 0.015248733572661877, + 0.022424792870879173, + 0.01766376569867134, + 0.027115730568766594, + 0.02421567775309086, + 0.004865174647420645, + 0.010861094109714031, + 0.004561337176710367, + 0.005688161123543978, + 0.013783284462988377, + 0.00582533422857523, + 0.008279229514300823, + 0.009693006984889507, + 0.0035215099342167377, + 0.004537164233624935, + 0.0022712855134159327, + 0.0066650426015257835, + 0.005216706078499556, + 0.004927437752485275, + 0.005049601197242737, + 0.004562594462186098, + 0.010768868029117584, + 0.14393910765647888, + 0.005456438288092613, + 0.005410533864051104, + 0.14950262010097504, + 0.01768135093152523, + 0.013242759741842747, + 0.010350827127695084, + 0.0033418647944927216, + 0.007818836718797684, + 0.003133464138954878, + 0.004793246276676655, + 0.008570647798478603, + 0.003987794276326895, + 0.004873926285654306, + 0.005160870961844921, + 0.0020409885328263044, + 0.003418226260691881, + 0.001617330126464367, + 0.004198217298835516, + 0.0027938743587583303, + 0.0031664655543863773, + 0.0042692432180047035, + 0.0040719518437981606, + 0.005894397385418415, + 0.007872366346418858, + 0.0033817696385085583, + 0.00614976417273283, + 0.00996819231659174, + 0.008733050897717476, + 0.014691525138914585, + 0.01608988083899021, + 0.0065080090425908566, + 0.039911359548568726, + 0.00447511114180088, + 0.006903134286403656, + 0.04204988107085228, + 0.010032976046204567, + 0.018010731786489487, + 0.01872829720377922, + 0.0022991043515503407, + 0.016565551981329918, + 0.0019518564222380519, + 0.0032023796811699867, + 0.017124220728874207, + 0.0039627524092793465, + 0.0066214026883244514, + 0.005088496021926403, + 0.0026773237623274326, + 0.008974901400506496, + 0.0021569307427853346, + 0.0032441639341413975, + 0.009824116714298725, + 0.003974560648202896, + 0.005782731808722019, + 0.006162412464618683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.012578792870044708 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210578985512257 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210578985512257, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210578985512257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24784968793392181, + 0.07643749564886093, + 0.027653768658638, + 0.11677689850330353, + 0.05137866362929344, + 0.02773982100188732, + 0.28043028712272644, + 0.05853112041950226, + 0.08526147156953812, + 0.027940871194005013 + ], + "credit_map": [ + 0.01762690395116806, + 0.06579740345478058, + 0.010616403073072433, + 0.019465485587716103, + 0.05550060048699379, + 0.024298084899783134, + 0.029971571639180183, + 0.024573195725679398, + 0.006834734231233597, + 0.010547531768679619, + 0.006420631892979145, + 0.007937556132674217, + 0.00888766162097454, + 0.008571026846766472, + 0.015021643601357937, + 0.012216702103614807, + 0.0029048940632492304, + 0.0049271672032773495, + 0.0014743030769750476, + 0.0033416280057281256, + 0.004461043979972601, + 0.0037718540988862514, + 0.004151628352701664, + 0.002621246036142111, + 0.004699817392975092, + 0.043227121233940125, + 0.0026266700588166714, + 0.005422996822744608, + 0.044214535504579544, + 0.006725131068378687, + 0.0049093118868768215, + 0.004951299633830786, + 0.005936706904321909, + 0.010172316804528236, + 0.004927340894937515, + 0.006004744675010443, + 0.010828332044184208, + 0.005665330681949854, + 0.00452797906473279, + 0.003315909765660763, + 0.002886564703658223, + 0.0035167064052075148, + 0.0025482208002358675, + 0.003370831022039056, + 0.0038903558161109686, + 0.004336440935730934, + 0.003774053417146206, + 0.003416643710806966, + 0.01832684502005577, + 0.0749664157629013, + 0.015800397843122482, + 0.01863005943596363, + 0.07315665483474731, + 0.022764291614294052, + 0.027717040851712227, + 0.02906855382025242, + 0.0023244735784828663, + 0.017508916556835175, + 0.0020267069339752197, + 0.0023773068096488714, + 0.018902119249105453, + 0.0041051702573895454, + 0.005562799982726574, + 0.005723621696233749, + 0.002998439595103264, + 0.03432584926486015, + 0.0015475378604605794, + 0.002029282273724675, + 0.03550886735320091, + 0.004180531483143568, + 0.0028145185206085443, + 0.001856433693319559, + 0.0019300252897664905, + 0.006686576176434755, + 0.0013940567150712013, + 0.0020055619534105062, + 0.006913917604833841, + 0.0031242752447724342, + 0.0034477675799280405, + 0.002438687952235341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.014210578985512257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067168444395065 + ], + "opposite_logprob_tokens": [ + -3.168067216873169 + ], + "logprob_reward": -0.043067168444395065, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067168444395065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24646836519241333, + 0.09677081555128098, + 0.0663193017244339, + 0.04136268422007561, + 0.07623074948787689, + 0.029774608090519905, + 0.19535887241363525, + 0.09076612442731857, + 0.14146223664283752, + 0.015486234799027443 + ], + "credit_map": [ + 0.027980862185359, + 0.02806970477104187, + 0.017839910462498665, + 0.028947418555617332, + 0.032044149935245514, + 0.027683893218636513, + 0.042084578424692154, + 0.04181784391403198, + 0.008178936317563057, + 0.019713710993528366, + 0.010087606497108936, + 0.00806264579296112, + 0.017123304307460785, + 0.011831761337816715, + 0.011382875964045525, + 0.010389965027570724, + 0.006662060972303152, + 0.012545410543680191, + 0.00485219107940793, + 0.008484472520649433, + 0.011122268624603748, + 0.00892522744834423, + 0.006897683255374432, + 0.006829980760812759, + 0.0033940414432436228, + 0.010084796696901321, + 0.001842077705077827, + 0.003022336633875966, + 0.009145237505435944, + 0.004555226769298315, + 0.004783072974532843, + 0.0045358967036008835, + 0.007022079546004534, + 0.014824995771050453, + 0.005598578602075577, + 0.006946513429284096, + 0.016838347539305687, + 0.011481259949505329, + 0.00606532720848918, + 0.007453643251210451, + 0.0037356859538704157, + 0.003230758011341095, + 0.004428823012858629, + 0.004734086338430643, + 0.003742335829883814, + 0.0035158239770680666, + 0.003323053242638707, + 0.003064042190089822, + 0.013837319798767567, + 0.05217673256993294, + 0.010133888572454453, + 0.013900130055844784, + 0.051611870527267456, + 0.017756445333361626, + 0.019316185265779495, + 0.016626300290226936, + 0.00347866746596992, + 0.02526922896504402, + 0.003994406666606665, + 0.004117119591683149, + 0.028033079579472542, + 0.005791186820715666, + 0.010174435563385487, + 0.009908011183142662, + 0.004923604428768158, + 0.05412071570754051, + 0.004559531342238188, + 0.004671448841691017, + 0.05633130297064781, + 0.007137274835258722, + 0.005768659990280867, + 0.00394969480112195, + 0.00148386403452605, + 0.0024691757280379534, + 0.00102735694963485, + 0.0014065884752199054, + 0.002839691238477826, + 0.0016133850440382957, + 0.0024513741955161095, + 0.0021947992499917746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.043067168444395065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene the interior of a high-speed maglev train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the seats light gray, streamlined, and made of a high-tech composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting a mix of bright sunlight from large windows and cool white LED strips on the ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the view outside the windows a blur of rapidly passing green and brown streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static with constant, high-frequency micro-vibrations visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255891114473343 + ], + "opposite_logprob_tokens": [ + -1.6005891561508179 + ], + "logprob_reward": -0.2255891114473343, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255891114473343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22440189123153687, + 0.061345066875219345, + 0.05861994996666908, + 0.449977844953537, + 0.03537707030773163, + 0.022305969148874283, + 0.045837074518203735, + 0.031338661909103394, + 0.04598323255777359, + 0.024813212454319 + ], + "credit_map": [ + 0.023950086906552315, + 0.03393879532814026, + 0.01543265487998724, + 0.022542038932442665, + 0.04006803035736084, + 0.03175143525004387, + 0.029192917048931122, + 0.027525953948497772, + 0.005645183380693197, + 0.011947853490710258, + 0.004204705357551575, + 0.00486911041662097, + 0.012892447412014008, + 0.008552209474146366, + 0.008113676682114601, + 0.005119886249303818, + 0.004140601027756929, + 0.015852954238653183, + 0.002894375938922167, + 0.006633758544921875, + 0.014143364503979683, + 0.004483845084905624, + 0.00524537730962038, + 0.005225683096796274, + 0.014027980156242847, + 0.18392249941825867, + 0.007288413122296333, + 0.0069017838686704636, + 0.19064763188362122, + 0.022676125168800354, + 0.01372699998319149, + 0.010786446742713451, + 0.0027800900861620903, + 0.004403304774314165, + 0.00265499041415751, + 0.004471710417419672, + 0.0048347353003919125, + 0.004079864360392094, + 0.005556777119636536, + 0.006595600862056017, + 0.0022821829188615084, + 0.0021514245308935642, + 0.001816397299990058, + 0.003308634040877223, + 0.002501292619854212, + 0.0034487845841795206, + 0.003973627462983131, + 0.0028236277867108583, + 0.004247923381626606, + 0.007704847026616335, + 0.002256433479487896, + 0.005275130737572908, + 0.007799352984875441, + 0.0058250222355127335, + 0.006087453104555607, + 0.006640920415520668, + 0.0033591545652598143, + 0.0035923258401453495, + 0.0023066720459610224, + 0.004722252022475004, + 0.0053040278144180775, + 0.005677817389369011, + 0.003024573205038905, + 0.0033518392592668533, + 0.0033775384072214365, + 0.00721319392323494, + 0.0034099083859473467, + 0.0050696455873548985, + 0.0076348125003278255, + 0.004533560946583748, + 0.006497316528111696, + 0.008247260004281998, + 0.0015513056423515081, + 0.004214836750179529, + 0.0016556497430428863, + 0.002499612746760249, + 0.004029801115393639, + 0.00291855214163661, + 0.0035494123585522175, + 0.004394044633954763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.2255891114473343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_12/index_tensor([1156, 1156])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_12/index_tensor([1156, 1156])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a1ee9b49d628828e4096cac879b64de330eeea0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_12/index_tensor([1156, 1156])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic and intense cinematic low-angle wide shot maintains a static-like view of a vast, sinister Gothic mansion, its form almost completely engulfed by the deep black of a stormy night. A fierce thunderstorm is in full swing, and thick rain pelts the building, obscuring the view. Suddenly, a massive flash of brilliant white lightning streaks across the turbulent, roiling purple-black sky, momentarily and violently illuminating the entire scene. In this fleeting, high-contrast flash, the mansion is revealed in stark detail: towering, needle-like spires reach toward the sky, weathered stone walls gleam damply, and eerie gargoyles seem to scream silently in the gusts of wind. Dark, tall lancet windows, like hollow eyes, reflect the searing flash before the scene returns to near-total darkness, leaving only a haunting, ominous silhouette against briefly ignited clouds. The intense atmosphere is filled with foreboding and the raw power of nature.", + "seed": 6671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_12/index_1156_seed_6671_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03497968886513263, + "vlm_logprob": -0.03497968886513263, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03497968886513263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015153642743825912 + ], + "opposite_logprob_tokens": [ + -5.515153408050537 + ], + "logprob_reward": -0.015153642743825912, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.015153642743825912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08369272202253342, + 0.17759042978286743, + 0.058862172067165375, + 0.0683029294013977, + 0.10827986150979996, + 0.08973581343889236, + 0.12759605050086975, + 0.11490891128778458, + 0.06287005543708801, + 0.10816101729869843 + ], + "credit_map": [ + 0.012353542260825634, + 0.008427738212049007, + 0.015274185687303543, + 0.008813371881842613, + 0.008901825174689293, + 0.009480981156229973, + 0.011349684558808804, + 0.009091387502849102, + 0.023619864135980606, + 0.02060568891465664, + 0.03367369994521141, + 0.02388775907456875, + 0.01007822621613741, + 0.02054727077484131, + 0.02409668080508709, + 0.021081244572997093, + 0.007148194592446089, + 0.006106102839112282, + 0.010027428157627583, + 0.004850056953728199, + 0.007582979276776314, + 0.006014994811266661, + 0.008785158395767212, + 0.008347255177795887, + 0.010396566241979599, + 0.0075506847351789474, + 0.010668737813830376, + 0.00807344913482666, + 0.009335723705589771, + 0.006877262610942125, + 0.0070130410604178905, + 0.008387461304664612, + 0.01253524236381054, + 0.014379877597093582, + 0.01122970413416624, + 0.012479469180107117, + 0.014618092216551304, + 0.010734682902693748, + 0.012214835733175278, + 0.020087955519557, + 0.012920735403895378, + 0.007814128883183002, + 0.009624695405364037, + 0.009964668191969395, + 0.007546667475253344, + 0.010590378195047379, + 0.00931580737233162, + 0.0219587292522192, + 0.010684981010854244, + 0.01854359172284603, + 0.016870133578777313, + 0.01294072438031435, + 0.01945558562874794, + 0.011092116124927998, + 0.019762203097343445, + 0.0182467233389616, + 0.009482544846832752, + 0.011245207861065865, + 0.01407327875494957, + 0.009154400788247585, + 0.017912378534674644, + 0.02117229253053665, + 0.017151106148958206, + 0.014717687852680683, + 0.005728740245103836, + 0.007780859712511301, + 0.006979873403906822, + 0.008483844809234142, + 0.007559166289865971, + 0.008918731473386288, + 0.009356525726616383, + 0.008062312379479408, + 0.011482421308755875, + 0.012170729227364063, + 0.013879774138331413, + 0.011455447413027287, + 0.014136589132249355, + 0.010623189620673656, + 0.01650364138185978, + 0.017909221351146698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.015153642743825912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062035512179136276 + ], + "opposite_logprob_tokens": [ + -2.81203556060791 + ], + "logprob_reward": -0.062035512179136276, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062035512179136276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16153021156787872, + 0.05081174895167351, + 0.08492586016654968, + 0.14863905310630798, + 0.10044067353010178, + 0.0383940227329731, + 0.18505889177322388, + 0.053226396441459656, + 0.12439414858818054, + 0.05257892236113548 + ], + "credit_map": [ + 0.027248678728938103, + 0.017144950106739998, + 0.018655264750123024, + 0.01646556705236435, + 0.02094261534512043, + 0.01795397512614727, + 0.025359436869621277, + 0.017759744077920914, + 0.007508135866373777, + 0.00457818154245615, + 0.0066857123747467995, + 0.005468803457915783, + 0.0072393338195979595, + 0.005940379109233618, + 0.005901105236262083, + 0.007490110117942095, + 0.00922281015664339, + 0.00660264678299427, + 0.014213288202881813, + 0.007243430707603693, + 0.01651526242494583, + 0.011018622666597366, + 0.01128754299134016, + 0.008822271600365639, + 0.016181278973817825, + 0.02247278392314911, + 0.01936046965420246, + 0.01819719932973385, + 0.0264076367020607, + 0.018346179276704788, + 0.01665179803967476, + 0.01102172490209341, + 0.016418980434536934, + 0.010672502219676971, + 0.012700570747256279, + 0.010326450690627098, + 0.00784361269325018, + 0.007544118911027908, + 0.012681708671152592, + 0.02225273661315441, + 0.005064074881374836, + 0.0038086322601884604, + 0.004122127778828144, + 0.002913998207077384, + 0.0036848532035946846, + 0.004971039015799761, + 0.005468372255563736, + 0.008360928855836391, + 0.01948215439915657, + 0.024721823632717133, + 0.028248971328139305, + 0.02015514299273491, + 0.020494580268859863, + 0.022112740203738213, + 0.030213486403226852, + 0.01963002048432827, + 0.0045313662849366665, + 0.00690424395725131, + 0.005321689881384373, + 0.004449483938515186, + 0.007402383256703615, + 0.0074693369679152966, + 0.006879972293972969, + 0.010267924517393112, + 0.008866795338690281, + 0.01724322885274887, + 0.011642087250947952, + 0.02053920179605484, + 0.017654458060860634, + 0.01319393701851368, + 0.016772277653217316, + 0.018482176586985588, + 0.006093580741435289, + 0.0054178424179553986, + 0.008344504982233047, + 0.006427108775824308, + 0.006252853665500879, + 0.005328504368662834, + 0.00763858575373888, + 0.00707594957202673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.062035512179136276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902736008167267 + ], + "opposite_logprob_tokens": [ + -2.4640274047851562 + ], + "logprob_reward": -0.08902736008167267, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902736008167267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11594844609498978, + 0.14421603083610535, + 0.07852419465780258, + 0.32908129692077637, + 0.06917259097099304, + 0.060439854860305786, + 0.10165426135063171, + 0.02320161834359169, + 0.04374374821782112, + 0.03401795029640198 + ], + "credit_map": [ + 0.016096891835331917, + 0.018777519464492798, + 0.015487784519791603, + 0.011630211025476456, + 0.014065368101000786, + 0.012983538210391998, + 0.016050852835178375, + 0.010856281034648418, + 0.022827982902526855, + 0.013547820039093494, + 0.024747159332036972, + 0.016836699098348618, + 0.01019024197012186, + 0.016223162412643433, + 0.020702242851257324, + 0.019140727818012238, + 0.01273068692535162, + 0.005596059374511242, + 0.008072219789028168, + 0.004537499509751797, + 0.00924998801201582, + 0.011897687800228596, + 0.00872542429715395, + 0.01771463453769684, + 0.037638839334249496, + 0.029039300978183746, + 0.05140621215105057, + 0.02841518446803093, + 0.07325010001659393, + 0.045962557196617126, + 0.03310101851820946, + 0.03026805818080902, + 0.009592396207153797, + 0.008631306700408459, + 0.007790368050336838, + 0.005886536091566086, + 0.010735394433140755, + 0.006142833270132542, + 0.0105860261246562, + 0.009807728230953217, + 0.006588968448340893, + 0.007979130372405052, + 0.007208760362118483, + 0.006353134289383888, + 0.008719212375581264, + 0.005870251916348934, + 0.007361740805208683, + 0.0103586595505476, + 0.008551346138119698, + 0.011403820477426052, + 0.01595933735370636, + 0.009429233148694038, + 0.01713022030889988, + 0.01237686350941658, + 0.01445669773966074, + 0.01234674733132124, + 0.002293424680829048, + 0.0031368692871183157, + 0.0026081781834363937, + 0.0025861291214823723, + 0.0031838295981287956, + 0.003117304528132081, + 0.0028267402667552233, + 0.0034491433762013912, + 0.004387151915580034, + 0.005647693295031786, + 0.004793580621480942, + 0.006438782438635826, + 0.006549112964421511, + 0.003963335417211056, + 0.006755863316357136, + 0.005208231043070555, + 0.00395596120506525, + 0.005874675698578358, + 0.004852001555263996, + 0.003263408550992608, + 0.003427257528528571, + 0.004287143237888813, + 0.004044807981699705, + 0.004312695469707251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.08902736008167267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019443194614723325 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019443194614723325, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019443194614723325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18276898562908173, + 0.19329528510570526, + 0.13479387760162354, + 0.0544034019112587, + 0.14963141083717346, + 0.04886285215616226, + 0.09157994389533997, + 0.06094646826386452, + 0.057192105799913406, + 0.02652570605278015 + ], + "credit_map": [ + 0.027120284736156464, + 0.02749965526163578, + 0.015608209185302258, + 0.016922924667596817, + 0.03570681810379028, + 0.022005189210176468, + 0.022905495017766953, + 0.015000405721366405, + 0.026446525007486343, + 0.020448878407478333, + 0.03520182520151138, + 0.02145698480308056, + 0.018558798357844353, + 0.022005818784236908, + 0.025297200307250023, + 0.02387923002243042, + 0.0191691555082798, + 0.009803935885429382, + 0.013178105466067791, + 0.007174486294388771, + 0.023152433335781097, + 0.021063022315502167, + 0.010831590741872787, + 0.030421148985624313, + 0.005550206173211336, + 0.004842018708586693, + 0.006350966170430183, + 0.0036917340476065874, + 0.014511204324662685, + 0.007030852604657412, + 0.006277371663600206, + 0.006149049382656813, + 0.01816609315574169, + 0.018627019599080086, + 0.01635713316500187, + 0.015588784590363503, + 0.019080832600593567, + 0.011601435020565987, + 0.020527228713035583, + 0.029682863503694534, + 0.004971950780600309, + 0.0069254436530172825, + 0.006025956943631172, + 0.00692036934196949, + 0.006980654317885637, + 0.00553869130089879, + 0.006708652712404728, + 0.004791128449141979, + 0.008364259265363216, + 0.0099640479311347, + 0.011994071304798126, + 0.007613769732415676, + 0.016116444021463394, + 0.009594808332622051, + 0.014308704063296318, + 0.013623842969536781, + 0.005945628043264151, + 0.00908591691404581, + 0.006388182286173105, + 0.006376537028700113, + 0.008235004730522633, + 0.006373445503413677, + 0.007508347276598215, + 0.011033406481146812, + 0.005331083200871944, + 0.007178897969424725, + 0.005942268762737513, + 0.008911024779081345, + 0.008166102692484856, + 0.004529375117272139, + 0.008992829360067844, + 0.008140523917973042, + 0.0033537193667143583, + 0.003586162580177188, + 0.004043462220579386, + 0.0027662268839776516, + 0.0028543181251734495, + 0.003002620767802, + 0.003466642228886485, + 0.0034525517839938402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0019443194614723325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673760985955596 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.00673760985955596, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673760985955596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12011045962572098, + 0.1579108089208603, + 0.10042323917150497, + 0.09117645770311356, + 0.21474038064479828, + 0.06517039239406586, + 0.06732359528541565, + 0.07238312065601349, + 0.050932321697473526, + 0.059829261153936386 + ], + "credit_map": [ + 0.01597652956843376, + 0.017469609156250954, + 0.009349477477371693, + 0.013559081591665745, + 0.01986055076122284, + 0.02041422761976719, + 0.012148236855864525, + 0.011332747526466846, + 0.02308211848139763, + 0.01676025055348873, + 0.029400477185845375, + 0.02095978893339634, + 0.009045873768627644, + 0.019036762416362762, + 0.02032587304711342, + 0.019299663603305817, + 0.018202703446149826, + 0.006267125252634287, + 0.00892061460763216, + 0.004910625983029604, + 0.01208069734275341, + 0.012998411431908607, + 0.011425074189901352, + 0.02561798505485058, + 0.01088891550898552, + 0.00862316507846117, + 0.015813805162906647, + 0.008602150715887547, + 0.01431706827133894, + 0.008854164741933346, + 0.014657962135970592, + 0.009419223293662071, + 0.031036904081702232, + 0.02763506770133972, + 0.023883944377303123, + 0.019772522151470184, + 0.02173880860209465, + 0.016632337123155594, + 0.02333701215684414, + 0.05070380121469498, + 0.007890605367720127, + 0.006478171795606613, + 0.007544929161667824, + 0.008041256107389927, + 0.005509377457201481, + 0.006940525956451893, + 0.007794071454554796, + 0.014971457421779633, + 0.006779188755899668, + 0.00800204835832119, + 0.008881918154656887, + 0.009497267194092274, + 0.01275589782744646, + 0.006761183496564627, + 0.007297842763364315, + 0.0073482501320540905, + 0.0055489311926066875, + 0.009389851242303848, + 0.006983737926930189, + 0.008803736418485641, + 0.01058080606162548, + 0.007515297271311283, + 0.011805303394794464, + 0.011755459941923618, + 0.004433646332472563, + 0.006374891381710768, + 0.006105974316596985, + 0.005293671041727066, + 0.006071134470403194, + 0.00803400482982397, + 0.008143860846757889, + 0.006475143134593964, + 0.0055927434004843235, + 0.00874558836221695, + 0.005958874709904194, + 0.0070615485310554504, + 0.008663876913487911, + 0.007731064222753048, + 0.0074872104451060295, + 0.008588350377976894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.00673760985955596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic and intense cinematic low-angle wide shot maintains a static-like view of a vast, sinister Gothic mansion, its form almost completely engulfed by the deep black of a stormy night. A fierce thunderstorm is in full swing, and thick rain pelts the building, obscuring the view. Suddenly, a massive flash of brilliant white lightning streaks across the turbulent, roiling purple-black sky, momentarily and violently illuminating the entire scene. In this fleeting, high-contrast flash, the mansion is revealed in stark detail: towering, needle-like spires reach toward the sky, weathered stone walls gleam damply, and eerie gargoyles seem to scream silently in the gusts of wind. Dark, tall lancet windows, like hollow eyes, reflect the searing flash before the scene returns to near-total darkness, leaving only a haunting, ominous silhouette against briefly ignited clouds. The intense atmosphere is filled with foreboding and the raw power of nature.", + "seed": 6672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_12/index_1156_seed_6672_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1496325488202274, + "vlm_logprob": -0.1496325488202274, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.1496325488202274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05397813767194748 + ], + "opposite_logprob_tokens": [ + -3.178978204727173 + ], + "logprob_reward": -0.05397813767194748, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05397813767194748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2323695421218872, + 0.1248144879937172, + 0.07269749045372009, + 0.1090422049164772, + 0.07950133085250854, + 0.0685640349984169, + 0.14634756743907928, + 0.05924856662750244, + 0.03052636981010437, + 0.07688841223716736 + ], + "credit_map": [ + 0.033467452973127365, + 0.02312454581260681, + 0.029568785801529884, + 0.03972281888127327, + 0.01750982739031315, + 0.041509877890348434, + 0.018733730539679527, + 0.028732512146234512, + 0.0330086424946785, + 0.010893329977989197, + 0.021174851804971695, + 0.012046091258525848, + 0.009037422947585583, + 0.014265517704188824, + 0.008877819404006004, + 0.015510808676481247, + 0.013641542755067348, + 0.007055155001580715, + 0.008877520449459553, + 0.006659830454736948, + 0.009825164452195168, + 0.01158459484577179, + 0.006498620845377445, + 0.008555058389902115, + 0.019514305517077446, + 0.010451377369463444, + 0.014123045839369297, + 0.010489210486412048, + 0.0143797742202878, + 0.01261522900313139, + 0.012912211008369923, + 0.014557050541043282, + 0.010539518669247627, + 0.006915128789842129, + 0.01341291144490242, + 0.010037513449788094, + 0.00982961431145668, + 0.007367220241576433, + 0.010728992521762848, + 0.010670436546206474, + 0.011707385070621967, + 0.00898468866944313, + 0.009782313369214535, + 0.007561416830867529, + 0.008090442977845669, + 0.004430614411830902, + 0.008659318089485168, + 0.00934786256402731, + 0.02039520815014839, + 0.01382860355079174, + 0.026076506823301315, + 0.01827366091310978, + 0.011061922647058964, + 0.013651778921484947, + 0.02165556140244007, + 0.021404318511486053, + 0.012072967365384102, + 0.004361653700470924, + 0.011349697597324848, + 0.006690970156341791, + 0.0037869513034820557, + 0.003672588849440217, + 0.007811268791556358, + 0.009502468630671501, + 0.006575834006071091, + 0.003741034772247076, + 0.003346045734360814, + 0.002862445544451475, + 0.0025135811883956194, + 0.0034407079219818115, + 0.004237310495227575, + 0.0038094092160463333, + 0.012691766954958439, + 0.013298694044351578, + 0.006652506534010172, + 0.007853920571506023, + 0.00826797354966402, + 0.0046465033665299416, + 0.011998211964964867, + 0.011478830128908157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.05397813767194748 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654209822416306 + ], + "opposite_logprob_tokens": [ + -3.048654317855835 + ], + "logprob_reward": -0.048654209822416306, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654209822416306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21349647641181946, + 0.10612817853689194, + 0.09450697153806686, + 0.09134766459465027, + 0.09138964116573334, + 0.14134669303894043, + 0.10212010145187378, + 0.06048055738210678, + 0.05262264609336853, + 0.04656105488538742 + ], + "credit_map": [ + 0.03208835422992706, + 0.021059373393654823, + 0.028495000675320625, + 0.03904931992292404, + 0.013446812517940998, + 0.031071066856384277, + 0.021468674764037132, + 0.026817869395017624, + 0.012257521972060204, + 0.006691988091915846, + 0.012868089601397514, + 0.010517210699617863, + 0.011550498194992542, + 0.019473116844892502, + 0.017250224947929382, + 0.015519536100327969, + 0.014717421494424343, + 0.005112845450639725, + 0.012390880845487118, + 0.010013890452682972, + 0.009838690049946308, + 0.01340168621391058, + 0.011709289625287056, + 0.017322272062301636, + 0.012003806419670582, + 0.009499426931142807, + 0.01140453852713108, + 0.010754186660051346, + 0.013285367749631405, + 0.011135324835777283, + 0.010818171314895153, + 0.012446838431060314, + 0.009244423359632492, + 0.008672161027789116, + 0.008564420975744724, + 0.011150489561259747, + 0.02004922367632389, + 0.00914692971855402, + 0.013527876697480679, + 0.011034121736884117, + 0.01232457160949707, + 0.020456871017813683, + 0.011541621759533882, + 0.018413208425045013, + 0.022896144539117813, + 0.010655784979462624, + 0.025295233353972435, + 0.019763240590691566, + 0.01394106075167656, + 0.011641575954854488, + 0.012003529816865921, + 0.013233691453933716, + 0.011528000235557556, + 0.009039822965860367, + 0.01666245050728321, + 0.014069970697164536, + 0.008953364565968513, + 0.0049469792284071445, + 0.011000431142747402, + 0.00682425033301115, + 0.004674091935157776, + 0.004775893874466419, + 0.009468336589634418, + 0.009837212972342968, + 0.009334898553788662, + 0.004919521044939756, + 0.011789643205702305, + 0.005732607562094927, + 0.003741795662790537, + 0.0036731273867189884, + 0.005136404652148485, + 0.008294651284813881, + 0.009385998360812664, + 0.008129067718982697, + 0.0046923658810555935, + 0.005191219504922628, + 0.0034479275345802307, + 0.002494976855814457, + 0.005679498892277479, + 0.0075399973429739475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.048654209822416306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327916383743286 + ], + "opposite_logprob_tokens": [ + -0.7577916383743286 + ], + "logprob_reward": -0.6327916383743286, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327916383743286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1768423169851303, + 0.28584879636764526, + 0.09096096456050873, + 0.08372659236192703, + 0.08506612479686737, + 0.09895798563957214, + 0.06605087965726852, + 0.021280929446220398, + 0.05755091831088066, + 0.03371444344520569 + ], + "credit_map": [ + 0.02301088720560074, + 0.014798359014093876, + 0.023338835686445236, + 0.03684479743242264, + 0.012516029179096222, + 0.02204173058271408, + 0.014233440160751343, + 0.030058253556489944, + 0.1002412810921669, + 0.02523273229598999, + 0.058684512972831726, + 0.017253093421459198, + 0.012648150324821472, + 0.01455761305987835, + 0.014133939519524574, + 0.043097492307424545, + 0.02440672554075718, + 0.007477851584553719, + 0.016102995723485947, + 0.007915211841464043, + 0.007369597442448139, + 0.006888454779982567, + 0.008058669045567513, + 0.012741461396217346, + 0.012749974615871906, + 0.006617275532335043, + 0.010839094407856464, + 0.009718691930174828, + 0.010868475772440434, + 0.00939980335533619, + 0.012025727890431881, + 0.011507549323141575, + 0.009279119782149792, + 0.008929712697863579, + 0.008832627907395363, + 0.01048341579735279, + 0.01645725965499878, + 0.008756312541663647, + 0.011644967831671238, + 0.010682720690965652, + 0.02333368919789791, + 0.012934738770127296, + 0.017814701423048973, + 0.008846729062497616, + 0.007859345525503159, + 0.004824159666895866, + 0.009132164530456066, + 0.01421246025711298, + 0.017604177817702293, + 0.0061902995221316814, + 0.008866565302014351, + 0.00621403194963932, + 0.004312705714255571, + 0.0032505737617611885, + 0.009501566179096699, + 0.010110960341989994, + 0.0030567466747015715, + 0.0027565935160964727, + 0.0027314734179526567, + 0.0024080711882561445, + 0.001698714797385037, + 0.0028301088605076075, + 0.002728402614593506, + 0.00307082268409431, + 0.011926491744816303, + 0.007414642721414566, + 0.006362439598888159, + 0.005019159056246281, + 0.003674197942018509, + 0.0035860033240169287, + 0.00878654234111309, + 0.010781444609165192, + 0.004965037107467651, + 0.0043173679150640965, + 0.004398711025714874, + 0.003589157946407795, + 0.0030797822400927544, + 0.002715602982789278, + 0.005225982051342726, + 0.005422803573310375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.6327916383743286 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099181387573481 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099181387573481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099181387573481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27130797505378723, + 0.2885919213294983, + 0.08142640441656113, + 0.06742605566978455, + 0.024974318221211433, + 0.04328792169690132, + 0.062165871262550354, + 0.027262821793556213, + 0.0905243307352066, + 0.04303237423300743 + ], + "credit_map": [ + 0.0614849254488945, + 0.021851468831300735, + 0.030159156769514084, + 0.04589210823178291, + 0.0197073295712471, + 0.025790754705667496, + 0.025943677872419357, + 0.04047856852412224, + 0.10345271229743958, + 0.02696363255381584, + 0.05366319790482521, + 0.018202796578407288, + 0.0135780218988657, + 0.013834265060722828, + 0.015343564562499523, + 0.04355374723672867, + 0.021035000681877136, + 0.008339675143361092, + 0.012186329811811447, + 0.00865994207561016, + 0.007303973194211721, + 0.007310796994715929, + 0.007610689848661423, + 0.008979997597634792, + 0.01640254259109497, + 0.00791163370013237, + 0.009587370790541172, + 0.00718945125117898, + 0.007064914796501398, + 0.0041669197380542755, + 0.006866743788123131, + 0.008236488327383995, + 0.0032268858049064875, + 0.002711112378165126, + 0.005674742162227631, + 0.0025981697253882885, + 0.002748124999925494, + 0.0014739569742232561, + 0.0028871428221464157, + 0.0036541817244142294, + 0.009826342575252056, + 0.004518205765634775, + 0.008067235350608826, + 0.004435848444700241, + 0.0025081790518015623, + 0.002375163836404681, + 0.004446772392839193, + 0.007110174745321274, + 0.00756335211917758, + 0.005790975876152515, + 0.01181719545274973, + 0.00797891616821289, + 0.005887835752218962, + 0.004987501539289951, + 0.008153100498020649, + 0.009986995719373226, + 0.004427261650562286, + 0.00279633654281497, + 0.00425406638532877, + 0.003376200795173645, + 0.002052054973319173, + 0.0021375934593379498, + 0.003672864055261016, + 0.004546441603451967, + 0.02413797564804554, + 0.011869789101183414, + 0.013552647083997726, + 0.007149219047278166, + 0.005664417520165443, + 0.005412989296019077, + 0.008782624267041683, + 0.013954665511846542, + 0.009644654579460621, + 0.007582665421068668, + 0.004047028720378876, + 0.004147181753069162, + 0.003962003625929356, + 0.0020233842078596354, + 0.0046363212168216705, + 0.006989140063524246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.004099181387573481 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21957460045814514, + 0.14660890400409698, + 0.07101867347955704, + 0.047602053731679916, + 0.051136087626218796, + 0.08473021537065506, + 0.06826614588499069, + 0.04492329806089401, + 0.04535919427871704, + 0.2207808941602707 + ], + "credit_map": [ + 0.0412028506398201, + 0.0198501143604517, + 0.02377123199403286, + 0.03921514004468918, + 0.014760411344468594, + 0.02755793184041977, + 0.024225501343607903, + 0.028991397470235825, + 0.04744097217917442, + 0.01189417764544487, + 0.027187008410692215, + 0.010223614983260632, + 0.008043735288083553, + 0.010519537143409252, + 0.009006998501718044, + 0.022292841225862503, + 0.010474598035216331, + 0.005077471025288105, + 0.01011300552636385, + 0.008670377545058727, + 0.006625561974942684, + 0.006790878716856241, + 0.009845254942774773, + 0.013421525247395039, + 0.006505852099508047, + 0.004060155246406794, + 0.005106550641357899, + 0.0057615311816334724, + 0.007096993271261454, + 0.00687007512897253, + 0.005754923913627863, + 0.0064459689892828465, + 0.006894285790622234, + 0.005133434198796749, + 0.005209657829254866, + 0.006214451976120472, + 0.009790024720132351, + 0.004506799392402172, + 0.007213745266199112, + 0.0061736879870295525, + 0.013009325601160526, + 0.00978157389909029, + 0.00935214664787054, + 0.010726488195359707, + 0.009366179816424847, + 0.004749957472085953, + 0.016855571419000626, + 0.010888967663049698, + 0.011865346692502499, + 0.006434986367821693, + 0.009891402907669544, + 0.00711661484092474, + 0.006564210169017315, + 0.004608370363712311, + 0.010264123789966106, + 0.011521084234118462, + 0.007009522523730993, + 0.00394506985321641, + 0.007717348635196686, + 0.005604325793683529, + 0.004543534014374018, + 0.0027753845788538456, + 0.006946970708668232, + 0.006381136365234852, + 0.01102471724152565, + 0.0033740231301635504, + 0.009687029756605625, + 0.0035864904057234526, + 0.0029471551533788443, + 0.0033274260349571705, + 0.004899452906101942, + 0.006512893829494715, + 0.032243262976408005, + 0.030478009954094887, + 0.0205906480550766, + 0.023011192679405212, + 0.02380390837788582, + 0.016632119193673134, + 0.03930743411183357, + 0.03471430391073227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008639576844871044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic and intense cinematic low-angle wide shot maintains a static-like view of a vast, sinister Gothic mansion, its form almost completely engulfed by the deep black of a stormy night. A fierce thunderstorm is in full swing, and thick rain pelts the building, obscuring the view. Suddenly, a massive flash of brilliant white lightning streaks across the turbulent, roiling purple-black sky, momentarily and violently illuminating the entire scene. In this fleeting, high-contrast flash, the mansion is revealed in stark detail: towering, needle-like spires reach toward the sky, weathered stone walls gleam damply, and eerie gargoyles seem to scream silently in the gusts of wind. Dark, tall lancet windows, like hollow eyes, reflect the searing flash before the scene returns to near-total darkness, leaving only a haunting, ominous silhouette against briefly ignited clouds. The intense atmosphere is filled with foreboding and the raw power of nature.", + "seed": 6673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_12/index_1156_seed_6673_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06042435779236257, + "vlm_logprob": -0.06042435779236257, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.06042435779236257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043615445494651794 + ], + "opposite_logprob_tokens": [ + -3.9186153411865234 + ], + "logprob_reward": -0.043615445494651794, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.043615445494651794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13511258363723755, + 0.03955621272325516, + 0.04121557995676994, + 0.49481692910194397, + 0.14788445830345154, + 0.019493937492370605, + 0.05066312849521637, + 0.02375158667564392, + 0.026425346732139587, + 0.021080171689391136 + ], + "credit_map": [ + 0.018267378211021423, + 0.01959739252924919, + 0.012833314947783947, + 0.014653502963483334, + 0.013040294870734215, + 0.017475321888923645, + 0.01477652508765459, + 0.024468857795000076, + 0.005998971872031689, + 0.0038952010218054056, + 0.005037964321672916, + 0.0032079441007226706, + 0.008340482600033283, + 0.004401056561619043, + 0.0050132665783166885, + 0.003661323804408312, + 0.004204099532216787, + 0.005896987859159708, + 0.004436288960278034, + 0.002807691227644682, + 0.006078121718019247, + 0.008264104835689068, + 0.004812311846762896, + 0.004715974908322096, + 0.11631283164024353, + 0.03575471043586731, + 0.04708189517259598, + 0.02205711230635643, + 0.04553687572479248, + 0.13126784563064575, + 0.028720617294311523, + 0.06808502227067947, + 0.0071042375639081, + 0.004693067632615566, + 0.006056797690689564, + 0.0047242408618330956, + 0.05474301800131798, + 0.052326008677482605, + 0.007779278792440891, + 0.010457812808454037, + 0.002204674296081066, + 0.0014402662636712193, + 0.0021497078705579042, + 0.0016943709924817085, + 0.0026144729927182198, + 0.003365885466337204, + 0.0030829920433461666, + 0.0029415686149150133, + 0.005394692532718182, + 0.005246052052825689, + 0.005076436325907707, + 0.005961822345852852, + 0.009083850309252739, + 0.005949568934738636, + 0.0066206250339746475, + 0.00733008049428463, + 0.003219638718292117, + 0.0018751700408756733, + 0.0020007190760225058, + 0.0020085980650037527, + 0.004214309621602297, + 0.005452642682939768, + 0.0027571942191570997, + 0.002223316114395857, + 0.0034746399614959955, + 0.0021037438418716192, + 0.00634721340611577, + 0.0033576206769794226, + 0.0023423549719154835, + 0.0031842149328440428, + 0.0032023994717746973, + 0.002413160167634487, + 0.0027932911179959774, + 0.001587621751241386, + 0.0033889999613165855, + 0.002281509805470705, + 0.0024796940851956606, + 0.0022841135505586863, + 0.0035762684419751167, + 0.0026886723935604095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.043615445494651794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865897819399834 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865897819399834, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865897819399834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3196401000022888, + 0.09194716811180115, + 0.08045273274183273, + 0.12278658151626587, + 0.060892254114151, + 0.07926910370588303, + 0.03773406893014908, + 0.06301204860210419, + 0.09890631586313248, + 0.04535958915948868 + ], + "credit_map": [ + 0.04658539593219757, + 0.05197331681847572, + 0.035889845341444016, + 0.030264990404248238, + 0.03422721102833748, + 0.041178274899721146, + 0.02928701788187027, + 0.05023405700922012, + 0.01524367369711399, + 0.00983935222029686, + 0.009638892486691475, + 0.0076495930552482605, + 0.01122664101421833, + 0.01710365153849125, + 0.009101128205657005, + 0.012144235894083977, + 0.007978075183928013, + 0.010502025485038757, + 0.012056753970682621, + 0.00812164694070816, + 0.00809619389474392, + 0.011853101663291454, + 0.012365083210170269, + 0.009479858912527561, + 0.026975499466061592, + 0.008427930064499378, + 0.013500687666237354, + 0.006331434473395348, + 0.012824933044612408, + 0.02904576249420643, + 0.008322343230247498, + 0.017357993870973587, + 0.005027738865464926, + 0.003410179167985916, + 0.006725895218551159, + 0.0035890196450054646, + 0.015127893537282944, + 0.015094766393303871, + 0.005844737868756056, + 0.006072022952139378, + 0.008670632727444172, + 0.005432615056633949, + 0.00805347878485918, + 0.010251530446112156, + 0.014846913516521454, + 0.010271843522787094, + 0.011887256987392902, + 0.009854829870164394, + 0.005788935348391533, + 0.003532146569341421, + 0.003322214586660266, + 0.004047770984470844, + 0.006630185525864363, + 0.005218661390244961, + 0.004336207173764706, + 0.004857948515564203, + 0.007824228145182133, + 0.006626092363148928, + 0.003657876281067729, + 0.005828411318361759, + 0.01593979075551033, + 0.011601868085563183, + 0.005449388176202774, + 0.006084398366510868, + 0.012310086749494076, + 0.01409235130995512, + 0.010997209697961807, + 0.010614396072924137, + 0.013097739778459072, + 0.01568487100303173, + 0.012153658084571362, + 0.009955997578799725, + 0.005087689030915499, + 0.005888377316296101, + 0.0044904835522174835, + 0.0048493011854588985, + 0.004456243012100458, + 0.0047955987975001335, + 0.009676055051386356, + 0.006115839816629887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.04865897819399834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158612728118896 + ], + "opposite_logprob_tokens": [ + -1.701586127281189 + ], + "logprob_reward": -0.20158612728118896, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158612728118896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08913884311914444, + 0.035744018852710724, + 0.1497204452753067, + 0.43592533469200134, + 0.11819907277822495, + 0.026974454522132874, + 0.03669169917702675, + 0.038850631564855576, + 0.04184480756521225, + 0.026910701766610146 + ], + "credit_map": [ + 0.010428495705127716, + 0.013408945873379707, + 0.009425118565559387, + 0.010306033305823803, + 0.012275070883333683, + 0.01161021925508976, + 0.01159591693431139, + 0.01008905004709959, + 0.005107831209897995, + 0.0040509975515306, + 0.0035848564002662897, + 0.0032983357086777687, + 0.0051238201558589935, + 0.006979512982070446, + 0.0035540133249014616, + 0.004044654779136181, + 0.01786845363676548, + 0.017885282635688782, + 0.014892476610839367, + 0.012524791993200779, + 0.02228102833032608, + 0.02479236200451851, + 0.022678257897496223, + 0.016797799617052078, + 0.10361719876527786, + 0.03157590702176094, + 0.04221133142709732, + 0.020033814013004303, + 0.04048748314380646, + 0.11361688375473022, + 0.02467878721654415, + 0.05970395356416702, + 0.007698670960962772, + 0.006319527979940176, + 0.00678324094042182, + 0.0057311528362333775, + 0.03793049976229668, + 0.035375043749809265, + 0.008765343576669693, + 0.009595599956810474, + 0.005141791421920061, + 0.003555145813152194, + 0.0026516185607761145, + 0.0022545617539435625, + 0.003412904217839241, + 0.0036259950138628483, + 0.0031943838112056255, + 0.0031380578875541687, + 0.004085126798599958, + 0.004771089646965265, + 0.0037492436822503805, + 0.0030226334929466248, + 0.006331780459731817, + 0.005864389706403017, + 0.004084371961653233, + 0.004783067852258682, + 0.004481546115130186, + 0.00293807708658278, + 0.0027030615601688623, + 0.002481048693880439, + 0.00952315516769886, + 0.010344339534640312, + 0.0023989800829440355, + 0.003980428446084261, + 0.004865902941673994, + 0.005866122432053089, + 0.00787085946649313, + 0.003911842126399279, + 0.00720644136890769, + 0.005180454347282648, + 0.0036211649421602488, + 0.0033220245968550444, + 0.00427382392808795, + 0.003460496198385954, + 0.002523699076846242, + 0.0020995470695197582, + 0.003885623998939991, + 0.004621349275112152, + 0.00311991386115551, + 0.002926248125731945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.20158612728118896 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001520430902019143 + ], + "opposite_logprob_tokens": [ + -6.50152063369751 + ], + "logprob_reward": -0.001520430902019143, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001520430902019143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06150076538324356, + 0.05977577716112137, + 0.12204113602638245, + 0.5286676287651062, + 0.1220492497086525, + 0.043084416538476944, + 0.018090764060616493, + 0.010761495679616928, + 0.023703740909695625, + 0.010325080715119839 + ], + "credit_map": [ + 0.007007280830293894, + 0.009044084697961807, + 0.0074159433133900166, + 0.008013997226953506, + 0.006171836983412504, + 0.007221338339149952, + 0.008585463277995586, + 0.008040815591812134, + 0.007099947426468134, + 0.0063507650047540665, + 0.007789588067680597, + 0.0064085377380251884, + 0.011202007532119751, + 0.006925723981112242, + 0.008084245957434177, + 0.00591495493426919, + 0.01601300574839115, + 0.01666911505162716, + 0.009511798620223999, + 0.010497365146875381, + 0.025194311514496803, + 0.019884567707777023, + 0.013170519843697548, + 0.01110044401139021, + 0.1228201612830162, + 0.03724047541618347, + 0.05010436475276947, + 0.022239534184336662, + 0.04902322590351105, + 0.14443619549274445, + 0.029178394004702568, + 0.07362522184848785, + 0.008936862461268902, + 0.005737835541367531, + 0.0061590890400111675, + 0.004897244740277529, + 0.044211771339178085, + 0.03660428151488304, + 0.007120454218238592, + 0.00838170200586319, + 0.0071846009232103825, + 0.005800130311399698, + 0.004803560208529234, + 0.003763761604204774, + 0.00652892142534256, + 0.005554290022701025, + 0.003933153580874205, + 0.005515997763723135, + 0.0025330327916890383, + 0.002862026449292898, + 0.0015732728643342853, + 0.0015008277259767056, + 0.002792673883959651, + 0.0036464177537709475, + 0.0015739452792331576, + 0.001608566613867879, + 0.0012952411780133843, + 0.0012163971550762653, + 0.0008689890382811427, + 0.0009182597277686, + 0.0024325239937752485, + 0.0021244040690362453, + 0.0007666415185667574, + 0.0011390381259843707, + 0.002906798617914319, + 0.0025719760451465845, + 0.004229035694152117, + 0.0024422972928732634, + 0.003034342313185334, + 0.0035519178491085768, + 0.00291724712587893, + 0.0020501252729445696, + 0.0015317213255912066, + 0.0009448035270906985, + 0.001127400086261332, + 0.00100501230917871, + 0.0012651875149458647, + 0.0015620243502780795, + 0.001460933592170477, + 0.0014279976021498442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.001520430902019143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1165880560874939, + 0.08150964975357056, + 0.05954704433679581, + 0.4186517298221588, + 0.10419599711894989, + 0.050328049808740616, + 0.04021468013525009, + 0.024211889132857323, + 0.06827998161315918, + 0.03647293895483017 + ], + "credit_map": [ + 0.012622685171663761, + 0.01713068038225174, + 0.013229680247604847, + 0.015708254650235176, + 0.01385669969022274, + 0.012823847122490406, + 0.01665719971060753, + 0.014558996073901653, + 0.009571492671966553, + 0.008007767610251904, + 0.009249650873243809, + 0.009097694419324398, + 0.014928518794476986, + 0.009935609064996243, + 0.01163558755069971, + 0.009083329699933529, + 0.006811744533479214, + 0.008080841973423958, + 0.00590538652613759, + 0.004720658995211124, + 0.008939451538026333, + 0.011378152295947075, + 0.007136547472327948, + 0.006574258673936129, + 0.09497957676649094, + 0.032045308500528336, + 0.04220292717218399, + 0.024385524913668633, + 0.04873722419142723, + 0.0831170529127121, + 0.02891022525727749, + 0.0642738789319992, + 0.005322822369635105, + 0.004322671797126532, + 0.004307306371629238, + 0.003617639420554042, + 0.0342821404337883, + 0.03905243054032326, + 0.004110279958695173, + 0.00918070413172245, + 0.0037096217274665833, + 0.00394295621663332, + 0.00590070616453886, + 0.005897517316043377, + 0.007640649564564228, + 0.006944466382265091, + 0.00926740188151598, + 0.007024727296084166, + 0.0048622083850204945, + 0.0045999870635569096, + 0.004093429539352655, + 0.003766557201743126, + 0.006233287509530783, + 0.0067154658026993275, + 0.004442465491592884, + 0.005501281004399061, + 0.0032148214522749186, + 0.004460596479475498, + 0.0020878631621599197, + 0.002216045744717121, + 0.004199107177555561, + 0.0027833550702780485, + 0.002802953589707613, + 0.002447144826874137, + 0.007364262826740742, + 0.007417342159897089, + 0.01047919038683176, + 0.00865188054740429, + 0.00875172484666109, + 0.007630711887031794, + 0.011128541082143784, + 0.006856322754174471, + 0.0045568617060780525, + 0.0028908380772918463, + 0.0063406312838196754, + 0.003926888573914766, + 0.0039802067913115025, + 0.003970766440033913, + 0.0056930482387542725, + 0.005113697145134211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.006740807089954615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic and intense cinematic low-angle wide shot maintains a static-like view of a vast, sinister Gothic mansion, its form almost completely engulfed by the deep black of a stormy night. A fierce thunderstorm is in full swing, and thick rain pelts the building, obscuring the view. Suddenly, a massive flash of brilliant white lightning streaks across the turbulent, roiling purple-black sky, momentarily and violently illuminating the entire scene. In this fleeting, high-contrast flash, the mansion is revealed in stark detail: towering, needle-like spires reach toward the sky, weathered stone walls gleam damply, and eerie gargoyles seem to scream silently in the gusts of wind. Dark, tall lancet windows, like hollow eyes, reflect the searing flash before the scene returns to near-total darkness, leaving only a haunting, ominous silhouette against briefly ignited clouds. The intense atmosphere is filled with foreboding and the raw power of nature.", + "seed": 6674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_12/index_1156_seed_6674_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2228562694042921, + "vlm_logprob": -0.2228562694042921, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.2228562694042921, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03205719590187073 + ], + "opposite_logprob_tokens": [ + -4.157057285308838 + ], + "logprob_reward": -0.03205719590187073, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03205719590187073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09260763972997665, + 0.032516151666641235, + 0.1415061503648758, + 0.10633522272109985, + 0.03722762316465378, + 0.05534506216645241, + 0.38065072894096375, + 0.06626305729150772, + 0.06994859129190445, + 0.01759977824985981 + ], + "credit_map": [ + 0.015007298439741135, + 0.007974687032401562, + 0.009281777776777744, + 0.008897939696907997, + 0.011766595765948296, + 0.017810599878430367, + 0.010277265682816505, + 0.011591475456953049, + 0.004855748265981674, + 0.003439111402258277, + 0.004364900756627321, + 0.0025841807946562767, + 0.00586319575086236, + 0.00503884581848979, + 0.0035277092829346657, + 0.002842458663508296, + 0.018694406375288963, + 0.02153698541224003, + 0.008001736365258694, + 0.01112570334225893, + 0.031476620584726334, + 0.017666731029748917, + 0.021831640973687172, + 0.011172324419021606, + 0.014019913971424103, + 0.013128024525940418, + 0.007467516697943211, + 0.011896313168108463, + 0.019385574385523796, + 0.014571390114724636, + 0.01671748049557209, + 0.009149021469056606, + 0.006630482152104378, + 0.005416185595095158, + 0.0038427533581852913, + 0.003981994930654764, + 0.004639865830540657, + 0.00593226682394743, + 0.002972665010020137, + 0.0038114122580736876, + 0.006730082910507917, + 0.005059374496340752, + 0.0093928137794137, + 0.006416816730052233, + 0.005033396650105715, + 0.007735061924904585, + 0.006385263986885548, + 0.008592255413532257, + 0.13552123308181763, + 0.023745769634842873, + 0.01914433017373085, + 0.019236279651522636, + 0.01470870990306139, + 0.028568409383296967, + 0.024239884689450264, + 0.11548613011837006, + 0.02319514751434326, + 0.0058125657960772514, + 0.003588890191167593, + 0.003209152491763234, + 0.0033807996660470963, + 0.0049043376930058, + 0.004399842116981745, + 0.01777232065796852, + 0.014686981216073036, + 0.006918895989656448, + 0.007842723280191422, + 0.006669313181191683, + 0.0058584269136190414, + 0.006485415622591972, + 0.011533396318554878, + 0.009953438304364681, + 0.004392601083964109, + 0.0013417219743132591, + 0.001543239806778729, + 0.001411059987731278, + 0.001575050293467939, + 0.0015829696785658598, + 0.0020997016690671444, + 0.0036534329410642385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.03205719590187073 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134394884109497 + ], + "opposite_logprob_tokens": [ + -1.3134394884109497 + ], + "logprob_reward": -0.3134394884109497, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134394884109497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0862966775894165, + 0.1322753131389618, + 0.11627451330423355, + 0.04891844466328621, + 0.09786569327116013, + 0.20562465488910675, + 0.22430956363677979, + 0.023636898025870323, + 0.055523116141557693, + 0.009275166317820549 + ], + "credit_map": [ + 0.012051880359649658, + 0.0072654131799936295, + 0.00963944848626852, + 0.007206298876553774, + 0.008270129561424255, + 0.018306046724319458, + 0.011484223417937756, + 0.01207323931157589, + 0.025753971189260483, + 0.009413816966116428, + 0.010282033123075962, + 0.0067823659628629684, + 0.015312272123992443, + 0.04079941287636757, + 0.011413834989070892, + 0.012517613358795643, + 0.015450578182935715, + 0.01952465996146202, + 0.013279054313898087, + 0.011178952641785145, + 0.020671561360359192, + 0.012702538631856441, + 0.0144654531031847, + 0.00900170486420393, + 0.005895471666008234, + 0.005307917948812246, + 0.005985709838569164, + 0.007474144455045462, + 0.005723634269088507, + 0.005860811565071344, + 0.007687120698392391, + 0.004983631428331137, + 0.01493883691728115, + 0.014317981898784637, + 0.009324872866272926, + 0.010744147934019566, + 0.008293744176626205, + 0.01738710328936577, + 0.012137418612837791, + 0.010721581988036633, + 0.03002052754163742, + 0.01924324221909046, + 0.027690039947628975, + 0.018479255959391594, + 0.03549693524837494, + 0.03201322257518768, + 0.020895713940262794, + 0.021785717457532883, + 0.07662896811962128, + 0.014175236225128174, + 0.012570059858262539, + 0.013183943927288055, + 0.009176547639071941, + 0.01703014224767685, + 0.0158078633248806, + 0.0657368153333664, + 0.006932676304131746, + 0.0020507806912064552, + 0.0023244949989020824, + 0.0017210778314620256, + 0.0014818395720794797, + 0.001984890317544341, + 0.0017863300163298845, + 0.005354809109121561, + 0.009214876219630241, + 0.006855919491499662, + 0.012767059728503227, + 0.005946456454694271, + 0.003969565965235233, + 0.004688140004873276, + 0.006189550273120403, + 0.005891548935323954, + 0.001975647872313857, + 0.0007556872442364693, + 0.0012664530659094453, + 0.000756875379011035, + 0.0007636996451765299, + 0.0011639150325208902, + 0.0008628562209196389, + 0.0017300316831097007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.3134394884109497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578034996986389 + ], + "opposite_logprob_tokens": [ + -0.6328034996986389 + ], + "logprob_reward": -0.7578034996986389, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578034996986389, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08211232721805573, + 0.1277783215045929, + 0.11347702145576477, + 0.17192910611629486, + 0.13767564296722412, + 0.09570135176181793, + 0.18998852372169495, + 0.03550038859248161, + 0.027348019182682037, + 0.01848929189145565 + ], + "credit_map": [ + 0.013277268037199974, + 0.015269915573298931, + 0.010719535872340202, + 0.008686508052051067, + 0.00877708662301302, + 0.009252444840967655, + 0.0075998455286026, + 0.00852971337735653, + 0.024750014767050743, + 0.011737562716007233, + 0.00984901562333107, + 0.00617334246635437, + 0.015695277601480484, + 0.03819834068417549, + 0.01064726710319519, + 0.01072748377919197, + 0.016580602154135704, + 0.01734456606209278, + 0.00967238750308752, + 0.008268876932561398, + 0.021540144458413124, + 0.012348179705440998, + 0.01872747763991356, + 0.008994780480861664, + 0.02352960966527462, + 0.02171740122139454, + 0.01065953727811575, + 0.01883704587817192, + 0.031527239829301834, + 0.020896371454000473, + 0.02781810238957405, + 0.016943776980042458, + 0.03689717873930931, + 0.012435032986104488, + 0.007529220078140497, + 0.009100322611629963, + 0.014064403250813484, + 0.01975702866911888, + 0.011670319363474846, + 0.026222117245197296, + 0.01582253910601139, + 0.006899765692651272, + 0.010914215818047523, + 0.00844530574977398, + 0.013367450796067715, + 0.00960290152579546, + 0.011986233294010162, + 0.01866293139755726, + 0.0657326802611351, + 0.012409216724336147, + 0.009210804477334023, + 0.009178485721349716, + 0.008517139591276646, + 0.01580449566245079, + 0.012919113039970398, + 0.05621657893061638, + 0.010703964158892632, + 0.004084327723830938, + 0.002482462441548705, + 0.002241698559373617, + 0.0020695137791335583, + 0.0028682618867605925, + 0.0024432125501334667, + 0.008606945164501667, + 0.0058034625835716724, + 0.0024330371525138617, + 0.003368127392604947, + 0.002205491531640291, + 0.002079644473269582, + 0.004660437814891338, + 0.002082488499581814, + 0.004715326242148876, + 0.005403767805546522, + 0.001540830358862877, + 0.001106279087252915, + 0.0012297277571633458, + 0.0009872405789792538, + 0.0022088123951107264, + 0.0013591600582003593, + 0.004653473384678364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.7578034996986389 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001190311275422573 + ], + "opposite_logprob_tokens": [ + -6.751190185546875 + ], + "logprob_reward": -0.001190311275422573, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001190311275422573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06691857427358627, + 0.2525223195552826, + 0.12338350713253021, + 0.17269770801067352, + 0.18599306046962738, + 0.03567797690629959, + 0.03889674320816994, + 0.061503298580646515, + 0.052406974136829376, + 0.009999776259064674 + ], + "credit_map": [ + 0.010334995575249195, + 0.010670206509530544, + 0.006261337082833052, + 0.008626489900052547, + 0.006941739935427904, + 0.007307692896574736, + 0.008759252727031708, + 0.008016861975193024, + 0.04845793917775154, + 0.02167271263897419, + 0.0190956499427557, + 0.012403149157762527, + 0.029625650495290756, + 0.08102446794509888, + 0.019236471503973007, + 0.021006284281611443, + 0.016131151467561722, + 0.013645743019878864, + 0.009615745395421982, + 0.012797178700566292, + 0.018062666058540344, + 0.02256513386964798, + 0.019307252019643784, + 0.011258639395236969, + 0.026150984689593315, + 0.02252877689898014, + 0.012620250694453716, + 0.010819828137755394, + 0.040134549140930176, + 0.029185162857174873, + 0.015181482769548893, + 0.016076678410172462, + 0.03842455521225929, + 0.01832674816250801, + 0.009847561828792095, + 0.01131993718445301, + 0.034325141459703445, + 0.03055853769183159, + 0.0160151869058609, + 0.02717541716992855, + 0.004495187662541866, + 0.0029016383923590183, + 0.00456974608823657, + 0.003249970730394125, + 0.00689922459423542, + 0.005218786187469959, + 0.004371762275695801, + 0.00397166283801198, + 0.007522956933826208, + 0.00480180699378252, + 0.0033381255343556404, + 0.0035068904981017113, + 0.0035709452349692583, + 0.006235731765627861, + 0.0035745101049542427, + 0.006345782428979874, + 0.02171683870255947, + 0.004217219538986683, + 0.003771686227992177, + 0.00295581785030663, + 0.0033339420333504677, + 0.0045835399068892, + 0.004620002117007971, + 0.0163042563945055, + 0.013605025596916676, + 0.0035210400819778442, + 0.006400883197784424, + 0.0037240544334053993, + 0.003508224617689848, + 0.0060255201533436775, + 0.004216259811073542, + 0.011405969969928265, + 0.001229270827025175, + 0.0007318155840039253, + 0.0014344523660838604, + 0.001103948918171227, + 0.0012495232513174415, + 0.0017455745255574584, + 0.0012210055720061064, + 0.001284186146222055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.001190311275422573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a wide shot showing the vastness of the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mansion engulfed by a stormy night with thick rain pelting the building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive flash of brilliant white lightning streak across a purple-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are towering needle-like spires and weathered stone walls visible during the lightning flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979085173457861 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.00979085173457861, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979085173457861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04734191671013832, + 0.19998258352279663, + 0.28159287571907043, + 0.05960804969072342, + 0.08427223563194275, + 0.020442422479391098, + 0.14717210829257965, + 0.07134091854095459, + 0.05952431261539459, + 0.028722578659653664 + ], + "credit_map": [ + 0.006302467547357082, + 0.006593633908778429, + 0.006067140493541956, + 0.004278445616364479, + 0.007065206300467253, + 0.005640789866447449, + 0.006242229603230953, + 0.0051520029082894325, + 0.037168871611356735, + 0.018241679295897484, + 0.016105055809020996, + 0.011942963115870953, + 0.022316042333841324, + 0.06107879430055618, + 0.016991203650832176, + 0.016137972474098206, + 0.03496013954281807, + 0.042941201478242874, + 0.016164541244506836, + 0.020918861031532288, + 0.06599461287260056, + 0.036879897117614746, + 0.04231363534927368, + 0.021419981494545937, + 0.008118411526083946, + 0.006817436311393976, + 0.006511457730084658, + 0.008055697195231915, + 0.00814012624323368, + 0.005846293643116951, + 0.00975024700164795, + 0.006368379574269056, + 0.013324971310794353, + 0.011624986305832863, + 0.006754450500011444, + 0.007165411952883005, + 0.012685001827776432, + 0.015701020136475563, + 0.008553095161914825, + 0.008463291451334953, + 0.002587144495919347, + 0.0016151699237525463, + 0.0029868073761463165, + 0.002488924190402031, + 0.0031694022472947836, + 0.0025257377419620752, + 0.002425143262371421, + 0.002644094405695796, + 0.046092599630355835, + 0.011705851182341576, + 0.014624213799834251, + 0.008590691722929478, + 0.008545388467609882, + 0.011135628446936607, + 0.00973702035844326, + 0.036740709096193314, + 0.02094862423837185, + 0.006614709738641977, + 0.004822925198823214, + 0.006247431039810181, + 0.004371080547571182, + 0.0052651166915893555, + 0.007362500298768282, + 0.01570853218436241, + 0.014422283507883549, + 0.0059713260270655155, + 0.006895977072417736, + 0.004911516793072224, + 0.003595806425437331, + 0.005579149816185236, + 0.0052693807519972324, + 0.012878868728876114, + 0.006892028264701366, + 0.002363454783335328, + 0.0029842383228242397, + 0.0027904522139579058, + 0.001953830709680915, + 0.0022704435978084803, + 0.003981329966336489, + 0.005486799404025078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.00979085173457861 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_tensor([1580, 1580])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_tensor([1580, 1580])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..33bb4fc714410fd11cfc6b7f9ae4941773afba68 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_tensor([1580, 1580])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a Ghibli-inspired art style, a high-speed slow-motion, low-angle, medium-length shot slowly zooms in on a majestic, ancient walnut tree, the centerpiece of a sun-dappled clearing. The tree's trunk is remarkably thick, richly painted with hand-painted textures, covered in deeply fissured gray-brown bark dotted with soft green moss. Its dense, gnarled branches spread outward like ancient arms, their vibrant green husks laden with clusters of walnuts. As the camera pans, the branches begin to vibrate gently and rhythmically, a slow yet powerful pulse derived not from the wind but from a magical, unseen energy. In exquisite slow motion, each lush leaf trembles and shimmers, capturing the warm, filtered sunlight, while the walnut itself begins to vibrate with a quiet intensity. A soft, inner golden light emanates from within, pulsating in sync with the tree's movement, revealing a quiet, powerful vitality within the scene and creating a serene, breathtakingly magical atmosphere.", + "seed": 6546, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_1580_seed_6546_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012918425258249044, + "vlm_logprob": -0.012918425258249044, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.012918425258249044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03741738572716713 + ], + "opposite_logprob_tokens": [ + -3.537417411804199 + ], + "logprob_reward": -0.03741738572716713, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03741738572716713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10253588855266571, + 0.07588661462068558, + 0.055528316646814346, + 0.036705661565065384, + 0.0797608494758606, + 0.017251785844564438, + 0.10273606330156326, + 0.418503075838089, + 0.022285455837845802, + 0.08880633115768433 + ], + "credit_map": [ + 0.013460791669785976, + 0.014867913909256458, + 0.007734461221843958, + 0.006414816714823246, + 0.016206732019782066, + 0.02562752179801464, + 0.010756394825875759, + 0.0074672517366707325, + 0.005255632568150759, + 0.016290908679366112, + 0.006349142175167799, + 0.004928919952362776, + 0.0092589370906353, + 0.02448447234928608, + 0.005305551458150148, + 0.00401304941624403, + 0.006678794045001268, + 0.0065278843976557255, + 0.007402637507766485, + 0.0060606347396969795, + 0.008313896134495735, + 0.0076120286248624325, + 0.007257246412336826, + 0.005675192456692457, + 0.005056047346442938, + 0.0043179308995604515, + 0.00329203181900084, + 0.004557653795927763, + 0.005077622830867767, + 0.006007709074765444, + 0.0047161695547401905, + 0.0036804971750825644, + 0.006037062965333462, + 0.004715808667242527, + 0.004507498349994421, + 0.004244130104780197, + 0.0077921729534864426, + 0.029702473431825638, + 0.01916654035449028, + 0.0035951603204011917, + 0.0016360003501176834, + 0.0030966640915721655, + 0.0013448027893900871, + 0.0031462344340980053, + 0.002034198259934783, + 0.002208983525633812, + 0.002616175217553973, + 0.0011687278747558594, + 0.01077873632311821, + 0.009384495206177235, + 0.019206037744879723, + 0.01699596457183361, + 0.012666582129895687, + 0.011910350993275642, + 0.013253627344965935, + 0.008540265262126923, + 0.03572860732674599, + 0.04670966416597366, + 0.03738640248775482, + 0.048993367701768875, + 0.06759396195411682, + 0.05081814527511597, + 0.08899877220392227, + 0.04227412864565849, + 0.0016272641951218247, + 0.0016613502521067858, + 0.0017608143389225006, + 0.0030534572433680296, + 0.004096725955605507, + 0.002921672770753503, + 0.004506194032728672, + 0.0026579778641462326, + 0.007004762999713421, + 0.005260645877569914, + 0.012189135886728764, + 0.01922035962343216, + 0.014486104249954224, + 0.011840930208563805, + 0.009532562457025051, + 0.009271833114326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03741738572716713 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647168796509504 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647168796509504, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647168796509504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15523479878902435, + 0.0626436322927475, + 0.11122769862413406, + 0.09694970399141312, + 0.11304411292076111, + 0.04551805555820465, + 0.07373945415019989, + 0.15287575125694275, + 0.06780777871608734, + 0.12095899134874344 + ], + "credit_map": [ + 0.014285796321928501, + 0.019968990236520767, + 0.016915401443839073, + 0.01113592367619276, + 0.029632670804858208, + 0.03489183634519577, + 0.014705752953886986, + 0.013698444701731205, + 0.00937903206795454, + 0.01104438491165638, + 0.008109834976494312, + 0.003984394017606974, + 0.00856119766831398, + 0.009777375496923923, + 0.005498414393514395, + 0.00628900621086359, + 0.010420880280435085, + 0.014398323372006416, + 0.012426658533513546, + 0.017235059291124344, + 0.013323506340384483, + 0.015244836919009686, + 0.01781328395009041, + 0.010365156456828117, + 0.018945762887597084, + 0.010784722864627838, + 0.00893399864435196, + 0.007843771018087864, + 0.010917345993220806, + 0.014188396744430065, + 0.010523420758545399, + 0.014812290668487549, + 0.010231471620500088, + 0.006267234683036804, + 0.004293690901249647, + 0.004383649677038193, + 0.01342698559165001, + 0.045000385493040085, + 0.018248507753014565, + 0.011192194186151028, + 0.0038336999714374542, + 0.013730352744460106, + 0.004683698993176222, + 0.005116681102663279, + 0.0038422264624387026, + 0.0062560876831412315, + 0.004877131897956133, + 0.00317817903123796, + 0.008133898489177227, + 0.008152724243700504, + 0.008592186495661736, + 0.009851492941379547, + 0.011107353493571281, + 0.011544576846063137, + 0.01215492095798254, + 0.004202303942292929, + 0.014736366458237171, + 0.01694067008793354, + 0.014111250638961792, + 0.016930025070905685, + 0.024967167526483536, + 0.020141631364822388, + 0.03009134531021118, + 0.014957304112613201, + 0.004348793998360634, + 0.004609584342688322, + 0.008100159466266632, + 0.009557303972542286, + 0.01023824792355299, + 0.009307444095611572, + 0.011786377988755703, + 0.009859871119260788, + 0.010878588072955608, + 0.008836725726723671, + 0.01876956969499588, + 0.019358858466148376, + 0.015492871403694153, + 0.017311112955212593, + 0.01633233018219471, + 0.013978949747979641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.004647168796509504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028228219598531723 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028228219598531723, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028228219598531723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11301930993795395, + 0.06014108285307884, + 0.065906822681427, + 0.05695081129670143, + 0.19601920247077942, + 0.11206797510385513, + 0.060663387179374695, + 0.10750026255846024, + 0.0628330186009407, + 0.164898082613945 + ], + "credit_map": [ + 0.010301739908754826, + 0.01393464207649231, + 0.013081533834338188, + 0.011115573346614838, + 0.017212433740496635, + 0.02503211237490177, + 0.008872730657458305, + 0.013468550518155098, + 0.005981823429465294, + 0.007051107008010149, + 0.008143451996147633, + 0.0053205532021820545, + 0.008708923123776913, + 0.008460016921162605, + 0.007696807384490967, + 0.008778400719165802, + 0.010047734715044498, + 0.013620657846331596, + 0.0073784273117780685, + 0.005397966597229242, + 0.009774439968168736, + 0.008627103641629219, + 0.005890359170734882, + 0.005170134361833334, + 0.013231717981398106, + 0.006714498158544302, + 0.005590250249952078, + 0.006251696962863207, + 0.005622954107820988, + 0.0056521170772612095, + 0.00642234506085515, + 0.007465233560651541, + 0.019130565226078033, + 0.009846449829638004, + 0.0060790530405938625, + 0.00663178414106369, + 0.021281680092215538, + 0.08240494877099991, + 0.03529472276568413, + 0.015350007452070713, + 0.009850450791418552, + 0.02996380813419819, + 0.010026124306023121, + 0.010539418086409569, + 0.010868743062019348, + 0.019243916496634483, + 0.011588783003389835, + 0.009986723773181438, + 0.004709102213382721, + 0.00858814176172018, + 0.004154708702117205, + 0.008793311193585396, + 0.010576847940683365, + 0.010222525335848331, + 0.008910881355404854, + 0.004707865417003632, + 0.008320153690874577, + 0.012146898545324802, + 0.010131091810762882, + 0.012905169278383255, + 0.01728527806699276, + 0.01192331500351429, + 0.023531701415777206, + 0.011256652884185314, + 0.004696091171354055, + 0.003849621396511793, + 0.007068379316478968, + 0.0074218567460775375, + 0.011186463758349419, + 0.01222843024879694, + 0.008078192360699177, + 0.008303986862301826, + 0.013619255274534225, + 0.012318260036408901, + 0.02402813360095024, + 0.027078228071331978, + 0.025257306173443794, + 0.020775744691491127, + 0.024647235870361328, + 0.01717393286526203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.0028228219598531723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072972789406776 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011072972789406776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072972789406776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10671154409646988, + 0.0651707574725151, + 0.0405883751809597, + 0.06721644103527069, + 0.07215751707553864, + 0.020881816744804382, + 0.08349800854921341, + 0.17306400835514069, + 0.034539032727479935, + 0.3361724317073822 + ], + "credit_map": [ + 0.0097898468375206, + 0.01691502518951893, + 0.008206907659769058, + 0.006234670057892799, + 0.02062583528459072, + 0.026327943429350853, + 0.011028137989342213, + 0.007583190221339464, + 0.0058525484055280685, + 0.008237842470407486, + 0.010768000036478043, + 0.005531148985028267, + 0.009005303494632244, + 0.010852483101189137, + 0.006282207556068897, + 0.008641233667731285, + 0.0037509340327233076, + 0.00542809721082449, + 0.006033269688487053, + 0.0047639948315918446, + 0.004936290439218283, + 0.004823838826268911, + 0.00518953287974, + 0.005662421230226755, + 0.013092972338199615, + 0.0071822223253548145, + 0.00667321914806962, + 0.004815137013792992, + 0.008433230221271515, + 0.01022628229111433, + 0.007627546787261963, + 0.009165842086076736, + 0.00786853488534689, + 0.0042951530776917934, + 0.0029760238248854876, + 0.003714895574375987, + 0.008367258124053478, + 0.028108198195695877, + 0.010237148962914944, + 0.00659031281247735, + 0.0019878011662513018, + 0.0021930367220193148, + 0.0020579220727086067, + 0.0032950222957879305, + 0.0030902789440006018, + 0.003857990959659219, + 0.002718877512961626, + 0.0016808888176456094, + 0.01070946641266346, + 0.009662941098213196, + 0.014138449914753437, + 0.01207241415977478, + 0.011097736656665802, + 0.010152571834623814, + 0.00974856223911047, + 0.005915873218327761, + 0.011433880776166916, + 0.019941648468375206, + 0.01868932880461216, + 0.02186349593102932, + 0.02756132185459137, + 0.017950160428881645, + 0.0360814705491066, + 0.019542718306183815, + 0.0026148466859012842, + 0.004389001056551933, + 0.003343178192153573, + 0.004405029583722353, + 0.003991453442722559, + 0.003267370630055666, + 0.008799399249255657, + 0.003728758543729782, + 0.028397783637046814, + 0.02639123797416687, + 0.054231975227594376, + 0.053280215710401535, + 0.045357976108789444, + 0.03984586149454117, + 0.0490025095641613, + 0.03966491296887398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.011072972789406776 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863177701830864 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.00863177701830864, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863177701830864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1577431708574295, + 0.10040519386529922, + 0.0513102188706398, + 0.06218433380126953, + 0.0808267891407013, + 0.17368660867214203, + 0.06870587170124054, + 0.15830367803573608, + 0.08573299646377563, + 0.06110115721821785 + ], + "credit_map": [ + 0.017483050003647804, + 0.01844426989555359, + 0.017857970669865608, + 0.016288727521896362, + 0.02629968710243702, + 0.03287792578339577, + 0.014667393639683723, + 0.013824142515659332, + 0.007513783872127533, + 0.023877672851085663, + 0.006841519381850958, + 0.004319183994084597, + 0.009684228338301182, + 0.03838896006345749, + 0.004718982148915529, + 0.005060852039605379, + 0.0064102946780622005, + 0.009106348268687725, + 0.006459863390773535, + 0.005660237278789282, + 0.005836840253323317, + 0.006433713715523481, + 0.006820566486567259, + 0.004582355730235577, + 0.008055347017943859, + 0.007368793711066246, + 0.0053405980579555035, + 0.010238965973258018, + 0.008639497682452202, + 0.008268552832305431, + 0.007669487036764622, + 0.0066030919551849365, + 0.007580386940389872, + 0.007646440528333187, + 0.008672048337757587, + 0.00672577228397131, + 0.008695065975189209, + 0.019302962347865105, + 0.008561909198760986, + 0.013642198406159878, + 0.01868472434580326, + 0.0523502342402935, + 0.014712406322360039, + 0.013279006816446781, + 0.016017913818359375, + 0.02614494413137436, + 0.017315514385700226, + 0.015181858092546463, + 0.007951409555971622, + 0.009596199728548527, + 0.00854327343404293, + 0.009855994954705238, + 0.00920794066041708, + 0.00960291177034378, + 0.009300310164690018, + 0.004647829569876194, + 0.009883216582238674, + 0.02157668024301529, + 0.021151216700673103, + 0.019101204350590706, + 0.020250733941793442, + 0.02043897658586502, + 0.030091429129242897, + 0.015810225158929825, + 0.008097860030829906, + 0.010232262313365936, + 0.01048788521438837, + 0.011753888800740242, + 0.011644314043223858, + 0.00875898078083992, + 0.013926173560321331, + 0.010831641033291817, + 0.00587200652807951, + 0.004423009697347879, + 0.009108981117606163, + 0.010350760072469711, + 0.008034661412239075, + 0.009176518768072128, + 0.0077758231200277805, + 0.006359399762004614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.00863177701830864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a Ghibli-inspired art style, a high-speed slow-motion, low-angle, medium-length shot slowly zooms in on a majestic, ancient walnut tree, the centerpiece of a sun-dappled clearing. The tree's trunk is remarkably thick, richly painted with hand-painted textures, covered in deeply fissured gray-brown bark dotted with soft green moss. Its dense, gnarled branches spread outward like ancient arms, their vibrant green husks laden with clusters of walnuts. As the camera pans, the branches begin to vibrate gently and rhythmically, a slow yet powerful pulse derived not from the wind but from a magical, unseen energy. In exquisite slow motion, each lush leaf trembles and shimmers, capturing the warm, filtered sunlight, while the walnut itself begins to vibrate with a quiet intensity. A soft, inner golden light emanates from within, pulsating in sync with the tree's movement, revealing a quiet, powerful vitality within the scene and creating a serene, breathtakingly magical atmosphere.", + "seed": 6547, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_1580_seed_6547_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009173234214540572, + "vlm_logprob": -0.009173234214540572, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.009173234214540572, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021300017833709717 + ], + "opposite_logprob_tokens": [ + -4.146299839019775 + ], + "logprob_reward": -0.021300017833709717, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021300017833709717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6118484735488892, + 0.02175276353955269, + 0.03100997768342495, + 0.0312744602560997, + 0.025177210569381714, + 0.07446271926164627, + 0.06477571278810501, + 0.025395041331648827, + 0.019849199801683426, + 0.09445448219776154 + ], + "credit_map": [ + 0.1955893337726593, + 0.07019170373678207, + 0.024351058527827263, + 0.03654925897717476, + 0.02929598093032837, + 0.20664964616298676, + 0.02557186223566532, + 0.023649612441658974, + 0.0028985324315726757, + 0.0027156888972967863, + 0.0013675427762791514, + 0.0035732146352529526, + 0.0035882845986634493, + 0.002327179303392768, + 0.003698519663885236, + 0.0015838021645322442, + 0.00429814588278532, + 0.00494497362524271, + 0.0025265098083764315, + 0.004773539490997791, + 0.0031080269254744053, + 0.004982333164662123, + 0.0036191511899232864, + 0.0027572973631322384, + 0.0053819105960428715, + 0.0068630799651145935, + 0.0038975754287093878, + 0.0033865540754050016, + 0.0025199041701853275, + 0.0039135245606303215, + 0.002874811878427863, + 0.002437099814414978, + 0.0025172156747430563, + 0.0034732792992144823, + 0.003577678231522441, + 0.0033372596371918917, + 0.002683900063857436, + 0.002979912795126438, + 0.0029983441345393658, + 0.0036096200346946716, + 0.009177204221487045, + 0.012742661871016026, + 0.009442239068448544, + 0.008275317028164864, + 0.00876135379076004, + 0.015687085688114166, + 0.006455790251493454, + 0.00392106594517827, + 0.005202222149819136, + 0.0077343289740383625, + 0.009389095939695835, + 0.008393430151045322, + 0.006736028008162975, + 0.004031323827803135, + 0.020006943494081497, + 0.0032823418732732534, + 0.002433499787002802, + 0.004391308408230543, + 0.002355024218559265, + 0.0038976031355559826, + 0.0019780199509114027, + 0.003770500887185335, + 0.005238417536020279, + 0.0013306664768606424, + 0.0017753608990460634, + 0.003218686208128929, + 0.002090025693178177, + 0.0029396014288067818, + 0.0026695907581597567, + 0.0025921158958226442, + 0.003397369058802724, + 0.0011664496269077063, + 0.005816207733005285, + 0.007899867370724678, + 0.026507558301091194, + 0.014808128587901592, + 0.004017294850200415, + 0.006376269273459911, + 0.02565736509859562, + 0.0033717916812747717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.021300017833709717 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067406888119876385 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.0067406888119876385, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067406888119876385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21946388483047485, + 0.02681436389684677, + 0.041624926030635834, + 0.008057275786995888, + 0.18829146027565002, + 0.028121668845415115, + 0.022653797641396523, + 0.24434788525104523, + 0.014759261161088943, + 0.20586542785167694 + ], + "credit_map": [ + 0.07654380053281784, + 0.024508144706487656, + 0.008437327109277248, + 0.011516821570694447, + 0.009595248848199844, + 0.07179450243711472, + 0.00849953480064869, + 0.008568505756556988, + 0.004256102722138166, + 0.004458564333617687, + 0.001768251764588058, + 0.0038165398873388767, + 0.0042389146983623505, + 0.002874263795092702, + 0.0032934146001935005, + 0.0021083110477775335, + 0.005774172022938728, + 0.006947777234017849, + 0.002901253988966346, + 0.005300705786794424, + 0.005312362220138311, + 0.006507260259240866, + 0.004977092146873474, + 0.003904301207512617, + 0.0016839888412505388, + 0.0010237274691462517, + 0.001304873381741345, + 0.0011800231877714396, + 0.0006890409858897328, + 0.0006160089396871626, + 0.0010604829294607043, + 0.0004991297028027475, + 0.018622569739818573, + 0.02302108332514763, + 0.028638090938329697, + 0.027875667437911034, + 0.021340737119317055, + 0.02188878133893013, + 0.02242579124867916, + 0.024478742852807045, + 0.0033435355871915817, + 0.005056528840214014, + 0.00475334282964468, + 0.0021793865598738194, + 0.002602464286610484, + 0.006044505629688501, + 0.0019926493987441063, + 0.0021492557134479284, + 0.0020844608079642057, + 0.004240710288286209, + 0.002980610355734825, + 0.0024287232663482428, + 0.0019657877273857594, + 0.0019313391530886292, + 0.005773829761892557, + 0.0012483353493735194, + 0.01092864852398634, + 0.02454974502325058, + 0.012567508965730667, + 0.08798286318778992, + 0.01082541048526764, + 0.007083033677190542, + 0.08323311060667038, + 0.007177552208304405, + 0.0012589569669216871, + 0.002024528570473194, + 0.0018880880670621991, + 0.002187101636081934, + 0.0017277520382776856, + 0.002288096584379673, + 0.002432632725685835, + 0.0009521027095615864, + 0.01310590747743845, + 0.015616721473634243, + 0.05697024613618851, + 0.03319452702999115, + 0.007875214330852032, + 0.01296989619731903, + 0.059152357280254364, + 0.006980560719966888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0067406888119876385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005615564878098667 + ], + "opposite_logprob_tokens": [ + -7.500561714172363 + ], + "logprob_reward": -0.0005615564878098667, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005615564878098667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1119111031293869, + 0.03782760724425316, + 0.09912711381912231, + 0.022813228890299797, + 0.023986466228961945, + 0.16553419828414917, + 0.051860593259334564, + 0.18297742307186127, + 0.07587847858667374, + 0.22808368504047394 + ], + "credit_map": [ + 0.023503975942730904, + 0.02428906410932541, + 0.007639727089554071, + 0.01105018611997366, + 0.011966164223849773, + 0.017504699528217316, + 0.00892581045627594, + 0.0070314835757017136, + 0.0055043441243469715, + 0.005603121593594551, + 0.0024591442197561264, + 0.005509600974619389, + 0.005643142387270927, + 0.0043412791565060616, + 0.0060460832901299, + 0.0027208938263356686, + 0.013023714534938335, + 0.01244866382330656, + 0.007471139542758465, + 0.013539925217628479, + 0.010605700314044952, + 0.018981309607625008, + 0.016086723655462265, + 0.006969950161874294, + 0.0044235894456505775, + 0.0040844641625881195, + 0.0021067841444164515, + 0.0025808101054280996, + 0.002151082968339324, + 0.003150070086121559, + 0.002033333759754896, + 0.002283097244799137, + 0.0027653833385556936, + 0.0027805038262158632, + 0.002265441929921508, + 0.004180162213742733, + 0.0026798746548593044, + 0.0022278206888586283, + 0.0036049627233296633, + 0.0034823191817849874, + 0.022523043677210808, + 0.028514588251709938, + 0.020626787096261978, + 0.011956446804106236, + 0.014126519672572613, + 0.03720219060778618, + 0.01362860482186079, + 0.016956036910414696, + 0.004998771473765373, + 0.007298145908862352, + 0.007396416738629341, + 0.006545422598719597, + 0.004648576956242323, + 0.0034939462784677744, + 0.014610232785344124, + 0.0028690809849649668, + 0.008439376950263977, + 0.018697623163461685, + 0.009655298665165901, + 0.06511541455984116, + 0.009427004493772984, + 0.008747601881623268, + 0.05758913978934288, + 0.0053059873171150684, + 0.0072790938429534435, + 0.010867966338992119, + 0.008695286698639393, + 0.011716523207724094, + 0.010551040060818195, + 0.009769641794264317, + 0.012888149358332157, + 0.004110781010240316, + 0.014623666182160378, + 0.016102558001875877, + 0.06919444352388382, + 0.032089587301015854, + 0.008320224471390247, + 0.014908700250089169, + 0.06610026210546494, + 0.006744252983480692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0005615564878098667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635558187961578 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635558187961578, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635558187961578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0732450932264328, + 0.03814753517508507, + 0.15662623941898346, + 0.03298189491033554, + 0.0960552990436554, + 0.05487433448433876, + 0.09366856515407562, + 0.2156912237405777, + 0.026746951043605804, + 0.21196284890174866 + ], + "credit_map": [ + 0.0111243836581707, + 0.01459207572042942, + 0.004715818911790848, + 0.006767455954104662, + 0.008670092560350895, + 0.016077930107712746, + 0.006612330209463835, + 0.004685006570070982, + 0.006363443098962307, + 0.009307833388447762, + 0.0029940989334136248, + 0.004154508002102375, + 0.004133146721869707, + 0.0060208323411643505, + 0.002815770683810115, + 0.0023579010739922523, + 0.023163096979260445, + 0.029322315007448196, + 0.011106675490736961, + 0.020856531336903572, + 0.017186453565955162, + 0.0229153074324131, + 0.017115669324994087, + 0.014960181899368763, + 0.004304319620132446, + 0.006620633881539106, + 0.003684745403006673, + 0.004915089346468449, + 0.002184166107326746, + 0.004209426231682301, + 0.00466324994340539, + 0.0024002643767744303, + 0.009308254346251488, + 0.010963983833789825, + 0.016440030187368393, + 0.012578493915498257, + 0.011946948245167732, + 0.01031279843300581, + 0.01153983548283577, + 0.012964962981641293, + 0.006199196446686983, + 0.009994872845709324, + 0.007716563995927572, + 0.0063376519829034805, + 0.005108517594635487, + 0.005977180786430836, + 0.007611312437802553, + 0.0059290421195328236, + 0.008078020066022873, + 0.013134211301803589, + 0.01368511002510786, + 0.014849214814603329, + 0.013531159609556198, + 0.00863510463386774, + 0.014352410100400448, + 0.007403337862342596, + 0.010608742013573647, + 0.022433258593082428, + 0.012575170025229454, + 0.07680493593215942, + 0.009426445700228214, + 0.00737181818112731, + 0.06966377049684525, + 0.00680709769949317, + 0.0028845123015344143, + 0.004136919975280762, + 0.003511699615046382, + 0.004160982556641102, + 0.0036057850811630487, + 0.0037131772842258215, + 0.003231640439480543, + 0.001502236002124846, + 0.013751570135354996, + 0.017737509682774544, + 0.06600792706012726, + 0.028735708445310593, + 0.008876356296241283, + 0.014659225009381771, + 0.05420614778995514, + 0.007988396100699902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.008635558187961578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008628349751234055 + ], + "opposite_logprob_tokens": [ + -4.7586283683776855 + ], + "logprob_reward": -0.008628349751234055, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008628349751234055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14833062887191772, + 0.03060426563024521, + 0.1457493156194687, + 0.026745636016130447, + 0.2681107223033905, + 0.07483925670385361, + 0.11622582376003265, + 0.09530583024024963, + 0.05778859555721283, + 0.03629998490214348 + ], + "credit_map": [ + 0.031188536435365677, + 0.033172015100717545, + 0.01073669083416462, + 0.014927254058420658, + 0.015716947615146637, + 0.022750040516257286, + 0.011214562691748142, + 0.008624586276710033, + 0.004863397683948278, + 0.00532497838139534, + 0.002123676473274827, + 0.003769627073779702, + 0.005334137007594109, + 0.004214828368276358, + 0.0033096105325967073, + 0.0016640092944726348, + 0.02277039736509323, + 0.027830665931105614, + 0.009980502538383007, + 0.019639529287815094, + 0.016807330772280693, + 0.020084766671061516, + 0.01540555153042078, + 0.013230560347437859, + 0.0057961898855865, + 0.0039004897698760033, + 0.0025140440557152033, + 0.0038347558584064245, + 0.0021170226391404867, + 0.0031082462519407272, + 0.003160761436447501, + 0.0023141263518482447, + 0.026748377829790115, + 0.030449124053120613, + 0.04119820147752762, + 0.0380035899579525, + 0.03342391178011894, + 0.02991551160812378, + 0.03292630985379219, + 0.0354456752538681, + 0.008180546574294567, + 0.010061273351311684, + 0.00983139593154192, + 0.008993030525743961, + 0.008457492105662823, + 0.013463851064443588, + 0.008985726162791252, + 0.006865943316370249, + 0.007608757819980383, + 0.015541809611022472, + 0.019825175404548645, + 0.01743142120540142, + 0.013302401639521122, + 0.0052589657716453075, + 0.03205426037311554, + 0.00520302914083004, + 0.004950340371578932, + 0.011379183270037174, + 0.005319320596754551, + 0.03121108189225197, + 0.004722749814391136, + 0.00461601885035634, + 0.029943620786070824, + 0.003163509303703904, + 0.004967713728547096, + 0.010084383189678192, + 0.006420237012207508, + 0.007458108477294445, + 0.006760927848517895, + 0.007721873931586742, + 0.011905105784535408, + 0.0024702458176761866, + 0.0028371866792440414, + 0.0031362995505332947, + 0.010399139486253262, + 0.00407306756824255, + 0.0023195806425064802, + 0.0026323110796511173, + 0.008630778640508652, + 0.0022716207895427942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.008628349751234055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a Ghibli-inspired art style, a high-speed slow-motion, low-angle, medium-length shot slowly zooms in on a majestic, ancient walnut tree, the centerpiece of a sun-dappled clearing. The tree's trunk is remarkably thick, richly painted with hand-painted textures, covered in deeply fissured gray-brown bark dotted with soft green moss. Its dense, gnarled branches spread outward like ancient arms, their vibrant green husks laden with clusters of walnuts. As the camera pans, the branches begin to vibrate gently and rhythmically, a slow yet powerful pulse derived not from the wind but from a magical, unseen energy. In exquisite slow motion, each lush leaf trembles and shimmers, capturing the warm, filtered sunlight, while the walnut itself begins to vibrate with a quiet intensity. A soft, inner golden light emanates from within, pulsating in sync with the tree's movement, revealing a quiet, powerful vitality within the scene and creating a serene, breathtakingly magical atmosphere.", + "seed": 6548, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_1580_seed_6548_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007041200553067029, + "vlm_logprob": -0.007041200553067029, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.007041200553067029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017796946689486504 + ], + "opposite_logprob_tokens": [ + -4.392796993255615 + ], + "logprob_reward": -0.017796946689486504, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017796946689486504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14987024664878845, + 0.051403872668743134, + 0.030583055689930916, + 0.0735371932387352, + 0.08210300654172897, + 0.02543041855096817, + 0.06742224097251892, + 0.07642966508865356, + 0.08693877607584, + 0.35628148913383484 + ], + "credit_map": [ + 0.02471896819770336, + 0.017680687829852104, + 0.011103294789791107, + 0.013574439100921154, + 0.03323417529463768, + 0.024759963154792786, + 0.013166191056370735, + 0.011632513254880905, + 0.00999145582318306, + 0.004636151250451803, + 0.0032276296988129616, + 0.006435131188482046, + 0.0061380499973893166, + 0.009393933229148388, + 0.005983682814985514, + 0.005597835406661034, + 0.0034873480908572674, + 0.003899573814123869, + 0.002484031720086932, + 0.004052328877151012, + 0.003736412152647972, + 0.003885151818394661, + 0.005453373771160841, + 0.003584835911169648, + 0.010335604660212994, + 0.014282715506851673, + 0.009136602282524109, + 0.008790450170636177, + 0.007938726805150509, + 0.00854562222957611, + 0.007728811353445053, + 0.006778656505048275, + 0.009856516495347023, + 0.010467981919646263, + 0.005653679836541414, + 0.019421983510255814, + 0.005118705332279205, + 0.0167506393045187, + 0.0067258914932608604, + 0.008107603527605534, + 0.0028778824489563704, + 0.0027482230216264725, + 0.002229368779808283, + 0.003316225716844201, + 0.00369405304081738, + 0.003566946368664503, + 0.0026401064824312925, + 0.004357611294835806, + 0.00619832519441843, + 0.023817984387278557, + 0.0043971948325634, + 0.011186229065060616, + 0.004560953006148338, + 0.009865215048193932, + 0.003370871301740408, + 0.004025470465421677, + 0.009494415484368801, + 0.010374393314123154, + 0.004029378294944763, + 0.024746200069785118, + 0.005589094944298267, + 0.004099192097783089, + 0.013312175869941711, + 0.004784814547747374, + 0.01683003455400467, + 0.007891783490777016, + 0.006193199194967747, + 0.007712989579886198, + 0.008496307767927647, + 0.023659398779273033, + 0.008764472790062428, + 0.00739058805629611, + 0.03199110925197601, + 0.03665642440319061, + 0.010500405915081501, + 0.01543784886598587, + 0.10594584047794342, + 0.12539511919021606, + 0.013433519750833511, + 0.016921229660511017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.017796946689486504 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203538315370679 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.002203538315370679, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203538315370679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10722334682941437, + 0.10228890180587769, + 0.04312070831656456, + 0.03795063495635986, + 0.17579445242881775, + 0.08571916818618774, + 0.07591278851032257, + 0.055878762155771255, + 0.04764893651008606, + 0.26846233010292053 + ], + "credit_map": [ + 0.013744322583079338, + 0.014273781329393387, + 0.008884413167834282, + 0.00920415110886097, + 0.021059399470686913, + 0.0213264599442482, + 0.008086162619292736, + 0.01064466591924429, + 0.019324997439980507, + 0.01809621974825859, + 0.005802484229207039, + 0.012864950112998486, + 0.01248355582356453, + 0.020358985289931297, + 0.00703319301828742, + 0.0063245221972465515, + 0.005304878577589989, + 0.005500799976289272, + 0.003939508926123381, + 0.006176746916025877, + 0.005709684919565916, + 0.005633621010929346, + 0.005945569835603237, + 0.004909903276711702, + 0.005644822493195534, + 0.004757506772875786, + 0.0068729342892766, + 0.00465260399505496, + 0.004044820088893175, + 0.004027439281344414, + 0.0037678119260817766, + 0.004182695411145687, + 0.022561222314834595, + 0.019190117716789246, + 0.009272050112485886, + 0.04657210409641266, + 0.01028949860483408, + 0.03857110068202019, + 0.01272282563149929, + 0.01661553978919983, + 0.010277575813233852, + 0.013243511319160461, + 0.005876325536519289, + 0.013287743553519249, + 0.00864141434431076, + 0.019588572904467583, + 0.00826296117156744, + 0.00654106680303812, + 0.01047410350292921, + 0.011147993616759777, + 0.006934949662536383, + 0.015694210305809975, + 0.0051426575519144535, + 0.011084165424108505, + 0.007300237659364939, + 0.008134477771818638, + 0.009233176708221436, + 0.008127995766699314, + 0.0031139571219682693, + 0.019460830837488174, + 0.002742048352956772, + 0.0035185927990823984, + 0.005064171273261309, + 0.0046179951168596745, + 0.00928712822496891, + 0.0031309088226407766, + 0.004490610212087631, + 0.004842832684516907, + 0.003813775023445487, + 0.013462215662002563, + 0.0045877075754106045, + 0.0040337624959647655, + 0.025077765807509422, + 0.02711726725101471, + 0.007804648019373417, + 0.013363149017095566, + 0.07885147631168365, + 0.0915728285908699, + 0.010340956971049309, + 0.014334243722259998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.002203538315370679 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013387774815782905 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013387774815782905, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013387774815782905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10292895138263702, + 0.08849518746137619, + 0.04144952818751335, + 0.04427766799926758, + 0.18279381096363068, + 0.11030969768762589, + 0.08206332474946976, + 0.03343556821346283, + 0.03287374600768089, + 0.2813724875450134 + ], + "credit_map": [ + 0.013635984621942043, + 0.009477640502154827, + 0.008775213733315468, + 0.011966630816459656, + 0.018704107031226158, + 0.020333128049969673, + 0.011832610704004765, + 0.00820363312959671, + 0.017634114250540733, + 0.016053995117545128, + 0.00525220250710845, + 0.008003346621990204, + 0.009386304765939713, + 0.01800490729510784, + 0.007753865327686071, + 0.006406445987522602, + 0.0057337102480232716, + 0.005373331252485514, + 0.004031732212752104, + 0.004506489261984825, + 0.005114527884870768, + 0.005694711580872536, + 0.0055221025831997395, + 0.005472924560308456, + 0.005480268504470587, + 0.007841723039746284, + 0.005462848115712404, + 0.004992645233869553, + 0.0037647150456905365, + 0.005294582340866327, + 0.006049822550266981, + 0.005391063168644905, + 0.023682817816734314, + 0.02079484611749649, + 0.009620999917387962, + 0.049545060843229294, + 0.01025603897869587, + 0.0387498177587986, + 0.012866200879216194, + 0.0172780342400074, + 0.014220855198800564, + 0.009815729223191738, + 0.006957918405532837, + 0.014611314982175827, + 0.01300093624740839, + 0.030695810914039612, + 0.006793313659727573, + 0.014213822782039642, + 0.014736607670783997, + 0.02217138558626175, + 0.007977690547704697, + 0.010783674195408821, + 0.004242727067321539, + 0.008284118957817554, + 0.006296979729086161, + 0.007570141926407814, + 0.0039208270609378815, + 0.00694958446547389, + 0.002708033425733447, + 0.006529540289193392, + 0.003075182903558016, + 0.0035887891426682472, + 0.0035701452288776636, + 0.0030934633687138557, + 0.0057622576132416725, + 0.002664920873939991, + 0.004311680793762207, + 0.0029980496037751436, + 0.002611269010230899, + 0.006611386314034462, + 0.0034550861455500126, + 0.004459095187485218, + 0.026870058849453926, + 0.02836531400680542, + 0.008196442387998104, + 0.014401351101696491, + 0.08259440213441849, + 0.09449601918458939, + 0.010247258469462395, + 0.016201646998524666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.0013387774815782905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12244957685470581, + 0.06361518800258636, + 0.0352456159889698, + 0.05672169849276543, + 0.21181854605674744, + 0.038262005895376205, + 0.05790798366069794, + 0.1434556096792221, + 0.06884166598320007, + 0.20168212056159973 + ], + "credit_map": [ + 0.018602319061756134, + 0.01275690644979477, + 0.011045384220778942, + 0.020776409655809402, + 0.014178413897752762, + 0.022844137623906136, + 0.012050171382725239, + 0.010195831768214703, + 0.0130534702911973, + 0.007058382499963045, + 0.004637222737073898, + 0.006765121594071388, + 0.006927695125341415, + 0.012228967621922493, + 0.006327552255243063, + 0.006616779137402773, + 0.0045942445285618305, + 0.0048433709889650345, + 0.002816596068441868, + 0.005239209625869989, + 0.0048590851947665215, + 0.004141315817832947, + 0.005996874067932367, + 0.0027549192309379578, + 0.008629160933196545, + 0.00876597035676241, + 0.008067202754318714, + 0.006444956641644239, + 0.0061934893019497395, + 0.006618059705942869, + 0.006490783765912056, + 0.005512075964361429, + 0.026582252234220505, + 0.02210451103746891, + 0.011369864456355572, + 0.05290259048342705, + 0.012911680154502392, + 0.050430625677108765, + 0.014276253990828991, + 0.021240759640932083, + 0.004798454232513905, + 0.004478348884731531, + 0.00359721458517015, + 0.004820162430405617, + 0.00548126082867384, + 0.007909780368208885, + 0.003402846632525325, + 0.003773939795792103, + 0.009016377851366997, + 0.011762836016714573, + 0.0038336501456797123, + 0.011145184747874737, + 0.003998335916548967, + 0.008024290204048157, + 0.005219362676143646, + 0.0049079409800469875, + 0.016065234318375587, + 0.0207788348197937, + 0.006425677333027124, + 0.04276150092482567, + 0.00943323690444231, + 0.008564659394323826, + 0.029275398701429367, + 0.010151064954698086, + 0.01051240973174572, + 0.0066291862167418, + 0.007734780665487051, + 0.011517267674207687, + 0.005654392298310995, + 0.009961284697055817, + 0.007257172837853432, + 0.009575171396136284, + 0.0167218204587698, + 0.021184157580137253, + 0.007552414201200008, + 0.007885945960879326, + 0.05532205477356911, + 0.06763053685426712, + 0.009728564880788326, + 0.015656623989343643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.004092651419341564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774088859558105 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774088859558105, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774088859558105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2783540189266205, + 0.051994211971759796, + 0.045659299939870834, + 0.055820342153310776, + 0.2302381545305252, + 0.11239563673734665, + 0.06360667198896408, + 0.06168993189930916, + 0.027759239077568054, + 0.07248250395059586 + ], + "credit_map": [ + 0.03544430434703827, + 0.02652800641953945, + 0.027584999799728394, + 0.02185131050646305, + 0.05923720449209213, + 0.03961386904120445, + 0.028791485354304314, + 0.03930283337831497, + 0.010395336896181107, + 0.008069686591625214, + 0.0032002869993448257, + 0.005324284080415964, + 0.007007501553744078, + 0.008847123011946678, + 0.004371141083538532, + 0.004778849892318249, + 0.006851889658719301, + 0.006634035147726536, + 0.0037829449865967035, + 0.005564108956605196, + 0.005763429682701826, + 0.00651182048022747, + 0.005317226983606815, + 0.005233848001807928, + 0.009407205507159233, + 0.007121071219444275, + 0.010185218416154385, + 0.006796775385737419, + 0.005489236209541559, + 0.005813083611428738, + 0.00630537373945117, + 0.0047023785300552845, + 0.030516650527715683, + 0.027130533009767532, + 0.012591666541993618, + 0.06468620151281357, + 0.012718193233013153, + 0.04468679800629616, + 0.017457418143749237, + 0.020450687035918236, + 0.010531697422266006, + 0.012202618643641472, + 0.006944234017282724, + 0.014409544877707958, + 0.018357181921601295, + 0.02851223573088646, + 0.009351921267807484, + 0.012086203321814537, + 0.00996224395930767, + 0.00858952198177576, + 0.006562466267496347, + 0.010471921414136887, + 0.005509600974619389, + 0.01089955773204565, + 0.004820982925593853, + 0.006790374405682087, + 0.00575088057667017, + 0.012995880097150803, + 0.005167894996702671, + 0.011309457942843437, + 0.005077009089291096, + 0.004786219913512468, + 0.01317534875124693, + 0.0034272437915205956, + 0.0038935034535825253, + 0.002705157268792391, + 0.0031066725496202707, + 0.0037282314151525497, + 0.0029748131055384874, + 0.003585920901969075, + 0.003994019236415625, + 0.003770921379327774, + 0.006692518014460802, + 0.007399481721222401, + 0.003072709310799837, + 0.004298001527786255, + 0.020352303981781006, + 0.020266763865947723, + 0.0037220409139990807, + 0.00667868135496974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.009774088859558105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a Ghibli-inspired art style, a high-speed slow-motion, low-angle, medium-length shot slowly zooms in on a majestic, ancient walnut tree, the centerpiece of a sun-dappled clearing. The tree's trunk is remarkably thick, richly painted with hand-painted textures, covered in deeply fissured gray-brown bark dotted with soft green moss. Its dense, gnarled branches spread outward like ancient arms, their vibrant green husks laden with clusters of walnuts. As the camera pans, the branches begin to vibrate gently and rhythmically, a slow yet powerful pulse derived not from the wind but from a magical, unseen energy. In exquisite slow motion, each lush leaf trembles and shimmers, capturing the warm, filtered sunlight, while the walnut itself begins to vibrate with a quiet intensity. A soft, inner golden light emanates from within, pulsating in sync with the tree's movement, revealing a quiet, powerful vitality within the scene and creating a serene, breathtakingly magical atmosphere.", + "seed": 6549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_13/index_1580_seed_6549_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014846390718594194, + "vlm_logprob": -0.014846390718594194, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.014846390718594194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044020723551511765 + ], + "opposite_logprob_tokens": [ + -3.294020652770996 + ], + "logprob_reward": -0.044020723551511765, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.044020723551511765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3098479211330414, + 0.12671612203121185, + 0.04839722439646721, + 0.06933772563934326, + 0.05019446834921837, + 0.1124914363026619, + 0.09398128092288971, + 0.06337146461009979, + 0.050395626574754715, + 0.07526674121618271 + ], + "credit_map": [ + 0.03195834904909134, + 0.037445131689310074, + 0.04143301397562027, + 0.039665382355451584, + 0.025998195633292198, + 0.03864898905158043, + 0.04653380811214447, + 0.04816504567861557, + 0.013445837423205376, + 0.017794949933886528, + 0.013282394967973232, + 0.020697668194770813, + 0.010566270910203457, + 0.01806267537176609, + 0.020835351198911667, + 0.012030965648591518, + 0.004833057522773743, + 0.005063394550234079, + 0.006757549941539764, + 0.005188832990825176, + 0.004690089728683233, + 0.007271172944456339, + 0.007465093396604061, + 0.007128030527383089, + 0.009990481659770012, + 0.006286729127168655, + 0.01044763345271349, + 0.007271235343068838, + 0.008859939873218536, + 0.009810586459934711, + 0.008651298470795155, + 0.008019820787012577, + 0.0036808375734835863, + 0.00598234822973609, + 0.0036860937252640724, + 0.007211494259536266, + 0.008068359456956387, + 0.007879848591983318, + 0.007667779456824064, + 0.006017710082232952, + 0.009907586500048637, + 0.014170453883707523, + 0.017371993511915207, + 0.014326532371342182, + 0.012429303489625454, + 0.019032087177038193, + 0.013904350809752941, + 0.01134912483394146, + 0.007429488934576511, + 0.009194024838507175, + 0.012329557910561562, + 0.011707816272974014, + 0.014953486621379852, + 0.013400532305240631, + 0.011807674542069435, + 0.013158705085515976, + 0.006913237273693085, + 0.006170301698148251, + 0.01242768857628107, + 0.00829765573143959, + 0.007520277053117752, + 0.004354836419224739, + 0.009725910611450672, + 0.007961556315422058, + 0.010109766386449337, + 0.004848933778703213, + 0.004788384307175875, + 0.00633902195841074, + 0.005305110942572355, + 0.0068037910386919975, + 0.004048634320497513, + 0.00815198291093111, + 0.005485036876052618, + 0.009431056678295135, + 0.005936251021921635, + 0.012859839014708996, + 0.021754104644060135, + 0.007914564572274685, + 0.005012464709579945, + 0.006873421370983124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.044020723551511765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052657718770205975 + ], + "opposite_logprob_tokens": [ + -5.255265712738037 + ], + "logprob_reward": -0.0052657718770205975, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052657718770205975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06947011500597, + 0.09078732877969742, + 0.050340138375759125, + 0.03948308527469635, + 0.04537038505077362, + 0.0780051127076149, + 0.13790298998355865, + 0.13360241055488586, + 0.18039070069789886, + 0.1746477484703064 + ], + "credit_map": [ + 0.007685973774641752, + 0.008199140429496765, + 0.007353249005973339, + 0.009341748431324959, + 0.009614917449653149, + 0.011050485074520111, + 0.007456416729837656, + 0.008768181316554546, + 0.007799197919666767, + 0.013356300070881844, + 0.010355189442634583, + 0.009866708889603615, + 0.010249681770801544, + 0.013716074638068676, + 0.013989130966365337, + 0.011455045081675053, + 0.0045369393192231655, + 0.004409042187035084, + 0.005380510818213224, + 0.00853973999619484, + 0.004322849679738283, + 0.007141902111470699, + 0.009158258326351643, + 0.006850897800177336, + 0.0050939577631652355, + 0.0035685475450009108, + 0.00638567516580224, + 0.0044780308380723, + 0.0055301194079220295, + 0.004372125957161188, + 0.0049499645829200745, + 0.005104664247483015, + 0.00445909146219492, + 0.006595877464860678, + 0.0044052256271243095, + 0.005970971658825874, + 0.004722461104393005, + 0.008610776625573635, + 0.005079785827547312, + 0.005526197142899036, + 0.008208448067307472, + 0.008663015440106392, + 0.0076752775348722935, + 0.007862010970711708, + 0.00874874833971262, + 0.008833692409098148, + 0.009743710979819298, + 0.018270207569003105, + 0.01449011079967022, + 0.017343226820230484, + 0.02478146366775036, + 0.016566326841711998, + 0.014935617335140705, + 0.02006402239203453, + 0.015064600855112076, + 0.014657631516456604, + 0.014444484375417233, + 0.022709771990776062, + 0.012485692277550697, + 0.017116349190473557, + 0.013210177421569824, + 0.020002009347081184, + 0.02089902199804783, + 0.012734909541904926, + 0.0268356055021286, + 0.015959840267896652, + 0.01927008107304573, + 0.02789425104856491, + 0.020529979839920998, + 0.028562502935528755, + 0.018835468217730522, + 0.022502977401018143, + 0.016632400453090668, + 0.013643047772347927, + 0.01063633244484663, + 0.05975189059972763, + 0.03984573483467102, + 0.01068340614438057, + 0.013554754666984081, + 0.009900173172354698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.0052657718770205975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013398488517850637 + ], + "opposite_logprob_tokens": [ + -6.626339912414551 + ], + "logprob_reward": -0.0013398488517850637, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013398488517850637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12897460162639618, + 0.0625605508685112, + 0.02789287455379963, + 0.08688393980264664, + 0.04311291128396988, + 0.32601651549339294, + 0.046976156532764435, + 0.09414015710353851, + 0.0689905658364296, + 0.11445168405771255 + ], + "credit_map": [ + 0.01355944387614727, + 0.018676534295082092, + 0.010231882333755493, + 0.018259119242429733, + 0.019902048632502556, + 0.0179134551435709, + 0.017119504511356354, + 0.013312620110809803, + 0.007279834244400263, + 0.00950031727552414, + 0.006693324074149132, + 0.0070014032535254955, + 0.006492397282272577, + 0.010260596871376038, + 0.008473671972751617, + 0.006859013810753822, + 0.003240433055907488, + 0.002905642380937934, + 0.003340867580845952, + 0.003207616973668337, + 0.0027915092650800943, + 0.0036748985294252634, + 0.004203462041914463, + 0.004528446588665247, + 0.011965176090598106, + 0.009961487725377083, + 0.010033906437456608, + 0.010720642283558846, + 0.010619888082146645, + 0.01335753221064806, + 0.011381222866475582, + 0.008844083175063133, + 0.0036768449936062098, + 0.005702648777514696, + 0.0038722683675587177, + 0.005801899824291468, + 0.005162883084267378, + 0.006603263318538666, + 0.005231714341789484, + 0.007061387877911329, + 0.02908940613269806, + 0.03835124149918556, + 0.04289385676383972, + 0.04137946665287018, + 0.036891814321279526, + 0.04546379670500755, + 0.034471407532691956, + 0.05747552588582039, + 0.004575065337121487, + 0.005880608223378658, + 0.006765828933566809, + 0.005662556737661362, + 0.005195087753236294, + 0.008453360758721828, + 0.004387953784316778, + 0.006055695004761219, + 0.008575888350605965, + 0.012461548671126366, + 0.01639966107904911, + 0.00974612683057785, + 0.011681966483592987, + 0.01114437822252512, + 0.014645322225987911, + 0.009485268034040928, + 0.013310675509274006, + 0.006269327364861965, + 0.005811017006635666, + 0.008680650033056736, + 0.006693619769066572, + 0.010650780983269215, + 0.004439252894371748, + 0.013135246932506561, + 0.011858036741614342, + 0.01303023099899292, + 0.009591984562575817, + 0.017416492104530334, + 0.026490839198231697, + 0.009161660447716713, + 0.010786274448037148, + 0.016116159036755562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0013398488517850637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538649141788483 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538649141788483, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538649141788483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1242767795920372, + 0.08106815069913864, + 0.028353415429592133, + 0.05204864591360092, + 0.0433524064719677, + 0.050566453486680984, + 0.07404318451881409, + 0.39474472403526306, + 0.05351557210087776, + 0.0980306938290596 + ], + "credit_map": [ + 0.013687453232705593, + 0.02275821566581726, + 0.010401279665529728, + 0.016826678067445755, + 0.018452879041433334, + 0.015958815813064575, + 0.01489270944148302, + 0.011298755183815956, + 0.0093478849157691, + 0.012781864032149315, + 0.010792738758027554, + 0.008079975843429565, + 0.007971063256263733, + 0.010711614042520523, + 0.010940748266875744, + 0.010442256927490234, + 0.0029220120050013065, + 0.0026341823395341635, + 0.0038645314052700996, + 0.0037579054478555918, + 0.0021575160790234804, + 0.0057731070555746555, + 0.0034860451705753803, + 0.0037581182550638914, + 0.006334097124636173, + 0.005772786680608988, + 0.010074376128613949, + 0.006660457234829664, + 0.006808620877563953, + 0.006037563551217318, + 0.005413245875388384, + 0.004947497509419918, + 0.0047125606797635555, + 0.005662683863192797, + 0.0042521292343735695, + 0.00507714319974184, + 0.006487224251031876, + 0.00557347945868969, + 0.006354245822876692, + 0.005232940427958965, + 0.004278241656720638, + 0.005535702686756849, + 0.006324784830212593, + 0.0048202937468886375, + 0.004975850693881512, + 0.00668419199064374, + 0.005537915974855423, + 0.012409470975399017, + 0.005010813940316439, + 0.00812672171741724, + 0.009725847281515598, + 0.008569592610001564, + 0.01309458538889885, + 0.010297744534909725, + 0.01065132487565279, + 0.008566553704440594, + 0.03210049495100975, + 0.03993622586131096, + 0.06964565813541412, + 0.0523836575448513, + 0.05000976845622063, + 0.03153545781970024, + 0.08637228608131409, + 0.03276117518544197, + 0.006698056124150753, + 0.006992876064032316, + 0.006420417223125696, + 0.007755659986287355, + 0.006396863609552383, + 0.006518970709294081, + 0.006524433381855488, + 0.006208288948982954, + 0.0082491934299469, + 0.009626484476029873, + 0.01174650527536869, + 0.018028125166893005, + 0.026278197765350342, + 0.009070527739822865, + 0.006040266249328852, + 0.00899139977991581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.012538649141788483 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a Ghibli-inspired art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient walnut tree with thick, gray-brown bark covered in green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the branches laden with clusters of walnuts and vibrant green husks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tree and its leaves vibrate gently with a soft inner golden light pulsating from within?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle medium-length shot that slowly zooms in on the tree?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066960170865059 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011066960170865059, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066960170865059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08094259351491928, + 0.05440417304635048, + 0.04088423773646355, + 0.06311403959989548, + 0.1353277713060379, + 0.18172863125801086, + 0.10813003778457642, + 0.10192662477493286, + 0.04819286987185478, + 0.1853490173816681 + ], + "credit_map": [ + 0.00833110511302948, + 0.007576535921543837, + 0.01259592454880476, + 0.008165073581039906, + 0.009752267971634865, + 0.011178090237081051, + 0.010743426159024239, + 0.012600166723132133, + 0.005863693542778492, + 0.005674085579812527, + 0.004309659358114004, + 0.005305891390889883, + 0.005890659987926483, + 0.012613460421562195, + 0.008993391878902912, + 0.005753330420702696, + 0.00426418986171484, + 0.0035666008479893208, + 0.007415715139359236, + 0.0038219550624489784, + 0.003973568323999643, + 0.0075463177636265755, + 0.004246456082910299, + 0.006049436051398516, + 0.008052004501223564, + 0.009100415743887424, + 0.009515331126749516, + 0.007004343904554844, + 0.006597066298127174, + 0.0106653468683362, + 0.00674492446705699, + 0.00543460575863719, + 0.012601095251739025, + 0.01476165372878313, + 0.021338315680623055, + 0.020708950236439705, + 0.011267749592661858, + 0.024439718574285507, + 0.015358904376626015, + 0.014851372689008713, + 0.015861790627241135, + 0.020520104095339775, + 0.02678024210035801, + 0.01917523331940174, + 0.023943839594721794, + 0.03268866613507271, + 0.02260368876159191, + 0.02015507221221924, + 0.014841177500784397, + 0.016802478581666946, + 0.014976262114942074, + 0.012320579960942268, + 0.010660231113433838, + 0.0113541129976511, + 0.015193307772278786, + 0.011981883086264133, + 0.010202616453170776, + 0.009805240668356419, + 0.024926692247390747, + 0.010760175995528698, + 0.007953896187245846, + 0.00868438370525837, + 0.016614260151982307, + 0.012979359365999699, + 0.007764978799968958, + 0.005887947045266628, + 0.005773322191089392, + 0.006653728894889355, + 0.006020981352776289, + 0.005538534373044968, + 0.0047345468774437904, + 0.005818831734359264, + 0.028444601222872734, + 0.019259601831436157, + 0.011750603094696999, + 0.046385787427425385, + 0.023509738966822624, + 0.015828169882297516, + 0.013967271894216537, + 0.02620323933660984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.011066960170865059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_14/index_tensor([1474, 1474])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_14/index_tensor([1474, 1474])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a31ca755ef5bbd812ddceca3e9d09a78041693b9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_14/index_tensor([1474, 1474])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with an extreme close-up shot from ground level, revealing an ultra-high-definition, photorealistic CGI image focused on a patch of moist, dark brown soil rich in organic matter. Soft morning light slants down, making tiny water droplets cling to the variously sized soil particles shimmer with a shimmering brilliance. Suddenly, the ground arches slightly, a tiny crack appears. A fresh, emerald green sprout, its curved stem thrusting upward with unstoppable vitality, pushes away the soil particles above it, sending clumps of moist earth tumbling to the side. As the translucent sprout fully emerges from the earth, its stem gradually straightens, and the two cotyledons curled at the top begin to carefully and nearly symmetrically unfold to the sides, revealing the fresher green and distinct veins within. Throughout this process, the camera rises in a smooth and steady rhythm from a close-up perspective, maintaining focus on the growing sprout, ultimately settling at a perspective parallel to the unfolding leaves. The whole process appears sacred and poetic under the warm light, creating a tranquil atmosphere of rebirth, resilience and hope.", + "seed": 2400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_14/index_1474_seed_2400_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040504826698452234, + "vlm_logprob": -0.040504826698452234, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.040504826698452234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005144928116351366 + ], + "opposite_logprob_tokens": [ + -6.255145072937012 + ], + "logprob_reward": -0.005144928116351366, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.005144928116351366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.167318657040596, + 0.202925443649292, + 0.1708543598651886, + 0.16046485304832458, + 0.03979995474219322, + 0.04899062588810921, + 0.043934471905231476, + 0.04652877897024155, + 0.0759301409125328, + 0.04325271397829056 + ], + "credit_map": [ + 0.01641022600233555, + 0.01834653690457344, + 0.023243913426995277, + 0.012785862199962139, + 0.041910093277692795, + 0.026037879288196564, + 0.017881236970424652, + 0.010702905245125294, + 0.06676647812128067, + 0.011385245248675346, + 0.01131031010299921, + 0.013070614077150822, + 0.008114065043628216, + 0.06862758845090866, + 0.00746572483330965, + 0.016185399144887924, + 0.061500802636146545, + 0.010934672318398952, + 0.009526167064905167, + 0.007137462496757507, + 0.005644374992698431, + 0.06454585492610931, + 0.005391254089772701, + 0.006173768080770969, + 0.02496723271906376, + 0.014139511622488499, + 0.01730441488325596, + 0.021640297025442123, + 0.010823365300893784, + 0.027679720893502235, + 0.011887647211551666, + 0.032022666186094284, + 0.005148809868842363, + 0.002696165582165122, + 0.004443677142262459, + 0.009534374810755253, + 0.0033776178024709225, + 0.004155564121901989, + 0.0036628139205276966, + 0.00678093358874321, + 0.00854541640728712, + 0.005469503812491894, + 0.005622171331197023, + 0.006731326691806316, + 0.006902152672410011, + 0.005458357743918896, + 0.004531762562692165, + 0.005729933734983206, + 0.0048880865797400475, + 0.0033967571798712015, + 0.007429242134094238, + 0.007988440804183483, + 0.00354683306068182, + 0.004693185444921255, + 0.004637337755411863, + 0.007354590576142073, + 0.006093052215874195, + 0.005219870712608099, + 0.006445724982768297, + 0.007316348608583212, + 0.004531560465693474, + 0.0062970384024083614, + 0.005042314063757658, + 0.005582870915532112, + 0.01065760012716055, + 0.0058647943660616875, + 0.006250183098018169, + 0.016112195327878, + 0.007866042666137218, + 0.010415096767246723, + 0.009512627497315407, + 0.009251597337424755, + 0.003920573275536299, + 0.004404118284583092, + 0.008260222151875496, + 0.0110380994156003, + 0.0033592921681702137, + 0.004978201352059841, + 0.003938468173146248, + 0.003353739855811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.005144928116351366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125654898583889 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.0125654898583889, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125654898583889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1338045299053192, + 0.33727243542671204, + 0.21826553344726562, + 0.0835263803601265, + 0.03354671597480774, + 0.05169624462723732, + 0.03293213993310928, + 0.03693441301584244, + 0.04849989712238312, + 0.023521607741713524 + ], + "credit_map": [ + 0.012161157093942165, + 0.012756435200572014, + 0.03167708218097687, + 0.010638224892318249, + 0.017857495695352554, + 0.026847105473279953, + 0.014273482374846935, + 0.0075935483910143375, + 0.11884549260139465, + 0.017485303804278374, + 0.020525244995951653, + 0.01784423552453518, + 0.016656357795000076, + 0.1127905324101448, + 0.015486365184187889, + 0.017638955265283585, + 0.0778108760714531, + 0.015110976994037628, + 0.01068695355206728, + 0.011330693028867245, + 0.00832012202590704, + 0.08041782677173615, + 0.007101239636540413, + 0.007486867718398571, + 0.008649931289255619, + 0.006191350985318422, + 0.008968048729002476, + 0.017664548009634018, + 0.004643389489501715, + 0.011042248457670212, + 0.0071894265711307526, + 0.019177444279193878, + 0.005684738513082266, + 0.003827808191999793, + 0.004325707908719778, + 0.0036269803531467915, + 0.005745804402977228, + 0.004221117589622736, + 0.002710578963160515, + 0.003403984708711505, + 0.005856691859662533, + 0.005312902852892876, + 0.005440024193376303, + 0.007054016925394535, + 0.008176008239388466, + 0.005292167421430349, + 0.004255251027643681, + 0.010309185832738876, + 0.0046245199628174305, + 0.003959707450121641, + 0.00331331929191947, + 0.0054896739311516285, + 0.0044115521013736725, + 0.004087458830326796, + 0.0024607304949313402, + 0.004585180897265673, + 0.0038643358275294304, + 0.0049708811566233635, + 0.0033367611467838287, + 0.004302996210753918, + 0.007421452086418867, + 0.005108666606247425, + 0.004594721831381321, + 0.003334601176902652, + 0.0056054433807730675, + 0.005941939540207386, + 0.003910253290086985, + 0.008649923838675022, + 0.005644389893859625, + 0.006339894607663155, + 0.005961692426353693, + 0.006446364335715771, + 0.0020987526513636112, + 0.0019872579723596573, + 0.004560503177344799, + 0.006546887569129467, + 0.0012962925247848034, + 0.002738314913585782, + 0.0023287045769393444, + 0.0019648976158350706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0125654898583889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371515870094299 + ], + "opposite_logprob_tokens": [ + -3.40871524810791 + ], + "logprob_reward": -0.03371515870094299, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371515870094299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06574742496013641, + 0.560893714427948, + 0.04537993296980858, + 0.0801418274641037, + 0.04968002811074257, + 0.04334856569766998, + 0.06380439549684525, + 0.03259690850973129, + 0.03632340580224991, + 0.022083798423409462 + ], + "credit_map": [ + 0.010228438302874565, + 0.007652332540601492, + 0.008063916116952896, + 0.0056936414912343025, + 0.00950124952942133, + 0.011986073106527328, + 0.0086366580799222, + 0.003985117189586163, + 0.2080257385969162, + 0.03007475845515728, + 0.03256061300635338, + 0.029097499325871468, + 0.01997358351945877, + 0.19736076891422272, + 0.019670486450195312, + 0.02413025125861168, + 0.008763900026679039, + 0.004738334566354752, + 0.004576210398226976, + 0.00466250441968441, + 0.0066677178256213665, + 0.007821392267942429, + 0.003509542439132929, + 0.004640331957489252, + 0.0091992923989892, + 0.005932961590588093, + 0.009515998885035515, + 0.017102761194109917, + 0.0038070541340857744, + 0.010708419606089592, + 0.006309568416327238, + 0.017565766349434853, + 0.006703478284180164, + 0.004796797409653664, + 0.007737386506050825, + 0.006676629185676575, + 0.007020771037787199, + 0.005424948874861002, + 0.004652116447687149, + 0.006667896639555693, + 0.0069771381095051765, + 0.004224421922117472, + 0.004229547921568155, + 0.004200815223157406, + 0.005851793568581343, + 0.004952059127390385, + 0.004016999155282974, + 0.008895792067050934, + 0.00808987021446228, + 0.005195143166929483, + 0.011786176823079586, + 0.00786780845373869, + 0.008845220319926739, + 0.007560746744275093, + 0.006443639285862446, + 0.008015794679522514, + 0.003891782835125923, + 0.0032814715523272753, + 0.004632903728634119, + 0.0037836867850273848, + 0.0043535372242331505, + 0.0045280153863132, + 0.003558333031833172, + 0.0045671784318983555, + 0.004487716127187014, + 0.0036283847875893116, + 0.003723232541233301, + 0.006256881635636091, + 0.0034299201797693968, + 0.004757470451295376, + 0.004949893802404404, + 0.005089905578643084, + 0.0020925302524119616, + 0.0019527587573975325, + 0.004087360110133886, + 0.00608476810157299, + 0.001283377641811967, + 0.002628812799230218, + 0.002145450795069337, + 0.0018087407806888223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.03371515870094299 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204873323440552 + ], + "opposite_logprob_tokens": [ + -2.8120486736297607 + ], + "logprob_reward": -0.06204873323440552, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204873323440552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12759698927402496, + 0.4397110641002655, + 0.1803997904062271, + 0.04650946334004402, + 0.023808326572179794, + 0.03001507930457592, + 0.03625423461198807, + 0.029114529490470886, + 0.0700257420539856, + 0.016564803197979927 + ], + "credit_map": [ + 0.016413813456892967, + 0.016250884160399437, + 0.015543859452009201, + 0.012058519758284092, + 0.025541173294186592, + 0.018688851967453957, + 0.01617804914712906, + 0.006921826861798763, + 0.16429957747459412, + 0.02263263426721096, + 0.027965346351265907, + 0.026238176971673965, + 0.017043711617588997, + 0.14856243133544922, + 0.016568731516599655, + 0.01640043780207634, + 0.06084650754928589, + 0.011390534229576588, + 0.010076001286506653, + 0.012079799547791481, + 0.007877158932387829, + 0.06313476711511612, + 0.006722757592797279, + 0.008272267878055573, + 0.009409645572304726, + 0.0050176046788692474, + 0.006698864512145519, + 0.004290487617254257, + 0.005485282279551029, + 0.00492840725928545, + 0.004423660226166248, + 0.006255511660128832, + 0.0038465713150799274, + 0.002936025382950902, + 0.0019204188138246536, + 0.0024022895377129316, + 0.0032805828377604485, + 0.00322892889380455, + 0.0030042200814932585, + 0.003189290640875697, + 0.004841839894652367, + 0.004436210263520479, + 0.0033372214529663324, + 0.00516731571406126, + 0.0034964345395565033, + 0.002854169812053442, + 0.002552065532654524, + 0.0033298202324658632, + 0.00439612427726388, + 0.0024992753751575947, + 0.0051262578926980495, + 0.005763071123510599, + 0.0044409227557480335, + 0.004397720564156771, + 0.00432571629062295, + 0.005305150058120489, + 0.003805364016443491, + 0.0033082771115005016, + 0.0034898193553090096, + 0.0027666334062814713, + 0.004164234735071659, + 0.003887672908604145, + 0.0030388704035431147, + 0.004653655458241701, + 0.009563552215695381, + 0.006922322791069746, + 0.005269235465675592, + 0.011593708768486977, + 0.0077147711999714375, + 0.009429429657757282, + 0.009067660197615623, + 0.01046505942940712, + 0.001584929646924138, + 0.0015904722968116403, + 0.002889864146709442, + 0.004007669631391764, + 0.0013920033816248178, + 0.0017983440775424242, + 0.001705625792965293, + 0.001595894107595086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.06204873323440552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890498235821724 + ], + "opposite_logprob_tokens": [ + -2.464049816131592 + ], + "logprob_reward": -0.0890498235821724, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890498235821724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11145425587892532, + 0.11087772995233536, + 0.37882670760154724, + 0.12663012742996216, + 0.05508193373680115, + 0.07171984016895294, + 0.048769302666187286, + 0.019641200080513954, + 0.03484645113348961, + 0.04215243458747864 + ], + "credit_map": [ + 0.007899243384599686, + 0.013403357937932014, + 0.020352888852357864, + 0.01386564876884222, + 0.02064441330730915, + 0.020902667194604874, + 0.009301047772169113, + 0.005084990989416838, + 0.023741211742162704, + 0.009113162755966187, + 0.009757086634635925, + 0.019258150830864906, + 0.00958581268787384, + 0.01679428108036518, + 0.008828883990645409, + 0.013799148611724377, + 0.14106719195842743, + 0.024478167295455933, + 0.019227176904678345, + 0.015066089108586311, + 0.011130750179290771, + 0.14523039758205414, + 0.01044787373393774, + 0.012179060839116573, + 0.02207004837691784, + 0.015109633095562458, + 0.020508665591478348, + 0.014063716866075993, + 0.01067180186510086, + 0.016874639317393303, + 0.011803497560322285, + 0.015528126619756222, + 0.0074617862701416016, + 0.005231753457337618, + 0.009912696667015553, + 0.007659869268536568, + 0.005304847843945026, + 0.0062933508306741714, + 0.004929360933601856, + 0.008288275450468063, + 0.012950819917023182, + 0.008547857403755188, + 0.009427591226994991, + 0.010197043418884277, + 0.00984996184706688, + 0.008019537664949894, + 0.005825771484524012, + 0.0069012632593512535, + 0.006564504001289606, + 0.004735462833195925, + 0.007900788448750973, + 0.009468081407248974, + 0.005005905404686928, + 0.005292524583637714, + 0.004746634513139725, + 0.0050554038025438786, + 0.002178708091378212, + 0.0019742019940167665, + 0.0037124401424080133, + 0.0027740816585719585, + 0.002099681878462434, + 0.002249521901831031, + 0.0023563571739941835, + 0.0022962104994803667, + 0.0044594379141926765, + 0.0027511923108249903, + 0.004342901054769754, + 0.005137420259416103, + 0.004540898837149143, + 0.004464726895093918, + 0.0041095903143286705, + 0.005040287505835295, + 0.004295671824365854, + 0.0037521629128605127, + 0.009012042544782162, + 0.010784082114696503, + 0.0026402294170111418, + 0.004770523402839899, + 0.003646112745627761, + 0.003251611953601241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0890498235821724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an extreme close-up shot from ground level, revealing an ultra-high-definition, photorealistic CGI image focused on a patch of moist, dark brown soil rich in organic matter. Soft morning light slants down, making tiny water droplets cling to the variously sized soil particles shimmer with a shimmering brilliance. Suddenly, the ground arches slightly, a tiny crack appears. A fresh, emerald green sprout, its curved stem thrusting upward with unstoppable vitality, pushes away the soil particles above it, sending clumps of moist earth tumbling to the side. As the translucent sprout fully emerges from the earth, its stem gradually straightens, and the two cotyledons curled at the top begin to carefully and nearly symmetrically unfold to the sides, revealing the fresher green and distinct veins within. Throughout this process, the camera rises in a smooth and steady rhythm from a close-up perspective, maintaining focus on the growing sprout, ultimately settling at a perspective parallel to the unfolding leaves. The whole process appears sacred and poetic under the warm light, creating a tranquil atmosphere of rebirth, resilience and hope.", + "seed": 2401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_14/index_1474_seed_2401_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01837722985073924, + "vlm_logprob": -0.01837722985073924, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.01837722985073924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003757323371246457 + ], + "opposite_logprob_tokens": [ + -7.503757476806641 + ], + "logprob_reward": -0.003757323371246457, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.003757323371246457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1764741986989975, + 0.060194097459316254, + 0.1192161813378334, + 0.048612065613269806, + 0.07496564835309982, + 0.042664334177970886, + 0.05414591729640961, + 0.13374345004558563, + 0.20546744763851166, + 0.08451663702726364 + ], + "credit_map": [ + 0.014766905456781387, + 0.030179547145962715, + 0.018834179267287254, + 0.021573949605226517, + 0.02133583277463913, + 0.020825954154133797, + 0.035735152661800385, + 0.013222686015069485, + 0.007702138274908066, + 0.008463060483336449, + 0.008811704814434052, + 0.008613442070782185, + 0.005889973137527704, + 0.007086505182087421, + 0.0083137396723032, + 0.005313536152243614, + 0.015889909118413925, + 0.014234068803489208, + 0.02155473455786705, + 0.022685149684548378, + 0.011891569942235947, + 0.008761596865952015, + 0.013008538633584976, + 0.011190615594387054, + 0.004815019201487303, + 0.005949944723397493, + 0.00621040677651763, + 0.005627645645290613, + 0.007366166450083256, + 0.005231491755694151, + 0.008844444528222084, + 0.004566947463899851, + 0.009417515248060226, + 0.011969359591603279, + 0.006894924212247133, + 0.006944372318685055, + 0.012146271765232086, + 0.01103716716170311, + 0.008975504897534847, + 0.007580531295388937, + 0.005252297036349773, + 0.005732919555157423, + 0.0037864367477595806, + 0.003665742464363575, + 0.004653571173548698, + 0.008546379394829273, + 0.006573152262717485, + 0.004453835543245077, + 0.006464747246354818, + 0.008143246173858643, + 0.005466122645884752, + 0.004345099441707134, + 0.010057938285171986, + 0.007266511209309101, + 0.004983703605830669, + 0.007418548222631216, + 0.019942469894886017, + 0.017873490229249, + 0.012879702262580395, + 0.015637310221791267, + 0.01580451987683773, + 0.020108025521039963, + 0.01977299526333809, + 0.011724925599992275, + 0.023431753739714622, + 0.023458706215023994, + 0.022127211093902588, + 0.02465926483273506, + 0.03149162605404854, + 0.024542180821299553, + 0.03263779729604721, + 0.023118911311030388, + 0.010104570537805557, + 0.016273081302642822, + 0.007795438170433044, + 0.00790622178465128, + 0.014567429199814796, + 0.011864488944411278, + 0.007030262611806393, + 0.008975142613053322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.003757323371246457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036286956164985895 + ], + "opposite_logprob_tokens": [ + -5.628628730773926 + ], + "logprob_reward": -0.0036286956164985895, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036286956164985895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08174291253089905, + 0.05223561450839043, + 0.031688012182712555, + 0.0362701341509819, + 0.040030475705862045, + 0.12474168092012405, + 0.15655654668807983, + 0.265707790851593, + 0.14424052834510803, + 0.06678630411624908 + ], + "credit_map": [ + 0.0053998371586203575, + 0.0065426514483988285, + 0.00872706901282072, + 0.010560673661530018, + 0.013838576152920723, + 0.011421499773859978, + 0.01734953559935093, + 0.007903067395091057, + 0.004712574649602175, + 0.010478692129254341, + 0.0066231489181518555, + 0.006824304815381765, + 0.009072909131646156, + 0.006190943531692028, + 0.004789779428392649, + 0.003543263068422675, + 0.0036777867935597897, + 0.005000374745577574, + 0.0038153789937496185, + 0.0036541747394949198, + 0.003941141068935394, + 0.004031812772154808, + 0.005075600929558277, + 0.0024917416740208864, + 0.003337766043841839, + 0.00418599508702755, + 0.006074988748878241, + 0.005505888257175684, + 0.004989318549633026, + 0.0034535329323261976, + 0.005394148174673319, + 0.0033284968230873346, + 0.004385540261864662, + 0.008195274509489536, + 0.0034115868620574474, + 0.0056261769495904446, + 0.004367683082818985, + 0.004660268779844046, + 0.005055277142673731, + 0.004328672308474779, + 0.009881372563540936, + 0.014271094463765621, + 0.009819844737648964, + 0.023171747103333473, + 0.016339758411049843, + 0.01531174685806036, + 0.026271510869264603, + 0.009674611501395702, + 0.02181071788072586, + 0.01865522563457489, + 0.012705828063189983, + 0.0200370941311121, + 0.017912400886416435, + 0.01777641661465168, + 0.03075801394879818, + 0.01690085045993328, + 0.030257150530815125, + 0.0359172523021698, + 0.0350877083837986, + 0.024299977347254753, + 0.039929021149873734, + 0.032977212220430374, + 0.043702591210603714, + 0.023536894470453262, + 0.016937974840402603, + 0.02126452326774597, + 0.017224431037902832, + 0.014944205060601234, + 0.02268839068710804, + 0.01695854403078556, + 0.01636451855301857, + 0.017857933416962624, + 0.007693117950111628, + 0.011109917424619198, + 0.00732787512242794, + 0.006421549711376429, + 0.009933801367878914, + 0.008277238346636295, + 0.009007759392261505, + 0.007015041541308165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.0036286956164985895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545594945549965 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545594945549965, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545594945549965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055657241493463516, + 0.06394185870885849, + 0.035012416541576385, + 0.031102504581212997, + 0.0489310584962368, + 0.047531262040138245, + 0.1121496632695198, + 0.22104346752166748, + 0.2227732092142105, + 0.16185727715492249 + ], + "credit_map": [ + 0.005485488101840019, + 0.0055427574552595615, + 0.007628460414707661, + 0.007653915323317051, + 0.005430338438600302, + 0.00469724228605628, + 0.012507072649896145, + 0.006711968686431646, + 0.005744848400354385, + 0.013361972756683826, + 0.00709948968142271, + 0.010607698932290077, + 0.009786088019609451, + 0.007583692669868469, + 0.005834297277033329, + 0.003923766314983368, + 0.0042409878224134445, + 0.006055144593119621, + 0.004063451196998358, + 0.00600670138373971, + 0.003371176542714238, + 0.003926327917724848, + 0.004066745284944773, + 0.0032818836625665426, + 0.002617674181237817, + 0.003436962841078639, + 0.0027182288467884064, + 0.0056198774836957455, + 0.004947693552821875, + 0.0035179767291992903, + 0.005431399215012789, + 0.0028126914985477924, + 0.003609789302572608, + 0.005376132670789957, + 0.005637832451611757, + 0.009624121710658073, + 0.007101999595761299, + 0.005689550656825304, + 0.00665245158597827, + 0.0052391826175153255, + 0.004137708805501461, + 0.0057966927997767925, + 0.0033267312683165073, + 0.0074039651080966, + 0.007951440289616585, + 0.006196469999849796, + 0.008219940587878227, + 0.004498317372053862, + 0.015665873885154724, + 0.014796570874750614, + 0.008650480769574642, + 0.014582165516912937, + 0.012524235062301159, + 0.013157567009329796, + 0.02212217077612877, + 0.01065059844404459, + 0.02427978627383709, + 0.030527066439390182, + 0.027512460947036743, + 0.022945666685700417, + 0.03609802946448326, + 0.02758646197617054, + 0.035154927521944046, + 0.01693907007575035, + 0.022874146699905396, + 0.03221551701426506, + 0.02419288456439972, + 0.021677536889910698, + 0.03324456140398979, + 0.025931062176823616, + 0.03281944617629051, + 0.029818058013916016, + 0.017232347279787064, + 0.028842145577073097, + 0.013790974393486977, + 0.015431652776896954, + 0.02606087177991867, + 0.021032067015767097, + 0.025234293192625046, + 0.014232936315238476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.012545594945549965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287491872906685 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287491872906685, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287491872906685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06910950690507889, + 0.18295705318450928, + 0.03024897910654545, + 0.04610472172498703, + 0.04027543589472771, + 0.06267859041690826, + 0.10332564264535904, + 0.26448363065719604, + 0.11285201460123062, + 0.08796446770429611 + ], + "credit_map": [ + 0.007306282874196768, + 0.006582684349268675, + 0.009089348837733269, + 0.00899410154670477, + 0.008172722533345222, + 0.006756361108273268, + 0.01425221748650074, + 0.007955791428685188, + 0.017658546566963196, + 0.04017181694507599, + 0.018077347427606583, + 0.02377587929368019, + 0.025279194116592407, + 0.01961914263665676, + 0.023764440789818764, + 0.014610687270760536, + 0.0034721444826573133, + 0.005168537609279156, + 0.003500457853078842, + 0.0035937773063778877, + 0.003266958286985755, + 0.00370011362247169, + 0.0032642518635839224, + 0.004282739479094744, + 0.005808532238006592, + 0.008118837140500546, + 0.006083889398723841, + 0.0048499335534870625, + 0.00664337957277894, + 0.005284793209284544, + 0.004492948763072491, + 0.004822409246116877, + 0.005110923200845718, + 0.0036305170506238937, + 0.004875264596194029, + 0.008347731083631516, + 0.005112986546009779, + 0.0039009740576148033, + 0.0064015379175543785, + 0.0028955028392374516, + 0.008891735225915909, + 0.009384415112435818, + 0.0059318807907402515, + 0.006308089476078749, + 0.007439216133207083, + 0.014410511590540409, + 0.005811614450067282, + 0.004501130897551775, + 0.012177536264061928, + 0.009788302704691887, + 0.011181463487446308, + 0.013944054953753948, + 0.010805655270814896, + 0.016401497647166252, + 0.016772309318184853, + 0.012254815548658371, + 0.04132630676031113, + 0.03019545041024685, + 0.033596597611904144, + 0.02850172109901905, + 0.025808293372392654, + 0.03148415684700012, + 0.0432579331099987, + 0.03031316213309765, + 0.011059125885367393, + 0.014062166213989258, + 0.011618449352681637, + 0.010151088237762451, + 0.014310532249510288, + 0.0132681243121624, + 0.01787654310464859, + 0.020505990833044052, + 0.010634265840053558, + 0.013396833091974258, + 0.009172983467578888, + 0.008521495386958122, + 0.013030056841671467, + 0.010618545114994049, + 0.009794761426746845, + 0.012795528396964073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.023287491872906685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866704344749451 + ], + "opposite_logprob_tokens": [ + -3.0486669540405273 + ], + "logprob_reward": -0.04866704344749451, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866704344749451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19604726135730743, + 0.05618071183562279, + 0.09587787091732025, + 0.05511808022856712, + 0.07998451590538025, + 0.09526071697473526, + 0.10661696642637253, + 0.10334131866693497, + 0.10757003724575043, + 0.10400260239839554 + ], + "credit_map": [ + 0.01734030805528164, + 0.023364009335637093, + 0.03077945113182068, + 0.029619473963975906, + 0.0182392206043005, + 0.01886732317507267, + 0.04267129302024841, + 0.015166164375841618, + 0.006060663145035505, + 0.00946277566254139, + 0.00840851292014122, + 0.004895634017884731, + 0.006859102286398411, + 0.007365379482507706, + 0.006116489879786968, + 0.007012147456407547, + 0.013784532435238361, + 0.02083643339574337, + 0.013474762439727783, + 0.007469957694411278, + 0.011623768135905266, + 0.01333052758127451, + 0.007191782351583242, + 0.008166099898517132, + 0.007536240387707949, + 0.00806973222643137, + 0.007683217525482178, + 0.007338786963373423, + 0.00679395766928792, + 0.005682528950273991, + 0.005940708797425032, + 0.006072901654988527, + 0.00913665909320116, + 0.01580647937953472, + 0.012496416456997395, + 0.006762356497347355, + 0.009102392010390759, + 0.010661394335329533, + 0.008821383118629456, + 0.007197422441095114, + 0.011443727649748325, + 0.012462026439607143, + 0.009233280085027218, + 0.015428585931658745, + 0.011136761866509914, + 0.015621146187186241, + 0.009240671060979366, + 0.010694506578147411, + 0.011032144539058208, + 0.01744445040822029, + 0.008100874722003937, + 0.014147618785500526, + 0.014984073117375374, + 0.014577187597751617, + 0.019866956397891045, + 0.00646365899592638, + 0.01719409041106701, + 0.012372155673801899, + 0.011499526910483837, + 0.010117358528077602, + 0.011225538328289986, + 0.011290105991065502, + 0.012738313525915146, + 0.01690422184765339, + 0.010256561450660229, + 0.015300322324037552, + 0.010013350285589695, + 0.013249831274151802, + 0.01395348273217678, + 0.012409653514623642, + 0.02050165645778179, + 0.011885164305567741, + 0.012507240287959576, + 0.017139790579676628, + 0.010172870010137558, + 0.008447990752756596, + 0.019397834315896034, + 0.013187360018491745, + 0.011195960454642773, + 0.011953546665608883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.04866704344749451 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an extreme close-up shot from ground level, revealing an ultra-high-definition, photorealistic CGI image focused on a patch of moist, dark brown soil rich in organic matter. Soft morning light slants down, making tiny water droplets cling to the variously sized soil particles shimmer with a shimmering brilliance. Suddenly, the ground arches slightly, a tiny crack appears. A fresh, emerald green sprout, its curved stem thrusting upward with unstoppable vitality, pushes away the soil particles above it, sending clumps of moist earth tumbling to the side. As the translucent sprout fully emerges from the earth, its stem gradually straightens, and the two cotyledons curled at the top begin to carefully and nearly symmetrically unfold to the sides, revealing the fresher green and distinct veins within. Throughout this process, the camera rises in a smooth and steady rhythm from a close-up perspective, maintaining focus on the growing sprout, ultimately settling at a perspective parallel to the unfolding leaves. The whole process appears sacred and poetic under the warm light, creating a tranquil atmosphere of rebirth, resilience and hope.", + "seed": 2402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_14/index_1474_seed_2402_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01586434510536492, + "vlm_logprob": -0.01586434510536492, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.01586434510536492, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003054716857150197 + ], + "opposite_logprob_tokens": [ + -6.378054618835449 + ], + "logprob_reward": -0.003054716857150197, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.003054716857150197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1476922333240509, + 0.06743692606687546, + 0.04199402779340744, + 0.037838492542505264, + 0.04381156712770462, + 0.027605971321463585, + 0.041089944541454315, + 0.0528852716088295, + 0.04128020629286766, + 0.4983654022216797 + ], + "credit_map": [ + 0.014855676330626011, + 0.022103557363152504, + 0.017963333055377007, + 0.014684541150927544, + 0.0311935693025589, + 0.01346027385443449, + 0.016413621604442596, + 0.017017647624015808, + 0.010599135421216488, + 0.006614247336983681, + 0.008163996040821075, + 0.00727473571896553, + 0.008908838033676147, + 0.01065848208963871, + 0.007088446523994207, + 0.008129036985337734, + 0.007438975851982832, + 0.004889151081442833, + 0.008631276898086071, + 0.005069998558610678, + 0.003888641716912389, + 0.006049952935427427, + 0.003098811022937298, + 0.0029272176325321198, + 0.007601786870509386, + 0.00821248721331358, + 0.0037735921796411276, + 0.003952347207814455, + 0.0037841058801859617, + 0.005045141093432903, + 0.002796206157654524, + 0.0026728215161710978, + 0.004687124863266945, + 0.00621134415268898, + 0.009040629491209984, + 0.007199419196695089, + 0.0038970476016402245, + 0.004455218091607094, + 0.00578454090282321, + 0.002536238869652152, + 0.003309028223156929, + 0.005530042573809624, + 0.002150831511244178, + 0.0031269886530935764, + 0.004278373438864946, + 0.0028177625499665737, + 0.0040492108091712, + 0.002343730302527547, + 0.010846599005162716, + 0.0027716713957488537, + 0.0035991864278912544, + 0.003930503968149424, + 0.002793462248519063, + 0.01174196694046259, + 0.002406548010185361, + 0.003000006778165698, + 0.005785062443464994, + 0.005922983400523663, + 0.00732076819986105, + 0.0071878149174153805, + 0.004975257907062769, + 0.007074464578181505, + 0.009352861903607845, + 0.005266052670776844, + 0.005515621043741703, + 0.006253029685467482, + 0.004515477456152439, + 0.00446253502741456, + 0.005148722790181637, + 0.00456346245482564, + 0.003136302111670375, + 0.00768505223095417, + 0.05949300155043602, + 0.06614596396684647, + 0.0987263172864914, + 0.06537119299173355, + 0.03667357191443443, + 0.06268752366304398, + 0.05543670803308487, + 0.05383108928799629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.003054716857150197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003627745434641838 + ], + "opposite_logprob_tokens": [ + -5.628627777099609 + ], + "logprob_reward": -0.003627745434641838, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003627745434641838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08383456617593765, + 0.13722452521324158, + 0.07253365963697433, + 0.06675399094820023, + 0.06777942180633545, + 0.020051753148436546, + 0.05305582284927368, + 0.06731017678976059, + 0.3239034116268158, + 0.1075526624917984 + ], + "credit_map": [ + 0.01360607985407114, + 0.010828578844666481, + 0.010308917611837387, + 0.006133571267127991, + 0.013168740086257458, + 0.01133685652166605, + 0.007308408617973328, + 0.011143417097628117, + 0.01315807644277811, + 0.014928311109542847, + 0.023501433432102203, + 0.02502671629190445, + 0.022334860637784004, + 0.010961022228002548, + 0.014160706661641598, + 0.01315339282155037, + 0.011208959855139256, + 0.009588749147951603, + 0.008877751417458057, + 0.012829015962779522, + 0.005249790847301483, + 0.010008684359490871, + 0.009247508831322193, + 0.005523198284208775, + 0.015304901637136936, + 0.009678887203335762, + 0.00508059561252594, + 0.006017711944878101, + 0.005543392617255449, + 0.013852689415216446, + 0.0066734193824231625, + 0.004602395463734865, + 0.007781114429235458, + 0.010640512220561504, + 0.011343983933329582, + 0.011911895126104355, + 0.00589751685038209, + 0.00733558414503932, + 0.008366212248802185, + 0.004502602852880955, + 0.0017617191188037395, + 0.0025838110595941544, + 0.002025206107646227, + 0.0040160538628697395, + 0.0020627614576369524, + 0.002462188946083188, + 0.0032172882929444313, + 0.0019227248849347234, + 0.007994682528078556, + 0.007648901082575321, + 0.004885987378656864, + 0.009760379791259766, + 0.004541388247162104, + 0.008764677681028843, + 0.006930036004632711, + 0.0025297715328633785, + 0.008371956646442413, + 0.005496560130268335, + 0.008818307891488075, + 0.017289290204644203, + 0.004873724188655615, + 0.007870224304497242, + 0.00737740145996213, + 0.007212715223431587, + 0.03652225434780121, + 0.02801896631717682, + 0.08046872168779373, + 0.05390016362071037, + 0.021465204656124115, + 0.04021307826042175, + 0.02645972929894924, + 0.03685528784990311, + 0.013927238062024117, + 0.015310842543840408, + 0.02203158475458622, + 0.01570822112262249, + 0.009898288175463676, + 0.011971727013587952, + 0.007457728497684002, + 0.01124703697860241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.003627745434641838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203409664332867 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203409664332867, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203409664332867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06684928387403488, + 0.0906423032283783, + 0.0516238696873188, + 0.08072115480899811, + 0.05645952373743057, + 0.03462989628314972, + 0.04984460771083832, + 0.04384854808449745, + 0.04758575186133385, + 0.4777950644493103 + ], + "credit_map": [ + 0.0066452925093472, + 0.012402313761413097, + 0.006025196984410286, + 0.005035551730543375, + 0.014574545435607433, + 0.007354324217885733, + 0.007770100142806768, + 0.007041951175779104, + 0.014463038183748722, + 0.007752064615488052, + 0.011366236954927444, + 0.011426921933889389, + 0.013746578246355057, + 0.013682099990546703, + 0.00853416882455349, + 0.009671185165643692, + 0.012878126464784145, + 0.005238196812570095, + 0.004439338576048613, + 0.006524479947984219, + 0.003075412707403302, + 0.011888754554092884, + 0.0052217766642570496, + 0.0023577811662107706, + 0.008373068645596504, + 0.011680412106215954, + 0.011587549932301044, + 0.008425361476838589, + 0.011429950594902039, + 0.0068989950232207775, + 0.014290844090282917, + 0.008034961298108101, + 0.006026432849466801, + 0.00868426263332367, + 0.009490971453487873, + 0.01034652441740036, + 0.005865671206265688, + 0.006177636329084635, + 0.006144373212009668, + 0.003723649075254798, + 0.004055527504533529, + 0.0041762045584619045, + 0.005899864714592695, + 0.004561874084174633, + 0.004621234722435474, + 0.005132012069225311, + 0.0031656476203352213, + 0.0030175289139151573, + 0.009040034376084805, + 0.003983625210821629, + 0.005429130978882313, + 0.009720277041196823, + 0.00321201141923666, + 0.009948731400072575, + 0.004992157686501741, + 0.003518636105582118, + 0.0045098536647856236, + 0.00506208511069417, + 0.006071936804801226, + 0.00918189063668251, + 0.003869415493682027, + 0.0053928084671497345, + 0.006154588423669338, + 0.003605965292081237, + 0.005638164468109608, + 0.006286164280027151, + 0.010322192683815956, + 0.00468285521492362, + 0.003999308217316866, + 0.005338761955499649, + 0.0034875960554927588, + 0.007830703631043434, + 0.05855107307434082, + 0.0629638209939003, + 0.10265302658081055, + 0.058834198862314224, + 0.03729065880179405, + 0.05664684623479843, + 0.04790202155709267, + 0.05295339971780777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.014203409664332867 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781643748283386 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781643748283386, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781643748283386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050400249660015106, + 0.0671970471739769, + 0.04418542981147766, + 0.04044553264975548, + 0.04003967344760895, + 0.03127939999103546, + 0.03760969638824463, + 0.06434841454029083, + 0.4748343527317047, + 0.149660125374794 + ], + "credit_map": [ + 0.005465322639793158, + 0.009790470823645592, + 0.00426253629848361, + 0.00327132036909461, + 0.014021256938576698, + 0.0061858780682086945, + 0.004203238990157843, + 0.003200231119990349, + 0.007543713785707951, + 0.008103505708277225, + 0.010868776589632034, + 0.009811041876673698, + 0.01067502610385418, + 0.008203951641917229, + 0.005541989579796791, + 0.0064490484073758125, + 0.006313901394605637, + 0.005950091872364283, + 0.011119766160845757, + 0.004513472318649292, + 0.003429445205256343, + 0.004357595462352037, + 0.004586497321724892, + 0.0039146700873970985, + 0.006493894848972559, + 0.00765643548220396, + 0.003998230677098036, + 0.0047293067909777164, + 0.004940871149301529, + 0.005856260657310486, + 0.0043940762989223, + 0.0023764632642269135, + 0.004541967995464802, + 0.004504594020545483, + 0.004503708332777023, + 0.007363526616245508, + 0.004971345420926809, + 0.00522291986271739, + 0.004612073302268982, + 0.004319541156291962, + 0.0036168501246720552, + 0.005007181316614151, + 0.0032350358087569475, + 0.003994222264736891, + 0.0047554196789860725, + 0.004664696753025055, + 0.0022468839306384325, + 0.0037591145373880863, + 0.008005766198039055, + 0.003465368878096342, + 0.004233939573168755, + 0.002769619692116976, + 0.0033640810288488865, + 0.01084673497825861, + 0.002451862907037139, + 0.0024723249953240156, + 0.007989879697561264, + 0.0052198502235114574, + 0.01121232658624649, + 0.015805520117282867, + 0.003847280750051141, + 0.006895368453115225, + 0.0066636912524700165, + 0.0067145065404474735, + 0.054890699684619904, + 0.04433354362845421, + 0.11075986921787262, + 0.08045609295368195, + 0.031013434752821922, + 0.06092483922839165, + 0.0373186431825161, + 0.05513725057244301, + 0.020128754898905754, + 0.018450316041707993, + 0.03153621777892113, + 0.018797049298882484, + 0.01028791069984436, + 0.019354429095983505, + 0.013574441894888878, + 0.0175310168415308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.009781643748283386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654209822416306 + ], + "opposite_logprob_tokens": [ + -3.048654317855835 + ], + "logprob_reward": -0.048654209822416306, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654209822416306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08601029217243195, + 0.02595086395740509, + 0.016942720860242844, + 0.010701518505811691, + 0.01168789155781269, + 0.011606783606112003, + 0.017233077436685562, + 0.016856396570801735, + 0.19941753149032593, + 0.6035929322242737 + ], + "credit_map": [ + 0.009340357035398483, + 0.015973808243870735, + 0.007868100889027119, + 0.006100933067500591, + 0.022990545257925987, + 0.009663525968790054, + 0.008413597010076046, + 0.005659416783601046, + 0.004088369198143482, + 0.0022369748912751675, + 0.0027529108338057995, + 0.0034382331650704145, + 0.0031350278295576572, + 0.004243762698024511, + 0.003032345324754715, + 0.0030232411809265614, + 0.004598679021000862, + 0.0016113437013700604, + 0.001380417961627245, + 0.001471473602578044, + 0.0012370283948257565, + 0.004711724352091551, + 0.001139304949901998, + 0.0007927485276013613, + 0.0019838125444948673, + 0.001693714875727892, + 0.0015231651486828923, + 0.0010813265107572079, + 0.0010308048222213984, + 0.0016943280352279544, + 0.0008137469412758946, + 0.0008806196856312454, + 0.001541406149044633, + 0.0017478227382525802, + 0.0014490234898403287, + 0.001102387672290206, + 0.0016887227538973093, + 0.0016506989486515522, + 0.0011129516642540693, + 0.001394876861013472, + 0.001606662874110043, + 0.0009939245646819472, + 0.001047009602189064, + 0.001289089792408049, + 0.0013469455298036337, + 0.002796083688735962, + 0.0011702736373990774, + 0.0013567935675382614, + 0.003989220596849918, + 0.0013031155103817582, + 0.001419131993316114, + 0.002080017002299428, + 0.0011770309647545218, + 0.004675219766795635, + 0.0011071953922510147, + 0.0014821466756984591, + 0.0019850688986480236, + 0.0018695450853556395, + 0.0016569953877478838, + 0.002882851054891944, + 0.0014238763833418489, + 0.002493779407814145, + 0.002807964337989688, + 0.001736316131427884, + 0.021099090576171875, + 0.019956398755311966, + 0.04969818517565727, + 0.03288838267326355, + 0.011501827277243137, + 0.023591037839651108, + 0.015404793433845043, + 0.025277819484472275, + 0.07435047626495361, + 0.07967416942119598, + 0.12064962089061737, + 0.0784955695271492, + 0.04314372316002846, + 0.07760949432849884, + 0.06364365667104721, + 0.06602619588375092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.048654209822416306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an extreme close-up shot from ground level, revealing an ultra-high-definition, photorealistic CGI image focused on a patch of moist, dark brown soil rich in organic matter. Soft morning light slants down, making tiny water droplets cling to the variously sized soil particles shimmer with a shimmering brilliance. Suddenly, the ground arches slightly, a tiny crack appears. A fresh, emerald green sprout, its curved stem thrusting upward with unstoppable vitality, pushes away the soil particles above it, sending clumps of moist earth tumbling to the side. As the translucent sprout fully emerges from the earth, its stem gradually straightens, and the two cotyledons curled at the top begin to carefully and nearly symmetrically unfold to the sides, revealing the fresher green and distinct veins within. Throughout this process, the camera rises in a smooth and steady rhythm from a close-up perspective, maintaining focus on the growing sprout, ultimately settling at a perspective parallel to the unfolding leaves. The whole process appears sacred and poetic under the warm light, creating a tranquil atmosphere of rebirth, resilience and hope.", + "seed": 2403, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_14/index_1474_seed_2403_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027233319962397216, + "vlm_logprob": -0.027233319962397216, + "vlm_yes_no_margin": 4.3249999523162845, + "vlm_token_credit": -0.027233319962397216, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00368095631711185 + ], + "opposite_logprob_tokens": [ + -6.003681182861328 + ], + "logprob_reward": -0.00368095631711185, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00368095631711185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09526478499174118, + 0.04880165308713913, + 0.25054389238357544, + 0.07438574731349945, + 0.2886281907558441, + 0.08780290931463242, + 0.046350155025720596, + 0.038591787219047546, + 0.0450628362596035, + 0.024568025022745132 + ], + "credit_map": [ + 0.010938530787825584, + 0.012931319884955883, + 0.020416731014847755, + 0.013294906355440617, + 0.009074851870536804, + 0.006937393452972174, + 0.014383702538907528, + 0.007287342566996813, + 0.003394882660359144, + 0.008769048377871513, + 0.013462682254612446, + 0.0062615224160254, + 0.0039941854774951935, + 0.003607714083045721, + 0.00431510154157877, + 0.004996519535779953, + 0.026854248717427254, + 0.056541237980127335, + 0.016827836632728577, + 0.09141194075345993, + 0.015672598034143448, + 0.009664488956332207, + 0.015037459321320057, + 0.01853409968316555, + 0.014149626716971397, + 0.005826791748404503, + 0.017129596322774887, + 0.010550185106694698, + 0.008560022339224815, + 0.005807250272482634, + 0.007914622314274311, + 0.004447652958333492, + 0.05132476985454559, + 0.014688743278384209, + 0.011370902881026268, + 0.009019747376441956, + 0.052794694900512695, + 0.1272851973772049, + 0.01093924418091774, + 0.011204907670617104, + 0.02757539413869381, + 0.007139613386243582, + 0.0037863822653889656, + 0.004363834857940674, + 0.0038921514060348272, + 0.03432083502411842, + 0.0028986947145313025, + 0.0038260030560195446, + 0.012611585669219494, + 0.0034547948744148016, + 0.0035448919516056776, + 0.003184169065207243, + 0.003345389384776354, + 0.014438573271036148, + 0.001976742409169674, + 0.0037940095644444227, + 0.007035504095256329, + 0.005234181880950928, + 0.004441044293344021, + 0.004205542150884867, + 0.0033329271245747805, + 0.007087105885148048, + 0.0032059885561466217, + 0.00404949439689517, + 0.005876168143004179, + 0.006039525382220745, + 0.008083295077085495, + 0.006178851239383221, + 0.003600476775318384, + 0.004282866604626179, + 0.006689061410725117, + 0.004312592558562756, + 0.003482737811282277, + 0.0029260728042572737, + 0.0047423928044736385, + 0.002232069382444024, + 0.002136638853698969, + 0.0025152305606752634, + 0.002236243337392807, + 0.004296640399843454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.00368095631711185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651796538382769 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004651796538382769, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651796538382769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05381910130381584, + 0.0585014782845974, + 0.33345675468444824, + 0.0984332263469696, + 0.12401797622442245, + 0.21512766182422638, + 0.018864892423152924, + 0.03645437955856323, + 0.02924945577979088, + 0.032075103372335434 + ], + "credit_map": [ + 0.008113878779113293, + 0.004679732955992222, + 0.010807338170707226, + 0.007012329064309597, + 0.006204404402524233, + 0.004468089435249567, + 0.005482160951942205, + 0.007051168475300074, + 0.0032954500056803226, + 0.011115841567516327, + 0.013941607438027859, + 0.00762920593842864, + 0.01048902329057455, + 0.002984169404953718, + 0.004873346071690321, + 0.004172831773757935, + 0.05528981611132622, + 0.05460255593061447, + 0.024213338270783424, + 0.12594886124134064, + 0.018249444663524628, + 0.01103937067091465, + 0.019365718588232994, + 0.024747639894485474, + 0.025412404909729958, + 0.006240360904484987, + 0.027774646878242493, + 0.007678493857383728, + 0.009958471171557903, + 0.005400529131293297, + 0.00866416934877634, + 0.007304149214178324, + 0.016478119418025017, + 0.005358736962080002, + 0.0070114415138959885, + 0.005878469906747341, + 0.02704235352575779, + 0.05024264007806778, + 0.006131814792752266, + 0.005874395836144686, + 0.06462845951318741, + 0.016461532562971115, + 0.008520426228642464, + 0.009544700384140015, + 0.011427530087530613, + 0.08668898046016693, + 0.007861688733100891, + 0.009994336403906345, + 0.0037853948306292295, + 0.002844213740900159, + 0.0019658650271594524, + 0.0011557815596461296, + 0.0023616314865648746, + 0.003785553155466914, + 0.0013054842129349709, + 0.0016609696904197335, + 0.00602267449721694, + 0.005247307475656271, + 0.004890983458608389, + 0.002994152018800378, + 0.003999934997409582, + 0.006058999337255955, + 0.0032959580421447754, + 0.00394437275826931, + 0.004060161765664816, + 0.0035463569220155478, + 0.005269866902381182, + 0.003794458694756031, + 0.002394112292677164, + 0.002869493793696165, + 0.004290884826332331, + 0.0030241194181144238, + 0.005196878220885992, + 0.0022700524423271418, + 0.006747058127075434, + 0.003684710245579481, + 0.0019075479358434677, + 0.00331104826182127, + 0.003378219436854124, + 0.005579588934779167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.004651796538382769 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181953813880682 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.0181953813880682, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181953813880682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050975024700164795, + 0.018213286995887756, + 0.24941225349903107, + 0.054413698613643646, + 0.277100145816803, + 0.168268620967865, + 0.029909618198871613, + 0.06866288930177689, + 0.03834185376763344, + 0.044702645391225815 + ], + "credit_map": [ + 0.006409498862922192, + 0.004518251400440931, + 0.006182838696986437, + 0.005576462484896183, + 0.009257381781935692, + 0.003597363829612732, + 0.00919108185917139, + 0.0062421471811831, + 0.0021109613589942455, + 0.0018952945247292519, + 0.002206620294600725, + 0.0023381575010716915, + 0.0047689164057374, + 0.0013925374951213598, + 0.0021785080898553133, + 0.0013222924899309874, + 0.03290485963225365, + 0.045611608773469925, + 0.01938997022807598, + 0.09685985743999481, + 0.01100468821823597, + 0.008104290813207626, + 0.014092146418988705, + 0.02144482731819153, + 0.011963454075157642, + 0.004193255212157965, + 0.014405635185539722, + 0.00413866713643074, + 0.006526952143758535, + 0.004913592245429754, + 0.004511066246777773, + 0.0037610759027302265, + 0.056356120854616165, + 0.01370592974126339, + 0.009522301144897938, + 0.005308924242854118, + 0.04784262180328369, + 0.12585540115833282, + 0.011161550879478455, + 0.007347275037318468, + 0.052748750895261765, + 0.013310359790921211, + 0.006293418351560831, + 0.007175026927143335, + 0.008202326484024525, + 0.06908798962831497, + 0.0049982089549303055, + 0.006452537607401609, + 0.00374627229757607, + 0.004032103810459375, + 0.003383895382285118, + 0.002909725299105048, + 0.005080294329673052, + 0.0049662464298307896, + 0.0024785965215414762, + 0.003312483662739396, + 0.010049640201032162, + 0.012948489747941494, + 0.007259003818035126, + 0.006930571049451828, + 0.0097308699041605, + 0.00920474249869585, + 0.006743685808032751, + 0.0057958862744271755, + 0.005250362679362297, + 0.004480248317122459, + 0.007714881561696529, + 0.004649997688829899, + 0.0029961089603602886, + 0.003974702209234238, + 0.005508382339030504, + 0.0037671697791665792, + 0.005376262124627829, + 0.00417201966047287, + 0.00790941808372736, + 0.007156072184443474, + 0.003791988827288151, + 0.004235697910189629, + 0.00585161941125989, + 0.006209565792232752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.0181953813880682 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058405987918377 + ], + "opposite_logprob_tokens": [ + -3.8955841064453125 + ], + "logprob_reward": -0.02058405987918377, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058405987918377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0267574954777956, + 0.01740616001188755, + 0.5365309119224548, + 0.08115572482347488, + 0.18889404833316803, + 0.03439183905720711, + 0.05607220530509949, + 0.016988884657621384, + 0.019913634285330772, + 0.021889053285121918 + ], + "credit_map": [ + 0.003205736167728901, + 0.0036018688697367907, + 0.004827225115150213, + 0.00470692990347743, + 0.0021547109354287386, + 0.002176090143620968, + 0.003140221582725644, + 0.0029447153210639954, + 0.0008951494819484651, + 0.002300390973687172, + 0.003557808231562376, + 0.004164487589150667, + 0.0022851412650197744, + 0.0010445944499224424, + 0.001735219033434987, + 0.0014233702095225453, + 0.03190681338310242, + 0.1593359112739563, + 0.03982596844434738, + 0.19948525726795197, + 0.023645512759685516, + 0.01909886673092842, + 0.029352577403187752, + 0.03388005495071411, + 0.022610345855355263, + 0.003709949553012848, + 0.025121619924902916, + 0.005494742188602686, + 0.0050961957313120365, + 0.004455321468412876, + 0.007902534678578377, + 0.006765021476894617, + 0.03909609466791153, + 0.008797061629593372, + 0.006864641793072224, + 0.006463099271059036, + 0.03257269784808159, + 0.0784279927611351, + 0.007398250978440046, + 0.009274208918213844, + 0.006573264021426439, + 0.004005494061857462, + 0.0022706708405166864, + 0.0032395031303167343, + 0.002493042964488268, + 0.011468514800071716, + 0.0019719889387488365, + 0.002369360765442252, + 0.011412872932851315, + 0.004161396995186806, + 0.005774139892309904, + 0.005479681305587292, + 0.003509070025756955, + 0.013334177434444427, + 0.0033579946029931307, + 0.009042870253324509, + 0.002674232004210353, + 0.0021269218996167183, + 0.002706624334678054, + 0.0015485403127968311, + 0.0018581660697236657, + 0.0018286434933543205, + 0.0013551617739722133, + 0.002890596864745021, + 0.002099385717883706, + 0.0024287644773721695, + 0.003200559178367257, + 0.003222095314413309, + 0.0015736873028799891, + 0.0020197632256895304, + 0.0034484872594475746, + 0.0019208937883377075, + 0.003277255455031991, + 0.0017171049257740378, + 0.004356604069471359, + 0.0026787491515278816, + 0.0015613483265042305, + 0.002301908563822508, + 0.0020810645073652267, + 0.003915017936378717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.02058405987918377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot an extreme close-up from ground level showing moist, dark brown soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a fresh emerald green sprout with a curved stem emerge from the soil and push away earth particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two cotyledons at the top of the sprout unfold symmetrically to reveal distinct veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rise smoothly from a close-up perspective to a view parallel to the unfolding leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light causing water droplets on the soil to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890544056892395 + ], + "opposite_logprob_tokens": [ + -2.4640543460845947 + ], + "logprob_reward": -0.0890544056892395, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890544056892395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07885615527629852, + 0.047962550073862076, + 0.24962124228477478, + 0.1403093785047531, + 0.10646981000900269, + 0.16523195803165436, + 0.06804729253053665, + 0.07652495056390762, + 0.03620954975485802, + 0.03076712228357792 + ], + "credit_map": [ + 0.01421637274324894, + 0.010930314660072327, + 0.010875745676457882, + 0.014046525582671165, + 0.007037954870611429, + 0.007159812841564417, + 0.009020108729600906, + 0.005569321569055319, + 0.006315970327705145, + 0.0046517495065927505, + 0.006144430488348007, + 0.007805761881172657, + 0.01040352787822485, + 0.003985583316534758, + 0.0052398149855434895, + 0.0034157107584178448, + 0.03098144941031933, + 0.08822928369045258, + 0.014555753208696842, + 0.06980951130390167, + 0.010735482908785343, + 0.010666081681847572, + 0.012028208002448082, + 0.012615463696420193, + 0.039658043533563614, + 0.006616227328777313, + 0.04313259944319725, + 0.010016288608312607, + 0.009320106357336044, + 0.007230693474411964, + 0.012795443646609783, + 0.011539969593286514, + 0.02070743776857853, + 0.00792425125837326, + 0.005425981245934963, + 0.005866077728569508, + 0.015464692376554012, + 0.041315723210573196, + 0.004529897123575211, + 0.00523575022816658, + 0.04886912181973457, + 0.012134681455790997, + 0.00771318981423974, + 0.009453579783439636, + 0.008172332309186459, + 0.06441319733858109, + 0.00675235316157341, + 0.0077235158532857895, + 0.00848650187253952, + 0.006534953135997057, + 0.009695232845842838, + 0.009169409982860088, + 0.007677157409489155, + 0.009088783524930477, + 0.005147866439074278, + 0.012247391976416111, + 0.012496892362833023, + 0.00791716668754816, + 0.009296140633523464, + 0.009406255558133125, + 0.00662500225007534, + 0.014301693066954613, + 0.00864509679377079, + 0.007836706005036831, + 0.004834706895053387, + 0.004769681021571159, + 0.006647796370089054, + 0.005024841520935297, + 0.00277875573374331, + 0.0035537034273147583, + 0.005221294239163399, + 0.0033787719439715147, + 0.004580505192279816, + 0.0029082417022436857, + 0.005233237985521555, + 0.003405332798138261, + 0.0026103800628334284, + 0.002936684526503086, + 0.003102265764027834, + 0.0059904735535383224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.0890544056892395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_15/index_tensor([1268, 1268])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_15/index_tensor([1268, 1268])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..68aa9bce91ae5fbde2f7185bc2014430a978bfd7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_15/index_tensor([1268, 1268])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot from the side, the camera steady, captures the concentration of an elderly craftsman in Steve McCurry's signature documentary style. A rich beam of afternoon sunlight filters through a dusty window, casting a sharp beam of light precisely upon the man and his work. The old craftsman's face, deeply wrinkled, his eyes narrowed, his expression solemn and serene. Dressed in a faded dark blue cotton work shirt, he bends over his workbench. His left hand, calloused yet remarkably steady, supports an unfinished boxwood Buddha statue. His right hand grips a traditional carving knife, its handle glistening with sweat. Carefully yet decisively, he advances, his precise pressure gracefully peeling a thin, curled shaving from the statue's smooth, curved surface. In the process, countless tiny particles of sawdust are stirred up, dancing in the golden light, shimmering like golden dust imbued with life. The entire picture is filled with warm earth tones and strong contrasts of light and shadow, creating an almost sacred atmosphere of tranquility and concentration, and profoundly conveying the beauty of the inheritance of handicrafts.", + "seed": 19664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_15/index_1268_seed_19664_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04952996391803026, + "vlm_logprob": -0.04952996391803026, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04952996391803026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04402482882142067 + ], + "opposite_logprob_tokens": [ + -5.16902494430542 + ], + "logprob_reward": -0.04402482882142067, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.04402482882142067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.324327290058136, + 0.062230516225099564, + 0.10419391095638275, + 0.07226663082838058, + 0.05712726339697838, + 0.06164585426449776, + 0.07246461510658264, + 0.0923948585987091, + 0.09901446104049683, + 0.05433463305234909 + ], + "credit_map": [ + 0.038082994520664215, + 0.11713983863592148, + 0.017951862886548042, + 0.036219578236341476, + 0.02940673939883709, + 0.045789800584316254, + 0.0181075781583786, + 0.021628856658935547, + 0.007998034358024597, + 0.007738856133073568, + 0.008695787750184536, + 0.00840817578136921, + 0.010518794879317284, + 0.009472740814089775, + 0.0036966388579458, + 0.005701480433344841, + 0.01623481884598732, + 0.01590804010629654, + 0.011265465058386326, + 0.007700099144130945, + 0.013773757964372635, + 0.016803544014692307, + 0.008310043253004551, + 0.01419813185930252, + 0.008383561857044697, + 0.009026707150042057, + 0.008791418746113777, + 0.008823464624583721, + 0.012148074805736542, + 0.011150612495839596, + 0.005530849099159241, + 0.008411942981183529, + 0.007148232776671648, + 0.008005412295460701, + 0.006017864681780338, + 0.005722596775740385, + 0.008736261166632175, + 0.008630549535155296, + 0.00513951713219285, + 0.007726829033344984, + 0.008223730139434338, + 0.009691210463643074, + 0.007149652112275362, + 0.007893798872828484, + 0.00905864778906107, + 0.00953845214098692, + 0.004086967557668686, + 0.006003387272357941, + 0.00826334860175848, + 0.012862314470112324, + 0.007257815450429916, + 0.008486353792250156, + 0.012516994029283524, + 0.011172364465892315, + 0.005944563075900078, + 0.005960854236036539, + 0.01823239028453827, + 0.013646707870066166, + 0.009012793190777302, + 0.009892255999147892, + 0.012889979407191277, + 0.013373712077736855, + 0.007904383353888988, + 0.007442637346684933, + 0.014382760040462017, + 0.015357688069343567, + 0.01124660111963749, + 0.012124438770115376, + 0.012002697214484215, + 0.013579939492046833, + 0.00899262074381113, + 0.011327708140015602, + 0.006719002965837717, + 0.009974365122616291, + 0.005801206920295954, + 0.007147616241127253, + 0.008546685799956322, + 0.008543390780687332, + 0.0029970062896609306, + 0.004605353809893131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.04402482882142067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428242176771164 + ], + "opposite_logprob_tokens": [ + -2.017824172973633 + ], + "logprob_reward": -0.1428242176771164, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428242176771164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19971831142902374, + 0.07950100302696228, + 0.13424235582351685, + 0.05688890069723129, + 0.13060876727104187, + 0.05145403370261192, + 0.12539616227149963, + 0.06020384281873703, + 0.11086132377386093, + 0.051125358790159225 + ], + "credit_map": [ + 0.020999427884817123, + 0.039964497089385986, + 0.018234483897686005, + 0.02611483447253704, + 0.02708195522427559, + 0.03421400114893913, + 0.01528404001146555, + 0.017825068905949593, + 0.007753782905638218, + 0.010457555763423443, + 0.010590152814984322, + 0.010619486682116985, + 0.014192218892276287, + 0.011588373221457005, + 0.006044169422239065, + 0.00825526937842369, + 0.020737508311867714, + 0.022546103224158287, + 0.01370485220104456, + 0.012522305361926556, + 0.018536362797021866, + 0.02043924666941166, + 0.009309596382081509, + 0.016446372494101524, + 0.010021149180829525, + 0.007985186763107777, + 0.005903341807425022, + 0.003842290723696351, + 0.00879133865237236, + 0.009880457073450089, + 0.003416953608393669, + 0.007048181723803282, + 0.011418544687330723, + 0.016932550817728043, + 0.011834729462862015, + 0.01627793349325657, + 0.020782502368092537, + 0.020909665152430534, + 0.009502225555479527, + 0.02295062318444252, + 0.008685210719704628, + 0.006508106365799904, + 0.006565652322024107, + 0.006285300012677908, + 0.007593252696096897, + 0.008088755421340466, + 0.0032057405915111303, + 0.004522013012319803, + 0.013070918619632721, + 0.020861370489001274, + 0.011824261397123337, + 0.016864385455846786, + 0.02095470391213894, + 0.018483879044651985, + 0.01186290755867958, + 0.011473730206489563, + 0.006804721429944038, + 0.0057481140829622746, + 0.0068334657698869705, + 0.01423115935176611, + 0.009135124273598194, + 0.009853750467300415, + 0.003500250168144703, + 0.0040972596034407616, + 0.011771232821047306, + 0.0142691470682621, + 0.010791538283228874, + 0.013068394735455513, + 0.01189902238547802, + 0.018501808866858482, + 0.01209321990609169, + 0.018466953188180923, + 0.0067748562432825565, + 0.009486019611358643, + 0.004859818611294031, + 0.007845859974622726, + 0.008308280259370804, + 0.006651623640209436, + 0.003007987281307578, + 0.00419091060757637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.1428242176771164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023313580080866814 + ], + "opposite_logprob_tokens": [ + -3.773313522338867 + ], + "logprob_reward": -0.023313580080866814, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023313580080866814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2705068290233612, + 0.1667831987142563, + 0.12336686253547668, + 0.05810392647981644, + 0.07307416200637817, + 0.03643824905157089, + 0.0698172003030777, + 0.13186761736869812, + 0.025201303884387016, + 0.04484070837497711 + ], + "credit_map": [ + 0.02523854188621044, + 0.06446129083633423, + 0.017450300976634026, + 0.025063546374440193, + 0.02982591651380062, + 0.04242662340402603, + 0.02352251298725605, + 0.04251807555556297, + 0.014067788608372211, + 0.022990597411990166, + 0.019400201737880707, + 0.019805921241641045, + 0.032979097217321396, + 0.026978662237524986, + 0.012118624523282051, + 0.018442293629050255, + 0.010542781092226505, + 0.019487587735056877, + 0.015120776370167732, + 0.012265472672879696, + 0.021752379834651947, + 0.02054920233786106, + 0.009644570760428905, + 0.014004074037075043, + 0.007524742279201746, + 0.006373930722475052, + 0.006866881623864174, + 0.007288588676601648, + 0.009528383612632751, + 0.008722404949367046, + 0.005529799498617649, + 0.006269186269491911, + 0.007249120157212019, + 0.009586342610418797, + 0.009559477679431438, + 0.008508184924721718, + 0.011776164174079895, + 0.01153591275215149, + 0.004990308079868555, + 0.009868641383945942, + 0.0033728359267115593, + 0.007879478856921196, + 0.0034078452736139297, + 0.0037934209685772657, + 0.004570038989186287, + 0.005858636926859617, + 0.003043166594579816, + 0.004512819927185774, + 0.007157889194786549, + 0.010334105230867863, + 0.007177663967013359, + 0.00928327813744545, + 0.010218414478003979, + 0.010671139694750309, + 0.007295617833733559, + 0.007679079193621874, + 0.027070892974734306, + 0.018624775111675262, + 0.011811614967882633, + 0.014194200746715069, + 0.016662470996379852, + 0.019966986030340195, + 0.011294824071228504, + 0.012241829186677933, + 0.002003922825679183, + 0.0020155617967247963, + 0.0020189129281789064, + 0.004209333565086126, + 0.0036159951705485582, + 0.004389472771435976, + 0.0025792638771235943, + 0.004368838854134083, + 0.003933754283934832, + 0.0047711655497550964, + 0.004357428755611181, + 0.006365485955029726, + 0.009790822863578796, + 0.007154111284762621, + 0.003974514547735453, + 0.004493418615311384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.023313580080866814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294596001505852 + ], + "opposite_logprob_tokens": [ + -3.773294687271118 + ], + "logprob_reward": -0.023294596001505852, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294596001505852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20663875341415405, + 0.04587027058005333, + 0.2141435593366623, + 0.043509598821401596, + 0.08470778167247772, + 0.05553444102406502, + 0.04426772519946098, + 0.2637844979763031, + 0.016564005985856056, + 0.02497943490743637 + ], + "credit_map": [ + 0.022830035537481308, + 0.037843216210603714, + 0.015860918909311295, + 0.02033991739153862, + 0.028693927451968193, + 0.036607421934604645, + 0.015417098999023438, + 0.029046189039945602, + 0.0063444762490689754, + 0.00803903117775917, + 0.00503759179264307, + 0.005639125593006611, + 0.006330227944999933, + 0.006842831149697304, + 0.0034143736120313406, + 0.004222608637064695, + 0.03030405193567276, + 0.04444009065628052, + 0.021830221638083458, + 0.023295992985367775, + 0.024973150342702866, + 0.03028496354818344, + 0.015424707904458046, + 0.02359037660062313, + 0.005729577504098415, + 0.00593212665989995, + 0.006029581651091576, + 0.0048268698155879974, + 0.006380888633430004, + 0.006819294299930334, + 0.0034413556568324566, + 0.004349902737885714, + 0.016001243144273758, + 0.0076596117578446865, + 0.013809849508106709, + 0.0108003756031394, + 0.00934735219925642, + 0.012449308298528194, + 0.005014697555452585, + 0.009625335223972797, + 0.007463698275387287, + 0.009640038013458252, + 0.006267930381000042, + 0.00610524183139205, + 0.007936036214232445, + 0.008558458648622036, + 0.0033759106881916523, + 0.006187121383845806, + 0.005947106052190065, + 0.006025637034326792, + 0.004093271680176258, + 0.006894678808748722, + 0.00550836930051446, + 0.006529640406370163, + 0.004094790667295456, + 0.005174228455871344, + 0.06328913569450378, + 0.0378422886133194, + 0.025868264958262444, + 0.028803333640098572, + 0.033894479274749756, + 0.03667351230978966, + 0.01873956434428692, + 0.01867389865219593, + 0.0031751426868140697, + 0.002109746914356947, + 0.0013618478551506996, + 0.002097593853250146, + 0.0024494656827300787, + 0.0025331934448331594, + 0.0009440492140129209, + 0.0018929659854620695, + 0.003616394242271781, + 0.004281011410057545, + 0.0023797592148184776, + 0.0036984265316277742, + 0.003500940976664424, + 0.003376159118488431, + 0.0018015207024291158, + 0.0023252188693732023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.023294596001505852 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37638741731643677, + 0.05915861949324608, + 0.08939549326896667, + 0.0800747498869896, + 0.07331707328557968, + 0.039435770362615585, + 0.10579333454370499, + 0.07338210195302963, + 0.06822167336940765, + 0.03483380004763603 + ], + "credit_map": [ + 0.04118269681930542, + 0.0908612459897995, + 0.023682309314608574, + 0.0368172712624073, + 0.04032347723841667, + 0.06917311251163483, + 0.024211764335632324, + 0.05013555660843849, + 0.006894726771861315, + 0.0070435223169624805, + 0.006200528237968683, + 0.006679029669612646, + 0.011316392570734024, + 0.009562846273183823, + 0.005556706804782152, + 0.005904863588511944, + 0.016513818874955177, + 0.013203396461904049, + 0.008125227876007557, + 0.011066957376897335, + 0.013113331981003284, + 0.013142170384526253, + 0.0059542227536439896, + 0.008276364766061306, + 0.008178412914276123, + 0.007597717922180891, + 0.00992008950561285, + 0.012893438339233398, + 0.012064805254340172, + 0.012904321774840355, + 0.006818781141191721, + 0.009697184897959232, + 0.008962265215814114, + 0.007442988455295563, + 0.010451864451169968, + 0.007890857756137848, + 0.010393994860351086, + 0.012538135051727295, + 0.0043900031596422195, + 0.011246965266764164, + 0.0035568070597946644, + 0.0054996563121676445, + 0.004580326844006777, + 0.005842859856784344, + 0.006614321377128363, + 0.0065560173243284225, + 0.002855052938684821, + 0.003930729813873768, + 0.007270227652043104, + 0.01271693967282772, + 0.01093610841780901, + 0.014840117655694485, + 0.013349962420761585, + 0.018208619207143784, + 0.013035176321864128, + 0.01543619018048048, + 0.01889161393046379, + 0.011193118058145046, + 0.007162306923419237, + 0.008251145482063293, + 0.007669148035347462, + 0.0100319255143404, + 0.0049720266833901405, + 0.0052108159288764, + 0.010429405607283115, + 0.010611477307975292, + 0.007292706985026598, + 0.008688746951520443, + 0.008655047044157982, + 0.008966032415628433, + 0.006676499266177416, + 0.0069017596542835236, + 0.004205067176371813, + 0.004989804700016975, + 0.004238722380250692, + 0.006233864929527044, + 0.0051899501122534275, + 0.004911723546683788, + 0.0018851085333153605, + 0.003179560648277402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.014192597009241581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from the side, the camera steady, captures the concentration of an elderly craftsman in Steve McCurry's signature documentary style. A rich beam of afternoon sunlight filters through a dusty window, casting a sharp beam of light precisely upon the man and his work. The old craftsman's face, deeply wrinkled, his eyes narrowed, his expression solemn and serene. Dressed in a faded dark blue cotton work shirt, he bends over his workbench. His left hand, calloused yet remarkably steady, supports an unfinished boxwood Buddha statue. His right hand grips a traditional carving knife, its handle glistening with sweat. Carefully yet decisively, he advances, his precise pressure gracefully peeling a thin, curled shaving from the statue's smooth, curved surface. In the process, countless tiny particles of sawdust are stirred up, dancing in the golden light, shimmering like golden dust imbued with life. The entire picture is filled with warm earth tones and strong contrasts of light and shadow, creating an almost sacred atmosphere of tranquility and concentration, and profoundly conveying the beauty of the inheritance of handicrafts.", + "seed": 19665, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_15/index_1268_seed_19665_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.043728973250836134, + "vlm_logprob": -0.043728973250836134, + "vlm_yes_no_margin": 4.074999976158142, + "vlm_token_credit": -0.043728973250836134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037796832621097565 + ], + "opposite_logprob_tokens": [ + -5.537796974182129 + ], + "logprob_reward": -0.037796832621097565, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.037796832621097565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2621869742870331, + 0.06376434862613678, + 0.13452289998531342, + 0.09745866805315018, + 0.06624174118041992, + 0.060564249753952026, + 0.0899885818362236, + 0.09709762781858444, + 0.06890714913606644, + 0.05926787108182907 + ], + "credit_map": [ + 0.0427766777575016, + 0.02902989462018013, + 0.03494624048471451, + 0.046064600348472595, + 0.014077305793762207, + 0.012669872492551804, + 0.02844219096004963, + 0.05418017506599426, + 0.010752803646028042, + 0.008464493788778782, + 0.011857510544359684, + 0.01037193275988102, + 0.003456663340330124, + 0.003731236094608903, + 0.008031745441257954, + 0.007097957190126181, + 0.026179300621151924, + 0.015119019895792007, + 0.024243297055363655, + 0.016132857650518417, + 0.008993910625576973, + 0.010916437022387981, + 0.01080410834401846, + 0.02213393896818161, + 0.019451232627034187, + 0.011417177505791187, + 0.019396742805838585, + 0.018862005323171616, + 0.004684702958911657, + 0.0037493191193789244, + 0.009149339981377125, + 0.010748130269348621, + 0.011736062355339527, + 0.008350050076842308, + 0.0117710642516613, + 0.009276246652007103, + 0.003920214716345072, + 0.004387761931866407, + 0.007182323839515448, + 0.009618003852665424, + 0.008656151592731476, + 0.008299116976559162, + 0.009380297735333443, + 0.008327607065439224, + 0.00412742281332612, + 0.004369740374386311, + 0.005358585622161627, + 0.012045318260788918, + 0.017039380967617035, + 0.01512963604182005, + 0.01585477590560913, + 0.008024441078305244, + 0.007646592333912849, + 0.007084019016474485, + 0.010013834573328495, + 0.009195888414978981, + 0.012628819793462753, + 0.01148282177746296, + 0.013334537856280804, + 0.01196626853197813, + 0.006050580646842718, + 0.009402601048350334, + 0.013359726406633854, + 0.01887226104736328, + 0.014517147094011307, + 0.012886385433375835, + 0.012917359359562397, + 0.007297837175428867, + 0.0043822103179991245, + 0.004764558281749487, + 0.004850002471357584, + 0.007291638292372227, + 0.007517063990235329, + 0.005453494377434254, + 0.009831001050770283, + 0.008039773441851139, + 0.0035946478601545095, + 0.005315195303410292, + 0.005877085495740175, + 0.013639603741466999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.037796832621097565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706243991851807 + ], + "opposite_logprob_tokens": [ + -2.1270623207092285 + ], + "logprob_reward": -0.12706243991851807, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706243991851807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20954950153827667, + 0.10269413888454437, + 0.15117008984088898, + 0.04571714997291565, + 0.04210573434829712, + 0.10283911973237991, + 0.07465622574090958, + 0.1589868813753128, + 0.0462481826543808, + 0.0660330206155777 + ], + "credit_map": [ + 0.032733574509620667, + 0.025715995579957962, + 0.02838788740336895, + 0.03565613552927971, + 0.013491022400557995, + 0.011708150617778301, + 0.02538485825061798, + 0.036471884697675705, + 0.01782817207276821, + 0.01157932635396719, + 0.014267781749367714, + 0.012722417712211609, + 0.005440939217805862, + 0.008932762779295444, + 0.00927182100713253, + 0.02265091799199581, + 0.018828939646482468, + 0.01570703089237213, + 0.022344395518302917, + 0.02278895489871502, + 0.0076097832061350346, + 0.0112833920866251, + 0.016470614820718765, + 0.03613698109984398, + 0.00856965035200119, + 0.006293777842074633, + 0.007487475872039795, + 0.008797393180429935, + 0.0032300367020070553, + 0.003600186202675104, + 0.0036351513117551804, + 0.004103475715965033, + 0.006835774052888155, + 0.005536115728318691, + 0.006926370784640312, + 0.008002462796866894, + 0.0031356753315776587, + 0.003506735200062394, + 0.004128117114305496, + 0.004034483805298805, + 0.01779092662036419, + 0.01574576646089554, + 0.017438430339097977, + 0.016399355605244637, + 0.006402492523193359, + 0.005832239054143429, + 0.009307949803769588, + 0.013921957463026047, + 0.01470091100782156, + 0.009513238444924355, + 0.011890638619661331, + 0.00662597268819809, + 0.008197368122637272, + 0.007122037932276726, + 0.008185593411326408, + 0.008420467376708984, + 0.02752612717449665, + 0.023406602442264557, + 0.030892398208379745, + 0.019916243851184845, + 0.008844325318932533, + 0.011258411221206188, + 0.01711818017065525, + 0.020024601370096207, + 0.009155601263046265, + 0.007479691877961159, + 0.008512269705533981, + 0.0044211349450051785, + 0.0032769436948001385, + 0.0028407506179064512, + 0.00483838003128767, + 0.005723411682993174, + 0.011985034681856632, + 0.011401931755244732, + 0.012208803556859493, + 0.008399787358939648, + 0.005929598119109869, + 0.006013310514390469, + 0.0036596483550965786, + 0.006434902548789978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.12706243991851807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026399023830890656 + ], + "opposite_logprob_tokens": [ + -3.6513991355895996 + ], + "logprob_reward": -0.026399023830890656, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026399023830890656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14444682002067566, + 0.15616200864315033, + 0.1327260583639145, + 0.11407037824392319, + 0.14756789803504944, + 0.0743703842163086, + 0.05319022759795189, + 0.03584764897823334, + 0.0441179983317852, + 0.09750059992074966 + ], + "credit_map": [ + 0.021713614463806152, + 0.02234962210059166, + 0.02152169682085514, + 0.025309942662715912, + 0.008621620014309883, + 0.01009125355631113, + 0.017344199120998383, + 0.017494872212409973, + 0.025230303406715393, + 0.023191340267658234, + 0.029162993654608727, + 0.016692476347088814, + 0.00908488966524601, + 0.007925531826913357, + 0.015054108574986458, + 0.029820358380675316, + 0.026432856917381287, + 0.0197575893253088, + 0.026645127683877945, + 0.015225930139422417, + 0.006924569606781006, + 0.008935671299695969, + 0.010928194969892502, + 0.017876122146844864, + 0.026752976700663567, + 0.015580668114125729, + 0.026732202619314194, + 0.01822040230035782, + 0.006157372612506151, + 0.0044811684638261795, + 0.008123020641505718, + 0.008022572845220566, + 0.021857857704162598, + 0.027287999168038368, + 0.021922651678323746, + 0.028623739257454872, + 0.008314896374940872, + 0.008412360213696957, + 0.01543066743761301, + 0.01571773551404476, + 0.012101302854716778, + 0.01309711579233408, + 0.01388486661016941, + 0.01414655800908804, + 0.005089540500193834, + 0.005119468085467815, + 0.005479475017637014, + 0.005452051758766174, + 0.010311991907656193, + 0.007813525386154652, + 0.010337643325328827, + 0.005515682976692915, + 0.004721849225461483, + 0.0035583386197686195, + 0.0051169139333069324, + 0.005814284086227417, + 0.005707543808966875, + 0.005244242958724499, + 0.0064188772812485695, + 0.006404105108231306, + 0.0025169632863253355, + 0.0021272236481308937, + 0.0039733960293233395, + 0.003455294296145439, + 0.007849532179534435, + 0.006733898539096117, + 0.007347707636654377, + 0.005161931272596121, + 0.0040698787197470665, + 0.003017457202076912, + 0.005536722019314766, + 0.004400869365781546, + 0.016127195209264755, + 0.018533406779170036, + 0.017202192917466164, + 0.016825396567583084, + 0.006028132047504187, + 0.004791834391653538, + 0.007537348195910454, + 0.010455100797116756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.026399023830890656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040998938493430614 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040998938493430614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040998938493430614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3458836078643799, + 0.11408735066652298, + 0.09411448240280151, + 0.040643755346536636, + 0.1127445250749588, + 0.0760679543018341, + 0.06708870083093643, + 0.04249369353055954, + 0.0625794306397438, + 0.04429653286933899 + ], + "credit_map": [ + 0.046282682567834854, + 0.048558566719293594, + 0.05084977671504021, + 0.039446018636226654, + 0.03770143538713455, + 0.03992174565792084, + 0.03689665347337723, + 0.04622671753168106, + 0.012355265207588673, + 0.01689332164824009, + 0.01896953582763672, + 0.011059232987463474, + 0.01483682356774807, + 0.01595078967511654, + 0.009563311003148556, + 0.01445906050503254, + 0.010515248402953148, + 0.012564183212816715, + 0.012569649145007133, + 0.009893612004816532, + 0.009821176528930664, + 0.010753990150988102, + 0.013480844907462597, + 0.014515777118504047, + 0.006670760922133923, + 0.004848791286349297, + 0.007916280999779701, + 0.007345964666455984, + 0.0023868016432970762, + 0.0021273838356137276, + 0.003396212123334408, + 0.005951561499387026, + 0.015974678099155426, + 0.014749603345990181, + 0.016408417373895645, + 0.01673378422856331, + 0.008473427034914494, + 0.010591363534331322, + 0.0148860402405262, + 0.014927217736840248, + 0.010990607552230358, + 0.008488098159432411, + 0.012138455174863338, + 0.006280905567109585, + 0.009605945087969303, + 0.01261747907847166, + 0.007183825597167015, + 0.008762635290622711, + 0.009674127213656902, + 0.01119642797857523, + 0.01191707607358694, + 0.006216326262801886, + 0.006438585929572582, + 0.006146853789687157, + 0.006106747779995203, + 0.009392554871737957, + 0.0058349925093352795, + 0.005712389014661312, + 0.0061428481712937355, + 0.005319562274962664, + 0.0032902222592383623, + 0.0031459489837288857, + 0.005152130499482155, + 0.007895603775978088, + 0.012855124659836292, + 0.00976756401360035, + 0.013040591962635517, + 0.01043721940368414, + 0.003658171510323882, + 0.0028357531409710646, + 0.005033144261687994, + 0.004951858893036842, + 0.006759705953299999, + 0.007434946019202471, + 0.007833858020603657, + 0.005870414432138205, + 0.004681907128542662, + 0.003915547393262386, + 0.0030014701187610626, + 0.004798683803528547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0040998938493430614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328667603433132 + ], + "opposite_logprob_tokens": [ + -3.7732865810394287 + ], + "logprob_reward": -0.02328667603433132, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328667603433132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14402849972248077, + 0.08063846081495285, + 0.31283095479011536, + 0.04334336891770363, + 0.07989084720611572, + 0.056121788918972015, + 0.15805190801620483, + 0.056385792791843414, + 0.04736766219139099, + 0.02134072035551071 + ], + "credit_map": [ + 0.02384037710726261, + 0.020174451172351837, + 0.022701751440763474, + 0.02130707912147045, + 0.011039676144719124, + 0.011711636558175087, + 0.018027551472187042, + 0.01522598136216402, + 0.011585220694541931, + 0.006874862592667341, + 0.010371988639235497, + 0.0077338553965091705, + 0.005122002679854631, + 0.00852521788328886, + 0.009170479141175747, + 0.02125483937561512, + 0.04257560521364212, + 0.037177275866270065, + 0.03954309970140457, + 0.027548514306545258, + 0.04105271399021149, + 0.045650847256183624, + 0.027578873559832573, + 0.05170401558279991, + 0.008127809502184391, + 0.0074100918136537075, + 0.008021154440939426, + 0.005843968130648136, + 0.0030073351226747036, + 0.0031729894690215588, + 0.0041155144572257996, + 0.0036445041187107563, + 0.010888750664889812, + 0.009330272674560547, + 0.010391589254140854, + 0.010990321636199951, + 0.006806729827076197, + 0.00707039050757885, + 0.011561211198568344, + 0.012851581908762455, + 0.008855322375893593, + 0.006935066543519497, + 0.008119301870465279, + 0.009026004001498222, + 0.004380411934107542, + 0.0049823676235973835, + 0.006092889234423637, + 0.007730425801128149, + 0.026614829897880554, + 0.02151510678231716, + 0.025701457634568214, + 0.015463510528206825, + 0.014227520674467087, + 0.012541420757770538, + 0.025631258264183998, + 0.016356807202100754, + 0.006122753489762545, + 0.006230267696082592, + 0.007964745163917542, + 0.0061730374582111835, + 0.0035088197328150272, + 0.004699543118476868, + 0.009488088078796864, + 0.01219854038208723, + 0.010096475481987, + 0.007186567410826683, + 0.00938375759869814, + 0.006633691489696503, + 0.0025930062402039766, + 0.0024475499521940947, + 0.0033282358199357986, + 0.0056983791291713715, + 0.004206296056509018, + 0.0026802027132362127, + 0.003657765919342637, + 0.004673629067838192, + 0.00191352772526443, + 0.001413495629094541, + 0.0013143260730430484, + 0.0014814779860898852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.02328667603433132 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from the side, the camera steady, captures the concentration of an elderly craftsman in Steve McCurry's signature documentary style. A rich beam of afternoon sunlight filters through a dusty window, casting a sharp beam of light precisely upon the man and his work. The old craftsman's face, deeply wrinkled, his eyes narrowed, his expression solemn and serene. Dressed in a faded dark blue cotton work shirt, he bends over his workbench. His left hand, calloused yet remarkably steady, supports an unfinished boxwood Buddha statue. His right hand grips a traditional carving knife, its handle glistening with sweat. Carefully yet decisively, he advances, his precise pressure gracefully peeling a thin, curled shaving from the statue's smooth, curved surface. In the process, countless tiny particles of sawdust are stirred up, dancing in the golden light, shimmering like golden dust imbued with life. The entire picture is filled with warm earth tones and strong contrasts of light and shadow, creating an almost sacred atmosphere of tranquility and concentration, and profoundly conveying the beauty of the inheritance of handicrafts.", + "seed": 19666, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_15/index_1268_seed_19666_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06925759054720401, + "vlm_logprob": -0.06925759054720401, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06925759054720401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03959508240222931 + ], + "opposite_logprob_tokens": [ + -4.664595127105713 + ], + "logprob_reward": -0.03959508240222931, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.03959508240222931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1789083182811737, + 0.1112702414393425, + 0.07738927006721497, + 0.06723741441965103, + 0.09326263517141342, + 0.06019444763660431, + 0.2439025342464447, + 0.04930305853486061, + 0.05331209674477577, + 0.06522000581026077 + ], + "credit_map": [ + 0.016825193539261818, + 0.021995101124048233, + 0.0202394500374794, + 0.020936144515872, + 0.02246781438589096, + 0.029650544747710228, + 0.0239948108792305, + 0.022799236699938774, + 0.013306427747011185, + 0.019278213381767273, + 0.012890167534351349, + 0.01572781428694725, + 0.00974442157894373, + 0.01635020785033703, + 0.012039139866828918, + 0.011933844536542892, + 0.010482849553227425, + 0.010630068369209766, + 0.015871118754148483, + 0.010300735011696815, + 0.009789010509848595, + 0.007225997280329466, + 0.004499457310885191, + 0.008590031415224075, + 0.009266496635973454, + 0.008415542542934418, + 0.014041624963283539, + 0.006992048118263483, + 0.008100184611976147, + 0.006983259692788124, + 0.00653097964823246, + 0.006907281000167131, + 0.013733135536313057, + 0.011939464136958122, + 0.016741320490837097, + 0.008770525455474854, + 0.015797104686498642, + 0.009515325538814068, + 0.007495548576116562, + 0.009270206093788147, + 0.006710758898407221, + 0.0107033122330904, + 0.009552771225571632, + 0.007351386826485395, + 0.004665167536586523, + 0.006693284492939711, + 0.003708373522385955, + 0.010809392668306828, + 0.04389509558677673, + 0.0167252104729414, + 0.05583614856004715, + 0.013780794106423855, + 0.06346491724252701, + 0.020571570843458176, + 0.014630503952503204, + 0.014998282305896282, + 0.005390958860516548, + 0.004449645057320595, + 0.007196385879069567, + 0.006896849721670151, + 0.00413492601364851, + 0.008577605709433556, + 0.006443830672651529, + 0.006212853826582432, + 0.006588041316717863, + 0.006868628319352865, + 0.006375193130224943, + 0.00444800965487957, + 0.00423940597102046, + 0.009084437973797321, + 0.006431967485696077, + 0.009276408702135086, + 0.008393542841076851, + 0.007517759222537279, + 0.008004000410437584, + 0.007524372544139624, + 0.004374729469418526, + 0.011369538493454456, + 0.010478965938091278, + 0.007557095494121313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.03959508240222931 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561518847942352 + ], + "opposite_logprob_tokens": [ + -1.60061514377594 + ], + "logprob_reward": -0.22561518847942352, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561518847942352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23774123191833496, + 0.07548008114099503, + 0.11535032838582993, + 0.10504655539989471, + 0.10381989181041718, + 0.08840867131948471, + 0.1082950159907341, + 0.06967949122190475, + 0.0632786974310875, + 0.032900046557188034 + ], + "credit_map": [ + 0.02455470710992813, + 0.03368918597698212, + 0.04426424950361252, + 0.024229755625128746, + 0.05275360867381096, + 0.02019120380282402, + 0.021898474544286728, + 0.016160057857632637, + 0.008824548684060574, + 0.012272622436285019, + 0.009392342530190945, + 0.007681368384510279, + 0.007462434004992247, + 0.009435596875846386, + 0.008097258396446705, + 0.012313911691308022, + 0.0151545200496912, + 0.01482640951871872, + 0.020487608388066292, + 0.02805391512811184, + 0.008060934953391552, + 0.010315541177988052, + 0.009066778235137463, + 0.00938461534678936, + 0.015480946749448776, + 0.010869709774851799, + 0.020958159118890762, + 0.009364177472889423, + 0.022442515939474106, + 0.010335913859307766, + 0.007567308377474546, + 0.008027827367186546, + 0.013952692039310932, + 0.013126600533723831, + 0.01574380323290825, + 0.013641415163874626, + 0.01711053028702736, + 0.01012601237744093, + 0.008881195448338985, + 0.011237645521759987, + 0.011143035255372524, + 0.013562245294451714, + 0.015916479751467705, + 0.007171208504587412, + 0.011440139263868332, + 0.009824227541685104, + 0.004993838258087635, + 0.01435750350356102, + 0.016849715262651443, + 0.008320963010191917, + 0.022445041686296463, + 0.008904965594410896, + 0.02557312324643135, + 0.011015895754098892, + 0.007922202348709106, + 0.007263097446411848, + 0.007568918168544769, + 0.0069383177906274796, + 0.009785301983356476, + 0.009017407894134521, + 0.00545903155580163, + 0.013914018869400024, + 0.007339136209338903, + 0.009657363407313824, + 0.007280346937477589, + 0.006541922688484192, + 0.008473435416817665, + 0.007683795876801014, + 0.004645864479243755, + 0.008212600834667683, + 0.009724409319460392, + 0.010716321878135204, + 0.004118524957448244, + 0.0038610175251960754, + 0.004951588809490204, + 0.0034371220972388983, + 0.0032256534323096275, + 0.0046744621358811855, + 0.004662811756134033, + 0.003968868404626846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.22561518847942352 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211636036634445 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211636036634445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211636036634445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21132755279541016, + 0.14688527584075928, + 0.062064822763204575, + 0.11640127003192902, + 0.11388325691223145, + 0.06814295053482056, + 0.09266586601734161, + 0.09827688336372375, + 0.04002269729971886, + 0.05032951384782791 + ], + "credit_map": [ + 0.022379236295819283, + 0.03627356141805649, + 0.03111160732805729, + 0.022136934101581573, + 0.03466804325580597, + 0.019444335252046585, + 0.0207393579185009, + 0.024574480950832367, + 0.017324041575193405, + 0.02411782741546631, + 0.021315064281225204, + 0.01921917125582695, + 0.025146296247839928, + 0.014032473787665367, + 0.013777748681604862, + 0.01195263396948576, + 0.007623538840562105, + 0.011265874840319157, + 0.009837054647505283, + 0.009770278818905354, + 0.00842801108956337, + 0.0054940092377364635, + 0.004838765598833561, + 0.004807287361472845, + 0.014066483825445175, + 0.014004923403263092, + 0.021467598155140877, + 0.021945949643850327, + 0.014673634432256222, + 0.009077095426619053, + 0.011576345190405846, + 0.009589237160980701, + 0.01584237813949585, + 0.013414637185633183, + 0.025485660880804062, + 0.01995314285159111, + 0.009945539757609367, + 0.01016732957214117, + 0.009683466516435146, + 0.009391091763973236, + 0.00977832917124033, + 0.012701458297669888, + 0.011372946202754974, + 0.0075542074628174305, + 0.005114633124321699, + 0.009283550083637238, + 0.005134805105626583, + 0.007203016430139542, + 0.01346397865563631, + 0.0076244850642979145, + 0.01856374926865101, + 0.009281961247324944, + 0.01836956851184368, + 0.009723217226564884, + 0.007394526153802872, + 0.008244376629590988, + 0.012228669598698616, + 0.01459403894841671, + 0.021000970155000687, + 0.011706139892339706, + 0.01190099585801363, + 0.008288695476949215, + 0.009212070144712925, + 0.00934530608355999, + 0.0051698023453354836, + 0.004479209426790476, + 0.009082028642296791, + 0.003182743676006794, + 0.0036164610646665096, + 0.004303419962525368, + 0.005412406288087368, + 0.004776621703058481, + 0.006921438965946436, + 0.007552587892860174, + 0.008552487939596176, + 0.004726130049675703, + 0.003933385014533997, + 0.0070754713378846645, + 0.004896981175988913, + 0.00667102774605155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014211636036634445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867487773299217 + ], + "opposite_logprob_tokens": [ + -3.0486748218536377 + ], + "logprob_reward": -0.04867487773299217, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867487773299217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16977867484092712, + 0.06724928319454193, + 0.14398527145385742, + 0.04499143362045288, + 0.15665259957313538, + 0.050017934292554855, + 0.06550271064043045, + 0.047767505049705505, + 0.12614938616752625, + 0.1279052346944809 + ], + "credit_map": [ + 0.02052980661392212, + 0.018232591450214386, + 0.03588489070534706, + 0.0175284706056118, + 0.030658267438411713, + 0.015314813703298569, + 0.015372708439826965, + 0.016257131472229958, + 0.009565259329974651, + 0.005761039908975363, + 0.013310534879565239, + 0.005189020652323961, + 0.016382090747356415, + 0.00782343465834856, + 0.0047553591430187225, + 0.004462545737624168, + 0.018549364060163498, + 0.02472967654466629, + 0.021691694855690002, + 0.03456917405128479, + 0.012552057392895222, + 0.01089488249272108, + 0.012129373848438263, + 0.008869045414030552, + 0.006291762460023165, + 0.00509199034422636, + 0.005948287900537252, + 0.00418113311752677, + 0.008780959993600845, + 0.00522454222664237, + 0.0036156189162284136, + 0.0058571393601596355, + 0.0250246562063694, + 0.018962301313877106, + 0.03050650656223297, + 0.02216377854347229, + 0.022877244278788567, + 0.015039360150694847, + 0.010762114077806473, + 0.011316641233861446, + 0.006396939512342215, + 0.0074234167113900185, + 0.011231821030378342, + 0.006110210437327623, + 0.004701673053205013, + 0.005583178251981735, + 0.003448164090514183, + 0.005122533068060875, + 0.00986990612000227, + 0.004738438408821821, + 0.014027658849954605, + 0.0040320963598787785, + 0.014104078523814678, + 0.007089535240083933, + 0.005682682618498802, + 0.00595830986276269, + 0.006511157378554344, + 0.005504684988409281, + 0.007018669508397579, + 0.007065605837851763, + 0.003865476232022047, + 0.005928568542003632, + 0.004659853409975767, + 0.007213490083813667, + 0.012370084412395954, + 0.01221458800137043, + 0.012860479764640331, + 0.014831554144620895, + 0.009322750382125378, + 0.018661078065633774, + 0.025782067328691483, + 0.020106781274080276, + 0.015009818598628044, + 0.014710045419633389, + 0.014031539671123028, + 0.013349060900509357, + 0.009918333031237125, + 0.022268394008278847, + 0.022029194980859756, + 0.016588851809501648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.04867487773299217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819116808474064 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819116808474064, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819116808474064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12433343380689621, + 0.034437648952007294, + 0.09326387196779251, + 0.13410484790802002, + 0.03607209399342537, + 0.06268735975027084, + 0.29424330592155457, + 0.034009676426649094, + 0.11526308208703995, + 0.07158466428518295 + ], + "credit_map": [ + 0.013577399775385857, + 0.021593086421489716, + 0.019757729023694992, + 0.016771920025348663, + 0.008459025993943214, + 0.013099973089993, + 0.014846866019070148, + 0.016227446496486664, + 0.0036549880169332027, + 0.005151469260454178, + 0.00430260831490159, + 0.005713785532861948, + 0.0022608640138059855, + 0.004075854085385799, + 0.004052281845360994, + 0.005225802771747112, + 0.013431062921881676, + 0.015064998529851437, + 0.017925214022397995, + 0.02011697180569172, + 0.009754164144396782, + 0.00633268104866147, + 0.00530069787055254, + 0.005338079296052456, + 0.02114696055650711, + 0.01340791117399931, + 0.02757769450545311, + 0.010859566740691662, + 0.03274610638618469, + 0.012499058619141579, + 0.007102617993950844, + 0.008764947764575481, + 0.005353158339858055, + 0.0059220995754003525, + 0.006399492267519236, + 0.004378402605652809, + 0.004972414113581181, + 0.0036884769797325134, + 0.002014470286667347, + 0.003343582386150956, + 0.007217260543256998, + 0.009628672152757645, + 0.00844829436391592, + 0.007603468373417854, + 0.005028884392231703, + 0.010553218424320221, + 0.005780940409749746, + 0.008426622487604618, + 0.05240700766444206, + 0.01964130997657776, + 0.06731277704238892, + 0.01710311509668827, + 0.0725506842136383, + 0.024742910638451576, + 0.020119192078709602, + 0.020366322249174118, + 0.004167360253632069, + 0.004162607714533806, + 0.005001090932637453, + 0.003437225241214037, + 0.0028678246308118105, + 0.004628775641322136, + 0.003441812936216593, + 0.0063029807060956955, + 0.012239254079759121, + 0.012379847466945648, + 0.013674779795110226, + 0.01097690686583519, + 0.009703528136014938, + 0.01525981817394495, + 0.023792432621121407, + 0.01723652519285679, + 0.008199996314942837, + 0.006659448146820068, + 0.008513404987752438, + 0.007895019836723804, + 0.004975525662302971, + 0.013448294252157211, + 0.011490767821669579, + 0.010402216576039791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.01819116808474064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from the side, the camera steady, captures the concentration of an elderly craftsman in Steve McCurry's signature documentary style. A rich beam of afternoon sunlight filters through a dusty window, casting a sharp beam of light precisely upon the man and his work. The old craftsman's face, deeply wrinkled, his eyes narrowed, his expression solemn and serene. Dressed in a faded dark blue cotton work shirt, he bends over his workbench. His left hand, calloused yet remarkably steady, supports an unfinished boxwood Buddha statue. His right hand grips a traditional carving knife, its handle glistening with sweat. Carefully yet decisively, he advances, his precise pressure gracefully peeling a thin, curled shaving from the statue's smooth, curved surface. In the process, countless tiny particles of sawdust are stirred up, dancing in the golden light, shimmering like golden dust imbued with life. The entire picture is filled with warm earth tones and strong contrasts of light and shadow, creating an almost sacred atmosphere of tranquility and concentration, and profoundly conveying the beauty of the inheritance of handicrafts.", + "seed": 19667, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_15/index_1268_seed_19667_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06552634462714195, + "vlm_logprob": -0.06552634462714195, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.06552634462714195, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058392442762851715 + ], + "opposite_logprob_tokens": [ + -4.683392524719238 + ], + "logprob_reward": -0.058392442762851715, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.058392442762851715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18479019403457642, + 0.034187495708465576, + 0.02696993201971054, + 0.03671417012810707, + 0.01961538754403591, + 0.025569496676325798, + 0.026449115946888924, + 0.03568489849567413, + 0.48625582456588745, + 0.12376350164413452 + ], + "credit_map": [ + 0.03801223635673523, + 0.029439564794301987, + 0.03882159665226936, + 0.022147150710225105, + 0.013494938611984253, + 0.013809829950332642, + 0.012532352469861507, + 0.016532529145479202, + 0.0030429502949118614, + 0.005158146843314171, + 0.004477634560316801, + 0.004493946209549904, + 0.00448491470888257, + 0.0042926957830786705, + 0.002975249430164695, + 0.005261960905045271, + 0.0028073368594050407, + 0.005016274284571409, + 0.0036918676923960447, + 0.003375288797542453, + 0.0019439742900431156, + 0.004139285534620285, + 0.0017656246200203896, + 0.004230281803756952, + 0.003831935115158558, + 0.0046204146929085255, + 0.00608791084960103, + 0.007419601548463106, + 0.002662567188963294, + 0.005236173514276743, + 0.0029151791241019964, + 0.003940391354262829, + 0.0013552861055359244, + 0.002954986412078142, + 0.002799190115183592, + 0.0023772153072059155, + 0.0015069545479491353, + 0.004138021729886532, + 0.001405162620358169, + 0.0030785715207457542, + 0.001987496856600046, + 0.003856763942167163, + 0.004106962587684393, + 0.003775231773033738, + 0.0016719403211027384, + 0.00415534945204854, + 0.002316704485565424, + 0.0036990472581237555, + 0.002258168300613761, + 0.004058942664414644, + 0.003955342806875706, + 0.004962834529578686, + 0.0022131733130663633, + 0.0035567216109484434, + 0.002493127714842558, + 0.0029508068691939116, + 0.00262760603800416, + 0.005858427379280329, + 0.006289373151957989, + 0.005048843100667, + 0.0028083063662052155, + 0.003149844706058502, + 0.00456552067771554, + 0.0053369770757853985, + 0.03184012323617935, + 0.057625651359558105, + 0.06707606464624405, + 0.08660127967596054, + 0.03484531491994858, + 0.03638887777924538, + 0.0888146385550499, + 0.08306391537189484, + 0.011174183338880539, + 0.023145049810409546, + 0.016962895169854164, + 0.017647024244070053, + 0.00722717959433794, + 0.009361478500068188, + 0.015621775761246681, + 0.022623907774686813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.058392442762851715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039927303791046 + ], + "opposite_logprob_tokens": [ + -1.910399317741394 + ], + "logprob_reward": -0.16039927303791046, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039927303791046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13586457073688507, + 0.04015887901186943, + 0.03193431347608566, + 0.018512176349759102, + 0.020490281283855438, + 0.01868639513850212, + 0.02018510363996029, + 0.024145355448126793, + 0.4200204014778137, + 0.2700025737285614 + ], + "credit_map": [ + 0.014409284107387066, + 0.023329326882958412, + 0.02470238320529461, + 0.025708599016070366, + 0.00859109964221716, + 0.012925181537866592, + 0.013483360409736633, + 0.012715333141386509, + 0.003060293383896351, + 0.006742402445524931, + 0.006678035017102957, + 0.008516823872923851, + 0.0036851142067462206, + 0.003333983477205038, + 0.003981301560997963, + 0.004160923883318901, + 0.00459595862776041, + 0.005219544284045696, + 0.005117526277899742, + 0.004299360793083906, + 0.0016440650215372443, + 0.0031809948850423098, + 0.002399889286607504, + 0.00547697301954031, + 0.0015055175172165036, + 0.003392992075532675, + 0.0031688869930803776, + 0.002455764915794134, + 0.0012762624537572265, + 0.002454095520079136, + 0.001906128483824432, + 0.0023525261785835028, + 0.001381475361995399, + 0.0034541564527899027, + 0.002690398134291172, + 0.0032674612011760473, + 0.0021447937469929457, + 0.0030377982184290886, + 0.0015052357921376824, + 0.0030089598149061203, + 0.0016291906358674169, + 0.002739007119089365, + 0.0028630055021494627, + 0.002238661516457796, + 0.0020301821641623974, + 0.0033159113954752684, + 0.001358809182420373, + 0.002511625410988927, + 0.001720189698971808, + 0.0036590537056326866, + 0.0031075943261384964, + 0.0028788892086595297, + 0.0012897466076537967, + 0.003188981208950281, + 0.0015577219892293215, + 0.0027829238679260015, + 0.0021487970370799303, + 0.00449032848700881, + 0.004621736705303192, + 0.003518751822412014, + 0.0016018974129110575, + 0.0024411457125097513, + 0.0019698787946254015, + 0.0033528166823089123, + 0.027193067595362663, + 0.04910846799612045, + 0.058055367320775986, + 0.07560712099075317, + 0.030005373060703278, + 0.029926296323537827, + 0.08069723844528198, + 0.06942744553089142, + 0.01926332898437977, + 0.04130055755376816, + 0.03322505205869675, + 0.04204503819346428, + 0.01589900255203247, + 0.023022696375846863, + 0.03314414247870445, + 0.06210274621844292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.16039927303791046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199946731328964 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018199946731328964, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199946731328964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16965365409851074, + 0.07539521157741547, + 0.1480797976255417, + 0.048526350408792496, + 0.035511359572410583, + 0.026509929448366165, + 0.021720340475440025, + 0.01724514178931713, + 0.15958480536937714, + 0.2977733910083771 + ], + "credit_map": [ + 0.033690232783555984, + 0.023379161953926086, + 0.035276006907224655, + 0.024313582107424736, + 0.011589271016418934, + 0.01280327420681715, + 0.009965219534933567, + 0.01863691583275795, + 0.005456394050270319, + 0.01038394682109356, + 0.011750764213502407, + 0.0129509586840868, + 0.007546077948063612, + 0.011144422926008701, + 0.006016444880515337, + 0.010146204382181168, + 0.02637634240090847, + 0.030708665028214455, + 0.030486036092042923, + 0.018818609416484833, + 0.00926882866770029, + 0.009499234147369862, + 0.00791255384683609, + 0.015009526163339615, + 0.003139319596812129, + 0.010430173948407173, + 0.008747829124331474, + 0.009295552968978882, + 0.00285200122743845, + 0.006326551549136639, + 0.0036349755246192217, + 0.004099948797374964, + 0.0031598559580743313, + 0.004873768426477909, + 0.005232326686382294, + 0.0069757443852722645, + 0.0034366059117019176, + 0.004806713201105595, + 0.0030044701416045427, + 0.004021876957267523, + 0.002162779215723276, + 0.004718071315437555, + 0.004114782437682152, + 0.004197603557258844, + 0.0019262571586295962, + 0.004587618168443441, + 0.002230491954833269, + 0.002572326920926571, + 0.0017793802544474602, + 0.0039863926358520985, + 0.0034984969533979893, + 0.00394932646304369, + 0.0018185590161010623, + 0.0024137317668646574, + 0.002017710590735078, + 0.0022567410487681627, + 0.0011980901472270489, + 0.003094145329669118, + 0.002816771389916539, + 0.0027176581788808107, + 0.001397889107465744, + 0.0022734201047569513, + 0.0015989396488294005, + 0.0021482279989868402, + 0.009942696429789066, + 0.01763530634343624, + 0.0217676293104887, + 0.028395792469382286, + 0.01176302507519722, + 0.013268297538161278, + 0.029848940670490265, + 0.026963114738464355, + 0.024352479726076126, + 0.04621225595474243, + 0.04215913265943527, + 0.04035433009266853, + 0.016235070303082466, + 0.027247436344623566, + 0.03077474795281887, + 0.07043793797492981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.018199946731328964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005471736192703 + ], + "opposite_logprob_tokens": [ + -2.695054769515991 + ], + "logprob_reward": -0.07005471736192703, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005471736192703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20166391134262085, + 0.1043165847659111, + 0.1459515392780304, + 0.09852790832519531, + 0.023903552442789078, + 0.03253373131155968, + 0.02692718803882599, + 0.03721395507454872, + 0.287408709526062, + 0.041552912443876266 + ], + "credit_map": [ + 0.038783054798841476, + 0.02953287586569786, + 0.03923702985048294, + 0.02245483174920082, + 0.01562218926846981, + 0.020759427919983864, + 0.011388647370040417, + 0.023885836824774742, + 0.016855504363775253, + 0.017250407487154007, + 0.023145437240600586, + 0.017545463517308235, + 0.0060777002945542336, + 0.009784768335521221, + 0.007033628411591053, + 0.006623666733503342, + 0.024467483162879944, + 0.026258470490574837, + 0.029659714549779892, + 0.02206471562385559, + 0.009256716817617416, + 0.011270135641098022, + 0.007139080669730902, + 0.015835216268897057, + 0.009939047507941723, + 0.018904002383351326, + 0.018835579976439476, + 0.016680963337421417, + 0.007605328690260649, + 0.010063149966299534, + 0.008975956588983536, + 0.007523884531110525, + 0.00196123612113297, + 0.004215608816593885, + 0.0037470951210707426, + 0.004715771414339542, + 0.001988352509215474, + 0.002344603883102536, + 0.0021082209423184395, + 0.002822664799168706, + 0.0024841215927153826, + 0.004279663786292076, + 0.005036463029682636, + 0.004898028448224068, + 0.003368895733729005, + 0.006291774567216635, + 0.0024380104150623083, + 0.0037367732729762793, + 0.002165060956031084, + 0.00446885684505105, + 0.004693527705967426, + 0.004803066607564688, + 0.0024717480409890413, + 0.003250504145398736, + 0.002085848478600383, + 0.0029885743279010057, + 0.003267658408731222, + 0.006453882437199354, + 0.008161012083292007, + 0.005650702863931656, + 0.002564122900366783, + 0.0038875385653227568, + 0.0032558583188802004, + 0.003973180428147316, + 0.018822940066456795, + 0.034667063504457474, + 0.04024064540863037, + 0.05263630300760269, + 0.020673196762800217, + 0.020613085478544235, + 0.0536077655851841, + 0.04614771530032158, + 0.003978487569838762, + 0.009457064792513847, + 0.006961341015994549, + 0.0059634847566485405, + 0.0033635906875133514, + 0.003459758125245571, + 0.0032465066760778427, + 0.005122678820043802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.07005471736192703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rich beam of afternoon sunlight filtering through a dusty window to illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny particles of sawdust visible dancing in the light while the craftsman carves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058534324169159 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.02058534324169159, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058534324169159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22092942893505096, + 0.10588004440069199, + 0.13234005868434906, + 0.06450636684894562, + 0.042744845151901245, + 0.03914720192551613, + 0.042512647807598114, + 0.02586805261671543, + 0.26882922649383545, + 0.057242151349782944 + ], + "credit_map": [ + 0.019110774621367455, + 0.04429258033633232, + 0.03564373776316643, + 0.033549562096595764, + 0.019739728420972824, + 0.03294830024242401, + 0.016031451523303986, + 0.01961328834295273, + 0.014541435986757278, + 0.016894767060875893, + 0.021363673731684685, + 0.019079705700278282, + 0.007024421356618404, + 0.009105280973017216, + 0.00808256957679987, + 0.009788187220692635, + 0.02796589769423008, + 0.018694857135415077, + 0.02210766077041626, + 0.016072209924459457, + 0.010756414383649826, + 0.015328706242144108, + 0.007259862497448921, + 0.014154450967907906, + 0.004093719646334648, + 0.009190509095788002, + 0.009198454208672047, + 0.009340554475784302, + 0.00438856054097414, + 0.014108125120401382, + 0.004673135932534933, + 0.00951331201940775, + 0.003223754232749343, + 0.006557614076882601, + 0.006147309206426144, + 0.005151982419192791, + 0.006872386205941439, + 0.006926819682121277, + 0.0029512501787394285, + 0.004913725424557924, + 0.0029668165370821953, + 0.005894371308386326, + 0.006534377578645945, + 0.0050050802528858185, + 0.004439890384674072, + 0.006136441603302956, + 0.0031121158972382545, + 0.0050581106916069984, + 0.004311596974730492, + 0.007610900793224573, + 0.006870571058243513, + 0.003892510896548629, + 0.003931533079594374, + 0.007448992226272821, + 0.0028383093886077404, + 0.005608235951513052, + 0.0021583347115665674, + 0.004533899016678333, + 0.004827933385968208, + 0.003217516466975212, + 0.0024496312253177166, + 0.0035876466426998377, + 0.0015210309065878391, + 0.0035720590967684984, + 0.018297219648957253, + 0.03518082574009895, + 0.03772551938891411, + 0.04873032867908478, + 0.019670646637678146, + 0.019744141027331352, + 0.04736780747771263, + 0.04211270436644554, + 0.005971272476017475, + 0.00807789247483015, + 0.010154899209737778, + 0.010872655548155308, + 0.00392154511064291, + 0.00556830083951354, + 0.005264835432171822, + 0.007410749327391386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.02058534324169159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_tensor([9880, 9880])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_tensor([9880, 9880])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d97edd3608209b54863ac2c6ecb1870f838982fc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_tensor([9880, 9880])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning low-angle close-up, almost completely immersed in the scalding hot pot, the camera trembles subtly with the bubbling liquid. Within this deep, viscous, roiling sea of crimson oil, countless dried red chilies, dark Sichuan peppercorns, and a variety of star anise and bay leaves, saturated with grease and glistening with oil, are swept up by a turbulent vortex. They tumble, collide, and tumble fiercely within the seething red oil, like a vast army engaged in an epic melee. A powerful wave of heat erupts from the pot's bottom, hurling peppers and spices into the air. In slow-motion detail, captured by high-speed photography, they streak across the camera, accompanied by splattering, flaming droplets of oil, before slamming back into the crimson purgatory. The entire picture is illuminated by a hellish fire that originates from the bottom of the pot and shines from the inside out, forming an extremely high contrast of light and shadow, locking the main color in a saturated and almost supernatural red. This is not just food, but also a visual feast full of Tsui Hark's martial arts fantasy style, full of wild and unrestrained energy and spectacular aesthetics of destruction, creating a grand and dangerous atmosphere.", + "seed": 6168, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_9880_seed_6168_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04343300675973296, + "vlm_logprob": -0.04343300675973296, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.04343300675973296, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018417302519083023 + ], + "opposite_logprob_tokens": [ + -4.6434173583984375 + ], + "logprob_reward": -0.018417302519083023, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.018417302519083023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14002130925655365, + 0.10510069876909256, + 0.10367999225854874, + 0.09833887219429016, + 0.25632625818252563, + 0.07920313626527786, + 0.0454876534640789, + 0.08845417946577072, + 0.0314350463449955, + 0.051952898502349854 + ], + "credit_map": [ + 0.02177535556256771, + 0.016969438642263412, + 0.014628682285547256, + 0.0164091307669878, + 0.019811080768704414, + 0.019744165241718292, + 0.016865843906998634, + 0.013817605562508106, + 0.012347312644124031, + 0.007366515696048737, + 0.014873385429382324, + 0.013827312737703323, + 0.013836520723998547, + 0.023567169904708862, + 0.010651095770299435, + 0.008631384931504726, + 0.009055162779986858, + 0.009804321452975273, + 0.012154100462794304, + 0.008727545849978924, + 0.026205524802207947, + 0.016868336126208305, + 0.0142869483679533, + 0.006578052416443825, + 0.014018627814948559, + 0.016199611127376556, + 0.009839370846748352, + 0.009814249351620674, + 0.015424610115587711, + 0.015895923599600792, + 0.00914085190743208, + 0.008005633018910885, + 0.033281806856393814, + 0.03697934001684189, + 0.02456456609070301, + 0.022939156740903854, + 0.05356011539697647, + 0.03261421248316765, + 0.03012971766293049, + 0.022257346659898758, + 0.008893811143934727, + 0.007634224835783243, + 0.009392405860126019, + 0.012446892447769642, + 0.01283365860581398, + 0.007435825187712908, + 0.012856436893343925, + 0.00770987942814827, + 0.008360791020095348, + 0.005518299527466297, + 0.006784691475331783, + 0.005491518881171942, + 0.004963625688105822, + 0.005278644151985645, + 0.0054473006166517735, + 0.003642782336100936, + 0.011823257431387901, + 0.011077703908085823, + 0.01366401743143797, + 0.01584043726325035, + 0.009933038614690304, + 0.007228134199976921, + 0.00972555298358202, + 0.009162036702036858, + 0.0035861560609191656, + 0.0041856542229652405, + 0.0034959791228175163, + 0.003263764549046755, + 0.0040437001734972, + 0.0039243088103830814, + 0.004577644169330597, + 0.004357841331511736, + 0.005644567310810089, + 0.005663035903126001, + 0.008492440916597843, + 0.004963839426636696, + 0.006612842902541161, + 0.006319817621260881, + 0.008146792650222778, + 0.006109560374170542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.018417302519083023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088063940405846 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011088063940405846, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088063940405846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15390492975711823, + 0.08101804554462433, + 0.20115119218826294, + 0.15372666716575623, + 0.08589670807123184, + 0.07244380563497543, + 0.056023385375738144, + 0.099410280585289, + 0.04752414673566818, + 0.04890080913901329 + ], + "credit_map": [ + 0.028648672625422478, + 0.02020999975502491, + 0.016648760065436363, + 0.010594821535050869, + 0.014521702192723751, + 0.03785024210810661, + 0.010881052352488041, + 0.014549674466252327, + 0.006325433496385813, + 0.008779747411608696, + 0.009840816259384155, + 0.012868279591202736, + 0.015451168641448021, + 0.010111699812114239, + 0.009864226914942265, + 0.007776672951877117, + 0.019221456721425056, + 0.01898614875972271, + 0.028032008558511734, + 0.015358924865722656, + 0.04326845332980156, + 0.033587925136089325, + 0.025811025872826576, + 0.016885247081518173, + 0.023865096271038055, + 0.01656257174909115, + 0.01419137604534626, + 0.012787751853466034, + 0.02429894544184208, + 0.028317909687757492, + 0.01575811766088009, + 0.017944900318980217, + 0.009515495039522648, + 0.012106706388294697, + 0.00819218996912241, + 0.0074969325214624405, + 0.0172476377338171, + 0.01202860102057457, + 0.009899569675326347, + 0.009409572929143906, + 0.009316764771938324, + 0.007583205588161945, + 0.009355694986879826, + 0.00898614339530468, + 0.011344369500875473, + 0.010136520490050316, + 0.008355368860065937, + 0.007365739438682795, + 0.011743629351258278, + 0.014660656452178955, + 0.00502630602568388, + 0.005199197679758072, + 0.0057554906234145164, + 0.004179284907877445, + 0.005461843218654394, + 0.003996975254267454, + 0.016118690371513367, + 0.011773455888032913, + 0.020467186346650124, + 0.015312756411731243, + 0.008992631919682026, + 0.007864207029342651, + 0.009164109826087952, + 0.009717246517539024, + 0.00519896624609828, + 0.007092177867889404, + 0.00367758865468204, + 0.004515802953392267, + 0.008798045106232166, + 0.00497079873457551, + 0.007907453924417496, + 0.00536331394687295, + 0.007038387469947338, + 0.005550949368625879, + 0.00749833881855011, + 0.004404824692755938, + 0.005028241313993931, + 0.0037842565216124058, + 0.010952360928058624, + 0.004643450025469065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.011088063940405846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732991896569729 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732991896569729, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732991896569729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19048954546451569, + 0.10650648921728134, + 0.08971180021762848, + 0.0734449103474617, + 0.07335132360458374, + 0.07514237612485886, + 0.06678718328475952, + 0.19367918372154236, + 0.02962835133075714, + 0.10125888884067535 + ], + "credit_map": [ + 0.02622929960489273, + 0.02737480401992798, + 0.02184448018670082, + 0.015894213691353798, + 0.023523524403572083, + 0.03193378821015358, + 0.02525992877781391, + 0.018429504707455635, + 0.011731785722076893, + 0.012299666181206703, + 0.014731019735336304, + 0.015455865301191807, + 0.016538644209504128, + 0.013004527427256107, + 0.01362061407417059, + 0.009124359115958214, + 0.008861362002789974, + 0.009749315679073334, + 0.014904137700796127, + 0.009148743003606796, + 0.014463056810200214, + 0.012221520766615868, + 0.012649543583393097, + 0.007714118342846632, + 0.009262826293706894, + 0.012656093575060368, + 0.008518099784851074, + 0.008093713782727718, + 0.009931185282766819, + 0.009747656993567944, + 0.009320460259914398, + 0.005914881359785795, + 0.010677813552320004, + 0.00868777185678482, + 0.006279970519244671, + 0.007889649830758572, + 0.01332565862685442, + 0.009491214528679848, + 0.00936529878526926, + 0.007633945904672146, + 0.012613959610462189, + 0.0069726272486150265, + 0.008758148178458214, + 0.008818832226097584, + 0.010476846247911453, + 0.012849138118326664, + 0.007932661101222038, + 0.006720166187733412, + 0.01506767701357603, + 0.01468699797987938, + 0.006662505678832531, + 0.005690300837159157, + 0.0058523910120129585, + 0.006429956294596195, + 0.006514112465083599, + 0.005883240140974522, + 0.026568956673145294, + 0.02389327622950077, + 0.0326569490134716, + 0.03215508535504341, + 0.021594634279608727, + 0.01715833880007267, + 0.019131192937493324, + 0.020520778372883797, + 0.003991775680333376, + 0.004245467949658632, + 0.0020447575952857733, + 0.0029938670340925455, + 0.006217816844582558, + 0.0032707997597754, + 0.0036828203592449427, + 0.0031810461077839136, + 0.014031178317964077, + 0.009914564900100231, + 0.008369284681975842, + 0.007254359312355518, + 0.008007017895579338, + 0.01064435113221407, + 0.03640337288379669, + 0.006634758319705725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.006732991896569729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988616228103638 + ], + "opposite_logprob_tokens": [ + -1.8048861026763916 + ], + "logprob_reward": -0.17988616228103638, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988616228103638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13902464509010315, + 0.09121254831552505, + 0.10327338427305222, + 0.08606640994548798, + 0.045381754636764526, + 0.05360029265284538, + 0.10767487436532974, + 0.25546586513519287, + 0.03112589195370674, + 0.08717433363199234 + ], + "credit_map": [ + 0.021275894716382027, + 0.01651255041360855, + 0.013388325460255146, + 0.011963903903961182, + 0.01770312897861004, + 0.02625311352312565, + 0.014072495512664318, + 0.017855241894721985, + 0.008509662933647633, + 0.0071654412895441055, + 0.010280551388859749, + 0.008344795554876328, + 0.015131485648453236, + 0.024430537596344948, + 0.009091398678719997, + 0.008258676156401634, + 0.009591136127710342, + 0.012882155366241932, + 0.010330455377697945, + 0.008881298825144768, + 0.028645915910601616, + 0.015857713297009468, + 0.011411541141569614, + 0.005673168692737818, + 0.014942574314773083, + 0.012581424787640572, + 0.007772570941597223, + 0.008368645794689655, + 0.008188237436115742, + 0.017473412677645683, + 0.009241377003490925, + 0.007498167920857668, + 0.006608322728425264, + 0.006676385644823313, + 0.004890897776931524, + 0.004243845120072365, + 0.007358815521001816, + 0.006035366095602512, + 0.006036638282239437, + 0.003531484631821513, + 0.006344005465507507, + 0.005341788288205862, + 0.006200207397341728, + 0.006577217951416969, + 0.008581053465604782, + 0.007141957990825176, + 0.008224141784012318, + 0.005189921706914902, + 0.018068034201860428, + 0.01744033396244049, + 0.011543923988938332, + 0.014818806201219559, + 0.01563371531665325, + 0.01110750250518322, + 0.01094815507531166, + 0.008114404045045376, + 0.03472912684082985, + 0.029746387153863907, + 0.04555294290184975, + 0.042355578392744064, + 0.026310335844755173, + 0.021098069846630096, + 0.02824040874838829, + 0.027433013543486595, + 0.003350421553477645, + 0.005553088616579771, + 0.0022213696502149105, + 0.0032568397000432014, + 0.006771848537027836, + 0.004216076340526342, + 0.0031618454959243536, + 0.0025944022927433252, + 0.011598393321037292, + 0.009566585533320904, + 0.004015487618744373, + 0.006102407351136208, + 0.009388890117406845, + 0.00944136455655098, + 0.03255484253168106, + 0.0045063612051308155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.17988616228103638 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010405131615698338 + ], + "opposite_logprob_tokens": [ + -6.876040458679199 + ], + "logprob_reward": -0.0010405131615698338, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010405131615698338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12163308262825012, + 0.09592494368553162, + 0.06056877225637436, + 0.1098392978310585, + 0.029075616970658302, + 0.025091873481869698, + 0.05266817659139633, + 0.08441735804080963, + 0.15376918017864227, + 0.2670116722583771 + ], + "credit_map": [ + 0.012273842468857765, + 0.011745278723537922, + 0.015233631245791912, + 0.015601083636283875, + 0.01865890435874462, + 0.013521027751266956, + 0.01676173321902752, + 0.01783759519457817, + 0.011349380947649479, + 0.005959702655673027, + 0.011391157284379005, + 0.00989905558526516, + 0.014643291011452675, + 0.025992482900619507, + 0.008447781205177307, + 0.008242100477218628, + 0.0062233381904661655, + 0.005507598165422678, + 0.006443822756409645, + 0.003549683839082718, + 0.013166033662855625, + 0.011098086833953857, + 0.0074743228033185005, + 0.0071058874018490314, + 0.01711318828165531, + 0.013648366555571556, + 0.00989964697510004, + 0.010038737207651138, + 0.01778024062514305, + 0.021415168419480324, + 0.010840176604688168, + 0.009103778749704361, + 0.0032948015723377466, + 0.005187930539250374, + 0.002424144884571433, + 0.0034002186730504036, + 0.004966472741216421, + 0.00437555368989706, + 0.0031740644481033087, + 0.002252432284876704, + 0.0024117191787809134, + 0.0020848880521953106, + 0.003661598078906536, + 0.004372022580355406, + 0.0037579163908958435, + 0.002990311710163951, + 0.0032668134663254023, + 0.0025466063525527716, + 0.006992338225245476, + 0.009739527478814125, + 0.006568163633346558, + 0.005549791269004345, + 0.006403145380318165, + 0.005930197890847921, + 0.00558483274653554, + 0.005900182761251926, + 0.014313913881778717, + 0.007482029497623444, + 0.007883717305958271, + 0.011844763532280922, + 0.0067292191088199615, + 0.008537068963050842, + 0.011618119664490223, + 0.01600852981209755, + 0.015439149923622608, + 0.022495845332741737, + 0.00826372392475605, + 0.015436726622283459, + 0.031822871416807175, + 0.02126367948949337, + 0.025152185931801796, + 0.013895006850361824, + 0.038065098226070404, + 0.03920616954565048, + 0.020038718357682228, + 0.022741680964827538, + 0.04113326966762543, + 0.023974992334842682, + 0.06775733083486557, + 0.01409443374723196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0010405131615698338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle close-up, almost completely immersed in the scalding hot pot, the camera trembles subtly with the bubbling liquid. Within this deep, viscous, roiling sea of crimson oil, countless dried red chilies, dark Sichuan peppercorns, and a variety of star anise and bay leaves, saturated with grease and glistening with oil, are swept up by a turbulent vortex. They tumble, collide, and tumble fiercely within the seething red oil, like a vast army engaged in an epic melee. A powerful wave of heat erupts from the pot's bottom, hurling peppers and spices into the air. In slow-motion detail, captured by high-speed photography, they streak across the camera, accompanied by splattering, flaming droplets of oil, before slamming back into the crimson purgatory. The entire picture is illuminated by a hellish fire that originates from the bottom of the pot and shines from the inside out, forming an extremely high contrast of light and shadow, locking the main color in a saturated and almost supernatural red. This is not just food, but also a visual feast full of Tsui Hark's martial arts fantasy style, full of wild and unrestrained energy and spectacular aesthetics of destruction, creating a grand and dangerous atmosphere.", + "seed": 6169, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_9880_seed_6169_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0320063341408968, + "vlm_logprob": -0.0320063341408968, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.0320063341408968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013582550920546055 + ], + "opposite_logprob_tokens": [ + -5.138582706451416 + ], + "logprob_reward": -0.013582550920546055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.013582550920546055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12946371734142303, + 0.10504304617643356, + 0.1820267289876938, + 0.06136779114603996, + 0.06891274452209473, + 0.06865768879652023, + 0.06696607172489166, + 0.07191465049982071, + 0.12234674394130707, + 0.12330083549022675 + ], + "credit_map": [ + 0.01622600294649601, + 0.015515217557549477, + 0.01723499596118927, + 0.025592749938368797, + 0.010219238698482513, + 0.010722838342189789, + 0.016348950564861298, + 0.017603706568479538, + 0.012960879132151604, + 0.015153126791119576, + 0.009902965277433395, + 0.016291646286845207, + 0.01617121882736683, + 0.010142617858946323, + 0.013810539618134499, + 0.010610038414597511, + 0.01847568340599537, + 0.014638258144259453, + 0.03778389096260071, + 0.03274144232273102, + 0.015877895057201385, + 0.011446123011410236, + 0.03195948898792267, + 0.019103936851024628, + 0.0065590571612119675, + 0.0067711565643548965, + 0.007509332615882158, + 0.010425765067338943, + 0.00741506926715374, + 0.00787555892020464, + 0.007034804206341505, + 0.007777040358632803, + 0.006886548828333616, + 0.006314453668892384, + 0.007773924618959427, + 0.010658558458089828, + 0.007489581126719713, + 0.009181556291878223, + 0.007885565049946308, + 0.012722551822662354, + 0.007377805188298225, + 0.006681878585368395, + 0.010293655097484589, + 0.007723609916865826, + 0.005644938908517361, + 0.008116256445646286, + 0.009523059241473675, + 0.013296477496623993, + 0.007393899839371443, + 0.008642011322081089, + 0.008996224962174892, + 0.011711747385561466, + 0.009805023670196533, + 0.006253607105463743, + 0.007348934654146433, + 0.006814621388912201, + 0.007307854015380144, + 0.010893582366406918, + 0.007725987117737532, + 0.009858625009655952, + 0.012355712242424488, + 0.006188615690916777, + 0.008102433755993843, + 0.009481831453740597, + 0.023119859397411346, + 0.011399581097066402, + 0.015532445162534714, + 0.010592124424874783, + 0.020023368299007416, + 0.012558253481984138, + 0.011640259064733982, + 0.017480844631791115, + 0.018950454890727997, + 0.015540500171482563, + 0.009697440080344677, + 0.014434775337576866, + 0.01637190766632557, + 0.02166106551885605, + 0.013524332083761692, + 0.013120346702635288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.013582550920546055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304495960474014 + ], + "opposite_logprob_tokens": [ + -3.7733044624328613 + ], + "logprob_reward": -0.023304495960474014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304495960474014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1056043729186058, + 0.1878986656665802, + 0.14144743978977203, + 0.095460906624794, + 0.07463911920785904, + 0.06600192934274673, + 0.09745873510837555, + 0.09339854121208191, + 0.08221450448036194, + 0.055875711143016815 + ], + "credit_map": [ + 0.013056284748017788, + 0.010669833980500698, + 0.017183534801006317, + 0.01649910770356655, + 0.009206589311361313, + 0.00893913023173809, + 0.009143807925283909, + 0.02090608887374401, + 0.02434602566063404, + 0.0391557514667511, + 0.017982052639126778, + 0.02720279060304165, + 0.022215306758880615, + 0.01895776018500328, + 0.020008772611618042, + 0.0180301945656538, + 0.015183514915406704, + 0.01211604941636324, + 0.025779277086257935, + 0.021811315789818764, + 0.017789078876376152, + 0.014999650418758392, + 0.019828638061881065, + 0.013939918950200081, + 0.01481572724878788, + 0.009583849459886551, + 0.008994977921247482, + 0.013761062175035477, + 0.009918212890625, + 0.020223645493388176, + 0.00914511363953352, + 0.009018320590257645, + 0.010013259947299957, + 0.008723033592104912, + 0.010080814361572266, + 0.009578797966241837, + 0.005134688224643469, + 0.013398841954767704, + 0.007389617618173361, + 0.010320071130990982, + 0.005354710388928652, + 0.00623223464936018, + 0.007624820340424776, + 0.008754141628742218, + 0.006083280313760042, + 0.00931496825069189, + 0.010738483630120754, + 0.011899293400347233, + 0.011779909022152424, + 0.008955313824117184, + 0.01553968247026205, + 0.014207355678081512, + 0.012932934798300266, + 0.007955829612910748, + 0.014654058031737804, + 0.011433654464781284, + 0.01816197670996189, + 0.007047847844660282, + 0.007551094517111778, + 0.0097800986841321, + 0.02424364537000656, + 0.008085435256361961, + 0.009766259230673313, + 0.008762177079916, + 0.012331350706517696, + 0.007474994286894798, + 0.007562743499875069, + 0.00563777657225728, + 0.011707242578268051, + 0.011347249150276184, + 0.007962098345160484, + 0.018191056326031685, + 0.008165369741618633, + 0.006581429857760668, + 0.0053854878060519695, + 0.007742504123598337, + 0.007274170406162739, + 0.00609586713835597, + 0.007977328263223171, + 0.006653555668890476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.023304495960474014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097782202064991 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.0097782202064991, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097782202064991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13199156522750854, + 0.10211304575204849, + 0.06336982548236847, + 0.09662959724664688, + 0.12256573885679245, + 0.09373841434717178, + 0.1507713347673416, + 0.09882289916276932, + 0.09045862406492233, + 0.049538932740688324 + ], + "credit_map": [ + 0.018448561429977417, + 0.014773499220609665, + 0.01773885264992714, + 0.02682461217045784, + 0.013964277692139149, + 0.010715502314269543, + 0.01532365009188652, + 0.014202610589563847, + 0.012672949582338333, + 0.014207655563950539, + 0.011215492151677608, + 0.014565529301762581, + 0.012704779393970966, + 0.010097412392497063, + 0.015363222919404507, + 0.011286012828350067, + 0.011271519586443901, + 0.0060243550688028336, + 0.005856623407453299, + 0.008452418260276318, + 0.009254171513020992, + 0.008618584834039211, + 0.006015368737280369, + 0.007876784540712833, + 0.024972639977931976, + 0.011301692575216293, + 0.008311293087899685, + 0.013098200783133507, + 0.010221804492175579, + 0.013980869203805923, + 0.007423712406307459, + 0.007319396361708641, + 0.016528597101569176, + 0.01243605837225914, + 0.015073759481310844, + 0.015512925572693348, + 0.010426383465528488, + 0.013219902291893959, + 0.01315549947321415, + 0.026212619617581367, + 0.0075474814511835575, + 0.012871161103248596, + 0.00905552413314581, + 0.011930020526051521, + 0.009430588223040104, + 0.012818271294236183, + 0.01232096366584301, + 0.017764408141374588, + 0.01682780124247074, + 0.01580730825662613, + 0.019286103546619415, + 0.021547120064496994, + 0.01606898009777069, + 0.014740061946213245, + 0.023316174745559692, + 0.023177793249487877, + 0.016243617981672287, + 0.006849678233265877, + 0.01038476824760437, + 0.015391210094094276, + 0.0190820824354887, + 0.007599954959005117, + 0.01160789467394352, + 0.011663693934679031, + 0.014650493860244751, + 0.007727845571935177, + 0.011145117692649364, + 0.007984290830790997, + 0.013903884217143059, + 0.01281408779323101, + 0.009430999867618084, + 0.012801908887922764, + 0.0064462171867489815, + 0.005686202552169561, + 0.005747412797063589, + 0.005400800611823797, + 0.00603869091719389, + 0.0047890846617519855, + 0.007176164072006941, + 0.008254366926848888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0097782202064991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292840540409088 + ], + "opposite_logprob_tokens": [ + -2.2379283905029297 + ], + "logprob_reward": -0.11292840540409088, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292840540409088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04417578876018524, + 0.05509583279490471, + 0.09880532324314117, + 0.07518504559993744, + 0.030155189335346222, + 0.03757387772202492, + 0.11000008136034012, + 0.10736244916915894, + 0.282075434923172, + 0.15957090258598328 + ], + "credit_map": [ + 0.0053169517777860165, + 0.005823797546327114, + 0.007004656363278627, + 0.006205834913998842, + 0.0039268131367862225, + 0.0028037428855895996, + 0.005638756789267063, + 0.007455242797732353, + 0.0054016406647861, + 0.006987453904002905, + 0.005287015810608864, + 0.00848477240651846, + 0.011757396161556244, + 0.004071812145411968, + 0.008356465958058834, + 0.004749278072267771, + 0.014953644014894962, + 0.009939468465745449, + 0.0140752037987113, + 0.015294809825718403, + 0.011054066941142082, + 0.00904221273958683, + 0.01287640631198883, + 0.011569526046514511, + 0.008860691450536251, + 0.008674778044223785, + 0.00796568300575018, + 0.00993705540895462, + 0.011676711961627007, + 0.010740700177848339, + 0.008581644855439663, + 0.00874777976423502, + 0.004164972808212042, + 0.0034411705564707518, + 0.003911437001079321, + 0.004862579517066479, + 0.003043118165805936, + 0.004293020814657211, + 0.0031506153754889965, + 0.003288277657702565, + 0.005320695694535971, + 0.0038086271379143, + 0.003845370141789317, + 0.004101651720702648, + 0.0033023871947079897, + 0.00415384117513895, + 0.0057050921022892, + 0.007336219307035208, + 0.01248801127076149, + 0.014318102970719337, + 0.013333368115127087, + 0.015462089329957962, + 0.014656552113592625, + 0.0121554434299469, + 0.015551537275314331, + 0.012034992687404156, + 0.012350844219326973, + 0.013634922914206982, + 0.01348920539021492, + 0.01712391898036003, + 0.01328391581773758, + 0.008900155313313007, + 0.013367568142712116, + 0.015211932361125946, + 0.04218341410160065, + 0.023681282997131348, + 0.03009389527142048, + 0.025625359266996384, + 0.045954629778862, + 0.04525528848171234, + 0.02971435710787773, + 0.03956722095608711, + 0.02607920952141285, + 0.017682865262031555, + 0.015441426075994968, + 0.016744334250688553, + 0.016558734700083733, + 0.02293681539595127, + 0.01441720500588417, + 0.02971033751964569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.11292840540409088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004379982128739357 + ], + "opposite_logprob_tokens": [ + -7.750438213348389 + ], + "logprob_reward": -0.0004379982128739357, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004379982128739357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18743079900741577, + 0.07306192815303802, + 0.11165468394756317, + 0.048030704259872437, + 0.03910456597805023, + 0.03553291782736778, + 0.14448955655097961, + 0.06690789014101028, + 0.22643618285655975, + 0.06735076010227203 + ], + "credit_map": [ + 0.029826553538441658, + 0.02266913466155529, + 0.024854347109794617, + 0.03324120491743088, + 0.015902599319815636, + 0.015277832746505737, + 0.01990228146314621, + 0.0257568396627903, + 0.011071980930864811, + 0.015512633137404919, + 0.007565744686871767, + 0.009397878311574459, + 0.01086752861738205, + 0.005254031158983707, + 0.008121269755065441, + 0.00527087040245533, + 0.012923698872327805, + 0.012646816670894623, + 0.021888907998800278, + 0.01728210598230362, + 0.011330521665513515, + 0.010111652314662933, + 0.014842403121292591, + 0.010628587566316128, + 0.00637768255546689, + 0.004180168733000755, + 0.006599017884582281, + 0.006387063767760992, + 0.004473517183214426, + 0.005290639121085405, + 0.0056977891363203526, + 0.009024831466376781, + 0.0047980668023228645, + 0.0038736648857593536, + 0.004642437677830458, + 0.006320265121757984, + 0.003953432198613882, + 0.004458798095583916, + 0.004475015681236982, + 0.006582887843251228, + 0.004793181549757719, + 0.005308727268129587, + 0.005519692320376635, + 0.0033741674851626158, + 0.00457301689311862, + 0.004603395238518715, + 0.003007066436111927, + 0.004353671334683895, + 0.01735462062060833, + 0.016329249367117882, + 0.021619677543640137, + 0.020282646641135216, + 0.01810268498957157, + 0.012661999091506004, + 0.02043449878692627, + 0.017704185098409653, + 0.006765275727957487, + 0.008254622109234333, + 0.007176574319601059, + 0.010660117492079735, + 0.007984695956110954, + 0.00589958718046546, + 0.010256120003759861, + 0.009910897351801395, + 0.04145388677716255, + 0.01763088069856167, + 0.016994241625070572, + 0.018279068171977997, + 0.047478530555963516, + 0.04145845025777817, + 0.0157198254019022, + 0.027421317994594574, + 0.009609687142074108, + 0.008067292161285877, + 0.006097170989960432, + 0.010643529705703259, + 0.0080754105001688, + 0.007614476606249809, + 0.00945693626999855, + 0.0077862548641860485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0004379982128739357 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle close-up, almost completely immersed in the scalding hot pot, the camera trembles subtly with the bubbling liquid. Within this deep, viscous, roiling sea of crimson oil, countless dried red chilies, dark Sichuan peppercorns, and a variety of star anise and bay leaves, saturated with grease and glistening with oil, are swept up by a turbulent vortex. They tumble, collide, and tumble fiercely within the seething red oil, like a vast army engaged in an epic melee. A powerful wave of heat erupts from the pot's bottom, hurling peppers and spices into the air. In slow-motion detail, captured by high-speed photography, they streak across the camera, accompanied by splattering, flaming droplets of oil, before slamming back into the crimson purgatory. The entire picture is illuminated by a hellish fire that originates from the bottom of the pot and shines from the inside out, forming an extremely high contrast of light and shadow, locking the main color in a saturated and almost supernatural red. This is not just food, but also a visual feast full of Tsui Hark's martial arts fantasy style, full of wild and unrestrained energy and spectacular aesthetics of destruction, creating a grand and dangerous atmosphere.", + "seed": 6170, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_9880_seed_6170_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06839759135618806, + "vlm_logprob": -0.06839759135618806, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.06839759135618806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02361796610057354 + ], + "opposite_logprob_tokens": [ + -4.398617744445801 + ], + "logprob_reward": -0.02361796610057354, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02361796610057354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21288812160491943, + 0.06512755155563354, + 0.14889340102672577, + 0.1278647631406784, + 0.056844718754291534, + 0.12620708346366882, + 0.058996524661779404, + 0.0506402924656868, + 0.05645488575100899, + 0.09608263522386551 + ], + "credit_map": [ + 0.020271066576242447, + 0.050773948431015015, + 0.02203080616891384, + 0.028340665623545647, + 0.028707413002848625, + 0.021812614053487778, + 0.023114768788218498, + 0.01783684454858303, + 0.008097795769572258, + 0.0104728564620018, + 0.0067012133076786995, + 0.009411224164068699, + 0.010582677088677883, + 0.007521302904933691, + 0.006109422538429499, + 0.006231057923287153, + 0.007477033883333206, + 0.050398822873830795, + 0.004690354224294424, + 0.006800240371376276, + 0.00992609839886427, + 0.055515844374895096, + 0.007593438494950533, + 0.006491573061794043, + 0.009873669594526291, + 0.016891855746507645, + 0.011722229421138763, + 0.015334156341850758, + 0.012409206479787827, + 0.02688840590417385, + 0.013294326141476631, + 0.021450906991958618, + 0.005074642598628998, + 0.006262482609599829, + 0.006079953163862228, + 0.01321122795343399, + 0.006239729002118111, + 0.005355929955840111, + 0.008225797675549984, + 0.0063949571922421455, + 0.012594245374202728, + 0.02649652771651745, + 0.015015536919236183, + 0.015215722844004631, + 0.015165532007813454, + 0.01952826976776123, + 0.010599826462566853, + 0.011591412127017975, + 0.0046568503603339195, + 0.008452050387859344, + 0.007037849631160498, + 0.006340343039482832, + 0.005963824689388275, + 0.008242230862379074, + 0.008963819593191147, + 0.009339554235339165, + 0.004717736970633268, + 0.008809736929833889, + 0.005290383473038673, + 0.005088318604975939, + 0.00716532114893198, + 0.006137722171843052, + 0.005902489647269249, + 0.007528583984822035, + 0.006976326461881399, + 0.007548830006271601, + 0.0046714055351912975, + 0.004600800573825836, + 0.00668956246227026, + 0.013104964978992939, + 0.007416299544274807, + 0.005446691997349262, + 0.008790044113993645, + 0.017957370728254318, + 0.010442308150231838, + 0.013245980255305767, + 0.013111361302435398, + 0.01368948258459568, + 0.007338458672165871, + 0.011507629416882992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.02361796610057354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302748799324036 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302748799324036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302748799324036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6638817191123962, + 0.03164440765976906, + 0.037859492003917694, + 0.052381377667188644, + 0.015939630568027496, + 0.06941545754671097, + 0.021448925137519836, + 0.033068373799324036, + 0.034971997141838074, + 0.039388589560985565 + ], + "credit_map": [ + 0.03780006244778633, + 0.28056877851486206, + 0.009213903918862343, + 0.02217058092355728, + 0.039114248007535934, + 0.2341383695602417, + 0.01981901004910469, + 0.021056801080703735, + 0.004341166466474533, + 0.004582359921187162, + 0.0026470443699508905, + 0.0033511531073600054, + 0.0034485533833503723, + 0.004838194232434034, + 0.0035604548174887896, + 0.004875485319644213, + 0.00213066628202796, + 0.011407071724534035, + 0.0021614893339574337, + 0.0033140776213258505, + 0.009485946968197823, + 0.0035562224220484495, + 0.00293388101272285, + 0.002870141062885523, + 0.005426747724413872, + 0.007790306583046913, + 0.005782855674624443, + 0.006795758847147226, + 0.006085141561925411, + 0.008427235297858715, + 0.0049203489907085896, + 0.007152984384447336, + 0.0014502438716590405, + 0.002423268975690007, + 0.0015266966074705124, + 0.003490435192361474, + 0.001917097601108253, + 0.0012813890352845192, + 0.0020739762112498283, + 0.0017765238881111145, + 0.006162314210087061, + 0.019929897040128708, + 0.00483625428751111, + 0.008366989903151989, + 0.010443856939673424, + 0.009381712414324284, + 0.004537882748991251, + 0.005756562575697899, + 0.001686470932327211, + 0.0024323768448084593, + 0.002466105157509446, + 0.0027381516993045807, + 0.0024992269463837147, + 0.0036270548589527607, + 0.0026617860421538353, + 0.0033377548679709435, + 0.0034981390926986933, + 0.004681539721786976, + 0.0030786956194788218, + 0.003285697428509593, + 0.0035926057025790215, + 0.0032189267221838236, + 0.0069762058556079865, + 0.004736563190817833, + 0.0032388311810791492, + 0.011723448522388935, + 0.001966360956430435, + 0.0030471670906990767, + 0.004265305586159229, + 0.005641699768602848, + 0.003115642350167036, + 0.0019735435489565134, + 0.002975126961246133, + 0.008173772133886814, + 0.0026529880706220865, + 0.004656532313674688, + 0.0046452064998447895, + 0.008940298110246658, + 0.0030315364710986614, + 0.00431313319131732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.023302748799324036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253676600754261 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253676600754261, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253676600754261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7664409875869751, + 0.023389426991343498, + 0.04226136952638626, + 0.023708846420049667, + 0.012042761780321598, + 0.02236550860106945, + 0.03165657818317413, + 0.03345245495438576, + 0.03158562630414963, + 0.013096421957015991 + ], + "credit_map": [ + 0.045025020837783813, + 0.3254346549510956, + 0.009900275617837906, + 0.02671854756772518, + 0.05082585662603378, + 0.2613675892353058, + 0.021372243762016296, + 0.02579684928059578, + 0.0023164444137364626, + 0.003103676252067089, + 0.00275101768784225, + 0.0022691236808896065, + 0.0029994421638548374, + 0.003543255152180791, + 0.0029815935995429754, + 0.0034248747397214174, + 0.0018994546262547374, + 0.014084725640714169, + 0.0024109252262860537, + 0.0031755745876580477, + 0.006047066766768694, + 0.01053105853497982, + 0.002209999365732074, + 0.0019025703659281135, + 0.002408899599686265, + 0.002621748950332403, + 0.0041960459202528, + 0.0025387320201843977, + 0.0028354970272630453, + 0.003286672756075859, + 0.0024667459074407816, + 0.003354504471644759, + 0.0010129715083166957, + 0.0015648703556507826, + 0.00179901416413486, + 0.0022567834239453077, + 0.0015900284051895142, + 0.001390475663356483, + 0.001382596674375236, + 0.001046022167429328, + 0.0022975506726652384, + 0.004276142921298742, + 0.0022396999411284924, + 0.002696158131584525, + 0.0032383257057517767, + 0.0033831337932497263, + 0.0018636401509866118, + 0.002370858797803521, + 0.0023296745494008064, + 0.002820143476128578, + 0.0035205441527068615, + 0.00439797667786479, + 0.0038724299520254135, + 0.00422494113445282, + 0.006722777616232634, + 0.0037680896930396557, + 0.002478631678968668, + 0.003879793453961611, + 0.0030590216629207134, + 0.0037079539615660906, + 0.004308272153139114, + 0.004995519295334816, + 0.007621895056217909, + 0.0034013697877526283, + 0.002606005407869816, + 0.010134116746485233, + 0.002006527967751026, + 0.0028636211063712835, + 0.0034076529555022717, + 0.005662658717483282, + 0.002620202489197254, + 0.002284844871610403, + 0.001098564942367375, + 0.002108335494995117, + 0.001124561415053904, + 0.0016372757963836193, + 0.001616902882233262, + 0.0023072920739650726, + 0.0010893455473706126, + 0.002114145318046212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.01253676600754261 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813507616519928 + ], + "opposite_logprob_tokens": [ + -1.4063507318496704 + ], + "logprob_reward": -0.2813507616519928, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813507616519928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4783470630645752, + 0.03844830021262169, + 0.05833565443754196, + 0.14106297492980957, + 0.030924078077077866, + 0.08585435897111893, + 0.03353390842676163, + 0.03605569526553154, + 0.06394995003938675, + 0.033488113433122635 + ], + "credit_map": [ + 0.028540655970573425, + 0.1798652559518814, + 0.012815127149224281, + 0.027002260088920593, + 0.03989266976714134, + 0.13711227476596832, + 0.02977040596306324, + 0.023348325863480568, + 0.003078027628362179, + 0.007650949992239475, + 0.003220781683921814, + 0.005568414926528931, + 0.005256099626421928, + 0.00411318289116025, + 0.0032713855616748333, + 0.006289450451731682, + 0.003451008116826415, + 0.010680685751140118, + 0.005153315607458353, + 0.0053308019414544106, + 0.006383197382092476, + 0.015238492749631405, + 0.006178548559546471, + 0.005919591058045626, + 0.010745481587946415, + 0.016933241859078407, + 0.010058829560875893, + 0.01894720084965229, + 0.012272139079868793, + 0.03340201452374458, + 0.017963821068406105, + 0.020740214735269547, + 0.00249061593785882, + 0.002928460482507944, + 0.004125728737562895, + 0.006446294020861387, + 0.004533631261438131, + 0.003327060490846634, + 0.003489215625450015, + 0.0035830663982778788, + 0.007183167617768049, + 0.011873671784996986, + 0.010640754364430904, + 0.010513854213058949, + 0.01576727069914341, + 0.010839789174497128, + 0.011253123171627522, + 0.007782703265547752, + 0.0026759819593280554, + 0.0036043054424226284, + 0.005209497641772032, + 0.004205410368740559, + 0.003378190565854311, + 0.004804456606507301, + 0.004540460649877787, + 0.00511559983715415, + 0.004090366885066032, + 0.006796482019126415, + 0.0032456754706799984, + 0.004004208836704493, + 0.003745183115825057, + 0.004063062835484743, + 0.005975825246423483, + 0.004134885501116514, + 0.007681414484977722, + 0.008680540136992931, + 0.006280372850596905, + 0.006223804783076048, + 0.008865595795214176, + 0.012949857860803604, + 0.005855084862560034, + 0.0074132708832621574, + 0.0024455594830214977, + 0.0054094730876386166, + 0.002377129392698407, + 0.0034329050686210394, + 0.0043497332371771336, + 0.007993226870894432, + 0.003180846106261015, + 0.0042992327362298965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.2813507616519928 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011797142215073109 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011797142215073109, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011797142215073109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4980327785015106, + 0.05551724135875702, + 0.10203520953655243, + 0.1620139479637146, + 0.016878556460142136, + 0.03199363872408867, + 0.026611290872097015, + 0.027788938954472542, + 0.051944948732852936, + 0.02718345820903778 + ], + "credit_map": [ + 0.032456010580062866, + 0.2088482826948166, + 0.009312870912253857, + 0.0220724456012249, + 0.03544647991657257, + 0.1555245965719223, + 0.015391681343317032, + 0.018980348482728004, + 0.006156364921480417, + 0.009415674023330212, + 0.00528835179284215, + 0.006240147165954113, + 0.008483219891786575, + 0.007486111484467983, + 0.006433541886508465, + 0.006013828795403242, + 0.004723209422081709, + 0.032429032027721405, + 0.003389588091522455, + 0.004645504057407379, + 0.005932589992880821, + 0.04250907897949219, + 0.00436254171654582, + 0.004043667111545801, + 0.01394166424870491, + 0.02262653037905693, + 0.012794521637260914, + 0.01634828932583332, + 0.01994481310248375, + 0.030712680891156197, + 0.01651736907660961, + 0.02912808023393154, + 0.0016206341097131371, + 0.0017700537573546171, + 0.002645880915224552, + 0.0036155125126242638, + 0.0021476077381521463, + 0.0015333942137658596, + 0.001676852465607226, + 0.0018686206312850118, + 0.0022117190528661013, + 0.004490816034376621, + 0.004948100075125694, + 0.003528767032548785, + 0.006183421239256859, + 0.0033022193238139153, + 0.00412625540047884, + 0.0032023387029767036, + 0.0019509204430505633, + 0.0027211422566324472, + 0.0024196752347052097, + 0.003501990344375372, + 0.0027286559343338013, + 0.003948961850255728, + 0.006391915492713451, + 0.002948026405647397, + 0.002944271545857191, + 0.005398393142968416, + 0.0023276801221072674, + 0.00291437073610723, + 0.003442595712840557, + 0.005130297504365444, + 0.0025760105345398188, + 0.0030553198885172606, + 0.004786195699125528, + 0.009340372867882252, + 0.003652094164863229, + 0.004948719870299101, + 0.006365008186548948, + 0.011020914651453495, + 0.004369997885078192, + 0.007461640052497387, + 0.0022325944155454636, + 0.0035768840461969376, + 0.0034583706874400377, + 0.0029728878289461136, + 0.0048788911662995815, + 0.004395908676087856, + 0.0020898093935102224, + 0.0035781085025519133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0011797142215073109 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle close-up, almost completely immersed in the scalding hot pot, the camera trembles subtly with the bubbling liquid. Within this deep, viscous, roiling sea of crimson oil, countless dried red chilies, dark Sichuan peppercorns, and a variety of star anise and bay leaves, saturated with grease and glistening with oil, are swept up by a turbulent vortex. They tumble, collide, and tumble fiercely within the seething red oil, like a vast army engaged in an epic melee. A powerful wave of heat erupts from the pot's bottom, hurling peppers and spices into the air. In slow-motion detail, captured by high-speed photography, they streak across the camera, accompanied by splattering, flaming droplets of oil, before slamming back into the crimson purgatory. The entire picture is illuminated by a hellish fire that originates from the bottom of the pot and shines from the inside out, forming an extremely high contrast of light and shadow, locking the main color in a saturated and almost supernatural red. This is not just food, but also a visual feast full of Tsui Hark's martial arts fantasy style, full of wild and unrestrained energy and spectacular aesthetics of destruction, creating a grand and dangerous atmosphere.", + "seed": 6171, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_2/index_9880_seed_6171_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04356171577237546, + "vlm_logprob": -0.04356171577237546, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.04356171577237546, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02118506468832493 + ], + "opposite_logprob_tokens": [ + -4.396184921264648 + ], + "logprob_reward": -0.02118506468832493, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02118506468832493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23077626526355743, + 0.47625088691711426, + 0.04929086193442345, + 0.02484753541648388, + 0.051015790551900864, + 0.032048746943473816, + 0.035823237150907516, + 0.016525866463780403, + 0.05855967104434967, + 0.024861132726073265 + ], + "credit_map": [ + 0.023616423830389977, + 0.048538435250520706, + 0.01778900995850563, + 0.021331245079636574, + 0.02952982857823372, + 0.02215534634888172, + 0.026140574365854263, + 0.041675399988889694, + 0.1557302325963974, + 0.029427388682961464, + 0.011336578987538815, + 0.013316184282302856, + 0.2070094496011734, + 0.034283388406038284, + 0.014257881790399551, + 0.010889767669141293, + 0.00867310632020235, + 0.005029730498790741, + 0.005698277149349451, + 0.0037706538569182158, + 0.01034511998295784, + 0.005680885631591082, + 0.004797997418791056, + 0.005295088980346918, + 0.0032270215451717377, + 0.0034346883185207844, + 0.0026406918186694384, + 0.0029612607322633266, + 0.00327373412437737, + 0.004154002759605646, + 0.0025755122769623995, + 0.002580623608082533, + 0.004753031302243471, + 0.005599577445536852, + 0.005887992214411497, + 0.005197235848754644, + 0.006031561177223921, + 0.00954430177807808, + 0.006604946684092283, + 0.007397147361189127, + 0.0022994845639914274, + 0.0047895824536681175, + 0.002125560073181987, + 0.0027081193402409554, + 0.005830050446093082, + 0.003843512386083603, + 0.003893174696713686, + 0.006559263449162245, + 0.0038590128533542156, + 0.006624811328947544, + 0.0034836868289858103, + 0.004524476360529661, + 0.006973255425691605, + 0.00331108714453876, + 0.003877597162500024, + 0.0031693060882389545, + 0.0016343811294063926, + 0.0019297153921797872, + 0.0019676019437611103, + 0.003235619282349944, + 0.0016213535564020276, + 0.002213200554251671, + 0.0017006299458444118, + 0.002223363146185875, + 0.004699613433331251, + 0.015014152973890305, + 0.005683384835720062, + 0.004660789389163256, + 0.012522144243121147, + 0.006597389001399279, + 0.004048862960189581, + 0.00533333420753479, + 0.002419698517769575, + 0.003974681720137596, + 0.0027954396791756153, + 0.0029359045438468456, + 0.003481282154098153, + 0.004166636150330305, + 0.0024715012405067682, + 0.0026159880217164755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02118506468832493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255195215344429 + ], + "opposite_logprob_tokens": [ + -4.387551784515381 + ], + "logprob_reward": -0.01255195215344429, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255195215344429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21390804648399353, + 0.2812293469905853, + 0.05698084458708763, + 0.04062819853425026, + 0.09688718616962433, + 0.15296196937561035, + 0.03406187891960144, + 0.025389060378074646, + 0.057223204523324966, + 0.040730275213718414 + ], + "credit_map": [ + 0.02006462775170803, + 0.030006248503923416, + 0.03299975022673607, + 0.038739610463380814, + 0.03508343547582626, + 0.021837539970874786, + 0.01678302139043808, + 0.01839381828904152, + 0.0583144947886467, + 0.018823696300387383, + 0.025140712037682533, + 0.018936317414045334, + 0.06842267513275146, + 0.028936898335814476, + 0.025989223271608353, + 0.036665335297584534, + 0.01075788401067257, + 0.006212481763213873, + 0.0046926215291023254, + 0.003703249152749777, + 0.012508055195212364, + 0.007110653445124626, + 0.006289707496762276, + 0.0057061901316046715, + 0.004231330007314682, + 0.003139181761071086, + 0.004212802276015282, + 0.00391321862116456, + 0.005148572847247124, + 0.00700205797329545, + 0.005843923892825842, + 0.007137114182114601, + 0.009357273578643799, + 0.011947054415941238, + 0.013163507916033268, + 0.013610457070171833, + 0.009460830129683018, + 0.012116562575101852, + 0.01188060361891985, + 0.015350905247032642, + 0.011640272103250027, + 0.02170494943857193, + 0.009333878755569458, + 0.010185172781348228, + 0.03090560995042324, + 0.021050870418548584, + 0.01771578937768936, + 0.03042541816830635, + 0.003679197281599045, + 0.0052655343897640705, + 0.004092949442565441, + 0.005887109786272049, + 0.0037627830170094967, + 0.0036956905387341976, + 0.003500761929899454, + 0.0041778492741286755, + 0.0022499901242554188, + 0.004090266302227974, + 0.002958919620141387, + 0.0029679788276553154, + 0.00435726810246706, + 0.004351869225502014, + 0.0017996685346588492, + 0.0026131020858883858, + 0.00789166335016489, + 0.006749396212399006, + 0.006049592047929764, + 0.008086888119578362, + 0.009585256688296795, + 0.00945808831602335, + 0.004230620805174112, + 0.005171701777726412, + 0.00432668998837471, + 0.005007301922887564, + 0.004377974197268486, + 0.0034684089478105307, + 0.006633454468101263, + 0.007093516644090414, + 0.004846782423555851, + 0.004976145923137665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.01255195215344429 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615036141127348 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003615036141127348, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615036141127348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08287610858678818, + 0.5289824604988098, + 0.1576172113418579, + 0.023008031770586967, + 0.0195249542593956, + 0.036589279770851135, + 0.029281646013259888, + 0.05055936798453331, + 0.05113956332206726, + 0.020421354100108147 + ], + "credit_map": [ + 0.008467165753245354, + 0.015039687044918537, + 0.007927709259092808, + 0.007242536637932062, + 0.01050218753516674, + 0.010828115977346897, + 0.01021348312497139, + 0.012655231170356274, + 0.16866356134414673, + 0.0334976427257061, + 0.015060427598655224, + 0.01762816123664379, + 0.22019504010677338, + 0.037746477872133255, + 0.021881666034460068, + 0.0143095338717103, + 0.03912712261080742, + 0.017199480906128883, + 0.011100503616034985, + 0.007759601343423128, + 0.050177186727523804, + 0.013998220674693584, + 0.00966624915599823, + 0.008588847704231739, + 0.0024168037343770266, + 0.0026608591433614492, + 0.0028931121341884136, + 0.0028341240249574184, + 0.003269482171162963, + 0.0033661024644970894, + 0.002992734545841813, + 0.002574814949184656, + 0.001651445054449141, + 0.00270154420286417, + 0.0026442050002515316, + 0.0024653119035065174, + 0.0019413911504670978, + 0.0029709050431847572, + 0.0023731307592242956, + 0.002777021611109376, + 0.0031422150786966085, + 0.0033099004067480564, + 0.003239128040149808, + 0.002651518676429987, + 0.005686986725777388, + 0.007113293278962374, + 0.003846249310299754, + 0.0075999898836016655, + 0.00344822253100574, + 0.0029539517126977444, + 0.0032482161186635494, + 0.004665041342377663, + 0.003797172335907817, + 0.0035316876601427794, + 0.003311601933091879, + 0.004325754474848509, + 0.004566873423755169, + 0.005037549417465925, + 0.006105460692197084, + 0.0086624501273036, + 0.00532948225736618, + 0.01042270753532648, + 0.004764880053699017, + 0.0056699649430811405, + 0.005847658962011337, + 0.006744070909917355, + 0.0050138989463448524, + 0.008492437191307545, + 0.008857986889779568, + 0.007636607624590397, + 0.004224112723022699, + 0.004322791006416082, + 0.0021070183720439672, + 0.003681089961901307, + 0.002826568903401494, + 0.0020018259529024363, + 0.002818398643285036, + 0.003308531129732728, + 0.0018487427150830626, + 0.0018291778396815062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.003615036141127348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989592254161835 + ], + "opposite_logprob_tokens": [ + -1.8048958778381348 + ], + "logprob_reward": -0.17989592254161835, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989592254161835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09229878336191177, + 0.6526552438735962, + 0.025839058682322502, + 0.01886136829853058, + 0.02369685284793377, + 0.06157796457409859, + 0.030804526060819626, + 0.017614848911762238, + 0.06852888315916061, + 0.008122444152832031 + ], + "credit_map": [ + 0.009664208628237247, + 0.01686616614460945, + 0.008443628437817097, + 0.008425358682870865, + 0.015399825759232044, + 0.013023214414715767, + 0.010142561048269272, + 0.01033382024616003, + 0.21469426155090332, + 0.039980001747608185, + 0.016780313104391098, + 0.019816357642412186, + 0.28132864832878113, + 0.04515954107046127, + 0.021070564165711403, + 0.013825606554746628, + 0.0038611122872680426, + 0.0032220042776316404, + 0.003150361590087414, + 0.002559803193435073, + 0.0035867965780198574, + 0.003982160240411758, + 0.0027463752776384354, + 0.002730444772168994, + 0.002106515923514962, + 0.0035429359413683414, + 0.001602501841261983, + 0.0019302876899018884, + 0.0022287906613200903, + 0.0021695580799132586, + 0.0019938317127525806, + 0.0032869460992515087, + 0.0023814558517187834, + 0.002867555245757103, + 0.003318606875836849, + 0.002808194374665618, + 0.004085715860128403, + 0.0031557518523186445, + 0.0027104683686047792, + 0.0023691034875810146, + 0.005662586539983749, + 0.012149789370596409, + 0.004558504559099674, + 0.006698145531117916, + 0.010643525049090385, + 0.009326539933681488, + 0.0057872263714671135, + 0.006751647684723139, + 0.003295540576800704, + 0.00556675111874938, + 0.002880391199141741, + 0.0034742620773613453, + 0.00415711710229516, + 0.004036782309412956, + 0.00357803818769753, + 0.0038156413938850164, + 0.001539301360026002, + 0.0036008977331221104, + 0.0014773770235478878, + 0.0018382011912763119, + 0.003761701751500368, + 0.002335969591513276, + 0.0013815894490107894, + 0.001679811510257423, + 0.00566872488707304, + 0.013245631940662861, + 0.007522400468587875, + 0.007667671423405409, + 0.011217224411666393, + 0.008697360754013062, + 0.006745422258973122, + 0.007764449343085289, + 0.0009277045028284192, + 0.0011602015001699328, + 0.0009841077262535691, + 0.0007698263507336378, + 0.0011277723824605346, + 0.0013071809662505984, + 0.0010119650978595018, + 0.0008336855680681765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.17989592254161835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a close-up view of a pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dried red chilies, dark Sichuan peppercorns, star anise, and bay leaves visible in the oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the ingredients tumble, collide, and swirl in a turbulent vortex within the roiling oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast light source originating from the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dominant color of the scene a saturated, supernatural red?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005606033373624086 + ], + "opposite_logprob_tokens": [ + -7.500560760498047 + ], + "logprob_reward": -0.0005606033373624086, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005606033373624086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09271393716335297, + 0.7525242567062378, + 0.02172514609992504, + 0.015272801741957664, + 0.035811152309179306, + 0.017313197255134583, + 0.016749540343880653, + 0.009450933896005154, + 0.02694201096892357, + 0.011496958322823048 + ], + "credit_map": [ + 0.009008397348225117, + 0.020270559936761856, + 0.007436336018145084, + 0.007721084635704756, + 0.014181135222315788, + 0.007998643442988396, + 0.01030577439814806, + 0.015792004764080048, + 0.24784241616725922, + 0.04697294160723686, + 0.018009761348366737, + 0.02147103287279606, + 0.328247994184494, + 0.0528595857322216, + 0.023511216044425964, + 0.0136093245819211, + 0.0027799822855740786, + 0.003353159874677658, + 0.002263061935082078, + 0.0019991120789200068, + 0.003980507608503103, + 0.00330149638466537, + 0.00203344807960093, + 0.00201437808573246, + 0.0018739434890449047, + 0.002087314845994115, + 0.001428012503311038, + 0.0016622734256088734, + 0.0018668122356757522, + 0.0021315247286111116, + 0.0019819699227809906, + 0.002240950707346201, + 0.004194419831037521, + 0.004305435810238123, + 0.0033281580545008183, + 0.0034554183948785067, + 0.005215520970523357, + 0.006129568442702293, + 0.004345573019236326, + 0.004837059415876865, + 0.0012560023460537195, + 0.00396202877163887, + 0.0015909195644780993, + 0.00132966588716954, + 0.0038062145467847586, + 0.0017367146210744977, + 0.0014729336835443974, + 0.002158718416467309, + 0.0018964798655360937, + 0.002049860544502735, + 0.001981954788789153, + 0.0021639850456267595, + 0.0026364733930677176, + 0.00166232674382627, + 0.0023919404484331608, + 0.0019665195140987635, + 0.0007771359523758292, + 0.001292638829909265, + 0.001015261048451066, + 0.0010309062199667096, + 0.001729137497022748, + 0.0017712131375446916, + 0.000808262440841645, + 0.0010263783624395728, + 0.0032580799888819456, + 0.003147584619000554, + 0.0035082416143268347, + 0.0030178329907357693, + 0.003986114636063576, + 0.004736642353236675, + 0.0026806630194187164, + 0.0026068510487675667, + 0.0013222318375483155, + 0.001519783167168498, + 0.0012135823490098119, + 0.0015278497012332082, + 0.0015767726581543684, + 0.0016345304902642965, + 0.0016188014997169375, + 0.0010834065033122897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0005606033373624086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_tensor([8394, 8394])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_tensor([8394, 8394])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2655a1a5349005b611cc81a21be9bfafa78eb775 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_tensor([8394, 8394])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, low-angle shot panning rightward, captured in Steve McCurry's signature documentary style, captures a scene on a wet city sidewalk late at night. In the frame, a highly saturated neon light casts a shadow from the entrance of an unseen bar. Deep reds and electric blues intertwine, creating psychedelic patches of light across the damp asphalt. At the center of the image is a ginger tabby cat, one ear slightly chipped, its fur slightly matted but warm in the multicolored light. Huddled in the shadows near the bar entrance, its body forming an elegant arc, it raises one front paw with focused and deliberate concentration, meticulously and methodically licking the dust from its paw pad with its pink tongue. The camera moves in incredibly slow steps, following the rhythm of its licking, capturing in close-up the way water droplets from its tongue touch its paw and the soft golden edges of its body outlined by the neon light. The entire scene is shrouded in a quiet and profound atmosphere. The strong color contrast forms a sharp contrast with the subject's quiet posture, creating a poetic moment full of dignity and loneliness found in a corner of a bustling city.", + "seed": 13001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_8394_seed_13001_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20252425235230476, + "vlm_logprob": -0.20252425235230476, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.20252425235230476, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7681388854980469 + ], + "opposite_logprob_tokens": [ + -0.6431388854980469 + ], + "logprob_reward": -0.7681388854980469, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7681388854980469, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17994113266468048, + 0.2245684564113617, + 0.17950940132141113, + 0.0557551272213459, + 0.07922068983316422, + 0.10268239676952362, + 0.02833147533237934, + 0.024955105036497116, + 0.10833698511123657, + 0.016699252650141716 + ], + "credit_map": [ + 0.023724334314465523, + 0.012190553359687328, + 0.028700707480311394, + 0.054679494351148605, + 0.011706206947565079, + 0.01563546061515808, + 0.01785106584429741, + 0.015453312546014786, + 0.03750878572463989, + 0.033803585916757584, + 0.052444618195295334, + 0.035508912056684494, + 0.01833854615688324, + 0.010411858558654785, + 0.023563986644148827, + 0.012988150119781494, + 0.014486648142337799, + 0.020191799849271774, + 0.04864298552274704, + 0.06316765397787094, + 0.009200675413012505, + 0.004451454617083073, + 0.012147746048867702, + 0.007220444269478321, + 0.0069665065966546535, + 0.006338243838399649, + 0.0074438173323869705, + 0.017302246764302254, + 0.004323121160268784, + 0.002397239673882723, + 0.005827070213854313, + 0.005156878847628832, + 0.013254174031317234, + 0.005879850592464209, + 0.006456681992858648, + 0.026331886649131775, + 0.005548740271478891, + 0.004972704220563173, + 0.008653885684907436, + 0.008122763596475124, + 0.012513881549239159, + 0.01079228799790144, + 0.01048447284847498, + 0.02301107719540596, + 0.0061996025033295155, + 0.010005001910030842, + 0.009820805862545967, + 0.019855262711644173, + 0.005505461245775223, + 0.003923707641661167, + 0.00357445958070457, + 0.005382627714425325, + 0.0017746068770065904, + 0.0018952653044834733, + 0.002743392251431942, + 0.0035319530870765448, + 0.004654623568058014, + 0.002533780410885811, + 0.0021655610762536526, + 0.0036303738597780466, + 0.0015375595539808273, + 0.0017918044468387961, + 0.003656175220385194, + 0.004985228646546602, + 0.010933337733149529, + 0.011766881681978703, + 0.01047718059271574, + 0.027768999338150024, + 0.0074561540968716145, + 0.005327665712684393, + 0.01873128116130829, + 0.01587548851966858, + 0.001353948493488133, + 0.0023563068825751543, + 0.0017983525758609176, + 0.004799657501280308, + 0.0011784438975155354, + 0.0011997746769338846, + 0.001474405755288899, + 0.0025383627507835627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.7681388854980469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600875839591026 + ], + "opposite_logprob_tokens": [ + -3.8956007957458496 + ], + "logprob_reward": -0.020600875839591026, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600875839591026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1703743189573288, + 0.11270742863416672, + 0.1437731832265854, + 0.09915321320295334, + 0.062069278210401535, + 0.09524238109588623, + 0.21203412115573883, + 0.04905611276626587, + 0.02754800021648407, + 0.02804199978709221 + ], + "credit_map": [ + 0.014611538499593735, + 0.013635280542075634, + 0.0415455587208271, + 0.05689917132258415, + 0.010874113067984581, + 0.007798695471137762, + 0.014110276475548744, + 0.010899687185883522, + 0.016266901046037674, + 0.018144622445106506, + 0.016153723001480103, + 0.01600520871579647, + 0.012902720831334591, + 0.005879129748791456, + 0.018065178766846657, + 0.0092899389564991, + 0.04447677358984947, + 0.00942350085824728, + 0.05119698867201805, + 0.014287726953625679, + 0.004694505129009485, + 0.0028782745357602835, + 0.009056830778717995, + 0.007758582942187786, + 0.018577706068754196, + 0.012359129264950752, + 0.010554366745054722, + 0.02911446988582611, + 0.007127425633370876, + 0.002960270969197154, + 0.01244364120066166, + 0.006016201339662075, + 0.008973539806902409, + 0.008450004272162914, + 0.007497506216168404, + 0.009801414795219898, + 0.004703923128545284, + 0.007021468132734299, + 0.00778229208663106, + 0.007839127443730831, + 0.01232131477445364, + 0.012697819620370865, + 0.01240343414247036, + 0.017782043665647507, + 0.007939302362501621, + 0.006877443287521601, + 0.012077675200998783, + 0.013143347576260567, + 0.0385805107653141, + 0.028087656944990158, + 0.025214314460754395, + 0.06020144745707512, + 0.011252844706177711, + 0.00792941264808178, + 0.023847442120313644, + 0.016920488327741623, + 0.008272193372249603, + 0.007488307077437639, + 0.005580398719757795, + 0.004373886156827211, + 0.0034377295523881912, + 0.004954581148922443, + 0.007246893364936113, + 0.007702123373746872, + 0.003230620874091983, + 0.004495501518249512, + 0.0029023061506450176, + 0.003473449032753706, + 0.002712666755542159, + 0.0016814690316095948, + 0.006190599408000708, + 0.0028613887261599302, + 0.0034130141139030457, + 0.003518505720421672, + 0.003256636206060648, + 0.006601032335311174, + 0.0017178112175315619, + 0.0027488693594932556, + 0.0033017799723893404, + 0.0034843527246266603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.020600875839591026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307287558913231 + ], + "opposite_logprob_tokens": [ + -3.1680729389190674 + ], + "logprob_reward": -0.04307287558913231, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307287558913231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04518258944153786, + 0.049847185611724854, + 0.09531503170728683, + 0.2476092278957367, + 0.06387219578027725, + 0.1327137053012848, + 0.16222430765628815, + 0.049916140735149384, + 0.019966095685958862, + 0.13335348665714264 + ], + "credit_map": [ + 0.006571202073246241, + 0.00492924964055419, + 0.008094373159110546, + 0.0066983443684875965, + 0.0035623887088149786, + 0.005637642461806536, + 0.004561546724289656, + 0.005127841606736183, + 0.005679606460034847, + 0.006956130266189575, + 0.007062049116939306, + 0.006262858863919973, + 0.0053217713721096516, + 0.0025944265071302652, + 0.012005206197500229, + 0.003965139389038086, + 0.022635584697127342, + 0.008268514648079872, + 0.032865360379219055, + 0.016299063339829445, + 0.003715048311278224, + 0.0023095300421118736, + 0.005168185103684664, + 0.00405373889952898, + 0.04089874401688576, + 0.02948310412466526, + 0.026674551889300346, + 0.07885834574699402, + 0.016498111188411713, + 0.009815518744289875, + 0.03171348571777344, + 0.01366735901683569, + 0.007879509590566158, + 0.0062756952829658985, + 0.00798859540373087, + 0.012400813400745392, + 0.004520648159086704, + 0.0056181494146585464, + 0.009762111119925976, + 0.009426672011613846, + 0.01848873868584633, + 0.023518700152635574, + 0.014217039570212364, + 0.012170166708528996, + 0.010609767399728298, + 0.00576047832146287, + 0.017325876280665398, + 0.03062293492257595, + 0.029151644557714462, + 0.023132184520363808, + 0.019724631682038307, + 0.04383842274546623, + 0.008022827096283436, + 0.00745404651388526, + 0.018038280308246613, + 0.012862270697951317, + 0.007839301601052284, + 0.006228778976947069, + 0.004911282565444708, + 0.005277037154883146, + 0.004163168836385012, + 0.004709009081125259, + 0.00813016016036272, + 0.008657403290271759, + 0.0020293162669986486, + 0.0036540019791573286, + 0.002389784436672926, + 0.0020653281826525927, + 0.0022953117731958628, + 0.0013357929419726133, + 0.003551879432052374, + 0.002644683001562953, + 0.009020893834531307, + 0.009980917908251286, + 0.015737246721982956, + 0.05418986454606056, + 0.00425252690911293, + 0.0027470202185213566, + 0.01011846773326397, + 0.027306560426950455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.04307287558913231 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988577485084534 + ], + "opposite_logprob_tokens": [ + -1.804885745048523 + ], + "logprob_reward": -0.17988577485084534, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988577485084534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05168897658586502, + 0.06942225992679596, + 0.21031977236270905, + 0.0642908439040184, + 0.08244213461875916, + 0.06871684640645981, + 0.05060749128460884, + 0.040867678821086884, + 0.024844789877533913, + 0.336799293756485 + ], + "credit_map": [ + 0.006931553594768047, + 0.006403185892850161, + 0.006904215551912785, + 0.008424063213169575, + 0.005045493599027395, + 0.006178617477416992, + 0.00639198487624526, + 0.005409860983490944, + 0.008819976821541786, + 0.010132143273949623, + 0.009427006356418133, + 0.010263231582939625, + 0.009883575141429901, + 0.005464809946715832, + 0.010241350159049034, + 0.0051901675760746, + 0.06925174593925476, + 0.014682523906230927, + 0.07489362359046936, + 0.01933770813047886, + 0.007014119997620583, + 0.0037833256646990776, + 0.011295966804027557, + 0.010060741566121578, + 0.01041815709322691, + 0.006497323513031006, + 0.007339021656662226, + 0.020942607894539833, + 0.004973178263753653, + 0.003141580382362008, + 0.007941885851323605, + 0.0030370871536433697, + 0.008102841675281525, + 0.009340524673461914, + 0.012816580012440681, + 0.0196188036352396, + 0.010560790076851845, + 0.004807854071259499, + 0.010448774322867393, + 0.006745954509824514, + 0.010732538066804409, + 0.009607240557670593, + 0.00819226261228323, + 0.012723066844046116, + 0.005999124143272638, + 0.005120885092765093, + 0.007025345228612423, + 0.009316377341747284, + 0.004898445680737495, + 0.009125692769885063, + 0.005299929063767195, + 0.00926693994551897, + 0.0042184581980109215, + 0.004831596277654171, + 0.008109119720757008, + 0.004857310559600592, + 0.004388159140944481, + 0.0053418478928506374, + 0.004117803182452917, + 0.005663316231220961, + 0.002961444202810526, + 0.003929645288735628, + 0.005567941814661026, + 0.008897517807781696, + 0.0037013806868344545, + 0.0036895284429192543, + 0.0031977465841919184, + 0.0032407392282038927, + 0.002161140087991953, + 0.0020094825886189938, + 0.0035897206980735064, + 0.0032550510950386524, + 0.023910952731966972, + 0.025129323825240135, + 0.03954242169857025, + 0.1359960436820984, + 0.010074860416352749, + 0.008032748475670815, + 0.025698738172650337, + 0.06841416656970978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.17988577485084534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000922849983908236 + ], + "opposite_logprob_tokens": [ + -7.000922679901123 + ], + "logprob_reward": -0.000922849983908236, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000922849983908236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12858818471431732, + 0.22662405669689178, + 0.18426929414272308, + 0.08955679833889008, + 0.062420669943094254, + 0.13436806201934814, + 0.041222259402275085, + 0.0375727042555809, + 0.06538717448711395, + 0.02999076247215271 + ], + "credit_map": [ + 0.01835015043616295, + 0.009866018779575825, + 0.02315381169319153, + 0.020444832742214203, + 0.01291689369827509, + 0.015325953252613544, + 0.013291031122207642, + 0.015239491127431393, + 0.02829929068684578, + 0.038832567632198334, + 0.02408229000866413, + 0.031469136476516724, + 0.033951036632061005, + 0.014987783506512642, + 0.03466634824872017, + 0.0203356072306633, + 0.017424732446670532, + 0.020686835050582886, + 0.05651795491576195, + 0.057051341980695724, + 0.009155144914984703, + 0.004540309775620699, + 0.011263118125498295, + 0.007629866246134043, + 0.0137496842071414, + 0.01681278832256794, + 0.008909408934414387, + 0.01817859709262848, + 0.008815987035632133, + 0.0035492018796503544, + 0.012588192708790302, + 0.006952932104468346, + 0.0065542422235012054, + 0.00945977121591568, + 0.007725163362920284, + 0.008396334014832973, + 0.006631768774241209, + 0.007089914754033089, + 0.00970277190208435, + 0.006860704161226749, + 0.016135787591338158, + 0.013392281718552113, + 0.01613333262503147, + 0.03050176613032818, + 0.008154836483299732, + 0.007226367946714163, + 0.014750919304788113, + 0.028072770684957504, + 0.0066057536751031876, + 0.007756667211651802, + 0.00469753285869956, + 0.006026451010257006, + 0.003076961264014244, + 0.0021435613743960857, + 0.006585580296814442, + 0.004329753573983908, + 0.004551459103822708, + 0.004840213805437088, + 0.005018899217247963, + 0.00535620329901576, + 0.0031071919947862625, + 0.002278669970110059, + 0.00816370826214552, + 0.0042563616298139095, + 0.007579419296234846, + 0.012456045486032963, + 0.006731835659593344, + 0.013238475657999516, + 0.004429263528436422, + 0.004119643941521645, + 0.009617727249860764, + 0.007214768324047327, + 0.002887127688154578, + 0.004570895340293646, + 0.0035593726206570864, + 0.005423692055046558, + 0.003700550878420472, + 0.0021683622617274523, + 0.004062309395521879, + 0.003618454560637474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.000922849983908236 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle shot panning rightward, captured in Steve McCurry's signature documentary style, captures a scene on a wet city sidewalk late at night. In the frame, a highly saturated neon light casts a shadow from the entrance of an unseen bar. Deep reds and electric blues intertwine, creating psychedelic patches of light across the damp asphalt. At the center of the image is a ginger tabby cat, one ear slightly chipped, its fur slightly matted but warm in the multicolored light. Huddled in the shadows near the bar entrance, its body forming an elegant arc, it raises one front paw with focused and deliberate concentration, meticulously and methodically licking the dust from its paw pad with its pink tongue. The camera moves in incredibly slow steps, following the rhythm of its licking, capturing in close-up the way water droplets from its tongue touch its paw and the soft golden edges of its body outlined by the neon light. The entire scene is shrouded in a quiet and profound atmosphere. The strong color contrast forms a sharp contrast with the subject's quiet posture, creating a poetic moment full of dignity and loneliness found in a corner of a bustling city.", + "seed": 13002, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_8394_seed_13002_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19315035056788474, + "vlm_logprob": -0.19315035056788474, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.19315035056788474, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7070873975753784 + ], + "opposite_logprob_tokens": [ + -0.7070873975753784 + ], + "logprob_reward": -0.7070873975753784, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7070873975753784, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15079884231090546, + 0.34854674339294434, + 0.1166345551609993, + 0.10299663245677948, + 0.036615870893001556, + 0.04233855381608009, + 0.02290087565779686, + 0.04597807675600052, + 0.10474042594432831, + 0.028449373319745064 + ], + "credit_map": [ + 0.010130457580089569, + 0.011048699729144573, + 0.028598766773939133, + 0.02331358939409256, + 0.01786084845662117, + 0.01448855921626091, + 0.029900304973125458, + 0.015457632020115852, + 0.015115533955395222, + 0.010749978944659233, + 0.07831794023513794, + 0.01911206915974617, + 0.028983628377318382, + 0.018212182447314262, + 0.13275696337223053, + 0.04529845342040062, + 0.01013741735368967, + 0.015724949538707733, + 0.01596374623477459, + 0.02035156637430191, + 0.01337895542383194, + 0.012104433961212635, + 0.016690779477357864, + 0.012282717041671276, + 0.008002030663192272, + 0.007922579534351826, + 0.023549772799015045, + 0.021276935935020447, + 0.00908996257930994, + 0.009313251823186874, + 0.009947077371180058, + 0.013895024545490742, + 0.0017400241922587156, + 0.003445591777563095, + 0.0037883420009166002, + 0.006462683901190758, + 0.00487139169126749, + 0.004610201343894005, + 0.006739708594977856, + 0.004957931116223335, + 0.0020525669679045677, + 0.0032773814164102077, + 0.004588018171489239, + 0.00881312694400549, + 0.004833058454096317, + 0.005654890555888414, + 0.007784467656165361, + 0.0053350478410720825, + 0.0018057017587125301, + 0.0030352207832038403, + 0.00251724268309772, + 0.004832178354263306, + 0.0025176650378853083, + 0.003019674215465784, + 0.0032267835922539234, + 0.0019464095821604133, + 0.004434532951563597, + 0.004773259162902832, + 0.004823845345526934, + 0.0058808717876672745, + 0.005229712929576635, + 0.00810468103736639, + 0.004971104208379984, + 0.007760072126984596, + 0.006226786877959967, + 0.006837939377874136, + 0.008681021630764008, + 0.009407466277480125, + 0.014222980476915836, + 0.02519921399652958, + 0.010412607342004776, + 0.023752423003315926, + 0.0016952495789155364, + 0.001882028067484498, + 0.004319267813116312, + 0.0036728247068822384, + 0.004477297887206078, + 0.00474844453856349, + 0.003780641360208392, + 0.0038736211135983467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.7070873975753784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369184255599976 + ], + "opposite_logprob_tokens": [ + -3.651369094848633 + ], + "logprob_reward": -0.026369184255599976, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369184255599976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07197496294975281, + 0.5733175873756409, + 0.09100339561700821, + 0.05527748912572861, + 0.06184104084968567, + 0.04809432476758957, + 0.020676080137491226, + 0.02450212836265564, + 0.020962845534086227, + 0.03235011175274849 + ], + "credit_map": [ + 0.004962774459272623, + 0.00566217303276062, + 0.01009354554116726, + 0.010906747542321682, + 0.00844576209783554, + 0.007181832101196051, + 0.013150379061698914, + 0.011571756564080715, + 0.024711396545171738, + 0.014783850871026516, + 0.1425483673810959, + 0.02849769778549671, + 0.042776141315698624, + 0.025016887113451958, + 0.22322838008403778, + 0.07175488770008087, + 0.005610543768852949, + 0.005984442308545113, + 0.008242911659181118, + 0.01215999573469162, + 0.010540589690208435, + 0.0139211630448699, + 0.022157659754157066, + 0.012386096641421318, + 0.0026304014027118683, + 0.0033523167949169874, + 0.008197277784347534, + 0.01279402244836092, + 0.005862400401383638, + 0.00706807104870677, + 0.007270116824656725, + 0.00810289103537798, + 0.0028436118736863136, + 0.006517133675515652, + 0.005833539646118879, + 0.011793571524322033, + 0.006736007984727621, + 0.009274096228182316, + 0.007920661009848118, + 0.010922422632575035, + 0.0028525034431368113, + 0.006208584178239107, + 0.0050039333291351795, + 0.008224382996559143, + 0.006924954708665609, + 0.006074388045817614, + 0.006572498474270105, + 0.00623308215290308, + 0.0014534114161506295, + 0.0017001396045088768, + 0.0024902771692723036, + 0.0031118530314415693, + 0.002631996525451541, + 0.0029698426369577646, + 0.0028515909798443317, + 0.003466969123110175, + 0.0023521410766988993, + 0.0026822895742952824, + 0.002626720815896988, + 0.00378957437351346, + 0.0026013271417468786, + 0.0037567769177258015, + 0.0030625653453171253, + 0.0036307345144450665, + 0.001641619484871626, + 0.0014800309436395764, + 0.0027974715922027826, + 0.0023944431450217962, + 0.004161560907959938, + 0.003031787695363164, + 0.0030715104658156633, + 0.0023844221141189337, + 0.00248932046815753, + 0.002362801693379879, + 0.003361394628882408, + 0.003931255079805851, + 0.0035793378483504057, + 0.007825089618563652, + 0.003049784107133746, + 0.0057511297054588795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.026369184255599976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808619990944862 + ], + "opposite_logprob_tokens": [ + -3.529808521270752 + ], + "logprob_reward": -0.029808619990944862, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808619990944862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0634659081697464, + 0.7378758788108826, + 0.025322698056697845, + 0.03124723955988884, + 0.02090281806886196, + 0.033230774104595184, + 0.014239134266972542, + 0.012302979826927185, + 0.02865706942975521, + 0.03275550901889801 + ], + "credit_map": [ + 0.002774012042209506, + 0.003553090151399374, + 0.014011770486831665, + 0.007107218727469444, + 0.007221238221973181, + 0.006713610142469406, + 0.0128340944647789, + 0.009250873699784279, + 0.030146168544888496, + 0.019247446209192276, + 0.16765044629573822, + 0.03470464050769806, + 0.05317799001932144, + 0.031501367688179016, + 0.2984631359577179, + 0.10298462957143784, + 0.0018652789294719696, + 0.0017314460128545761, + 0.003107154741883278, + 0.003867755876854062, + 0.003316815011203289, + 0.0030540365260094404, + 0.0052731772884726524, + 0.003107032971456647, + 0.0013689654879271984, + 0.001881563337519765, + 0.006295148283243179, + 0.0037644128315150738, + 0.004041773267090321, + 0.0042135524563491344, + 0.005326761864125729, + 0.004355057142674923, + 0.0013463530922308564, + 0.002009747549891472, + 0.0028126004617661238, + 0.0037675516214221716, + 0.001749662565998733, + 0.0028066998347640038, + 0.002572886412963271, + 0.0038373160641640425, + 0.0018404924776405096, + 0.0036002625711262226, + 0.004296994302421808, + 0.005843462888151407, + 0.004644135944545269, + 0.004197938367724419, + 0.0042101494036614895, + 0.004597333259880543, + 0.0010317573323845863, + 0.0014304075157269835, + 0.0016484949737787247, + 0.0020593185909092426, + 0.001549663720652461, + 0.0029800673946738243, + 0.0012810442131012678, + 0.0022583792451769114, + 0.0009906126651912928, + 0.0013030838454142213, + 0.001276505645364523, + 0.002237529493868351, + 0.0020003512036055326, + 0.001352804247289896, + 0.0017167902551591396, + 0.0014253006083890796, + 0.0017823234666138887, + 0.0016186339780688286, + 0.002066216431558132, + 0.002629664959385991, + 0.002926086075603962, + 0.007456119637936354, + 0.002908055204898119, + 0.007269968278706074, + 0.0024229036644101143, + 0.0025641126558184624, + 0.003026072634384036, + 0.0048720957711339, + 0.004836909472942352, + 0.005843284539878368, + 0.004561458248645067, + 0.004628666676580906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.029808619990944862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156370103359222 + ], + "opposite_logprob_tokens": [ + -1.7015637159347534 + ], + "logprob_reward": -0.20156370103359222, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156370103359222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05793875828385353, + 0.06401325762271881, + 0.07008910179138184, + 0.09536989778280258, + 0.06013917177915573, + 0.03499937430024147, + 0.08516313135623932, + 0.0452423095703125, + 0.16189667582511902, + 0.3251482844352722 + ], + "credit_map": [ + 0.0037965443916618824, + 0.0036166070494800806, + 0.009861469268798828, + 0.007733484730124474, + 0.009060843847692013, + 0.005768598057329655, + 0.010983243584632874, + 0.00711797084659338, + 0.0051277861930429935, + 0.006013332866132259, + 0.00809507817029953, + 0.008507158607244492, + 0.010890217497944832, + 0.007901579141616821, + 0.011212839744985104, + 0.006265268661081791, + 0.006447300314903259, + 0.005398232489824295, + 0.010480545461177826, + 0.010598724707961082, + 0.008961771614849567, + 0.008868776261806488, + 0.010366199538111687, + 0.008967556990683079, + 0.006157799623906612, + 0.005615591071546078, + 0.024529289454221725, + 0.018622836098074913, + 0.009915510192513466, + 0.009251999668776989, + 0.009196051396429539, + 0.012080824002623558, + 0.00469370000064373, + 0.0061977216973900795, + 0.007891295477747917, + 0.010087640956044197, + 0.004351065959781408, + 0.0069983904249966145, + 0.00617581233382225, + 0.013743549585342407, + 0.0023449459113180637, + 0.0034781419672071934, + 0.0045813312754035, + 0.004864768590778112, + 0.004006455652415752, + 0.0060410043224692345, + 0.004286953713744879, + 0.005395775660872459, + 0.0072554233483970165, + 0.009021095931529999, + 0.013105092570185661, + 0.014507715590298176, + 0.010116576217114925, + 0.011901465244591236, + 0.0094297481700778, + 0.009826025925576687, + 0.0022926880046725273, + 0.0021833167411386967, + 0.0042430320754647255, + 0.0047181895934045315, + 0.003610753919929266, + 0.013062640093266964, + 0.004324123729020357, + 0.010807567276060581, + 0.012740518897771835, + 0.01109306700527668, + 0.01875999942421913, + 0.028201140463352203, + 0.02054017223417759, + 0.02807854674756527, + 0.016369355842471123, + 0.02611389383673668, + 0.03198105841875076, + 0.023558348417282104, + 0.03416512534022331, + 0.0488455630838871, + 0.035407423973083496, + 0.06642527133226395, + 0.0332835428416729, + 0.05148197337985039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.20156370103359222 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000922849983908236 + ], + "opposite_logprob_tokens": [ + -7.000922679901123 + ], + "logprob_reward": -0.000922849983908236, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000922849983908236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1246267706155777, + 0.1098778173327446, + 0.08884456008672714, + 0.13380905985832214, + 0.08517768979072571, + 0.06302745640277863, + 0.03096044436097145, + 0.07093320041894913, + 0.1757958084344864, + 0.1169472336769104 + ], + "credit_map": [ + 0.009866667911410332, + 0.010407266207039356, + 0.023809310048818588, + 0.022334182634949684, + 0.01260190550237894, + 0.009862447157502174, + 0.0216121356934309, + 0.014132854528725147, + 0.006457069888710976, + 0.007806925568729639, + 0.012509931810200214, + 0.011212607845664024, + 0.023964278399944305, + 0.013519885018467903, + 0.02220798097550869, + 0.012199141085147858, + 0.0070828404277563095, + 0.01019056886434555, + 0.01178012602031231, + 0.016558829694986343, + 0.009352318942546844, + 0.010949688032269478, + 0.012969139963388443, + 0.009961046278476715, + 0.007824471220374107, + 0.008553629741072655, + 0.032333895564079285, + 0.031805988401174545, + 0.010865057818591595, + 0.010736766271293163, + 0.012660574167966843, + 0.019028695300221443, + 0.0035395370796322823, + 0.008648844435811043, + 0.00768617307767272, + 0.015336527489125729, + 0.01063139084726572, + 0.009710438549518585, + 0.014123858883976936, + 0.015500916168093681, + 0.004935132339596748, + 0.005716902669519186, + 0.0063565620221197605, + 0.011400836519896984, + 0.007490738295018673, + 0.009035155177116394, + 0.00868894811719656, + 0.009403177537024021, + 0.0022332703229039907, + 0.002620466286316514, + 0.004054922144860029, + 0.004407003056257963, + 0.004740895703434944, + 0.004393189679831266, + 0.0043238983489573, + 0.004186796955764294, + 0.0064192828722298145, + 0.0072182598523795605, + 0.007248392794281244, + 0.010498046875, + 0.008629181422293186, + 0.012361000292003155, + 0.007583112921565771, + 0.010975918732583523, + 0.01021980494260788, + 0.012411840260028839, + 0.014962545596063137, + 0.014510034583508968, + 0.024305421859025955, + 0.044771645218133926, + 0.013547774404287338, + 0.04106674715876579, + 0.006979444529861212, + 0.008302050642669201, + 0.011245010420680046, + 0.01677553541958332, + 0.02194199524819851, + 0.01836949959397316, + 0.017741579562425613, + 0.015592117793858051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.000922849983908236 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle shot panning rightward, captured in Steve McCurry's signature documentary style, captures a scene on a wet city sidewalk late at night. In the frame, a highly saturated neon light casts a shadow from the entrance of an unseen bar. Deep reds and electric blues intertwine, creating psychedelic patches of light across the damp asphalt. At the center of the image is a ginger tabby cat, one ear slightly chipped, its fur slightly matted but warm in the multicolored light. Huddled in the shadows near the bar entrance, its body forming an elegant arc, it raises one front paw with focused and deliberate concentration, meticulously and methodically licking the dust from its paw pad with its pink tongue. The camera moves in incredibly slow steps, following the rhythm of its licking, capturing in close-up the way water droplets from its tongue touch its paw and the soft golden edges of its body outlined by the neon light. The entire scene is shrouded in a quiet and profound atmosphere. The strong color contrast forms a sharp contrast with the subject's quiet posture, creating a poetic moment full of dignity and loneliness found in a corner of a bustling city.", + "seed": 13003, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_8394_seed_13003_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07537092523416504, + "vlm_logprob": -0.07537092523416504, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.07537092523416504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23681548237800598 + ], + "opposite_logprob_tokens": [ + -1.6118154525756836 + ], + "logprob_reward": -0.23681548237800598, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23681548237800598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16503913700580597, + 0.12515400350093842, + 0.23128800094127655, + 0.05371393263339996, + 0.1516173630952835, + 0.04765167459845543, + 0.02476244792342186, + 0.025028029456734657, + 0.03470126539468765, + 0.14104418456554413 + ], + "credit_map": [ + 0.012010088190436363, + 0.021225949749350548, + 0.023834755644202232, + 0.020624345168471336, + 0.028660034760832787, + 0.019425179809331894, + 0.01944865472614765, + 0.019810112193226814, + 0.008801659569144249, + 0.014383714646100998, + 0.019276631996035576, + 0.017697205767035484, + 0.013334481976926327, + 0.017178736627101898, + 0.015711167827248573, + 0.01877039484679699, + 0.0169307179749012, + 0.015114905312657356, + 0.059876393526792526, + 0.03912043944001198, + 0.014018584042787552, + 0.028530271723866463, + 0.02577347308397293, + 0.031923189759254456, + 0.003414714941754937, + 0.005745421163737774, + 0.0074136992916464806, + 0.007020543795078993, + 0.005414912942796946, + 0.00719454325735569, + 0.0076798610389232635, + 0.009830230847001076, + 0.006779455579817295, + 0.013131082989275455, + 0.00830046646296978, + 0.022949904203414917, + 0.01722084730863571, + 0.02588769607245922, + 0.03339943289756775, + 0.023948457092046738, + 0.002511924132704735, + 0.00519154779613018, + 0.004442958626896143, + 0.007931399159133434, + 0.005719015374779701, + 0.005136595573276281, + 0.009255883283913136, + 0.007462347391992807, + 0.0016332768136635423, + 0.0026105260476469994, + 0.0030524474568665028, + 0.004097068682312965, + 0.0032949158921837807, + 0.003346224082633853, + 0.0030430257320404053, + 0.0036849621683359146, + 0.0014364337548613548, + 0.00253338273614645, + 0.0026029394939541817, + 0.004161445423960686, + 0.002820269437506795, + 0.004241599701344967, + 0.003165661357343197, + 0.004066294990479946, + 0.0023832274600863457, + 0.004324373789131641, + 0.003928542602807283, + 0.004487947095185518, + 0.0058306860737502575, + 0.004203729797154665, + 0.004091828595846891, + 0.0054509276524186134, + 0.011351723223924637, + 0.01400443259626627, + 0.016454661265015602, + 0.020456623286008835, + 0.02489745244383812, + 0.016616089269518852, + 0.01697181537747383, + 0.020291374996304512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.23681548237800598 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819726943969727 + ], + "opposite_logprob_tokens": [ + -3.5298197269439697 + ], + "logprob_reward": -0.029819726943969727, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819726943969727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4316307008266449, + 0.019704658538103104, + 0.03441565856337547, + 0.02223445475101471, + 0.11627384275197983, + 0.04632330313324928, + 0.029665613546967506, + 0.05340246111154556, + 0.046648599207401276, + 0.19970069825649261 + ], + "credit_map": [ + 0.017382482066750526, + 0.040177009999752045, + 0.025003843009471893, + 0.056631311774253845, + 0.13313725590705872, + 0.05911470949649811, + 0.06046328321099281, + 0.039720822125673294, + 0.0010999282822012901, + 0.0026928780134767294, + 0.0018908488564193249, + 0.003352536354213953, + 0.0018003948498517275, + 0.002669937675818801, + 0.0035205145832151175, + 0.0026776185259222984, + 0.002362937433645129, + 0.002625806024298072, + 0.009422935545444489, + 0.005725730210542679, + 0.0018009126652032137, + 0.003786270972341299, + 0.0038918068166822195, + 0.004799259826540947, + 0.0012512257089838386, + 0.002810862846672535, + 0.0016449268441647291, + 0.0026511806063354015, + 0.0026766322553157806, + 0.0038278407882899046, + 0.0035453494638204575, + 0.003826433327049017, + 0.004972061142325401, + 0.010750544257462025, + 0.007551578804850578, + 0.022957343608140945, + 0.015274236910045147, + 0.020991545170545578, + 0.01662055402994156, + 0.017155975103378296, + 0.0020921221002936363, + 0.005692841485142708, + 0.00332463881932199, + 0.007142550311982632, + 0.008311389945447445, + 0.007126349024474621, + 0.00496134627610445, + 0.007672065403312445, + 0.0017150185303762555, + 0.002851843601092696, + 0.003371061757206917, + 0.0063243736512959, + 0.003772386582568288, + 0.0028416200075298548, + 0.003878949908539653, + 0.004910358693450689, + 0.0030805650167167187, + 0.005461303982883692, + 0.0041811647824943066, + 0.005828523542732, + 0.004664439242333174, + 0.015938295051455498, + 0.0048633478581905365, + 0.009384824894368649, + 0.002597725484520197, + 0.004492642357945442, + 0.005844929721206427, + 0.005949707236140966, + 0.007739502005279064, + 0.006291151978075504, + 0.00610380107536912, + 0.007629140745848417, + 0.012672347016632557, + 0.014027253724634647, + 0.01915108412504196, + 0.02690676413476467, + 0.05600539594888687, + 0.017810702323913574, + 0.024540744721889496, + 0.028586382046341896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.029819726943969727 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496637150645256 + ], + "opposite_logprob_tokens": [ + -2.929966449737549 + ], + "logprob_reward": -0.05496637150645256, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496637150645256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6480847001075745, + 0.025889115408062935, + 0.023024391382932663, + 0.01819617860019207, + 0.0500495508313179, + 0.07753018289804459, + 0.025567226111888885, + 0.05518396571278572, + 0.022711018100380898, + 0.05376369506120682 + ], + "credit_map": [ + 0.023903174325823784, + 0.05968274548649788, + 0.03593295067548752, + 0.08668804168701172, + 0.20014944672584534, + 0.0910174623131752, + 0.09082861244678497, + 0.059882283210754395, + 0.0014823070960119367, + 0.0017945419531315565, + 0.002017901511862874, + 0.004509578458964825, + 0.0027449261397123337, + 0.005479051265865564, + 0.004515248816460371, + 0.003345560748130083, + 0.0018695698818191886, + 0.0018089876975864172, + 0.0021391070913523436, + 0.0031675908248871565, + 0.002034575678408146, + 0.004379699006676674, + 0.004077422432601452, + 0.0035474388860166073, + 0.0010264776647090912, + 0.0017120452830567956, + 0.0015187718672677875, + 0.0025338614359498024, + 0.0015771827893331647, + 0.0037498390302062035, + 0.0026875529438257217, + 0.003390447236597538, + 0.0018153875134885311, + 0.004355786368250847, + 0.003673953004181385, + 0.01078549213707447, + 0.005274493247270584, + 0.007689611986279488, + 0.008494945243000984, + 0.0079598817974329, + 0.0036835449282079935, + 0.008780572563409805, + 0.005597084295004606, + 0.01483098417520523, + 0.011717161163687706, + 0.010061923414468765, + 0.010094557888805866, + 0.012764355167746544, + 0.001462138257920742, + 0.0027185557410120964, + 0.0024339111987501383, + 0.004897147882729769, + 0.0036082137376070023, + 0.0037309397011995316, + 0.003067149082198739, + 0.003649170510470867, + 0.002191298408433795, + 0.006385473068803549, + 0.006207720842212439, + 0.011487076990306377, + 0.007736867759376764, + 0.009094261564314365, + 0.004590029362589121, + 0.007491236086934805, + 0.001526769483461976, + 0.0023118203971534967, + 0.002503457944840193, + 0.003210637718439102, + 0.0037081162445247173, + 0.003641426796093583, + 0.002679130295291543, + 0.0031296578235924244, + 0.0026454466860741377, + 0.003743237815797329, + 0.006001794710755348, + 0.0074179829098284245, + 0.012183299288153648, + 0.008384788408875465, + 0.006684067193418741, + 0.006703077349811792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.05496637150645256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054946962743997574 + ], + "opposite_logprob_tokens": [ + -2.9299468994140625 + ], + "logprob_reward": -0.054946962743997574, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054946962743997574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.290859192609787, + 0.04273494333028793, + 0.06832685321569443, + 0.07281988859176636, + 0.11172901839017868, + 0.160233736038208, + 0.030764739960432053, + 0.1397867053747177, + 0.01571415178477764, + 0.06703081727027893 + ], + "credit_map": [ + 0.010372217744588852, + 0.026487061753869057, + 0.01654091477394104, + 0.038281816989183426, + 0.08723863959312439, + 0.04288759455084801, + 0.04200389236211777, + 0.027047041803598404, + 0.0035376392770558596, + 0.004786214791238308, + 0.007358839735388756, + 0.005608301144093275, + 0.004013990517705679, + 0.006320275366306305, + 0.004540951922535896, + 0.006568727549165487, + 0.004948464222252369, + 0.0046133073046803474, + 0.015975521877408028, + 0.010976970195770264, + 0.005940015893429518, + 0.009524282068014145, + 0.007971108891069889, + 0.00837718602269888, + 0.004113157279789448, + 0.007435258012264967, + 0.010589511133730412, + 0.007983943447470665, + 0.007901588454842567, + 0.011979755014181137, + 0.01171433087438345, + 0.011102345772087574, + 0.004942034371197224, + 0.01024181954562664, + 0.008302231319248676, + 0.01687772199511528, + 0.015677504241466522, + 0.023969879373908043, + 0.015851477161049843, + 0.01586635410785675, + 0.009250960312783718, + 0.018536873161792755, + 0.014263234101235867, + 0.02305976301431656, + 0.020643936470150948, + 0.024128878489136696, + 0.025350330397486687, + 0.024999752640724182, + 0.002574802376329899, + 0.002879783045500517, + 0.004633528180420399, + 0.00409963121637702, + 0.003965269308537245, + 0.004206790588796139, + 0.0035486873239278793, + 0.004856247920542955, + 0.012500802986323833, + 0.01470544096082449, + 0.01514781266450882, + 0.017661482095718384, + 0.019371114671230316, + 0.0242997445166111, + 0.013074859976768494, + 0.023025449365377426, + 0.0011706134537234902, + 0.0017553364159539342, + 0.0021321012172847986, + 0.002413003472611308, + 0.0024182708002626896, + 0.0016774744726717472, + 0.002006631577387452, + 0.002140721306204796, + 0.003692369442433119, + 0.0043607740662992, + 0.005901006516069174, + 0.01016163732856512, + 0.023015689104795456, + 0.003845388535410166, + 0.006452473811805248, + 0.009601476602256298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.054946962743997574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003060825983993709 + ], + "opposite_logprob_tokens": [ + -8.125306129455566 + ], + "logprob_reward": -0.0003060825983993709, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003060825983993709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8276332020759583, + 0.012244471348822117, + 0.01491796039044857, + 0.008981473743915558, + 0.017175478860735893, + 0.01615300215780735, + 0.012245096266269684, + 0.025362424552440643, + 0.02879353240132332, + 0.036493316292762756 + ], + "credit_map": [ + 0.0279645174741745, + 0.07483640313148499, + 0.0430615209043026, + 0.11133965104818344, + 0.25330501794815063, + 0.1226668432354927, + 0.11580946296453476, + 0.07864981889724731, + 0.0005710445111617446, + 0.0008041611290536821, + 0.001453787088394165, + 0.002334591932594776, + 0.0010507535189390182, + 0.0024156118743121624, + 0.0019509615376591682, + 0.0016635595820844173, + 0.0013749670470133424, + 0.0011513610370457172, + 0.002120576798915863, + 0.0021969981025904417, + 0.001492993556894362, + 0.0017020924715325236, + 0.0027345868293195963, + 0.0021443841978907585, + 0.0005513396463356912, + 0.0006144189974293113, + 0.0008400665828958154, + 0.000997409108094871, + 0.0011365743121132255, + 0.0020316599402576685, + 0.0009529890958219767, + 0.0018570167012512684, + 0.0009248625719919801, + 0.001682488713413477, + 0.0017875228077173233, + 0.0022858348675072193, + 0.0018427652539685369, + 0.0031335053499788046, + 0.0029437674675136805, + 0.0025747318286448717, + 0.0006911004311405122, + 0.0017806646646931767, + 0.0014392182929441333, + 0.002469855360686779, + 0.0020523343700915575, + 0.0018082951428368688, + 0.003448269795626402, + 0.002463264623656869, + 0.0007129971636459231, + 0.0012025393079966307, + 0.0012003207812085748, + 0.0014487870503216982, + 0.0019844418857246637, + 0.0015270825242623687, + 0.0021581577602773905, + 0.0020107701420783997, + 0.001915296190418303, + 0.0030447118915617466, + 0.0026995057705789804, + 0.004009978845715523, + 0.002517047105357051, + 0.004457397386431694, + 0.003165054600685835, + 0.0035534307826310396, + 0.0019306311151012778, + 0.0029360451735556126, + 0.004170896019786596, + 0.0033227894455194473, + 0.005116583313792944, + 0.003790353424847126, + 0.003187054768204689, + 0.004339178558439016, + 0.00216052052564919, + 0.0029441595543175936, + 0.003555209841579199, + 0.004657570738345385, + 0.00812496431171894, + 0.005094533320516348, + 0.005236188415437937, + 0.004720169119536877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0003060825983993709 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle shot panning rightward, captured in Steve McCurry's signature documentary style, captures a scene on a wet city sidewalk late at night. In the frame, a highly saturated neon light casts a shadow from the entrance of an unseen bar. Deep reds and electric blues intertwine, creating psychedelic patches of light across the damp asphalt. At the center of the image is a ginger tabby cat, one ear slightly chipped, its fur slightly matted but warm in the multicolored light. Huddled in the shadows near the bar entrance, its body forming an elegant arc, it raises one front paw with focused and deliberate concentration, meticulously and methodically licking the dust from its paw pad with its pink tongue. The camera moves in incredibly slow steps, following the rhythm of its licking, capturing in close-up the way water droplets from its tongue touch its paw and the soft golden edges of its body outlined by the neon light. The entire scene is shrouded in a quiet and profound atmosphere. The strong color contrast forms a sharp contrast with the subject's quiet posture, creating a poetic moment full of dignity and loneliness found in a corner of a bustling city.", + "seed": 13004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_3/index_8394_seed_13004_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0768389891134575, + "vlm_logprob": -0.0768389891134575, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.0768389891134575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09000913798809052 + ], + "opposite_logprob_tokens": [ + -2.7150092124938965 + ], + "logprob_reward": -0.09000913798809052, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09000913798809052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47524431347846985, + 0.11553028225898743, + 0.1828548014163971, + 0.026429083198308945, + 0.017984382808208466, + 0.04794500395655632, + 0.047982197254896164, + 0.017991622909903526, + 0.04220970720052719, + 0.025828663259744644 + ], + "credit_map": [ + 0.021549450233578682, + 0.033597785979509354, + 0.04340030252933502, + 0.07720904797315598, + 0.04281945899128914, + 0.029555827379226685, + 0.14198635518550873, + 0.08512607216835022, + 0.008054258301854134, + 0.01437226589769125, + 0.014228418469429016, + 0.021974189206957817, + 0.009761277586221695, + 0.008428622968494892, + 0.023760156705975533, + 0.01495109312236309, + 0.01230346504598856, + 0.019051261246204376, + 0.01983722671866417, + 0.025352947413921356, + 0.044878363609313965, + 0.01925758086144924, + 0.019341519102454185, + 0.022832438349723816, + 0.0018965482013300061, + 0.0034890705719590187, + 0.0034530232660472393, + 0.0032400127965956926, + 0.0033281075302511454, + 0.002693394897505641, + 0.004019829444587231, + 0.004309094976633787, + 0.0014711427502334118, + 0.0013764009345322847, + 0.003043706063181162, + 0.002569406758993864, + 0.001767707639373839, + 0.0012561437906697392, + 0.0034491282422095537, + 0.0030507470946758986, + 0.0035346599761396646, + 0.0036039697006344795, + 0.012949532829225063, + 0.0072845579124987125, + 0.0035766763612627983, + 0.004768308252096176, + 0.006221980322152376, + 0.006005318369716406, + 0.0021868585608899593, + 0.004048713017255068, + 0.004810658283531666, + 0.008878473192453384, + 0.0049138059839606285, + 0.0037219238001853228, + 0.0098568769171834, + 0.00956488586962223, + 0.001213424257002771, + 0.0015052075032144785, + 0.001624232972972095, + 0.0022475281730294228, + 0.0016690229531377554, + 0.0012082590255886316, + 0.004475514870136976, + 0.004048433154821396, + 0.002385403960943222, + 0.003669978817924857, + 0.004472602624446154, + 0.0037171405274420977, + 0.0035516726784408092, + 0.00193607562687248, + 0.011185666546225548, + 0.011291165836155415, + 0.0019297190010547638, + 0.0020743694622069597, + 0.0038330936804413795, + 0.002904894994571805, + 0.001876034657470882, + 0.0013489763950929046, + 0.005793087650090456, + 0.006068485789000988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.09000913798809052 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308144003152847 + ], + "opposite_logprob_tokens": [ + -3.168081521987915 + ], + "logprob_reward": -0.04308144003152847, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308144003152847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09382898360490799, + 0.08913493156433105, + 0.13778533041477203, + 0.06477194279432297, + 0.06450150907039642, + 0.06159602478146553, + 0.17161624133586884, + 0.1081419587135315, + 0.04368865117430687, + 0.1649344116449356 + ], + "credit_map": [ + 0.005742118693888187, + 0.0104439128190279, + 0.011184519156813622, + 0.014468294568359852, + 0.008467674255371094, + 0.008324233815073967, + 0.02251441963016987, + 0.012683807872235775, + 0.0076627423986792564, + 0.011369475163519382, + 0.007822922430932522, + 0.014573439955711365, + 0.010881738737225533, + 0.006593063939362764, + 0.01848595030605793, + 0.011745598167181015, + 0.007584038190543652, + 0.011779929511249065, + 0.00775973591953516, + 0.01819009520113468, + 0.025174425914883614, + 0.011578264646232128, + 0.03343842923641205, + 0.022280411794781685, + 0.0027978725265711546, + 0.004466035403311253, + 0.003894685534760356, + 0.006655458360910416, + 0.005238119978457689, + 0.003845196682959795, + 0.01904614455997944, + 0.018828433007001877, + 0.002898682374507189, + 0.004613972268998623, + 0.0027990664821118116, + 0.006912572775036097, + 0.009544657543301582, + 0.005077103618532419, + 0.01743878237903118, + 0.015216671861708164, + 0.004041328094899654, + 0.004451073240488768, + 0.008985163643956184, + 0.00780975678935647, + 0.0048207687214016914, + 0.003838764037936926, + 0.01395595632493496, + 0.01369321160018444, + 0.00611798046156764, + 0.004284568130970001, + 0.005942784249782562, + 0.005400759633630514, + 0.007386944256722927, + 0.005549014080315828, + 0.0697755366563797, + 0.06715865433216095, + 0.005742904730141163, + 0.0071999491192400455, + 0.009018790908157825, + 0.009843798354268074, + 0.009537308476865292, + 0.006085217464715242, + 0.031080802902579308, + 0.029633192345499992, + 0.002297760220244527, + 0.0016310660867020488, + 0.0028247758746147156, + 0.00199555023573339, + 0.002737718867138028, + 0.001923565985634923, + 0.014965689741075039, + 0.015312528237700462, + 0.006996401585638523, + 0.005116980988532305, + 0.004572907462716103, + 0.008425476960837841, + 0.010287035256624222, + 0.006485568359494209, + 0.06127612292766571, + 0.06177390366792679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.04308144003152847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003937661647797 + ], + "opposite_logprob_tokens": [ + -2.6950392723083496 + ], + "logprob_reward": -0.07003937661647797, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003937661647797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0415024496614933, + 0.07230661064386368, + 0.05402793735265732, + 0.07545006275177002, + 0.13955679535865784, + 0.06623949855566025, + 0.3825782835483551, + 0.07899108529090881, + 0.03792641684412956, + 0.05142088234424591 + ], + "credit_map": [ + 0.0025861533358693123, + 0.00458219088613987, + 0.007111852988600731, + 0.007092155981808901, + 0.004862516652792692, + 0.003847433952614665, + 0.006330945063382387, + 0.005089201033115387, + 0.0047698793932795525, + 0.008562239818274975, + 0.007562557235360146, + 0.013838591054081917, + 0.006635178346186876, + 0.004136052448302507, + 0.01589873991906643, + 0.010903368704020977, + 0.004387999884784222, + 0.005896168760955334, + 0.005675620399415493, + 0.012781268917024136, + 0.005649798084050417, + 0.0038990492466837168, + 0.008510996587574482, + 0.00722702918574214, + 0.005185709800571203, + 0.008390454575419426, + 0.011099264957010746, + 0.009125218726694584, + 0.01414505485445261, + 0.006903181318193674, + 0.010809814557433128, + 0.009791354648768902, + 0.005207620095461607, + 0.004016030579805374, + 0.006888630799949169, + 0.007542961742728949, + 0.0059122079983353615, + 0.005062295123934746, + 0.05593924969434738, + 0.04898779094219208, + 0.00478653609752655, + 0.005231531802564859, + 0.015518929809331894, + 0.010546759702265263, + 0.005811978597193956, + 0.006234649568796158, + 0.009238234721124172, + 0.008870875462889671, + 0.013515835627913475, + 0.012691846117377281, + 0.016957564279437065, + 0.016924502328038216, + 0.015916695818305016, + 0.011604133993387222, + 0.1513599008321762, + 0.14360779523849487, + 0.005463532637804747, + 0.00588770117610693, + 0.011371859349310398, + 0.009345135651528835, + 0.00826304405927658, + 0.00522775249555707, + 0.01774316281080246, + 0.01568889245390892, + 0.0021104111801832914, + 0.0024802491534501314, + 0.003250626614317298, + 0.0029383732471615076, + 0.003127096453681588, + 0.0017302760388702154, + 0.01114728394895792, + 0.011142096482217312, + 0.0025210343301296234, + 0.0031803229358047247, + 0.006817121058702469, + 0.005532507319003344, + 0.004554677288979292, + 0.002633490599691868, + 0.013286268338561058, + 0.012895455583930016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.07003937661647797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798797994852066 + ], + "opposite_logprob_tokens": [ + -1.8048797845840454 + ], + "logprob_reward": -0.1798797994852066, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798797994852066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06992675364017487, + 0.053813837468624115, + 0.042706042528152466, + 0.06983430683612823, + 0.03185086324810982, + 0.02497449889779091, + 0.32631000876426697, + 0.04058339446783066, + 0.026855722069740295, + 0.31314462423324585 + ], + "credit_map": [ + 0.0035291651729494333, + 0.005398159846663475, + 0.009682298637926579, + 0.014283891767263412, + 0.005758565850555897, + 0.004516391549259424, + 0.015810174867510796, + 0.01094810664653778, + 0.0032600269187241793, + 0.004687591455876827, + 0.006238617468625307, + 0.008876591920852661, + 0.008430046029388905, + 0.0048761069774627686, + 0.009754670783877373, + 0.007690188940614462, + 0.0026338212192058563, + 0.003545906161889434, + 0.004760769661515951, + 0.00672902213409543, + 0.006692654453217983, + 0.004042987711727619, + 0.007547472603619099, + 0.006753405090421438, + 0.004253672901540995, + 0.006551813799887896, + 0.005463641136884689, + 0.007445742376148701, + 0.011691270396113396, + 0.005037237424403429, + 0.015799041837453842, + 0.013591884635388851, + 0.001509415335021913, + 0.0012975811259821057, + 0.0037580118514597416, + 0.0030818316154181957, + 0.0021953354589641094, + 0.0011875999625772238, + 0.00999322347342968, + 0.00882786512374878, + 0.0011510306503623724, + 0.003095583524554968, + 0.0016128539573401213, + 0.0037851142697036266, + 0.002858043648302555, + 0.0018292200984433293, + 0.004762209486216307, + 0.005880442447960377, + 0.011983818374574184, + 0.008827532641589642, + 0.01466493308544159, + 0.012756936252117157, + 0.015232713893055916, + 0.01075383834540844, + 0.12863114476203918, + 0.1234591007232666, + 0.0029246860649436712, + 0.004249188583344221, + 0.005622353870421648, + 0.004657328594475985, + 0.004918096121400595, + 0.003229734720662236, + 0.007946663536131382, + 0.007035345770418644, + 0.0011821654625236988, + 0.002196789951995015, + 0.002434255788102746, + 0.0031247069127857685, + 0.0032836622558534145, + 0.002004997804760933, + 0.006292469799518585, + 0.006336673628538847, + 0.012364634312689304, + 0.010220437310636044, + 0.016529833897948265, + 0.018553439527750015, + 0.01945849321782589, + 0.012890035286545753, + 0.111959308385849, + 0.11116845160722733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.1798797994852066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a wet city sidewalk late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a ginger tabby cat with one slightly chipped ear visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat raising one front paw to lick dust from its paw pad with a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow pan rightward while capturing close-up details of water droplets and neon light edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the colors highly saturated with deep reds and electric blues creating psychedelic patches of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011851914459839463 + ], + "opposite_logprob_tokens": [ + -6.751185417175293 + ], + "logprob_reward": -0.0011851914459839463, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011851914459839463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25630462169647217, + 0.14373046159744263, + 0.16264550387859344, + 0.08159606903791428, + 0.11621852219104767, + 0.03604355454444885, + 0.08453312516212463, + 0.03824634850025177, + 0.017748411744832993, + 0.06293340027332306 + ], + "credit_map": [ + 0.010600751265883446, + 0.02102505788207054, + 0.021983683109283447, + 0.035913873463869095, + 0.024731386452913284, + 0.018951261416077614, + 0.07733242213726044, + 0.04576617479324341, + 0.010263700038194656, + 0.014999567531049252, + 0.024468641728162766, + 0.035475969314575195, + 0.01052151806652546, + 0.007645095232874155, + 0.021125927567481995, + 0.019230032339692116, + 0.007561902049928904, + 0.016657043248414993, + 0.01145353727042675, + 0.017536384984850883, + 0.04373607039451599, + 0.017841294407844543, + 0.026241425424814224, + 0.02161785028874874, + 0.005703838542103767, + 0.005857531912624836, + 0.009610982611775398, + 0.007553444243967533, + 0.006222357042133808, + 0.005304761230945587, + 0.019400401040911674, + 0.021942753344774246, + 0.003859078511595726, + 0.0030392420012503862, + 0.004967132583260536, + 0.006227268371731043, + 0.005177627317607403, + 0.0037067080847918987, + 0.047543715685606, + 0.04169774800539017, + 0.002043665386736393, + 0.002521972870454192, + 0.006750196218490601, + 0.0035168416798114777, + 0.005081410054117441, + 0.002702418016269803, + 0.006731893867254257, + 0.006695156916975975, + 0.005611080676317215, + 0.004050140734761953, + 0.014518462121486664, + 0.008633654564619064, + 0.0054529462940990925, + 0.0040524681098759174, + 0.02091151848435402, + 0.021302856504917145, + 0.0032052556052803993, + 0.004516215063631535, + 0.0065372842364013195, + 0.004843696486204863, + 0.0036144007463008165, + 0.0026960864197462797, + 0.006207626312971115, + 0.006625781301409006, + 0.0015717969508841634, + 0.001634382177144289, + 0.0014287569792941213, + 0.0024683168157935143, + 0.0021202396601438522, + 0.0014383718371391296, + 0.0035373421851545572, + 0.003549205604940653, + 0.0036366586573421955, + 0.0029923925176262856, + 0.005200735758990049, + 0.003753936616703868, + 0.003338449401780963, + 0.002817576052621007, + 0.020295582711696625, + 0.020898068323731422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0011851914459839463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_4/index_tensor([5716, 5716])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_4/index_tensor([5716, 5716])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2875852be0be13d120641fd449ad9abb663c2c4a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_4/index_tensor([5716, 5716])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low, wide-angle shot, shot from above, slowly pans upward through a lush bamboo forest shrouded in morning mist. Soft, diffused light penetrates the mist and dense bamboo foliage, casting visible beams of light in the air. The camera finally settles on a woman in a traditional Chinese costume, dressed in an elegant aquamarine silk robe. Her long, cascading black hair is held together only partially by a simple jade hairpin. Her face is clear and beautiful, her eyes focused and ethereal. With a light touch of the ground, her entire body rises as if weightless. The camera instantly switches to a fast, dynamic tracking shot, following her graceful figure as she nimbly weaves between the towering bamboo branches. Her wide sleeves flutter like butterfly wings as she moves swiftly, her skirt carving graceful arcs through the mist. A series of rapidly alternating close-ups follows: the tip of her shoe, delicately embroidered with cloud patterns, lightly steps on a dewy bamboo leaf, which trembles only slightly before returning to its composure; a strand of hair brushes across her focused and serene profile, revealing her perfectly defined jawline; her slender fingers form an elegant sword-like gesture in the air, steadily and calmly directing her body. The entire sequence is seamless, the camera language sharp and precise, capturing every moment of her ethereal, ethereal beauty. Amidst the cool tones of light and the lingering mist, an ethereal atmosphere is created, imbued with both martial arts power and poetic imagery.", + "seed": 17757, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_4/index_5716_seed_17757_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027810224145650864, + "vlm_logprob": -0.027810224145650864, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.027810224145650864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08142706751823425 + ], + "opposite_logprob_tokens": [ + -2.5814270973205566 + ], + "logprob_reward": -0.08142706751823425, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08142706751823425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14620473980903625, + 0.04129651561379433, + 0.1665949523448944, + 0.021945806220173836, + 0.3804764747619629, + 0.021654775366187096, + 0.13528798520565033, + 0.012130232527852058, + 0.030787324532866478, + 0.043621163815259933 + ], + "credit_map": [ + 0.0120698818936944, + 0.011151861399412155, + 0.016435882076621056, + 0.015483337454497814, + 0.02283940277993679, + 0.024254124611616135, + 0.02848956175148487, + 0.015480675734579563, + 0.0028545402456074953, + 0.004965680651366711, + 0.004722830839455128, + 0.005029503256082535, + 0.0031676548533141613, + 0.0038411696441471577, + 0.005223512649536133, + 0.0114916255697608, + 0.03174903988838196, + 0.012722117826342583, + 0.029667813330888748, + 0.012105366215109825, + 0.010385005734860897, + 0.03894089162349701, + 0.010566475801169872, + 0.02045823633670807, + 0.002413010224699974, + 0.0023982455022633076, + 0.0028262354899197817, + 0.004491141065955162, + 0.0019859238527715206, + 0.002278102096170187, + 0.003459931816905737, + 0.002093218034133315, + 0.015321156941354275, + 0.02447650209069252, + 0.021461617201566696, + 0.1183767095208168, + 0.13800066709518433, + 0.017795134335756302, + 0.02628815360367298, + 0.01875651255249977, + 0.0025775243993848562, + 0.0023894987534731627, + 0.0026145153678953648, + 0.0021454179659485817, + 0.0033316989429295063, + 0.0028182838577777147, + 0.001969902077689767, + 0.0038079351652413607, + 0.004820754751563072, + 0.050180431455373764, + 0.0032610499765723944, + 0.020489903166890144, + 0.008381674066185951, + 0.005970163736492395, + 0.005863455589860678, + 0.03632056340575218, + 0.0008468261221423745, + 0.00214293017052114, + 0.0008970575872808695, + 0.001620565657503903, + 0.0017984851729124784, + 0.0021758773364126682, + 0.0015531902899965644, + 0.0010953003074973822, + 0.0026306493673473597, + 0.0032765341456979513, + 0.0017011883901432157, + 0.005877129267901182, + 0.0054304632358253, + 0.004301026929169893, + 0.004618050064891577, + 0.0029522862751036882, + 0.0037115244194865227, + 0.005120902322232723, + 0.0032884585671126842, + 0.003521124366670847, + 0.010461674071848392, + 0.004904747940599918, + 0.008212549611926079, + 0.00440018018707633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.08142706751823425 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747911684215069 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747911684215069, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747911684215069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08126577734947205, + 0.05831017717719078, + 0.09356626868247986, + 0.04219374805688858, + 0.31751343607902527, + 0.05199563875794411, + 0.04936429113149643, + 0.07341159135103226, + 0.13680298626422882, + 0.09557611495256424 + ], + "credit_map": [ + 0.009765446186065674, + 0.0070517598651349545, + 0.01745687983930111, + 0.006003346759825945, + 0.011526333168148994, + 0.011158560402691364, + 0.006173846311867237, + 0.012129608541727066, + 0.006747570354491472, + 0.0068686301819980145, + 0.008003630675375462, + 0.00740642799064517, + 0.005062428768724203, + 0.006042376160621643, + 0.006865325849503279, + 0.0113137885928154, + 0.017168069258332253, + 0.006871814373880625, + 0.016400611028075218, + 0.006460024043917656, + 0.0065686060115695, + 0.02194979600608349, + 0.006634093821048737, + 0.01151326298713684, + 0.006979864090681076, + 0.0045798905193805695, + 0.006130777299404144, + 0.005892335902899504, + 0.004508710466325283, + 0.003874853951856494, + 0.00580567866563797, + 0.004421638324856758, + 0.011072296649217606, + 0.020597582682967186, + 0.013471435755491257, + 0.11460065096616745, + 0.10994134843349457, + 0.012751930393278599, + 0.02141486294567585, + 0.013663318939507008, + 0.004671023692935705, + 0.009363796561956406, + 0.006263655610382557, + 0.006677644792944193, + 0.009281638078391552, + 0.004530938807874918, + 0.005489404313266277, + 0.005717533174902201, + 0.0018676039762794971, + 0.018517661839723587, + 0.0015601550694555044, + 0.007679780013859272, + 0.004242164082825184, + 0.002758591203019023, + 0.0040732878260314465, + 0.008665047585964203, + 0.008234417997300625, + 0.008092929609119892, + 0.010073048062622547, + 0.016988221555948257, + 0.008896895684301853, + 0.007741160225123167, + 0.0083816172555089, + 0.005003299564123154, + 0.009668753482401371, + 0.014447905123233795, + 0.0069843316450715065, + 0.02519736997783184, + 0.027969976887106895, + 0.01815192960202694, + 0.02151980623602867, + 0.012862914241850376, + 0.007902898825705051, + 0.010293596424162388, + 0.006043725181370974, + 0.009557140059769154, + 0.01793568581342697, + 0.018274348229169846, + 0.01628687046468258, + 0.009281855076551437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.006747911684215069 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370075300335884 + ], + "opposite_logprob_tokens": [ + -3.408700704574585 + ], + "logprob_reward": -0.03370075300335884, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370075300335884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08308301120996475, + 0.1073627844452858, + 0.046506352722644806, + 0.0391797237098217, + 0.3231569528579712, + 0.07249286025762558, + 0.03315151855349541, + 0.05231819674372673, + 0.0994398444890976, + 0.14330872893333435 + ], + "credit_map": [ + 0.009776405990123749, + 0.007262619212269783, + 0.013261807151138783, + 0.006047763861715794, + 0.013479354791343212, + 0.012339328415691853, + 0.009915429167449474, + 0.011000312864780426, + 0.011865059845149517, + 0.015049992129206657, + 0.012264298275113106, + 0.013969998806715012, + 0.007632219698280096, + 0.010255473665893078, + 0.009328687563538551, + 0.026997052133083344, + 0.006393615622073412, + 0.004484456032514572, + 0.006920510437339544, + 0.004111781716346741, + 0.004320272710174322, + 0.00872907042503357, + 0.006078766193240881, + 0.005467884708195925, + 0.006539738271385431, + 0.005521300248801708, + 0.004505505785346031, + 0.006083245854824781, + 0.00393365602940321, + 0.003690405050292611, + 0.005107610020786524, + 0.0037982668727636337, + 0.010898442938923836, + 0.023847049102187157, + 0.009693462401628494, + 0.09683660417795181, + 0.11856257170438766, + 0.010943255387246609, + 0.04072026163339615, + 0.011655337177217007, + 0.00675303814932704, + 0.009476317092776299, + 0.008291884325444698, + 0.008094819262623787, + 0.012720334343612194, + 0.008164839819073677, + 0.007932599633932114, + 0.011059037409722805, + 0.0018682401860132813, + 0.009690681472420692, + 0.0016208712477236986, + 0.003988287411630154, + 0.0023949455935508013, + 0.0023461831733584404, + 0.0029093821067363024, + 0.008332927711308002, + 0.006685552187263966, + 0.008880458772182465, + 0.008629219606518745, + 0.006499945186078548, + 0.006137178745120764, + 0.006711021065711975, + 0.00470993435010314, + 0.004064890556037426, + 0.005741416476666927, + 0.009384924545884132, + 0.0064151594415307045, + 0.0233016274869442, + 0.013449685648083687, + 0.012426701374351978, + 0.019926635548472404, + 0.00879370141774416, + 0.013993985950946808, + 0.013663937337696552, + 0.0104750981554389, + 0.013185957446694374, + 0.04114251583814621, + 0.01688992790877819, + 0.02076905034482479, + 0.013188272714614868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.03370075300335884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638269543647766 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638269543647766, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638269543647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09431415051221848, + 0.0488211028277874, + 0.07169324159622192, + 0.031396713107824326, + 0.3126031756401062, + 0.065078504383564, + 0.13891956210136414, + 0.06682657450437546, + 0.046809472143650055, + 0.12353755533695221 + ], + "credit_map": [ + 0.012404782697558403, + 0.006841482128947973, + 0.01751790940761566, + 0.006199502851814032, + 0.01488405466079712, + 0.016931189224123955, + 0.007299668621271849, + 0.012235550209879875, + 0.0049162511713802814, + 0.004608016926795244, + 0.005403600633144379, + 0.006361060310155153, + 0.004020014312118292, + 0.004712889902293682, + 0.010679579339921474, + 0.008119688369333744, + 0.01012235227972269, + 0.005971930455416441, + 0.009424117393791676, + 0.006876931060105562, + 0.006948061753064394, + 0.014689985662698746, + 0.010350223630666733, + 0.00730963284149766, + 0.002645252738147974, + 0.00457885954529047, + 0.00259851454757154, + 0.007061991840600967, + 0.0038653837982565165, + 0.0030806793365627527, + 0.004288660828024149, + 0.003277367912232876, + 0.012844662182033062, + 0.019606636837124825, + 0.015266782604157925, + 0.10627874732017517, + 0.11183079332113266, + 0.013183183036744595, + 0.021179167553782463, + 0.012413173913955688, + 0.006581586319953203, + 0.0076179527677595615, + 0.007230211980640888, + 0.008532319217920303, + 0.01047305017709732, + 0.0069294762797653675, + 0.006822295021265745, + 0.010891608893871307, + 0.0044206613674759865, + 0.05234605073928833, + 0.0035808063112199306, + 0.021047426387667656, + 0.009070240892469883, + 0.006800749339163303, + 0.0069712563417851925, + 0.03468235209584236, + 0.006459074094891548, + 0.0077105010859668255, + 0.006586388219147921, + 0.015527302399277687, + 0.008658762089908123, + 0.009117506444454193, + 0.007293124683201313, + 0.005473910365253687, + 0.004380451515316963, + 0.0042348699644207954, + 0.004387842956930399, + 0.008208747021853924, + 0.007772261742502451, + 0.005974957719445229, + 0.007527102250605822, + 0.0043232315219938755, + 0.009997474029660225, + 0.011649488471448421, + 0.009183738380670547, + 0.012085958383977413, + 0.018223794177174568, + 0.028892116621136665, + 0.0186574999243021, + 0.014847478829324245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.004638269543647766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253711897879839 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253711897879839, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253711897879839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10150013864040375, + 0.0895349457859993, + 0.11146540939807892, + 0.07454860210418701, + 0.15910831093788147, + 0.09265471994876862, + 0.06611236929893494, + 0.04779552295804024, + 0.16754557192325592, + 0.08973439037799835 + ], + "credit_map": [ + 0.011537989601492882, + 0.006863121408969164, + 0.020644206553697586, + 0.010945034213364124, + 0.01074237935245037, + 0.011299113743007183, + 0.01419233251363039, + 0.015275968238711357, + 0.008285983465611935, + 0.00992943812161684, + 0.012359445914626122, + 0.016443030908703804, + 0.007423185743391514, + 0.008032343350350857, + 0.00983610562980175, + 0.017225418239831924, + 0.01914827711880207, + 0.009694506414234638, + 0.01614370383322239, + 0.009404615499079227, + 0.009170560166239738, + 0.023396318778395653, + 0.012051636353135109, + 0.012455792166292667, + 0.014224824495613575, + 0.0119331581518054, + 0.010273921303451061, + 0.01004865299910307, + 0.007071126252412796, + 0.0074119665659964085, + 0.0058636292815208435, + 0.0077213263139128685, + 0.011471613310277462, + 0.017233818769454956, + 0.009716366417706013, + 0.022593053057789803, + 0.04135251045227051, + 0.00969499908387661, + 0.034403737634420395, + 0.012642232701182365, + 0.008229779079556465, + 0.012454402633011341, + 0.00929311104118824, + 0.009527955204248428, + 0.016440052539110184, + 0.011121310293674469, + 0.009668764658272266, + 0.015919344499707222, + 0.003932780586183071, + 0.021457666531205177, + 0.0028715243097394705, + 0.008325780741870403, + 0.00400823587551713, + 0.004119688179343939, + 0.003765569068491459, + 0.017631126567721367, + 0.004787906538695097, + 0.006778793875128031, + 0.0051146941259503365, + 0.011975064873695374, + 0.004537022206932306, + 0.004807436838746071, + 0.005707575939595699, + 0.004087033681571484, + 0.009641687385737896, + 0.016861559823155403, + 0.009926144033670425, + 0.03503386303782463, + 0.023318910971283913, + 0.023677131161093712, + 0.0324021652340889, + 0.016684111207723618, + 0.007936988025903702, + 0.012492645531892776, + 0.008758795447647572, + 0.01040111668407917, + 0.014073306694626808, + 0.015853503718972206, + 0.0100295040756464, + 0.01018853485584259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.01253711897879839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, wide-angle shot, shot from above, slowly pans upward through a lush bamboo forest shrouded in morning mist. Soft, diffused light penetrates the mist and dense bamboo foliage, casting visible beams of light in the air. The camera finally settles on a woman in a traditional Chinese costume, dressed in an elegant aquamarine silk robe. Her long, cascading black hair is held together only partially by a simple jade hairpin. Her face is clear and beautiful, her eyes focused and ethereal. With a light touch of the ground, her entire body rises as if weightless. The camera instantly switches to a fast, dynamic tracking shot, following her graceful figure as she nimbly weaves between the towering bamboo branches. Her wide sleeves flutter like butterfly wings as she moves swiftly, her skirt carving graceful arcs through the mist. A series of rapidly alternating close-ups follows: the tip of her shoe, delicately embroidered with cloud patterns, lightly steps on a dewy bamboo leaf, which trembles only slightly before returning to its composure; a strand of hair brushes across her focused and serene profile, revealing her perfectly defined jawline; her slender fingers form an elegant sword-like gesture in the air, steadily and calmly directing her body. The entire sequence is seamless, the camera language sharp and precise, capturing every moment of her ethereal, ethereal beauty. Amidst the cool tones of light and the lingering mist, an ethereal atmosphere is created, imbued with both martial arts power and poetic imagery.", + "seed": 17758, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_4/index_5716_seed_17758_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.053768701106309894, + "vlm_logprob": -0.053768701106309894, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.053768701106309894, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08487202227115631 + ], + "opposite_logprob_tokens": [ + -2.584872007369995 + ], + "logprob_reward": -0.08487202227115631, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08487202227115631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1717430204153061, + 0.20119009912014008, + 0.15199434757232666, + 0.05799781158566475, + 0.027915148064494133, + 0.022854629904031754, + 0.04651269316673279, + 0.03532067686319351, + 0.04834001511335373, + 0.2361314743757248 + ], + "credit_map": [ + 0.01617027446627617, + 0.015397131443023682, + 0.022978980094194412, + 0.026671726256608963, + 0.02007695659995079, + 0.026378348469734192, + 0.0241246335208416, + 0.01994497887790203, + 0.024294022470712662, + 0.014614981599152088, + 0.03383859619498253, + 0.020519081503152847, + 0.01624441333115101, + 0.03764786943793297, + 0.024758758023381233, + 0.029272396117448807, + 0.015503444708883762, + 0.022785991430282593, + 0.015888262540102005, + 0.0233165230602026, + 0.027544286102056503, + 0.012981527484953403, + 0.019190624356269836, + 0.014783707447350025, + 0.010201513767242432, + 0.0057610683143138885, + 0.011447062715888023, + 0.00632221344858408, + 0.005978498142212629, + 0.005576272960752249, + 0.006389201618731022, + 0.006321985740214586, + 0.003576455870643258, + 0.0035714865662157536, + 0.0034053348936140537, + 0.003800970036536455, + 0.0033950582146644592, + 0.004635518882423639, + 0.0033478555269539356, + 0.002182472264394164, + 0.0031405179761350155, + 0.0027514470275491476, + 0.003655574284493923, + 0.0029194781091064215, + 0.0023433994501829147, + 0.0026563953142613173, + 0.001551600988022983, + 0.00383621989749372, + 0.0032898944336920977, + 0.007526909001171589, + 0.004890065174549818, + 0.0060792965814471245, + 0.0058589656837284565, + 0.0071121808141469955, + 0.005844097584486008, + 0.005911287851631641, + 0.0018406715244054794, + 0.0038706345949321985, + 0.003109679091721773, + 0.0028872769325971603, + 0.007403355557471514, + 0.004357744008302689, + 0.005232726223766804, + 0.006618591956794262, + 0.005522257648408413, + 0.006188919767737389, + 0.004415942821651697, + 0.004185632802546024, + 0.00550822215154767, + 0.008850550279021263, + 0.005040292628109455, + 0.008628200739622116, + 0.017784399911761284, + 0.013829744420945644, + 0.015201268717646599, + 0.014040099456906319, + 0.04112642630934715, + 0.06276005506515503, + 0.0226446483284235, + 0.04874484986066818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.08487202227115631 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221273362636566 + ], + "opposite_logprob_tokens": [ + -4.26422119140625 + ], + "logprob_reward": -0.014221273362636566, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221273362636566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31366804242134094, + 0.13412152230739594, + 0.07230022549629211, + 0.040737684816122055, + 0.06414463371038437, + 0.08555422723293304, + 0.05188656970858574, + 0.03628933057188988, + 0.06118890270590782, + 0.1401088684797287 + ], + "credit_map": [ + 0.03850001469254494, + 0.04813806340098381, + 0.04947192221879959, + 0.07002429664134979, + 0.02698906511068344, + 0.022762959823012352, + 0.03420739248394966, + 0.02357431873679161, + 0.013737523928284645, + 0.013158870860934258, + 0.02437784895300865, + 0.026715058833360672, + 0.014859924092888832, + 0.011595825664699078, + 0.015033213421702385, + 0.014643258415162563, + 0.010549436323344707, + 0.007771828677505255, + 0.016229266300797462, + 0.0074041844345629215, + 0.0035587914753705263, + 0.007980875670909882, + 0.007688511628657579, + 0.011117327958345413, + 0.005637787282466888, + 0.003597778035327792, + 0.008413450792431831, + 0.003781785722821951, + 0.005478391423821449, + 0.0031629696022719145, + 0.007149201352149248, + 0.0035163203720003366, + 0.009396919049322605, + 0.007792919408529997, + 0.008632030338048935, + 0.008412107825279236, + 0.008571824058890343, + 0.009059792384505272, + 0.005259246565401554, + 0.007019799668341875, + 0.005936503876000643, + 0.013600308448076248, + 0.007972684688866138, + 0.017976466566324234, + 0.01645495742559433, + 0.006790202111005783, + 0.005848032888025045, + 0.010975067503750324, + 0.003975089639425278, + 0.00845787487924099, + 0.005497231148183346, + 0.006606951355934143, + 0.007482816930860281, + 0.007351598236709833, + 0.006160443648695946, + 0.0063545661978423595, + 0.003341489005833864, + 0.006064798682928085, + 0.004113742616027594, + 0.0035376420710235834, + 0.004708595108240843, + 0.0037211612798273563, + 0.005170807242393494, + 0.005631096661090851, + 0.004023020155727863, + 0.009420672431588173, + 0.005266415420919657, + 0.01441547367721796, + 0.006664169952273369, + 0.007063701748847961, + 0.009823953732848167, + 0.004511493723839521, + 0.011715142987668514, + 0.015081338584423065, + 0.012472542002797127, + 0.016212979331612587, + 0.025974057614803314, + 0.02590988576412201, + 0.013297745026648045, + 0.01944517344236374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.014221273362636566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031711310148239 + ], + "opposite_logprob_tokens": [ + -2.3503170013427734 + ], + "logprob_reward": -0.10031711310148239, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031711310148239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17288771271705627, + 0.05235958844423294, + 0.04940799996256828, + 0.07883462309837341, + 0.0985783040523529, + 0.12514929473400116, + 0.08729874342679977, + 0.11861182004213333, + 0.09774895012378693, + 0.11912290751934052 + ], + "credit_map": [ + 0.015651876106858253, + 0.0176568403840065, + 0.022560469806194305, + 0.028246572241187096, + 0.024260111153125763, + 0.018563445657491684, + 0.026889700442552567, + 0.019058708101511, + 0.005925761070102453, + 0.004156437236815691, + 0.005303555633872747, + 0.007955714128911495, + 0.004043941386044025, + 0.008077199570834637, + 0.010092389769852161, + 0.006804590579122305, + 0.007056109607219696, + 0.0051783062517642975, + 0.008521512150764465, + 0.005001924466341734, + 0.005714442580938339, + 0.006150180008262396, + 0.005853826180100441, + 0.005931696388870478, + 0.011248378083109856, + 0.009173525497317314, + 0.013571804389357567, + 0.010423921048641205, + 0.008664259687066078, + 0.007191158831119537, + 0.01297811884433031, + 0.005583456717431545, + 0.010637305676937103, + 0.012532755732536316, + 0.015331112779676914, + 0.014869464561343193, + 0.010440489277243614, + 0.012628797441720963, + 0.012219643220305443, + 0.009918737225234509, + 0.011846939101815224, + 0.013203305192291737, + 0.015301299281418324, + 0.02002824656665325, + 0.01719082146883011, + 0.013477986678481102, + 0.009545319713652134, + 0.024555377662181854, + 0.0086856409907341, + 0.009157259948551655, + 0.008562644012272358, + 0.0110926553606987, + 0.011271810159087181, + 0.011886193417012691, + 0.011676833033561707, + 0.014965708367526531, + 0.01021854393184185, + 0.021364670246839523, + 0.01236986368894577, + 0.012069444172084332, + 0.014187186025083065, + 0.013403593562543392, + 0.020078597590327263, + 0.014919917099177837, + 0.010586552321910858, + 0.01055337954312563, + 0.008774609304964542, + 0.018489643931388855, + 0.011578050442039967, + 0.0143953338265419, + 0.010943718254566193, + 0.012427669018507004, + 0.011601969599723816, + 0.013754689134657383, + 0.010565567761659622, + 0.019464444369077682, + 0.01667780615389347, + 0.013115857727825642, + 0.02043764479458332, + 0.013504932634532452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.10031711310148239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308132827281952 + ], + "opposite_logprob_tokens": [ + -3.168081283569336 + ], + "logprob_reward": -0.04308132827281952, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308132827281952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.127890482544899, + 0.07006958872079849, + 0.1343708038330078, + 0.056691642850637436, + 0.10449254512786865, + 0.19652311503887177, + 0.10428042709827423, + 0.049993082880973816, + 0.05891365930438042, + 0.09677469730377197 + ], + "credit_map": [ + 0.014827764593064785, + 0.016787610948085785, + 0.015732724219560623, + 0.026534246280789375, + 0.012844477780163288, + 0.008400185965001583, + 0.020488198846578598, + 0.012275267392396927, + 0.006658296566456556, + 0.006476263515651226, + 0.00984395481646061, + 0.013380107469856739, + 0.008748685009777546, + 0.009381291456520557, + 0.008834685198962688, + 0.0067463042214512825, + 0.017632339149713516, + 0.01566462591290474, + 0.025205831974744797, + 0.013973587192595005, + 0.009339334443211555, + 0.018577679991722107, + 0.016560586169362068, + 0.017416812479496002, + 0.007484773639589548, + 0.007534698583185673, + 0.006671043112874031, + 0.0066603017039597034, + 0.006643705070018768, + 0.006686805281788111, + 0.00789673812687397, + 0.0071135773323476315, + 0.009534352459013462, + 0.015141287818551064, + 0.01175647508352995, + 0.01931796781718731, + 0.012162999249994755, + 0.013683991506695747, + 0.010519130155444145, + 0.012376335449516773, + 0.01965963840484619, + 0.020443247631192207, + 0.024357596412301064, + 0.029069457203149796, + 0.029785996302962303, + 0.026838185265660286, + 0.01265732292085886, + 0.03371167555451393, + 0.009421131573617458, + 0.021537330001592636, + 0.012201344594359398, + 0.013716747984290123, + 0.01051727868616581, + 0.010435930453240871, + 0.012202870100736618, + 0.014247790910303593, + 0.003987334202975035, + 0.007572927512228489, + 0.004875028505921364, + 0.004784636199474335, + 0.00605496671050787, + 0.005271896719932556, + 0.008334060199558735, + 0.009112230502068996, + 0.003703533438965678, + 0.009130042046308517, + 0.00495333643630147, + 0.013615032657980919, + 0.006777159869670868, + 0.006550766061991453, + 0.009389122016727924, + 0.004794668406248093, + 0.008571193553507328, + 0.00893038883805275, + 0.007490736432373524, + 0.012067358009517193, + 0.019464101642370224, + 0.014900828711688519, + 0.01055027823895216, + 0.014799810945987701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.04308132827281952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351768523454666 + ], + "opposite_logprob_tokens": [ + -3.6513516902923584 + ], + "logprob_reward": -0.026351768523454666, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351768523454666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22396066784858704, + 0.05834191292524338, + 0.1289076805114746, + 0.08239760249853134, + 0.10028079897165298, + 0.1796436458826065, + 0.04901673272252083, + 0.05593346059322357, + 0.03741230070590973, + 0.0841052457690239 + ], + "credit_map": [ + 0.022794917225837708, + 0.02789844200015068, + 0.03336585685610771, + 0.04726474732160568, + 0.02011706493794918, + 0.02584272250533104, + 0.03204872086644173, + 0.014628192409873009, + 0.005157070700079203, + 0.006612440571188927, + 0.009515427052974701, + 0.008840253576636314, + 0.006334953010082245, + 0.008279874920845032, + 0.007236187811940908, + 0.006365708541125059, + 0.011620364151895046, + 0.016604453325271606, + 0.023314066231250763, + 0.01671893335878849, + 0.010568948462605476, + 0.01251588761806488, + 0.017767634242773056, + 0.019797392189502716, + 0.0121181420981884, + 0.011762254871428013, + 0.012489058077335358, + 0.008938504382967949, + 0.008401188999414444, + 0.010242867283523083, + 0.008663428016006947, + 0.00978215504437685, + 0.011859889142215252, + 0.012605582363903522, + 0.02496315911412239, + 0.01553725078701973, + 0.007268448825925589, + 0.01232922449707985, + 0.008350427262485027, + 0.007366816513240337, + 0.013697012327611446, + 0.023396363481879234, + 0.015606195665895939, + 0.042996279895305634, + 0.03725321963429451, + 0.016857462003827095, + 0.010920276865363121, + 0.018916841596364975, + 0.0046362983994185925, + 0.006691472604870796, + 0.00575575977563858, + 0.006288193166255951, + 0.006979084108024836, + 0.006256739608943462, + 0.006497425027191639, + 0.005911759100854397, + 0.006362478714436293, + 0.008653140626847744, + 0.005824172403663397, + 0.006023832596838474, + 0.00782877579331398, + 0.006331826560199261, + 0.009441415779292583, + 0.005467815324664116, + 0.003132663667201996, + 0.005459223873913288, + 0.002583116991445422, + 0.005486456211656332, + 0.004524183925241232, + 0.006133616901934147, + 0.004439764656126499, + 0.005653273314237595, + 0.0064489878714084625, + 0.005727611016482115, + 0.0061652157455682755, + 0.006300469394773245, + 0.012929243966937065, + 0.021475717425346375, + 0.007668782491236925, + 0.017389213666319847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.026351768523454666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, wide-angle shot, shot from above, slowly pans upward through a lush bamboo forest shrouded in morning mist. Soft, diffused light penetrates the mist and dense bamboo foliage, casting visible beams of light in the air. The camera finally settles on a woman in a traditional Chinese costume, dressed in an elegant aquamarine silk robe. Her long, cascading black hair is held together only partially by a simple jade hairpin. Her face is clear and beautiful, her eyes focused and ethereal. With a light touch of the ground, her entire body rises as if weightless. The camera instantly switches to a fast, dynamic tracking shot, following her graceful figure as she nimbly weaves between the towering bamboo branches. Her wide sleeves flutter like butterfly wings as she moves swiftly, her skirt carving graceful arcs through the mist. A series of rapidly alternating close-ups follows: the tip of her shoe, delicately embroidered with cloud patterns, lightly steps on a dewy bamboo leaf, which trembles only slightly before returning to its composure; a strand of hair brushes across her focused and serene profile, revealing her perfectly defined jawline; her slender fingers form an elegant sword-like gesture in the air, steadily and calmly directing her body. The entire sequence is seamless, the camera language sharp and precise, capturing every moment of her ethereal, ethereal beauty. Amidst the cool tones of light and the lingering mist, an ethereal atmosphere is created, imbued with both martial arts power and poetic imagery.", + "seed": 17759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_4/index_5716_seed_17759_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05906847845762968, + "vlm_logprob": -0.05906847845762968, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.05906847845762968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11746488511562347 + ], + "opposite_logprob_tokens": [ + -2.242464780807495 + ], + "logprob_reward": -0.11746488511562347, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11746488511562347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12800565361976624, + 0.12490765005350113, + 0.06323345750570297, + 0.09167307615280151, + 0.11612801253795624, + 0.04650510102510452, + 0.2104499787092209, + 0.07551483809947968, + 0.08818119019269943, + 0.05540105700492859 + ], + "credit_map": [ + 0.01264592632651329, + 0.014997838996350765, + 0.017899328842759132, + 0.023069897666573524, + 0.011990328319370747, + 0.010294702835381031, + 0.021101098507642746, + 0.016006534919142723, + 0.012312364764511585, + 0.014789222739636898, + 0.016597425565123558, + 0.01863500103354454, + 0.015204085037112236, + 0.01401759684085846, + 0.018902938812971115, + 0.014449015259742737, + 0.008782163262367249, + 0.006584147457033396, + 0.007834810763597488, + 0.007722815498709679, + 0.0058732531033456326, + 0.007399281486868858, + 0.011058030650019646, + 0.007978958077728748, + 0.00924496166408062, + 0.009755037724971771, + 0.009212939068675041, + 0.014177938923239708, + 0.013969683088362217, + 0.01183968037366867, + 0.014726979658007622, + 0.008745855651795864, + 0.01483374834060669, + 0.011686444282531738, + 0.011477111838757992, + 0.016993407160043716, + 0.017763003706932068, + 0.01625729352235794, + 0.013059117831289768, + 0.01405788492411375, + 0.005339085124433041, + 0.005335690453648567, + 0.00570497615262866, + 0.005175808444619179, + 0.006250835955142975, + 0.006451415363699198, + 0.005615314468741417, + 0.006631977390497923, + 0.015570063143968582, + 0.012546763755381107, + 0.06900467723608017, + 0.009040778502821922, + 0.009155732579529285, + 0.07853808254003525, + 0.007297208998352289, + 0.00929667241871357, + 0.007945000194013119, + 0.0053544496186077595, + 0.008534325286746025, + 0.007034428883343935, + 0.013596275821328163, + 0.013429860584437847, + 0.011181080713868141, + 0.008439422585070133, + 0.00935292523354292, + 0.009437140077352524, + 0.011031483300030231, + 0.008219893090426922, + 0.0164943840354681, + 0.01642880029976368, + 0.009930884465575218, + 0.007285673171281815, + 0.007800825871527195, + 0.00488112960010767, + 0.006511338520795107, + 0.011365577578544617, + 0.0032223593443632126, + 0.006275486201047897, + 0.008813244290649891, + 0.006531094666570425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.11746488511562347 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014220803044736385 + ], + "opposite_logprob_tokens": [ + -4.264220714569092 + ], + "logprob_reward": -0.014220803044736385, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014220803044736385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0508771650493145, + 0.042781051248311996, + 0.04071368649601936, + 0.04497595876455307, + 0.05765707045793533, + 0.04654866084456444, + 0.5052230358123779, + 0.1390032023191452, + 0.03716135770082474, + 0.03505885601043701 + ], + "credit_map": [ + 0.004854653961956501, + 0.004354164935648441, + 0.007348271086812019, + 0.009157173335552216, + 0.006289473734796047, + 0.004438831936568022, + 0.008553717285394669, + 0.005880876909941435, + 0.0045034270733594894, + 0.005180032458156347, + 0.005973574239760637, + 0.008417571894824505, + 0.0033681688364595175, + 0.004200444091111422, + 0.005951684433966875, + 0.005186148453503847, + 0.0044401478953659534, + 0.004171045962721109, + 0.006099601276218891, + 0.006948462687432766, + 0.0033666000235825777, + 0.0038964259438216686, + 0.005337871611118317, + 0.006453527137637138, + 0.003285552142187953, + 0.003569670720025897, + 0.00660164887085557, + 0.0066814846359193325, + 0.0058255973272025585, + 0.008604432456195354, + 0.006289212964475155, + 0.004118357785046101, + 0.0037381364963948727, + 0.006570865400135517, + 0.007389503996819258, + 0.010673749260604382, + 0.008412393741309643, + 0.004754689987748861, + 0.00824563018977642, + 0.007872100919485092, + 0.005443709436804056, + 0.0032748416997492313, + 0.008559522219002247, + 0.004329268354922533, + 0.00571046955883503, + 0.006834457162767649, + 0.00606793025508523, + 0.006328456103801727, + 0.04223273694515228, + 0.03066324256360531, + 0.16213466227054596, + 0.022271722555160522, + 0.021646764129400253, + 0.18847298622131348, + 0.0182312224060297, + 0.019569668918848038, + 0.014002721756696701, + 0.009335693903267384, + 0.032963745296001434, + 0.010297310538589954, + 0.04119672253727913, + 0.01363812293857336, + 0.011762317270040512, + 0.005806553643196821, + 0.003256630850955844, + 0.0043263002298772335, + 0.007011899258941412, + 0.0041041625663638115, + 0.003702678019180894, + 0.003348412923514843, + 0.005749039817601442, + 0.005662232171744108, + 0.004354254342615604, + 0.004429315682500601, + 0.0042333099991083145, + 0.006276300176978111, + 0.002800338203087449, + 0.005410837009549141, + 0.003757044905796647, + 0.0037974554579705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.014220803044736385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890444815158844 + ], + "opposite_logprob_tokens": [ + -2.4640445709228516 + ], + "logprob_reward": -0.0890444815158844, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890444815158844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03670324757695198, + 0.07378412783145905, + 0.04849785566329956, + 0.03631014749407768, + 0.025227991864085197, + 0.0376371331512928, + 0.0791250690817833, + 0.5961180925369263, + 0.03547004982829094, + 0.03112631104886532 + ], + "credit_map": [ + 0.002865962916985154, + 0.0035947351716458797, + 0.00632186233997345, + 0.00491334730759263, + 0.0056497990153729916, + 0.0034065493382513523, + 0.006566282827407122, + 0.0033847084268927574, + 0.00731481472030282, + 0.008636267855763435, + 0.009060054086148739, + 0.011660823598504066, + 0.008731904439628124, + 0.008613776415586472, + 0.011157373897731304, + 0.008609116077423096, + 0.005128452088683844, + 0.003751041367650032, + 0.0063182213343679905, + 0.005729284603148699, + 0.0053148348815739155, + 0.006011098623275757, + 0.009275894612073898, + 0.006969029083848, + 0.003913366701453924, + 0.0028551663272082806, + 0.007468667812645435, + 0.003547267522662878, + 0.0037956612650305033, + 0.008195193484425545, + 0.003414265811443329, + 0.0031205571722239256, + 0.0028471110854297876, + 0.002977384952828288, + 0.0031605386175215244, + 0.004002042580395937, + 0.00388842960819602, + 0.002756314119324088, + 0.0026715572457760572, + 0.0029246131889522076, + 0.0044645327143371105, + 0.003902347758412361, + 0.006748005282133818, + 0.004214930813759565, + 0.004314863122999668, + 0.0042717838659882545, + 0.004705058876425028, + 0.005015611182898283, + 0.007201148197054863, + 0.005063351709395647, + 0.022544072940945625, + 0.006021757144480944, + 0.005212393123656511, + 0.02358315885066986, + 0.005006361287087202, + 0.004492828622460365, + 0.06513530761003494, + 0.037037670612335205, + 0.1528950035572052, + 0.036553628742694855, + 0.178669273853302, + 0.057556167244911194, + 0.03916754201054573, + 0.02910349704325199, + 0.0036068514455109835, + 0.00442842161282897, + 0.004234475549310446, + 0.004460232798010111, + 0.0029562818817794323, + 0.0025745038874447346, + 0.006911416072398424, + 0.006297866348177195, + 0.0036360344383865595, + 0.004617040045559406, + 0.004283612128347158, + 0.004429538268595934, + 0.002479902934283018, + 0.00377332279458642, + 0.003868918167427182, + 0.00403794227167964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.0890444815158844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206497922539711 + ], + "opposite_logprob_tokens": [ + -2.8120648860931396 + ], + "logprob_reward": -0.06206497922539711, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206497922539711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10294944047927856, + 0.15787947177886963, + 0.06410658359527588, + 0.04514386132359505, + 0.03123036026954651, + 0.020274333655834198, + 0.38099032640457153, + 0.15871085226535797, + 0.01705222949385643, + 0.021662510931491852 + ], + "credit_map": [ + 0.010762592777609825, + 0.009108785539865494, + 0.017814485356211662, + 0.020175226032733917, + 0.008126966655254364, + 0.00896933488547802, + 0.01777397096157074, + 0.010218068026006222, + 0.015443561598658562, + 0.018199212849140167, + 0.019802892580628395, + 0.026467537507414818, + 0.016858315095305443, + 0.01666305400431156, + 0.026072664186358452, + 0.018372219055891037, + 0.006483258679509163, + 0.005475870333611965, + 0.0072364285588264465, + 0.008843466639518738, + 0.006513445638120174, + 0.0074398391880095005, + 0.011712313629686832, + 0.010401951149106026, + 0.004427074454724789, + 0.005037013906985521, + 0.0060928515158593655, + 0.011726250872015953, + 0.004051155876368284, + 0.005516947712749243, + 0.004019075073301792, + 0.00427348492667079, + 0.002397108357399702, + 0.003688698634505272, + 0.004321932327002287, + 0.006331100128591061, + 0.003919923212379217, + 0.002311923773959279, + 0.004123047459870577, + 0.004136622883379459, + 0.0023510882165282965, + 0.0014264930505305529, + 0.0037140962667763233, + 0.001634247717447579, + 0.002562018809840083, + 0.0035664320457726717, + 0.0024667612742632627, + 0.002553194761276245, + 0.032248467206954956, + 0.02362627536058426, + 0.12142733484506607, + 0.01673118583858013, + 0.016327936202287674, + 0.14340168237686157, + 0.0130078149959445, + 0.014219637960195541, + 0.01748412288725376, + 0.010311335325241089, + 0.04018406942486763, + 0.009817088954150677, + 0.04681304469704628, + 0.015027280896902084, + 0.010006633587181568, + 0.00906726997345686, + 0.001614506240002811, + 0.002574445214122534, + 0.004418041557073593, + 0.002484781201928854, + 0.0017806331161409616, + 0.0013172798790037632, + 0.0015638885088264942, + 0.0012986521469429135, + 0.002589169191196561, + 0.002800490939989686, + 0.002723070327192545, + 0.004209514241665602, + 0.001396911684423685, + 0.0021608101669698954, + 0.003015556838363409, + 0.002766986843198538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.06206497922539711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547243386507034 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547243386507034, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547243386507034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1357266753911972, + 0.21008367836475372, + 0.19254541397094727, + 0.06606949865818024, + 0.04016745090484619, + 0.04483662545681, + 0.06206128001213074, + 0.13224217295646667, + 0.034587420523166656, + 0.08167978376150131 + ], + "credit_map": [ + 0.011081267148256302, + 0.012476465664803982, + 0.021026892587542534, + 0.02088186703622341, + 0.017132334411144257, + 0.013486536219716072, + 0.025447776541113853, + 0.014193542301654816, + 0.019155234098434448, + 0.031031109392642975, + 0.026227137073874474, + 0.03268555551767349, + 0.024521037936210632, + 0.022499168291687965, + 0.030218616127967834, + 0.023745838552713394, + 0.014246064238250256, + 0.013367428444325924, + 0.025733767077326775, + 0.025556977838277817, + 0.02227315865457058, + 0.026130398735404015, + 0.037661533802747726, + 0.02757609635591507, + 0.007763526868075132, + 0.006995128933340311, + 0.006284775212407112, + 0.009943061508238316, + 0.00846380926668644, + 0.007042654789984226, + 0.012136014178395271, + 0.007440533488988876, + 0.004181987140327692, + 0.003413435537368059, + 0.005343206226825714, + 0.006010344717651606, + 0.004139088559895754, + 0.007304731756448746, + 0.0038316671270877123, + 0.005942992866039276, + 0.006150557193905115, + 0.004865884315222502, + 0.008756226859986782, + 0.005994708742946386, + 0.004957257304340601, + 0.005852339323610067, + 0.0037128308322280645, + 0.004546825308352709, + 0.005121821071952581, + 0.006569582503288984, + 0.011412198655307293, + 0.010033228434622288, + 0.00465019978582859, + 0.013506569899618626, + 0.006407366134226322, + 0.004360314458608627, + 0.0136447474360466, + 0.009754229336977005, + 0.03192727640271187, + 0.009514795616269112, + 0.03502173349261284, + 0.013252654112875462, + 0.009985839016735554, + 0.009140904992818832, + 0.003345061093568802, + 0.0038594165816903114, + 0.0038153815548866987, + 0.004322728607803583, + 0.005098685622215271, + 0.005585918202996254, + 0.005343500524759293, + 0.003216733690351248, + 0.008528903126716614, + 0.008749132975935936, + 0.010227886028587818, + 0.019128387793898582, + 0.005201913882046938, + 0.00717172771692276, + 0.012373695150017738, + 0.010298140347003937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.012547243386507034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, wide-angle shot, shot from above, slowly pans upward through a lush bamboo forest shrouded in morning mist. Soft, diffused light penetrates the mist and dense bamboo foliage, casting visible beams of light in the air. The camera finally settles on a woman in a traditional Chinese costume, dressed in an elegant aquamarine silk robe. Her long, cascading black hair is held together only partially by a simple jade hairpin. Her face is clear and beautiful, her eyes focused and ethereal. With a light touch of the ground, her entire body rises as if weightless. The camera instantly switches to a fast, dynamic tracking shot, following her graceful figure as she nimbly weaves between the towering bamboo branches. Her wide sleeves flutter like butterfly wings as she moves swiftly, her skirt carving graceful arcs through the mist. A series of rapidly alternating close-ups follows: the tip of her shoe, delicately embroidered with cloud patterns, lightly steps on a dewy bamboo leaf, which trembles only slightly before returning to its composure; a strand of hair brushes across her focused and serene profile, revealing her perfectly defined jawline; her slender fingers form an elegant sword-like gesture in the air, steadily and calmly directing her body. The entire sequence is seamless, the camera language sharp and precise, capturing every moment of her ethereal, ethereal beauty. Amidst the cool tones of light and the lingering mist, an ethereal atmosphere is created, imbued with both martial arts power and poetic imagery.", + "seed": 17760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_4/index_5716_seed_17760_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04631451666355133, + "vlm_logprob": -0.04631451666355133, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04631451666355133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1343500316143036 + ], + "opposite_logprob_tokens": [ + -2.134350061416626 + ], + "logprob_reward": -0.1343500316143036, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1343500316143036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1673036366701126, + 0.07690437883138657, + 0.029852956533432007, + 0.04115753248333931, + 0.36148956418037415, + 0.15097872912883759, + 0.06784064322710037, + 0.03793007507920265, + 0.023224851116538048, + 0.04331771284341812 + ], + "credit_map": [ + 0.01486151572316885, + 0.026778612285852432, + 0.021357174962759018, + 0.015706170350313187, + 0.030803510919213295, + 0.02324552834033966, + 0.017746267840266228, + 0.016804855316877365, + 0.009572356007993221, + 0.011617438867688179, + 0.008183127269148827, + 0.01182743813842535, + 0.012140157632529736, + 0.010730098001658916, + 0.00608658604323864, + 0.006747175939381123, + 0.0028923009522259235, + 0.002786702709272504, + 0.0037564297672361135, + 0.0027736504562199116, + 0.0032065771520137787, + 0.003908208571374416, + 0.0033937166444957256, + 0.007135370280593634, + 0.004045053850859404, + 0.009630713611841202, + 0.0068347700871527195, + 0.006541281472891569, + 0.004159317351877689, + 0.0021762519609183073, + 0.004345097113400698, + 0.003425046568736434, + 0.012251095846295357, + 0.12593932449817657, + 0.010926637798547745, + 0.08257314562797546, + 0.014442270621657372, + 0.03716162219643593, + 0.05604816973209381, + 0.02214730717241764, + 0.006753642577677965, + 0.015610569156706333, + 0.006704484112560749, + 0.04849020764231682, + 0.006656985729932785, + 0.010433130897581577, + 0.0472005270421505, + 0.0091291768476367, + 0.0028037710580974817, + 0.021306758746504784, + 0.0025584117975085974, + 0.020820824429392815, + 0.004042689222842455, + 0.006433735601603985, + 0.0037924584466964006, + 0.0060819885693490505, + 0.0039262110367417336, + 0.006123182829469442, + 0.003846795530989766, + 0.00507426867261529, + 0.004641185514628887, + 0.005682843271642923, + 0.004763349890708923, + 0.003872241824865341, + 0.0028672514017671347, + 0.004660253878682852, + 0.00262849940918386, + 0.0032246478367596865, + 0.002273974474519491, + 0.002264975570142269, + 0.003144313581287861, + 0.002160935662686825, + 0.008173885755240917, + 0.004265924915671349, + 0.00572711369022727, + 0.004630018956959248, + 0.005085047334432602, + 0.007625462021678686, + 0.0032544643618166447, + 0.004555794410407543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.1343500316143036 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800177067518234 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.009800177067518234, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800177067518234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07649224251508713, + 0.05605543032288551, + 0.05935795232653618, + 0.053402259945869446, + 0.32037076354026794, + 0.1920202523469925, + 0.05259038880467415, + 0.028898553922772408, + 0.040611471980810165, + 0.1202007383108139 + ], + "credit_map": [ + 0.005316874943673611, + 0.015071374364197254, + 0.00815735012292862, + 0.008046621456742287, + 0.016235249117016792, + 0.007975012995302677, + 0.007988144643604755, + 0.007701612543314695, + 0.007122533395886421, + 0.007774723693728447, + 0.0068657356314361095, + 0.009777599014341831, + 0.007651707623153925, + 0.006847023032605648, + 0.005309005733579397, + 0.004707101732492447, + 0.005524241365492344, + 0.005295843817293644, + 0.010555705055594444, + 0.006375742144882679, + 0.00592130608856678, + 0.007259187288582325, + 0.008836129680275917, + 0.009589795023202896, + 0.0062995171174407005, + 0.008389334194362164, + 0.010190047323703766, + 0.0068841115571558475, + 0.0060385181568562984, + 0.0036754063330590725, + 0.008336047641932964, + 0.0035892766900360584, + 0.011648854240775108, + 0.1213231086730957, + 0.008165060542523861, + 0.04614727571606636, + 0.01765538565814495, + 0.01175038330256939, + 0.09229274094104767, + 0.011387963779270649, + 0.010755648836493492, + 0.03335943445563316, + 0.009825707413256168, + 0.06982647627592087, + 0.00932107400149107, + 0.00804661400616169, + 0.042071059346199036, + 0.008814232423901558, + 0.005279624834656715, + 0.006604490336030722, + 0.004132307134568691, + 0.005542593076825142, + 0.006225690245628357, + 0.013577098958194256, + 0.00326896901242435, + 0.00795961543917656, + 0.0025559114292263985, + 0.003220323706045747, + 0.0023711316753178835, + 0.004115401301532984, + 0.0030952049419283867, + 0.005047947634011507, + 0.0044736480340361595, + 0.004018985666334629, + 0.0043815020471811295, + 0.007216171827167273, + 0.006628799717873335, + 0.005036122631281614, + 0.004548748955130577, + 0.0032224298920482397, + 0.006386862136423588, + 0.003190837102010846, + 0.016125796362757683, + 0.016785118728876114, + 0.015604328364133835, + 0.015664663165807724, + 0.02300921268761158, + 0.011908095329999924, + 0.00978916883468628, + 0.011314354836940765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.009800177067518234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048655346035957336 + ], + "opposite_logprob_tokens": [ + -3.0486552715301514 + ], + "logprob_reward": -0.048655346035957336, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048655346035957336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23572410643100739, + 0.12137007713317871, + 0.09992758929729462, + 0.04228810593485832, + 0.07647919654846191, + 0.2403646558523178, + 0.02718471549451351, + 0.03633486479520798, + 0.03597218543291092, + 0.08435448259115219 + ], + "credit_map": [ + 0.008252468891441822, + 0.0854022428393364, + 0.008858547545969486, + 0.009238462895154953, + 0.08213653415441513, + 0.017190098762512207, + 0.011734331957995892, + 0.012911430560052395, + 0.011624136008322239, + 0.020969590172171593, + 0.008862989954650402, + 0.018180187791585922, + 0.021185513585805893, + 0.0176800899207592, + 0.013009469024837017, + 0.009858105331659317, + 0.008524765260517597, + 0.009996659122407436, + 0.012914106249809265, + 0.009682868607342243, + 0.01067525427788496, + 0.009629957377910614, + 0.013764817267656326, + 0.024739159271121025, + 0.005728133954107761, + 0.005150271113961935, + 0.007649065461009741, + 0.004019470419734716, + 0.005593332927674055, + 0.004494678229093552, + 0.006286080460995436, + 0.0033670691773295403, + 0.004048939794301987, + 0.02535121515393257, + 0.004996899515390396, + 0.013492401689291, + 0.004595430102199316, + 0.0034335264936089516, + 0.01673063263297081, + 0.0038301576860249043, + 0.0066313971765339375, + 0.02704768069088459, + 0.009817526675760746, + 0.08558224141597748, + 0.018239600583910942, + 0.006007558200508356, + 0.07958371937274933, + 0.007454922888427973, + 0.0029299466405063868, + 0.004412981681525707, + 0.0031571409199386835, + 0.004829965066164732, + 0.0033256623428314924, + 0.002460578689351678, + 0.0027145189233124256, + 0.0033539235591888428, + 0.0027884389273822308, + 0.004693703260272741, + 0.0030608135275542736, + 0.005795192904770374, + 0.004817288368940353, + 0.005111596547067165, + 0.005703633651137352, + 0.004364197142422199, + 0.0056390524841845036, + 0.005734650418162346, + 0.005032447166740894, + 0.005365196615457535, + 0.00324233528226614, + 0.003140445798635483, + 0.004921493120491505, + 0.0028965631499886513, + 0.01054384559392929, + 0.013333468697965145, + 0.01334423292428255, + 0.013290032744407654, + 0.007624628487974405, + 0.007794646080583334, + 0.008109977468848228, + 0.010313652455806732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.048655346035957336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09812620282173157, + 0.04473942145705223, + 0.047974735498428345, + 0.01926865428686142, + 0.04635218530893326, + 0.4708788990974426, + 0.09804288297891617, + 0.0542953796684742, + 0.04903801530599594, + 0.07128367573022842 + ], + "credit_map": [ + 0.005975034087896347, + 0.026238251477479935, + 0.007550162263214588, + 0.010461450554430485, + 0.027876760810613632, + 0.007845135405659676, + 0.0055650705471634865, + 0.00661433394998312, + 0.004613195080310106, + 0.004175527952611446, + 0.004524413961917162, + 0.010257012210786343, + 0.004936400335282087, + 0.007136960979551077, + 0.005336961708962917, + 0.0037589494604617357, + 0.0036952963564544916, + 0.004263606388121843, + 0.009383833967149258, + 0.00530574657022953, + 0.005418018903583288, + 0.007454127538949251, + 0.006108166184276342, + 0.006345940288156271, + 0.002264626557007432, + 0.002643765415996313, + 0.0025833791587501764, + 0.002578588668256998, + 0.002646190579980612, + 0.001353742554783821, + 0.0031633558683097363, + 0.0020350047852844, + 0.0020100341644138098, + 0.005480648018419743, + 0.002280167071148753, + 0.009051689878106117, + 0.002653643488883972, + 0.006491003092378378, + 0.013756442815065384, + 0.004628559108823538, + 0.009760561399161816, + 0.05389833450317383, + 0.009756987914443016, + 0.19316159188747406, + 0.017886843532323837, + 0.010434750467538834, + 0.16210946440696716, + 0.01387035008519888, + 0.006648999638855457, + 0.01701941341161728, + 0.005837793927639723, + 0.014256964437663555, + 0.01079504657536745, + 0.025510944426059723, + 0.005793971475213766, + 0.012179750017821789, + 0.005009528249502182, + 0.0056192297488451, + 0.004068652633577585, + 0.006508498918265104, + 0.005882590543478727, + 0.01202393602579832, + 0.00533787626773119, + 0.0098450668156147, + 0.0055837482213974, + 0.013168836943805218, + 0.006936986465007067, + 0.006025786977261305, + 0.003774399636313319, + 0.003080375026911497, + 0.006312992889434099, + 0.004154889844357967, + 0.00884433463215828, + 0.00909302569925785, + 0.010448413901031017, + 0.011294757016003132, + 0.008629055693745613, + 0.007604559417814016, + 0.005842928774654865, + 0.009526599198579788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.02058546058833599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a woman wearing an elegant aquamarine silk robe with long black hair held by a simple jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman rise weightlessly from the ground and then weave nimbly between the bamboo branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include a sequence of rapidly alternating close-ups showing her shoe stepping on a dewy leaf, her hair brushing her face, and her fingers forming a sword-like gesture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low wide-angle shot panning upward, then switch to a fast dynamic tracking shot, and finally execute rapidly alternating close-ups?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181568011641502 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181568011641502, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181568011641502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2224586308002472, + 0.1169440969824791, + 0.14906729757785797, + 0.03557467460632324, + 0.06518137454986572, + 0.1886463612318039, + 0.07463083416223526, + 0.042911428958177567, + 0.04072338715195656, + 0.0638618990778923 + ], + "credit_map": [ + 0.013358300551772118, + 0.07631406933069229, + 0.015168916434049606, + 0.014954323880374432, + 0.07248804718255997, + 0.011515519581735134, + 0.010126215405762196, + 0.008533229120075703, + 0.007980205118656158, + 0.020348655059933662, + 0.011379801668226719, + 0.022135745733976364, + 0.021148882806301117, + 0.012846008874475956, + 0.011955655179917812, + 0.009149148128926754, + 0.013771122321486473, + 0.015302990563213825, + 0.019127896055579185, + 0.014395184814929962, + 0.01699262671172619, + 0.014095901511609554, + 0.01927373930811882, + 0.03610783815383911, + 0.0037248770240694284, + 0.004286457784473896, + 0.006254200823605061, + 0.005185261368751526, + 0.004390583839267492, + 0.0031256915535777807, + 0.005364211276173592, + 0.003243392100557685, + 0.0035496584605425596, + 0.014639750123023987, + 0.002992053749039769, + 0.019172942265868187, + 0.0033339534420520067, + 0.00972906593233347, + 0.0057906401343643665, + 0.005973309278488159, + 0.005733618978410959, + 0.04184220731258392, + 0.004606351722031832, + 0.07906650751829147, + 0.0057108779437839985, + 0.004763829987496138, + 0.041145920753479004, + 0.005777034442871809, + 0.007640156429260969, + 0.013752346858382225, + 0.007755542639642954, + 0.014090311713516712, + 0.006269850302487612, + 0.012441893108189106, + 0.004642263054847717, + 0.00803847424685955, + 0.002374005503952503, + 0.0041761985048651695, + 0.003214586293324828, + 0.005838099401444197, + 0.008236708119511604, + 0.006129998713731766, + 0.008202601224184036, + 0.004739230964332819, + 0.004822115879505873, + 0.006536821834743023, + 0.007113292347639799, + 0.0049013979732990265, + 0.004729162901639938, + 0.003910972271114588, + 0.005864719394594431, + 0.0028449094388633966, + 0.00964299589395523, + 0.005050606559962034, + 0.010823797434568405, + 0.007755472790449858, + 0.008172862231731415, + 0.008175242692232132, + 0.007490483578294516, + 0.006750438828021288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.018181568011641502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_5/index_tensor([8683, 8683])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_5/index_tensor([8683, 8683])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..41d1fb838a75d5695e9d2b1851e53554bd0fada7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_5/index_tensor([8683, 8683])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, high-speed tracking shot, as if shot from an accompanying aircraft, captures the flight of a vibrant, candy-apple-red biplane, its smooth surface reflecting brilliant sunlight. The camera maintains an eye-level perspective in the single, open cockpit, revealing the faint outline of the pilot. The aircraft's propellers spin at such a rapid rate that the nose becomes a mesmerizingly transparent blur. Against the expansive backdrop of a flawless, deep blue sky, the biplane playfully executes daring maneuvers, weaving between and through enormous, fluffy, white cumulus clouds that appear as soft as cotton. It executes a swift and graceful barrel roll, its wings swiveling 360 degrees with thrilling precision, before sharply pulling up in a perfect, gravity-defying loop. The entire scene is bathed in crisp, high-key sunlight, creating brilliant highlights on the aircraft's fuselage and enlivening the saturated primary colors of red, white, and blue, perfectly capturing the festive atmosphere of pure, unadulterated joy in flight.", + "seed": 5587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_5/index_8683_seed_5587_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20498520126566291, + "vlm_logprob": -0.20498520126566291, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.20498520126566291, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13050180673599243 + ], + "opposite_logprob_tokens": [ + -2.1305017471313477 + ], + "logprob_reward": -0.13050180673599243, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13050180673599243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2904781699180603, + 0.06495414674282074, + 0.3226933479309082, + 0.07539691030979156, + 0.0709613561630249, + 0.019146552309393883, + 0.045080725103616714, + 0.03467443212866783, + 0.017703264951705933, + 0.05891110748052597 + ], + "credit_map": [ + 0.03472498431801796, + 0.03007357567548752, + 0.05261360481381416, + 0.07338114082813263, + 0.03181120380759239, + 0.03541899099946022, + 0.018566103652119637, + 0.013888588175177574, + 0.008641660213470459, + 0.00924059096723795, + 0.010129190050065517, + 0.014234614558517933, + 0.006150368135422468, + 0.009963814169168472, + 0.004121978767216206, + 0.0024719280190765858, + 0.028705699369311333, + 0.0982322171330452, + 0.027575545012950897, + 0.018565194681286812, + 0.10150876641273499, + 0.028931930661201477, + 0.011793872341513634, + 0.007380126044154167, + 0.008553632535040379, + 0.016495438292622566, + 0.009232141077518463, + 0.009999246336519718, + 0.015539231710135937, + 0.007384902331978083, + 0.0045186858624219894, + 0.0036736330948770046, + 0.007231722120195627, + 0.0075683435425162315, + 0.009139851666986942, + 0.012159815058112144, + 0.01314580999314785, + 0.01028081588447094, + 0.006870997603982687, + 0.004564002621918917, + 0.002104953397065401, + 0.0028763865120708942, + 0.002159950789064169, + 0.0023346508387476206, + 0.003896759357303381, + 0.0021401119884103537, + 0.0019002200569957495, + 0.0017335187876597047, + 0.008723217993974686, + 0.006359911523759365, + 0.005180403124541044, + 0.007532166782766581, + 0.006382747087627649, + 0.0039298743940889835, + 0.003365671494975686, + 0.0036067338660359383, + 0.003964023664593697, + 0.004557568579912186, + 0.004492452833801508, + 0.0048617408610880375, + 0.00515027204528451, + 0.006445921026170254, + 0.002864109119400382, + 0.0023383446969091892, + 0.0020141664426773787, + 0.002366994507610798, + 0.0024433238431811333, + 0.002875340636819601, + 0.002383304527029395, + 0.0029776364099234343, + 0.0014970045303925872, + 0.0011454937048256397, + 0.007984974421560764, + 0.01289514359086752, + 0.008038653060793877, + 0.004748299717903137, + 0.010817677713930607, + 0.006454178132116795, + 0.004290325567126274, + 0.00368185481056571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.13050180673599243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207875907421112 + ], + "opposite_logprob_tokens": [ + -2.8120787143707275 + ], + "logprob_reward": -0.06207875907421112, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207875907421112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3391459286212921, + 0.07360754907131195, + 0.16235999763011932, + 0.04289596155285835, + 0.1493004411458969, + 0.03743372857570648, + 0.03639841452240944, + 0.06230747699737549, + 0.029153695330023766, + 0.06739675998687744 + ], + "credit_map": [ + 0.045770928263664246, + 0.04185941070318222, + 0.05193215608596802, + 0.06629391759634018, + 0.04864346981048584, + 0.040081433951854706, + 0.024991514161229134, + 0.01957312598824501, + 0.008924892172217369, + 0.010133687406778336, + 0.011690428480505943, + 0.017401453107595444, + 0.006074205040931702, + 0.010595273226499557, + 0.005477182101458311, + 0.003310428000986576, + 0.012134449556469917, + 0.03830559924244881, + 0.014394012279808521, + 0.018513621762394905, + 0.04260240122675896, + 0.018742943182587624, + 0.01116623543202877, + 0.006500738672912121, + 0.005398349370807409, + 0.006633029785007238, + 0.004968344233930111, + 0.004765461664646864, + 0.005621753633022308, + 0.0068373903632164, + 0.005266473162919283, + 0.003405158407986164, + 0.015382355079054832, + 0.01673155464231968, + 0.01816430315375328, + 0.02080339752137661, + 0.024458637461066246, + 0.02073199674487114, + 0.019477667286992073, + 0.01355052087455988, + 0.00571522768586874, + 0.005872839596122503, + 0.003815604606643319, + 0.006399764213711023, + 0.004884316585958004, + 0.0047730035148561, + 0.0037162750959396362, + 0.002256697742268443, + 0.005863791797310114, + 0.003266754560172558, + 0.004314708523452282, + 0.009198109619319439, + 0.00532993720844388, + 0.0034545643720775843, + 0.002715046051889658, + 0.0022555035538971424, + 0.007819649763405323, + 0.009005535393953323, + 0.007659073919057846, + 0.010440428741276264, + 0.007518839556723833, + 0.008565382100641727, + 0.006807001773267984, + 0.004491565749049187, + 0.006373122800141573, + 0.005792695563286543, + 0.003531597787514329, + 0.0036421508993953466, + 0.003137757536023855, + 0.003143256064504385, + 0.0017762241186574101, + 0.0017568912589922547, + 0.008286923170089722, + 0.017226533964276314, + 0.009385088458657265, + 0.005083953961730003, + 0.01482270285487175, + 0.00652523385360837, + 0.0035095661878585815, + 0.0025567582342773676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.06207875907421112 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761404037475586 + ], + "opposite_logprob_tokens": [ + -0.8261404037475586 + ], + "logprob_reward": -0.5761404037475586, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761404037475586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3191893994808197, + 0.09050434082746506, + 0.11472790688276291, + 0.16904044151306152, + 0.14922353625297546, + 0.015291200019419193, + 0.05480392277240753, + 0.04902902618050575, + 0.012709459289908409, + 0.025480782613158226 + ], + "credit_map": [ + 0.035184841603040695, + 0.03771352022886276, + 0.04571116715669632, + 0.0605909489095211, + 0.0566507987678051, + 0.046024177223443985, + 0.02041555754840374, + 0.016898365691304207, + 0.009695012122392654, + 0.009982476942241192, + 0.011925850994884968, + 0.015534711070358753, + 0.012330688536167145, + 0.012468106113374233, + 0.011087327264249325, + 0.007480161730200052, + 0.010007338598370552, + 0.028113219887018204, + 0.010052799247205257, + 0.009230644442141056, + 0.029991351068019867, + 0.013546221889555454, + 0.008326779119670391, + 0.005459554493427277, + 0.018743367865681648, + 0.036559924483299255, + 0.02140682004392147, + 0.025215784087777138, + 0.035042159259319305, + 0.016467267647385597, + 0.009004034101963043, + 0.006601090542972088, + 0.01292031817138195, + 0.01446679700165987, + 0.018734849989414215, + 0.02031761407852173, + 0.030427629128098488, + 0.02001773752272129, + 0.021118177101016045, + 0.01122040580958128, + 0.00209203758276999, + 0.002567555056884885, + 0.0014304636279121041, + 0.0021202198695391417, + 0.001973100006580353, + 0.0018370525212958455, + 0.002012415323406458, + 0.0012583564966917038, + 0.008903460577130318, + 0.007272173184901476, + 0.0057224091142416, + 0.0075445277616381645, + 0.010318984277546406, + 0.006181395146995783, + 0.004915740340948105, + 0.003945230506360531, + 0.007911358959972858, + 0.006667653098702431, + 0.005522176623344421, + 0.007981345988810062, + 0.006091511342674494, + 0.007121141068637371, + 0.004373427480459213, + 0.0033604097552597523, + 0.0009372102213092148, + 0.0010771524393931031, + 0.0015234355814754963, + 0.00209622154943645, + 0.0016190322348847985, + 0.0022108762059360743, + 0.0017836404731497169, + 0.0014618909917771816, + 0.0026714836712926626, + 0.0054810610599815845, + 0.003608131082728505, + 0.002370017347857356, + 0.005052751395851374, + 0.0027644464280456305, + 0.002041586209088564, + 0.0014913036720827222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.5761404037475586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252111554145813 + ], + "opposite_logprob_tokens": [ + -1.502111554145813 + ], + "logprob_reward": -0.252111554145813, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252111554145813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47987493872642517, + 0.07879061996936798, + 0.06420357525348663, + 0.1343001276254654, + 0.032263100147247314, + 0.020093996077775955, + 0.06459607928991318, + 0.025324076414108276, + 0.03247985988855362, + 0.0680735856294632 + ], + "credit_map": [ + 0.06776552647352219, + 0.05616997927427292, + 0.07344464957714081, + 0.09205299615859985, + 0.066435307264328, + 0.06543444842100143, + 0.03197411820292473, + 0.026597950607538223, + 0.008036975748836994, + 0.010626349598169327, + 0.009760274551808834, + 0.011246644891798496, + 0.013922006823122501, + 0.009789288975298405, + 0.008870234712958336, + 0.006538849789649248, + 0.004856206476688385, + 0.013401782140135765, + 0.006318757310509682, + 0.006692160852253437, + 0.014699534513056278, + 0.0068997228518128395, + 0.007257685996592045, + 0.0040777334943413734, + 0.016695303842425346, + 0.03227247670292854, + 0.015679314732551575, + 0.01623367704451084, + 0.028253551572561264, + 0.012938636355102062, + 0.006680286955088377, + 0.005546902772039175, + 0.00320310122333467, + 0.004188954830169678, + 0.0042625246569514275, + 0.004795715678483248, + 0.005932840518653393, + 0.004852007608860731, + 0.0029507498256862164, + 0.002077208599075675, + 0.0028351638466119766, + 0.0035476540215313435, + 0.0021653531584888697, + 0.002393333939835429, + 0.002978438278660178, + 0.0028860734310001135, + 0.0019586263224482536, + 0.0013293555239215493, + 0.01046371553093195, + 0.009377236478030682, + 0.006786611396819353, + 0.008931238204240799, + 0.012332156300544739, + 0.00703647593036294, + 0.005222452338784933, + 0.004446201957762241, + 0.0040523395873606205, + 0.0038502737879753113, + 0.0026796141173690557, + 0.003227722831070423, + 0.00315084564499557, + 0.0034986333921551704, + 0.0024473683442920446, + 0.0024172780103981495, + 0.007717190310359001, + 0.006115505937486887, + 0.004234392661601305, + 0.0037857552524656057, + 0.0036902118008583784, + 0.003440295346081257, + 0.0019714883528649807, + 0.0015250194119289517, + 0.008559878915548325, + 0.011286039836704731, + 0.009052705019712448, + 0.007093291264027357, + 0.011242331936955452, + 0.00852633360773325, + 0.006748402491211891, + 0.005564612336456776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.252111554145813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040934826247394085 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040934826247394085, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040934826247394085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2008270025253296, + 0.07084546238183975, + 0.253961980342865, + 0.27397510409355164, + 0.02254948578774929, + 0.034448835998773575, + 0.031510885804891586, + 0.03317227587103844, + 0.015150381252169609, + 0.06355857849121094 + ], + "credit_map": [ + 0.03530514985322952, + 0.0203071441501379, + 0.0312594510614872, + 0.03724723681807518, + 0.026762301102280617, + 0.023431112989783287, + 0.015118397772312164, + 0.011396200396120548, + 0.008146029897034168, + 0.00972660630941391, + 0.011038324795663357, + 0.01582544483244419, + 0.008175772614777088, + 0.009855329059064388, + 0.004819991532713175, + 0.0032579617109149694, + 0.022521579638123512, + 0.07893698662519455, + 0.02271205186843872, + 0.013025685213506222, + 0.08166620135307312, + 0.021158400923013687, + 0.00832211971282959, + 0.0056189619936048985, + 0.03218090161681175, + 0.06491053849458694, + 0.032645463943481445, + 0.034503523260354996, + 0.05945136770606041, + 0.02612679824233055, + 0.013018524274230003, + 0.011137990280985832, + 0.002576726023107767, + 0.003186670131981373, + 0.002533495891839266, + 0.003002060577273369, + 0.004691562615334988, + 0.002110787434503436, + 0.0027854449581354856, + 0.0016627387376502156, + 0.005331561900675297, + 0.006319125648587942, + 0.004091664217412472, + 0.0052137961611151695, + 0.006603304762393236, + 0.0033795221243053675, + 0.0017586263129487634, + 0.0017512338235974312, + 0.003232111455872655, + 0.004503475036472082, + 0.003685672301799059, + 0.005049723666161299, + 0.00587191479280591, + 0.003951340913772583, + 0.0026002658996731043, + 0.0026163836009800434, + 0.0033530984073877335, + 0.0052823759615421295, + 0.0038473443128168583, + 0.004133434034883976, + 0.005650853272527456, + 0.00477803498506546, + 0.0036502776201814413, + 0.002476859139278531, + 0.0016977416817098856, + 0.0022072060965001583, + 0.001886958722025156, + 0.002162182005122304, + 0.0018963924376294017, + 0.0025729460176080465, + 0.0014845781261101365, + 0.0012423761654645205, + 0.005722046364098787, + 0.013203158974647522, + 0.008541460148990154, + 0.007299019489437342, + 0.014031387865543365, + 0.007942725904285908, + 0.0036416822113096714, + 0.003177100792527199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0040934826247394085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, high-speed tracking shot, as if shot from an accompanying aircraft, captures the flight of a vibrant, candy-apple-red biplane, its smooth surface reflecting brilliant sunlight. The camera maintains an eye-level perspective in the single, open cockpit, revealing the faint outline of the pilot. The aircraft's propellers spin at such a rapid rate that the nose becomes a mesmerizingly transparent blur. Against the expansive backdrop of a flawless, deep blue sky, the biplane playfully executes daring maneuvers, weaving between and through enormous, fluffy, white cumulus clouds that appear as soft as cotton. It executes a swift and graceful barrel roll, its wings swiveling 360 degrees with thrilling precision, before sharply pulling up in a perfect, gravity-defying loop. The entire scene is bathed in crisp, high-key sunlight, creating brilliant highlights on the aircraft's fuselage and enlivening the saturated primary colors of red, white, and blue, perfectly capturing the festive atmosphere of pure, unadulterated joy in flight.", + "seed": 5588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_5/index_8683_seed_5588_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0940842003794387, + "vlm_logprob": -0.0940842003794387, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.0940842003794387, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09214778244495392 + ], + "opposite_logprob_tokens": [ + -2.4671478271484375 + ], + "logprob_reward": -0.09214778244495392, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09214778244495392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2211596667766571, + 0.04273807629942894, + 0.06090438738465309, + 0.04767286777496338, + 0.05317287892103195, + 0.04318266734480858, + 0.17625170946121216, + 0.05364200100302696, + 0.08934140205383301, + 0.21193434298038483 + ], + "credit_map": [ + 0.02397608943283558, + 0.025986718013882637, + 0.026474541053175926, + 0.0386357381939888, + 0.036536797881126404, + 0.033235274255275726, + 0.016886701807379723, + 0.01942780241370201, + 0.003029716433957219, + 0.0037335630040615797, + 0.004011279437690973, + 0.006648701149970293, + 0.009418649598956108, + 0.007959207519888878, + 0.0034932622220367193, + 0.00444369949400425, + 0.005333161447197199, + 0.013392322696745396, + 0.006911642383784056, + 0.006815183907747269, + 0.012557041831314564, + 0.008442709222435951, + 0.0035256941337138414, + 0.003926630597561598, + 0.00667057279497385, + 0.00800247024744749, + 0.013793280348181725, + 0.005605782847851515, + 0.004480123519897461, + 0.004465664736926556, + 0.0019146938575431705, + 0.0027402795385569334, + 0.008500317111611366, + 0.007921596057713032, + 0.017747988924384117, + 0.005265871528536081, + 0.003886078717187047, + 0.0036808105651289225, + 0.002482104115188122, + 0.003688113298267126, + 0.004393581300973892, + 0.004527631215751171, + 0.005927583668380976, + 0.006109557580202818, + 0.007506852503865957, + 0.004821322858333588, + 0.003847404383122921, + 0.0060487366281449795, + 0.01345978956669569, + 0.015313718467950821, + 0.027117999270558357, + 0.022703183814883232, + 0.020150117576122284, + 0.02275044098496437, + 0.017870906740427017, + 0.03688553720712662, + 0.004728358704596758, + 0.007509167306125164, + 0.007532512303441763, + 0.006840846501290798, + 0.008675607852637768, + 0.0075847734697163105, + 0.00425005191937089, + 0.006520679220557213, + 0.00711782043799758, + 0.008819903247058392, + 0.014942076057195663, + 0.018499912694096565, + 0.008471358567476273, + 0.012858877889811993, + 0.008087040856480598, + 0.010544410906732082, + 0.01570807956159115, + 0.04417286813259125, + 0.04563247039914131, + 0.02788262814283371, + 0.023710332810878754, + 0.01922672800719738, + 0.012750969268381596, + 0.02285025641322136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.09214778244495392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207606941461563 + ], + "opposite_logprob_tokens": [ + -2.8120760917663574 + ], + "logprob_reward": -0.06207606941461563, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207606941461563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1815674602985382, + 0.0896490216255188, + 0.10532195121049881, + 0.05353758484125137, + 0.057419657707214355, + 0.05559737980365753, + 0.07832938432693481, + 0.08530758321285248, + 0.09071478992700577, + 0.20255516469478607 + ], + "credit_map": [ + 0.01386148389428854, + 0.01594223827123642, + 0.016213875263929367, + 0.015984367579221725, + 0.04584590718150139, + 0.043038081377744675, + 0.010506629943847656, + 0.020174894481897354, + 0.00954597257077694, + 0.010558667592704296, + 0.013544600456953049, + 0.0195268914103508, + 0.008128244429826736, + 0.008878514170646667, + 0.008589359931647778, + 0.01087677851319313, + 0.007857774384319782, + 0.023723186925053596, + 0.01430953573435545, + 0.013309327885508537, + 0.020865822210907936, + 0.01327530574053526, + 0.005459903739392757, + 0.006521101109683514, + 0.00724538741633296, + 0.009851891547441483, + 0.011665673926472664, + 0.006214015185832977, + 0.00532202422618866, + 0.006004453636705875, + 0.003628687933087349, + 0.0036054556258022785, + 0.007427687291055918, + 0.007232293952256441, + 0.01562321837991476, + 0.007493937853723764, + 0.005384963471442461, + 0.0047865440137684345, + 0.003882777178660035, + 0.005588239524513483, + 0.006103576626628637, + 0.0069130538031458855, + 0.010285534895956516, + 0.005821438040584326, + 0.010784156620502472, + 0.006298568099737167, + 0.003184862667694688, + 0.006206192541867495, + 0.008308705873787403, + 0.008906523697078228, + 0.010599696077406406, + 0.008092356845736504, + 0.01116509921848774, + 0.01290053129196167, + 0.00638708146288991, + 0.011969395913183689, + 0.005602574907243252, + 0.010532153770327568, + 0.01044404786080122, + 0.012284746393561363, + 0.011222092434763908, + 0.01132549811154604, + 0.007876265794038773, + 0.016020212322473526, + 0.007951102219522, + 0.008897013030946255, + 0.010137049481272697, + 0.017285050824284554, + 0.011078347451984882, + 0.013597655110061169, + 0.007480186875909567, + 0.014288387261331081, + 0.020124688744544983, + 0.023843832314014435, + 0.023667972534894943, + 0.03324712812900543, + 0.02946317382156849, + 0.032293420284986496, + 0.016544746235013008, + 0.02337021380662918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.06207606941461563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015525996685028 + ], + "opposite_logprob_tokens": [ + -1.7015526294708252 + ], + "logprob_reward": -0.2015525996685028, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015525996685028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17664022743701935, + 0.0775989219546318, + 0.15678465366363525, + 0.06873355060815811, + 0.07410401850938797, + 0.04219478368759155, + 0.05963543429970741, + 0.05513254553079605, + 0.1215135008096695, + 0.16766242682933807 + ], + "credit_map": [ + 0.016846293583512306, + 0.016857851296663284, + 0.01935564912855625, + 0.025108518078923225, + 0.03638318553566933, + 0.03326816484332085, + 0.011439901776611805, + 0.01738065294921398, + 0.005140433553606272, + 0.0072458526119589806, + 0.01023730169981718, + 0.013007511384785175, + 0.01291565876454115, + 0.010750588960945606, + 0.010117068886756897, + 0.008184500969946384, + 0.015113987028598785, + 0.02288847230374813, + 0.0164010189473629, + 0.02061116136610508, + 0.04295174032449722, + 0.023504726588726044, + 0.006200622767210007, + 0.009112924337387085, + 0.010165012441575527, + 0.01201261579990387, + 0.02220512367784977, + 0.004877024330198765, + 0.006189511623233557, + 0.006395245436578989, + 0.0027491406071931124, + 0.004139878787100315, + 0.012228899635374546, + 0.010932182893157005, + 0.023067811504006386, + 0.006753353867679834, + 0.005922781303524971, + 0.005568015854805708, + 0.003992124460637569, + 0.005638850387185812, + 0.004846382886171341, + 0.0062081231735646725, + 0.0059352838434278965, + 0.004832268226891756, + 0.007804089225828648, + 0.004972266033291817, + 0.003464170964434743, + 0.004132200498133898, + 0.006257761735469103, + 0.008310324512422085, + 0.011124100536108017, + 0.008337315171957016, + 0.006138207856565714, + 0.0074041252955794334, + 0.004151389934122562, + 0.007912203669548035, + 0.0038510526064783335, + 0.0059862034395337105, + 0.005691437516361475, + 0.0068427277728915215, + 0.006739996373653412, + 0.008006428368389606, + 0.004927698988467455, + 0.013087003491818905, + 0.013289088383316994, + 0.013831224292516708, + 0.02121138945221901, + 0.02080736868083477, + 0.0139029361307621, + 0.020822938531637192, + 0.008384862914681435, + 0.009263685904443264, + 0.014405743218958378, + 0.027099929749965668, + 0.030566634610295296, + 0.021960312500596046, + 0.024660436436533928, + 0.023461133241653442, + 0.00900894682854414, + 0.016499271616339684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.2015525996685028 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293202638626099 + ], + "opposite_logprob_tokens": [ + -2.237931966781616 + ], + "logprob_reward": -0.11293202638626099, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293202638626099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19946709275245667, + 0.1730850636959076, + 0.12988072633743286, + 0.03157389909029007, + 0.0456143394112587, + 0.04634209722280502, + 0.03715502843260765, + 0.14869529008865356, + 0.09695552289485931, + 0.09123089164495468 + ], + "credit_map": [ + 0.019018234685063362, + 0.024193501099944115, + 0.02588644064962864, + 0.0427943654358387, + 0.02572154812514782, + 0.02211906388401985, + 0.020478136837482452, + 0.019255826249718666, + 0.010689379647374153, + 0.017535341903567314, + 0.016785940155386925, + 0.016967125236988068, + 0.0505446121096611, + 0.03641766309738159, + 0.010076351463794708, + 0.0140686659142375, + 0.011832302436232567, + 0.022994333878159523, + 0.009014216251671314, + 0.014394505880773067, + 0.04199628904461861, + 0.01849312148988247, + 0.004729093052446842, + 0.006426870822906494, + 0.0061753299087285995, + 0.006271619349718094, + 0.006212134845554829, + 0.003505559405311942, + 0.0032564762514084578, + 0.002984939841553569, + 0.0012707107234746218, + 0.0018971288809552789, + 0.005787828471511602, + 0.005714768543839455, + 0.011758455075323582, + 0.006108563859015703, + 0.004453697241842747, + 0.0042203208431601524, + 0.002663316670805216, + 0.004907392431050539, + 0.005540212616324425, + 0.005519690923392773, + 0.010341471061110497, + 0.005099658854305744, + 0.006912372540682554, + 0.004674783907830715, + 0.0024028278421610594, + 0.005851083900779486, + 0.0026992715429514647, + 0.0043246508575975895, + 0.0045630065724253654, + 0.004450373351573944, + 0.004222744144499302, + 0.005631268490105867, + 0.004355520009994507, + 0.006908196955919266, + 0.012369680218398571, + 0.0268921609967947, + 0.013355563394725323, + 0.016893206164240837, + 0.030684679746627808, + 0.016861505806446075, + 0.009060327894985676, + 0.022578177973628044, + 0.010506078600883484, + 0.010760334320366383, + 0.01351204514503479, + 0.014500564895570278, + 0.012369688600301743, + 0.017675276845693588, + 0.00817720964550972, + 0.009454323910176754, + 0.00740949297323823, + 0.01810123585164547, + 0.018907126039266586, + 0.01313865277916193, + 0.01131997536867857, + 0.011543788947165012, + 0.0049034119583666325, + 0.00590722169727087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.11293202638626099 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001712523982860148 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.001712523982860148, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001712523982860148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3073078691959381, + 0.10894433408975601, + 0.15488925576210022, + 0.04339686781167984, + 0.042322151362895966, + 0.06548348069190979, + 0.06479720026254654, + 0.11874173581600189, + 0.022113090381026268, + 0.07200399786233902 + ], + "credit_map": [ + 0.0316561684012413, + 0.03272765874862671, + 0.035097796469926834, + 0.049330636858940125, + 0.054811231791973114, + 0.051278289407491684, + 0.027767954394221306, + 0.024638136848807335, + 0.007587679661810398, + 0.011359699070453644, + 0.011507864110171795, + 0.019242754206061363, + 0.0234422218054533, + 0.015668993815779686, + 0.00952195469290018, + 0.01061316393315792, + 0.015974171459674835, + 0.020338160917162895, + 0.020749902352690697, + 0.02529001422226429, + 0.02906535379588604, + 0.02490381710231304, + 0.00733334431424737, + 0.011234492063522339, + 0.006853275001049042, + 0.007169206161051989, + 0.007386029697954655, + 0.005855100229382515, + 0.004573264624923468, + 0.0048826709389686584, + 0.0027041074354201555, + 0.003973216284066439, + 0.006133147049695253, + 0.005765113513916731, + 0.007887489162385464, + 0.004021044354885817, + 0.003677483880892396, + 0.0038399200420826674, + 0.0043614585883915424, + 0.006636494770646095, + 0.007025311700999737, + 0.007990946061909199, + 0.013252582401037216, + 0.00836327113211155, + 0.00915586855262518, + 0.006279651541262865, + 0.004506658762693405, + 0.008909190073609352, + 0.0058726840652525425, + 0.010914335958659649, + 0.010478837415575981, + 0.005132640711963177, + 0.006669569294899702, + 0.009758072905242443, + 0.005635315086692572, + 0.010335742495954037, + 0.008553532883524895, + 0.024510828778147697, + 0.010149333626031876, + 0.010927572846412659, + 0.025182975456118584, + 0.018502097576856613, + 0.006900839973241091, + 0.014014552347362041, + 0.0016268565086647868, + 0.002429386368021369, + 0.0036164189223200083, + 0.003145736875012517, + 0.003038808237761259, + 0.0027741480153054, + 0.0019783189054578543, + 0.0035034166648983955, + 0.0070013627409935, + 0.00909163523465395, + 0.011956886388361454, + 0.010439754463732243, + 0.012028027325868607, + 0.010776876471936703, + 0.004087245557457209, + 0.0066222152672708035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.001712523982860148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, high-speed tracking shot, as if shot from an accompanying aircraft, captures the flight of a vibrant, candy-apple-red biplane, its smooth surface reflecting brilliant sunlight. The camera maintains an eye-level perspective in the single, open cockpit, revealing the faint outline of the pilot. The aircraft's propellers spin at such a rapid rate that the nose becomes a mesmerizingly transparent blur. Against the expansive backdrop of a flawless, deep blue sky, the biplane playfully executes daring maneuvers, weaving between and through enormous, fluffy, white cumulus clouds that appear as soft as cotton. It executes a swift and graceful barrel roll, its wings swiveling 360 degrees with thrilling precision, before sharply pulling up in a perfect, gravity-defying loop. The entire scene is bathed in crisp, high-key sunlight, creating brilliant highlights on the aircraft's fuselage and enlivening the saturated primary colors of red, white, and blue, perfectly capturing the festive atmosphere of pure, unadulterated joy in flight.", + "seed": 5589, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_5/index_8683_seed_5589_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09382004518993199, + "vlm_logprob": -0.09382004518993199, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09382004518993199, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10429118573665619 + ], + "opposite_logprob_tokens": [ + -2.3542912006378174 + ], + "logprob_reward": -0.10429118573665619, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10429118573665619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06138069927692413, + 0.008076905272901058, + 0.008985447697341442, + 0.42872968316078186, + 0.008742336183786392, + 0.2961956560611725, + 0.029825866222381592, + 0.0624082051217556, + 0.049612920731306076, + 0.04604228958487511 + ], + "credit_map": [ + 0.003325603436678648, + 0.00835475418716669, + 0.010779276490211487, + 0.01613888144493103, + 0.005791772622615099, + 0.004416267853230238, + 0.009178326465189457, + 0.0033958216663450003, + 0.0009701316594146192, + 0.0010877185268327594, + 0.0010902205249294639, + 0.001134294900111854, + 0.0008744214428588748, + 0.0012051424710080028, + 0.0009729546727612615, + 0.000742021482437849, + 0.0007764866459183395, + 0.0010731543879956007, + 0.0011764811351895332, + 0.0017479765228927135, + 0.0014620519941672683, + 0.0007932934095151722, + 0.0013727030018344522, + 0.0005832999595440924, + 0.05292699486017227, + 0.14802587032318115, + 0.15249839425086975, + 0.03816661238670349, + 0.01727338321506977, + 0.013139276765286922, + 0.0038922415114939213, + 0.002806916134431958, + 0.0006766818696632981, + 0.0010916783940047026, + 0.0010493877343833447, + 0.0012958284933120012, + 0.001452243304811418, + 0.001449935487471521, + 0.0012704251566901803, + 0.0004561559180729091, + 0.017002563923597336, + 0.0897449478507042, + 0.013543137349188328, + 0.005723351147025824, + 0.045104898512363434, + 0.11709822714328766, + 0.003993675112724304, + 0.003984861075878143, + 0.0031413869000971317, + 0.0036962912417948246, + 0.003359990892931819, + 0.003629920305684209, + 0.0040794662199914455, + 0.007237880025058985, + 0.0029142398852854967, + 0.001766689121723175, + 0.0023563727736473083, + 0.003296659793704748, + 0.003747065318748355, + 0.0073170289397239685, + 0.019466977566480637, + 0.01951824128627777, + 0.0025142072699964046, + 0.004191652871668339, + 0.0023274393752217293, + 0.00806763581931591, + 0.005199195351451635, + 0.008218095637857914, + 0.007253440096974373, + 0.006345189176499844, + 0.006759129464626312, + 0.00544279208406806, + 0.002282758941873908, + 0.006438322365283966, + 0.005520309321582317, + 0.009026818908751011, + 0.008243615739047527, + 0.008939416147768497, + 0.0036781588569283485, + 0.001912888023070991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.10429118573665619 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004838436841965 + ], + "opposite_logprob_tokens": [ + -2.6950483322143555 + ], + "logprob_reward": -0.07004838436841965, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004838436841965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03189889341592789, + 0.008232010528445244, + 0.005359359085559845, + 0.7547426223754883, + 0.006794498302042484, + 0.08424017578363419, + 0.01605568639934063, + 0.033328887075185776, + 0.045846521854400635, + 0.013501298613846302 + ], + "credit_map": [ + 0.0017819724744185805, + 0.00344616686925292, + 0.00578724779188633, + 0.00883558951318264, + 0.0032797532621771097, + 0.002269269898533821, + 0.004694763571023941, + 0.0018041349248960614, + 0.0004909286508336663, + 0.0008205389603972435, + 0.0014901186805218458, + 0.0017564520239830017, + 0.0012680533109232783, + 0.001119887689128518, + 0.0008562472648918629, + 0.0004297849372960627, + 0.00044937050552107394, + 0.0006059617735445499, + 0.0007600424578413367, + 0.0012462029699236155, + 0.0006351694464683533, + 0.0005231374525465071, + 0.0008002005633898079, + 0.0003392741782590747, + 0.09218431264162064, + 0.2597927749156952, + 0.27040019631385803, + 0.06762269884347916, + 0.0303927231580019, + 0.0231034979224205, + 0.006447776686400175, + 0.0047987159341573715, + 0.0006625628448091447, + 0.0007506095571443439, + 0.0012617933098226786, + 0.0014743417268618941, + 0.000645465508569032, + 0.0010533202439546585, + 0.0006571143167093396, + 0.0002892915508709848, + 0.004179454408586025, + 0.038942545652389526, + 0.0020300366450101137, + 0.002304007764905691, + 0.01926162838935852, + 0.015418906696140766, + 0.0013769970973953605, + 0.0007266147877089679, + 0.0017585641471669078, + 0.0025443951599299908, + 0.0019447335507720709, + 0.002019534120336175, + 0.0020718693267554045, + 0.0029415490571409464, + 0.0016886871308088303, + 0.0010863551869988441, + 0.0015805953880771995, + 0.005633848253637552, + 0.0011880481615662575, + 0.0020464793778955936, + 0.01327971275895834, + 0.007572214584797621, + 0.0011644978076219559, + 0.0008634921978227794, + 0.0029860856011509895, + 0.0076065948233008385, + 0.0038416977040469646, + 0.005614263936877251, + 0.005890270229429007, + 0.01287827454507351, + 0.004160413984209299, + 0.0028689284808933735, + 0.0005960535490885377, + 0.003473985707387328, + 0.0014649326913058758, + 0.002084258710965514, + 0.003156940918415785, + 0.0012855627574026585, + 0.0008821671945042908, + 0.000557398539967835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.07004838436841965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988955974578857 + ], + "opposite_logprob_tokens": [ + -1.8048895597457886 + ], + "logprob_reward": -0.17988955974578857, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988955974578857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021337777376174927, + 0.007288802415132523, + 0.004502749536186457, + 0.3163222372531891, + 0.006636595353484154, + 0.4894355833530426, + 0.018230291083455086, + 0.07768513262271881, + 0.04209986329078674, + 0.01646093837916851 + ], + "credit_map": [ + 0.0018237384501844645, + 0.00224373210221529, + 0.00350951892323792, + 0.004591285716742277, + 0.002724012592807412, + 0.0026055246125906706, + 0.002169504761695862, + 0.001670460100285709, + 0.0005395383341237903, + 0.001071273465640843, + 0.0012284850236028433, + 0.0013495513703674078, + 0.0009001653525047004, + 0.0011151233920827508, + 0.0006194736924953759, + 0.00046519265742972493, + 0.0004676423850469291, + 0.0007455868180841208, + 0.00052312680054456, + 0.0007508466369472444, + 0.0006146889645606279, + 0.0004977614153176546, + 0.000542751862667501, + 0.0003603450022637844, + 0.03847939893603325, + 0.10867387056350708, + 0.11340587586164474, + 0.028463836759328842, + 0.012537115253508091, + 0.009859859943389893, + 0.002803710987791419, + 0.0020985726732760668, + 0.000694835209287703, + 0.0008952168864198029, + 0.0007447239477187395, + 0.0010545373661443591, + 0.0010505346581339836, + 0.0010331617668271065, + 0.0007909353007562459, + 0.00037265042192302644, + 0.0320981964468956, + 0.09080304205417633, + 0.014224092476069927, + 0.009365512989461422, + 0.12387902289628983, + 0.20502246916294098, + 0.00675596809014678, + 0.007287279702723026, + 0.0010131624294444919, + 0.004807775374501944, + 0.0013966019032523036, + 0.0012036403641104698, + 0.0020658622961491346, + 0.006453194655478001, + 0.0007511911680921912, + 0.0005388638819567859, + 0.002063250169157982, + 0.009258585050702095, + 0.002146405167877674, + 0.0032813549041748047, + 0.033200524747371674, + 0.02436762861907482, + 0.0018176587764173746, + 0.0015497278654947877, + 0.0020504151470959187, + 0.012987715192139149, + 0.0019941905047744513, + 0.00288248504512012, + 0.003964181523770094, + 0.014958156272768974, + 0.0019057287136092782, + 0.001356994966045022, + 0.0009872972732409835, + 0.003356915432959795, + 0.003167513757944107, + 0.003370067570358515, + 0.0015325722051784396, + 0.0012753078481182456, + 0.001317432033829391, + 0.001453833538107574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.17988955974578857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293213069438934 + ], + "opposite_logprob_tokens": [ + -2.2379322052001953 + ], + "logprob_reward": -0.11293213069438934, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293213069438934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03270135819911957, + 0.015494661405682564, + 0.007509009446948767, + 0.26209622621536255, + 0.01136521715670824, + 0.5381448268890381, + 0.015337721444666386, + 0.055278051644563675, + 0.02600756101310253, + 0.03606538847088814 + ], + "credit_map": [ + 0.002145936479791999, + 0.004120842088013887, + 0.005436239298433065, + 0.007348921615630388, + 0.0048047928139567375, + 0.0033059180714190006, + 0.003728818614035845, + 0.0018098907312378287, + 0.001001547439955175, + 0.0017914929194375873, + 0.0025920290499925613, + 0.00252744578756392, + 0.0026329432148486376, + 0.0017917163204401731, + 0.0021996800787746906, + 0.0009578054887242615, + 0.0007269852212630212, + 0.001306690159253776, + 0.001017359085381031, + 0.0013483274960890412, + 0.0007834680145606399, + 0.001005654688924551, + 0.0009055070113390684, + 0.00041501756641082466, + 0.032986998558044434, + 0.09076189994812012, + 0.09242484718561172, + 0.023160459473729134, + 0.01060749962925911, + 0.007706081960350275, + 0.0025607249699532986, + 0.0018877219408750534, + 0.001335682813078165, + 0.0015647149411961436, + 0.0012757770018652081, + 0.0021553649567067623, + 0.0016727527836337686, + 0.0018460099818184972, + 0.001032174681313336, + 0.0004827399388886988, + 0.028784574940800667, + 0.24760904908180237, + 0.012321998365223408, + 0.013163317926228046, + 0.08961982280015945, + 0.132341668009758, + 0.008462338708341122, + 0.005842047743499279, + 0.0012789515312761068, + 0.002971251495182514, + 0.0016074771992862225, + 0.0012899144785478711, + 0.0019135030452162027, + 0.0033011857885867357, + 0.0012423557927832007, + 0.0017330816481262445, + 0.0015927734784781933, + 0.010407507419586182, + 0.0017588341142982244, + 0.003988219890743494, + 0.02200455591082573, + 0.012621751055121422, + 0.00154989049769938, + 0.0013545213732868433, + 0.001905719400383532, + 0.0051987385377287865, + 0.0022332388907670975, + 0.0035730788949877024, + 0.004628001246601343, + 0.003392778569832444, + 0.002680819248780608, + 0.0023951877374202013, + 0.001938905450515449, + 0.0033397898077964783, + 0.006204962730407715, + 0.009786811657249928, + 0.0038688902277499437, + 0.0037226162385195494, + 0.0028136232867836952, + 0.0043897912837564945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.11293213069438934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00193896540440619 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.00193896540440619, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00193896540440619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017721407115459442, + 0.008731921203434467, + 0.022493356838822365, + 0.7873002290725708, + 0.0015260922955349088, + 0.0405905582010746, + 0.029456034302711487, + 0.07155180722475052, + 0.007816262543201447, + 0.012812312692403793 + ], + "credit_map": [ + 0.0013570813462138176, + 0.0021967352367937565, + 0.0027955046389251947, + 0.003717045532539487, + 0.0028514517471194267, + 0.0020717105362564325, + 0.001686735195107758, + 0.0010451455600559711, + 0.0007078700000420213, + 0.0011115365196019411, + 0.0014666874194517732, + 0.0014458914520218968, + 0.0010006232187151909, + 0.0014135370729491115, + 0.0010492528090253472, + 0.0005365231772884727, + 0.0019054338335990906, + 0.003227668348699808, + 0.0032915628980845213, + 0.005473427940160036, + 0.0025119215715676546, + 0.0026319115422666073, + 0.0025875004939734936, + 0.0008639323641546071, + 0.09679143875837326, + 0.27163606882095337, + 0.28122830390930176, + 0.07034360617399216, + 0.0316329263150692, + 0.023943791165947914, + 0.006722037680447102, + 0.005002055782824755, + 0.0001387654192512855, + 0.00024059429415501654, + 0.00022118938795756549, + 0.0002750552084762603, + 0.00019773072563111782, + 0.00020777390454895794, + 0.00015855590754654258, + 8.642764441901818e-05, + 0.0032048767898231745, + 0.01712253876030445, + 0.0021009386982768774, + 0.0014013313921168447, + 0.0088484613224864, + 0.006937189493328333, + 0.0005171082448214293, + 0.0004581178945954889, + 0.0020381025969982147, + 0.005139379762113094, + 0.0030135491397231817, + 0.0029492974281311035, + 0.0036556688137352467, + 0.008297786116600037, + 0.0022717509418725967, + 0.0020905027631670237, + 0.0025341480504721403, + 0.008785732090473175, + 0.003964514005929232, + 0.008074648678302765, + 0.024671828374266624, + 0.01662256196141243, + 0.003377107437700033, + 0.003521271515637636, + 0.0004528384015429765, + 0.002497619716450572, + 0.0005395137704908848, + 0.0008361911750398576, + 0.0006054269033484161, + 0.002286322182044387, + 0.0003682237584143877, + 0.00023012651945464313, + 0.0007040421478450298, + 0.0013472954742610455, + 0.0022521894425153732, + 0.0033896430395543575, + 0.0017655093688517809, + 0.0016219470417127013, + 0.0009741787798702717, + 0.0007575086783617735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.00193896540440619 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, high-speed tracking shot, as if shot from an accompanying aircraft, captures the flight of a vibrant, candy-apple-red biplane, its smooth surface reflecting brilliant sunlight. The camera maintains an eye-level perspective in the single, open cockpit, revealing the faint outline of the pilot. The aircraft's propellers spin at such a rapid rate that the nose becomes a mesmerizingly transparent blur. Against the expansive backdrop of a flawless, deep blue sky, the biplane playfully executes daring maneuvers, weaving between and through enormous, fluffy, white cumulus clouds that appear as soft as cotton. It executes a swift and graceful barrel roll, its wings swiveling 360 degrees with thrilling precision, before sharply pulling up in a perfect, gravity-defying loop. The entire scene is bathed in crisp, high-key sunlight, creating brilliant highlights on the aircraft's fuselage and enlivening the saturated primary colors of red, white, and blue, perfectly capturing the festive atmosphere of pure, unadulterated joy in flight.", + "seed": 5590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_5/index_8683_seed_5590_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09642552477307617, + "vlm_logprob": -0.09642552477307617, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.09642552477307617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029957987368106842 + ], + "opposite_logprob_tokens": [ + -3.6549580097198486 + ], + "logprob_reward": -0.029957987368106842, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029957987368106842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07044810801744461, + 0.02014942467212677, + 0.20859389007091522, + 0.013295693323016167, + 0.04797283187508583, + 0.05788177624344826, + 0.45964816212654114, + 0.014816946350038052, + 0.022908249869942665, + 0.0842849612236023 + ], + "credit_map": [ + 0.007627990562468767, + 0.016645383089780807, + 0.009698060341179371, + 0.003483385778963566, + 0.008050773292779922, + 0.017574505880475044, + 0.0022681739646941423, + 0.005099823698401451, + 0.0020037898793816566, + 0.0020845576655119658, + 0.0022523878142237663, + 0.0017398529453203082, + 0.0036030467599630356, + 0.003072750521823764, + 0.001738205086439848, + 0.00365483108907938, + 0.006007684860378504, + 0.07405408471822739, + 0.004387343768030405, + 0.004612100310623646, + 0.09398002922534943, + 0.017963603138923645, + 0.0033823198173195124, + 0.004206713289022446, + 0.0018118122825399041, + 0.0014108672039583325, + 0.0016412411350756884, + 0.0010051692370325327, + 0.0018055845284834504, + 0.0014985038433223963, + 0.0015497711719945073, + 0.0025727429892867804, + 0.005510499235242605, + 0.004120448138564825, + 0.004868194926530123, + 0.004327632952481508, + 0.007439484354108572, + 0.007909516803920269, + 0.006923350039869547, + 0.006873698905110359, + 0.006286787334829569, + 0.003618877148255706, + 0.010988970287144184, + 0.0029467048589140177, + 0.008157960139214993, + 0.02189893089234829, + 0.002037347061559558, + 0.0019461952615529299, + 0.08389943838119507, + 0.02996780350804329, + 0.0575837679207325, + 0.012429755181074142, + 0.06332619488239288, + 0.19412468373775482, + 0.006310677621513605, + 0.01200582180172205, + 0.0015556482831016183, + 0.00180557940620929, + 0.0018538720905780792, + 0.0025214538909494877, + 0.0017242910107597709, + 0.001741771469824016, + 0.0015431882347911596, + 0.0020711396355181932, + 0.0018846397288143635, + 0.0024024364538490772, + 0.003286037826910615, + 0.002883020555600524, + 0.0031558622140437365, + 0.0036803577095270157, + 0.0024402427952736616, + 0.0031756500247865915, + 0.004652366042137146, + 0.011737971566617489, + 0.022627145051956177, + 0.00674936268478632, + 0.0030932242516428232, + 0.020016541704535484, + 0.0023979893885552883, + 0.013010356575250626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.029957987368106842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256060540676117 + ], + "opposite_logprob_tokens": [ + -1.600606083869934 + ], + "logprob_reward": -0.2256060540676117, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256060540676117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09811437129974365, + 0.034359533339738846, + 0.20757898688316345, + 0.030178505927324295, + 0.06825853884220123, + 0.1475699096918106, + 0.3170718550682068, + 0.030260402709245682, + 0.03808995336294174, + 0.0285179503262043 + ], + "credit_map": [ + 0.01300651952624321, + 0.013285131193697453, + 0.02555396966636181, + 0.010203042067587376, + 0.007998442277312279, + 0.013621937483549118, + 0.007216319441795349, + 0.007229014299809933, + 0.0035416067112237215, + 0.002798258326947689, + 0.009774565696716309, + 0.0032347096130251884, + 0.003939340356737375, + 0.003712431527674198, + 0.002624472137540579, + 0.00473414734005928, + 0.005586070939898491, + 0.07403522729873657, + 0.003093085018917918, + 0.003853447735309601, + 0.09476911276578903, + 0.018899904564023018, + 0.0031964746303856373, + 0.004145665559917688, + 0.004908096976578236, + 0.0048802862875163555, + 0.0035082236863672733, + 0.0028880550526082516, + 0.005404975265264511, + 0.0039380700327456, + 0.0019330892246216536, + 0.0027177101001143456, + 0.00671377731487155, + 0.006231570616364479, + 0.009300251491367817, + 0.0070992689579725266, + 0.009133038111031055, + 0.00861862488090992, + 0.0123946787789464, + 0.008767331950366497, + 0.00957527756690979, + 0.010281682945787907, + 0.018484745174646378, + 0.006102437153458595, + 0.034856922924518585, + 0.05821339786052704, + 0.005497658625245094, + 0.004557775799185038, + 0.05142072215676308, + 0.033489733934402466, + 0.09823331981897354, + 0.016791783273220062, + 0.018303724005818367, + 0.07926317304372787, + 0.006852257531136274, + 0.012717110104858875, + 0.004032652359455824, + 0.004287983290851116, + 0.004614565521478653, + 0.005191844888031483, + 0.0029827493708580732, + 0.0027829688042402267, + 0.0033518504351377487, + 0.0030157887376844883, + 0.0037033793050795794, + 0.005462832748889923, + 0.004742793273180723, + 0.0051057059317827225, + 0.004794747103005648, + 0.004321882501244545, + 0.0039426893927156925, + 0.006015924271196127, + 0.0024331500753760338, + 0.007759202271699905, + 0.002082121791318059, + 0.0032130747567862272, + 0.0018303910037502646, + 0.007042004261165857, + 0.002053013304248452, + 0.0021049934439361095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.2256060540676117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988745868206024 + ], + "opposite_logprob_tokens": [ + -1.8048874139785767 + ], + "logprob_reward": -0.17988745868206024, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988745868206024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2979661226272583, + 0.033301688730716705, + 0.14424432814121246, + 0.024484846740961075, + 0.03470220789313316, + 0.05355769768357277, + 0.3043612837791443, + 0.008072531782090664, + 0.017013372853398323, + 0.08229590207338333 + ], + "credit_map": [ + 0.01241342630237341, + 0.11797638982534409, + 0.011835400015115738, + 0.010607168078422546, + 0.01461377926170826, + 0.11967507749795914, + 0.004006173927336931, + 0.006838694680482149, + 0.004481669515371323, + 0.0032385955564677715, + 0.0067937616258859634, + 0.0031864377669990063, + 0.005000318866223097, + 0.0042955102398991585, + 0.002176899230107665, + 0.004128493834286928, + 0.0040921312756836414, + 0.052156392484903336, + 0.0020534496288746595, + 0.002735461574047804, + 0.06594701111316681, + 0.012542600743472576, + 0.001972370082512498, + 0.002744919154793024, + 0.0036320704966783524, + 0.004187980201095343, + 0.0025910919066518545, + 0.0021515414118766785, + 0.005154815502464771, + 0.003327930113300681, + 0.0014814489986747503, + 0.001957968110218644, + 0.004031990189105272, + 0.003001532284542918, + 0.0023884514812380075, + 0.002707496052607894, + 0.0046056718565523624, + 0.006883502937853336, + 0.00514438608661294, + 0.005939173046499491, + 0.0046254489570856094, + 0.005555625539273024, + 0.0037683574482798576, + 0.004147642757743597, + 0.01649213396012783, + 0.01404950488358736, + 0.0027342578396201134, + 0.002184726530686021, + 0.044460371136665344, + 0.017746057361364365, + 0.055005498230457306, + 0.008161712437868118, + 0.03744608163833618, + 0.1302231103181839, + 0.0031381985172629356, + 0.00818026065826416, + 0.0008296029409393668, + 0.0007182528497651219, + 0.0012103600893169641, + 0.0015087249921634793, + 0.0007459829794242978, + 0.0010318956337869167, + 0.0008213800028897822, + 0.0012063324684277177, + 0.0017555360682308674, + 0.001943307463079691, + 0.0023528444580733776, + 0.0025612055324018, + 0.0022792944218963385, + 0.001957676140591502, + 0.0017630679067224264, + 0.0024004399310797453, + 0.004618829116225243, + 0.015781959518790245, + 0.014502855949103832, + 0.006970474496483803, + 0.0037812015507370234, + 0.022852003574371338, + 0.005968325771391392, + 0.007820247672498226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.17988745868206024 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306180030107498 + ], + "opposite_logprob_tokens": [ + -3.1680617332458496 + ], + "logprob_reward": -0.04306180030107498, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306180030107498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20925332605838776, + 0.018002139404416084, + 0.10520743578672409, + 0.036903396248817444, + 0.04849063977599144, + 0.04505731537938118, + 0.33267009258270264, + 0.03187694028019905, + 0.03816574811935425, + 0.13437294960021973 + ], + "credit_map": [ + 0.014047428034245968, + 0.07880600541830063, + 0.010704251937568188, + 0.00786664616316557, + 0.015638992190361023, + 0.07081791758537292, + 0.004140707664191723, + 0.00723139476031065, + 0.0029955455102026463, + 0.002059388905763626, + 0.002152408240363002, + 0.0015976662980392575, + 0.0033029355108737946, + 0.0021781153045594692, + 0.0014470743481069803, + 0.0022690072655677795, + 0.003197425976395607, + 0.036849334836006165, + 0.002340993145480752, + 0.001990612829104066, + 0.04746260493993759, + 0.009458024054765701, + 0.0016705066664144397, + 0.0022379381116479635, + 0.006213546264916658, + 0.005603852681815624, + 0.0052712117321789265, + 0.0033748175483196974, + 0.004464173223823309, + 0.003325783647596836, + 0.0029519794043153524, + 0.005698037799447775, + 0.004691950511187315, + 0.004855913110077381, + 0.005580666940659285, + 0.004588708747178316, + 0.006363560911267996, + 0.007140238303691149, + 0.00835131760686636, + 0.006918285973370075, + 0.004501675721257925, + 0.004941724240779877, + 0.010308856144547462, + 0.0039544133469462395, + 0.007289424538612366, + 0.009388173930346966, + 0.002913704374805093, + 0.0017593513475731015, + 0.07542026787996292, + 0.030739106237888336, + 0.020330607891082764, + 0.008987846784293652, + 0.049328453838825226, + 0.13365615904331207, + 0.005315249320119619, + 0.008892449550330639, + 0.003905958728864789, + 0.003732899436727166, + 0.003863708348944783, + 0.00483968760818243, + 0.003634826513007283, + 0.004332189913839102, + 0.0031861611641943455, + 0.00438151229172945, + 0.0026809934061020613, + 0.0047368984669446945, + 0.004479189403355122, + 0.0034511396661400795, + 0.004627757240086794, + 0.005011984147131443, + 0.006025225389748812, + 0.007152564357966185, + 0.007819848135113716, + 0.01179331075400114, + 0.04295358806848526, + 0.011632885783910751, + 0.0044876509346067905, + 0.02469012327492237, + 0.00422762893140316, + 0.026767922565340996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.04306180030107498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a biplane painted in a vibrant candy-apple-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned inside the single open cockpit at an eye-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pilot visible as a faint outline in the cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the biplane execute a barrel roll followed by a sharp pull-up into a loop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue sky with enormous fluffy white cumulus clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036143234465271235 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036143234465271235, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036143234465271235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041439566761255264, + 0.013281106948852539, + 0.23932980000972748, + 0.010184821672737598, + 0.008342646062374115, + 0.07092244178056717, + 0.5139527320861816, + 0.022906053811311722, + 0.011954069137573242, + 0.06768667697906494 + ], + "credit_map": [ + 0.004249319434165955, + 0.011127202771604061, + 0.006508082617074251, + 0.0025139658246189356, + 0.0042985412292182446, + 0.008438743650913239, + 0.0016303864540532231, + 0.002673323964700103, + 0.0013601669343188405, + 0.001038096146658063, + 0.003645590040832758, + 0.001286991871893406, + 0.001685024704784155, + 0.0014789815759286284, + 0.0009639265481382608, + 0.0018223286606371403, + 0.006853779312223196, + 0.08693180233240128, + 0.004778922535479069, + 0.005021644290536642, + 0.10813207179307938, + 0.019510310143232346, + 0.003354349173605442, + 0.004746907390654087, + 0.0019448065431788564, + 0.0016883318312466145, + 0.0013326882617548108, + 0.0007612191257067025, + 0.001373866805806756, + 0.0009198696934618056, + 0.0006550652906298637, + 0.0015089737717062235, + 0.0009311860194429755, + 0.0008908815216273069, + 0.00113132712431252, + 0.0006679669604636729, + 0.0014984506415203214, + 0.0012972484109923244, + 0.0009412432555109262, + 0.0009843416046351194, + 0.0051392484456300735, + 0.003859529737383127, + 0.011767801828682423, + 0.002502980176359415, + 0.014301775954663754, + 0.03009827807545662, + 0.001620408147573471, + 0.0016324268653988838, + 0.08553355187177658, + 0.030824115499854088, + 0.07739938050508499, + 0.012539606541395187, + 0.06823883205652237, + 0.22062592208385468, + 0.005632004234939814, + 0.013159340247511864, + 0.002993253292515874, + 0.0017782723298296332, + 0.004068860784173012, + 0.004346717614680529, + 0.0020048064179718494, + 0.0035659486893564463, + 0.0018514960538595915, + 0.0022966996766626835, + 0.0018115673447027802, + 0.0010657700477167964, + 0.0016022503841668367, + 0.0022341227158904076, + 0.001173379016108811, + 0.0018232624279335141, + 0.000989146064966917, + 0.0012545717181637883, + 0.005901941563934088, + 0.009447024203836918, + 0.022582529112696648, + 0.006430258508771658, + 0.0020624410826712847, + 0.0050939396023750305, + 0.001970476470887661, + 0.014198068529367447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.0036143234465271235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_tensor([7019, 7019])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_tensor([7019, 7019])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a5452d8542d8ff9bf8d3fb432aa6046055477db --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_tensor([7019, 7019])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle tracking shot follows a column of heavily armored cavalry on the march, capturing their epic journey across the vast wasteland. The group, comprised of approximately ten knights, is clad in weathered steel plate armor. Their closed helmets, with only narrow eye slits, obscure their features, revealing only a steely resolve. The metal of their armor, stained with battle scratches and a touch of rust, reflects a dazzling, sharp white light in the glaring afternoon sun. Their towering warhorses, similarly clad in metal armor, march with steady, powerful strides. Each hooves' tread on the cracked yellow earth and scattered gravel stirs up powerful plumes of yellow dust. This dust, illuminated by the sun, forms a golden mist that shrouds the cavalry, blurring the distant horizon and imbuing the scene with a profound sense of historical weight and vicissitudes. The leading knight occasionally raises his gauntleted hand and points into the distance, as if to determine the direction. The entire team maintains a compact and vigilant formation, steadily exploring forward. The main colors of the picture are ochre and the pale sky, punctuated by the silver-grey of the armor and the occasional crimson of the edge of the battle robe, creating a solemn, tenacious, and epic atmosphere of exploration.", + "seed": 1458, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_7019_seed_1458_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10628439383581281, + "vlm_logprob": -0.10628439383581281, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.10628439383581281, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2650326192378998 + ], + "opposite_logprob_tokens": [ + -1.7650326490402222 + ], + "logprob_reward": -0.2650326192378998, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2650326192378998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08493811637163162, + 0.03594483807682991, + 0.38678863644599915, + 0.17411202192306519, + 0.012506172992289066, + 0.03693457692861557, + 0.11750909686088562, + 0.030430998653173447, + 0.09995919466018677, + 0.02087632566690445 + ], + "credit_map": [ + 0.012356093153357506, + 0.0146385682746768, + 0.006923414766788483, + 0.009236409328877926, + 0.00788566842675209, + 0.01130703091621399, + 0.013244024477899075, + 0.00934690609574318, + 0.005733426194638014, + 0.0060378373600542545, + 0.003491688286885619, + 0.003250595647841692, + 0.005318670067936182, + 0.0046072290278971195, + 0.003866926534101367, + 0.0036384672857820988, + 0.016277538612484932, + 0.12017697095870972, + 0.012386499904096127, + 0.02442196197807789, + 0.03634786233305931, + 0.04350315406918526, + 0.09869885444641113, + 0.03497577831149101, + 0.005328409373760223, + 0.05563410744071007, + 0.0049534207209944725, + 0.013659275136888027, + 0.007331029511988163, + 0.050768546760082245, + 0.02535448782145977, + 0.011082751676440239, + 0.0012724027037620544, + 0.0021782091353088617, + 0.0010152236791327596, + 0.001300758682191372, + 0.0012269912986084819, + 0.0029238394927233458, + 0.0014335412997752428, + 0.0011552071664482355, + 0.0032984649296849966, + 0.007631197106093168, + 0.003117377171292901, + 0.005421872716397047, + 0.0032423397060483694, + 0.0050125750713050365, + 0.005080908536911011, + 0.004129843320697546, + 0.0071783107705414295, + 0.025713464245200157, + 0.010292530991137028, + 0.006953434552997351, + 0.0073163160122931, + 0.005514450371265411, + 0.0393526628613472, + 0.015187922865152359, + 0.001486198976635933, + 0.007948772981762886, + 0.0021642348729074, + 0.0035618068650364876, + 0.002922603627666831, + 0.0022812834940850735, + 0.004664035048335791, + 0.00540206441655755, + 0.02613712288439274, + 0.016889281570911407, + 0.013249184936285019, + 0.007178808096796274, + 0.007766393478959799, + 0.010297875851392746, + 0.006742488592863083, + 0.011698035523295403, + 0.0021152235567569733, + 0.003039910225197673, + 0.007252168841660023, + 0.0019855655264109373, + 0.0011047032894566655, + 0.0011814391473308206, + 0.0028325028251856565, + 0.0013648133026435971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.2650326192378998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179933562874794 + ], + "opposite_logprob_tokens": [ + -1.8049335479736328 + ], + "logprob_reward": -0.179933562874794, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179933562874794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09665217250585556, + 0.03038644604384899, + 0.31493279337882996, + 0.3652116656303406, + 0.014252807013690472, + 0.014245918020606041, + 0.05604005604982376, + 0.0692593902349472, + 0.01370767317712307, + 0.025311028584837914 + ], + "credit_map": [ + 0.01062098704278469, + 0.014977356418967247, + 0.007283218204975128, + 0.010564894415438175, + 0.011307451874017715, + 0.011530762538313866, + 0.015803584828972816, + 0.014563916251063347, + 0.005225409287959337, + 0.004525500349700451, + 0.002302222652360797, + 0.002504720352590084, + 0.0026136301457881927, + 0.005514232441782951, + 0.0031986632384359837, + 0.0045020682737231255, + 0.01921204663813114, + 0.09151925891637802, + 0.013308864086866379, + 0.020134277641773224, + 0.030388368293642998, + 0.03320007398724556, + 0.07853095978498459, + 0.02863897942006588, + 0.014494890347123146, + 0.057154037058353424, + 0.01351588498800993, + 0.026812920346856117, + 0.018514307215809822, + 0.10743635147809982, + 0.09212958067655563, + 0.035153694450855255, + 0.0015301327221095562, + 0.0035312550608068705, + 0.0010741706937551498, + 0.001583831268362701, + 0.0012953702826052904, + 0.0021885433234274387, + 0.0014686601934954524, + 0.001580843236297369, + 0.0017333213472738862, + 0.002042574342340231, + 0.0015744689153507352, + 0.0016815629787743092, + 0.001628414960578084, + 0.001628287136554718, + 0.002158949850127101, + 0.0017983390716835856, + 0.0035132255870848894, + 0.012439781799912453, + 0.004785817116498947, + 0.003473291639238596, + 0.0037320340052247047, + 0.002720763674005866, + 0.018126964569091797, + 0.00724817905575037, + 0.0036581438034772873, + 0.02338210493326187, + 0.005456127692013979, + 0.0035249716602265835, + 0.008043844252824783, + 0.005307222716510296, + 0.006476649548858404, + 0.01341033075004816, + 0.0024061817675828934, + 0.0019095824100077152, + 0.001851963927038014, + 0.0013105550315231085, + 0.0016327530611306429, + 0.0014647948555648327, + 0.0014749951660633087, + 0.0016568476567044854, + 0.0019495465094223619, + 0.0045842560939490795, + 0.00881241075694561, + 0.0022691278718411922, + 0.0013314749812707305, + 0.001243753358721733, + 0.0035789383109658957, + 0.0015415193047374487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.179933562874794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381406769156456 + ], + "opposite_logprob_tokens": [ + -3.2881407737731934 + ], + "logprob_reward": -0.0381406769156456, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381406769156456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06536552309989929, + 0.06778963655233383, + 0.3015426695346832, + 0.2689303159713745, + 0.018544116988778114, + 0.020901376381516457, + 0.16794441640377045, + 0.017448201775550842, + 0.02285565435886383, + 0.048678137362003326 + ], + "credit_map": [ + 0.006328633986413479, + 0.008980457670986652, + 0.007283679209649563, + 0.008593648672103882, + 0.006635358091443777, + 0.007961338385939598, + 0.01123674213886261, + 0.008345657028257847, + 0.010165348649024963, + 0.011787759140133858, + 0.006040899083018303, + 0.006353521719574928, + 0.011657637543976307, + 0.010061186738312244, + 0.005777465179562569, + 0.005945814307779074, + 0.01544925756752491, + 0.09081335365772247, + 0.01128883846104145, + 0.019580205902457237, + 0.028217781335115433, + 0.03178269416093826, + 0.07729090005159378, + 0.02711958810687065, + 0.010123210027813911, + 0.05395222827792168, + 0.009120779111981392, + 0.01737571880221367, + 0.011230288073420525, + 0.08474962413311005, + 0.05922740697860718, + 0.023151036351919174, + 0.001441247295588255, + 0.0043074204586446285, + 0.0011090455809608102, + 0.0016033161664381623, + 0.0017479663947597146, + 0.004274306353181601, + 0.0012982862535864115, + 0.0027625272050499916, + 0.0020796088501811028, + 0.0033180052414536476, + 0.0022694398649036884, + 0.0021175285801291466, + 0.002078054007142782, + 0.0027515096589922905, + 0.0038932464085519314, + 0.0023939821403473616, + 0.010113687254488468, + 0.03592049330472946, + 0.015182849019765854, + 0.009760682471096516, + 0.010392274707555771, + 0.0061958483420312405, + 0.057733748108148575, + 0.022644823417067528, + 0.0017209927318617702, + 0.003942286130040884, + 0.0022688726894557476, + 0.0013437087181955576, + 0.0023013425525277853, + 0.001533068367280066, + 0.0018862447468563914, + 0.002451686654239893, + 0.005473331548273563, + 0.003276843810454011, + 0.003160256426781416, + 0.0020595460664480925, + 0.002075849799439311, + 0.0024785238783806562, + 0.0018125241622328758, + 0.002518776571378112, + 0.004897278267890215, + 0.008609877899289131, + 0.007309265434741974, + 0.007352215703576803, + 0.004011881072074175, + 0.00455633457750082, + 0.006596056278795004, + 0.00534522021189332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.0381406769156456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06055998057126999, + 0.046999242156744, + 0.18373961746692657, + 0.27098900079727173, + 0.07601215690374374, + 0.05426611378788948, + 0.2050548493862152, + 0.029438180848956108, + 0.02457629330456257, + 0.048364561051130295 + ], + "credit_map": [ + 0.008573982864618301, + 0.00807489175349474, + 0.008704145438969135, + 0.007223851978778839, + 0.0073517318814992905, + 0.006695972755551338, + 0.008157728239893913, + 0.005777672864496708, + 0.009829837828874588, + 0.005378570407629013, + 0.004333547782152891, + 0.003537530777975917, + 0.005322340875864029, + 0.008360971696674824, + 0.00411242013797164, + 0.006124022416770458, + 0.015501672402024269, + 0.04879824444651604, + 0.011124232783913612, + 0.012317666783928871, + 0.01591387204825878, + 0.01713082194328308, + 0.04604693874716759, + 0.01690615341067314, + 0.010366435162723064, + 0.06840099394321442, + 0.010076814331114292, + 0.025906285271048546, + 0.011596946977078915, + 0.051434535533189774, + 0.06725636124610901, + 0.025950629264116287, + 0.005964117590337992, + 0.021807270124554634, + 0.0036876925732940435, + 0.006868945434689522, + 0.008220965974032879, + 0.017535395920276642, + 0.004714666865766048, + 0.007213099393993616, + 0.003633724059909582, + 0.011089867912232876, + 0.005250472109764814, + 0.0045555103570222855, + 0.004846267402172089, + 0.014790697023272514, + 0.003822580212727189, + 0.006276996340602636, + 0.01257417257875204, + 0.04508858546614647, + 0.018862076103687286, + 0.010501250624656677, + 0.012499039992690086, + 0.007663361728191376, + 0.07132195681333542, + 0.026544399559497833, + 0.001674280734732747, + 0.009052260778844357, + 0.0025253016501665115, + 0.002782593248412013, + 0.0031645179260522127, + 0.002276062499731779, + 0.003025348763912916, + 0.00493781641125679, + 0.00566759891808033, + 0.0031281625851988792, + 0.0035089892335236073, + 0.0025388647336512804, + 0.0018179566832259297, + 0.00240168790332973, + 0.0025495027657598257, + 0.0029635310638695955, + 0.005642564035952091, + 0.0076320297084748745, + 0.011713175103068352, + 0.007941285148262978, + 0.0030992156825959682, + 0.0029732296243309975, + 0.006402571219950914, + 0.0029604919254779816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.005256047938019037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305906221270561 + ], + "opposite_logprob_tokens": [ + -3.1680591106414795 + ], + "logprob_reward": -0.04305906221270561, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305906221270561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08747294545173645, + 0.029245438054203987, + 0.18686950206756592, + 0.45660632848739624, + 0.012388902716338634, + 0.025439105927944183, + 0.12469359487295151, + 0.028764864429831505, + 0.025876237079501152, + 0.02264307625591755 + ], + "credit_map": [ + 0.010665315203368664, + 0.011224415153265, + 0.010082527995109558, + 0.008939499035477638, + 0.008567982353270054, + 0.008749258704483509, + 0.019789082929491997, + 0.009454868733882904, + 0.0063096908852458, + 0.003580015152692795, + 0.003920453600585461, + 0.003160004736855626, + 0.002978638978675008, + 0.0033736033365130424, + 0.003030437510460615, + 0.0028925950173288584, + 0.007944402284920216, + 0.05823662132024765, + 0.005994057282805443, + 0.01159529760479927, + 0.018429400399327278, + 0.021723920479416847, + 0.04608051851391792, + 0.016865281388163567, + 0.019493749365210533, + 0.06525328755378723, + 0.01829214207828045, + 0.03587259724736214, + 0.02355094440281391, + 0.12924593687057495, + 0.11995615810155869, + 0.044941533356904984, + 0.0013121262891218066, + 0.001924207084812224, + 0.001203862251713872, + 0.0012552556581795216, + 0.0010171718895435333, + 0.0016371208475902677, + 0.002298027276992798, + 0.0017411329317837954, + 0.002206937875598669, + 0.0033381872344762087, + 0.002874653320759535, + 0.003345987293869257, + 0.002326817950233817, + 0.0037334898952394724, + 0.004511489998549223, + 0.003101543290540576, + 0.007544330321252346, + 0.027676083147525787, + 0.011230560950934887, + 0.006454445421695709, + 0.007848501205444336, + 0.0049252682365477085, + 0.04302183538675308, + 0.01599256321787834, + 0.001655466970987618, + 0.007127912249416113, + 0.0024151126854121685, + 0.002875120146200061, + 0.0027369994204491377, + 0.0019431881373748183, + 0.004778746049851179, + 0.005232318304479122, + 0.002420594450086355, + 0.004248881246894598, + 0.0041850232519209385, + 0.004434568807482719, + 0.0032344567589461803, + 0.0021692162845283747, + 0.0025239200331270695, + 0.0026595755480229855, + 0.0022023096680641174, + 0.0034265045542269945, + 0.005230289418250322, + 0.002883218927308917, + 0.0016691654454916716, + 0.0017912192270159721, + 0.003361172042787075, + 0.0020791967399418354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.04305906221270561 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle tracking shot follows a column of heavily armored cavalry on the march, capturing their epic journey across the vast wasteland. The group, comprised of approximately ten knights, is clad in weathered steel plate armor. Their closed helmets, with only narrow eye slits, obscure their features, revealing only a steely resolve. The metal of their armor, stained with battle scratches and a touch of rust, reflects a dazzling, sharp white light in the glaring afternoon sun. Their towering warhorses, similarly clad in metal armor, march with steady, powerful strides. Each hooves' tread on the cracked yellow earth and scattered gravel stirs up powerful plumes of yellow dust. This dust, illuminated by the sun, forms a golden mist that shrouds the cavalry, blurring the distant horizon and imbuing the scene with a profound sense of historical weight and vicissitudes. The leading knight occasionally raises his gauntleted hand and points into the distance, as if to determine the direction. The entire team maintains a compact and vigilant formation, steadily exploring forward. The main colors of the picture are ochre and the pale sky, punctuated by the silver-grey of the armor and the occasional crimson of the edge of the battle robe, creating a solemn, tenacious, and epic atmosphere of exploration.", + "seed": 1459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_7019_seed_1459_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09258906366303563, + "vlm_logprob": -0.09258906366303563, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.09258906366303563, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.27479609847068787 + ], + "opposite_logprob_tokens": [ + -1.6497961282730103 + ], + "logprob_reward": -0.27479609847068787, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.27479609847068787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08165039122104645, + 0.055935800075531006, + 0.0925353392958641, + 0.05379950627684593, + 0.0315142497420311, + 0.08302440494298935, + 0.06605181843042374, + 0.04919389635324478, + 0.43229058384895325, + 0.05400400608778 + ], + "credit_map": [ + 0.006908157840371132, + 0.00827177707105875, + 0.01254239957779646, + 0.009191698394715786, + 0.006266000680625439, + 0.009055175818502903, + 0.01833903230726719, + 0.01107614953070879, + 0.004751692991703749, + 0.00368939689360559, + 0.009766634553670883, + 0.011435596272349358, + 0.004201212432235479, + 0.005865811370313168, + 0.00945163331925869, + 0.006773816887289286, + 0.00646463967859745, + 0.00563637912273407, + 0.018709229305386543, + 0.020915985107421875, + 0.005587990395724773, + 0.007862831465899944, + 0.014361930079758167, + 0.012996356934309006, + 0.0073310271836817265, + 0.004817393142729998, + 0.007306115701794624, + 0.01111358217895031, + 0.0026675022672861814, + 0.0027428646571934223, + 0.008815951645374298, + 0.009005073457956314, + 0.0034223096445202827, + 0.0029130459297448397, + 0.004812352824956179, + 0.007761568296700716, + 0.00199898611754179, + 0.0025763805024325848, + 0.004994817078113556, + 0.0030347874853760004, + 0.01086517982184887, + 0.010756698437035084, + 0.013770046643912792, + 0.011212784796953201, + 0.006934571545571089, + 0.010724775493144989, + 0.00948358979076147, + 0.009276754222810268, + 0.008395196869969368, + 0.005292969290167093, + 0.01200832985341549, + 0.01722770743072033, + 0.00404997356235981, + 0.005451504606753588, + 0.00828105490654707, + 0.005345084238797426, + 0.004880886524915695, + 0.006080495193600655, + 0.009599098935723305, + 0.00921028945595026, + 0.004356129560619593, + 0.0049200477078557014, + 0.0045517487451434135, + 0.005595198832452297, + 0.08728303760290146, + 0.015055779367685318, + 0.03668108955025673, + 0.01728578843176365, + 0.05404755100607872, + 0.17097845673561096, + 0.01797524280846119, + 0.03298366069793701, + 0.009098867885768414, + 0.005771711003035307, + 0.006847691256552935, + 0.00722908740863204, + 0.00878774095326662, + 0.005591625347733498, + 0.004317123908549547, + 0.006360162515193224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.27479609847068787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709760665893555 + ], + "opposite_logprob_tokens": [ + -2.1270976066589355 + ], + "logprob_reward": -0.12709760665893555, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709760665893555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06656241416931152, + 0.07324337214231491, + 0.07376974821090698, + 0.04142211750149727, + 0.04572553187608719, + 0.04940667375922203, + 0.01750245876610279, + 0.04242713749408722, + 0.5428738594055176, + 0.047066718339920044 + ], + "credit_map": [ + 0.005814853589981794, + 0.007101558148860931, + 0.010785635560750961, + 0.007197429891675711, + 0.0063313585706055164, + 0.008889303542673588, + 0.011993267573416233, + 0.008449005894362926, + 0.006411541253328323, + 0.00728184636682272, + 0.009061111137270927, + 0.013070936314761639, + 0.005128686781972647, + 0.006977814715355635, + 0.01677936501801014, + 0.008532073348760605, + 0.004810746759176254, + 0.004873502533882856, + 0.015222713351249695, + 0.019379854202270508, + 0.003471804317086935, + 0.006195408292114735, + 0.012734542600810528, + 0.007081173360347748, + 0.003619144670665264, + 0.004080134443938732, + 0.005983418319374323, + 0.01136985793709755, + 0.0029986754525452852, + 0.0031983249355107546, + 0.006596817169338465, + 0.0035757438745349646, + 0.003983147908002138, + 0.0032967212609946728, + 0.008295984007418156, + 0.014123505912721157, + 0.0020463252440094948, + 0.0034110962878912687, + 0.006148232612758875, + 0.004420517943799496, + 0.005424748174846172, + 0.004785420838743448, + 0.009100265800952911, + 0.007649027742445469, + 0.003976818174123764, + 0.005783546715974808, + 0.006661131978034973, + 0.006025711540132761, + 0.0023337306920439005, + 0.0015892070950940251, + 0.003813616232946515, + 0.0019193764310330153, + 0.001539314165711403, + 0.002666554180905223, + 0.0019653430208563805, + 0.001675317296758294, + 0.010655974969267845, + 0.005826162174344063, + 0.00585921248421073, + 0.00443039508536458, + 0.00523481797426939, + 0.0031481182668358088, + 0.00251727644354105, + 0.004755175672471523, + 0.10703381896018982, + 0.019687648862600327, + 0.049433983862400055, + 0.022603480145335197, + 0.06644988805055618, + 0.20891794562339783, + 0.024673866108059883, + 0.04407322034239769, + 0.009986880235373974, + 0.003428560681641102, + 0.004857649561017752, + 0.005603349301964045, + 0.010870127938687801, + 0.004143484402447939, + 0.004091894254088402, + 0.0040847742930054665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.12709760665893555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033737633377313614 + ], + "opposite_logprob_tokens": [ + -3.4087376594543457 + ], + "logprob_reward": -0.033737633377313614, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033737633377313614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1661766916513443, + 0.09919828176498413, + 0.09054561704397202, + 0.060141146183013916, + 0.07662536948919296, + 0.07169867306947708, + 0.03123553656041622, + 0.04570161923766136, + 0.29379311203956604, + 0.06488398462533951 + ], + "credit_map": [ + 0.013079605996608734, + 0.015565929003059864, + 0.026646291837096214, + 0.018356848508119583, + 0.012809551320970058, + 0.01852935552597046, + 0.03935804218053818, + 0.02183106355369091, + 0.009373610839247704, + 0.008745000697672367, + 0.017801623791456223, + 0.015610788017511368, + 0.009094084613025188, + 0.012156957760453224, + 0.015046668238937855, + 0.01136954315006733, + 0.006446326617151499, + 0.006575333420187235, + 0.015679985284805298, + 0.02296009287238121, + 0.0045805578120052814, + 0.00691804476082325, + 0.017445923760533333, + 0.009939349256455898, + 0.004168807063251734, + 0.005349206738173962, + 0.008670790120959282, + 0.015163528732955456, + 0.004411345347762108, + 0.006223788019269705, + 0.008322499692440033, + 0.007831184193491936, + 0.005728587973862886, + 0.008397028781473637, + 0.011305831372737885, + 0.016489291563630104, + 0.0051029641181230545, + 0.009466256946325302, + 0.011489208787679672, + 0.008646202273666859, + 0.007857182063162327, + 0.008558172732591629, + 0.012263298965990543, + 0.010289758443832397, + 0.006829067133367062, + 0.009756872430443764, + 0.007561853155493736, + 0.00858246348798275, + 0.0035940445959568024, + 0.0029213924426585436, + 0.005308802239596844, + 0.004802784416824579, + 0.0031276103109121323, + 0.004933644086122513, + 0.0037768767215311527, + 0.002770381048321724, + 0.009653913788497448, + 0.006050601601600647, + 0.008352261036634445, + 0.0043128663673996925, + 0.004968699067831039, + 0.0030677674803882837, + 0.00384509121067822, + 0.005450419150292873, + 0.05739224702119827, + 0.010763773694634438, + 0.027201281860470772, + 0.01506136730313301, + 0.03442336618900299, + 0.10882479697465897, + 0.015041698701679707, + 0.02508455514907837, + 0.00804917048662901, + 0.00805240124464035, + 0.00928716454654932, + 0.009558891877532005, + 0.00706065446138382, + 0.008275930769741535, + 0.007789877709001303, + 0.006809892598539591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.033737633377313614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738912779837847 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006738912779837847, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738912779837847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27293500304222107, + 0.2159123569726944, + 0.12395325303077698, + 0.09801720827817917, + 0.07826825231313705, + 0.04716706648468971, + 0.01226833090186119, + 0.0577419251203537, + 0.06658513844013214, + 0.027151523157954216 + ], + "credit_map": [ + 0.0234147310256958, + 0.012449266389012337, + 0.06333441287279129, + 0.04198700189590454, + 0.01613425835967064, + 0.015589544549584389, + 0.07196611166000366, + 0.02805968001484871, + 0.017194410786032677, + 0.021126728504896164, + 0.024927405640482903, + 0.03777166083455086, + 0.016682809218764305, + 0.02115456759929657, + 0.04740944504737854, + 0.029645321890711784, + 0.008007680997252464, + 0.008494168519973755, + 0.02540646493434906, + 0.032952774316072464, + 0.005558296572417021, + 0.0057571642100811005, + 0.027042318135499954, + 0.01073438674211502, + 0.008272262290120125, + 0.009914637543261051, + 0.013605758547782898, + 0.022040361538529396, + 0.006865793839097023, + 0.009215078316628933, + 0.013798460364341736, + 0.014304857701063156, + 0.008391893468797207, + 0.006304051727056503, + 0.012979986146092415, + 0.019164610654115677, + 0.0039116861298680305, + 0.011043934151530266, + 0.009573184885084629, + 0.006898901890963316, + 0.006978913210332394, + 0.005123199429363012, + 0.008461385034024715, + 0.005934853572398424, + 0.0043514203280210495, + 0.006578180473297834, + 0.0043736593797802925, + 0.005365457385778427, + 0.001239365548826754, + 0.0014160549035295844, + 0.0019345161272212863, + 0.001822988037019968, + 0.0012649839045479894, + 0.0012423282023519278, + 0.001880883937701583, + 0.001467210124246776, + 0.009412537328898907, + 0.005417038220912218, + 0.010459290817379951, + 0.006750994361937046, + 0.005078661721199751, + 0.005315545946359634, + 0.008628611452877522, + 0.006679244805127382, + 0.0124185336753726, + 0.002545597730204463, + 0.0057858871296048164, + 0.0035169641487300396, + 0.008651339448988438, + 0.02596038393676281, + 0.0029048421420156956, + 0.004801587667316198, + 0.003335388144478202, + 0.004056139383465052, + 0.0036636621225625277, + 0.004017147701233625, + 0.0031741613056510687, + 0.003565812250599265, + 0.002842784859240055, + 0.002496426459401846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006738912779837847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575067028403282 + ], + "opposite_logprob_tokens": [ + -3.8955750465393066 + ], + "logprob_reward": -0.020575067028403282, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575067028403282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08644317835569382, + 0.091189906001091, + 0.12919999659061432, + 0.08481024950742722, + 0.04790464788675308, + 0.07450210303068161, + 0.07066017389297485, + 0.04628388211131096, + 0.27538734674453735, + 0.09361854940652847 + ], + "credit_map": [ + 0.007930362597107887, + 0.009017733857035637, + 0.012881583534181118, + 0.013176891021430492, + 0.010325944051146507, + 0.01133275032043457, + 0.012273142114281654, + 0.009504759684205055, + 0.006629845127463341, + 0.006127544213086367, + 0.012339653447270393, + 0.014081403613090515, + 0.006587275769561529, + 0.010646147653460503, + 0.02215665765106678, + 0.012621373869478703, + 0.011883918195962906, + 0.009497132152318954, + 0.02269641123712063, + 0.02442210726439953, + 0.007725301664322615, + 0.008432365953922272, + 0.025112083181738853, + 0.019430672749876976, + 0.009245803579688072, + 0.008412014693021774, + 0.012752823531627655, + 0.020387601107358932, + 0.005735678598284721, + 0.007033079396933317, + 0.013127307407557964, + 0.008115924894809723, + 0.004658460151404142, + 0.006055700127035379, + 0.006130507681518793, + 0.009075571782886982, + 0.0037377283442765474, + 0.006344346329569817, + 0.006043126340955496, + 0.005859198980033398, + 0.006717006675899029, + 0.009169729426503181, + 0.01181122474372387, + 0.009864547289907932, + 0.007062525488436222, + 0.012883784249424934, + 0.007917766459286213, + 0.009075517766177654, + 0.011145426891744137, + 0.005714550614356995, + 0.016753194853663445, + 0.01607460528612137, + 0.0043350583873689175, + 0.0037296838127076626, + 0.0076516252011060715, + 0.005256017204374075, + 0.006809381768107414, + 0.004726041108369827, + 0.008885014802217484, + 0.005892501212656498, + 0.004040735773742199, + 0.0044725146144628525, + 0.005170916672796011, + 0.006286769174039364, + 0.056786756962537766, + 0.011947105638682842, + 0.022702358663082123, + 0.011341622099280357, + 0.03335554525256157, + 0.10649356245994568, + 0.01220922265201807, + 0.020551150664687157, + 0.019379036501049995, + 0.010165790095925331, + 0.011799666099250317, + 0.01166134886443615, + 0.018150903284549713, + 0.007452991791069508, + 0.005940549075603485, + 0.009068255312740803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.020575067028403282 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle tracking shot follows a column of heavily armored cavalry on the march, capturing their epic journey across the vast wasteland. The group, comprised of approximately ten knights, is clad in weathered steel plate armor. Their closed helmets, with only narrow eye slits, obscure their features, revealing only a steely resolve. The metal of their armor, stained with battle scratches and a touch of rust, reflects a dazzling, sharp white light in the glaring afternoon sun. Their towering warhorses, similarly clad in metal armor, march with steady, powerful strides. Each hooves' tread on the cracked yellow earth and scattered gravel stirs up powerful plumes of yellow dust. This dust, illuminated by the sun, forms a golden mist that shrouds the cavalry, blurring the distant horizon and imbuing the scene with a profound sense of historical weight and vicissitudes. The leading knight occasionally raises his gauntleted hand and points into the distance, as if to determine the direction. The entire team maintains a compact and vigilant formation, steadily exploring forward. The main colors of the picture are ochre and the pale sky, punctuated by the silver-grey of the armor and the occasional crimson of the edge of the battle robe, creating a solemn, tenacious, and epic atmosphere of exploration.", + "seed": 1460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_7019_seed_1460_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09106536423787474, + "vlm_logprob": -0.09106536423787474, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.09106536423787474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21553948521614075 + ], + "opposite_logprob_tokens": [ + -1.8405394554138184 + ], + "logprob_reward": -0.21553948521614075, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21553948521614075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2147585153579712, + 0.07187739759683609, + 0.08933144062757492, + 0.08686202764511108, + 0.12429767847061157, + 0.051400452852249146, + 0.1343289166688919, + 0.09017980098724365, + 0.06887780129909515, + 0.06808602809906006 + ], + "credit_map": [ + 0.027400432154536247, + 0.020505009219050407, + 0.02916615828871727, + 0.03854241594672203, + 0.021243879571557045, + 0.021038400009274483, + 0.0359640009701252, + 0.020898202434182167, + 0.007685652934014797, + 0.009812719188630581, + 0.009436160326004028, + 0.013113781809806824, + 0.007465267553925514, + 0.007040109485387802, + 0.009577898308634758, + 0.007745807059109211, + 0.0062490967102348804, + 0.007360426709055901, + 0.013278765603899956, + 0.0197769645601511, + 0.006049474701285362, + 0.00965139176696539, + 0.015832260251045227, + 0.011133052408695221, + 0.0074820187874138355, + 0.008685152046382427, + 0.01080075092613697, + 0.01245225127786398, + 0.007274247705936432, + 0.007224499247968197, + 0.019323740154504776, + 0.013619355857372284, + 0.011953304521739483, + 0.012337987311184406, + 0.017662128433585167, + 0.024005461484193802, + 0.008579942397773266, + 0.008960696868598461, + 0.020671585574746132, + 0.020126551389694214, + 0.006299782078713179, + 0.00403281906619668, + 0.007075809873640537, + 0.011267908848822117, + 0.004199200309813023, + 0.00591622618958354, + 0.006733749061822891, + 0.005874953232705593, + 0.017905469983816147, + 0.015592670999467373, + 0.032834265381097794, + 0.02427847683429718, + 0.00804976187646389, + 0.008618362247943878, + 0.018262475728988647, + 0.008787424303591251, + 0.019106771796941757, + 0.0070052496157586575, + 0.022408366203308105, + 0.009899010881781578, + 0.005843474064022303, + 0.010955934412777424, + 0.007673801388591528, + 0.007287188433110714, + 0.009693695232272148, + 0.010359964333474636, + 0.014544298872351646, + 0.010302111506462097, + 0.005952020175755024, + 0.006221081595867872, + 0.0061468868516385555, + 0.005657740402966738, + 0.008051163516938686, + 0.010080324485898018, + 0.015871943905949593, + 0.006936152931302786, + 0.007431369740515947, + 0.0067082298919558525, + 0.006530298851430416, + 0.006476543378084898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.21553948521614075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993326485157013 + ], + "opposite_logprob_tokens": [ + -1.8049333095550537 + ], + "logprob_reward": -0.17993326485157013, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993326485157013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13683542609214783, + 0.051057688891887665, + 0.08982447534799576, + 0.09160953015089035, + 0.15853090584278107, + 0.08456413447856903, + 0.07554163038730621, + 0.12421295046806335, + 0.17339131236076355, + 0.014431984163820744 + ], + "credit_map": [ + 0.015264919027686119, + 0.012821978889405727, + 0.019631756469607353, + 0.028453022241592407, + 0.011447152122855186, + 0.01035391166806221, + 0.021163543686270714, + 0.017699146643280983, + 0.006792461033910513, + 0.004153464920818806, + 0.013258817605674267, + 0.008753122761845589, + 0.003895202185958624, + 0.003330664476379752, + 0.005630682688206434, + 0.0052432711236178875, + 0.006435591261833906, + 0.007722205016762018, + 0.01575012132525444, + 0.023553898558020592, + 0.0056350273080170155, + 0.00680522620677948, + 0.015352887101471424, + 0.008569509722292423, + 0.008961583487689495, + 0.009218874387443066, + 0.014530784450471401, + 0.019582100212574005, + 0.007040334865450859, + 0.005079707596451044, + 0.01718728616833687, + 0.010008848272264004, + 0.015581972897052765, + 0.014196107164025307, + 0.02421269193291664, + 0.04739644005894661, + 0.008285739459097385, + 0.011481905356049538, + 0.0232697781175375, + 0.014106261543929577, + 0.00748424232006073, + 0.0070640756748616695, + 0.011441503651440144, + 0.012656308710575104, + 0.008578269742429256, + 0.01314061414450407, + 0.013168438337743282, + 0.011030673049390316, + 0.010810146108269691, + 0.008727394975721836, + 0.014978109858930111, + 0.011807050555944443, + 0.004587066825479269, + 0.006584422197192907, + 0.011531958356499672, + 0.0065154749900102615, + 0.02313179336488247, + 0.011568563990294933, + 0.020159898325800896, + 0.013362282887101173, + 0.01075282134115696, + 0.01760801114141941, + 0.01653684303164482, + 0.011092728935182095, + 0.022228119894862175, + 0.02249119244515896, + 0.021448791027069092, + 0.029058067128062248, + 0.020426098257303238, + 0.01772919110953808, + 0.02068137377500534, + 0.01932847872376442, + 0.0019380932208150625, + 0.00237685046158731, + 0.002559986896812916, + 0.002371501410380006, + 0.0011864406988024712, + 0.0013096341863274574, + 0.0014776706229895353, + 0.00121180503629148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.17993326485157013 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374754637479782 + ], + "opposite_logprob_tokens": [ + -3.408747434616089 + ], + "logprob_reward": -0.03374754637479782, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374754637479782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1830952912569046, + 0.05281301215291023, + 0.06101337820291519, + 0.06505271047353745, + 0.07257265597581863, + 0.02417939342558384, + 0.1682959645986557, + 0.16990894079208374, + 0.17827224731445312, + 0.02479645423591137 + ], + "credit_map": [ + 0.02057930827140808, + 0.018420517444610596, + 0.03280269354581833, + 0.03693464398384094, + 0.02130666747689247, + 0.014539714902639389, + 0.023924333974719048, + 0.014587403275072575, + 0.006018196698278189, + 0.006200543139129877, + 0.007829259149730206, + 0.008231552317738533, + 0.005813101772218943, + 0.004377446603029966, + 0.008339683525264263, + 0.006003234535455704, + 0.005847492720931768, + 0.00919526070356369, + 0.008161164820194244, + 0.012865171767771244, + 0.0038301204331219196, + 0.00479531567543745, + 0.00954484473913908, + 0.00677400641143322, + 0.00544793251901865, + 0.007552469614893198, + 0.008459904231131077, + 0.01060259249061346, + 0.003368879435583949, + 0.005468469578772783, + 0.009739848785102367, + 0.01441261451691389, + 0.007024879101663828, + 0.00745745375752449, + 0.010622764006257057, + 0.018047846853733063, + 0.004938341211527586, + 0.00578507874161005, + 0.011537113226950169, + 0.007159177679568529, + 0.0021139460150152445, + 0.0026126170996576548, + 0.0032553670462220907, + 0.004460504744201899, + 0.002059337217360735, + 0.0029808273538947105, + 0.0034755365923047066, + 0.0032212568912655115, + 0.016966206952929497, + 0.014702335931360722, + 0.02951928786933422, + 0.029037266969680786, + 0.011165986768901348, + 0.01213102601468563, + 0.03772889822721481, + 0.017044950276613235, + 0.031123366206884384, + 0.014950055629014969, + 0.03961753472685814, + 0.014422692358493805, + 0.01704166643321514, + 0.026902997866272926, + 0.013976069167256355, + 0.011874553747475147, + 0.029160160571336746, + 0.02557612955570221, + 0.023487303406000137, + 0.03836796060204506, + 0.01133047416806221, + 0.013043639250099659, + 0.020083146169781685, + 0.017223438248038292, + 0.003161005210131407, + 0.004023520275950432, + 0.004458122421056032, + 0.002938987920060754, + 0.0033401791006326675, + 0.0020673314575105906, + 0.0023007404524832964, + 0.0025065678637474775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03374754637479782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17582345008850098, + 0.13417811691761017, + 0.08830735832452774, + 0.16111670434474945, + 0.08060060441493988, + 0.07140415161848068, + 0.09278397262096405, + 0.06436537206172943, + 0.0824313834309578, + 0.048988956958055496 + ], + "credit_map": [ + 0.018494121730327606, + 0.016183845698833466, + 0.02437354438006878, + 0.03688541799783707, + 0.01744050160050392, + 0.02005727030336857, + 0.024811258539557457, + 0.017577486112713814, + 0.016960328444838524, + 0.014098922722041607, + 0.020699163898825645, + 0.03206607326865196, + 0.010089725255966187, + 0.007330228108912706, + 0.017430346459150314, + 0.015503332018852234, + 0.00862210988998413, + 0.007865105755627155, + 0.013979923911392689, + 0.019769441336393356, + 0.006814859341830015, + 0.009206128306686878, + 0.011379504576325417, + 0.01067028846591711, + 0.013304337859153748, + 0.013879915699362755, + 0.028447818011045456, + 0.034360870718955994, + 0.008673779666423798, + 0.010417900048196316, + 0.02517540752887726, + 0.026856673881411552, + 0.010339410044252872, + 0.008638694882392883, + 0.010259018279612064, + 0.01768440753221512, + 0.004832729697227478, + 0.006131702568382025, + 0.010852107778191566, + 0.011862529441714287, + 0.0064980508759617805, + 0.005977267399430275, + 0.009212655946612358, + 0.00962432473897934, + 0.0071497512981295586, + 0.009837478399276733, + 0.010497502982616425, + 0.012607117183506489, + 0.013297750614583492, + 0.008903224021196365, + 0.021410992369055748, + 0.01839538663625717, + 0.004626552574336529, + 0.006183137185871601, + 0.013153410516679287, + 0.006813515443354845, + 0.008803088217973709, + 0.007269140798598528, + 0.009850265458226204, + 0.007072092965245247, + 0.007911475375294685, + 0.011655045673251152, + 0.005782643333077431, + 0.006021620240062475, + 0.011706555262207985, + 0.012631540186703205, + 0.012297077104449272, + 0.008407487533986568, + 0.009533431380987167, + 0.0100664421916008, + 0.009662661701440811, + 0.008126184344291687, + 0.006595837417989969, + 0.006924809422343969, + 0.006904560141265392, + 0.005332691594958305, + 0.00714667746797204, + 0.006500374060124159, + 0.00404774397611618, + 0.005536262411624193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002819731365889311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286793380975723 + ], + "opposite_logprob_tokens": [ + -3.773286819458008 + ], + "logprob_reward": -0.023286793380975723, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286793380975723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2191496193408966, + 0.09115993976593018, + 0.0694853737950325, + 0.0819736197590828, + 0.09874777495861053, + 0.08405797183513641, + 0.1143164187669754, + 0.07659070193767548, + 0.13975492119789124, + 0.024763604626059532 + ], + "credit_map": [ + 0.02662985399365425, + 0.021892184391617775, + 0.03895745053887367, + 0.0485486164689064, + 0.016492662951350212, + 0.014394234865903854, + 0.030828258022665977, + 0.02140638418495655, + 0.011609488166868687, + 0.00692795543000102, + 0.02410864643752575, + 0.02161191962659359, + 0.004277670290321112, + 0.005718749016523361, + 0.009317358024418354, + 0.007588162552565336, + 0.0062796613201498985, + 0.0066494569182395935, + 0.010509494692087173, + 0.016601093113422394, + 0.00432417681440711, + 0.007115931715816259, + 0.010374370962381363, + 0.007631187327206135, + 0.007418085355311632, + 0.007432354148477316, + 0.015350278466939926, + 0.02195183001458645, + 0.005328892730176449, + 0.003787339199334383, + 0.013658895157277584, + 0.007045954931527376, + 0.010083042085170746, + 0.008059228770434856, + 0.01557468343526125, + 0.025594772771000862, + 0.007269740104675293, + 0.007400113623589277, + 0.014892831444740295, + 0.009873364120721817, + 0.007385422475636005, + 0.007500925101339817, + 0.010199817828834057, + 0.010745800100266933, + 0.008851015008985996, + 0.013753336854279041, + 0.012354188598692417, + 0.013267471455037594, + 0.021033788099884987, + 0.00792821403592825, + 0.038672950118780136, + 0.019152361899614334, + 0.005433235317468643, + 0.007255493197590113, + 0.008829748257994652, + 0.006010645069181919, + 0.009689068421721458, + 0.006388378329575062, + 0.01264369860291481, + 0.008936815895140171, + 0.007852106355130672, + 0.012840642593801022, + 0.009306513704359531, + 0.008933492004871368, + 0.018741225823760033, + 0.01329015288501978, + 0.02082708664238453, + 0.03502001240849495, + 0.010713032446801662, + 0.01062762551009655, + 0.015772296115756035, + 0.01476350985467434, + 0.0028953843284398317, + 0.0028872026596218348, + 0.004083257168531418, + 0.0033865761943161488, + 0.0034787505865097046, + 0.0023200593423098326, + 0.0030438865069299936, + 0.0026684910990297794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.023286793380975723 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle tracking shot follows a column of heavily armored cavalry on the march, capturing their epic journey across the vast wasteland. The group, comprised of approximately ten knights, is clad in weathered steel plate armor. Their closed helmets, with only narrow eye slits, obscure their features, revealing only a steely resolve. The metal of their armor, stained with battle scratches and a touch of rust, reflects a dazzling, sharp white light in the glaring afternoon sun. Their towering warhorses, similarly clad in metal armor, march with steady, powerful strides. Each hooves' tread on the cracked yellow earth and scattered gravel stirs up powerful plumes of yellow dust. This dust, illuminated by the sun, forms a golden mist that shrouds the cavalry, blurring the distant horizon and imbuing the scene with a profound sense of historical weight and vicissitudes. The leading knight occasionally raises his gauntleted hand and points into the distance, as if to determine the direction. The entire team maintains a compact and vigilant formation, steadily exploring forward. The main colors of the picture are ochre and the pale sky, punctuated by the silver-grey of the armor and the occasional crimson of the edge of the battle robe, creating a solemn, tenacious, and epic atmosphere of exploration.", + "seed": 1461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_6/index_7019_seed_1461_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08611603449098766, + "vlm_logprob": -0.08611603449098766, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.08611603449098766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1552080512046814 + ], + "opposite_logprob_tokens": [ + -2.155208110809326 + ], + "logprob_reward": -0.1552080512046814, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1552080512046814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17442934215068817, + 0.07464386522769928, + 0.17668618261814117, + 0.08573056012392044, + 0.12277259677648544, + 0.11620838195085526, + 0.04631074145436287, + 0.13336938619613647, + 0.03937256336212158, + 0.030476374551653862 + ], + "credit_map": [ + 0.019620530307292938, + 0.014406335540115833, + 0.031281985342502594, + 0.024968694895505905, + 0.0123104527592659, + 0.011427730321884155, + 0.03198574110865593, + 0.028427865356206894, + 0.007854878902435303, + 0.010653615929186344, + 0.011341662146151066, + 0.015027949586510658, + 0.005813307128846645, + 0.006046430207788944, + 0.010613544844090939, + 0.007292482536286116, + 0.018497953191399574, + 0.018300333991646767, + 0.02800433151423931, + 0.05163617432117462, + 0.009819111786782742, + 0.012553418055176735, + 0.023755894973874092, + 0.014118972234427929, + 0.014170468784868717, + 0.0088204275816679, + 0.015218953602015972, + 0.015314305201172829, + 0.005677695851773024, + 0.008274050429463387, + 0.00914635881781578, + 0.009108295664191246, + 0.013670406304299831, + 0.009376896545290947, + 0.025734826922416687, + 0.03844516724348068, + 0.006766865495592356, + 0.006757838651537895, + 0.011038855649530888, + 0.01098173949867487, + 0.019318368285894394, + 0.019817067310214043, + 0.013339278288185596, + 0.023205852136015892, + 0.007607694715261459, + 0.009136457927525043, + 0.01483897864818573, + 0.008944679982960224, + 0.008453643880784512, + 0.005215620622038841, + 0.005538186989724636, + 0.0075472272001206875, + 0.002873786725103855, + 0.00440557673573494, + 0.006438022013753653, + 0.005838679615408182, + 0.037021372467279434, + 0.030934665352106094, + 0.018530840054154396, + 0.013024898245930672, + 0.01155445259064436, + 0.007958035916090012, + 0.007170061115175486, + 0.007175063714385033, + 0.006079443730413914, + 0.00966835767030716, + 0.0037445295602083206, + 0.003924585413187742, + 0.003343819407746196, + 0.006074988283216953, + 0.0031308038160204887, + 0.0034060333855450153, + 0.003717813640832901, + 0.002947085304185748, + 0.003809708636254072, + 0.004448702558875084, + 0.0030721225775778294, + 0.004055419471114874, + 0.003049293765798211, + 0.005376228131353855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.1552080512046814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255953997373581 + ], + "opposite_logprob_tokens": [ + -1.6005953550338745 + ], + "logprob_reward": -0.2255953997373581, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255953997373581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17455404996871948, + 0.07140027731657028, + 0.09366220235824585, + 0.2503263056278229, + 0.07437795400619507, + 0.07044168561697006, + 0.07449143379926682, + 0.05098731070756912, + 0.030617911368608475, + 0.10914085805416107 + ], + "credit_map": [ + 0.01351610105484724, + 0.012200194410979748, + 0.021885549649596214, + 0.026928739622235298, + 0.024966515600681305, + 0.018300751224160194, + 0.03478478267788887, + 0.02197142504155636, + 0.007842163555324078, + 0.011957560665905476, + 0.009131981059908867, + 0.009693359024822712, + 0.008121452294290066, + 0.007825074717402458, + 0.011242624372243881, + 0.005586056504398584, + 0.010353037156164646, + 0.00999012216925621, + 0.016099918633699417, + 0.01596967689692974, + 0.007723596412688494, + 0.011721945367753506, + 0.011106488294899464, + 0.01069742813706398, + 0.048077043145895004, + 0.018299229443073273, + 0.06650438904762268, + 0.041093479841947556, + 0.01431022398173809, + 0.019618097692728043, + 0.023799359798431396, + 0.01862446218729019, + 0.0066101523116230965, + 0.004901386331766844, + 0.017723936587572098, + 0.016947396099567413, + 0.0047792308032512665, + 0.007312479428946972, + 0.008185350336134434, + 0.00791802816092968, + 0.009517994709312916, + 0.012151939794421196, + 0.009623016230762005, + 0.01373126357793808, + 0.005769954528659582, + 0.00416222307831049, + 0.009694213978946209, + 0.0057910773903131485, + 0.0069108991883695126, + 0.006919286213815212, + 0.009683110751211643, + 0.01228464674204588, + 0.005997809581458569, + 0.009045657701790333, + 0.01352065708488226, + 0.01012936420738697, + 0.007444662973284721, + 0.007812762632966042, + 0.005478286650031805, + 0.005295997951179743, + 0.005368590820580721, + 0.006391817703843117, + 0.006900448352098465, + 0.006294741295278072, + 0.004904583562165499, + 0.0032723559997975826, + 0.005266611929982901, + 0.004848448559641838, + 0.0017401785589754581, + 0.0027697295881807804, + 0.004605196416378021, + 0.0032108083833009005, + 0.01332993246614933, + 0.00868673250079155, + 0.024356698617339134, + 0.015963114798069, + 0.00966508686542511, + 0.013888908550143242, + 0.009418373927474022, + 0.013832015916705132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.2255953997373581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638125978410244 + ], + "opposite_logprob_tokens": [ + -3.651381254196167 + ], + "logprob_reward": -0.02638125978410244, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638125978410244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461203396320343, + 0.09111448377370834, + 0.10434141010046005, + 0.11915935575962067, + 0.06368288397789001, + 0.08561097830533981, + 0.0464787557721138, + 0.14351537823677063, + 0.08424092084169388, + 0.11573544889688492 + ], + "credit_map": [ + 0.01683102548122406, + 0.013843835331499577, + 0.0205095112323761, + 0.019960567355155945, + 0.009705029428005219, + 0.0106106111779809, + 0.0296919085085392, + 0.0249678585678339, + 0.01063975878059864, + 0.014687861315906048, + 0.015413261950016022, + 0.017090486362576485, + 0.00673330994322896, + 0.006146654952317476, + 0.011243827641010284, + 0.009159320965409279, + 0.010807101614773273, + 0.009033135138452053, + 0.021023109555244446, + 0.021765299141407013, + 0.006789589766412973, + 0.012258002534508705, + 0.011438540183007717, + 0.011226641945540905, + 0.024529103189706802, + 0.012325643561780453, + 0.02825447916984558, + 0.01811610907316208, + 0.00878184661269188, + 0.008220129646360874, + 0.010617182590067387, + 0.008314860053360462, + 0.0053838337771594524, + 0.004368837457150221, + 0.015076340176165104, + 0.01461157388985157, + 0.004027043469250202, + 0.006427382118999958, + 0.006556886713951826, + 0.007230985444039106, + 0.011091060005128384, + 0.015820153057575226, + 0.015176715329289436, + 0.013762264512479305, + 0.008540419861674309, + 0.005571609828621149, + 0.009175905957818031, + 0.006472848821431398, + 0.0044929590076208115, + 0.0060592712834477425, + 0.005966403055936098, + 0.006635547615587711, + 0.003958297427743673, + 0.005541761871427298, + 0.008905100636184216, + 0.004919412545859814, + 0.02934451401233673, + 0.026625217869877815, + 0.028524532914161682, + 0.01572355069220066, + 0.010153783485293388, + 0.008286158554255962, + 0.010954344645142555, + 0.013903277926146984, + 0.00896948017179966, + 0.013662072829902172, + 0.018294380977749825, + 0.012446594424545765, + 0.007579834200441837, + 0.007559358142316341, + 0.006493349093943834, + 0.009235848672688007, + 0.013971658423542976, + 0.00943949818611145, + 0.021488850936293602, + 0.019235163927078247, + 0.008231975138187408, + 0.019276566803455353, + 0.00956494640558958, + 0.014526786282658577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.02638125978410244 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817591419443488 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.002817591419443488, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817591419443488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23812898993492126, + 0.09187095612287521, + 0.06112657114863396, + 0.08105232566595078, + 0.10650800168514252, + 0.06793566793203354, + 0.020696930587291718, + 0.18442067503929138, + 0.08558359742164612, + 0.0626763254404068 + ], + "credit_map": [ + 0.025376413017511368, + 0.029797138646245003, + 0.045386310666799545, + 0.04061292111873627, + 0.018833473324775696, + 0.02213784120976925, + 0.030777504667639732, + 0.025207383558154106, + 0.011155698448419571, + 0.010942027904093266, + 0.01723947748541832, + 0.020190002396702766, + 0.005726032890379429, + 0.005096468608826399, + 0.012992138974368572, + 0.00852910801768303, + 0.007261164952069521, + 0.006790236569941044, + 0.011368822306394577, + 0.00960287544876337, + 0.0044761402532458305, + 0.0073212143033742905, + 0.00755701120942831, + 0.006749101914465427, + 0.016575006768107414, + 0.006310191936790943, + 0.01996491476893425, + 0.014234179630875587, + 0.0033999572042375803, + 0.005618471186608076, + 0.007507698144763708, + 0.007441902067512274, + 0.00840572826564312, + 0.009455855004489422, + 0.029261602088809013, + 0.035736583173274994, + 0.003600252326577902, + 0.005274463910609484, + 0.007711984682828188, + 0.007061521522700787, + 0.008598113432526588, + 0.010626641102135181, + 0.008906343951821327, + 0.0131760835647583, + 0.006560721900314093, + 0.005505154840648174, + 0.009779919870197773, + 0.004782683216035366, + 0.0027345933485776186, + 0.0037020223680883646, + 0.003098792629316449, + 0.0029073746409267187, + 0.0018342537805438042, + 0.002107785316184163, + 0.002680398290976882, + 0.0016317085828632116, + 0.04938958212733269, + 0.041398029774427414, + 0.026551399379968643, + 0.017741352319717407, + 0.01552763395011425, + 0.013113640248775482, + 0.010899033397436142, + 0.009799988009035587, + 0.01525135338306427, + 0.020249392837285995, + 0.011207710020244122, + 0.009309504181146622, + 0.008538280613720417, + 0.007600909098982811, + 0.005234894808381796, + 0.008191550150513649, + 0.006558676715940237, + 0.0063593401573598385, + 0.011472082696855068, + 0.008502093143761158, + 0.008124162442982197, + 0.007882960140705109, + 0.007916721515357494, + 0.0058602821081876755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.002817591419443488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene featuring a column of approximately ten knights in weathered steel plate armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights' helmets closed with only narrow eye slits visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the knights riding warhorses clad in metal armor across cracked yellow earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a wide-angle tracking shot following the cavalry's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.149463951587677, + 0.07945071905851364, + 0.060094717890024185, + 0.14553943276405334, + 0.03231237828731537, + 0.05571639910340309, + 0.06986469030380249, + 0.1970783919095993, + 0.18241654336452484, + 0.02806275524199009 + ], + "credit_map": [ + 0.01669348031282425, + 0.014891903847455978, + 0.01710391975939274, + 0.017448391765356064, + 0.024648763239383698, + 0.017027296125888824, + 0.02022215537726879, + 0.02142803929746151, + 0.0077927010133862495, + 0.008124406449496746, + 0.01276242733001709, + 0.013643101789057255, + 0.008052702061831951, + 0.009171046316623688, + 0.012652752920985222, + 0.007251577917486429, + 0.007857391610741615, + 0.0071989139541983604, + 0.010241495445370674, + 0.011134734377264977, + 0.005015444476157427, + 0.006915274076163769, + 0.005426862742751837, + 0.006304601673036814, + 0.028143271803855896, + 0.011103719472885132, + 0.038229286670684814, + 0.025035113096237183, + 0.007805618457496166, + 0.01130062434822321, + 0.01208385732024908, + 0.011837942525744438, + 0.003404855728149414, + 0.003070298582315445, + 0.004476465750485659, + 0.007050430867820978, + 0.002299163956195116, + 0.0031512698624283075, + 0.0041540078818798065, + 0.004705884959548712, + 0.008749593049287796, + 0.013321501202881336, + 0.006812931504100561, + 0.008674945682287216, + 0.004130947403609753, + 0.0043408810161054134, + 0.006034144200384617, + 0.0036514578387141228, + 0.00895850732922554, + 0.006760311312973499, + 0.0090125547721982, + 0.009372345171868801, + 0.005022414028644562, + 0.008688020519912243, + 0.011930303648114204, + 0.01012023538351059, + 0.034177832305431366, + 0.03800433874130249, + 0.03131501004099846, + 0.022898975759744644, + 0.02136273868381977, + 0.014835761860013008, + 0.019708484411239624, + 0.014775256626307964, + 0.02217480167746544, + 0.01996682584285736, + 0.05060886591672897, + 0.021711306646466255, + 0.013871935196220875, + 0.02014666609466076, + 0.018680671229958534, + 0.015255472622811794, + 0.005182701628655195, + 0.0029835135210305452, + 0.004268066957592964, + 0.00301259383559227, + 0.0036356153432279825, + 0.003995394799858332, + 0.0026957374066114426, + 0.0022891315165907145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.020577870309352875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_tensor([2825, 2825])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_tensor([2825, 2825])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..53067b543bdd90fbc9c63a4e522d8dc9c962f920 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_tensor([2825, 2825])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "surreal, static, extreme macro close-up, tightly framed from a slightly low angle, captures the raw moment of the confrontation with breathtaking clarity. A small, sleek black scorpion dominates the frame, its jet-black, carapace-like exoskeleton gleaming with a menacing sheen. Its body tenses, its pincers slightly open, poised, its distinctive, segmented tail arched high over its back in a characteristically menacing pose. At its tip, a sharp, yellow stinger gleams, poised to strike with deadly precision. Its target is a larger beetle larva, helplessly wriggling directly beneath it. The larva's soft, milky-white, C-shaped body, with its fleshy, translucent sections, twitches in slow, desperate waves. This dramatic encounter unfolds on a piece of rotting, damp wood, where the intricate texture of the dark veins and shredded fibers is clearly rendered. A soft, diffuse light, as if filtering through a forest canopy, meticulously captures the tiny, almost invisible hairs on the scorpion's legs and highlights the larva's damp, glistening surface, creating a scene of intense, suspended tension. The overall atmosphere is clean and calm, a raw and focused look at the predator-prey dynamic frozen in the moments before its violent end.", + "seed": 30778, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30778_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012680201278999447, + "vlm_logprob": -0.012680201278999447, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.012680201278999447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01742764189839363 + ], + "opposite_logprob_tokens": [ + -4.142427444458008 + ], + "logprob_reward": -0.01742764189839363, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01742764189839363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1830807328224182, + 0.07418207079172134, + 0.064702108502388, + 0.028590044006705284, + 0.04796202480792999, + 0.06054510548710823, + 0.04608316719532013, + 0.30613940954208374, + 0.036025527864694595, + 0.1526898443698883 + ], + "credit_map": [ + 0.022440265864133835, + 0.028740014880895615, + 0.01706734113395214, + 0.02935514785349369, + 0.032260820269584656, + 0.017535462975502014, + 0.018327994272112846, + 0.017353668808937073, + 0.009176380932331085, + 0.00831227470189333, + 0.009453341364860535, + 0.00849903468042612, + 0.00887697096914053, + 0.008912600576877594, + 0.007091888226568699, + 0.013859573751688004, + 0.009416552260518074, + 0.006902259308844805, + 0.00675978185608983, + 0.007648882921785116, + 0.012754380702972412, + 0.008187949657440186, + 0.006254727952182293, + 0.006777564063668251, + 0.0038710697554051876, + 0.0037481326144188643, + 0.0029394137673079967, + 0.004215049557387829, + 0.004419347736984491, + 0.0024896375834941864, + 0.003077693283557892, + 0.003829697612673044, + 0.008927267044782639, + 0.006238102912902832, + 0.005073711276054382, + 0.006554449908435345, + 0.005732411053031683, + 0.005892013665288687, + 0.0044948342256248, + 0.0050492314621806145, + 0.011519324965775013, + 0.0051542241126298904, + 0.0073713334277272224, + 0.007456591818481684, + 0.0054077273234725, + 0.008831250481307507, + 0.008320876397192478, + 0.006483776494860649, + 0.0058784266002476215, + 0.004118987359106541, + 0.004654639400541782, + 0.004490450490266085, + 0.006954971235245466, + 0.005210447125136852, + 0.008756590075790882, + 0.0060186488553881645, + 0.05459969490766525, + 0.030967501923441887, + 0.032494205981492996, + 0.029029302299022675, + 0.04758550971746445, + 0.038331352174282074, + 0.04376700520515442, + 0.029364828020334244, + 0.007187497802078724, + 0.002745435107499361, + 0.004740062635391951, + 0.003401434049010277, + 0.0029959212988615036, + 0.007550694979727268, + 0.0035728116054087877, + 0.003831665264442563, + 0.017399825155735016, + 0.019268367439508438, + 0.015284423716366291, + 0.015581892803311348, + 0.01668792963027954, + 0.016504807397723198, + 0.029070306569337845, + 0.022892272099852562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.01742764189839363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820240542292595 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.01820240542292595, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820240542292595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28591880202293396, + 0.1058020070195198, + 0.09177681803703308, + 0.04116824269294739, + 0.04616580903530121, + 0.02740757167339325, + 0.037312645465135574, + 0.16462504863739014, + 0.05132266879081726, + 0.14850036799907684 + ], + "credit_map": [ + 0.04361162707209587, + 0.045073967427015305, + 0.02879774197936058, + 0.046279724687337875, + 0.04129588231444359, + 0.02570400759577751, + 0.02805057354271412, + 0.027105262503027916, + 0.012178348377346992, + 0.013595362193882465, + 0.01184967439621687, + 0.013274789787828922, + 0.01549456175416708, + 0.013433927670121193, + 0.008191884495317936, + 0.017783453688025475, + 0.011482831090688705, + 0.012544319964945316, + 0.011388059705495834, + 0.011442575603723526, + 0.01265342254191637, + 0.012310250662267208, + 0.007537979166954756, + 0.012417379766702652, + 0.005379782989621162, + 0.006497868336737156, + 0.004377060104161501, + 0.00647631986066699, + 0.003003104589879513, + 0.0037860984448343515, + 0.004772154614329338, + 0.006875853054225445, + 0.007056598085910082, + 0.006468087900429964, + 0.004925370682030916, + 0.00657676113769412, + 0.005323419347405434, + 0.004896506201475859, + 0.005499653518199921, + 0.0054194130934774876, + 0.0040188804268836975, + 0.0032483977265655994, + 0.003109862795099616, + 0.0030385989230126143, + 0.0035624047741293907, + 0.0034899876918643713, + 0.003499050857499242, + 0.0034403903409838676, + 0.004900684114545584, + 0.004576539155095816, + 0.003868681378662586, + 0.005580123979598284, + 0.004136685281991959, + 0.004526849370449781, + 0.004388013854622841, + 0.005335069727152586, + 0.024625316262245178, + 0.018405573442578316, + 0.019523179158568382, + 0.018662096932530403, + 0.017864303663372993, + 0.019774332642555237, + 0.025656405836343765, + 0.02011384814977646, + 0.008827488869428635, + 0.006079785991460085, + 0.004302807152271271, + 0.005184896290302277, + 0.009882465936243534, + 0.00558116752654314, + 0.007072580046951771, + 0.004391476046293974, + 0.015004457905888557, + 0.02186959981918335, + 0.01311536505818367, + 0.024057665839791298, + 0.01214102003723383, + 0.014638726599514484, + 0.022584397345781326, + 0.025089146569371223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.01820240542292595 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605910062789917 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.01605910062789917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605910062789917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.290701299905777, + 0.19279634952545166, + 0.14070382714271545, + 0.049089644104242325, + 0.07840660959482193, + 0.04094671830534935, + 0.03286997228860855, + 0.05720379948616028, + 0.04717708379030228, + 0.07010468095541 + ], + "credit_map": [ + 0.04140586033463478, + 0.03980423882603645, + 0.029246507212519646, + 0.038723014295101166, + 0.05490417033433914, + 0.030536090955138206, + 0.017687920480966568, + 0.03839349374175072, + 0.020184708759188652, + 0.024452833458781242, + 0.03555457666516304, + 0.021999280899763107, + 0.019177287817001343, + 0.024202851578593254, + 0.01841222308576107, + 0.028812583535909653, + 0.01596469059586525, + 0.016918187960982323, + 0.01490102056413889, + 0.018153425306081772, + 0.02341408096253872, + 0.014350572600960732, + 0.012826763093471527, + 0.02417508326470852, + 0.00825701653957367, + 0.003950863610953093, + 0.007538248784840107, + 0.003890428924933076, + 0.00797287281602621, + 0.007132751401513815, + 0.0043290285393595695, + 0.006018429528921843, + 0.013637473806738853, + 0.006217320915311575, + 0.009847727604210377, + 0.00783278327435255, + 0.009991656057536602, + 0.010334250517189503, + 0.009786531329154968, + 0.010758865624666214, + 0.004048717673867941, + 0.004192032385617495, + 0.005098850000649691, + 0.006672319956123829, + 0.004024246707558632, + 0.003992433659732342, + 0.0065031335689127445, + 0.006414986215531826, + 0.0035122602712363005, + 0.003556329058483243, + 0.004033198114484549, + 0.003510925453156233, + 0.0036132161039859056, + 0.003207962727174163, + 0.006593890022486448, + 0.004842192865908146, + 0.008009183220565319, + 0.0045069120824337006, + 0.007586495950818062, + 0.005677372682839632, + 0.0061754644848406315, + 0.005829563830047846, + 0.011484832502901554, + 0.007933976128697395, + 0.007310845889151096, + 0.003928143531084061, + 0.00549916410818696, + 0.003448289819061756, + 0.0071299076080322266, + 0.005269505549222231, + 0.008670791052281857, + 0.005920438561588526, + 0.008462926372885704, + 0.009803680703043938, + 0.007574687711894512, + 0.008649667724967003, + 0.009727615863084793, + 0.008047064766287804, + 0.009040127508342266, + 0.00879890937358141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01605910062789917 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019370617810636759 + ], + "opposite_logprob_tokens": [ + -6.251936912536621 + ], + "logprob_reward": -0.0019370617810636759, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019370617810636759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.630001425743103, + 0.07079500705003738, + 0.04185111075639725, + 0.04638351872563362, + 0.034978121519088745, + 0.028654640540480614, + 0.029906494542956352, + 0.05713987350463867, + 0.020976627245545387, + 0.03931310400366783 + ], + "credit_map": [ + 0.0991826206445694, + 0.06680856645107269, + 0.09968351572751999, + 0.10577211529016495, + 0.07010918110609055, + 0.05223458260297775, + 0.05000508204102516, + 0.08620581775903702, + 0.00852272380143404, + 0.010207361541688442, + 0.008881919085979462, + 0.01045987755060196, + 0.007630303036421537, + 0.00909736379981041, + 0.006303638219833374, + 0.009691823273897171, + 0.00549704534932971, + 0.0032102421391755342, + 0.008839537389576435, + 0.004875402431935072, + 0.005108642391860485, + 0.0073911952786147594, + 0.002675200579687953, + 0.004253853112459183, + 0.006688733119517565, + 0.004045120906084776, + 0.005880418233573437, + 0.006367737893015146, + 0.005916133057326078, + 0.005636068992316723, + 0.003994552418589592, + 0.007854757830500603, + 0.003217016812413931, + 0.0030555794946849346, + 0.0041032941080629826, + 0.00586437713354826, + 0.004404304549098015, + 0.004420978017151356, + 0.002905413042753935, + 0.007007163017988205, + 0.0038483068346977234, + 0.002608509501442313, + 0.004437125287950039, + 0.0036575456615537405, + 0.003201819024980068, + 0.003990004770457745, + 0.002385512227192521, + 0.004525821190327406, + 0.004155453760176897, + 0.00265562254935503, + 0.0039126258343458176, + 0.0032296546269208193, + 0.0038090103771537542, + 0.004112627822905779, + 0.003933691885322332, + 0.004097809083759785, + 0.008839530870318413, + 0.006337893195450306, + 0.006937143858522177, + 0.00623535830527544, + 0.004803062882274389, + 0.007824871689081192, + 0.008895530365407467, + 0.007266486529260874, + 0.003914588131010532, + 0.0021589500829577446, + 0.002066195709630847, + 0.001835198374465108, + 0.003793966257944703, + 0.0022845303174108267, + 0.003009492065757513, + 0.00191370933316648, + 0.005861471872776747, + 0.003502361476421356, + 0.0051200068555772305, + 0.0026636209804564714, + 0.0029093921184539795, + 0.005178332328796387, + 0.006045647896826267, + 0.008032270707190037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0019370617810636759 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774796664714813 + ], + "opposite_logprob_tokens": [ + -4.634774684906006 + ], + "logprob_reward": -0.009774796664714813, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774796664714813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25486674904823303, + 0.11095472425222397, + 0.08900174498558044, + 0.09276924282312393, + 0.04284370318055153, + 0.05812755599617958, + 0.10834784060716629, + 0.11214511096477509, + 0.0840049758553505, + 0.04693843424320221 + ], + "credit_map": [ + 0.04498305544257164, + 0.032177411019802094, + 0.03815054893493652, + 0.03227647393941879, + 0.02946646139025688, + 0.029046248644590378, + 0.01499740406870842, + 0.03376912325620651, + 0.011338204145431519, + 0.014035047963261604, + 0.015430528670549393, + 0.012704146094620228, + 0.009626099839806557, + 0.015372888185083866, + 0.009047660045325756, + 0.023400144651532173, + 0.008561983704566956, + 0.010693038813769817, + 0.010145424865186214, + 0.011151734739542007, + 0.0136353624984622, + 0.012016365304589272, + 0.009686322882771492, + 0.013111508451402187, + 0.011323780752718449, + 0.009270978160202503, + 0.010249452665448189, + 0.01575259491801262, + 0.009786292910575867, + 0.008614710532128811, + 0.010557994246482849, + 0.017213433980941772, + 0.004379819612950087, + 0.004532036371529102, + 0.005455502774566412, + 0.006424380000680685, + 0.004241142421960831, + 0.005997356027364731, + 0.005506193730980158, + 0.006307271309196949, + 0.008043092675507069, + 0.007905538193881512, + 0.006108176428824663, + 0.009604284539818764, + 0.006097005680203438, + 0.006460454780608416, + 0.005449250806123018, + 0.008459757082164288, + 0.01951771043241024, + 0.011940233409404755, + 0.014961921609938145, + 0.01224129181355238, + 0.009136675857007504, + 0.015165046788752079, + 0.012372625060379505, + 0.013012328185141087, + 0.021052440628409386, + 0.01082656066864729, + 0.011280187405645847, + 0.011330237612128258, + 0.019858235493302345, + 0.012772422283887863, + 0.012267247773706913, + 0.012757775373756886, + 0.013754482381045818, + 0.005993225611746311, + 0.010146691463887691, + 0.006386055611073971, + 0.014221818186342716, + 0.012318277731537819, + 0.010109814815223217, + 0.011074608191847801, + 0.007687099743634462, + 0.00785084255039692, + 0.004782368429005146, + 0.00813851784914732, + 0.004798867739737034, + 0.004763302858918905, + 0.004066093824803829, + 0.004851344041526318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.009774796664714813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal, static, extreme macro close-up, tightly framed from a slightly low angle, captures the raw moment of the confrontation with breathtaking clarity. A small, sleek black scorpion dominates the frame, its jet-black, carapace-like exoskeleton gleaming with a menacing sheen. Its body tenses, its pincers slightly open, poised, its distinctive, segmented tail arched high over its back in a characteristically menacing pose. At its tip, a sharp, yellow stinger gleams, poised to strike with deadly precision. Its target is a larger beetle larva, helplessly wriggling directly beneath it. The larva's soft, milky-white, C-shaped body, with its fleshy, translucent sections, twitches in slow, desperate waves. This dramatic encounter unfolds on a piece of rotting, damp wood, where the intricate texture of the dark veins and shredded fibers is clearly rendered. A soft, diffuse light, as if filtering through a forest canopy, meticulously captures the tiny, almost invisible hairs on the scorpion's legs and highlights the larva's damp, glistening surface, creating a scene of intense, suspended tension. The overall atmosphere is clean and calm, a raw and focused look at the predator-prey dynamic frozen in the moments before its violent end.", + "seed": 30779, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30779_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.033959916327148676, + "vlm_logprob": -0.033959916327148676, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.033959916327148676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11422501504421234 + ], + "opposite_logprob_tokens": [ + -2.239224910736084 + ], + "logprob_reward": -0.11422501504421234, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11422501504421234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11416399478912354, + 0.1293458491563797, + 0.023376449942588806, + 0.016531184315681458, + 0.033785391598939896, + 0.04653456434607506, + 0.2488684356212616, + 0.055768031626939774, + 0.28512242436408997, + 0.04650362953543663 + ], + "credit_map": [ + 0.01125025749206543, + 0.012576091103255749, + 0.008991426788270473, + 0.009025754407048225, + 0.01521306298673153, + 0.029618406668305397, + 0.008729936555027962, + 0.018759068101644516, + 0.013551386073231697, + 0.012869124300777912, + 0.018067926168441772, + 0.01746492087841034, + 0.019882341846823692, + 0.020852945744991302, + 0.01074946392327547, + 0.01590774394571781, + 0.004632387775927782, + 0.0019290567142888904, + 0.0026329453103244305, + 0.003192012431100011, + 0.0018476262921467423, + 0.004657554905861616, + 0.0026215186808258295, + 0.0018633499275892973, + 0.0017360906349495053, + 0.0020156551618129015, + 0.0013370433589443564, + 0.0019100930076092482, + 0.001797573990188539, + 0.0028676772490143776, + 0.0023957917001098394, + 0.002471259329468012, + 0.005031241569668055, + 0.006513677071779966, + 0.004029013682156801, + 0.002789859427139163, + 0.002689848653972149, + 0.006928286515176296, + 0.002239265013486147, + 0.0035642017610371113, + 0.00520993722602725, + 0.0052350424230098724, + 0.003686667885631323, + 0.002964779967442155, + 0.002847259398549795, + 0.014111603610217571, + 0.0025339333806186914, + 0.009945344179868698, + 0.026573920622467995, + 0.022636303678154945, + 0.025167055428028107, + 0.044005315750837326, + 0.01771238073706627, + 0.08176105469465256, + 0.014389324001967907, + 0.016623083502054214, + 0.006904476787894964, + 0.004071621224284172, + 0.0028540529310703278, + 0.0030924954917281866, + 0.0024175914004445076, + 0.02020997926592827, + 0.0021997366566210985, + 0.014018089510500431, + 0.04415841028094292, + 0.012619122862815857, + 0.019436083734035492, + 0.08986245095729828, + 0.006809063255786896, + 0.03035971336066723, + 0.012087684124708176, + 0.06978991627693176, + 0.007957489229738712, + 0.0028968672268092632, + 0.0039812372997403145, + 0.0030364070553332567, + 0.0032234969548881054, + 0.017159434035420418, + 0.0017464804695919156, + 0.006502219475805759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.11422501504421234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594803616404533 + ], + "opposite_logprob_tokens": [ + -3.895594835281372 + ], + "logprob_reward": -0.020594803616404533, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594803616404533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05069156363606453, + 0.037532977759838104, + 0.010838877409696579, + 0.009985776618123055, + 0.052666813135147095, + 0.127213254570961, + 0.057083774358034134, + 0.14775072038173676, + 0.39575281739234924, + 0.11048340797424316 + ], + "credit_map": [ + 0.005730938632041216, + 0.006424832157790661, + 0.0038752013351768255, + 0.0030234972946345806, + 0.018103845417499542, + 0.004425655584782362, + 0.003143342910334468, + 0.005964247509837151, + 0.004079766571521759, + 0.0037147346884012222, + 0.005350152961909771, + 0.005212312564253807, + 0.005899451207369566, + 0.005935439839959145, + 0.0027670490089803934, + 0.004574068356305361, + 0.001173872733488679, + 0.0009501447784714401, + 0.00140533281955868, + 0.0014124636072665453, + 0.0016486002132296562, + 0.001371027436107397, + 0.001669601770117879, + 0.0012078339932486415, + 0.0010214602807536721, + 0.0010194209171459079, + 0.000841112167108804, + 0.0012468044878914952, + 0.0014956255909055471, + 0.00103989418130368, + 0.0017532792408019304, + 0.0015681792283430696, + 0.007937331683933735, + 0.008757401257753372, + 0.00543971499428153, + 0.0046267383731901646, + 0.0037994664162397385, + 0.012609015218913555, + 0.0033221649937331676, + 0.006174978334456682, + 0.014150664210319519, + 0.013063660822808743, + 0.008379337377846241, + 0.0036152913235127926, + 0.005701782647520304, + 0.04871442914009094, + 0.004694010131061077, + 0.02889407053589821, + 0.012656785547733307, + 0.003618104849010706, + 0.004356029909104109, + 0.007509838789701462, + 0.0036913754884153605, + 0.01870548166334629, + 0.00303996866568923, + 0.003506192471832037, + 0.018816569820046425, + 0.011093608103692532, + 0.007531376555562019, + 0.007543592713773251, + 0.0064287991262972355, + 0.05446102097630501, + 0.005346944555640221, + 0.036528799682855606, + 0.072597436606884, + 0.01673509180545807, + 0.024175193160772324, + 0.10519033670425415, + 0.009168748743832111, + 0.06266747415065765, + 0.014217354357242584, + 0.09100119769573212, + 0.019518842920660973, + 0.007034987676888704, + 0.007426590658724308, + 0.0034699251409620047, + 0.0065661282278597355, + 0.043362341821193695, + 0.0032001961953938007, + 0.019904397428035736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.020594803616404533 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776095859706402 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776095859706402, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776095859706402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014388352632522583, + 0.028334498405456543, + 0.009028302505612373, + 0.0035574142821133137, + 0.010403873398900032, + 0.04281831532716751, + 0.045148491859436035, + 0.7292380332946777, + 0.06833834201097488, + 0.04874439910054207 + ], + "credit_map": [ + 0.0016685966402292252, + 0.001877897186204791, + 0.001117967301979661, + 0.0010688534239307046, + 0.003536185482516885, + 0.0021361385006457567, + 0.0009227936388924718, + 0.002059921156615019, + 0.0029026195406913757, + 0.002657441422343254, + 0.004021978005766869, + 0.0039007680024951696, + 0.004287390969693661, + 0.004786010831594467, + 0.0021895503159612417, + 0.0035887400154024363, + 0.0016671711346134543, + 0.0007557934732176363, + 0.001111628138460219, + 0.001251197885721922, + 0.0008845266420394182, + 0.0013649336760863662, + 0.0012093216646462679, + 0.0007837312296032906, + 0.00039614285924471915, + 0.0003245589032303542, + 0.0005041621625423431, + 0.0005334629095159471, + 0.00030300288926810026, + 0.0005299318581819534, + 0.0005138313863426447, + 0.00045232154661789536, + 0.001214058487676084, + 0.0018680640496313572, + 0.0016394851263612509, + 0.001179755199700594, + 0.001089723315089941, + 0.0010013919090852141, + 0.001272224122658372, + 0.001139171072281897, + 0.005097875837236643, + 0.004101000260561705, + 0.0029645580798387527, + 0.0013643903657793999, + 0.0018681335495784879, + 0.01639828085899353, + 0.0016589422011747956, + 0.009365134872496128, + 0.005214773118495941, + 0.0038113209884613752, + 0.0037335383240133524, + 0.008854866959154606, + 0.002335811033844948, + 0.016562744975090027, + 0.0020143326837569475, + 0.0026211063377559185, + 0.08962516486644745, + 0.05521271750330925, + 0.03554137051105499, + 0.03959178552031517, + 0.029697787016630173, + 0.2651609182357788, + 0.026968514546751976, + 0.18743981420993805, + 0.012375466525554657, + 0.0030339681543409824, + 0.0040544127114117146, + 0.016631001606583595, + 0.0017102971905842423, + 0.01243218220770359, + 0.002350266557186842, + 0.01575075462460518, + 0.007873494178056717, + 0.0035639991983771324, + 0.0036172103136777878, + 0.00245429459027946, + 0.0031452954281121492, + 0.01656046137213707, + 0.002425390761345625, + 0.009104253724217415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.009776095859706402 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632336553186178 + ], + "opposite_logprob_tokens": [ + -5.379632472991943 + ], + "logprob_reward": -0.004632336553186178, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632336553186178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04616045579314232, + 0.03408148139715195, + 0.009852415882050991, + 0.0047018444165587425, + 0.10849373042583466, + 0.13188867270946503, + 0.16574382781982422, + 0.35285308957099915, + 0.133702352643013, + 0.012522071599960327 + ], + "credit_map": [ + 0.005567439831793308, + 0.005765682086348534, + 0.003696793457493186, + 0.002627594629302621, + 0.015868445858359337, + 0.005929343868046999, + 0.0032845381647348404, + 0.0034206171985715628, + 0.004051665775477886, + 0.003785104025155306, + 0.005413189064711332, + 0.004607208073139191, + 0.005458042491227388, + 0.004716308321803808, + 0.0023952077608555555, + 0.0036547547206282616, + 0.0013390234671533108, + 0.0009001491125673056, + 0.0010489406995475292, + 0.0015190572012215853, + 0.0009239646606147289, + 0.00192149612121284, + 0.0010075620375573635, + 0.001192222349345684, + 0.0005418749642558396, + 0.000547736999578774, + 0.00046757192467339337, + 0.0005860757664777339, + 0.0006697172648273408, + 0.0005812909221276641, + 0.0007220100960694253, + 0.0005855667404830456, + 0.015380538068711758, + 0.01849864050745964, + 0.011088310740888119, + 0.009527591057121754, + 0.007776868529617786, + 0.025583606213331223, + 0.007212235126644373, + 0.013425948098301888, + 0.015446348115801811, + 0.013351261615753174, + 0.00874184537678957, + 0.003457210259512067, + 0.005662853363901377, + 0.051143769174814224, + 0.004623208660632372, + 0.02946218103170395, + 0.027960781008005142, + 0.012982750311493874, + 0.013945022597908974, + 0.024999216198921204, + 0.00826641172170639, + 0.061611682176589966, + 0.006222953554242849, + 0.00975501723587513, + 0.043575625866651535, + 0.026868892833590508, + 0.01707410253584385, + 0.019153006374835968, + 0.014339959248900414, + 0.12837442755699158, + 0.013019711710512638, + 0.09044736623764038, + 0.02178550884127617, + 0.006119343917816877, + 0.008367210626602173, + 0.03825371712446213, + 0.0030238868203014135, + 0.018299823626875877, + 0.00512909609824419, + 0.03272377699613571, + 0.0018669813871383667, + 0.0014327896060422063, + 0.0015278600621968508, + 0.001188970054499805, + 0.001311544794589281, + 0.0028133855666965246, + 0.0009527493384666741, + 0.0014277907321229577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.004632336553186178 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571330562233925 + ], + "opposite_logprob_tokens": [ + -3.895571231842041 + ], + "logprob_reward": -0.020571330562233925, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571330562233925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.008559065870940685, + 0.009364611469209194, + 0.0068308450281620026, + 0.006271353457123041, + 0.035917509347200394, + 0.06917635351419449, + 0.04031549394130707, + 0.6482952237129211, + 0.1178504079580307, + 0.057419173419475555 + ], + "credit_map": [ + 0.0007911510765552521, + 0.0011866091517731547, + 0.0009320054086856544, + 0.0008798142662271857, + 0.0013219723477959633, + 0.00128437508828938, + 0.0010788814397528768, + 0.001084257964976132, + 0.001118962187319994, + 0.001160358777269721, + 0.0016476111486554146, + 0.0009030151995830238, + 0.0017245120834559202, + 0.0012001491850242019, + 0.0007074564346112311, + 0.0009025464532896876, + 0.001139611005783081, + 0.0007744646281935275, + 0.0007923773373477161, + 0.0007111300947144628, + 0.0009513974655419588, + 0.0010105108376592398, + 0.0007208746974356472, + 0.0007304790196940303, + 0.0007287290063686669, + 0.0011026259744539857, + 0.0005943233845755458, + 0.0007373755215667188, + 0.0006033409154042602, + 0.0008261154289357364, + 0.0008644995396025479, + 0.0008143433951772749, + 0.005338897928595543, + 0.006144124548882246, + 0.0037772944197058678, + 0.003066413104534149, + 0.002606347668915987, + 0.008493592031300068, + 0.0022682552225887775, + 0.004222582560032606, + 0.008186540566384792, + 0.0067799994722008705, + 0.0047624739818274975, + 0.002159446245059371, + 0.0029891622252762318, + 0.026379698887467384, + 0.0026288314256817102, + 0.01529020071029663, + 0.0030603904742747545, + 0.003810533322393894, + 0.0039320397190749645, + 0.009613714180886745, + 0.0023457761853933334, + 0.013196536339819431, + 0.0021218103356659412, + 0.002234697574749589, + 0.07991808652877808, + 0.04928253963589668, + 0.031524356454610825, + 0.03529490530490875, + 0.02634664811193943, + 0.23551329970359802, + 0.023953473195433617, + 0.16646189987659454, + 0.022198714315891266, + 0.006749656051397324, + 0.006550043821334839, + 0.013063660822808743, + 0.004714159294962883, + 0.0360909067094326, + 0.0033710033167153597, + 0.025112265720963478, + 0.008699735626578331, + 0.005010955035686493, + 0.006061116699129343, + 0.00463761854916811, + 0.00506033468991518, + 0.014913833700120449, + 0.004053767304867506, + 0.008981817401945591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.020571330562233925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal, static, extreme macro close-up, tightly framed from a slightly low angle, captures the raw moment of the confrontation with breathtaking clarity. A small, sleek black scorpion dominates the frame, its jet-black, carapace-like exoskeleton gleaming with a menacing sheen. Its body tenses, its pincers slightly open, poised, its distinctive, segmented tail arched high over its back in a characteristically menacing pose. At its tip, a sharp, yellow stinger gleams, poised to strike with deadly precision. Its target is a larger beetle larva, helplessly wriggling directly beneath it. The larva's soft, milky-white, C-shaped body, with its fleshy, translucent sections, twitches in slow, desperate waves. This dramatic encounter unfolds on a piece of rotting, damp wood, where the intricate texture of the dark veins and shredded fibers is clearly rendered. A soft, diffuse light, as if filtering through a forest canopy, meticulously captures the tiny, almost invisible hairs on the scorpion's legs and highlights the larva's damp, glistening surface, creating a scene of intense, suspended tension. The overall atmosphere is clean and calm, a raw and focused look at the predator-prey dynamic frozen in the moments before its violent end.", + "seed": 30780, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30780_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012107368465512991, + "vlm_logprob": -0.012107368465512991, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.012107368465512991, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013894733041524887 + ], + "opposite_logprob_tokens": [ + -4.388894557952881 + ], + "logprob_reward": -0.013894733041524887, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013894733041524887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1327052265405655, + 0.45021218061447144, + 0.05840536579489708, + 0.02564997971057892, + 0.05346107482910156, + 0.01827193610370159, + 0.08975119888782501, + 0.059820543974637985, + 0.06174897402524948, + 0.04997348412871361 + ], + "credit_map": [ + 0.030357759445905685, + 0.015128743834793568, + 0.01310215424746275, + 0.012564252130687237, + 0.020209377631545067, + 0.020475847646594048, + 0.010899899527430534, + 0.009967192076146603, + 0.04329518973827362, + 0.07472054660320282, + 0.043561480939388275, + 0.06413043290376663, + 0.08360527455806732, + 0.05133349448442459, + 0.056849297136068344, + 0.03271647170186043, + 0.00980411283671856, + 0.0071090105921030045, + 0.0076495627872645855, + 0.0062926034443080425, + 0.006802079267799854, + 0.006613901350647211, + 0.008498193696141243, + 0.005635898094624281, + 0.004659438505768776, + 0.0031440742313861847, + 0.0024741075467318296, + 0.0030167440418154, + 0.003363550640642643, + 0.002892852760851383, + 0.0034574544988572598, + 0.00264175608754158, + 0.013054211623966694, + 0.0048021795228123665, + 0.005742538254708052, + 0.003882428165525198, + 0.004380608443170786, + 0.014356480911374092, + 0.004179093055427074, + 0.0030635357834398746, + 0.003005129285156727, + 0.0020313444547355175, + 0.0020990546327084303, + 0.0022081532515585423, + 0.0017679734155535698, + 0.002225752454251051, + 0.0031690155155956745, + 0.0017655132105574012, + 0.010558209381997585, + 0.008860020898282528, + 0.01812790334224701, + 0.011252671480178833, + 0.011208786629140377, + 0.011508811265230179, + 0.007755439728498459, + 0.010479358024895191, + 0.0087796151638031, + 0.005575582385063171, + 0.01094284001737833, + 0.005074466113001108, + 0.006991575006395578, + 0.007646835874766111, + 0.006558755878359079, + 0.008250873535871506, + 0.011375488713383675, + 0.006466094870120287, + 0.011770352721214294, + 0.003765180939808488, + 0.006822486408054829, + 0.00836609210819006, + 0.005567533895373344, + 0.0076157450675964355, + 0.009872524999082088, + 0.007282396778464317, + 0.005723882000893354, + 0.00514239864423871, + 0.006260111462324858, + 0.0065981550142169, + 0.004616673104465008, + 0.004477338399738073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.013894733041524887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301351815462112 + ], + "opposite_logprob_tokens": [ + -3.773301362991333 + ], + "logprob_reward": -0.023301351815462112, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301351815462112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2427159547805786, + 0.09231913089752197, + 0.059067774564027786, + 0.09712465107440948, + 0.14110411703586578, + 0.0480436235666275, + 0.13023439049720764, + 0.025637539103627205, + 0.14358310401439667, + 0.020169703289866447 + ], + "credit_map": [ + 0.05069641396403313, + 0.024828430265188217, + 0.032335080206394196, + 0.02488642744719982, + 0.02921309880912304, + 0.030021946877241135, + 0.02753099985420704, + 0.023203587159514427, + 0.012784533202648163, + 0.013340676203370094, + 0.009722772054374218, + 0.01165841706097126, + 0.01531139761209488, + 0.011982987634837627, + 0.009031083434820175, + 0.008487261831760406, + 0.009242888540029526, + 0.0072355917654931545, + 0.007333310320973396, + 0.007455141749233007, + 0.005913148168474436, + 0.007074116263538599, + 0.010089116171002388, + 0.004724464379251003, + 0.02160879597067833, + 0.01130102202296257, + 0.011119109578430653, + 0.010143779218196869, + 0.011749466881155968, + 0.013786518014967442, + 0.008702605031430721, + 0.008713353425264359, + 0.04334640130400658, + 0.0072078038938343525, + 0.009189750999212265, + 0.0053586652502417564, + 0.010550135746598244, + 0.05500537529587746, + 0.004641486331820488, + 0.005804494023323059, + 0.00962673407047987, + 0.006209243554621935, + 0.005977800115942955, + 0.003910582512617111, + 0.004968729335814714, + 0.010218205861747265, + 0.0038904445245862007, + 0.0032418789342045784, + 0.021105285733938217, + 0.013619480654597282, + 0.0314457006752491, + 0.009888106025755405, + 0.013433142565190792, + 0.018199816346168518, + 0.00706193782389164, + 0.015480907633900642, + 0.003784951288253069, + 0.0027458008844405413, + 0.0019602407701313496, + 0.0023754292633384466, + 0.006808230187743902, + 0.003677285509184003, + 0.002570497104898095, + 0.00171510293148458, + 0.024455975741147995, + 0.02121785841882229, + 0.009741929359734058, + 0.013379757292568684, + 0.019685693085193634, + 0.01723172329366207, + 0.02227831818163395, + 0.015591860748827457, + 0.002745093544945121, + 0.002740744734182954, + 0.0020826507825404406, + 0.002066665329039097, + 0.0032463849056512117, + 0.002474667737260461, + 0.0024695107713341713, + 0.002343983855098486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.023301351815462112 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031911423429846764 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031911423429846764, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031911423429846764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17267228662967682, + 0.24708083271980286, + 0.14179229736328125, + 0.04719771817326546, + 0.1467050164937973, + 0.021907299757003784, + 0.060423143208026886, + 0.08263862878084183, + 0.047253724187612534, + 0.03232898935675621 + ], + "credit_map": [ + 0.025484606623649597, + 0.014181525446474552, + 0.01833251863718033, + 0.016944028437137604, + 0.02220339886844158, + 0.021825598552823067, + 0.034801144152879715, + 0.018899474292993546, + 0.02251649647951126, + 0.03322116285562515, + 0.031006598845124245, + 0.03192639350891113, + 0.044412270188331604, + 0.02685309387743473, + 0.036021359264850616, + 0.021123463287949562, + 0.015580118633806705, + 0.014183578081429005, + 0.01836872473359108, + 0.01573919504880905, + 0.015766924247145653, + 0.018789559602737427, + 0.028680704534053802, + 0.014683506451547146, + 0.009847951121628284, + 0.0052794707007706165, + 0.007287094835191965, + 0.0048956493847072124, + 0.005538389086723328, + 0.0062137870118021965, + 0.0037296863738447428, + 0.004405694548040628, + 0.044227976351976395, + 0.006009039934724569, + 0.011709138751029968, + 0.005823873914778233, + 0.010484584607183933, + 0.057053372263908386, + 0.005334693938493729, + 0.006062336266040802, + 0.0031928077805787325, + 0.0019929951522499323, + 0.004766426980495453, + 0.0019127854611724615, + 0.0021823844872415066, + 0.003323678160086274, + 0.002240358851850033, + 0.002295861719176173, + 0.008954346179962158, + 0.005335357505828142, + 0.01377631165087223, + 0.005180038511753082, + 0.008105644956231117, + 0.008323793299496174, + 0.004354248289018869, + 0.0063933986239135265, + 0.009784029796719551, + 0.011127867735922337, + 0.009052460081875324, + 0.009813157841563225, + 0.01347140409052372, + 0.012025398202240467, + 0.007551279850304127, + 0.009813029319047928, + 0.007154006045311689, + 0.004628919064998627, + 0.011182782240211964, + 0.0034487047232687473, + 0.0050766426138579845, + 0.005279680714011192, + 0.0043844254687428474, + 0.0060985637828707695, + 0.0050102341920137405, + 0.004137995652854443, + 0.005040299613028765, + 0.0030283411033451557, + 0.004004065413028002, + 0.003900136100128293, + 0.003970768768340349, + 0.0032371487468481064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0031911423429846764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090514499694109 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.004090514499694109, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090514499694109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09418856352567673, + 0.058463845402002335, + 0.0952935516834259, + 0.06067037209868431, + 0.0617758110165596, + 0.07911132276058197, + 0.048603955656290054, + 0.3890417516231537, + 0.08689647167921066, + 0.025954395532608032 + ], + "credit_map": [ + 0.01959959603846073, + 0.008587679825723171, + 0.013739801943302155, + 0.008945642970502377, + 0.01229353342205286, + 0.010943086817860603, + 0.009277243167161942, + 0.010801981203258038, + 0.006398126017302275, + 0.011393341235816479, + 0.006665219087153673, + 0.009311335161328316, + 0.008155674673616886, + 0.0064791725017130375, + 0.0050154635682702065, + 0.0050455122254788876, + 0.015245304442942142, + 0.00947397481650114, + 0.014031934551894665, + 0.008083360269665718, + 0.011107923462986946, + 0.010280395857989788, + 0.012161590158939362, + 0.014909067191183567, + 0.014411031268537045, + 0.007495551370084286, + 0.00667906180024147, + 0.006333813536912203, + 0.00864955224096775, + 0.008122905157506466, + 0.004883973393589258, + 0.004094487056136131, + 0.010110561735928059, + 0.004193177446722984, + 0.013161317445337772, + 0.006575644947588444, + 0.004708017688244581, + 0.013040709309279919, + 0.005162594374269247, + 0.004823789000511169, + 0.014703478664159775, + 0.009994088672101498, + 0.009675301611423492, + 0.006847567390650511, + 0.010446101427078247, + 0.013044730760157108, + 0.008828576654195786, + 0.005571470595896244, + 0.007782929111272097, + 0.005051441956311464, + 0.010155054740607738, + 0.004279313143342733, + 0.005131491459906101, + 0.0064735738560557365, + 0.0032342823687940836, + 0.006495869252830744, + 0.06783369928598404, + 0.04039411246776581, + 0.055449120700359344, + 0.026055751368403435, + 0.06286796927452087, + 0.05364012345671654, + 0.027896655723452568, + 0.05490429699420929, + 0.011152115650475025, + 0.008758710697293282, + 0.012684090994298458, + 0.007297706790268421, + 0.009238730184733868, + 0.009170693345367908, + 0.01660376600921154, + 0.011990662664175034, + 0.002730581210926175, + 0.002185339340940118, + 0.0056663379073143005, + 0.0030605511274188757, + 0.0026338214520365, + 0.002832138678058982, + 0.003289557062089443, + 0.003556069452315569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.004090514499694109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605910062789917 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.01605910062789917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605910062789917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11538409441709518, + 0.10838343948125839, + 0.0682041347026825, + 0.07004406303167343, + 0.20619624853134155, + 0.08486772328615189, + 0.10058977454900742, + 0.09390804171562195, + 0.09840861707925797, + 0.05401381477713585 + ], + "credit_map": [ + 0.01763646863400936, + 0.015325658954679966, + 0.011275527998805046, + 0.01508315559476614, + 0.013666371814906597, + 0.012255118228495121, + 0.01976381056010723, + 0.010378002189099789, + 0.013045438565313816, + 0.014175689779222012, + 0.014088134281337261, + 0.014129137620329857, + 0.015878910198807716, + 0.012762312777340412, + 0.015238642692565918, + 0.009065184742212296, + 0.01690858043730259, + 0.008153483271598816, + 0.005603978410363197, + 0.005648768041282892, + 0.007899108342826366, + 0.009929993189871311, + 0.0066475775092840195, + 0.007412643171846867, + 0.012572224251925945, + 0.007612130604684353, + 0.008823038078844547, + 0.007482441142201424, + 0.00861295498907566, + 0.008752541616559029, + 0.006438766606152058, + 0.009749970398843288, + 0.06594947725534439, + 0.010595153085887432, + 0.007537435740232468, + 0.007260199636220932, + 0.014828626997768879, + 0.08437764644622803, + 0.006843696814030409, + 0.008804021403193474, + 0.01673487015068531, + 0.011059173382818699, + 0.00832398608326912, + 0.008293593302369118, + 0.008785394951701164, + 0.021082060411572456, + 0.005646777804940939, + 0.004941868595778942, + 0.010373503901064396, + 0.0203096941113472, + 0.008088729344308376, + 0.010593526065349579, + 0.019583730027079582, + 0.013698210939764977, + 0.010031217709183693, + 0.007911167107522488, + 0.015226701274514198, + 0.01104787364602089, + 0.008284424431622028, + 0.008672880008816719, + 0.02128411829471588, + 0.014098158106207848, + 0.007388192228972912, + 0.007905696518719196, + 0.02146266959607601, + 0.01398826390504837, + 0.012340578250586987, + 0.007653775624930859, + 0.012963899411261082, + 0.012108419090509415, + 0.008395301178097725, + 0.009495720267295837, + 0.008922191336750984, + 0.006707407534122467, + 0.007492653094232082, + 0.005500308703631163, + 0.0060857245698571205, + 0.006801648065447807, + 0.00591065501794219, + 0.006593230180442333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.01605910062789917 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal, static, extreme macro close-up, tightly framed from a slightly low angle, captures the raw moment of the confrontation with breathtaking clarity. A small, sleek black scorpion dominates the frame, its jet-black, carapace-like exoskeleton gleaming with a menacing sheen. Its body tenses, its pincers slightly open, poised, its distinctive, segmented tail arched high over its back in a characteristically menacing pose. At its tip, a sharp, yellow stinger gleams, poised to strike with deadly precision. Its target is a larger beetle larva, helplessly wriggling directly beneath it. The larva's soft, milky-white, C-shaped body, with its fleshy, translucent sections, twitches in slow, desperate waves. This dramatic encounter unfolds on a piece of rotting, damp wood, where the intricate texture of the dark veins and shredded fibers is clearly rendered. A soft, diffuse light, as if filtering through a forest canopy, meticulously captures the tiny, almost invisible hairs on the scorpion's legs and highlights the larva's damp, glistening surface, creating a scene of intense, suspended tension. The overall atmosphere is clean and calm, a raw and focused look at the predator-prey dynamic frozen in the moments before its violent end.", + "seed": 30781, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_7/index_2825_seed_30781_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.008286567381583154, + "vlm_logprob": -0.008286567381583154, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.008286567381583154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00513733783736825 + ], + "opposite_logprob_tokens": [ + -5.7551374435424805 + ], + "logprob_reward": -0.00513733783736825, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00513733783736825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16455166041851044, + 0.08105221390724182, + 0.19210049510002136, + 0.04358125850558281, + 0.11087794601917267, + 0.11627291887998581, + 0.2086712270975113, + 0.025018993765115738, + 0.022087564691901207, + 0.0357857383787632 + ], + "credit_map": [ + 0.01873171329498291, + 0.021329212933778763, + 0.016969405114650726, + 0.028751036152243614, + 0.015157578513026237, + 0.0221414752304554, + 0.015895776450634003, + 0.02557547017931938, + 0.009643434546887875, + 0.009489073418080807, + 0.010763023048639297, + 0.011356856673955917, + 0.010022355243563652, + 0.01093363855034113, + 0.00944626796990633, + 0.009397564455866814, + 0.02386360615491867, + 0.013133613392710686, + 0.03873610123991966, + 0.01701905019581318, + 0.02101360261440277, + 0.03196469694375992, + 0.020007148385047913, + 0.026362674310803413, + 0.005459748208522797, + 0.00420784018933773, + 0.00597518403083086, + 0.0060766879469156265, + 0.004609156399965286, + 0.00806941743940115, + 0.004712576046586037, + 0.004470652900636196, + 0.0142299123108387, + 0.011026744730770588, + 0.017050785943865776, + 0.011432639323174953, + 0.01403547078371048, + 0.0221633929759264, + 0.009383666329085827, + 0.011555337347090244, + 0.01201451476663351, + 0.019685544073581696, + 0.01873432844877243, + 0.019419854506850243, + 0.010354677215218544, + 0.010613517835736275, + 0.01445514801889658, + 0.01099533773958683, + 0.02572478912770748, + 0.03454984724521637, + 0.021234741434454918, + 0.030828848481178284, + 0.018443549051880836, + 0.024479428306221962, + 0.026669079437851906, + 0.02674095705151558, + 0.00406932458281517, + 0.0016253333305940032, + 0.004669824615120888, + 0.0019830577075481415, + 0.001987595111131668, + 0.004887051414698362, + 0.0024346031714230776, + 0.0033622048795223236, + 0.0038053046446293592, + 0.0021718102507293224, + 0.0023856398183852434, + 0.00282678403891623, + 0.002278366591781378, + 0.002812339924275875, + 0.0024975244887173176, + 0.003309795167297125, + 0.005137462168931961, + 0.004320371430367231, + 0.0036292902659624815, + 0.003524123691022396, + 0.004148670006543398, + 0.004472946282476187, + 0.005387488286942244, + 0.005165387410670519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.00513733783736825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201176352798939 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201176352798939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201176352798939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24392321705818176, + 0.08096762746572495, + 0.11838595569133759, + 0.04675837233662605, + 0.09102986007928848, + 0.09577606618404388, + 0.10533870756626129, + 0.09046325087547302, + 0.05131727084517479, + 0.07603970170021057 + ], + "credit_map": [ + 0.024259846657514572, + 0.018446184694767, + 0.03131149709224701, + 0.07715688645839691, + 0.01370904128998518, + 0.013666313141584396, + 0.021088335663080215, + 0.044285088777542114, + 0.008418033830821514, + 0.01012616790831089, + 0.0111650750041008, + 0.010606501251459122, + 0.011725434102118015, + 0.009251467883586884, + 0.010573063977062702, + 0.009101876057684422, + 0.016107043251395226, + 0.015483745373785496, + 0.016182443127036095, + 0.01640103943645954, + 0.012587404809892178, + 0.0142366336658597, + 0.012461218051612377, + 0.014926417730748653, + 0.004691319540143013, + 0.006346350070089102, + 0.005615700501948595, + 0.005239738617092371, + 0.005977793596684933, + 0.008329663425683975, + 0.005490052979439497, + 0.0050677526742219925, + 0.010358509607613087, + 0.01000948529690504, + 0.014304540120065212, + 0.009126025252044201, + 0.009204122237861156, + 0.02050120383501053, + 0.007631028536707163, + 0.009894953109323978, + 0.009439061395823956, + 0.020078755915164948, + 0.010776865296065807, + 0.015415754169225693, + 0.007717594970017672, + 0.008593578822910786, + 0.012522352859377861, + 0.011232111603021622, + 0.015115673653781414, + 0.014792876318097115, + 0.013787276111543179, + 0.014283943921327591, + 0.009337104856967926, + 0.014864404685795307, + 0.010968944989144802, + 0.012188487686216831, + 0.011346769519150257, + 0.0137378154322505, + 0.011423719115555286, + 0.00978902354836464, + 0.009487435221672058, + 0.010922675020992756, + 0.010665449313819408, + 0.01309036836028099, + 0.007113868370652199, + 0.005236304365098476, + 0.005997586064040661, + 0.007596054580062628, + 0.005134850274771452, + 0.008353905752301216, + 0.006000163033604622, + 0.0058845337480306625, + 0.009178263135254383, + 0.009653134271502495, + 0.009007994085550308, + 0.00966411828994751, + 0.00772802485153079, + 0.012108344584703445, + 0.008748620748519897, + 0.009951197542250156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.014201176352798939 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067322817631065845 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067322817631065845, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067322817631065845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17345274984836578, + 0.22657553851604462, + 0.129860058426857, + 0.06427159160375595, + 0.09088364988565445, + 0.12785901129245758, + 0.06621982157230377, + 0.01763627678155899, + 0.0720510482788086, + 0.031190237030386925 + ], + "credit_map": [ + 0.0290437750518322, + 0.015585060231387615, + 0.025490855798125267, + 0.04248238354921341, + 0.014096343889832497, + 0.012193761765956879, + 0.011057132855057716, + 0.023503432050347328, + 0.022857610136270523, + 0.021209316328167915, + 0.027339601889252663, + 0.025645926594734192, + 0.020900247618556023, + 0.027947675436735153, + 0.03592434898018837, + 0.04475081339478493, + 0.016110841184854507, + 0.010370801202952862, + 0.024350130930542946, + 0.013277844525873661, + 0.012462832033634186, + 0.023706752806901932, + 0.013442476280033588, + 0.016138385981321335, + 0.007383239921182394, + 0.005154980346560478, + 0.011198586784303188, + 0.005854005925357342, + 0.006953911390155554, + 0.01318140048533678, + 0.005293661262840033, + 0.009251806885004044, + 0.011028885841369629, + 0.011232256889343262, + 0.011739527806639671, + 0.01111355610191822, + 0.010353652760386467, + 0.015068462118506432, + 0.00895218551158905, + 0.01139511726796627, + 0.015241295099258423, + 0.015893014147877693, + 0.021966559812426567, + 0.018176589161157608, + 0.013297252357006073, + 0.013035675510764122, + 0.017693674191832542, + 0.012554952874779701, + 0.008409329690039158, + 0.009965432807803154, + 0.008277587592601776, + 0.00979873351752758, + 0.005422055721282959, + 0.007401136681437492, + 0.008777384646236897, + 0.008168162778019905, + 0.0028360995929688215, + 0.0021410395856946707, + 0.0021772144827991724, + 0.0020164938177913427, + 0.0015850429190322757, + 0.0026928249280899763, + 0.002373450668528676, + 0.0018141113687306643, + 0.009230740368366241, + 0.008278079330921173, + 0.009546348825097084, + 0.010889669880270958, + 0.006981366779655218, + 0.010154662653803825, + 0.008893613703548908, + 0.008076565340161324, + 0.0049410792998969555, + 0.0037588742561638355, + 0.0050387633964419365, + 0.003022886347025633, + 0.003252545604482293, + 0.0035893376916646957, + 0.004826332908123732, + 0.0027604177594184875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0067322817631065845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011762612266466022 + ], + "opposite_logprob_tokens": [ + -6.751176357269287 + ], + "logprob_reward": -0.0011762612266466022, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011762612266466022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1802881807088852, + 0.120389923453331, + 0.19097954034805298, + 0.09972351789474487, + 0.08526904881000519, + 0.1058378741145134, + 0.023064931854605675, + 0.05726673826575279, + 0.08440754562616348, + 0.0527726411819458 + ], + "credit_map": [ + 0.020082339644432068, + 0.025067301467061043, + 0.02846050076186657, + 0.019821643829345703, + 0.0170306209474802, + 0.02101645991206169, + 0.013580692932009697, + 0.03522862493991852, + 0.012832296080887318, + 0.011537354439496994, + 0.01504611223936081, + 0.011619663797318935, + 0.017555834725499153, + 0.025372928008437157, + 0.012010355480015278, + 0.014415371231734753, + 0.023763706907629967, + 0.015368417836725712, + 0.03860701993107796, + 0.020145395770668983, + 0.01713648810982704, + 0.03237941488623619, + 0.018695969134569168, + 0.024883130565285683, + 0.012207133695483208, + 0.007244700565934181, + 0.02125961147248745, + 0.009323417209088802, + 0.010251619853079319, + 0.02221287600696087, + 0.006726681720465422, + 0.01049747597426176, + 0.010314816609025002, + 0.0076433508656919, + 0.013284314423799515, + 0.007266136351972818, + 0.010923666879534721, + 0.01981588639318943, + 0.006597170140594244, + 0.009423704817891121, + 0.01109227817505598, + 0.016074163839221, + 0.012317635118961334, + 0.015507959760725498, + 0.009490478783845901, + 0.009648144245147705, + 0.014561912044882774, + 0.017145304009318352, + 0.004042507614940405, + 0.0020870943553745747, + 0.0023561748676002026, + 0.0027611153200268745, + 0.0023359465412795544, + 0.004503676667809486, + 0.0023489894811064005, + 0.002629427704960108, + 0.01099899411201477, + 0.005800014827400446, + 0.007752006407827139, + 0.005102517548948526, + 0.004553890321403742, + 0.009905974380671978, + 0.006453811656683683, + 0.006699529476463795, + 0.011730898171663284, + 0.008303863927721977, + 0.01173087302595377, + 0.009064164012670517, + 0.008291141130030155, + 0.012389645911753178, + 0.009787997230887413, + 0.013108963146805763, + 0.007369270548224449, + 0.005854190327227116, + 0.00848939549177885, + 0.004883093293756247, + 0.005818331614136696, + 0.007050401996821165, + 0.008761227130889893, + 0.004546726122498512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0011762612266466022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a small, sleek black scorpion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scorpion's body tensed with pincers slightly open and its segmented tail arched high over its back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a larger beetle larva visible directly beneath the scorpion, wriggling in a soft, milky-white C-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a piece of rotting, damp wood with clearly rendered dark veins and shredded fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera framed from a slightly low angle in a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014185779727995396 + ], + "opposite_logprob_tokens": [ + -4.264185905456543 + ], + "logprob_reward": -0.014185779727995396, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014185779727995396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17372038960456848, + 0.16379398107528687, + 0.15094785392284393, + 0.055511459708213806, + 0.12472300976514816, + 0.07709395885467529, + 0.0691210925579071, + 0.03546648472547531, + 0.09943798184394836, + 0.05018378421664238 + ], + "credit_map": [ + 0.02733236737549305, + 0.013329452835023403, + 0.019410906359553337, + 0.04207456111907959, + 0.013395470567047596, + 0.01553276740014553, + 0.00943843089044094, + 0.03320644795894623, + 0.017407266423106194, + 0.016550812870264053, + 0.023972932249307632, + 0.014436068944633007, + 0.024199238047003746, + 0.030961774289608, + 0.016402533277869225, + 0.019863374531269073, + 0.01669527031481266, + 0.01953902468085289, + 0.018193941563367844, + 0.022626059129834175, + 0.01784164272248745, + 0.01639746129512787, + 0.01712373085319996, + 0.022530732676386833, + 0.0059471880085766315, + 0.005910780746489763, + 0.00672107981517911, + 0.007050432730466127, + 0.005851003807038069, + 0.010418949648737907, + 0.006435612682253122, + 0.007176410872489214, + 0.014640060253441334, + 0.012971269898116589, + 0.019581789150834084, + 0.011535844765603542, + 0.014702929183840752, + 0.026254769414663315, + 0.010686743073165417, + 0.014349604956805706, + 0.010761665180325508, + 0.00778074050322175, + 0.014768420718610287, + 0.009191536344587803, + 0.00447316188365221, + 0.008207869715988636, + 0.008491603657603264, + 0.013418969698250294, + 0.009042243473231792, + 0.010239735245704651, + 0.010059312917292118, + 0.010054518468677998, + 0.004904760979115963, + 0.00826194230467081, + 0.009278623387217522, + 0.0072799548506736755, + 0.004869250115007162, + 0.005396324209868908, + 0.004415679723024368, + 0.0034986159298568964, + 0.00297456793487072, + 0.005331764928996563, + 0.004540317226201296, + 0.004439965356141329, + 0.013169370591640472, + 0.011219299398362637, + 0.010814276523888111, + 0.0161066222935915, + 0.01059397030621767, + 0.013682064600288868, + 0.012779227457940578, + 0.011073162779211998, + 0.005345479119569063, + 0.007814882323145866, + 0.006505679804831743, + 0.007078514900058508, + 0.00491191353648901, + 0.004194429609924555, + 0.009585891850292683, + 0.004746993537992239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014185779727995396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_8/index_tensor([3933, 3933])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_8/index_tensor([3933, 3933])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a2eb51bd5fb3f6259bbf57e08e4ff591bf254e7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_8/index_tensor([3933, 3933])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "At night, a dramatic low-angle shot pans slowly sideways across an abandoned outdoor basketball court, focusing on a lone athlete in the center of the frame. The athlete is a highly focused figure, with short, sweat-soaked hair and dressed in a simple, dark, sleeveless sweatshirt and matching shorts that cling slightly to their muscular frame. They maintain a low, dribbling stance, dribbling the ball rhythmically with steady, powerful hands. Every impact of the orange ball on the worn, painted concrete is deliberate and controlled. The towering, bright stadium floodlights pierce the inky night sky, casting an intense, high-contrast glow that sculpts the athletes' forms and captures the sheen of sweat on their skin. This intense, direct light casts distinctly elongated and crisp shadows of the players and the bouncing ball, a dark twin that stretches far across the empty court. The surrounding stands are engulfed in a vast, silent darkness, amplifying the profound sense of solitude and creating a stark, focused atmosphere of pure, meditative devotion.", + "seed": 7574, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_8/index_3933_seed_7574_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03743065083399415, + "vlm_logprob": -0.03743065083399415, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03743065083399415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04220516234636307 + ], + "opposite_logprob_tokens": [ + -3.5422050952911377 + ], + "logprob_reward": -0.04220516234636307, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04220516234636307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1491335779428482, + 0.04032600671052933, + 0.4587760269641876, + 0.06300348788499832, + 0.035986851900815964, + 0.09449636936187744, + 0.04314461350440979, + 0.04086286947131157, + 0.03680027648806572, + 0.037469957023859024 + ], + "credit_map": [ + 0.01418221928179264, + 0.01866786926984787, + 0.014000567607581615, + 0.019977107644081116, + 0.026133853942155838, + 0.021216347813606262, + 0.013558140955865383, + 0.02139747515320778, + 0.0031523120123893023, + 0.004818223882466555, + 0.0029595615342259407, + 0.007589382119476795, + 0.008568276651203632, + 0.0068023549392819405, + 0.0028800186701118946, + 0.0035558741074055433, + 0.0464356504380703, + 0.03820641338825226, + 0.0828825831413269, + 0.04071715474128723, + 0.04806595668196678, + 0.055784087628126144, + 0.06029630824923515, + 0.08638784289360046, + 0.007074587978422642, + 0.007761059328913689, + 0.007433747407048941, + 0.008138651959598064, + 0.007927659898996353, + 0.007890252396464348, + 0.010188806802034378, + 0.0065887197852134705, + 0.005789563991129398, + 0.005599457770586014, + 0.004018154926598072, + 0.003325114492326975, + 0.0036752987653017044, + 0.004814947489649057, + 0.006338509265333414, + 0.0024258012417703867, + 0.014827159233391285, + 0.009808795526623726, + 0.011095168069005013, + 0.012227104976773262, + 0.007637610659003258, + 0.008256764151155949, + 0.015189366415143013, + 0.015454384498298168, + 0.006902496330440044, + 0.003587554907426238, + 0.005711620207875967, + 0.004255706910043955, + 0.003173715202137828, + 0.007101267110556364, + 0.00727060716599226, + 0.005141642410308123, + 0.006041411310434341, + 0.0038802067283540964, + 0.00403574900701642, + 0.004608332645148039, + 0.006851965095847845, + 0.00559718394652009, + 0.007235018070787191, + 0.002612998941913247, + 0.00614035502076149, + 0.004020323511213064, + 0.006790854036808014, + 0.004631167743355036, + 0.0033680933993309736, + 0.004065179731696844, + 0.004558497574180365, + 0.0032258026767522097, + 0.0051639499142766, + 0.004164998419582844, + 0.0032993280328810215, + 0.004555935971438885, + 0.002866365946829319, + 0.004753841087222099, + 0.0060968464240431786, + 0.0065686870366334915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.04220516234636307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030611604452133 + ], + "opposite_logprob_tokens": [ + -2.3503060340881348 + ], + "logprob_reward": -0.10030611604452133, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030611604452133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11682973802089691, + 0.08230449259281158, + 0.15522019565105438, + 0.08897776156663895, + 0.044040143489837646, + 0.17913363873958588, + 0.05661574751138687, + 0.07419836521148682, + 0.14245975017547607, + 0.0602201409637928 + ], + "credit_map": [ + 0.015302513726055622, + 0.014800237491726875, + 0.007217682432383299, + 0.012917851097881794, + 0.028459656983613968, + 0.018463317304849625, + 0.00730513408780098, + 0.012363339774310589, + 0.007913830690085888, + 0.01214474905282259, + 0.005434100981801748, + 0.02302241511642933, + 0.00946187786757946, + 0.01289207860827446, + 0.005721234250813723, + 0.005714201368391514, + 0.013850376941263676, + 0.01764179952442646, + 0.025131378322839737, + 0.018622199073433876, + 0.015056355856359005, + 0.018592143431305885, + 0.02262316644191742, + 0.023702770471572876, + 0.008513801731169224, + 0.017463697120547295, + 0.008329681120812893, + 0.010020696558058262, + 0.014029036276042461, + 0.010394618846476078, + 0.011523491702973843, + 0.008702735416591167, + 0.00698491558432579, + 0.003282403340563178, + 0.004021842498332262, + 0.004794843960553408, + 0.004838623106479645, + 0.004542523063719273, + 0.011358986608684063, + 0.004216006025671959, + 0.025486137717962265, + 0.01764230988919735, + 0.020449670031666756, + 0.020255981013178825, + 0.01626981422305107, + 0.01420010719448328, + 0.030272066593170166, + 0.034557558596134186, + 0.008067958988249302, + 0.006122993305325508, + 0.005494034383445978, + 0.005079867783933878, + 0.004726266488432884, + 0.008949752897024155, + 0.009974815882742405, + 0.00820006150752306, + 0.011520462110638618, + 0.008203120902180672, + 0.007432539481669664, + 0.00779837416484952, + 0.010485921986401081, + 0.013128254562616348, + 0.010277148336172104, + 0.005352544132620096, + 0.02242223545908928, + 0.017999863252043724, + 0.017116867005825043, + 0.017837271094322205, + 0.012645144015550613, + 0.0143509516492486, + 0.028893545269966125, + 0.011193881742656231, + 0.009803497232496738, + 0.005116206593811512, + 0.006418070290237665, + 0.011203852482140064, + 0.004550418816506863, + 0.005415487103164196, + 0.010003751143813133, + 0.007708858232945204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.10030611604452133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635037526488304 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635037526488304, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635037526488304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07321780174970627, + 0.05065729841589928, + 0.580143928527832, + 0.05561424046754837, + 0.03545859828591347, + 0.08799651265144348, + 0.0337284691631794, + 0.017663471400737762, + 0.03131565451622009, + 0.03420398384332657 + ], + "credit_map": [ + 0.007167044095695019, + 0.011725486256182194, + 0.006467949133366346, + 0.01526285707950592, + 0.01013258472084999, + 0.010541792958974838, + 0.005005508195608854, + 0.0069145867601037025, + 0.0043094647116959095, + 0.006548066157847643, + 0.002977399155497551, + 0.008928153663873672, + 0.011993899010121822, + 0.007859046570956707, + 0.004113893955945969, + 0.003927378915250301, + 0.06813585013151169, + 0.05291459709405899, + 0.0974062830209732, + 0.05064532905817032, + 0.06712402403354645, + 0.06989187747240067, + 0.07372363656759262, + 0.10030242800712585, + 0.006932923570275307, + 0.008258342742919922, + 0.0045898472890257835, + 0.005339447408914566, + 0.00896789226680994, + 0.007448860444128513, + 0.010395160876214504, + 0.0036817684303969145, + 0.006025763228535652, + 0.00397760234773159, + 0.003294514026492834, + 0.0028060548938810825, + 0.005754311103373766, + 0.0047569735907018185, + 0.006238001864403486, + 0.0026053772307932377, + 0.012284288182854652, + 0.008060338906943798, + 0.009925330057740211, + 0.010554990731179714, + 0.008174331858754158, + 0.007525753695517778, + 0.017596140503883362, + 0.013875351287424564, + 0.005703010130673647, + 0.003588212188333273, + 0.003761283354833722, + 0.0024877183604985476, + 0.003869002452120185, + 0.00405078986659646, + 0.006029024254530668, + 0.004239432979375124, + 0.002684303093701601, + 0.0017689790111035109, + 0.002250306773930788, + 0.002272027777507901, + 0.001223916420713067, + 0.0027444299776107073, + 0.002873636083677411, + 0.001845873543061316, + 0.0050642360001802444, + 0.0038921113591641188, + 0.0024011428467929363, + 0.0034397905692458153, + 0.0029317073058336973, + 0.005912010557949543, + 0.00508861243724823, + 0.002586046466603875, + 0.0051684873178601265, + 0.003662821836769581, + 0.003271482652053237, + 0.005722167436033487, + 0.0034457289148122072, + 0.003549147630110383, + 0.005078062880784273, + 0.004306088201701641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.02635037526488304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198355376720428 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198355376720428, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198355376720428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09930025041103363, + 0.060960881412029266, + 0.37443897128105164, + 0.10094282776117325, + 0.05098559334874153, + 0.07111380994319916, + 0.0350206233561039, + 0.052663709968328476, + 0.052763961255550385, + 0.10180942714214325 + ], + "credit_map": [ + 0.00922468863427639, + 0.01653187908232212, + 0.009920384734869003, + 0.017257539555430412, + 0.014718898572027683, + 0.013481875881552696, + 0.008601442910730839, + 0.009563538245856762, + 0.005724946036934853, + 0.005820204969495535, + 0.0040133316069841385, + 0.015793154016137123, + 0.008906647562980652, + 0.01003508735448122, + 0.004418659955263138, + 0.006248846650123596, + 0.03870833292603493, + 0.04069725424051285, + 0.06441197544336319, + 0.03363938257098198, + 0.037933360785245895, + 0.044014301151037216, + 0.050433456897735596, + 0.06460090726613998, + 0.010596651583909988, + 0.01541069708764553, + 0.008767839521169662, + 0.009965642355382442, + 0.017498057335615158, + 0.012239846400916576, + 0.017073210328817368, + 0.00939087849110365, + 0.0073905629105865955, + 0.008573489263653755, + 0.004703037440776825, + 0.004471120424568653, + 0.0066496748477220535, + 0.00608048215508461, + 0.010077000595629215, + 0.003040228271856904, + 0.010349572636187077, + 0.00652040122076869, + 0.007255025673657656, + 0.007426066789776087, + 0.007002922706305981, + 0.007315644063055515, + 0.013442284427583218, + 0.011801885440945625, + 0.004917331505566835, + 0.004525606986135244, + 0.0036652982234954834, + 0.005647518206387758, + 0.0030211552511900663, + 0.0035211846698075533, + 0.00608000997453928, + 0.003642519237473607, + 0.008181228302419186, + 0.005362814292311668, + 0.004821000620722771, + 0.00533239496871829, + 0.007992837578058243, + 0.007661300245672464, + 0.008624804206192493, + 0.004687332548201084, + 0.007370933424681425, + 0.0053259022533893585, + 0.006083197426050901, + 0.007018618751317263, + 0.005265514366328716, + 0.008669103495776653, + 0.007037503644824028, + 0.005993186961859465, + 0.016507146880030632, + 0.011549980379641056, + 0.013818129897117615, + 0.01563231274485588, + 0.007710834499448538, + 0.01133287325501442, + 0.015670295804739, + 0.009587850421667099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.014198355376720428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040932451374828815 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.0040932451374828815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040932451374828815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0951712504029274, + 0.04935525357723236, + 0.45239147543907166, + 0.03387577831745148, + 0.03160502016544342, + 0.07637515664100647, + 0.03595134988427162, + 0.03802375867962837, + 0.09830053150653839, + 0.08895041793584824 + ], + "credit_map": [ + 0.010980125516653061, + 0.013734693638980389, + 0.00736598577350378, + 0.011421097442507744, + 0.022023066878318787, + 0.013913466595113277, + 0.0071389758959412575, + 0.0085938461124897, + 0.0044721802696585655, + 0.007229669950902462, + 0.0038153021596372128, + 0.01184859685599804, + 0.006777124013751745, + 0.00764743285253644, + 0.004070363473147154, + 0.0034945921506732702, + 0.04334191977977753, + 0.037052400410175323, + 0.08515076339244843, + 0.03975952789187431, + 0.04752079024910927, + 0.05497952550649643, + 0.05539395287632942, + 0.08919264376163483, + 0.004474019631743431, + 0.004423464182764292, + 0.003378666937351227, + 0.00407467782497406, + 0.004989486653357744, + 0.004656425677239895, + 0.005069802049547434, + 0.002809238387271762, + 0.005476056132465601, + 0.0025886406656354666, + 0.003870852990075946, + 0.0037107926327735186, + 0.0026794876903295517, + 0.003798473160713911, + 0.006296373438090086, + 0.0031843455508351326, + 0.010949441231787205, + 0.009909434244036674, + 0.00816870667040348, + 0.010195272974669933, + 0.00552021898329258, + 0.007621914613991976, + 0.011547044850885868, + 0.012463126331567764, + 0.005160822998732328, + 0.0035372814163565636, + 0.004011139739304781, + 0.0026225289329886436, + 0.003680683672428131, + 0.005460523534566164, + 0.007564788684248924, + 0.003913584165275097, + 0.0055042956955730915, + 0.0034775843378156424, + 0.005506107117980719, + 0.0043774922378361225, + 0.003712724894285202, + 0.004975308198481798, + 0.005213538650423288, + 0.005256712436676025, + 0.014351435005664825, + 0.011387302540242672, + 0.010334775783121586, + 0.011169820092618465, + 0.009986027143895626, + 0.015163982287049294, + 0.01585993729531765, + 0.010047267191112041, + 0.013808272778987885, + 0.010633116587996483, + 0.007914852350950241, + 0.00841410830616951, + 0.0084981769323349, + 0.0132295535877347, + 0.01485404185950756, + 0.011598301120102406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.0040932451374828815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "At night, a dramatic low-angle shot pans slowly sideways across an abandoned outdoor basketball court, focusing on a lone athlete in the center of the frame. The athlete is a highly focused figure, with short, sweat-soaked hair and dressed in a simple, dark, sleeveless sweatshirt and matching shorts that cling slightly to their muscular frame. They maintain a low, dribbling stance, dribbling the ball rhythmically with steady, powerful hands. Every impact of the orange ball on the worn, painted concrete is deliberate and controlled. The towering, bright stadium floodlights pierce the inky night sky, casting an intense, high-contrast glow that sculpts the athletes' forms and captures the sheen of sweat on their skin. This intense, direct light casts distinctly elongated and crisp shadows of the players and the bouncing ball, a dark twin that stretches far across the empty court. The surrounding stands are engulfed in a vast, silent darkness, amplifying the profound sense of solitude and creating a stark, focused atmosphere of pure, meditative devotion.", + "seed": 7575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_8/index_3933_seed_7575_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06372042391449213, + "vlm_logprob": -0.06372042391449213, + "vlm_yes_no_margin": 3.249999976158142, + "vlm_token_credit": -0.06372042391449213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11752180010080338 + ], + "opposite_logprob_tokens": [ + -2.3675217628479004 + ], + "logprob_reward": -0.11752180010080338, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11752180010080338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12997128069400787, + 0.10693024098873138, + 0.07914801687002182, + 0.18366678059101105, + 0.2303900122642517, + 0.056709397584199905, + 0.053774766623973846, + 0.04284973442554474, + 0.036629460752010345, + 0.07993024587631226 + ], + "credit_map": [ + 0.011745169758796692, + 0.021252047270536423, + 0.013388046994805336, + 0.01660086400806904, + 0.01481385063380003, + 0.016788963228464127, + 0.014147184789180756, + 0.021235156804323196, + 0.005903376266360283, + 0.011938165873289108, + 0.013953855261206627, + 0.00867595337331295, + 0.013792671263217926, + 0.01596028171479702, + 0.013623571023344994, + 0.023082366213202477, + 0.006578860338777304, + 0.009846658445894718, + 0.0098513662815094, + 0.008291860111057758, + 0.012600792571902275, + 0.011567024514079094, + 0.011026469059288502, + 0.009384984150528908, + 0.015473893843591213, + 0.018971066921949387, + 0.017661582678556442, + 0.01457966584712267, + 0.03471601381897926, + 0.0270609799772501, + 0.02745404653251171, + 0.027749529108405113, + 0.024034136906266212, + 0.034635577350854874, + 0.03911302983760834, + 0.03230834752321243, + 0.02926304005086422, + 0.031938910484313965, + 0.020071525126695633, + 0.01902545429766178, + 0.0056553613394498825, + 0.006229621823877096, + 0.00963681936264038, + 0.007421664893627167, + 0.007012187037616968, + 0.006936844903975725, + 0.00603861641138792, + 0.007778277155011892, + 0.005150616634637117, + 0.00388311012648046, + 0.005797607824206352, + 0.006186065264046192, + 0.008460895158350468, + 0.008651753887534142, + 0.008148022927343845, + 0.007496695965528488, + 0.006267199292778969, + 0.007572203408926725, + 0.0068962834775447845, + 0.002456753049045801, + 0.004888270981609821, + 0.007177471648901701, + 0.003745738649740815, + 0.003845813451334834, + 0.002129526110365987, + 0.0035026906989514828, + 0.003026495687663555, + 0.004548678640276194, + 0.005238722078502178, + 0.005163062829524279, + 0.00919728446751833, + 0.003822999307885766, + 0.004915164317935705, + 0.01078073587268591, + 0.008821710012853146, + 0.010380305349826813, + 0.009194489568471909, + 0.011816550977528095, + 0.012882198207080364, + 0.011139092035591602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.11752180010080338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280375838279724 + ], + "opposite_logprob_tokens": [ + -2.01780366897583 + ], + "logprob_reward": -0.14280375838279724, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280375838279724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17825806140899658, + 0.11272145062685013, + 0.17259913682937622, + 0.09137385338544846, + 0.042998649179935455, + 0.09136201441287994, + 0.1378166824579239, + 0.05068386718630791, + 0.08228715509176254, + 0.0398990698158741 + ], + "credit_map": [ + 0.011159304529428482, + 0.027576565742492676, + 0.014775813557207584, + 0.026105983182787895, + 0.022561946883797646, + 0.02597617171704769, + 0.02577780932188034, + 0.024324478581547737, + 0.007132209371775389, + 0.013413180597126484, + 0.013042963109910488, + 0.01516671758145094, + 0.014441742561757565, + 0.0169201847165823, + 0.01478979829698801, + 0.01781466044485569, + 0.010762045159935951, + 0.012850729748606682, + 0.012932397425174713, + 0.024876227602362633, + 0.03241295740008354, + 0.026612378656864166, + 0.03233577683568001, + 0.019816627725958824, + 0.008577338419854641, + 0.007410914171487093, + 0.011201849207282066, + 0.008851810358464718, + 0.014240455813705921, + 0.01141396351158619, + 0.012490618973970413, + 0.017186898738145828, + 0.003868622239679098, + 0.005316130816936493, + 0.005883633159101009, + 0.005876675248146057, + 0.005659857764840126, + 0.0054293242283165455, + 0.005480804480612278, + 0.00548360263928771, + 0.008938565850257874, + 0.005624669138342142, + 0.01110184844583273, + 0.007626661099493504, + 0.014796039089560509, + 0.011767814867198467, + 0.016457153484225273, + 0.015049268491566181, + 0.01050762739032507, + 0.005746854934841394, + 0.01443204004317522, + 0.01989506557583809, + 0.0248605664819479, + 0.019214827567338943, + 0.023415587842464447, + 0.01974412053823471, + 0.004144751932471991, + 0.0055388351902365685, + 0.003818307537585497, + 0.005653328262269497, + 0.008088646456599236, + 0.008700027130544186, + 0.008977307006716728, + 0.005762670189142227, + 0.004931597039103508, + 0.010246641002595425, + 0.0067842151038348675, + 0.007647922728210688, + 0.012736772187054157, + 0.012761364690959454, + 0.016899349167943, + 0.01027930062264204, + 0.0018699662759900093, + 0.0031841362360864878, + 0.003143687266856432, + 0.004592461045831442, + 0.008230844512581825, + 0.0066816252656280994, + 0.008092124946415424, + 0.004104228224605322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.14280375838279724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349911466240883 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026349911466240883, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349911466240883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2808774411678314, + 0.22025279700756073, + 0.04376988485455513, + 0.05431725084781647, + 0.08242888003587723, + 0.059728216379880905, + 0.07408861070871353, + 0.0307619608938694, + 0.07726866751909256, + 0.07650632411241531 + ], + "credit_map": [ + 0.02105201780796051, + 0.03692621365189552, + 0.02078378200531006, + 0.02562115527689457, + 0.056891459971666336, + 0.056621864438056946, + 0.024410437792539597, + 0.03857051581144333, + 0.022642523050308228, + 0.03097280114889145, + 0.018667981028556824, + 0.023816317319869995, + 0.03573768213391304, + 0.03744072467088699, + 0.029132705181837082, + 0.02184206061065197, + 0.0032163711730390787, + 0.0032248124480247498, + 0.004609429743140936, + 0.005175402853637934, + 0.007638157811015844, + 0.006471862085163593, + 0.006488120183348656, + 0.006945729721337557, + 0.004695540759712458, + 0.003611518768593669, + 0.005501600448042154, + 0.005105952266603708, + 0.010672677308321, + 0.008775431662797928, + 0.007878162898123264, + 0.00807636696845293, + 0.007596320938318968, + 0.010390183888375759, + 0.010303054936230183, + 0.009537816978991032, + 0.012178528122603893, + 0.009904097765684128, + 0.011467745527625084, + 0.011051133275032043, + 0.0058855279348790646, + 0.003961730282753706, + 0.006605240982025862, + 0.005219793878495693, + 0.009786417707800865, + 0.007726782001554966, + 0.010824442841112614, + 0.009718281216919422, + 0.005431086756289005, + 0.005094013176858425, + 0.007170735392719507, + 0.009042450226843357, + 0.009178919717669487, + 0.012446602806448936, + 0.013583220541477203, + 0.012141582556068897, + 0.0023547590244561434, + 0.002332797273993492, + 0.003555804258212447, + 0.0027793175540864468, + 0.004614657256752253, + 0.005194098223000765, + 0.005999046843498945, + 0.003931480459868908, + 0.005604575388133526, + 0.009881418198347092, + 0.005358704831451178, + 0.0074159810319542885, + 0.01514105498790741, + 0.012621237896382809, + 0.01294414047151804, + 0.008301557041704655, + 0.0044915154576301575, + 0.006487867329269648, + 0.005892771761864424, + 0.007597220595926046, + 0.01626710221171379, + 0.014448495581746101, + 0.013863028027117252, + 0.007458321750164032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.026349911466240883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291800171136856 + ], + "opposite_logprob_tokens": [ + -3.773291826248169 + ], + "logprob_reward": -0.023291800171136856, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291800171136856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17844755947589874, + 0.08710659295320511, + 0.06567167490720749, + 0.08807019889354706, + 0.07250414043664932, + 0.13944637775421143, + 0.034054845571517944, + 0.04161803796887398, + 0.08708839863538742, + 0.20599214732646942 + ], + "credit_map": [ + 0.010948825627565384, + 0.027451831847429276, + 0.013314740732312202, + 0.022193850949406624, + 0.030941464006900787, + 0.033098045736551285, + 0.022564103826880455, + 0.01793469488620758, + 0.00728606479242444, + 0.014629751443862915, + 0.008233019150793552, + 0.011699600145220757, + 0.010476989671587944, + 0.012200817465782166, + 0.010326704010367393, + 0.012253651395440102, + 0.00419161980971694, + 0.004663222003728151, + 0.006081426050513983, + 0.008786573074758053, + 0.012141398154199123, + 0.010007104836404324, + 0.011426616460084915, + 0.008373717777431011, + 0.006920871324837208, + 0.004557167645543814, + 0.005971656180918217, + 0.006250102538615465, + 0.02339390479028225, + 0.015748094767332077, + 0.013987626880407333, + 0.011240772902965546, + 0.006840900052338839, + 0.006953482516109943, + 0.008119008503854275, + 0.006193855777382851, + 0.014496362768113613, + 0.009425661526620388, + 0.010795596055686474, + 0.009679273702204227, + 0.012877385132014751, + 0.010100511834025383, + 0.01798371970653534, + 0.011591905727982521, + 0.025750862434506416, + 0.016385486349463463, + 0.023286154493689537, + 0.02147035486996174, + 0.002502904972061515, + 0.0022220872342586517, + 0.003622230840846896, + 0.0033367376308888197, + 0.004234086722135544, + 0.005780360195785761, + 0.005909837782382965, + 0.006446600425988436, + 0.002689671702682972, + 0.0033574423287063837, + 0.005126748699694872, + 0.004057855345308781, + 0.005210530944168568, + 0.00623345747590065, + 0.008205419406294823, + 0.006736908107995987, + 0.006348846480250359, + 0.00982227735221386, + 0.00675527798011899, + 0.010549207217991352, + 0.013328366912901402, + 0.013286144472658634, + 0.016545183956623077, + 0.010453092865645885, + 0.01131803821772337, + 0.02191230095922947, + 0.01231254730373621, + 0.02102391980588436, + 0.04128669574856758, + 0.03566007688641548, + 0.04372459277510643, + 0.018753984943032265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.023291800171136856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1312510222196579, + 0.06121126934885979, + 0.07885608822107315, + 0.08281973749399185, + 0.06409928947687149, + 0.06004206836223602, + 0.1537017971277237, + 0.2875957489013672, + 0.05094154551625252, + 0.029481390491127968 + ], + "credit_map": [ + 0.008215575478971004, + 0.017973866313695908, + 0.012172280810773373, + 0.014465022832155228, + 0.020316995680332184, + 0.02349446527659893, + 0.01118859276175499, + 0.02342422492802143, + 0.0047156596556305885, + 0.009795496240258217, + 0.005237638019025326, + 0.009275651536881924, + 0.007521538063883781, + 0.00920459907501936, + 0.008112558163702488, + 0.0073481276631355286, + 0.010404658503830433, + 0.013465940952301025, + 0.006857200525701046, + 0.008574641309678555, + 0.010221458971500397, + 0.011797276325523853, + 0.008262000977993011, + 0.009272906929254532, + 0.01056615263223648, + 0.006841046269983053, + 0.009527520276606083, + 0.006202736869454384, + 0.015865927562117577, + 0.01244677510112524, + 0.009912820532917976, + 0.011456762440502644, + 0.005449777934700251, + 0.007081868592649698, + 0.012076249346137047, + 0.010244788601994514, + 0.0075019956566393375, + 0.007837041281163692, + 0.0056898025795817375, + 0.008217768743634224, + 0.006195410154759884, + 0.006811831146478653, + 0.0075685097835958, + 0.008151370100677013, + 0.006965762935578823, + 0.008151077665388584, + 0.007689374033361673, + 0.008508730679750443, + 0.014870826154947281, + 0.009617745876312256, + 0.0173195730894804, + 0.0192318893969059, + 0.02420654147863388, + 0.021893741562962532, + 0.021916525438427925, + 0.024644961580634117, + 0.03597666323184967, + 0.0688038021326065, + 0.027390271425247192, + 0.01823187991976738, + 0.034532345831394196, + 0.0458751879632473, + 0.025947002694010735, + 0.030838606879115105, + 0.0026941094547510147, + 0.007758514955639839, + 0.003093344857916236, + 0.0040443274192512035, + 0.008004089817404747, + 0.008110685274004936, + 0.011739330366253853, + 0.005497146863490343, + 0.0018841868732124567, + 0.0038988664746284485, + 0.0028179511427879333, + 0.0021768412552773952, + 0.004959172569215298, + 0.004606598522514105, + 0.00512637197971344, + 0.004011399112641811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.008634849451482296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "At night, a dramatic low-angle shot pans slowly sideways across an abandoned outdoor basketball court, focusing on a lone athlete in the center of the frame. The athlete is a highly focused figure, with short, sweat-soaked hair and dressed in a simple, dark, sleeveless sweatshirt and matching shorts that cling slightly to their muscular frame. They maintain a low, dribbling stance, dribbling the ball rhythmically with steady, powerful hands. Every impact of the orange ball on the worn, painted concrete is deliberate and controlled. The towering, bright stadium floodlights pierce the inky night sky, casting an intense, high-contrast glow that sculpts the athletes' forms and captures the sheen of sweat on their skin. This intense, direct light casts distinctly elongated and crisp shadows of the players and the bouncing ball, a dark twin that stretches far across the empty court. The surrounding stands are engulfed in a vast, silent darkness, amplifying the profound sense of solitude and creating a stark, focused atmosphere of pure, meditative devotion.", + "seed": 7576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_8/index_3933_seed_7576_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06537008807063102, + "vlm_logprob": -0.06537008807063102, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.06537008807063102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13889162242412567 + ], + "opposite_logprob_tokens": [ + -2.1388916969299316 + ], + "logprob_reward": -0.13889162242412567, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13889162242412567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18658988177776337, + 0.09397099167108536, + 0.027655484154820442, + 0.051065534353256226, + 0.08746659010648727, + 0.03409411758184433, + 0.09236946702003479, + 0.3130906820297241, + 0.05174373462796211, + 0.061953578144311905 + ], + "credit_map": [ + 0.023912103846669197, + 0.019484469667077065, + 0.04339071735739708, + 0.017927832901477814, + 0.0166965052485466, + 0.021072492003440857, + 0.023466777056455612, + 0.020638976246118546, + 0.010721549391746521, + 0.011111436411738396, + 0.012943984940648079, + 0.010507035069167614, + 0.011204753071069717, + 0.01627928949892521, + 0.012425944209098816, + 0.00877698790282011, + 0.0038379800971597433, + 0.003453942248597741, + 0.0037033427506685257, + 0.00281615206040442, + 0.0034323662985116243, + 0.005739140789955854, + 0.0026222700253129005, + 0.0020502854604274035, + 0.008875385858118534, + 0.005244821775704622, + 0.006274065468460321, + 0.003617404494434595, + 0.0036661229096353054, + 0.0052831438370049, + 0.005375373642891645, + 0.012729211710393429, + 0.01577889919281006, + 0.012076647020876408, + 0.01667174883186817, + 0.009206300601363182, + 0.008420283906161785, + 0.008319948799908161, + 0.00741760665550828, + 0.009575147181749344, + 0.0035560938995331526, + 0.005114264320582151, + 0.007747867610305548, + 0.003780144965276122, + 0.0030884933657944202, + 0.0019306809408590198, + 0.0047173332422971725, + 0.0041592358611524105, + 0.013724833726882935, + 0.010272814892232418, + 0.018931008875370026, + 0.008712198585271835, + 0.012580410577356815, + 0.008519216440618038, + 0.010634834878146648, + 0.00899413600564003, + 0.06127673760056496, + 0.06075415015220642, + 0.04855256527662277, + 0.015160441398620605, + 0.056801579892635345, + 0.02113344334065914, + 0.022843128070235252, + 0.026568591594696045, + 0.008971323259174824, + 0.006156341638416052, + 0.01289804745465517, + 0.004055453464388847, + 0.004457105882465839, + 0.003859936958178878, + 0.004040080588310957, + 0.007305438630282879, + 0.007611869368702173, + 0.007371342275291681, + 0.01097048632800579, + 0.007046112325042486, + 0.006771821994334459, + 0.010647467337548733, + 0.005914687644690275, + 0.005619778297841549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.13889162242412567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898459583520889 + ], + "opposite_logprob_tokens": [ + -2.578984498977661 + ], + "logprob_reward": -0.07898459583520889, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898459583520889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12195827811956406, + 0.033217571675777435, + 0.0622127503156662, + 0.08666583150625229, + 0.0744473859667778, + 0.10307484120130539, + 0.26427316665649414, + 0.11214090138673782, + 0.0729367658495903, + 0.06907254457473755 + ], + "credit_map": [ + 0.013105032965540886, + 0.016852719709277153, + 0.01893816702067852, + 0.015607367269694805, + 0.012200101278722286, + 0.013783521950244904, + 0.019048068672418594, + 0.012423282489180565, + 0.0047264848835766315, + 0.0040765078738331795, + 0.005844948347657919, + 0.002949880203232169, + 0.0037960740737617016, + 0.005284917540848255, + 0.0036989650689065456, + 0.0028397920541465282, + 0.007022309582680464, + 0.008066228590905666, + 0.007452393416315317, + 0.00787772610783577, + 0.008673709817230701, + 0.008584373630583286, + 0.008896549232304096, + 0.005639451555907726, + 0.015408162027597427, + 0.008427921682596207, + 0.010120206512510777, + 0.005997055675834417, + 0.006529357749968767, + 0.008178560063242912, + 0.010499656200408936, + 0.021504905074834824, + 0.009707704186439514, + 0.0104673458263278, + 0.013605421409010887, + 0.011241561733186245, + 0.00937015563249588, + 0.006377661135047674, + 0.006774156354367733, + 0.006903374567627907, + 0.014957345090806484, + 0.01522009912878275, + 0.022892413660883904, + 0.01070413924753666, + 0.008015754632651806, + 0.005449904128909111, + 0.01122488733381033, + 0.0146102886646986, + 0.051340192556381226, + 0.027184659615159035, + 0.06309114396572113, + 0.018845688551664352, + 0.045812226831912994, + 0.016635866835713387, + 0.020027175545692444, + 0.021336203441023827, + 0.014299801550805569, + 0.019423965364694595, + 0.015047564171254635, + 0.008796943351626396, + 0.01813511736690998, + 0.01109015941619873, + 0.01135658472776413, + 0.013990744948387146, + 0.011004484258592129, + 0.00799157191067934, + 0.0139576755464077, + 0.006357357371598482, + 0.007975783199071884, + 0.006931597366929054, + 0.007272814866155386, + 0.011445475742220879, + 0.007691237144172192, + 0.008885691873729229, + 0.010358607396483421, + 0.007039749529212713, + 0.012394106015563011, + 0.007461926434189081, + 0.009667470119893551, + 0.005573748145252466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.07898459583520889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584875717759132 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.020584875717759132, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584875717759132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19300252199172974, + 0.08190949261188507, + 0.05973634868860245, + 0.10596401989459991, + 0.0859922543168068, + 0.04276806861162186, + 0.1119755282998085, + 0.09492350369691849, + 0.13435083627700806, + 0.08937747031450272 + ], + "credit_map": [ + 0.022126564756035805, + 0.025787269696593285, + 0.03635821118950844, + 0.024026311933994293, + 0.018340975046157837, + 0.020093144848942757, + 0.029448332265019417, + 0.01682172156870365, + 0.011180493049323559, + 0.010190562345087528, + 0.01549961231648922, + 0.008884912356734276, + 0.009896866977214813, + 0.009348489344120026, + 0.009966052137315273, + 0.0069425068795681, + 0.006070595700293779, + 0.006976115982979536, + 0.008046540431678295, + 0.009315126575529575, + 0.008064799010753632, + 0.006094865966588259, + 0.009818254970014095, + 0.005350056570023298, + 0.011650356464087963, + 0.011996107175946236, + 0.01840050332248211, + 0.010303479619324207, + 0.014935113489627838, + 0.013121350668370724, + 0.013667675666511059, + 0.011889425106346607, + 0.009635056369006634, + 0.00882237683981657, + 0.013124952092766762, + 0.012501483783125877, + 0.012495875358581543, + 0.006070966366678476, + 0.01485513150691986, + 0.008486416190862656, + 0.005078540183603764, + 0.005263817496597767, + 0.005003572441637516, + 0.005292101763188839, + 0.005292339716106653, + 0.003652737243101001, + 0.005996943451464176, + 0.007188018411397934, + 0.018932167440652847, + 0.013513805344700813, + 0.01729576848447323, + 0.009576714597642422, + 0.022139182314276695, + 0.009705653414130211, + 0.011650859378278255, + 0.009161383844912052, + 0.008111247792840004, + 0.009242754429578781, + 0.017958035692572594, + 0.007195213809609413, + 0.011057654395699501, + 0.009244220331311226, + 0.013785889372229576, + 0.018328487873077393, + 0.017537551000714302, + 0.020767705515027046, + 0.015627890825271606, + 0.012275178916752338, + 0.015430080704391003, + 0.02094772830605507, + 0.013047084212303162, + 0.018717607483267784, + 0.009755824692547321, + 0.009002829901874065, + 0.01920558698475361, + 0.01000294741243124, + 0.010978218168020248, + 0.008967645466327667, + 0.009786659851670265, + 0.011677758768200874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.020584875717759132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263564120978117 + ], + "opposite_logprob_tokens": [ + -3.6513564586639404 + ], + "logprob_reward": -0.0263564120978117, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263564120978117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11917082965373993, + 0.05416272580623627, + 0.05573449656367302, + 0.20651166141033173, + 0.06516856700181961, + 0.054343145340681076, + 0.10000072419643402, + 0.09132363647222519, + 0.19416508078575134, + 0.059419114142656326 + ], + "credit_map": [ + 0.013627341017127037, + 0.01764588989317417, + 0.017820004373788834, + 0.015607492066919804, + 0.01135481707751751, + 0.014203815720975399, + 0.018188074231147766, + 0.01072339341044426, + 0.006321734748780727, + 0.006398472469300032, + 0.0063997907564044, + 0.005945014767348766, + 0.005349605344235897, + 0.008654637262225151, + 0.009180574677884579, + 0.00591289671137929, + 0.0078917620703578, + 0.006237174849957228, + 0.011463330127298832, + 0.005789698101580143, + 0.006228534039109945, + 0.006775332614779472, + 0.006613331846892834, + 0.004735336638987064, + 0.03975076228380203, + 0.023014230653643608, + 0.02389989048242569, + 0.014937731437385082, + 0.011061915196478367, + 0.016816185787320137, + 0.021085327491164207, + 0.05594563111662865, + 0.007142635062336922, + 0.007940433919429779, + 0.009298955090343952, + 0.008168446831405163, + 0.009342363104224205, + 0.004518107045441866, + 0.0123347332701087, + 0.0064228917472064495, + 0.010229288600385189, + 0.005155445542186499, + 0.013032861985266209, + 0.005480656400322914, + 0.004330578725785017, + 0.002842768793925643, + 0.0065510948188602924, + 0.006720451638102531, + 0.010985919274389744, + 0.012438523583114147, + 0.02389463596045971, + 0.009393606334924698, + 0.014301593415439129, + 0.007858827710151672, + 0.010745249688625336, + 0.01038236916065216, + 0.01624954678118229, + 0.013484321534633636, + 0.008906216360628605, + 0.007617852184921503, + 0.014962000772356987, + 0.008017964661121368, + 0.012674345634877682, + 0.00941138993948698, + 0.032630886882543564, + 0.014106453396379948, + 0.06605733931064606, + 0.013273202814161777, + 0.010326341725885868, + 0.01344098336994648, + 0.012198966927826405, + 0.03213090822100639, + 0.005792235489934683, + 0.00761648453772068, + 0.013275884091854095, + 0.007572679314762354, + 0.006331696640700102, + 0.005022789351642132, + 0.006455846596509218, + 0.007351499050855637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0263564120978117 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203293427824974 + ], + "opposite_logprob_tokens": [ + -2.81203293800354 + ], + "logprob_reward": -0.06203293427824974, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203293427824974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18584060668945312, + 0.12630990147590637, + 0.04526310786604881, + 0.04191029816865921, + 0.18554256856441498, + 0.030994940549135208, + 0.09674199670553207, + 0.12969212234020233, + 0.05448222532868385, + 0.1032223030924797 + ], + "credit_map": [ + 0.017741531133651733, + 0.022977976128458977, + 0.027986539527773857, + 0.019512314349412918, + 0.01769639551639557, + 0.014473214745521545, + 0.042505066841840744, + 0.02294757403433323, + 0.014868594706058502, + 0.013731497339904308, + 0.016561977565288544, + 0.014398124068975449, + 0.017937129363417625, + 0.023817598819732666, + 0.013310875743627548, + 0.011684096418321133, + 0.005767071153968573, + 0.005577853415161371, + 0.009684097953140736, + 0.005915183573961258, + 0.004928522277623415, + 0.0046041496098041534, + 0.005170099437236786, + 0.003616128582507372, + 0.003930815029889345, + 0.005058608017861843, + 0.007374147418886423, + 0.006561988033354282, + 0.005077784415334463, + 0.004765406250953674, + 0.005697729531675577, + 0.0034438178408890963, + 0.027650650590658188, + 0.02385251596570015, + 0.04390377551317215, + 0.01698067970573902, + 0.01849282532930374, + 0.016956832259893417, + 0.01564040593802929, + 0.02206488884985447, + 0.004730294458568096, + 0.003775249468162656, + 0.008615732192993164, + 0.0028496915474534035, + 0.0024259337224066257, + 0.0014924467541277409, + 0.00366747728548944, + 0.0034381160512566566, + 0.01048306841403246, + 0.009565072134137154, + 0.027149155735969543, + 0.007726828567683697, + 0.01478729210793972, + 0.0063143884763121605, + 0.008840736001729965, + 0.011875455267727375, + 0.022369911894202232, + 0.02309386059641838, + 0.02117728814482689, + 0.006796113681048155, + 0.019428884610533714, + 0.011194838210940361, + 0.00984054896980524, + 0.015790676698088646, + 0.008452218025922775, + 0.007204733788967133, + 0.011785067617893219, + 0.004986714571714401, + 0.004743851255625486, + 0.00654560374096036, + 0.0033071767538785934, + 0.007456861436367035, + 0.010584034025669098, + 0.011885999701917171, + 0.01798613741993904, + 0.010097797960042953, + 0.012374942190945148, + 0.014014250598847866, + 0.012233388610184193, + 0.01404575351625681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.06203293427824974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "At night, a dramatic low-angle shot pans slowly sideways across an abandoned outdoor basketball court, focusing on a lone athlete in the center of the frame. The athlete is a highly focused figure, with short, sweat-soaked hair and dressed in a simple, dark, sleeveless sweatshirt and matching shorts that cling slightly to their muscular frame. They maintain a low, dribbling stance, dribbling the ball rhythmically with steady, powerful hands. Every impact of the orange ball on the worn, painted concrete is deliberate and controlled. The towering, bright stadium floodlights pierce the inky night sky, casting an intense, high-contrast glow that sculpts the athletes' forms and captures the sheen of sweat on their skin. This intense, direct light casts distinctly elongated and crisp shadows of the players and the bouncing ball, a dark twin that stretches far across the empty court. The surrounding stands are engulfed in a vast, silent darkness, amplifying the profound sense of solitude and creating a stark, focused atmosphere of pure, meditative devotion.", + "seed": 7577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_8/index_3933_seed_7577_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.040308872051537034, + "vlm_logprob": -0.040308872051537034, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.040308872051537034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03299707919359207 + ], + "opposite_logprob_tokens": [ + -3.6579971313476562 + ], + "logprob_reward": -0.03299707919359207, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03299707919359207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1723354309797287, + 0.0752217173576355, + 0.17501167953014374, + 0.04054316133260727, + 0.14128723740577698, + 0.04639453440904617, + 0.03898994252085686, + 0.13029690086841583, + 0.056172262877225876, + 0.12374713271856308 + ], + "credit_map": [ + 0.026081446558237076, + 0.02404930256307125, + 0.022954655811190605, + 0.01613122969865799, + 0.014931551180779934, + 0.03385990858078003, + 0.017212633043527603, + 0.017114724963903427, + 0.009873603470623493, + 0.007717709988355637, + 0.009704690426588058, + 0.011713981628417969, + 0.007591577712446451, + 0.009075840003788471, + 0.009696234948933125, + 0.009848086163401604, + 0.05487806722521782, + 0.0730239674448967, + 0.01227931771427393, + 0.006031129974871874, + 0.009219957515597343, + 0.005990895908325911, + 0.009229413233697414, + 0.00435895100235939, + 0.0071640945971012115, + 0.00836732517927885, + 0.0035275265108793974, + 0.005408422090113163, + 0.0036398074589669704, + 0.002304704627022147, + 0.006204843055456877, + 0.003926441073417664, + 0.03505541756749153, + 0.045398514717817307, + 0.010506357066333294, + 0.009026029147207737, + 0.012365697883069515, + 0.008901153691112995, + 0.009133967570960522, + 0.010900112800300121, + 0.005441758316010237, + 0.006508167367428541, + 0.004368952475488186, + 0.004187019541859627, + 0.006323265377432108, + 0.010223215445876122, + 0.006136995740234852, + 0.003205160377547145, + 0.005651624407619238, + 0.005765796639025211, + 0.005469560157507658, + 0.0033531608060002327, + 0.006818960886448622, + 0.004253040999174118, + 0.004423987120389938, + 0.003253814997151494, + 0.015634572133421898, + 0.0140030886977911, + 0.01123994030058384, + 0.01362263597548008, + 0.017680173739790916, + 0.01837058551609516, + 0.02199964039027691, + 0.017746271565556526, + 0.0048925369046628475, + 0.006020628847181797, + 0.007881718687713146, + 0.005063493270426989, + 0.0052743712440133095, + 0.00816667266190052, + 0.01068856380879879, + 0.00818428210914135, + 0.011466003023087978, + 0.04010771960020065, + 0.005651588086038828, + 0.004355904646217823, + 0.009555668570101261, + 0.042672984302043915, + 0.005526473745703697, + 0.004410802852362394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.03299707919359207 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127033993601799 + ], + "opposite_logprob_tokens": [ + -2.1270339488983154 + ], + "logprob_reward": -0.127033993601799, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127033993601799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09048411250114441, + 0.038968514651060104, + 0.10265569388866425, + 0.06136884167790413, + 0.13879963755607605, + 0.16279813647270203, + 0.09382888674736023, + 0.09373049437999725, + 0.10467346012592316, + 0.1126922145485878 + ], + "credit_map": [ + 0.014992141164839268, + 0.009167756885290146, + 0.009275130927562714, + 0.00802894402295351, + 0.008787008002400398, + 0.0175398588180542, + 0.012925318442285061, + 0.009767954237759113, + 0.005465058144181967, + 0.004145039711147547, + 0.004376702941954136, + 0.00644144369289279, + 0.0035642078146338463, + 0.003806763095781207, + 0.006293208803981543, + 0.0048760888166725636, + 0.029604144394397736, + 0.03893584758043289, + 0.00654164282605052, + 0.006359611637890339, + 0.0051529137417674065, + 0.003642365336418152, + 0.008082840591669083, + 0.004336333367973566, + 0.008667170070111752, + 0.00998910702764988, + 0.005982178263366222, + 0.011612922884523869, + 0.004398406017571688, + 0.0050092837773263454, + 0.010601097717881203, + 0.005108680576086044, + 0.029100855812430382, + 0.03906559199094772, + 0.012954742647707462, + 0.014859875664114952, + 0.009942430071532726, + 0.006497503723949194, + 0.016563771292567253, + 0.009814861230552197, + 0.012993605807423592, + 0.028967421501874924, + 0.012385503388941288, + 0.01094415970146656, + 0.04669922962784767, + 0.01672409288585186, + 0.024562278762459755, + 0.009521851316094398, + 0.009868266992270947, + 0.014147662557661533, + 0.006989915855228901, + 0.00794121902436018, + 0.027317890897393227, + 0.009472837671637535, + 0.011149080470204353, + 0.006942017935216427, + 0.00844365544617176, + 0.010335092432796955, + 0.007632141932845116, + 0.010934824123978615, + 0.011445127427577972, + 0.015824489295482635, + 0.014176125638186932, + 0.014939040876924992, + 0.012637334875762463, + 0.014503403566777706, + 0.017311353236436844, + 0.01056351512670517, + 0.01252251397818327, + 0.014659461565315723, + 0.012299326248466969, + 0.010176551528275013, + 0.008850036188960075, + 0.02856493555009365, + 0.0062087783589959145, + 0.008228297345340252, + 0.011918598785996437, + 0.030152561143040657, + 0.011793441139161587, + 0.006975573021918535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.127033993601799 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06920546293258667, + 0.06759688258171082, + 0.11820313334465027, + 0.03365836665034294, + 0.04020174965262413, + 0.2715953588485718, + 0.11572698503732681, + 0.08514063060283661, + 0.04242975637316704, + 0.15624170005321503 + ], + "credit_map": [ + 0.009620347060263157, + 0.009754152968525887, + 0.006737585179507732, + 0.004780737683176994, + 0.008414041250944138, + 0.011376963928341866, + 0.011088277213275433, + 0.007433352526277304, + 0.009121655486524105, + 0.00794618297368288, + 0.011294134892523289, + 0.011052372865378857, + 0.003993420396000147, + 0.005594027228653431, + 0.00839967466890812, + 0.010195409879088402, + 0.03490055724978447, + 0.04639332741498947, + 0.007483026944100857, + 0.007074309978634119, + 0.006150208879262209, + 0.003946051932871342, + 0.008335988968610764, + 0.003919654991477728, + 0.004704195540398359, + 0.005531870760023594, + 0.0036078565753996372, + 0.005032413639128208, + 0.0038957023061811924, + 0.003318921197205782, + 0.004252144601196051, + 0.00331525644287467, + 0.004944784101098776, + 0.006719675846397877, + 0.004249574616551399, + 0.006561115384101868, + 0.004215175751596689, + 0.0030784166883677244, + 0.006983655504882336, + 0.0034493496641516685, + 0.0257564727216959, + 0.04597366973757744, + 0.016785241663455963, + 0.021212037652730942, + 0.0727638378739357, + 0.03145793080329895, + 0.04208947718143463, + 0.01555666234344244, + 0.01769328862428665, + 0.018462367355823517, + 0.005812636110931635, + 0.007729080505669117, + 0.03630078583955765, + 0.011062207631766796, + 0.010922710411250591, + 0.007743898779153824, + 0.009517110884189606, + 0.00997776910662651, + 0.011155166663229465, + 0.009947598911821842, + 0.009977074339985847, + 0.011762318201363087, + 0.009497592225670815, + 0.013305994682013988, + 0.005213161930441856, + 0.007280092686414719, + 0.003392012557014823, + 0.004136017989367247, + 0.006104705855250359, + 0.006384193431586027, + 0.006166254635900259, + 0.0037533126305788755, + 0.010420745238661766, + 0.05609734728932381, + 0.0053804293274879456, + 0.004476524889469147, + 0.011205118149518967, + 0.056780971586704254, + 0.006426320876926184, + 0.005454236641526222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.02058347500860691 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197885990142822 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197885990142822, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197885990142822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046985924243927, + 0.034901756793260574, + 0.06455818563699722, + 0.04266008362174034, + 0.046211954206228256, + 0.03548378497362137, + 0.035299304872751236, + 0.053473349660634995, + 0.05255475267767906, + 0.5878709554672241 + ], + "credit_map": [ + 0.007463445421308279, + 0.004042614251375198, + 0.010467668995261192, + 0.00427799578756094, + 0.0037027199286967516, + 0.007140712812542915, + 0.005798412952572107, + 0.0040923529304564, + 0.005035631358623505, + 0.004291623365134001, + 0.0057623740285634995, + 0.008288001641631126, + 0.00185013294685632, + 0.0024828952737152576, + 0.003532406175509095, + 0.0036586918868124485, + 0.019467422738671303, + 0.02530861087143421, + 0.0033151183743029833, + 0.004442545119673014, + 0.003413058817386627, + 0.002363508800044656, + 0.0037421914748847485, + 0.0025057299062609673, + 0.00798041746020317, + 0.009991326369345188, + 0.005082053132355213, + 0.00597050366923213, + 0.0029108014423400164, + 0.0019965653773397207, + 0.004588662646710873, + 0.00413975166156888, + 0.009474574588239193, + 0.012213459238409996, + 0.004710914101451635, + 0.005071929190307856, + 0.004381123930215836, + 0.0025853412225842476, + 0.0044782729819417, + 0.003296337090432644, + 0.0034668100997805595, + 0.005232522264122963, + 0.0025746957398951054, + 0.003864167956635356, + 0.007298492360860109, + 0.003862649668008089, + 0.006769130006432533, + 0.0024153164122253656, + 0.004325545858591795, + 0.004328795708715916, + 0.003448045812547207, + 0.003313660155981779, + 0.009680687449872494, + 0.0037402440793812275, + 0.0038428367115557194, + 0.0026194897945970297, + 0.005794762633740902, + 0.006528632715344429, + 0.006754559930413961, + 0.006413131952285767, + 0.006855762097984552, + 0.006254247389733791, + 0.00706877326592803, + 0.007803476881235838, + 0.006348288152366877, + 0.008796114474534988, + 0.004547996446490288, + 0.005922713782638311, + 0.005578852258622646, + 0.010099649429321289, + 0.00715677160769701, + 0.004104367457330227, + 0.03366217389702797, + 0.23269017040729523, + 0.01312097068876028, + 0.010467217303812504, + 0.027173608541488647, + 0.24161003530025482, + 0.012999366968870163, + 0.016147391870617867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.014197885990142822 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an abandoned outdoor basketball court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow sideways pan from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete wearing a dark sleeveless sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the athlete dribbling an orange ball with steady, powerful hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright stadium floodlights cast intense, elongated shadows across the court?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731926463544369 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006731926463544369, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731926463544369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08404713124036789, + 0.04418031871318817, + 0.4474767744541168, + 0.1550336480140686, + 0.05395551770925522, + 0.05569208040833473, + 0.05947776883840561, + 0.05089360475540161, + 0.02396189048886299, + 0.0252812709659338 + ], + "credit_map": [ + 0.013848433271050453, + 0.009746105410158634, + 0.008307473734021187, + 0.007556663826107979, + 0.006790203042328358, + 0.014987912029027939, + 0.013996455818414688, + 0.008813882246613503, + 0.005685463082045317, + 0.005322069860994816, + 0.0060507203452289104, + 0.0066479165107011795, + 0.0031676865182816982, + 0.0058972593396902084, + 0.006768152117729187, + 0.004641052335500717, + 0.13853448629379272, + 0.184833362698555, + 0.0318012572824955, + 0.016718143597245216, + 0.02260860800743103, + 0.0155857028439641, + 0.025102367624640465, + 0.012292865663766861, + 0.01802798919379711, + 0.02343836799263954, + 0.01393179502338171, + 0.01940632238984108, + 0.019000882282853127, + 0.024812348186969757, + 0.01889464072883129, + 0.017521295696496964, + 0.011066792532801628, + 0.014571155421435833, + 0.004491891711950302, + 0.004079618025571108, + 0.005469451658427715, + 0.006463084369897842, + 0.003317211288958788, + 0.004496307577937841, + 0.0062196762301027775, + 0.009367726743221283, + 0.00745210750028491, + 0.004845055751502514, + 0.009966754354536533, + 0.005178716965019703, + 0.008303124457597733, + 0.004358916077762842, + 0.007729026023298502, + 0.007867848500609398, + 0.007966174744069576, + 0.006414790637791157, + 0.00860358215868473, + 0.008594623766839504, + 0.005166022107005119, + 0.007135700434446335, + 0.005284593906253576, + 0.005671374499797821, + 0.00782814435660839, + 0.005730258300900459, + 0.005600090604275465, + 0.00677745183929801, + 0.00623283488675952, + 0.007768853101879358, + 0.002610139548778534, + 0.004361626226454973, + 0.002388706197962165, + 0.0029165539890527725, + 0.002894093282520771, + 0.0030712320003658533, + 0.0034793352242559195, + 0.0022402049507945776, + 0.003916667774319649, + 0.003437608014792204, + 0.0020781680941581726, + 0.0024350902531296015, + 0.0029917326755821705, + 0.0033512336667627096, + 0.0029917419888079166, + 0.00407902617007494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.006731926463544369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_9/index_tensor([6940, 6940])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_9/index_tensor([6940, 6940])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c96742e522eac19ec270f224f0b86a2c0b097c9c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_9/index_tensor([6940, 6940])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic confrontation across the vast, dusty expanse of the Australian outback. Two male red kangaroos engage in a ritual boxing match, their powerful bodies rendered in stark, dark silhouettes. They rear up on their massive hind legs, balancing on their thick, muscular tails against the arid, red-brown earth. The setting sun, low below the horizon, ignites the sky into a breathtaking canvas of fiery red, deep scarlet, and faded violet. This intense backlighting perfectly delineates their forms, capturing the edges of their fur with a subtle, luminous rim light. As they struggle and strike each other—jabbing with their front paws and delivering powerful, deliberate kicks—wisps of fine red dust are kicked up from the dry ground, swirling around their feet and catching the surrounding light, creating a glowing, ethereal mist. The entire scene is imbued with a primal, breathtaking atmosphere, contrasting the raw, untamed energy of the battle with the serene, majestic beauty of the sunset.", + "seed": 3146, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_9/index_6940_seed_3146_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030869109882041813, + "vlm_logprob": -0.030869109882041813, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.030869109882041813, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042200133204460144 + ], + "opposite_logprob_tokens": [ + -3.2922000885009766 + ], + "logprob_reward": -0.042200133204460144, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.042200133204460144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10960684716701508, + 0.059627093374729156, + 0.2859881520271301, + 0.04923570528626442, + 0.03537600487470627, + 0.15091338753700256, + 0.05229916051030159, + 0.09850221127271652, + 0.055669207125902176, + 0.10278227925300598 + ], + "credit_map": [ + 0.006519780494272709, + 0.011720012873411179, + 0.01627298817038536, + 0.01828821934759617, + 0.01033197995275259, + 0.008758299984037876, + 0.022348498925566673, + 0.015367058105766773, + 0.00545521778985858, + 0.005072820000350475, + 0.008704762905836105, + 0.00975754763931036, + 0.007155335042625666, + 0.0041381255723536015, + 0.01036805845797062, + 0.008975222706794739, + 0.01791982911527157, + 0.015130520798265934, + 0.03983907774090767, + 0.062083952128887177, + 0.018389597535133362, + 0.029855726286768913, + 0.052493926137685776, + 0.050275519490242004, + 0.004577729385346174, + 0.003357294714078307, + 0.008531429804861546, + 0.006785391829907894, + 0.005906375590711832, + 0.004260365851223469, + 0.007660218980163336, + 0.008156896568834782, + 0.0025432934053242207, + 0.002957151969894767, + 0.006609856616705656, + 0.006244572810828686, + 0.004777833819389343, + 0.0037456911522895098, + 0.00540745398029685, + 0.0030901457648724318, + 0.024253832176327705, + 0.015182710252702236, + 0.020575514063239098, + 0.010770095512270927, + 0.005543402396142483, + 0.013569463044404984, + 0.013918371871113777, + 0.047099970281124115, + 0.004378765355795622, + 0.0041662584990262985, + 0.010208592750132084, + 0.01159772090613842, + 0.004667559172958136, + 0.004570006392896175, + 0.007203693501651287, + 0.00550656346604228, + 0.009060871787369251, + 0.003048571990802884, + 0.03419947996735573, + 0.0116098178550601, + 0.003737807273864746, + 0.007969597354531288, + 0.006892451550811529, + 0.02198360674083233, + 0.0035389752592891455, + 0.0035082451067864895, + 0.00825219601392746, + 0.016572391614317894, + 0.004662991035729647, + 0.004059165716171265, + 0.007951844483613968, + 0.007123393937945366, + 0.015299545601010323, + 0.007436607498675585, + 0.01549792941659689, + 0.01230056956410408, + 0.006843986921012402, + 0.008752151392400265, + 0.00982253160327673, + 0.0268289502710104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.042200133204460144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202821999788284 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202821999788284, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202821999788284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20049242675304413, + 0.0728391706943512, + 0.1181865930557251, + 0.10982514917850494, + 0.06918235123157501, + 0.1668286770582199, + 0.10499163717031479, + 0.042792901396751404, + 0.046774912625551224, + 0.06808620691299438 + ], + "credit_map": [ + 0.015967845916748047, + 0.018787721171975136, + 0.022481240332126617, + 0.03171983361244202, + 0.01849079504609108, + 0.015364154241979122, + 0.050763633102178574, + 0.02691720798611641, + 0.0060912081971764565, + 0.005805466789752245, + 0.009561920538544655, + 0.012801652774214745, + 0.008479153737425804, + 0.005453080870211124, + 0.014422468841075897, + 0.010224218480288982, + 0.009503803215920925, + 0.008682704530656338, + 0.016025254502892494, + 0.027654629200696945, + 0.009716000407934189, + 0.009921807795763016, + 0.02041812054812908, + 0.016264261677861214, + 0.016165249049663544, + 0.007227436173707247, + 0.015268666669726372, + 0.012637756764888763, + 0.0077089606784284115, + 0.010327029973268509, + 0.013793154619634151, + 0.026696892455220222, + 0.005042854230850935, + 0.004570319317281246, + 0.0072009689174592495, + 0.012682216241955757, + 0.007003155071288347, + 0.004754017572849989, + 0.014848295599222183, + 0.013080519624054432, + 0.025559483096003532, + 0.015388895757496357, + 0.024005619809031487, + 0.01575879007577896, + 0.008103357627987862, + 0.015302867628633976, + 0.01645554043352604, + 0.04625412076711655, + 0.00874207355082035, + 0.00920343678444624, + 0.018199516460299492, + 0.01376750972121954, + 0.01164113637059927, + 0.00960333738476038, + 0.01168792974203825, + 0.022146696224808693, + 0.0029406920075416565, + 0.0019612915348261595, + 0.011566044762730598, + 0.003986062481999397, + 0.0025307603646069765, + 0.0034346100874245167, + 0.006041170097887516, + 0.01033227238804102, + 0.0036059990525245667, + 0.0030795412603765726, + 0.011562229134142399, + 0.006751975975930691, + 0.0032733436673879623, + 0.0036218110471963882, + 0.0054472885094583035, + 0.00943271815776825, + 0.008163788355886936, + 0.005199480801820755, + 0.01144701149314642, + 0.009450790472328663, + 0.00302052590996027, + 0.005091710016131401, + 0.006073294207453728, + 0.01963960938155651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.014202821999788284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370640054345131 + ], + "opposite_logprob_tokens": [ + -3.4087064266204834 + ], + "logprob_reward": -0.03370640054345131, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370640054345131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.095610111951828, + 0.10394701361656189, + 0.0638674795627594, + 0.06499750912189484, + 0.031128952279686928, + 0.20632268488407135, + 0.10500460118055344, + 0.17536300420761108, + 0.04682127758860588, + 0.10693731904029846 + ], + "credit_map": [ + 0.006690378300845623, + 0.010750303976237774, + 0.015851451084017754, + 0.01060529612004757, + 0.009842932224273682, + 0.007816753350198269, + 0.019580191001296043, + 0.014472804963588715, + 0.006080921273678541, + 0.0049369740299880505, + 0.012645505368709564, + 0.015950320288538933, + 0.00973867904394865, + 0.007159355096518993, + 0.03057641163468361, + 0.016858844086527824, + 0.006128543987870216, + 0.0038098397199064493, + 0.010886022821068764, + 0.013109449297189713, + 0.004410129971802235, + 0.005060387775301933, + 0.010616060346364975, + 0.009847047738730907, + 0.005479080136865377, + 0.005181687884032726, + 0.014660817570984364, + 0.009123950265347958, + 0.005279868841171265, + 0.005865281447768211, + 0.01158530917018652, + 0.007821513339877129, + 0.0018723987741395831, + 0.002700126264244318, + 0.005021923221647739, + 0.006163440179079771, + 0.0037788525223731995, + 0.003116682870313525, + 0.005004165228456259, + 0.0034713642671704292, + 0.03385781869292259, + 0.0219428613781929, + 0.027207041159272194, + 0.014650911092758179, + 0.007329186424612999, + 0.018509455025196075, + 0.019000429660081863, + 0.06382498145103455, + 0.009085152298212051, + 0.009969678707420826, + 0.02008778229355812, + 0.016422517597675323, + 0.011586525477468967, + 0.009390663355588913, + 0.010131721384823322, + 0.018330564722418785, + 0.014911506325006485, + 0.004691742360591888, + 0.06522880494594574, + 0.014988690614700317, + 0.00426369346678257, + 0.015589054673910141, + 0.011121140792965889, + 0.04456838592886925, + 0.0027462472207844257, + 0.003174839774146676, + 0.007958473637700081, + 0.01285872794687748, + 0.003757408820092678, + 0.0034149307757616043, + 0.006216367240995169, + 0.006694277282804251, + 0.015901844948530197, + 0.007799343205988407, + 0.016744038090109825, + 0.012918599881231785, + 0.00538903521373868, + 0.009216316044330597, + 0.00982503592967987, + 0.029143106192350388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.03370640054345131 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204133853316307 + ], + "opposite_logprob_tokens": [ + -2.8120412826538086 + ], + "logprob_reward": -0.06204133853316307, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204133853316307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12494417279958725, + 0.11615508794784546, + 0.10264716297388077, + 0.11166580021381378, + 0.09066405892372131, + 0.08652105927467346, + 0.03602317348122597, + 0.08124405145645142, + 0.04450170323252678, + 0.20563367009162903 + ], + "credit_map": [ + 0.0076675512827932835, + 0.011346233077347279, + 0.030028579756617546, + 0.013370227999985218, + 0.0162358395755291, + 0.011432403698563576, + 0.018544569611549377, + 0.016318781301379204, + 0.00779131893068552, + 0.006443416699767113, + 0.015023047104477882, + 0.022595729678869247, + 0.011665037833154202, + 0.008880672045052052, + 0.025786815211176872, + 0.017969055101275444, + 0.006983751896768808, + 0.005879873875528574, + 0.016571903601288795, + 0.02111229859292507, + 0.006664824206382036, + 0.009757875464856625, + 0.018538296222686768, + 0.01713833585381508, + 0.014506593346595764, + 0.008321857079863548, + 0.015133881941437721, + 0.012951541692018509, + 0.007981203496456146, + 0.011733329854905605, + 0.017191041260957718, + 0.02384635992348194, + 0.006239418871700764, + 0.009204906411468983, + 0.009102222509682178, + 0.01844159886240959, + 0.009089131839573383, + 0.010804125107824802, + 0.018656695261597633, + 0.009125962853431702, + 0.01429544948041439, + 0.00848168134689331, + 0.01251576840877533, + 0.006544058211147785, + 0.003939945250749588, + 0.007718575652688742, + 0.008323104120790958, + 0.024702487513422966, + 0.005128112155944109, + 0.0022176108323037624, + 0.006795729976147413, + 0.004203580319881439, + 0.0029359187465161085, + 0.005204394925385714, + 0.004148012958467007, + 0.005389814730733633, + 0.006389588117599487, + 0.003658606670796871, + 0.023746704682707787, + 0.0073564923368394375, + 0.003097317647188902, + 0.006449319422245026, + 0.007276440970599651, + 0.023269586265087128, + 0.0038228954654186964, + 0.0020297139417380095, + 0.015274361707270145, + 0.004394484218209982, + 0.001556551200337708, + 0.004437297582626343, + 0.0036187025252729654, + 0.009367699734866619, + 0.03429205343127251, + 0.016716856509447098, + 0.03128793463110924, + 0.022120235487818718, + 0.007981748320162296, + 0.017717130482196808, + 0.020694315433502197, + 0.05482339486479759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.06204133853316307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021948551293462515 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021948551293462515, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021948551293462515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17659991979599, + 0.0634680762887001, + 0.09563855081796646, + 0.05310511589050293, + 0.05713991820812225, + 0.05792740359902382, + 0.029458722099661827, + 0.29487738013267517, + 0.07814235985279083, + 0.09364251792430878 + ], + "credit_map": [ + 0.028433647006750107, + 0.013346726074814796, + 0.02119680680334568, + 0.02494722418487072, + 0.016929298639297485, + 0.023496175184845924, + 0.02590436302125454, + 0.022345684468746185, + 0.00622126180678606, + 0.0043585533276200294, + 0.01187179610133171, + 0.010152095928788185, + 0.005112505983561277, + 0.0060532549396157265, + 0.011585870757699013, + 0.008112740702927113, + 0.006275998428463936, + 0.005265901796519756, + 0.01342334970831871, + 0.020641334354877472, + 0.00606114137917757, + 0.010690515860915184, + 0.01817777194082737, + 0.01510254479944706, + 0.005540721118450165, + 0.003170263022184372, + 0.007252967916429043, + 0.006122019607573748, + 0.004326061345636845, + 0.003970627672970295, + 0.00857223104685545, + 0.014150220900774002, + 0.0047277179546654224, + 0.0061089517548680305, + 0.006931900046765804, + 0.008460833691060543, + 0.005366554018110037, + 0.009795909747481346, + 0.00839507207274437, + 0.00735298078507185, + 0.007357861381024122, + 0.005177630577236414, + 0.008698888123035431, + 0.007508099544793367, + 0.0024154323618859053, + 0.0057690260000526905, + 0.005573416594415903, + 0.015427048318088055, + 0.003802102291956544, + 0.0023253564722836018, + 0.004414915572851896, + 0.0044669038616120815, + 0.0034570302814245224, + 0.003900721436366439, + 0.0039566862396895885, + 0.003135006409138441, + 0.03137846663594246, + 0.009462577290832996, + 0.10930387675762177, + 0.02736266329884529, + 0.007943134754896164, + 0.024973822757601738, + 0.019857944920659065, + 0.06459487974643707, + 0.007634708657860756, + 0.004627943970263004, + 0.01896737515926361, + 0.011671529151499271, + 0.006086238659918308, + 0.007886329665780067, + 0.008965320885181427, + 0.012302915565669537, + 0.014050292782485485, + 0.008824734017252922, + 0.013741676695644855, + 0.00785585306584835, + 0.005052132997661829, + 0.00846357923001051, + 0.008597787469625473, + 0.027056459337472916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0021948551293462515 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic confrontation across the vast, dusty expanse of the Australian outback. Two male red kangaroos engage in a ritual boxing match, their powerful bodies rendered in stark, dark silhouettes. They rear up on their massive hind legs, balancing on their thick, muscular tails against the arid, red-brown earth. The setting sun, low below the horizon, ignites the sky into a breathtaking canvas of fiery red, deep scarlet, and faded violet. This intense backlighting perfectly delineates their forms, capturing the edges of their fur with a subtle, luminous rim light. As they struggle and strike each other—jabbing with their front paws and delivering powerful, deliberate kicks—wisps of fine red dust are kicked up from the dry ground, swirling around their feet and catching the surrounding light, creating a glowing, ethereal mist. The entire scene is imbued with a primal, breathtaking atmosphere, contrasting the raw, untamed energy of the battle with the serene, majestic beauty of the sunset.", + "seed": 3147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_9/index_6940_seed_3147_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0497276094276458, + "vlm_logprob": -0.0497276094276458, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.0497276094276458, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0444062240421772 + ], + "opposite_logprob_tokens": [ + -3.1694061756134033 + ], + "logprob_reward": -0.0444062240421772, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0444062240421772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24672433733940125, + 0.08016213774681091, + 0.07998480647802353, + 0.0941019207239151, + 0.0659690648317337, + 0.12472167611122131, + 0.04663151502609253, + 0.08749455213546753, + 0.08158499747514725, + 0.09262502193450928 + ], + "credit_map": [ + 0.021939847618341446, + 0.022611849009990692, + 0.027024025097489357, + 0.03080328181385994, + 0.0678650364279747, + 0.022627463564276695, + 0.026755638420581818, + 0.027097180485725403, + 0.010917303152382374, + 0.008367780596017838, + 0.009725500829517841, + 0.009225301444530487, + 0.009766065515577793, + 0.008968678303062916, + 0.010687926784157753, + 0.012503579258918762, + 0.010736547410488129, + 0.015850689262151718, + 0.0058619435876607895, + 0.009699222631752491, + 0.011536221019923687, + 0.008864881470799446, + 0.0071576982736587524, + 0.010277598164975643, + 0.01182867307215929, + 0.011985711753368378, + 0.01643056422472, + 0.009533443488180637, + 0.010010184720158577, + 0.011593152768909931, + 0.011561841703951359, + 0.011158342473208904, + 0.010540959425270557, + 0.011858438141644001, + 0.007777236867696047, + 0.0052633569575846195, + 0.009154451079666615, + 0.008183086290955544, + 0.006990557536482811, + 0.006200974341481924, + 0.016109151765704155, + 0.015339626930654049, + 0.015090525150299072, + 0.020188095048069954, + 0.012257120572030544, + 0.01377832144498825, + 0.01606360636651516, + 0.01589522697031498, + 0.004802234936505556, + 0.0045075914822518826, + 0.009148024953901768, + 0.005179664120078087, + 0.003930742386728525, + 0.005356382112950087, + 0.005677024833858013, + 0.008029847405850887, + 0.01650489866733551, + 0.010600320994853973, + 0.010258393362164497, + 0.008828536607325077, + 0.006647674832493067, + 0.017352690920233727, + 0.009373601526021957, + 0.007928429171442986, + 0.008358222432434559, + 0.009859227575361729, + 0.010753929615020752, + 0.016150517389178276, + 0.012710525654256344, + 0.006931200623512268, + 0.008279218338429928, + 0.008542150259017944, + 0.011046275496482849, + 0.01490472350269556, + 0.008912578225135803, + 0.016497794538736343, + 0.011237407103180885, + 0.011049747467041016, + 0.011316073127090931, + 0.00766041548922658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0444062240421772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548773549497128 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012548773549497128, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548773549497128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1537076085805893, + 0.11853460222482681, + 0.090659961104393, + 0.09017378836870193, + 0.07301096618175507, + 0.14794233441352844, + 0.05601610615849495, + 0.07838156074285507, + 0.046284377574920654, + 0.14528879523277283 + ], + "credit_map": [ + 0.017318248748779297, + 0.01700286753475666, + 0.022815130650997162, + 0.019786246120929718, + 0.02700437232851982, + 0.017001885920763016, + 0.017116203904151917, + 0.015662649646401405, + 0.012822110205888748, + 0.012802924029529095, + 0.020446453243494034, + 0.01521434634923935, + 0.01805875264108181, + 0.010059443302452564, + 0.014967966824769974, + 0.014162599109113216, + 0.01107968669384718, + 0.015640832483768463, + 0.008523387834429741, + 0.010205542668700218, + 0.012589704245328903, + 0.01332857832312584, + 0.007627397309988737, + 0.01166483573615551, + 0.011945909820497036, + 0.012891796417534351, + 0.01276506669819355, + 0.008926773443818092, + 0.011100029572844505, + 0.009914418682456017, + 0.012742246501147747, + 0.00988754816353321, + 0.009412587620317936, + 0.014166262932121754, + 0.008905828930437565, + 0.009452508762478828, + 0.0076500289142131805, + 0.008950317278504372, + 0.006544401403516531, + 0.007929030805826187, + 0.020101893693208694, + 0.015552614815533161, + 0.01825013756752014, + 0.026370776817202568, + 0.013247302733361721, + 0.0161945428699255, + 0.018206188455224037, + 0.02001887932419777, + 0.005154978483915329, + 0.005945494398474693, + 0.013016407378017902, + 0.0062790182419121265, + 0.005281446501612663, + 0.006449810694903135, + 0.006975265685468912, + 0.006913685705512762, + 0.013920800760388374, + 0.0080624520778656, + 0.011732056736946106, + 0.007732105441391468, + 0.007494443561881781, + 0.015251163393259048, + 0.008324800990521908, + 0.005863736383616924, + 0.006207313388586044, + 0.007361582946032286, + 0.00587286613881588, + 0.005810271482914686, + 0.006396061275154352, + 0.0055570756085217, + 0.004582804627716541, + 0.004496400710195303, + 0.01602928154170513, + 0.018288642168045044, + 0.01756320707499981, + 0.02672477439045906, + 0.020683566108345985, + 0.01475706696510315, + 0.018023330718278885, + 0.013218918815255165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.012548773549497128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029804222285747528 + ], + "opposite_logprob_tokens": [ + -3.529804229736328 + ], + "logprob_reward": -0.029804222285747528, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029804222285747528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35208630561828613, + 0.057014282792806625, + 0.0937851220369339, + 0.05028865858912468, + 0.04523691162467003, + 0.08175598084926605, + 0.03215028718113899, + 0.08285067975521088, + 0.13892748951911926, + 0.06590433418750763 + ], + "credit_map": [ + 0.03700347617268562, + 0.03092697076499462, + 0.03196612000465393, + 0.03715166077017784, + 0.1078125387430191, + 0.031185217201709747, + 0.03713440150022507, + 0.038905899971723557, + 0.007341839838773012, + 0.006657315883785486, + 0.00602752948179841, + 0.006414140574634075, + 0.009466229937970638, + 0.00526300398632884, + 0.007429503370076418, + 0.00841471552848816, + 0.012849877588450909, + 0.019678717479109764, + 0.008286471478641033, + 0.008566982112824917, + 0.011596065014600754, + 0.009790531359612942, + 0.0077828592620790005, + 0.015233608894050121, + 0.006634481716901064, + 0.0069388123229146, + 0.0065681468695402145, + 0.005856379400938749, + 0.006385096348822117, + 0.006818790920078754, + 0.005328588653355837, + 0.005758356302976608, + 0.004653809126466513, + 0.005841144826263189, + 0.006460550241172314, + 0.0046722921542823315, + 0.006510929204523563, + 0.005077594891190529, + 0.007152028381824493, + 0.004868559539318085, + 0.009105777367949486, + 0.011754398234188557, + 0.011196819134056568, + 0.012073507532477379, + 0.008493483997881413, + 0.008768444880843163, + 0.011068085208535194, + 0.009295462630689144, + 0.0034394084941595793, + 0.0032324271742254496, + 0.0056738643907010555, + 0.004036957863718271, + 0.003828578395769, + 0.0036571864038705826, + 0.003793153678998351, + 0.004488710779696703, + 0.017349136993288994, + 0.009689310565590858, + 0.012116570957005024, + 0.006947318557649851, + 0.00588681036606431, + 0.019086521118879318, + 0.005540665704756975, + 0.006234330590814352, + 0.01578710973262787, + 0.020666353404521942, + 0.021323390305042267, + 0.01902959682047367, + 0.022425193339586258, + 0.011056512594223022, + 0.013883103616535664, + 0.014756218530237675, + 0.008324503898620605, + 0.010707056149840355, + 0.006193872075527906, + 0.00798864383250475, + 0.008980442769825459, + 0.010743615217506886, + 0.007094545755535364, + 0.005871648900210857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.029804222285747528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036637127399445 + ], + "opposite_logprob_tokens": [ + -1.910366415977478 + ], + "logprob_reward": -0.16036637127399445, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036637127399445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10356797277927399, + 0.08298972994089127, + 0.061513498425483704, + 0.10922374576330185, + 0.12655025720596313, + 0.3344082534313202, + 0.04773896560072899, + 0.05883953720331192, + 0.03211002051830292, + 0.043058037757873535 + ], + "credit_map": [ + 0.012386507354676723, + 0.011297334916889668, + 0.02118505723774433, + 0.012818916700780392, + 0.013062468729913235, + 0.008099055849015713, + 0.014873522333800793, + 0.009845105931162834, + 0.010668812319636345, + 0.009564682841300964, + 0.009436259977519512, + 0.009553076699376106, + 0.0069690425880253315, + 0.007907991297543049, + 0.014558538794517517, + 0.014331324957311153, + 0.008719885721802711, + 0.007435437757521868, + 0.005980520509183407, + 0.011286378838121891, + 0.0061833206564188, + 0.007240608334541321, + 0.006950141862034798, + 0.007717202417552471, + 0.016311900690197945, + 0.012743167579174042, + 0.0185866616666317, + 0.013058329932391644, + 0.00881711021065712, + 0.011858705431222916, + 0.01311683189123869, + 0.014731031842529774, + 0.023137889802455902, + 0.018424618989229202, + 0.016646306961774826, + 0.01213904283940792, + 0.011713907122612, + 0.016743047162890434, + 0.011815758422017097, + 0.015929678454995155, + 0.05190816894173622, + 0.02871113084256649, + 0.04948632791638374, + 0.056516923010349274, + 0.026896575465798378, + 0.03824487328529358, + 0.03674992546439171, + 0.0458943136036396, + 0.004957256838679314, + 0.005339926108717918, + 0.010696827434003353, + 0.005505007226020098, + 0.0033637695014476776, + 0.005877407733350992, + 0.005810000468045473, + 0.006188768427819014, + 0.01667606644332409, + 0.00365057704038918, + 0.007454601116478443, + 0.004486770834773779, + 0.0026308936066925526, + 0.016550596803426743, + 0.003384791547432542, + 0.004005237948149443, + 0.004213791340589523, + 0.0058279866352677345, + 0.004715065471827984, + 0.0050871726125478745, + 0.004098320845514536, + 0.002666605170816183, + 0.002671225694939494, + 0.0028298483230173588, + 0.0068001761101186275, + 0.004472718574106693, + 0.007258409634232521, + 0.008153017610311508, + 0.006240674760192633, + 0.003356083994731307, + 0.00418980373069644, + 0.00258714915253222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.16036637127399445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015124559868127108 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015124559868127108, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015124559868127108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25974005460739136, + 0.06408093124628067, + 0.16309231519699097, + 0.0514189749956131, + 0.05500614270567894, + 0.0970410406589508, + 0.06266485899686813, + 0.09752409160137177, + 0.08724154531955719, + 0.06219005584716797 + ], + "credit_map": [ + 0.03479772433638573, + 0.02777327597141266, + 0.0665455013513565, + 0.027742216363549232, + 0.02216939628124237, + 0.01951023004949093, + 0.03157982975244522, + 0.029621880501508713, + 0.007337416056543589, + 0.007139286957681179, + 0.008874922059476376, + 0.007887925021350384, + 0.009974957443773746, + 0.007670262828469276, + 0.00943227019160986, + 0.005763900000602007, + 0.022980501875281334, + 0.022586291655898094, + 0.017990702763199806, + 0.02126041054725647, + 0.019598351791501045, + 0.025566641241312027, + 0.01253065187484026, + 0.020578784868121147, + 0.009886382147669792, + 0.005831677466630936, + 0.0054106037132442, + 0.007900072261691093, + 0.005373634863644838, + 0.006288320757448673, + 0.005045759957283735, + 0.0056825256906449795, + 0.008249223232269287, + 0.005513741634786129, + 0.004816100001335144, + 0.01321179885417223, + 0.0037403537426143885, + 0.005933951586484909, + 0.006404607091099024, + 0.0071363747119903564, + 0.014627574943006039, + 0.01263746153563261, + 0.01060296781361103, + 0.018501706421375275, + 0.007413938175886869, + 0.01097562350332737, + 0.011316641233861446, + 0.010965137742459774, + 0.007157119456678629, + 0.008581755682826042, + 0.014137697406113148, + 0.006239351350814104, + 0.007403065916150808, + 0.007013518363237381, + 0.00746481167152524, + 0.0046675377525389194, + 0.027621988207101822, + 0.007165289018303156, + 0.006053838413208723, + 0.006805102340877056, + 0.005003851838409901, + 0.030832067131996155, + 0.006825956515967846, + 0.007216012571007013, + 0.011068833991885185, + 0.013271171599626541, + 0.016530614346265793, + 0.012156059965491295, + 0.008647005073726177, + 0.00858613196760416, + 0.008192475885152817, + 0.008789259940385818, + 0.013554735109210014, + 0.010869522579014301, + 0.0104508176445961, + 0.006232732906937599, + 0.006914587691426277, + 0.0054158736020326614, + 0.004562633112072945, + 0.004189159255474806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0015124559868127108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic confrontation across the vast, dusty expanse of the Australian outback. Two male red kangaroos engage in a ritual boxing match, their powerful bodies rendered in stark, dark silhouettes. They rear up on their massive hind legs, balancing on their thick, muscular tails against the arid, red-brown earth. The setting sun, low below the horizon, ignites the sky into a breathtaking canvas of fiery red, deep scarlet, and faded violet. This intense backlighting perfectly delineates their forms, capturing the edges of their fur with a subtle, luminous rim light. As they struggle and strike each other—jabbing with their front paws and delivering powerful, deliberate kicks—wisps of fine red dust are kicked up from the dry ground, swirling around their feet and catching the surrounding light, creating a glowing, ethereal mist. The entire scene is imbued with a primal, breathtaking atmosphere, contrasting the raw, untamed energy of the battle with the serene, majestic beauty of the sunset.", + "seed": 3148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_9/index_6940_seed_3148_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02191038290038705, + "vlm_logprob": -0.02191038290038705, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02191038290038705, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030729349702596664 + ], + "opposite_logprob_tokens": [ + -3.530729293823242 + ], + "logprob_reward": -0.030729349702596664, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.030729349702596664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27970388531684875, + 0.22149018943309784, + 0.0924692451953888, + 0.06051891669631004, + 0.13695570826530457, + 0.04616621509194374, + 0.07301132380962372, + 0.021130554378032684, + 0.02576695755124092, + 0.04278700426220894 + ], + "credit_map": [ + 0.02907143346965313, + 0.04245595633983612, + 0.03323423117399216, + 0.021314172074198723, + 0.03853430971503258, + 0.047334618866443634, + 0.026433849707245827, + 0.04132530838251114, + 0.028329521417617798, + 0.025781281292438507, + 0.022158710286021233, + 0.03301968798041344, + 0.020194880664348602, + 0.040383439511060715, + 0.025367438793182373, + 0.026255235075950623, + 0.009391159750521183, + 0.013608909212052822, + 0.008852263912558556, + 0.012953532859683037, + 0.009309317916631699, + 0.011857131496071815, + 0.013884237967431545, + 0.01261268649250269, + 0.00796658918261528, + 0.006953649688512087, + 0.006027962546795607, + 0.007835776545107365, + 0.005786995869129896, + 0.0072387102991342545, + 0.011972292326390743, + 0.006736941169947386, + 0.012551954947412014, + 0.01876242272555828, + 0.015039381571114063, + 0.023195529356598854, + 0.017122892662882805, + 0.015291051007807255, + 0.023919034749269485, + 0.011073438450694084, + 0.005224673543125391, + 0.0065580252557992935, + 0.00483827106654644, + 0.007843539118766785, + 0.00601678853854537, + 0.005530616734176874, + 0.004711188841611147, + 0.00544311385601759, + 0.01323211845010519, + 0.007505892775952816, + 0.008988697081804276, + 0.0076302955858409405, + 0.005983549635857344, + 0.01150326244533062, + 0.007944099605083466, + 0.01022341102361679, + 0.003307935781776905, + 0.0021646181121468544, + 0.002055897144600749, + 0.0025990235153585672, + 0.003169440431520343, + 0.003197765676304698, + 0.002093446208164096, + 0.0025424291379749775, + 0.004567256197333336, + 0.002423506695777178, + 0.0036056239623576403, + 0.0034580498468130827, + 0.003189910901710391, + 0.0038073903415352106, + 0.0024032893124967813, + 0.0023119314573705196, + 0.00531834876164794, + 0.004352212883532047, + 0.005982863716781139, + 0.005621729418635368, + 0.00426875613629818, + 0.007022991310805082, + 0.005075438879430294, + 0.005144662223756313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.030729349702596664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072005033493042 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016072005033493042, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072005033493042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16185225546360016, + 0.2149597406387329, + 0.09885355830192566, + 0.14684221148490906, + 0.08896271884441376, + 0.0874137431383133, + 0.046137358993291855, + 0.06142837554216385, + 0.035438381135463715, + 0.05811165273189545 + ], + "credit_map": [ + 0.019688133150339127, + 0.02146845869719982, + 0.021900422871112823, + 0.01439463160932064, + 0.022851936519145966, + 0.02729804441332817, + 0.015417136251926422, + 0.018833495676517487, + 0.025183744728565216, + 0.039737578481435776, + 0.019104018807411194, + 0.02646200731396675, + 0.021666375920176506, + 0.023312417790293694, + 0.023051567375659943, + 0.03644203394651413, + 0.008136295713484287, + 0.01182363461703062, + 0.009753430262207985, + 0.013456328772008419, + 0.008974408730864525, + 0.014660290442407131, + 0.01612965390086174, + 0.01591951586306095, + 0.015114923007786274, + 0.0148550346493721, + 0.012565110810101032, + 0.019365791231393814, + 0.01816384680569172, + 0.02146298997104168, + 0.03236623480916023, + 0.012948285788297653, + 0.006923322565853596, + 0.008567108772695065, + 0.008614493533968925, + 0.021260546520352364, + 0.009628021158277988, + 0.011769634671509266, + 0.014224022626876831, + 0.007975561544299126, + 0.012072121724486351, + 0.006045433226972818, + 0.007793758530169725, + 0.01876775547862053, + 0.006586261093616486, + 0.013842885382473469, + 0.01262472104281187, + 0.009680806659162045, + 0.00801828969269991, + 0.003684480907395482, + 0.004930975381284952, + 0.004449190106242895, + 0.00405619852244854, + 0.011245684698224068, + 0.004153783433139324, + 0.005598757416009903, + 0.008483403362333775, + 0.004950053058564663, + 0.007035885006189346, + 0.007957193069159985, + 0.006374756805598736, + 0.011703618802130222, + 0.0073834671638906, + 0.007539999205619097, + 0.003989867866039276, + 0.0034544002264738083, + 0.004197016358375549, + 0.004485174547880888, + 0.0037160080391913652, + 0.0063808634877204895, + 0.0029446263797581196, + 0.006270424462854862, + 0.008826250210404396, + 0.005446483381092548, + 0.0064929150976240635, + 0.00692362803965807, + 0.004670495167374611, + 0.013426309451460838, + 0.005428188946098089, + 0.006897382438182831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.016072005033493042 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188944086432457 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.018188944086432457, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188944086432457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14480994641780853, + 0.18214809894561768, + 0.1737075299024582, + 0.27231961488723755, + 0.04675086960196495, + 0.02642485685646534, + 0.03065262734889984, + 0.021704232320189476, + 0.06539670377969742, + 0.03608550503849983 + ], + "credit_map": [ + 0.01375881489366293, + 0.021782545372843742, + 0.016739370301365852, + 0.014440923929214478, + 0.020045708864927292, + 0.01932857185602188, + 0.014730899594724178, + 0.023983119055628777, + 0.013130986131727695, + 0.044469572603702545, + 0.01771528087556362, + 0.021555516868829727, + 0.021060161292552948, + 0.020067309960722923, + 0.01992889679968357, + 0.024220377206802368, + 0.012367863208055496, + 0.024402467533946037, + 0.01689755916595459, + 0.025203874334692955, + 0.0175932664424181, + 0.021764930337667465, + 0.03503274917602539, + 0.020444819703698158, + 0.027789896354079247, + 0.02492649294435978, + 0.024968203157186508, + 0.03773361071944237, + 0.02998245134949684, + 0.03925797715783119, + 0.06128465011715889, + 0.026376325637102127, + 0.003920749295502901, + 0.0043341778218746185, + 0.004935416392982006, + 0.0095058623701334, + 0.006317838095128536, + 0.00522360485047102, + 0.008583176881074905, + 0.003930046223104, + 0.0030037686228752136, + 0.002494067419320345, + 0.0023946701548993587, + 0.006272502243518829, + 0.002838544547557831, + 0.0034711139742285013, + 0.0033231100533157587, + 0.002627080073580146, + 0.0051605356857180595, + 0.003609968349337578, + 0.003094576997682452, + 0.0032047824934124947, + 0.004130842629820108, + 0.004007370676845312, + 0.002967523643746972, + 0.004477027803659439, + 0.0025818042922765017, + 0.002517045708373189, + 0.002479070331901312, + 0.0026296121068298817, + 0.0027409312315285206, + 0.0030817894730716944, + 0.0030097540002316236, + 0.0026642244774848223, + 0.009078619070351124, + 0.004685061983764172, + 0.007997102104127407, + 0.010168006643652916, + 0.007415176369249821, + 0.009025821462273598, + 0.009080884046852589, + 0.007946029305458069, + 0.0033565855119377375, + 0.003507522866129875, + 0.00440643448382616, + 0.006350673735141754, + 0.004065435379743576, + 0.004685365129262209, + 0.004842562135308981, + 0.004870925098657608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.018188944086432457 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043051183223724365 + ], + "opposite_logprob_tokens": [ + -3.168051242828369 + ], + "logprob_reward": -0.043051183223724365, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043051183223724365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31621742248535156, + 0.09314408153295517, + 0.07276806980371475, + 0.0627230703830719, + 0.127201646566391, + 0.1520725041627884, + 0.04254141449928284, + 0.07247774302959442, + 0.028922561556100845, + 0.031931474804878235 + ], + "credit_map": [ + 0.04211125150322914, + 0.04534154385328293, + 0.041931312531232834, + 0.023275110870599747, + 0.043059077113866806, + 0.053195640444755554, + 0.024496622383594513, + 0.04280683398246765, + 0.009601941332221031, + 0.01801462285220623, + 0.012288753874599934, + 0.010009272955358028, + 0.008695228956639767, + 0.010542926378548145, + 0.010177123360335827, + 0.013814212754368782, + 0.01617998257279396, + 0.007618119474500418, + 0.00772649934515357, + 0.005915460176765919, + 0.007803880609571934, + 0.011919925920665264, + 0.006651121191680431, + 0.008953079581260681, + 0.01377832517027855, + 0.005768787115812302, + 0.0083044758066535, + 0.005144548136740923, + 0.005512741860002279, + 0.01148828025907278, + 0.006452413275837898, + 0.006273495499044657, + 0.014914390631020069, + 0.01413891464471817, + 0.011988705955445766, + 0.027393624186515808, + 0.013277662917971611, + 0.01705750823020935, + 0.017357967793941498, + 0.011072881519794464, + 0.018385203555226326, + 0.010930577293038368, + 0.013728988356888294, + 0.03557000681757927, + 0.011459658853709698, + 0.025580955669283867, + 0.020079772919416428, + 0.01633734069764614, + 0.0071278223767876625, + 0.004524612333625555, + 0.004218964837491512, + 0.006253397092223167, + 0.004671453032642603, + 0.006469110958278179, + 0.003828774206340313, + 0.0054472824558615685, + 0.01338710356503725, + 0.005173336714506149, + 0.009837416000664234, + 0.01148679293692112, + 0.006420420948415995, + 0.010622337460517883, + 0.007616972550749779, + 0.007933359593153, + 0.004023358225822449, + 0.002585623413324356, + 0.003822009777650237, + 0.003916710149496794, + 0.00225620623677969, + 0.005073429550975561, + 0.004189118277281523, + 0.0030561049934476614, + 0.005976781249046326, + 0.0030537538696080446, + 0.004104877356439829, + 0.0027394029311835766, + 0.002092940267175436, + 0.007862632162868977, + 0.0024471720680594444, + 0.0036539146676659584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.043051183223724365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510432455688715 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510432455688715, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510432455688715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33918800950050354, + 0.13381843268871307, + 0.07035113126039505, + 0.06703196465969086, + 0.06781574338674545, + 0.10077282786369324, + 0.08369556814432144, + 0.044108811765909195, + 0.045084398239851, + 0.048133090138435364 + ], + "credit_map": [ + 0.05563117563724518, + 0.060962777584791183, + 0.04405580461025238, + 0.03860802575945854, + 0.028712833300232887, + 0.05344128981232643, + 0.02302701026201248, + 0.0347491130232811, + 0.016323478892445564, + 0.02994617074728012, + 0.013051803223788738, + 0.01243306789547205, + 0.013642444275319576, + 0.01594291254878044, + 0.017308015376329422, + 0.015170547179877758, + 0.009982463903725147, + 0.00889419112354517, + 0.00841602310538292, + 0.008290545083582401, + 0.007734035607427359, + 0.010652593336999416, + 0.007687927223742008, + 0.00869335699826479, + 0.012289655394852161, + 0.007913543842732906, + 0.008255193941295147, + 0.007560661528259516, + 0.006441017612814903, + 0.00849790871143341, + 0.009641599841415882, + 0.006432391237467527, + 0.00808793120086193, + 0.008554128929972649, + 0.006751297041773796, + 0.01285234559327364, + 0.0065984162501990795, + 0.0102261146530509, + 0.009882932528853416, + 0.0048625837080180645, + 0.018488174304366112, + 0.007159342523664236, + 0.010792716406285763, + 0.017651241272687912, + 0.007818530313670635, + 0.01916452683508396, + 0.009109550155699253, + 0.0105887521058321, + 0.011948428116738796, + 0.0092289038002491, + 0.008965096436440945, + 0.00988609716296196, + 0.008235117420554161, + 0.012496472336351871, + 0.008347645401954651, + 0.014587809331715107, + 0.006409736815840006, + 0.004293326288461685, + 0.004680159501731396, + 0.005761250853538513, + 0.004316851031035185, + 0.007627010345458984, + 0.005419866181910038, + 0.005600608419626951, + 0.006915793754160404, + 0.004248124081641436, + 0.005441534798592329, + 0.005354701075702906, + 0.003744689282029867, + 0.007628262974321842, + 0.004562717862427235, + 0.007188576273620129, + 0.010743677616119385, + 0.0055657424964010715, + 0.006115881260484457, + 0.005473878234624863, + 0.0032039424404501915, + 0.008827975019812584, + 0.0031534405425190926, + 0.005048557184636593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.001510432455688715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic confrontation across the vast, dusty expanse of the Australian outback. Two male red kangaroos engage in a ritual boxing match, their powerful bodies rendered in stark, dark silhouettes. They rear up on their massive hind legs, balancing on their thick, muscular tails against the arid, red-brown earth. The setting sun, low below the horizon, ignites the sky into a breathtaking canvas of fiery red, deep scarlet, and faded violet. This intense backlighting perfectly delineates their forms, capturing the edges of their fur with a subtle, luminous rim light. As they struggle and strike each other—jabbing with their front paws and delivering powerful, deliberate kicks—wisps of fine red dust are kicked up from the dry ground, swirling around their feet and catching the surrounding light, creating a glowing, ethereal mist. The entire scene is imbued with a primal, breathtaking atmosphere, contrasting the raw, untamed energy of the battle with the serene, majestic beauty of the sunset.", + "seed": 3149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000018/group_9/index_6940_seed_3149_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022213595104403794, + "vlm_logprob": -0.022213595104403794, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.022213595104403794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025506602600216866 + ], + "opposite_logprob_tokens": [ + -3.7755064964294434 + ], + "logprob_reward": -0.025506602600216866, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025506602600216866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27052927017211914, + 0.14321787655353546, + 0.1353524625301361, + 0.08277592808008194, + 0.04034075140953064, + 0.029963361099362373, + 0.07879883050918579, + 0.06203436478972435, + 0.07143689692020416, + 0.08555028587579727 + ], + "credit_map": [ + 0.032925546169281006, + 0.06969430297613144, + 0.02226901799440384, + 0.017276301980018616, + 0.020192228257656097, + 0.06799504160881042, + 0.01773471012711525, + 0.022442124783992767, + 0.03148791193962097, + 0.020748529583215714, + 0.010119331069290638, + 0.014123620465397835, + 0.012050328776240349, + 0.024039559066295624, + 0.012262103147804737, + 0.0183864813297987, + 0.03582439199090004, + 0.0221547968685627, + 0.009933196008205414, + 0.014729473739862442, + 0.01252109743654728, + 0.021867288276553154, + 0.007652636617422104, + 0.010669572278857231, + 0.015415416099131107, + 0.010151010937988758, + 0.010024802759289742, + 0.010454566217958927, + 0.0056430380791425705, + 0.009279840625822544, + 0.008172190748155117, + 0.013635074719786644, + 0.006128111854195595, + 0.0036193716805428267, + 0.004629763308912516, + 0.004752758890390396, + 0.0031754490919411182, + 0.004834463354200125, + 0.0055278330110013485, + 0.007673000916838646, + 0.003675696440041065, + 0.002799518872052431, + 0.0028732509817928076, + 0.00396473566070199, + 0.0037435265257954597, + 0.004214902874082327, + 0.004464531317353249, + 0.0042271981947124004, + 0.00817491952329874, + 0.006825522519648075, + 0.0076220170594751835, + 0.012280761264264584, + 0.008256846107542515, + 0.0113506019115448, + 0.01050612423568964, + 0.01378202997148037, + 0.005053999368101358, + 0.0049335649237036705, + 0.007139463443309069, + 0.008652704767882824, + 0.006264330353587866, + 0.007319468539208174, + 0.010466789826750755, + 0.012204041704535484, + 0.005652604624629021, + 0.006679339334368706, + 0.008123677223920822, + 0.011091131716966629, + 0.00905640609562397, + 0.009835440665483475, + 0.008145095780491829, + 0.012853199616074562, + 0.010646570473909378, + 0.0075336629524827, + 0.00877515971660614, + 0.010201489552855492, + 0.01238359697163105, + 0.012169977650046349, + 0.010325334034860134, + 0.013514497317373753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.025506602600216866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084762401878834 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084762401878834, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084762401878834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21345345675945282, + 0.11701537668704987, + 0.18416638672351837, + 0.09756998717784882, + 0.04683661833405495, + 0.06328554451465607, + 0.06614270061254501, + 0.07222843170166016, + 0.07055874168872833, + 0.0687427818775177 + ], + "credit_map": [ + 0.016001101583242416, + 0.047237593680620193, + 0.019675523042678833, + 0.0198760274797678, + 0.016375793144106865, + 0.04706934094429016, + 0.015078495256602764, + 0.03213958442211151, + 0.011915205046534538, + 0.009540906175971031, + 0.012076118029654026, + 0.016908373683691025, + 0.009344443678855896, + 0.01802998222410679, + 0.01595310866832733, + 0.023247243836522102, + 0.043758898973464966, + 0.02771342732012272, + 0.01399659737944603, + 0.023344535380601883, + 0.017867881804704666, + 0.026483234018087387, + 0.014545289799571037, + 0.016456522047519684, + 0.015214974991977215, + 0.009483570232987404, + 0.012057801708579063, + 0.011026429943740368, + 0.007233334705233574, + 0.011641210876405239, + 0.01099354587495327, + 0.019919121637940407, + 0.004019211977720261, + 0.004265029914677143, + 0.006342156790196896, + 0.00674488814547658, + 0.0033552972599864006, + 0.007026828825473785, + 0.0050473264418542385, + 0.010035878047347069, + 0.006709368899464607, + 0.0038345009088516235, + 0.006426907144486904, + 0.009077919647097588, + 0.00584682822227478, + 0.008814247325062752, + 0.01304894220083952, + 0.009526827372610569, + 0.007394829764962196, + 0.005667424760758877, + 0.006518764887005091, + 0.009694504551589489, + 0.007578542456030846, + 0.009709738194942474, + 0.008279426954686642, + 0.011299475096166134, + 0.005276827607303858, + 0.006476207170635462, + 0.010802271775901318, + 0.011426248587667942, + 0.007819118909537792, + 0.010670892894268036, + 0.007942922413349152, + 0.011813940480351448, + 0.005453928839415312, + 0.006654775235801935, + 0.00741207879036665, + 0.007266045082360506, + 0.0066875494085252285, + 0.010276076383888721, + 0.00790590513497591, + 0.018902381882071495, + 0.009115024469792843, + 0.006898984778672457, + 0.010408484376966953, + 0.009105455130338669, + 0.010014563798904419, + 0.009456079453229904, + 0.006498571019619703, + 0.007245613262057304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.011084762401878834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190348520874977 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.018190348520874977, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190348520874977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24256007373332977, + 0.17648938298225403, + 0.058879561722278595, + 0.040644221007823944, + 0.049736157059669495, + 0.04651830717921257, + 0.21516945958137512, + 0.05755103752017021, + 0.04578910022974014, + 0.06666268408298492 + ], + "credit_map": [ + 0.01999649405479431, + 0.0659177228808403, + 0.017236553132534027, + 0.014637010172009468, + 0.020062129944562912, + 0.06757993251085281, + 0.01573249138891697, + 0.02139774151146412, + 0.03255720064043999, + 0.022188955917954445, + 0.015372436493635178, + 0.02279430441558361, + 0.015949491411447525, + 0.020600324496626854, + 0.025584347546100616, + 0.02144232578575611, + 0.00539644667878747, + 0.005277989432215691, + 0.0077047934755682945, + 0.010269347578287125, + 0.007320634555071592, + 0.005927461199462414, + 0.007621699012815952, + 0.009361191652715206, + 0.004210320767015219, + 0.0037379562854766846, + 0.005107831675559282, + 0.006212259642779827, + 0.003743042703717947, + 0.004396306350827217, + 0.006722947116941214, + 0.006513558328151703, + 0.006865424569696188, + 0.0043468112125992775, + 0.00472736731171608, + 0.007173897698521614, + 0.003597644157707691, + 0.007390000857412815, + 0.005862024612724781, + 0.009772984310984612, + 0.005898688919842243, + 0.004060579929500818, + 0.0051799193024635315, + 0.0068784961476922035, + 0.004812662489712238, + 0.00673849368467927, + 0.006010779179632664, + 0.006938684266060591, + 0.01703805848956108, + 0.01566498726606369, + 0.026734599843621254, + 0.032651618123054504, + 0.019570516422390938, + 0.024082094430923462, + 0.030004659667611122, + 0.04942292720079422, + 0.008317455649375916, + 0.007151851896196604, + 0.005139159504324198, + 0.007992955856025219, + 0.005138217005878687, + 0.007848883047699928, + 0.008917667903006077, + 0.00704484898597002, + 0.0034438890870660543, + 0.006318947765976191, + 0.004394093994051218, + 0.004153149668127298, + 0.005737288389354944, + 0.008754132315516472, + 0.003870762651786208, + 0.009116836823523045, + 0.009231111966073513, + 0.007353400811553001, + 0.005973141174763441, + 0.008159828372299671, + 0.008783412165939808, + 0.00962071679532528, + 0.006841421127319336, + 0.010699652135372162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.018190348520874977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495046079158783 + ], + "opposite_logprob_tokens": [ + -2.929950475692749 + ], + "logprob_reward": -0.05495046079158783, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495046079158783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2662985324859619, + 0.056325484067201614, + 0.1647043377161026, + 0.11011004447937012, + 0.033756669610738754, + 0.05321981757879257, + 0.04481860622763634, + 0.06615705043077469, + 0.08883548527956009, + 0.11577395349740982 + ], + "credit_map": [ + 0.032786522060632706, + 0.06326870620250702, + 0.024238934740424156, + 0.017537860199809074, + 0.014229926280677319, + 0.06246999651193619, + 0.01652657985687256, + 0.03523999825119972, + 0.008868535980582237, + 0.007567072752863169, + 0.005730089731514454, + 0.007214933633804321, + 0.005871948320418596, + 0.00652502104640007, + 0.006688535213470459, + 0.007859352044761181, + 0.04220099747180939, + 0.02487776055932045, + 0.01513065304607153, + 0.016564760357141495, + 0.014191032387316227, + 0.026074644178152084, + 0.009263948537409306, + 0.016400543972849846, + 0.024398677051067352, + 0.014988679438829422, + 0.011049632914364338, + 0.011753574945032597, + 0.008205484598875046, + 0.013668976724147797, + 0.00901060737669468, + 0.01703440025448799, + 0.004785240162163973, + 0.00406981585547328, + 0.0032219025306403637, + 0.004574380349367857, + 0.004187292419373989, + 0.004318860359489918, + 0.00381727353669703, + 0.0047819046303629875, + 0.008625160902738571, + 0.003495733719319105, + 0.006065938621759415, + 0.005341133568435907, + 0.005024651996791363, + 0.007993328385055065, + 0.00814130436629057, + 0.0085325688123703, + 0.006849423050880432, + 0.004162076395004988, + 0.006664914079010487, + 0.004071488510817289, + 0.004413349088281393, + 0.007995571941137314, + 0.0030932652298361063, + 0.0075685190968215466, + 0.01008624117821455, + 0.008764808066189289, + 0.005532706156373024, + 0.0059641283005476, + 0.0059293173253536224, + 0.00848014373332262, + 0.009766100905835629, + 0.01163360383361578, + 0.006483770906925201, + 0.00790487602353096, + 0.009620346128940582, + 0.010505239479243755, + 0.0067474255338311195, + 0.013113490305840969, + 0.010605670511722565, + 0.02385466732084751, + 0.01448678970336914, + 0.010344281792640686, + 0.013688845559954643, + 0.014675389975309372, + 0.015722760930657387, + 0.015874017030000687, + 0.011683976277709007, + 0.019297892227768898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.05495046079158783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two male red kangaroos visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the kangaroos rearing up on their hind legs and balancing on their tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the kangaroos engaged in a boxing match involving jabs with front paws and kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment the Australian outback with red-brown earth and wisps of red dust swirling around the kangaroos' feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky colored with fiery red, deep scarlet, and faded violet due to a low setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335801207460463 + ], + "opposite_logprob_tokens": [ + -6.626335620880127 + ], + "logprob_reward": -0.001335801207460463, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335801207460463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29085037112236023, + 0.08316801488399506, + 0.1515592783689499, + 0.117977075278759, + 0.04549847170710564, + 0.0481809601187706, + 0.07689428329467773, + 0.06651611626148224, + 0.053113218396902084, + 0.06624223291873932 + ], + "credit_map": [ + 0.04693453758955002, + 0.05498291179537773, + 0.025721650570631027, + 0.030476830899715424, + 0.01962246373295784, + 0.04638652876019478, + 0.026933422312140465, + 0.0397920086979866, + 0.018916696310043335, + 0.012799237854778767, + 0.0074663348495960236, + 0.01010639313608408, + 0.005987051408737898, + 0.013673106208443642, + 0.005701421294361353, + 0.00851777009665966, + 0.042183831334114075, + 0.026063716039061546, + 0.009911016561090946, + 0.017313485965132713, + 0.013754787854850292, + 0.023139921948313713, + 0.008425631560385227, + 0.010766878724098206, + 0.025584017857909203, + 0.014162345789372921, + 0.012078366242349148, + 0.013965192250907421, + 0.00813420582562685, + 0.015220845118165016, + 0.008438432589173317, + 0.02039366401731968, + 0.0054292697459459305, + 0.0036445772275328636, + 0.0064526149071753025, + 0.005670846439898014, + 0.0034136585891246796, + 0.005678622052073479, + 0.0068340725265443325, + 0.008374808356165886, + 0.007138036657124758, + 0.0035570107866078615, + 0.005579367745667696, + 0.00766020780429244, + 0.004237061832100153, + 0.0063807060942053795, + 0.00647320318967104, + 0.00715536018833518, + 0.006091293878853321, + 0.005355514585971832, + 0.00961207039654255, + 0.011568240821361542, + 0.006123808678239584, + 0.008852190338075161, + 0.01222024392336607, + 0.017070917412638664, + 0.008546130731701851, + 0.006960269995033741, + 0.010609213262796402, + 0.005891323555260897, + 0.005869586020708084, + 0.00901753455400467, + 0.008578585460782051, + 0.011043473146855831, + 0.0031071624252945185, + 0.005029124207794666, + 0.007013235706835985, + 0.005949772894382477, + 0.005223271436989307, + 0.006089698523283005, + 0.006419525947421789, + 0.014281424693763256, + 0.008314836770296097, + 0.00636576022952795, + 0.008523798547685146, + 0.008092382922768593, + 0.010371877811849117, + 0.01000465638935566, + 0.005880211945623159, + 0.00868870411068201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.001335801207460463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_0/index_tensor([5415, 5415])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_0/index_tensor([5415, 5415])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ee0175529219bf60fd7c91d2b5eb8f9aa88015d1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_0/index_tensor([5415, 5415])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the vibrant style of Disney Pixar animation, a static, slightly low-angled, full-body shot captures a captivating and determined Indian cobra. Its upper body is supported by protective coils on the ground, its signature thorn-patterned hood fully open, revealing bold black markings that contrast starkly against its smooth, golden-brown animated scales, which have a soft, plasticine-like sheen. Its large, expressive dark brown eyes, wide with genuine concern, gaze forward, guarding a small clutch of glossy, off-white eggs nestled in a depression on the mossy jungle floor. It sways with a deliberate, rhythmic movement, a classic defensive dance humorously undercut by an oversized, comedic sweatdrop that meticulously rolls down its temples, catching a glint of light as it follows a path along its smooth scales. The scene is set in a colorful jungle clearing, bathed in sunlight, with warm golden shafts piercing a dense canopy of oversized, waxy green leaves. The backdrop is a soft-focus tapestry of lush foliage and vibrant tropical flowers in shades of fuchsia and sunny yellow, creating a warm, lovely atmosphere filled with easy tension and parental devotion.", + "seed": 11619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_0/index_5415_seed_11619_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05549765843898058, + "vlm_logprob": -0.05549765843898058, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.05549765843898058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09072728455066681 + ], + "opposite_logprob_tokens": [ + -2.4657273292541504 + ], + "logprob_reward": -0.09072728455066681, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09072728455066681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2641557455062866, + 0.100968137383461, + 0.1079457700252533, + 0.0390876829624176, + 0.07898207753896713, + 0.088164322078228, + 0.06362766772508621, + 0.07586496323347092, + 0.04540914297103882, + 0.135794535279274 + ], + "credit_map": [ + 0.02806548774242401, + 0.030091414228081703, + 0.014162729494273663, + 0.02678624726831913, + 0.050772637128829956, + 0.031995613127946854, + 0.05048752203583717, + 0.031794093549251556, + 0.007381001953035593, + 0.008180982433259487, + 0.010157352313399315, + 0.01986672729253769, + 0.018266193568706512, + 0.011612755246460438, + 0.013552737422287464, + 0.011950389482080936, + 0.013237951323390007, + 0.011959103867411613, + 0.01208201702684164, + 0.009827785193920135, + 0.019130690023303032, + 0.011931431479752064, + 0.016668541356921196, + 0.013108246959745884, + 0.002515150234103203, + 0.005254616029560566, + 0.0035379219334572554, + 0.0059088594280183315, + 0.006624969653785229, + 0.003934830892831087, + 0.006702783051878214, + 0.004608550108969212, + 0.008414918556809425, + 0.006650016643106937, + 0.005712755955755711, + 0.011698340997099876, + 0.013963916338980198, + 0.007539441343396902, + 0.009842506609857082, + 0.015160183422267437, + 0.009525872766971588, + 0.007915596477687359, + 0.007300626486539841, + 0.012093770317733288, + 0.0186243187636137, + 0.009940487332642078, + 0.010648595169186592, + 0.012115053832530975, + 0.005039438139647245, + 0.0055010621435940266, + 0.0062609040178358555, + 0.011100858449935913, + 0.009800932370126247, + 0.007969453930854797, + 0.010032846592366695, + 0.007922174409031868, + 0.006599039770662785, + 0.009446217678487301, + 0.00875464640557766, + 0.009919057600200176, + 0.009888574481010437, + 0.009562482126057148, + 0.013039586134254932, + 0.008655359968543053, + 0.005180566105991602, + 0.005125220865011215, + 0.0033271154388785362, + 0.00552355544641614, + 0.0057481564581394196, + 0.0057996525429189205, + 0.008177830837666988, + 0.006527046672999859, + 0.013446273282170296, + 0.015221759676933289, + 0.00922473706305027, + 0.016380853950977325, + 0.023700809106230736, + 0.018469320610165596, + 0.020236440002918243, + 0.019114337861537933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.09072728455066681 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744596175849438 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744596175849438, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744596175849438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1652008593082428, + 0.12387044727802277, + 0.08990488946437836, + 0.07366432994604111, + 0.14577960968017578, + 0.08334873616695404, + 0.04790583997964859, + 0.027005555108189583, + 0.13556969165802002, + 0.10775000602006912 + ], + "credit_map": [ + 0.014183295890688896, + 0.01816382259130478, + 0.012840943410992622, + 0.015940701588988304, + 0.03262385353446007, + 0.02155141718685627, + 0.033496566116809845, + 0.01640027016401291, + 0.015031132847070694, + 0.010007024742662907, + 0.01170463114976883, + 0.012403049506247044, + 0.01809008978307247, + 0.01625155471265316, + 0.024857625365257263, + 0.015525354072451591, + 0.008384833112359047, + 0.008708341978490353, + 0.007068166509270668, + 0.010686216875910759, + 0.015050530433654785, + 0.009768609888851643, + 0.016103949397802353, + 0.01413425151258707, + 0.007203494664281607, + 0.009061582386493683, + 0.005870783235877752, + 0.007617711555212736, + 0.013150662183761597, + 0.008701509796082973, + 0.012471945956349373, + 0.00958664808422327, + 0.009962278418242931, + 0.021804798394441605, + 0.010620309971272945, + 0.019561992958188057, + 0.02906266786158085, + 0.01592062972486019, + 0.02527444250881672, + 0.01357249915599823, + 0.007143180817365646, + 0.007978525944054127, + 0.008396930992603302, + 0.019786905497312546, + 0.00878907460719347, + 0.009284254163503647, + 0.007755839265882969, + 0.014214036986231804, + 0.0060370611026883125, + 0.005946172866970301, + 0.004221615847200155, + 0.004029461182653904, + 0.009014392271637917, + 0.005780711770057678, + 0.0064657023176550865, + 0.0064107272773981094, + 0.002563217654824257, + 0.003111323807388544, + 0.001914464752189815, + 0.0032918392680585384, + 0.005387625191360712, + 0.0029956544749438763, + 0.0049807364121079445, + 0.0027606955263763666, + 0.012712312862277031, + 0.013284253887832165, + 0.010274858213961124, + 0.02068646065890789, + 0.017264872789382935, + 0.017996562644839287, + 0.026819774881005287, + 0.016530603170394897, + 0.008716938085854053, + 0.013032684102654457, + 0.006903538014739752, + 0.009190854616463184, + 0.018730074167251587, + 0.012650759890675545, + 0.020802397280931473, + 0.01772276870906353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.006744596175849438 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033703088760376 + ], + "opposite_logprob_tokens": [ + -1.910336971282959 + ], + "logprob_reward": -0.16033703088760376, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033703088760376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2119535207748413, + 0.16463616490364075, + 0.12964308261871338, + 0.07199535518884659, + 0.08287917822599411, + 0.041127242147922516, + 0.10351788252592087, + 0.05888456851243973, + 0.10363957285881042, + 0.03172343224287033 + ], + "credit_map": [ + 0.020025813952088356, + 0.022426607087254524, + 0.015867939218878746, + 0.018542233854532242, + 0.039512746036052704, + 0.03176744282245636, + 0.03881678357720375, + 0.024993956089019775, + 0.01594681665301323, + 0.015213665552437305, + 0.013856152072548866, + 0.02816307730972767, + 0.021413784474134445, + 0.023936213925480843, + 0.025441603735089302, + 0.020664867013692856, + 0.011873963288962841, + 0.013666000217199326, + 0.012829888612031937, + 0.016172068193554878, + 0.0209113247692585, + 0.013540636748075485, + 0.019536886364221573, + 0.02111232280731201, + 0.0070595708675682545, + 0.008098484016954899, + 0.004650732968002558, + 0.009979347698390484, + 0.010593044571578503, + 0.008634570986032486, + 0.012344037182629108, + 0.010635566897690296, + 0.008471725508570671, + 0.011519748717546463, + 0.00791481975466013, + 0.013242345303297043, + 0.013137923553586006, + 0.009172149933874607, + 0.010288764722645283, + 0.009131712839007378, + 0.004150875378400087, + 0.00398851465433836, + 0.004177155438810587, + 0.006765538360923529, + 0.005697435233741999, + 0.004614023491740227, + 0.005882852710783482, + 0.005850849207490683, + 0.008567185141146183, + 0.011607732623815536, + 0.009127235040068626, + 0.018572304397821426, + 0.015042967163026333, + 0.013569037429988384, + 0.015341035090386868, + 0.011690392158925533, + 0.004300262313336134, + 0.005647079087793827, + 0.005507712252438068, + 0.009272395633161068, + 0.007231866475194693, + 0.00751207210123539, + 0.010132706724107265, + 0.009280476719141006, + 0.009050378575921059, + 0.009671549312770367, + 0.010565351694822311, + 0.01617877371609211, + 0.010942448861896992, + 0.01427503488957882, + 0.017245637252926826, + 0.015710406005382538, + 0.004093687981367111, + 0.003827288979664445, + 0.0026877436321228743, + 0.002980798250064254, + 0.00510677183046937, + 0.004967892542481422, + 0.00449387775734067, + 0.0035653722006827593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.16033703088760376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067076474428177 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016067076474428177, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067076474428177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.231583371758461, + 0.09763386845588684, + 0.07416808605194092, + 0.05763708055019379, + 0.21421407163143158, + 0.09189312905073166, + 0.03527979925274849, + 0.06615084409713745, + 0.10051252692937851, + 0.030927224084734917 + ], + "credit_map": [ + 0.02489883080124855, + 0.030265258625149727, + 0.019360072910785675, + 0.022696398198604584, + 0.03826957195997238, + 0.025313708931207657, + 0.035719133913517, + 0.03506040573120117, + 0.009996410459280014, + 0.009319029748439789, + 0.008765185251832008, + 0.011699309572577477, + 0.01320208702236414, + 0.013643252663314342, + 0.019611217081546783, + 0.011397374793887138, + 0.007157705258578062, + 0.005784578621387482, + 0.006041456013917923, + 0.011046383529901505, + 0.012808019295334816, + 0.007889524102210999, + 0.014034140855073929, + 0.009406276978552341, + 0.006730587221682072, + 0.004623968619853258, + 0.007329326588660479, + 0.008259274065494537, + 0.007555512711405754, + 0.006098573561757803, + 0.008017078042030334, + 0.009022760204970837, + 0.018173670396208763, + 0.03386462852358818, + 0.015229295939207077, + 0.022648965939879417, + 0.040891382843256, + 0.02469531260430813, + 0.0384228490293026, + 0.020287949591875076, + 0.007273146882653236, + 0.007507952395826578, + 0.006259181536734104, + 0.013153922744095325, + 0.02275684103369713, + 0.010797184891998768, + 0.0133529556915164, + 0.010791948065161705, + 0.0032003233209252357, + 0.0036350495647639036, + 0.0028551893774420023, + 0.0060453047044575214, + 0.006682773120701313, + 0.003964537288993597, + 0.004273909144103527, + 0.004622712265700102, + 0.006459589581936598, + 0.004432071931660175, + 0.007044812198728323, + 0.01561372634023428, + 0.009532547555863857, + 0.007126904092729092, + 0.005626040510833263, + 0.010315152816474438, + 0.010035854764282703, + 0.009429906494915485, + 0.009143037721514702, + 0.01982145756483078, + 0.014327660202980042, + 0.011234848760068417, + 0.013853427954018116, + 0.012666339986026287, + 0.0035847455728799105, + 0.004320061299949884, + 0.0034248328302055597, + 0.003242270555347204, + 0.0051773483864963055, + 0.0038833296857774258, + 0.0037110422272235155, + 0.003583593061193824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.016067076474428177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036123041063547134 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.0036123041063547134, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036123041063547134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21171820163726807, + 0.1845691055059433, + 0.07817845046520233, + 0.031224068254232407, + 0.0851806029677391, + 0.09022129327058792, + 0.09667830914258957, + 0.0509670116007328, + 0.06946990638971329, + 0.1017930805683136 + ], + "credit_map": [ + 0.03243136405944824, + 0.019181903451681137, + 0.018623346462845802, + 0.025994552299380302, + 0.031105220317840576, + 0.025703618302941322, + 0.029477981850504875, + 0.02920021116733551, + 0.02165828086435795, + 0.013277421705424786, + 0.020678725093603134, + 0.04302261397242546, + 0.015610773116350174, + 0.026069797575473785, + 0.01660991460084915, + 0.02764158509671688, + 0.010666624642908573, + 0.007193358149379492, + 0.010903664864599705, + 0.012955042533576488, + 0.006850233767181635, + 0.011095499619841576, + 0.00858373288065195, + 0.009930295869708061, + 0.0038831443525850773, + 0.0030221124179661274, + 0.0034933651331812143, + 0.006191383581608534, + 0.004585334099829197, + 0.0028035766445100307, + 0.003943176008760929, + 0.003301978809759021, + 0.0082634212449193, + 0.008567549288272858, + 0.00904370378702879, + 0.017111094668507576, + 0.01359700970351696, + 0.009660501033067703, + 0.008259446360170841, + 0.01067788153886795, + 0.01394775789231062, + 0.006385261192917824, + 0.010242624208331108, + 0.012351911514997482, + 0.012507515959441662, + 0.009500794112682343, + 0.009756987914443016, + 0.015528438612818718, + 0.010751032270491123, + 0.008591248653829098, + 0.007963664829730988, + 0.016330115497112274, + 0.01779559813439846, + 0.010822105221450329, + 0.011953276582062244, + 0.012471278198063374, + 0.007851019501686096, + 0.0034369397908449173, + 0.005662008188664913, + 0.009298346936702728, + 0.005805269815027714, + 0.006374191492795944, + 0.004374668467789888, + 0.008164572529494762, + 0.005471587181091309, + 0.012394852936267853, + 0.004523054696619511, + 0.0064237997867167, + 0.014181804843246937, + 0.008036548271775246, + 0.013239906169474125, + 0.005198351573199034, + 0.011621121317148209, + 0.015405033715069294, + 0.007815408520400524, + 0.01380822155624628, + 0.015583045780658722, + 0.011531973257660866, + 0.012883969582617283, + 0.013144298456609249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0036123041063547134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the vibrant style of Disney Pixar animation, a static, slightly low-angled, full-body shot captures a captivating and determined Indian cobra. Its upper body is supported by protective coils on the ground, its signature thorn-patterned hood fully open, revealing bold black markings that contrast starkly against its smooth, golden-brown animated scales, which have a soft, plasticine-like sheen. Its large, expressive dark brown eyes, wide with genuine concern, gaze forward, guarding a small clutch of glossy, off-white eggs nestled in a depression on the mossy jungle floor. It sways with a deliberate, rhythmic movement, a classic defensive dance humorously undercut by an oversized, comedic sweatdrop that meticulously rolls down its temples, catching a glint of light as it follows a path along its smooth scales. The scene is set in a colorful jungle clearing, bathed in sunlight, with warm golden shafts piercing a dense canopy of oversized, waxy green leaves. The backdrop is a soft-focus tapestry of lush foliage and vibrant tropical flowers in shades of fuchsia and sunny yellow, creating a warm, lovely atmosphere filled with easy tension and parental devotion.", + "seed": 11620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_0/index_5415_seed_11620_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02870723595842719, + "vlm_logprob": -0.02870723595842719, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02870723595842719, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022123442962765694 + ], + "opposite_logprob_tokens": [ + -3.897123336791992 + ], + "logprob_reward": -0.022123442962765694, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022123442962765694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15139515697956085, + 0.024389665573835373, + 0.06530921161174774, + 0.037208475172519684, + 0.19363383948802948, + 0.11578353494405746, + 0.20663341879844666, + 0.06180894747376442, + 0.10958454012870789, + 0.034253232181072235 + ], + "credit_map": [ + 0.009978676214814186, + 0.008564673364162445, + 0.02694077603518963, + 0.012497521936893463, + 0.018787533044815063, + 0.02252371422946453, + 0.01801205240190029, + 0.034090202301740646, + 0.0017437805654481053, + 0.002203158335760236, + 0.0036944274324923754, + 0.0026915050111711025, + 0.004089720547199249, + 0.0034377656411379576, + 0.0029426703695207834, + 0.0035866356920450926, + 0.006957857869565487, + 0.004669912625104189, + 0.013556902296841145, + 0.008212887682020664, + 0.00827476754784584, + 0.005025725811719894, + 0.009137257933616638, + 0.009473896585404873, + 0.002407015999779105, + 0.0029947427101433277, + 0.008241048082709312, + 0.008475725539028645, + 0.004765729419887066, + 0.0036715762689709663, + 0.002682832535356283, + 0.003969804383814335, + 0.013164832256734371, + 0.01296977512538433, + 0.02836349979043007, + 0.020493246614933014, + 0.03767882660031319, + 0.021081095561385155, + 0.03326477110385895, + 0.026617800816893578, + 0.008560337126255035, + 0.015078810043632984, + 0.012190197594463825, + 0.014505513943731785, + 0.02423635497689247, + 0.012275386601686478, + 0.01549617387354374, + 0.013440761715173721, + 0.0156814344227314, + 0.02367275021970272, + 0.024809055030345917, + 0.02428644709289074, + 0.04226510971784592, + 0.02586504817008972, + 0.025199158117175102, + 0.024854391813278198, + 0.005127199459820986, + 0.004116899333894253, + 0.015483763068914413, + 0.008995766751468182, + 0.007505947258323431, + 0.00710611417889595, + 0.005467699375003576, + 0.008005557581782341, + 0.005479196086525917, + 0.008821288123726845, + 0.012809589505195618, + 0.012756248004734516, + 0.025147980079054832, + 0.01257419865578413, + 0.01200234517455101, + 0.019993694499135017, + 0.0022981830406934023, + 0.003577956696972251, + 0.005632070358842611, + 0.003701836336404085, + 0.004046150948852301, + 0.006153167691081762, + 0.002546052448451519, + 0.0062978132627904415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.022123442962765694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085351929068565 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085351929068565, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085351929068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09439220279455185, + 0.08417228609323502, + 0.23550835251808167, + 0.034573547542095184, + 0.1179586723446846, + 0.12310667335987091, + 0.18245653808116913, + 0.06361246854066849, + 0.04426034912467003, + 0.019958941265940666 + ], + "credit_map": [ + 0.008126890286803246, + 0.006725572980940342, + 0.01740395836532116, + 0.008632590062916279, + 0.010012823157012463, + 0.014210295863449574, + 0.010190007276833057, + 0.01909005455672741, + 0.005651514977216721, + 0.0058110561221838, + 0.015549788251519203, + 0.010733487084507942, + 0.017292937263846397, + 0.008917341940104961, + 0.01081631425768137, + 0.009399835020303726, + 0.04571061581373215, + 0.018176060169935226, + 0.03469965606927872, + 0.02924186922609806, + 0.024694178253412247, + 0.041439808905124664, + 0.01816914603114128, + 0.02337699942290783, + 0.003452673554420471, + 0.003593248315155506, + 0.005969068966805935, + 0.007694877218455076, + 0.0034732185304164886, + 0.0029917494393885136, + 0.0034479405730962753, + 0.003950767684727907, + 0.007492606993764639, + 0.008319229818880558, + 0.016411855816841125, + 0.012280449271202087, + 0.02360900677740574, + 0.012087828479707241, + 0.021555660292506218, + 0.01620202511548996, + 0.0085498271510005, + 0.015202533453702927, + 0.013321309350430965, + 0.015519383363425732, + 0.030870120972394943, + 0.009408104233443737, + 0.017644189298152924, + 0.012591194361448288, + 0.015266625210642815, + 0.01577463373541832, + 0.024910829961299896, + 0.024114329367876053, + 0.038036003708839417, + 0.021475670859217644, + 0.021346386522054672, + 0.021532053127884865, + 0.004498677793890238, + 0.005218031816184521, + 0.010123909451067448, + 0.010914446786046028, + 0.011582212522625923, + 0.006425677333027124, + 0.00799776241183281, + 0.00685174111276865, + 0.003283186350017786, + 0.004447362385690212, + 0.004999429918825626, + 0.005625189282000065, + 0.008868912234902382, + 0.006789719685912132, + 0.005228402093052864, + 0.005018142517656088, + 0.0016231624176725745, + 0.001744658686220646, + 0.003296641632914543, + 0.0025920618791133165, + 0.0032730328384786844, + 0.003232687944546342, + 0.0018119693268090487, + 0.00238472456112504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.011085351929068565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890105664730072 + ], + "opposite_logprob_tokens": [ + -2.46401047706604 + ], + "logprob_reward": -0.0890105664730072, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890105664730072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10417555272579193, + 0.1576469987630844, + 0.06696505844593048, + 0.1468224823474884, + 0.13327381014823914, + 0.1430397927761078, + 0.053188711404800415, + 0.07063288986682892, + 0.08653350174427032, + 0.037721164524555206 + ], + "credit_map": [ + 0.010096613317728043, + 0.008909597061574459, + 0.020323654636740685, + 0.014454258605837822, + 0.014466647990047932, + 0.011082153767347336, + 0.013534708879888058, + 0.01130792684853077, + 0.010057296603918076, + 0.01238204725086689, + 0.027666913345456123, + 0.0185053963214159, + 0.027535170316696167, + 0.0210496187210083, + 0.022262806072831154, + 0.018187765032052994, + 0.0058146449737250805, + 0.004954630974680185, + 0.012330970726907253, + 0.011744329705834389, + 0.006926099769771099, + 0.005223515909165144, + 0.010763755068182945, + 0.00920711737126112, + 0.011776172555983067, + 0.008864020928740501, + 0.04277297109365463, + 0.01983598619699478, + 0.01250856090337038, + 0.013392559252679348, + 0.014366641640663147, + 0.023305587470531464, + 0.008635720238089561, + 0.008852346800267696, + 0.021753204986453056, + 0.01431821659207344, + 0.022508809342980385, + 0.013892408460378647, + 0.026369905099272728, + 0.016943220049142838, + 0.012104550376534462, + 0.021770795807242393, + 0.016414402052760124, + 0.01571609638631344, + 0.025318827480077744, + 0.01898001693189144, + 0.018092399463057518, + 0.014642726629972458, + 0.005859602242708206, + 0.004213876556605101, + 0.007323618978261948, + 0.007445335853844881, + 0.009036724455654621, + 0.006866964045912027, + 0.0062909675762057304, + 0.006151627749204636, + 0.004789639730006456, + 0.004508738871663809, + 0.019902415573596954, + 0.012078973464667797, + 0.006208871025592089, + 0.00845300406217575, + 0.005336167756468058, + 0.009355085901916027, + 0.00709045585244894, + 0.00790633074939251, + 0.009966820478439331, + 0.00844409316778183, + 0.023242540657520294, + 0.0069425106048583984, + 0.010222071781754494, + 0.012718685902655125, + 0.0031862843316048384, + 0.004124651197344065, + 0.004861743189394474, + 0.0034726327285170555, + 0.006092006806284189, + 0.005980536807328463, + 0.0034160215873271227, + 0.006587291602045298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.0890105664730072 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069071367383003 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.016069071367383003, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069071367383003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2348545342683792, + 0.11907599121332169, + 0.1832764595746994, + 0.029041865840554237, + 0.15111112594604492, + 0.056368637830019, + 0.08912897109985352, + 0.02624836564064026, + 0.05183037742972374, + 0.059063736349344254 + ], + "credit_map": [ + 0.017521491274237633, + 0.015199549496173859, + 0.03595627099275589, + 0.017735902220010757, + 0.034143634140491486, + 0.03672237694263458, + 0.026796801015734673, + 0.05077848583459854, + 0.008011779747903347, + 0.009845875203609467, + 0.016285434365272522, + 0.012706981971859932, + 0.023120878264307976, + 0.016695130616426468, + 0.01939019188284874, + 0.01301970612257719, + 0.024846021085977554, + 0.0106185432523489, + 0.03871754929423332, + 0.017277728766202927, + 0.0185360349714756, + 0.026361895725131035, + 0.0243967417627573, + 0.022521940991282463, + 0.0027737452182918787, + 0.002253703773021698, + 0.00894232839345932, + 0.003954032901674509, + 0.0034213776234537363, + 0.002697223098948598, + 0.0026115200016647577, + 0.002387933200225234, + 0.010535011067986488, + 0.011010437272489071, + 0.01798330806195736, + 0.015054984018206596, + 0.03318481519818306, + 0.015364940278232098, + 0.027244171127676964, + 0.020733444020152092, + 0.005004627164453268, + 0.006977545563131571, + 0.006378686986863613, + 0.007293679751455784, + 0.011219002306461334, + 0.004704850725829601, + 0.009996847249567509, + 0.004793390166014433, + 0.007979444228112698, + 0.007704117335379124, + 0.011777007021009922, + 0.011093977838754654, + 0.01816391944885254, + 0.01151529885828495, + 0.009464098140597343, + 0.011431103572249413, + 0.0019884887151420116, + 0.0022401176393032074, + 0.005424343049526215, + 0.0031756518874317408, + 0.004062840715050697, + 0.0024578573647886515, + 0.0033351234160363674, + 0.0035639421548694372, + 0.0037368913181126118, + 0.005049420986324549, + 0.005220203194767237, + 0.005259821657091379, + 0.010593029670417309, + 0.0076695941388607025, + 0.00787122081965208, + 0.006430191453546286, + 0.0038635041564702988, + 0.0034267662558704615, + 0.008078319951891899, + 0.005856125615537167, + 0.00804732646793127, + 0.010549712926149368, + 0.006302206311374903, + 0.012939768843352795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.016069071367383003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052477470599114895 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052477470599114895, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052477470599114895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24689029157161713, + 0.07894498854875565, + 0.28108304738998413, + 0.07625648379325867, + 0.04227737709879875, + 0.0763770341873169, + 0.041401930153369904, + 0.07205238193273544, + 0.05606141686439514, + 0.028655050322413445 + ], + "credit_map": [ + 0.019698727875947952, + 0.016828153282403946, + 0.0395398773252964, + 0.02092660218477249, + 0.03738291934132576, + 0.03272729739546776, + 0.03322668373584747, + 0.04656000807881355, + 0.007255160249769688, + 0.0071779340505599976, + 0.017272301018238068, + 0.008201523683965206, + 0.012690228410065174, + 0.011097376234829426, + 0.00828113965690136, + 0.006969321519136429, + 0.05730796232819557, + 0.02275189384818077, + 0.04408421367406845, + 0.027504976838827133, + 0.03013499826192856, + 0.04451510310173035, + 0.024789167568087578, + 0.029994724318385124, + 0.0067109521478414536, + 0.007968083955347538, + 0.015955254435539246, + 0.017383849248290062, + 0.009295336902141571, + 0.007980934344232082, + 0.0064454032108187675, + 0.00451666209846735, + 0.0029568688478320837, + 0.004141987767070532, + 0.006228133104741573, + 0.0040974728763103485, + 0.007264204788953066, + 0.004892910365015268, + 0.005673110485076904, + 0.007022685371339321, + 0.010049937292933464, + 0.008257192559540272, + 0.01597944274544716, + 0.009765900671482086, + 0.00789870135486126, + 0.007614892907440662, + 0.008549445308744907, + 0.008261507377028465, + 0.005576749797910452, + 0.005418635439127684, + 0.005737281404435635, + 0.004366283770650625, + 0.005357736721634865, + 0.004905474837869406, + 0.004248021170496941, + 0.005791742354631424, + 0.006083258893340826, + 0.003807768225669861, + 0.022741368040442467, + 0.011710400693118572, + 0.0070455558598041534, + 0.0080451974645257, + 0.004177003633230925, + 0.008441819809377193, + 0.0037164289969950914, + 0.004321555607020855, + 0.009867504239082336, + 0.01232588104903698, + 0.006345184985548258, + 0.006607046350836754, + 0.0055435760878026485, + 0.0073342290706932545, + 0.0020738926250487566, + 0.002137853065505624, + 0.0057112909853458405, + 0.003489090595394373, + 0.0030274344608187675, + 0.005011238623410463, + 0.001964067341759801, + 0.005240181460976601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0052477470599114895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the vibrant style of Disney Pixar animation, a static, slightly low-angled, full-body shot captures a captivating and determined Indian cobra. Its upper body is supported by protective coils on the ground, its signature thorn-patterned hood fully open, revealing bold black markings that contrast starkly against its smooth, golden-brown animated scales, which have a soft, plasticine-like sheen. Its large, expressive dark brown eyes, wide with genuine concern, gaze forward, guarding a small clutch of glossy, off-white eggs nestled in a depression on the mossy jungle floor. It sways with a deliberate, rhythmic movement, a classic defensive dance humorously undercut by an oversized, comedic sweatdrop that meticulously rolls down its temples, catching a glint of light as it follows a path along its smooth scales. The scene is set in a colorful jungle clearing, bathed in sunlight, with warm golden shafts piercing a dense canopy of oversized, waxy green leaves. The backdrop is a soft-focus tapestry of lush foliage and vibrant tropical flowers in shades of fuchsia and sunny yellow, creating a warm, lovely atmosphere filled with easy tension and parental devotion.", + "seed": 11621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_0/index_5415_seed_11621_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05192983066663146, + "vlm_logprob": -0.05192983066663146, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05192983066663146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1444341540336609 + ], + "opposite_logprob_tokens": [ + -2.0194342136383057 + ], + "logprob_reward": -0.1444341540336609, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1444341540336609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09798567742109299, + 0.07731233537197113, + 0.0330047570168972, + 0.285392165184021, + 0.2852086126804352, + 0.039474185556173325, + 0.06252741068601608, + 0.05550893023610115, + 0.025145672261714935, + 0.03844022750854492 + ], + "credit_map": [ + 0.029125653207302094, + 0.010809261351823807, + 0.005956667009741068, + 0.005965432617813349, + 0.014116926118731499, + 0.017017606645822525, + 0.009421770460903645, + 0.005572368390858173, + 0.01072526816278696, + 0.011066198348999023, + 0.011177318170666695, + 0.010517855174839497, + 0.008840023539960384, + 0.0072224861942231655, + 0.0087891211733222, + 0.008974069729447365, + 0.005525663495063782, + 0.004615679383277893, + 0.004233839921653271, + 0.0051911878399550915, + 0.0034343507140874863, + 0.0032701108139008284, + 0.0027647968381643295, + 0.00396913243457675, + 0.04010099545121193, + 0.03740311041474342, + 0.03605350852012634, + 0.05067920312285423, + 0.026297586038708687, + 0.0266579557210207, + 0.03218981251120567, + 0.03601004183292389, + 0.04786430671811104, + 0.034634772688150406, + 0.033870913088321686, + 0.045332882553339005, + 0.026886146515607834, + 0.03222844749689102, + 0.026479359716176987, + 0.0379117876291275, + 0.005759185645729303, + 0.004633692558854818, + 0.003853320609778166, + 0.006191566586494446, + 0.004214078187942505, + 0.0035743308253586292, + 0.0056727128103375435, + 0.005575302056968212, + 0.00864316988736391, + 0.010885144583880901, + 0.00594876566901803, + 0.006938556674867868, + 0.009425297379493713, + 0.005028199404478073, + 0.008686676621437073, + 0.006971604656428099, + 0.006328716408461332, + 0.006770409177988768, + 0.005383114796131849, + 0.005579991731792688, + 0.011803952977061272, + 0.00580547284334898, + 0.009537283331155777, + 0.004299988970160484, + 0.004829640034586191, + 0.0032629140187054873, + 0.0025193423498421907, + 0.0027090541552752256, + 0.003125535324215889, + 0.002582009183242917, + 0.003421936184167862, + 0.0026952431071549654, + 0.005378554109483957, + 0.0037600956857204437, + 0.00329226185567677, + 0.004165718797594309, + 0.007981013506650925, + 0.0036452077329158783, + 0.004760682117193937, + 0.005456696730107069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.1444341540336609 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790261276066303 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.009790261276066303, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790261276066303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4288122057914734, + 0.03791452571749687, + 0.04947841167449951, + 0.11779423803091049, + 0.11278036236763, + 0.13720175623893738, + 0.02104572393000126, + 0.04150918126106262, + 0.03154492750763893, + 0.02191869542002678 + ], + "credit_map": [ + 0.1697404533624649, + 0.03596653789281845, + 0.014990600757300854, + 0.012357977218925953, + 0.043134208768606186, + 0.1146286129951477, + 0.018783101812005043, + 0.0192106980830431, + 0.005601806566119194, + 0.004802174400538206, + 0.0038128818850964308, + 0.0044380249455571175, + 0.006414704490453005, + 0.004427220206707716, + 0.004197392612695694, + 0.004220318049192429, + 0.007584843784570694, + 0.005811487324535847, + 0.008182903751730919, + 0.008057215251028538, + 0.004898985382169485, + 0.004476957023143768, + 0.005016917362809181, + 0.005449104588478804, + 0.01673063263297081, + 0.01629238948225975, + 0.01517490204423666, + 0.02065826579928398, + 0.01071914378553629, + 0.01139006856828928, + 0.012451129034161568, + 0.01437770202755928, + 0.016649384051561356, + 0.01393407117575407, + 0.01306123286485672, + 0.019688351079821587, + 0.010453402064740658, + 0.01232934184372425, + 0.012290123850107193, + 0.01437446754425764, + 0.021092157810926437, + 0.021647006273269653, + 0.015920700505375862, + 0.020429346710443497, + 0.012238130904734135, + 0.013477507047355175, + 0.014014804735779762, + 0.018382109701633453, + 0.002798915607854724, + 0.0030656210146844387, + 0.004045761190354824, + 0.00290408986620605, + 0.002045685425400734, + 0.0018596486188471317, + 0.0021626807283610106, + 0.0021633224096149206, + 0.005913237109780312, + 0.009091468527913094, + 0.003757619531825185, + 0.003838501637801528, + 0.005737382918596268, + 0.0036089331842958927, + 0.005462241359055042, + 0.004099796060472727, + 0.003973094746470451, + 0.004805519711226225, + 0.0032761970069259405, + 0.003504399908706546, + 0.0040945312939584255, + 0.0029673550743609667, + 0.004880350083112717, + 0.004043480381369591, + 0.0025383203756064177, + 0.0030132883694022894, + 0.0021213404834270477, + 0.0020398579072207212, + 0.003963741008192301, + 0.0020212666131556034, + 0.0035192915238440037, + 0.002701588673517108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.009790261276066303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895165681838989 + ], + "opposite_logprob_tokens": [ + -2.578951597213745 + ], + "logprob_reward": -0.07895165681838989, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895165681838989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2595987915992737, + 0.08331320434808731, + 0.15150880813598633, + 0.07372072339057922, + 0.2632845938205719, + 0.05605189874768257, + 0.04374971240758896, + 0.03230433538556099, + 0.018267035484313965, + 0.01820089854300022 + ], + "credit_map": [ + 0.09317344427108765, + 0.02218065783381462, + 0.015020476654171944, + 0.011807066388428211, + 0.028639109805226326, + 0.0636911541223526, + 0.01193130761384964, + 0.013155586086213589, + 0.011987597681581974, + 0.011124520562589169, + 0.010786780156195164, + 0.01126067340373993, + 0.010061664506793022, + 0.011480839923024178, + 0.00784627627581358, + 0.008764849044382572, + 0.02143111266195774, + 0.020270010456442833, + 0.02118591032922268, + 0.023951923474669456, + 0.014553462155163288, + 0.01721877045929432, + 0.015869395807385445, + 0.017028218135237694, + 0.00937665905803442, + 0.009679464623332024, + 0.008489353582262993, + 0.009340101853013039, + 0.009041061624884605, + 0.007056108210235834, + 0.011442238464951515, + 0.009295739233493805, + 0.04122485965490341, + 0.03484119102358818, + 0.03087940625846386, + 0.04120279476046562, + 0.025536589324474335, + 0.029299043118953705, + 0.025723861530423164, + 0.03457685187458992, + 0.007617019582539797, + 0.006079535000026226, + 0.005485870409756899, + 0.007969125173985958, + 0.006863285321742296, + 0.004589881282299757, + 0.009336364455521107, + 0.008110817521810532, + 0.005653425585478544, + 0.007317063864320517, + 0.0045931353233754635, + 0.005133811850100756, + 0.0052888537757098675, + 0.004055103287100792, + 0.006531014572829008, + 0.005177306942641735, + 0.004080287180840969, + 0.007795894984155893, + 0.0036246078088879585, + 0.003107052529230714, + 0.0034694683272391558, + 0.003117702202871442, + 0.0039114393293857574, + 0.0031978825572878122, + 0.002233846578747034, + 0.00411846861243248, + 0.0021457350812852383, + 0.00186220183968544, + 0.002150582615286112, + 0.0016450965777039528, + 0.0023048287257552147, + 0.0018062763847410679, + 0.0021347971633076668, + 0.0030006777960807085, + 0.0017191976075991988, + 0.002052072901278734, + 0.002676864853128791, + 0.0017282165354117751, + 0.0024249658454209566, + 0.0024641056079417467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07895165681838989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328586019575596 + ], + "opposite_logprob_tokens": [ + -3.7732858657836914 + ], + "logprob_reward": -0.02328586019575596, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328586019575596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42307424545288086, + 0.06365327537059784, + 0.02575894631445408, + 0.16717004776000977, + 0.08523441851139069, + 0.04268104210495949, + 0.031608156859874725, + 0.03655821084976196, + 0.033695288002491, + 0.09056639671325684 + ], + "credit_map": [ + 0.1678355187177658, + 0.034667715430259705, + 0.01549595221877098, + 0.01313862670212984, + 0.04266085848212242, + 0.11432330310344696, + 0.017158815637230873, + 0.01779341697692871, + 0.011154897511005402, + 0.006477578077465296, + 0.007352156564593315, + 0.008038868196308613, + 0.007694676518440247, + 0.005682096816599369, + 0.010282691568136215, + 0.006970317102968693, + 0.00275311223231256, + 0.0038320429157465696, + 0.0035271805245429277, + 0.003422464244067669, + 0.003109431592747569, + 0.003363937372341752, + 0.0029569517355412245, + 0.0027938252314925194, + 0.022785993292927742, + 0.025622544810175896, + 0.021187933161854744, + 0.026502402499318123, + 0.016331586986780167, + 0.01641642116010189, + 0.017888477072119713, + 0.020434707403182983, + 0.013788510113954544, + 0.01045696996152401, + 0.011177822016179562, + 0.013660618104040623, + 0.007821697741746902, + 0.009273192845284939, + 0.007914613001048565, + 0.011140989139676094, + 0.006064228713512421, + 0.008645961061120033, + 0.004885803442448378, + 0.005387002136558294, + 0.003830677829682827, + 0.003852901980280876, + 0.004701210185885429, + 0.005313258618116379, + 0.004695902578532696, + 0.0057007973082363605, + 0.0040137204341590405, + 0.003318225732073188, + 0.003904961282387376, + 0.0025648369919508696, + 0.0041382815688848495, + 0.0032714330591261387, + 0.005028334446251392, + 0.004017247352749109, + 0.003083662362769246, + 0.002927987603470683, + 0.005532364826649427, + 0.004115993622690439, + 0.006512260064482689, + 0.005340360105037689, + 0.004584538284689188, + 0.004020089283585548, + 0.0027956245467066765, + 0.00319105782546103, + 0.005125241819769144, + 0.0032110661268234253, + 0.005888305604457855, + 0.004879364278167486, + 0.011061664670705795, + 0.012093358673155308, + 0.007552393712103367, + 0.007672288455069065, + 0.021407799795269966, + 0.007836652919650078, + 0.013848990201950073, + 0.009093253873288631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.02328586019575596 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003187221009284258 + ], + "opposite_logprob_tokens": [ + -5.75318717956543 + ], + "logprob_reward": -0.003187221009284258, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003187221009284258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4890146255493164, + 0.07663168013095856, + 0.06310439854860306, + 0.10496938973665237, + 0.0907934159040451, + 0.027627738192677498, + 0.029155220836400986, + 0.052612241357564926, + 0.034702010452747345, + 0.03138931468129158 + ], + "credit_map": [ + 0.1971537321805954, + 0.04062927886843681, + 0.01638384722173214, + 0.013610823079943657, + 0.0483405739068985, + 0.1325891613960266, + 0.01985318586230278, + 0.02045399881899357, + 0.011844723485410213, + 0.009404892101883888, + 0.009859581477940083, + 0.010202085599303246, + 0.007798823993653059, + 0.00692706648260355, + 0.011456182226538658, + 0.009138324297964573, + 0.009887405671179295, + 0.010571720078587532, + 0.006246211472898722, + 0.00997078325599432, + 0.0068185459822416306, + 0.006296847946941853, + 0.00504478020593524, + 0.008268101140856743, + 0.016312578693032265, + 0.013989923521876335, + 0.012757213786244392, + 0.018111666664481163, + 0.009559310041368008, + 0.010258634574711323, + 0.010693158023059368, + 0.01328690443187952, + 0.014293725602328777, + 0.010868320241570473, + 0.011782776564359665, + 0.014276239089667797, + 0.0084679601714015, + 0.009819555096328259, + 0.009765957482159138, + 0.011518887244164944, + 0.003819158999249339, + 0.005146760027855635, + 0.003465978428721428, + 0.0036165539640933275, + 0.002359428210183978, + 0.0028601274825632572, + 0.003296967362985015, + 0.003062763251364231, + 0.003686714917421341, + 0.004178048577159643, + 0.003186484333127737, + 0.004167635925114155, + 0.005558108910918236, + 0.00242237257771194, + 0.0033338540233671665, + 0.0026220025029033422, + 0.006580558605492115, + 0.010909926146268845, + 0.0045097582042217255, + 0.005819343030452728, + 0.007478015497326851, + 0.0049308002926409245, + 0.00639757513999939, + 0.005986266769468784, + 0.004056454170495272, + 0.004953670781105757, + 0.00426524830982089, + 0.004151572939008474, + 0.005188371054828167, + 0.003400015877559781, + 0.004188103135675192, + 0.004498576745390892, + 0.004264669958502054, + 0.006587133277207613, + 0.003887234954163432, + 0.003467214060947299, + 0.003870348446071148, + 0.002483958611264825, + 0.003395102685317397, + 0.0034336524549871683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.003187221009284258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the vibrant style of Disney Pixar animation, a static, slightly low-angled, full-body shot captures a captivating and determined Indian cobra. Its upper body is supported by protective coils on the ground, its signature thorn-patterned hood fully open, revealing bold black markings that contrast starkly against its smooth, golden-brown animated scales, which have a soft, plasticine-like sheen. Its large, expressive dark brown eyes, wide with genuine concern, gaze forward, guarding a small clutch of glossy, off-white eggs nestled in a depression on the mossy jungle floor. It sways with a deliberate, rhythmic movement, a classic defensive dance humorously undercut by an oversized, comedic sweatdrop that meticulously rolls down its temples, catching a glint of light as it follows a path along its smooth scales. The scene is set in a colorful jungle clearing, bathed in sunlight, with warm golden shafts piercing a dense canopy of oversized, waxy green leaves. The backdrop is a soft-focus tapestry of lush foliage and vibrant tropical flowers in shades of fuchsia and sunny yellow, creating a warm, lovely atmosphere filled with easy tension and parental devotion.", + "seed": 11622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_0/index_5415_seed_11622_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04366519004106521, + "vlm_logprob": -0.04366519004106521, + "vlm_yes_no_margin": 3.725000023841858, + "vlm_token_credit": -0.04366519004106521, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1467534601688385 + ], + "opposite_logprob_tokens": [ + -2.0217535495758057 + ], + "logprob_reward": -0.1467534601688385, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1467534601688385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07096703350543976, + 0.057063594460487366, + 0.020078718662261963, + 0.09383565932512283, + 0.16889071464538574, + 0.17370504140853882, + 0.044124241918325424, + 0.22312124073505402, + 0.04329893738031387, + 0.10491479188203812 + ], + "credit_map": [ + 0.008206529542803764, + 0.005156353581696749, + 0.00885282177478075, + 0.008870103396475315, + 0.007185627240687609, + 0.008604717440903187, + 0.014829461462795734, + 0.009261412546038628, + 0.010606808587908745, + 0.0046602580696344376, + 0.007527154870331287, + 0.005800266284495592, + 0.007816853933036327, + 0.005059913266450167, + 0.007273645140230656, + 0.008318699896335602, + 0.003984758164733648, + 0.001995641505345702, + 0.0027833771891891956, + 0.0019181418465450406, + 0.0025482834316790104, + 0.0017229473451152444, + 0.0029729846864938736, + 0.0021525851916521788, + 0.021265890449285507, + 0.010079012252390385, + 0.008708175271749496, + 0.009249783121049404, + 0.011878058314323425, + 0.010036511346697807, + 0.007589406333863735, + 0.015028821304440498, + 0.02600383758544922, + 0.02151809260249138, + 0.01885146088898182, + 0.018685000017285347, + 0.028705066069960594, + 0.01775471866130829, + 0.012747866101562977, + 0.0246246587485075, + 0.046297963708639145, + 0.00874944869428873, + 0.012612931430339813, + 0.006819017231464386, + 0.015708904713392258, + 0.012821263633668423, + 0.01106741651892662, + 0.059628091752529144, + 0.011752437800168991, + 0.002720241667702794, + 0.003278560470789671, + 0.0018533187685534358, + 0.0044823396019637585, + 0.0035466481931507587, + 0.0035296985879540443, + 0.012960996478796005, + 0.0441056564450264, + 0.01479709055274725, + 0.015394947491586208, + 0.01672957092523575, + 0.04171570762991905, + 0.01919984631240368, + 0.03708608075976372, + 0.034092336893081665, + 0.005288939923048019, + 0.005798955913633108, + 0.004182837437838316, + 0.004078293219208717, + 0.008649593219161034, + 0.005329899024218321, + 0.006400917656719685, + 0.003569501917809248, + 0.011229344643652439, + 0.010813307948410511, + 0.0048136040568351746, + 0.005626017693430185, + 0.013278566300868988, + 0.02660716138780117, + 0.016910338774323463, + 0.01563645340502262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.1467534601688385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254948042333126 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.01254948042333126, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254948042333126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27505165338516235, + 0.07423971593379974, + 0.020121848210692406, + 0.05780676379799843, + 0.044648922979831696, + 0.18433089554309845, + 0.017925577238202095, + 0.07846619933843613, + 0.04388001561164856, + 0.20352837443351746 + ], + "credit_map": [ + 0.0357329323887825, + 0.017826281487941742, + 0.032489702105522156, + 0.03824048116803169, + 0.021989019587635994, + 0.03133029118180275, + 0.03093801997601986, + 0.06650490313768387, + 0.007223837077617645, + 0.005670221522450447, + 0.008919905871152878, + 0.00702710309997201, + 0.007318469695746899, + 0.010289721190929413, + 0.012004457414150238, + 0.01578599587082863, + 0.0023216118570417166, + 0.002542322501540184, + 0.0025366570334881544, + 0.002717643743380904, + 0.003561241552233696, + 0.0018828556640073657, + 0.0024594746064394712, + 0.0021000404376536608, + 0.008574428036808968, + 0.0071000587195158005, + 0.006655599921941757, + 0.007013368420302868, + 0.008604067377746105, + 0.006031168159097433, + 0.008359548635780811, + 0.0054685198701918125, + 0.007431130390614271, + 0.004945070017129183, + 0.004384027794003487, + 0.004992225207388401, + 0.00657588941976428, + 0.004071562550961971, + 0.005420382600277662, + 0.006828632205724716, + 0.05634456127882004, + 0.009140890091657639, + 0.010741962119936943, + 0.0056358929723501205, + 0.016074735671281815, + 0.014056102372705936, + 0.009289678186178207, + 0.06304707378149033, + 0.0021857640240341425, + 0.0019970270805060863, + 0.003045434597879648, + 0.0017411389853805304, + 0.002781007206067443, + 0.0016524867387488484, + 0.002188177779316902, + 0.0023345404770225286, + 0.020302316173911095, + 0.004729944281280041, + 0.005243072751909494, + 0.0042924764566123486, + 0.012631994672119617, + 0.006390994880348444, + 0.00852873269468546, + 0.016346663236618042, + 0.006067908834666014, + 0.0069423858076334, + 0.004437212832272053, + 0.0044377390295267105, + 0.006690152920782566, + 0.004759529605507851, + 0.0058139897882938385, + 0.004731092136353254, + 0.044849295169115067, + 0.018553229048848152, + 0.007438276428729296, + 0.009055905975401402, + 0.02170051448047161, + 0.042045578360557556, + 0.028297292068600655, + 0.03158828616142273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.01254948042333126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794966802001 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794966802001, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794966802001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13341562449932098, + 0.04825160279870033, + 0.05463087186217308, + 0.1052093356847763, + 0.06467660516500473, + 0.13173484802246094, + 0.07830923795700073, + 0.20094497501850128, + 0.09391582012176514, + 0.08891109377145767 + ], + "credit_map": [ + 0.01803835853934288, + 0.012112392112612724, + 0.017235664650797844, + 0.02130860649049282, + 0.019739245995879173, + 0.013679596595466137, + 0.011412441730499268, + 0.01988932490348816, + 0.0070909010246396065, + 0.005172157194465399, + 0.008346069604158401, + 0.004874423146247864, + 0.007922516204416752, + 0.005433704238384962, + 0.00509957130998373, + 0.004312258679419756, + 0.007267969660460949, + 0.005884503945708275, + 0.0048025064170360565, + 0.006503112614154816, + 0.008006910793483257, + 0.0046174428425729275, + 0.008032291196286678, + 0.00951613299548626, + 0.014115500263869762, + 0.01301546674221754, + 0.011582519859075546, + 0.012179180048406124, + 0.020070845261216164, + 0.011910730972886086, + 0.01211840845644474, + 0.01021667756140232, + 0.00895520206540823, + 0.006419079843908548, + 0.0064351544715464115, + 0.008143389597535133, + 0.012392954900860786, + 0.005230552516877651, + 0.006220303941518068, + 0.010879965499043465, + 0.03671633079648018, + 0.006761691067367792, + 0.007642263080924749, + 0.004619356244802475, + 0.013323510065674782, + 0.009258218109607697, + 0.008366488851606846, + 0.045046981424093246, + 0.013970048166811466, + 0.007442723959684372, + 0.007695258595049381, + 0.00618920149281621, + 0.011282582767307758, + 0.006175078451633453, + 0.011263963766396046, + 0.014290383085608482, + 0.04134600982069969, + 0.013253123499453068, + 0.017279915511608124, + 0.017241856083273888, + 0.03486420586705208, + 0.0160371083766222, + 0.033401526510715485, + 0.027521217241883278, + 0.013463785871863365, + 0.009139124304056168, + 0.012132133357226849, + 0.0149563979357481, + 0.01384421344846487, + 0.011761156842112541, + 0.011265245266258717, + 0.007353768218308687, + 0.011053942143917084, + 0.009445509873330593, + 0.004513957072049379, + 0.005693742539733648, + 0.014505310915410519, + 0.018587283790111542, + 0.012672209180891514, + 0.012439130805432796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.029794966802001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285627365112305 + ], + "opposite_logprob_tokens": [ + -3.7732856273651123 + ], + "logprob_reward": -0.023285627365112305, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285627365112305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08339522778987885, + 0.061384592205286026, + 0.022962886840105057, + 0.0869537815451622, + 0.13830126821994781, + 0.26785510778427124, + 0.01853293366730213, + 0.08594810962677002, + 0.04053341597318649, + 0.1941327005624771 + ], + "credit_map": [ + 0.01120869629085064, + 0.007211209274828434, + 0.009902187623083591, + 0.008700798265635967, + 0.008901406079530716, + 0.009882102720439434, + 0.008739548735320568, + 0.018849268555641174, + 0.009842758998274803, + 0.007153297308832407, + 0.006374891381710768, + 0.006032758392393589, + 0.012507520616054535, + 0.005239284131675959, + 0.006141898222267628, + 0.008092178963124752, + 0.0027507238555699587, + 0.002412916859611869, + 0.004116644151508808, + 0.0025357375852763653, + 0.0027882838621735573, + 0.0025442461483180523, + 0.0020637831185013056, + 0.003750548930838704, + 0.01680358499288559, + 0.011904053390026093, + 0.008365784771740437, + 0.008677964098751545, + 0.014849627390503883, + 0.006997197400778532, + 0.008694919757544994, + 0.010660638101398945, + 0.01677229255437851, + 0.01399966236203909, + 0.017065750434994698, + 0.014735600911080837, + 0.02093527652323246, + 0.014297875575721264, + 0.01097814179956913, + 0.029516668990254402, + 0.07477103173732758, + 0.0130692757666111, + 0.01729075238108635, + 0.010001640766859055, + 0.02637387253344059, + 0.019291676580905914, + 0.015499269589781761, + 0.0915575698018074, + 0.0021444472949951887, + 0.0018430361524224281, + 0.002408830216154456, + 0.001620166702196002, + 0.003588038496673107, + 0.001840541954152286, + 0.0024784088600426912, + 0.002609463408589363, + 0.014334098435938358, + 0.006096349097788334, + 0.007183842360973358, + 0.007102382369339466, + 0.01782270520925522, + 0.008422589860856533, + 0.014832626096904278, + 0.010153505019843578, + 0.0054510002955794334, + 0.0038571988698095083, + 0.0028746624011546373, + 0.003955083899199963, + 0.009086577221751213, + 0.005038734525442123, + 0.006453787442296743, + 0.003816368989646435, + 0.03686093911528587, + 0.006314042955636978, + 0.0046657780185341835, + 0.009350141510367393, + 0.01742188073694706, + 0.00704527273774147, + 0.08400774747133255, + 0.02846688963472843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.023285627365112305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Indian cobra with a thorn-patterned hood fully open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cobra's body color smooth and golden-brown with a soft, plasticine-like sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cobra swaying with a deliberate, rhythmic movement while guarding a clutch of glossy, off-white eggs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an oversized, comedic sweatdrop rolling down the cobra's temples along its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a colorful jungle clearing with warm golden sunlight shafts and a background of fuchsia and sunny yellow flowers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059424154460430145 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.0059424154460430145, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059424154460430145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11458712816238403, + 0.07013200223445892, + 0.01631312444806099, + 0.04265372455120087, + 0.0420849435031414, + 0.1360515058040619, + 0.02928828075528145, + 0.043541666120290756, + 0.04439152032136917, + 0.4609561264514923 + ], + "credit_map": [ + 0.021143775433301926, + 0.006147186737507582, + 0.012972707860171795, + 0.012627555988729, + 0.008855478838086128, + 0.01678319089114666, + 0.011393661610782146, + 0.024663565680384636, + 0.008759691379964352, + 0.006150806322693825, + 0.009168323129415512, + 0.005918435286730528, + 0.010296626947820187, + 0.0068329400382936, + 0.007294211536645889, + 0.01571096107363701, + 0.0019024451030418277, + 0.0013122882228344679, + 0.0034193790052086115, + 0.0013248296454548836, + 0.0016569712897762656, + 0.0016828462248668075, + 0.001470001065172255, + 0.0035443631932139397, + 0.006284757982939482, + 0.003179129445925355, + 0.0034010959789156914, + 0.005245249252766371, + 0.006689900066703558, + 0.005846748128533363, + 0.005486152600497007, + 0.006520692724734545, + 0.007880497723817825, + 0.004090854432433844, + 0.004061862826347351, + 0.004387354012578726, + 0.00654623145237565, + 0.005421679001301527, + 0.0032326625660061836, + 0.006463801953941584, + 0.03992607444524765, + 0.0070303878746926785, + 0.007967772893607616, + 0.0046911886893212795, + 0.012312978506088257, + 0.01090059895068407, + 0.007618406321853399, + 0.045604098588228226, + 0.003396540880203247, + 0.003174610435962677, + 0.0028605018742382526, + 0.0023926880676299334, + 0.005086418706923723, + 0.0037285618018358946, + 0.005779956933110952, + 0.0028690032195299864, + 0.009757524356245995, + 0.002642446430400014, + 0.0029358069878071547, + 0.0032036856282502413, + 0.007406699471175671, + 0.003035572124645114, + 0.006988251581788063, + 0.0075716786086559296, + 0.006821999792009592, + 0.004842783790081739, + 0.003154152538627386, + 0.003260972909629345, + 0.008560195565223694, + 0.006608155556023121, + 0.005058678332716227, + 0.006084583234041929, + 0.10876213014125824, + 0.016876062378287315, + 0.015138166956603527, + 0.02369876578450203, + 0.040607769042253494, + 0.016363728791475296, + 0.19335515797138214, + 0.04615432769060135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0059424154460430145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_tensor([148, 148])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_tensor([148, 148])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..773c2d0c2cff647cb5915ed5fd7510c5006be009 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_tensor([148, 148])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low, wide-angle shot, shot from above, slowly pulls back to reveal a moment of monumental and devastating collapse. At the center of the frame is an ancient, massive stone wall perched on the edge of a precipitous cliff. Composed of massive, weathered granite blocks of varying sizes, its surface covered in dark green moss and withered vines, its cracks brimming with weathered mortar. Under a dark, overcast sky, the cool, evenly diffused light accentuates the wall's weathered texture. Suddenly, with a deep, teeth-grinding crack, a crucial upper support stone, under immense pressure, thrusts outward, setting off an unstoppable chain reaction. Tons of boulders seemed to have broken free from the constraints of thousands of years and were violently peeled off from the wall. They rolled heavily in the air and collided violently with each other, emitting sporadic sparks, and then carried a cloud of yellow-brown smoke and dust that swallowed the sky and the earth, rolling down the cliff with the force of thunder. The smoke and dust rose up in an instant, blurring the outline of the valley. The whole scene was full of awe-inspiring, primitive destructive power.", + "seed": 22588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_148_seed_22588_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12164246886968613, + "vlm_logprob": -0.12164246886968613, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.12164246886968613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19302552938461304 + ], + "opposite_logprob_tokens": [ + -1.8180255889892578 + ], + "logprob_reward": -0.19302552938461304, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19302552938461304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3240399658679962, + 0.10002646595239639, + 0.07344312220811844, + 0.1585204005241394, + 0.053051840513944626, + 0.13404971361160278, + 0.060907285660505295, + 0.029081977903842926, + 0.021100753918290138, + 0.045778531581163406 + ], + "credit_map": [ + 0.023135624825954437, + 0.03032008372247219, + 0.08204802125692368, + 0.026554835960268974, + 0.03454677760601044, + 0.037237055599689484, + 0.03144288808107376, + 0.05875466763973236, + 0.008223515935242176, + 0.02467445097863674, + 0.011554375290870667, + 0.010077223181724548, + 0.011379293166100979, + 0.01128541398793459, + 0.008481954224407673, + 0.014350234530866146, + 0.007307704072445631, + 0.006524394266307354, + 0.014496959745883942, + 0.00860168132930994, + 0.006562578026205301, + 0.007562211249023676, + 0.01323220320045948, + 0.00915539264678955, + 0.014611097984015942, + 0.017323622480034828, + 0.022927386686205864, + 0.02051425911486149, + 0.016173986718058586, + 0.014529630541801453, + 0.030972180888056755, + 0.021468238905072212, + 0.0055685280822217464, + 0.005675779189914465, + 0.009033734910190105, + 0.005653653759509325, + 0.0053711854852736, + 0.004763915203511715, + 0.009452604688704014, + 0.007532438263297081, + 0.014882572926580906, + 0.01285008154809475, + 0.014425300061702728, + 0.014044409617781639, + 0.02358905039727688, + 0.013552623800933361, + 0.0223966334015131, + 0.018309032544493675, + 0.004272731952369213, + 0.005724737420678139, + 0.007793546654284, + 0.006732235196977854, + 0.0038802395574748516, + 0.0063151526264846325, + 0.020233210176229477, + 0.005955432076007128, + 0.002211331855505705, + 0.0028679901733994484, + 0.0036120612639933825, + 0.0044086831621825695, + 0.0037495307624340057, + 0.003412460209801793, + 0.006168043706566095, + 0.002651875838637352, + 0.001976764528080821, + 0.002457135822623968, + 0.00275903707370162, + 0.0026616468094289303, + 0.0017467403085902333, + 0.0034489608369767666, + 0.002940750913694501, + 0.003109718207269907, + 0.004154687747359276, + 0.004818693269044161, + 0.009404855780303478, + 0.0052069989033043385, + 0.008334985002875328, + 0.0038166320882737637, + 0.005616325419396162, + 0.004425352439284325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.19302552938461304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313489705324173 + ], + "opposite_logprob_tokens": [ + -1.3134896755218506 + ], + "logprob_reward": -0.313489705324173, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.313489705324173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2089250683784485, + 0.21427123248577118, + 0.12478610128164291, + 0.06437495350837708, + 0.05615341663360596, + 0.09566233307123184, + 0.1405823975801468, + 0.04026122763752937, + 0.024205364286899567, + 0.03077785111963749 + ], + "credit_map": [ + 0.01343735121190548, + 0.019459007307887077, + 0.053188927471637726, + 0.015542847104370594, + 0.02618175745010376, + 0.02516942471265793, + 0.018958641216158867, + 0.036987096071243286, + 0.013408325612545013, + 0.043627284467220306, + 0.035399094223976135, + 0.015241077169775963, + 0.03160609304904938, + 0.019762378185987473, + 0.019754307344555855, + 0.03547268360853195, + 0.01370549201965332, + 0.014167646877467632, + 0.01868576928973198, + 0.015227959491312504, + 0.014795582741498947, + 0.013895351439714432, + 0.016536789014935493, + 0.01777150109410286, + 0.00632822560146451, + 0.009269287809729576, + 0.009649686515331268, + 0.008735629729926586, + 0.00583935109898448, + 0.00531590124592185, + 0.011413292959332466, + 0.007823582738637924, + 0.008060128428041935, + 0.004664999432861805, + 0.011737732216715813, + 0.006201739422976971, + 0.003628346836194396, + 0.006108648609369993, + 0.007484069559723139, + 0.008267750963568687, + 0.009726151823997498, + 0.011162618175148964, + 0.011771217919886112, + 0.009852021001279354, + 0.012745456770062447, + 0.008520337752997875, + 0.0188409686088562, + 0.013043561019003391, + 0.010394205339252949, + 0.01593261957168579, + 0.017964355647563934, + 0.01658622734248638, + 0.008685193955898285, + 0.01205336395651102, + 0.045033641159534454, + 0.013932776637375355, + 0.0037130555137991905, + 0.005941020790487528, + 0.0048782783560454845, + 0.007526400499045849, + 0.0028557397890836, + 0.00439267884939909, + 0.006144147366285324, + 0.004809906706213951, + 0.0025059243198484182, + 0.0028911279514431953, + 0.003201009938493371, + 0.003074483945965767, + 0.0019458496244624257, + 0.003428050549700856, + 0.0038304347544908524, + 0.0033284826204180717, + 0.002670314395800233, + 0.0037981204222887754, + 0.004688096232712269, + 0.004217991139739752, + 0.004854621831327677, + 0.004515682812780142, + 0.003270143875852227, + 0.0027628825046122074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.313489705324173 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085351929068565 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085351929068565, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085351929068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.285288542509079, + 0.12511689960956573, + 0.1281551569700241, + 0.08094020187854767, + 0.09525708854198456, + 0.07826929539442062, + 0.0697493776679039, + 0.053409550338983536, + 0.05276946350932121, + 0.031044388189911842 + ], + "credit_map": [ + 0.02763097919523716, + 0.03592947497963905, + 0.04200100898742676, + 0.025099271908402443, + 0.034932415932416916, + 0.033895496279001236, + 0.02585742063820362, + 0.05994250252842903, + 0.006807755678892136, + 0.027656521648168564, + 0.013724529184401035, + 0.01652272790670395, + 0.020717665553092957, + 0.009419644251465797, + 0.013680098578333855, + 0.01658797450363636, + 0.012931729666888714, + 0.02134919911623001, + 0.017078831791877747, + 0.018297208473086357, + 0.01692216843366623, + 0.013097389601171017, + 0.01102985069155693, + 0.0174487866461277, + 0.004873926751315594, + 0.007224162109196186, + 0.012141445651650429, + 0.01426298264414072, + 0.007560002151876688, + 0.008839784190058708, + 0.014104539528489113, + 0.011933363042771816, + 0.012849465943872929, + 0.010314807295799255, + 0.015377871692180634, + 0.00920720212161541, + 0.009531710296869278, + 0.011045199818909168, + 0.014361814595758915, + 0.012569031678140163, + 0.0074626081623137, + 0.00899870228022337, + 0.009532492607831955, + 0.010142189450562, + 0.013887527398765087, + 0.00853357370942831, + 0.00791181530803442, + 0.011800390668213367, + 0.004897401202470064, + 0.008158898912370205, + 0.009674136526882648, + 0.006560333538800478, + 0.006602672394365072, + 0.0058304863050580025, + 0.022202398627996445, + 0.005823059473186731, + 0.0045798164792358875, + 0.005586491897702217, + 0.007923892699182034, + 0.006475744768977165, + 0.010119914077222347, + 0.004608750343322754, + 0.008535705506801605, + 0.005579239688813686, + 0.005594255868345499, + 0.006689326837658882, + 0.007036789320409298, + 0.00913739949464798, + 0.004687138367444277, + 0.006879633758217096, + 0.006964401807636023, + 0.005780518986284733, + 0.0032121192198246717, + 0.003379153786227107, + 0.003753450233489275, + 0.005125894211232662, + 0.002332344651222229, + 0.003693418810144067, + 0.005092557519674301, + 0.00445545045658946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.011085351929068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583825185894966 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583825185894966, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583825185894966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.178763747215271, + 0.15485142171382904, + 0.294592946767807, + 0.12941524386405945, + 0.04105338826775551, + 0.06667643785476685, + 0.02972646802663803, + 0.04713018983602524, + 0.02176784537732601, + 0.03602232784032822 + ], + "credit_map": [ + 0.013107714243233204, + 0.016870032995939255, + 0.042050935328006744, + 0.01184347365051508, + 0.023135622963309288, + 0.02511328086256981, + 0.017740892246365547, + 0.028901780024170876, + 0.008582260459661484, + 0.03845847770571709, + 0.015128254890441895, + 0.016260668635368347, + 0.022452104836702347, + 0.014994080178439617, + 0.015583845786750317, + 0.023391734808683395, + 0.02320687100291252, + 0.04407934099435806, + 0.04024665057659149, + 0.04624635726213455, + 0.0378066748380661, + 0.027441982179880142, + 0.034891463816165924, + 0.040673620998859406, + 0.012172945775091648, + 0.0147892190143466, + 0.02318003587424755, + 0.017649738118052483, + 0.013606213964521885, + 0.010923882946372032, + 0.020103340968489647, + 0.01698986254632473, + 0.006094734184443951, + 0.004869252908974886, + 0.006437800358980894, + 0.0052360002882778645, + 0.0036332732997834682, + 0.0043352097272872925, + 0.005374466069042683, + 0.005072650033980608, + 0.00801892764866352, + 0.007119718939065933, + 0.012090535834431648, + 0.006968051660805941, + 0.004179494455456734, + 0.008245976641774178, + 0.010217666625976562, + 0.009836063720285892, + 0.003886006074026227, + 0.0038597281090915203, + 0.0035918783396482468, + 0.0044365497305989265, + 0.0023186742328107357, + 0.0031786479521542788, + 0.004684024956077337, + 0.0037709574680775404, + 0.004423373844474554, + 0.006960987579077482, + 0.005173759534955025, + 0.005994860082864761, + 0.005420397967100143, + 0.00382866058498621, + 0.010636579245328903, + 0.004691570997238159, + 0.002453812165185809, + 0.002759629162028432, + 0.0029976004734635353, + 0.0033153172116726637, + 0.0016758244019001722, + 0.002618643222376704, + 0.0032218676060438156, + 0.0027251499705016613, + 0.003237638156861067, + 0.00449525099247694, + 0.00542102288454771, + 0.0054367524571716785, + 0.004960459657013416, + 0.004214130807667971, + 0.00402138102799654, + 0.004235692322254181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.020583825185894966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002793252468109 + ], + "opposite_logprob_tokens": [ + -2.6950278282165527 + ], + "logprob_reward": -0.07002793252468109, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002793252468109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21244174242019653, + 0.19276385009288788, + 0.1465214341878891, + 0.06267715990543365, + 0.0641680508852005, + 0.09498801827430725, + 0.05447062849998474, + 0.04191324859857559, + 0.04275573045015335, + 0.0873001292347908 + ], + "credit_map": [ + 0.018795402720570564, + 0.024796677753329277, + 0.035316068679094315, + 0.01594613678753376, + 0.027819430455565453, + 0.041870422661304474, + 0.018403466790914536, + 0.029494158923625946, + 0.017898470163345337, + 0.03837240859866142, + 0.02568952739238739, + 0.0364307202398777, + 0.01818416267633438, + 0.009410332888364792, + 0.022714093327522278, + 0.024064138531684875, + 0.015955697745084763, + 0.021415453404188156, + 0.02335995063185692, + 0.01802859455347061, + 0.01789245940744877, + 0.016026845201849937, + 0.014377698302268982, + 0.01946474425494671, + 0.008720487356185913, + 0.005537932273000479, + 0.012897027656435966, + 0.006489843595772982, + 0.005508147645741701, + 0.004655048251152039, + 0.010925491340458393, + 0.007943181321024895, + 0.009785310365259647, + 0.006340880878269672, + 0.01233716681599617, + 0.006461549550294876, + 0.0054932404309511185, + 0.007816563360393047, + 0.007724255323410034, + 0.008209089748561382, + 0.008699432015419006, + 0.014331845566630363, + 0.01368111651390791, + 0.010301374830305576, + 0.00763298012316227, + 0.011443930678069592, + 0.018062910065054893, + 0.010834439657628536, + 0.004794170614331961, + 0.007180608343333006, + 0.0064170765690505505, + 0.007147739641368389, + 0.004589985124766827, + 0.008639172650873661, + 0.00882693100720644, + 0.006874950136989355, + 0.004629978910088539, + 0.004787600599229336, + 0.005867464933544397, + 0.005884090904146433, + 0.004716790746897459, + 0.0037758280523121357, + 0.008177739568054676, + 0.004073759075254202, + 0.004513334017246962, + 0.004831213038414717, + 0.005631437059491873, + 0.006294609513133764, + 0.0073346360586583614, + 0.003832034533843398, + 0.005243356805294752, + 0.005075113382190466, + 0.008836455643177032, + 0.010846194811165333, + 0.015269480645656586, + 0.007788774557411671, + 0.014610904268920422, + 0.010721839033067226, + 0.011301704682409763, + 0.007924780249595642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07002793252468109 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, wide-angle shot, shot from above, slowly pulls back to reveal a moment of monumental and devastating collapse. At the center of the frame is an ancient, massive stone wall perched on the edge of a precipitous cliff. Composed of massive, weathered granite blocks of varying sizes, its surface covered in dark green moss and withered vines, its cracks brimming with weathered mortar. Under a dark, overcast sky, the cool, evenly diffused light accentuates the wall's weathered texture. Suddenly, with a deep, teeth-grinding crack, a crucial upper support stone, under immense pressure, thrusts outward, setting off an unstoppable chain reaction. Tons of boulders seemed to have broken free from the constraints of thousands of years and were violently peeled off from the wall. They rolled heavily in the air and collided violently with each other, emitting sporadic sparks, and then carried a cloud of yellow-brown smoke and dust that swallowed the sky and the earth, rolling down the cliff with the force of thunder. The smoke and dust rose up in an instant, blurring the outline of the valley. The whole scene was full of awe-inspiring, primitive destructive power.", + "seed": 22589, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_148_seed_22589_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12777997860684992, + "vlm_logprob": -0.12777997860684992, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.12777997860684992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22571296989917755 + ], + "opposite_logprob_tokens": [ + -1.7257130146026611 + ], + "logprob_reward": -0.22571296989917755, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22571296989917755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22985361516475677, + 0.04464208334684372, + 0.09040064364671707, + 0.0912436842918396, + 0.25775158405303955, + 0.03227531537413597, + 0.016716739162802696, + 0.016553737223148346, + 0.01914677768945694, + 0.2014157772064209 + ], + "credit_map": [ + 0.007430741563439369, + 0.017802812159061432, + 0.01503528468310833, + 0.010273586958646774, + 0.013743575662374496, + 0.08170147985219955, + 0.010422107763588428, + 0.07344401627779007, + 0.002476857975125313, + 0.006129472982138395, + 0.006477345246821642, + 0.005988372955471277, + 0.0048595936968922615, + 0.006943044252693653, + 0.006000992376357317, + 0.005766402930021286, + 0.0063062310218811035, + 0.02150723710656166, + 0.010396554134786129, + 0.010923796333372593, + 0.009707476012408733, + 0.01247462909668684, + 0.010583343915641308, + 0.00850137509405613, + 0.005154803860932589, + 0.014470185153186321, + 0.010075640864670277, + 0.01614196226000786, + 0.00864252820611, + 0.013763060793280602, + 0.013541696593165398, + 0.009453818202018738, + 0.018494559451937675, + 0.033697038888931274, + 0.030679624527692795, + 0.03048912063241005, + 0.02449781447649002, + 0.034670159220695496, + 0.03258170560002327, + 0.05264157056808472, + 0.001377634471282363, + 0.0014681457541882992, + 0.003078919369727373, + 0.005230956710875034, + 0.004178490024060011, + 0.006829599849879742, + 0.004073630552738905, + 0.006037939805537462, + 0.0011864108964800835, + 0.0017201866721734405, + 0.0027722057420760393, + 0.0027283737435936928, + 0.0020595695823431015, + 0.002228457946330309, + 0.002132388995960355, + 0.0018891457002609968, + 0.0009191324352286756, + 0.0008365210378542542, + 0.0023655404802411795, + 0.0022075744345784187, + 0.0017544394358992577, + 0.002554333768785, + 0.0031029698438942432, + 0.0028132256120443344, + 0.001326371799223125, + 0.001379810506477952, + 0.0028453711420297623, + 0.0023198712151497602, + 0.0025894984137266874, + 0.0026128245517611504, + 0.0034307970199733973, + 0.002642232459038496, + 0.010744869709014893, + 0.005949526093900204, + 0.009051238186657429, + 0.0229048952460289, + 0.013867586851119995, + 0.045508112758398056, + 0.08100369572639465, + 0.012385852634906769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.22571296989917755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521422207355499 + ], + "opposite_logprob_tokens": [ + -1.5021421909332275 + ], + "logprob_reward": -0.2521422207355499, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521422207355499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10651592910289764, + 0.04702630266547203, + 0.03983090817928314, + 0.021579567342996597, + 0.06676358729600906, + 0.03460657596588135, + 0.04856041818857193, + 0.046001508831977844, + 0.037007905542850494, + 0.5521073341369629 + ], + "credit_map": [ + 0.0032495460473001003, + 0.004505496472120285, + 0.007338722702115774, + 0.00885745882987976, + 0.007990040816366673, + 0.034892525523900986, + 0.007776049897074699, + 0.03190608695149422, + 0.003941324073821306, + 0.007509559392929077, + 0.005134380888193846, + 0.009195567108690739, + 0.004371349234133959, + 0.0069764163345098495, + 0.0050581363029778, + 0.004839565139263868, + 0.0026689788792282343, + 0.007125246338546276, + 0.004046381916850805, + 0.004478165414184332, + 0.004295103717595339, + 0.006508821155875921, + 0.006525720469653606, + 0.004182487726211548, + 0.0013027817476540804, + 0.0027277895715087652, + 0.0025773386005312204, + 0.004357467405498028, + 0.0021831581834703684, + 0.0034826132468879223, + 0.0025972456205636263, + 0.002351172734051943, + 0.003793854732066393, + 0.004932735115289688, + 0.008849356323480606, + 0.00884009338915348, + 0.007584029342979193, + 0.009139863774180412, + 0.011350308544933796, + 0.012273341417312622, + 0.001868946012109518, + 0.002280776621773839, + 0.0028693354688584805, + 0.00562988081946969, + 0.002957076532766223, + 0.00644017243757844, + 0.007289264351129532, + 0.0052711209282279015, + 0.002551252255216241, + 0.005943590309470892, + 0.007417385466396809, + 0.005965694319456816, + 0.008341317996382713, + 0.0062909373082220554, + 0.006211814470589161, + 0.005838420242071152, + 0.0023698003496974707, + 0.003247616346925497, + 0.007293487433344126, + 0.006135911215096712, + 0.005194639787077904, + 0.005986534990370274, + 0.009536531753838062, + 0.006236981134861708, + 0.0029107790905982256, + 0.00351644866168499, + 0.0038485692348331213, + 0.0045397356152534485, + 0.004923113156110048, + 0.004492561332881451, + 0.007036077789962292, + 0.005740621592849493, + 0.03019261732697487, + 0.017814766615629196, + 0.023230472579598427, + 0.06304457038640976, + 0.038992077112197876, + 0.12476521730422974, + 0.2209034115076065, + 0.033164165914058685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.2521422207355499 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076352194882929325 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.0076352194882929325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076352194882929325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5219877362251282, + 0.08855888992547989, + 0.07894323766231537, + 0.024277733638882637, + 0.03912145644426346, + 0.029829762876033783, + 0.02229158580303192, + 0.025688372552394867, + 0.032192591577768326, + 0.13710865378379822 + ], + "credit_map": [ + 0.012580456212162971, + 0.015914591029286385, + 0.023914095014333725, + 0.019170276820659637, + 0.02401094138622284, + 0.20745213329792023, + 0.019061513245105743, + 0.19988375902175903, + 0.006978183053433895, + 0.011657294817268848, + 0.010526003316044807, + 0.009819543920457363, + 0.008804043754935265, + 0.015469449572265148, + 0.012468001805245876, + 0.012836367823183537, + 0.006382501218467951, + 0.018793189898133278, + 0.007929805666208267, + 0.008546828292310238, + 0.010519317351281643, + 0.010834221728146076, + 0.008120118640363216, + 0.007817260921001434, + 0.0013122378150001168, + 0.0023620540741831064, + 0.0023419063072651625, + 0.006936094257980585, + 0.002345574786886573, + 0.003776933066546917, + 0.0027695628814399242, + 0.0024333710316568613, + 0.0029420487117022276, + 0.0028799534775316715, + 0.004542783834040165, + 0.005307398270815611, + 0.005159782711416483, + 0.005777142476290464, + 0.00474684638902545, + 0.007765504065901041, + 0.001981814159080386, + 0.0014985373709350824, + 0.002862896304577589, + 0.003454636549577117, + 0.0032335754949599504, + 0.006385868415236473, + 0.004395187832415104, + 0.006017246749252081, + 0.0015498825814574957, + 0.0023434869945049286, + 0.003865145379677415, + 0.0024414188228547573, + 0.0036747935228049755, + 0.0031561912037432194, + 0.0026499503292143345, + 0.0026107165031135082, + 0.0016862987540662289, + 0.0018334623891860247, + 0.004624898079782724, + 0.002720889635384083, + 0.004308830015361309, + 0.003419201821088791, + 0.003928858321160078, + 0.0031659340020269156, + 0.0021424831356853247, + 0.0034788325428962708, + 0.004759312141686678, + 0.003697785083204508, + 0.00486427778378129, + 0.0037317827809602022, + 0.004639691207557917, + 0.0048784250393509865, + 0.00828811340034008, + 0.007100080139935017, + 0.007234020158648491, + 0.016201993450522423, + 0.011517778038978577, + 0.02936353161931038, + 0.04901145398616791, + 0.008391681127250195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0076352194882929325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347355917096138 + ], + "opposite_logprob_tokens": [ + -3.6513473987579346 + ], + "logprob_reward": -0.026347355917096138, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347355917096138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19348855316638947, + 0.040042851120233536, + 0.061126966029405594, + 0.03085017390549183, + 0.04308390989899635, + 0.0696127638220787, + 0.06969375163316727, + 0.03966377303004265, + 0.10733242332935333, + 0.3451049029827118 + ], + "credit_map": [ + 0.007229642476886511, + 0.01037534512579441, + 0.012367215007543564, + 0.01211980264633894, + 0.01798611879348755, + 0.06430038064718246, + 0.010555987246334553, + 0.058554038405418396, + 0.003059892449527979, + 0.005139882210642099, + 0.004062962252646685, + 0.00741736451163888, + 0.0033430999610573053, + 0.007069719489663839, + 0.004769304767251015, + 0.005180620588362217, + 0.004223284777253866, + 0.015012047253549099, + 0.006582231260836124, + 0.007250380702316761, + 0.006811428349465132, + 0.008869102224707603, + 0.006580302957445383, + 0.005798181518912315, + 0.0016277865506708622, + 0.004847235046327114, + 0.003277735086157918, + 0.005471549462527037, + 0.0031835732515901327, + 0.005513747688382864, + 0.002972174435853958, + 0.003956370055675507, + 0.002296034712344408, + 0.0033645909279584885, + 0.0049149817787110806, + 0.005673299543559551, + 0.004145022016018629, + 0.0063712019473314285, + 0.007572947535663843, + 0.008745827712118626, + 0.0025405697524547577, + 0.003238104050979018, + 0.005984016228467226, + 0.01128825917840004, + 0.008117320947349072, + 0.01279586274176836, + 0.015097237192094326, + 0.0105513921007514, + 0.004022692795842886, + 0.006180665455758572, + 0.009664897806942463, + 0.011462857946753502, + 0.011973950080573559, + 0.009825930930674076, + 0.00852219294756651, + 0.008040559478104115, + 0.002642153762280941, + 0.00273396959528327, + 0.007272638846188784, + 0.006656889338046312, + 0.004966456443071365, + 0.004457802511751652, + 0.006488146260380745, + 0.0044457134790718555, + 0.006970859132707119, + 0.008226226083934307, + 0.010988250374794006, + 0.015890680253505707, + 0.017865687608718872, + 0.014565027318894863, + 0.01976132020354271, + 0.013064366765320301, + 0.017681412398815155, + 0.010560552589595318, + 0.016869211569428444, + 0.03786531090736389, + 0.023538347333669662, + 0.07829614728689194, + 0.13906151056289673, + 0.021232392638921738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.026347355917096138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127062126994133 + ], + "opposite_logprob_tokens": [ + -2.1270620822906494 + ], + "logprob_reward": -0.127062126994133, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127062126994133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2015516757965088, + 0.03983957692980766, + 0.13919563591480255, + 0.05433828383684158, + 0.08174668997526169, + 0.14849801361560822, + 0.07518314570188522, + 0.03455545753240585, + 0.030281420797109604, + 0.19481003284454346 + ], + "credit_map": [ + 0.004784226883202791, + 0.006258559878915548, + 0.009038377553224564, + 0.0075247022323310375, + 0.008271841332316399, + 0.08221623301506042, + 0.0077066076919436455, + 0.07575114071369171, + 0.003167433198541403, + 0.0039788284339010715, + 0.005263825878500938, + 0.006789924576878548, + 0.0036208091769367456, + 0.0060922675766050816, + 0.006213874090462923, + 0.004712613765150309, + 0.010816751979291439, + 0.030994560569524765, + 0.016019269824028015, + 0.01459771953523159, + 0.018283003941178322, + 0.019926249980926514, + 0.014445701614022255, + 0.014112388715147972, + 0.0038774372078478336, + 0.004022042267024517, + 0.007941937074065208, + 0.010808488354086876, + 0.004548852331936359, + 0.010411843657493591, + 0.0054582576267421246, + 0.007269425317645073, + 0.007260683458298445, + 0.010572089813649654, + 0.010080046951770782, + 0.010332105681300163, + 0.009071857668459415, + 0.011002088896930218, + 0.009605246596038342, + 0.013822565786540508, + 0.005156028550118208, + 0.004230061080306768, + 0.004862899426370859, + 0.007896017283201218, + 0.005566468928009272, + 0.057234201580286026, + 0.0065630218014121056, + 0.056989315897226334, + 0.004158106632530689, + 0.010137601755559444, + 0.009682385250926018, + 0.011485318653285503, + 0.016265712678432465, + 0.007382670883089304, + 0.00865787360817194, + 0.007413480430841446, + 0.0024452274665236473, + 0.0028180356603115797, + 0.005345309618860483, + 0.004376673139631748, + 0.0043158396147191525, + 0.004508329089730978, + 0.0065290615893900394, + 0.004216983448714018, + 0.0020761319901794195, + 0.0023218903224915266, + 0.0038773021660745144, + 0.0031223217956721783, + 0.003598702372983098, + 0.005452664103358984, + 0.004872229881584644, + 0.004960177466273308, + 0.011385153979063034, + 0.0082097128033638, + 0.008314434438943863, + 0.021993981674313545, + 0.01692807674407959, + 0.042468663305044174, + 0.0738268792629242, + 0.011683134362101555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.127062126994133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, wide-angle shot, shot from above, slowly pulls back to reveal a moment of monumental and devastating collapse. At the center of the frame is an ancient, massive stone wall perched on the edge of a precipitous cliff. Composed of massive, weathered granite blocks of varying sizes, its surface covered in dark green moss and withered vines, its cracks brimming with weathered mortar. Under a dark, overcast sky, the cool, evenly diffused light accentuates the wall's weathered texture. Suddenly, with a deep, teeth-grinding crack, a crucial upper support stone, under immense pressure, thrusts outward, setting off an unstoppable chain reaction. Tons of boulders seemed to have broken free from the constraints of thousands of years and were violently peeled off from the wall. They rolled heavily in the air and collided violently with each other, emitting sporadic sparks, and then carried a cloud of yellow-brown smoke and dust that swallowed the sky and the earth, rolling down the cliff with the force of thunder. The smoke and dust rose up in an instant, blurring the outline of the valley. The whole scene was full of awe-inspiring, primitive destructive power.", + "seed": 22590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_148_seed_22590_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07625395059585571, + "vlm_logprob": -0.07625395059585571, + "vlm_yes_no_margin": 2.975000023841858, + "vlm_token_credit": -0.07625395059585571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13527561724185944 + ], + "opposite_logprob_tokens": [ + -2.2602756023406982 + ], + "logprob_reward": -0.13527561724185944, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13527561724185944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16698497533798218, + 0.07385893166065216, + 0.07506882399320602, + 0.15374042093753815, + 0.12263119965791702, + 0.08275214582681656, + 0.07999736815690994, + 0.05262632668018341, + 0.11247435212135315, + 0.0798654779791832 + ], + "credit_map": [ + 0.015332558192312717, + 0.015161706134676933, + 0.02014748938381672, + 0.02891717292368412, + 0.015641048550605774, + 0.016879098489880562, + 0.02752239629626274, + 0.02738349512219429, + 0.007927600294351578, + 0.0068910568952560425, + 0.014601110480725765, + 0.010345470160245895, + 0.006886071991175413, + 0.008447892963886261, + 0.009489087387919426, + 0.009270640090107918, + 0.0068201408721506596, + 0.00884219165891409, + 0.018082408234477043, + 0.011301031336188316, + 0.007055431138724089, + 0.0062449099496006966, + 0.008940007537603378, + 0.0077827018685638905, + 0.018288472667336464, + 0.016603833064436913, + 0.022339267656207085, + 0.019976772367954254, + 0.011415109969675541, + 0.027342813089489937, + 0.01856549270451069, + 0.01920866034924984, + 0.01736980862915516, + 0.01268852036446333, + 0.011665884405374527, + 0.020457228645682335, + 0.009737761691212654, + 0.016704201698303223, + 0.02225033938884735, + 0.011757449246942997, + 0.011317181400954723, + 0.011912120506167412, + 0.009675226174294949, + 0.012415874749422073, + 0.00771266408264637, + 0.0096573606133461, + 0.010181557387113571, + 0.009880159981548786, + 0.010688078589737415, + 0.010495760478079319, + 0.007684851996600628, + 0.009065745398402214, + 0.007100950460880995, + 0.01107491459697485, + 0.011552647687494755, + 0.012334425002336502, + 0.0064385863952338696, + 0.0063382526859641075, + 0.006823748350143433, + 0.006984054110944271, + 0.00558306323364377, + 0.0057264408096671104, + 0.00873142946511507, + 0.0060007511638104916, + 0.013794406317174435, + 0.018675372004508972, + 0.010780919343233109, + 0.01188233308494091, + 0.013522210530936718, + 0.012240364216268063, + 0.017129894345998764, + 0.014448852278292179, + 0.006539239548146725, + 0.007750657852739096, + 0.008828343823552132, + 0.008317692205309868, + 0.009826648980379105, + 0.00911533273756504, + 0.0185173861682415, + 0.010970184579491615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.13527561724185944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283673465251923 + ], + "opposite_logprob_tokens": [ + -2.017836809158325 + ], + "logprob_reward": -0.14283673465251923, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14283673465251923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21920958161354065, + 0.04988763481378555, + 0.038129691034555435, + 0.13795897364616394, + 0.1027095764875412, + 0.10317006707191467, + 0.1206548884510994, + 0.048761505633592606, + 0.058851517736911774, + 0.12066657841205597 + ], + "credit_map": [ + 0.015925971791148186, + 0.018175754696130753, + 0.03295114263892174, + 0.041503969579935074, + 0.01976441591978073, + 0.02313772775232792, + 0.042965274304151535, + 0.024785324931144714, + 0.0035050311125814915, + 0.004450129810720682, + 0.008127600885927677, + 0.0074158585630357265, + 0.004304023459553719, + 0.007162195630371571, + 0.009936803951859474, + 0.004985990468412638, + 0.003039781004190445, + 0.0037439526058733463, + 0.00874463003128767, + 0.0056422315537929535, + 0.0033220723271369934, + 0.004145747981965542, + 0.004904703237116337, + 0.004586571361869574, + 0.012825868092477322, + 0.014122942462563515, + 0.025663305073976517, + 0.024608759209513664, + 0.01557757519185543, + 0.016997095197439194, + 0.015550847165286541, + 0.012612581253051758, + 0.011563954874873161, + 0.011700539849698544, + 0.012478183023631573, + 0.018084483221173286, + 0.008420328609645367, + 0.016086352989077568, + 0.016832513734698296, + 0.007543223910033703, + 0.013243144378066063, + 0.011761772446334362, + 0.01168499980121851, + 0.014231123961508274, + 0.009481962770223618, + 0.017903631553053856, + 0.012229240499436855, + 0.01263419073075056, + 0.013070371001958847, + 0.012734765186905861, + 0.011889934539794922, + 0.015610823407769203, + 0.01320610661059618, + 0.013005513697862625, + 0.01795484870672226, + 0.023182524368166924, + 0.003559533040970564, + 0.0054804133251309395, + 0.004459704738110304, + 0.005869976710528135, + 0.005020953249186277, + 0.00508493697270751, + 0.011397331953048706, + 0.007888654246926308, + 0.00696603674441576, + 0.00895620882511139, + 0.006840848829597235, + 0.0053293537348508835, + 0.006748431362211704, + 0.0057646711356937885, + 0.012052583508193493, + 0.006193380802869797, + 0.010656770318746567, + 0.011181658133864403, + 0.0137107502669096, + 0.009640820324420929, + 0.016813812777400017, + 0.013055260293185711, + 0.029427258297801018, + 0.016180245205760002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.14283673465251923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546654790639877 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546654790639877, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546654790639877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37808823585510254, + 0.08921469002962112, + 0.04595956206321716, + 0.08375132828950882, + 0.05459916591644287, + 0.05422858148813248, + 0.06539852172136307, + 0.09823707491159439, + 0.08761217445135117, + 0.042910635471343994 + ], + "credit_map": [ + 0.02547943964600563, + 0.028681190684437752, + 0.04647926613688469, + 0.07453525066375732, + 0.03806362301111221, + 0.04706922173500061, + 0.06939704716205597, + 0.0483832024037838, + 0.00887291505932808, + 0.006296177860349417, + 0.010988954454660416, + 0.012241837568581104, + 0.00820090901106596, + 0.016947560012340546, + 0.015420924872159958, + 0.010245412588119507, + 0.003945274278521538, + 0.00443731714040041, + 0.007080265320837498, + 0.007880626246333122, + 0.006575696635991335, + 0.005181670654565096, + 0.006258935667574406, + 0.004599773790687323, + 0.005242842249572277, + 0.005370726343244314, + 0.013481125235557556, + 0.02336745150387287, + 0.009508411400020123, + 0.00884704664349556, + 0.007952864281833172, + 0.009980862028896809, + 0.0068460493348538876, + 0.006440664175897837, + 0.0060278004966676235, + 0.009067114442586899, + 0.004914046730846167, + 0.007861973717808723, + 0.007625491824001074, + 0.005816024262458086, + 0.0058099813759326935, + 0.006713805254548788, + 0.005779877305030823, + 0.008870510384440422, + 0.004279520362615585, + 0.010885579511523247, + 0.0068705580197274685, + 0.005018750671297312, + 0.006390908267349005, + 0.0075736199505627155, + 0.007710690144449472, + 0.010516583919525146, + 0.005936452187597752, + 0.008770003914833069, + 0.007600368466228247, + 0.010899892076849937, + 0.009410674683749676, + 0.014073546975851059, + 0.007937151938676834, + 0.012292380444705486, + 0.015813855454325676, + 0.014228676445782185, + 0.012545767240226269, + 0.01193502452224493, + 0.01014427375048399, + 0.008795802481472492, + 0.010402862913906574, + 0.009982289746403694, + 0.009620944038033485, + 0.008990121074020863, + 0.020949004217982292, + 0.008726870641112328, + 0.003663326846435666, + 0.0032057510688900948, + 0.0056626866571605206, + 0.004576018545776606, + 0.003610889660194516, + 0.004791805054992437, + 0.01073520164936781, + 0.0066649578511714935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.012546654790639877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581256598234177 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581256598234177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581256598234177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23212379217147827, + 0.057736482471227646, + 0.049352873116731644, + 0.11233390867710114, + 0.09099528193473816, + 0.05367830768227577, + 0.04997192323207855, + 0.22532279789447784, + 0.07666381448507309, + 0.051820822060108185 + ], + "credit_map": [ + 0.016680672764778137, + 0.01947232149541378, + 0.02275845780968666, + 0.04000046104192734, + 0.0234173983335495, + 0.02996787056326866, + 0.04077163338661194, + 0.039054956287145615, + 0.004593414720147848, + 0.004443197511136532, + 0.007678832393139601, + 0.009923544712364674, + 0.004534252919256687, + 0.007983755320310593, + 0.012555059045553207, + 0.006024419330060482, + 0.003757340367883444, + 0.0051690987311303616, + 0.007860742509365082, + 0.006665872409939766, + 0.005234288517385721, + 0.006889801472425461, + 0.006567472126334906, + 0.00720825232565403, + 0.010744048282504082, + 0.009170004166662693, + 0.022935260087251663, + 0.013336827978491783, + 0.008262736722826958, + 0.01889704167842865, + 0.01252060104161501, + 0.016467375680804253, + 0.013013777323067188, + 0.011198433116078377, + 0.0091450996696949, + 0.013938674703240395, + 0.008497113361954689, + 0.009029838256537914, + 0.011950120329856873, + 0.014222216792404652, + 0.005851894151419401, + 0.0069255284033715725, + 0.006006953772157431, + 0.008537663146853447, + 0.004799088928848505, + 0.008136674761772156, + 0.006386096589267254, + 0.0070344023406505585, + 0.006649733055382967, + 0.005772068630903959, + 0.005191681440919638, + 0.006477513816207647, + 0.004782631527632475, + 0.006992207374423742, + 0.0068709515035152435, + 0.0072351316921412945, + 0.024614060297608376, + 0.021275360137224197, + 0.020559674128890038, + 0.023643966764211655, + 0.03257119283080101, + 0.03181051090359688, + 0.04531900957226753, + 0.025529006496071815, + 0.008889442309737206, + 0.006335445214062929, + 0.01160622201859951, + 0.010645896196365356, + 0.008106729947030544, + 0.0073120309971272945, + 0.010184094309806824, + 0.01358394231647253, + 0.004645842593163252, + 0.004599893465638161, + 0.0062700072303414345, + 0.006649806629866362, + 0.005202858243137598, + 0.006246872246265411, + 0.010834882035851479, + 0.007370653562247753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.020581256598234177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002948969602585 + ], + "opposite_logprob_tokens": [ + -2.6950294971466064 + ], + "logprob_reward": -0.07002948969602585, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002948969602585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34760352969169617, + 0.23257897794246674, + 0.12630900740623474, + 0.05829453095793724, + 0.03143083676695824, + 0.013570141047239304, + 0.058163322508335114, + 0.040939200669527054, + 0.05203798785805702, + 0.039072539657354355 + ], + "credit_map": [ + 0.026684436947107315, + 0.033072009682655334, + 0.04200047254562378, + 0.06396836787462234, + 0.030057214200496674, + 0.04364398866891861, + 0.06147943437099457, + 0.04669751599431038, + 0.02773374691605568, + 0.03208940476179123, + 0.035131826996803284, + 0.04182633012533188, + 0.017510855570435524, + 0.016177542507648468, + 0.038035571575164795, + 0.024073665961623192, + 0.009985560551285744, + 0.012014216743409634, + 0.014548893086612225, + 0.02481685020029545, + 0.009950891137123108, + 0.01177388709038496, + 0.017302682623267174, + 0.025915993377566338, + 0.005587426014244556, + 0.00493193743750453, + 0.005480744410306215, + 0.012199328280985355, + 0.008434128016233444, + 0.008350744843482971, + 0.006842072121798992, + 0.006468141917139292, + 0.004066933877766132, + 0.0034448730293661356, + 0.002653402043506503, + 0.0045567466877400875, + 0.0035664390306919813, + 0.005508149042725563, + 0.004980419296771288, + 0.0026538679376244545, + 0.0011993420775979757, + 0.0015630517154932022, + 0.0016619950765743852, + 0.002648895140737295, + 0.0015224834205582738, + 0.0011859082151204348, + 0.0018353594932705164, + 0.0019531029975041747, + 0.007294181268662214, + 0.006066123489290476, + 0.004831934347748756, + 0.005227237939834595, + 0.005990755278617144, + 0.008049658499658108, + 0.009377136826515198, + 0.011326282285153866, + 0.0038111195899546146, + 0.00404087221249938, + 0.0042583332397043705, + 0.004134143702685833, + 0.005892875604331493, + 0.005954541265964508, + 0.008210713975131512, + 0.004636595956981182, + 0.005054306238889694, + 0.0062492298893630505, + 0.005640148650854826, + 0.006697321310639381, + 0.007922514341771603, + 0.006228548940271139, + 0.009017829783260822, + 0.005228076595813036, + 0.0032095802016556263, + 0.003392307087779045, + 0.005104627925902605, + 0.004861896391957998, + 0.005299399606883526, + 0.004075719974935055, + 0.0068534864112734795, + 0.006275511812418699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07002948969602585 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, wide-angle shot, shot from above, slowly pulls back to reveal a moment of monumental and devastating collapse. At the center of the frame is an ancient, massive stone wall perched on the edge of a precipitous cliff. Composed of massive, weathered granite blocks of varying sizes, its surface covered in dark green moss and withered vines, its cracks brimming with weathered mortar. Under a dark, overcast sky, the cool, evenly diffused light accentuates the wall's weathered texture. Suddenly, with a deep, teeth-grinding crack, a crucial upper support stone, under immense pressure, thrusts outward, setting off an unstoppable chain reaction. Tons of boulders seemed to have broken free from the constraints of thousands of years and were violently peeled off from the wall. They rolled heavily in the air and collided violently with each other, emitting sporadic sparks, and then carried a cloud of yellow-brown smoke and dust that swallowed the sky and the earth, rolling down the cliff with the force of thunder. The smoke and dust rose up in an instant, blurring the outline of the valley. The whole scene was full of awe-inspiring, primitive destructive power.", + "seed": 22591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_1/index_148_seed_22591_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11101232869550585, + "vlm_logprob": -0.11101232869550585, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.11101232869550585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20780831575393677 + ], + "opposite_logprob_tokens": [ + -1.832808256149292 + ], + "logprob_reward": -0.20780831575393677, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20780831575393677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0798446536064148, + 0.024524910375475883, + 0.049061331897974014, + 0.07107630372047424, + 0.02597692608833313, + 0.017118314281105995, + 0.5993893146514893, + 0.059793177992105484, + 0.04613759368658066, + 0.02707749791443348 + ], + "credit_map": [ + 0.0069815004244446754, + 0.011812014505267143, + 0.00758093548938632, + 0.015278293751180172, + 0.00861271657049656, + 0.005229076836258173, + 0.013150845654308796, + 0.011199267581105232, + 0.002104711253196001, + 0.0033150524832308292, + 0.002839828608557582, + 0.00355151086114347, + 0.0026441311929374933, + 0.0017032172763720155, + 0.00435144966468215, + 0.004015009384602308, + 0.004670321010053158, + 0.004167495761066675, + 0.0047358181327581406, + 0.006321198306977749, + 0.0050911568105220795, + 0.006621102336794138, + 0.010830078274011612, + 0.006624163594096899, + 0.00631509255617857, + 0.00563350273296237, + 0.013086953200399876, + 0.007882739417254925, + 0.006178960204124451, + 0.007027848158031702, + 0.010269487276673317, + 0.014681725762784481, + 0.0026310565881431103, + 0.004217400215566158, + 0.002388355555012822, + 0.002671885536983609, + 0.0036026977468281984, + 0.0037181973457336426, + 0.0038014238234609365, + 0.0029459097422659397, + 0.001679645967669785, + 0.0021677925251424313, + 0.0019591059535741806, + 0.0028403049800544977, + 0.001736355829052627, + 0.001662042224779725, + 0.0033411155454814434, + 0.001731951255351305, + 0.01820002682507038, + 0.02707185596227646, + 0.02514304406940937, + 0.1813998520374298, + 0.011327022686600685, + 0.022655854001641273, + 0.05331550911068916, + 0.2602761387825012, + 0.004396936856210232, + 0.005631488747894764, + 0.00617648009210825, + 0.003871280699968338, + 0.01031710859388113, + 0.01582341454923153, + 0.00682490598410368, + 0.006751565728336573, + 0.002611588453873992, + 0.0014457034412771463, + 0.01325253490358591, + 0.0029425756074488163, + 0.0027928135823458433, + 0.0025432815309613943, + 0.002756598172709346, + 0.01779249496757984, + 0.0036455299705266953, + 0.0032274937257170677, + 0.004366779234260321, + 0.003509840462356806, + 0.0022145977709442377, + 0.0029257747810333967, + 0.003451179713010788, + 0.003736302023753524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.20780831575393677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558854520320892 + ], + "opposite_logprob_tokens": [ + -1.6005885601043701 + ], + "logprob_reward": -0.22558854520320892, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558854520320892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05406498536467552, + 0.011757138185203075, + 0.030899183824658394, + 0.1762123852968216, + 0.02044687047600746, + 0.030995283275842667, + 0.5786933898925781, + 0.04612908512353897, + 0.024457909166812897, + 0.026343807578086853 + ], + "credit_map": [ + 0.00585185457020998, + 0.008431701920926571, + 0.007142811082303524, + 0.008507066406309605, + 0.0035844738595187664, + 0.003223337698727846, + 0.00964123010635376, + 0.007682506926357746, + 0.0013810953823849559, + 0.0012309389421716332, + 0.001732939388602972, + 0.0013890754198655486, + 0.0012944679474458098, + 0.0012176325544714928, + 0.0016572107560932636, + 0.0018537779105827212, + 0.003643637290224433, + 0.0037053273990750313, + 0.002923196880146861, + 0.0042360094375908375, + 0.0033526744227856398, + 0.0043886457569897175, + 0.004663328640162945, + 0.003986363764852285, + 0.015097740106284618, + 0.005346679128706455, + 0.05456947907805443, + 0.01101442240178585, + 0.00495811365544796, + 0.007904080674052238, + 0.014715530909597874, + 0.0626063197851181, + 0.0015523218316957355, + 0.0033488390035927296, + 0.0020678271539509296, + 0.002451709471642971, + 0.0022463835775852203, + 0.0036646071821451187, + 0.002558631356805563, + 0.002556550782173872, + 0.004874819424003363, + 0.0032887363340705633, + 0.0042640333995223045, + 0.00378173659555614, + 0.0028463087510317564, + 0.002343178493902087, + 0.005153610371053219, + 0.004442858509719372, + 0.01735115423798561, + 0.026305541396141052, + 0.02334527112543583, + 0.1744282841682434, + 0.011030795983970165, + 0.022017329931259155, + 0.051774200052022934, + 0.25244078040122986, + 0.003489665687084198, + 0.00549292890354991, + 0.005176031030714512, + 0.003297405084595084, + 0.006760864518582821, + 0.008640664629638195, + 0.006511093117296696, + 0.006760432384908199, + 0.0021113206166774035, + 0.0014682867331430316, + 0.005633937660604715, + 0.002338987309485674, + 0.0018481593579053879, + 0.0017179163405671716, + 0.0023791282437741756, + 0.006960172206163406, + 0.002259925240650773, + 0.0030351595487445593, + 0.0025694214273244143, + 0.0032826962415128946, + 0.004318780265748501, + 0.004159601870924234, + 0.002910420997068286, + 0.0038078029174357653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.22558854520320892 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258656572550535 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258656572550535, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258656572550535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16455379128456116, + 0.07269621640443802, + 0.03549698367714882, + 0.17316190898418427, + 0.03458039462566376, + 0.047414280474185944, + 0.20287184417247772, + 0.08055388182401657, + 0.12918232381343842, + 0.05948839709162712 + ], + "credit_map": [ + 0.014772244729101658, + 0.027183393016457558, + 0.019533880054950714, + 0.027275756001472473, + 0.01375114917755127, + 0.010940869338810444, + 0.028469182550907135, + 0.022627312690019608, + 0.007570969872176647, + 0.010515792295336723, + 0.0099980803206563, + 0.008684790693223476, + 0.00629728427156806, + 0.004864135757088661, + 0.011986958794295788, + 0.012778200209140778, + 0.0035144805442541838, + 0.004935929551720619, + 0.0035415254533290863, + 0.00428262073546648, + 0.0034357942640781403, + 0.004693904891610146, + 0.006323269102722406, + 0.004769457038491964, + 0.01442723162472248, + 0.0052495719864964485, + 0.05392523482441902, + 0.01090321596711874, + 0.00384943257085979, + 0.006365257780998945, + 0.01381885539740324, + 0.06462310999631882, + 0.0031700972467660904, + 0.0039994907565414906, + 0.0043906569480896, + 0.004188730847090483, + 0.0050034718587994576, + 0.005235059652477503, + 0.004831606987863779, + 0.0037612775340676308, + 0.00796483550220728, + 0.005220369901508093, + 0.005434269551187754, + 0.0067900498397648335, + 0.005026916973292828, + 0.002812198596075177, + 0.006789588835090399, + 0.007376053370535374, + 0.006920863874256611, + 0.00921158492565155, + 0.00986936129629612, + 0.05758671462535858, + 0.004629926290363073, + 0.007793269120156765, + 0.01971616968512535, + 0.08714394271373749, + 0.00649265618994832, + 0.008167319931089878, + 0.008849184028804302, + 0.005533373448997736, + 0.013768551871180534, + 0.02116684801876545, + 0.008270922116935253, + 0.00830503087490797, + 0.007174992933869362, + 0.0034277918748557568, + 0.039066921919584274, + 0.007100129500031471, + 0.007725130766630173, + 0.006510122679173946, + 0.007570128887891769, + 0.050607092678546906, + 0.006909737829118967, + 0.004542127717286348, + 0.008208290673792362, + 0.010325698181986809, + 0.006714475341141224, + 0.008434535004198551, + 0.006594039499759674, + 0.007759498432278633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.005258656572550535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071652993559837 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016071652993559837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071652993559837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051962923258543015, + 0.022009775042533875, + 0.01291139516979456, + 0.18199191987514496, + 0.018919315189123154, + 0.048061955720186234, + 0.13841766119003296, + 0.10422462970018387, + 0.3034929931163788, + 0.1180073618888855 + ], + "credit_map": [ + 0.006886761169880629, + 0.007044379599392414, + 0.006851394195109606, + 0.006397204473614693, + 0.006684130989015102, + 0.0035744761116802692, + 0.00807429663836956, + 0.006450281944125891, + 0.002193908207118511, + 0.0025798643473535776, + 0.0033208047971129417, + 0.0028119233902543783, + 0.0020652494858950377, + 0.0016432830598205328, + 0.0035185711458325386, + 0.003876170376315713, + 0.0014728860696777701, + 0.002001302782446146, + 0.0013656864175572991, + 0.0021019065752625465, + 0.0014842887176200747, + 0.0013880239566788077, + 0.0013383299810811877, + 0.001758970902301371, + 0.015255377627909184, + 0.00482177035883069, + 0.05751274153590202, + 0.011381302028894424, + 0.004444883204996586, + 0.006850880570709705, + 0.014637667685747147, + 0.06708730012178421, + 0.001932744518853724, + 0.0029628234915435314, + 0.001427596784196794, + 0.0022671895567327738, + 0.002366907661780715, + 0.0027098548598587513, + 0.0025274751242250204, + 0.0027247234247624874, + 0.009004161693155766, + 0.006211029831320047, + 0.005564502906054258, + 0.006698087323457003, + 0.005304180551320314, + 0.003468044102191925, + 0.005946654360741377, + 0.005865295883268118, + 0.004537811037153006, + 0.007122959475964308, + 0.0062028770335018635, + 0.04132453352212906, + 0.0033493665978312492, + 0.005271852016448975, + 0.012407620437443256, + 0.05820063501596451, + 0.008540383540093899, + 0.010641893371939659, + 0.011603902094066143, + 0.008012058213353157, + 0.017807766795158386, + 0.025555435568094254, + 0.011829071678221226, + 0.010234126821160316, + 0.016469743102788925, + 0.008547966368496418, + 0.09138280153274536, + 0.017535636201500893, + 0.01757078431546688, + 0.015233148820698261, + 0.018707500770688057, + 0.11804541200399399, + 0.013994972221553326, + 0.011749720200896263, + 0.014992566779255867, + 0.013327733613550663, + 0.01154306624084711, + 0.018280738964676857, + 0.015104906633496284, + 0.019013661891222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.016071652993559837 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, wide-angle view taken from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an ancient massive stone wall made of weathered granite blocks covered in dark green moss and withered vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wall collapse with a crucial upper support stone thrusting outward, causing boulders to break free, roll, collide with sparks, and generate yellow-brown smoke and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and evenly diffused under a dark overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033447295427322 + ], + "opposite_logprob_tokens": [ + -2.350334405899048 + ], + "logprob_reward": -0.10033447295427322, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033447295427322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07575198262929916, + 0.017817221581935883, + 0.14028124511241913, + 0.23217326402664185, + 0.014703246764838696, + 0.044753871858119965, + 0.32144972681999207, + 0.019584838300943375, + 0.0845307931303978, + 0.048953812569379807 + ], + "credit_map": [ + 0.009366385638713837, + 0.009047702886164188, + 0.00944042019546032, + 0.010873965919017792, + 0.00608257157728076, + 0.005567801650613546, + 0.014532594010233879, + 0.010840537957847118, + 0.0016452374402433634, + 0.002142665907740593, + 0.002270278986543417, + 0.0029244478791952133, + 0.0017398146446794271, + 0.0012586974771693349, + 0.0035563886631280184, + 0.0022796890698373318, + 0.011984417214989662, + 0.009204194881021976, + 0.012345435097813606, + 0.01690996252000332, + 0.010238440707325935, + 0.021272730082273483, + 0.04000748321413994, + 0.01831856556236744, + 0.019355695694684982, + 0.0059927282854914665, + 0.0721844807267189, + 0.014352616854012012, + 0.005250732414424419, + 0.009032337926328182, + 0.019661350175738335, + 0.08634332567453384, + 0.0018227316904813051, + 0.0020584885496646166, + 0.0015632209833711386, + 0.0015133122215047479, + 0.0017145968740805984, + 0.0019925322849303484, + 0.001592386281117797, + 0.0024459767155349255, + 0.007363345939666033, + 0.0055757854133844376, + 0.005608777981251478, + 0.006457736250013113, + 0.005771531257778406, + 0.003628683974966407, + 0.005797185003757477, + 0.0045508211478590965, + 0.010005179792642593, + 0.014891669154167175, + 0.013414038345217705, + 0.09409313648939133, + 0.006790232844650745, + 0.012308021076023579, + 0.030373580753803253, + 0.1395738571882248, + 0.0018668933771550655, + 0.0019656445365399122, + 0.001965574687346816, + 0.0023806607350707054, + 0.003144878661260009, + 0.0034128017723560333, + 0.0026819584891200066, + 0.002166423946619034, + 0.005169680342078209, + 0.0026324247010052204, + 0.023954249918460846, + 0.004986769985407591, + 0.006038694642484188, + 0.004214786924421787, + 0.005718003958463669, + 0.03181617334485054, + 0.006212032865732908, + 0.005228492897003889, + 0.0070221018977463245, + 0.007633173372596502, + 0.0038488968275487423, + 0.005612015258520842, + 0.006312426179647446, + 0.007084672339260578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.10033447295427322 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_tensor([7886, 7886])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_tensor([7886, 7886])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..43c3d99f2afd495604937d9b31827b01db7df764 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_tensor([7886, 7886])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This static, eye-level close-up captures a crystal-clear image of a downy duck adorably struggling in the calm waters of a sun-dappled pond. Its downy feathers, a vibrant lemon yellow, are rendered in surreal detail, with tufts of feathers slightly tangled in the water. Its beady black eyes widen with earnest concentration as it attempts to swim. Its movements are a series of clumsy micro-movements: it kicks its tiny webbed feet in a frantic, uncoordinated pattern, causing its entire body to sway precariously. Then, it stumbles through the water, pitching heavily to one side, flapping its crude, undeveloped wings in a futile, unbalanced attempt to steady itself before tentatively regaining its footing. The entire scene is illuminated by the warm golden glow of the afternoon sun, which filters through unseen trees, creating shimmering caustics on the water's surface and a soft, glowing rimlight on the duckling's body. The soft-focus blur of deep green water lilies and dark, reflective water in the background creates a tender and endearing atmosphere, focusing entirely on the young creature's innocent yet comical efforts.", + "seed": 1910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_7886_seed_1910_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10372464172542095, + "vlm_logprob": -0.10372464172542095, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.10372464172542095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10425316542387009 + ], + "opposite_logprob_tokens": [ + -2.354253053665161 + ], + "logprob_reward": -0.10425316542387009, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10425316542387009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0534554123878479, + 0.13971783220767975, + 0.08122585713863373, + 0.0641707107424736, + 0.0363430380821228, + 0.03191372752189636, + 0.2181071937084198, + 0.06185714900493622, + 0.23383747041225433, + 0.0793716162443161 + ], + "credit_map": [ + 0.006580322049558163, + 0.00519547238945961, + 0.004856992978602648, + 0.005561681464314461, + 0.007049313746392727, + 0.005613463930785656, + 0.009821289218962193, + 0.008776875212788582, + 0.015126881189644337, + 0.01326084230095148, + 0.01680522970855236, + 0.017990240827202797, + 0.02409748174250126, + 0.015453987754881382, + 0.020647365599870682, + 0.01633579097688198, + 0.014180253259837627, + 0.006481282878667116, + 0.007785674184560776, + 0.006455432157963514, + 0.010529868304729462, + 0.01707979291677475, + 0.00879267044365406, + 0.009920875541865826, + 0.007788724731653929, + 0.00962678249925375, + 0.004784446209669113, + 0.006259585730731487, + 0.0123885627835989, + 0.011369841173291206, + 0.006200338248163462, + 0.005752424709498882, + 0.0029111173935234547, + 0.0039211162365973, + 0.004860574845224619, + 0.0038066112902015448, + 0.003985328134149313, + 0.005878831725567579, + 0.006386504974216223, + 0.004592950455844402, + 0.00447651045396924, + 0.004099379293620586, + 0.00320781534537673, + 0.005897646304219961, + 0.003994799219071865, + 0.0036648313980549574, + 0.0036392242182046175, + 0.0029335208237171173, + 0.03333084657788277, + 0.009320872835814953, + 0.010965338908135891, + 0.009410231374204159, + 0.0174198467284441, + 0.08268845826387405, + 0.011782900430262089, + 0.043188679963350296, + 0.006311068776994944, + 0.007941557094454765, + 0.007308153435587883, + 0.009198903106153011, + 0.010600592009723186, + 0.006927653681486845, + 0.008417629636824131, + 0.005151587072759867, + 0.017641915008425713, + 0.016676904633641243, + 0.01851676031947136, + 0.013638789765536785, + 0.06553468108177185, + 0.06509652733802795, + 0.021563522517681122, + 0.015168357640504837, + 0.008973353542387486, + 0.010492153465747833, + 0.010843892581760883, + 0.008478650823235512, + 0.010355542413890362, + 0.009342517703771591, + 0.013666294515132904, + 0.007219207473099232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.10425316542387009 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344151496887207 + ], + "opposite_logprob_tokens": [ + -1.313441514968872 + ], + "logprob_reward": -0.31344151496887207, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344151496887207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05914856493473053, + 0.05929269641637802, + 0.04936869814991951, + 0.2273952215909958, + 0.14959102869033813, + 0.04514830932021141, + 0.0943249762058258, + 0.052984267473220825, + 0.1948944628238678, + 0.06785177439451218 + ], + "credit_map": [ + 0.006320867221802473, + 0.008059819228947163, + 0.005184173583984375, + 0.0048073516227304935, + 0.008189061656594276, + 0.011239740997552872, + 0.008578686974942684, + 0.006768863648176193, + 0.007377453614026308, + 0.005132691003382206, + 0.005606869701296091, + 0.008459644392132759, + 0.011324581690132618, + 0.00603262335062027, + 0.008475190959870815, + 0.006883643567562103, + 0.005987243726849556, + 0.005502461455762386, + 0.004387271124869585, + 0.0045850458554923534, + 0.011972243897616863, + 0.007752155885100365, + 0.006072653457522392, + 0.003109622048214078, + 0.026035789400339127, + 0.03700951114296913, + 0.017961319535970688, + 0.024467458948493004, + 0.041247352957725525, + 0.040241360664367676, + 0.02161998860538006, + 0.01881243847310543, + 0.017455613240599632, + 0.0080863730981946, + 0.010332346893846989, + 0.006324618123471737, + 0.013622846454381943, + 0.0529661662876606, + 0.009004509076476097, + 0.031798552721738815, + 0.007449304219335318, + 0.005109073594212532, + 0.00407761987298727, + 0.007877292111515999, + 0.005483055487275124, + 0.007653909735381603, + 0.004400060512125492, + 0.0030979891307651997, + 0.012453592382371426, + 0.006752598099410534, + 0.006999679375439882, + 0.010068325325846672, + 0.009718733839690685, + 0.02652212418615818, + 0.007200154941529036, + 0.014609773643314838, + 0.0054664043709635735, + 0.00745370052754879, + 0.006297969724982977, + 0.005536735523492098, + 0.010217745788395405, + 0.006226874887943268, + 0.0076555898413062096, + 0.004129248205572367, + 0.014217018149793148, + 0.0213942788541317, + 0.011900702491402626, + 0.011561734601855278, + 0.05825235694646835, + 0.05635299161076546, + 0.011881065554916859, + 0.009334306232631207, + 0.008487497456371784, + 0.008501001633703709, + 0.012376056984066963, + 0.007786542642861605, + 0.007631220854818821, + 0.008490190841257572, + 0.009161355905234814, + 0.0054179104045033455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.31344151496887207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890357568860054 + ], + "opposite_logprob_tokens": [ + -2.464035749435425 + ], + "logprob_reward": -0.0890357568860054, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890357568860054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0400254912674427, + 0.0732269361615181, + 0.04811311885714531, + 0.20238327980041504, + 0.11939544975757599, + 0.029946908354759216, + 0.253849059343338, + 0.05350378155708313, + 0.1272592842578888, + 0.05229669809341431 + ], + "credit_map": [ + 0.004764115903526545, + 0.005674122367054224, + 0.003305386286228895, + 0.00390760600566864, + 0.005513880401849747, + 0.006235809065401554, + 0.00690452940762043, + 0.0037200413644313812, + 0.011017311364412308, + 0.006745758932083845, + 0.007449731230735779, + 0.00989408977329731, + 0.01194891706109047, + 0.00727996788918972, + 0.008958655409514904, + 0.009932505898177624, + 0.007187454495579004, + 0.005104049574583769, + 0.004189983010292053, + 0.004274897743016481, + 0.008958166465163231, + 0.008535153232514858, + 0.005761235021054745, + 0.004102177452296019, + 0.02266916260123253, + 0.03336947783827782, + 0.01620856299996376, + 0.022830625995993614, + 0.037192195653915405, + 0.03544744476675987, + 0.01929401233792305, + 0.015371796675026417, + 0.013703130185604095, + 0.006473064422607422, + 0.00887261237949133, + 0.00699265394359827, + 0.011272931471467018, + 0.041102759540081024, + 0.008977875113487244, + 0.022000424563884735, + 0.0038411025889217854, + 0.003998637665063143, + 0.002374309115111828, + 0.0059877256862819195, + 0.003366556018590927, + 0.0034845578484237194, + 0.00406227633357048, + 0.002831743098795414, + 0.037201084196567535, + 0.012527844868600368, + 0.01335854735225439, + 0.015533960424363613, + 0.021364176645874977, + 0.09114862978458405, + 0.014065887778997421, + 0.0486488975584507, + 0.005178330931812525, + 0.007468872703611851, + 0.006131492089480162, + 0.0059736561961472034, + 0.010405251756310463, + 0.006280410103499889, + 0.0079755624756217, + 0.004090206231921911, + 0.009996694512665272, + 0.01580473780632019, + 0.007320106960833073, + 0.007825545966625214, + 0.03634864464402199, + 0.034430019557476044, + 0.009300288744270802, + 0.006233253050595522, + 0.006193104665726423, + 0.006413248833268881, + 0.0068774656392633915, + 0.006935247220098972, + 0.0086392005905509, + 0.0062054917216300964, + 0.007316288538277149, + 0.0037166462279856205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.0890357568860054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948103964328766 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948103964328766, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948103964328766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14787772297859192, + 0.07486190646886826, + 0.052181996405124664, + 0.034247878938913345, + 0.2149609923362732, + 0.013645049184560776, + 0.16928772628307343, + 0.018962031230330467, + 0.09498536586761475, + 0.17898929119110107 + ], + "credit_map": [ + 0.01592530868947506, + 0.014526312239468098, + 0.013993024826049805, + 0.012009581550955772, + 0.025565654039382935, + 0.013407768681645393, + 0.023078611120581627, + 0.029371460899710655, + 0.009617398492991924, + 0.007730469107627869, + 0.009805831126868725, + 0.008778433315455914, + 0.012364083901047707, + 0.00765842804685235, + 0.009924276731908321, + 0.008982982486486435, + 0.008653795346617699, + 0.004057723097503185, + 0.0061120521277189255, + 0.0056036850437521935, + 0.008273659273982048, + 0.009282653219997883, + 0.005969794001430273, + 0.004228635225445032, + 0.004549573175609112, + 0.004835034720599651, + 0.002838551066815853, + 0.003837019670754671, + 0.00606962013989687, + 0.004096153657883406, + 0.004625420086085796, + 0.0033965101465582848, + 0.025449827313423157, + 0.009350262582302094, + 0.0142943374812603, + 0.006984720006585121, + 0.0202658548951149, + 0.07857217639684677, + 0.014665108174085617, + 0.04537869989871979, + 0.0017944066785275936, + 0.0014811091823503375, + 0.0014068495947867632, + 0.0020567115861922503, + 0.0014028216246515512, + 0.0019140340154990554, + 0.002195230685174465, + 0.0013938862830400467, + 0.02511882409453392, + 0.0076172915287315845, + 0.008850307203829288, + 0.009751064702868462, + 0.013129401952028275, + 0.06056802347302437, + 0.010635804384946823, + 0.03361700102686882, + 0.002236627973616123, + 0.0029243917670100927, + 0.0019524300005286932, + 0.0020783853251487017, + 0.0033433486241847277, + 0.0019945406820625067, + 0.0025190357118844986, + 0.0019132725428789854, + 0.008002779446542263, + 0.0069089592434465885, + 0.005897475406527519, + 0.005999694112688303, + 0.02640240266919136, + 0.02421046979725361, + 0.010705473832786083, + 0.006858114618808031, + 0.018144669011235237, + 0.025760037824511528, + 0.03271856531500816, + 0.018194548785686493, + 0.021956399083137512, + 0.020330457016825676, + 0.025523371994495392, + 0.01636124588549137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.005948103964328766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944667384028435 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944667384028435, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944667384028435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07383602857589722, + 0.0966358408331871, + 0.14326593279838562, + 0.068802610039711, + 0.09456147998571396, + 0.05373010039329529, + 0.3505215048789978, + 0.08521667867898941, + 0.012485944665968418, + 0.020943794399499893 + ], + "credit_map": [ + 0.01078438013792038, + 0.006767659913748503, + 0.005548569839447737, + 0.00816698931157589, + 0.009131764061748981, + 0.008824827149510384, + 0.012120863422751427, + 0.012490982189774513, + 0.015572968870401382, + 0.008781415410339832, + 0.009933206252753735, + 0.011537620797753334, + 0.01341773010790348, + 0.010254394263029099, + 0.013127106241881847, + 0.014011415652930737, + 0.021257376298308372, + 0.015707803890109062, + 0.014667161740362644, + 0.01564251072704792, + 0.021748658269643784, + 0.02277234196662903, + 0.015746643766760826, + 0.015723442658782005, + 0.010647709481418133, + 0.008326489478349686, + 0.005004027858376503, + 0.0073738694190979, + 0.013488240540027618, + 0.010872743092477322, + 0.006753932684659958, + 0.006335599813610315, + 0.011550189927220345, + 0.0067415134981274605, + 0.0072525134310126305, + 0.004513971041887999, + 0.009826394729316235, + 0.031643327325582504, + 0.008160525001585484, + 0.014873052947223186, + 0.006826114375144243, + 0.0071168141439557076, + 0.004878593143075705, + 0.011147585697472095, + 0.005086474120616913, + 0.005989220459014177, + 0.008683021180331707, + 0.004002280533313751, + 0.05340644717216492, + 0.016219189390540123, + 0.019221080467104912, + 0.01697005145251751, + 0.02891463041305542, + 0.1304033100605011, + 0.01776350848376751, + 0.0676233097910881, + 0.008372592739760876, + 0.010671389289200306, + 0.011523454450070858, + 0.010258348658680916, + 0.015259470790624619, + 0.00967414677143097, + 0.012170020490884781, + 0.007287264801561832, + 0.0016341260634362698, + 0.001565634156577289, + 0.0019172638421878219, + 0.0015241907676681876, + 0.0017038618680089712, + 0.0013287013862282038, + 0.0016777548007667065, + 0.0011344143422320485, + 0.002228934783488512, + 0.002492589643225074, + 0.004061955492943525, + 0.002740641124546528, + 0.0019229883328080177, + 0.0026532222982496023, + 0.0030046412721276283, + 0.0018388228490948677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.005944667384028435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level close-up captures a crystal-clear image of a downy duck adorably struggling in the calm waters of a sun-dappled pond. Its downy feathers, a vibrant lemon yellow, are rendered in surreal detail, with tufts of feathers slightly tangled in the water. Its beady black eyes widen with earnest concentration as it attempts to swim. Its movements are a series of clumsy micro-movements: it kicks its tiny webbed feet in a frantic, uncoordinated pattern, causing its entire body to sway precariously. Then, it stumbles through the water, pitching heavily to one side, flapping its crude, undeveloped wings in a futile, unbalanced attempt to steady itself before tentatively regaining its footing. The entire scene is illuminated by the warm golden glow of the afternoon sun, which filters through unseen trees, creating shimmering caustics on the water's surface and a soft, glowing rimlight on the duckling's body. The soft-focus blur of deep green water lilies and dark, reflective water in the background creates a tender and endearing atmosphere, focusing entirely on the young creature's innocent yet comical efforts.", + "seed": 1911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_7886_seed_1911_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14077561143785716, + "vlm_logprob": -0.14077561143785716, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.14077561143785716, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22810491919517517 + ], + "opposite_logprob_tokens": [ + -1.6031049489974976 + ], + "logprob_reward": -0.22810491919517517, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22810491919517517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43806514143943787, + 0.08792746067047119, + 0.019755709916353226, + 0.05448354035615921, + 0.03845585510134697, + 0.030933158472180367, + 0.07584010064601898, + 0.050111789256334305, + 0.0897330567240715, + 0.11469423025846481 + ], + "credit_map": [ + 0.046357911080121994, + 0.05553058162331581, + 0.04258807376027107, + 0.04718993976712227, + 0.0654691830277443, + 0.04990236461162567, + 0.08693042397499084, + 0.04409665986895561, + 0.010856245644390583, + 0.015077464282512665, + 0.006800355855375528, + 0.01335162203758955, + 0.012712649069726467, + 0.01042685005813837, + 0.009742840193212032, + 0.008959433063864708, + 0.0028598990757018328, + 0.0023607497569173574, + 0.002612013602629304, + 0.002736572874709964, + 0.00258953170850873, + 0.0017838187050074339, + 0.0024018408730626106, + 0.002411283552646637, + 0.006967386696487665, + 0.006017642095685005, + 0.0044255442917346954, + 0.004625894129276276, + 0.013998821377754211, + 0.006716899108141661, + 0.006409660447388887, + 0.0053216926753520966, + 0.0036956442054361105, + 0.0036438554525375366, + 0.002825808012858033, + 0.005461910739541054, + 0.0048158010467886925, + 0.005965157877653837, + 0.006719574797898531, + 0.005328102502971888, + 0.002942490391433239, + 0.0033452885691076517, + 0.0031098879408091307, + 0.003335576271638274, + 0.004454039037227631, + 0.0042987409979105, + 0.0031605095136910677, + 0.0062866248190402985, + 0.01034245640039444, + 0.0067107584327459335, + 0.01059346366673708, + 0.008069314993917942, + 0.00987197458744049, + 0.01114089135080576, + 0.009384134784340858, + 0.009727108292281628, + 0.0057596489787101746, + 0.005679579451680183, + 0.006154161412268877, + 0.0053168912418186665, + 0.006160181015729904, + 0.00782568845897913, + 0.0061005717143416405, + 0.007115068379789591, + 0.00869812723249197, + 0.007562091574072838, + 0.006151887122541666, + 0.005485903471708298, + 0.02866535447537899, + 0.01905905082821846, + 0.0055511146783828735, + 0.008559530600905418, + 0.010419392958283424, + 0.03452414646744728, + 0.014999917708337307, + 0.006382275838404894, + 0.009306404739618301, + 0.008673974312841892, + 0.02304014563560486, + 0.007347969338297844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.22810491919517517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42891013622283936 + ], + "opposite_logprob_tokens": [ + -1.0539101362228394 + ], + "logprob_reward": -0.42891013622283936, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42891013622283936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.222251296043396, + 0.05298624560236931, + 0.025541972368955612, + 0.05754955857992172, + 0.05767884477972984, + 0.0843304917216301, + 0.06260006129741669, + 0.2378854900598526, + 0.03751104697585106, + 0.16166503727436066 + ], + "credit_map": [ + 0.02736169658601284, + 0.029843153432011604, + 0.02398529462516308, + 0.029914233833551407, + 0.030364079400897026, + 0.02843928337097168, + 0.025492943823337555, + 0.026850629597902298, + 0.006828832905739546, + 0.008622735738754272, + 0.006457078270614147, + 0.007538372650742531, + 0.0059812129475176334, + 0.007776489946991205, + 0.005157032981514931, + 0.00462449062615633, + 0.0028630197048187256, + 0.0035361212212592363, + 0.0028292543720453978, + 0.0032980693504214287, + 0.004017014056444168, + 0.002608615905046463, + 0.0028132344596087933, + 0.0035766437649726868, + 0.007120802067220211, + 0.006597515661269426, + 0.006087609566748142, + 0.006527361460030079, + 0.012484905309975147, + 0.006246341858059168, + 0.005807869601994753, + 0.006677150726318359, + 0.0041548642329871655, + 0.006321107503026724, + 0.005267574917525053, + 0.006184106692671776, + 0.0078184949234128, + 0.012980179861187935, + 0.005553032737225294, + 0.009399483911693096, + 0.0072692567482590675, + 0.008145474828779697, + 0.006954310927540064, + 0.00939379632472992, + 0.011549368500709534, + 0.015806129202246666, + 0.008409740403294563, + 0.016802413389086723, + 0.006776537746191025, + 0.005656328983604908, + 0.011207790113985538, + 0.007310714107006788, + 0.00802008155733347, + 0.007899054326117039, + 0.01035852637141943, + 0.005371033679693937, + 0.02094840258359909, + 0.02012072131037712, + 0.015792682766914368, + 0.014954040758311749, + 0.07291825860738754, + 0.055061716586351395, + 0.01653486117720604, + 0.021554792299866676, + 0.004006768576800823, + 0.0037277902010828257, + 0.00476033752784133, + 0.003890765830874443, + 0.005242291372269392, + 0.0065439543686807156, + 0.0036129476502537727, + 0.00572618842124939, + 0.011332942172884941, + 0.05585174635052681, + 0.008075081743299961, + 0.007436084561049938, + 0.01598713919520378, + 0.010998884215950966, + 0.04562922194600105, + 0.006353926844894886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.42891013622283936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810085728764534 + ], + "opposite_logprob_tokens": [ + -3.2881009578704834 + ], + "logprob_reward": -0.03810085728764534, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810085728764534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2193867564201355, + 0.06089683994650841, + 0.03706026449799538, + 0.053932588547468185, + 0.058551087975502014, + 0.08365541696548462, + 0.031367748975753784, + 0.18137353658676147, + 0.04303887486457825, + 0.2307368665933609 + ], + "credit_map": [ + 0.030582530423998833, + 0.02660030871629715, + 0.02553281933069229, + 0.028087500482797623, + 0.026271983981132507, + 0.027545040473341942, + 0.026681261137127876, + 0.02808530442416668, + 0.00824983324855566, + 0.009376749396324158, + 0.008846042677760124, + 0.008324810303747654, + 0.006336968857795, + 0.006482755299657583, + 0.006581756751984358, + 0.006697927135974169, + 0.004849825520068407, + 0.006974852178245783, + 0.003832899732515216, + 0.003961042035371065, + 0.005648563150316477, + 0.0050603533163666725, + 0.0033803482074290514, + 0.0033523805905133486, + 0.005898506846278906, + 0.005110791884362698, + 0.007030532229691744, + 0.006413118448108435, + 0.008287238888442516, + 0.004862767178565264, + 0.007920237258076668, + 0.008409399539232254, + 0.004457609262317419, + 0.006976557429879904, + 0.005605731625109911, + 0.006755354348570108, + 0.008134444244205952, + 0.012291744351387024, + 0.00566375907510519, + 0.008665893226861954, + 0.007377252448350191, + 0.008119977079331875, + 0.00638470146805048, + 0.009194457903504372, + 0.012562129646539688, + 0.016301926225423813, + 0.00756343686953187, + 0.01615152880549431, + 0.0034669239539653063, + 0.0033957124687731266, + 0.002989002503454685, + 0.004031685180962086, + 0.0029807290993630886, + 0.004659818019717932, + 0.005751428194344044, + 0.004092448856681585, + 0.016301210969686508, + 0.01610918529331684, + 0.013393067754805088, + 0.012418078258633614, + 0.052033502608537674, + 0.0394820012152195, + 0.014362338930368423, + 0.017274152487516403, + 0.004482722375541925, + 0.004434891510754824, + 0.004815602675080299, + 0.00400211475789547, + 0.007506285794079304, + 0.007812191266566515, + 0.003885295009240508, + 0.006099769379943609, + 0.016152171418070793, + 0.08277745544910431, + 0.009805099107325077, + 0.011025648564100266, + 0.02403578720986843, + 0.015883581712841988, + 0.06221846118569374, + 0.008838672190904617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.03810085728764534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028190179727971554 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028190179727971554, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028190179727971554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18466129899024963, + 0.0673837661743164, + 0.024686742573976517, + 0.10438210517168045, + 0.07108991593122482, + 0.04521877318620682, + 0.03066580556333065, + 0.1998681277036667, + 0.22632426023483276, + 0.04571918398141861 + ], + "credit_map": [ + 0.01350073330104351, + 0.024694962427020073, + 0.010855644010007381, + 0.019691230729222298, + 0.02899792604148388, + 0.024258315563201904, + 0.03772738575935364, + 0.024935096502304077, + 0.009945147670805454, + 0.014065008610486984, + 0.003799122292548418, + 0.007118097972124815, + 0.009465583600103855, + 0.010679135099053383, + 0.006166297011077404, + 0.006145375315099955, + 0.0038960292004048824, + 0.0029026991687715054, + 0.0021722321398556232, + 0.0029511833563447, + 0.002943015191704035, + 0.0039871446788311005, + 0.002380464458838105, + 0.0034539743792265654, + 0.01591300033032894, + 0.013839539140462875, + 0.012774264439940453, + 0.0111893555149436, + 0.015122641809284687, + 0.009846823289990425, + 0.011916871182620525, + 0.013779609464108944, + 0.007631597109138966, + 0.008004368282854557, + 0.005570317152887583, + 0.00942231249064207, + 0.008684313856065273, + 0.010046672075986862, + 0.012391866184771061, + 0.009338472969830036, + 0.006336066871881485, + 0.004207677207887173, + 0.004389138426631689, + 0.004942247178405523, + 0.006735641043633223, + 0.008231188170611858, + 0.003642881056293845, + 0.0067339325323700905, + 0.003891562344506383, + 0.003309225896373391, + 0.0048309555277228355, + 0.003271490568295121, + 0.003976341336965561, + 0.004447587765753269, + 0.003044100245460868, + 0.003894544206559658, + 0.01639198325574398, + 0.015371894463896751, + 0.012030222453176975, + 0.011920101009309292, + 0.06736809015274048, + 0.049066200852394104, + 0.011107384227216244, + 0.01661224290728569, + 0.024418961256742477, + 0.02013363689184189, + 0.023604795336723328, + 0.020405499264597893, + 0.020506609231233597, + 0.056984346359968185, + 0.021567195653915405, + 0.03870321065187454, + 0.004131953697651625, + 0.003760081250220537, + 0.00825770664960146, + 0.003389761783182621, + 0.0034323674626648426, + 0.009526596404612064, + 0.009487529285252094, + 0.0037331855855882168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0028190179727971554 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059431265108287334 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.0059431265108287334, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059431265108287334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20162266492843628, + 0.07081662118434906, + 0.05279111489653587, + 0.07370079308748245, + 0.12264731526374817, + 0.08265161514282227, + 0.04406232759356499, + 0.11334485560655594, + 0.04239693656563759, + 0.19596576690673828 + ], + "credit_map": [ + 0.02756314165890217, + 0.018414093181490898, + 0.02417094074189663, + 0.022845657542347908, + 0.020444460213184357, + 0.02246742881834507, + 0.04043597728013992, + 0.025280959904193878, + 0.009901183657348156, + 0.014409718103706837, + 0.005453449208289385, + 0.008568146266043186, + 0.00769816106185317, + 0.011750593781471252, + 0.004916623700410128, + 0.008118746802210808, + 0.007929329760372639, + 0.00458214245736599, + 0.006403768435120583, + 0.005638904869556427, + 0.006280003115534782, + 0.007230384740978479, + 0.006600167602300644, + 0.008126414380967617, + 0.008175265043973923, + 0.008718862198293209, + 0.006434298120439053, + 0.008413995616137981, + 0.010112837888300419, + 0.013628183864057064, + 0.008837053552269936, + 0.00938030518591404, + 0.010711089707911015, + 0.011123438365757465, + 0.010565555654466152, + 0.013909618370234966, + 0.01653459295630455, + 0.02234772779047489, + 0.021949362009763718, + 0.015505928546190262, + 0.009548718109726906, + 0.008738134987652302, + 0.006297048646956682, + 0.009262770414352417, + 0.018085850402712822, + 0.014332747086882591, + 0.006111156661063433, + 0.01027519442141056, + 0.004221135284751654, + 0.004666607361286879, + 0.008619110099971294, + 0.0043937163427472115, + 0.0034120320342481136, + 0.00525525351986289, + 0.007675205357372761, + 0.005819273181259632, + 0.009497018530964851, + 0.009199974127113819, + 0.009441445581614971, + 0.008183147758245468, + 0.03165331110358238, + 0.024413947016000748, + 0.009598138742148876, + 0.011357877403497696, + 0.004488140810281038, + 0.003531757276505232, + 0.0034320109989494085, + 0.0030461973510682583, + 0.0107542360201478, + 0.009977017529308796, + 0.002494035754352808, + 0.004673544783145189, + 0.014471336267888546, + 0.06868483871221542, + 0.00982548575848341, + 0.00842673983424902, + 0.02028508670628071, + 0.011762414127588272, + 0.05493631213903427, + 0.007573570590466261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0059431265108287334 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level close-up captures a crystal-clear image of a downy duck adorably struggling in the calm waters of a sun-dappled pond. Its downy feathers, a vibrant lemon yellow, are rendered in surreal detail, with tufts of feathers slightly tangled in the water. Its beady black eyes widen with earnest concentration as it attempts to swim. Its movements are a series of clumsy micro-movements: it kicks its tiny webbed feet in a frantic, uncoordinated pattern, causing its entire body to sway precariously. Then, it stumbles through the water, pitching heavily to one side, flapping its crude, undeveloped wings in a futile, unbalanced attempt to steady itself before tentatively regaining its footing. The entire scene is illuminated by the warm golden glow of the afternoon sun, which filters through unseen trees, creating shimmering caustics on the water's surface and a soft, glowing rimlight on the duckling's body. The soft-focus blur of deep green water lilies and dark, reflective water in the background creates a tender and endearing atmosphere, focusing entirely on the young creature's innocent yet comical efforts.", + "seed": 1912, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_7886_seed_1912_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.22633206900209188, + "vlm_logprob": -0.22633206900209188, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.22633206900209188, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2845105528831482 + ], + "opposite_logprob_tokens": [ + -1.409510612487793 + ], + "logprob_reward": -0.2845105528831482, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2845105528831482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.197044238448143, + 0.06131907179951668, + 0.05656809359788895, + 0.2434578835964203, + 0.09083154797554016, + 0.11063586175441742, + 0.11194970458745956, + 0.02556784823536873, + 0.05691545084118843, + 0.045710302889347076 + ], + "credit_map": [ + 0.01846309006214142, + 0.026505203917622566, + 0.025534745305776596, + 0.020130401477217674, + 0.036791060119867325, + 0.021551482379436493, + 0.03130461648106575, + 0.01676364056766033, + 0.010196014307439327, + 0.005721185356378555, + 0.0070502422749996185, + 0.006204408127814531, + 0.007781043648719788, + 0.006371515803039074, + 0.009266329929232597, + 0.0087283318862319, + 0.006295714993029833, + 0.007686989847570658, + 0.009944459423422813, + 0.005568238906562328, + 0.00646437332034111, + 0.005223578307777643, + 0.008193793706595898, + 0.007190943695604801, + 0.029391111806035042, + 0.04385875165462494, + 0.026500113308429718, + 0.02250455506145954, + 0.036917127668857574, + 0.02892565354704857, + 0.021803604438900948, + 0.033556971698999405, + 0.010719714686274529, + 0.009725736454129219, + 0.008144807070493698, + 0.01917826384305954, + 0.010135948657989502, + 0.009865839965641499, + 0.007852730341255665, + 0.015208510681986809, + 0.016928084194660187, + 0.02956067956984043, + 0.008702537044882774, + 0.011779868975281715, + 0.012337255291640759, + 0.013025530613958836, + 0.01137219462543726, + 0.006929714698344469, + 0.004654749296605587, + 0.0066917366348207, + 0.004425282124429941, + 0.004242497030645609, + 0.04166879132390022, + 0.0390147902071476, + 0.006408327724784613, + 0.004843526519834995, + 0.002864112379029393, + 0.0033578379079699516, + 0.004235441330820322, + 0.0021106989588588476, + 0.0046691689640283585, + 0.0035491471644490957, + 0.0022563517559319735, + 0.002525088842958212, + 0.006850800476968288, + 0.0071805864572525024, + 0.006553194019943476, + 0.010344808921217918, + 0.007317465730011463, + 0.007853611372411251, + 0.006454888265579939, + 0.004360097460448742, + 0.007946197874844074, + 0.0049842894077301025, + 0.005509448237717152, + 0.004663222469389439, + 0.00312680471688509, + 0.004070543684065342, + 0.0093830656260252, + 0.00602673040702939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.2845105528831482 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328494548797607 + ], + "opposite_logprob_tokens": [ + -0.7578494548797607 + ], + "logprob_reward": -0.6328494548797607, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328494548797607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07963632792234421, + 0.058738235384225845, + 0.12368956953287125, + 0.17194361984729767, + 0.17113307118415833, + 0.16336478292942047, + 0.03804470971226692, + 0.029139844700694084, + 0.10857783257961273, + 0.055732037872076035 + ], + "credit_map": [ + 0.008936366066336632, + 0.008978620171546936, + 0.00849063042551279, + 0.008976574055850506, + 0.01423867791891098, + 0.012597627937793732, + 0.011955352500081062, + 0.005462480243295431, + 0.007187264040112495, + 0.005882113240659237, + 0.007719609420746565, + 0.00691610062494874, + 0.010100035928189754, + 0.00670995656400919, + 0.00667713675647974, + 0.007546016480773687, + 0.014956269413232803, + 0.015441637486219406, + 0.01440091896802187, + 0.012337268330156803, + 0.024104950949549675, + 0.012109925039112568, + 0.014614932239055634, + 0.01572366990149021, + 0.01608501933515072, + 0.024978093802928925, + 0.015331386588513851, + 0.01553112268447876, + 0.02306397631764412, + 0.021052099764347076, + 0.032719071954488754, + 0.02318284846842289, + 0.0125277079641819, + 0.016658293083310127, + 0.013765585608780384, + 0.04132550582289696, + 0.023693596944212914, + 0.021381787955760956, + 0.020041586831212044, + 0.02173900417983532, + 0.027029775083065033, + 0.027304532006382942, + 0.019305754452943802, + 0.020429110154509544, + 0.0198187418282032, + 0.017009245231747627, + 0.018489299342036247, + 0.013978325761854649, + 0.0030323914252221584, + 0.004592366982251406, + 0.0024935828987509012, + 0.0037802793085575104, + 0.008913961239159107, + 0.007872971706092358, + 0.0036007105372846127, + 0.0037584444507956505, + 0.003089181613177061, + 0.003945866599678993, + 0.005023536738008261, + 0.003056525718420744, + 0.0038957514334470034, + 0.0037449277006089687, + 0.0025358095299452543, + 0.0038482453674077988, + 0.011671032756567001, + 0.011156602762639523, + 0.014854764565825462, + 0.028393959626555443, + 0.009276670403778553, + 0.011918873526155949, + 0.011911775916814804, + 0.009394154883921146, + 0.006107473745942116, + 0.004061630927026272, + 0.01782509870827198, + 0.0038361744955182076, + 0.003938162233680487, + 0.0038983558770269156, + 0.004153511021286249, + 0.011911633424460888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.6328494548797607 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161272585391998 + ], + "opposite_logprob_tokens": [ + -1.7016127109527588 + ], + "logprob_reward": -0.20161272585391998, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161272585391998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306866556406021, + 0.05049777030944824, + 0.10362745821475983, + 0.3085036277770996, + 0.10713398456573486, + 0.13390709459781647, + 0.030346684157848358, + 0.02577207051217556, + 0.05597108602523804, + 0.05355353280901909 + ], + "credit_map": [ + 0.02065705694258213, + 0.015682922676205635, + 0.012940539047122002, + 0.015769364312291145, + 0.015183374285697937, + 0.01722453162074089, + 0.01856539398431778, + 0.014663465321063995, + 0.008295286446809769, + 0.006855084095150232, + 0.005453846417367458, + 0.005020723678171635, + 0.005745759233832359, + 0.005946972407400608, + 0.0066015939228236675, + 0.006578501313924789, + 0.017204903066158295, + 0.010268722660839558, + 0.010279249399900436, + 0.010197428986430168, + 0.019670741632580757, + 0.011039823293685913, + 0.011672740802168846, + 0.013293847441673279, + 0.03012639284133911, + 0.047479357570409775, + 0.0298045352101326, + 0.026040829718112946, + 0.04305882751941681, + 0.03560316935181618, + 0.05097484588623047, + 0.04541569948196411, + 0.00862160325050354, + 0.009467246010899544, + 0.008623824454843998, + 0.026439432054758072, + 0.013631261885166168, + 0.012307590804994106, + 0.013377328403294086, + 0.014665700495243073, + 0.020559711381793022, + 0.014499693177640438, + 0.01633034273982048, + 0.02081879787147045, + 0.01538820005953312, + 0.01381803210824728, + 0.01662464812397957, + 0.015867678448557854, + 0.002998678246513009, + 0.004420696757733822, + 0.0030667963437736034, + 0.0028363268356770277, + 0.005733390338718891, + 0.0044062091037631035, + 0.0033163626212626696, + 0.0035682248417288065, + 0.0023840656504034996, + 0.005042718257755041, + 0.0024777986109256744, + 0.0025048619136214256, + 0.004059961065649986, + 0.0036424826830625534, + 0.0028384539764374495, + 0.0028217281214892864, + 0.005990443751215935, + 0.0084515530616045, + 0.006999197416007519, + 0.006386092398315668, + 0.008787931874394417, + 0.0071939388290047646, + 0.0066711315885186195, + 0.0054907952435314655, + 0.008640105836093426, + 0.005481833592057228, + 0.014281835407018661, + 0.005486279726028442, + 0.0035540948156267405, + 0.004224193282425404, + 0.005412738304585218, + 0.006472449284046888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.20161272585391998 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951895844191313 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951895844191313, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951895844191313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07103089988231659, + 0.04952871799468994, + 0.08561956882476807, + 0.3427046239376068, + 0.03848954290151596, + 0.10559435933828354, + 0.03374405577778816, + 0.023296663537621498, + 0.03836940973997116, + 0.2116221785545349 + ], + "credit_map": [ + 0.0075121377594769, + 0.007600088603794575, + 0.005518841557204723, + 0.008565469644963741, + 0.011468096636235714, + 0.006760018412023783, + 0.015897488221526146, + 0.007708761841058731, + 0.00980111863464117, + 0.0036408391315490007, + 0.004238464869558811, + 0.005076475907117128, + 0.004132668487727642, + 0.00863601453602314, + 0.00761757418513298, + 0.006385560613125563, + 0.014517714269459248, + 0.007860875688493252, + 0.00826379843056202, + 0.008193941786885262, + 0.011956973932683468, + 0.008358730003237724, + 0.012765812687575817, + 0.01370171457529068, + 0.037576861679553986, + 0.05955580249428749, + 0.036536626517772675, + 0.030089646577835083, + 0.0517716147005558, + 0.04020676016807556, + 0.03833886235952377, + 0.048628438264131546, + 0.003900517476722598, + 0.003370356047526002, + 0.002961050719022751, + 0.009390756487846375, + 0.005051970947533846, + 0.0045560370199382305, + 0.003971535246819258, + 0.0052873194217681885, + 0.015550157055258751, + 0.01203474123030901, + 0.012933808378875256, + 0.016016732901334763, + 0.011950121261179447, + 0.010489492677152157, + 0.01402864046394825, + 0.012590662576258183, + 0.0025393995456397533, + 0.002699335105717182, + 0.0017626561457291245, + 0.0020257022697478533, + 0.011075105518102646, + 0.009693101979792118, + 0.0020545439328998327, + 0.0018942130263894796, + 0.0029178671538829803, + 0.00316817918792367, + 0.0025784007739275694, + 0.001575165311805904, + 0.005765870213508606, + 0.0031729908660054207, + 0.0019680033437907696, + 0.002150188200175762, + 0.002497120061889291, + 0.004620997700840235, + 0.004447644576430321, + 0.008381719700992107, + 0.004642426501959562, + 0.005948810372501612, + 0.004104687366634607, + 0.003726004622876644, + 0.04320789873600006, + 0.014852015301585197, + 0.06767717748880386, + 0.016524937003850937, + 0.014314146712422371, + 0.0166827030479908, + 0.020598741248250008, + 0.01776457577943802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.005951895844191313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11679727584123611, + 0.1094355657696724, + 0.3266332745552063, + 0.12953463196754456, + 0.05163627117872238, + 0.059862930327653885, + 0.08194387704133987, + 0.053944818675518036, + 0.03274695947766304, + 0.037464309483766556 + ], + "credit_map": [ + 0.011042183265089989, + 0.014412193559110165, + 0.012938897125422955, + 0.014502233825623989, + 0.015917005017399788, + 0.012899111956357956, + 0.02531270496547222, + 0.00977296568453312, + 0.03057711385190487, + 0.00963553786277771, + 0.008835751563310623, + 0.008969702757894993, + 0.009932615794241428, + 0.012369410134851933, + 0.013689697720110416, + 0.015425747260451317, + 0.08477143198251724, + 0.028226438909769058, + 0.03280215337872505, + 0.029105037450790405, + 0.03327290341258049, + 0.03430211916565895, + 0.04074223339557648, + 0.04341098293662071, + 0.020852219313383102, + 0.02171982266008854, + 0.011769196949899197, + 0.00938196200877428, + 0.016729295253753662, + 0.01759953238070011, + 0.012497182004153728, + 0.018985427916049957, + 0.00448582274839282, + 0.00666684377938509, + 0.004160280339419842, + 0.010301207192242146, + 0.008719632402062416, + 0.00679053645581007, + 0.006201804615557194, + 0.004310146439820528, + 0.00926969014108181, + 0.008858616463840008, + 0.0060317181050777435, + 0.006949863862246275, + 0.005663028918206692, + 0.006444515194743872, + 0.009116030298173428, + 0.007529473397880793, + 0.005907965358346701, + 0.008150402456521988, + 0.0054784612730145454, + 0.006138704717159271, + 0.027109086513519287, + 0.015684587880969048, + 0.008566725067794323, + 0.004907949361950159, + 0.005317214410752058, + 0.0057001239620149136, + 0.0049910820089280605, + 0.003353806911036372, + 0.014569251798093319, + 0.007413138169795275, + 0.005300379358232021, + 0.007299831137061119, + 0.003264095401391387, + 0.0038437116891145706, + 0.0032060248777270317, + 0.005532660521566868, + 0.005040274001657963, + 0.004557689186185598, + 0.0030389127787202597, + 0.004263595212250948, + 0.006563511677086353, + 0.0026804725639522076, + 0.01255561038851738, + 0.0028948995750397444, + 0.0025404903572052717, + 0.0030623669736087322, + 0.003138677217066288, + 0.004028284456580877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.006735715549439192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level close-up captures a crystal-clear image of a downy duck adorably struggling in the calm waters of a sun-dappled pond. Its downy feathers, a vibrant lemon yellow, are rendered in surreal detail, with tufts of feathers slightly tangled in the water. Its beady black eyes widen with earnest concentration as it attempts to swim. Its movements are a series of clumsy micro-movements: it kicks its tiny webbed feet in a frantic, uncoordinated pattern, causing its entire body to sway precariously. Then, it stumbles through the water, pitching heavily to one side, flapping its crude, undeveloped wings in a futile, unbalanced attempt to steady itself before tentatively regaining its footing. The entire scene is illuminated by the warm golden glow of the afternoon sun, which filters through unseen trees, creating shimmering caustics on the water's surface and a soft, glowing rimlight on the duckling's body. The soft-focus blur of deep green water lilies and dark, reflective water in the background creates a tender and endearing atmosphere, focusing entirely on the young creature's innocent yet comical efforts.", + "seed": 1913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_10/index_7886_seed_1913_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1548560786060989, + "vlm_logprob": -0.1548560786060989, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1548560786060989, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31528928875923157 + ], + "opposite_logprob_tokens": [ + -1.3152892589569092 + ], + "logprob_reward": -0.31528928875923157, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31528928875923157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18526846170425415, + 0.09094876796007156, + 0.2085290253162384, + 0.36750397086143494, + 0.020569268614053726, + 0.04709074646234512, + 0.036912862211465836, + 0.018151670694351196, + 0.011213340796530247, + 0.013811920769512653 + ], + "credit_map": [ + 0.01016428042203188, + 0.008096366189420223, + 0.012581760063767433, + 0.00482126185670495, + 0.013517390936613083, + 0.07217951864004135, + 0.00926404818892479, + 0.05464382469654083, + 0.014868004247546196, + 0.009397684596478939, + 0.01771044172346592, + 0.007880710996687412, + 0.007272700779139996, + 0.016324281692504883, + 0.009598936885595322, + 0.0078959995880723, + 0.058428216725587845, + 0.01415778137743473, + 0.07194118201732635, + 0.025302046909928322, + 0.012233447283506393, + 0.008673561736941338, + 0.008571700192987919, + 0.009221096523106098, + 0.04499300569295883, + 0.03267890587449074, + 0.037810686975717545, + 0.06595594435930252, + 0.04106864705681801, + 0.06158875301480293, + 0.04603007808327675, + 0.03737795352935791, + 0.002779558999463916, + 0.0024967873468995094, + 0.0025986225809901953, + 0.003821670077741146, + 0.002575754886493087, + 0.0016489740228280425, + 0.0022690703626722097, + 0.0023788316175341606, + 0.006742460187524557, + 0.004478281829506159, + 0.00477983895689249, + 0.005694106221199036, + 0.005341945216059685, + 0.008269251324236393, + 0.005101450718939304, + 0.006683411076664925, + 0.00578808831050992, + 0.005214898381382227, + 0.003449514275416732, + 0.004938443191349506, + 0.00501737417653203, + 0.004040484316647053, + 0.004552144557237625, + 0.003911916632205248, + 0.0025466810911893845, + 0.0024983759503811598, + 0.0018652691505849361, + 0.0025058516766875982, + 0.0020327018573880196, + 0.0017753083957359195, + 0.0020632222294807434, + 0.002864259760826826, + 0.0013632558984681964, + 0.0014136971440166235, + 0.001181318424642086, + 0.0017502015689387918, + 0.00149646180216223, + 0.0014426226262003183, + 0.0015206836396828294, + 0.001045098528265953, + 0.0015052059898152947, + 0.0022839894518256187, + 0.0014877932844683528, + 0.0015624025836586952, + 0.0014027468860149384, + 0.001523685990832746, + 0.0023213280364871025, + 0.0017247684299945831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.31528928875923157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706183433532715 + ], + "opposite_logprob_tokens": [ + -1.1370618343353271 + ], + "logprob_reward": -0.38706183433532715, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38706183433532715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16052667796611786, + 0.18799470365047455, + 0.09111355245113373, + 0.15814749896526337, + 0.0627024695277214, + 0.07910425961017609, + 0.12729844450950623, + 0.03996705263853073, + 0.025269998237490654, + 0.06787537038326263 + ], + "credit_map": [ + 0.012024903669953346, + 0.010547990910708904, + 0.011880127713084221, + 0.004593969788402319, + 0.015356822870671749, + 0.06251107901334763, + 0.009213663637638092, + 0.03439812734723091, + 0.036379601806402206, + 0.02144445851445198, + 0.044922422617673874, + 0.015391555614769459, + 0.014874013140797615, + 0.02192450687289238, + 0.017145143821835518, + 0.01591300033032894, + 0.016023455187678337, + 0.010855010710656643, + 0.017966464161872864, + 0.00991547480225563, + 0.009095278568565845, + 0.009145543910562992, + 0.007389359641820192, + 0.010722965002059937, + 0.020211640745401382, + 0.015172786079347134, + 0.01556890644133091, + 0.020767470821738243, + 0.017238305881619453, + 0.022081982344388962, + 0.02652394399046898, + 0.020582472905516624, + 0.0076141320168972015, + 0.00510043790563941, + 0.006482969969511032, + 0.010885670781135559, + 0.006996530573815107, + 0.012194957584142685, + 0.00595754524692893, + 0.007470228709280491, + 0.012740439735352993, + 0.00949096865952015, + 0.0069203549064695835, + 0.008215894922614098, + 0.0059937965124845505, + 0.016908427700400352, + 0.0069885822013020515, + 0.011845793575048447, + 0.020950274541974068, + 0.01895180158317089, + 0.013216209597885609, + 0.016122356057167053, + 0.017174633219838142, + 0.013279025442898273, + 0.01441803015768528, + 0.013186128810048103, + 0.005154102575033903, + 0.003533882088959217, + 0.005027786362916231, + 0.004222316201776266, + 0.0056135947816073895, + 0.00655247550457716, + 0.0038034419994801283, + 0.006059451960027218, + 0.003283239435404539, + 0.0031702660489827394, + 0.004366055130958557, + 0.002872144803404808, + 0.0034036184661090374, + 0.003550751367583871, + 0.0024963116738945246, + 0.0021276138722896576, + 0.0069657075218856335, + 0.0041832514107227325, + 0.01589890755712986, + 0.009896324016153812, + 0.0037024246994405985, + 0.004506327677518129, + 0.00728378351777792, + 0.015438640490174294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.38706183433532715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045034021139145 + ], + "opposite_logprob_tokens": [ + -2.812045097351074 + ], + "logprob_reward": -0.062045034021139145, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045034021139145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11599241942167282, + 0.1841220259666443, + 0.18729376792907715, + 0.12216175347566605, + 0.08111609518527985, + 0.06118978187441826, + 0.10264551639556885, + 0.04454103857278824, + 0.02930454909801483, + 0.07163305580615997 + ], + "credit_map": [ + 0.010229743085801601, + 0.00904487632215023, + 0.010809475556015968, + 0.006078708916902542, + 0.013119415380060673, + 0.017490282654762268, + 0.006005220580846071, + 0.043214693665504456, + 0.03553494065999985, + 0.019693052396178246, + 0.04323337972164154, + 0.016427138820290565, + 0.014681937173008919, + 0.023316094651818275, + 0.017368651926517487, + 0.013866831548511982, + 0.05044014751911163, + 0.013192686252295971, + 0.06387213617563248, + 0.02255970425903797, + 0.011189168319106102, + 0.008455227129161358, + 0.00658064940944314, + 0.011004047468304634, + 0.015926336869597435, + 0.009623739868402481, + 0.011470536701381207, + 0.01593334786593914, + 0.01307039987295866, + 0.01872124895453453, + 0.01867547817528248, + 0.018740663304924965, + 0.010364885441958904, + 0.006741201505064964, + 0.008838573470711708, + 0.013899678364396095, + 0.00915187131613493, + 0.015574807301163673, + 0.007865064777433872, + 0.00868001114577055, + 0.00924256443977356, + 0.006201025098562241, + 0.005254592280834913, + 0.0075286636129021645, + 0.005213236436247826, + 0.011391496285796165, + 0.006736982613801956, + 0.009621222503483295, + 0.016971657052636147, + 0.015164869837462902, + 0.010460489429533482, + 0.013847670517861843, + 0.01353354100137949, + 0.010477634146809578, + 0.012130741961300373, + 0.010058910585939884, + 0.005963359493762255, + 0.004362402018159628, + 0.005785305984318256, + 0.00478245597332716, + 0.00495131267234683, + 0.007154215592890978, + 0.004501426592469215, + 0.007040560245513916, + 0.004541458562016487, + 0.003595172194764018, + 0.004473038949072361, + 0.003976809326559305, + 0.0032284853514283895, + 0.003080469323322177, + 0.0036943708546459675, + 0.002714745234698057, + 0.007350622210651636, + 0.005972839891910553, + 0.01570422574877739, + 0.010125518776476383, + 0.004052961710840464, + 0.004706692416220903, + 0.00830453634262085, + 0.015415665693581104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.062045034021139145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046355403028428555 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.0046355403028428555, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046355403028428555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23026129603385925, + 0.2767674922943115, + 0.2462908774614334, + 0.06994982063770294, + 0.06188047304749489, + 0.023687241598963737, + 0.014695161022245884, + 0.022839583456516266, + 0.03209234029054642, + 0.02153571881353855 + ], + "credit_map": [ + 0.014209696091711521, + 0.015456623397767544, + 0.01568736881017685, + 0.00949503667652607, + 0.01975221373140812, + 0.08406852930784225, + 0.012740357778966427, + 0.0588514544069767, + 0.06261183321475983, + 0.02607862278819084, + 0.0796528086066246, + 0.0217263363301754, + 0.023397047072649002, + 0.019952967762947083, + 0.018463583663105965, + 0.02488427609205246, + 0.06601767987012863, + 0.01909451186656952, + 0.07987039536237717, + 0.03138304501771927, + 0.016377586871385574, + 0.01013494934886694, + 0.009674333035945892, + 0.013738405890762806, + 0.00928872637450695, + 0.009827779605984688, + 0.006943027023226023, + 0.01097817998379469, + 0.006969778332859278, + 0.00834619253873825, + 0.00889833178371191, + 0.008697808720171452, + 0.008455660194158554, + 0.00836235098540783, + 0.006922292057424784, + 0.011183935217559338, + 0.007054684218019247, + 0.004592869430780411, + 0.007025988306850195, + 0.008282688446342945, + 0.00331555656157434, + 0.003448061877861619, + 0.0021634805016219616, + 0.0034122902434319258, + 0.002746134763583541, + 0.002348849084228277, + 0.002790368627756834, + 0.0034624978434294462, + 0.0022635438945144415, + 0.001795757794752717, + 0.001630865503102541, + 0.0015932191163301468, + 0.001771394396200776, + 0.0017606255132704973, + 0.0018068902427330613, + 0.002072864444926381, + 0.0037838800344616175, + 0.0029917252250015736, + 0.0023143740836530924, + 0.0025455858558416367, + 0.0033805673010647297, + 0.0031649963930249214, + 0.0018751409370452166, + 0.002783314324915409, + 0.005762248300015926, + 0.0048609282821416855, + 0.003380686044692993, + 0.003509261878207326, + 0.0037443635519593954, + 0.004969393834471703, + 0.0027919833082705736, + 0.003073475556448102, + 0.003140344750136137, + 0.003728630021214485, + 0.0020517779048532248, + 0.0025085839442908764, + 0.0028665070421993732, + 0.001989656360819936, + 0.0026459423825144768, + 0.0026042775716632605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0046355403028428555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a downy duck with vibrant lemon yellow feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the duck performing clumsy micro-movements such as kicking tiny webbed feet and swaying precariously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the duck flap crude, undeveloped wings in an unbalanced attempt to steady itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow creating shimmering caustics on the water and a soft rimlight on the duck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a static, eye-level close-up with a soft-focus background of deep green water lilies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248695611953735 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248695611953735, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248695611953735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28587502241134644, + 0.10229399800300598, + 0.1623649150133133, + 0.07817865908145905, + 0.04096594452857971, + 0.054396625608205795, + 0.07632369548082352, + 0.022040890529751778, + 0.057349175214767456, + 0.12021111696958542 + ], + "credit_map": [ + 0.03284190967679024, + 0.017954953014850616, + 0.02013700269162655, + 0.009604083374142647, + 0.03742118552327156, + 0.07857833057641983, + 0.01604868285357952, + 0.07328885048627853, + 0.018573418259620667, + 0.010297242552042007, + 0.02393462136387825, + 0.009915693663060665, + 0.009646633639931679, + 0.009427147917449474, + 0.00845517497509718, + 0.012044064700603485, + 0.03784165531396866, + 0.013912474736571312, + 0.042924802750349045, + 0.01963784359395504, + 0.014675994403660297, + 0.010525439865887165, + 0.010504164732992649, + 0.012342534959316254, + 0.012013341300189495, + 0.016753600910305977, + 0.0061272201128304005, + 0.006802581250667572, + 0.007598778232932091, + 0.007085044402629137, + 0.010876989923417568, + 0.010921099223196507, + 0.006501382682472467, + 0.0038517273496836424, + 0.003239405807107687, + 0.005468897055834532, + 0.004319286439567804, + 0.006938425358384848, + 0.004926696419715881, + 0.005720121785998344, + 0.007481906563043594, + 0.006931686773896217, + 0.005098940338939428, + 0.006723223719745874, + 0.005860887933522463, + 0.009923756122589111, + 0.005137793254107237, + 0.007238424848765135, + 0.01135321892797947, + 0.011443502269685268, + 0.008431429974734783, + 0.011072876863181591, + 0.012303069233894348, + 0.0047061024233698845, + 0.00992988795042038, + 0.007083605043590069, + 0.002896767109632492, + 0.002104054670780897, + 0.0025615966878831387, + 0.002848928328603506, + 0.003348186146467924, + 0.0028647028375416994, + 0.002498861402273178, + 0.0029177931137382984, + 0.0074748629704117775, + 0.007778565865010023, + 0.005909564904868603, + 0.00819795485585928, + 0.008370894007384777, + 0.008634300902485847, + 0.0055391788482666016, + 0.0054438551887869835, + 0.013129997067153454, + 0.008217066526412964, + 0.02862778678536415, + 0.018236270174384117, + 0.005108767654746771, + 0.0076833926141262054, + 0.011336494237184525, + 0.02787134051322937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.005248695611953735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_11/index_tensor([7039, 7039])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_11/index_tensor([7039, 7039])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a7bdc6a130a8be9a7f7bcbb0e7d003757f7d3565 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_11/index_tensor([7039, 7039])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely slow-motion macro close-up, captured from a slightly overhead perspective, focuses on a pair of slender, white hands carefully cradling a creamy white, pebble-shaped electric hand warmer. The hands, with their neatly manicured nails and gentle gestures, seem to be protecting a treasure. The hand warmer's smooth, rounded, matte finish reflects a soft, warm glow under the warm yellow light of the cafe's vintage tungsten lamps, which cast a sideways cast. As the camera zooms in at an almost imperceptible pace, the heat radiating from the hand warmer's surface becomes concrete in the extreme slow motion: the air around it undergoes a visible, transparent distortion, forming gentle ripples that slowly and gracefully rise and dissipate, like heat waves on an asphalt road on a summer afternoon. The background is rendered extremely blurred by a shallow depth of field, barely discernible enough to distinguish the texture of the dark wood tabletop and the curl of steam from a latte in the distance. Together, they create an intimate atmosphere of profound tranquility, warmth, healing, and exquisite detail.", + "seed": 23289, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_11/index_7039_seed_23289_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06784196132794022, + "vlm_logprob": -0.06784196132794022, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06784196132794022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029597083106637 + ], + "opposite_logprob_tokens": [ + -3.654597043991089 + ], + "logprob_reward": -0.029597083106637, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029597083106637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15453456342220306, + 0.17993950843811035, + 0.08486895263195038, + 0.06455668061971664, + 0.05498379468917847, + 0.11999108642339706, + 0.10719732940196991, + 0.06124482676386833, + 0.1132858544588089, + 0.05939742922782898 + ], + "credit_map": [ + 0.016180314123630524, + 0.021063970401883125, + 0.02133338525891304, + 0.019980311393737793, + 0.01758619025349617, + 0.021651143208146095, + 0.020482636988162994, + 0.016256608068943024, + 0.013801496475934982, + 0.0207817479968071, + 0.01551542803645134, + 0.03749668598175049, + 0.018417878076434135, + 0.024067867547273636, + 0.03284164518117905, + 0.017016755416989326, + 0.00890160072594881, + 0.007757933344691992, + 0.012465249747037888, + 0.01939500868320465, + 0.006713668815791607, + 0.011168200522661209, + 0.008618899621069431, + 0.009848390705883503, + 0.011086692102253437, + 0.006598584819585085, + 0.007045659702271223, + 0.010430851951241493, + 0.006414864677935839, + 0.008906832896173, + 0.0076517206616699696, + 0.0064214738085865974, + 0.005271569825708866, + 0.005958696361631155, + 0.008953115902841091, + 0.008741962723433971, + 0.005014374852180481, + 0.009705725125968456, + 0.005873560905456543, + 0.005464787129312754, + 0.008535774424672127, + 0.011802591383457184, + 0.017719022929668427, + 0.024075593799352646, + 0.015280726365745068, + 0.016267715021967888, + 0.011936136521399021, + 0.014373531565070152, + 0.015390380285680294, + 0.015776406973600388, + 0.017259862273931503, + 0.01778811402618885, + 0.009997467510402203, + 0.00725487619638443, + 0.011354451067745686, + 0.012375764548778534, + 0.006801902316510677, + 0.007572336588054895, + 0.006769063416868448, + 0.009248350746929646, + 0.006852766498923302, + 0.008712337352335453, + 0.007128075230866671, + 0.008159998804330826, + 0.015062595717608929, + 0.013436686247587204, + 0.017749439924955368, + 0.016069814562797546, + 0.013587896712124348, + 0.009826666675508022, + 0.014817326329648495, + 0.012735427357256413, + 0.0054269409738481045, + 0.005512910429388285, + 0.008865362964570522, + 0.007117432542145252, + 0.007996829226613045, + 0.008114484138786793, + 0.008561530150473118, + 0.0078019374050199986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.029597083106637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804308414459229 + ], + "opposite_logprob_tokens": [ + -4.6348042488098145 + ], + "logprob_reward": -0.009804308414459229, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804308414459229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42854925990104675, + 0.14561811089515686, + 0.038448113948106766, + 0.029540449380874634, + 0.04637369513511658, + 0.051771026104688644, + 0.08164869248867035, + 0.10599970072507858, + 0.04182419180870056, + 0.030226731672883034 + ], + "credit_map": [ + 0.027318935841321945, + 0.022777972742915154, + 0.02388397790491581, + 0.024243751540780067, + 0.1348278522491455, + 0.04117993637919426, + 0.12962563335895538, + 0.024691199883818626, + 0.013178205117583275, + 0.017837928608059883, + 0.02227919176220894, + 0.018931597471237183, + 0.015668362379074097, + 0.017537515610456467, + 0.02968311496078968, + 0.010502197779715061, + 0.004020003601908684, + 0.004069188144057989, + 0.005347910802811384, + 0.006934762932360172, + 0.00370948389172554, + 0.007438320200890303, + 0.0030193396378308535, + 0.0039091045036911964, + 0.0033669148106127977, + 0.003354326356202364, + 0.004185912199318409, + 0.005041891243308783, + 0.00228477967903018, + 0.00509473355486989, + 0.003507564775645733, + 0.0027043248992413282, + 0.0034060014877468348, + 0.004355728626251221, + 0.006005587987601757, + 0.007852938026189804, + 0.004121560603380203, + 0.011562293395400047, + 0.0046965694054961205, + 0.004373013973236084, + 0.0032989780884236097, + 0.0057184454053640366, + 0.006900602485984564, + 0.011394575238227844, + 0.004504783544689417, + 0.00608446542173624, + 0.006521040108054876, + 0.007348136976361275, + 0.01110929623246193, + 0.009596984833478928, + 0.015775520354509354, + 0.011543815955519676, + 0.007124221418052912, + 0.007975456304848194, + 0.008629340678453445, + 0.009894052520394325, + 0.012967634946107864, + 0.015270796604454517, + 0.013438688591122627, + 0.01563606970012188, + 0.011517917737364769, + 0.010640251450240612, + 0.012558852322399616, + 0.013969489373266697, + 0.003449114039540291, + 0.004974879790097475, + 0.004425130784511566, + 0.01059030368924141, + 0.0032350528053939342, + 0.00293723470531404, + 0.006117631681263447, + 0.006094844546169043, + 0.00412754574790597, + 0.003009927226230502, + 0.004161507356911898, + 0.0050122844986617565, + 0.0026269841473549604, + 0.0035691976081579924, + 0.004024534486234188, + 0.0036947503685951233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.009804308414459229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132250905036926 + ], + "opposite_logprob_tokens": [ + -1.4063224792480469 + ], + "logprob_reward": -0.28132250905036926, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132250905036926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.228829488158226, + 0.36738529801368713, + 0.09745454043149948, + 0.032312504947185516, + 0.08893996477127075, + 0.02504408359527588, + 0.022801442071795464, + 0.07352454960346222, + 0.034934017807245255, + 0.028774095699191093 + ], + "credit_map": [ + 0.016216842457652092, + 0.014706728979945183, + 0.014270681887865067, + 0.01800990104675293, + 0.06364746391773224, + 0.02573332004249096, + 0.06361091136932373, + 0.01263363752514124, + 0.03449447453022003, + 0.03804314509034157, + 0.05311504378914833, + 0.04875311627984047, + 0.04087945073843002, + 0.04229259118437767, + 0.08277236670255661, + 0.027035105973482132, + 0.010063148103654385, + 0.00941853690892458, + 0.013662269338965416, + 0.01841835491359234, + 0.0078450171276927, + 0.013742532581090927, + 0.014504935592412949, + 0.009799739345908165, + 0.005292060319334269, + 0.004111848771572113, + 0.004701429978013039, + 0.00502549996599555, + 0.002614987548440695, + 0.005064336583018303, + 0.0030272607691586018, + 0.0024750777520239353, + 0.0070971217937767506, + 0.012820719741284847, + 0.009882654063403606, + 0.01379877608269453, + 0.009668519720435143, + 0.019382137805223465, + 0.008460032753646374, + 0.00782999861985445, + 0.00159369723405689, + 0.0024480426218360662, + 0.002681718673557043, + 0.005766912829130888, + 0.0028511390555649996, + 0.002695167437195778, + 0.003325555706396699, + 0.0036818478256464005, + 0.0024736972991377115, + 0.0035706646740436554, + 0.002804154995828867, + 0.0035929065197706223, + 0.00236130109988153, + 0.0027857439126819372, + 0.002557787811383605, + 0.002655183197930455, + 0.007091951090842485, + 0.01074803527444601, + 0.009392213076353073, + 0.012250667437911034, + 0.008354997262358665, + 0.005095057655125856, + 0.009423200972378254, + 0.011168422177433968, + 0.004284520633518696, + 0.004254275932908058, + 0.003994774539023638, + 0.004854065831750631, + 0.0033092519734054804, + 0.005543975159525871, + 0.00456280680373311, + 0.0041303434409201145, + 0.0033415076322853565, + 0.002515004249289632, + 0.0031755464151501656, + 0.0047051990404725075, + 0.003081343602389097, + 0.003413344733417034, + 0.004177651833742857, + 0.0043644956313073635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.28132250905036926 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541357427835464 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541357427835464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541357427835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3605625331401825, + 0.13069702684879303, + 0.038960546255111694, + 0.039890386164188385, + 0.08480555564165115, + 0.14518573880195618, + 0.05950881540775299, + 0.03669373691082001, + 0.07601776719093323, + 0.02767792157828808 + ], + "credit_map": [ + 0.019208837300539017, + 0.020984778180718422, + 0.022327682003378868, + 0.017498420551419258, + 0.12068536877632141, + 0.029419347643852234, + 0.10620251297950745, + 0.02423560805618763, + 0.00925361830741167, + 0.0227767713367939, + 0.01091746799647808, + 0.01582232490181923, + 0.015029958449304104, + 0.01932467892765999, + 0.02550949528813362, + 0.012062711641192436, + 0.006076621823012829, + 0.004564336035400629, + 0.0042502386495471, + 0.00561829749494791, + 0.003662211587652564, + 0.006030445918440819, + 0.005667198449373245, + 0.0030911974608898163, + 0.0065689110197126865, + 0.004388553090393543, + 0.005838763900101185, + 0.0054914518259465694, + 0.0031011137180030346, + 0.0061734504997730255, + 0.005486570298671722, + 0.0028415729757398367, + 0.0065003070048987865, + 0.00859123095870018, + 0.011266602203249931, + 0.013174563646316528, + 0.008428344503045082, + 0.021269118413329124, + 0.0081162229180336, + 0.007459165528416634, + 0.013345842249691486, + 0.018579646944999695, + 0.021718749776482582, + 0.02127853035926819, + 0.018697841092944145, + 0.01600581966340542, + 0.01878942735493183, + 0.016769878566265106, + 0.005012117326259613, + 0.0068712919019162655, + 0.008372802287340164, + 0.010804315097630024, + 0.00604779040440917, + 0.007153212558478117, + 0.006695958320051432, + 0.008551329374313354, + 0.004312518518418074, + 0.004635212942957878, + 0.004778121132403612, + 0.004818249959498644, + 0.002841674955561757, + 0.006443814840167761, + 0.004411232657730579, + 0.004452911671251059, + 0.008060688152909279, + 0.009472592733800411, + 0.008005678653717041, + 0.012047821655869484, + 0.007293489295989275, + 0.01107637956738472, + 0.010676192119717598, + 0.009384920820593834, + 0.0029433195013552904, + 0.0027209233958274126, + 0.004061087965965271, + 0.0032302618492394686, + 0.0027692047879099846, + 0.005493388045579195, + 0.0036265109665691853, + 0.002833223668858409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.012541357427835464 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944548640400171 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944548640400171, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944548640400171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2450893074274063, + 0.3444780707359314, + 0.08610985428094864, + 0.0347139947116375, + 0.09101385623216629, + 0.03986397385597229, + 0.044851936399936676, + 0.04213101789355278, + 0.031421128660440445, + 0.040326934307813644 + ], + "credit_map": [ + 0.030863720923662186, + 0.023354899138212204, + 0.026003394275903702, + 0.02256445586681366, + 0.04471391439437866, + 0.029315296560525894, + 0.04968171566724777, + 0.018591884523630142, + 0.037620894610881805, + 0.042475052177906036, + 0.065921850502491, + 0.043261539191007614, + 0.035860612988471985, + 0.030808234587311745, + 0.06445187330245972, + 0.02407803200185299, + 0.009575542993843555, + 0.011272173374891281, + 0.00900748185813427, + 0.010194688104093075, + 0.006901451852172613, + 0.018640674650669098, + 0.013623801060020924, + 0.006894041784107685, + 0.004937571939080954, + 0.004083202686160803, + 0.003666725941002369, + 0.004727661609649658, + 0.0022121001966297626, + 0.006988764740526676, + 0.005622274242341518, + 0.002475692890584469, + 0.007810617331415415, + 0.014714160934090614, + 0.009662505239248276, + 0.012246858328580856, + 0.010549233295023441, + 0.020104162395000458, + 0.00848165899515152, + 0.00744465971365571, + 0.0027436057571321726, + 0.003966364543884993, + 0.005075313616544008, + 0.009337887167930603, + 0.0035832442808896303, + 0.004395774099975824, + 0.004828306380659342, + 0.005933479405939579, + 0.005444920156151056, + 0.004789477214217186, + 0.007119572255760431, + 0.006599386688321829, + 0.004151672590523958, + 0.004744516219943762, + 0.005174112971872091, + 0.006828277371823788, + 0.005781661719083786, + 0.005555742885917425, + 0.005351930391043425, + 0.005238022189587355, + 0.003849061904475093, + 0.006374651100486517, + 0.004981586709618568, + 0.0049983602948486805, + 0.003483015578240156, + 0.003239220008254051, + 0.005636240355670452, + 0.004836762323975563, + 0.0035530973691493273, + 0.0031770160421729088, + 0.004138060379773378, + 0.003357715904712677, + 0.004757765214890242, + 0.005517615005373955, + 0.007541115861386061, + 0.004526033997535706, + 0.0030084485188126564, + 0.0058840117417275906, + 0.004584863781929016, + 0.004507079720497131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.005944548640400171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely slow-motion macro close-up, captured from a slightly overhead perspective, focuses on a pair of slender, white hands carefully cradling a creamy white, pebble-shaped electric hand warmer. The hands, with their neatly manicured nails and gentle gestures, seem to be protecting a treasure. The hand warmer's smooth, rounded, matte finish reflects a soft, warm glow under the warm yellow light of the cafe's vintage tungsten lamps, which cast a sideways cast. As the camera zooms in at an almost imperceptible pace, the heat radiating from the hand warmer's surface becomes concrete in the extreme slow motion: the air around it undergoes a visible, transparent distortion, forming gentle ripples that slowly and gracefully rise and dissipate, like heat waves on an asphalt road on a summer afternoon. The background is rendered extremely blurred by a shallow depth of field, barely discernible enough to distinguish the texture of the dark wood tabletop and the curl of steam from a latte in the distance. Together, they create an intimate atmosphere of profound tranquility, warmth, healing, and exquisite detail.", + "seed": 23290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_11/index_7039_seed_23290_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027375034894794224, + "vlm_logprob": -0.027375034894794224, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.027375034894794224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008212133310735226 + ], + "opposite_logprob_tokens": [ + -5.508212089538574 + ], + "logprob_reward": -0.008212133310735226, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008212133310735226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09053785353899002, + 0.13540196418762207, + 0.03813911974430084, + 0.029228582978248596, + 0.13990791141986847, + 0.19723206758499146, + 0.045688457787036896, + 0.12103646993637085, + 0.038105934858322144, + 0.16472162306308746 + ], + "credit_map": [ + 0.008645898662507534, + 0.00880629476159811, + 0.011031257919967175, + 0.017976822331547737, + 0.011122262105345726, + 0.007437604945152998, + 0.01477782428264618, + 0.010739892721176147, + 0.00692404992878437, + 0.005469241179525852, + 0.04562000930309296, + 0.006488977000117302, + 0.007491504307836294, + 0.03424205631017685, + 0.007025789935141802, + 0.02214032970368862, + 0.00364419212564826, + 0.0034192027524113655, + 0.0035932764876633883, + 0.007699436042457819, + 0.0037028100341558456, + 0.004057074896991253, + 0.00626886822283268, + 0.005754255689680576, + 0.0025378752034157515, + 0.001665456686168909, + 0.0021491728257387877, + 0.0031038704328238964, + 0.0057760681957006454, + 0.0037464264314621687, + 0.006110836751759052, + 0.0041388776153326035, + 0.016622178256511688, + 0.028196195140480995, + 0.01494209561496973, + 0.018834972754120827, + 0.01607055775821209, + 0.012029089033603668, + 0.01789422146975994, + 0.015318604186177254, + 0.006893927697092295, + 0.01219113264232874, + 0.00805213674902916, + 0.00810807105153799, + 0.07715465873479843, + 0.029405344277620316, + 0.03943667560815811, + 0.0159901212900877, + 0.002516517648473382, + 0.006413084454834461, + 0.004185410216450691, + 0.007459230720996857, + 0.006613445933908224, + 0.006335696205496788, + 0.007127297576516867, + 0.005037776660174131, + 0.005156404804438353, + 0.005786601919680834, + 0.006846047937870026, + 0.006283482536673546, + 0.013967045582830906, + 0.04511899873614311, + 0.030381981283426285, + 0.007495901547372341, + 0.0033093986567109823, + 0.0045973253436386585, + 0.007361856754869223, + 0.00715667987242341, + 0.003334594890475273, + 0.004735218361020088, + 0.0036424417048692703, + 0.003968419972807169, + 0.010729743167757988, + 0.017822610214352608, + 0.019440392032265663, + 0.03327717259526253, + 0.01527878176420927, + 0.024504326283931732, + 0.0240933820605278, + 0.01957521215081215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.008212133310735226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609804667532444 + ], + "opposite_logprob_tokens": [ + -4.1410980224609375 + ], + "logprob_reward": -0.01609804667532444, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609804667532444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06502638012170792, + 0.047724299132823944, + 0.03387445583939552, + 0.029100405052304268, + 0.08786569535732269, + 0.43610912561416626, + 0.10021073371171951, + 0.05948866903781891, + 0.040688153356313705, + 0.0999121367931366 + ], + "credit_map": [ + 0.006194295361638069, + 0.0056289248168468475, + 0.006617274135351181, + 0.01623343490064144, + 0.006270142737776041, + 0.005898232571780682, + 0.009777857922017574, + 0.008406221866607666, + 0.002313855802640319, + 0.003835899755358696, + 0.013105093501508236, + 0.0034615874756127596, + 0.004228185396641493, + 0.014995237812399864, + 0.002216269262135029, + 0.0035681710578501225, + 0.0027980480808764696, + 0.0027466528117656708, + 0.003450856078416109, + 0.007444456219673157, + 0.0024604229256510735, + 0.004117299802601337, + 0.006369602866470814, + 0.004487118683755398, + 0.002193571301177144, + 0.0022345802281051874, + 0.004078099969774485, + 0.004656375851482153, + 0.008277229964733124, + 0.002476668683812022, + 0.0032719317823648453, + 0.0019119458738714457, + 0.009766703471541405, + 0.016541756689548492, + 0.009461037814617157, + 0.012057837098836899, + 0.011356158182024956, + 0.007895718328654766, + 0.011627528816461563, + 0.009158956818282604, + 0.013187795877456665, + 0.025409525260329247, + 0.016981802880764008, + 0.015457836911082268, + 0.17071034014225006, + 0.06240520626306534, + 0.10332557559013367, + 0.028631024062633514, + 0.005485640838742256, + 0.009290114045143127, + 0.005664162803441286, + 0.009214822202920914, + 0.011450069025158882, + 0.019631460309028625, + 0.028041791170835495, + 0.011432671919465065, + 0.004190520383417606, + 0.007123173214495182, + 0.007821859791874886, + 0.007456284482032061, + 0.009985151700675488, + 0.008868849836289883, + 0.007775884587317705, + 0.006266945973038673, + 0.0030672820284962654, + 0.004502945113927126, + 0.0062256306409835815, + 0.010209618136286736, + 0.002721636788919568, + 0.0038803499191999435, + 0.00478568859398365, + 0.0052950019016861916, + 0.008152661845088005, + 0.017289256677031517, + 0.008482451550662518, + 0.014929212629795074, + 0.015545009635388851, + 0.010879223234951496, + 0.01473873108625412, + 0.009895588271319866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.01609804667532444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029932111501694 + ], + "opposite_logprob_tokens": [ + -2.35029935836792 + ], + "logprob_reward": -0.10029932111501694, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029932111501694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19810593128204346, + 0.1270519644021988, + 0.05283663794398308, + 0.06192294508218765, + 0.1531287580728531, + 0.2191472351551056, + 0.03214092180132866, + 0.06753427535295486, + 0.055450208485126495, + 0.0326811708509922 + ], + "credit_map": [ + 0.02007489651441574, + 0.017921607941389084, + 0.021344993263483047, + 0.0501333624124527, + 0.01779894158244133, + 0.01773458905518055, + 0.02871652878820896, + 0.024380998685956, + 0.005859746132045984, + 0.0074988133274018764, + 0.03890747204422951, + 0.006573787424713373, + 0.008712394163012505, + 0.05120163783431053, + 0.0032393136061728, + 0.005058795213699341, + 0.006275390274822712, + 0.004278842359781265, + 0.005986566189676523, + 0.00876399502158165, + 0.006018474698066711, + 0.008345270529389381, + 0.007828593254089355, + 0.005339503288269043, + 0.005240716505795717, + 0.0036559284199029207, + 0.005741047207266092, + 0.006894788239151239, + 0.012402021326124668, + 0.01657569780945778, + 0.0043773408979177475, + 0.007035402115434408, + 0.013285430148243904, + 0.023089461028575897, + 0.015306109562516212, + 0.021475166082382202, + 0.025829825550317764, + 0.016688484698534012, + 0.021772947162389755, + 0.015681330114603043, + 0.00984940119087696, + 0.01224225852638483, + 0.015173065476119518, + 0.016238447278738022, + 0.05810261145234108, + 0.03292642906308174, + 0.031773392111063004, + 0.04284161701798439, + 0.002263549482449889, + 0.0039432900957763195, + 0.0049775876104831696, + 0.005324735306203365, + 0.004547731019556522, + 0.003913313150405884, + 0.003646476659923792, + 0.0035242359153926373, + 0.005580507684499025, + 0.005847935099154711, + 0.010018616914749146, + 0.009177248924970627, + 0.005372822284698486, + 0.008884876035153866, + 0.014308471232652664, + 0.008343789726495743, + 0.006737780757248402, + 0.006215726491063833, + 0.007195296231657267, + 0.010158606804907322, + 0.005452234763652086, + 0.0068956236355006695, + 0.0064604440703988075, + 0.006334492936730385, + 0.0028533870354294777, + 0.0031742402352392673, + 0.005843369755893946, + 0.006245353724807501, + 0.0025505544617772102, + 0.003760160179808736, + 0.003942620009183884, + 0.004311482887715101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.10029932111501694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05380873754620552, + 0.10688556730747223, + 0.04608972370624542, + 0.09551785886287689, + 0.04186422377824783, + 0.23899267613887787, + 0.049368828535079956, + 0.23560211062431335, + 0.0991029292345047, + 0.03276733309030533 + ], + "credit_map": [ + 0.003442628076300025, + 0.0069983298890292645, + 0.009862874634563923, + 0.009805752895772457, + 0.006268758326768875, + 0.006949444767087698, + 0.004492111969739199, + 0.0059888409450650215, + 0.005622768774628639, + 0.0052915578708052635, + 0.03596695885062218, + 0.007201979868113995, + 0.005621254909783602, + 0.023298636078834534, + 0.006001324858516455, + 0.01788109913468361, + 0.002272889716550708, + 0.003107306081801653, + 0.004811997525393963, + 0.009210053831338882, + 0.004478262271732092, + 0.005670428276062012, + 0.009583327919244766, + 0.006955462507903576, + 0.009270913898944855, + 0.0066255792044103146, + 0.010859483852982521, + 0.009968184866011143, + 0.028810910880565643, + 0.010938645340502262, + 0.006715983618050814, + 0.01232815906405449, + 0.0028671997133642435, + 0.005117553286254406, + 0.003831614274531603, + 0.0054718428291380405, + 0.009892217814922333, + 0.004819792229682207, + 0.006338232196867466, + 0.003525771666318178, + 0.007060179952532053, + 0.010888626798987389, + 0.007866764441132545, + 0.008440008386969566, + 0.07907059788703918, + 0.024543577805161476, + 0.07610104978084564, + 0.025021877139806747, + 0.0037633059546351433, + 0.003259137040004134, + 0.004148497711867094, + 0.007856565527617931, + 0.004158956930041313, + 0.00643272465094924, + 0.013465261086821556, + 0.006284384522587061, + 0.010182843543589115, + 0.016381196677684784, + 0.014549636282026768, + 0.01508807297796011, + 0.0290543083101511, + 0.08295749872922897, + 0.052822940051555634, + 0.014565633609890938, + 0.010282929055392742, + 0.007946115918457508, + 0.017511291429400444, + 0.022961610928177834, + 0.006855281535536051, + 0.011492778547108173, + 0.012574443593621254, + 0.009478486143052578, + 0.002637803088873625, + 0.0034023409243673086, + 0.004582740366458893, + 0.006015160121023655, + 0.0030426844023168087, + 0.005125040654093027, + 0.0045553166419267654, + 0.003406251547858119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635659046471119 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635659046471119, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635659046471119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0830211490392685, + 0.04469599202275276, + 0.044652342796325684, + 0.22654259204864502, + 0.08730068057775497, + 0.34219956398010254, + 0.018929537385702133, + 0.0687159076333046, + 0.02556818723678589, + 0.05837403982877731 + ], + "credit_map": [ + 0.007327314931899309, + 0.00849608052521944, + 0.01387828215956688, + 0.015404727309942245, + 0.00899680145084858, + 0.010584664531052113, + 0.009008491411805153, + 0.009324777871370316, + 0.004272722639143467, + 0.004794175736606121, + 0.005961776245385408, + 0.006197470240294933, + 0.004324155393987894, + 0.006139949429780245, + 0.0043584806844592094, + 0.008647255599498749, + 0.00521340686827898, + 0.004316391889005899, + 0.004855315666645765, + 0.005190784111618996, + 0.00713816424831748, + 0.011067945510149002, + 0.0035917856730520725, + 0.0032785418443381786, + 0.029016323387622833, + 0.01615934632718563, + 0.029393376782536507, + 0.02480211853981018, + 0.027722759172320366, + 0.041838329285383224, + 0.017876237630844116, + 0.03973407670855522, + 0.010936525650322437, + 0.013478442095220089, + 0.006280980072915554, + 0.01091235876083374, + 0.013949008658528328, + 0.008507026359438896, + 0.010153084993362427, + 0.013083244673907757, + 0.011111246421933174, + 0.016460897400975227, + 0.01149074174463749, + 0.01289919763803482, + 0.11816833168268204, + 0.03870827704668045, + 0.10301654040813446, + 0.03034430555999279, + 0.0012573815183714032, + 0.0014252853579819202, + 0.0020640806760638952, + 0.0035298524890094995, + 0.0034315830562263727, + 0.0020167322363704443, + 0.003373912535607815, + 0.001830707537010312, + 0.0023681537713855505, + 0.0026015823241323233, + 0.002864239737391472, + 0.003845451632514596, + 0.0050816042348742485, + 0.02436726726591587, + 0.021607082337141037, + 0.0059805260971188545, + 0.0018775769276544452, + 0.0027777364011853933, + 0.003378048073500395, + 0.004506464581936598, + 0.0026678245048969984, + 0.0032925468403846025, + 0.0029654819518327713, + 0.004102504812180996, + 0.0038534251507371664, + 0.005074840970337391, + 0.009008556604385376, + 0.010059024207293987, + 0.00498082023113966, + 0.00966127309948206, + 0.01064756978303194, + 0.005088521167635918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.004635659046471119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely slow-motion macro close-up, captured from a slightly overhead perspective, focuses on a pair of slender, white hands carefully cradling a creamy white, pebble-shaped electric hand warmer. The hands, with their neatly manicured nails and gentle gestures, seem to be protecting a treasure. The hand warmer's smooth, rounded, matte finish reflects a soft, warm glow under the warm yellow light of the cafe's vintage tungsten lamps, which cast a sideways cast. As the camera zooms in at an almost imperceptible pace, the heat radiating from the hand warmer's surface becomes concrete in the extreme slow motion: the air around it undergoes a visible, transparent distortion, forming gentle ripples that slowly and gracefully rise and dissipate, like heat waves on an asphalt road on a summer afternoon. The background is rendered extremely blurred by a shallow depth of field, barely discernible enough to distinguish the texture of the dark wood tabletop and the curl of steam from a latte in the distance. Together, they create an intimate atmosphere of profound tranquility, warmth, healing, and exquisite detail.", + "seed": 23291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_11/index_7039_seed_23291_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013796374714002014, + "vlm_logprob": -0.013796374714002014, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.013796374714002014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006859445478767157 + ], + "opposite_logprob_tokens": [ + -5.381859302520752 + ], + "logprob_reward": -0.006859445478767157, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006859445478767157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17940719425678253, + 0.3811172842979431, + 0.07009214162826538, + 0.02689589001238346, + 0.08023034781217575, + 0.11752407252788544, + 0.012065153568983078, + 0.025571465492248535, + 0.07963688671588898, + 0.027459600940346718 + ], + "credit_map": [ + 0.022693820297718048, + 0.02043992280960083, + 0.014904252253472805, + 0.015396058559417725, + 0.03419812396168709, + 0.030362581834197044, + 0.018027734011411667, + 0.023384708911180496, + 0.03843478485941887, + 0.08111622929573059, + 0.021006228402256966, + 0.042182572185993195, + 0.03855164349079132, + 0.04469424858689308, + 0.0482260026037693, + 0.06690556555986404, + 0.006697430741041899, + 0.011058765463531017, + 0.007500709034502506, + 0.004320929292589426, + 0.008388482965528965, + 0.00626019760966301, + 0.005829459987580776, + 0.02003617025911808, + 0.003424659138545394, + 0.0032895400654524565, + 0.0027875066734850407, + 0.002934866352006793, + 0.0028488938696682453, + 0.004174314439296722, + 0.003145448165014386, + 0.004290660843253136, + 0.00816067960113287, + 0.006710451561957598, + 0.0076541779562830925, + 0.01446892786771059, + 0.01355461310595274, + 0.011125083081424236, + 0.010025438852608204, + 0.008530975319445133, + 0.01241734903305769, + 0.011876247823238373, + 0.006445698905736208, + 0.005476293154060841, + 0.04572705179452896, + 0.015344033017754555, + 0.010367521084845066, + 0.00986988004297018, + 0.0018775351345539093, + 0.001766964211128652, + 0.001247276784852147, + 0.0017281678738072515, + 0.0012991279363632202, + 0.0014585034223273396, + 0.0010478203184902668, + 0.0016397580038756132, + 0.0023879341315478086, + 0.0031976248137652874, + 0.0038482900708913803, + 0.0032457306515425444, + 0.0045706043019890785, + 0.0029314220882952213, + 0.0023797140456736088, + 0.003010144457221031, + 0.007775483187288046, + 0.008752075955271721, + 0.009449433535337448, + 0.010983508080244064, + 0.008021230809390545, + 0.009853994473814964, + 0.01033314410597086, + 0.014468011446297169, + 0.0027268934063613415, + 0.0033843982964754105, + 0.002791473176330328, + 0.0034264237619936466, + 0.0037095355801284313, + 0.0033857726957648993, + 0.0037081087939441204, + 0.004326994996517897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.006859445478767157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076436190865933895 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.0076436190865933895, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076436190865933895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06556674838066101, + 0.14933039247989655, + 0.1732042282819748, + 0.07987080514431, + 0.09925082325935364, + 0.16263657808303833, + 0.02215181104838848, + 0.0332447811961174, + 0.14485381543636322, + 0.06989003717899323 + ], + "credit_map": [ + 0.008304060436785221, + 0.006470437627285719, + 0.005513670854270458, + 0.006280655041337013, + 0.012953540310263634, + 0.010951503179967403, + 0.005845760926604271, + 0.00924711674451828, + 0.014461078681051731, + 0.027015766128897667, + 0.0076898387633264065, + 0.019756993278861046, + 0.014852924272418022, + 0.017226682975888252, + 0.01852947659790516, + 0.02979763224720955, + 0.01599700003862381, + 0.021211648359894753, + 0.018705880269408226, + 0.015436169691383839, + 0.01471683755517006, + 0.017417261376976967, + 0.017288831993937492, + 0.05243061110377312, + 0.008053784258663654, + 0.007524171378463507, + 0.009359843097627163, + 0.008266125805675983, + 0.010601701214909554, + 0.011138217523694038, + 0.009896653704345226, + 0.015030310489237309, + 0.008824947290122509, + 0.0209923405200243, + 0.006002210546284914, + 0.010687918402254581, + 0.020396016538143158, + 0.010532481595873833, + 0.01094464398920536, + 0.010870259255170822, + 0.016653524711728096, + 0.01708301343023777, + 0.008031286299228668, + 0.00815571565181017, + 0.06182301044464111, + 0.022368475794792175, + 0.015462690964341164, + 0.013058852404356003, + 0.0028011270333081484, + 0.0027852815110236406, + 0.003626578487455845, + 0.0030253943987190723, + 0.002712582005187869, + 0.0025533728767186403, + 0.002291839336976409, + 0.0023556349333375692, + 0.00332092703320086, + 0.005633446387946606, + 0.0036128871142864227, + 0.0038089279551059008, + 0.006354424636811018, + 0.003239566693082452, + 0.003511910792440176, + 0.003762691980227828, + 0.011584948748350143, + 0.02153247967362404, + 0.012428897432982922, + 0.02014247700572014, + 0.01958073116838932, + 0.017756665125489235, + 0.021567452698946, + 0.0202601570636034, + 0.0072525665163993835, + 0.012281406670808792, + 0.007988154888153076, + 0.007806054782122374, + 0.0072049363516271114, + 0.01120076235383749, + 0.010007456876337528, + 0.006148704793304205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.0076436190865933895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038090988993644714 + ], + "opposite_logprob_tokens": [ + -3.288090944290161 + ], + "logprob_reward": -0.038090988993644714, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038090988993644714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37151020765304565, + 0.053755033761262894, + 0.06356865167617798, + 0.06681463122367859, + 0.08353257179260254, + 0.21059384942054749, + 0.02575966902077198, + 0.042062755674123764, + 0.03332464024424553, + 0.04907801002264023 + ], + "credit_map": [ + 0.04933924227952957, + 0.039602555334568024, + 0.02874729596078396, + 0.03399277478456497, + 0.07310274243354797, + 0.07735839486122131, + 0.03349470719695091, + 0.035872504115104675, + 0.005793220829218626, + 0.007518058642745018, + 0.0042447615414857864, + 0.006313486956059933, + 0.006866442039608955, + 0.0062584299594163895, + 0.005466864909976721, + 0.011293767020106316, + 0.005678566638380289, + 0.011817693710327148, + 0.00370068964548409, + 0.0034352492075413465, + 0.007401514798402786, + 0.013223587535321712, + 0.0038692874368280172, + 0.01444206852465868, + 0.0074082897044718266, + 0.01025047805160284, + 0.006241625174880028, + 0.007465366739779711, + 0.011124861426651478, + 0.010574882850050926, + 0.007726948242634535, + 0.006022176239639521, + 0.008832581341266632, + 0.016691066324710846, + 0.005469431169331074, + 0.00592456292361021, + 0.02232874557375908, + 0.00815325602889061, + 0.006716842297464609, + 0.009416085667908192, + 0.022032978013157845, + 0.018962619826197624, + 0.012278023175895214, + 0.02120201475918293, + 0.060486070811748505, + 0.023663785308599472, + 0.021364355459809303, + 0.03060399554669857, + 0.0030241876374930143, + 0.002736405236646533, + 0.002588353818282485, + 0.003576331539079547, + 0.0032805358059704304, + 0.004740288946777582, + 0.003478708676993847, + 0.0023348561953753233, + 0.004581085406243801, + 0.005997036583721638, + 0.005228324327617884, + 0.005996020045131445, + 0.005736981984227896, + 0.005046066828072071, + 0.004507887177169323, + 0.004969354718923569, + 0.002805339638143778, + 0.00440469104796648, + 0.0025418559089303017, + 0.004492361564189196, + 0.008156687952578068, + 0.003896788228303194, + 0.003663331503048539, + 0.00336358486674726, + 0.005629661027342081, + 0.004650563932955265, + 0.006043183617293835, + 0.007365524303168058, + 0.004878236446529627, + 0.00672161066904664, + 0.00589169654995203, + 0.007897532545030117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.038090988993644714 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13062791526317596, + 0.1456594169139862, + 0.08410163968801498, + 0.04464711993932724, + 0.10220300406217575, + 0.3603489100933075, + 0.033740296959877014, + 0.021836085245013237, + 0.043764445930719376, + 0.033071208745241165 + ], + "credit_map": [ + 0.01311669871211052, + 0.016240645200014114, + 0.014607015997171402, + 0.01475905254483223, + 0.021591249853372574, + 0.014680912718176842, + 0.020747002214193344, + 0.014885329641401768, + 0.014041838236153126, + 0.02674453891813755, + 0.009277617558836937, + 0.019902504980564117, + 0.021143252030014992, + 0.014102152548730373, + 0.017467256635427475, + 0.022980250418186188, + 0.007731021847575903, + 0.013904274441301823, + 0.005531067959964275, + 0.0047345939092338085, + 0.0058171129785478115, + 0.023051347583532333, + 0.0059204925782978535, + 0.017411720007658005, + 0.004316018428653479, + 0.003688819007948041, + 0.002811422571539879, + 0.007366156671196222, + 0.005912661552429199, + 0.007198152132332325, + 0.005787070374935865, + 0.007566816173493862, + 0.010356122627854347, + 0.007924216799438, + 0.009725715033710003, + 0.02206186205148697, + 0.010975031182169914, + 0.013842880725860596, + 0.015298583544790745, + 0.012018583714962006, + 0.038030073046684265, + 0.033573251217603683, + 0.022837895900011063, + 0.021495822817087173, + 0.1338510662317276, + 0.04728594422340393, + 0.034315336495637894, + 0.028959495946764946, + 0.005794888362288475, + 0.002966222120448947, + 0.003323975019156933, + 0.003651713253930211, + 0.0030907229520380497, + 0.0052195400930941105, + 0.0029205861501395702, + 0.006772646214812994, + 0.002721810946241021, + 0.0034559962805360556, + 0.0018161131301894784, + 0.002775357337668538, + 0.004176902584731579, + 0.0027153361588716507, + 0.002278337487950921, + 0.0018962295725941658, + 0.004107067361474037, + 0.006846805568784475, + 0.0048964316956698895, + 0.004247605800628662, + 0.006890422664582729, + 0.008749401196837425, + 0.004512520506978035, + 0.0035141874104738235, + 0.003853379050269723, + 0.003937766421586275, + 0.004636587109416723, + 0.004885772243142128, + 0.003640634473413229, + 0.004577813670039177, + 0.003412828780710697, + 0.004126423969864845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.01419165637344122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021961636375635862 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021961636375635862, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021961636375635862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3719930052757263, + 0.1334153264760971, + 0.11588672548532486, + 0.039294153451919556, + 0.10618215799331665, + 0.09199134260416031, + 0.02413948066532612, + 0.01963994652032852, + 0.06159330531954765, + 0.03586451709270477 + ], + "credit_map": [ + 0.038758307695388794, + 0.04694152623414993, + 0.028547203168272972, + 0.047493405640125275, + 0.06279462575912476, + 0.0463201105594635, + 0.0580032654106617, + 0.04313460364937782, + 0.015077625401318073, + 0.02602749690413475, + 0.010320357978343964, + 0.01300530880689621, + 0.014243558049201965, + 0.018611211329698563, + 0.014818833209574223, + 0.02131093665957451, + 0.009211384691298008, + 0.022944362834095955, + 0.015036332421004772, + 0.007569232024252415, + 0.008260749280452728, + 0.023750029504299164, + 0.007172766141593456, + 0.021941885352134705, + 0.0049174330197274685, + 0.004096024204045534, + 0.00590712483972311, + 0.004776155110448599, + 0.004078020807355642, + 0.0037103795912116766, + 0.004315781872719526, + 0.007493237499147654, + 0.01066067349165678, + 0.00809562113136053, + 0.024161603301763535, + 0.01713082194328308, + 0.00848968606442213, + 0.01077640987932682, + 0.012796102091670036, + 0.014071251265704632, + 0.011967984028160572, + 0.010586506687104702, + 0.014994730241596699, + 0.008642693050205708, + 0.015167689882218838, + 0.010900091379880905, + 0.01101988647133112, + 0.00871177390217781, + 0.003221596824005246, + 0.0025871845427900553, + 0.003410497447475791, + 0.003051589708775282, + 0.0032265791669487953, + 0.0030665427912026644, + 0.0026045357808470726, + 0.002970956265926361, + 0.0026162280701100826, + 0.002503602998331189, + 0.0030786050483584404, + 0.002417857525870204, + 0.0024442654103040695, + 0.002288536401465535, + 0.0022520902566611767, + 0.002038760809227824, + 0.0052735162898898125, + 0.009228103794157505, + 0.005620407871901989, + 0.007187242619693279, + 0.015022475272417068, + 0.008206839673221111, + 0.00685880659148097, + 0.00419591786339879, + 0.00506030535325408, + 0.004591953940689564, + 0.0039702728390693665, + 0.005116966087371111, + 0.003819939447566867, + 0.005435430444777012, + 0.004348087124526501, + 0.003521564183756709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0021961636375635862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely slow-motion macro close-up, captured from a slightly overhead perspective, focuses on a pair of slender, white hands carefully cradling a creamy white, pebble-shaped electric hand warmer. The hands, with their neatly manicured nails and gentle gestures, seem to be protecting a treasure. The hand warmer's smooth, rounded, matte finish reflects a soft, warm glow under the warm yellow light of the cafe's vintage tungsten lamps, which cast a sideways cast. As the camera zooms in at an almost imperceptible pace, the heat radiating from the hand warmer's surface becomes concrete in the extreme slow motion: the air around it undergoes a visible, transparent distortion, forming gentle ripples that slowly and gracefully rise and dissipate, like heat waves on an asphalt road on a summer afternoon. The background is rendered extremely blurred by a shallow depth of field, barely discernible enough to distinguish the texture of the dark wood tabletop and the curl of steam from a latte in the distance. Together, they create an intimate atmosphere of profound tranquility, warmth, healing, and exquisite detail.", + "seed": 23292, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_11/index_7039_seed_23292_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02479604948312044, + "vlm_logprob": -0.02479604948312044, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.02479604948312044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004149755463004112 + ], + "opposite_logprob_tokens": [ + -6.129149913787842 + ], + "logprob_reward": -0.004149755463004112, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004149755463004112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3630583584308624, + 0.026686392724514008, + 0.031961001455783844, + 0.016914529725909233, + 0.023181607946753502, + 0.04644704982638359, + 0.3637998402118683, + 0.022355226799845695, + 0.08108209818601608, + 0.024514049291610718 + ], + "credit_map": [ + 0.01498199999332428, + 0.01851513795554638, + 0.040836818516254425, + 0.01246589981019497, + 0.08863777667284012, + 0.15425251424312592, + 0.019029155373573303, + 0.014338976703584194, + 0.0024838936515152454, + 0.001960313180461526, + 0.00258030672557652, + 0.0031628399156033993, + 0.002553783357143402, + 0.00356221292167902, + 0.006050598341971636, + 0.004332439042627811, + 0.0020863167010247707, + 0.0014624993782490492, + 0.0032429599668830633, + 0.009106270968914032, + 0.006288633681833744, + 0.0037276819348335266, + 0.0035548510495573282, + 0.002491776831448078, + 0.0008810050203464925, + 0.0012652610894292593, + 0.0015468839555978775, + 0.00339562282897532, + 0.004850545432418585, + 0.0015804540598765016, + 0.0021561572793871164, + 0.0012385952286422253, + 0.0017215970437973738, + 0.0037047413643449545, + 0.002564293099567294, + 0.0027128360234200954, + 0.003095766995102167, + 0.002654504729434848, + 0.004500157665461302, + 0.0022277061361819506, + 0.002911474322900176, + 0.004814003594219685, + 0.00747799314558506, + 0.010883774608373642, + 0.006684138439595699, + 0.003879590891301632, + 0.005425179842859507, + 0.004370886832475662, + 0.030025161802768707, + 0.0360490120947361, + 0.030729657039046288, + 0.06597261130809784, + 0.07261956483125687, + 0.023634662851691246, + 0.07940845191478729, + 0.02536063641309738, + 0.0017214474501088262, + 0.002510039135813713, + 0.002384781837463379, + 0.00341782602481544, + 0.0029319878667593002, + 0.0022160212974995375, + 0.004260439891368151, + 0.00291267829015851, + 0.004131317604333162, + 0.00693144928663969, + 0.008078794926404953, + 0.011368748731911182, + 0.01632513664662838, + 0.006998955272138119, + 0.018838470801711082, + 0.008409210480749607, + 0.001712096738629043, + 0.004065802786499262, + 0.003078977344557643, + 0.002689956920221448, + 0.0027880719862878323, + 0.0025706738233566284, + 0.004773405380547047, + 0.0028350583743304014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.004149755463004112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656359277665615 + ], + "opposite_logprob_tokens": [ + -4.7586565017700195 + ], + "logprob_reward": -0.008656359277665615, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656359277665615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5174273252487183, + 0.06137530878186226, + 0.09626932442188263, + 0.049583856016397476, + 0.023066872730851173, + 0.06483183056116104, + 0.0353732593357563, + 0.027097545564174652, + 0.031898170709609985, + 0.09307651966810226 + ], + "credit_map": [ + 0.02161656692624092, + 0.024388018995523453, + 0.054767053574323654, + 0.014187393710017204, + 0.14131176471710205, + 0.22431841492652893, + 0.020641997456550598, + 0.0161961130797863, + 0.0047332849353551865, + 0.006229509599506855, + 0.009162659756839275, + 0.008587367832660675, + 0.008289969526231289, + 0.007696668151766062, + 0.008994868025183678, + 0.007680980488657951, + 0.003632252337411046, + 0.004462459124624729, + 0.005562246311455965, + 0.03770401328802109, + 0.012148329056799412, + 0.018180515617132187, + 0.010777472518384457, + 0.003802041057497263, + 0.0020235346164554358, + 0.003380412235856056, + 0.003273939248174429, + 0.01104013342410326, + 0.018364738672971725, + 0.005648065824061632, + 0.003938533831387758, + 0.0019144952530041337, + 0.001485638553276658, + 0.0031061985064297915, + 0.0021526247728616, + 0.002971965353935957, + 0.0033568257931619883, + 0.0027171268593519926, + 0.004666613880544901, + 0.002609879244118929, + 0.0042844596318900585, + 0.005950265098363161, + 0.013603122904896736, + 0.007597258780151606, + 0.005860134493559599, + 0.0057608019560575485, + 0.012883500196039677, + 0.008892284706234932, + 0.0027416739612817764, + 0.003547814441844821, + 0.003480662126094103, + 0.007216768804937601, + 0.004525105003267527, + 0.0022273596841841936, + 0.008296011947095394, + 0.003337862901389599, + 0.001531551475636661, + 0.004713919945061207, + 0.00385666242800653, + 0.0038320431485772133, + 0.0029744773637503386, + 0.0030504781752824783, + 0.004748293198645115, + 0.0023901190143078566, + 0.0024958569556474686, + 0.003572610439732671, + 0.003184163710102439, + 0.006255814339965582, + 0.005538355093449354, + 0.002941545331850648, + 0.004368342459201813, + 0.003541484009474516, + 0.005043019540607929, + 0.01481879223138094, + 0.009606612846255302, + 0.01135577168315649, + 0.01493298914283514, + 0.009067904204130173, + 0.018334439024329185, + 0.009916985407471657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.008656359277665615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890301913022995 + ], + "opposite_logprob_tokens": [ + -2.4640302658081055 + ], + "logprob_reward": -0.0890301913022995, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890301913022995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7425853610038757, + 0.022482359781861305, + 0.04121321439743042, + 0.016421250998973846, + 0.042147744446992874, + 0.06219356507062912, + 0.029934920370578766, + 0.011963024735450745, + 0.01667499914765358, + 0.014383658766746521 + ], + "credit_map": [ + 0.0310773104429245, + 0.03529173508286476, + 0.07835663110017776, + 0.01971040479838848, + 0.2045915126800537, + 0.3233424723148346, + 0.028194883838295937, + 0.022020360454916954, + 0.0015719899674877524, + 0.0013444563373923302, + 0.0036878897808492184, + 0.0027385479770600796, + 0.003525745589286089, + 0.004199047572910786, + 0.0035201460123062134, + 0.0018945364281535149, + 0.0014400562504306436, + 0.0027252419386059046, + 0.001993410522118211, + 0.012393336743116379, + 0.015069176442921162, + 0.002441978082060814, + 0.002951271366328001, + 0.0021987396758049726, + 0.0010893391445279121, + 0.001200733706355095, + 0.0015299482038244605, + 0.002880660118535161, + 0.004205717705190182, + 0.0019221643451601267, + 0.0024568261578679085, + 0.0011358618503436446, + 0.0021633747965097427, + 0.005305412225425243, + 0.004333559889346361, + 0.00641992874443531, + 0.005722121801227331, + 0.0043404544703662395, + 0.009142491035163403, + 0.004720402415841818, + 0.003678493667393923, + 0.0055765425786376, + 0.009425917640328407, + 0.013830709271132946, + 0.009070344269275665, + 0.004831559956073761, + 0.009572078473865986, + 0.006207916419953108, + 0.0023395782336592674, + 0.002978360280394554, + 0.002483660588040948, + 0.003776267869397998, + 0.0038658594712615013, + 0.001912420499138534, + 0.009426593780517578, + 0.003152181627228856, + 0.0006097711739130318, + 0.0012987599475309253, + 0.0018930413061752915, + 0.0016404372872784734, + 0.0011327689280733466, + 0.0011586104519665241, + 0.002767082303762436, + 0.0014625524636358023, + 0.0007870528497733176, + 0.0013699569972231984, + 0.001167454058304429, + 0.0034765969030559063, + 0.0038454472087323666, + 0.001500127837061882, + 0.0031221937388181686, + 0.0014061700785532594, + 0.0017600276041775942, + 0.002427775179967284, + 0.0015655518509447575, + 0.0015899167628958821, + 0.00220372318290174, + 0.0016051275888457894, + 0.0017442384269088507, + 0.0014872982865199447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0890301913022995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107415184378624 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.01107415184378624, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107415184378624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.625822901725769, + 0.028687577694654465, + 0.07033036649227142, + 0.0242533590644598, + 0.027125127613544464, + 0.01965785212814808, + 0.13470079004764557, + 0.012393799610435963, + 0.01667928881943226, + 0.04034895822405815 + ], + "credit_map": [ + 0.026106877252459526, + 0.03162211924791336, + 0.07427918910980225, + 0.02243584208190441, + 0.13739308714866638, + 0.27948063611984253, + 0.029879791662096977, + 0.024625353515148163, + 0.0029664558824151754, + 0.0018736869096755981, + 0.003940610680729151, + 0.0036104877945035696, + 0.004133434500545263, + 0.0036411078181117773, + 0.005899281240999699, + 0.0026225142646580935, + 0.002383763901889324, + 0.004083232022821903, + 0.004095886368304491, + 0.013415751047432423, + 0.026771709322929382, + 0.010495424270629883, + 0.003557695774361491, + 0.0055269054137170315, + 0.0008815006003715098, + 0.0014126686146482825, + 0.0013887613313272595, + 0.004565507173538208, + 0.007729252800345421, + 0.0025266394950449467, + 0.003674509935081005, + 0.0020745189394801855, + 0.0019227743614464998, + 0.004868431482464075, + 0.004209918901324272, + 0.0031998511403799057, + 0.00407154206186533, + 0.0024902343284338713, + 0.003740596817806363, + 0.002621779916808009, + 0.001772782881744206, + 0.0024378395173698664, + 0.0029900209046900272, + 0.002335540484637022, + 0.002614642959088087, + 0.0018665236420929432, + 0.0034861064050346613, + 0.0021543968468904495, + 0.011699480004608631, + 0.016151711344718933, + 0.012810157611966133, + 0.029797689989209175, + 0.0329006053507328, + 0.008885740302503109, + 0.014094296842813492, + 0.008361108601093292, + 0.0006053484394215047, + 0.001301588024944067, + 0.0014294099528342485, + 0.001725366455502808, + 0.0009581337217241526, + 0.0010853854473680258, + 0.0037397206760942936, + 0.0015488456701859832, + 0.0009202665532939136, + 0.0012433548690751195, + 0.0019916987512260675, + 0.002032674616202712, + 0.002422264777123928, + 0.0016149242874234915, + 0.004418862517923117, + 0.0020352425053715706, + 0.0028309496119618416, + 0.005984598305076361, + 0.0034789443016052246, + 0.0054146223701536655, + 0.006827133242040873, + 0.005089910235255957, + 0.006997713819146156, + 0.003725090529769659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.01107415184378624 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, white hands with neatly manicured nails visible cradling an object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the object being held a creamy white, pebble-shaped electric hand warmer with a smooth, rounded, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot in extreme slow motion showing visible, transparent air ripples rising from the hand warmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in at an almost imperceptible pace from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background extremely blurred with a shallow depth of field, revealing only a dark wood tabletop and distant steam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106978952884674 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.01106978952884674, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106978952884674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9436900019645691, + 0.007080147974193096, + 0.008819518610835075, + 0.002049484523013234, + 0.0036706498358398676, + 0.005315691698342562, + 0.016072962433099747, + 0.003078712848946452, + 0.006543281022459269, + 0.003679475747048855 + ], + "credit_map": [ + 0.03810913488268852, + 0.0462510772049427, + 0.10393878817558289, + 0.028061294928193092, + 0.23741520941257477, + 0.41955050826072693, + 0.039429329335689545, + 0.03093484416604042, + 0.00040578143671154976, + 0.00033194164279848337, + 0.0008263984345830977, + 0.0008410029113292694, + 0.0012779379030689597, + 0.0016810351517051458, + 0.0010265219025313854, + 0.000689529872033745, + 0.0003850051143672317, + 0.0003574721049517393, + 0.0007254312513396144, + 0.00033594146952964365, + 0.002924279309809208, + 0.002942568389698863, + 0.00041750274249352515, + 0.0007313202368095517, + 0.00011126682511530817, + 0.00015517749125137925, + 0.0002313173026777804, + 0.0003926271456293762, + 0.00040808465564623475, + 0.000238447668380104, + 0.0002526639145798981, + 0.0002598999999463558, + 0.0002807870623655617, + 0.0005022001569159329, + 0.0003640562354121357, + 0.0005950885824859142, + 0.0005064954748377204, + 0.0003890931257046759, + 0.0006566386437043548, + 0.00037629116559401155, + 0.00032372004352509975, + 0.0006551507394760847, + 0.0008724766667000949, + 0.0011773634469136596, + 0.0008535991073586047, + 0.0003863836464006454, + 0.0005768128903582692, + 0.000470186147140339, + 0.0012405836023390293, + 0.0015683198580518365, + 0.0015129769453778863, + 0.0027039535343647003, + 0.002933458425104618, + 0.0010132839670404792, + 0.004003977403044701, + 0.0010964099783450365, + 0.00018971330428030342, + 0.0004355176934041083, + 0.0004777393478434533, + 0.0003578625328373164, + 0.0005586720653809607, + 0.00029987606103532016, + 0.00048431899631395936, + 0.00027501347358338535, + 0.00026725706993602216, + 0.0005842987447977066, + 0.000370923982700333, + 0.0015436334069818258, + 0.0020768428221344948, + 0.0003775048244278878, + 0.0008920990512706339, + 0.00043072275002487004, + 0.00039867375744506717, + 0.0007524198153987527, + 0.000372526264982298, + 0.00046762818237766623, + 0.0005763316294178367, + 0.00040515742148272693, + 0.00039476773235946894, + 0.00031197184580378234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.01106978952884674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_12/index_tensor([1891, 1891])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_12/index_tensor([1891, 1891])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9c8210e3be6c13fc4486cc52a594ea58101ed228 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_12/index_tensor([1891, 1891])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "hyperreal 3D rendering captures a time-lapse sequence from a static, low-angle macro shot focused on a desert rose (Adenium obesum). The plant's gnarled, swollen stem base, resembling aged, rugged bark, is firmly rooted in fine, pale, sandy soil. A single, powerful light source moves across the scene, creating dramatic, deep shadows that constantly shift and lengthen, emphasizing the plant's form. As the time-lapse progresses, a single, magnificent flower begins to unfold from a tight bud; its petals, boasting a realistic waxy texture, steadily and gracefully open, revealing a vibrant color gradient from a saturated, deep pink at the edges to a soft, pure white near the flower's throat. Dynamic light flickers from the petals' waxy surface, capturing tiny grains of crystallized sand scattered along the stem and highlighting every minute detail, creating a focused and dramatic atmosphere that celebrates the slow, beautiful spectacle of life.", + "seed": 18721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_12/index_1891_seed_18721_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03093636375851929, + "vlm_logprob": -0.03093636375851929, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.03093636375851929, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01742037944495678 + ], + "opposite_logprob_tokens": [ + -4.892420291900635 + ], + "logprob_reward": -0.01742037944495678, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01742037944495678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023357810452580452, + 0.03634999319911003, + 0.007964649237692356, + 0.09319216012954712, + 0.07140977680683136, + 0.03943949192762375, + 0.4024874269962311, + 0.17921310663223267, + 0.10352954268455505, + 0.043056048452854156 + ], + "credit_map": [ + 0.003553497837856412, + 0.002285616472363472, + 0.0036461574491113424, + 0.002058439888060093, + 0.0025553428567945957, + 0.003800614271312952, + 0.0024427506141364574, + 0.003015389433130622, + 0.005440572742372751, + 0.00818171538412571, + 0.00398870836943388, + 0.004032272379845381, + 0.004985006991773844, + 0.004429512191563845, + 0.002830212004482746, + 0.002461993135511875, + 0.0011630437802523375, + 0.0010157719952985644, + 0.0011099817929789424, + 0.000840117281768471, + 0.0009404500597156584, + 0.0011862918036058545, + 0.0010327015770599246, + 0.0006762906559742987, + 0.016830282285809517, + 0.008027228526771069, + 0.017341872677206993, + 0.008885943330824375, + 0.0067754085175693035, + 0.019727738574147224, + 0.00696387467905879, + 0.008639812469482422, + 0.010863464325666428, + 0.0031896759755909443, + 0.0263707023113966, + 0.0037636435590684414, + 0.002080461010336876, + 0.014166049659252167, + 0.004080876242369413, + 0.006894905120134354, + 0.011420969851315022, + 0.002599888015538454, + 0.0035740372259169817, + 0.002617388730868697, + 0.0018447611946612597, + 0.012569521553814411, + 0.0025214231573045254, + 0.0022915047593414783, + 0.08987867832183838, + 0.04147745668888092, + 0.050785571336746216, + 0.029712781310081482, + 0.05444609746336937, + 0.08644868433475494, + 0.02405915968120098, + 0.025679007172584534, + 0.05143463984131813, + 0.008141472935676575, + 0.07237747311592102, + 0.01594386249780655, + 0.004827974829822779, + 0.012446703389286995, + 0.005987103562802076, + 0.008053879253566265, + 0.03685172274708748, + 0.004073975142091513, + 0.04463230073451996, + 0.005501348990947008, + 0.0039390092715620995, + 0.0037397483829408884, + 0.0022063138894736767, + 0.0025851137470453978, + 0.009570575319230556, + 0.0038970522582530975, + 0.0035703897010535, + 0.0029677615966647863, + 0.013490746729075909, + 0.0036003224086016417, + 0.003153037279844284, + 0.0028061657212674618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.01742037944495678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015257871709764004 + ], + "opposite_logprob_tokens": [ + -6.50152587890625 + ], + "logprob_reward": -0.0015257871709764004, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015257871709764004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031083455309271812, + 0.05670157074928284, + 0.01888321153819561, + 0.06241399049758911, + 0.051010001450777054, + 0.05554725602269173, + 0.42830461263656616, + 0.15800638496875763, + 0.06536784023046494, + 0.07268169522285461 + ], + "credit_map": [ + 0.004213850479573011, + 0.0033428797032684088, + 0.004319822881370783, + 0.0033617247827351093, + 0.004138553515076637, + 0.00501698674634099, + 0.0036384782288223505, + 0.003051157109439373, + 0.009013893082737923, + 0.009564586915075779, + 0.009998615831136703, + 0.006072100251913071, + 0.005824698135256767, + 0.008656885474920273, + 0.0038897425401955843, + 0.003681046422570944, + 0.0026912398170679808, + 0.002461017109453678, + 0.0025167479179799557, + 0.0023404182866215706, + 0.0026530746836215258, + 0.002233909908682108, + 0.002379036508500576, + 0.0016077656764537096, + 0.010621792636811733, + 0.0061217742040753365, + 0.011096743866801262, + 0.005734129808843136, + 0.005053849890828133, + 0.013565169647336006, + 0.004549982491880655, + 0.005670546554028988, + 0.005652743391692638, + 0.0031599425710737705, + 0.016028258949518204, + 0.003238259581848979, + 0.0021709343418478966, + 0.009084434248507023, + 0.004161484073847532, + 0.007513946387916803, + 0.017680460587143898, + 0.0030756692867726088, + 0.004669387824833393, + 0.0027129040099680424, + 0.00200340966694057, + 0.01882975734770298, + 0.0033188294619321823, + 0.0032568341121077538, + 0.09419028460979462, + 0.04463685676455498, + 0.05582890659570694, + 0.03256150707602501, + 0.059162404388189316, + 0.08889903873205185, + 0.025975612923502922, + 0.027049953117966652, + 0.04252927005290985, + 0.007616721093654633, + 0.06173251196742058, + 0.014332564547657967, + 0.005306263919919729, + 0.012536205351352692, + 0.00658769067376852, + 0.007365145720541477, + 0.022142747417092323, + 0.003035970265045762, + 0.026623908430337906, + 0.004282820504158735, + 0.0027281981892883778, + 0.0024985801428556442, + 0.0018813913920894265, + 0.002174221910536289, + 0.021877989172935486, + 0.004274613689631224, + 0.015507039614021778, + 0.00408670911565423, + 0.01506463997066021, + 0.004352121613919735, + 0.003730720141902566, + 0.003787856549024582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0015257871709764004 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036305959802120924 + ], + "opposite_logprob_tokens": [ + -5.628630638122559 + ], + "logprob_reward": -0.0036305959802120924, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036305959802120924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04009559005498886, + 0.10545807331800461, + 0.08306697010993958, + 0.06257369369268417, + 0.06815239787101746, + 0.03321496769785881, + 0.27167242765426636, + 0.19020885229110718, + 0.0680340975522995, + 0.0775228887796402 + ], + "credit_map": [ + 0.0053536961786448956, + 0.005301433615386486, + 0.00488709332421422, + 0.00573735311627388, + 0.006099465768784285, + 0.0042682671919465065, + 0.0047821542248129845, + 0.0036661257036030293, + 0.017280444502830505, + 0.010750452987849712, + 0.023132838308811188, + 0.013678357936441898, + 0.008677162230014801, + 0.01838851347565651, + 0.007162910420447588, + 0.0063873883336782455, + 0.01388282049447298, + 0.006965958047658205, + 0.015080517157912254, + 0.00953708030283451, + 0.007494017016142607, + 0.01551448367536068, + 0.007638849783688784, + 0.006953246891498566, + 0.008506719954311848, + 0.009516255930066109, + 0.008661274798214436, + 0.005552861839532852, + 0.0070210229605436325, + 0.011550224386155605, + 0.006085666827857494, + 0.005679671186953783, + 0.010197142139077187, + 0.003512366907671094, + 0.02405715361237526, + 0.004245156887918711, + 0.002580129075795412, + 0.013467839919030666, + 0.0036128945648670197, + 0.006479715462774038, + 0.009617311879992485, + 0.0022788229398429394, + 0.0032673454843461514, + 0.0019864016212522984, + 0.0014914149651303887, + 0.010260854847729206, + 0.002512677339836955, + 0.0018001374555751681, + 0.06022679805755615, + 0.027888013049960136, + 0.03502267599105835, + 0.021369706839323044, + 0.038086749613285065, + 0.05509795993566513, + 0.016843285411596298, + 0.01713724620640278, + 0.05155399814248085, + 0.009243478067219257, + 0.07481549680233002, + 0.01714986190199852, + 0.006109706591814756, + 0.014692166820168495, + 0.007725388742983341, + 0.008918753825128078, + 0.01948145218193531, + 0.004887507297098637, + 0.021848592907190323, + 0.005807146895676851, + 0.002974779810756445, + 0.00386583199724555, + 0.0042266626842319965, + 0.004942124709486961, + 0.02287079021334648, + 0.005386190488934517, + 0.013750839047133923, + 0.004097414202988148, + 0.01781575009226799, + 0.0049162134528160095, + 0.004372821655124426, + 0.00431287894025445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.0036305959802120924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306374117732048 + ], + "opposite_logprob_tokens": [ + -3.1680636405944824 + ], + "logprob_reward": -0.04306374117732048, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306374117732048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06483101099729538, + 0.10471320152282715, + 0.10594075918197632, + 0.1772083044052124, + 0.08509714156389236, + 0.058182135224342346, + 0.0724702700972557, + 0.10210826247930527, + 0.19953075051307678, + 0.029918139800429344 + ], + "credit_map": [ + 0.010003176517784595, + 0.008600390516221523, + 0.005847175605595112, + 0.007732430472970009, + 0.01261825766414404, + 0.007482228334993124, + 0.006646336056292057, + 0.005901014897972345, + 0.015735596418380737, + 0.015535078011453152, + 0.019094111397862434, + 0.016221188008785248, + 0.009882221929728985, + 0.014433905482292175, + 0.008680891245603561, + 0.005130205303430557, + 0.019340118393301964, + 0.009825965389609337, + 0.020025067031383514, + 0.011752553284168243, + 0.008907408453524113, + 0.019168810918927193, + 0.008647006936371326, + 0.008273833431303501, + 0.030513742938637733, + 0.016844745725393295, + 0.03283675014972687, + 0.017397798597812653, + 0.013031978160142899, + 0.036569297313690186, + 0.01283178199082613, + 0.017182214185595512, + 0.01931609958410263, + 0.007631060667335987, + 0.023786302655935287, + 0.005577027797698975, + 0.004121830686926842, + 0.012517223134636879, + 0.006468098144978285, + 0.005679496563971043, + 0.015956450253725052, + 0.005115493666380644, + 0.004251799080520868, + 0.004454507492482662, + 0.0037672696635127068, + 0.01618001237511635, + 0.004662786610424519, + 0.00379381887614727, + 0.013649261556565762, + 0.004249969031661749, + 0.010088825598359108, + 0.004327893257141113, + 0.006326981820166111, + 0.019381413236260414, + 0.004753238055855036, + 0.009692691266536713, + 0.021925510838627815, + 0.007240304257720709, + 0.03349122777581215, + 0.011087057180702686, + 0.00584944523870945, + 0.009420858696103096, + 0.006512606516480446, + 0.006581243593245745, + 0.06807894259691238, + 0.009042523801326752, + 0.08191610872745514, + 0.013180636800825596, + 0.007879755459725857, + 0.007048536092042923, + 0.005896474700421095, + 0.006487771403044462, + 0.004452177323400974, + 0.005031422711908817, + 0.0025808908976614475, + 0.004184491001069546, + 0.003039235481992364, + 0.004607377573847771, + 0.002853808458894491, + 0.003168735420331359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.04306374117732048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890413150191307 + ], + "opposite_logprob_tokens": [ + -2.464041233062744 + ], + "logprob_reward": -0.0890413150191307, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890413150191307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024777984246611595, + 0.15643242001533508, + 0.057619016617536545, + 0.13470593094825745, + 0.08031166344881058, + 0.0324457511305809, + 0.14275699853897095, + 0.08323130756616592, + 0.17976754903793335, + 0.10795144736766815 + ], + "credit_map": [ + 0.003434737678617239, + 0.0031472158152610064, + 0.003929357044398785, + 0.0031480558682233095, + 0.0030148408841341734, + 0.0030703251250088215, + 0.0028958746697753668, + 0.0021375759970396757, + 0.026641270145773888, + 0.014906433410942554, + 0.03545377776026726, + 0.01996534690260887, + 0.0121697299182415, + 0.02768132835626602, + 0.009892144240438938, + 0.009722385555505753, + 0.009731685742735863, + 0.005808873567730188, + 0.010625746101140976, + 0.00551649508997798, + 0.005115410313010216, + 0.011091673746705055, + 0.00474948575720191, + 0.004979648161679506, + 0.024646885693073273, + 0.009912329725921154, + 0.025880910456180573, + 0.01427531149238348, + 0.00969762448221445, + 0.02622326835989952, + 0.011766569688916206, + 0.012303020805120468, + 0.011020736768841743, + 0.0028731501661241055, + 0.0303092859685421, + 0.00388261117041111, + 0.001990396063774824, + 0.01722809672355652, + 0.0040102615021169186, + 0.008997130207717419, + 0.006730583496391773, + 0.0030247094109654427, + 0.0041832770220935345, + 0.0032038188073784113, + 0.00159416813403368, + 0.00829046405851841, + 0.0026441351510584354, + 0.002774592023342848, + 0.030221901834011078, + 0.014989475719630718, + 0.017947524785995483, + 0.01303406897932291, + 0.021436138078570366, + 0.02630492113530636, + 0.0095831248909235, + 0.009239835664629936, + 0.023103773593902588, + 0.003915747627615929, + 0.03165895491838455, + 0.006796961650252342, + 0.0028402935713529587, + 0.006186377722769976, + 0.0030212823767215014, + 0.0057079182006418705, + 0.06258510798215866, + 0.007389049977064133, + 0.07635961472988129, + 0.009682728908956051, + 0.0070793768391013145, + 0.007142308633774519, + 0.004302655346691608, + 0.005226692650467157, + 0.0400363951921463, + 0.005965751595795155, + 0.03717110678553581, + 0.005854940973222256, + 0.007601679302752018, + 0.004097628407180309, + 0.0036778643261641264, + 0.0035460777580738068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.0890413150191307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "hyperreal 3D rendering captures a time-lapse sequence from a static, low-angle macro shot focused on a desert rose (Adenium obesum). The plant's gnarled, swollen stem base, resembling aged, rugged bark, is firmly rooted in fine, pale, sandy soil. A single, powerful light source moves across the scene, creating dramatic, deep shadows that constantly shift and lengthen, emphasizing the plant's form. As the time-lapse progresses, a single, magnificent flower begins to unfold from a tight bud; its petals, boasting a realistic waxy texture, steadily and gracefully open, revealing a vibrant color gradient from a saturated, deep pink at the edges to a soft, pure white near the flower's throat. Dynamic light flickers from the petals' waxy surface, capturing tiny grains of crystallized sand scattered along the stem and highlighting every minute detail, creating a focused and dramatic atmosphere that celebrates the slow, beautiful spectacle of life.", + "seed": 18722, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_12/index_1891_seed_18722_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04005455996375531, + "vlm_logprob": -0.04005455996375531, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.04005455996375531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010458542965352535 + ], + "opposite_logprob_tokens": [ + -5.885458469390869 + ], + "logprob_reward": -0.010458542965352535, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.010458542965352535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06658229976892471, + 0.3196103572845459, + 0.10656634718179703, + 0.10846664011478424, + 0.035071566700935364, + 0.03686142712831497, + 0.24041031301021576, + 0.029245059937238693, + 0.033721186220645905, + 0.023464757949113846 + ], + "credit_map": [ + 0.010001879185438156, + 0.003638122696429491, + 0.003922820556908846, + 0.003688289551064372, + 0.0030550919473171234, + 0.02291753701865673, + 0.005438604392111301, + 0.013919951394200325, + 0.039416175335645676, + 0.005487234331667423, + 0.00796500127762556, + 0.007355811074376106, + 0.00898035429418087, + 0.12813900411128998, + 0.006634600926190615, + 0.11563218384981155, + 0.0106747942045331, + 0.0035621868446469307, + 0.0032825374510139227, + 0.0037703856360167265, + 0.003951392602175474, + 0.04001867398619652, + 0.0031098974868655205, + 0.03819647803902626, + 0.03308677673339844, + 0.003950851503759623, + 0.007665290497243404, + 0.005584102123975754, + 0.0031739447731524706, + 0.012133869342505932, + 0.0032944469712674618, + 0.03957735747098923, + 0.005734679289162159, + 0.0021535465493798256, + 0.004771479405462742, + 0.002561465371400118, + 0.001626406330615282, + 0.005563084501773119, + 0.00340132019482553, + 0.009259587153792381, + 0.0052913338877260685, + 0.0026476543862372637, + 0.00477716326713562, + 0.004381207749247551, + 0.002807888202369213, + 0.005881267134100199, + 0.003505631582811475, + 0.007569280453026295, + 0.07009823620319366, + 0.017602618783712387, + 0.016800427809357643, + 0.014212217181921005, + 0.014515803195536137, + 0.020220279693603516, + 0.009168572723865509, + 0.07779216766357422, + 0.006246848963201046, + 0.0017215540865436196, + 0.004987596068531275, + 0.0018312111496925354, + 0.0012509828666225076, + 0.0032160745467990637, + 0.0015957988798618317, + 0.008394993841648102, + 0.004187498241662979, + 0.003599831135943532, + 0.004454980604350567, + 0.002951017813757062, + 0.003834341885522008, + 0.006129606626927853, + 0.002720627700909972, + 0.005843285471200943, + 0.004102690145373344, + 0.002400569384917617, + 0.004144884645938873, + 0.0019763975869864225, + 0.002429185202345252, + 0.0034318079706281424, + 0.0023023164831101894, + 0.002676907693967223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.010458542965352535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001053969725035131 + ], + "opposite_logprob_tokens": [ + -6.876053810119629 + ], + "logprob_reward": -0.001053969725035131, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001053969725035131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39864909648895264, + 0.061940666288137436, + 0.2808123528957367, + 0.10302076488733292, + 0.0369974747300148, + 0.03224163129925728, + 0.038770709186792374, + 0.017462575808167458, + 0.015851419419050217, + 0.014253386296331882 + ], + "credit_map": [ + 0.055899620056152344, + 0.007163258269429207, + 0.009203352965414524, + 0.008987494744360447, + 0.00900875311344862, + 0.1671116203069687, + 0.008513273671269417, + 0.13276167213916779, + 0.006498151924461126, + 0.002371091628447175, + 0.002211879938840866, + 0.002515861764550209, + 0.0019052042625844479, + 0.023195773363113403, + 0.0021130989771336317, + 0.02112959697842598, + 0.026476841419935226, + 0.006073388271033764, + 0.012364329770207405, + 0.011537731625139713, + 0.0074133737944066525, + 0.10555597394704819, + 0.006997333839535713, + 0.1043933779001236, + 0.031922969967126846, + 0.003700487781316042, + 0.008390040136873722, + 0.005899875424802303, + 0.0022919815964996815, + 0.010981188155710697, + 0.003015707479789853, + 0.036818504333496094, + 0.006862849462777376, + 0.003054010448977351, + 0.005348837003111839, + 0.004893769044429064, + 0.0035005740355700254, + 0.004460153169929981, + 0.003566583851352334, + 0.005310693755745888, + 0.008525699377059937, + 0.001960878260433674, + 0.0024165413342416286, + 0.0022197673097252846, + 0.0026899264194071293, + 0.0032924083061516285, + 0.002673865295946598, + 0.008462540805339813, + 0.008257738314568996, + 0.003679104382172227, + 0.003243762068450451, + 0.0037094566505402327, + 0.0036174089182168245, + 0.00482200225815177, + 0.002158781047910452, + 0.009282450191676617, + 0.005041256546974182, + 0.0007641713600605726, + 0.0014734864234924316, + 0.0007521514780819416, + 0.0007669449551030993, + 0.001921993331052363, + 0.0008738106116652489, + 0.005868758074939251, + 0.002744601108133793, + 0.0020035579800605774, + 0.0017272245604544878, + 0.001654918654821813, + 0.0022945301607251167, + 0.0020423338282853365, + 0.0016629839083179832, + 0.001721266657114029, + 0.002931965049356222, + 0.0014172315131872892, + 0.0015923155006021261, + 0.0012500694720074534, + 0.0020864037796854973, + 0.002066971268504858, + 0.0013062277575954795, + 0.0016021996270865202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.001053969725035131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206347987055779 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206347987055779, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206347987055779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43327972292900085, + 0.09329486638307571, + 0.23467634618282318, + 0.08836699277162552, + 0.030558641999959946, + 0.04407421872019768, + 0.027769796550273895, + 0.018215106800198555, + 0.01644989475607872, + 0.013314441777765751 + ], + "credit_map": [ + 0.05693312734365463, + 0.006659894250333309, + 0.008889471180737019, + 0.008745826780796051, + 0.009366442449390888, + 0.18283671140670776, + 0.008527752012014389, + 0.15132048726081848, + 0.010405127890408039, + 0.001605072757229209, + 0.001910015824250877, + 0.0017819453496485949, + 0.0026168995536863804, + 0.037639547139406204, + 0.001976984553039074, + 0.03535926714539528, + 0.022704705595970154, + 0.005220694467425346, + 0.00839306227862835, + 0.00865766778588295, + 0.0066948155872523785, + 0.08990621566772461, + 0.005726905539631844, + 0.08737223595380783, + 0.027628779411315918, + 0.003593643894419074, + 0.007495853118598461, + 0.004679723177105188, + 0.0032128102611750364, + 0.008829010650515556, + 0.0030467535834759474, + 0.029880408197641373, + 0.006474462803453207, + 0.0024378602392971516, + 0.002899467246606946, + 0.003251882502809167, + 0.0035416376776993275, + 0.002992036985233426, + 0.002960208337754011, + 0.006001082714647055, + 0.012972251512110233, + 0.002214265987277031, + 0.0027769922744482756, + 0.002463478595018387, + 0.002847291063517332, + 0.004081255756318569, + 0.0029285233467817307, + 0.013790154829621315, + 0.003848956897854805, + 0.004086038563400507, + 0.004785254597663879, + 0.004171327687799931, + 0.002864629030227661, + 0.002830896060913801, + 0.0018835937371477485, + 0.0032990979962050915, + 0.004564960021525621, + 0.0010779771255329251, + 0.002172329928725958, + 0.0009905609767884016, + 0.0009546818910166621, + 0.001557359704747796, + 0.001198309357278049, + 0.005698926746845245, + 0.0018419537227600813, + 0.002355779055505991, + 0.001804207218810916, + 0.0020467687863856554, + 0.001917484332807362, + 0.0017678204458206892, + 0.0016677158419042826, + 0.003048163140192628, + 0.0025483588688075542, + 0.0017177251866087317, + 0.0013994068140164018, + 0.001342757954262197, + 0.0016488643595948815, + 0.0018261487130075693, + 0.0012933325488120317, + 0.0015378467505797744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014206347987055779 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195182360708714 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195182360708714, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195182360708714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19895382225513458, + 0.1072382926940918, + 0.21403075754642487, + 0.013259121216833591, + 0.014845771715044975, + 0.026409519836306572, + 0.34665435552597046, + 0.017454516142606735, + 0.017086992040276527, + 0.04406691715121269 + ], + "credit_map": [ + 0.030259111896157265, + 0.003654307220131159, + 0.003986272029578686, + 0.0039096432738006115, + 0.004721455276012421, + 0.08354382961988449, + 0.004130143206566572, + 0.06474903970956802, + 0.013467038050293922, + 0.0017413784516975284, + 0.002713203663006425, + 0.0019243238493800163, + 0.0028403736650943756, + 0.04385247454047203, + 0.0019931329879909754, + 0.03870635852217674, + 0.021155742928385735, + 0.004010884091258049, + 0.0071637751534581184, + 0.005827391054481268, + 0.005644690245389938, + 0.0838649570941925, + 0.005227716639637947, + 0.0811355784535408, + 0.0012757101794704795, + 0.0011214526602998376, + 0.0014871270395815372, + 0.0008959288243204355, + 0.0020151573698967695, + 0.0024560661986470222, + 0.00094279256882146, + 0.0030648864340037107, + 0.0020530824549496174, + 0.0009485652553848922, + 0.0014272996922954917, + 0.0018283125245943666, + 0.0016651825280860066, + 0.0030203512869775295, + 0.0018699058564379811, + 0.0020330704282969236, + 0.007303517311811447, + 0.0018573000561445951, + 0.001726278685964644, + 0.0019720331765711308, + 0.001993149984627962, + 0.002387228887528181, + 0.0023217422422021627, + 0.006848270073533058, + 0.10026191920042038, + 0.025468919426202774, + 0.025729846209287643, + 0.0205769594758749, + 0.021639235317707062, + 0.029334675520658493, + 0.013257493264973164, + 0.11038530617952347, + 0.003881391603499651, + 0.001433771918527782, + 0.0014572239015251398, + 0.0010384350316599011, + 0.0016916606109589338, + 0.0021016576793044806, + 0.0014398485654965043, + 0.004410523921251297, + 0.0017249885713681579, + 0.002451761858537793, + 0.0018542882753536105, + 0.0023070164024829865, + 0.0021788631565868855, + 0.0020842396188527346, + 0.00211139558814466, + 0.0023744397331029177, + 0.004148317966610193, + 0.003569781081750989, + 0.005352622363716364, + 0.003703785827383399, + 0.005714474245905876, + 0.006721676327288151, + 0.004675996024161577, + 0.010180264711380005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.014195182360708714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603587567806244 + ], + "opposite_logprob_tokens": [ + -1.9103587865829468 + ], + "logprob_reward": -0.1603587567806244, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603587567806244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2204606980085373, + 0.12169401347637177, + 0.06976170837879181, + 0.05771064758300781, + 0.06572851538658142, + 0.017444614320993423, + 0.3601985573768616, + 0.032256826758384705, + 0.039013609290122986, + 0.015730824321508408 + ], + "credit_map": [ + 0.029639998450875282, + 0.004349462687969208, + 0.004757289309054613, + 0.005586574785411358, + 0.005282471422106028, + 0.0913303941488266, + 0.005449612624943256, + 0.07406490296125412, + 0.020775016397237778, + 0.0026819545309990644, + 0.002472413470968604, + 0.0024390630424022675, + 0.0034528819378465414, + 0.049988266080617905, + 0.0024621025659143925, + 0.037422310560941696, + 0.006679581943899393, + 0.001989382319152355, + 0.0024054700043052435, + 0.002039822516962886, + 0.002353629097342491, + 0.026580531150102615, + 0.002037441823631525, + 0.025675849989056587, + 0.014386686496436596, + 0.0033438976388424635, + 0.008242132142186165, + 0.004705019295215607, + 0.002149965614080429, + 0.00563223147764802, + 0.0026948354206979275, + 0.01655587926506996, + 0.011672692373394966, + 0.0047468398697674274, + 0.008804368786513805, + 0.008330202661454678, + 0.005496211349964142, + 0.007351941894739866, + 0.007016815710812807, + 0.01230944599956274, + 0.003036936279386282, + 0.001434887177310884, + 0.0019629360176622868, + 0.002318501938134432, + 0.0013079748023301363, + 0.002318798331543803, + 0.0012222648365423083, + 0.003842313541099429, + 0.10474392026662827, + 0.027655955404043198, + 0.0250614695250988, + 0.022450849413871765, + 0.02323700673878193, + 0.02911527082324028, + 0.013805374503135681, + 0.11412869393825531, + 0.00804218277335167, + 0.0021175534930080175, + 0.0031515921000391245, + 0.0018147930968552828, + 0.0024727743584662676, + 0.0028252555057406425, + 0.002246292307972908, + 0.00958638172596693, + 0.004638672806322575, + 0.002929716370999813, + 0.006748337298631668, + 0.002926397603005171, + 0.004806621000170708, + 0.007176392711699009, + 0.005610826890915632, + 0.004176643677055836, + 0.0016461493214592338, + 0.0014682423789054155, + 0.0032526778522878885, + 0.0013973851455375552, + 0.0014683662448078394, + 0.002193815540522337, + 0.0022515146993100643, + 0.0020526719745248556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.1603587567806244 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "hyperreal 3D rendering captures a time-lapse sequence from a static, low-angle macro shot focused on a desert rose (Adenium obesum). The plant's gnarled, swollen stem base, resembling aged, rugged bark, is firmly rooted in fine, pale, sandy soil. A single, powerful light source moves across the scene, creating dramatic, deep shadows that constantly shift and lengthen, emphasizing the plant's form. As the time-lapse progresses, a single, magnificent flower begins to unfold from a tight bud; its petals, boasting a realistic waxy texture, steadily and gracefully open, revealing a vibrant color gradient from a saturated, deep pink at the edges to a soft, pure white near the flower's throat. Dynamic light flickers from the petals' waxy surface, capturing tiny grains of crystallized sand scattered along the stem and highlighting every minute detail, creating a focused and dramatic atmosphere that celebrates the slow, beautiful spectacle of life.", + "seed": 18723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_12/index_1891_seed_18723_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02390286438167095, + "vlm_logprob": -0.02390286438167095, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.02390286438167095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018290545791387558 + ], + "opposite_logprob_tokens": [ + -5.5182905197143555 + ], + "logprob_reward": -0.018290545791387558, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.018290545791387558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4131137728691101, + 0.01684132218360901, + 0.010214355774223804, + 0.17911161482334137, + 0.26162809133529663, + 0.011988678015768528, + 0.0469534657895565, + 0.023276526480913162, + 0.01550566591322422, + 0.021366538479924202 + ], + "credit_map": [ + 0.16588596999645233, + 0.011555048637092113, + 0.013452211394906044, + 0.011728060431778431, + 0.013435610570013523, + 0.173341304063797, + 0.009791935794055462, + 0.01392363477498293, + 0.002551032230257988, + 0.0018659117631614208, + 0.0023514984641224146, + 0.0018287806306034327, + 0.0020593239460140467, + 0.002960227895528078, + 0.001744816661812365, + 0.001479730592109263, + 0.0021728521678596735, + 0.0016720453277230263, + 0.0007847605156712234, + 0.0009155551670119166, + 0.0011697878362610936, + 0.0017774306470528245, + 0.000805916846729815, + 0.0009160064510069788, + 0.06553052365779877, + 0.011321394704282284, + 0.006091854069381952, + 0.006946095265448093, + 0.010718992911279202, + 0.06606100499629974, + 0.005286127328872681, + 0.007155623286962509, + 0.05177531763911247, + 0.021362626925110817, + 0.05378631129860878, + 0.01829822175204754, + 0.021337540820240974, + 0.026568477973341942, + 0.019372081384062767, + 0.04912750795483589, + 0.0017178619746118784, + 0.0014239426236599684, + 0.0011742353672161698, + 0.0013199214590713382, + 0.0016445579240098596, + 0.0016177186043933034, + 0.0014411879237741232, + 0.001649251556955278, + 0.004929446615278721, + 0.0024406781885772943, + 0.012170900590717793, + 0.002015582052990794, + 0.003238316858187318, + 0.014386697672307491, + 0.0023570270277559757, + 0.005414817947894335, + 0.0035522279795259237, + 0.0039930762723088264, + 0.0021309612784534693, + 0.002690626308321953, + 0.0024442633148282766, + 0.002926269080489874, + 0.0020861285738646984, + 0.0034529739059507847, + 0.00118070503231138, + 0.001750026480294764, + 0.0013754393439739943, + 0.0035113398917019367, + 0.0013901104684919119, + 0.002154937479645014, + 0.0017393177840858698, + 0.002403789199888706, + 0.002136687282472849, + 0.0028033035341650248, + 0.0022974966559559107, + 0.0035338597372174263, + 0.0030962093733251095, + 0.0031891262624412775, + 0.002179661300033331, + 0.002130194567143917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.018290545791387558 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022082962095737457 + ], + "opposite_logprob_tokens": [ + -6.127208232879639 + ], + "logprob_reward": -0.0022082962095737457, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022082962095737457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8103842735290527, + 0.02341175451874733, + 0.006243899464607239, + 0.0988721027970314, + 0.006034551188349724, + 0.007682991214096546, + 0.022499775514006615, + 0.008053798228502274, + 0.003929559141397476, + 0.012887280434370041 + ], + "credit_map": [ + 0.3347678780555725, + 0.0226015355437994, + 0.017080416902899742, + 0.020702963694930077, + 0.02625400759279728, + 0.3486005961894989, + 0.01751302368938923, + 0.022863760590553284, + 0.0057392786256968975, + 0.0015661661745980382, + 0.0015793724451214075, + 0.0013533835299313068, + 0.001491697272285819, + 0.006195760332047939, + 0.001863033277913928, + 0.0036230587866157293, + 0.0013235218357294798, + 0.0008553255465812981, + 0.0007104624528437853, + 0.0004601091786753386, + 0.0008572700317017734, + 0.001150123425759375, + 0.0004110859299544245, + 0.0004760002775583416, + 0.04054443538188934, + 0.00414513424038887, + 0.002302107634022832, + 0.00241878186352551, + 0.003576381132006645, + 0.04031071811914444, + 0.002482146490365267, + 0.0030923811718821526, + 0.0010114433243870735, + 0.0006544645875692368, + 0.0010565542615950108, + 0.0006975200376473367, + 0.000621455954387784, + 0.0005964839947409928, + 0.0005459741805680096, + 0.0008506537997163832, + 0.000940190686378628, + 0.0013358879368752241, + 0.0008147222688421607, + 0.0009402030846104026, + 0.001176960184238851, + 0.0009874814422801137, + 0.0007219771505333483, + 0.0007655672961845994, + 0.00217198533937335, + 0.0014481308171525598, + 0.0058380113914608955, + 0.001168292248621583, + 0.001489895861595869, + 0.006964819971472025, + 0.0009805718436837196, + 0.0024380660615861416, + 0.0012941983295604587, + 0.0014250832609832287, + 0.0006700793164782226, + 0.0009618783951736987, + 0.0008269752142950892, + 0.000964779348578304, + 0.0006728675798512995, + 0.0012379356194287539, + 0.0005611921078525484, + 0.0005884450511075556, + 0.0004888080875389278, + 0.0004675667150877416, + 0.0005098497495055199, + 0.0005062050768174231, + 0.0003982420894317329, + 0.0004092497401870787, + 0.0015563605120405555, + 0.001464587403461337, + 0.0011275818105787039, + 0.0014911359176039696, + 0.002014164114370942, + 0.0025636013597249985, + 0.0012837424874305725, + 0.0013861061306670308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0022082962095737457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595959834754467 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.00595959834754467, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595959834754467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06484175473451614, + 0.018675724044442177, + 0.01472298800945282, + 0.32344767451286316, + 0.3369297683238983, + 0.028023798018693924, + 0.07003095746040344, + 0.043572261929512024, + 0.02767014130949974, + 0.07208491861820221 + ], + "credit_map": [ + 0.022432005032896996, + 0.003300068201497197, + 0.0025326055474579334, + 0.003180423052981496, + 0.0034438427537679672, + 0.024047276005148888, + 0.0028276739176362753, + 0.0030778623186051846, + 0.003418814856559038, + 0.0020926871802657843, + 0.002351760398596525, + 0.0018764120759442449, + 0.0017667703796178102, + 0.0034704087302088737, + 0.0014803789090365171, + 0.0022184900008141994, + 0.0026573173236101866, + 0.0017202043673023582, + 0.0017731826519593596, + 0.0011583872837945819, + 0.001411237521097064, + 0.0030227270908653736, + 0.001044725300744176, + 0.0019352049566805363, + 0.1213277131319046, + 0.01895754598081112, + 0.010101626627147198, + 0.011170844547450542, + 0.019485529512166977, + 0.12130938470363617, + 0.009350979700684547, + 0.011744039133191109, + 0.0648254007101059, + 0.02609139122068882, + 0.07084749639034271, + 0.024468686431646347, + 0.024848898872733116, + 0.035218745470047, + 0.024534139782190323, + 0.06609498709440231, + 0.0032902532257139683, + 0.005488989409059286, + 0.0024956639390438795, + 0.0033348079305142164, + 0.0048216586001217365, + 0.003661243710666895, + 0.002354737836867571, + 0.002576443599537015, + 0.007391708437353373, + 0.004349027294665575, + 0.017064498737454414, + 0.0037806397303938866, + 0.004928164184093475, + 0.020901553332805634, + 0.0033700019121170044, + 0.00824536569416523, + 0.007071126718074083, + 0.005300013814121485, + 0.0037066438235342503, + 0.004666995722800493, + 0.004792532883584499, + 0.006184061523526907, + 0.004689345136284828, + 0.007161544170230627, + 0.003660521237179637, + 0.004118205048143864, + 0.0029200573917478323, + 0.0038591031916439533, + 0.0028833968099206686, + 0.003744998248293996, + 0.002238457091152668, + 0.004245399497449398, + 0.01027116272598505, + 0.00752638652920723, + 0.007988132536411285, + 0.007686258759349585, + 0.009333538822829723, + 0.014727634377777576, + 0.006986867170780897, + 0.007564937230199575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.00595959834754467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097184151411057 + ], + "opposite_logprob_tokens": [ + -3.2880971431732178 + ], + "logprob_reward": -0.038097184151411057, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097184151411057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4767264425754547, + 0.024081546813249588, + 0.01226416602730751, + 0.09626293182373047, + 0.18888956308364868, + 0.02693168632686138, + 0.04115535318851471, + 0.030350837856531143, + 0.019114665687084198, + 0.08422282338142395 + ], + "credit_map": [ + 0.19736750423908234, + 0.013445897959172726, + 0.010195426642894745, + 0.012057177722454071, + 0.01517426036298275, + 0.20476356148719788, + 0.010475990362465382, + 0.013246634975075722, + 0.0037338121328502893, + 0.00371932378038764, + 0.002361741615459323, + 0.003032837761566043, + 0.003246359061449766, + 0.003585407743230462, + 0.0022540695499628782, + 0.0021479963324964046, + 0.0017300277249887586, + 0.0018304355908185244, + 0.0016301613068208098, + 0.0010480930795893073, + 0.0017841716762632132, + 0.0018764170818030834, + 0.001110532321035862, + 0.001254327711649239, + 0.038045432418584824, + 0.004257700871676207, + 0.0030099316500127316, + 0.0030366925057023764, + 0.004200001712888479, + 0.03786838799715042, + 0.0025912069249898195, + 0.0032535758800804615, + 0.037506990134716034, + 0.015690570697188377, + 0.038441047072410583, + 0.012215496972203255, + 0.01540827564895153, + 0.02023899368941784, + 0.013835315592586994, + 0.03555287793278694, + 0.0035268182400614023, + 0.004337436985224485, + 0.002891472540795803, + 0.00372143997810781, + 0.004351063631474972, + 0.003296853741630912, + 0.002538491738960147, + 0.002268110401928425, + 0.004474921617656946, + 0.0038280219305306673, + 0.009836502373218536, + 0.003108202712610364, + 0.0033611857797950506, + 0.010584190487861633, + 0.002029815223067999, + 0.003932511433959007, + 0.004114464856684208, + 0.00663188099861145, + 0.0027539816219359636, + 0.003963756840676069, + 0.00310046155937016, + 0.003367919707670808, + 0.0026246006600558758, + 0.003793771378695965, + 0.0020708225201815367, + 0.0022717048414051533, + 0.001707435236312449, + 0.002144008409231901, + 0.003409056691452861, + 0.0025235081557184458, + 0.0025719203986227512, + 0.002416208852082491, + 0.00855020247399807, + 0.00905546173453331, + 0.009345118887722492, + 0.010573631152510643, + 0.012750624679028988, + 0.016019703820347786, + 0.008469460532069206, + 0.009458622895181179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.038097184151411057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495869740843773 + ], + "opposite_logprob_tokens": [ + -2.9299585819244385 + ], + "logprob_reward": -0.05495869740843773, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495869740843773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21107690036296844, + 0.004402554128319025, + 0.007355120498687029, + 0.5239436030387878, + 0.22384756803512573, + 0.00691137695685029, + 0.009177599102258682, + 0.005235735792666674, + 0.005878554657101631, + 0.002170971129089594 + ], + "credit_map": [ + 0.08635060489177704, + 0.005853680893778801, + 0.0052308328449726105, + 0.005703547038137913, + 0.00689404783770442, + 0.09026926010847092, + 0.004650735761970282, + 0.0061242044903337955, + 0.000730280124116689, + 0.0005495349760167301, + 0.0005401961971074343, + 0.00039137116982601583, + 0.0006257877103053033, + 0.00069043057737872, + 0.0003951959952246398, + 0.0004797572910320014, + 0.0008909401949495077, + 0.0012469502398744226, + 0.000980905257165432, + 0.0006989799439907074, + 0.0008714321302250028, + 0.0009153116843663156, + 0.0007105944096110761, + 0.0010400066385045648, + 0.1968853622674942, + 0.028920751065015793, + 0.0163965355604887, + 0.01847086474299431, + 0.02950163558125496, + 0.19817271828651428, + 0.014938630163669586, + 0.02065713331103325, + 0.04288613423705101, + 0.01950596459209919, + 0.04562053084373474, + 0.014820928685367107, + 0.020078133791685104, + 0.02328047715127468, + 0.01609400473535061, + 0.041561391204595566, + 0.0010919036576524377, + 0.001040968461893499, + 0.0008583232993260026, + 0.0010036982130259275, + 0.0010432782582938671, + 0.0005526794120669365, + 0.000799378554802388, + 0.0005211468087509274, + 0.0007799291051924229, + 0.0005522872088477015, + 0.0023671670351177454, + 0.0004830069374293089, + 0.00073747627902776, + 0.002936109434813261, + 0.0003775796212721616, + 0.000944044440984726, + 0.0010207218583673239, + 0.0006715909112244844, + 0.0005170422955416143, + 0.00043854914838448167, + 0.00036594655830413103, + 0.0008017353247851133, + 0.00039863670826889575, + 0.001021512900479138, + 0.0007403583731502295, + 0.0007415086147375405, + 0.0007422463386319578, + 0.0009699063957668841, + 0.0006560470210388303, + 0.0007384273922070861, + 0.0007686783210374415, + 0.0005213823169469833, + 0.0003682780952658504, + 0.0002719427866395563, + 0.0002471617190167308, + 0.0003545404179021716, + 0.0002226842480013147, + 0.0002761933137662709, + 0.00017416082846466452, + 0.0002560096327215433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.05495869740843773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "hyperreal 3D rendering captures a time-lapse sequence from a static, low-angle macro shot focused on a desert rose (Adenium obesum). The plant's gnarled, swollen stem base, resembling aged, rugged bark, is firmly rooted in fine, pale, sandy soil. A single, powerful light source moves across the scene, creating dramatic, deep shadows that constantly shift and lengthen, emphasizing the plant's form. As the time-lapse progresses, a single, magnificent flower begins to unfold from a tight bud; its petals, boasting a realistic waxy texture, steadily and gracefully open, revealing a vibrant color gradient from a saturated, deep pink at the edges to a soft, pure white near the flower's throat. Dynamic light flickers from the petals' waxy surface, capturing tiny grains of crystallized sand scattered along the stem and highlighting every minute detail, creating a focused and dramatic atmosphere that celebrates the slow, beautiful spectacle of life.", + "seed": 18724, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_12/index_1891_seed_18724_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03082695915363729, + "vlm_logprob": -0.03082695915363729, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03082695915363729, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0184018537402153 + ], + "opposite_logprob_tokens": [ + -4.768401622772217 + ], + "logprob_reward": -0.0184018537402153, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0184018537402153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2868424952030182, + 0.08560023456811905, + 0.07982108741998672, + 0.10982681810855865, + 0.04826981574296951, + 0.07027643918991089, + 0.08591559529304504, + 0.11361810564994812, + 0.06382860988378525, + 0.05600083991885185 + ], + "credit_map": [ + 0.04249376431107521, + 0.044193532317876816, + 0.03539257496595383, + 0.042071424424648285, + 0.02368483692407608, + 0.02589099295437336, + 0.023464344441890717, + 0.04965101554989815, + 0.009720769710838795, + 0.012422967702150345, + 0.009872092865407467, + 0.007470062468200922, + 0.01110947784036398, + 0.01291775330901146, + 0.008993183262646198, + 0.013093930669128895, + 0.0176199059933424, + 0.0077729178592562675, + 0.008213931694626808, + 0.008305108174681664, + 0.00703768664970994, + 0.019585449248552322, + 0.0049349465407431126, + 0.0063511403277516365, + 0.01567867398262024, + 0.007645040284842253, + 0.01632419042289257, + 0.007957124151289463, + 0.009346176870167255, + 0.015695886686444283, + 0.007043114863336086, + 0.03013661503791809, + 0.007751064840704203, + 0.004995882045477629, + 0.00667797215282917, + 0.005861489102244377, + 0.005685695447027683, + 0.00691887317225337, + 0.0034862582106143236, + 0.006892581004649401, + 0.00933810044080019, + 0.0057992031797766685, + 0.013319078832864761, + 0.008604771457612514, + 0.005907257087528706, + 0.006596227642148733, + 0.006039103493094444, + 0.01467269565910101, + 0.006681549828499556, + 0.01854776218533516, + 0.009303412400186062, + 0.00941929966211319, + 0.014545704238116741, + 0.009260980412364006, + 0.010675134137272835, + 0.007481748703867197, + 0.012445073574781418, + 0.01094381045550108, + 0.03116697631776333, + 0.010105262510478497, + 0.008692958392202854, + 0.011472745798528194, + 0.010171926580369473, + 0.018619343638420105, + 0.00972974393516779, + 0.00557452067732811, + 0.008885348215699196, + 0.013709494844079018, + 0.0069459350779652596, + 0.004051248077303171, + 0.007493758108466864, + 0.0074385544285178185, + 0.007188256364315748, + 0.011087028309702873, + 0.005449110176414251, + 0.006242436822503805, + 0.007442377973347902, + 0.004943151958286762, + 0.007832661271095276, + 0.0058158161118626595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0184018537402153 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028278145473450422 + ], + "opposite_logprob_tokens": [ + -5.8778276443481445 + ], + "logprob_reward": -0.0028278145473450422, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028278145473450422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22757983207702637, + 0.0906737893819809, + 0.12436534464359283, + 0.09130413085222244, + 0.04258705675601959, + 0.16458266973495483, + 0.04616392031311989, + 0.06176020950078964, + 0.12424293160438538, + 0.026740089058876038 + ], + "credit_map": [ + 0.02505131997168064, + 0.04364802688360214, + 0.02955670841038227, + 0.031331323087215424, + 0.023754065856337547, + 0.02387974224984646, + 0.024355746805667877, + 0.026002895087003708, + 0.016756724566221237, + 0.011074504815042019, + 0.011214641854166985, + 0.007150245830416679, + 0.012171042151749134, + 0.012880997732281685, + 0.008142379112541676, + 0.011283254250884056, + 0.03800215944647789, + 0.006585270632058382, + 0.015760265290737152, + 0.005643041804432869, + 0.006417477037757635, + 0.03214126080274582, + 0.004983000922948122, + 0.014832873828709126, + 0.02514059841632843, + 0.0064590247347950935, + 0.01234684232622385, + 0.004437671974301338, + 0.005865420680493116, + 0.01732674613595009, + 0.004548023454844952, + 0.015179799869656563, + 0.0060976711101830006, + 0.004517446272075176, + 0.006418677978217602, + 0.004678704310208559, + 0.005719610955566168, + 0.005482233129441738, + 0.0035117389634251595, + 0.00616097915917635, + 0.01385462749749422, + 0.008706046268343925, + 0.05694403126835823, + 0.011197846382856369, + 0.006645048502832651, + 0.011960964649915695, + 0.010011708363890648, + 0.0452624075114727, + 0.006214073859155178, + 0.005675195716321468, + 0.00877237506210804, + 0.0051046377047896385, + 0.004747940227389336, + 0.0061485422775149345, + 0.004064327571541071, + 0.0054368251003324986, + 0.006724075879901648, + 0.005623417906463146, + 0.019654953852295876, + 0.0047221798449754715, + 0.003486824221909046, + 0.00491617526859045, + 0.004951410926878452, + 0.01168117206543684, + 0.025896767154335976, + 0.011130056343972683, + 0.020230280235409737, + 0.022083431482315063, + 0.012532366439700127, + 0.008691919036209583, + 0.012699799612164497, + 0.010978310368955135, + 0.0030233096331357956, + 0.00556153291836381, + 0.002604040550068021, + 0.0026190713979303837, + 0.004260610789060593, + 0.0023187154438346624, + 0.0026793733704835176, + 0.0036734347231686115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0028278145473450422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300768807530403 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023300768807530403, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300768807530403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13718466460704803, + 0.05364856868982315, + 0.06126948818564415, + 0.06549546122550964, + 0.07262963801622391, + 0.3183627128601074, + 0.06362660229206085, + 0.03555905446410179, + 0.14295543730258942, + 0.04926835373044014 + ], + "credit_map": [ + 0.01704700104892254, + 0.01729007251560688, + 0.022657176479697227, + 0.016040511429309845, + 0.012493520975112915, + 0.01985730417072773, + 0.011074280366301537, + 0.02072479948401451, + 0.0083342669531703, + 0.007168106269091368, + 0.0067862868309021, + 0.00633261539041996, + 0.004651661962270737, + 0.007048118859529495, + 0.00653717340901494, + 0.006790337152779102, + 0.00739305792376399, + 0.006147188600152731, + 0.007659312337636948, + 0.0065659391693770885, + 0.005980462767183781, + 0.006594185251742601, + 0.005418951157480478, + 0.015510386787354946, + 0.007262098602950573, + 0.006474614143371582, + 0.013113991357386112, + 0.005733024794608355, + 0.006600700784474611, + 0.010703439824283123, + 0.0053428299725055695, + 0.010264763608574867, + 0.00687901908531785, + 0.010177113115787506, + 0.007349628023803234, + 0.012358037754893303, + 0.01087059173732996, + 0.009753094986081123, + 0.00950960349291563, + 0.005732549354434013, + 0.033172376453876495, + 0.015132217667996883, + 0.11222957074642181, + 0.021835079416632652, + 0.012900358997285366, + 0.022552911192178726, + 0.019307758659124374, + 0.0812324583530426, + 0.008890406228601933, + 0.013334705494344234, + 0.00905807688832283, + 0.005204635672271252, + 0.009217943996191025, + 0.003843535203486681, + 0.006877122446894646, + 0.007200177293270826, + 0.0032279975712299347, + 0.004710529465228319, + 0.004865000024437904, + 0.004939827136695385, + 0.004376855678856373, + 0.0037203868851065636, + 0.005093196872621775, + 0.004625261295586824, + 0.029395457357168198, + 0.012049151584506035, + 0.0228725615888834, + 0.02664063312113285, + 0.013694975525140762, + 0.01019784715026617, + 0.01485482882708311, + 0.013249990530312061, + 0.004975533112883568, + 0.011211230419576168, + 0.003819478442892432, + 0.006117573007941246, + 0.009550937451422215, + 0.003852458903566003, + 0.004908609203994274, + 0.0048325336538255215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.023300768807530403 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11688879877328873, + 0.08433444797992706, + 0.04820895940065384, + 0.07255306094884872, + 0.14434322714805603, + 0.18672043085098267, + 0.12046060711145401, + 0.02855423279106617, + 0.1576913446187973, + 0.04024485498666763 + ], + "credit_map": [ + 0.020918935537338257, + 0.01391284167766571, + 0.0144354784861207, + 0.011701961979269981, + 0.01122687105089426, + 0.01493498682975769, + 0.008459311909973621, + 0.021298421546816826, + 0.016795264557003975, + 0.008507943712174892, + 0.01185820996761322, + 0.005904289428144693, + 0.009599233977496624, + 0.009722555987536907, + 0.008659854531288147, + 0.013287102803587914, + 0.008425610139966011, + 0.006164760794490576, + 0.004446160048246384, + 0.005661848932504654, + 0.005965021438896656, + 0.008087954483926296, + 0.004841919057071209, + 0.00461568683385849, + 0.012587742879986763, + 0.009197776205837727, + 0.0070874434895813465, + 0.004815545864403248, + 0.007532159332185984, + 0.01815365068614483, + 0.005200729705393314, + 0.007978018373250961, + 0.02043777145445347, + 0.01888013631105423, + 0.016587311401963234, + 0.017803022637963295, + 0.0201956145465374, + 0.02430427446961403, + 0.013978300616145134, + 0.012156805023550987, + 0.02691972441971302, + 0.012459427118301392, + 0.06291721016168594, + 0.011800170876085758, + 0.012594766914844513, + 0.010283670388162136, + 0.009029537439346313, + 0.040715958923101425, + 0.011465267278254032, + 0.01996912807226181, + 0.011868268251419067, + 0.015581488609313965, + 0.0203092023730278, + 0.014808138832449913, + 0.015692301094532013, + 0.010766824707388878, + 0.002437407849356532, + 0.0026186571922153234, + 0.007608642335981131, + 0.002350004157051444, + 0.0027646480593830347, + 0.004030486568808556, + 0.0020777112804353237, + 0.004666678141802549, + 0.03322266787290573, + 0.01254959125071764, + 0.029450489208102226, + 0.02651369199156761, + 0.015220160596072674, + 0.011337303556501865, + 0.014620578847825527, + 0.01477687805891037, + 0.00500614196062088, + 0.009886397048830986, + 0.00452085118740797, + 0.004480164032429457, + 0.006409012246876955, + 0.002690703608095646, + 0.003604827681556344, + 0.0036467614118009806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.020577870309352875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle macro view focused on a desert rose (Adenium obesum)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the plant's gnarled, swollen stem base resemble aged, rugged bark rooted in fine, pale, sandy soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single flower unfold from a tight bud, revealing petals with a realistic waxy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flower petals display a color gradient from saturated deep pink at the edges to soft pure white near the throat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single powerful light source move across the scene, creating shifting deep shadows and highlighting crystallized sand grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902648836374283 + ], + "opposite_logprob_tokens": [ + -2.46402645111084 + ], + "logprob_reward": -0.08902648836374283, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902648836374283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22017689049243927, + 0.08744613826274872, + 0.027957230806350708, + 0.10118556767702103, + 0.035649292171001434, + 0.06496141105890274, + 0.044996026903390884, + 0.06574767082929611, + 0.3253513276576996, + 0.026528483256697655 + ], + "credit_map": [ + 0.03178117796778679, + 0.0358128547668457, + 0.02868831530213356, + 0.02985566295683384, + 0.019706396386027336, + 0.019419163465499878, + 0.0190165713429451, + 0.03589673712849617, + 0.015468531288206577, + 0.00824745837599039, + 0.012520195916295052, + 0.009463449940085411, + 0.008487064391374588, + 0.009097689762711525, + 0.008066956885159016, + 0.016094790771603584, + 0.00337430858053267, + 0.0039383480325341225, + 0.004069171380251646, + 0.003415499348193407, + 0.004350582603365183, + 0.00300597888417542, + 0.003165026195347309, + 0.0026383160147815943, + 0.023780155926942825, + 0.010432356968522072, + 0.009853488765656948, + 0.007934190332889557, + 0.012258545495569706, + 0.01951323263347149, + 0.006649770308285952, + 0.010763830505311489, + 0.0043264878913760185, + 0.004489603918045759, + 0.004434001166373491, + 0.004236398730427027, + 0.004495799541473389, + 0.007017435040324926, + 0.003160842228680849, + 0.003488726681098342, + 0.006894751917570829, + 0.00793064758181572, + 0.008152754046022892, + 0.009715101681649685, + 0.008406964130699635, + 0.008428639732301235, + 0.008232333697378635, + 0.00720022339373827, + 0.006307574920356274, + 0.006395154632627964, + 0.008549661375582218, + 0.003562299069017172, + 0.0051843468099832535, + 0.0035934674087911844, + 0.004502872936427593, + 0.006900649983435869, + 0.006150180008262396, + 0.006067014764994383, + 0.021908963099122047, + 0.005349529907107353, + 0.003513470059260726, + 0.005272797308862209, + 0.004727950785309076, + 0.012757768854498863, + 0.06629219651222229, + 0.024786198511719704, + 0.0513484813272953, + 0.06630910933017731, + 0.03058379888534546, + 0.021423617377877235, + 0.03232073783874512, + 0.03228719159960747, + 0.0030937865376472473, + 0.0048104035668075085, + 0.002429607789963484, + 0.0033401004038751125, + 0.005180131644010544, + 0.0022215929348021746, + 0.002568677766248584, + 0.002884181449189782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.08902648836374283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_tensor([4883, 4883])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_tensor([4883, 4883])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c66c48040ecfc14f2d41ce2d11e14e1364f3c3f7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_tensor([4883, 4883])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "full-body, horizontal tracking shot follows a sleek, multi-rotor manned aircraft as it travels at a steady pace, on a straight course, from left to right through a nighttime urban canyon shrouded in torrential rain. Towering skyscrapers, with a distinctly pixelated texture, appear to be composed of giant, glowing blocks, their surfaces gleaming with low-resolution, animated advertisements. The aircraft's dark, composite hue gleams clammy in the rain, its outline sharply illuminated by the bright blue and magenta neon strips etched across its fuselage. As it advances, these neon lights cast long, digital-like trails through the humid air, hovering briefly between the pixelated buildings. The dense stream of rain streaks across the camera, intertwining with the equally pixelated, flickering lights of billboards atop the massive buildings in the background, blurring distant details. The scene is dominated by deep blue and purple hues, creating a cyberpunk atmosphere that feels both retro and futuristic, steeped in the aesthetics of 8-bit video games.", + "seed": 5496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_4883_seed_5496_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.045031317556276915, + "vlm_logprob": -0.045031317556276915, + "vlm_yes_no_margin": 3.874999976158142, + "vlm_token_credit": -0.045031317556276915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04130672663450241 + ], + "opposite_logprob_tokens": [ + -3.416306734085083 + ], + "logprob_reward": -0.04130672663450241, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04130672663450241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14977851510047913, + 0.06054249778389931, + 0.09567983448505402, + 0.11019320785999298, + 0.07586079090833664, + 0.14442451298236847, + 0.1108376756310463, + 0.06946628540754318, + 0.10259634256362915, + 0.08062027394771576 + ], + "credit_map": [ + 0.020883291959762573, + 0.011136515997350216, + 0.02726772241294384, + 0.015107442624866962, + 0.018358977511525154, + 0.02477392926812172, + 0.014831609092652798, + 0.017419029027223587, + 0.004144403617829084, + 0.003895327914506197, + 0.008575508370995522, + 0.007855875417590141, + 0.007339875679463148, + 0.01363219041377306, + 0.007644804660230875, + 0.007454514037817717, + 0.00966123677790165, + 0.006736129987984896, + 0.008994380943477154, + 0.009749664925038815, + 0.012422480620443821, + 0.01881343498826027, + 0.01089635118842125, + 0.018406154587864876, + 0.010965920984745026, + 0.011732381768524647, + 0.011798709630966187, + 0.01478351466357708, + 0.013952350243926048, + 0.01757834665477276, + 0.017734024673700333, + 0.011647962033748627, + 0.0035463557578623295, + 0.005827669985592365, + 0.0027494849637150764, + 0.004828510340303183, + 0.006101815029978752, + 0.02456607297062874, + 0.0034146413672715425, + 0.02482624351978302, + 0.012553339824080467, + 0.02255277894437313, + 0.011038986034691334, + 0.012012586928904057, + 0.02587699703872204, + 0.02601582184433937, + 0.017313601449131966, + 0.017060406506061554, + 0.01292507629841566, + 0.014422261156141758, + 0.005834612064063549, + 0.00722862733528018, + 0.014867524616420269, + 0.01999673806130886, + 0.01244993694126606, + 0.02311289682984352, + 0.010865094140172005, + 0.009543822146952152, + 0.005768157541751862, + 0.006089319474995136, + 0.009760327637195587, + 0.009481157176196575, + 0.00976814515888691, + 0.008190260268747807, + 0.013191360048949718, + 0.011030246503651142, + 0.008736265823245049, + 0.010911665856838226, + 0.00884736143052578, + 0.0248122401535511, + 0.011710969731211662, + 0.01335623487830162, + 0.013645217753946781, + 0.01547851413488388, + 0.008823138661682606, + 0.006069163326174021, + 0.016835641115903854, + 0.009198284707963467, + 0.004099547863006592, + 0.00647076778113842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.04130672663450241 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543240562081337 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543240562081337, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543240562081337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40653282403945923, + 0.09104202687740326, + 0.046427298337221146, + 0.06113961338996887, + 0.05486414581537247, + 0.09736998379230499, + 0.056601088494062424, + 0.04205785319209099, + 0.06085776537656784, + 0.08310738950967789 + ], + "credit_map": [ + 0.031094564124941826, + 0.036785658448934555, + 0.07143375277519226, + 0.05106772482395172, + 0.05427737534046173, + 0.04589034244418144, + 0.05247756093740463, + 0.06350582093000412, + 0.006946104113012552, + 0.007737429812550545, + 0.010347041301429272, + 0.011098518036305904, + 0.008765293285250664, + 0.02074558474123478, + 0.01108589582145214, + 0.01431615836918354, + 0.005539930425584316, + 0.004865975119173527, + 0.003969020675867796, + 0.0069454326294362545, + 0.006183599587529898, + 0.00685671903192997, + 0.004327833186835051, + 0.00773877976462245, + 0.004953397437930107, + 0.005138916429132223, + 0.004852265119552612, + 0.008329526521265507, + 0.009813274256885052, + 0.009473290294408798, + 0.008898498490452766, + 0.009680435992777348, + 0.0026528150774538517, + 0.003527884604409337, + 0.0021970593370497227, + 0.002952767303213477, + 0.005122425500303507, + 0.01824568398296833, + 0.001961115514859557, + 0.01820438727736473, + 0.011785470880568027, + 0.010761723853647709, + 0.006514563690871, + 0.005812142975628376, + 0.016744161024689674, + 0.01919591799378395, + 0.013084981590509415, + 0.013471012003719807, + 0.006397433113306761, + 0.008536451496183872, + 0.004666656721383333, + 0.004968579392880201, + 0.008055990561842918, + 0.008935569785535336, + 0.005951935891062021, + 0.009088466875255108, + 0.0059333317913115025, + 0.00532596604898572, + 0.0035855956375598907, + 0.0034222512040287256, + 0.00541042722761631, + 0.009484602138400078, + 0.005199848674237728, + 0.003695824882015586, + 0.010071732103824615, + 0.00694860378280282, + 0.004046385642141104, + 0.005950573831796646, + 0.005270480178296566, + 0.01302833016961813, + 0.006904344540089369, + 0.008637307211756706, + 0.013374329544603825, + 0.01595224440097809, + 0.010064706206321716, + 0.008168075233697891, + 0.01214414183050394, + 0.00968790240585804, + 0.004799359943717718, + 0.00891661923378706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.012543240562081337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346776634454727 + ], + "opposite_logprob_tokens": [ + -3.6513466835021973 + ], + "logprob_reward": -0.026346776634454727, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346776634454727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16248176991939545, + 0.17816530168056488, + 0.08166008442640305, + 0.11397991329431534, + 0.11583947390317917, + 0.0976729542016983, + 0.0854761078953743, + 0.05096388980746269, + 0.04581418260931969, + 0.06794627755880356 + ], + "credit_map": [ + 0.019107066094875336, + 0.014575611799955368, + 0.02296793833374977, + 0.012941453605890274, + 0.0232989601790905, + 0.022938257083296776, + 0.019840747117996216, + 0.026811745017766953, + 0.01568816974759102, + 0.013730257749557495, + 0.016162198036909103, + 0.020654898136854172, + 0.016717424616217613, + 0.029814882203936577, + 0.02617552876472473, + 0.03922194615006447, + 0.009144161827862263, + 0.005673438310623169, + 0.007110620383173227, + 0.009161504916846752, + 0.014235800132155418, + 0.013519738800823689, + 0.01017522718757391, + 0.012639593333005905, + 0.008575215935707092, + 0.01656772941350937, + 0.009434320032596588, + 0.011309735476970673, + 0.01712099276483059, + 0.023787256330251694, + 0.011475971899926662, + 0.015708696097135544, + 0.004428521264344454, + 0.007457248866558075, + 0.003569471649825573, + 0.005438969470560551, + 0.0102192098274827, + 0.03948013857007027, + 0.005536844953894615, + 0.03970906883478165, + 0.008853048086166382, + 0.010288400575518608, + 0.008671551942825317, + 0.00824175588786602, + 0.013538341969251633, + 0.018079552799463272, + 0.016505153849720955, + 0.013495153747498989, + 0.008420467376708984, + 0.011886860243976116, + 0.0065888939425349236, + 0.008939653635025024, + 0.012107721529901028, + 0.012783369049429893, + 0.008420547470450401, + 0.016328591853380203, + 0.0054414779879152775, + 0.006689159199595451, + 0.005313805304467678, + 0.004548846744000912, + 0.007530647329986095, + 0.006615476682782173, + 0.010138590820133686, + 0.00468588899821043, + 0.0045550111681222916, + 0.004281180445104837, + 0.004013085272163153, + 0.004292369354516268, + 0.00484807975590229, + 0.007621950935572386, + 0.005783443804830313, + 0.01041906327009201, + 0.011874368414282799, + 0.011103342287242413, + 0.008027283474802971, + 0.00809482205659151, + 0.011149073950946331, + 0.006759470794349909, + 0.004863313399255276, + 0.006074607837945223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.026346776634454727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200207905843854 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200207905843854, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200207905843854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11485223472118378, + 0.07808248698711395, + 0.06455887854099274, + 0.10479280352592468, + 0.1656484603881836, + 0.09831073880195618, + 0.08420853316783905, + 0.06089717149734497, + 0.14492283761501312, + 0.08372581005096436 + ], + "credit_map": [ + 0.015221071429550648, + 0.011517765000462532, + 0.01373757142573595, + 0.008816970512270927, + 0.02461126260459423, + 0.018356550484895706, + 0.010378982871770859, + 0.012212060391902924, + 0.006318304222077131, + 0.00788524467498064, + 0.008465005084872246, + 0.007333285175263882, + 0.012390979565680027, + 0.019200138747692108, + 0.007768607698380947, + 0.008720926940441132, + 0.0062856534495949745, + 0.0058948746882379055, + 0.005421011243015528, + 0.007652969099581242, + 0.006628498900681734, + 0.01390005648136139, + 0.006430429872125387, + 0.012345395982265472, + 0.010558199137449265, + 0.017096636816859245, + 0.01290152221918106, + 0.012963386252522469, + 0.015553182922303677, + 0.0105135478079319, + 0.013218661770224571, + 0.011987673118710518, + 0.008880391716957092, + 0.010459676384925842, + 0.01104184240102768, + 0.009638172574341297, + 0.014682632870972157, + 0.04948742687702179, + 0.012309172190725803, + 0.04914914816617966, + 0.009849345311522484, + 0.013638212345540524, + 0.008282741531729698, + 0.011927666142582893, + 0.015696154907345772, + 0.01431115623563528, + 0.010833813808858395, + 0.013771659694612026, + 0.007801679894328117, + 0.013314112089574337, + 0.0068108756095170975, + 0.009906960651278496, + 0.011874576099216938, + 0.013633627444505692, + 0.008260825648903847, + 0.012605876661837101, + 0.008551607839763165, + 0.008235232904553413, + 0.005413365084677935, + 0.006642391439527273, + 0.007748317439109087, + 0.009230497293174267, + 0.009652871638536453, + 0.005422895774245262, + 0.011164119467139244, + 0.018529387190937996, + 0.012942511588335037, + 0.026380423456430435, + 0.021764084696769714, + 0.019902270287275314, + 0.018725981935858727, + 0.015514077618718147, + 0.011981483548879623, + 0.013514920137822628, + 0.007638007402420044, + 0.006411934271454811, + 0.020860183984041214, + 0.007863592356443405, + 0.007271587383002043, + 0.008184108883142471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.002200207905843854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14275963604450226 + ], + "opposite_logprob_tokens": [ + -2.0177595615386963 + ], + "logprob_reward": -0.14275963604450226, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14275963604450226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14467626810073853, + 0.10169614106416702, + 0.09664171934127808, + 0.09226568043231964, + 0.14180554449558258, + 0.09769172221422195, + 0.1303732544183731, + 0.03841904178261757, + 0.12601391971111298, + 0.03041667491197586 + ], + "credit_map": [ + 0.01511865109205246, + 0.011328976601362228, + 0.02829921804368496, + 0.013172965496778488, + 0.019842000678181648, + 0.02140338160097599, + 0.014592044055461884, + 0.020919034257531166, + 0.009220999665558338, + 0.011707277037203312, + 0.014914887957274914, + 0.012443347834050655, + 0.011228660121560097, + 0.01510818861424923, + 0.012808971107006073, + 0.014263803139328957, + 0.013273321092128754, + 0.008391495794057846, + 0.01146195363253355, + 0.014968423172831535, + 0.010110903531312943, + 0.014245226047933102, + 0.009352436289191246, + 0.014837957918643951, + 0.006278957240283489, + 0.01008227001875639, + 0.007252684328705072, + 0.008376771584153175, + 0.013487355783581734, + 0.025107985362410545, + 0.007034274749457836, + 0.014645380899310112, + 0.006911366246640682, + 0.00811101496219635, + 0.010900304652750492, + 0.006800771225243807, + 0.007507944945245981, + 0.0477127879858017, + 0.006130784284323454, + 0.04773057997226715, + 0.009579956531524658, + 0.013639891520142555, + 0.00918265525251627, + 0.007520980667322874, + 0.013990323059260845, + 0.016238553449511528, + 0.015734078362584114, + 0.011805283837020397, + 0.013480579480528831, + 0.02334771864116192, + 0.011605721898376942, + 0.009935876354575157, + 0.020076284185051918, + 0.02217152714729309, + 0.013013185001909733, + 0.016742365434765816, + 0.004988923668861389, + 0.005513823591172695, + 0.004060832317918539, + 0.003993107937276363, + 0.0038696955889463425, + 0.006438010837882757, + 0.0052800849080085754, + 0.004274562932550907, + 0.016631944105029106, + 0.01586165279150009, + 0.010584875009953976, + 0.014870828948915005, + 0.011110426858067513, + 0.02648942731320858, + 0.016884278506040573, + 0.01358047779649496, + 0.0052102054469287395, + 0.004951239563524723, + 0.0036557491403073072, + 0.0020018410868942738, + 0.004911008756607771, + 0.003892438253387809, + 0.0033890975173562765, + 0.002405093517154455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.14275963604450226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "full-body, horizontal tracking shot follows a sleek, multi-rotor manned aircraft as it travels at a steady pace, on a straight course, from left to right through a nighttime urban canyon shrouded in torrential rain. Towering skyscrapers, with a distinctly pixelated texture, appear to be composed of giant, glowing blocks, their surfaces gleaming with low-resolution, animated advertisements. The aircraft's dark, composite hue gleams clammy in the rain, its outline sharply illuminated by the bright blue and magenta neon strips etched across its fuselage. As it advances, these neon lights cast long, digital-like trails through the humid air, hovering briefly between the pixelated buildings. The dense stream of rain streaks across the camera, intertwining with the equally pixelated, flickering lights of billboards atop the massive buildings in the background, blurring distant details. The scene is dominated by deep blue and purple hues, creating a cyberpunk atmosphere that feels both retro and futuristic, steeped in the aesthetics of 8-bit video games.", + "seed": 5497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_4883_seed_5497_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10116087943315506, + "vlm_logprob": -0.10116087943315506, + "vlm_yes_no_margin": 3.450000011920929, + "vlm_token_credit": -0.10116087943315506, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35591548681259155 + ], + "opposite_logprob_tokens": [ + -1.2309155464172363 + ], + "logprob_reward": -0.35591548681259155, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.35591548681259155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21469435095787048, + 0.18486568331718445, + 0.11675269156694412, + 0.14009560644626617, + 0.05890641734004021, + 0.10957972705364227, + 0.10842961817979813, + 0.028411636129021645, + 0.017545457929372787, + 0.020718805491924286 + ], + "credit_map": [ + 0.012621048837900162, + 0.009511617943644524, + 0.01481569278985262, + 0.011731975711882114, + 0.01428278349339962, + 0.0623578280210495, + 0.029430899769067764, + 0.05994250997900963, + 0.009166168980300426, + 0.011927338317036629, + 0.013548105023801327, + 0.01702473871409893, + 0.015426252968609333, + 0.050874583423137665, + 0.05015082657337189, + 0.016747670248150826, + 0.005236248951405287, + 0.007217211648821831, + 0.007420887239277363, + 0.00651436485350132, + 0.008114807307720184, + 0.036440037190914154, + 0.00779780512675643, + 0.0380113422870636, + 0.0112849036231637, + 0.012227975763380527, + 0.02013247087597847, + 0.01951260305941105, + 0.012060756795108318, + 0.026955002918839455, + 0.018999479711055756, + 0.01892242208123207, + 0.003955845721065998, + 0.006329665891826153, + 0.006827856879681349, + 0.008311017416417599, + 0.006253731437027454, + 0.006322760600596666, + 0.014130774885416031, + 0.006774767767637968, + 0.006707407999783754, + 0.006642584688961506, + 0.01483840961009264, + 0.017222575843334198, + 0.015542491339147091, + 0.011090067215263844, + 0.024156644940376282, + 0.01337954681366682, + 0.011487352661788464, + 0.011829870752990246, + 0.008373986929655075, + 0.012977683916687965, + 0.011914758011698723, + 0.019875822588801384, + 0.016444221138954163, + 0.015525927767157555, + 0.00202270969748497, + 0.0019367877393960953, + 0.005262247286736965, + 0.002980473916977644, + 0.00258392421528697, + 0.008108125068247318, + 0.003196889068931341, + 0.0023204819299280643, + 0.00199407571926713, + 0.0017968588508665562, + 0.0015837366227060556, + 0.0030685567762702703, + 0.0026875706389546394, + 0.0026222371961921453, + 0.001975142629817128, + 0.0018172822892665863, + 0.004033099859952927, + 0.0015511802630499005, + 0.002429146086797118, + 0.001342329545877874, + 0.0015513140242546797, + 0.0019979411736130714, + 0.0014412322780117393, + 0.006372561678290367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.35591548681259155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3134981691837311, + 0.09111550450325012, + 0.06271392107009888, + 0.05077885836362839, + 0.018964320421218872, + 0.10798591375350952, + 0.07364903390407562, + 0.057790160179138184, + 0.024381237104535103, + 0.19912290573120117 + ], + "credit_map": [ + 0.016254669055342674, + 0.00803188793361187, + 0.02330472506582737, + 0.015260439366102219, + 0.03549858182668686, + 0.03882645443081856, + 0.12939530611038208, + 0.046926096081733704, + 0.0066741411574184895, + 0.007077429909259081, + 0.005436331499367952, + 0.00783595535904169, + 0.006078672129660845, + 0.02689654752612114, + 0.022395150735974312, + 0.00872126966714859, + 0.0022524436935782433, + 0.0015676917973905802, + 0.00321659236215055, + 0.0027453815564513206, + 0.0022646375000476837, + 0.022461822256445885, + 0.003948437515646219, + 0.024256911128759384, + 0.003880067029967904, + 0.004294056911021471, + 0.008046035654842854, + 0.005333272740244865, + 0.0038563900161534548, + 0.008709874004125595, + 0.008958800695836544, + 0.007700353395193815, + 0.0012449449859559536, + 0.001121220993809402, + 0.0029847314581274986, + 0.003836223157122731, + 0.001825448009185493, + 0.0017343852669000626, + 0.0033681418281048536, + 0.0028492207638919353, + 0.009208704344928265, + 0.00824667327105999, + 0.020555859431624413, + 0.021593594923615456, + 0.006113911978900433, + 0.010821046307682991, + 0.01753355748951435, + 0.013912560418248177, + 0.0077766538597643375, + 0.007217868231236935, + 0.006304878741502762, + 0.00890202447772026, + 0.008148248307406902, + 0.01256073173135519, + 0.012448658235371113, + 0.010289966128766537, + 0.004289666656404734, + 0.0027345679700374603, + 0.016076980158686638, + 0.006508951541036367, + 0.004877907689660788, + 0.01071494072675705, + 0.008344744332134724, + 0.004242401570081711, + 0.0023075020872056484, + 0.0032552543561905622, + 0.0020786381792277098, + 0.003427817951887846, + 0.004308168776333332, + 0.0031322790309786797, + 0.002866878407076001, + 0.003004697384312749, + 0.03538329899311066, + 0.010449891909956932, + 0.031364113092422485, + 0.010246972553431988, + 0.007737808395177126, + 0.014115563593804836, + 0.012023579329252243, + 0.0778016671538353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.02058546058833599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419964898377657 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.01419964898377657, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419964898377657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13623732328414917, + 0.11189589649438858, + 0.09753942489624023, + 0.17273926734924316, + 0.10640640556812286, + 0.09950155019760132, + 0.08353926986455917, + 0.06242089718580246, + 0.03359769657254219, + 0.09612226486206055 + ], + "credit_map": [ + 0.008365028537809849, + 0.0068335277028381824, + 0.011289816349744797, + 0.009177196770906448, + 0.01871977373957634, + 0.03074975684285164, + 0.019828898832201958, + 0.03127333149313927, + 0.007733906153589487, + 0.00965244509279728, + 0.01774503104388714, + 0.015929095447063446, + 0.009420137852430344, + 0.019057538360357285, + 0.02026236243546009, + 0.012095380574464798, + 0.008913733996450901, + 0.007894632406532764, + 0.0105722201988101, + 0.00843752920627594, + 0.014003459364175797, + 0.014636813662946224, + 0.01134234294295311, + 0.021738702431321144, + 0.01478246133774519, + 0.017457572743296623, + 0.02209702879190445, + 0.023723628371953964, + 0.015039840713143349, + 0.03195992484688759, + 0.025750264525413513, + 0.02192855067551136, + 0.007714810315519571, + 0.008400797843933105, + 0.015888523310422897, + 0.011126347817480564, + 0.013586076907813549, + 0.012591601349413395, + 0.026647036895155907, + 0.010451210662722588, + 0.005819824989885092, + 0.005800160113722086, + 0.014370597898960114, + 0.022195618599653244, + 0.009952123276889324, + 0.013340678997337818, + 0.016133612021803856, + 0.01188893523067236, + 0.008812191896140575, + 0.00836174190044403, + 0.006161803845316172, + 0.008265855722129345, + 0.010125827975571156, + 0.01602066680788994, + 0.013793057762086391, + 0.011998126283288002, + 0.004834326449781656, + 0.005413790233433247, + 0.011740252375602722, + 0.006658181548118591, + 0.006894446909427643, + 0.013843988068401814, + 0.00675151776522398, + 0.006284395698457956, + 0.0028143972158432007, + 0.004661131650209427, + 0.0026651432272046804, + 0.00402066670358181, + 0.0056931618601083755, + 0.0047768838703632355, + 0.0047139450907707214, + 0.004252367187291384, + 0.01628134399652481, + 0.005247384309768677, + 0.014334121719002724, + 0.005620580166578293, + 0.005025411490350962, + 0.006725947838276625, + 0.005695439409464598, + 0.03719204291701317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.01419964898377657 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022011594846844673 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.0022011594846844673, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022011594846844673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22002673149108887, + 0.04987214878201485, + 0.07320937514305115, + 0.19401922821998596, + 0.06243147328495979, + 0.10705582052469254, + 0.061333876103162766, + 0.041896481066942215, + 0.03461313247680664, + 0.15554174780845642 + ], + "credit_map": [ + 0.011200401932001114, + 0.009436164051294327, + 0.012277442030608654, + 0.006930092349648476, + 0.02028617262840271, + 0.07154558598995209, + 0.0655287429690361, + 0.022822123020887375, + 0.004926583264023066, + 0.005646031349897385, + 0.003924997057765722, + 0.0046364725567400455, + 0.008090656250715256, + 0.009073437191545963, + 0.009104132652282715, + 0.004469841718673706, + 0.005719552282243967, + 0.009465518407523632, + 0.007329820189625025, + 0.007321519777178764, + 0.014509554021060467, + 0.010846237652003765, + 0.008126357570290565, + 0.009890817105770111, + 0.01832897961139679, + 0.01580986939370632, + 0.03138525038957596, + 0.02647026814520359, + 0.01578092947602272, + 0.042663753032684326, + 0.021036408841609955, + 0.022543756291270256, + 0.00369135825894773, + 0.007399940863251686, + 0.00975303165614605, + 0.007413399871438742, + 0.006385039072483778, + 0.008259076625108719, + 0.011200631968677044, + 0.008328991942107677, + 0.013464083895087242, + 0.015570121817290783, + 0.012176236137747765, + 0.01304418221116066, + 0.013538473285734653, + 0.011912865564227104, + 0.01607613079249859, + 0.011273727752268314, + 0.005359640344977379, + 0.007939576171338558, + 0.006577600259333849, + 0.005865035578608513, + 0.008711153641343117, + 0.008368668146431446, + 0.010219165123999119, + 0.008293038234114647, + 0.003768421243876219, + 0.004090440459549427, + 0.006726098712533712, + 0.004210442770272493, + 0.008671030402183533, + 0.00690600136294961, + 0.003680963534861803, + 0.003843079786747694, + 0.004165750928223133, + 0.0047308532521128654, + 0.003087603719905019, + 0.003716830164194107, + 0.007498509716242552, + 0.004121654201298952, + 0.004050256684422493, + 0.0032416721805930138, + 0.025146692991256714, + 0.007074407301843166, + 0.026052678003907204, + 0.009077898226678371, + 0.006745966617017984, + 0.011507008224725723, + 0.00951139535754919, + 0.06042570248246193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0022011594846844673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290264129638672 + ], + "opposite_logprob_tokens": [ + -2.2379026412963867 + ], + "logprob_reward": -0.11290264129638672, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290264129638672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1652873456478119, + 0.07529241591691971, + 0.054844554513692856, + 0.07719197869300842, + 0.09838219732046127, + 0.03620389848947525, + 0.12618228793144226, + 0.21493403613567352, + 0.04810694605112076, + 0.10357440263032913 + ], + "credit_map": [ + 0.008102862164378166, + 0.006771352607756853, + 0.012322385795414448, + 0.006500752177089453, + 0.021563969552516937, + 0.06192251667380333, + 0.011698869988322258, + 0.036404628306627274, + 0.006244682241231203, + 0.008531016297638416, + 0.00580238364636898, + 0.008290963247418404, + 0.008157077245414257, + 0.01678352616727352, + 0.013174605555832386, + 0.008308161981403828, + 0.005803609266877174, + 0.005282134748995304, + 0.005121462978422642, + 0.0063599636778235435, + 0.008619396016001701, + 0.011793495155870914, + 0.00549289071932435, + 0.006371609400957823, + 0.00899259652942419, + 0.007819760590791702, + 0.010682076215744019, + 0.010151440277695656, + 0.005852473899722099, + 0.011397424153983593, + 0.011251057498157024, + 0.011045143939554691, + 0.008233286440372467, + 0.006454513408243656, + 0.013665764592587948, + 0.013099570758640766, + 0.007330021355301142, + 0.00966655369848013, + 0.01673773303627968, + 0.023194750770926476, + 0.0034034308046102524, + 0.004358455538749695, + 0.004621184431016445, + 0.005110215395689011, + 0.0038785492070019245, + 0.005148045718669891, + 0.004552836529910564, + 0.005131181329488754, + 0.009248844347894192, + 0.014605792239308357, + 0.014653019607067108, + 0.01647288352251053, + 0.015483704395592213, + 0.011867226101458073, + 0.024937905371189117, + 0.01891290582716465, + 0.01665167696774006, + 0.013881727121770382, + 0.06744935363531113, + 0.026471193879842758, + 0.02258072979748249, + 0.024066099897027016, + 0.03140522912144661, + 0.012428018264472485, + 0.006256481632590294, + 0.004989999812096357, + 0.004866220057010651, + 0.006330517586320639, + 0.007325383834540844, + 0.005179041996598244, + 0.006634841673076153, + 0.006524459458887577, + 0.019399389624595642, + 0.004829983692616224, + 0.01573115400969982, + 0.005211632698774338, + 0.0034010775852948427, + 0.007654337212443352, + 0.006266417913138866, + 0.04108041524887085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.11290264129638672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "full-body, horizontal tracking shot follows a sleek, multi-rotor manned aircraft as it travels at a steady pace, on a straight course, from left to right through a nighttime urban canyon shrouded in torrential rain. Towering skyscrapers, with a distinctly pixelated texture, appear to be composed of giant, glowing blocks, their surfaces gleaming with low-resolution, animated advertisements. The aircraft's dark, composite hue gleams clammy in the rain, its outline sharply illuminated by the bright blue and magenta neon strips etched across its fuselage. As it advances, these neon lights cast long, digital-like trails through the humid air, hovering briefly between the pixelated buildings. The dense stream of rain streaks across the camera, intertwining with the equally pixelated, flickering lights of billboards atop the massive buildings in the background, blurring distant details. The scene is dominated by deep blue and purple hues, creating a cyberpunk atmosphere that feels both retro and futuristic, steeped in the aesthetics of 8-bit video games.", + "seed": 5498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_4883_seed_5498_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.178434129152447, + "vlm_logprob": -0.178434129152447, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.178434129152447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.582190990447998 + ], + "opposite_logprob_tokens": [ + -0.832190990447998 + ], + "logprob_reward": -0.582190990447998, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.582190990447998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08666170388460159, + 0.03652026504278183, + 0.055499497801065445, + 0.06310424953699112, + 0.22664089500904083, + 0.20333881676197052, + 0.18797796964645386, + 0.04554029181599617, + 0.0654066875576973, + 0.029309580102562904 + ], + "credit_map": [ + 0.009877865202724934, + 0.012820916250348091, + 0.008919628337025642, + 0.008998226374387741, + 0.012607814744114876, + 0.011660692282021046, + 0.010543087497353554, + 0.011233472265303135, + 0.0052719032391905785, + 0.004240498412400484, + 0.0050502498634159565, + 0.0036135301925241947, + 0.0037246649153530598, + 0.004587533418089151, + 0.004932979121804237, + 0.005098903551697731, + 0.0048324973322451115, + 0.004348293412476778, + 0.0103157265111804, + 0.010293702594935894, + 0.006184167228639126, + 0.005359860137104988, + 0.0077484166249632835, + 0.006416834890842438, + 0.0041625057347118855, + 0.004852190613746643, + 0.008697465993463993, + 0.01062703412026167, + 0.005886273458600044, + 0.004418820142745972, + 0.014537202194333076, + 0.009922759607434273, + 0.019736601039767265, + 0.032668791711330414, + 0.040160130709409714, + 0.03141787648200989, + 0.028466561809182167, + 0.02570127695798874, + 0.024851912632584572, + 0.02363775111734867, + 0.014492638409137726, + 0.009298073127865791, + 0.012970204465091228, + 0.0232289619743824, + 0.020343324169516563, + 0.019400710240006447, + 0.048464659601449966, + 0.055140238255262375, + 0.013653390109539032, + 0.010786459781229496, + 0.020672958344221115, + 0.04998796433210373, + 0.01388455182313919, + 0.022663237527012825, + 0.021873580291867256, + 0.034455835819244385, + 0.004729458596557379, + 0.004776091314852238, + 0.005467207171022892, + 0.007512110751122236, + 0.0027324515394866467, + 0.0063497633673250675, + 0.00767409335821867, + 0.006299117114394903, + 0.006349500734359026, + 0.007962694391608238, + 0.009997851215302944, + 0.008504770696163177, + 0.008531288243830204, + 0.008557404391467571, + 0.00898395013064146, + 0.006519223563373089, + 0.003174198791384697, + 0.003004638943821192, + 0.005095198284834623, + 0.004499349277466536, + 0.002586132613942027, + 0.0022793293464928865, + 0.003783918684348464, + 0.004886812996119261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.582190990447998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370778262615204 + ], + "opposite_logprob_tokens": [ + -3.408707857131958 + ], + "logprob_reward": -0.03370778262615204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370778262615204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39817309379577637, + 0.14351125061511993, + 0.06923691183328629, + 0.05404628813266754, + 0.05804898962378502, + 0.06023545190691948, + 0.08469594269990921, + 0.02704881690442562, + 0.059551458805799484, + 0.04545174911618233 + ], + "credit_map": [ + 0.03968373313546181, + 0.07383549213409424, + 0.028856968507170677, + 0.03039604239165783, + 0.06618085503578186, + 0.07969184964895248, + 0.028734354302287102, + 0.050793811678886414, + 0.014191488735377789, + 0.018168114125728607, + 0.025417594239115715, + 0.012950882315635681, + 0.009689741767942905, + 0.03575177863240242, + 0.014087643474340439, + 0.013254020363092422, + 0.006153715308755636, + 0.008726577274501324, + 0.011429054662585258, + 0.012953365221619606, + 0.009292224422097206, + 0.008909672498703003, + 0.005681515671312809, + 0.006090790499001741, + 0.003652708837762475, + 0.004919269122183323, + 0.010020889341831207, + 0.009898287244141102, + 0.005107994191348553, + 0.003757826518267393, + 0.010323415510356426, + 0.006365903187543154, + 0.004528422839939594, + 0.006092472467571497, + 0.0051514869555830956, + 0.006744974292814732, + 0.010346817784011364, + 0.006770177278667688, + 0.0097142169252038, + 0.008700428530573845, + 0.005493082106113434, + 0.00267071183770895, + 0.006185867823660374, + 0.007370848208665848, + 0.004043007735162973, + 0.005137723870575428, + 0.013362999074161053, + 0.015971219167113304, + 0.005353886168450117, + 0.004769738297909498, + 0.005989858880639076, + 0.02541177347302437, + 0.004303130321204662, + 0.010310813784599304, + 0.0068249874748289585, + 0.02173176035284996, + 0.0025688062887638807, + 0.004088140558451414, + 0.0038514852058142424, + 0.005434798542410135, + 0.0019992736633867025, + 0.0033044235315173864, + 0.0028360167052596807, + 0.00296587310731411, + 0.005047344136983156, + 0.005434764549136162, + 0.006908228155225515, + 0.007990522310137749, + 0.006887238472700119, + 0.011039502918720245, + 0.00887112133204937, + 0.007372740656137466, + 0.008566256612539291, + 0.004563913680613041, + 0.008375294506549835, + 0.005022590979933739, + 0.0024113550316542387, + 0.0018034606473520398, + 0.005157028324902058, + 0.009551852010190487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.03370778262615204 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586511120200157 + ], + "opposite_logprob_tokens": [ + -3.8955864906311035 + ], + "logprob_reward": -0.020586511120200157, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586511120200157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13240565359592438, + 0.06645534932613373, + 0.046937305480241776, + 0.10517175495624542, + 0.2184152752161026, + 0.16493946313858032, + 0.15739408135414124, + 0.038864314556121826, + 0.02824585884809494, + 0.041171010583639145 + ], + "credit_map": [ + 0.014214727096259594, + 0.02479872666299343, + 0.013860823586583138, + 0.010153139010071754, + 0.019095195457339287, + 0.019183607771992683, + 0.01450286153703928, + 0.016596568748354912, + 0.007538034114986658, + 0.0052029830403625965, + 0.01110508106648922, + 0.006011588498950005, + 0.005545236635953188, + 0.011466673575341702, + 0.010345198214054108, + 0.009240549989044666, + 0.0036612062249332666, + 0.004788575693964958, + 0.006368005648255348, + 0.00825212337076664, + 0.006024512462317944, + 0.006399136036634445, + 0.005891282577067614, + 0.005552464164793491, + 0.008641944266855717, + 0.014642838388681412, + 0.01593630202114582, + 0.0170741006731987, + 0.013309229165315628, + 0.010089995339512825, + 0.014172788709402084, + 0.011304549872875214, + 0.015248727053403854, + 0.024313200265169144, + 0.025338005274534225, + 0.023514319211244583, + 0.022488946095108986, + 0.052573222666978836, + 0.019124934449791908, + 0.03581392392516136, + 0.013272172771394253, + 0.009812652133405209, + 0.008702083490788937, + 0.016709772869944572, + 0.022694140672683716, + 0.01554685179144144, + 0.03881760314106941, + 0.03938417509198189, + 0.011793287470936775, + 0.008977914229035378, + 0.011194419115781784, + 0.04366035386919975, + 0.011700007133185863, + 0.021525727584958076, + 0.013952124863862991, + 0.034590255469083786, + 0.0037536905147135258, + 0.005468230228871107, + 0.004274959675967693, + 0.006501102354377508, + 0.0036628383677452803, + 0.004889884032309055, + 0.005766540300101042, + 0.004547070246189833, + 0.0030001697596162558, + 0.0030279033817350864, + 0.0031317216344177723, + 0.0037264302372932434, + 0.002644685795530677, + 0.004157554358243942, + 0.005109807942062616, + 0.003447586437687278, + 0.00585838221013546, + 0.004769372288137674, + 0.006888618692755699, + 0.005185888148844242, + 0.0029003729578107595, + 0.0034100832417607307, + 0.0056390706449747086, + 0.006519221235066652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.020586511120200157 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036183618940413 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036183618940413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036183618940413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08576672524213791, + 0.3632962703704834, + 0.04413461685180664, + 0.07652302086353302, + 0.1055576503276825, + 0.048302263021469116, + 0.1163129210472107, + 0.027151083573698997, + 0.07167128473520279, + 0.061284139752388 + ], + "credit_map": [ + 0.008384164422750473, + 0.014184530824422836, + 0.011273372918367386, + 0.008162280544638634, + 0.013987749814987183, + 0.009900018572807312, + 0.011026252992451191, + 0.008848351426422596, + 0.0302328672260046, + 0.011859524995088577, + 0.09878150373697281, + 0.015532424673438072, + 0.01717226207256317, + 0.12497177720069885, + 0.026854662224650383, + 0.03789125382900238, + 0.0034939623437821865, + 0.004316238686442375, + 0.008300087414681911, + 0.0071439011953771114, + 0.004890723153948784, + 0.005674877669662237, + 0.005747630726546049, + 0.004567195661365986, + 0.0062903291545808315, + 0.00851036049425602, + 0.010418223217129707, + 0.008836755529046059, + 0.008813366293907166, + 0.00663888081908226, + 0.01287570409476757, + 0.014139404520392418, + 0.007808720227330923, + 0.01433409284800291, + 0.01734320819377899, + 0.010410859249532223, + 0.014003491960465908, + 0.016042543575167656, + 0.010191788896918297, + 0.01542295515537262, + 0.0035867588594555855, + 0.0028710528276860714, + 0.004961815197020769, + 0.00922725535929203, + 0.0040998258627951145, + 0.004435284994542599, + 0.008754804730415344, + 0.01036546565592289, + 0.007364646065980196, + 0.006902690976858139, + 0.009265871718525887, + 0.033312272280454636, + 0.007845803163945675, + 0.014594463631510735, + 0.010260319337248802, + 0.02676684781908989, + 0.0031233816407620907, + 0.0030371048487722874, + 0.0033027383033186197, + 0.004124713130295277, + 0.0023062615655362606, + 0.0037662575487047434, + 0.004328080918639898, + 0.0031625444535166025, + 0.0071948510594666, + 0.006798967719078064, + 0.011980191804468632, + 0.01045695785433054, + 0.006033952813595533, + 0.011604992672801018, + 0.009653011336922646, + 0.007948365993797779, + 0.008085054345428944, + 0.00836160872131586, + 0.009467282332479954, + 0.006884961389005184, + 0.0038879283238202333, + 0.005578176584094763, + 0.00819358229637146, + 0.010825550183653831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.0036183618940413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520669996738434 + ], + "opposite_logprob_tokens": [ + -1.502066969871521 + ], + "logprob_reward": -0.2520669996738434, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520669996738434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12784413993358612, + 0.17338109016418457, + 0.03252984955906868, + 0.125406414270401, + 0.08899203687906265, + 0.23450146615505219, + 0.05092061311006546, + 0.049939073622226715, + 0.06544104218482971, + 0.05104432255029678 + ], + "credit_map": [ + 0.009949962608516216, + 0.016802486032247543, + 0.01478113979101181, + 0.015009011141955853, + 0.02527208812534809, + 0.018061337992548943, + 0.012990341521799564, + 0.014977776445448399, + 0.01398391928523779, + 0.007570294197648764, + 0.04885590076446533, + 0.01036363560706377, + 0.008009609766304493, + 0.055818721652030945, + 0.01354117039591074, + 0.015237824991345406, + 0.0034087402746081352, + 0.0033697441685944796, + 0.006288738921284676, + 0.004394835792481899, + 0.004143774043768644, + 0.003928743302822113, + 0.0033904940355569124, + 0.003604779252782464, + 0.007123047020286322, + 0.00812414102256298, + 0.019016437232494354, + 0.02128053829073906, + 0.009794062934815884, + 0.014005940407514572, + 0.02645168825984001, + 0.019610559567809105, + 0.005796501878648996, + 0.005530105903744698, + 0.007035703398287296, + 0.007825951091945171, + 0.011784011498093605, + 0.024638036265969276, + 0.0064516509883105755, + 0.019930079579353333, + 0.018075652420520782, + 0.008677707985043526, + 0.015391947701573372, + 0.024750521406531334, + 0.015599451027810574, + 0.021577980369329453, + 0.06165008619427681, + 0.06877811253070831, + 0.004377331119030714, + 0.0033372349571436644, + 0.005724163260310888, + 0.00971965491771698, + 0.005100847687572241, + 0.008811051025986671, + 0.005968935787677765, + 0.007881388999521732, + 0.006605146918445826, + 0.004782926291227341, + 0.007718002889305353, + 0.005949348211288452, + 0.002899865387007594, + 0.007603115402162075, + 0.007265533786267042, + 0.007115133572369814, + 0.0074006845243275166, + 0.006250352133065462, + 0.008711921982467175, + 0.009338047355413437, + 0.005048404447734356, + 0.011141940951347351, + 0.010751398280262947, + 0.006798289716243744, + 0.010076760314404964, + 0.0034864963963627815, + 0.009571650065481663, + 0.0059822904877364635, + 0.0025350432842969894, + 0.0028879891615360975, + 0.00530700059607625, + 0.011197091080248356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.2520669996738434 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "full-body, horizontal tracking shot follows a sleek, multi-rotor manned aircraft as it travels at a steady pace, on a straight course, from left to right through a nighttime urban canyon shrouded in torrential rain. Towering skyscrapers, with a distinctly pixelated texture, appear to be composed of giant, glowing blocks, their surfaces gleaming with low-resolution, animated advertisements. The aircraft's dark, composite hue gleams clammy in the rain, its outline sharply illuminated by the bright blue and magenta neon strips etched across its fuselage. As it advances, these neon lights cast long, digital-like trails through the humid air, hovering briefly between the pixelated buildings. The dense stream of rain streaks across the camera, intertwining with the equally pixelated, flickering lights of billboards atop the massive buildings in the background, blurring distant details. The scene is dominated by deep blue and purple hues, creating a cyberpunk atmosphere that feels both retro and futuristic, steeped in the aesthetics of 8-bit video games.", + "seed": 5499, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_13/index_4883_seed_5499_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08640675134956836, + "vlm_logprob": -0.08640675134956836, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.08640675134956836, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14422860741615295 + ], + "opposite_logprob_tokens": [ + -2.14422869682312 + ], + "logprob_reward": -0.14422860741615295, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14422860741615295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.621065080165863, + 0.02730081044137478, + 0.0703902319073677, + 0.030655359849333763, + 0.026798851788043976, + 0.022569304332137108, + 0.029383569955825806, + 0.11389335244894028, + 0.03733935207128525, + 0.020604124292731285 + ], + "credit_map": [ + 0.04870788753032684, + 0.03233225271105766, + 0.01653960533440113, + 0.025062132626771927, + 0.035583075135946274, + 0.2449273318052292, + 0.022915253415703773, + 0.19499750435352325, + 0.0026254134718328714, + 0.004137409385293722, + 0.0020500323735177517, + 0.003264337545260787, + 0.004149124026298523, + 0.0026832600124180317, + 0.004541114903986454, + 0.0038501156959682703, + 0.007917998358607292, + 0.006064258050173521, + 0.0068016862496733665, + 0.0113524179905653, + 0.0058548906818032265, + 0.00802771095186472, + 0.014229635708034039, + 0.010141627863049507, + 0.0033221060875803232, + 0.004405295476317406, + 0.002840845612809062, + 0.002842166693881154, + 0.0042578838765621185, + 0.005083118099719286, + 0.004589413292706013, + 0.0033145269844681025, + 0.0025021762121468782, + 0.004260513931512833, + 0.0023982636630535126, + 0.0027663609944283962, + 0.004332088865339756, + 0.005542398430407047, + 0.002331759547814727, + 0.0026652885135263205, + 0.0028187960851937532, + 0.002587544033303857, + 0.0019288144540041685, + 0.0027466611936688423, + 0.0030978480353951454, + 0.00447074742987752, + 0.00240282341837883, + 0.0025160699151456356, + 0.00333948596380651, + 0.0030355339404195547, + 0.0027157471049576998, + 0.005207054782658815, + 0.0031455776188522577, + 0.003922836855053902, + 0.00372145208530128, + 0.004295878577977419, + 0.016102410852909088, + 0.014777880162000656, + 0.011823245324194431, + 0.012942534871399403, + 0.013224737718701363, + 0.016501957550644875, + 0.015394608490169048, + 0.013125970959663391, + 0.0065300981514155865, + 0.003826969536021352, + 0.0030788590665906668, + 0.004427487030625343, + 0.003838752629235387, + 0.008020828478038311, + 0.004244996700435877, + 0.003371360246092081, + 0.0028840084560215473, + 0.0035976471845060587, + 0.0015193800209090114, + 0.0032257228158414364, + 0.0023049558512866497, + 0.0017367902910336852, + 0.0024948192294687033, + 0.002840799745172262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.14422860741615295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057880349457264 + ], + "opposite_logprob_tokens": [ + -3.8955788612365723 + ], + "logprob_reward": -0.02057880349457264, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057880349457264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33980047702789307, + 0.06008153408765793, + 0.13250182569026947, + 0.0893358513712883, + 0.030710065737366676, + 0.02359921671450138, + 0.03230283036828041, + 0.1002480685710907, + 0.07839509099721909, + 0.11302506923675537 + ], + "credit_map": [ + 0.031159350648522377, + 0.0254873838275671, + 0.014700154773890972, + 0.01783602684736252, + 0.028303561732172966, + 0.11371494829654694, + 0.01919044367969036, + 0.08940859884023666, + 0.006034030579030514, + 0.007742590270936489, + 0.004932926967740059, + 0.008066080510616302, + 0.009847202338278294, + 0.007280620746314526, + 0.009403259493410587, + 0.006774825509637594, + 0.012175045907497406, + 0.01007460430264473, + 0.012070529162883759, + 0.015267559327185154, + 0.018434645608067513, + 0.029188336804509163, + 0.0195851419121027, + 0.015705963596701622, + 0.009654583409428596, + 0.010275039821863174, + 0.0080105010420084, + 0.009843993932008743, + 0.012947316281497478, + 0.013517930172383785, + 0.015466926619410515, + 0.009619561024010181, + 0.003584271064028144, + 0.004641111008822918, + 0.0024319773074239492, + 0.003124752314761281, + 0.005277116782963276, + 0.005997716449201107, + 0.0023302908521145582, + 0.003322829492390156, + 0.004496101289987564, + 0.002213091356679797, + 0.002501094015315175, + 0.002792943036183715, + 0.002042346866801381, + 0.0035874752793461084, + 0.0030011499766260386, + 0.0029650148935616016, + 0.0033331867307424545, + 0.0030830788891762495, + 0.0031697885133326054, + 0.0035326217766851187, + 0.004812812898308039, + 0.004586915019899607, + 0.003892179112881422, + 0.0058922478929162025, + 0.016461433842778206, + 0.0071303267031908035, + 0.010669930838048458, + 0.00688220513984561, + 0.01340092159807682, + 0.02251625992357731, + 0.012716460041701794, + 0.010470528155565262, + 0.014481108635663986, + 0.008384662680327892, + 0.005146157927811146, + 0.009378280490636826, + 0.009769531898200512, + 0.017344865947961807, + 0.007029274944216013, + 0.006861208006739616, + 0.017988575622439384, + 0.009715108200907707, + 0.009341157041490078, + 0.010362870059907436, + 0.019628925248980522, + 0.026780523359775543, + 0.007277755998075008, + 0.011930159293115139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02057880349457264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809867799282074 + ], + "opposite_logprob_tokens": [ + -3.2880985736846924 + ], + "logprob_reward": -0.03809867799282074, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809867799282074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42791131138801575, + 0.03186822682619095, + 0.10082190483808517, + 0.06433455646038055, + 0.03773089870810509, + 0.07718982547521591, + 0.023471679538488388, + 0.15203048288822174, + 0.04266630485653877, + 0.04197480157017708 + ], + "credit_map": [ + 0.026068447157740593, + 0.021416978910565376, + 0.008786287158727646, + 0.011313915252685547, + 0.020593244582414627, + 0.18607006967067719, + 0.011381886899471283, + 0.1422804743051529, + 0.003566265571862459, + 0.005292378831654787, + 0.0027216491289436817, + 0.0037219920195639133, + 0.00430424464866519, + 0.0032814263831824064, + 0.003784578526392579, + 0.005195688921958208, + 0.00849873200058937, + 0.0070037138648331165, + 0.00736432708799839, + 0.009471195749938488, + 0.015531777404248714, + 0.022808287292718887, + 0.011757771484553814, + 0.018386101350188255, + 0.009417650289833546, + 0.01488327793776989, + 0.00624195346608758, + 0.005096441600471735, + 0.008061862550675869, + 0.00830898154526949, + 0.005547656677663326, + 0.006776728667318821, + 0.005320788361132145, + 0.004902503453195095, + 0.0033251529093831778, + 0.004533611238002777, + 0.004922874737530947, + 0.00591792119666934, + 0.003810581983998418, + 0.0049974690191447735, + 0.013483441434800625, + 0.006685709580779076, + 0.0071584018878638744, + 0.008305853232741356, + 0.008307595737278461, + 0.015588360838592052, + 0.009755047969520092, + 0.007905411534011364, + 0.0026611487846821547, + 0.002450324362143874, + 0.0020687351934611797, + 0.0034736813977360725, + 0.002919849008321762, + 0.0030882705468684435, + 0.002900457940995693, + 0.003909212537109852, + 0.02449607104063034, + 0.011906732805073261, + 0.013907981105148792, + 0.01004659291356802, + 0.020690003409981728, + 0.03691108524799347, + 0.01754601299762726, + 0.016526000574231148, + 0.005629837978631258, + 0.004332419019192457, + 0.005897541530430317, + 0.0034223971888422966, + 0.003930605016648769, + 0.009895789436995983, + 0.003982396796345711, + 0.005575317889451981, + 0.006553094368427992, + 0.005118429660797119, + 0.0031283129937946796, + 0.004857797175645828, + 0.006431455258280039, + 0.006989252753555775, + 0.003678968409076333, + 0.005217488389462233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.03809867799282074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616936504840851 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616936504840851, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616936504840851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2171282172203064, + 0.059251781553030014, + 0.0691954717040062, + 0.06230119615793228, + 0.06258900463581085, + 0.052407748997211456, + 0.05473551154136658, + 0.28482672572135925, + 0.08965767174959183, + 0.047906696796417236 + ], + "credit_map": [ + 0.01963800936937332, + 0.017693229019641876, + 0.011848649010062218, + 0.01631242223083973, + 0.013356600888073444, + 0.06180783361196518, + 0.01605794206261635, + 0.060413554310798645, + 0.0068935309536755085, + 0.008049432188272476, + 0.005661399103701115, + 0.006548278499394655, + 0.008867323398590088, + 0.00768633559346199, + 0.008303479291498661, + 0.007242000196129084, + 0.005729897879064083, + 0.006699279882013798, + 0.006111558061093092, + 0.009146521799266338, + 0.009823350235819817, + 0.013407979160547256, + 0.009885260835289955, + 0.008391626179218292, + 0.008109593763947487, + 0.012302098795771599, + 0.0051799737848341465, + 0.0055821542628109455, + 0.007518311496824026, + 0.009509214200079441, + 0.007406897842884064, + 0.0066929529421031475, + 0.007165308576077223, + 0.00894603505730629, + 0.005931285209953785, + 0.007228769827634096, + 0.009549028240144253, + 0.007281475700438023, + 0.007585862651467323, + 0.008901240304112434, + 0.008233794011175632, + 0.0057872445322573185, + 0.004662970546633005, + 0.009873299859464169, + 0.005010093096643686, + 0.005226915702223778, + 0.006633978337049484, + 0.006979451980441809, + 0.006957908160984516, + 0.007479934021830559, + 0.006827258970588446, + 0.00801398791372776, + 0.007496609818190336, + 0.006312396842986345, + 0.006027733441442251, + 0.005619680974632502, + 0.05840737745165825, + 0.031259626150131226, + 0.044772520661354065, + 0.021977083757519722, + 0.026267969980835915, + 0.036360181868076324, + 0.035494573414325714, + 0.030287403613328934, + 0.01392030157148838, + 0.014284160919487476, + 0.006555681116878986, + 0.012141784653067589, + 0.008102483116090298, + 0.017196018248796463, + 0.009070937521755695, + 0.00838630273938179, + 0.007935803383588791, + 0.0074103642255067825, + 0.0035419014748185873, + 0.0048915366642177105, + 0.007815394550561905, + 0.006857344880700111, + 0.004514037631452084, + 0.004940314684063196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.003616936504840851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, multi-rotor manned aircraft traveling at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a nighttime urban canyon with towering skyscrapers featuring a distinctly pixelated texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene characterized by torrential rain with streaks crossing the camera and deep blue and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the aircraft display bright blue and magenta neon strips on its fuselage casting long digital-like trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22551073133945465 + ], + "opposite_logprob_tokens": [ + -1.6005107164382935 + ], + "logprob_reward": -0.22551073133945465, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22551073133945465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5292235016822815, + 0.03574094921350479, + 0.08027761429548264, + 0.07076101750135422, + 0.05572693049907684, + 0.05471130087971687, + 0.03142409399151802, + 0.07499957084655762, + 0.0343913696706295, + 0.03274368494749069 + ], + "credit_map": [ + 0.03666410222649574, + 0.024164490401744843, + 0.012819793075323105, + 0.01685338281095028, + 0.029916783794760704, + 0.21931175887584686, + 0.015935199335217476, + 0.17355792224407196, + 0.0043314737267792225, + 0.004785248544067144, + 0.0030210791155695915, + 0.004236295353621244, + 0.004895184654742479, + 0.006041944492608309, + 0.004578432068228722, + 0.0038512905593961477, + 0.009115472435951233, + 0.007623882032930851, + 0.005288150627166033, + 0.012178998440504074, + 0.009470011107623577, + 0.01213656086474657, + 0.009650890715420246, + 0.01481364481151104, + 0.010131211951375008, + 0.013129203580319881, + 0.006562197580933571, + 0.006090947892516851, + 0.008353646844625473, + 0.011176728643476963, + 0.009190110489726067, + 0.006126965396106243, + 0.007040810771286488, + 0.011708707548677921, + 0.003179792780429125, + 0.004073416814208031, + 0.008248855359852314, + 0.009653407149016857, + 0.00437147356569767, + 0.007450459990650415, + 0.008251281455159187, + 0.007313266862183809, + 0.00565353874117136, + 0.006812050472944975, + 0.006167253479361534, + 0.00684448704123497, + 0.006217528600245714, + 0.007451893296092749, + 0.003130133030936122, + 0.003668384160846472, + 0.002301248488947749, + 0.0041762725450098515, + 0.005958362016826868, + 0.004615116398781538, + 0.004039018414914608, + 0.003535556374117732, + 0.010305906645953655, + 0.007677745074033737, + 0.00724371150135994, + 0.005825385916978121, + 0.008101376704871655, + 0.013054653070867062, + 0.012539815157651901, + 0.010250977240502834, + 0.004432187415659428, + 0.0044059548527002335, + 0.0035533711779862642, + 0.006100006401538849, + 0.003191552357748151, + 0.004601750522851944, + 0.004742918070405722, + 0.0033636249136179686, + 0.00471495883539319, + 0.003258056938648224, + 0.003083752002567053, + 0.0036238939501345158, + 0.003916456829756498, + 0.00335825071670115, + 0.00303761288523674, + 0.007750700693577528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.22551073133945465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_14/index_tensor([287, 287])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_14/index_tensor([287, 287])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8177b0a4e48dcf36beedc593cc09b6deaa34501a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_14/index_tensor([287, 287])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of Ghibli animation, a sweeping wide shot, like that from a low-flying drone, glides gracefully across a vast expanse of sun-drenched, rolling meadows. The camera slowly pans right, revealing a massive, hill-sized pumpkin perched in the center of the field, its vibrant orange rind deeply grooved and textured with a soft, painterly texture. The pumpkin is meticulously carved into a rustic, whimsical house, with a small, round wooden door and several circular windows. Flanked by dense, leafy green vines, they curl around a curving, chimney-like stem at the top. The entire scene is bathed in warm, golden afternoon light, casting long, soft shadows across the tall, emerald green grass, which sways hypnotically in the breeze. A massive, sea urchin-like creature, curled up asleep on the pumpkin's roof; its spines are not sharp but soft and plush, like velvety tendrils covered in a layer of delicate green moss. As the camera settles into position, the creature begins to stir, slowly and deliberately emerging from its slumber. Its mossy spine trembles gently, and then its body unfolds section by section, a gentle and peaceful awakening that perfectly captures the serene and magical atmosphere.", + "seed": 11781, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_14/index_287_seed_11781_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.7030491009354591, + "vlm_logprob": -0.7030491009354591, + "vlm_yes_no_margin": 0.45, + "vlm_token_credit": -0.7030491009354591, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7032719850540161 + ], + "opposite_logprob_tokens": [ + -0.20327197015285492 + ], + "logprob_reward": -1.7032719850540161, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7032719850540161, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.27109476923942566, + 0.06514520198106766, + 0.12934525310993195, + 0.15490755438804626, + 0.04347731173038483, + 0.06124543771147728, + 0.09535328298807144, + 0.08366286754608154, + 0.045748401433229446, + 0.050019945949316025 + ], + "credit_map": [ + 0.02294865995645523, + 0.025245176628232002, + 0.025697946548461914, + 0.06457231938838959, + 0.01943904720246792, + 0.03557819500565529, + 0.04937221482396126, + 0.028241202235221863, + 0.005862030200660229, + 0.0057724579237401485, + 0.0057496968656778336, + 0.01830313540995121, + 0.004950729664415121, + 0.008392401970922947, + 0.010891460813581944, + 0.005223285872489214, + 0.0240583885461092, + 0.011255268007516861, + 0.006813022308051586, + 0.025974294170737267, + 0.009189720265567303, + 0.027548696845769882, + 0.011513647623360157, + 0.012992213480174541, + 0.049376748502254486, + 0.00650026835501194, + 0.003645475720986724, + 0.01392887532711029, + 0.006595999468117952, + 0.05971967428922653, + 0.006820725277066231, + 0.008319794200360775, + 0.005707574542611837, + 0.0033464417792856693, + 0.003688274184241891, + 0.008730401284992695, + 0.0035836962051689625, + 0.0061346134170889854, + 0.005480829160660505, + 0.006805481389164925, + 0.00796652864664793, + 0.005576514173299074, + 0.004762342199683189, + 0.008270226418972015, + 0.010701894760131836, + 0.010159268975257874, + 0.006187818478792906, + 0.007620842661708593, + 0.008423862047493458, + 0.010305185802280903, + 0.008215777575969696, + 0.024286536499857903, + 0.009283830411732197, + 0.011050427332520485, + 0.010550221428275108, + 0.013237442821264267, + 0.01222398690879345, + 0.007723493501543999, + 0.007755662780255079, + 0.014444543048739433, + 0.011940057389438152, + 0.009521462954580784, + 0.00724693713709712, + 0.012806722894310951, + 0.0038438786286860704, + 0.005645032040774822, + 0.004667812027037144, + 0.010111363604664803, + 0.007155013736337423, + 0.005365160759538412, + 0.0046880668960511684, + 0.004272075369954109, + 0.006210261955857277, + 0.0035584592260420322, + 0.004620908759534359, + 0.00794666912406683, + 0.004690322559326887, + 0.005490959156304598, + 0.004264927469193935, + 0.013237436302006245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -1.7032719850540161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329197287559509 + ], + "opposite_logprob_tokens": [ + -0.7579197287559509 + ], + "logprob_reward": -0.6329197287559509, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329197287559509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2807672917842865, + 0.05416868254542351, + 0.0556487999856472, + 0.34549862146377563, + 0.04477584362030029, + 0.0782458484172821, + 0.022630026564002037, + 0.04607006162405014, + 0.03694294020533562, + 0.03525187447667122 + ], + "credit_map": [ + 0.02553365007042885, + 0.02858361229300499, + 0.02286371961236, + 0.07170084863901138, + 0.019980648532509804, + 0.033285580575466156, + 0.05849360302090645, + 0.020325614139437675, + 0.0062731667421758175, + 0.007230639457702637, + 0.0033981497399508953, + 0.01523214764893055, + 0.004026118200272322, + 0.007275385782122612, + 0.0076289488933980465, + 0.0031041197944432497, + 0.007427331991493702, + 0.0054266746155917645, + 0.005466111470013857, + 0.00787246786057949, + 0.005391118116676807, + 0.00767146609723568, + 0.010196687653660774, + 0.006196936592459679, + 0.09894368052482605, + 0.017306677997112274, + 0.011371208354830742, + 0.04931538179516792, + 0.013964633457362652, + 0.12296607345342636, + 0.01642671413719654, + 0.015204251743853092, + 0.005204044282436371, + 0.004033494275063276, + 0.005232498049736023, + 0.011009824462234974, + 0.0033970598597079515, + 0.006648487411439419, + 0.004038158338516951, + 0.005212272051721811, + 0.00991360004991293, + 0.00769062852486968, + 0.006912943907082081, + 0.013023766689002514, + 0.01649544946849346, + 0.010035701096057892, + 0.006498563569039106, + 0.007675185799598694, + 0.003020655596628785, + 0.002339772880077362, + 0.002636349992826581, + 0.0047716302797198296, + 0.0021299216896295547, + 0.003385954536497593, + 0.0020830051507800817, + 0.00226273643784225, + 0.00718406867235899, + 0.004139363300055265, + 0.003832616377621889, + 0.008179053664207458, + 0.006789002567529678, + 0.005323890596628189, + 0.004031653981655836, + 0.00659040967002511, + 0.004275694955140352, + 0.004345634486526251, + 0.0036836471408605576, + 0.0078359330072999, + 0.0052073048427701, + 0.004194528795778751, + 0.003795542987063527, + 0.0036046530585736036, + 0.0052488502115011215, + 0.0027562628965824842, + 0.0029574956279248, + 0.004452989902347326, + 0.003388597397133708, + 0.003988610580563545, + 0.0032172189094126225, + 0.009241849184036255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.6329197287559509 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905707508325577 + ], + "opposite_logprob_tokens": [ + -2.579056978225708 + ], + "logprob_reward": -0.07905707508325577, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905707508325577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09747452288866043, + 0.05100950598716736, + 0.2661832273006439, + 0.25383496284484863, + 0.06418754160404205, + 0.07468585669994354, + 0.04171326383948326, + 0.03707730397582054, + 0.0700446292757988, + 0.04378921538591385 + ], + "credit_map": [ + 0.01069572102278471, + 0.011968422681093216, + 0.008135554380714893, + 0.018378842622041702, + 0.007907427847385406, + 0.011184385046362877, + 0.016041863709688187, + 0.013162297196686268, + 0.008237446658313274, + 0.006168541964143515, + 0.00404987670481205, + 0.010285920463502407, + 0.004822759423404932, + 0.006653012707829475, + 0.006643851287662983, + 0.004148097243160009, + 0.046271804720163345, + 0.024012668058276176, + 0.014418882317841053, + 0.04967765137553215, + 0.022212067618966103, + 0.05230144038796425, + 0.02831801399588585, + 0.028970707207918167, + 0.08408837020397186, + 0.009834117256104946, + 0.006502220407128334, + 0.017914773896336555, + 0.009720120579004288, + 0.10235319286584854, + 0.01030422281473875, + 0.013117932714521885, + 0.009553746320307255, + 0.006664527114480734, + 0.004536909982562065, + 0.01573297381401062, + 0.005727119278162718, + 0.00985159631818533, + 0.006525722332298756, + 0.005594940856099129, + 0.01164363231509924, + 0.005757600534707308, + 0.004147710278630257, + 0.006745025981217623, + 0.015113303437829018, + 0.01643327809870243, + 0.0053255874663591385, + 0.00951971672475338, + 0.0041914465837180614, + 0.0035821085330098867, + 0.005223558284342289, + 0.010261589661240578, + 0.003947989083826542, + 0.005372939165681601, + 0.005293223541229963, + 0.003840405959635973, + 0.004427062813192606, + 0.0038091614842414856, + 0.0035316655412316322, + 0.0061765871942043304, + 0.006310882978141308, + 0.004116436466574669, + 0.0030741447117179632, + 0.005631363019347191, + 0.0059783668257296085, + 0.008497589267790318, + 0.007955159991979599, + 0.016952840611338615, + 0.00855383649468422, + 0.008355626836419106, + 0.00691933324560523, + 0.006831870414316654, + 0.00406221067532897, + 0.004377954639494419, + 0.004737401381134987, + 0.009168541058897972, + 0.004959087818861008, + 0.00508036557585001, + 0.003887882921844721, + 0.00751577178016305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.07905707508325577 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20165744423866272 + ], + "opposite_logprob_tokens": [ + -1.7016574144363403 + ], + "logprob_reward": -0.20165744423866272, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20165744423866272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09167953580617905, + 0.09674616903066635, + 0.22152917087078094, + 0.21448886394500732, + 0.08089318871498108, + 0.0784650593996048, + 0.06840916723012924, + 0.02501390501856804, + 0.11161588877439499, + 0.011159101501107216 + ], + "credit_map": [ + 0.010809913277626038, + 0.011924484744668007, + 0.008645813912153244, + 0.021961795166134834, + 0.007805009838193655, + 0.011668600142002106, + 0.01116902008652687, + 0.007694898638874292, + 0.013111449778079987, + 0.013473072089254856, + 0.005609840154647827, + 0.023363376036286354, + 0.011030502617359161, + 0.013036957941949368, + 0.011995818465948105, + 0.005125144962221384, + 0.03292241320014, + 0.021474644541740417, + 0.013592567294836044, + 0.043852150440216064, + 0.022071922197937965, + 0.037773679941892624, + 0.02483913116157055, + 0.025002650916576385, + 0.048304203897714615, + 0.012389134615659714, + 0.008284692652523518, + 0.035229094326496124, + 0.017641475424170494, + 0.05767897516489029, + 0.01952187716960907, + 0.015439403243362904, + 0.011329918168485165, + 0.007796185091137886, + 0.006093725096434355, + 0.02213464304804802, + 0.008625342510640621, + 0.012024462223052979, + 0.007727297488600016, + 0.005161614622920752, + 0.008092901669442654, + 0.006436347030103207, + 0.005888460669666529, + 0.012684822082519531, + 0.016158325597643852, + 0.009297311305999756, + 0.005758896470069885, + 0.014147997833788395, + 0.006634835619479418, + 0.006244586780667305, + 0.006681382190436125, + 0.018916307017207146, + 0.006416114512830973, + 0.008920756168663502, + 0.00814723689109087, + 0.006447950378060341, + 0.0034321516286581755, + 0.0025958435144275427, + 0.0018605109071359038, + 0.004111568909138441, + 0.0049066427163779736, + 0.0027130835223942995, + 0.002344585256651044, + 0.003049519844353199, + 0.010801125317811966, + 0.013202786445617676, + 0.013051401823759079, + 0.027698898687958717, + 0.011900614015758038, + 0.013017304241657257, + 0.011002849787473679, + 0.010940906591713428, + 0.0008962807478383183, + 0.0009600315825082362, + 0.0013753786915913224, + 0.002593287033960223, + 0.0011258850572630763, + 0.001344638760201633, + 0.0011217505671083927, + 0.001741848886013031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.20165744423866272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983392715454102 + ], + "opposite_logprob_tokens": [ + -0.5233392715454102 + ], + "logprob_reward": -0.8983392715454102, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983392715454102, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11648215353488922, + 0.05975175276398659, + 0.26397940516471863, + 0.08459631353616714, + 0.07631765305995941, + 0.13470664620399475, + 0.08700178563594818, + 0.043922387063503265, + 0.08547642827033997, + 0.04776551201939583 + ], + "credit_map": [ + 0.011481033638119698, + 0.013934263959527016, + 0.01100257970392704, + 0.02378404140472412, + 0.009620051831007004, + 0.01361449621617794, + 0.012735900469124317, + 0.02030978351831436, + 0.008827748708426952, + 0.0071749915368855, + 0.0040955631993710995, + 0.012906145304441452, + 0.0056269122287631035, + 0.00788176991045475, + 0.008556677028536797, + 0.004681940656155348, + 0.04454025626182556, + 0.02454964630305767, + 0.014465841464698315, + 0.052599407732486725, + 0.02173401415348053, + 0.04983324557542801, + 0.027202244848012924, + 0.029054740443825722, + 0.016578223556280136, + 0.00535361235961318, + 0.004845533519983292, + 0.01417685579508543, + 0.007458431646227837, + 0.018802622333168983, + 0.008844751864671707, + 0.008536276407539845, + 0.01002996601164341, + 0.005469407886266708, + 0.00914894137531519, + 0.016338404268026352, + 0.008619597181677818, + 0.009743832983076572, + 0.005393967963755131, + 0.011573543772101402, + 0.01258454192429781, + 0.014110629446804523, + 0.008707473054528236, + 0.03335702419281006, + 0.022166293114423752, + 0.016693979501724243, + 0.011738327331840992, + 0.01534837856888771, + 0.00901583582162857, + 0.010442247614264488, + 0.006893895100802183, + 0.019488109275698662, + 0.00960296019911766, + 0.009968297556042671, + 0.010308535769581795, + 0.011281903833150864, + 0.007018632255494595, + 0.004453141707926989, + 0.0032969829626381397, + 0.005880437791347504, + 0.006923155393451452, + 0.004896608646959066, + 0.004125962499529123, + 0.007327467203140259, + 0.007639672607183456, + 0.01039471011608839, + 0.0096229063346982, + 0.02096850425004959, + 0.009611431509256363, + 0.009777161292731762, + 0.008773934096097946, + 0.008688103407621384, + 0.0052311657927930355, + 0.00343874073587358, + 0.005452125798910856, + 0.010691849514842033, + 0.004893189761787653, + 0.004810992628335953, + 0.0043054367415606976, + 0.008942008949816227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.8983392715454102 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In the style of Ghibli animation, a sweeping wide shot, like that from a low-flying drone, glides gracefully across a vast expanse of sun-drenched, rolling meadows. The camera slowly pans right, revealing a massive, hill-sized pumpkin perched in the center of the field, its vibrant orange rind deeply grooved and textured with a soft, painterly texture. The pumpkin is meticulously carved into a rustic, whimsical house, with a small, round wooden door and several circular windows. Flanked by dense, leafy green vines, they curl around a curving, chimney-like stem at the top. The entire scene is bathed in warm, golden afternoon light, casting long, soft shadows across the tall, emerald green grass, which sways hypnotically in the breeze. A massive, sea urchin-like creature, curled up asleep on the pumpkin's roof; its spines are not sharp but soft and plush, like velvety tendrils covered in a layer of delicate green moss. As the camera settles into position, the creature begins to stir, slowly and deliberately emerging from its slumber. Its mossy spine trembles gently, and then its body unfolds section by section, a gentle and peaceful awakening that perfectly captures the serene and magical atmosphere.", + "seed": 11782, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_14/index_287_seed_11782_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2686973962932825, + "vlm_logprob": -0.2686973962932825, + "vlm_yes_no_margin": 2.1249999523162844, + "vlm_token_credit": -0.2686973962932825, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019538093358278275 + ], + "opposite_logprob_tokens": [ + -4.019537925720215 + ], + "logprob_reward": -0.019538093358278275, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.019538093358278275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11608660966157913, + 0.16803514957427979, + 0.03570730239152908, + 0.04783981293439865, + 0.09411623328924179, + 0.18836407363414764, + 0.04780234396457672, + 0.2294340282678604, + 0.0456320084631443, + 0.026982396841049194 + ], + "credit_map": [ + 0.018109818920493126, + 0.010483498685061932, + 0.013826911337673664, + 0.009680266492068768, + 0.013507382944226265, + 0.021024400368332863, + 0.0122528076171875, + 0.017201529815793037, + 0.05492469668388367, + 0.008983420208096504, + 0.007213970646262169, + 0.008138199336826801, + 0.012631518766283989, + 0.06143394112586975, + 0.007955475710332394, + 0.006753914523869753, + 0.006548101548105478, + 0.003989992197602987, + 0.004007427487522364, + 0.003239076817408204, + 0.005167722702026367, + 0.005438163410872221, + 0.003987488802522421, + 0.003329330123960972, + 0.0065546464174985886, + 0.006843884941190481, + 0.004153883550316095, + 0.008465653285384178, + 0.0060716778971254826, + 0.0042077419348061085, + 0.007050590589642525, + 0.004491731524467468, + 0.02656487002968788, + 0.011375019326806068, + 0.00708726653829217, + 0.005919644609093666, + 0.009146509692072868, + 0.012086077593266964, + 0.009371628984808922, + 0.012565218843519688, + 0.021406035870313644, + 0.019160829484462738, + 0.018338195979595184, + 0.029476284980773926, + 0.025298092514276505, + 0.022016683593392372, + 0.03281860426068306, + 0.019849352538585663, + 0.0067543149925768375, + 0.007645426783710718, + 0.003988961223512888, + 0.006602861452847719, + 0.005326299462467432, + 0.005367591045796871, + 0.007375979796051979, + 0.004740910604596138, + 0.02604248933494091, + 0.031345292925834656, + 0.021359574049711227, + 0.039992716163396835, + 0.024075254797935486, + 0.029835455119609833, + 0.03756237030029297, + 0.019220875576138496, + 0.005142178386449814, + 0.007606327533721924, + 0.005609550513327122, + 0.007857018150389194, + 0.004403020720928907, + 0.0059875184670090675, + 0.0050844307988882065, + 0.003941967152059078, + 0.004315873607993126, + 0.003594702575355768, + 0.0021589777898043394, + 0.004476137459278107, + 0.0034254209604114294, + 0.0028050432447344065, + 0.003039934439584613, + 0.003166306996718049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.019538093358278275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207405403256416 + ], + "opposite_logprob_tokens": [ + -2.8120739459991455 + ], + "logprob_reward": -0.06207405403256416, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207405403256416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15218046307563782, + 0.17430521547794342, + 0.035110168159008026, + 0.22374248504638672, + 0.12176758795976639, + 0.08353432267904282, + 0.03867533057928085, + 0.04276552423834801, + 0.09578058868646622, + 0.03213829547166824 + ], + "credit_map": [ + 0.02442883886396885, + 0.01451192144304514, + 0.017123620957136154, + 0.01898961327970028, + 0.018243912607431412, + 0.020496133714914322, + 0.019731227308511734, + 0.018655190244317055, + 0.04906328395009041, + 0.01327090710401535, + 0.00870386604219675, + 0.011086393147706985, + 0.013266504742205143, + 0.05404719337821007, + 0.01873919554054737, + 0.006127878557890654, + 0.004561563953757286, + 0.0034281855914741755, + 0.004697447642683983, + 0.00460017891600728, + 0.005263767205178738, + 0.0036963720340281725, + 0.005040914285928011, + 0.0038217371329665184, + 0.028665922582149506, + 0.021392257884144783, + 0.026384500786662102, + 0.03667561337351799, + 0.018058424815535545, + 0.02215791866183281, + 0.05058945342898369, + 0.01981838047504425, + 0.015612372197210789, + 0.014628597535192966, + 0.011320553719997406, + 0.023986853659152985, + 0.011374818161129951, + 0.013624368235468864, + 0.01863095909357071, + 0.012589073739945889, + 0.010647257789969444, + 0.008082421496510506, + 0.00810295157134533, + 0.017819369211792946, + 0.009990852326154709, + 0.0101658059284091, + 0.011087020859122276, + 0.007638643961399794, + 0.004453282803297043, + 0.003613333450630307, + 0.005100756883621216, + 0.006788484752178192, + 0.003705390030518174, + 0.0041125002317130566, + 0.006326924078166485, + 0.00457465834915638, + 0.004369623027741909, + 0.00495833670720458, + 0.0039440481923520565, + 0.0057556829415261745, + 0.005396710708737373, + 0.004415511153638363, + 0.008997397497296333, + 0.0049282158724963665, + 0.011165152303874493, + 0.018314821645617485, + 0.010091882199048996, + 0.01681816764175892, + 0.008154514245688915, + 0.01169079914689064, + 0.010654754936695099, + 0.008890497498214245, + 0.006173641886562109, + 0.004220043774694204, + 0.0025240599643439054, + 0.00281506753526628, + 0.005546384956687689, + 0.0035613374784588814, + 0.0035371941048651934, + 0.003760569728910923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.06207405403256416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981058694422245 + ], + "opposite_logprob_tokens": [ + -3.529810667037964 + ], + "logprob_reward": -0.02981058694422245, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981058694422245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05081253871321678, + 0.3944588303565979, + 0.06101461499929428, + 0.07015252858400345, + 0.051995012909173965, + 0.14374616742134094, + 0.03714877367019653, + 0.04997660219669342, + 0.11251827329397202, + 0.02817661128938198 + ], + "credit_map": [ + 0.008874866180121899, + 0.004480634815990925, + 0.00530615309253335, + 0.005771920550614595, + 0.00565290218219161, + 0.008371198549866676, + 0.0056773582473397255, + 0.0066775050945580006, + 0.12572619318962097, + 0.02203669585287571, + 0.021539047360420227, + 0.01874309964478016, + 0.031598590314388275, + 0.1349506974220276, + 0.02191852778196335, + 0.01794601045548916, + 0.01444372907280922, + 0.0049699279479682446, + 0.01061997003853321, + 0.00456004124134779, + 0.005167391616851091, + 0.010430107824504375, + 0.00407975260168314, + 0.00674369465559721, + 0.0075560687109827995, + 0.007441012188792229, + 0.006006864830851555, + 0.013292692601680756, + 0.008720304816961288, + 0.007458413485437632, + 0.015035883523523808, + 0.004641287960112095, + 0.007831405848264694, + 0.005630094092339277, + 0.004854111932218075, + 0.00969890970736742, + 0.004825477954000235, + 0.005999299697577953, + 0.008540491573512554, + 0.004615224432200193, + 0.01461866870522499, + 0.0157591812312603, + 0.013277275487780571, + 0.016339318826794624, + 0.020699331536889076, + 0.016644755378365517, + 0.029329942539334297, + 0.017077701166272163, + 0.005170889664441347, + 0.005606507882475853, + 0.0030430175829678774, + 0.004467683378607035, + 0.004440354183316231, + 0.004205686971545219, + 0.005787907633930445, + 0.004426727071404457, + 0.006023542955517769, + 0.00612871628254652, + 0.004059046972543001, + 0.007063011173158884, + 0.004990957211703062, + 0.005295249633491039, + 0.00903298705816269, + 0.007383089978247881, + 0.01377808302640915, + 0.02326233685016632, + 0.010192874819040298, + 0.014961832202970982, + 0.012026913464069366, + 0.014117174781858921, + 0.014001128263771534, + 0.010177930817008018, + 0.004316795151680708, + 0.004592603538185358, + 0.0020068218000233173, + 0.0038383740466088057, + 0.0030216339509934187, + 0.0032183637376874685, + 0.004096215590834618, + 0.003085804171860218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.02981058694422245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742923974990845 + ], + "opposite_logprob_tokens": [ + -0.9742923974990845 + ], + "logprob_reward": -0.4742923974990845, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742923974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20887121558189392, + 0.16732469201087952, + 0.061051059514284134, + 0.19921451807022095, + 0.05216836556792259, + 0.1147378534078598, + 0.020637057721614838, + 0.08053687959909439, + 0.07270166277885437, + 0.022756725549697876 + ], + "credit_map": [ + 0.030843092128634453, + 0.020678864791989326, + 0.020974930375814438, + 0.03295448049902916, + 0.025008399039506912, + 0.02538393996655941, + 0.02817530184984207, + 0.02485220693051815, + 0.027255075052380562, + 0.020880719646811485, + 0.015553129836916924, + 0.0173801239579916, + 0.01828855648636818, + 0.028380965813994408, + 0.028085874393582344, + 0.011500246822834015, + 0.013105545192956924, + 0.00745547516271472, + 0.0061521935276687145, + 0.004924736451357603, + 0.00881072785705328, + 0.010153906419873238, + 0.005565505009144545, + 0.004882965702563524, + 0.02416244149208069, + 0.02436227910220623, + 0.023774925619363785, + 0.02818388119339943, + 0.021858960390090942, + 0.01846332475543022, + 0.03882058709859848, + 0.019588109105825424, + 0.007538252510130405, + 0.006864503491669893, + 0.004795042332261801, + 0.010330804623663425, + 0.00504123093560338, + 0.006007064133882523, + 0.007498528808355331, + 0.004092932678759098, + 0.015057556331157684, + 0.012639694847166538, + 0.010303138755261898, + 0.02700752019882202, + 0.009645854122936726, + 0.01266405638307333, + 0.016797665506601334, + 0.010622357949614525, + 0.002826330251991749, + 0.0037091572303324938, + 0.0016561715165153146, + 0.00270267017185688, + 0.0026554798241704702, + 0.0024929887149482965, + 0.0030943371821194887, + 0.001499920734204352, + 0.009952844120562077, + 0.010505631566047668, + 0.007078008260577917, + 0.01127556525170803, + 0.008382796309888363, + 0.009013337083160877, + 0.017722109332680702, + 0.006606581620872021, + 0.008081472478806973, + 0.009491978213191032, + 0.009586824104189873, + 0.013142473995685577, + 0.0076553914695978165, + 0.009375626221299171, + 0.008912380784749985, + 0.006455510854721069, + 0.0035920157097280025, + 0.004081130027770996, + 0.0016365968622267246, + 0.002866703551262617, + 0.002301588887348771, + 0.0024980821181088686, + 0.002277289517223835, + 0.0035033165477216244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.4742923974990845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577718496322632 + ], + "opposite_logprob_tokens": [ + -0.6327718496322632 + ], + "logprob_reward": -0.7577718496322632, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577718496322632, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05407628417015076, + 0.06866802275180817, + 0.064134880900383, + 0.1892615109682083, + 0.07855299860239029, + 0.26867222785949707, + 0.03469521179795265, + 0.22092878818511963, + 0.016587868332862854, + 0.00442213099449873 + ], + "credit_map": [ + 0.010669449344277382, + 0.004157237242907286, + 0.005347843281924725, + 0.00350682414136827, + 0.005324812605977058, + 0.01288523618131876, + 0.004695865325629711, + 0.007489012088626623, + 0.022518308833241463, + 0.004108196124434471, + 0.0030232486315071583, + 0.0029941119719296694, + 0.005220732651650906, + 0.025089425966143608, + 0.003227995242923498, + 0.0024860051926225424, + 0.012933654710650444, + 0.007461761590093374, + 0.007495845668017864, + 0.00450986297801137, + 0.009501681663095951, + 0.010087759234011173, + 0.0065186601132154465, + 0.00562565540894866, + 0.01867455430328846, + 0.02833562158048153, + 0.021107066422700882, + 0.02560037560760975, + 0.024298781529068947, + 0.016853779554367065, + 0.0380738265812397, + 0.016317501664161682, + 0.012408614158630371, + 0.011276996694505215, + 0.006794861517846584, + 0.012735741212964058, + 0.008250921964645386, + 0.009225722402334213, + 0.010041928850114346, + 0.007818212732672691, + 0.03143828362226486, + 0.03220339119434357, + 0.021824542433023453, + 0.0379917211830616, + 0.03725746273994446, + 0.0310856644064188, + 0.04770601913332939, + 0.029165156185626984, + 0.005381631199270487, + 0.005357022862881422, + 0.002914713229984045, + 0.004847982432693243, + 0.00394435552880168, + 0.003849125001579523, + 0.005685437470674515, + 0.0027149454690515995, + 0.022946981713175774, + 0.026767633855342865, + 0.020761672407388687, + 0.03711845725774765, + 0.024882426485419273, + 0.027110066264867783, + 0.04266009479761124, + 0.018681468442082405, + 0.0022649371530860662, + 0.0014578913105651736, + 0.002100354293361306, + 0.0033735521137714386, + 0.00165047578047961, + 0.0015508017968386412, + 0.002093905583024025, + 0.002095951233059168, + 0.0006871496443636715, + 0.0005313857109285891, + 0.00026120577240362763, + 0.0006067616050131619, + 0.0006303535774350166, + 0.00048530561616644263, + 0.0006058572907932103, + 0.0006141120684333146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.7577718496322632 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In the style of Ghibli animation, a sweeping wide shot, like that from a low-flying drone, glides gracefully across a vast expanse of sun-drenched, rolling meadows. The camera slowly pans right, revealing a massive, hill-sized pumpkin perched in the center of the field, its vibrant orange rind deeply grooved and textured with a soft, painterly texture. The pumpkin is meticulously carved into a rustic, whimsical house, with a small, round wooden door and several circular windows. Flanked by dense, leafy green vines, they curl around a curving, chimney-like stem at the top. The entire scene is bathed in warm, golden afternoon light, casting long, soft shadows across the tall, emerald green grass, which sways hypnotically in the breeze. A massive, sea urchin-like creature, curled up asleep on the pumpkin's roof; its spines are not sharp but soft and plush, like velvety tendrils covered in a layer of delicate green moss. As the camera settles into position, the creature begins to stir, slowly and deliberately emerging from its slumber. Its mossy spine trembles gently, and then its body unfolds section by section, a gentle and peaceful awakening that perfectly captures the serene and magical atmosphere.", + "seed": 11783, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_14/index_287_seed_11783_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.22335703130811452, + "vlm_logprob": -0.22335703130811452, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.22335703130811452, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006301296874880791 + ], + "opposite_logprob_tokens": [ + -5.256301403045654 + ], + "logprob_reward": -0.006301296874880791, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006301296874880791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12438859790563583, + 0.06796187907457352, + 0.06561890244483948, + 0.20774324238300323, + 0.04986586049199104, + 0.0797906145453453, + 0.15156279504299164, + 0.03918793052434921, + 0.05298406258225441, + 0.16089606285095215 + ], + "credit_map": [ + 0.01680627278983593, + 0.015153148211538792, + 0.016940422356128693, + 0.01297538261860609, + 0.016553252935409546, + 0.017402801662683487, + 0.016831988468766212, + 0.011725344695150852, + 0.00977658573538065, + 0.0076078008860349655, + 0.005926590878516436, + 0.009262247942388058, + 0.0071193440817296505, + 0.007654625456780195, + 0.013926856219768524, + 0.006687830202281475, + 0.01185383927077055, + 0.006843763403594494, + 0.007564140018075705, + 0.009042002260684967, + 0.007263263687491417, + 0.007876090705394745, + 0.009023934602737427, + 0.00615187780931592, + 0.022704709321260452, + 0.018994588404893875, + 0.036933086812496185, + 0.02236694097518921, + 0.01743662729859352, + 0.021661581471562386, + 0.023795679211616516, + 0.043850045651197433, + 0.005838283337652683, + 0.005438818130642176, + 0.00563277630135417, + 0.007342727389186621, + 0.0052561149932444096, + 0.005728115327656269, + 0.008598202839493752, + 0.006030827760696411, + 0.013094153255224228, + 0.005515050143003464, + 0.008747730404138565, + 0.010250359773635864, + 0.005278643220663071, + 0.018655454739928246, + 0.012001006864011288, + 0.006248221732676029, + 0.05712957680225372, + 0.005730270408093929, + 0.007323643192648888, + 0.0033306421246379614, + 0.004775592591613531, + 0.06165899336338043, + 0.004988153465092182, + 0.006625938694924116, + 0.008206508122384548, + 0.0028945067897439003, + 0.004606231581419706, + 0.004438549745827913, + 0.00281313294544816, + 0.006890361197292805, + 0.004672108218073845, + 0.004666536580771208, + 0.007804648485034704, + 0.004307272844016552, + 0.006138138473033905, + 0.0077520934864878654, + 0.004542181268334389, + 0.006880934815853834, + 0.0067287711426615715, + 0.008830028586089611, + 0.026549700647592545, + 0.0074430182576179504, + 0.03177042677998543, + 0.011316457763314247, + 0.006162770092487335, + 0.05103926360607147, + 0.008954605087637901, + 0.017659839242696762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.006301296874880791 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956635810434818 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956635810434818, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956635810434818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11801557242870331, + 0.1754169911146164, + 0.08349255472421646, + 0.17105302214622498, + 0.03319445997476578, + 0.04585970565676689, + 0.16238439083099365, + 0.04894855618476868, + 0.02657703496515751, + 0.13505768775939941 + ], + "credit_map": [ + 0.01426386833190918, + 0.016399124637246132, + 0.011152384802699089, + 0.011300665326416492, + 0.02068084478378296, + 0.015976229682564735, + 0.01528807170689106, + 0.012954371981322765, + 0.017678434029221535, + 0.021165234968066216, + 0.01436591800302267, + 0.02594493329524994, + 0.02263559028506279, + 0.019855212420225143, + 0.03497616946697235, + 0.018795505166053772, + 0.010392097756266594, + 0.010008864104747772, + 0.0076487427577376366, + 0.01408641692250967, + 0.01037079468369484, + 0.00916488841176033, + 0.0158726517111063, + 0.005948096979409456, + 0.019529247656464577, + 0.01443484891206026, + 0.02994166500866413, + 0.01713152974843979, + 0.014902239665389061, + 0.01870587095618248, + 0.019328419119119644, + 0.037079207599163055, + 0.003889982122927904, + 0.002737488131970167, + 0.006885610055178404, + 0.003845605533570051, + 0.002724254271015525, + 0.0036523151211440563, + 0.004677204415202141, + 0.004782001953572035, + 0.006884493865072727, + 0.003629392944276333, + 0.004565514158457518, + 0.006001359783113003, + 0.004105292726308107, + 0.00781466905027628, + 0.008367524482309818, + 0.004491457715630531, + 0.05868037790060043, + 0.0061533162370324135, + 0.009894165210425854, + 0.00306163402274251, + 0.005601827055215836, + 0.06589765101671219, + 0.005382006987929344, + 0.0077134124003350735, + 0.014982623979449272, + 0.002114734845235944, + 0.003277694107964635, + 0.0019963926170021296, + 0.0026312903501093388, + 0.01739008165895939, + 0.002638214500620961, + 0.003917521331459284, + 0.003221259219571948, + 0.002405969426035881, + 0.0037459528539329767, + 0.004450618755072355, + 0.002160683972761035, + 0.002550214296206832, + 0.004925085697323084, + 0.0031172505114227533, + 0.02605677768588066, + 0.005786105524748564, + 0.026980359107255936, + 0.005247367080301046, + 0.003985567484050989, + 0.04700386896729469, + 0.00741189531981945, + 0.01258573867380619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.005956635810434818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349563151597977 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349563151597977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349563151597977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0588812530040741, + 0.023852000012993813, + 0.035699889063835144, + 0.3669019043445587, + 0.014981590211391449, + 0.042947378009557724, + 0.0902031809091568, + 0.1050935909152031, + 0.055315904319286346, + 0.20612329244613647 + ], + "credit_map": [ + 0.006625274661928415, + 0.008015560917556286, + 0.006225558463484049, + 0.007245394866913557, + 0.010258115828037262, + 0.006376569624990225, + 0.009558158926665783, + 0.004576624371111393, + 0.003801874117925763, + 0.001991097116842866, + 0.0031857185531407595, + 0.0021627985406666994, + 0.0022573289461433887, + 0.00307334098033607, + 0.0034888661466538906, + 0.003890975611284375, + 0.005412801634520292, + 0.004056221339851618, + 0.003226945409551263, + 0.005665866192430258, + 0.004232417326420546, + 0.0039056111127138138, + 0.006610314827412367, + 0.0025897135492414236, + 0.041760753840208054, + 0.02962125837802887, + 0.05480137839913368, + 0.03207996115088463, + 0.037338681519031525, + 0.04554898664355278, + 0.04872123897075653, + 0.07702969014644623, + 0.0019167782738804817, + 0.001614104607142508, + 0.002393909730017185, + 0.001990755321457982, + 0.0013955524191260338, + 0.00174325134139508, + 0.0019450164400041103, + 0.0019822227768599987, + 0.008266250602900982, + 0.0024858107790350914, + 0.0043058861047029495, + 0.005195261910557747, + 0.0026639243587851524, + 0.00975286029279232, + 0.005993291735649109, + 0.004284094087779522, + 0.020429177209734917, + 0.004261768888682127, + 0.012337855063378811, + 0.006029836367815733, + 0.004341954831033945, + 0.02916555665433407, + 0.007076998241245747, + 0.006560044828802347, + 0.03294533118605614, + 0.0044471388682723045, + 0.006261490285396576, + 0.005174317862838507, + 0.005255149211734533, + 0.03782197833061218, + 0.0060764458030462265, + 0.007111754734069109, + 0.008209732361137867, + 0.004352982621639967, + 0.0063352519646286964, + 0.009825797751545906, + 0.004887931980192661, + 0.006735080387443304, + 0.00767948292195797, + 0.0072896466590464115, + 0.038086917251348495, + 0.00843135081231594, + 0.04373437166213989, + 0.008431735448539257, + 0.006188570987433195, + 0.07051832973957062, + 0.0104312589392066, + 0.020300770178437233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.026349563151597977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207969546318054 + ], + "opposite_logprob_tokens": [ + -1.502079725265503 + ], + "logprob_reward": -0.25207969546318054, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207969546318054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03134313225746155, + 0.030916262418031693, + 0.011360589414834976, + 0.5508843660354614, + 0.05181380733847618, + 0.07677619159221649, + 0.05117432028055191, + 0.010192913003265858, + 0.03505546227097511, + 0.15048295259475708 + ], + "credit_map": [ + 0.00275283376686275, + 0.00481130788102746, + 0.0024723964743316174, + 0.00354204629547894, + 0.006037910468876362, + 0.0038273523095995188, + 0.005181366577744484, + 0.0027179215103387833, + 0.003014639252796769, + 0.003184006316587329, + 0.002705175895243883, + 0.004874622914940119, + 0.0037565305829048157, + 0.0034192719031125307, + 0.006647339556366205, + 0.003314678091555834, + 0.0015553116099908948, + 0.001279265503399074, + 0.0009666936821304262, + 0.0014700123574584723, + 0.0014728523092344403, + 0.0015341980615630746, + 0.0019500815542414784, + 0.0011321750935167074, + 0.06402982026338577, + 0.04528127610683441, + 0.08087193965911865, + 0.04731087014079094, + 0.056001484394073486, + 0.06972208619117737, + 0.07375671714544296, + 0.11391016840934753, + 0.005832475610077381, + 0.005651224870234728, + 0.005817243363708258, + 0.0074022128246724606, + 0.0051677050068974495, + 0.006283024325966835, + 0.008621909655630589, + 0.007038016803562641, + 0.010987634770572186, + 0.0055421506986021996, + 0.008704033680260181, + 0.010541022755205631, + 0.005519253667443991, + 0.0159178264439106, + 0.013056362047791481, + 0.00650791497901082, + 0.006039006169885397, + 0.004417047835886478, + 0.009500754065811634, + 0.0076164547353982925, + 0.004043844062834978, + 0.007470411714166403, + 0.007361209951341152, + 0.004725595470517874, + 0.001851397566497326, + 0.0007966991979628801, + 0.0012575689470395446, + 0.0012180557241663337, + 0.0007974599138833582, + 0.0017850031144917011, + 0.0014525232836604118, + 0.0010342050809413195, + 0.004521054215729237, + 0.0028977985493838787, + 0.005132361315190792, + 0.007785473484545946, + 0.0022780923172831535, + 0.0035505075938999653, + 0.00491391122341156, + 0.003976267296820879, + 0.02702726237475872, + 0.006204553414136171, + 0.03177335113286972, + 0.0066045429557561874, + 0.0052055539563298225, + 0.049800001084804535, + 0.0078502856194973, + 0.016017409041523933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.25207969546318054 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8260979652404785 + ], + "opposite_logprob_tokens": [ + -0.5760979652404785 + ], + "logprob_reward": -0.8260979652404785, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8260979652404785, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.040249597281217575, + 0.02142605558037758, + 0.021970367059111595, + 0.7802909016609192, + 0.020520733669400215, + 0.021233366802334785, + 0.024256592616438866, + 0.020076068118214607, + 0.018966523930430412, + 0.031009813770651817 + ], + "credit_map": [ + 0.004778019618242979, + 0.005241324193775654, + 0.0042427899315953255, + 0.0058647277764976025, + 0.004974601790308952, + 0.004454746376723051, + 0.006742809899151325, + 0.003950578160583973, + 0.00287772249430418, + 0.002226721029728651, + 0.0019359030993655324, + 0.0031024280469864607, + 0.002714455360546708, + 0.002366500673815608, + 0.0035660334397107363, + 0.0026362896896898746, + 0.006035219877958298, + 0.0024585023056715727, + 0.0028788726776838303, + 0.0018179997568950057, + 0.001985713839530945, + 0.002955361269414425, + 0.0017995884409174323, + 0.0020391091238707304, + 0.09118665009737015, + 0.06487463414669037, + 0.11429285258054733, + 0.06893723458051682, + 0.08037110418081284, + 0.09739463031291962, + 0.10345328599214554, + 0.15978044271469116, + 0.0021881668362766504, + 0.0019749468192458153, + 0.0030098611023277044, + 0.0025412144605070353, + 0.0019971849396824837, + 0.0024747850839048624, + 0.003480575978755951, + 0.0028539958875626326, + 0.0029301426839083433, + 0.0014424044638872147, + 0.002390046138316393, + 0.0027664494700729847, + 0.0016464146319776773, + 0.00419146753847599, + 0.0037524974904954433, + 0.0021139425225555897, + 0.007371149025857449, + 0.0011557529214769602, + 0.0016951473662629724, + 0.0016124325338751078, + 0.0010682668071240187, + 0.008503050543367863, + 0.0015117889270186424, + 0.0013390040257945657, + 0.007521330378949642, + 0.0006969079840928316, + 0.0009490707307122648, + 0.0006223886157386005, + 0.0006136196898296475, + 0.0077957576140761375, + 0.0007672584615647793, + 0.0011097327806055546, + 0.0036873877979815006, + 0.0014837944181635976, + 0.001910537132062018, + 0.002855639671906829, + 0.0016198822995647788, + 0.0031814617104828358, + 0.002128828316926956, + 0.0020989906042814255, + 0.004202228505164385, + 0.0019270592601969838, + 0.004524601623415947, + 0.003386403899639845, + 0.0019202816765755415, + 0.0071953157894313335, + 0.003015710972249508, + 0.004838209133595228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.8260979652404785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In the style of Ghibli animation, a sweeping wide shot, like that from a low-flying drone, glides gracefully across a vast expanse of sun-drenched, rolling meadows. The camera slowly pans right, revealing a massive, hill-sized pumpkin perched in the center of the field, its vibrant orange rind deeply grooved and textured with a soft, painterly texture. The pumpkin is meticulously carved into a rustic, whimsical house, with a small, round wooden door and several circular windows. Flanked by dense, leafy green vines, they curl around a curving, chimney-like stem at the top. The entire scene is bathed in warm, golden afternoon light, casting long, soft shadows across the tall, emerald green grass, which sways hypnotically in the breeze. A massive, sea urchin-like creature, curled up asleep on the pumpkin's roof; its spines are not sharp but soft and plush, like velvety tendrils covered in a layer of delicate green moss. As the camera settles into position, the creature begins to stir, slowly and deliberately emerging from its slumber. Its mossy spine trembles gently, and then its body unfolds section by section, a gentle and peaceful awakening that perfectly captures the serene and magical atmosphere.", + "seed": 11784, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_14/index_287_seed_11784_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.329130226932466, + "vlm_logprob": -0.329130226932466, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.329130226932466, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06401707977056503 + ], + "opposite_logprob_tokens": [ + -2.8140170574188232 + ], + "logprob_reward": -0.06401707977056503, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06401707977056503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13444790244102478, + 0.07178465276956558, + 0.047761742025613785, + 0.02113218791782856, + 0.07471036165952682, + 0.18678660690784454, + 0.07887105643749237, + 0.06873016059398651, + 0.054556481540203094, + 0.2612188458442688 + ], + "credit_map": [ + 0.027064083144068718, + 0.011084840632975101, + 0.026450805366039276, + 0.016357440501451492, + 0.009189105592668056, + 0.012217448092997074, + 0.016789929941296577, + 0.01529424823820591, + 0.02401111088693142, + 0.003490854287520051, + 0.020037420094013214, + 0.004946380387991667, + 0.005896825343370438, + 0.005022523459047079, + 0.002793150721117854, + 0.005586391780525446, + 0.008989200927317142, + 0.0036117974668741226, + 0.006991947535425425, + 0.006469107232987881, + 0.003950832411646843, + 0.0044640409760177135, + 0.007501725107431412, + 0.005783087108284235, + 0.004295709077268839, + 0.0018641476053744555, + 0.002338525140658021, + 0.0019653288181871176, + 0.002154557965695858, + 0.0023469605948776007, + 0.003566286526620388, + 0.002600673586130142, + 0.010180425830185413, + 0.00886293314397335, + 0.008785624988377094, + 0.013955671340227127, + 0.008743642829358578, + 0.008544579148292542, + 0.00944871548563242, + 0.006188770290464163, + 0.024682477116584778, + 0.01687312312424183, + 0.021075110882520676, + 0.026275737211108208, + 0.03410718962550163, + 0.01702364720404148, + 0.02880994603037834, + 0.017939368262887, + 0.019220100715756416, + 0.006869346369057894, + 0.011889795772731304, + 0.008409020490944386, + 0.006846569944173098, + 0.01205400936305523, + 0.0081007806584239, + 0.005481428001075983, + 0.018363306298851967, + 0.004809512756764889, + 0.020837783813476562, + 0.006294497288763523, + 0.0035566354636102915, + 0.0035753133706748486, + 0.0075255464762449265, + 0.0037675637286156416, + 0.00500608142465353, + 0.008237937465310097, + 0.004353319760411978, + 0.008944042958319187, + 0.008994760923087597, + 0.004248116165399551, + 0.007044680416584015, + 0.007727542892098427, + 0.038403868675231934, + 0.023914296180009842, + 0.030794303864240646, + 0.04093426093459129, + 0.027203869074583054, + 0.026616323739290237, + 0.04300493374466896, + 0.030346985906362534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.06401707977056503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289182126522064 + ], + "opposite_logprob_tokens": [ + -1.0539182424545288 + ], + "logprob_reward": -0.4289182126522064, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289182126522064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2704247832298279, + 0.03483913838863373, + 0.042205244302749634, + 0.07516323775053024, + 0.08872025460004807, + 0.12054181098937988, + 0.06467879563570023, + 0.08732524514198303, + 0.12319020181894302, + 0.0929112359881401 + ], + "credit_map": [ + 0.077569879591465, + 0.02062666229903698, + 0.06414967775344849, + 0.021995237097144127, + 0.016414545476436615, + 0.018261456862092018, + 0.02189471200108528, + 0.029512634500861168, + 0.005616442300379276, + 0.0037728522438555956, + 0.005289814434945583, + 0.004664883017539978, + 0.0030018524266779423, + 0.004698197823017836, + 0.004061054438352585, + 0.0037340412382036448, + 0.007498720195144415, + 0.0031977000180631876, + 0.006872921250760555, + 0.0036702011711895466, + 0.0035405622329562902, + 0.00593792786821723, + 0.004202075768262148, + 0.007285136729478836, + 0.00863182358443737, + 0.006693194154649973, + 0.0075067984871566296, + 0.007266514468938112, + 0.008127582259476185, + 0.006773372646421194, + 0.022330624982714653, + 0.0078333280980587, + 0.013965866528451443, + 0.007333244197070599, + 0.01070748083293438, + 0.009472801350057125, + 0.013190328143537045, + 0.010563847608864307, + 0.016027089208364487, + 0.007459594868123531, + 0.01607702672481537, + 0.013893715105950832, + 0.01897335797548294, + 0.014017270877957344, + 0.015165152959525585, + 0.018738197162747383, + 0.012654346413910389, + 0.01102274376899004, + 0.017680034041404724, + 0.005779489409178495, + 0.010182473808526993, + 0.008101718500256538, + 0.004881103057414293, + 0.006036624312400818, + 0.007510393857955933, + 0.004506963770836592, + 0.01780962571501732, + 0.01213619764894247, + 0.0209499578922987, + 0.008111968636512756, + 0.006322177592664957, + 0.00804942101240158, + 0.008713580667972565, + 0.005232306197285652, + 0.017686139792203903, + 0.01825891248881817, + 0.011770838871598244, + 0.013648995198309422, + 0.015360310673713684, + 0.013844224624335766, + 0.016285954043269157, + 0.01633482240140438, + 0.01705457828938961, + 0.00761406309902668, + 0.006178935524076223, + 0.011614780873060226, + 0.012116330675780773, + 0.009987886063754559, + 0.019869593903422356, + 0.008475067093968391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.4289182126522064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791205637156963 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.009791205637156963, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791205637156963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10180220752954483, + 0.0610671304166317, + 0.03322085365653038, + 0.027079837396740913, + 0.0272044837474823, + 0.110268734395504, + 0.0392216295003891, + 0.5202802419662476, + 0.0505761094391346, + 0.029278790578246117 + ], + "credit_map": [ + 0.023595813661813736, + 0.008508497849106789, + 0.020483005791902542, + 0.0113105783239007, + 0.007115749176591635, + 0.00814396608620882, + 0.012242830358445644, + 0.010401756502687931, + 0.018332941457629204, + 0.0030026377644389868, + 0.014792478643357754, + 0.004552995320409536, + 0.004264406394213438, + 0.004876955412328243, + 0.006087240297347307, + 0.005157469771802425, + 0.004680586978793144, + 0.002942111110314727, + 0.0057465750724077225, + 0.0035074527841061354, + 0.0030068745836615562, + 0.002789140213280916, + 0.006698697805404663, + 0.003849413013085723, + 0.004119988530874252, + 0.0024701375514268875, + 0.0033953124657273293, + 0.00334052462130785, + 0.0033890572376549244, + 0.0029446131084114313, + 0.004641828127205372, + 0.0027783727273344994, + 0.005257840733975172, + 0.003093054750934243, + 0.004096088465303183, + 0.002646268345415592, + 0.002626798814162612, + 0.004435228183865547, + 0.0020953023340553045, + 0.0029539018869400024, + 0.01314103975892067, + 0.011402466334402561, + 0.01833411678671837, + 0.014007615856826305, + 0.01473463885486126, + 0.01533320639282465, + 0.01291434932500124, + 0.010401284322142601, + 0.005669540725648403, + 0.0035176558885723352, + 0.005291034933179617, + 0.0059493170119822025, + 0.004420503042638302, + 0.005323949735611677, + 0.005522705614566803, + 0.0035269230138510466, + 0.1361396610736847, + 0.03232191875576973, + 0.16087840497493744, + 0.04855635017156601, + 0.02964220754802227, + 0.025489799678325653, + 0.05398193374276161, + 0.033269915729761124, + 0.006456016097217798, + 0.006778504699468613, + 0.004378132522106171, + 0.010625896975398064, + 0.006785500328987837, + 0.0044235046952962875, + 0.0059184967540204525, + 0.005210051778703928, + 0.003985944204032421, + 0.0032653415109962225, + 0.0022753512021154165, + 0.00400103023275733, + 0.004994499962776899, + 0.002845147857442498, + 0.004418157506734133, + 0.0034933157730847597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.009791205637156963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904382586479187 + ], + "opposite_logprob_tokens": [ + -2.4640438556671143 + ], + "logprob_reward": -0.08904382586479187, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904382586479187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04421873390674591, + 0.040850576013326645, + 0.034596897661685944, + 0.05383112654089928, + 0.04535016044974327, + 0.2084021419286728, + 0.10618829727172852, + 0.12681086361408234, + 0.13956166803836823, + 0.200189471244812 + ], + "credit_map": [ + 0.0065985131077468395, + 0.0044164820574223995, + 0.0063518560491502285, + 0.006562883034348488, + 0.0034317064564675093, + 0.004226451739668846, + 0.005960379261523485, + 0.0066704642958939075, + 0.007732497528195381, + 0.0038807138334959745, + 0.005853105802088976, + 0.0050722761079669, + 0.004458925686776638, + 0.0036706174723803997, + 0.006082139443606138, + 0.004100299905985594, + 0.004798859357833862, + 0.002644085790961981, + 0.0045265150256454945, + 0.004658110905438662, + 0.003001041477546096, + 0.0031818600837141275, + 0.0074481843039393425, + 0.004338242579251528, + 0.006695093121379614, + 0.00495125912129879, + 0.006676021497696638, + 0.005790297873318195, + 0.005354527849704027, + 0.005509517155587673, + 0.012954652309417725, + 0.005899757146835327, + 0.008292712271213531, + 0.0035305041819810867, + 0.004463579040020704, + 0.003963925410062075, + 0.004185485187917948, + 0.00687553733587265, + 0.009076807647943497, + 0.004961609374731779, + 0.027622710913419724, + 0.017210442572832108, + 0.018149953335523605, + 0.02920452505350113, + 0.04034506902098656, + 0.018863195553421974, + 0.03605703264474869, + 0.020949210971593857, + 0.017155315726995468, + 0.010344469919800758, + 0.011961204931139946, + 0.018826480954885483, + 0.015000964514911175, + 0.009092320688068867, + 0.015003806911408901, + 0.008803737349808216, + 0.023047052323818207, + 0.015171128325164318, + 0.029933203011751175, + 0.011191901750862598, + 0.010384085588157177, + 0.015144191682338715, + 0.011491281911730766, + 0.010448009707033634, + 0.01967116817831993, + 0.016071336343884468, + 0.012588754296302795, + 0.01983722671866417, + 0.02105080336332321, + 0.014004462398588657, + 0.02161075919866562, + 0.014727165922522545, + 0.02806245908141136, + 0.017373185604810715, + 0.02579059824347496, + 0.031506530940532684, + 0.024757953360676765, + 0.02048102580010891, + 0.02677799202501774, + 0.02543971873819828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.08904382586479187 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of sun-drenched, rolling meadows with tall emerald green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, hill-sized pumpkin carved into a rustic house with a wooden door and circular windows visible in the center of the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, sea urchin-like creature with soft, plush, moss-covered spines sleeping on the pumpkin's roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slowly emerge from its slumber by unfolding its body section by section?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a sweeping wide shot and slowly pan right before settling into position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538808107376099 + ], + "opposite_logprob_tokens": [ + -0.4288807809352875 + ], + "logprob_reward": -1.0538808107376099, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538808107376099, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11373735964298248, + 0.09850439429283142, + 0.06732133030891418, + 0.03691254183650017, + 0.04775075614452362, + 0.09935782104730606, + 0.08611786365509033, + 0.1611787974834442, + 0.19550678133964539, + 0.09361240267753601 + ], + "credit_map": [ + 0.017661413177847862, + 0.012525377795100212, + 0.018726937472820282, + 0.015918830409646034, + 0.00832448247820139, + 0.012760900892317295, + 0.015368521213531494, + 0.012450897134840488, + 0.015361132100224495, + 0.007716248277574778, + 0.014256712980568409, + 0.012624466791749, + 0.008865880779922009, + 0.010644232854247093, + 0.017739448696374893, + 0.011296273209154606, + 0.007330754306167364, + 0.005898390430957079, + 0.010300926864147186, + 0.0074193449690938, + 0.005406567826867104, + 0.008168373256921768, + 0.01399956364184618, + 0.008797405287623405, + 0.005649569444358349, + 0.0029720489401370287, + 0.003770874347537756, + 0.003517249831929803, + 0.005689792800694704, + 0.0037299597170203924, + 0.0085009029135108, + 0.00308214221149683, + 0.007762137800455093, + 0.004646771587431431, + 0.008171121589839458, + 0.004710232373327017, + 0.004265282768756151, + 0.008651729673147202, + 0.0044450643472373486, + 0.005098416935652494, + 0.011984196491539478, + 0.009082919918000698, + 0.010147740133106709, + 0.013184200040996075, + 0.0178210511803627, + 0.009244592860341072, + 0.016805564984679222, + 0.011087549850344658, + 0.01267600804567337, + 0.006947291549295187, + 0.008415603078901768, + 0.016188859939575195, + 0.010372564196586609, + 0.007815202698111534, + 0.01524288672953844, + 0.008459446020424366, + 0.0390811450779438, + 0.013426220044493675, + 0.04316050559282303, + 0.016371142119169235, + 0.009641229175031185, + 0.010737666860222816, + 0.018685121089220047, + 0.010075767524540424, + 0.020456332713365555, + 0.031127505004405975, + 0.016456518322229385, + 0.0230236928910017, + 0.035461340099573135, + 0.016188517212867737, + 0.02502061240375042, + 0.02777227573096752, + 0.014545612968504429, + 0.011315912008285522, + 0.009726428426802158, + 0.016897983849048615, + 0.011056432500481606, + 0.00856451690196991, + 0.014508300460875034, + 0.006997215561568737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -1.0538808107376099 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_15/index_tensor([7943, 7943])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_15/index_tensor([7943, 7943])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bb261a20cbcedd77e02b7cf3220d9c14d13e5dfd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_15/index_tensor([7943, 7943])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot meticulously depicts the dangling end of a thin, flexible, translucent plastic ruler, resting on the smooth, rounded edge of a flat wooden table with a rich, dark oak finish. Warm, soft light, as if from a nearby desk lamp, illuminates the scene from the side, capturing the subtle sheen of the polished texture and creating bright, reflective highlights along the ruler's length, revealing faint wear marks and precise black measurement markings. The movement begins when gravity deliberately pulls downward on the unsupported end, causing the plastic to gracefully bend into a smooth, continuous curve. Upon reaching its lowest point, the ruler's tip generates a brief, rapid vertical oscillation, quivering up and down as its energy wanes. Captured in hyper-detail, the movement shows the plastic vibrating before steadily and deliberately settling into its final, stable, drooping position, its elegantly curved form held in silent tension, creating a focused, almost meditative atmosphere centered on the simple physics of movement.", + "seed": 2336, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_15/index_7943_seed_2336_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07174707250669599, + "vlm_logprob": -0.07174707250669599, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.07174707250669599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015730109065771103 + ], + "opposite_logprob_tokens": [ + -5.390729904174805 + ], + "logprob_reward": -0.015730109065771103, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.015730109065771103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12161353975534439, + 0.04580369591712952, + 0.076805979013443, + 0.19819869101047516, + 0.06994282454252243, + 0.08623377233743668, + 0.12570324540138245, + 0.037590865045785904, + 0.11851389706134796, + 0.1195935606956482 + ], + "credit_map": [ + 0.019568858668208122, + 0.020550202578306198, + 0.007639170158654451, + 0.011956596747040749, + 0.01895628497004509, + 0.019536901265382767, + 0.01308518461883068, + 0.010320342145860195, + 0.004410122986882925, + 0.005465052090585232, + 0.00240328349173069, + 0.006990230642259121, + 0.00584644079208374, + 0.007624223828315735, + 0.009185483679175377, + 0.003878858871757984, + 0.007763921283185482, + 0.01054896879941225, + 0.005384098272770643, + 0.008127311244606972, + 0.009594139643013477, + 0.009261119179427624, + 0.014110531657934189, + 0.012015891261398792, + 0.01976739428937435, + 0.024945534765720367, + 0.012402019463479519, + 0.018884984776377678, + 0.029775463044643402, + 0.0316617377102375, + 0.030888061970472336, + 0.029873497784137726, + 0.005857258103787899, + 0.013942337594926357, + 0.004103674087673426, + 0.006233941297978163, + 0.009385316632688046, + 0.012272955849766731, + 0.011204089969396591, + 0.006943248677998781, + 0.005862152203917503, + 0.016611117869615555, + 0.0034382834564894438, + 0.005588218569755554, + 0.027310756966471672, + 0.0096515491604805, + 0.012173098511993885, + 0.005598590709269047, + 0.006670439150184393, + 0.033385198563337326, + 0.0019626787398010492, + 0.0028048239182680845, + 0.054426055401563644, + 0.01664215512573719, + 0.006136117968708277, + 0.0036757781635969877, + 0.0020849776919931173, + 0.0077774180099368095, + 0.0010050397831946611, + 0.0011648412328213453, + 0.013883336447179317, + 0.004834841936826706, + 0.00436303298920393, + 0.0024773776531219482, + 0.0062933717854321, + 0.03262227773666382, + 0.0022100519854575396, + 0.00349328201264143, + 0.04971941187977791, + 0.01379402820020914, + 0.006886743940412998, + 0.003494719974696636, + 0.006483603268861771, + 0.03183100372552872, + 0.0018990515964105725, + 0.003765691537410021, + 0.04999981075525284, + 0.014463094063103199, + 0.0062255049124360085, + 0.0049258009530603886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.015730109065771103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081226170063019 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081226170063019, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081226170063019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23110496997833252, + 0.03966359794139862, + 0.10543288290500641, + 0.058944221585989, + 0.21453844010829926, + 0.040848616510629654, + 0.0736471489071846, + 0.09150457382202148, + 0.09644979983568192, + 0.0478658452630043 + ], + "credit_map": [ + 0.04655012860894203, + 0.05367700010538101, + 0.014180559664964676, + 0.018860990181565285, + 0.026153041049838066, + 0.031061915680766106, + 0.023293286561965942, + 0.017328010872006416, + 0.00429740734398365, + 0.005374986212700605, + 0.002329405164346099, + 0.00589447095990181, + 0.005523239728063345, + 0.005422207526862621, + 0.006884870585054159, + 0.0039369999431073666, + 0.009336704388260841, + 0.012635698541998863, + 0.006924462039023638, + 0.010789865627884865, + 0.019120214506983757, + 0.01682497002184391, + 0.018163591623306274, + 0.011637363582849503, + 0.006509669125080109, + 0.008312585763633251, + 0.0031055337749421597, + 0.004610343836247921, + 0.009834232740104198, + 0.009101624600589275, + 0.01060243509709835, + 0.006867788266390562, + 0.01907457783818245, + 0.032655686140060425, + 0.010734712705016136, + 0.018754836171865463, + 0.027741843834519386, + 0.03625445440411568, + 0.040136002004146576, + 0.029186295345425606, + 0.0034780583810061216, + 0.007001214195042849, + 0.002220593858510256, + 0.002791027072817087, + 0.009242139756679535, + 0.006415109615772963, + 0.005752854980528355, + 0.0039476132951676846, + 0.004735979717224836, + 0.01690101996064186, + 0.0032809346448630095, + 0.004323948640376329, + 0.02352079376578331, + 0.008935432881116867, + 0.008304325863718987, + 0.0036447024904191494, + 0.006258554756641388, + 0.0196058452129364, + 0.0036245007067918777, + 0.004603894427418709, + 0.027108756825327873, + 0.014186951331794262, + 0.009695974178612232, + 0.006420087069272995, + 0.00511575024574995, + 0.02740456350147724, + 0.0024370267055928707, + 0.004454992711544037, + 0.037778958678245544, + 0.010285650379955769, + 0.0053162178955972195, + 0.0036566199269145727, + 0.0027727223932743073, + 0.011020851321518421, + 0.0018024976598098874, + 0.002523851115256548, + 0.017004163935780525, + 0.005996454507112503, + 0.0038731987588107586, + 0.0028720975387841463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.011081226170063019 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956872832030058 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956872832030058, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956872832030058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23579362034797668, + 0.1548294574022293, + 0.10465557128190994, + 0.18781347572803497, + 0.030998962000012398, + 0.04498646780848503, + 0.07865943014621735, + 0.023911049589514732, + 0.10314412415027618, + 0.035207800567150116 + ], + "credit_map": [ + 0.04741685464978218, + 0.04866607114672661, + 0.015137833543121815, + 0.01622800901532173, + 0.02537916786968708, + 0.0358683206140995, + 0.021715698763728142, + 0.02538166381418705, + 0.014153446070849895, + 0.021174076944589615, + 0.012227332219481468, + 0.028726894408464432, + 0.02558094821870327, + 0.019529927521944046, + 0.02271444723010063, + 0.010722375474870205, + 0.009997005574405193, + 0.025221513584256172, + 0.005491544492542744, + 0.011378804221749306, + 0.01249658688902855, + 0.016559790819883347, + 0.01188401598483324, + 0.011626309715211391, + 0.020774714648723602, + 0.02205800451338291, + 0.012182655744254589, + 0.017929520457983017, + 0.025318007916212082, + 0.03293032944202423, + 0.031218815594911575, + 0.025401419028639793, + 0.0027826970908790827, + 0.004937681369483471, + 0.002744338009506464, + 0.0033233603462576866, + 0.004768918734043837, + 0.0037446236237883568, + 0.005616231355816126, + 0.003081112401559949, + 0.0050173671916127205, + 0.006170235108584166, + 0.002158551476895809, + 0.004545058589428663, + 0.006731139030307531, + 0.00784390140324831, + 0.007511510979384184, + 0.005008702632039785, + 0.004816392436623573, + 0.020792223513126373, + 0.0023604014422744513, + 0.003918631933629513, + 0.024398932233452797, + 0.007463586982339621, + 0.008489319123327732, + 0.006419943179935217, + 0.0019519763300195336, + 0.0035037058405578136, + 0.0012927789939567447, + 0.0016174024203792214, + 0.0054390570148825645, + 0.0032516273204237223, + 0.00429152837023139, + 0.002562973415479064, + 0.005060776136815548, + 0.02741178683936596, + 0.0028430556412786245, + 0.005618559196591377, + 0.038989197462797165, + 0.010358125902712345, + 0.008763417601585388, + 0.004099206533282995, + 0.003332443768158555, + 0.005839336197823286, + 0.0016933578299358487, + 0.0024628827814012766, + 0.009364248253405094, + 0.005322040989995003, + 0.004306687507778406, + 0.0028868073131889105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005956872832030058 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14118584990501404, + 0.07349465042352676, + 0.03495008870959282, + 0.0716286450624466, + 0.0492430217564106, + 0.08130401372909546, + 0.12538112699985504, + 0.0512840673327446, + 0.19405914843082428, + 0.17746944725513458 + ], + "credit_map": [ + 0.01844155602157116, + 0.019985096529126167, + 0.008257155306637287, + 0.013554941862821579, + 0.023303603753447533, + 0.026854518800973892, + 0.016618231311440468, + 0.014170750975608826, + 0.008258851245045662, + 0.010430056601762772, + 0.0065468414686620235, + 0.0110913822427392, + 0.009365180507302284, + 0.010732106864452362, + 0.010970287024974823, + 0.006099941208958626, + 0.0034933844581246376, + 0.0053957197815179825, + 0.002216641791164875, + 0.004029229748994112, + 0.004470110405236483, + 0.004685306455940008, + 0.006081826984882355, + 0.004577871412038803, + 0.009040332399308681, + 0.008920524269342422, + 0.004353476222604513, + 0.006337231956422329, + 0.009648991748690605, + 0.012361434288322926, + 0.011649168096482754, + 0.009317482821643353, + 0.00407380610704422, + 0.008331673219799995, + 0.0026206406764686108, + 0.0048596905544400215, + 0.010102033615112305, + 0.0056447130627930164, + 0.008471292443573475, + 0.005139170214533806, + 0.005481922999024391, + 0.012605838477611542, + 0.004368647001683712, + 0.009033053182065487, + 0.017934182658791542, + 0.008431858383119106, + 0.01724233478307724, + 0.006206181365996599, + 0.005746954120695591, + 0.036011580377817154, + 0.002946769818663597, + 0.006183242425322533, + 0.04442733898758888, + 0.009511114098131657, + 0.014731920324265957, + 0.005822205450385809, + 0.0031131040304899216, + 0.009712195955216885, + 0.0018079696455970407, + 0.0035008906852453947, + 0.015534180216491222, + 0.0050260466523468494, + 0.009752840735018253, + 0.0028368360362946987, + 0.009446586482226849, + 0.04869215562939644, + 0.006221499294042587, + 0.012714847922325134, + 0.06649288535118103, + 0.017619628459215164, + 0.022034749388694763, + 0.010836807079613209, + 0.012103508226573467, + 0.04046017676591873, + 0.005415754858404398, + 0.007957397028803825, + 0.0648883581161499, + 0.0236502718180418, + 0.010969248600304127, + 0.012024733237922192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.012537590228021145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134295642375946 + ], + "opposite_logprob_tokens": [ + -1.313429594039917 + ], + "logprob_reward": -0.3134295642375946, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134295642375946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.147443488240242, + 0.08492568135261536, + 0.08514933288097382, + 0.41691407561302185, + 0.12581005692481995, + 0.030938202515244484, + 0.037963833659887314, + 0.01688026636838913, + 0.027709349989891052, + 0.026265639811754227 + ], + "credit_map": [ + 0.022380677983164787, + 0.027638040482997894, + 0.009580093435943127, + 0.009596717543900013, + 0.016279375180602074, + 0.0324292816221714, + 0.015105069614946842, + 0.014434241689741611, + 0.009175023064017296, + 0.01149030402302742, + 0.004690849222242832, + 0.00931862648576498, + 0.01711091957986355, + 0.01376311108469963, + 0.010763571597635746, + 0.008613275364041328, + 0.0072936718352139, + 0.017903069034218788, + 0.004793176893144846, + 0.010423635132610798, + 0.01323810126632452, + 0.011499556712806225, + 0.012916434556245804, + 0.007081683725118637, + 0.0463765524327755, + 0.04724744334816933, + 0.0237202737480402, + 0.039566438645124435, + 0.05482263118028641, + 0.0686323493719101, + 0.07124587893486023, + 0.0653025433421135, + 0.011255697347223759, + 0.02188735269010067, + 0.007519908249378204, + 0.01157169695943594, + 0.01725977659225464, + 0.021747274324297905, + 0.020984068512916565, + 0.013584278523921967, + 0.0026108541060239077, + 0.005514387972652912, + 0.0014316111337393522, + 0.0029156310483813286, + 0.006741601508110762, + 0.0033433123026043177, + 0.005410289857536554, + 0.002970515051856637, + 0.002319827675819397, + 0.011973929591476917, + 0.0009270274895243347, + 0.00201726914383471, + 0.011941553093492985, + 0.002572166034951806, + 0.0036004688590765, + 0.0026115926448255777, + 0.0012736456701532006, + 0.0018034911481663585, + 0.0011974755907431245, + 0.0016858145827427506, + 0.00270478380843997, + 0.0026203005108982325, + 0.0033931639045476913, + 0.002201590919867158, + 0.002387424698099494, + 0.004362227860838175, + 0.0017870682058855891, + 0.00289342412725091, + 0.00603740056976676, + 0.003443051129579544, + 0.0035875055473297834, + 0.003211249364539981, + 0.0021563745103776455, + 0.004134605638682842, + 0.0019326655892655253, + 0.00300382892601192, + 0.005295155569911003, + 0.002716898452490568, + 0.004373790230602026, + 0.0026523214764893055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.3134295642375946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot meticulously depicts the dangling end of a thin, flexible, translucent plastic ruler, resting on the smooth, rounded edge of a flat wooden table with a rich, dark oak finish. Warm, soft light, as if from a nearby desk lamp, illuminates the scene from the side, capturing the subtle sheen of the polished texture and creating bright, reflective highlights along the ruler's length, revealing faint wear marks and precise black measurement markings. The movement begins when gravity deliberately pulls downward on the unsupported end, causing the plastic to gracefully bend into a smooth, continuous curve. Upon reaching its lowest point, the ruler's tip generates a brief, rapid vertical oscillation, quivering up and down as its energy wanes. Captured in hyper-detail, the movement shows the plastic vibrating before steadily and deliberately settling into its final, stable, drooping position, its elegantly curved form held in silent tension, creating a focused, almost meditative atmosphere centered on the simple physics of movement.", + "seed": 2337, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_15/index_7943_seed_2337_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11375234434381128, + "vlm_logprob": -0.11375234434381128, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.11375234434381128, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016532452777028084 + ], + "opposite_logprob_tokens": [ + -5.0165324211120605 + ], + "logprob_reward": -0.016532452777028084, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016532452777028084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1378972828388214, + 0.08075765520334244, + 0.05264702066779137, + 0.15903595089912415, + 0.18928207457065582, + 0.12963096797466278, + 0.09376700222492218, + 0.06702500581741333, + 0.059094417840242386, + 0.030862584710121155 + ], + "credit_map": [ + 0.012514343485236168, + 0.03595731034874916, + 0.00913330726325512, + 0.012110920622944832, + 0.019362296909093857, + 0.01703464798629284, + 0.02036469429731369, + 0.011419779621064663, + 0.008113216608762741, + 0.022968750447034836, + 0.003388035809621215, + 0.004184772726148367, + 0.013569778762757778, + 0.01349250040948391, + 0.008333981968462467, + 0.0067066289484500885, + 0.00921633280813694, + 0.008312180638313293, + 0.00438664760440588, + 0.0038287590723484755, + 0.004039597697556019, + 0.008433898910880089, + 0.006137643475085497, + 0.00829196348786354, + 0.031863920390605927, + 0.049603089690208435, + 0.0075241620652377605, + 0.0062883696518838406, + 0.0168619267642498, + 0.02748158574104309, + 0.011070432141423225, + 0.008342474699020386, + 0.022498421370983124, + 0.03792651370167732, + 0.006720428355038166, + 0.009015466086566448, + 0.022768206894397736, + 0.04387097805738449, + 0.030975276604294777, + 0.015506808646023273, + 0.025711730122566223, + 0.02489108219742775, + 0.007664958480745554, + 0.006645000074058771, + 0.012477977201342583, + 0.02602279558777809, + 0.011725639924407005, + 0.014491801150143147, + 0.02151501178741455, + 0.024658868089318275, + 0.00595251377671957, + 0.005903535056859255, + 0.0056432513520121574, + 0.012263022363185883, + 0.008448013104498386, + 0.009382784366607666, + 0.014603717252612114, + 0.019361872225999832, + 0.003616360481828451, + 0.004541615955531597, + 0.006460077594965696, + 0.009349451400339603, + 0.004175607580691576, + 0.0049163103103637695, + 0.01305561326444149, + 0.01872788555920124, + 0.0033677625469863415, + 0.004075336270034313, + 0.0040109287947416306, + 0.0082487678155303, + 0.004756538663059473, + 0.0028515940066426992, + 0.0052241375669837, + 0.008883193135261536, + 0.0018784408457577229, + 0.002287093549966812, + 0.0025277752429246902, + 0.004495528992265463, + 0.003636966459453106, + 0.0019294500816613436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.016532452777028084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083701625466347 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083701625466347, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083701625466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16002973914146423, + 0.06209852173924446, + 0.07722001522779465, + 0.250650554895401, + 0.2703476846218109, + 0.04668142646551132, + 0.04945814609527588, + 0.049802087247371674, + 0.019204577431082726, + 0.014507228508591652 + ], + "credit_map": [ + 0.01617630198597908, + 0.034528546035289764, + 0.009764445014297962, + 0.014125066809356213, + 0.025064939633011818, + 0.023700566962361336, + 0.021430790424346924, + 0.015239078551530838, + 0.00773263443261385, + 0.013579350896179676, + 0.002995129209011793, + 0.003974527586251497, + 0.012390431016683578, + 0.008364250883460045, + 0.006912377662956715, + 0.0061498237773776054, + 0.010472004301846027, + 0.01667960174381733, + 0.00549687072634697, + 0.005719632375985384, + 0.010751630179584026, + 0.009932772256433964, + 0.01046836469322443, + 0.007699142675846815, + 0.04556551203131676, + 0.05102965608239174, + 0.01624882221221924, + 0.017137888818979263, + 0.022888625040650368, + 0.058801695704460144, + 0.020635541528463364, + 0.018342794850468636, + 0.040239982306957245, + 0.04995190352201462, + 0.007292957045137882, + 0.00826325360685587, + 0.034624624997377396, + 0.10730346292257309, + 0.01050126738846302, + 0.012170235626399517, + 0.007343254052102566, + 0.007019568234682083, + 0.002955271163955331, + 0.00378797366283834, + 0.003378798021003604, + 0.00848724041134119, + 0.006219054572284222, + 0.007490269374102354, + 0.010671539232134819, + 0.014744961634278297, + 0.0027443389408290386, + 0.002631265204399824, + 0.004404666367918253, + 0.007496090140193701, + 0.003831772832199931, + 0.0029335112776607275, + 0.010441376827657223, + 0.014500600285828114, + 0.00298806419596076, + 0.0030652182176709175, + 0.003741707419976592, + 0.006684782914817333, + 0.004259584937244654, + 0.0041207545436918736, + 0.003246156731620431, + 0.003865388222038746, + 0.0015377018135040998, + 0.0011907123262062669, + 0.0015829617623239756, + 0.0036425807047635317, + 0.002094426192343235, + 0.0020446497946977615, + 0.0026905874256044626, + 0.004006299190223217, + 0.0009882764425128698, + 0.000900922401342541, + 0.0012957563158124685, + 0.0022597359493374825, + 0.0013798829168081284, + 0.0009857683908194304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.011083701625466347 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264111328870058 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.005264111328870058, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264111328870058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12349624931812286, + 0.050016116350889206, + 0.06424778699874878, + 0.25310218334198, + 0.20989738404750824, + 0.10123343765735626, + 0.07799392938613892, + 0.06204830855131149, + 0.026646122336387634, + 0.03131846711039543 + ], + "credit_map": [ + 0.015698526054620743, + 0.02209020033478737, + 0.010342936031520367, + 0.010426288470625877, + 0.015105769969522953, + 0.015668364241719246, + 0.02083028107881546, + 0.013333886861801147, + 0.0057971603237092495, + 0.014761678874492645, + 0.0018175621517002583, + 0.0022752860095351934, + 0.007806805428117514, + 0.008291167207062244, + 0.0053745596669614315, + 0.003891899948939681, + 0.009230499155819416, + 0.010272960178554058, + 0.005524105858057737, + 0.006574949715286493, + 0.008064397610723972, + 0.01210161205381155, + 0.007294628769159317, + 0.005184633657336235, + 0.03455805033445358, + 0.06715710461139679, + 0.018434539437294006, + 0.019397523254156113, + 0.025168834254145622, + 0.05825881287455559, + 0.017334342002868652, + 0.01279296912252903, + 0.019666846841573715, + 0.03612278774380684, + 0.00705766212195158, + 0.008988969959318638, + 0.027190586552023888, + 0.0656338557600975, + 0.029335690662264824, + 0.01590101607143879, + 0.025811055675148964, + 0.028581760823726654, + 0.005764137953519821, + 0.0056914035230875015, + 0.006729037035256624, + 0.0136829549446702, + 0.006250691134482622, + 0.008722398430109024, + 0.01951056532561779, + 0.021963223814964294, + 0.0036936714313924313, + 0.0029634470120072365, + 0.006678937468677759, + 0.01717263273894787, + 0.0027678427286446095, + 0.003243619343265891, + 0.012853426858782768, + 0.02113046497106552, + 0.0034871213138103485, + 0.004061534535139799, + 0.004827405791729689, + 0.008089915849268436, + 0.004744161851704121, + 0.002854278078302741, + 0.0038776672445237637, + 0.00615012738853693, + 0.0018797590164467692, + 0.002361293649300933, + 0.0036849440075457096, + 0.0037968370597809553, + 0.002695157891139388, + 0.002200337126851082, + 0.006597820203751326, + 0.00842861458659172, + 0.001963381888344884, + 0.0018170442199334502, + 0.002637595171108842, + 0.005727039184421301, + 0.0019053204450756311, + 0.002241651527583599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.005264111328870058 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537237256765366 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537237256765366, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537237256765366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0710514634847641, + 0.045731522142887115, + 0.07586503028869629, + 0.13113893568515778, + 0.4464509189128876, + 0.12079436331987381, + 0.039778951555490494, + 0.03954659774899483, + 0.017606118693947792, + 0.01203610748052597 + ], + "credit_map": [ + 0.00815782230347395, + 0.01472428534179926, + 0.004501553252339363, + 0.0035080963280051947, + 0.005909965373575687, + 0.011620261706411839, + 0.013939105905592442, + 0.008690371178090572, + 0.004964545369148254, + 0.013371152803301811, + 0.002041955478489399, + 0.002940497128292918, + 0.007495115045458078, + 0.00684775086119771, + 0.004822760354727507, + 0.003247746266424656, + 0.008007013238966465, + 0.017357535660266876, + 0.004726221784949303, + 0.005596848204731941, + 0.008598259650170803, + 0.009520143270492554, + 0.013428361155092716, + 0.008630648255348206, + 0.010849547572433949, + 0.01832064986228943, + 0.006906978785991669, + 0.007109974976629019, + 0.033698875457048416, + 0.02007228508591652, + 0.025514518842101097, + 0.00866609439253807, + 0.056219302117824554, + 0.06825977563858032, + 0.011020464822649956, + 0.010044457390904427, + 0.05945836007595062, + 0.1705733984708786, + 0.043521832674741745, + 0.027353286743164062, + 0.02878335304558277, + 0.03764791414141655, + 0.006628690753132105, + 0.006561060901731253, + 0.00965138804167509, + 0.012435968033969402, + 0.008770573884248734, + 0.01031540147960186, + 0.0079047791659832, + 0.010027864016592503, + 0.0022841074969619513, + 0.002344275126233697, + 0.0029593377839773893, + 0.008689903654158115, + 0.0027183841448277235, + 0.00285029667429626, + 0.007064058445394039, + 0.009346210397779942, + 0.0022913413122296333, + 0.0025850317906588316, + 0.0032321468461304903, + 0.006993358489125967, + 0.0038200528360903263, + 0.004214402288198471, + 0.0019349148496985435, + 0.003916338551789522, + 0.0013174416963011026, + 0.001923020463436842, + 0.0021543216425925493, + 0.002063480205833912, + 0.002792823128402233, + 0.001503777108155191, + 0.0025410770904272795, + 0.0029692514799535275, + 0.0006798336398787796, + 0.0008462806581519544, + 0.0009593332652002573, + 0.0024350706953555346, + 0.0008894674247130752, + 0.0007157928193919361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.012537237256765366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233442187309265 + ], + "opposite_logprob_tokens": [ + -0.8983442187309265 + ], + "logprob_reward": -0.5233442187309265, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233442187309265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057716239243745804, + 0.05560792610049248, + 0.07704145461320877, + 0.1341804563999176, + 0.3478311598300934, + 0.10066517442464828, + 0.058328669518232346, + 0.061321038752794266, + 0.05452486500144005, + 0.05278301611542702 + ], + "credit_map": [ + 0.008745463564991951, + 0.009490050375461578, + 0.004909890703856945, + 0.005684175528585911, + 0.005986568983644247, + 0.008060721680521965, + 0.008441807702183723, + 0.006397562101483345, + 0.004148148000240326, + 0.014860667288303375, + 0.0021786673460155725, + 0.003495360491797328, + 0.009758168831467628, + 0.009222648106515408, + 0.007915688678622246, + 0.004028577357530594, + 0.012890868820250034, + 0.0136032709851861, + 0.005328713450580835, + 0.005447993520647287, + 0.007560743018984795, + 0.019413482397794724, + 0.007115654647350311, + 0.005680724512785673, + 0.018312903121113777, + 0.029198095202445984, + 0.011289124377071857, + 0.012987284921109676, + 0.011244482360780239, + 0.02040540985763073, + 0.019569743424654007, + 0.011173422448337078, + 0.03850514814257622, + 0.06648709625005722, + 0.008793446235358715, + 0.013309258967638016, + 0.04381771758198738, + 0.12765441834926605, + 0.030692337080836296, + 0.01857174001634121, + 0.016416363418102264, + 0.019413208588957787, + 0.00835704617202282, + 0.008826100267469883, + 0.011130833066999912, + 0.013998561538755894, + 0.012111796997487545, + 0.010411256924271584, + 0.008821224793791771, + 0.012380308471620083, + 0.004467148333787918, + 0.004064259119331837, + 0.007609084714204073, + 0.007713360246270895, + 0.005697084590792656, + 0.007576201111078262, + 0.012491131201386452, + 0.014705653302371502, + 0.004541612695902586, + 0.0031451266258955, + 0.00341083318926394, + 0.008075525984168053, + 0.00558463716879487, + 0.009366519749164581, + 0.008040306158363819, + 0.011366313323378563, + 0.003273547627031803, + 0.003980359528213739, + 0.005174824967980385, + 0.009727232158184052, + 0.007705131080001593, + 0.005257151555269957, + 0.011653508059680462, + 0.0117220189422369, + 0.0031107948161661625, + 0.0036681329365819693, + 0.004947836976498365, + 0.01072354894131422, + 0.004028041381388903, + 0.0029291303362697363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.5233442187309265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot meticulously depicts the dangling end of a thin, flexible, translucent plastic ruler, resting on the smooth, rounded edge of a flat wooden table with a rich, dark oak finish. Warm, soft light, as if from a nearby desk lamp, illuminates the scene from the side, capturing the subtle sheen of the polished texture and creating bright, reflective highlights along the ruler's length, revealing faint wear marks and precise black measurement markings. The movement begins when gravity deliberately pulls downward on the unsupported end, causing the plastic to gracefully bend into a smooth, continuous curve. Upon reaching its lowest point, the ruler's tip generates a brief, rapid vertical oscillation, quivering up and down as its energy wanes. Captured in hyper-detail, the movement shows the plastic vibrating before steadily and deliberately settling into its final, stable, drooping position, its elegantly curved form held in silent tension, creating a focused, almost meditative atmosphere centered on the simple physics of movement.", + "seed": 2338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_15/index_7943_seed_2338_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.179877657443285, + "vlm_logprob": -0.179877657443285, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.179877657443285, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015415798872709274 + ], + "opposite_logprob_tokens": [ + -4.765415668487549 + ], + "logprob_reward": -0.015415798872709274, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015415798872709274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2726031243801117, + 0.06139356270432472, + 0.07039130479097366, + 0.08080240339040756, + 0.18056578934192657, + 0.1342213898897171, + 0.10308302193880081, + 0.04159523919224739, + 0.03973155468702316, + 0.015612668357789516 + ], + "credit_map": [ + 0.027225298807024956, + 0.0431518517434597, + 0.011779862456023693, + 0.012124880217015743, + 0.0874287560582161, + 0.03763362020254135, + 0.034020159393548965, + 0.019238688051700592, + 0.005458033178001642, + 0.008076823316514492, + 0.0041233631782233715, + 0.005326501559466124, + 0.009596091695129871, + 0.009998682886362076, + 0.011780256405472755, + 0.0070338076911866665, + 0.007598265539854765, + 0.014892354607582092, + 0.003194210585206747, + 0.0035831788554787636, + 0.009171898476779461, + 0.018610067665576935, + 0.0047981091775000095, + 0.008543216623365879, + 0.008902888745069504, + 0.01419882383197546, + 0.006197564769536257, + 0.008168105036020279, + 0.01051894761621952, + 0.01659199595451355, + 0.007589228451251984, + 0.008634842000901699, + 0.021050140261650085, + 0.05265384167432785, + 0.005970832891762257, + 0.008242554031312466, + 0.02199297770857811, + 0.04732346907258034, + 0.007812215946614742, + 0.015519747510552406, + 0.010216963477432728, + 0.012793650850653648, + 0.013682296499609947, + 0.015994271263480186, + 0.019036000594496727, + 0.01722016930580139, + 0.023596040904521942, + 0.02168198861181736, + 0.013835086487233639, + 0.01783435419201851, + 0.007644436322152615, + 0.009525596164166927, + 0.01741880550980568, + 0.011570869944989681, + 0.016728533431887627, + 0.00852533895522356, + 0.006052853539586067, + 0.005702937953174114, + 0.0030555720441043377, + 0.0038172395434230566, + 0.00587974488735199, + 0.00655616307631135, + 0.005770734511315823, + 0.004759991075843573, + 0.004942439496517181, + 0.007174528203904629, + 0.0037475775461643934, + 0.0026542525738477707, + 0.006084496155381203, + 0.005276539362967014, + 0.005129251629114151, + 0.004722472745925188, + 0.0017150914063677192, + 0.0019054862204939127, + 0.0015555564314126968, + 0.0011362945660948753, + 0.0027598701417446136, + 0.0019281718414276838, + 0.001911426312290132, + 0.0027007709722965956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.015415798872709274 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813595116138458 + ], + "opposite_logprob_tokens": [ + -3.529813528060913 + ], + "logprob_reward": -0.029813595116138458, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813595116138458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18875326216220856, + 0.053886543959379196, + 0.028424803167581558, + 0.04433425888419151, + 0.060718316584825516, + 0.07195016741752625, + 0.17438754439353943, + 0.1654054820537567, + 0.04937770962715149, + 0.16276191174983978 + ], + "credit_map": [ + 0.015111712738871574, + 0.027337921783328056, + 0.013628094457089901, + 0.01897694729268551, + 0.04010891541838646, + 0.023257609456777573, + 0.022135311737656593, + 0.02819673717021942, + 0.005825086496770382, + 0.006431774236261845, + 0.004701552912592888, + 0.004448786377906799, + 0.008244529366493225, + 0.0067289755679667, + 0.01041316706687212, + 0.007092666812241077, + 0.0025131222791969776, + 0.003786521265283227, + 0.002574552781879902, + 0.00292263925075531, + 0.0065560657531023026, + 0.003973882179707289, + 0.003499662270769477, + 0.0025983548257499933, + 0.0066018919460475445, + 0.00855737179517746, + 0.0023933572228997946, + 0.002624030690640211, + 0.005883473437279463, + 0.00953756459057331, + 0.004299344029277563, + 0.00443722540512681, + 0.004852249287068844, + 0.013104372657835484, + 0.003861095756292343, + 0.004486424848437309, + 0.007083060219883919, + 0.014336914755403996, + 0.005937650799751282, + 0.007056543603539467, + 0.01112573966383934, + 0.008297247812151909, + 0.004673705901950598, + 0.004895756021142006, + 0.012096493504941463, + 0.00963561050593853, + 0.009698919951915741, + 0.011526689864695072, + 0.017118871212005615, + 0.047915179282426834, + 0.004882398992776871, + 0.01032306905835867, + 0.014036936685442924, + 0.0509980209171772, + 0.010668398812413216, + 0.018444670364260674, + 0.015335182659327984, + 0.020772580057382584, + 0.013627215288579464, + 0.01340267714112997, + 0.02696545608341694, + 0.03021644614636898, + 0.017316915094852448, + 0.027768990024924278, + 0.006004896480590105, + 0.009212388657033443, + 0.004329324699938297, + 0.004624438937753439, + 0.008137005381286144, + 0.005658342037349939, + 0.005911553744226694, + 0.005499755498021841, + 0.013493661768734455, + 0.022994989529252052, + 0.012745747342705727, + 0.013066170737147331, + 0.0287795253098011, + 0.026756979525089264, + 0.01933937519788742, + 0.02558545209467411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.029813595116138458 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979415699839592 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.00979415699839592, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979415699839592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2649339735507965, + 0.06391824036836624, + 0.09479421377182007, + 0.06528172641992569, + 0.2046077698469162, + 0.1140110120177269, + 0.06025247275829315, + 0.0602860301733017, + 0.03658655658364296, + 0.035328011959791183 + ], + "credit_map": [ + 0.020843440666794777, + 0.037357933819293976, + 0.012132002972066402, + 0.01436013262718916, + 0.07077912986278534, + 0.04509634152054787, + 0.03892456367611885, + 0.02544042095541954, + 0.008214833214879036, + 0.012522057630121708, + 0.005411028396338224, + 0.004910435061901808, + 0.009275832213461399, + 0.007792674470692873, + 0.008024794980883598, + 0.0077665806747972965, + 0.009568298235535622, + 0.012048294767737389, + 0.007321711163967848, + 0.007800158113241196, + 0.021588774397969246, + 0.012692230753600597, + 0.012400618754327297, + 0.011374133639037609, + 0.004714679904282093, + 0.011178932152688503, + 0.004225504118949175, + 0.005032738670706749, + 0.01342182420194149, + 0.012664616107940674, + 0.007617811672389507, + 0.0064256158657372, + 0.02363036386668682, + 0.060189343988895416, + 0.006218285765498877, + 0.006797146052122116, + 0.019162531942129135, + 0.05975373089313507, + 0.00803422462195158, + 0.020822152495384216, + 0.012121433392167091, + 0.010412157513201237, + 0.010509250685572624, + 0.01211897935718298, + 0.01675000973045826, + 0.014529649168252945, + 0.021450798958539963, + 0.01611873134970665, + 0.007606200408190489, + 0.014426963403820992, + 0.0028715659864246845, + 0.003582801204174757, + 0.006306149531155825, + 0.014638497494161129, + 0.0052171992138028145, + 0.0056030950509011745, + 0.006045893765985966, + 0.006724658887833357, + 0.00483705522492528, + 0.006319215055555105, + 0.009960463270545006, + 0.009756465442478657, + 0.009156263433396816, + 0.007486013695597649, + 0.0027948543429374695, + 0.006245820317417383, + 0.0027763794641941786, + 0.0018546010833233595, + 0.006922152359038591, + 0.006178658455610275, + 0.004609897267073393, + 0.0052041951566934586, + 0.0033851228654384613, + 0.004289648029953241, + 0.003258855314925313, + 0.0031430150847882032, + 0.004884618800133467, + 0.005908408667892218, + 0.003983288072049618, + 0.006475056987255812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.00979415699839592 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186602741479874 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186602741479874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186602741479874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18028056621551514, + 0.049371007829904556, + 0.06678775697946548, + 0.06431299448013306, + 0.3613823354244232, + 0.1429252028465271, + 0.04912011697888374, + 0.031482428312301636, + 0.04279225319623947, + 0.01154534425586462 + ], + "credit_map": [ + 0.019068747758865356, + 0.029450273141264915, + 0.006789382081478834, + 0.007466055452823639, + 0.055176276713609695, + 0.027762796729803085, + 0.02076226845383644, + 0.013804767280817032, + 0.0060897753573954105, + 0.006515996530652046, + 0.003301037475466728, + 0.0035379326436668634, + 0.008283082395792007, + 0.0065664928406476974, + 0.009175145998597145, + 0.0059015462175011635, + 0.0062097725458443165, + 0.011355515569448471, + 0.004307335242629051, + 0.006146696861833334, + 0.014109957031905651, + 0.009048191830515862, + 0.009478164836764336, + 0.00613212026655674, + 0.009013636969029903, + 0.012317240238189697, + 0.004130587913095951, + 0.004538008477538824, + 0.0071345241740345955, + 0.015610085800290108, + 0.00497759273275733, + 0.0065913209691643715, + 0.04260280355811119, + 0.1113857701420784, + 0.00993858277797699, + 0.012751324102282524, + 0.03274379298090935, + 0.10489001870155334, + 0.012467396445572376, + 0.03460264950990677, + 0.013130051083862782, + 0.013520424254238605, + 0.010313684120774269, + 0.01222993154078722, + 0.025175895541906357, + 0.019563140347599983, + 0.024327490478754044, + 0.024664584547281265, + 0.005781438667327166, + 0.009804879315197468, + 0.002733775647357106, + 0.0039284187369048595, + 0.00576607882976532, + 0.009213189594447613, + 0.006519854534417391, + 0.005372480023652315, + 0.0036469907499849796, + 0.004293621983379126, + 0.003280546050518751, + 0.002957117510959506, + 0.005013312213122845, + 0.004368605557829142, + 0.003643411910161376, + 0.004278823267668486, + 0.003964378498494625, + 0.00512014189735055, + 0.004051325377076864, + 0.0034290822222828865, + 0.008217115886509418, + 0.006905543617904186, + 0.004612946417182684, + 0.006491720676422119, + 0.002040623454377055, + 0.0016380434390157461, + 0.001158104743808508, + 0.0006679593352600932, + 0.0014292934210970998, + 0.0018587929662317038, + 0.0012204557424411178, + 0.0015320704551413655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.018186602741479874 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261781334877014 + ], + "opposite_logprob_tokens": [ + -0.5761781334877014 + ], + "logprob_reward": -0.8261781334877014, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261781334877014, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11249960958957672, + 0.057511813938617706, + 0.055552754551172256, + 0.04605139046907425, + 0.2183084785938263, + 0.0649140477180481, + 0.1183985024690628, + 0.09850487112998962, + 0.12975861132144928, + 0.09849993884563446 + ], + "credit_map": [ + 0.011251016519963741, + 0.016980474814772606, + 0.004823159892112017, + 0.004936908837407827, + 0.029468433931469917, + 0.018682651221752167, + 0.015818912535905838, + 0.010538043454289436, + 0.005066473502665758, + 0.007727485615760088, + 0.005053520668298006, + 0.004009889904409647, + 0.007615937851369381, + 0.01389000192284584, + 0.006423498969525099, + 0.007725002244114876, + 0.007159809116274118, + 0.008661270141601562, + 0.006245041266083717, + 0.006411299575120211, + 0.008061185479164124, + 0.006853772327303886, + 0.006231074221432209, + 0.005929299630224705, + 0.00486013013869524, + 0.005769066512584686, + 0.005015057045966387, + 0.005041926633566618, + 0.007205849513411522, + 0.00625113258138299, + 0.006914290599524975, + 0.004993931856006384, + 0.02196640893816948, + 0.04887055233120918, + 0.010655655525624752, + 0.015618180856108665, + 0.0432237833738327, + 0.04260803014039993, + 0.018837863579392433, + 0.016527997329831123, + 0.005060173571109772, + 0.008782558143138885, + 0.004603412002325058, + 0.0055261775851249695, + 0.009206932038068771, + 0.010322144255042076, + 0.012896533124148846, + 0.00851611141115427, + 0.01302425842732191, + 0.030232438817620277, + 0.004377051722258329, + 0.007497872691601515, + 0.009633314795792103, + 0.03390955924987793, + 0.0070991525426507, + 0.012624848634004593, + 0.009625553153455257, + 0.010933014564216137, + 0.011539660394191742, + 0.010738599114120007, + 0.013838451355695724, + 0.015138505026698112, + 0.011960823088884354, + 0.014730258844792843, + 0.01666932925581932, + 0.015994014218449593, + 0.014604136347770691, + 0.012712366878986359, + 0.018736397847533226, + 0.017081238329410553, + 0.018881654366850853, + 0.015079475939273834, + 0.010824691504240036, + 0.012853274121880531, + 0.009613072499632835, + 0.005752079654484987, + 0.019086482003331184, + 0.017266716808080673, + 0.01364465244114399, + 0.009458960965275764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.8261781334877014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, static shot meticulously depicts the dangling end of a thin, flexible, translucent plastic ruler, resting on the smooth, rounded edge of a flat wooden table with a rich, dark oak finish. Warm, soft light, as if from a nearby desk lamp, illuminates the scene from the side, capturing the subtle sheen of the polished texture and creating bright, reflective highlights along the ruler's length, revealing faint wear marks and precise black measurement markings. The movement begins when gravity deliberately pulls downward on the unsupported end, causing the plastic to gracefully bend into a smooth, continuous curve. Upon reaching its lowest point, the ruler's tip generates a brief, rapid vertical oscillation, quivering up and down as its energy wanes. Captured in hyper-detail, the movement shows the plastic vibrating before steadily and deliberately settling into its final, stable, drooping position, its elegantly curved form held in silent tension, creating a focused, almost meditative atmosphere centered on the simple physics of movement.", + "seed": 2339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_15/index_7943_seed_2339_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08296067975461482, + "vlm_logprob": -0.08296067975461482, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.08296067975461482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02673833630979061 + ], + "opposite_logprob_tokens": [ + -4.401738166809082 + ], + "logprob_reward": -0.02673833630979061, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02673833630979061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20712395012378693, + 0.03423566743731499, + 0.03139602765440941, + 0.047465600073337555, + 0.044567007571458817, + 0.03801557049155235, + 0.07015497237443924, + 0.14590312540531158, + 0.32762032747268677, + 0.05351778864860535 + ], + "credit_map": [ + 0.018630068749189377, + 0.025391828268766403, + 0.01760559156537056, + 0.02298405021429062, + 0.03429776802659035, + 0.028231695294380188, + 0.024888575077056885, + 0.035094380378723145, + 0.0035862319637089968, + 0.0033564032055437565, + 0.002805715426802635, + 0.004714593291282654, + 0.004357645753771067, + 0.004706256557255983, + 0.005817451048642397, + 0.004891370423138142, + 0.0025355301331728697, + 0.0038072552997618914, + 0.0024213329888880253, + 0.0043936497531831264, + 0.005107915494590998, + 0.0037490162067115307, + 0.005029079504311085, + 0.004352246876806021, + 0.004428865388035774, + 0.004250768572092056, + 0.004940493497997522, + 0.0070614079013466835, + 0.005786008667200804, + 0.005879712291061878, + 0.00894369836896658, + 0.006174643989652395, + 0.003037135349586606, + 0.006683239247649908, + 0.003753752913326025, + 0.006179849151521921, + 0.0054615577682852745, + 0.0063631911762058735, + 0.0055442084558308125, + 0.007544071413576603, + 0.0032872026786208153, + 0.005627936217933893, + 0.0023204658646136522, + 0.0033383083064109087, + 0.007873216643929482, + 0.005717871710658073, + 0.004337435122579336, + 0.005513132084161043, + 0.00526574207469821, + 0.017554622143507004, + 0.0049540940672159195, + 0.007466492708772421, + 0.011246826499700546, + 0.009240515530109406, + 0.007841709069907665, + 0.006584971211850643, + 0.01044555101543665, + 0.012860932387411594, + 0.009314561262726784, + 0.01768917217850685, + 0.021281156688928604, + 0.020322158932685852, + 0.03552000969648361, + 0.01846957951784134, + 0.019409403204917908, + 0.06587445735931396, + 0.015881352126598358, + 0.04169730842113495, + 0.052788492292165756, + 0.041920069605112076, + 0.05752311274409294, + 0.032526105642318726, + 0.004812977742403746, + 0.007116473745554686, + 0.0043482668697834015, + 0.0054773446172475815, + 0.006929248571395874, + 0.008677794598042965, + 0.007377798203378916, + 0.008777876384556293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.02673833630979061 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058931440114975 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058931440114975, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058931440114975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22289881110191345, + 0.10024581104516983, + 0.08200915902853012, + 0.07856155186891556, + 0.1535947620868683, + 0.08027350157499313, + 0.034762904047966, + 0.10313542187213898, + 0.09459774941205978, + 0.049920324236154556 + ], + "credit_map": [ + 0.02311532199382782, + 0.028297631070017815, + 0.016090258955955505, + 0.0198291577398777, + 0.01870894990861416, + 0.03205399960279465, + 0.02652752958238125, + 0.05827596038579941, + 0.010196306742727757, + 0.011996000073850155, + 0.011495650745928288, + 0.012531602755188942, + 0.012224486097693443, + 0.013433486223220825, + 0.01538101676851511, + 0.012987256050109863, + 0.008633522316813469, + 0.009812487289309502, + 0.0060579488053917885, + 0.010180705226957798, + 0.013410341925919056, + 0.009695310145616531, + 0.012440696358680725, + 0.011778152547776699, + 0.007243067026138306, + 0.00850636139512062, + 0.007170418743044138, + 0.011003625579178333, + 0.011193622834980488, + 0.008911852724850178, + 0.0157039612531662, + 0.008828643709421158, + 0.0158435869961977, + 0.030406726524233818, + 0.01507620606571436, + 0.01886431686580181, + 0.022162465378642082, + 0.019616393372416496, + 0.013629050925374031, + 0.01799602061510086, + 0.00766847375780344, + 0.014400395564734936, + 0.00850636512041092, + 0.00908904429525137, + 0.008230103179812431, + 0.010468713007867336, + 0.009038939140737057, + 0.012871469371020794, + 0.003553706919774413, + 0.0032744635827839375, + 0.003334776498377323, + 0.0044060200452804565, + 0.006265711504966021, + 0.004337170626968145, + 0.004301100969314575, + 0.005289955530315638, + 0.00813873391598463, + 0.013497209176421165, + 0.009950626641511917, + 0.013095584698021412, + 0.008334308862686157, + 0.015115640126168728, + 0.015458972193300724, + 0.019544342532753944, + 0.013000629842281342, + 0.008398925885558128, + 0.007201310247182846, + 0.010664500296115875, + 0.00925538968294859, + 0.01244631502777338, + 0.014034712687134743, + 0.01959596388041973, + 0.0038620850536972284, + 0.0051619564183056355, + 0.0028675347566604614, + 0.006618469953536987, + 0.0076790545135736465, + 0.006531495600938797, + 0.00802637916058302, + 0.009173346683382988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.02058931440114975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647525027394295 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647525027394295, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647525027394295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20522651076316833, + 0.11584398150444031, + 0.06463105231523514, + 0.03999873995780945, + 0.0765729546546936, + 0.04298083111643791, + 0.08740519732236862, + 0.15046516060829163, + 0.14553774893283844, + 0.07133781164884567 + ], + "credit_map": [ + 0.025079092010855675, + 0.018134713172912598, + 0.014714314602315426, + 0.028118059039115906, + 0.022003984078764915, + 0.0279386006295681, + 0.03429531306028366, + 0.03494244068861008, + 0.011047235690057278, + 0.013333026319742203, + 0.006382911931723356, + 0.012656217440962791, + 0.016604630276560783, + 0.014906208962202072, + 0.021778808906674385, + 0.019134940579533577, + 0.005511458031833172, + 0.008430557325482368, + 0.004486540332436562, + 0.00909940805286169, + 0.009630079381167889, + 0.00810190662741661, + 0.009465192444622517, + 0.009905911982059479, + 0.003332445165142417, + 0.005333305802196264, + 0.0028612655587494373, + 0.006152864079922438, + 0.00542871979996562, + 0.0049461983144283295, + 0.00701987836509943, + 0.004924063105136156, + 0.007171926088631153, + 0.008228009566664696, + 0.006076105870306492, + 0.009005130268633366, + 0.007265318650752306, + 0.012301071546971798, + 0.005379968788474798, + 0.021145423874258995, + 0.003698108484968543, + 0.004927525762468576, + 0.003984772600233555, + 0.0042475126683712006, + 0.006164520978927612, + 0.005729652475565672, + 0.006342852953821421, + 0.007885883562266827, + 0.007497913204133511, + 0.010850726626813412, + 0.006803098134696484, + 0.009949772618710995, + 0.009105658158659935, + 0.011652832850813866, + 0.015429182909429073, + 0.016116010025143623, + 0.013323890045285225, + 0.01318054273724556, + 0.006093721836805344, + 0.015102830715477467, + 0.017644694074988365, + 0.02489556558430195, + 0.024990195408463478, + 0.03523371368646622, + 0.009728376753628254, + 0.025377394631505013, + 0.0067507377825677395, + 0.020608501508831978, + 0.023195305839180946, + 0.01675071194767952, + 0.02861378900706768, + 0.014512924477458, + 0.0074030025862157345, + 0.007397208828479052, + 0.005615966860204935, + 0.009786938317120075, + 0.008523669093847275, + 0.009231924079358578, + 0.009815561585128307, + 0.013563540764153004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.004647525027394295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20206591486930847, + 0.07895742356777191, + 0.10211178660392761, + 0.050289884209632874, + 0.08468230813741684, + 0.03806569427251816, + 0.07623465359210968, + 0.12063731253147125, + 0.13472552597522736, + 0.11222948879003525 + ], + "credit_map": [ + 0.016274765133857727, + 0.02852020412683487, + 0.013647468760609627, + 0.021784616634249687, + 0.042410627007484436, + 0.02951131761074066, + 0.02259499579668045, + 0.027321917936205864, + 0.006776349153369665, + 0.00863894447684288, + 0.005504138767719269, + 0.009649575687944889, + 0.014183254912495613, + 0.010183225385844707, + 0.012470229528844357, + 0.01155171450227499, + 0.008915126323699951, + 0.0134512884542346, + 0.00667734956368804, + 0.013867039233446121, + 0.017870765179395676, + 0.013227752409875393, + 0.014128336682915688, + 0.013974124565720558, + 0.0037383874878287315, + 0.0068095833994448185, + 0.00464426726102829, + 0.007420214358717203, + 0.00787822064012289, + 0.00624719588086009, + 0.008296309970319271, + 0.005255705676972866, + 0.006510929204523563, + 0.009107780642807484, + 0.005507977679371834, + 0.009055027738213539, + 0.012938931584358215, + 0.012798492796719074, + 0.019705118611454964, + 0.009058047086000443, + 0.004034781828522682, + 0.008576116524636745, + 0.00299341743811965, + 0.0044171069748699665, + 0.004548761062324047, + 0.005735321901738644, + 0.004375055897980928, + 0.0033851303160190582, + 0.005355871748179197, + 0.01527718361467123, + 0.004883109126240015, + 0.00852574035525322, + 0.010455996729433537, + 0.00911868829280138, + 0.012555446475744247, + 0.010062613524496555, + 0.011395939625799656, + 0.011210770346224308, + 0.009414467960596085, + 0.013774690218269825, + 0.015931587666273117, + 0.018728846684098244, + 0.02632034383714199, + 0.013860664330422878, + 0.010020535439252853, + 0.01693226583302021, + 0.006902242079377174, + 0.016598839312791824, + 0.01879083923995495, + 0.016555119305849075, + 0.02801010198891163, + 0.02091558463871479, + 0.008517193607985973, + 0.01552501879632473, + 0.006347839254885912, + 0.013868738897144794, + 0.01715230755507946, + 0.015507538802921772, + 0.017638275399804115, + 0.017672579735517502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014195887371897697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a thin, flexible, translucent plastic ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ruler resting on the smooth, rounded edge of a flat wooden table with a dark oak finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm, soft side lighting that creates bright reflective highlights and reveals black measurement markings on the ruler?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ruler bend into a smooth curve, quiver vertically at its lowest point, and then settle into a stable drooping position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486323356628418 + ], + "opposite_logprob_tokens": [ + -1.2236323356628418 + ], + "logprob_reward": -0.3486323356628418, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486323356628418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09999105334281921, + 0.024367056787014008, + 0.08894070237874985, + 0.08265211433172226, + 0.06869219243526459, + 0.06986287981271744, + 0.07204873859882355, + 0.062196191400289536, + 0.1645345538854599, + 0.2667144238948822 + ], + "credit_map": [ + 0.009713335894048214, + 0.013178514316678047, + 0.010391028597950935, + 0.010124455206096172, + 0.01936309225857258, + 0.013798722065985203, + 0.012560704723000526, + 0.010861208662390709, + 0.0022688736207783222, + 0.0031785573810338974, + 0.0027134879492223263, + 0.003153834491968155, + 0.0026755398139357567, + 0.0035520256496965885, + 0.003966628108173609, + 0.0028581121005117893, + 0.00759859848767519, + 0.011177693493664265, + 0.006207476370036602, + 0.012096603401005268, + 0.014882278628647327, + 0.011222492903470993, + 0.01247214525938034, + 0.013283425010740757, + 0.006347288843244314, + 0.011738880537450314, + 0.0060666399076581, + 0.012667425908148289, + 0.011186148971319199, + 0.011135471053421497, + 0.010570226237177849, + 0.012940042652189732, + 0.006571485660970211, + 0.009787931106984615, + 0.004879094194620848, + 0.0075768595561385155, + 0.00965591799467802, + 0.009147786535322666, + 0.00884446781128645, + 0.012228649109601974, + 0.008225769735872746, + 0.013923488557338715, + 0.0049317567609250546, + 0.007409426383674145, + 0.008521662093698978, + 0.009542197920382023, + 0.009140191599726677, + 0.008168394677340984, + 0.008031384088099003, + 0.008816435933113098, + 0.005726027302443981, + 0.007798876613378525, + 0.007259793113917112, + 0.00985120888799429, + 0.011059362441301346, + 0.013505665585398674, + 0.006337513215839863, + 0.010031133890151978, + 0.005363359581679106, + 0.007143748924136162, + 0.007724988739937544, + 0.00815710797905922, + 0.010549012571573257, + 0.006889329291880131, + 0.011634600348770618, + 0.023836683481931686, + 0.010371960699558258, + 0.01886008493602276, + 0.028264803811907768, + 0.01759590022265911, + 0.035631489008665085, + 0.018339062109589577, + 0.02404550090432167, + 0.0415479838848114, + 0.018217122182250023, + 0.036938928067684174, + 0.042335137724876404, + 0.04457559436559677, + 0.03214602917432785, + 0.026908142492175102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.3486323356628418 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_tensor([8949, 8949])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_tensor([8949, 8949])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e726639fbd889e04e23dcba87efaf20a9aab2afd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_tensor([8949, 8949])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, slow-motion, tracking shot captures a bizarre fashion show taking place in a surreal hotel. Amidst a maze of intense beams of light, alternating crimson, royal blue, and eerie green, several tall, androgynous models slowly march along an endlessly long corridor with a vaulted ceiling. Clad in avant-garde, dramatically structured garments in vibrant, saturated colors, some embellished with shimmering metal, others crafted from translucent tulle, they move steadily forward in a ritualistic, almost hypnotic rhythm, their faces expressionless, their eyes vacant and piercing. As they move, light streaming through the gaps between the ceiling and floor dramatically elongates and distorts their figures, casting sharp, dancing shadows across the smooth, obsidian floor. The walls of the corridor are made up of countless mirrors of different sizes and angles, which continuously reflect the fragmented and overlapping images of the models, as if countless clones are moving simultaneously in different dimensions. The whole scene is full of Jodorowsky-style dramatic tension and visual symbolism, creating a dream atmosphere that is both gorgeous and unsettling.", + "seed": 21966, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_8949_seed_21966_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04694093046709895, + "vlm_logprob": -0.04694093046709895, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04694093046709895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10176757723093033 + ], + "opposite_logprob_tokens": [ + -2.6017675399780273 + ], + "logprob_reward": -0.10176757723093033, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10176757723093033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09067786484956741, + 0.22762033343315125, + 0.05996491014957428, + 0.05472995713353157, + 0.12695980072021484, + 0.2629736065864563, + 0.027946528047323227, + 0.06883198022842407, + 0.03111608698964119, + 0.04917898774147034 + ], + "credit_map": [ + 0.014129064045846462, + 0.010365759022533894, + 0.010993191041052341, + 0.011212634854018688, + 0.008991261944174767, + 0.009081335738301277, + 0.010463415645062923, + 0.015441196970641613, + 0.01731727086007595, + 0.025251397863030434, + 0.03318174183368683, + 0.041768912225961685, + 0.019655754789710045, + 0.023084362968802452, + 0.042467307299375534, + 0.02489357441663742, + 0.00718591408804059, + 0.006017947103828192, + 0.007899824529886246, + 0.01205681636929512, + 0.005903442855924368, + 0.00674414774402976, + 0.006663713604211807, + 0.007493098732084036, + 0.007365959230810404, + 0.006416828837245703, + 0.006145227234810591, + 0.0062067280523478985, + 0.009592804126441479, + 0.006329372059553862, + 0.005571914836764336, + 0.007101121824234724, + 0.021369725465774536, + 0.01538038719445467, + 0.011541938409209251, + 0.016581503674387932, + 0.015024371445178986, + 0.01664641499519348, + 0.014445051550865173, + 0.015970401465892792, + 0.020974021404981613, + 0.03795412555336952, + 0.019537683576345444, + 0.037452612072229385, + 0.03714218735694885, + 0.03662826865911484, + 0.03837236016988754, + 0.03491232916712761, + 0.002930566668510437, + 0.0030817261431366205, + 0.0030997293069958687, + 0.002483166754245758, + 0.0020622126758098602, + 0.0036849475000053644, + 0.004166808910667896, + 0.006437370087951422, + 0.009131564758718014, + 0.00971642043441534, + 0.010680794715881348, + 0.006439810153096914, + 0.008114811033010483, + 0.007499268744140863, + 0.008205587975680828, + 0.009043719619512558, + 0.0028599051292985678, + 0.0029352251440286636, + 0.005017912480980158, + 0.0031461231410503387, + 0.0037975425366312265, + 0.00340280425734818, + 0.004819436464458704, + 0.005137136671692133, + 0.004865965805947781, + 0.007882120087742805, + 0.010153365321457386, + 0.003771010087803006, + 0.004943144042044878, + 0.005437757354229689, + 0.005390150938183069, + 0.006735467351973057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.10176757723093033 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869940131902695 + ], + "opposite_logprob_tokens": [ + -3.048699378967285 + ], + "logprob_reward": -0.04869940131902695, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869940131902695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09617526829242706, + 0.3193361163139343, + 0.10169214010238647, + 0.05917976424098015, + 0.059349045157432556, + 0.09149264544248581, + 0.043393369764089584, + 0.06274434924125671, + 0.039080649614334106, + 0.12755656242370605 + ], + "credit_map": [ + 0.01668296940624714, + 0.008458632975816727, + 0.011215586215257645, + 0.01006962638348341, + 0.00988861732184887, + 0.015840835869312286, + 0.008829615078866482, + 0.01518938783556223, + 0.02473008818924427, + 0.0382910892367363, + 0.04334096983075142, + 0.05162494257092476, + 0.031153958290815353, + 0.03369767218828201, + 0.06345891207456589, + 0.03303851932287216, + 0.011491773650050163, + 0.012273735366761684, + 0.013190094381570816, + 0.02243639901280403, + 0.012089296244084835, + 0.009629933163523674, + 0.00999846588820219, + 0.010582439601421356, + 0.00921517051756382, + 0.009298315271735191, + 0.008145390078425407, + 0.005563103593885899, + 0.005339484661817551, + 0.00981432106345892, + 0.003971936646848917, + 0.007832041010260582, + 0.008771570399403572, + 0.00756139075383544, + 0.006653024349361658, + 0.0072891442105174065, + 0.006429251749068499, + 0.008576039224863052, + 0.006742360536009073, + 0.007326266262680292, + 0.008200976997613907, + 0.013633050955832005, + 0.0069022225216031075, + 0.011439193971455097, + 0.013384518213570118, + 0.01677572913467884, + 0.009725221432745457, + 0.011431739665567875, + 0.005247221793979406, + 0.005338510498404503, + 0.0051582506857812405, + 0.0056742336601018906, + 0.00410833302885294, + 0.004973606206476688, + 0.005915070418268442, + 0.006978144869208336, + 0.008699215948581696, + 0.007978755980730057, + 0.010417328216135502, + 0.00598913012072444, + 0.005263917613774538, + 0.007269820664077997, + 0.008026786148548126, + 0.009099394083023071, + 0.004915888421237469, + 0.004467944148927927, + 0.006562419701367617, + 0.0040510897524654865, + 0.0028738281689584255, + 0.006548104342073202, + 0.0033170348033308983, + 0.006344339810311794, + 0.010428349487483501, + 0.021726910024881363, + 0.017525434494018555, + 0.013259515166282654, + 0.01454215683043003, + 0.022123172879219055, + 0.01250412967056036, + 0.01544689480215311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.04869940131902695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005973107647150755 + ], + "opposite_logprob_tokens": [ + -5.1309733390808105 + ], + "logprob_reward": -0.005973107647150755, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005973107647150755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0888310968875885, + 0.15882350504398346, + 0.08923517167568207, + 0.18163995444774628, + 0.05998639762401581, + 0.07866474986076355, + 0.07956687361001968, + 0.06696957349777222, + 0.08311397582292557, + 0.11316870898008347 + ], + "credit_map": [ + 0.013682051561772823, + 0.009703394956886768, + 0.008835886605083942, + 0.012648460455238819, + 0.010851576924324036, + 0.009600836783647537, + 0.013306179083883762, + 0.010202710516750813, + 0.014358852989971638, + 0.01654687523841858, + 0.022511281073093414, + 0.019452309235930443, + 0.020618531852960587, + 0.013639623299241066, + 0.03825976327061653, + 0.013436278328299522, + 0.012246133759617805, + 0.009050742723047733, + 0.0121492063626647, + 0.014046098105609417, + 0.007218737155199051, + 0.010161575861275196, + 0.008637463673949242, + 0.015725214034318924, + 0.018541088327765465, + 0.022177422419190407, + 0.033812426030635834, + 0.013834775425493717, + 0.018674137070775032, + 0.026344168931245804, + 0.014070549979805946, + 0.034185394644737244, + 0.009176849387586117, + 0.005861405748873949, + 0.00594831770285964, + 0.007100865710526705, + 0.007814113050699234, + 0.010654691606760025, + 0.005001792684197426, + 0.008428358472883701, + 0.008813575841486454, + 0.008236361667513847, + 0.010792373679578304, + 0.011178969405591488, + 0.009289824403822422, + 0.011127745732665062, + 0.00945600587874651, + 0.009769890457391739, + 0.009993533603847027, + 0.009793019853532314, + 0.009079589508473873, + 0.011053458787500858, + 0.006107845809310675, + 0.010095648467540741, + 0.010245502926409245, + 0.01319827325642109, + 0.006702996790409088, + 0.01063966378569603, + 0.008976319804787636, + 0.0071648419834673405, + 0.006170734763145447, + 0.008249606937170029, + 0.010145622305572033, + 0.008919787593185902, + 0.008814282715320587, + 0.010041937232017517, + 0.009797163307666779, + 0.007184644695371389, + 0.00685791065916419, + 0.011850672774016857, + 0.013866289518773556, + 0.014701074920594692, + 0.010684398002922535, + 0.021899281069636345, + 0.01591513864696026, + 0.010943613015115261, + 0.014057121239602566, + 0.015298301354050636, + 0.011107989586889744, + 0.013262860476970673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.005973107647150755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497393012046814 + ], + "opposite_logprob_tokens": [ + -2.929973840713501 + ], + "logprob_reward": -0.05497393012046814, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497393012046814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18354342877864838, + 0.11681366711854935, + 0.05723099038004875, + 0.07529636472463608, + 0.10692567378282547, + 0.19765733182430267, + 0.05021289363503456, + 0.0440005287528038, + 0.05974362790584564, + 0.1085754856467247 + ], + "credit_map": [ + 0.024032611399888992, + 0.015436558052897453, + 0.020413856953382492, + 0.024033134803175926, + 0.01919199898838997, + 0.02284403145313263, + 0.024707084521651268, + 0.03288416564464569, + 0.011121702380478382, + 0.013200611807405949, + 0.01903577707707882, + 0.014566372148692608, + 0.010937025770545006, + 0.013704992830753326, + 0.021096428856253624, + 0.013150755316019058, + 0.00774742104113102, + 0.005642965901643038, + 0.00833094958215952, + 0.014487508684396744, + 0.005172864533960819, + 0.004052810370922089, + 0.006193183362483978, + 0.005603288300335407, + 0.010187317617237568, + 0.009803411550819874, + 0.008136012591421604, + 0.01076801773160696, + 0.011717920191586018, + 0.00854138657450676, + 0.0064290775917470455, + 0.00971322413533926, + 0.016885345801711082, + 0.012520966120064259, + 0.012219441123306751, + 0.014028478413820267, + 0.010256826877593994, + 0.011626027524471283, + 0.015379366464912891, + 0.014009227976202965, + 0.015003548935055733, + 0.028544725850224495, + 0.014860671013593674, + 0.028603358194231987, + 0.026575371623039246, + 0.027932841330766678, + 0.029678381979465485, + 0.026458434760570526, + 0.0051297591999173164, + 0.005071142688393593, + 0.005454293917864561, + 0.005065322853624821, + 0.0028153022285550833, + 0.0071764616295695305, + 0.007312868256121874, + 0.012187742628157139, + 0.0056039635092020035, + 0.004427547566592693, + 0.00865423958748579, + 0.004329071380198002, + 0.004189610481262207, + 0.0047729844227433205, + 0.004922135733067989, + 0.007100972346961498, + 0.005275342147797346, + 0.006705236155539751, + 0.006838660221546888, + 0.005222391337156296, + 0.005819838959723711, + 0.007287682965397835, + 0.01189566683024168, + 0.01069880835711956, + 0.019036294892430305, + 0.010334682650864124, + 0.02136835642158985, + 0.009552739560604095, + 0.01112639531493187, + 0.011623387224972248, + 0.010175934061408043, + 0.015357697382569313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.05497393012046814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290636017918587 + ], + "opposite_logprob_tokens": [ + -3.7732906341552734 + ], + "logprob_reward": -0.023290636017918587, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290636017918587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09527761489152908, + 0.2259245663881302, + 0.08666560798883438, + 0.06115005165338516, + 0.08025290071964264, + 0.22111891210079193, + 0.031179610639810562, + 0.05995527282357216, + 0.05465938523411751, + 0.08381613343954086 + ], + "credit_map": [ + 0.011302201077342033, + 0.012054054997861385, + 0.01138186827301979, + 0.012539449147880077, + 0.01298592984676361, + 0.010133696720004082, + 0.012228704057633877, + 0.0126517154276371, + 0.016621870920062065, + 0.02733626589179039, + 0.03258216753602028, + 0.03964671492576599, + 0.018877791240811348, + 0.025792822241783142, + 0.038612619042396545, + 0.026454316452145576, + 0.010012711398303509, + 0.0058050816878676414, + 0.014357388019561768, + 0.022065294906497, + 0.004395367112010717, + 0.005121143534779549, + 0.010386927984654903, + 0.01452169194817543, + 0.008591744117438793, + 0.005962511990219355, + 0.010152940638363361, + 0.008001274429261684, + 0.004740568343549967, + 0.008476166054606438, + 0.004376379307359457, + 0.010848463512957096, + 0.012697216123342514, + 0.010491307824850082, + 0.007809292525053024, + 0.011045927181839943, + 0.007945872843265533, + 0.011097115464508533, + 0.008509745821356773, + 0.010656419210135937, + 0.017140284180641174, + 0.03177411109209061, + 0.016312718391418457, + 0.03213601931929588, + 0.03146081045269966, + 0.0300211813300848, + 0.032875578850507736, + 0.029398197308182716, + 0.0027356515638530254, + 0.004103653598576784, + 0.0032610632479190826, + 0.0027675398159772158, + 0.0032219632994383574, + 0.004048328846693039, + 0.004332637879997492, + 0.006708771921694279, + 0.008943349123001099, + 0.007123176008462906, + 0.009869761765003204, + 0.006412450224161148, + 0.005522081628441811, + 0.006154829636216164, + 0.0074975453317165375, + 0.00843207910656929, + 0.0051453267224133015, + 0.006562970578670502, + 0.007068185601383448, + 0.00528395501896739, + 0.005029129330068827, + 0.006009327247738838, + 0.01021191943436861, + 0.009348572231829166, + 0.01210810337215662, + 0.011810076422989368, + 0.016223665326833725, + 0.009278254583477974, + 0.006850116886198521, + 0.008551725186407566, + 0.0074350712820887566, + 0.01155911199748516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.023290636017918587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slow-motion, tracking shot captures a bizarre fashion show taking place in a surreal hotel. Amidst a maze of intense beams of light, alternating crimson, royal blue, and eerie green, several tall, androgynous models slowly march along an endlessly long corridor with a vaulted ceiling. Clad in avant-garde, dramatically structured garments in vibrant, saturated colors, some embellished with shimmering metal, others crafted from translucent tulle, they move steadily forward in a ritualistic, almost hypnotic rhythm, their faces expressionless, their eyes vacant and piercing. As they move, light streaming through the gaps between the ceiling and floor dramatically elongates and distorts their figures, casting sharp, dancing shadows across the smooth, obsidian floor. The walls of the corridor are made up of countless mirrors of different sizes and angles, which continuously reflect the fragmented and overlapping images of the models, as if countless clones are moving simultaneously in different dimensions. The whole scene is full of Jodorowsky-style dramatic tension and visual symbolism, creating a dream atmosphere that is both gorgeous and unsettling.", + "seed": 21967, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_8949_seed_21967_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07429156713187694, + "vlm_logprob": -0.07429156713187694, + "vlm_yes_no_margin": 3.5249999523162843, + "vlm_token_credit": -0.07429156713187694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2957985997200012 + ], + "opposite_logprob_tokens": [ + -1.5457985401153564 + ], + "logprob_reward": -0.2957985997200012, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2957985997200012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11898448318243027, + 0.04111480340361595, + 0.20362700521945953, + 0.02722158096730709, + 0.27743634581565857, + 0.15572765469551086, + 0.04925159737467766, + 0.051186636090278625, + 0.038717832416296005, + 0.03673209995031357 + ], + "credit_map": [ + 0.014943958260118961, + 0.015474804677069187, + 0.010836709290742874, + 0.023309100419282913, + 0.012902733869850636, + 0.013995335437357426, + 0.016176413744688034, + 0.01134543213993311, + 0.005789360962808132, + 0.004897823557257652, + 0.004716686904430389, + 0.005311647895723581, + 0.006949302740395069, + 0.003737661987543106, + 0.007027465384453535, + 0.0026848530396819115, + 0.06557326763868332, + 0.009702462702989578, + 0.008735666051506996, + 0.008195276372134686, + 0.040361519902944565, + 0.01847299560904503, + 0.016742797568440437, + 0.035843003541231155, + 0.003945412579923868, + 0.00256990734487772, + 0.0030148618388921022, + 0.002528299344703555, + 0.004928299691528082, + 0.005022501572966576, + 0.003414316102862358, + 0.0017979826079681516, + 0.025142647325992584, + 0.09343712776899338, + 0.008145089261233807, + 0.02084178663790226, + 0.019056059420108795, + 0.08121933788061142, + 0.017633236944675446, + 0.011961059644818306, + 0.025071516633033752, + 0.017287323251366615, + 0.008973951451480389, + 0.023713193833827972, + 0.028398267924785614, + 0.023939210921525955, + 0.01761343702673912, + 0.010730748996138573, + 0.004391433671116829, + 0.005398268345743418, + 0.0037018675357103348, + 0.008011688478291035, + 0.005613763816654682, + 0.00649805273860693, + 0.01008567214012146, + 0.00555085064843297, + 0.008442445658147335, + 0.007051672320812941, + 0.0037904728669673204, + 0.0074745966121554375, + 0.005698390770703554, + 0.005419082008302212, + 0.00473179342225194, + 0.008578184992074966, + 0.0038986681029200554, + 0.005002876278012991, + 0.002043589483946562, + 0.005883068777620792, + 0.0039545451290905476, + 0.0053709023632109165, + 0.007303327322006226, + 0.005260854959487915, + 0.0028147296980023384, + 0.006195649039000273, + 0.0023814314045011997, + 0.008401553146541119, + 0.00443365890532732, + 0.004210446495562792, + 0.003761146916076541, + 0.004533485043793917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.2957985997200012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016109425574541092 + ], + "opposite_logprob_tokens": [ + -4.141109466552734 + ], + "logprob_reward": -0.016109425574541092, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016109425574541092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08071418106555939, + 0.10778980702161789, + 0.34712928533554077, + 0.049528494477272034, + 0.10359743982553482, + 0.0696563795208931, + 0.04342586174607277, + 0.03525952994823456, + 0.06312031298875809, + 0.09977873414754868 + ], + "credit_map": [ + 0.012920808047056198, + 0.013130339793860912, + 0.008684933185577393, + 0.013961979188024998, + 0.01015537604689598, + 0.00892628449946642, + 0.006867921445518732, + 0.006066541187465191, + 0.019060736522078514, + 0.010667971335351467, + 0.013368014246225357, + 0.017362264916300774, + 0.018490778282284737, + 0.008061732165515423, + 0.014001206494867802, + 0.006777103990316391, + 0.12242808938026428, + 0.01750510185956955, + 0.05882861837744713, + 0.028353093191981316, + 0.027370942756533623, + 0.019765732809901237, + 0.028445076197385788, + 0.04443264380097389, + 0.004918419755995274, + 0.003258747048676014, + 0.002409445121884346, + 0.003462861990556121, + 0.007939022965729237, + 0.01803184114396572, + 0.005011293571442366, + 0.0044968645088374615, + 0.008786614052951336, + 0.03360830247402191, + 0.002851921133697033, + 0.007584294304251671, + 0.008005000650882721, + 0.03194694593548775, + 0.006106461398303509, + 0.004707907792180777, + 0.01301624532788992, + 0.007816300727427006, + 0.005749951582401991, + 0.009522730484604836, + 0.01093823742121458, + 0.009569975547492504, + 0.005078797694295645, + 0.00796414539217949, + 0.0044681415893137455, + 0.004977551754564047, + 0.00312807597219944, + 0.006449838634580374, + 0.005876016337424517, + 0.006192948669195175, + 0.007060644682496786, + 0.005272645503282547, + 0.0035602517891675234, + 0.005543801002204418, + 0.0025462547782808542, + 0.004816543310880661, + 0.004370781127363443, + 0.005184687674045563, + 0.004315685946494341, + 0.004921523854136467, + 0.005154712591320276, + 0.007979423739016056, + 0.0033151889219880104, + 0.00813309196382761, + 0.009557580575346947, + 0.011111151427030563, + 0.010202026925981045, + 0.007667132653295994, + 0.012187926098704338, + 0.01667998544871807, + 0.007160680368542671, + 0.01485315803438425, + 0.012871265411376953, + 0.013187509961426258, + 0.012093434110283852, + 0.010744770057499409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.016109425574541092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076508354395627975 + ], + "opposite_logprob_tokens": [ + -4.882650852203369 + ], + "logprob_reward": -0.0076508354395627975, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076508354395627975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10712884366512299, + 0.04868169128894806, + 0.3256218731403351, + 0.1326955258846283, + 0.06945871561765671, + 0.14883776009082794, + 0.036991190165281296, + 0.05746297538280487, + 0.03109050914645195, + 0.04203089699149132 + ], + "credit_map": [ + 0.017177658155560493, + 0.013429408892989159, + 0.00972739141434431, + 0.026314932852983475, + 0.012638437561690807, + 0.010037637315690517, + 0.01100221462547779, + 0.006801160983741283, + 0.008814305067062378, + 0.004454275127500296, + 0.006527800112962723, + 0.006943674758076668, + 0.007534859701991081, + 0.005141717381775379, + 0.005733228754252195, + 0.003531831782311201, + 0.06036340072751045, + 0.016089117154479027, + 0.03143305331468582, + 0.012790119275450706, + 0.1126522645354271, + 0.050440724939107895, + 0.024564296007156372, + 0.01728888601064682, + 0.01189765427261591, + 0.008003553375601768, + 0.0042288051918148994, + 0.007507803849875927, + 0.02077048271894455, + 0.05142123997211456, + 0.012522352859377861, + 0.016343627125024796, + 0.005224794149398804, + 0.016527704894542694, + 0.0029927666764706373, + 0.006146860308945179, + 0.00640846136957407, + 0.021560752764344215, + 0.006417625118046999, + 0.004179754760116339, + 0.022195013239979744, + 0.016475150361657143, + 0.008485437370836735, + 0.02096201293170452, + 0.029303256422281265, + 0.02284069173038006, + 0.01842327043414116, + 0.010152940638363361, + 0.003315429436042905, + 0.004944347310811281, + 0.003003225661814213, + 0.004614155273884535, + 0.006179888267070055, + 0.005607640836387873, + 0.004368271678686142, + 0.004958232864737511, + 0.00710113812237978, + 0.007629765663295984, + 0.004771742504090071, + 0.005830217152833939, + 0.006436982192099094, + 0.007884512655436993, + 0.0068084439262747765, + 0.011000169441103935, + 0.0031212635803967714, + 0.004225629847496748, + 0.0021214315202087164, + 0.005912070162594318, + 0.0043007200583815575, + 0.0036734004970639944, + 0.004516023211181164, + 0.003219971200451255, + 0.0030578509904444218, + 0.007183305453509092, + 0.004125640261918306, + 0.007804890628904104, + 0.0052400310523808, + 0.00465581938624382, + 0.004844264127314091, + 0.005119092296808958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.0076508354395627975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371366113424301 + ], + "opposite_logprob_tokens": [ + -3.4087135791778564 + ], + "logprob_reward": -0.03371366113424301, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371366113424301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09106277674436569, + 0.042638957500457764, + 0.19993357360363007, + 0.02307056076824665, + 0.3028131127357483, + 0.1188819482922554, + 0.09325160831212997, + 0.038267284631729126, + 0.056272320449352264, + 0.033807940781116486 + ], + "credit_map": [ + 0.01242784783244133, + 0.009789668023586273, + 0.009300731122493744, + 0.012241383083164692, + 0.017048664391040802, + 0.013130951672792435, + 0.011871689930558205, + 0.005251828581094742, + 0.005927937105298042, + 0.006636044010519981, + 0.0039049305487424135, + 0.00668790377676487, + 0.0041979337111115456, + 0.005696422420442104, + 0.006643188651651144, + 0.0029445947147905827, + 0.023934606462717056, + 0.010852474719285965, + 0.026246625930070877, + 0.012028910219669342, + 0.0737585499882698, + 0.03178239241242409, + 0.015552462078630924, + 0.005777549929916859, + 0.0038208994083106518, + 0.0018520148005336523, + 0.0018676701001822948, + 0.0018797998782247305, + 0.0035237823612987995, + 0.005191350355744362, + 0.003018334275111556, + 0.00191670807544142, + 0.026390254497528076, + 0.10013327747583389, + 0.00843232311308384, + 0.021206626668572426, + 0.02238941751420498, + 0.09242658317089081, + 0.017252804711461067, + 0.01458179485052824, + 0.021702930331230164, + 0.01481739804148674, + 0.009348772466182709, + 0.015101063065230846, + 0.019903330132365227, + 0.01571130007505417, + 0.009954862296581268, + 0.012342283502221107, + 0.009417074732482433, + 0.011356189846992493, + 0.003816754324361682, + 0.008382179774343967, + 0.010842789895832539, + 0.014008536003530025, + 0.021308623254299164, + 0.014119453728199005, + 0.003350228304043412, + 0.004631425254046917, + 0.00219699926674366, + 0.005605157930403948, + 0.0035169951152056456, + 0.00708750868216157, + 0.006197735667228699, + 0.0056812334805727005, + 0.006155448965728283, + 0.007174002006649971, + 0.003076090943068266, + 0.009264127351343632, + 0.008284869603812695, + 0.010042684152722359, + 0.007101557683199644, + 0.005173534620553255, + 0.0029472645837813616, + 0.005298805423080921, + 0.0025843980256468058, + 0.0034837224520742893, + 0.006648307200521231, + 0.004405476152896881, + 0.0044820550829172134, + 0.003957908600568771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.03371366113424301 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185313791036606 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185313791036606, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018185313791036606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10638020932674408, + 0.18056468665599823, + 0.17677946388721466, + 0.023017272353172302, + 0.1941702961921692, + 0.15883708000183105, + 0.047800809144973755, + 0.04832841083407402, + 0.03435854613780975, + 0.029763175174593925 + ], + "credit_map": [ + 0.012900952249765396, + 0.013553354889154434, + 0.011997701600193977, + 0.022436492145061493, + 0.015407472848892212, + 0.01025580707937479, + 0.013041039928793907, + 0.006787394173443317, + 0.03272313252091408, + 0.020589901134371758, + 0.019056659191846848, + 0.02048184536397457, + 0.032178401947021484, + 0.014771086163818836, + 0.031588491052389145, + 0.009175175800919533, + 0.029699983075261116, + 0.00877731665968895, + 0.028646761551499367, + 0.009568064473569393, + 0.05580173432826996, + 0.018874282017350197, + 0.013528530485928059, + 0.011882802471518517, + 0.0030008209869265556, + 0.0024518638383597136, + 0.001777459867298603, + 0.002781926654279232, + 0.003370333928614855, + 0.0023196726106107235, + 0.005244057159870863, + 0.002071140566840768, + 0.015671217814087868, + 0.058466095477342606, + 0.005097865127027035, + 0.013924572616815567, + 0.015488292090594769, + 0.06291811913251877, + 0.012287978082895279, + 0.010316186584532261, + 0.01982935518026352, + 0.018885333091020584, + 0.008419203571975231, + 0.023039640858769417, + 0.0331202931702137, + 0.02190122939646244, + 0.02146492339670658, + 0.012177130207419395, + 0.007825043052434921, + 0.006152955815196037, + 0.0030784697737544775, + 0.0048289368860423565, + 0.006569426041096449, + 0.006333612371236086, + 0.00618017977103591, + 0.0068321917206048965, + 0.007722022477537394, + 0.007185718510299921, + 0.00342154735699296, + 0.005078080575913191, + 0.006023094989359379, + 0.0055463616736233234, + 0.004721896722912788, + 0.008629695512354374, + 0.0032269067596644163, + 0.004948711954057217, + 0.002237563719972968, + 0.005216691642999649, + 0.006348446477204561, + 0.005160271655768156, + 0.0038031060248613358, + 0.0034168518614023924, + 0.0033003760036081076, + 0.0035598042886704206, + 0.0026664575561881065, + 0.0032605610322207212, + 0.005852417089045048, + 0.003888725535944104, + 0.004061152692884207, + 0.003173684701323509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.018185313791036606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slow-motion, tracking shot captures a bizarre fashion show taking place in a surreal hotel. Amidst a maze of intense beams of light, alternating crimson, royal blue, and eerie green, several tall, androgynous models slowly march along an endlessly long corridor with a vaulted ceiling. Clad in avant-garde, dramatically structured garments in vibrant, saturated colors, some embellished with shimmering metal, others crafted from translucent tulle, they move steadily forward in a ritualistic, almost hypnotic rhythm, their faces expressionless, their eyes vacant and piercing. As they move, light streaming through the gaps between the ceiling and floor dramatically elongates and distorts their figures, casting sharp, dancing shadows across the smooth, obsidian floor. The walls of the corridor are made up of countless mirrors of different sizes and angles, which continuously reflect the fragmented and overlapping images of the models, as if countless clones are moving simultaneously in different dimensions. The whole scene is full of Jodorowsky-style dramatic tension and visual symbolism, creating a dream atmosphere that is both gorgeous and unsettling.", + "seed": 21968, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_8949_seed_21968_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030673728650435807, + "vlm_logprob": -0.030673728650435807, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.030673728650435807, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07105928659439087 + ], + "opposite_logprob_tokens": [ + -2.946059226989746 + ], + "logprob_reward": -0.07105928659439087, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07105928659439087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09561555087566376, + 0.14718353748321533, + 0.06299301236867905, + 0.1900729089975357, + 0.18061594665050507, + 0.0746622309088707, + 0.054675012826919556, + 0.05755874142050743, + 0.06935887038707733, + 0.0672641396522522 + ], + "credit_map": [ + 0.00844112690538168, + 0.00907182041555643, + 0.014342236332595348, + 0.017044642940163612, + 0.009789842180907726, + 0.009040392935276031, + 0.0096775246784091, + 0.01820797473192215, + 0.0114304693415761, + 0.005876089911907911, + 0.05065915733575821, + 0.009634407237172127, + 0.007311500608921051, + 0.008677414618432522, + 0.007863027974963188, + 0.04573148116469383, + 0.006065111607313156, + 0.004510240163654089, + 0.008446716703474522, + 0.008557078428566456, + 0.01434538047760725, + 0.005354305729269981, + 0.009075811132788658, + 0.0066383760422468185, + 0.013390537351369858, + 0.006442919839173555, + 0.06543255597352982, + 0.0114830881357193, + 0.012097547762095928, + 0.01050610188394785, + 0.008729697205126286, + 0.061990462243556976, + 0.013141771778464317, + 0.013209095224738121, + 0.02708817459642887, + 0.026784827932715416, + 0.014453813433647156, + 0.017963597550988197, + 0.02423836477100849, + 0.04373631998896599, + 0.005384017713367939, + 0.00581749714910984, + 0.010268247686326504, + 0.010613836348056793, + 0.008495298214256763, + 0.010583223775029182, + 0.013060886412858963, + 0.010439228266477585, + 0.0073513733223080635, + 0.006652029696851969, + 0.008550730533897877, + 0.007354598492383957, + 0.005282630678266287, + 0.005309632048010826, + 0.0074636125937104225, + 0.006710411980748177, + 0.008019953966140747, + 0.0072620343416929245, + 0.007320643402636051, + 0.010007715784013271, + 0.0063325478695333, + 0.0046472507528960705, + 0.008139564655721188, + 0.005829032976180315, + 0.008024015463888645, + 0.006778342183679342, + 0.0086539750918746, + 0.008636880666017532, + 0.007946287281811237, + 0.00881042517721653, + 0.011548749171197414, + 0.00896020419895649, + 0.005694539751857519, + 0.00455345818772912, + 0.011421628296375275, + 0.010344126261770725, + 0.009476625360548496, + 0.008397242985665798, + 0.008424686267971992, + 0.008951842784881592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.07105928659439087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026379402726888657 + ], + "opposite_logprob_tokens": [ + -3.651379346847534 + ], + "logprob_reward": -0.026379402726888657, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026379402726888657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1733519434928894, + 0.10176556557416916, + 0.049913134425878525, + 0.07084809988737106, + 0.11152452975511551, + 0.0484163761138916, + 0.11953967809677124, + 0.08044003695249557, + 0.08532822877168655, + 0.15887245535850525 + ], + "credit_map": [ + 0.014215162955224514, + 0.009599839337170124, + 0.01989883929491043, + 0.041162312030792236, + 0.011933114379644394, + 0.014940970577299595, + 0.015186626464128494, + 0.04641508311033249, + 0.008456030860543251, + 0.004903484135866165, + 0.025183234363794327, + 0.013412849977612495, + 0.007580313831567764, + 0.009778049774467945, + 0.009870719164609909, + 0.022580880671739578, + 0.005269091576337814, + 0.004859393462538719, + 0.007145807147026062, + 0.0049899788573384285, + 0.00828576646745205, + 0.0052532316185534, + 0.007635672576725483, + 0.006474194582551718, + 0.0071944547817111015, + 0.005545942112803459, + 0.016354912891983986, + 0.009418587200343609, + 0.004672582261264324, + 0.00524162920191884, + 0.0056730397045612335, + 0.01674695312976837, + 0.0066963257268071175, + 0.007006704341620207, + 0.016012871637940407, + 0.014007172547280788, + 0.013765187934041023, + 0.007056063506752253, + 0.02930505760014057, + 0.01767515018582344, + 0.00547484215348959, + 0.004932000767439604, + 0.008843830786645412, + 0.004778279922902584, + 0.005912733729928732, + 0.005066593177616596, + 0.005461266729980707, + 0.007946829311549664, + 0.013695972971618176, + 0.009962369687855244, + 0.024755405262112617, + 0.015480334870517254, + 0.009519778192043304, + 0.014107086695730686, + 0.016662027686834335, + 0.015356707386672497, + 0.01076419372111559, + 0.0077465856447815895, + 0.008390608243644238, + 0.01337751280516386, + 0.007344275247305632, + 0.007808216381818056, + 0.008737712167203426, + 0.01627092994749546, + 0.009364991448819637, + 0.005738080479204655, + 0.012048914097249508, + 0.009072737768292427, + 0.010331649333238602, + 0.012143552303314209, + 0.01430434174835682, + 0.012323962524533272, + 0.01611907035112381, + 0.012750229798257351, + 0.01978861726820469, + 0.020544057711958885, + 0.023536313325166702, + 0.0223550945520401, + 0.024599215015769005, + 0.019179854542016983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.026379402726888657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003631665138527751 + ], + "opposite_logprob_tokens": [ + -5.628631591796875 + ], + "logprob_reward": -0.003631665138527751, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003631665138527751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12571944296360016, + 0.142698734998703, + 0.08575557172298431, + 0.19177092611789703, + 0.06351003795862198, + 0.03310282528400421, + 0.1578013002872467, + 0.06050603836774826, + 0.06690246611833572, + 0.07223266363143921 + ], + "credit_map": [ + 0.014648335985839367, + 0.010157722979784012, + 0.01774110458791256, + 0.020093150436878204, + 0.011194170452654362, + 0.014830190688371658, + 0.015692148357629776, + 0.02136261947453022, + 0.0140483183786273, + 0.007322073448449373, + 0.03835237771272659, + 0.011998594738543034, + 0.008908592164516449, + 0.017084596678614616, + 0.009311534464359283, + 0.0356726348400116, + 0.010320373810827732, + 0.0067665837705135345, + 0.0103991674259305, + 0.007578528951853514, + 0.02285779081285, + 0.008068634197115898, + 0.010550696402788162, + 0.009213799610733986, + 0.012576023116707802, + 0.004532049410045147, + 0.07251286506652832, + 0.011098271235823631, + 0.005076650995761156, + 0.01104495208710432, + 0.006315192207694054, + 0.06861492991447449, + 0.004314566031098366, + 0.003377099521458149, + 0.01140762958675623, + 0.0077657876536250114, + 0.007573076523840427, + 0.00485524395480752, + 0.014182396233081818, + 0.010034237056970596, + 0.0029618777334690094, + 0.0026596856769174337, + 0.00391724705696106, + 0.004056509118527174, + 0.0043128603138029575, + 0.003365990938618779, + 0.007199530489742756, + 0.004629124421626329, + 0.01687856949865818, + 0.009295722469687462, + 0.03617464005947113, + 0.018608298152685165, + 0.008453224785625935, + 0.0182303749024868, + 0.020425518974661827, + 0.029734954237937927, + 0.005894313100725412, + 0.005689338315278292, + 0.006836313288658857, + 0.01200158428400755, + 0.004945425316691399, + 0.005131834652274847, + 0.008208140730857849, + 0.011799088679254055, + 0.009087707847356796, + 0.004403069615364075, + 0.008795497007668018, + 0.007338083349168301, + 0.007688028737902641, + 0.007167466916143894, + 0.012289467267692089, + 0.010133147239685059, + 0.008987238630652428, + 0.006739340722560883, + 0.00937980879098177, + 0.008943975903093815, + 0.011745098978281021, + 0.009198760613799095, + 0.006244312040507793, + 0.010994130745530128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.003631665138527751 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810510411858559 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.03810510411858559, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810510411858559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11917484551668167, + 0.1735556721687317, + 0.08161265403032303, + 0.2040686011314392, + 0.07992953062057495, + 0.04633607715368271, + 0.044476620852947235, + 0.11189617216587067, + 0.08046949654817581, + 0.058480292558670044 + ], + "credit_map": [ + 0.008109202608466148, + 0.006437240168452263, + 0.014517069794237614, + 0.02809225581586361, + 0.0076975952833890915, + 0.008747429586946964, + 0.015682479366660118, + 0.029891574755311012, + 0.015441453084349632, + 0.007558005396276712, + 0.05438762530684471, + 0.015503192320466042, + 0.006520927418023348, + 0.012939447537064552, + 0.01041214819997549, + 0.0507928691804409, + 0.009797794744372368, + 0.007726714480668306, + 0.008477555587887764, + 0.0064121694304049015, + 0.02263314090669155, + 0.008213748224079609, + 0.01048434991389513, + 0.00786717887967825, + 0.014026643708348274, + 0.004480324219912291, + 0.07656638324260712, + 0.011151209473609924, + 0.005793652031570673, + 0.011185741052031517, + 0.009009208530187607, + 0.07185544818639755, + 0.006102989427745342, + 0.005668717436492443, + 0.010569246485829353, + 0.009849006310105324, + 0.009956187568604946, + 0.008593960665166378, + 0.016650788486003876, + 0.012538635171949863, + 0.004795050714164972, + 0.004431702196598053, + 0.00828656367957592, + 0.005191170610487461, + 0.004197233356535435, + 0.0046790409833192825, + 0.007477068807929754, + 0.007278244011104107, + 0.0055642868392169476, + 0.005861969664692879, + 0.0074381157755851746, + 0.006162141915410757, + 0.00502435676753521, + 0.004396692384034395, + 0.005886398255825043, + 0.00414265738800168, + 0.011901594698429108, + 0.010624026879668236, + 0.01227949745953083, + 0.022347358986735344, + 0.010054677724838257, + 0.009372599422931671, + 0.013971296139061451, + 0.02134513109922409, + 0.01671570912003517, + 0.009005199186503887, + 0.010651638731360435, + 0.01088722050189972, + 0.007646195590496063, + 0.00932345725595951, + 0.009939844720065594, + 0.006300237029790878, + 0.0050661140121519566, + 0.005168190225958824, + 0.006986058317124844, + 0.008770259097218513, + 0.009947661310434341, + 0.006672008894383907, + 0.00869612954556942, + 0.007173870224505663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.03810510411858559 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09182239323854446, + 0.09421487897634506, + 0.4404556155204773, + 0.09732085466384888, + 0.06110159680247307, + 0.01435086876153946, + 0.027917426079511642, + 0.030658408999443054, + 0.08160660415887833, + 0.06055138632655144 + ], + "credit_map": [ + 0.01004927046597004, + 0.006154188420623541, + 0.012182402424514294, + 0.01688913069665432, + 0.00706147775053978, + 0.007999065332114697, + 0.013139848597347736, + 0.01834700256586075, + 0.011306478641927242, + 0.004191613756120205, + 0.02241654321551323, + 0.011226216331124306, + 0.006256499327719212, + 0.00781423132866621, + 0.009626908227801323, + 0.021376388147473335, + 0.05455984175205231, + 0.027581267058849335, + 0.07435324788093567, + 0.0446929968893528, + 0.08374147117137909, + 0.03381152078509331, + 0.05523945391178131, + 0.06647578626871109, + 0.007213427219539881, + 0.0032069676090031862, + 0.033456385135650635, + 0.0068180374801158905, + 0.005555036943405867, + 0.005537049844861031, + 0.0050211395137012005, + 0.03051280602812767, + 0.0030082648154348135, + 0.00437137670814991, + 0.009883468970656395, + 0.007860389538109303, + 0.008748173713684082, + 0.004475579597055912, + 0.01340470090508461, + 0.009349641390144825, + 0.0022410748060792685, + 0.0014953567879274487, + 0.002249832032248378, + 0.0016318841371685266, + 0.0014418528880923986, + 0.0015742082614451647, + 0.0017239573644474149, + 0.0019927024841308594, + 0.0030827855225652456, + 0.0014381343498826027, + 0.006554818712174892, + 0.004239573143422604, + 0.0014655734412372112, + 0.0037051101680845022, + 0.0025335189420729876, + 0.004897912032902241, + 0.0034490751568228006, + 0.002547804731875658, + 0.0035519073717296124, + 0.0053391228429973125, + 0.0021480340510606766, + 0.0034054676070809364, + 0.005217435769736767, + 0.004999559838324785, + 0.008282121270895004, + 0.006141189951449633, + 0.011589735746383667, + 0.00949057936668396, + 0.009762454777956009, + 0.009684018790721893, + 0.015694236382842064, + 0.01096226368099451, + 0.006922237575054169, + 0.006298834457993507, + 0.0062336488626897335, + 0.007335883565247059, + 0.00932640302926302, + 0.008200375363230705, + 0.008495887741446495, + 0.007738117128610611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.014193184673786163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slow-motion, tracking shot captures a bizarre fashion show taking place in a surreal hotel. Amidst a maze of intense beams of light, alternating crimson, royal blue, and eerie green, several tall, androgynous models slowly march along an endlessly long corridor with a vaulted ceiling. Clad in avant-garde, dramatically structured garments in vibrant, saturated colors, some embellished with shimmering metal, others crafted from translucent tulle, they move steadily forward in a ritualistic, almost hypnotic rhythm, their faces expressionless, their eyes vacant and piercing. As they move, light streaming through the gaps between the ceiling and floor dramatically elongates and distorts their figures, casting sharp, dancing shadows across the smooth, obsidian floor. The walls of the corridor are made up of countless mirrors of different sizes and angles, which continuously reflect the fragmented and overlapping images of the models, as if countless clones are moving simultaneously in different dimensions. The whole scene is full of Jodorowsky-style dramatic tension and visual symbolism, creating a dream atmosphere that is both gorgeous and unsettling.", + "seed": 21969, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_2/index_8949_seed_21969_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.042694459389895204, + "vlm_logprob": -0.042694459389895204, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.042694459389895204, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09301376342773438 + ], + "opposite_logprob_tokens": [ + -2.7180137634277344 + ], + "logprob_reward": -0.09301376342773438, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09301376342773438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13639794290065765, + 0.11376088112592697, + 0.12288253009319305, + 0.1268603503704071, + 0.12100960314273834, + 0.11979484558105469, + 0.07576123625040054, + 0.06768271327018738, + 0.05030183121562004, + 0.06554808467626572 + ], + "credit_map": [ + 0.016658548265695572, + 0.017526358366012573, + 0.016024822369217873, + 0.016140202060341835, + 0.02128269523382187, + 0.01173512265086174, + 0.018632065504789352, + 0.01839812844991684, + 0.009449723176658154, + 0.015509501099586487, + 0.009227722883224487, + 0.012098566628992558, + 0.02541319653391838, + 0.009997199289500713, + 0.019370250403881073, + 0.01269472949206829, + 0.014567340724170208, + 0.010500896722078323, + 0.020659998059272766, + 0.0157938189804554, + 0.010310069657862186, + 0.017560629174113274, + 0.013354887254536152, + 0.02013489417731762, + 0.023861663416028023, + 0.010066371411085129, + 0.018669096753001213, + 0.01269654929637909, + 0.0134978536516428, + 0.014104201458394527, + 0.011400298215448856, + 0.02256431244313717, + 0.010791461914777756, + 0.010201339609920979, + 0.011461286805570126, + 0.01762319542467594, + 0.017806818708777428, + 0.01652112416923046, + 0.0182400643825531, + 0.01836431585252285, + 0.007849989458918571, + 0.008162437938153744, + 0.01295949425548315, + 0.015141986310482025, + 0.024120353162288666, + 0.013140851631760597, + 0.013498904183506966, + 0.024920834228396416, + 0.0065314737148582935, + 0.005395715124905109, + 0.009258919395506382, + 0.00849119108170271, + 0.013268324546515942, + 0.010826087556779385, + 0.01042365375906229, + 0.011565874330699444, + 0.010230048559606075, + 0.005313225090503693, + 0.010841415263712406, + 0.00779940839856863, + 0.008347487077116966, + 0.008605183102190495, + 0.005929740611463785, + 0.010616202838718891, + 0.0057113501243293285, + 0.0043007563799619675, + 0.006115075666457415, + 0.005884617567062378, + 0.007086168508976698, + 0.004875513259321451, + 0.007793453987687826, + 0.008534891530871391, + 0.00918571650981903, + 0.01002664864063263, + 0.00766487792134285, + 0.008188143372535706, + 0.0061678276397287846, + 0.006900360807776451, + 0.008667572401463985, + 0.008746937848627567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.09301376342773438 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500303953886032 + ], + "opposite_logprob_tokens": [ + -2.9300029277801514 + ], + "logprob_reward": -0.05500303953886032, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500303953886032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18033020198345184, + 0.17304380238056183, + 0.09610141813755035, + 0.14548124372959137, + 0.052413079887628555, + 0.09534081816673279, + 0.07127410173416138, + 0.06378605216741562, + 0.07717834413051605, + 0.04505094513297081 + ], + "credit_map": [ + 0.023424185812473297, + 0.02367463707923889, + 0.020446620881557465, + 0.02561742067337036, + 0.027149543166160583, + 0.01622915081679821, + 0.02606694959104061, + 0.01772167906165123, + 0.00955118052661419, + 0.010891212150454521, + 0.009472288191318512, + 0.013095946051180363, + 0.022832803428173065, + 0.010791733860969543, + 0.051948972046375275, + 0.044459644705057144, + 0.010559998452663422, + 0.010576694272458553, + 0.015146943740546703, + 0.01149153895676136, + 0.011778103187680244, + 0.010575540363788605, + 0.011856495402753353, + 0.014116098172962666, + 0.01815032958984375, + 0.01662043109536171, + 0.01994236931204796, + 0.009843608364462852, + 0.02857297472655773, + 0.013995571061968803, + 0.01905377022922039, + 0.019302191212773323, + 0.006086498964577913, + 0.005472235381603241, + 0.006115770898759365, + 0.006820254493504763, + 0.007110971491783857, + 0.006651800125837326, + 0.004983867052942514, + 0.009171673096716404, + 0.009574005380272865, + 0.009731153957545757, + 0.01121126115322113, + 0.02317415364086628, + 0.011412105523049831, + 0.006004065275192261, + 0.007898552343249321, + 0.016335517168045044, + 0.009826282039284706, + 0.007152933161705732, + 0.007156084757298231, + 0.008201593533158302, + 0.013797147199511528, + 0.00771561311557889, + 0.0089090745896101, + 0.008515369147062302, + 0.01023027766495943, + 0.005895162466913462, + 0.010664772242307663, + 0.008249573409557343, + 0.0080407140776515, + 0.005983847659081221, + 0.007327876519411802, + 0.007393822073936462, + 0.008654683828353882, + 0.0060622477903962135, + 0.014380235224962234, + 0.009463193826377392, + 0.00972489733248949, + 0.008546605706214905, + 0.010502497665584087, + 0.0098439771682024, + 0.0049050115048885345, + 0.004515341017395258, + 0.005149387754499912, + 0.006966910790652037, + 0.006807661149650812, + 0.004446386359632015, + 0.006588763557374477, + 0.005671478342264891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.05500303953886032 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006762356963008642 + ], + "opposite_logprob_tokens": [ + -5.006762504577637 + ], + "logprob_reward": -0.006762356963008642, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006762356963008642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.237155482172966, + 0.15434156358242035, + 0.1085394099354744, + 0.10556195676326752, + 0.06733816862106323, + 0.04384489357471466, + 0.05228423699736595, + 0.07812754809856415, + 0.0311618410050869, + 0.12164483219385147 + ], + "credit_map": [ + 0.027704976499080658, + 0.025771724060177803, + 0.020837871357798576, + 0.024111660197377205, + 0.0559471920132637, + 0.026307111606001854, + 0.030668439343571663, + 0.025806529447436333, + 0.007740093395113945, + 0.009593447670340538, + 0.009175126440823078, + 0.011988403275609016, + 0.015403490513563156, + 0.009545300155878067, + 0.04874200001358986, + 0.042153723537921906, + 0.009945429861545563, + 0.015102020464837551, + 0.01741856336593628, + 0.017477568238973618, + 0.010615668259561062, + 0.011292453855276108, + 0.011260653845965862, + 0.015427065081894398, + 0.016205182299017906, + 0.011586467735469341, + 0.012330315075814724, + 0.009828121401369572, + 0.016754424199461937, + 0.011509846895933151, + 0.014198562130331993, + 0.013149049133062363, + 0.007925711572170258, + 0.006048285402357578, + 0.006943357642740011, + 0.012306317687034607, + 0.0063317446038126945, + 0.008581634610891342, + 0.009462160989642143, + 0.009738960303366184, + 0.0038894733879715204, + 0.003715739119797945, + 0.005240923259407282, + 0.00809422880411148, + 0.007003087550401688, + 0.006155955605208874, + 0.0035212072543799877, + 0.006224281154572964, + 0.009475158527493477, + 0.004965101834386587, + 0.007036617025732994, + 0.007064957171678543, + 0.0063615832477808, + 0.005195505451411009, + 0.00463104760274291, + 0.007554271258413792, + 0.016067851334810257, + 0.006746182683855295, + 0.01652565412223339, + 0.010982492938637733, + 0.005893720779567957, + 0.005826013162732124, + 0.007719893008470535, + 0.0083657531067729, + 0.003641051473096013, + 0.003448593895882368, + 0.005671228282153606, + 0.003916904795914888, + 0.0049808090552687645, + 0.0025374279357492924, + 0.0035958734806627035, + 0.0033699562773108482, + 0.012490044347941875, + 0.010335301980376244, + 0.010944700799882412, + 0.011961305513978004, + 0.017546476796269417, + 0.014724000357091427, + 0.02401781640946865, + 0.019625211134552956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.006762356963008642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811153024435043 + ], + "opposite_logprob_tokens": [ + -3.288111448287964 + ], + "logprob_reward": -0.03811153024435043, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811153024435043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2010316252708435, + 0.10310953855514526, + 0.07362934201955795, + 0.0682210698723793, + 0.04422876238822937, + 0.04898820072412491, + 0.1443653553724289, + 0.06648504734039307, + 0.06919898837804794, + 0.18074209988117218 + ], + "credit_map": [ + 0.027006184682250023, + 0.02655637636780739, + 0.016002653166651726, + 0.01685367152094841, + 0.04538828134536743, + 0.02342790551483631, + 0.027461599558591843, + 0.018334941938519478, + 0.005844220519065857, + 0.007832584902644157, + 0.007006387691944838, + 0.00901489146053791, + 0.009580768644809723, + 0.010552220046520233, + 0.02753794938325882, + 0.02574051357805729, + 0.00607069069519639, + 0.00788169912993908, + 0.01088146585971117, + 0.009761180728673935, + 0.009679163806140423, + 0.006674482952803373, + 0.008443874306976795, + 0.014236778020858765, + 0.006946070585399866, + 0.006841451860964298, + 0.007383068557828665, + 0.005878397263586521, + 0.01262793131172657, + 0.010347881354391575, + 0.007222216110676527, + 0.010974057950079441, + 0.005827638786286116, + 0.004368610680103302, + 0.0053356001153588295, + 0.007269255816936493, + 0.004942631348967552, + 0.004263594746589661, + 0.005717931315302849, + 0.006503502372652292, + 0.00340754771605134, + 0.0037728771567344666, + 0.004452813882380724, + 0.004231397528201342, + 0.013600435107946396, + 0.007120965048670769, + 0.005258628632873297, + 0.007143535651266575, + 0.018746251240372658, + 0.011667047627270222, + 0.017288541421294212, + 0.011739572510123253, + 0.02827201411128044, + 0.019886203110218048, + 0.019278634339571, + 0.017487075179815292, + 0.012245316058397293, + 0.005866196937859058, + 0.012332946993410587, + 0.00803081039339304, + 0.009443597868084908, + 0.00547613063827157, + 0.005939206574112177, + 0.0071508451364934444, + 0.009303230792284012, + 0.006924804300069809, + 0.011889958754181862, + 0.00648818863555789, + 0.011877194978296757, + 0.005818240810185671, + 0.008020389825105667, + 0.00887698121368885, + 0.02044217847287655, + 0.017221691086888313, + 0.020958885550498962, + 0.02945012040436268, + 0.020902244374155998, + 0.019873255863785744, + 0.028516609221696854, + 0.023377109318971634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.03811153024435043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a slow-motion tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall, androgynous models marching in a corridor with a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the models wearing avant-garde garments with shimmering metal or translucent tulle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the walls consist of countless mirrors of different sizes and angles reflecting the models?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth and obsidian black with sharp, dancing shadows cast by the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581606775522232 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581606775522232, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581606775522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15823161602020264, + 0.1594223529100418, + 0.11519600450992584, + 0.12874899804592133, + 0.1651260256767273, + 0.04022890329360962, + 0.060395900160074234, + 0.10050134360790253, + 0.03583911061286926, + 0.036309730261564255 + ], + "credit_map": [ + 0.016627220436930656, + 0.019104544073343277, + 0.01318700984120369, + 0.020120954141020775, + 0.03169441968202591, + 0.015677044168114662, + 0.022603580728173256, + 0.01921684294939041, + 0.014322487637400627, + 0.020257286727428436, + 0.017773333936929703, + 0.01998656801879406, + 0.02753346972167492, + 0.01522339228540659, + 0.020481402054429054, + 0.023844420909881592, + 0.009013031609356403, + 0.011569233611226082, + 0.009907761588692665, + 0.011467153206467628, + 0.01568354107439518, + 0.020133430138230324, + 0.015024490654468536, + 0.022397363558411598, + 0.02000681683421135, + 0.011420433409512043, + 0.01569277234375477, + 0.012468644417822361, + 0.01712174341082573, + 0.016262458637356758, + 0.013166744261980057, + 0.02260938659310341, + 0.01382841169834137, + 0.014191899448633194, + 0.014525717124342918, + 0.012842485681176186, + 0.043998993933200836, + 0.021755404770374298, + 0.02091650292277336, + 0.02306661382317543, + 0.003652275074273348, + 0.00408079894259572, + 0.004671590402722359, + 0.004874940495938063, + 0.007644892670214176, + 0.004193226341158152, + 0.0048044477589428425, + 0.006306730210781097, + 0.009562398307025433, + 0.005619921255856752, + 0.00596868060529232, + 0.007098588161170483, + 0.007646422833204269, + 0.006873898673802614, + 0.009977663867175579, + 0.007648330181837082, + 0.019502975046634674, + 0.008741756901144981, + 0.020008815452456474, + 0.013752920553088188, + 0.008993665687739849, + 0.009089089930057526, + 0.008132756687700748, + 0.012279363349080086, + 0.0037480394821614027, + 0.003614593530073762, + 0.005920469760894775, + 0.0037327143363654613, + 0.006871071178466082, + 0.0033024379517883062, + 0.004713095258921385, + 0.003936687484383583, + 0.004272995516657829, + 0.003757036756724119, + 0.004591843578964472, + 0.00443673599511385, + 0.005628383252769709, + 0.003881675424054265, + 0.005240465514361858, + 0.0045005944557487965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020581606775522232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_tensor([1957, 1957])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_tensor([1957, 1957])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff9a51bb4274208915d8d854c1eb1e6152e9f013 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_tensor([1957, 1957])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, capturing a female warrior clad in a full suit of crimson armor. She stands at the edge of a windswept rocky cliff, a vast, blazing sunset sinking into the distant mountains behind her. The intense backlight casts a dazzling golden rim around her entire form, making the edges of her crimson lacquer armor gleam like molten gold, while the front of the armor is plunged into deep shadow. Her long, raven hair, braided into a thick braid, swirls in the air as she moves, revealing scattered strands. She wields a slender katana, the blade reflecting the last rays of the setting sun with each swing, slicing brilliant arcs. Her sword dance is fluid and powerful: she begins in a low, calm stance, her body like a drawn bow, then swivels swiftly and gracefully, her wide red culottes unfurling, followed by a series of precise and deadly slashes and blocks, each movement seamlessly connected, full of rhythm and control, as if she were engaging in a solemn duel with an invisible opponent. The entire scene is painted in warm and saturated tones, with the sky stained in a gradient of deep purple and fiery orange, creating an epic, majestic and lonely atmosphere.", + "seed": 30413, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_1957_seed_30413_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1667057746089995, + "vlm_logprob": -0.1667057746089995, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.1667057746089995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11719264090061188 + ], + "opposite_logprob_tokens": [ + -2.7421927452087402 + ], + "logprob_reward": -0.11719264090061188, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.11719264090061188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03603368252515793, + 0.03099984861910343, + 0.030739977955818176, + 0.7811733484268188, + 0.02616560459136963, + 0.015701700001955032, + 0.01374632865190506, + 0.01054076012223959, + 0.042392387986183167, + 0.012506406754255295 + ], + "credit_map": [ + 0.003927220590412617, + 0.00867795292288065, + 0.00358585431240499, + 0.004519809037446976, + 0.0034024666529148817, + 0.004447006154805422, + 0.004062817431986332, + 0.0034105577506124973, + 0.00313067389652133, + 0.0029561487026512623, + 0.003023847471922636, + 0.005747250746935606, + 0.003488108515739441, + 0.003180636093020439, + 0.005578326061367989, + 0.0038948592264205217, + 0.005443085916340351, + 0.003957192413508892, + 0.003377563552930951, + 0.0020778595935553312, + 0.004062287509441376, + 0.004298374056816101, + 0.0024570394307374954, + 0.005066573619842529, + 0.08523190021514893, + 0.25751668214797974, + 0.04181318357586861, + 0.027319544926285744, + 0.030872298404574394, + 0.2555297315120697, + 0.020036041736602783, + 0.06285396963357925, + 0.0032372947316616774, + 0.0032405357342213392, + 0.0025397527497261763, + 0.004045434761792421, + 0.003094841493293643, + 0.0030866609886288643, + 0.004737734794616699, + 0.002183350035920739, + 0.00201316736638546, + 0.0016498711192980409, + 0.000841129629407078, + 0.002260133856907487, + 0.0017973465146496892, + 0.0037679902743548155, + 0.002218027366325259, + 0.001154033001512289, + 0.0015768351731821895, + 0.002907129703089595, + 0.0011321769561618567, + 0.002011998789384961, + 0.0017413920722901821, + 0.0018428716575726867, + 0.0016953516751527786, + 0.0008385723922401667, + 0.0015374920330941677, + 0.0011393357999622822, + 0.001886397018097341, + 0.00122383504640311, + 0.0014824139652773738, + 0.0012194541050121188, + 0.0013374097179621458, + 0.0007144225528463721, + 0.013683543540537357, + 0.00570240942761302, + 0.0019444788340479136, + 0.0013496836181730032, + 0.0024301311932504177, + 0.013970017433166504, + 0.0013190747704356909, + 0.001993052661418915, + 0.0017405567923560739, + 0.0017602033913135529, + 0.0012090171221643686, + 0.0014580368297174573, + 0.002439464209601283, + 0.0020035160705447197, + 0.0010836111614480615, + 0.0008120001875795424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.11719264090061188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328253149986267 + ], + "opposite_logprob_tokens": [ + -0.7578253149986267 + ], + "logprob_reward": -0.6328253149986267, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328253149986267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02477944828569889, + 0.07543396204710007, + 0.0124721834436059, + 0.7073301672935486, + 0.11483290046453476, + 0.011876746080815792, + 0.008837352506816387, + 0.013157871551811695, + 0.015405087731778622, + 0.01587427593767643 + ], + "credit_map": [ + 0.0035824787337332964, + 0.004808010067790747, + 0.002990037202835083, + 0.0031467624939978123, + 0.002443843288347125, + 0.0033389015588909388, + 0.002462976146489382, + 0.002006437163800001, + 0.013143657706677914, + 0.008750972338020802, + 0.010305250063538551, + 0.007531291805207729, + 0.006806605961173773, + 0.009911676868796349, + 0.005243283696472645, + 0.01374121569097042, + 0.0014688927913084626, + 0.0021709941793233156, + 0.0016432484844699502, + 0.001143355155363679, + 0.001462975749745965, + 0.0017393047455698252, + 0.0012954238336533308, + 0.001547987456433475, + 0.07793129980564117, + 0.23559291660785675, + 0.03585660085082054, + 0.02372799441218376, + 0.02891736663877964, + 0.2320163995027542, + 0.01786748878657818, + 0.05542004480957985, + 0.022243304178118706, + 0.01381560880690813, + 0.013732586055994034, + 0.007891277782619, + 0.009221584536135197, + 0.021705105900764465, + 0.007578495889902115, + 0.018644927069544792, + 0.0021345855202525854, + 0.0014388307463377714, + 0.0009050667867995799, + 0.00151633913628757, + 0.0017454508924856782, + 0.0017851060256361961, + 0.001570071210153401, + 0.0007812940166331828, + 0.0010280204005539417, + 0.002051633084192872, + 0.0007074617897160351, + 0.0007621778640896082, + 0.0011426574783399701, + 0.0011344879167154431, + 0.001496638054959476, + 0.000514275161549449, + 0.0023091172333806753, + 0.0017454385524615645, + 0.0016917933244258165, + 0.0014563336735591292, + 0.0014556670794263482, + 0.0011927930172532797, + 0.0025200005620718002, + 0.0007867257227189839, + 0.003457249142229557, + 0.003505319356918335, + 0.0011636924464255571, + 0.0010445634834468365, + 0.0012262605596333742, + 0.003279732307419181, + 0.0008759446791373193, + 0.0008523249998688698, + 0.002222332637757063, + 0.0026778930332511663, + 0.0013433576095849276, + 0.0016217882512137294, + 0.0034418110735714436, + 0.0022682459093630314, + 0.0014339765766635537, + 0.0008648688672110438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.6328253149986267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054975058883428574 + ], + "opposite_logprob_tokens": [ + -2.9299750328063965 + ], + "logprob_reward": -0.054975058883428574, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054975058883428574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035544492304325104, + 0.0414019413292408, + 0.060785889625549316, + 0.5474596619606018, + 0.2529396116733551, + 0.020786819979548454, + 0.013998632319271564, + 0.009132388979196548, + 0.010334816761314869, + 0.007615706883370876 + ], + "credit_map": [ + 0.004911692813038826, + 0.008792462758719921, + 0.0027856575325131416, + 0.0036044178996235132, + 0.0042701964266598225, + 0.004972992464900017, + 0.0031935505103319883, + 0.003013522597029805, + 0.007981178350746632, + 0.005274975206702948, + 0.004320702515542507, + 0.004622085951268673, + 0.004675859585404396, + 0.00505842175334692, + 0.003980237990617752, + 0.005488479044288397, + 0.009121405892074108, + 0.00803059246391058, + 0.008209788240492344, + 0.0034145773388445377, + 0.007928145118057728, + 0.00799909234046936, + 0.00425910297781229, + 0.011823185719549656, + 0.06016764044761658, + 0.18223145604133606, + 0.02809467539191246, + 0.018818441778421402, + 0.02209867723286152, + 0.1798711121082306, + 0.013633042573928833, + 0.04254463315010071, + 0.05210687592625618, + 0.03140624612569809, + 0.030573781579732895, + 0.016740577295422554, + 0.021261164918541908, + 0.04850657656788826, + 0.014162140898406506, + 0.038182225078344345, + 0.002615375677123666, + 0.00203202897682786, + 0.0016973887104541063, + 0.0032056686468422413, + 0.004978243261575699, + 0.002440370386466384, + 0.0027559802401810884, + 0.0010617631487548351, + 0.0019626447465270758, + 0.003359786933287978, + 0.0008941477281041443, + 0.0014416382182389498, + 0.0015808033058419824, + 0.002353041199967265, + 0.0015214028535410762, + 0.0008851676830090582, + 0.0015915855765342712, + 0.0010867509990930557, + 0.0011786911636590958, + 0.0009537591249682009, + 0.0012465149629861116, + 0.0010117653291672468, + 0.0014931216137483716, + 0.0005702000926248729, + 0.002483101561665535, + 0.0016979133943095803, + 0.0007359556620940566, + 0.0008493574569001794, + 0.0009253019816242158, + 0.002359437756240368, + 0.0007349526858888566, + 0.0005487968446686864, + 0.0009865512838587165, + 0.0010157614015042782, + 0.0005897009395994246, + 0.0006908160867169499, + 0.0013331189984455705, + 0.0017310123657807708, + 0.0006264146068133414, + 0.0006423309678211808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.054975058883428574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525640370324254 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.00525640370324254, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525640370324254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049545396119356155, + 0.06882812827825546, + 0.14117567241191864, + 0.4069240689277649, + 0.11429481953382492, + 0.019162043929100037, + 0.03941720351576805, + 0.042369648814201355, + 0.10752062499523163, + 0.010762413963675499 + ], + "credit_map": [ + 0.005137491971254349, + 0.011797930113971233, + 0.005915751680731773, + 0.005310664884746075, + 0.005989413242787123, + 0.006129360292106867, + 0.0047433217987418175, + 0.004521459341049194, + 0.012142656370997429, + 0.00857716053724289, + 0.00802635494619608, + 0.008345306850969791, + 0.0066161006689071655, + 0.008318036794662476, + 0.006119491066783667, + 0.010683019645512104, + 0.018062323331832886, + 0.023544268682599068, + 0.018727876245975494, + 0.00779731897637248, + 0.02500045672059059, + 0.01683996059000492, + 0.010331473313272, + 0.02087198942899704, + 0.045013684779405594, + 0.13455365598201752, + 0.020931927487254143, + 0.014017906039953232, + 0.0171850323677063, + 0.1320083737373352, + 0.010921102948486805, + 0.032292380928993225, + 0.02284780517220497, + 0.0143393250182271, + 0.01461996603757143, + 0.007186602335423231, + 0.009297512471675873, + 0.021638918668031693, + 0.0062161837704479694, + 0.018148506060242653, + 0.0024223814252763987, + 0.002263487782329321, + 0.0019066815730184317, + 0.0027034974191337824, + 0.004189488012343645, + 0.0026406180113554, + 0.0015657612821087241, + 0.0014701291220262647, + 0.004549873527139425, + 0.006957574747502804, + 0.002959255129098892, + 0.005217981059104204, + 0.004765011370182037, + 0.005170612595975399, + 0.00765615189447999, + 0.0021407443564385176, + 0.0059178052470088005, + 0.0062564644031226635, + 0.007412723731249571, + 0.0038685891777276993, + 0.005277471151202917, + 0.004801312927156687, + 0.005476826336234808, + 0.003358454443514347, + 0.037447240203619, + 0.009538287296891212, + 0.00434362655505538, + 0.0026140727568417788, + 0.005348424892872572, + 0.03990934044122696, + 0.0028425450436770916, + 0.0054770829156041145, + 0.0016310029895976186, + 0.001887064310722053, + 0.0006927308277226985, + 0.000980884418822825, + 0.001866954960860312, + 0.0022150001022964716, + 0.00076123425969854, + 0.0007275423849932849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.00525640370324254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279454559087753 + ], + "opposite_logprob_tokens": [ + -3.7732794284820557 + ], + "logprob_reward": -0.023279454559087753, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279454559087753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03589538112282753, + 0.03488868847489357, + 0.06345944106578827, + 0.47910845279693604, + 0.34026822447776794, + 0.004801545757800341, + 0.0049605765379965305, + 0.009034094400703907, + 0.018184242770075798, + 0.009399410337209702 + ], + "credit_map": [ + 0.004789176397025585, + 0.007818557322025299, + 0.004628944210708141, + 0.0038645179010927677, + 0.0034872146788984537, + 0.0045310575515031815, + 0.0028788044583052397, + 0.003897105110809207, + 0.00560449343174696, + 0.00373468897305429, + 0.004383486695587635, + 0.0049893599934875965, + 0.0026801256462931633, + 0.00395906250923872, + 0.003705706913024187, + 0.005831759888678789, + 0.009839797392487526, + 0.007615718059241772, + 0.008411206305027008, + 0.00358909391798079, + 0.008686232380568981, + 0.008797049522399902, + 0.003811978967860341, + 0.012708357535302639, + 0.05157089605927467, + 0.1582772582769394, + 0.025392906740307808, + 0.01649421826004982, + 0.01951337233185768, + 0.15713801980018616, + 0.01290576346218586, + 0.03781599551439285, + 0.06692977249622345, + 0.04391501098871231, + 0.044895097613334656, + 0.021467534825205803, + 0.026616455987095833, + 0.06359124928712845, + 0.017081981524825096, + 0.05577113851904869, + 0.0007376940920948982, + 0.000391920271795243, + 0.0005531708593480289, + 0.000640593352727592, + 0.0006400751881301403, + 0.0009460630244575441, + 0.0005352087318897247, + 0.00035682006273418665, + 0.000656709133181721, + 0.0011226178612560034, + 0.0003628921986091882, + 0.0005119483103044331, + 0.0005892590852454305, + 0.0007393343257717788, + 0.0006383282016031444, + 0.000339487538440153, + 0.0020855257753282785, + 0.0012765279971063137, + 0.0011079912073910236, + 0.0007654690416529775, + 0.0010675230296328664, + 0.001287200953811407, + 0.000825365714263171, + 0.0006184898666106164, + 0.005871082190424204, + 0.0024296194314956665, + 0.000778642890509218, + 0.000613214448094368, + 0.0008462172700092196, + 0.006192449945956469, + 0.0005824511754326522, + 0.0008705627988092601, + 0.0013195869978517294, + 0.0013480109628289938, + 0.000753046537283808, + 0.0008742198697291315, + 0.0015205217059701681, + 0.002024468034505844, + 0.0006194809102453291, + 0.0009400751441717148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.023279454559087753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, capturing a female warrior clad in a full suit of crimson armor. She stands at the edge of a windswept rocky cliff, a vast, blazing sunset sinking into the distant mountains behind her. The intense backlight casts a dazzling golden rim around her entire form, making the edges of her crimson lacquer armor gleam like molten gold, while the front of the armor is plunged into deep shadow. Her long, raven hair, braided into a thick braid, swirls in the air as she moves, revealing scattered strands. She wields a slender katana, the blade reflecting the last rays of the setting sun with each swing, slicing brilliant arcs. Her sword dance is fluid and powerful: she begins in a low, calm stance, her body like a drawn bow, then swivels swiftly and gracefully, her wide red culottes unfurling, followed by a series of precise and deadly slashes and blocks, each movement seamlessly connected, full of rhythm and control, as if she were engaging in a solemn duel with an invisible opponent. The entire scene is painted in warm and saturated tones, with the sky stained in a gradient of deep purple and fiery orange, creating an epic, majestic and lonely atmosphere.", + "seed": 30414, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_1957_seed_30414_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06483065714128315, + "vlm_logprob": -0.06483065714128315, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.06483065714128315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13050296902656555 + ], + "opposite_logprob_tokens": [ + -2.755502939224243 + ], + "logprob_reward": -0.13050296902656555, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.13050296902656555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0699775367975235, + 0.03041907772421837, + 0.03180055320262909, + 0.23030395805835724, + 0.07331226766109467, + 0.38931721448898315, + 0.042508091777563095, + 0.022310175001621246, + 0.051822513341903687, + 0.058228522539138794 + ], + "credit_map": [ + 0.012119656428694725, + 0.007539683021605015, + 0.013290457427501678, + 0.007742455694824457, + 0.004682897590100765, + 0.007557517383247614, + 0.007954646833240986, + 0.009090235456824303, + 0.003829916473478079, + 0.002625198569148779, + 0.006335489917546511, + 0.004438558593392372, + 0.0020321037154644728, + 0.0031758672557771206, + 0.004185881931334734, + 0.003796063829213381, + 0.004660330712795258, + 0.006182048004120588, + 0.0032920464873313904, + 0.004823377821594477, + 0.002704397076740861, + 0.003576826537027955, + 0.0039308564737439156, + 0.0026306724175810814, + 0.06313259154558182, + 0.013117064721882343, + 0.02220103144645691, + 0.010879947803914547, + 0.011915776878595352, + 0.009497343562543392, + 0.03118312917649746, + 0.06837709248065948, + 0.010433215647935867, + 0.0034552859142422676, + 0.006347500253468752, + 0.0053785708732903, + 0.0032717897556722164, + 0.005064667668193579, + 0.01718197949230671, + 0.022179266437888145, + 0.11246418952941895, + 0.01907411962747574, + 0.03626852110028267, + 0.025922859087586403, + 0.08781688660383224, + 0.023302119225263596, + 0.026631304994225502, + 0.057837266474962234, + 0.0089589087292552, + 0.0032606685999780893, + 0.0029919478110969067, + 0.003918518777936697, + 0.006129786837846041, + 0.008854750543832779, + 0.004799394868314266, + 0.0035941200330853462, + 0.0030095090623944998, + 0.002137405565008521, + 0.0034823690075427294, + 0.002247368451207876, + 0.0024405657313764095, + 0.0027568016666918993, + 0.0025298853870481253, + 0.003706274088472128, + 0.005231352988630533, + 0.0031658397056162357, + 0.01051473617553711, + 0.004310863558202982, + 0.004502071533352137, + 0.007712335791438818, + 0.007692479994148016, + 0.008692840114235878, + 0.009671919047832489, + 0.0040707066655159, + 0.005761052016168833, + 0.005475387442857027, + 0.004312078468501568, + 0.017959648743271828, + 0.004610313102602959, + 0.006367422174662352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.13050296902656555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282391965389252 + ], + "opposite_logprob_tokens": [ + -2.0178239345550537 + ], + "logprob_reward": -0.14282391965389252, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282391965389252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05975302308797836, + 0.04092942923307419, + 0.042853742837905884, + 0.28046974539756775, + 0.06760014593601227, + 0.2621568441390991, + 0.06596273928880692, + 0.016357921063899994, + 0.11379773914813995, + 0.05011869594454765 + ], + "credit_map": [ + 0.006652006879448891, + 0.006946645677089691, + 0.00870831310749054, + 0.00784330815076828, + 0.0037968431133776903, + 0.004794166423380375, + 0.008866067044436932, + 0.012145670130848885, + 0.004592372570186853, + 0.004728465806692839, + 0.006287990137934685, + 0.006579862907528877, + 0.0033604903146624565, + 0.004316722042858601, + 0.006201296579092741, + 0.004862228408455849, + 0.005259274039417505, + 0.006202824413776398, + 0.005673701874911785, + 0.008322327397763729, + 0.0030571636743843555, + 0.0033059476409107447, + 0.00574411079287529, + 0.005288395565003157, + 0.07805536687374115, + 0.01741958223283291, + 0.025879032909870148, + 0.016035234555602074, + 0.01553183514624834, + 0.01184119749814272, + 0.036774035543203354, + 0.07893344759941101, + 0.014218754135072231, + 0.0029833605512976646, + 0.00533743342384696, + 0.0041156611405313015, + 0.01213054545223713, + 0.004267320968210697, + 0.013106596656143665, + 0.011440475471317768, + 0.0734647735953331, + 0.012640384025871754, + 0.023959828540682793, + 0.018382424488663673, + 0.04829256236553192, + 0.015201072208583355, + 0.019274218007922173, + 0.05094157159328461, + 0.01353360339999199, + 0.006450537126511335, + 0.005660575348883867, + 0.00579424574971199, + 0.007312098052352667, + 0.01251121237874031, + 0.00682624988257885, + 0.007874219678342342, + 0.0016381588065996766, + 0.001441653585061431, + 0.002026557456701994, + 0.0024162682238966227, + 0.0024188836105167866, + 0.001604336779564619, + 0.0020072259940207005, + 0.002804837189614773, + 0.011305213905870914, + 0.0075605399906635284, + 0.027683231979608536, + 0.009456864558160305, + 0.011756295338273048, + 0.015833549201488495, + 0.013910804875195026, + 0.016291240230202675, + 0.0058761900290846825, + 0.003378227585926652, + 0.004716968629509211, + 0.005624308250844479, + 0.004484397359192371, + 0.014463577419519424, + 0.005967373494058847, + 0.005607650149613619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.14282391965389252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018205681815743446 + ], + "opposite_logprob_tokens": [ + -4.018205642700195 + ], + "logprob_reward": -0.018205681815743446, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018205681815743446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05075295269489288, + 0.03496760502457619, + 0.12083026766777039, + 0.249736487865448, + 0.05899518355727196, + 0.10941899567842484, + 0.06340677291154861, + 0.08289897441864014, + 0.05839390680193901, + 0.17059877514839172 + ], + "credit_map": [ + 0.00634766323491931, + 0.00894328486174345, + 0.004187163896858692, + 0.009946359321475029, + 0.0038025116082280874, + 0.003730462398380041, + 0.008096011355519295, + 0.0056995004415512085, + 0.00387053401209414, + 0.0034384813625365496, + 0.004859591368585825, + 0.0066452957689762115, + 0.003098072251304984, + 0.003326230915263295, + 0.0056286766193807125, + 0.004100725520402193, + 0.013695798814296722, + 0.011139326728880405, + 0.013750672340393066, + 0.02683505415916443, + 0.009823182597756386, + 0.007675976492464542, + 0.023202644661068916, + 0.014707619324326515, + 0.06906682252883911, + 0.015021250583231449, + 0.02359895594418049, + 0.014095347374677658, + 0.0130007928237319, + 0.010163309052586555, + 0.033419255167245865, + 0.07137075811624527, + 0.01576412469148636, + 0.00400202302262187, + 0.005252281669527292, + 0.004461984615772963, + 0.015167174860835075, + 0.004451295360922813, + 0.004625719506293535, + 0.005270585883408785, + 0.031386781483888626, + 0.005396100226789713, + 0.010342016816139221, + 0.00803810078650713, + 0.024143505841493607, + 0.006941231898963451, + 0.007801415864378214, + 0.015369849279522896, + 0.01568496786057949, + 0.004497913643717766, + 0.004316744860261679, + 0.004791312385350466, + 0.006666551344096661, + 0.015818871557712555, + 0.00572166545316577, + 0.00590874906629324, + 0.008838016539812088, + 0.006920638494193554, + 0.011527874507009983, + 0.010530147701501846, + 0.010646190494298935, + 0.009672342799603939, + 0.012036912143230438, + 0.012726857326924801, + 0.007067713886499405, + 0.00339034222997725, + 0.0075029791332781315, + 0.004545132163912058, + 0.006389806512743235, + 0.010219885967671871, + 0.006425080355256796, + 0.012852976098656654, + 0.019768279045820236, + 0.008942232467234135, + 0.01749434322118759, + 0.019783420488238335, + 0.016393940895795822, + 0.05387355759739876, + 0.01880030147731304, + 0.015542713925242424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.018205681815743446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805263504385948 + ], + "opposite_logprob_tokens": [ + -3.5298051834106445 + ], + "logprob_reward": -0.029805263504385948, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805263504385948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10072606056928635, + 0.022122208029031754, + 0.03010333701968193, + 0.6074265837669373, + 0.025622060522437096, + 0.07576461881399155, + 0.018631236627697945, + 0.020676270127296448, + 0.08417633920907974, + 0.014751393347978592 + ], + "credit_map": [ + 0.012467184104025364, + 0.013161815702915192, + 0.016206804662942886, + 0.016622943803668022, + 0.006233738269656897, + 0.009130464866757393, + 0.014421005733311176, + 0.012482093647122383, + 0.002625211840495467, + 0.0030470862984657288, + 0.0032156677916646004, + 0.0031739845871925354, + 0.0019296022364869714, + 0.002053661970421672, + 0.0031005502678453922, + 0.0029764401260763407, + 0.00363959395326674, + 0.005825664848089218, + 0.0025983667001128197, + 0.005675078835338354, + 0.0022750955540686846, + 0.0025832215324044228, + 0.004491216503083706, + 0.0030150930397212505, + 0.17113837599754333, + 0.03741209954023361, + 0.057104866951704025, + 0.032172318547964096, + 0.03301648423075676, + 0.024579668417572975, + 0.08000942319631577, + 0.17199324071407318, + 0.003251333488151431, + 0.0012772271875292063, + 0.0025708863977342844, + 0.003635827684774995, + 0.002945337211713195, + 0.0016185871791094542, + 0.005102074705064297, + 0.005220784340053797, + 0.019998865202069283, + 0.00440051918849349, + 0.006380726583302021, + 0.005067141260951757, + 0.022351562976837158, + 0.006262141279876232, + 0.00498554902151227, + 0.006318101193755865, + 0.0017598735867068172, + 0.002378036268055439, + 0.0018929069628939033, + 0.002152456436306238, + 0.003281790530309081, + 0.0025692235212773085, + 0.0027415938675403595, + 0.0018553525442257524, + 0.002685347106307745, + 0.0023195473477244377, + 0.0028395329136401415, + 0.001876864698715508, + 0.0021974470000714064, + 0.0023065265268087387, + 0.0025726505555212498, + 0.0038783519994467497, + 0.013393731787800789, + 0.007337512914091349, + 0.014101327396929264, + 0.008080903440713882, + 0.008440038189291954, + 0.009382747113704681, + 0.013981330208480358, + 0.009458735585212708, + 0.0015127732185646892, + 0.001200372353196144, + 0.0019320620922371745, + 0.0017138023395091295, + 0.0014006864512339234, + 0.003392969025298953, + 0.00180740375071764, + 0.0017913205083459616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.029805263504385948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815451705828309 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815451705828309, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815451705828309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18288518488407135, + 0.05352792516350746, + 0.10526739805936813, + 0.09793788939714432, + 0.038859155029058456, + 0.1252923458814621, + 0.11327842622995377, + 0.13682495057582855, + 0.08863337337970734, + 0.057493310421705246 + ], + "credit_map": [ + 0.025485070422291756, + 0.021799582988023758, + 0.01941472478210926, + 0.027566183358430862, + 0.014322725124657154, + 0.019792718812823296, + 0.024022100493311882, + 0.030482091009616852, + 0.006178284529596567, + 0.00865493156015873, + 0.005101289600133896, + 0.008610733784735203, + 0.004571738187223673, + 0.005631076171994209, + 0.007950445637106895, + 0.006829423829913139, + 0.009637309238314629, + 0.009815770201385021, + 0.013115840032696724, + 0.0222933292388916, + 0.009963207878172398, + 0.009158882312476635, + 0.019405491650104523, + 0.011877576820552349, + 0.0247882641851902, + 0.006924238987267017, + 0.009624205529689789, + 0.009439782239496708, + 0.007374012842774391, + 0.006606203969568014, + 0.01118460949510336, + 0.021996578201651573, + 0.008342902176082134, + 0.0027116506826132536, + 0.004261199850589037, + 0.0034735167864710093, + 0.0070397742092609406, + 0.0052106259390711784, + 0.004164544399827719, + 0.0036549416836351156, + 0.015304754488170147, + 0.007105284370481968, + 0.012260387651622295, + 0.012502583675086498, + 0.014745152555406094, + 0.011269967071712017, + 0.013905322179198265, + 0.03819888457655907, + 0.02778755873441696, + 0.007705304771661758, + 0.008105534128844738, + 0.007517701014876366, + 0.0073033892549574375, + 0.03467685729265213, + 0.008528069593012333, + 0.011654015630483627, + 0.015251288190484047, + 0.012139721773564816, + 0.02896525338292122, + 0.012155975215137005, + 0.011456325650215149, + 0.012448646128177643, + 0.019992532208561897, + 0.024415208026766777, + 0.014657288789749146, + 0.006985742598772049, + 0.01625295728445053, + 0.006470474414527416, + 0.007432803511619568, + 0.009478087536990643, + 0.011569887399673462, + 0.015786129981279373, + 0.01129113044589758, + 0.005658336915075779, + 0.005205793771892786, + 0.006094043608754873, + 0.00463719479739666, + 0.014233880676329136, + 0.0043311710469424725, + 0.006041758693754673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.002815451705828309 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, capturing a female warrior clad in a full suit of crimson armor. She stands at the edge of a windswept rocky cliff, a vast, blazing sunset sinking into the distant mountains behind her. The intense backlight casts a dazzling golden rim around her entire form, making the edges of her crimson lacquer armor gleam like molten gold, while the front of the armor is plunged into deep shadow. Her long, raven hair, braided into a thick braid, swirls in the air as she moves, revealing scattered strands. She wields a slender katana, the blade reflecting the last rays of the setting sun with each swing, slicing brilliant arcs. Her sword dance is fluid and powerful: she begins in a low, calm stance, her body like a drawn bow, then swivels swiftly and gracefully, her wide red culottes unfurling, followed by a series of precise and deadly slashes and blocks, each movement seamlessly connected, full of rhythm and control, as if she were engaging in a solemn duel with an invisible opponent. The entire scene is painted in warm and saturated tones, with the sky stained in a gradient of deep purple and fiery orange, creating an epic, majestic and lonely atmosphere.", + "seed": 30415, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_1957_seed_30415_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13356928154826164, + "vlm_logprob": -0.13356928154826164, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.13356928154826164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07238324731588364 + ], + "opposite_logprob_tokens": [ + -3.197383165359497 + ], + "logprob_reward": -0.07238324731588364, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.07238324731588364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1847582459449768, + 0.05318830907344818, + 0.04729503393173218, + 0.10969816148281097, + 0.15277788043022156, + 0.24015216529369354, + 0.04752596840262413, + 0.12199976295232773, + 0.014367149211466312, + 0.02823738567531109 + ], + "credit_map": [ + 0.011110021732747555, + 0.06430381536483765, + 0.0052095637656748295, + 0.004543682094663382, + 0.010622317902743816, + 0.07595023512840271, + 0.009392742067575455, + 0.003625851357355714, + 0.004878148436546326, + 0.00825467985123396, + 0.004874071571975946, + 0.003544541308656335, + 0.00628138380125165, + 0.009803317487239838, + 0.008907560259103775, + 0.006644605193287134, + 0.00470308680087328, + 0.00692923367023468, + 0.004764209967106581, + 0.0030331688467413187, + 0.005874571390450001, + 0.009810451418161392, + 0.006755077745765448, + 0.005425228271633387, + 0.012096717022359371, + 0.013965555466711521, + 0.008764505386352539, + 0.01149204932153225, + 0.011471396312117577, + 0.018665125593543053, + 0.01742326281964779, + 0.015819545835256577, + 0.013041583821177483, + 0.052317988127470016, + 0.003919250797480345, + 0.0038249471690505743, + 0.004978574346750975, + 0.063302181661129, + 0.00306209409609437, + 0.008331245742738247, + 0.013619929552078247, + 0.07565799355506897, + 0.018694542348384857, + 0.007292455993592739, + 0.010033808648586273, + 0.08172747492790222, + 0.008239923976361752, + 0.024886012077331543, + 0.005726356990635395, + 0.005278448108583689, + 0.0032122924458235502, + 0.0033861780539155006, + 0.010317700915038586, + 0.005859320051968098, + 0.009206977672874928, + 0.0045386929996311665, + 0.01620672456920147, + 0.009550582617521286, + 0.0076112281531095505, + 0.010407401248812675, + 0.012559711001813412, + 0.015954282134771347, + 0.03190981224179268, + 0.01780001074075699, + 0.0025449723470956087, + 0.0023767875973135233, + 0.0008751738350838423, + 0.0013135159388184547, + 0.00197458709590137, + 0.002185904188081622, + 0.0015994051937013865, + 0.0014968017349019647, + 0.0046980115585029125, + 0.003465458517894149, + 0.0017384940292686224, + 0.0026509284507483244, + 0.003117397427558899, + 0.005920697934925556, + 0.0040461537428200245, + 0.0026002409867942333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.07238324731588364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233258008956909 + ], + "opposite_logprob_tokens": [ + -0.8983258008956909 + ], + "logprob_reward": -0.5233258008956909, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233258008956909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1326064020395279, + 0.07034848630428314, + 0.03605547919869423, + 0.2173340618610382, + 0.05134464427828789, + 0.27342498302459717, + 0.06692013889551163, + 0.07083729654550552, + 0.056670866906642914, + 0.02445761300623417 + ], + "credit_map": [ + 0.012018699198961258, + 0.02739543281495571, + 0.006252068094909191, + 0.007949369959533215, + 0.015396679751574993, + 0.02521679922938347, + 0.028180915862321854, + 0.010196452960371971, + 0.006875633727759123, + 0.010356583632528782, + 0.008235841989517212, + 0.0048764911480247974, + 0.008247755467891693, + 0.008235077373683453, + 0.015959272161126137, + 0.00756183173507452, + 0.003300060983747244, + 0.0033364396076649427, + 0.003238992067053914, + 0.004997017327696085, + 0.004459704272449017, + 0.004431602079421282, + 0.00882815197110176, + 0.0034635139163583517, + 0.040371306240558624, + 0.030885443091392517, + 0.010851560160517693, + 0.012127718888223171, + 0.015761857852339745, + 0.06745965778827667, + 0.013260608538985252, + 0.026615910232067108, + 0.0038998594973236322, + 0.013003598898649216, + 0.003330767620354891, + 0.0030015537049621344, + 0.0034467990044504404, + 0.0159281138330698, + 0.002930820919573307, + 0.005803134758025408, + 0.015816474333405495, + 0.08523350208997726, + 0.020985905081033707, + 0.008929494768381119, + 0.013378254137933254, + 0.0919092670083046, + 0.010461729019880295, + 0.026710350066423416, + 0.008214225992560387, + 0.008221820928156376, + 0.00523708900436759, + 0.004319950472563505, + 0.017468448728322983, + 0.010844813659787178, + 0.00778726814314723, + 0.004826532211154699, + 0.008511325344443321, + 0.007624360732734203, + 0.005674069747328758, + 0.005579376593232155, + 0.009486217983067036, + 0.011147052980959415, + 0.012191550806164742, + 0.01062334980815649, + 0.00570523040369153, + 0.007910892367362976, + 0.0035192754585295916, + 0.006897172424942255, + 0.009077630005776882, + 0.0060122027061879635, + 0.011984475888311863, + 0.005563993472605944, + 0.0038790996186435223, + 0.002929022768512368, + 0.002273046877235174, + 0.002526849042624235, + 0.00414643669500947, + 0.0029326954390853643, + 0.0030405574943870306, + 0.002729907399043441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.5233258008956909 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033711355179548264 + ], + "opposite_logprob_tokens": [ + -3.4087114334106445 + ], + "logprob_reward": -0.033711355179548264, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033711355179548264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4118155837059021, + 0.07592944800853729, + 0.02702963724732399, + 0.0203105416148901, + 0.11952060461044312, + 0.19585074484348297, + 0.03525405377149582, + 0.05919678881764412, + 0.017131902277469635, + 0.037960778921842575 + ], + "credit_map": [ + 0.022274522110819817, + 0.15265236794948578, + 0.0076415822841227055, + 0.005764903966337442, + 0.020390206947922707, + 0.17533966898918152, + 0.019994767382740974, + 0.007757562678307295, + 0.0048469980247318745, + 0.01925858110189438, + 0.0040852706879377365, + 0.00406848918646574, + 0.007856759242713451, + 0.019632164388895035, + 0.011116676032543182, + 0.005064505618065596, + 0.0019344516331329942, + 0.003511089365929365, + 0.0018500018632039428, + 0.0028656329959630966, + 0.0036885233130306005, + 0.005065965000540018, + 0.006341908592730761, + 0.001772063784301281, + 0.002640367951244116, + 0.002206994453445077, + 0.001679580076597631, + 0.0017807908589020371, + 0.0016208312008529902, + 0.004524833522737026, + 0.0018490732181817293, + 0.004008069634437561, + 0.010401671752333641, + 0.038697488605976105, + 0.0038130497559905052, + 0.003794499207288027, + 0.004587344825267792, + 0.04822301119565964, + 0.004099165555089712, + 0.005904378369450569, + 0.010510976426303387, + 0.06645200401544571, + 0.012613312341272831, + 0.005750731565058231, + 0.007532848510891199, + 0.07096172124147415, + 0.005325963255017996, + 0.016703186556696892, + 0.0045742760412395, + 0.0047512357123196125, + 0.002923527965322137, + 0.00210550962947309, + 0.008677629753947258, + 0.0056650228798389435, + 0.0042610084637999535, + 0.002295840997248888, + 0.0068344068713486195, + 0.004829278215765953, + 0.0037411446683108807, + 0.005867753643542528, + 0.006223461125046015, + 0.006811217870563269, + 0.016754211857914925, + 0.008135316893458366, + 0.0023657840210944414, + 0.0019707249011844397, + 0.001119502354413271, + 0.002379691693931818, + 0.0020750316325575113, + 0.002170887077227235, + 0.003177977865561843, + 0.0018723024986684322, + 0.004606541711837053, + 0.005677966866642237, + 0.0029885258991271257, + 0.0034768711775541306, + 0.006621453911066055, + 0.004657060839235783, + 0.006939716171473265, + 0.002992641879245639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.033711355179548264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640758693218231 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640758693218231, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640758693218231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22122128307819366, + 0.028043493628501892, + 0.08573342859745026, + 0.03410287946462631, + 0.03512922301888466, + 0.32595375180244446, + 0.08757247775793076, + 0.08039305359125137, + 0.07206830382347107, + 0.029782159253954887 + ], + "credit_map": [ + 0.011192016303539276, + 0.08843740075826645, + 0.0029839135240763426, + 0.004938057158142328, + 0.008019263856112957, + 0.09257997572422028, + 0.006648406386375427, + 0.006422234233468771, + 0.0020181608851999044, + 0.0060783568769693375, + 0.0019223925191909075, + 0.0017888547154143453, + 0.002826306037604809, + 0.006398964673280716, + 0.004534279927611351, + 0.0024761788081377745, + 0.00406656926497817, + 0.0292483139783144, + 0.004823107738047838, + 0.004067913629114628, + 0.0032059764489531517, + 0.030468827113509178, + 0.006437299307435751, + 0.003415424609556794, + 0.005517573561519384, + 0.0045542060397565365, + 0.0024957717396318913, + 0.0024884312879294157, + 0.001924533280543983, + 0.007312797009944916, + 0.002796715358272195, + 0.0070128487423062325, + 0.0029143299907445908, + 0.009288731962442398, + 0.0024972271639853716, + 0.002023363020271063, + 0.0026546025183051825, + 0.009280581958591938, + 0.00257296534255147, + 0.0038974189665168524, + 0.01727423444390297, + 0.11011816561222076, + 0.02055596560239792, + 0.009046552702784538, + 0.014412847347557545, + 0.11700702458620071, + 0.010188287124037743, + 0.027350660413503647, + 0.01042866986244917, + 0.00988014880567789, + 0.009366448037326336, + 0.005881073884665966, + 0.01800832897424698, + 0.01296963170170784, + 0.01166102010756731, + 0.009377150796353817, + 0.009580323472619057, + 0.007098297588527203, + 0.0034207263961434364, + 0.009098337963223457, + 0.010750376619398594, + 0.007997962646186352, + 0.024110887199640274, + 0.008336140774190426, + 0.012563110329210758, + 0.008344225585460663, + 0.005419797729700804, + 0.008944211527705193, + 0.007690646219998598, + 0.009881888516247272, + 0.010623864829540253, + 0.008600556291639805, + 0.003453587181866169, + 0.003981454763561487, + 0.002109312918037176, + 0.002856258302927017, + 0.006229350809007883, + 0.0034537126775830984, + 0.005134392995387316, + 0.0025640882086008787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.008640758693218231 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029785245656967163 + ], + "opposite_logprob_tokens": [ + -3.52978515625 + ], + "logprob_reward": -0.029785245656967163, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029785245656967163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05605505034327507, + 0.031233523041009903, + 0.061870697885751724, + 0.2647598385810852, + 0.36893972754478455, + 0.10093497484922409, + 0.036201369017362595, + 0.04772266373038292, + 0.009931440465152264, + 0.02235075831413269 + ], + "credit_map": [ + 0.0073515367694199085, + 0.00986744835972786, + 0.005605895537883043, + 0.004097572062164545, + 0.00644489424303174, + 0.010784047655761242, + 0.00655605411157012, + 0.00534760020673275, + 0.0025716666132211685, + 0.004077430348843336, + 0.003974956460297108, + 0.00262640044093132, + 0.0033150524832308292, + 0.006027784198522568, + 0.0039766402915120125, + 0.00466358894482255, + 0.0031948359683156013, + 0.022099586203694344, + 0.0025776007678359747, + 0.001838021562434733, + 0.002594023710116744, + 0.02472538687288761, + 0.002257065149024129, + 0.0025841728784143925, + 0.04564549773931503, + 0.03776372969150543, + 0.01592022366821766, + 0.014240505173802376, + 0.01817503571510315, + 0.07113524526357651, + 0.015512173995375633, + 0.04636739566922188, + 0.03152380883693695, + 0.11605663597583771, + 0.01755049079656601, + 0.010712658055126667, + 0.012398365885019302, + 0.14621537923812866, + 0.010719694197177887, + 0.023762688040733337, + 0.007191997952759266, + 0.021309848874807358, + 0.01198725774884224, + 0.005299688316881657, + 0.007798403967171907, + 0.024435603991150856, + 0.006586404051631689, + 0.01632576994597912, + 0.0055036768317222595, + 0.0040767802856862545, + 0.003570373635739088, + 0.003124805400148034, + 0.0033986149355769157, + 0.004887667950242758, + 0.0031811220105737448, + 0.008458325639367104, + 0.006922577973455191, + 0.004074038006365299, + 0.003186022164300084, + 0.004730659537017345, + 0.004821251146495342, + 0.007268406916409731, + 0.01113737840205431, + 0.005582325626164675, + 0.0011522360146045685, + 0.0014984760200604796, + 0.0009977760491892695, + 0.0008946175221353769, + 0.0011050888570025563, + 0.001181736239232123, + 0.0021560871973633766, + 0.0009454214014112949, + 0.003443841589614749, + 0.002649979665875435, + 0.0018759086960926652, + 0.0022049471735954285, + 0.0037332014180719852, + 0.0030016808304935694, + 0.003504241816699505, + 0.001936954795382917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.029785245656967163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, capturing a female warrior clad in a full suit of crimson armor. She stands at the edge of a windswept rocky cliff, a vast, blazing sunset sinking into the distant mountains behind her. The intense backlight casts a dazzling golden rim around her entire form, making the edges of her crimson lacquer armor gleam like molten gold, while the front of the armor is plunged into deep shadow. Her long, raven hair, braided into a thick braid, swirls in the air as she moves, revealing scattered strands. She wields a slender katana, the blade reflecting the last rays of the setting sun with each swing, slicing brilliant arcs. Her sword dance is fluid and powerful: she begins in a low, calm stance, her body like a drawn bow, then swivels swiftly and gracefully, her wide red culottes unfurling, followed by a series of precise and deadly slashes and blocks, each movement seamlessly connected, full of rhythm and control, as if she were engaging in a solemn duel with an invisible opponent. The entire scene is painted in warm and saturated tones, with the sky stained in a gradient of deep purple and fiery orange, creating an epic, majestic and lonely atmosphere.", + "seed": 30416, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_3/index_1957_seed_30416_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1294535337947309, + "vlm_logprob": -0.1294535337947309, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.1294535337947309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08605960756540298 + ], + "opposite_logprob_tokens": [ + -3.0860595703125 + ], + "logprob_reward": -0.08605960756540298, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.08605960756540298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2724073529243469, + 0.10366082936525345, + 0.11586635559797287, + 0.09275685995817184, + 0.05542683228850365, + 0.06778807193040848, + 0.07746792584657669, + 0.1310955435037613, + 0.048444073647260666, + 0.035086195915937424 + ], + "credit_map": [ + 0.04190570116043091, + 0.04172462224960327, + 0.025230275467038155, + 0.03268963098526001, + 0.04438085854053497, + 0.04044216126203537, + 0.025364961475133896, + 0.020669136196374893, + 0.018588360399007797, + 0.00692600617185235, + 0.00957636721432209, + 0.017149174585938454, + 0.00899474322795868, + 0.01795058697462082, + 0.015863558277487755, + 0.008612026460468769, + 0.022887125611305237, + 0.014639780856668949, + 0.009325025603175163, + 0.015107299201190472, + 0.008752014487981796, + 0.0253077931702137, + 0.011710242368280888, + 0.008137064054608345, + 0.0142403244972229, + 0.014990250580012798, + 0.010177107527852058, + 0.009446959011256695, + 0.009880036115646362, + 0.012361124157905579, + 0.013157614506781101, + 0.00850343145430088, + 0.0073427679017186165, + 0.007018240634351969, + 0.004875154234468937, + 0.006053337827324867, + 0.008218073286116123, + 0.006059153936803341, + 0.009788029827177525, + 0.006072071380913258, + 0.01226793508976698, + 0.00731702009215951, + 0.0045026568695902824, + 0.005060853902250528, + 0.006167765706777573, + 0.01834871619939804, + 0.003267440479248762, + 0.010855679400265217, + 0.008038251660764217, + 0.010612361133098602, + 0.008264760486781597, + 0.013732637278735638, + 0.008666010573506355, + 0.010792791843414307, + 0.008618452586233616, + 0.008742651902139187, + 0.01948891207575798, + 0.0156776811927557, + 0.014239182695746422, + 0.014664587564766407, + 0.014931833371520042, + 0.023677416145801544, + 0.010659018531441689, + 0.017756912857294083, + 0.005923360586166382, + 0.005223394371569157, + 0.006078343838453293, + 0.010420656763017178, + 0.004608409944921732, + 0.006562450900673866, + 0.005450102966278791, + 0.00417734868824482, + 0.00442236103117466, + 0.0040999422781169415, + 0.004744553007185459, + 0.004259342793375254, + 0.004090858157724142, + 0.005245613399893045, + 0.005007569212466478, + 0.0032159537076950073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.08605960756540298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233311653137207 + ], + "opposite_logprob_tokens": [ + -0.8983311653137207 + ], + "logprob_reward": -0.5233311653137207, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233311653137207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09938226640224457, + 0.07626142352819443, + 0.12652117013931274, + 0.12591350078582764, + 0.05983555316925049, + 0.05204717442393303, + 0.10808482766151428, + 0.22859273850917816, + 0.09429866820573807, + 0.029062598943710327 + ], + "credit_map": [ + 0.01209369208663702, + 0.011786325834691525, + 0.011610344983637333, + 0.013796661049127579, + 0.01669004186987877, + 0.009583726525306702, + 0.013841995969414711, + 0.009979486465454102, + 0.0147033566609025, + 0.008427065797150135, + 0.006943524349480867, + 0.008718117140233517, + 0.0106923533603549, + 0.01299749594181776, + 0.009251875802874565, + 0.004527647513896227, + 0.01872718334197998, + 0.020979145541787148, + 0.012411095201969147, + 0.02021668292582035, + 0.01084419246762991, + 0.018209639936685562, + 0.013903998769819736, + 0.011229232884943485, + 0.015841424465179443, + 0.014447354711592197, + 0.01057955902069807, + 0.02579646371304989, + 0.013941318728029728, + 0.0156503077596426, + 0.01794576272368431, + 0.01171132829040289, + 0.007524705957621336, + 0.004485019017010927, + 0.004502345807850361, + 0.010020769201219082, + 0.004911844152957201, + 0.011352422647178173, + 0.012273535132408142, + 0.004764915909618139, + 0.008166348561644554, + 0.0048339711502194405, + 0.0071025448851287365, + 0.009541087783873081, + 0.005680771078914404, + 0.007836681790649891, + 0.004365756642073393, + 0.004520016722381115, + 0.011681715957820415, + 0.013859586790204048, + 0.009324967861175537, + 0.02216486446559429, + 0.010585299693048, + 0.01658620499074459, + 0.01129783783107996, + 0.012584361247718334, + 0.03408673405647278, + 0.02752341516315937, + 0.027151817455887794, + 0.03656301647424698, + 0.02083643525838852, + 0.035720039159059525, + 0.015324114821851254, + 0.031387194991111755, + 0.01328201498836279, + 0.008672921918332577, + 0.012645632959902287, + 0.016485171392560005, + 0.010629751719534397, + 0.012919687665998936, + 0.010974382050335407, + 0.008689112029969692, + 0.0029055315535515547, + 0.00442713126540184, + 0.0028889060486108065, + 0.006072752643376589, + 0.003020733594894409, + 0.0035988555755466223, + 0.0038672161754220724, + 0.0022814758121967316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.5233311653137207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298904299736023 + ], + "opposite_logprob_tokens": [ + -3.773298978805542 + ], + "logprob_reward": -0.023298904299736023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298904299736023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12127137929201126, + 0.06692053377628326, + 0.16593791544437408, + 0.07379013299942017, + 0.06307937949895859, + 0.12375334650278091, + 0.1411275565624237, + 0.1637989580631256, + 0.04498568922281265, + 0.035335052758455276 + ], + "credit_map": [ + 0.016256697475910187, + 0.016871223226189613, + 0.012881739996373653, + 0.021060766652226448, + 0.01597706414759159, + 0.015358533710241318, + 0.014167712070047855, + 0.008697646670043468, + 0.012836920097470284, + 0.0056359716691076756, + 0.007179123815149069, + 0.01084411982446909, + 0.005549546796828508, + 0.013444790616631508, + 0.007012245710939169, + 0.0044178105890750885, + 0.04777422547340393, + 0.008796805515885353, + 0.016283296048641205, + 0.01260759774595499, + 0.008560552261769772, + 0.0493600033223629, + 0.00952413771301508, + 0.01303130853921175, + 0.008779308758676052, + 0.014547695405781269, + 0.007951210252940655, + 0.010298890992999077, + 0.008222970180213451, + 0.008433000184595585, + 0.009998563677072525, + 0.005558486096560955, + 0.00812842883169651, + 0.0057844799011945724, + 0.006683902814984322, + 0.00902466755360365, + 0.004736048169434071, + 0.008828292600810528, + 0.01176070049405098, + 0.008132859133183956, + 0.01901664212346077, + 0.011379161849617958, + 0.011745459400117397, + 0.008449445478618145, + 0.010602159425616264, + 0.03842949494719505, + 0.004709458444267511, + 0.019421515986323357, + 0.019842129200696945, + 0.026171289384365082, + 0.02029627189040184, + 0.02512168325483799, + 0.01323268935084343, + 0.010606633499264717, + 0.01716754585504532, + 0.008689315058290958, + 0.04559898003935814, + 0.01948946714401245, + 0.011708168312907219, + 0.01316545344889164, + 0.013109643012285233, + 0.03969490900635719, + 0.008490776643157005, + 0.012541581876575947, + 0.005799163598567247, + 0.005394418723881245, + 0.00647337269037962, + 0.007812031544744968, + 0.005483613815158606, + 0.0042601460590958595, + 0.005602158140391111, + 0.004160786513239145, + 0.004166021011769772, + 0.005161674693226814, + 0.00347241316922009, + 0.0066612581722438335, + 0.004006552509963512, + 0.004728603642433882, + 0.004708355758339167, + 0.002430176129564643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.023298904299736023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949051585048437 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949051585048437, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949051585048437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08680994063615799, + 0.08566072583198547, + 0.15327946841716766, + 0.08396398276090622, + 0.07354177534580231, + 0.10214248299598694, + 0.05780626833438873, + 0.2950708270072937, + 0.04026642069220543, + 0.021458076313138008 + ], + "credit_map": [ + 0.015027681365609169, + 0.009709990583360195, + 0.00779367471113801, + 0.00831235758960247, + 0.01079254224896431, + 0.02024264633655548, + 0.008306304924190044, + 0.006624745205044746, + 0.015348654240369797, + 0.008168262429535389, + 0.007794544100761414, + 0.012874042615294456, + 0.010554603300988674, + 0.011585604399442673, + 0.01056395098567009, + 0.008771070279181004, + 0.040627509355545044, + 0.014578609727323055, + 0.01035264041274786, + 0.012008026242256165, + 0.01100679486989975, + 0.046683426946401596, + 0.008348478935658932, + 0.00967397540807724, + 0.010505298152565956, + 0.014809129759669304, + 0.007347702980041504, + 0.01036517322063446, + 0.00796656683087349, + 0.012102196924388409, + 0.011934202164411545, + 0.008933713659644127, + 0.010550830513238907, + 0.008924037218093872, + 0.0061111669056117535, + 0.012704554945230484, + 0.0070499638095498085, + 0.009066890925168991, + 0.008292587473988533, + 0.010841744020581245, + 0.020423805341124535, + 0.01181146688759327, + 0.006433449685573578, + 0.00884451624006033, + 0.007861318066716194, + 0.02561126835644245, + 0.0058128489181399345, + 0.015343804843723774, + 0.006435661111027002, + 0.00706465682014823, + 0.005626195576041937, + 0.012678975239396095, + 0.005510041024535894, + 0.00783354602754116, + 0.007042353041470051, + 0.005614840891212225, + 0.061411045491695404, + 0.02196010760962963, + 0.03274096921086311, + 0.03460019826889038, + 0.020382044836878777, + 0.09124059230089188, + 0.016840189695358276, + 0.015895672142505646, + 0.005449558608233929, + 0.005444907583296299, + 0.005380619782954454, + 0.005741183180361986, + 0.004310799762606621, + 0.004866943694651127, + 0.0053111482411623, + 0.0037612614687532187, + 0.00410360237583518, + 0.002807766664773226, + 0.0019535927567631006, + 0.0028635861817747355, + 0.0018775928765535355, + 0.004184532444924116, + 0.002247437136247754, + 0.001419965410605073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.005949051585048437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle in a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a full suit of crimson armor with a thick braid of raven hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the warrior performing a fluid sword dance with a katana, including a low stance, swivels, and precise slashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a windswept rocky cliff with a vast sunset over distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00862894020974636 + ], + "opposite_logprob_tokens": [ + -4.758628845214844 + ], + "logprob_reward": -0.00862894020974636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00862894020974636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16296499967575073, + 0.05522996932268143, + 0.1847606748342514, + 0.042293813079595566, + 0.03992639109492302, + 0.08644049614667892, + 0.05157856270670891, + 0.3418740928173065, + 0.016493601724505424, + 0.018437474966049194 + ], + "credit_map": [ + 0.02344898134469986, + 0.02031571976840496, + 0.015850862488150597, + 0.017657632008194923, + 0.022974351420998573, + 0.026505419984459877, + 0.01937251165509224, + 0.01683950051665306, + 0.008839800022542477, + 0.006516835652291775, + 0.004832889419049025, + 0.00698894914239645, + 0.005901090335100889, + 0.007958873175084591, + 0.007427326403558254, + 0.0067642005160450935, + 0.02319263480603695, + 0.02339751459658146, + 0.023316361010074615, + 0.02028520405292511, + 0.0128516536206007, + 0.03376846760511398, + 0.018383724614977837, + 0.02956509031355381, + 0.007012669462710619, + 0.004971904680132866, + 0.006078703328967094, + 0.004174894653260708, + 0.003462444292381406, + 0.006711376830935478, + 0.004955944139510393, + 0.004925870802253485, + 0.0056925248354673386, + 0.004569694399833679, + 0.0035609749611467123, + 0.005820718128234148, + 0.0031804582104086876, + 0.006878625135868788, + 0.00497458828613162, + 0.0052488017827272415, + 0.011471150442957878, + 0.008259364403784275, + 0.008525462821125984, + 0.008771940134465694, + 0.007401236332952976, + 0.02324816584587097, + 0.005166675895452499, + 0.013596485368907452, + 0.005391011014580727, + 0.007051426451653242, + 0.004702884703874588, + 0.012766465544700623, + 0.005577827338129282, + 0.005576011259108782, + 0.005566346924751997, + 0.004946578294038773, + 0.12668271362781525, + 0.0171409510076046, + 0.011187594383955002, + 0.017008507624268532, + 0.01209338754415512, + 0.12961474061012268, + 0.010440574027597904, + 0.01770557463169098, + 0.0029325003270059824, + 0.0015670310240238905, + 0.001946526113897562, + 0.0029206438921391964, + 0.0012834841618314385, + 0.002833129372447729, + 0.0017168036429211497, + 0.0012934831902384758, + 0.002364962827414274, + 0.0024822347331792116, + 0.0022301780991256237, + 0.001963441027328372, + 0.0019879909232258797, + 0.002758115530014038, + 0.0029154843650758266, + 0.0017350655980408192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.00862894020974636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_4/index_tensor([1745, 1745])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_4/index_tensor([1745, 1745])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d1d6c2cc69fa6aab41f8948619d7d981fce26db8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_4/index_tensor([1745, 1745])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, the camera gently and fluidly tilting downward to follow the subject's movements, captures a vibrant orange tabby cat taking a swift and graceful leap from the top of a towering, dark wood bookshelf. In a warm study bathed in soft afternoon sunlight, light slanted in through a side window, illuminating tiny dust particles floating in the air and outlining a golden rim across the cat's short, dense fur. With a sudden kick of its hind legs, its body stretched out in mid-air, a fluid orange arc. Its emerald green eyes fixed intently on its target below, its tail balancing steadily behind it. In its leap, its hind paws inevitably swept across several misaligned books on the edge of the shelf, causing the heavy hardcovers to sway violently back and forth. One volume even slipped halfway out, its pages scattering in mid-air, as if poised to fall at any moment, adding a sense of chaotic tension to the graceful leap. The whole picture is painted in warm tones, juxtaposing the agility of cats with the tranquility of the home environment, full of dynamic beauty and storytelling.", + "seed": 12472, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_4/index_1745_seed_12472_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.30440432876348494, + "vlm_logprob": -0.30440432876348494, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.30440432876348494, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016258155927062035 + ], + "opposite_logprob_tokens": [ + -4.766258239746094 + ], + "logprob_reward": -0.016258155927062035, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.016258155927062035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05432945862412453, + 0.27389290928840637, + 0.054393526166677475, + 0.06017686799168587, + 0.028793688863515854, + 0.3844786584377289, + 0.026762159541249275, + 0.05810442194342613, + 0.026386793702840805, + 0.03268157318234444 + ], + "credit_map": [ + 0.004257152322679758, + 0.004515470005571842, + 0.009242590516805649, + 0.007182762026786804, + 0.0052345022559165955, + 0.008358747698366642, + 0.007685480173677206, + 0.007852754555642605, + 0.10802973806858063, + 0.017979372292757034, + 0.1107831820845604, + 0.010615979321300983, + 0.005691453814506531, + 0.006306458730250597, + 0.007132887374609709, + 0.00735383341088891, + 0.005167514085769653, + 0.006131935864686966, + 0.005352772306650877, + 0.007544012740254402, + 0.008383417502045631, + 0.009193578734993935, + 0.007324370089918375, + 0.005295927170664072, + 0.009320041164755821, + 0.0071184346452355385, + 0.005838560406118631, + 0.007153482176363468, + 0.008647394366562366, + 0.006908492185175419, + 0.006469917483627796, + 0.00872054323554039, + 0.0033817461226135492, + 0.004413682501763105, + 0.003228398272767663, + 0.003484079847112298, + 0.005330533254891634, + 0.0022405057679861784, + 0.003113859798759222, + 0.003600883996114135, + 0.023161202669143677, + 0.019225411117076874, + 0.14406980574131012, + 0.026055915281176567, + 0.01047985628247261, + 0.027426300570368767, + 0.011814948171377182, + 0.12224520742893219, + 0.002964031882584095, + 0.004333266522735357, + 0.004974002949893475, + 0.0035354583524167538, + 0.004508188460022211, + 0.0016905503580346704, + 0.0021013710647821426, + 0.0026552879717200994, + 0.006941383238881826, + 0.005055651534348726, + 0.013892591930925846, + 0.0039716181345283985, + 0.006804474163800478, + 0.004559364169836044, + 0.007090603932738304, + 0.009788734838366508, + 0.003060905961319804, + 0.0033019983675330877, + 0.003505968488752842, + 0.002075127325952053, + 0.004281066358089447, + 0.0033866818994283676, + 0.0034545615781098604, + 0.0033204841893166304, + 0.002726186066865921, + 0.005243074614554644, + 0.0026015317998826504, + 0.007415219210088253, + 0.003791258903220296, + 0.0044967420399188995, + 0.0032139720860868692, + 0.003193589625880122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.016258155927062035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635768987238407 + ], + "opposite_logprob_tokens": [ + -3.651357650756836 + ], + "logprob_reward": -0.02635768987238407, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635768987238407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07325059175491333, + 0.12580180168151855, + 0.19263112545013428, + 0.054122209548950195, + 0.09059298038482666, + 0.05724431946873665, + 0.07112043350934982, + 0.13751013576984406, + 0.08715052157640457, + 0.11057586967945099 + ], + "credit_map": [ + 0.006444886792451143, + 0.006446796003729105, + 0.01186992134898901, + 0.010016996413469315, + 0.006778262555599213, + 0.009995467029511929, + 0.012815273366868496, + 0.008882978931069374, + 0.04127885028719902, + 0.007929211482405663, + 0.04344527795910835, + 0.007075674366205931, + 0.006933452095836401, + 0.004969942849129438, + 0.007453810423612595, + 0.006715584080666304, + 0.020516566932201385, + 0.024523582309484482, + 0.020130455493927002, + 0.031547147780656815, + 0.026032108813524246, + 0.02866137959063053, + 0.02536703273653984, + 0.015852859243750572, + 0.007591458037495613, + 0.00743189686909318, + 0.00444921525195241, + 0.00777097325772047, + 0.008743797428905964, + 0.007488604169338942, + 0.005138951353728771, + 0.005507312715053558, + 0.005944803822785616, + 0.008057262748479843, + 0.021722346544265747, + 0.00952569954097271, + 0.020923467352986336, + 0.007765255868434906, + 0.008213573135435581, + 0.008440573699772358, + 0.0050368704833090305, + 0.006130490452051163, + 0.009061831049621105, + 0.009100062772631645, + 0.005999717861413956, + 0.009491831064224243, + 0.005001474637538195, + 0.007422039285302162, + 0.006275696214288473, + 0.009748060256242752, + 0.016692735254764557, + 0.008756043389439583, + 0.012441419064998627, + 0.0040581123903393745, + 0.006385223474353552, + 0.0067631397396326065, + 0.014362337067723274, + 0.020434895530343056, + 0.024558689445257187, + 0.011514556594192982, + 0.021029839292168617, + 0.00953611545264721, + 0.019325153902173042, + 0.016748545691370964, + 0.01209244318306446, + 0.00904684979468584, + 0.010439139790832996, + 0.005128785036504269, + 0.010651933029294014, + 0.015672756358981133, + 0.008009325712919235, + 0.01610928401350975, + 0.010373309254646301, + 0.016095343977212906, + 0.014878773130476475, + 0.017986072227358818, + 0.015252734534442425, + 0.017514856532216072, + 0.008913121186196804, + 0.009561659768223763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.02635768987238407 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203954666852951 + ], + "opposite_logprob_tokens": [ + -2.812039613723755 + ], + "logprob_reward": -0.06203954666852951, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203954666852951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08663927763700485, + 0.28889378905296326, + 0.060602206736803055, + 0.0284323301166296, + 0.04868841543793678, + 0.25088629126548767, + 0.08912230283021927, + 0.05629667639732361, + 0.07611381262540817, + 0.01432486530393362 + ], + "credit_map": [ + 0.0072931027971208096, + 0.007510575465857983, + 0.01094632688909769, + 0.007874621078372002, + 0.008467912673950195, + 0.014674318954348564, + 0.012214194983243942, + 0.017658229917287827, + 0.11404966562986374, + 0.019536279141902924, + 0.11797388643026352, + 0.01167979184538126, + 0.006302003748714924, + 0.005569222383201122, + 0.006280284374952316, + 0.007502687629312277, + 0.00678882235661149, + 0.00764432642608881, + 0.006081674713641405, + 0.008795113302767277, + 0.008337913081049919, + 0.009271368384361267, + 0.008368996903300285, + 0.005313995759934187, + 0.0019651709590107203, + 0.003986788913607597, + 0.0022793046664446592, + 0.0047849551774561405, + 0.006220483686774969, + 0.0029133919160813093, + 0.0034889313392341137, + 0.002793304855003953, + 0.004915561527013779, + 0.004792716354131699, + 0.006590953562408686, + 0.005433853715658188, + 0.009771679528057575, + 0.0062135145999491215, + 0.005296994000673294, + 0.005673149600625038, + 0.015484945848584175, + 0.012698530219495296, + 0.09441309422254562, + 0.01724611595273018, + 0.00717434287071228, + 0.018375754356384277, + 0.007438731845468283, + 0.07805478572845459, + 0.01740826666355133, + 0.006124506238847971, + 0.024324260652065277, + 0.013776467181742191, + 0.008136777207255363, + 0.0057494076900184155, + 0.005993566010147333, + 0.007609054446220398, + 0.00615109596401453, + 0.0068181655369699, + 0.011001567356288433, + 0.004827694036066532, + 0.007508410606533289, + 0.00466985534876585, + 0.008372245356440544, + 0.006947647780179977, + 0.00970427691936493, + 0.008099601604044437, + 0.0081425029784441, + 0.004662152845412493, + 0.012685061432421207, + 0.011684905737638474, + 0.008895115926861763, + 0.012240209616720676, + 0.0011542777065187693, + 0.0016661161789670587, + 0.0013041463680565357, + 0.002727033570408821, + 0.0015749726444482803, + 0.002048986032605171, + 0.0021908970084041357, + 0.0016584355616942048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.06203954666852951 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062937498092651 + ], + "opposite_logprob_tokens": [ + -0.28129372000694275 + ], + "logprob_reward": -1.4062937498092651, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062937498092651, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14687693119049072, + 0.2035273313522339, + 0.11044621467590332, + 0.033609867095947266, + 0.05046232417225838, + 0.27300122380256653, + 0.07652286440134048, + 0.032618407160043716, + 0.03531718999147415, + 0.037617601454257965 + ], + "credit_map": [ + 0.011978846974670887, + 0.014197328127920628, + 0.019526423886418343, + 0.024795034900307655, + 0.013801843859255314, + 0.022688938304781914, + 0.024125462397933006, + 0.0157630555331707, + 0.07908526808023453, + 0.013093593530356884, + 0.08182264864444733, + 0.008628688752651215, + 0.0049057318829, + 0.0042467680759727955, + 0.0061080120503902435, + 0.005636624060571194, + 0.011122999712824821, + 0.01467609591782093, + 0.011139508336782455, + 0.016239384189248085, + 0.016327714547514915, + 0.01679408736526966, + 0.014678801409900188, + 0.009467623196542263, + 0.00263598351739347, + 0.006292886100709438, + 0.002381153404712677, + 0.004008406773209572, + 0.005458798259496689, + 0.005924756173044443, + 0.0036473756190389395, + 0.0032605049200356007, + 0.003945840522646904, + 0.007249996066093445, + 0.010028267279267311, + 0.005111017730087042, + 0.010178954340517521, + 0.004049181938171387, + 0.006675805896520615, + 0.0032232599332928658, + 0.016744662076234818, + 0.011845638044178486, + 0.10169227421283722, + 0.020005926489830017, + 0.0072022988460958, + 0.021142464131116867, + 0.0077541545033454895, + 0.08661381900310516, + 0.008863956667482853, + 0.011298987083137035, + 0.01649329625070095, + 0.01006973348557949, + 0.012418324127793312, + 0.004178809467703104, + 0.008089005947113037, + 0.005110755562782288, + 0.0038836542516946793, + 0.00399056077003479, + 0.004571777768433094, + 0.0022704151924699545, + 0.005669900216162205, + 0.002955472096800804, + 0.0043531558476388454, + 0.004923471249639988, + 0.004053632263094187, + 0.0037123369984328747, + 0.003966521006077528, + 0.002731264103204012, + 0.005997845903038979, + 0.0049613346345722675, + 0.004538875538855791, + 0.005355379078537226, + 0.0032807281240820885, + 0.004103719722479582, + 0.004697938449680805, + 0.0063734035938978195, + 0.005889114458113909, + 0.005404714494943619, + 0.004050364252179861, + 0.003817617427557707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -1.4062937498092651 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072501540184021 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072501540184021, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072501540184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1064172014594078, + 0.3810296952724457, + 0.09046077728271484, + 0.02418556809425354, + 0.03294490650296211, + 0.10657573491334915, + 0.056049980223178864, + 0.11018628627061844, + 0.04989863187074661, + 0.042251165956258774 + ], + "credit_map": [ + 0.012596475891768932, + 0.010235980153083801, + 0.014042729511857033, + 0.012461444362998009, + 0.015091209672391415, + 0.012656508944928646, + 0.01109298039227724, + 0.018239885568618774, + 0.15295027196407318, + 0.02464260533452034, + 0.1577930599451065, + 0.014459347352385521, + 0.006684411317110062, + 0.005183557514101267, + 0.00960898119956255, + 0.009707490913569927, + 0.009994255378842354, + 0.011956148780882359, + 0.009177123196423054, + 0.01491217128932476, + 0.011859185062348843, + 0.01402360387146473, + 0.010440588928759098, + 0.008097706362605095, + 0.0029864846728742123, + 0.003461055690422654, + 0.0019107598345726728, + 0.0031357293482869864, + 0.003870355663821101, + 0.004385499283671379, + 0.0023765889927744865, + 0.0020590967033058405, + 0.0029358393512666225, + 0.003923717886209488, + 0.007451054640114307, + 0.0030807601287961006, + 0.00780560402199626, + 0.002733963541686535, + 0.0030499976128339767, + 0.001963972579687834, + 0.004753405693918467, + 0.005028146784752607, + 0.02604174055159092, + 0.010950551368296146, + 0.006599364802241325, + 0.015165022574365139, + 0.004224862437695265, + 0.03381264582276344, + 0.008518198505043983, + 0.00860664714127779, + 0.011632137931883335, + 0.006236082408577204, + 0.00797372218221426, + 0.0033395313657820225, + 0.005675750318914652, + 0.0040679145604372025, + 0.01643926277756691, + 0.006280241534113884, + 0.027165653184056282, + 0.007691959850490093, + 0.009048976004123688, + 0.013341248966753483, + 0.01195360254496336, + 0.01826534792780876, + 0.005865113344043493, + 0.005079524125903845, + 0.0054262550547719, + 0.0030201999470591545, + 0.007091473788022995, + 0.008884607814252377, + 0.005651498679071665, + 0.008879965171217918, + 0.003644218435510993, + 0.004378448240458965, + 0.005187618546187878, + 0.005548281129449606, + 0.007090835366398096, + 0.00672892713919282, + 0.00424755085259676, + 0.005425287876278162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.011072501540184021 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "medium shot, the camera gently and fluidly tilting downward to follow the subject's movements, captures a vibrant orange tabby cat taking a swift and graceful leap from the top of a towering, dark wood bookshelf. In a warm study bathed in soft afternoon sunlight, light slanted in through a side window, illuminating tiny dust particles floating in the air and outlining a golden rim across the cat's short, dense fur. With a sudden kick of its hind legs, its body stretched out in mid-air, a fluid orange arc. Its emerald green eyes fixed intently on its target below, its tail balancing steadily behind it. In its leap, its hind paws inevitably swept across several misaligned books on the edge of the shelf, causing the heavy hardcovers to sway violently back and forth. One volume even slipped halfway out, its pages scattering in mid-air, as if poised to fall at any moment, adding a sense of chaotic tension to the graceful leap. The whole picture is painted in warm tones, juxtaposing the agility of cats with the tranquility of the home environment, full of dynamic beauty and storytelling.", + "seed": 12473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_4/index_1745_seed_12473_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4783975835889578, + "vlm_logprob": -0.4783975835889578, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.4783975835889578, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028169890865683556 + ], + "opposite_logprob_tokens": [ + -3.903169870376587 + ], + "logprob_reward": -0.028169890865683556, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.028169890865683556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2980262041091919, + 0.14331841468811035, + 0.06701216101646423, + 0.08667522668838501, + 0.03673716261982918, + 0.13148365914821625, + 0.06577015668153763, + 0.031184112653136253, + 0.06697596609592438, + 0.07281696051359177 + ], + "credit_map": [ + 0.01715647242963314, + 0.014234106056392193, + 0.01913318783044815, + 0.013284118846058846, + 0.020711548626422882, + 0.09662126004695892, + 0.02145935781300068, + 0.09542614966630936, + 0.0044762794859707355, + 0.005461288150399923, + 0.005607216153293848, + 0.0058631873689591885, + 0.004464162979274988, + 0.05429752916097641, + 0.007954842410981655, + 0.055193912237882614, + 0.006210390944033861, + 0.010939151048660278, + 0.00787295401096344, + 0.009065353311598301, + 0.008512903936207294, + 0.005849792622029781, + 0.011178499087691307, + 0.007383116986602545, + 0.008257444947957993, + 0.015459167771041393, + 0.011749865487217903, + 0.00776716647669673, + 0.010852690786123276, + 0.009579741396009922, + 0.010700823739171028, + 0.012308323755860329, + 0.003860877361148596, + 0.004481730051338673, + 0.004928849171847105, + 0.003759209532290697, + 0.004685246851295233, + 0.0046784523874521255, + 0.00496639683842659, + 0.005376399494707584, + 0.015878990292549133, + 0.016787996515631676, + 0.013709335587918758, + 0.013281552121043205, + 0.02047482132911682, + 0.01667935401201248, + 0.01570642925798893, + 0.018965184688568115, + 0.011627646163105965, + 0.008389363065361977, + 0.005282234400510788, + 0.006743451114743948, + 0.009712349623441696, + 0.0059545328840613365, + 0.00989518128335476, + 0.00816540326923132, + 0.004097201395779848, + 0.0031460942700505257, + 0.0030862975399941206, + 0.004303373396396637, + 0.00631503202021122, + 0.0037688890006393194, + 0.0034904300700873137, + 0.0029767956584692, + 0.010141045786440372, + 0.006340556312352419, + 0.00569765642285347, + 0.01044842042028904, + 0.00997522845864296, + 0.006702297832816839, + 0.00864342413842678, + 0.009027338586747646, + 0.00953582488000393, + 0.00820363499224186, + 0.01063449028879404, + 0.01168767735362053, + 0.010556468740105629, + 0.006839999463409185, + 0.005508692003786564, + 0.009850170463323593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.028169890865683556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596671849489212 + ], + "opposite_logprob_tokens": [ + -3.895596742630005 + ], + "logprob_reward": -0.020596671849489212, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596671849489212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1559595912694931, + 0.1713457703590393, + 0.049766331911087036, + 0.06627314537763596, + 0.05041279271245003, + 0.15277807414531708, + 0.18528985977172852, + 0.06503476947546005, + 0.05936096981167793, + 0.04377865791320801 + ], + "credit_map": [ + 0.012655148282647133, + 0.00861111655831337, + 0.016752304509282112, + 0.014036695472896099, + 0.01733466051518917, + 0.03410197049379349, + 0.019948629662394524, + 0.03251907229423523, + 0.010691465809941292, + 0.01145385205745697, + 0.012402325868606567, + 0.010872073471546173, + 0.010496368631720543, + 0.05009252205491066, + 0.014421283267438412, + 0.050915878266096115, + 0.0053139920346438885, + 0.008172018453478813, + 0.0071083009243011475, + 0.00457064900547266, + 0.005160454660654068, + 0.00473877415060997, + 0.007729475386440754, + 0.006972666829824448, + 0.007728371303528547, + 0.00937116239219904, + 0.008917490020394325, + 0.006195093970745802, + 0.011290637776255608, + 0.007821552455425262, + 0.007350074592977762, + 0.007598761469125748, + 0.005686174146831036, + 0.0050932797603309155, + 0.0060110935010015965, + 0.0046263327822089195, + 0.006109516602009535, + 0.006594354286789894, + 0.008467394858598709, + 0.007824644446372986, + 0.020973294973373413, + 0.01760062761604786, + 0.015936538577079773, + 0.015953024849295616, + 0.021120943129062653, + 0.018402298912405968, + 0.01865057274699211, + 0.024140769615769386, + 0.027175065129995346, + 0.018886564299464226, + 0.016823723912239075, + 0.015636848285794258, + 0.03535555303096771, + 0.018869295716285706, + 0.035199519246816635, + 0.017343299463391304, + 0.008903546258807182, + 0.0081847058609128, + 0.007582756690680981, + 0.008517788723111153, + 0.00941553432494402, + 0.006681972648948431, + 0.009572932496666908, + 0.00617553386837244, + 0.006108228582888842, + 0.005540590267628431, + 0.007395573891699314, + 0.00996498204767704, + 0.007773355580866337, + 0.004581938963383436, + 0.011796951293945312, + 0.0061993468552827835, + 0.005186142865568399, + 0.00290988408960402, + 0.009385643526911736, + 0.006427540443837643, + 0.005504641216248274, + 0.005243006162345409, + 0.003861665492877364, + 0.005260133650153875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.020596671849489212 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789676308631897 + ], + "opposite_logprob_tokens": [ + -2.578967571258545 + ], + "logprob_reward": -0.0789676308631897, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789676308631897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13438616693019867, + 0.25347158312797546, + 0.052494585514068604, + 0.12322688847780228, + 0.0689840242266655, + 0.10499529540538788, + 0.04736355319619179, + 0.04321175441145897, + 0.05017384514212608, + 0.12169232219457626 + ], + "credit_map": [ + 0.008562282659113407, + 0.008343388326466084, + 0.013212718069553375, + 0.013743214309215546, + 0.013874591328203678, + 0.03161672502756119, + 0.015565253794193268, + 0.029467999935150146, + 0.010631754994392395, + 0.012749195098876953, + 0.012694678269326687, + 0.013404217548668385, + 0.012850371189415455, + 0.08744215220212936, + 0.014550864696502686, + 0.08914834260940552, + 0.005944742821156979, + 0.006568322889506817, + 0.006183518562465906, + 0.005858507007360458, + 0.006343729328364134, + 0.00567995710298419, + 0.008957765065133572, + 0.006958041340112686, + 0.007919053547084332, + 0.01494116522371769, + 0.013875773176550865, + 0.01078603882342577, + 0.026820890605449677, + 0.017162255942821503, + 0.016961684450507164, + 0.01476003136485815, + 0.006786643527448177, + 0.008006162941455841, + 0.008199796080589294, + 0.007572008762508631, + 0.008920690044760704, + 0.00891330186277628, + 0.010732904076576233, + 0.009852517396211624, + 0.012362966313958168, + 0.015972383320331573, + 0.011860368773341179, + 0.013551388867199421, + 0.01243364717811346, + 0.010241777636110783, + 0.0158359557390213, + 0.012736815959215164, + 0.009341107681393623, + 0.004519064910709858, + 0.005225118715316057, + 0.004835326690226793, + 0.006927162874490023, + 0.004326349589973688, + 0.00768757751211524, + 0.004501842427998781, + 0.006160255987197161, + 0.004648434929549694, + 0.00566829415038228, + 0.005139808636158705, + 0.005645659752190113, + 0.003722816938534379, + 0.006457487586885691, + 0.005768998526036739, + 0.006568229757249355, + 0.0036880732513964176, + 0.007203687448054552, + 0.009347346611320972, + 0.005930190905928612, + 0.004211821593344212, + 0.005982083268463612, + 0.00724240904673934, + 0.015204314142465591, + 0.00901997834444046, + 0.026198100298643112, + 0.014897752553224564, + 0.019293377175927162, + 0.01690291427075863, + 0.006305856630206108, + 0.013870018534362316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0789676308631897 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.237898588180542 + ], + "opposite_logprob_tokens": [ + -0.11289859563112259 + ], + "logprob_reward": -2.237898588180542, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.237898588180542, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0655020996928215, + 0.13020430505275726, + 0.06853177398443222, + 0.08581250160932541, + 0.06717047840356827, + 0.07281842827796936, + 0.05777484551072121, + 0.08226282894611359, + 0.23825420439243317, + 0.13166849315166473 + ], + "credit_map": [ + 0.0060412827879190445, + 0.005723880138248205, + 0.006962009705603123, + 0.006720388773828745, + 0.009245846420526505, + 0.011293312534689903, + 0.008615083061158657, + 0.010900303721427917, + 0.005404183175414801, + 0.005496181081980467, + 0.008429305627942085, + 0.00923831108957529, + 0.006073426455259323, + 0.04356222599744797, + 0.007277734111994505, + 0.044722944498062134, + 0.008492371998727322, + 0.008851712569594383, + 0.008367246948182583, + 0.00745391845703125, + 0.007685286924242973, + 0.007771951612085104, + 0.008752451278269291, + 0.0111568383872509, + 0.0099417045712471, + 0.013067441992461681, + 0.012007173150777817, + 0.010887780226767063, + 0.01029124390333891, + 0.007796243764460087, + 0.014492733404040337, + 0.0073281871154904366, + 0.00951019674539566, + 0.007875969633460045, + 0.007639129646122456, + 0.007026040460914373, + 0.009450545534491539, + 0.007828766480088234, + 0.00915274117141962, + 0.008687090128660202, + 0.006398934405297041, + 0.008418262004852295, + 0.007215606048703194, + 0.007517002988606691, + 0.012138751335442066, + 0.008579191751778126, + 0.01467868871986866, + 0.007871998474001884, + 0.010926445946097374, + 0.005775871220976114, + 0.00669477041810751, + 0.004844718147069216, + 0.008007094264030457, + 0.005195843521505594, + 0.009842446073889732, + 0.006487657781690359, + 0.010304282419383526, + 0.00912320613861084, + 0.008872042410075665, + 0.009070859290659428, + 0.014339030720293522, + 0.007995855063199997, + 0.014037525281310081, + 0.008520036935806274, + 0.027818383648991585, + 0.03292836248874664, + 0.028495702892541885, + 0.04238252341747284, + 0.03095981478691101, + 0.02101803943514824, + 0.03491470590233803, + 0.019736699759960175, + 0.01601404882967472, + 0.018905574455857277, + 0.017858194187283516, + 0.017557894811034203, + 0.01506600808352232, + 0.01193571463227272, + 0.018511367961764336, + 0.015819711610674858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -2.237898588180542 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355136185884476 + ], + "opposite_logprob_tokens": [ + -3.651355028152466 + ], + "logprob_reward": -0.026355136185884476, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355136185884476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04787546768784523, + 0.08625839650630951, + 0.06244221329689026, + 0.06897672265768051, + 0.049649059772491455, + 0.12920762598514557, + 0.3447829484939575, + 0.05986703559756279, + 0.1006132960319519, + 0.05032726004719734 + ], + "credit_map": [ + 0.0066544655710458755, + 0.005511692725121975, + 0.0065964581444859505, + 0.007082910742610693, + 0.007378729060292244, + 0.004336201120167971, + 0.006017577368766069, + 0.004297431092709303, + 0.003616203786805272, + 0.0039109596982598305, + 0.004042375832796097, + 0.003518495475873351, + 0.002795382170006633, + 0.03206159919500351, + 0.003680808236822486, + 0.032632581889629364, + 0.011372001841664314, + 0.007835118100047112, + 0.008519468829035759, + 0.009702428244054317, + 0.005721661262214184, + 0.005107781384140253, + 0.007786874659359455, + 0.006396877113729715, + 0.008646948263049126, + 0.013346576131880283, + 0.011016900651156902, + 0.004605516791343689, + 0.007898074574768543, + 0.005696612875908613, + 0.01235648151487112, + 0.00540960393846035, + 0.008857793174684048, + 0.008188472129404545, + 0.006432080175727606, + 0.004185512196272612, + 0.0054037123918533325, + 0.005417348816990852, + 0.005667179357260466, + 0.005496960133314133, + 0.020856691524386406, + 0.013892735354602337, + 0.016285670921206474, + 0.016619596630334854, + 0.01277029700577259, + 0.013325052335858345, + 0.01943030208349228, + 0.016027284786105156, + 0.06785088032484055, + 0.05521182343363762, + 0.04683539643883705, + 0.05448514223098755, + 0.02915184199810028, + 0.022233761847019196, + 0.04290780425071716, + 0.026106305420398712, + 0.009197896346449852, + 0.0083346301689744, + 0.008001735433936119, + 0.007650004234164953, + 0.0064653088338673115, + 0.0046197958290576935, + 0.00802917592227459, + 0.00756848743185401, + 0.012655808590352535, + 0.01079644076526165, + 0.012138946913182735, + 0.019081102684140205, + 0.014565180987119675, + 0.00918769184499979, + 0.011545281857252121, + 0.010642839595675468, + 0.006457251496613026, + 0.0055901650339365005, + 0.00781453400850296, + 0.00458445493131876, + 0.006977310869842768, + 0.005770984571427107, + 0.006031892262399197, + 0.007100666407495737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.026355136185884476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "medium shot, the camera gently and fluidly tilting downward to follow the subject's movements, captures a vibrant orange tabby cat taking a swift and graceful leap from the top of a towering, dark wood bookshelf. In a warm study bathed in soft afternoon sunlight, light slanted in through a side window, illuminating tiny dust particles floating in the air and outlining a golden rim across the cat's short, dense fur. With a sudden kick of its hind legs, its body stretched out in mid-air, a fluid orange arc. Its emerald green eyes fixed intently on its target below, its tail balancing steadily behind it. In its leap, its hind paws inevitably swept across several misaligned books on the edge of the shelf, causing the heavy hardcovers to sway violently back and forth. One volume even slipped halfway out, its pages scattering in mid-air, as if poised to fall at any moment, adding a sense of chaotic tension to the graceful leap. The whole picture is painted in warm tones, juxtaposing the agility of cats with the tranquility of the home environment, full of dynamic beauty and storytelling.", + "seed": 12474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_4/index_1745_seed_12474_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.25022341273725035, + "vlm_logprob": -0.25022341273725035, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.25022341273725035, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017254596576094627 + ], + "opposite_logprob_tokens": [ + -4.767254829406738 + ], + "logprob_reward": -0.017254596576094627, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017254596576094627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14601483941078186, + 0.13211461901664734, + 0.03604438528418541, + 0.06327498704195023, + 0.05046981945633888, + 0.12155123800039291, + 0.08215155452489853, + 0.08125931024551392, + 0.08278471976518631, + 0.20433451235294342 + ], + "credit_map": [ + 0.013254676014184952, + 0.018613362684845924, + 0.013668183237314224, + 0.024033408612012863, + 0.029382552951574326, + 0.018792642280459404, + 0.019306743517518044, + 0.008963272906839848, + 0.007964002899825573, + 0.014273149892687798, + 0.00908356811851263, + 0.018655693158507347, + 0.016765177249908447, + 0.021583586931228638, + 0.01689605414867401, + 0.02689339406788349, + 0.003075083252042532, + 0.004642169456928968, + 0.0052830069325864315, + 0.004913312382996082, + 0.004524200223386288, + 0.003970014397054911, + 0.005026796367019415, + 0.0046098059974610806, + 0.005051988177001476, + 0.010338333435356617, + 0.003202604129910469, + 0.007366030011326075, + 0.01002107746899128, + 0.011483866721391678, + 0.006376426666975021, + 0.009434658102691174, + 0.005110866855829954, + 0.005964899435639381, + 0.006724371109157801, + 0.005703320261090994, + 0.007214165758341551, + 0.007371457759290934, + 0.004473965615034103, + 0.007906774990260601, + 0.011521216481924057, + 0.012217783369123936, + 0.01668030582368374, + 0.020032033324241638, + 0.015446184203028679, + 0.013287314213812351, + 0.01939556747674942, + 0.012970834039151669, + 0.011139786802232265, + 0.010913210920989513, + 0.009685839526355267, + 0.010527177713811398, + 0.012053709477186203, + 0.008851537480950356, + 0.00793035514652729, + 0.011049933731555939, + 0.006882600020617247, + 0.009246178902685642, + 0.011671931482851505, + 0.004864593967795372, + 0.014196776784956455, + 0.016714293509721756, + 0.009117336943745613, + 0.00856560468673706, + 0.011700168251991272, + 0.011495662853121758, + 0.008567950688302517, + 0.009665329940617085, + 0.01181393675506115, + 0.009179570712149143, + 0.011532716453075409, + 0.008829382248222828, + 0.025533054023981094, + 0.02135133370757103, + 0.024306802079081535, + 0.030037855729460716, + 0.029391329735517502, + 0.024038754403591156, + 0.03327764943242073, + 0.016397733241319656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.017254596576094627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076697036623955 + ], + "opposite_logprob_tokens": [ + -4.141076564788818 + ], + "logprob_reward": -0.016076697036623955, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076697036623955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050983820110559464, + 0.09046991914510727, + 0.05070778355002403, + 0.10241557657718658, + 0.049491021782159805, + 0.16242924332618713, + 0.13403451442718506, + 0.08241169899702072, + 0.09225323051214218, + 0.18480318784713745 + ], + "credit_map": [ + 0.008067955262959003, + 0.005292840301990509, + 0.007019218057394028, + 0.0056732152588665485, + 0.006797873880714178, + 0.0060419924557209015, + 0.006172063760459423, + 0.005918663460761309, + 0.009148070588707924, + 0.009866872802376747, + 0.006985501851886511, + 0.008748071268200874, + 0.015719065442681313, + 0.013619428500533104, + 0.01135988812893629, + 0.015023019164800644, + 0.005478843115270138, + 0.0074539692141115665, + 0.006646429188549519, + 0.00613782461732626, + 0.006536215078085661, + 0.006076519377529621, + 0.006261340342462063, + 0.006116645410656929, + 0.010340509936213493, + 0.01590067334473133, + 0.007948906160891056, + 0.012744005769491196, + 0.016859784722328186, + 0.022554442286491394, + 0.008840967901051044, + 0.00722628366202116, + 0.004887802992016077, + 0.00853448361158371, + 0.004772950895130634, + 0.005439088214188814, + 0.008271940052509308, + 0.005904135759919882, + 0.005694488529115915, + 0.005986131727695465, + 0.017837101593613625, + 0.021755240857601166, + 0.019943667575716972, + 0.020354559645056725, + 0.0226245429366827, + 0.019103851169347763, + 0.022699736058712006, + 0.018110541626811028, + 0.013069952838122845, + 0.019089171662926674, + 0.015650875866413116, + 0.012459547258913517, + 0.024246029555797577, + 0.017853762954473495, + 0.013428016565740108, + 0.0182371586561203, + 0.01100899651646614, + 0.008843799121677876, + 0.01237082201987505, + 0.005888854619115591, + 0.007886546663939953, + 0.019926253706216812, + 0.009173132479190826, + 0.007313298527151346, + 0.011153257451951504, + 0.01498870737850666, + 0.006840765010565519, + 0.010610876604914665, + 0.013153974898159504, + 0.010421782732009888, + 0.014805817976593971, + 0.010278042405843735, + 0.01628783904016018, + 0.017467612400650978, + 0.01651054061949253, + 0.019982105121016502, + 0.040564317256212234, + 0.029796456918120384, + 0.028475094586610794, + 0.015719225630164146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.016076697036623955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603248417377472 + ], + "opposite_logprob_tokens": [ + -1.9103248119354248 + ], + "logprob_reward": -0.1603248417377472, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603248417377472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061338216066360474, + 0.05277872830629349, + 0.07476108521223068, + 0.08746416866779327, + 0.06312582641839981, + 0.4239475727081299, + 0.08629382401704788, + 0.03861323744058609, + 0.05192527174949646, + 0.05975203216075897 + ], + "credit_map": [ + 0.008343914523720741, + 0.007296928204596043, + 0.007971062324941158, + 0.006239025853574276, + 0.006587853189557791, + 0.007809872273355722, + 0.008763588033616543, + 0.008325975388288498, + 0.004137678071856499, + 0.006247614976018667, + 0.005372642073780298, + 0.007239528466016054, + 0.009670133702456951, + 0.007180222775787115, + 0.007240491919219494, + 0.005690424237400293, + 0.012520275078713894, + 0.008627175353467464, + 0.010109738446772099, + 0.007337186019867659, + 0.006657162215560675, + 0.007134764920920134, + 0.011458917520940304, + 0.010915874503552914, + 0.006003242917358875, + 0.010692414827644825, + 0.007436500862240791, + 0.010583385825157166, + 0.012405399233102798, + 0.02063244953751564, + 0.009333756752312183, + 0.010377022437751293, + 0.006107732653617859, + 0.009655294939875603, + 0.007160516455769539, + 0.006450820714235306, + 0.009361648000776768, + 0.009004789404571056, + 0.006294120568782091, + 0.009090909734368324, + 0.04346560314297676, + 0.038779277354478836, + 0.062303006649017334, + 0.0709681585431099, + 0.038935586810112, + 0.04837331175804138, + 0.07763073593378067, + 0.04349193722009659, + 0.007817321456968784, + 0.013946419581770897, + 0.008379435166716576, + 0.010433902032673359, + 0.01780383288860321, + 0.010361719876527786, + 0.0065736169926822186, + 0.010977580212056637, + 0.00574378902092576, + 0.0030917061958462, + 0.005811778828501701, + 0.0030374443158507347, + 0.0036931305658072233, + 0.00852485653012991, + 0.0033502669539302588, + 0.005360271316021681, + 0.006931733805686235, + 0.006523956079035997, + 0.0064351363107562065, + 0.005294242408126593, + 0.005942783784121275, + 0.00540511729195714, + 0.0077834646217525005, + 0.007608842104673386, + 0.006468940991908312, + 0.004870154894888401, + 0.00525174755603075, + 0.00555064482614398, + 0.014691614545881748, + 0.008955689147114754, + 0.00919161643832922, + 0.004771627951413393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.1603248417377472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538486242294312 + ], + "opposite_logprob_tokens": [ + -0.42884865403175354 + ], + "logprob_reward": -1.0538486242294312, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538486242294312, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09154010564088821, + 0.08935030549764633, + 0.05322796851396561, + 0.09216966480016708, + 0.03660211339592934, + 0.08553710579872131, + 0.19350875914096832, + 0.16893048584461212, + 0.0671200230717659, + 0.12201346457004547 + ], + "credit_map": [ + 0.009043527767062187, + 0.01271008513867855, + 0.008053592406213284, + 0.014741719700396061, + 0.01448312122374773, + 0.01072825025767088, + 0.01525868009775877, + 0.006521133240312338, + 0.006915448233485222, + 0.014245938509702682, + 0.00572222052142024, + 0.010769457556307316, + 0.020031098276376724, + 0.013862096704542637, + 0.009855026379227638, + 0.007949021644890308, + 0.0055519817396998405, + 0.008177550509572029, + 0.005363945383578539, + 0.006386480759829283, + 0.009077124297618866, + 0.006085317116230726, + 0.006689158733934164, + 0.005896411370486021, + 0.006039279047399759, + 0.025610921904444695, + 0.005351292435079813, + 0.01366463489830494, + 0.014088073745369911, + 0.013647104613482952, + 0.008042651228606701, + 0.005725707393139601, + 0.0036304383538663387, + 0.0065115392208099365, + 0.0031077093444764614, + 0.004388392437249422, + 0.005952764768153429, + 0.004152751062065363, + 0.004059151280671358, + 0.004799365997314453, + 0.005358162801712751, + 0.013181217014789581, + 0.009595627896487713, + 0.016502391546964645, + 0.011902736499905586, + 0.007561904843896627, + 0.013257007114589214, + 0.008178060874342918, + 0.02150743082165718, + 0.028999870643019676, + 0.01640244573354721, + 0.021632011979818344, + 0.0354154109954834, + 0.026894113048911095, + 0.01567918248474598, + 0.02697829343378544, + 0.014296920970082283, + 0.019958045333623886, + 0.02217097021639347, + 0.009689195081591606, + 0.03211364150047302, + 0.029537126421928406, + 0.020250661298632622, + 0.020913921296596527, + 0.011559046804904938, + 0.009339739568531513, + 0.0068320236168801785, + 0.008017037995159626, + 0.008853758685290813, + 0.007942697033286095, + 0.008215468376874924, + 0.006360252387821674, + 0.009817440994083881, + 0.00992144737392664, + 0.012324035167694092, + 0.01597457565367222, + 0.024061745032668114, + 0.0194370336830616, + 0.020010946318507195, + 0.010466242209076881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -1.0538486242294312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036123041063547134 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.0036123041063547134, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036123041063547134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11253049969673157, + 0.16311778128147125, + 0.029695656150579453, + 0.0479397289454937, + 0.09698628634214401, + 0.1674528270959854, + 0.1695300191640854, + 0.08049609512090683, + 0.034240782260894775, + 0.09801033139228821 + ], + "credit_map": [ + 0.01408479269593954, + 0.010569566860795021, + 0.015429018996655941, + 0.010997903533279896, + 0.013658975251019001, + 0.013226469978690147, + 0.014981028623878956, + 0.019582748413085938, + 0.01677943952381611, + 0.022693078964948654, + 0.011833496391773224, + 0.02298746630549431, + 0.026262979954481125, + 0.023911358788609505, + 0.019992182031273842, + 0.018657786771655083, + 0.0021952076349407434, + 0.005217432975769043, + 0.0037021012976765633, + 0.004303955007344484, + 0.004653296433389187, + 0.0034068170934915543, + 0.0027910240460187197, + 0.003425822826102376, + 0.003022475866600871, + 0.007934974506497383, + 0.0042316969484090805, + 0.006449846550822258, + 0.006442422978579998, + 0.009007287211716175, + 0.005437969230115414, + 0.0054130530916154385, + 0.015150095336139202, + 0.00992226880043745, + 0.01628333516418934, + 0.009613235481083393, + 0.013260525651276112, + 0.01127606350928545, + 0.009129918180406094, + 0.012350845150649548, + 0.015087462961673737, + 0.030778517946600914, + 0.013161374256014824, + 0.016624996438622475, + 0.03720640018582344, + 0.018012860789895058, + 0.015630116686224937, + 0.02095109224319458, + 0.019940821453928947, + 0.025409545749425888, + 0.015356807969510555, + 0.019161436706781387, + 0.0290730781853199, + 0.02144467458128929, + 0.014972340315580368, + 0.024171313270926476, + 0.007057310547679663, + 0.009926076047122478, + 0.01078956387937069, + 0.004860435612499714, + 0.014863822609186172, + 0.014741878025233746, + 0.009073333814740181, + 0.009183669462800026, + 0.004738458897918463, + 0.004489386919885874, + 0.004454223904758692, + 0.0038849071133881807, + 0.004481935407966375, + 0.004258028231561184, + 0.004198846872895956, + 0.0037349930498749018, + 0.0077987187542021275, + 0.008892207406461239, + 0.009147998876869678, + 0.01330989133566618, + 0.01916150376200676, + 0.015312882140278816, + 0.015595808625221252, + 0.008791319094598293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.0036123041063547134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "medium shot, the camera gently and fluidly tilting downward to follow the subject's movements, captures a vibrant orange tabby cat taking a swift and graceful leap from the top of a towering, dark wood bookshelf. In a warm study bathed in soft afternoon sunlight, light slanted in through a side window, illuminating tiny dust particles floating in the air and outlining a golden rim across the cat's short, dense fur. With a sudden kick of its hind legs, its body stretched out in mid-air, a fluid orange arc. Its emerald green eyes fixed intently on its target below, its tail balancing steadily behind it. In its leap, its hind paws inevitably swept across several misaligned books on the edge of the shelf, causing the heavy hardcovers to sway violently back and forth. One volume even slipped halfway out, its pages scattering in mid-air, as if poised to fall at any moment, adding a sense of chaotic tension to the graceful leap. The whole picture is painted in warm tones, juxtaposing the agility of cats with the tranquility of the home environment, full of dynamic beauty and storytelling.", + "seed": 12475, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_4/index_1745_seed_12475_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2819882843643427, + "vlm_logprob": -0.2819882843643427, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.2819882843643427, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014269223436713219 + ], + "opposite_logprob_tokens": [ + -5.7642693519592285 + ], + "logprob_reward": -0.014269223436713219, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.014269223436713219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19557417929172516, + 0.06580281257629395, + 0.05985052138566971, + 0.03633327782154083, + 0.04867148399353027, + 0.05971397086977959, + 0.2005254030227661, + 0.0879019945859909, + 0.0980740562081337, + 0.1475522667169571 + ], + "credit_map": [ + 0.028958512470126152, + 0.01783170737326145, + 0.028472570702433586, + 0.03044874779880047, + 0.02674706280231476, + 0.017010610550642014, + 0.023857204243540764, + 0.02224777638912201, + 0.008177409879863262, + 0.004924967885017395, + 0.007985304109752178, + 0.012625967152416706, + 0.009568221867084503, + 0.006580736953765154, + 0.00944590661674738, + 0.006494294852018356, + 0.01017497107386589, + 0.0064939772710204124, + 0.008642378263175488, + 0.008160898461937904, + 0.007604071870446205, + 0.004624761175364256, + 0.008319112472236156, + 0.0058303517289459705, + 0.0065103261731565, + 0.002937408396974206, + 0.00670517235994339, + 0.005085316486656666, + 0.003977362532168627, + 0.0022575699258595705, + 0.0038968666922301054, + 0.0049632517620921135, + 0.006895763799548149, + 0.004253856837749481, + 0.00745435431599617, + 0.0063399155624210835, + 0.0050304061733186245, + 0.0043839686550199986, + 0.005464499816298485, + 0.008848720230162144, + 0.009226868860423565, + 0.008022302761673927, + 0.006976758129894733, + 0.006632121279835701, + 0.006610405165702105, + 0.006913867779076099, + 0.008501480333507061, + 0.006830167956650257, + 0.015693020075559616, + 0.022947514429688454, + 0.01725948601961136, + 0.04648945480585098, + 0.021776610985398293, + 0.020964941009879112, + 0.042634449899196625, + 0.012759915553033352, + 0.012474069371819496, + 0.012723598629236221, + 0.007643064018338919, + 0.016981245949864388, + 0.01030625682324171, + 0.007166007533669472, + 0.009951393119990826, + 0.01065636146813631, + 0.01301499456167221, + 0.01436632964760065, + 0.016258079558610916, + 0.017255226150155067, + 0.009626074694097042, + 0.006264699622988701, + 0.01062692329287529, + 0.010661720298230648, + 0.014880544506013393, + 0.01759740151464939, + 0.025899872183799744, + 0.021190231665968895, + 0.01656353659927845, + 0.01651071570813656, + 0.017491048201918602, + 0.01741892658174038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.014269223436713219 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160837359726429 + ], + "opposite_logprob_tokens": [ + -4.141083717346191 + ], + "logprob_reward": -0.0160837359726429, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160837359726429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1476331204175949, + 0.05631521716713905, + 0.0758720263838768, + 0.05528557300567627, + 0.046541519463062286, + 0.06505854427814484, + 0.23739981651306152, + 0.15671169757843018, + 0.08323928713798523, + 0.07594313472509384 + ], + "credit_map": [ + 0.017360497266054153, + 0.0278250053524971, + 0.017550619319081306, + 0.017473537474870682, + 0.0254437904804945, + 0.00981785636395216, + 0.01927952840924263, + 0.012882301583886147, + 0.0061746686697006226, + 0.005072198808193207, + 0.008148033171892166, + 0.012826579622924328, + 0.008612973615527153, + 0.00387711264193058, + 0.006168913561850786, + 0.005434751510620117, + 0.010784732177853584, + 0.006387616973370314, + 0.010575037449598312, + 0.008530238643288612, + 0.010958232916891575, + 0.008761288598179817, + 0.009548290632665157, + 0.01032660435885191, + 0.008212758228182793, + 0.003996826242655516, + 0.010867795906960964, + 0.012970560230314732, + 0.009746051393449306, + 0.0022309678606688976, + 0.004339728970080614, + 0.002920896513387561, + 0.0071846600621938705, + 0.004247195087373257, + 0.009282751008868217, + 0.006542122922837734, + 0.007034567184746265, + 0.002839782042428851, + 0.005550631787627935, + 0.00385981984436512, + 0.009914800524711609, + 0.006817361805588007, + 0.006157230585813522, + 0.007170860189944506, + 0.007602885365486145, + 0.007811558432877064, + 0.007629771716892719, + 0.011954092420637608, + 0.021746501326560974, + 0.03242192417383194, + 0.02236192673444748, + 0.042991332709789276, + 0.042253728955984116, + 0.023160694167017937, + 0.03918146342039108, + 0.013282288797199726, + 0.02658090740442276, + 0.024910273030400276, + 0.014332382939755917, + 0.021657532081007957, + 0.020550599321722984, + 0.009267428889870644, + 0.02367737889289856, + 0.01573522388935089, + 0.014358790591359138, + 0.013465196825563908, + 0.012621371075510979, + 0.009128023870289326, + 0.009662272408604622, + 0.006631057243794203, + 0.008684537373483181, + 0.008688055910170078, + 0.00648255180567503, + 0.010018664412200451, + 0.012565935030579567, + 0.013199500739574432, + 0.014404954388737679, + 0.005020069424062967, + 0.008235327899456024, + 0.006016151513904333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0160837359726429 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292190849781036 + ], + "opposite_logprob_tokens": [ + -2.237921953201294 + ], + "logprob_reward": -0.11292190849781036, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292190849781036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18923938274383545, + 0.12953859567642212, + 0.26784369349479675, + 0.08351245522499084, + 0.05344253405928612, + 0.040853314101696014, + 0.03407168388366699, + 0.0650644302368164, + 0.04255186766386032, + 0.09388203173875809 + ], + "credit_map": [ + 0.02876185066998005, + 0.034383561462163925, + 0.019639912992715836, + 0.020321074873209, + 0.024708086624741554, + 0.01841883547604084, + 0.0239826999604702, + 0.01902337372303009, + 0.01801016554236412, + 0.01058994885534048, + 0.013870440423488617, + 0.01933668553829193, + 0.014656870625913143, + 0.01763560064136982, + 0.019325032830238342, + 0.016113845631480217, + 0.037359949201345444, + 0.023750921711325645, + 0.05570598691701889, + 0.035513926297426224, + 0.03664655610918999, + 0.021021079272031784, + 0.03209489956498146, + 0.025750353932380676, + 0.014270983636379242, + 0.006346795707941055, + 0.018191134557127953, + 0.014606201089918613, + 0.01331119891256094, + 0.004343126434832811, + 0.008274947293102741, + 0.004168071784079075, + 0.007632599677890539, + 0.005164008121937513, + 0.009595314040780067, + 0.007230740040540695, + 0.00610346207395196, + 0.0032541928812861443, + 0.008126801811158657, + 0.006335416808724403, + 0.006238998379558325, + 0.004674122668802738, + 0.006057939026504755, + 0.006313113961368799, + 0.004803764168173075, + 0.0034389421343803406, + 0.006189175881445408, + 0.0031372560188174248, + 0.0023498537484556437, + 0.004151014145463705, + 0.0033001990523189306, + 0.004376368131488562, + 0.004899867810308933, + 0.007955455221235752, + 0.0048711360432207584, + 0.002167788567021489, + 0.008782191202044487, + 0.010143440216779709, + 0.0058221472427248955, + 0.01002497598528862, + 0.007831399329006672, + 0.004787639249116182, + 0.010327457450330257, + 0.007345180958509445, + 0.006996951997280121, + 0.006195432040840387, + 0.007016477175056934, + 0.006595917511731386, + 0.0034705810248851776, + 0.004552830476313829, + 0.004721943289041519, + 0.0030017350800335407, + 0.009795673191547394, + 0.012327007949352264, + 0.017389975488185883, + 0.010189834050834179, + 0.013613931834697723, + 0.007757990621030331, + 0.012378176674246788, + 0.010429440066218376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.11292190849781036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236037254333496 + ], + "opposite_logprob_tokens": [ + -0.348603755235672 + ], + "logprob_reward": -1.2236037254333496, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236037254333496, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10649917274713516, + 0.07018209248781204, + 0.09218177199363708, + 0.05060689151287079, + 0.03006955236196518, + 0.02964271977543831, + 0.1690666526556015, + 0.10073243081569672, + 0.09398738294839859, + 0.2570313513278961 + ], + "credit_map": [ + 0.021953174844384193, + 0.009683335199952126, + 0.01964171789586544, + 0.017135482281446457, + 0.010997401550412178, + 0.007478037383407354, + 0.011665214784443378, + 0.007944808341562748, + 0.011128989048302174, + 0.0064656613394618034, + 0.008426214568316936, + 0.01009686291217804, + 0.006550879683345556, + 0.009753609076142311, + 0.010521724820137024, + 0.007238149642944336, + 0.01430094800889492, + 0.009738047607243061, + 0.01618840917944908, + 0.013355329632759094, + 0.016180219128727913, + 0.005923271644860506, + 0.009515645913779736, + 0.0069798934273421764, + 0.009027257561683655, + 0.004931611940264702, + 0.010389959439635277, + 0.007701660506427288, + 0.006853707600384951, + 0.003134077647700906, + 0.0048265703953802586, + 0.003742044558748603, + 0.00454773660749197, + 0.0030141391325742006, + 0.003967789467424154, + 0.0045190476812422276, + 0.003989725839346647, + 0.002176238689571619, + 0.003927651327103376, + 0.0039272247813642025, + 0.003686116309836507, + 0.0027848894242197275, + 0.004922783467918634, + 0.005667227786034346, + 0.002805565483868122, + 0.0020310990512371063, + 0.004946674685925245, + 0.002798361936584115, + 0.014135407283902168, + 0.019588887691497803, + 0.017016520723700523, + 0.03609499707818031, + 0.018839558586478233, + 0.016034582629799843, + 0.03730235993862152, + 0.010054348036646843, + 0.016505591571331024, + 0.015752557665109634, + 0.01025894284248352, + 0.012703733518719673, + 0.01138374675065279, + 0.0069197919219732285, + 0.013821647502481937, + 0.013386416248977184, + 0.014700794592499733, + 0.015298556536436081, + 0.014117605984210968, + 0.01474684476852417, + 0.010235968045890331, + 0.00825689546763897, + 0.008688061498105526, + 0.00794264953583479, + 0.03189600631594658, + 0.03251098096370697, + 0.05178019776940346, + 0.03475053608417511, + 0.03808999061584473, + 0.02066771872341633, + 0.02884327992796898, + 0.01849261485040188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -1.2236037254333496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of an orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently and fluidly tilt downward to follow the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat leaping from the top of a towering dark wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are books on the shelf shown swaying violently and one volume slipping with pages scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a warm study with soft afternoon sunlight streaming through a side window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306282848119736 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306282848119736, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306282848119736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5382872223854065, + 0.05538596957921982, + 0.04049453139305115, + 0.011139245703816414, + 0.03322918340563774, + 0.10907530039548874, + 0.05936064571142197, + 0.036500364542007446, + 0.03757477551698685, + 0.07895278185606003 + ], + "credit_map": [ + 0.07256556302309036, + 0.11192633956670761, + 0.062100037932395935, + 0.06241803243756294, + 0.10290101915597916, + 0.037148457020521164, + 0.051005907356739044, + 0.03822183609008789, + 0.008938617072999477, + 0.0072184097953140736, + 0.006710878573358059, + 0.007125318981707096, + 0.006926101166754961, + 0.00521789351478219, + 0.007255767937749624, + 0.005992984399199486, + 0.005360277369618416, + 0.0035502631217241287, + 0.008447840809822083, + 0.005951001308858395, + 0.004655678290873766, + 0.0034309430047869682, + 0.004501488525420427, + 0.0045970408245921135, + 0.0014422052772715688, + 0.0009161856723949313, + 0.002102513564750552, + 0.0023223154712468386, + 0.0017388071864843369, + 0.0007712453952990472, + 0.0009770654141902924, + 0.0008689069654792547, + 0.005305835045874119, + 0.0028203369583934546, + 0.007111331447958946, + 0.004819806199520826, + 0.003563341684639454, + 0.002420845441520214, + 0.004104091785848141, + 0.0030835922807455063, + 0.018417207524180412, + 0.014565971679985523, + 0.011206910014152527, + 0.012166796252131462, + 0.013723797164857388, + 0.011823535896837711, + 0.012277505360543728, + 0.014893580228090286, + 0.004289492964744568, + 0.006464504636824131, + 0.004844793118536472, + 0.012537947855889797, + 0.007050391286611557, + 0.007702578790485859, + 0.011750754900276661, + 0.004720181226730347, + 0.006661195773631334, + 0.006122736260294914, + 0.004028816241770983, + 0.004129359964281321, + 0.005287530831992626, + 0.0020633346866816282, + 0.004156833048909903, + 0.004050559364259243, + 0.007100334856659174, + 0.006207638885825872, + 0.005358790047466755, + 0.00443634856492281, + 0.0039863851852715015, + 0.003968207631260157, + 0.0036192750558257103, + 0.0028977973852306604, + 0.007424370385706425, + 0.010698260739445686, + 0.012702769599854946, + 0.008485672064125538, + 0.011812595650553703, + 0.0071869464591145515, + 0.010839208960533142, + 0.00980295892804861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.04306282848119736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_5/index_tensor([8541, 8541])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_5/index_tensor([8541, 8541])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1fa909b95e8cd839bbd0cb5a6308a31dd6313ce0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_5/index_tensor([8541, 8541])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle drone shot unfolds across a multi-lane highway, its dark, wet asphalt carving a steep path across the landscape. The camera then begins a slow, graceful vertical ascent to reveal a wider bird's-eye view. The entire scene is blanketed by a thick, low-hanging white mist, creating an ethereal, diffuse glow that eliminates harsh shadows. This mist lingers over the surrounding fields, softening the outlines of distant trees and rendering the background a soft, desaturated palette of greens and browns. A steady, rhythmic stream of vehicles, including large semi-trucks with long trailers and various cars, moves smoothly in both directions along the highway. Their headlights and taillights pierce the dense air, appearing as soft, glowing white and red spheres gliding along the pavement. As the camera steadily and deliberately pans upward, the individual vehicles shrink, their orderly procession becoming part of a larger, hypnotic pattern, intensifying the profound tranquility and slightly misty atmosphere of the still morning.", + "seed": 11452, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_5/index_8541_seed_11452_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01869963500648737, + "vlm_logprob": -0.01869963500648737, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.01869963500648737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014259468764066696 + ], + "opposite_logprob_tokens": [ + -5.764259338378906 + ], + "logprob_reward": -0.014259468764066696, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.014259468764066696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13637615740299225, + 0.07001298666000366, + 0.1373601108789444, + 0.09714998304843903, + 0.04443035274744034, + 0.22182519733905792, + 0.05755521357059479, + 0.07538207620382309, + 0.06052388623356819, + 0.09938409924507141 + ], + "credit_map": [ + 0.018017936497926712, + 0.023958712816238403, + 0.01164817251265049, + 0.011412279680371284, + 0.017315907403826714, + 0.022999923676252365, + 0.01637999527156353, + 0.014643223956227303, + 0.011052129790186882, + 0.009024389088153839, + 0.008333828300237656, + 0.007579578552395105, + 0.007111717946827412, + 0.015141299925744534, + 0.004798494279384613, + 0.006971541326493025, + 0.016115745529532433, + 0.01669958233833313, + 0.014704725705087185, + 0.009785659611225128, + 0.016680745407938957, + 0.03308768942952156, + 0.01715387962758541, + 0.0131320646032691, + 0.01318486500531435, + 0.019887836650013924, + 0.009885489009320736, + 0.007343406789004803, + 0.009054737165570259, + 0.02334747649729252, + 0.005901708733290434, + 0.008544460870325565, + 0.0051196846179664135, + 0.006603179033845663, + 0.007828821428120136, + 0.0060888733714818954, + 0.004245424177497625, + 0.006471872795373201, + 0.0031785471364855766, + 0.004893944598734379, + 0.009701978415250778, + 0.07119667530059814, + 0.010931349359452724, + 0.0161135196685791, + 0.014568348415195942, + 0.08091754466295242, + 0.007469748612493277, + 0.010926008224487305, + 0.004448265768587589, + 0.006654622498899698, + 0.0083332359790802, + 0.007655991706997156, + 0.007606384344398975, + 0.010638118721544743, + 0.0041798558086156845, + 0.008038732223212719, + 0.005257938988506794, + 0.009283991530537605, + 0.011561556719243526, + 0.009417401626706123, + 0.010424300096929073, + 0.01171807013452053, + 0.0074450853280723095, + 0.010273727588355541, + 0.004954094998538494, + 0.008446325547993183, + 0.007614707574248314, + 0.005580781027674675, + 0.011723197065293789, + 0.010089859366416931, + 0.0056792013347148895, + 0.006435716059058905, + 0.012064891867339611, + 0.014188784174621105, + 0.020009580999612808, + 0.009975459426641464, + 0.01202448271214962, + 0.012017155066132545, + 0.009813692420721054, + 0.009290050715208054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.014259468764066696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639005314558744 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007639005314558744, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639005314558744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09700972586870193, + 0.04371185228228569, + 0.12263964116573334, + 0.4240381717681885, + 0.0229206345975399, + 0.03685859590768814, + 0.021351132541894913, + 0.018137341365218163, + 0.029826274141669273, + 0.18350665271282196 + ], + "credit_map": [ + 0.012613046914339066, + 0.02310038171708584, + 0.007983257062733173, + 0.009523618035018444, + 0.009629188105463982, + 0.017678087577223778, + 0.007858380675315857, + 0.008623762987554073, + 0.005163886584341526, + 0.007699682842940092, + 0.003891241503879428, + 0.005209764000028372, + 0.007544531021267176, + 0.00557987904176116, + 0.005225075408816338, + 0.0033977932762354612, + 0.019517051056027412, + 0.013373399153351784, + 0.01764734648168087, + 0.007634364999830723, + 0.012509074993431568, + 0.017346784472465515, + 0.01938311941921711, + 0.015228494070470333, + 0.05552612990140915, + 0.09874901175498962, + 0.045171353965997696, + 0.04438049718737602, + 0.04286912456154823, + 0.07737969607114792, + 0.02353115938603878, + 0.0364312119781971, + 0.0017571705393493176, + 0.004799322225153446, + 0.003058838425204158, + 0.002111932495608926, + 0.0017928376328200102, + 0.00505472207441926, + 0.002026388654485345, + 0.002319422783330083, + 0.002974986331537366, + 0.00580566143617034, + 0.0043225665576756, + 0.006478321738541126, + 0.0045614647679030895, + 0.0060282135382294655, + 0.0027904172893613577, + 0.003896963782608509, + 0.0018216893076896667, + 0.002735961927101016, + 0.0026187673211097717, + 0.002114835660904646, + 0.003354255575686693, + 0.003972757142037153, + 0.00241572386585176, + 0.002317142440006137, + 0.0016543036326766014, + 0.002052042633295059, + 0.0024400814436376095, + 0.0020430744625627995, + 0.002047057496383786, + 0.00369153986684978, + 0.0019004902569577098, + 0.002308751456439495, + 0.0027800530660897493, + 0.004547934513539076, + 0.0030379523523151875, + 0.0026120042894035578, + 0.005003476049751043, + 0.006228297483175993, + 0.002537894994020462, + 0.0030786616262048483, + 0.015414461493492126, + 0.042522214353084564, + 0.013346747495234013, + 0.0101016191765666, + 0.0348094217479229, + 0.04107910022139549, + 0.015385829843580723, + 0.010847259312868118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.007639005314558744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264704581350088 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264704581350088, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264704581350088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23984168469905853, + 0.11462447047233582, + 0.15725892782211304, + 0.04724259674549103, + 0.03564465045928955, + 0.10527880489826202, + 0.08568327128887177, + 0.05791058763861656, + 0.05510443076491356, + 0.10141055285930634 + ], + "credit_map": [ + 0.023353368043899536, + 0.032244570553302765, + 0.02762359194457531, + 0.02617182396352291, + 0.028446786105632782, + 0.03210782632231712, + 0.040819864720106125, + 0.029073869809508324, + 0.022890929132699966, + 0.012244036421179771, + 0.006809360813349485, + 0.0067573669366538525, + 0.015557576902210712, + 0.04108025133609772, + 0.0039662932977080345, + 0.005318664945662022, + 0.035811807960271835, + 0.016789359971880913, + 0.010172843001782894, + 0.0069857388734817505, + 0.013524237088859081, + 0.057168539613485336, + 0.007990834303200245, + 0.008815581910312176, + 0.005481956526637077, + 0.006549963727593422, + 0.006066272035241127, + 0.004685261286795139, + 0.005024702288210392, + 0.009078707545995712, + 0.004046199843287468, + 0.006309536285698414, + 0.002817997010424733, + 0.00607184786349535, + 0.005877671297639608, + 0.003680879482999444, + 0.003217533929273486, + 0.005868888460099697, + 0.003480657236650586, + 0.004629176110029221, + 0.008229049853980541, + 0.017057286575436592, + 0.014869270846247673, + 0.019331136718392372, + 0.00799141637980938, + 0.01860588602721691, + 0.006609535776078701, + 0.012585225515067577, + 0.006199875380843878, + 0.009714657440781593, + 0.012761391699314117, + 0.015395010821521282, + 0.009674705564975739, + 0.013065565377473831, + 0.0069345226511359215, + 0.011937548406422138, + 0.004798564128577709, + 0.0070473807863891125, + 0.007610640488564968, + 0.005708122160285711, + 0.009525524452328682, + 0.009407845325767994, + 0.006307921838015318, + 0.007504589390009642, + 0.005146564915776253, + 0.006860164925456047, + 0.005982260685414076, + 0.005941376090049744, + 0.009949296712875366, + 0.007731080520898104, + 0.007552781607955694, + 0.005940909497439861, + 0.009006867185235023, + 0.02539648860692978, + 0.010506317019462585, + 0.005543283186852932, + 0.018331235274672508, + 0.0203081127256155, + 0.005375503562390804, + 0.0069427490234375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005264704581350088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286327719688416 + ], + "opposite_logprob_tokens": [ + -3.7732863426208496 + ], + "logprob_reward": -0.023286327719688416, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286327719688416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1817239224910736, + 0.0785665512084961, + 0.16278009116649628, + 0.12107568979263306, + 0.042310554534196854, + 0.15240342915058136, + 0.07754084467887878, + 0.03886302933096886, + 0.02947636879980564, + 0.11525952816009521 + ], + "credit_map": [ + 0.016956591978669167, + 0.03308522701263428, + 0.014419172890484333, + 0.018644334748387337, + 0.03187820315361023, + 0.026662427932024002, + 0.022951338440179825, + 0.017126621678471565, + 0.013380606658756733, + 0.007915213704109192, + 0.0039667487144470215, + 0.005210044328123331, + 0.01089965645223856, + 0.029792236164212227, + 0.0036800173111259937, + 0.003722026012837887, + 0.03490612655878067, + 0.018517419695854187, + 0.014149476774036884, + 0.008755938149988651, + 0.011388775892555714, + 0.047158949077129364, + 0.015003595501184464, + 0.012899818830192089, + 0.014785165898501873, + 0.026776468381285667, + 0.01873623952269554, + 0.011322882026433945, + 0.010203090496361256, + 0.01598285883665085, + 0.008192842826247215, + 0.015076138079166412, + 0.004879732616245747, + 0.005983497016131878, + 0.004224068950861692, + 0.003252581460401416, + 0.00631828885525465, + 0.009178318083286285, + 0.004319427534937859, + 0.004154643043875694, + 0.01034675631672144, + 0.03842261806130409, + 0.017289048060774803, + 0.014447520487010479, + 0.00978287123143673, + 0.04146754741668701, + 0.007544351276010275, + 0.013102725148200989, + 0.006497913505882025, + 0.009356309659779072, + 0.011284361593425274, + 0.014801890589296818, + 0.008954706601798534, + 0.008351879194378853, + 0.008713502436876297, + 0.009580272249877453, + 0.003555316012352705, + 0.004424617625772953, + 0.004265984520316124, + 0.004205205477774143, + 0.0068807778880000114, + 0.006001408211886883, + 0.004795053042471409, + 0.004734665155410767, + 0.0034679837990552187, + 0.00396488094702363, + 0.0030304896645247936, + 0.0029710556846112013, + 0.004112853668630123, + 0.004720109514892101, + 0.003760767634958029, + 0.0034482264891266823, + 0.00963797140866518, + 0.026537606492638588, + 0.00975645612925291, + 0.006406061351299286, + 0.02337982878088951, + 0.02068742737174034, + 0.01065196841955185, + 0.008202215656638145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.023286327719688416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430486686527729 + ], + "opposite_logprob_tokens": [ + -3.168048620223999 + ], + "logprob_reward": -0.0430486686527729, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430486686527729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1501394659280777, + 0.04589039459824562, + 0.11043606698513031, + 0.1134377270936966, + 0.02891269512474537, + 0.07917071133852005, + 0.06944413483142853, + 0.09409511834383011, + 0.053162045776844025, + 0.25531160831451416 + ], + "credit_map": [ + 0.01778932847082615, + 0.032840877771377563, + 0.009884696453809738, + 0.012464802712202072, + 0.012084105983376503, + 0.043495241552591324, + 0.012184325605630875, + 0.009396092034876347, + 0.0056014154106378555, + 0.008029483258724213, + 0.005486395210027695, + 0.004253712948411703, + 0.006102473940700293, + 0.0068389722146093845, + 0.004583544563502073, + 0.004994397051632404, + 0.011436555534601212, + 0.013587893918156624, + 0.01507661584764719, + 0.010319479741156101, + 0.012999866157770157, + 0.017343612387776375, + 0.01638377457857132, + 0.01328827440738678, + 0.013237294740974903, + 0.02265685610473156, + 0.011638624593615532, + 0.011533643119037151, + 0.01246113795787096, + 0.021258099004626274, + 0.009033022448420525, + 0.011619041673839092, + 0.00230873329564929, + 0.004693657625466585, + 0.0032020604703575373, + 0.002837789012119174, + 0.0030793272890150547, + 0.006648132111877203, + 0.003185893641784787, + 0.002957100747153163, + 0.005678222514688969, + 0.021507628262043, + 0.005454314406961203, + 0.0041965097188949585, + 0.009663579985499382, + 0.024304445832967758, + 0.0035025214310735464, + 0.004863497335463762, + 0.005194401368498802, + 0.012698928825557232, + 0.009776941500604153, + 0.006214455235749483, + 0.00836759153753519, + 0.012013321742415428, + 0.006101885810494423, + 0.009076605550944805, + 0.0060238586738705635, + 0.009193014353513718, + 0.014261649921536446, + 0.014711050316691399, + 0.012410497292876244, + 0.01778886467218399, + 0.009207552298903465, + 0.010498630814254284, + 0.005071515683084726, + 0.005968369543552399, + 0.0074188546277582645, + 0.00518029136583209, + 0.007717665750533342, + 0.008390193805098534, + 0.0053632925264537334, + 0.008051859214901924, + 0.026518484577536583, + 0.03446740657091141, + 0.055900439620018005, + 0.02861451916396618, + 0.028766505420207977, + 0.03294725716114044, + 0.021189114078879356, + 0.02690788172185421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0430486686527729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot unfolds across a multi-lane highway, its dark, wet asphalt carving a steep path across the landscape. The camera then begins a slow, graceful vertical ascent to reveal a wider bird's-eye view. The entire scene is blanketed by a thick, low-hanging white mist, creating an ethereal, diffuse glow that eliminates harsh shadows. This mist lingers over the surrounding fields, softening the outlines of distant trees and rendering the background a soft, desaturated palette of greens and browns. A steady, rhythmic stream of vehicles, including large semi-trucks with long trailers and various cars, moves smoothly in both directions along the highway. Their headlights and taillights pierce the dense air, appearing as soft, glowing white and red spheres gliding along the pavement. As the camera steadily and deliberately pans upward, the individual vehicles shrink, their orderly procession becoming part of a larger, hypnotic pattern, intensifying the profound tranquility and slightly misty atmosphere of the still morning.", + "seed": 11453, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_5/index_8541_seed_11453_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031937805190682414, + "vlm_logprob": -0.031937805190682414, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.031937805190682414, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026511404663324356 + ], + "opposite_logprob_tokens": [ + -3.9015114307403564 + ], + "logprob_reward": -0.026511404663324356, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026511404663324356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1452014446258545, + 0.03007591888308525, + 0.021164726465940475, + 0.01372804306447506, + 0.02665955014526844, + 0.16075390577316284, + 0.13482578098773956, + 0.39953118562698364, + 0.04394904896616936, + 0.024110380560159683 + ], + "credit_map": [ + 0.013107799924910069, + 0.020190194249153137, + 0.0207896139472723, + 0.01656399480998516, + 0.018869424238801003, + 0.02806372381746769, + 0.013678792864084244, + 0.01393789704889059, + 0.0026700689923018217, + 0.004037556704133749, + 0.003759415354579687, + 0.0024575458373874426, + 0.004283836577087641, + 0.005186706781387329, + 0.004212050698697567, + 0.003468737006187439, + 0.0017259682063013315, + 0.0018647455144673586, + 0.003745841793715954, + 0.0038916324265301228, + 0.0021365811116993427, + 0.0022698710672557354, + 0.0027206959202885628, + 0.0028093901928514242, + 0.001426183502189815, + 0.001543134218081832, + 0.0015148492529988289, + 0.001250151195563376, + 0.0018272004090249538, + 0.002247824799269438, + 0.0022062784992158413, + 0.0017124213045462966, + 0.0019808923825621605, + 0.0027285246178507805, + 0.003726756665855646, + 0.00487201614305377, + 0.003092518774792552, + 0.0026666310150176287, + 0.005102709401398897, + 0.002489499980583787, + 0.016303224489092827, + 0.0242210254073143, + 0.020794352516531944, + 0.020345168188214302, + 0.013821042142808437, + 0.027373652905225754, + 0.01701764203608036, + 0.02087780088186264, + 0.013209249824285507, + 0.02082330919802189, + 0.017873290926218033, + 0.013284134678542614, + 0.011537393555045128, + 0.022887688130140305, + 0.021145226433873177, + 0.014065493829548359, + 0.03479001671075821, + 0.05148036777973175, + 0.04296158254146576, + 0.08526325225830078, + 0.03976410999894142, + 0.04677087441086769, + 0.054347582161426544, + 0.04415339604020119, + 0.0034223543480038643, + 0.003673091297969222, + 0.0054306937381625175, + 0.01030620839446783, + 0.0048273541033267975, + 0.004735478200018406, + 0.006497645750641823, + 0.005056225229054689, + 0.0018381966510787606, + 0.0026919387746602297, + 0.0029333324637264013, + 0.002975860610604286, + 0.002320039551705122, + 0.0036258012987673283, + 0.003809988498687744, + 0.003915221896022558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.026511404663324356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108570583164692 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.01108570583164692, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108570583164692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2449800670146942, + 0.05550624057650566, + 0.06415040791034698, + 0.07864216715097427, + 0.050025396049022675, + 0.16992461681365967, + 0.13556332886219025, + 0.08430308848619461, + 0.09474721550941467, + 0.022157426923513412 + ], + "credit_map": [ + 0.0215793214738369, + 0.044500019401311874, + 0.031318794935941696, + 0.019307876005768776, + 0.04227456450462341, + 0.047740038484334946, + 0.019186757504940033, + 0.01907268352806568, + 0.0063433051109313965, + 0.007003377191722393, + 0.007588725071400404, + 0.005247393157333136, + 0.011065995320677757, + 0.007414440158754587, + 0.005804590415209532, + 0.005038418807089329, + 0.005928223021328449, + 0.006563330069184303, + 0.008499741554260254, + 0.006652092095464468, + 0.012820890173316002, + 0.007237941492348909, + 0.007696077693253756, + 0.008752110414206982, + 0.006114681717008352, + 0.008808132261037827, + 0.009387542493641376, + 0.011847791261970997, + 0.010347286239266396, + 0.01078831683844328, + 0.013176021166145802, + 0.008172390051186085, + 0.0029790301341563463, + 0.0039478475227952, + 0.006452771369367838, + 0.007095413748174906, + 0.005048973951488733, + 0.0065215155482292175, + 0.012440265156328678, + 0.0055395811796188354, + 0.011190238408744335, + 0.020147569477558136, + 0.022816507145762444, + 0.02734721079468727, + 0.013981001451611519, + 0.02452440746128559, + 0.030962523072957993, + 0.01895517110824585, + 0.010485658422112465, + 0.024537421762943268, + 0.01719954051077366, + 0.01868963986635208, + 0.012155449017882347, + 0.023147134110331535, + 0.016214951872825623, + 0.013133530505001545, + 0.009199826046824455, + 0.010874547995626926, + 0.009223894216120243, + 0.01604451797902584, + 0.010067524388432503, + 0.01056580152362585, + 0.010014505125582218, + 0.008312471210956573, + 0.010186037048697472, + 0.012687455862760544, + 0.00975867547094822, + 0.008539819158613682, + 0.008434773422777653, + 0.02616449072957039, + 0.00788787193596363, + 0.011088092811405659, + 0.0035860200878232718, + 0.003942403942346573, + 0.002433192217722535, + 0.0021782508119940758, + 0.002071038354188204, + 0.002239399356767535, + 0.002585608744993806, + 0.0031215143389999866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.01108570583164692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371458128094673 + ], + "opposite_logprob_tokens": [ + -3.408714532852173 + ], + "logprob_reward": -0.03371458128094673, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371458128094673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19432535767555237, + 0.06447591632604599, + 0.07834507524967194, + 0.05688784271478653, + 0.025180339813232422, + 0.24074910581111908, + 0.08811413496732712, + 0.11087886989116669, + 0.06231248751282692, + 0.07873085141181946 + ], + "credit_map": [ + 0.018294574692845345, + 0.028298523277044296, + 0.023947767913341522, + 0.016368180513381958, + 0.024371249601244926, + 0.03607403114438057, + 0.016388963907957077, + 0.030582070350646973, + 0.005623089149594307, + 0.007492904085665941, + 0.009563883766531944, + 0.006742514669895172, + 0.010793713852763176, + 0.009942060336470604, + 0.0064672078005969524, + 0.007850538939237595, + 0.0082404101267457, + 0.005914596375077963, + 0.01854396052658558, + 0.010333988815546036, + 0.011508642695844173, + 0.007583098951727152, + 0.00918568018823862, + 0.007034692447632551, + 0.00495462492108345, + 0.005366731435060501, + 0.0073148151859641075, + 0.006923805922269821, + 0.008061608299612999, + 0.008337270468473434, + 0.00830113049596548, + 0.007627853658050299, + 0.0021012192592024803, + 0.0026580598205327988, + 0.0033710631541907787, + 0.003798707388341427, + 0.0024267463013529778, + 0.0029737644363194704, + 0.00413413904607296, + 0.0037166401743888855, + 0.014501153491437435, + 0.02400510199368, + 0.032675690948963165, + 0.04150368645787239, + 0.021550875157117844, + 0.028046635910868645, + 0.05132737383246422, + 0.0271386057138443, + 0.005833896342664957, + 0.010043837130069733, + 0.012114867568016052, + 0.014800602570176125, + 0.011450924910604954, + 0.009554489515721798, + 0.015574978664517403, + 0.008740543387830257, + 0.012012279592454433, + 0.011908856220543385, + 0.014050252735614777, + 0.014731314964592457, + 0.009486312977969646, + 0.012161482125520706, + 0.014962694607675076, + 0.021565673872828484, + 0.004155714996159077, + 0.006134032737463713, + 0.008054317906498909, + 0.012513427995145321, + 0.005952086299657822, + 0.005714002531021833, + 0.009992500767111778, + 0.00979641079902649, + 0.0058184522204101086, + 0.00925140269100666, + 0.010592461563646793, + 0.019723858684301376, + 0.007816320285201073, + 0.008910153061151505, + 0.009347084909677505, + 0.007271117065101862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.03371458128094673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346080005168915 + ], + "opposite_logprob_tokens": [ + -3.65134596824646 + ], + "logprob_reward": -0.026346080005168915, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346080005168915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16681592166423798, + 0.262808620929718, + 0.09267602115869522, + 0.0718798041343689, + 0.04727713391184807, + 0.06315495073795319, + 0.07051683962345123, + 0.09930052608251572, + 0.09472083300352097, + 0.030849406495690346 + ], + "credit_map": [ + 0.017116887494921684, + 0.0254321601241827, + 0.02374676801264286, + 0.014125622808933258, + 0.02697744593024254, + 0.025307485833764076, + 0.01539356354624033, + 0.01871596649289131, + 0.027326421812176704, + 0.03036307543516159, + 0.03585168346762657, + 0.03126201778650284, + 0.03572189435362816, + 0.04185936599969864, + 0.028113268315792084, + 0.032310858368873596, + 0.011160848662257195, + 0.010843555442988873, + 0.011023415252566338, + 0.010555531829595566, + 0.01533727627247572, + 0.014434228651225567, + 0.010311584919691086, + 0.009009572677314281, + 0.005311582703143358, + 0.006448791362345219, + 0.010378038510680199, + 0.012089165858924389, + 0.008488411083817482, + 0.008024499751627445, + 0.012103685177862644, + 0.009035619907081127, + 0.003807907924056053, + 0.0037142131477594376, + 0.006437950301915407, + 0.0058645205572247505, + 0.004076748620718718, + 0.0044846199452877045, + 0.010164679028093815, + 0.008726486936211586, + 0.006614187732338905, + 0.008959120139479637, + 0.007375393528491259, + 0.008405827917158604, + 0.005093580111861229, + 0.010087466798722744, + 0.008740326389670372, + 0.007879039272665977, + 0.004993606358766556, + 0.007914389483630657, + 0.009982400573790073, + 0.01135689951479435, + 0.007136438507586718, + 0.006849282421171665, + 0.013456270098686218, + 0.008827543817460537, + 0.008259757421910763, + 0.010485106147825718, + 0.011718329042196274, + 0.02279273420572281, + 0.009161901660263538, + 0.012019898742437363, + 0.012041511945426464, + 0.012821272946894169, + 0.00998519454151392, + 0.011326766572892666, + 0.011020787991583347, + 0.012288358062505722, + 0.009066387079656124, + 0.019624290987849236, + 0.008506056852638721, + 0.012902975082397461, + 0.002635068027302623, + 0.0030075295362621546, + 0.0036166405770927668, + 0.0049769217148423195, + 0.0026796215679496527, + 0.0048194946721196175, + 0.005006225313991308, + 0.004107903689146042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.026346080005168915 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062031254172325134 + ], + "opposite_logprob_tokens": [ + -2.8120312690734863 + ], + "logprob_reward": -0.062031254172325134, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062031254172325134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10521639138460159, + 0.04009036719799042, + 0.06530874967575073, + 0.04951503127813339, + 0.054965145885944366, + 0.20245826244354248, + 0.22829772531986237, + 0.10077416896820068, + 0.10520215332508087, + 0.048172011971473694 + ], + "credit_map": [ + 0.00948023796081543, + 0.01465913187712431, + 0.021271992474794388, + 0.008258350193500519, + 0.009841048158705235, + 0.019020721316337585, + 0.010886005125939846, + 0.011798909865319729, + 0.00507833668962121, + 0.006347212940454483, + 0.004900474101305008, + 0.0032273600809276104, + 0.00597727345302701, + 0.007009665947407484, + 0.0033783172257244587, + 0.004171726759523153, + 0.0066642253659665585, + 0.007653889711946249, + 0.010485845617949963, + 0.007710267789661884, + 0.008600819855928421, + 0.008695113472640514, + 0.008035236969590187, + 0.007463349960744381, + 0.003223561681807041, + 0.004474399145692587, + 0.006540203001350164, + 0.0077427285723388195, + 0.006260834634304047, + 0.0077086398378014565, + 0.006477723829448223, + 0.0070869410410523415, + 0.003530891379341483, + 0.003988091368228197, + 0.008527246303856373, + 0.009392233565449715, + 0.005250308196991682, + 0.004470985848456621, + 0.011983728036284447, + 0.00782165676355362, + 0.012480937875807285, + 0.020811276510357857, + 0.028210356831550598, + 0.035354603081941605, + 0.018141936510801315, + 0.025014203041791916, + 0.04151967540383339, + 0.020925285294651985, + 0.018064457923173904, + 0.04465237632393837, + 0.02816130593419075, + 0.03722484037280083, + 0.016553547233343124, + 0.04021413251757622, + 0.024128245189785957, + 0.01929880864918232, + 0.008014422841370106, + 0.014177696779370308, + 0.010958441533148289, + 0.01594407856464386, + 0.00883420929312706, + 0.012493520975112915, + 0.01411592960357666, + 0.016235865652561188, + 0.011570905335247517, + 0.0139153515920043, + 0.011164460331201553, + 0.009800656698644161, + 0.008715864270925522, + 0.0254422165453434, + 0.008341701701283455, + 0.01625099405646324, + 0.0032139199320226908, + 0.007373721804469824, + 0.005900349002331495, + 0.0037669481243938208, + 0.004294141195714474, + 0.007313532289117575, + 0.006125164683908224, + 0.010184240527451038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.062031254172325134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot unfolds across a multi-lane highway, its dark, wet asphalt carving a steep path across the landscape. The camera then begins a slow, graceful vertical ascent to reveal a wider bird's-eye view. The entire scene is blanketed by a thick, low-hanging white mist, creating an ethereal, diffuse glow that eliminates harsh shadows. This mist lingers over the surrounding fields, softening the outlines of distant trees and rendering the background a soft, desaturated palette of greens and browns. A steady, rhythmic stream of vehicles, including large semi-trucks with long trailers and various cars, moves smoothly in both directions along the highway. Their headlights and taillights pierce the dense air, appearing as soft, glowing white and red spheres gliding along the pavement. As the camera steadily and deliberately pans upward, the individual vehicles shrink, their orderly procession becoming part of a larger, hypnotic pattern, intensifying the profound tranquility and slightly misty atmosphere of the still morning.", + "seed": 11454, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_5/index_8541_seed_11454_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030357948411256074, + "vlm_logprob": -0.030357948411256074, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.030357948411256074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01572154276072979 + ], + "opposite_logprob_tokens": [ + -5.765721321105957 + ], + "logprob_reward": -0.01572154276072979, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.01572154276072979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08653122931718826, + 0.05120747163891792, + 0.07739944756031036, + 0.2710292935371399, + 0.09193408489227295, + 0.1320411115884781, + 0.08709369599819183, + 0.12331152707338333, + 0.034993696957826614, + 0.04445842653512955 + ], + "credit_map": [ + 0.007455458864569664, + 0.010817124508321285, + 0.011728665791451931, + 0.005345781333744526, + 0.0158687774091959, + 0.019952697679400444, + 0.005334078334271908, + 0.010028647258877754, + 0.010589485988020897, + 0.005370184313505888, + 0.004292025696486235, + 0.002287398325279355, + 0.008257080800831318, + 0.006213456392288208, + 0.003661713795736432, + 0.010536122135818005, + 0.006563560571521521, + 0.02517864853143692, + 0.0074669597670435905, + 0.00514471810311079, + 0.01892095059156418, + 0.005716697312891483, + 0.0036117746494710445, + 0.004796134773641825, + 0.021827593445777893, + 0.08956892788410187, + 0.013943873345851898, + 0.01639026217162609, + 0.07091180980205536, + 0.022015169262886047, + 0.020931672304868698, + 0.015439978800714016, + 0.015955742448568344, + 0.012463771738111973, + 0.005513518117368221, + 0.010227231308817863, + 0.014990709722042084, + 0.014583105221390724, + 0.010957656428217888, + 0.0072423494420945644, + 0.019999511539936066, + 0.018896490335464478, + 0.007197507657110691, + 0.010613824240863323, + 0.022413315251469612, + 0.0269834753125906, + 0.01590905338525772, + 0.010027929209172726, + 0.007173630408942699, + 0.023872220888733864, + 0.005533706862479448, + 0.002765487413853407, + 0.028446057811379433, + 0.010222962126135826, + 0.005200021900236607, + 0.0038796155713498592, + 0.00819694809615612, + 0.02488415502011776, + 0.00829441286623478, + 0.007022571284323931, + 0.028955135494470596, + 0.02565721794962883, + 0.009271680377423763, + 0.011029409244656563, + 0.0034020603634417057, + 0.004775273147970438, + 0.0036429690662771463, + 0.004919285885989666, + 0.003399732057005167, + 0.004163691774010658, + 0.006918556988239288, + 0.0037721276748925447, + 0.003923763986676931, + 0.006814789026975632, + 0.004727379884570837, + 0.0054176426492631435, + 0.006569959223270416, + 0.00676317373290658, + 0.006603768561035395, + 0.003637948539108038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.01572154276072979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041077290661633015 + ], + "opposite_logprob_tokens": [ + -5.50410795211792 + ], + "logprob_reward": -0.0041077290661633015, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041077290661633015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07136522978544235, + 0.026121225208044052, + 0.05210583284497261, + 0.2927650213241577, + 0.36882561445236206, + 0.06591152399778366, + 0.02829706110060215, + 0.042305268347263336, + 0.04055311158299446, + 0.011750116012990475 + ], + "credit_map": [ + 0.014255992136895657, + 0.007308566011488438, + 0.005313983652740717, + 0.0037412468809634447, + 0.01898152567446232, + 0.011552741751074791, + 0.003692523343488574, + 0.006518653593957424, + 0.00464955298230052, + 0.0028310944326221943, + 0.0031792703084647655, + 0.0015068574575707316, + 0.003199697704985738, + 0.0046355533413589, + 0.00210483162663877, + 0.004014370031654835, + 0.006663583219051361, + 0.01688738726079464, + 0.0036276972386986017, + 0.0024010532069951296, + 0.011173652485013008, + 0.005718165077269077, + 0.0021373697090893984, + 0.003496926510706544, + 0.031076651066541672, + 0.08189181983470917, + 0.018499813973903656, + 0.021450908854603767, + 0.060267865657806396, + 0.03110475093126297, + 0.02822485938668251, + 0.02024834044277668, + 0.06502117961645126, + 0.05137389153242111, + 0.020731214433908463, + 0.0432514064013958, + 0.07107891887426376, + 0.046310730278491974, + 0.044162817299366, + 0.026895485818386078, + 0.008696849457919598, + 0.007926627062261105, + 0.004632044583559036, + 0.004678758326917887, + 0.00968064647167921, + 0.014702882617712021, + 0.009504908695816994, + 0.006088814232498407, + 0.003061021212488413, + 0.004756168462336063, + 0.0038173315115273, + 0.0016760679427534342, + 0.005551424343138933, + 0.0033475467935204506, + 0.0036415597423911095, + 0.002445943420752883, + 0.0031763329170644283, + 0.009983872063457966, + 0.0025121066719293594, + 0.002842297311872244, + 0.010203221812844276, + 0.007896815426647663, + 0.002608060836791992, + 0.0030825650319457054, + 0.003485205117613077, + 0.006592334248125553, + 0.003870496992021799, + 0.006275448482483625, + 0.004704887978732586, + 0.004774934146553278, + 0.0071303448639810085, + 0.003719460451975465, + 0.001256748684681952, + 0.002034092554822564, + 0.0009135284344665706, + 0.0012727394932880998, + 0.0020608901977539062, + 0.0012518236180767417, + 0.001800260622985661, + 0.0011600328143686056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0041077290661633015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059515379369259 + ], + "opposite_logprob_tokens": [ + -3.895595073699951 + ], + "logprob_reward": -0.02059515379369259, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059515379369259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20682837069034576, + 0.054400376975536346, + 0.08962243050336838, + 0.35020843148231506, + 0.02108139730989933, + 0.054158926010131836, + 0.08132940530776978, + 0.07918279618024826, + 0.045331209897994995, + 0.01785670779645443 + ], + "credit_map": [ + 0.04941856488585472, + 0.016570180654525757, + 0.01295968983322382, + 0.011424629017710686, + 0.04611164331436157, + 0.030638480558991432, + 0.008902023546397686, + 0.030803151428699493, + 0.00752036040648818, + 0.006439253222197294, + 0.006273906212300062, + 0.0038817422464489937, + 0.006851882673799992, + 0.009315974079072475, + 0.004239392466843128, + 0.009877864271402359, + 0.009105113334953785, + 0.029329929500818253, + 0.007253316231071949, + 0.0036948772612959146, + 0.022108688950538635, + 0.00969123374670744, + 0.0031357111874967813, + 0.005303555633872747, + 0.03625022992491722, + 0.09721476584672928, + 0.023628586903214455, + 0.0270934347063303, + 0.0711008757352829, + 0.03345198556780815, + 0.036724064499139786, + 0.02474445290863514, + 0.0026552455965429544, + 0.003141196444630623, + 0.0018932850798591971, + 0.0020724961068481207, + 0.0032974230125546455, + 0.003123454749584198, + 0.002937189768999815, + 0.001961103640496731, + 0.006213400047272444, + 0.00632574874907732, + 0.0034260975662618876, + 0.004120663274079561, + 0.007283058017492294, + 0.01357316505163908, + 0.008013839833438396, + 0.005202948115766048, + 0.008428375236690044, + 0.013536305166780949, + 0.009894580580294132, + 0.004443603567779064, + 0.01885613426566124, + 0.010278075933456421, + 0.009291844442486763, + 0.006600480526685715, + 0.0066846199333667755, + 0.009687962010502815, + 0.006762344855815172, + 0.007268284913152456, + 0.011334684677422047, + 0.0211633313447237, + 0.00721437344327569, + 0.009067182429134846, + 0.004116615280508995, + 0.006887651979923248, + 0.0035034900065511465, + 0.005711746402084827, + 0.005072263069450855, + 0.008106771856546402, + 0.007361959666013718, + 0.004570708144456148, + 0.00177671923302114, + 0.002306076930835843, + 0.0014621547888964415, + 0.0014625450130552053, + 0.0029590728227049112, + 0.0033072587102651596, + 0.0025199290830641985, + 0.0020629512146115303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.02059515379369259 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107332669198513 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.01107332669198513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107332669198513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06296469271183014, + 0.04119051247835159, + 0.04095815494656563, + 0.46837663650512695, + 0.056050002574920654, + 0.03134416416287422, + 0.09548038244247437, + 0.0745042935013771, + 0.099209725856781, + 0.029921384528279305 + ], + "credit_map": [ + 0.007148799020797014, + 0.007134373765438795, + 0.008064297027885914, + 0.003946603741496801, + 0.0141524663195014, + 0.012801824137568474, + 0.0038978292141109705, + 0.005818501114845276, + 0.005661528557538986, + 0.004923161119222641, + 0.004800480790436268, + 0.0027640904299914837, + 0.005609050393104553, + 0.009101755917072296, + 0.0035534314811229706, + 0.004777016583830118, + 0.003105308162048459, + 0.008715846575796604, + 0.005415653809905052, + 0.003910453990101814, + 0.009633611887693405, + 0.0034245101269334555, + 0.002265277784317732, + 0.0044874935410916805, + 0.04433891922235489, + 0.13898418843746185, + 0.029500510543584824, + 0.033453311771154404, + 0.1045411229133606, + 0.04168154299259186, + 0.0447750948369503, + 0.031101979315280914, + 0.01020217314362526, + 0.007209461648017168, + 0.0033556402195245028, + 0.006846658419817686, + 0.009740746580064297, + 0.007097717374563217, + 0.0072924247942864895, + 0.004305187147110701, + 0.004386797547340393, + 0.004556145053356886, + 0.0024880494456738234, + 0.0013847077498212457, + 0.0038248912896960974, + 0.0100589320063591, + 0.002134630223736167, + 0.0025100172497332096, + 0.0063597350381314754, + 0.02912801317870617, + 0.007074404042214155, + 0.0037854823749512434, + 0.027676694095134735, + 0.010024343617260456, + 0.006016509607434273, + 0.005415207706391811, + 0.005463850684463978, + 0.018213432282209396, + 0.004189526662230492, + 0.004457884933799505, + 0.018901251256465912, + 0.012796943075954914, + 0.005029045511037111, + 0.005452364217489958, + 0.008702518418431282, + 0.012617064639925957, + 0.009384925477206707, + 0.015124849043786526, + 0.010024218820035458, + 0.013767153024673462, + 0.019561639055609703, + 0.010027368552982807, + 0.0031242112163454294, + 0.004165856633335352, + 0.002645439701154828, + 0.003725528484210372, + 0.004406998865306377, + 0.004419631790369749, + 0.0048304470255970955, + 0.002603272907435894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.01107332669198513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029198974370956 + ], + "opposite_logprob_tokens": [ + -2.3502919673919678 + ], + "logprob_reward": -0.10029198974370956, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029198974370956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12206269055604935, + 0.040479373186826706, + 0.03713072091341019, + 0.31034380197525024, + 0.27760687470436096, + 0.06750036776065826, + 0.027997372671961784, + 0.022533025592565536, + 0.05502412095665932, + 0.03932172432541847 + ], + "credit_map": [ + 0.031266774982213974, + 0.009642506949603558, + 0.006724623031914234, + 0.006765210535377264, + 0.02766505628824234, + 0.015992408618330956, + 0.005173469427973032, + 0.01883263699710369, + 0.0074644917622208595, + 0.004377701785415411, + 0.004733766429126263, + 0.002048684749752283, + 0.005169716197997332, + 0.008611992932856083, + 0.0032289668451994658, + 0.00484405318275094, + 0.005282547324895859, + 0.0053693195804953575, + 0.00450108852237463, + 0.0027301416266709566, + 0.00453369552269578, + 0.0069229863584041595, + 0.0036242236383259296, + 0.004166719503700733, + 0.028351353481411934, + 0.09266969561576843, + 0.01918819546699524, + 0.02204526588320732, + 0.07079748064279556, + 0.026881946250796318, + 0.030034247785806656, + 0.02037559449672699, + 0.05123061686754227, + 0.03720534220337868, + 0.01653786189854145, + 0.03431016579270363, + 0.049681518226861954, + 0.0333402045071125, + 0.03433860465884209, + 0.020962553098797798, + 0.0076958960853517056, + 0.008624738082289696, + 0.004828054457902908, + 0.004751480650156736, + 0.007282264996320009, + 0.017532475292682648, + 0.009968291968107224, + 0.006817166693508625, + 0.0034748888574540615, + 0.0034544351510703564, + 0.003494189353659749, + 0.0017080720281228423, + 0.005521564278751612, + 0.004974124953150749, + 0.002865225076675415, + 0.0025048749521374702, + 0.002712859772145748, + 0.003254799172282219, + 0.002221934264525771, + 0.0021247428376227617, + 0.004449594300240278, + 0.003887053346261382, + 0.0022997313644737005, + 0.0015823084395378828, + 0.005244739819318056, + 0.006416229996830225, + 0.005281925667077303, + 0.008469931781291962, + 0.005354167893528938, + 0.007295860908925533, + 0.011400628834962845, + 0.0055606430396437645, + 0.003730385797098279, + 0.0044915457256138325, + 0.004183190409094095, + 0.005398359149694443, + 0.004527998622506857, + 0.004823734983801842, + 0.007878650911152363, + 0.004287857562303543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.10029198974370956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot unfolds across a multi-lane highway, its dark, wet asphalt carving a steep path across the landscape. The camera then begins a slow, graceful vertical ascent to reveal a wider bird's-eye view. The entire scene is blanketed by a thick, low-hanging white mist, creating an ethereal, diffuse glow that eliminates harsh shadows. This mist lingers over the surrounding fields, softening the outlines of distant trees and rendering the background a soft, desaturated palette of greens and browns. A steady, rhythmic stream of vehicles, including large semi-trucks with long trailers and various cars, moves smoothly in both directions along the highway. Their headlights and taillights pierce the dense air, appearing as soft, glowing white and red spheres gliding along the pavement. As the camera steadily and deliberately pans upward, the individual vehicles shrink, their orderly procession becoming part of a larger, hypnotic pattern, intensifying the profound tranquility and slightly misty atmosphere of the still morning.", + "seed": 11455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_5/index_8541_seed_11455_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011718990840017795, + "vlm_logprob": -0.011718990840017795, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.011718990840017795, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016310816630721092 + ], + "opposite_logprob_tokens": [ + -5.266310691833496 + ], + "logprob_reward": -0.016310816630721092, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.016310816630721092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16424088180065155, + 0.08235279470682144, + 0.05888861417770386, + 0.08451874554157257, + 0.12005043029785156, + 0.08578738570213318, + 0.09168346971273422, + 0.10425769537687302, + 0.07946689426898956, + 0.12875309586524963 + ], + "credit_map": [ + 0.019705897197127342, + 0.029000038281083107, + 0.011193887330591679, + 0.01784978248178959, + 0.02000436745584011, + 0.03740480914711952, + 0.01783934235572815, + 0.011242765001952648, + 0.007801315747201443, + 0.007986955344676971, + 0.006848715711385012, + 0.012247796170413494, + 0.009957097470760345, + 0.014262845739722252, + 0.01577332615852356, + 0.007474740967154503, + 0.010084211826324463, + 0.007531492039561272, + 0.006146077997982502, + 0.007167532574385405, + 0.009460091590881348, + 0.008557206951081753, + 0.003727005096152425, + 0.006214997265487909, + 0.01890954189002514, + 0.013258817605674267, + 0.007658565882593393, + 0.011898169293999672, + 0.0097893625497818, + 0.009298156946897507, + 0.005549976136535406, + 0.008156155236065388, + 0.010917146690189838, + 0.034944064915180206, + 0.007513913791626692, + 0.008925247937440872, + 0.027637701481580734, + 0.015395868569612503, + 0.007281691301614046, + 0.00743479747325182, + 0.010884682647883892, + 0.017509231343865395, + 0.009011195041239262, + 0.003375023603439331, + 0.01722458191215992, + 0.014529548585414886, + 0.0055700019001960754, + 0.007683124393224716, + 0.012633788399398327, + 0.016597004607319832, + 0.010372371412813663, + 0.00376794277690351, + 0.01845896616578102, + 0.014386958442628384, + 0.005904719233512878, + 0.00956171378493309, + 0.009227719157934189, + 0.029273679479956627, + 0.011411435902118683, + 0.0038654827512800694, + 0.03015475533902645, + 0.007228957489132881, + 0.004407047294080257, + 0.008688618429005146, + 0.010599362663924694, + 0.016234684735536575, + 0.007094251457601786, + 0.007833807729184628, + 0.01550685428082943, + 0.012055283412337303, + 0.003758865175768733, + 0.00638378644362092, + 0.014076834544539452, + 0.020281877368688583, + 0.008766243234276772, + 0.015948180109262466, + 0.02129555307328701, + 0.03185955062508583, + 0.008707677945494652, + 0.007817180827260017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.016310816630721092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646931774914265 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646931774914265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646931774914265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04772791638970375, + 0.02549664117395878, + 0.04323084279894829, + 0.04031147062778473, + 0.07942145317792892, + 0.04975815489888191, + 0.043689578771591187, + 0.3897897005081177, + 0.14093135297298431, + 0.13964292407035828 + ], + "credit_map": [ + 0.008551300503313541, + 0.008330912329256535, + 0.002849831711500883, + 0.004373371601104736, + 0.004518243949860334, + 0.008037731051445007, + 0.006658766884356737, + 0.004407758824527264, + 0.0033026051241904497, + 0.003433158388361335, + 0.001647808589041233, + 0.0024117545690387487, + 0.006113963667303324, + 0.003766753477975726, + 0.003031027503311634, + 0.001789570669643581, + 0.006778445560485125, + 0.00747278518974781, + 0.0033238015603274107, + 0.003130103927105665, + 0.0063830092549324036, + 0.00954416673630476, + 0.0024432737845927477, + 0.004155258648097515, + 0.006051057484000921, + 0.009835895150899887, + 0.0034163862001150846, + 0.003227664390578866, + 0.00542029831558466, + 0.005984894465655088, + 0.0025577079504728317, + 0.0038175650406628847, + 0.007021347526460886, + 0.017523590475320816, + 0.005210469011217356, + 0.008347609080374241, + 0.020406575873494148, + 0.008748156018555164, + 0.005751533899456263, + 0.006412168499082327, + 0.006027966737747192, + 0.007636545225977898, + 0.005114790517836809, + 0.003273091511800885, + 0.010048492811620235, + 0.008689368143677711, + 0.0035853837616741657, + 0.005382516887038946, + 0.006734118331223726, + 0.007011862006038427, + 0.004957620054483414, + 0.0027077808044850826, + 0.0055917096324265, + 0.009001455269753933, + 0.0029190329369157553, + 0.004765999037772417, + 0.02133900858461857, + 0.15461350977420807, + 0.01054662186652422, + 0.009822013787925243, + 0.11224006116390228, + 0.06252966076135635, + 0.008854905143380165, + 0.0098439184948802, + 0.014687053859233856, + 0.025204619392752647, + 0.012854966334998608, + 0.01851731166243553, + 0.0256429985165596, + 0.018380843102931976, + 0.012660831212997437, + 0.012982740066945553, + 0.015831103548407555, + 0.028439393267035484, + 0.011007468216121197, + 0.015580732375383377, + 0.030576974153518677, + 0.021223757416009903, + 0.00816506240516901, + 0.008818431757390499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.004646931774914265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262095481157303 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005262095481157303, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262095481157303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1310436874628067, + 0.054465390741825104, + 0.039496567100286484, + 0.04178229719400406, + 0.09980935603380203, + 0.1560118943452835, + 0.13523554801940918, + 0.15090139210224152, + 0.0825127512216568, + 0.10874113440513611 + ], + "credit_map": [ + 0.01709025911986828, + 0.024139780551195145, + 0.008537881076335907, + 0.018624162301421165, + 0.014205500483512878, + 0.02019358240067959, + 0.014057512395083904, + 0.014194999821484089, + 0.006038902793079615, + 0.006523075979202986, + 0.005463240668177605, + 0.00981399416923523, + 0.0059659299440681934, + 0.008939174935221672, + 0.006465698592364788, + 0.005255375988781452, + 0.004376412369310856, + 0.0043060313910245895, + 0.0038926966954022646, + 0.00569481635466218, + 0.005342683754861355, + 0.007449304684996605, + 0.004391293972730637, + 0.0040433285757899284, + 0.005197202321141958, + 0.006081725936383009, + 0.004658204969018698, + 0.005338226445019245, + 0.005186461843550205, + 0.006936354096978903, + 0.0035409971605986357, + 0.004843126051127911, + 0.009490236639976501, + 0.024664757773280144, + 0.008709946647286415, + 0.0068662031553685665, + 0.012157075107097626, + 0.020272299647331238, + 0.00868424866348505, + 0.00896458514034748, + 0.011939266696572304, + 0.05073365196585655, + 0.008726786822080612, + 0.0070024034939706326, + 0.04542654752731323, + 0.015838297083973885, + 0.008631573058664799, + 0.007713367231190205, + 0.012544004246592522, + 0.031036118045449257, + 0.01104339212179184, + 0.00655230600386858, + 0.03928393870592117, + 0.017016274854540825, + 0.008644222281873226, + 0.009115285240113735, + 0.01115215104073286, + 0.05244360491633415, + 0.0067932563833892345, + 0.0054027000442147255, + 0.05368087440729141, + 0.010571318678557873, + 0.005626700818538666, + 0.005230783950537443, + 0.009880085475742817, + 0.01429903507232666, + 0.007842734456062317, + 0.012338666245341301, + 0.012603910639882088, + 0.013098244555294514, + 0.006299592554569244, + 0.006150487344712019, + 0.011210759170353413, + 0.018086379393935204, + 0.0096199419349432, + 0.01354971807450056, + 0.02102707512676716, + 0.018119443207979202, + 0.007471374701708555, + 0.009656441397964954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.005262095481157303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14082136750221252, + 0.044120289385318756, + 0.10118507593870163, + 0.042793918401002884, + 0.04971679300069809, + 0.05136556178331375, + 0.08954668045043945, + 0.2883572280406952, + 0.15496313571929932, + 0.037129953503608704 + ], + "credit_map": [ + 0.017296867445111275, + 0.03117208741605282, + 0.007983827963471413, + 0.017360374331474304, + 0.017772948369383812, + 0.02394837886095047, + 0.016884000971913338, + 0.008402873761951923, + 0.005275979172438383, + 0.0052366615273058414, + 0.004362280946224928, + 0.004830632358789444, + 0.0053892130963504314, + 0.006905821152031422, + 0.005993426311761141, + 0.006126279477030039, + 0.018747271969914436, + 0.019049109891057014, + 0.008986197412014008, + 0.00748112378641963, + 0.014393914490938187, + 0.016239238902926445, + 0.00842088833451271, + 0.007867326959967613, + 0.00670624990016222, + 0.009769884869456291, + 0.002674828050658107, + 0.0032646318431943655, + 0.007184249814599752, + 0.008271063677966595, + 0.00279413303360343, + 0.002128878142684698, + 0.0063873641192913055, + 0.005627557169646025, + 0.005225991830229759, + 0.005015759263187647, + 0.008188126608729362, + 0.009784916415810585, + 0.003920113202184439, + 0.005566962994635105, + 0.004592713434249163, + 0.01389263104647398, + 0.003917105495929718, + 0.0023214842658489943, + 0.015481270849704742, + 0.004563219845294952, + 0.0026984382420778275, + 0.00389869580976665, + 0.007820171304047108, + 0.020131222903728485, + 0.006204793695360422, + 0.0037672147154808044, + 0.02846696600317955, + 0.011977083049714565, + 0.0059149786829948425, + 0.0052642496302723885, + 0.015209811739623547, + 0.11299000680446625, + 0.009946387261152267, + 0.007678911555558443, + 0.09106248617172241, + 0.03719373419880867, + 0.007368046324700117, + 0.006907849106937647, + 0.018996840342879295, + 0.03193503990769386, + 0.014900374226272106, + 0.013724289834499359, + 0.032540298998355865, + 0.025920627638697624, + 0.0062163472175598145, + 0.010729318484663963, + 0.004262077156454325, + 0.009480641223490238, + 0.00261776358820498, + 0.0033595911227166653, + 0.006477493327111006, + 0.006283257622271776, + 0.002280840650200844, + 0.002368292538449168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.014194712042808533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-lane highway with dark, wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment blanketed by a thick, low-hanging white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large semi-trucks with long trailers and various cars visible moving in both directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle headlights and taillights appear as soft, glowing white and red spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful vertical ascent to reveal a wider bird's-eye view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10354311764240265, + 0.0326058492064476, + 0.07609028369188309, + 0.0660693496465683, + 0.09852965176105499, + 0.13067492842674255, + 0.14684253931045532, + 0.10537158697843552, + 0.13802500069141388, + 0.10224761068820953 + ], + "credit_map": [ + 0.01339723914861679, + 0.01868116669356823, + 0.006230000406503677, + 0.009749013930559158, + 0.014497067779302597, + 0.017168693244457245, + 0.015129998326301575, + 0.008689936250448227, + 0.0035822312347590923, + 0.0034086720552295446, + 0.00325135700404644, + 0.004887001123279333, + 0.005054614972323179, + 0.005840627010911703, + 0.0038383430801331997, + 0.0027430045884102583, + 0.009503785520792007, + 0.010320193134248257, + 0.00674719363451004, + 0.013399169780313969, + 0.009829238057136536, + 0.01182925421744585, + 0.0068826014176011086, + 0.007578844670206308, + 0.013492023572325706, + 0.010648406110703945, + 0.005943163298070431, + 0.00825970433652401, + 0.008742053061723709, + 0.010903147980570793, + 0.003642357187345624, + 0.004438501782715321, + 0.013458979316055775, + 0.012509092688560486, + 0.009916364215314388, + 0.014020011760294437, + 0.01912212185561657, + 0.012043926864862442, + 0.0072572678327560425, + 0.010201890021562576, + 0.015306918881833553, + 0.031441591680049896, + 0.013161959126591682, + 0.004737775307148695, + 0.026071378961205482, + 0.02025216445326805, + 0.007620193529874086, + 0.012082953937351704, + 0.018134402111172676, + 0.033199492841959, + 0.009450908750295639, + 0.006439219694584608, + 0.0395597480237484, + 0.02227533981204033, + 0.007836584001779556, + 0.009946845471858978, + 0.012426179833710194, + 0.02245698682963848, + 0.008966631256043911, + 0.006064855959266424, + 0.02469305321574211, + 0.018290270119905472, + 0.004136331379413605, + 0.008337283506989479, + 0.017217271029949188, + 0.031327951699495316, + 0.009582272730767727, + 0.01547933742403984, + 0.027712257578969002, + 0.016213737428188324, + 0.009692751802504063, + 0.010799422860145569, + 0.010608992539346218, + 0.019399581477046013, + 0.007393138948827982, + 0.009543504565954208, + 0.0217601265758276, + 0.0221713874489069, + 0.005590588785707951, + 0.0057802945375442505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.018180398270487785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_tensor([9698, 9698])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_tensor([9698, 9698])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..951c39d2d03b6e01331865a5ee19d9b6dfedb1f8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_tensor([9698, 9698])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, static, low-angle, tightly framed shot captures the scene of a massive natural force with stark documentary realism, captured with a high-speed camera. The sky is plunged into the mysterious twilight of a total solar eclipse, a deep, ominous indigo. The sun is a perfect black disk, haloed by the ghostly, shimmering white tendrils of the corona. A massive bolt of lightning tears through this backdrop, descending in breathtaking slow motion. Its core channel is a searing, overexposed white, from which intricate filaments of purple plasma branch out, like crystalline structures, deliberately crawling across the dark sky. The bolt's trajectory terminates precisely on the slender metal antenna spire of a towering skyscraper, its steel structure silhouetted against the strange light. The moment the main tendril makes contact, a violent and silent eruption of fiery white and orange sparks erupts from the tip, cascading down like a fiery waterfall, landing on the building's dark facade, each molten particle rendered in surreal clarity. The entire event unfolds with a sense of detached observation, creating an atmosphere of awe-inspiring and terrifying beauty that underscores the raw, untamed physics of the phenomenon.", + "seed": 14948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_9698_seed_14948_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1366426463238895, + "vlm_logprob": -0.1366426463238895, + "vlm_yes_no_margin": 3.025000011920929, + "vlm_token_credit": -0.1366426463238895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4665284752845764 + ], + "opposite_logprob_tokens": [ + -1.2165284156799316 + ], + "logprob_reward": -0.4665284752845764, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.4665284752845764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6593500375747681, + 0.02779833972454071, + 0.0865737721323967, + 0.03241807222366333, + 0.022797534242272377, + 0.011544794775545597, + 0.009647573344409466, + 0.011589712463319302, + 0.12660646438598633, + 0.011673646047711372 + ], + "credit_map": [ + 0.008064964786171913, + 0.011764240451157093, + 0.016754955053329468, + 0.0042998334392905235, + 0.3058917820453644, + 0.28381574153900146, + 0.01196292694658041, + 0.016795609146356583, + 0.004085589665919542, + 0.0017256656428799033, + 0.002136382507160306, + 0.002170413266867399, + 0.005286280531436205, + 0.004398830700665712, + 0.0038972804322838783, + 0.004097900353372097, + 0.006376409437507391, + 0.005675035994499922, + 0.004235904198139906, + 0.005417936015874147, + 0.027661075815558434, + 0.020009633153676987, + 0.008081109263002872, + 0.009116673842072487, + 0.0017865312984213233, + 0.0016003318596631289, + 0.0013702514115720987, + 0.0011412070598453283, + 0.011755729094147682, + 0.009656170383095741, + 0.0024945540353655815, + 0.0026132985949516296, + 0.0014025205746293068, + 0.0011900983517989516, + 0.001559196156449616, + 0.0011543622240424156, + 0.003783544059842825, + 0.002975172596052289, + 0.006583275739103556, + 0.004149369429796934, + 0.001487540197558701, + 0.0008841297822073102, + 0.0017209891229867935, + 0.0012200771598145366, + 0.001286093844100833, + 0.0012332601472735405, + 0.0023089551832526922, + 0.0014037499204277992, + 0.0010704000014811754, + 0.0011883623665198684, + 0.0009648743434809148, + 0.0010000241454690695, + 0.0010955550242215395, + 0.0009091335232369602, + 0.0019303373992443085, + 0.0014888878213241696, + 0.0012180703924968839, + 0.0011853304458782077, + 0.0013967800186946988, + 0.00113517208956182, + 0.0023400711361318827, + 0.0016675230581313372, + 0.0010244526201859117, + 0.0016223132843151689, + 0.0030483491718769073, + 0.0029819507617503405, + 0.003277076408267021, + 0.0017392357112839818, + 0.05828192085027695, + 0.052004553377628326, + 0.002517213812097907, + 0.0027561758179217577, + 0.0008799602510407567, + 0.0006406522006727755, + 0.0014636230189353228, + 0.0007213408243842423, + 0.002064795233309269, + 0.002157491398975253, + 0.0012238622875884175, + 0.002521922579035163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.4665284752845764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01425782311707735 + ], + "opposite_logprob_tokens": [ + -4.264257907867432 + ], + "logprob_reward": -0.01425782311707735, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01425782311707735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7790960669517517, + 0.032725777477025986, + 0.044446755200624466, + 0.015136992558836937, + 0.008836282417178154, + 0.0038254016544669867, + 0.004834541119635105, + 0.01158662885427475, + 0.07255160063505173, + 0.026959896087646484 + ], + "credit_map": [ + 0.009909949265420437, + 0.01409296877682209, + 0.020043710246682167, + 0.004479007795453072, + 0.36514514684677124, + 0.33758848905563354, + 0.011602838523685932, + 0.01623399928212166, + 0.00437961146235466, + 0.002319695893675089, + 0.0025369818322360516, + 0.0033659515902400017, + 0.006350015755742788, + 0.005449499469250441, + 0.003417652565985918, + 0.004906372167170048, + 0.006776970345526934, + 0.0032821116037666798, + 0.00275382399559021, + 0.0032518254593014717, + 0.012884783558547497, + 0.00890280306339264, + 0.0030465053860098124, + 0.0035479385405778885, + 0.0012601452181115746, + 0.0011505698785185814, + 0.0007928271079435945, + 0.0010304991155862808, + 0.004804724827408791, + 0.0037071495316922665, + 0.0013335092226043344, + 0.0010575690539553761, + 0.0010340390726923943, + 0.000898657541256398, + 0.0010164505802094936, + 0.0010954375611618161, + 0.000979623873718083, + 0.0008710877154953778, + 0.0017086159205064178, + 0.0012323702685534954, + 0.0005558175034821033, + 0.0004917489131912589, + 0.000600198341999203, + 0.0004193946660961956, + 0.0005020801909267902, + 0.0004544119874481112, + 0.0003742020344361663, + 0.00042754842434078455, + 0.0005406326963566244, + 0.00037019848241470754, + 0.0004195035435259342, + 0.00037889540544711053, + 0.0010243748547509313, + 0.0007965287077240646, + 0.0005145990871824324, + 0.0007898086332716048, + 0.0010343946050852537, + 0.0009488231735303998, + 0.0009930599480867386, + 0.0008968750480562449, + 0.0025729697663336992, + 0.0020737096201628447, + 0.0021023035515099764, + 0.0009644932579249144, + 0.013986255042254925, + 0.0050329361110925674, + 0.006245696917176247, + 0.0042641013860702515, + 0.015161318704485893, + 0.016624923795461655, + 0.005057233851402998, + 0.00617913668975234, + 0.0023572039790451527, + 0.0010274924570694566, + 0.002351514995098114, + 0.001189177855849266, + 0.008070266805589199, + 0.00786406546831131, + 0.001541233155876398, + 0.002558943349868059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.01425782311707735 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651796538382769 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004651796538382769, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651796538382769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46789756417274475, + 0.0403270348906517, + 0.0274871364235878, + 0.059691403061151505, + 0.019745048135519028, + 0.031222857534885406, + 0.03385702148079872, + 0.10754048079252243, + 0.15613478422164917, + 0.0560966357588768 + ], + "credit_map": [ + 0.0127007020637393, + 0.008319219574332237, + 0.01206244807690382, + 0.007113546133041382, + 0.20910561084747314, + 0.19633764028549194, + 0.010636121965944767, + 0.011622303165495396, + 0.007826631888747215, + 0.0026870882138609886, + 0.003967864438891411, + 0.0035443799570202827, + 0.006437011528760195, + 0.005623042583465576, + 0.005240547005087137, + 0.005000470206141472, + 0.003579121781513095, + 0.0019054714357480407, + 0.0020893895998597145, + 0.0017375452443957329, + 0.003942529205232859, + 0.003597347065806389, + 0.0047742100432515144, + 0.005861521698534489, + 0.009210062213242054, + 0.008294781669974327, + 0.0057252622209489346, + 0.005610241089016199, + 0.01031853910535574, + 0.007501471787691116, + 0.00651462422683835, + 0.006516418885439634, + 0.002622018102556467, + 0.0018926183693110943, + 0.002111525507643819, + 0.0031809438951313496, + 0.002292866352945566, + 0.001880949828773737, + 0.003171405056491494, + 0.0025927210226655006, + 0.003421164583414793, + 0.006128925364464521, + 0.003966860473155975, + 0.005096732173115015, + 0.004038793966174126, + 0.0031887090299278498, + 0.003073418280109763, + 0.0023082545958459377, + 0.006851612590253353, + 0.003002334153279662, + 0.004059843719005585, + 0.002008845331147313, + 0.004342248197644949, + 0.0036965543404221535, + 0.004658285994082689, + 0.005237296689301729, + 0.010394264943897724, + 0.0073099285364151, + 0.007426541764289141, + 0.00706647289916873, + 0.03271745145320892, + 0.027381692081689835, + 0.00905575230717659, + 0.00618838332593441, + 0.019100472331047058, + 0.009326719678938389, + 0.013175681233406067, + 0.009366843849420547, + 0.05097123607993126, + 0.04108226299285889, + 0.004812741652131081, + 0.008298824541270733, + 0.005383457988500595, + 0.003422939218580723, + 0.0077809784561395645, + 0.0037791964132338762, + 0.010566341690719128, + 0.00984527263790369, + 0.0067359977401793, + 0.008582448586821556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.004651796538382769 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054963212460279465 + ], + "opposite_logprob_tokens": [ + -2.9299631118774414 + ], + "logprob_reward": -0.054963212460279465, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054963212460279465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.76085364818573, + 0.020330043509602547, + 0.01631281152367592, + 0.02034873142838478, + 0.02803506888449192, + 0.0062194811180233955, + 0.028648698702454567, + 0.04272305965423584, + 0.045164890587329865, + 0.03136362135410309 + ], + "credit_map": [ + 0.00965110957622528, + 0.013954663649201393, + 0.019813131541013718, + 0.0044329664669930935, + 0.35653987526893616, + 0.3281642496585846, + 0.012116008438169956, + 0.016181670129299164, + 0.0017724764766171575, + 0.0008048660238273442, + 0.0012751497561112046, + 0.0010064279194921255, + 0.004976608790457249, + 0.004416892770677805, + 0.003414483042433858, + 0.0026631366927176714, + 0.0014110748888924718, + 0.0011979371774941683, + 0.0012311727041378617, + 0.0011557417456060648, + 0.0028999082278460264, + 0.0021765148267149925, + 0.002733358647674322, + 0.0035071035381406546, + 0.002110553439706564, + 0.002258486580103636, + 0.0013945619575679302, + 0.002018675673753023, + 0.003865891369059682, + 0.0031768360640853643, + 0.002639478538185358, + 0.0028842464089393616, + 0.001348858349956572, + 0.001306246966123581, + 0.0020216242410242558, + 0.0012328249868005514, + 0.004052570089697838, + 0.0041074673645198345, + 0.009686448611319065, + 0.004279027692973614, + 0.0006860932917334139, + 0.00042628918890841305, + 0.0007072592270560563, + 0.0005084457225166261, + 0.0009216238977387547, + 0.0006353829521685839, + 0.001101117697544396, + 0.0012332689948379993, + 0.004342020954936743, + 0.002223796444013715, + 0.002812884049490094, + 0.0019313765224069357, + 0.005965563002973795, + 0.004694719798862934, + 0.002719418378546834, + 0.003958920016884804, + 0.004486347548663616, + 0.004080779850482941, + 0.005074354354292154, + 0.00417583342641592, + 0.005700123030692339, + 0.0046297721564769745, + 0.010037688538432121, + 0.004538159817457199, + 0.006226004101336002, + 0.0028035191353410482, + 0.00345207704231143, + 0.002492963569238782, + 0.014098990708589554, + 0.011726993136107922, + 0.001864802441559732, + 0.0024995445273816586, + 0.0030138003639876842, + 0.0015919619472697377, + 0.0033979127183556557, + 0.0019777948036789894, + 0.00557839497923851, + 0.006241627503186464, + 0.005120746325701475, + 0.004441383294761181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.054963212460279465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281192421913147 + ], + "opposite_logprob_tokens": [ + -2.0178120136260986 + ], + "logprob_reward": -0.14281192421913147, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14281192421913147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6380680799484253, + 0.02467784658074379, + 0.035161182284355164, + 0.014621228910982609, + 0.008127733133733273, + 0.005344236735254526, + 0.01824929751455784, + 0.04887247830629349, + 0.17964375019073486, + 0.027234196662902832 + ], + "credit_map": [ + 0.009324410930275917, + 0.011598872020840645, + 0.015581737272441387, + 0.0033622391056269407, + 0.29821261763572693, + 0.27661973237991333, + 0.009380179457366467, + 0.013988290913403034, + 0.00446715485304594, + 0.0018253903836011887, + 0.0023719428572803736, + 0.002811110345646739, + 0.004162850324064493, + 0.0033359399531036615, + 0.0022996393963694572, + 0.0034038180019706488, + 0.00451402785256505, + 0.002242614282295108, + 0.0021008194889873266, + 0.0020774572622030973, + 0.010780135169625282, + 0.007711844053119421, + 0.0029152613133192062, + 0.0028190240263938904, + 0.0013158840592950583, + 0.0008322575595229864, + 0.0009804280707612634, + 0.0010157681535929441, + 0.003957421984523535, + 0.003260269993916154, + 0.0016543621895834804, + 0.001604837249033153, + 0.0006180216441862285, + 0.0007566826534457505, + 0.0005930220941081643, + 0.0006813189829699695, + 0.000907580484636128, + 0.000975675939116627, + 0.0020143105648458004, + 0.0015811201883479953, + 0.0007054779562167823, + 0.0006930529489181936, + 0.0006697001517750323, + 0.0005513965152204037, + 0.000947087537497282, + 0.0006770342006348073, + 0.0005751188145950437, + 0.000525368785019964, + 0.001369147328659892, + 0.0017664993647485971, + 0.0020213674288243055, + 0.0013710161438211799, + 0.003756148274987936, + 0.002616075100377202, + 0.0023285308852791786, + 0.003020512405782938, + 0.004347915295511484, + 0.0057433913461863995, + 0.005102450493723154, + 0.005931769497692585, + 0.008498243056237698, + 0.007206857670098543, + 0.007959585636854172, + 0.004082267172634602, + 0.006022875662893057, + 0.005238603334873915, + 0.006566884461790323, + 0.003123820060864091, + 0.08052872866392136, + 0.07094781845808029, + 0.003266712184995413, + 0.003948302008211613, + 0.001987685449421406, + 0.001952286111190915, + 0.0028018178418278694, + 0.001995485508814454, + 0.004813942592591047, + 0.005067809950560331, + 0.0033168140798807144, + 0.005298354662954807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.14281192421913147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, low-angle, tightly framed shot captures the scene of a massive natural force with stark documentary realism, captured with a high-speed camera. The sky is plunged into the mysterious twilight of a total solar eclipse, a deep, ominous indigo. The sun is a perfect black disk, haloed by the ghostly, shimmering white tendrils of the corona. A massive bolt of lightning tears through this backdrop, descending in breathtaking slow motion. Its core channel is a searing, overexposed white, from which intricate filaments of purple plasma branch out, like crystalline structures, deliberately crawling across the dark sky. The bolt's trajectory terminates precisely on the slender metal antenna spire of a towering skyscraper, its steel structure silhouetted against the strange light. The moment the main tendril makes contact, a violent and silent eruption of fiery white and orange sparks erupts from the tip, cascading down like a fiery waterfall, landing on the building's dark facade, each molten particle rendered in surreal clarity. The entire event unfolds with a sense of detached observation, creating an atmosphere of awe-inspiring and terrifying beauty that underscores the raw, untamed physics of the phenomenon.", + "seed": 14949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_9698_seed_14949_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10706812120042741, + "vlm_logprob": -0.10706812120042741, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.10706812120042741, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2916959226131439 + ], + "opposite_logprob_tokens": [ + -1.7916959524154663 + ], + "logprob_reward": -0.2916959226131439, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2916959226131439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07104542851448059, + 0.03556286171078682, + 0.02281997539103031, + 0.13371975719928741, + 0.07424161583185196, + 0.03567473962903023, + 0.03547446429729462, + 0.2239226996898651, + 0.23684106767177582, + 0.13069738447666168 + ], + "credit_map": [ + 0.005560866557061672, + 0.004575610160827637, + 0.008541048504412174, + 0.008207691833376884, + 0.005848124157637358, + 0.011164144612848759, + 0.017458628863096237, + 0.009689312428236008, + 0.0038191142957657576, + 0.003372920211404562, + 0.004924665205180645, + 0.0029545053839683533, + 0.004652339965105057, + 0.005293256137520075, + 0.006778582464903593, + 0.00376748014241457, + 0.002291117561981082, + 0.0021619750186800957, + 0.00230883969925344, + 0.0016973743913695216, + 0.00347654870711267, + 0.0032131350599229336, + 0.004795329179614782, + 0.002875657519325614, + 0.009227929636836052, + 0.012776268646121025, + 0.02009701170027256, + 0.011807172559201717, + 0.03205291926860809, + 0.01649618335068226, + 0.018919790163636208, + 0.012342477217316628, + 0.006924718152731657, + 0.004037938546389341, + 0.007037753239274025, + 0.008311755023896694, + 0.007151474244892597, + 0.009336200542747974, + 0.02129511535167694, + 0.010146662592887878, + 0.004263087175786495, + 0.003148499643430114, + 0.003632565028965473, + 0.004446227103471756, + 0.0034278773237019777, + 0.0036807998549193144, + 0.00828082486987114, + 0.004794859327375889, + 0.004655009601265192, + 0.0035760817117989063, + 0.003534614806994796, + 0.00488560926169157, + 0.004946285858750343, + 0.004660353064537048, + 0.005412583705037832, + 0.0038039267528802156, + 0.030161725357174873, + 0.028940051794052124, + 0.04598236829042435, + 0.01973380707204342, + 0.03793390840291977, + 0.028573883697390556, + 0.016675226390361786, + 0.015921728685498238, + 0.028728310018777847, + 0.027645936235785484, + 0.03702450916171074, + 0.02249615453183651, + 0.03973701596260071, + 0.030594125390052795, + 0.02457292005419731, + 0.02604210190474987, + 0.012950274161994457, + 0.0217270664870739, + 0.006185068748891354, + 0.006349707487970591, + 0.04767643287777901, + 0.015833625569939613, + 0.012302014045417309, + 0.007673194631934166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.2916959226131439 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004715987481176853 + ], + "opposite_logprob_tokens": [ + -5.379715919494629 + ], + "logprob_reward": -0.004715987481176853, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004715987481176853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26712271571159363, + 0.04772350564599037, + 0.03411746770143509, + 0.05490472540259361, + 0.03445287421345711, + 0.12457729876041412, + 0.04002075269818306, + 0.05329820141196251, + 0.1722540706396103, + 0.17152826488018036 + ], + "credit_map": [ + 0.05458610877394676, + 0.016775434836745262, + 0.03794075548648834, + 0.028983185067772865, + 0.028149208053946495, + 0.03536559268832207, + 0.038443803787231445, + 0.026878666132688522, + 0.008208464831113815, + 0.004568728152662516, + 0.00528356246650219, + 0.004268286749720573, + 0.006861405447125435, + 0.006520512979477644, + 0.006216748617589474, + 0.005795801058411598, + 0.0038678806740790606, + 0.0031316825188696384, + 0.0026318510062992573, + 0.00253266841173172, + 0.004147348925471306, + 0.0042631193064153194, + 0.009584723971784115, + 0.003958196844905615, + 0.004169296473264694, + 0.0055420841090381145, + 0.008567304350435734, + 0.005631805397570133, + 0.011893211863934994, + 0.006906021386384964, + 0.007470260839909315, + 0.004724741447716951, + 0.004725219216197729, + 0.004551737103611231, + 0.0046020252630114555, + 0.003325881902128458, + 0.0035386614035815, + 0.0039346348494291306, + 0.006223025731742382, + 0.0035516940988600254, + 0.014208889566361904, + 0.008282954804599285, + 0.01576204225420952, + 0.014609502628445625, + 0.013189559802412987, + 0.010571645572781563, + 0.021406972780823708, + 0.026545748114585876, + 0.005576370749622583, + 0.003397677093744278, + 0.004939730744808912, + 0.00672947010025382, + 0.0038759885355830193, + 0.005752336233854294, + 0.004417703952640295, + 0.0053314766846597195, + 0.009077254682779312, + 0.005864390172064304, + 0.010081786662340164, + 0.00563120236620307, + 0.006051730830222368, + 0.009374809451401234, + 0.003063185140490532, + 0.004153844900429249, + 0.019053766503930092, + 0.023473886772990227, + 0.024881185963749886, + 0.014673609286546707, + 0.030358456075191498, + 0.02693965658545494, + 0.016517912968993187, + 0.01635560765862465, + 0.02109210379421711, + 0.022914361208677292, + 0.017459144815802574, + 0.01097816787660122, + 0.03684268519282341, + 0.03307782858610153, + 0.015591357834637165, + 0.013572633266448975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.004715987481176853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003628814360126853 + ], + "opposite_logprob_tokens": [ + -5.628628730773926 + ], + "logprob_reward": -0.003628814360126853, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003628814360126853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20845527946949005, + 0.0771026536822319, + 0.06109696254134178, + 0.10116560012102127, + 0.07573778182268143, + 0.08816538006067276, + 0.053359344601631165, + 0.0707288607954979, + 0.10623878985643387, + 0.1579493135213852 + ], + "credit_map": [ + 0.03624136745929718, + 0.02282368578016758, + 0.019945573061704636, + 0.01588805764913559, + 0.015015062876045704, + 0.029420766979455948, + 0.03481383994221687, + 0.03430692479014397, + 0.006955771706998348, + 0.005035995971411467, + 0.0075132823549211025, + 0.013342826627194881, + 0.009947627782821655, + 0.009767696261405945, + 0.01708114519715309, + 0.007458303589373827, + 0.007540558930486441, + 0.006275995168834925, + 0.007455380167812109, + 0.004750813823193312, + 0.006188093218952417, + 0.007928448729217052, + 0.01377764344215393, + 0.007180029526352882, + 0.00949607603251934, + 0.01038101315498352, + 0.013088357634842396, + 0.00876635406166315, + 0.01604265347123146, + 0.011987611651420593, + 0.020992208272218704, + 0.010411320254206657, + 0.0072305165231227875, + 0.004651876166462898, + 0.008023437112569809, + 0.005462282337248325, + 0.008044371381402016, + 0.014164062216877937, + 0.018630793318152428, + 0.0095304474234581, + 0.011639640666544437, + 0.006121670361608267, + 0.012663329020142555, + 0.010994290001690388, + 0.008660892024636269, + 0.008286871947348118, + 0.011219502426683903, + 0.018579183146357536, + 0.007475715130567551, + 0.004688487853854895, + 0.013651938177645206, + 0.008599000982940197, + 0.004586349707096815, + 0.005667796358466148, + 0.004344823770225048, + 0.004345233552157879, + 0.009549906477332115, + 0.006909313146024942, + 0.01095994096249342, + 0.006076051853597164, + 0.011500122025609016, + 0.011877470649778843, + 0.006073071155697107, + 0.007782982662320137, + 0.013332182541489601, + 0.014280256815254688, + 0.012612526305019855, + 0.010430230759084225, + 0.018235355615615845, + 0.015149667859077454, + 0.011618395335972309, + 0.01058017835021019, + 0.017653774470090866, + 0.022067731246352196, + 0.013454393483698368, + 0.01234442088752985, + 0.04174421355128288, + 0.016165105625987053, + 0.022791102528572083, + 0.011728579178452492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.003628814360126853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370870277285576 + ], + "opposite_logprob_tokens": [ + -3.4087088108062744 + ], + "logprob_reward": -0.03370870277285576, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370870277285576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07519636303186417, + 0.04875103756785393, + 0.026504475623369217, + 0.051371876150369644, + 0.0773182213306427, + 0.06612855195999146, + 0.03779593110084534, + 0.2926706075668335, + 0.10482221096754074, + 0.219440758228302 + ], + "credit_map": [ + 0.01432293001562357, + 0.0069388835690915585, + 0.00771303428336978, + 0.0071815918199718, + 0.0056487335823476315, + 0.011070559732615948, + 0.009423183277249336, + 0.012897447682917118, + 0.005216220859438181, + 0.002609999617561698, + 0.004409013316035271, + 0.0035106204450130463, + 0.006466367281973362, + 0.006275108549743891, + 0.012554948218166828, + 0.007708759978413582, + 0.002779334783554077, + 0.0021518163848668337, + 0.0021659706253558397, + 0.003054147120565176, + 0.003742404980584979, + 0.0032558676321059465, + 0.005601474083960056, + 0.003753459779545665, + 0.0053589255549013615, + 0.0038036799523979425, + 0.00671849399805069, + 0.0037261508405208588, + 0.006571328267455101, + 0.0064829932525753975, + 0.009918720461428165, + 0.008791579864919186, + 0.007398704066872597, + 0.004904081579297781, + 0.0085732601583004, + 0.0076014515943825245, + 0.008361364714801311, + 0.011299469508230686, + 0.020701274275779724, + 0.008478616364300251, + 0.009227119386196136, + 0.0036370335146784782, + 0.0076342797838151455, + 0.008445496670901775, + 0.005711222533136606, + 0.009394708089530468, + 0.01014281902462244, + 0.01193587388843298, + 0.005168671254068613, + 0.002840836998075247, + 0.007649191655218601, + 0.005373201332986355, + 0.003444913076236844, + 0.004910015966743231, + 0.003883764613419771, + 0.004525337368249893, + 0.03959192708134651, + 0.03714631870388985, + 0.05280318856239319, + 0.024498578161001205, + 0.05399612709879875, + 0.040027376264333725, + 0.02293262630701065, + 0.02167445980012417, + 0.01190109271556139, + 0.013201694935560226, + 0.015006212517619133, + 0.013665340840816498, + 0.014503339305520058, + 0.01742783933877945, + 0.009243008680641651, + 0.009873677976429462, + 0.018513936549425125, + 0.035398393869400024, + 0.012267246842384338, + 0.014301779679954052, + 0.07335628569126129, + 0.023280447348952293, + 0.0293990857899189, + 0.012923592701554298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.03370870277285576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159117877483368 + ], + "opposite_logprob_tokens": [ + -1.70159113407135 + ], + "logprob_reward": -0.20159117877483368, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159117877483368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1920541524887085, + 0.07669616490602493, + 0.03760085627436638, + 0.03930411487817764, + 0.07712507992982864, + 0.10900364071130753, + 0.06587156653404236, + 0.1112108901143074, + 0.18605703115463257, + 0.10507640987634659 + ], + "credit_map": [ + 0.03283718228340149, + 0.014455611817538738, + 0.016116322949528694, + 0.022798150777816772, + 0.013593524694442749, + 0.027709241956472397, + 0.036400407552719116, + 0.028143716976046562, + 0.006233782973140478, + 0.005516039207577705, + 0.007990922778844833, + 0.007265293039381504, + 0.006138933356851339, + 0.01023631077259779, + 0.018638426437973976, + 0.014676466584205627, + 0.0034502926282584667, + 0.00335428724065423, + 0.0034375691320747137, + 0.0037315369118005037, + 0.006846387404948473, + 0.004798117093741894, + 0.005705193150788546, + 0.006277475506067276, + 0.0031684450805187225, + 0.0031155950855463743, + 0.004497865214943886, + 0.00316184782423079, + 0.006938963197171688, + 0.005485767498612404, + 0.008731534704566002, + 0.0042040990665555, + 0.00900721549987793, + 0.007056014146655798, + 0.007837062701582909, + 0.0054200817830860615, + 0.007415317464619875, + 0.012559131719172001, + 0.020468957722187042, + 0.007361308205872774, + 0.014460447244346142, + 0.006062104366719723, + 0.012658697552978992, + 0.012929846532642841, + 0.009091544896364212, + 0.01861751265823841, + 0.010306226089596748, + 0.02487727254629135, + 0.00925532728433609, + 0.005646572448313236, + 0.008309467695653439, + 0.012778978794813156, + 0.0057294415310025215, + 0.008722553960978985, + 0.005997953005135059, + 0.009431280195713043, + 0.0120372474193573, + 0.013621621765196323, + 0.01609988696873188, + 0.011459939181804657, + 0.02027587592601776, + 0.014758206903934479, + 0.010588428005576134, + 0.012369687668979168, + 0.018551604822278023, + 0.025870470330119133, + 0.026407655328512192, + 0.018176278099417686, + 0.0327366478741169, + 0.024945003911852837, + 0.021232113242149353, + 0.018137291073799133, + 0.00994067545980215, + 0.014265866950154305, + 0.01183987781405449, + 0.01181640662252903, + 0.022325295954942703, + 0.01153880450874567, + 0.016390204429626465, + 0.006959294434636831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.20159117877483368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, low-angle, tightly framed shot captures the scene of a massive natural force with stark documentary realism, captured with a high-speed camera. The sky is plunged into the mysterious twilight of a total solar eclipse, a deep, ominous indigo. The sun is a perfect black disk, haloed by the ghostly, shimmering white tendrils of the corona. A massive bolt of lightning tears through this backdrop, descending in breathtaking slow motion. Its core channel is a searing, overexposed white, from which intricate filaments of purple plasma branch out, like crystalline structures, deliberately crawling across the dark sky. The bolt's trajectory terminates precisely on the slender metal antenna spire of a towering skyscraper, its steel structure silhouetted against the strange light. The moment the main tendril makes contact, a violent and silent eruption of fiery white and orange sparks erupts from the tip, cascading down like a fiery waterfall, landing on the building's dark facade, each molten particle rendered in surreal clarity. The entire event unfolds with a sense of detached observation, creating an atmosphere of awe-inspiring and terrifying beauty that underscores the raw, untamed physics of the phenomenon.", + "seed": 14950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_9698_seed_14950_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15539581989869475, + "vlm_logprob": -0.15539581989869475, + "vlm_yes_no_margin": 2.874999976158142, + "vlm_token_credit": -0.15539581989869475, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4532882869243622 + ], + "opposite_logprob_tokens": [ + -1.3282883167266846 + ], + "logprob_reward": -0.4532882869243622, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.4532882869243622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8934197425842285, + 0.015865469351410866, + 0.022703148424625397, + 0.01173989474773407, + 0.008456921204924583, + 0.007718035951256752, + 0.008954702876508236, + 0.010608423501253128, + 0.012297184206545353, + 0.008236479014158249 + ], + "credit_map": [ + 0.05038439854979515, + 0.011098714545369148, + 0.3767802119255066, + 0.020328616723418236, + 0.019740229472517967, + 0.3839997351169586, + 0.00912370253354311, + 0.021964143961668015, + 0.0020708944648504257, + 0.0011919757816940546, + 0.002551398705691099, + 0.0011116891400888562, + 0.0009076267015188932, + 0.0023763461504131556, + 0.002455822192132473, + 0.003199715632945299, + 0.0017939901445060968, + 0.0032858995255082846, + 0.0028632166795432568, + 0.0020853078458458185, + 0.002160195726901293, + 0.0027407680172473192, + 0.004910258576273918, + 0.0028635135386139154, + 0.0010977911297231913, + 0.0021396572701632977, + 0.0016465812223032117, + 0.0014025650452822447, + 0.001933533581905067, + 0.0011584440944716334, + 0.0011855106567963958, + 0.0011758117470890284, + 0.0007550464943051338, + 0.0009649335988797247, + 0.0016825825441628695, + 0.0006461245357058942, + 0.0011512745404615998, + 0.0012901081936433911, + 0.0009357294184155762, + 0.0010311214718967676, + 0.0006328494055196643, + 0.0008923780987970531, + 0.001013323082588613, + 0.0007264543091878295, + 0.0011385270627215505, + 0.0007859144825488329, + 0.001441929372958839, + 0.0010866604279726744, + 0.0010823849588632584, + 0.0005824341787956655, + 0.0015057913260534406, + 0.0008215695270337164, + 0.0005725756054744124, + 0.0011091711930930614, + 0.0012558838352560997, + 0.002024892019107938, + 0.0013865540968254209, + 0.0011733433930203319, + 0.001845022663474083, + 0.000808997021522373, + 0.0012235856847837567, + 0.0012760163517668843, + 0.0008691982948221266, + 0.002025705762207508, + 0.0010996776400133967, + 0.001984048169106245, + 0.0014326460659503937, + 0.0012699459912255406, + 0.0012210210552439094, + 0.001256519928574562, + 0.0013183669652789831, + 0.002714957809075713, + 0.0008382013766095042, + 0.000862888409756124, + 0.0012249014107510448, + 0.0012541075702756643, + 0.0006153230788186193, + 0.0008231375832110643, + 0.001068617100827396, + 0.001549302483908832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.4532882869243622 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01260845735669136 + ], + "opposite_logprob_tokens": [ + -4.387608528137207 + ], + "logprob_reward": -0.01260845735669136, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01260845735669136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7943162322044373, + 0.03521629050374031, + 0.03417246416211128, + 0.018154650926589966, + 0.06927470862865448, + 0.016305264085531235, + 0.007446587551385164, + 0.010131621733307838, + 0.006329938303679228, + 0.008652331307530403 + ], + "credit_map": [ + 0.04419150948524475, + 0.010133751668035984, + 0.3336906433105469, + 0.018361130729317665, + 0.018228838220238686, + 0.34202659130096436, + 0.0076922099106013775, + 0.01999150589108467, + 0.005429428070783615, + 0.004194986540824175, + 0.005732100456953049, + 0.00222750473767519, + 0.003005838254466653, + 0.005545062478631735, + 0.0035910424776375294, + 0.0054903216660022736, + 0.002568055409938097, + 0.0035687547642737627, + 0.006430116482079029, + 0.0021706982515752316, + 0.0035529863089323044, + 0.005692348815500736, + 0.005056480411440134, + 0.00513302069157362, + 0.0020933030173182487, + 0.003686382668092847, + 0.0024073594249784946, + 0.0022528208792209625, + 0.002229103120043874, + 0.0019178467337042093, + 0.0017752222483977675, + 0.0017926105065271258, + 0.005972739774733782, + 0.007930302061140537, + 0.014619347639381886, + 0.004610452800989151, + 0.011439904570579529, + 0.01094917580485344, + 0.005665529519319534, + 0.008087248541414738, + 0.0017306372756138444, + 0.0025182568933814764, + 0.002673117443919182, + 0.0014425832778215408, + 0.001922589261084795, + 0.0018405010923743248, + 0.0022963539231568575, + 0.0018812214257195592, + 0.0007570161833427846, + 0.0011830083094537258, + 0.0012171347625553608, + 0.0006183436489664018, + 0.0011768046533688903, + 0.0009641657234169543, + 0.0007725624600425363, + 0.0007575513445772231, + 0.0010787488427013159, + 0.0014444611733779311, + 0.0013624908169731498, + 0.0009215466561727226, + 0.001614321838133037, + 0.001011642045341432, + 0.0009874271927401423, + 0.0017109812470152974, + 0.0006854765815660357, + 0.0011235394049435854, + 0.0009311960893683136, + 0.0006503266049548984, + 0.0005839804071001709, + 0.0006989496177993715, + 0.0006986989174038172, + 0.0009577699820511043, + 0.0010626950534060597, + 0.000962564256042242, + 0.001280060620047152, + 0.0010931959841400385, + 0.0008362099179066718, + 0.0010448115644976497, + 0.0009884899482131004, + 0.0013843027409166098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01260845735669136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004109866451472044 + ], + "opposite_logprob_tokens": [ + -5.504109859466553 + ], + "logprob_reward": -0.004109866451472044, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004109866451472044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9564903974533081, + 0.004003329202532768, + 0.005562267731875181, + 0.006744964513927698, + 0.007756166625767946, + 0.004631835501641035, + 0.004207275342196226, + 0.0027612652629613876, + 0.0014780148630961776, + 0.006364439148455858 + ], + "credit_map": [ + 0.05333079397678375, + 0.010912036523222923, + 0.403835654258728, + 0.02213282696902752, + 0.0214166771620512, + 0.4130706191062927, + 0.007878651842474937, + 0.023913223296403885, + 0.0005135126993991435, + 0.00034104110090993345, + 0.0004963231040164828, + 0.000279922503978014, + 0.00022457406157627702, + 0.0005893083871342242, + 0.0006238727364689112, + 0.000934775045607239, + 0.00035148567985743284, + 0.0006547083030454814, + 0.0013547901762649417, + 0.0003811754868365824, + 0.0005537827964872122, + 0.0011110822670161724, + 0.00079325184924528, + 0.0003619919589255005, + 0.0007105289259925485, + 0.0011247273068875074, + 0.0011370592983439565, + 0.000660630757920444, + 0.0007778704748488963, + 0.0007889231783337891, + 0.0009115738794207573, + 0.0006336510414257646, + 0.000673614616971463, + 0.0009149437537416816, + 0.0017227033386006951, + 0.0004916726029478014, + 0.001266344333998859, + 0.0013082962250337005, + 0.0006030005170032382, + 0.0007755918195471168, + 0.0004403450002428144, + 0.0003720402892213315, + 0.0007197693921625614, + 0.00040347472531720996, + 0.0004900617059320211, + 0.0005284021608531475, + 0.0006815602537244558, + 0.0009961826726794243, + 0.0004336140409577638, + 0.00021081110753584653, + 0.0009270862210541964, + 0.00033584912307560444, + 0.00034845538903027773, + 0.0007339271833188832, + 0.0004336872370913625, + 0.00078384525841102, + 0.0003126148658338934, + 0.0003326798905618489, + 0.0003255608899053186, + 0.00026408405392430723, + 0.0003244360559619963, + 0.0002999971038661897, + 0.0002761481737252325, + 0.0006257446948438883, + 0.00020968094759155065, + 0.00012893657549284399, + 0.0002987038460560143, + 0.00015207381511572748, + 0.00011125293531222269, + 0.0001837616291595623, + 0.0001346808421658352, + 0.00025892435223795474, + 0.000657126132864505, + 0.0005438737571239471, + 0.0010102404048666358, + 0.0008308659307658672, + 0.0005683978670276701, + 0.0007736996049061418, + 0.0007835988653823733, + 0.0011966369347646832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.004109866451472044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991414666175842 + ], + "opposite_logprob_tokens": [ + -1.804914116859436 + ], + "logprob_reward": -0.17991414666175842, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991414666175842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7663924694061279, + 0.030315058305859566, + 0.012414844706654549, + 0.021651390939950943, + 0.02038593590259552, + 0.03219034522771835, + 0.02188107930123806, + 0.020169556140899658, + 0.013769854791462421, + 0.06082940101623535 + ], + "credit_map": [ + 0.043268170207738876, + 0.00956075917929411, + 0.3229422867298126, + 0.017662763595581055, + 0.01716364547610283, + 0.3298138678073883, + 0.0065627070143818855, + 0.019418315961956978, + 0.004564224276691675, + 0.0021994642447680235, + 0.004704229533672333, + 0.002313873264938593, + 0.0018323534168303013, + 0.0042726933024823666, + 0.003846232546493411, + 0.00658198818564415, + 0.0016435967991128564, + 0.0014275374123826623, + 0.0022898870520293713, + 0.0009024767205119133, + 0.0010829258244484663, + 0.0019642652478069067, + 0.0017382600344717503, + 0.0013658973621204495, + 0.0024896711111068726, + 0.003348343074321747, + 0.003287679050117731, + 0.0022531694266945124, + 0.002431186381727457, + 0.0023459584917873144, + 0.0032409788109362125, + 0.0022544055245816708, + 0.0016809870721772313, + 0.0019148691790178418, + 0.0039022129494696856, + 0.0015381472185254097, + 0.002764721168205142, + 0.0030307446140795946, + 0.0021266748663038015, + 0.0034275820944458246, + 0.002647198736667633, + 0.0023453941103070974, + 0.004288796801120043, + 0.0029722843319177628, + 0.0022058358881622553, + 0.0039201099425554276, + 0.005418278742581606, + 0.008392448537051678, + 0.002881772583350539, + 0.0014862035168334842, + 0.0031372157391160727, + 0.00228286674246192, + 0.0016184658743441105, + 0.0024885237216949463, + 0.002666665241122246, + 0.005319368094205856, + 0.0023933113552629948, + 0.0019686806481331587, + 0.002624048851430416, + 0.0016657435335218906, + 0.0017547575989738107, + 0.002248578704893589, + 0.002269353251904249, + 0.005245083477348089, + 0.0012989218812435865, + 0.0008580366265960038, + 0.002294418402016163, + 0.0013447694946080446, + 0.001103217015042901, + 0.001743765315040946, + 0.001380082336254418, + 0.0037466445937752724, + 0.005643113516271114, + 0.005445858463644981, + 0.011608392931520939, + 0.007477849721908569, + 0.005270316265523434, + 0.008107139728963375, + 0.006653206422924995, + 0.010623527690768242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.17991414666175842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705834209918976 + ], + "opposite_logprob_tokens": [ + -2.127058267593384 + ], + "logprob_reward": -0.12705834209918976, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12705834209918976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8080735802650452, + 0.024962570518255234, + 0.03967618569731712, + 0.019542211666703224, + 0.012699640356004238, + 0.022238709032535553, + 0.015683401376008987, + 0.009926286526024342, + 0.01130096334964037, + 0.035896480083465576 + ], + "credit_map": [ + 0.0455196313560009, + 0.010214188136160374, + 0.34061941504478455, + 0.018862497061491013, + 0.017954697832465172, + 0.34676870703697205, + 0.007887073792517185, + 0.020247310400009155, + 0.0031751750502735376, + 0.0033862823620438576, + 0.003438758896663785, + 0.002270290395244956, + 0.001323650823906064, + 0.003344316966831684, + 0.0029536494985222816, + 0.005070443265140057, + 0.002790674567222595, + 0.005833490751683712, + 0.00558392284438014, + 0.0032168887555599213, + 0.005657754372805357, + 0.005261846352368593, + 0.007271673996001482, + 0.004059927072376013, + 0.0019190760795027018, + 0.003744323505088687, + 0.002809656085446477, + 0.0019924247171729803, + 0.002578810788691044, + 0.0023420988582074642, + 0.0022293932270258665, + 0.001926427474245429, + 0.001049022190272808, + 0.0015292706666514277, + 0.0020908552687615156, + 0.0008858320070430636, + 0.0018347242148593068, + 0.001639827503822744, + 0.001381470006890595, + 0.002288636052981019, + 0.001699369866400957, + 0.002198785310611129, + 0.003051137551665306, + 0.0019005321664735675, + 0.001903404714539647, + 0.002817901549860835, + 0.0029176250100135803, + 0.005749951116740704, + 0.0020066644065082073, + 0.0012543239863589406, + 0.0026499927043914795, + 0.0015014661476016045, + 0.0012729926966130733, + 0.001959778368473053, + 0.0017896961653605103, + 0.003248486202210188, + 0.001005112542770803, + 0.0007398843881674111, + 0.0011387808481231332, + 0.0007915618480183184, + 0.0011247446527704597, + 0.0010849741520360112, + 0.0014111382188275456, + 0.0026300891768187284, + 0.0011143385199829936, + 0.0009167951648123562, + 0.0017791169229894876, + 0.0010021405760198832, + 0.0007616322254762053, + 0.0012994947610422969, + 0.0008529653423465788, + 0.003574478905647993, + 0.0035234568640589714, + 0.0029819810297340155, + 0.006354091223329306, + 0.0046293847262859344, + 0.0033151896204799414, + 0.004592239391058683, + 0.004232210572808981, + 0.006267925258725882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.12705834209918976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, low-angle, tightly framed shot captures the scene of a massive natural force with stark documentary realism, captured with a high-speed camera. The sky is plunged into the mysterious twilight of a total solar eclipse, a deep, ominous indigo. The sun is a perfect black disk, haloed by the ghostly, shimmering white tendrils of the corona. A massive bolt of lightning tears through this backdrop, descending in breathtaking slow motion. Its core channel is a searing, overexposed white, from which intricate filaments of purple plasma branch out, like crystalline structures, deliberately crawling across the dark sky. The bolt's trajectory terminates precisely on the slender metal antenna spire of a towering skyscraper, its steel structure silhouetted against the strange light. The moment the main tendril makes contact, a violent and silent eruption of fiery white and orange sparks erupts from the tip, cascading down like a fiery waterfall, landing on the building's dark facade, each molten particle rendered in surreal clarity. The entire event unfolds with a sense of detached observation, creating an atmosphere of awe-inspiring and terrifying beauty that underscores the raw, untamed physics of the phenomenon.", + "seed": 14951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_6/index_9698_seed_14951_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2409661902114749, + "vlm_logprob": -0.2409661902114749, + "vlm_yes_no_margin": 2.400000011920929, + "vlm_token_credit": -0.2409661902114749, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.40703076124191284 + ], + "opposite_logprob_tokens": [ + -1.2820308208465576 + ], + "logprob_reward": -0.40703076124191284, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.40703076124191284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4430021643638611, + 0.04765751585364342, + 0.030111555010080338, + 0.05344086512923241, + 0.030353285372257233, + 0.11680220067501068, + 0.014608100056648254, + 0.22621287405490875, + 0.019199935719370842, + 0.018611498177051544 + ], + "credit_map": [ + 0.011952237226068974, + 0.007874003611505032, + 0.0063187782652676105, + 0.006706657819449902, + 0.052927084267139435, + 0.1967732012271881, + 0.016873588785529137, + 0.14357662200927734, + 0.009007635526359081, + 0.0017664211336523294, + 0.00964414980262518, + 0.0019785291515290737, + 0.003610417479649186, + 0.010226069018244743, + 0.003191183088347316, + 0.008233110420405865, + 0.0024568040389567614, + 0.002033863915130496, + 0.0017448753351345658, + 0.001904772943817079, + 0.0022464676294475794, + 0.009175668470561504, + 0.002408689586445689, + 0.008140410296618938, + 0.0019238860113546252, + 0.0038011730648577213, + 0.002683038590475917, + 0.002178759314119816, + 0.0019699386321008205, + 0.019989361986517906, + 0.00240423739887774, + 0.018490467220544815, + 0.0024906392209231853, + 0.003123170929029584, + 0.0034080781042575836, + 0.0025166673585772514, + 0.0018608561949804425, + 0.006590244825929403, + 0.002491381485015154, + 0.007872246205806732, + 0.006274232175201178, + 0.0026237587444484234, + 0.04770214855670929, + 0.00436187069863081, + 0.0029312651604413986, + 0.045981649309396744, + 0.0019530223216861486, + 0.004974250681698322, + 0.0017041235696524382, + 0.0014676598366349936, + 0.003103571478277445, + 0.00269692693836987, + 0.0008864139672368765, + 0.002319413935765624, + 0.0009545821812935174, + 0.0014754090225324035, + 0.022510381415486336, + 0.03692274168133736, + 0.03111535869538784, + 0.03275701403617859, + 0.046686749905347824, + 0.01851695403456688, + 0.02243795618414879, + 0.015265707857906818, + 0.0026313818525522947, + 0.0023479308001697063, + 0.00291137071326375, + 0.0037212055176496506, + 0.001376907341182232, + 0.0014644184848293662, + 0.002894969657063484, + 0.001851752633228898, + 0.0025671168696135283, + 0.0020873229950666428, + 0.0025430447421967983, + 0.0022566684056073427, + 0.002341269748285413, + 0.0015850551426410675, + 0.002854755846783519, + 0.002376265125349164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.40703076124191284 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632892906665802 + ], + "opposite_logprob_tokens": [ + -0.757892906665802 + ], + "logprob_reward": -0.632892906665802, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632892906665802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5121238827705383, + 0.053476933389902115, + 0.10905700922012329, + 0.17807184159755707, + 0.009803283028304577, + 0.03720321133732796, + 0.040052711963653564, + 0.01684729941189289, + 0.023481236770749092, + 0.01988254301249981 + ], + "credit_map": [ + 0.013953340239822865, + 0.008915632963180542, + 0.007712851744145155, + 0.006973260082304478, + 0.058378785848617554, + 0.22804908454418182, + 0.01961832493543625, + 0.16852258145809174, + 0.004918437916785479, + 0.0014973141951486468, + 0.005103433504700661, + 0.0014049469027668238, + 0.003664502641186118, + 0.018126441165804863, + 0.002283712849020958, + 0.016478147357702255, + 0.003077613888308406, + 0.00266234390437603, + 0.0034903588239103556, + 0.002112465910613537, + 0.006065191701054573, + 0.0446043498814106, + 0.0034758669789880514, + 0.04356881231069565, + 0.006538873538374901, + 0.005044323857873678, + 0.00897279754281044, + 0.005144751165062189, + 0.007237093523144722, + 0.07039234042167664, + 0.00902472622692585, + 0.0657169371843338, + 0.0009476635023020208, + 0.001588522456586361, + 0.001641265582293272, + 0.000802655122242868, + 0.0006987562519498169, + 0.0013137829955667257, + 0.0010323718888685107, + 0.001778265810571611, + 0.002118909964337945, + 0.0007673776126466691, + 0.01461479626595974, + 0.0016177620273083448, + 0.0010056927567347884, + 0.014194616116583347, + 0.0011441911337897182, + 0.0017398670315742493, + 0.004041171167045832, + 0.0056745437905192375, + 0.004943937063217163, + 0.0066538164392113686, + 0.0075898985378444195, + 0.003969832789152861, + 0.00377298379316926, + 0.003406529314815998, + 0.0022395323030650616, + 0.00242406758479774, + 0.0028918730095028877, + 0.0024757946375757456, + 0.002899778774008155, + 0.0017856648191809654, + 0.0014056413201615214, + 0.0007249446935020387, + 0.0030716678593307734, + 0.0021394139621406794, + 0.003116054693236947, + 0.004645929206162691, + 0.001608084887266159, + 0.002177563961595297, + 0.004335709381848574, + 0.002386812586337328, + 0.002996731549501419, + 0.002713174093514681, + 0.002764922333881259, + 0.0028357456903904676, + 0.0024439722765237093, + 0.0018982574110850692, + 0.0024471883662045, + 0.0017825508257374167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.632892906665802 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025009801611304283 + ], + "opposite_logprob_tokens": [ + -6.002501010894775 + ], + "logprob_reward": -0.0025009801611304283, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025009801611304283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30023670196533203, + 0.09079135209321976, + 0.11821851879358292, + 0.03333147615194321, + 0.0862719863653183, + 0.09095602482557297, + 0.09396049380302429, + 0.14474618434906006, + 0.02691296860575676, + 0.014574291184544563 + ], + "credit_map": [ + 0.008316334336996078, + 0.0054219928570091724, + 0.005417268723249435, + 0.004011485259979963, + 0.034003958106040955, + 0.13339276611804962, + 0.010773065499961376, + 0.09889984875917435, + 0.008298540487885475, + 0.002531967591494322, + 0.00769283901900053, + 0.0023220181465148926, + 0.007661786396056414, + 0.032663390040397644, + 0.0036673613358289003, + 0.025953445583581924, + 0.0044387890957295895, + 0.0030532325617969036, + 0.004404537379741669, + 0.0029679059516638517, + 0.0067884065210819244, + 0.04787416383624077, + 0.004635349847376347, + 0.044056132435798645, + 0.006987651810050011, + 0.0029194564558565617, + 0.003867710242047906, + 0.0037491372786462307, + 0.002654347801581025, + 0.0045022000558674335, + 0.0035184924490749836, + 0.00513248099014163, + 0.007699991576373577, + 0.006710599176585674, + 0.01036341954022646, + 0.006791606079787016, + 0.004716521594673395, + 0.02194688655436039, + 0.007520653307437897, + 0.020522307604551315, + 0.00738304341211915, + 0.0028834850527346134, + 0.025925898924469948, + 0.005873614456504583, + 0.003775774734094739, + 0.029857546091079712, + 0.004762300290167332, + 0.010494360700249672, + 0.009696283377707005, + 0.012333676218986511, + 0.01126166619360447, + 0.01614152267575264, + 0.017453089356422424, + 0.009196423925459385, + 0.009551852010190487, + 0.008325978182256222, + 0.014901295304298401, + 0.023740511387586594, + 0.02087395079433918, + 0.02087872289121151, + 0.02934645302593708, + 0.012113136239349842, + 0.014644368551671505, + 0.00824774894863367, + 0.0025153381284326315, + 0.004316855221986771, + 0.0033489945344626904, + 0.005013483110815287, + 0.0037494308780878782, + 0.002425548154860735, + 0.003939161077141762, + 0.0016041591297835112, + 0.0016283098375424743, + 0.0018468780908733606, + 0.002362180268391967, + 0.002740543568506837, + 0.0013462698552757502, + 0.0013784553157165647, + 0.0021026385948061943, + 0.001169015420600772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0025009801611304283 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206643581390381 + ], + "opposite_logprob_tokens": [ + -2.8120665550231934 + ], + "logprob_reward": -0.06206643581390381, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206643581390381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09465756267309189, + 0.02909856289625168, + 0.051433153450489044, + 0.05934491753578186, + 0.017933962866663933, + 0.028828447684645653, + 0.06509657949209213, + 0.25483036041259766, + 0.20921705663204193, + 0.18955950438976288 + ], + "credit_map": [ + 0.0036736582405865192, + 0.002819349756464362, + 0.003880241420120001, + 0.002417993266135454, + 0.011291513219475746, + 0.03675777092576027, + 0.004822454880923033, + 0.02899458445608616, + 0.004013935104012489, + 0.0017152944346889853, + 0.0031778078991919756, + 0.00203697825782001, + 0.0055242194794118404, + 0.006056183949112892, + 0.00335208373144269, + 0.0032220561988651752, + 0.003586651524528861, + 0.0031998208723962307, + 0.004128383472561836, + 0.0031663228292018175, + 0.006243990268558264, + 0.01492474414408207, + 0.0029780028853565454, + 0.013205230236053467, + 0.00917254202067852, + 0.007838049903512001, + 0.006074967794120312, + 0.006715301424264908, + 0.004817011766135693, + 0.00892292894423008, + 0.006600324995815754, + 0.009203786961734295, + 0.002626814879477024, + 0.0013470550766214728, + 0.0026163391303271055, + 0.0018156798323616385, + 0.0014839988434687257, + 0.003431146265938878, + 0.0018097462598234415, + 0.002803180366754532, + 0.003350300481542945, + 0.0015949273947626352, + 0.00813022069633007, + 0.002703394042328, + 0.0013284237356856465, + 0.005450414959341288, + 0.0027383328415453434, + 0.0035324308555573225, + 0.007427678909152746, + 0.008229123428463936, + 0.00822352897375822, + 0.013308080844581127, + 0.010179349221289158, + 0.006320311222225428, + 0.006193582434207201, + 0.0052149235270917416, + 0.02731270343065262, + 0.04320286214351654, + 0.037489812821149826, + 0.03461676836013794, + 0.05423715338110924, + 0.022739902138710022, + 0.022142676636576653, + 0.013088466599583626, + 0.022660601884126663, + 0.015077918767929077, + 0.02787870727479458, + 0.04855450615286827, + 0.012604301795363426, + 0.019401725381612778, + 0.04994521662592888, + 0.01309405080974102, + 0.027576955035328865, + 0.01783890277147293, + 0.02996559627354145, + 0.03190276026725769, + 0.016850197687745094, + 0.016545334830880165, + 0.02742931805551052, + 0.021450437605381012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.06206643581390381 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and tightly framed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a deep indigo color with a perfect black disk sun surrounded by a white corona?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bolt of lightning shown descending in slow motion with a white core and purple plasma filaments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lightning strike a slender metal antenna spire on a towering skyscraper, causing an eruption of white and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with stark documentary realism and surreal clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003398671746254 + ], + "opposite_logprob_tokens": [ + -2.350339889526367 + ], + "logprob_reward": -0.1003398671746254, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003398671746254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7178289890289307, + 0.03546934947371483, + 0.09803982824087143, + 0.04566112160682678, + 0.018615495413541794, + 0.023575251922011375, + 0.014960642904043198, + 0.01134391501545906, + 0.017230499535799026, + 0.017274867743253708 + ], + "credit_map": [ + 0.01909036748111248, + 0.012439233250916004, + 0.010172937996685505, + 0.010159255936741829, + 0.0832696482539177, + 0.3210107088088989, + 0.02501937933266163, + 0.23666751384735107, + 0.0022657914087176323, + 0.0007372031104750931, + 0.0025184806436300278, + 0.0006723595433868468, + 0.0023357602767646313, + 0.01365593634545803, + 0.001376379863359034, + 0.011907440610229969, + 0.002620530314743519, + 0.001655779080465436, + 0.002922668121755123, + 0.0014673675177618861, + 0.0037616624031215906, + 0.04325307905673981, + 0.001733456621877849, + 0.04062529280781746, + 0.001513693598099053, + 0.0015889022033661604, + 0.002284182934090495, + 0.0012534554116427898, + 0.0018025358440354466, + 0.018211979418992996, + 0.0019381741294637322, + 0.017068205401301384, + 0.0023028876166790724, + 0.0006939376471564174, + 0.002525210613384843, + 0.0015329099260270596, + 0.0008749691769480705, + 0.005050018895417452, + 0.0012236353941261768, + 0.004411926958709955, + 0.0015110652893781662, + 0.0005755490274168551, + 0.008984715677797794, + 0.001435779733583331, + 0.0004638493701349944, + 0.008683806285262108, + 0.0006730463355779648, + 0.0012474448885768652, + 0.0016791907837614417, + 0.0020636836998164654, + 0.0019768308848142624, + 0.002400309080258012, + 0.0026665604673326015, + 0.001525019179098308, + 0.0014448377769440413, + 0.0012042137095704675, + 0.001162990927696228, + 0.0012740156380459666, + 0.0015881923027336597, + 0.0022202038671821356, + 0.0010162617545574903, + 0.0010404745116829872, + 0.0017409614520147443, + 0.0013008161913603544, + 0.002010633237659931, + 0.0013914568116888404, + 0.002212371677160263, + 0.0033516455441713333, + 0.001388372271321714, + 0.001560222008265555, + 0.0038245541509240866, + 0.0014912436017766595, + 0.0019017677986994386, + 0.0017805879469960928, + 0.0026018102653324604, + 0.0035750120878219604, + 0.001389806275255978, + 0.0012965661007910967, + 0.0029656030237674713, + 0.0017637151759117842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1003398671746254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_tensor([8665, 8665])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_tensor([8665, 8665])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a647fb7ac25787bc9a60c8046535a15d3a58739 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_tensor([8665, 8665])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up, the camera rapidly tilts upward, capturing a flag made of heavy, faded crimson cotton and linen fabric. Initially still and drooping, the flag is coated in a thick layer of grayish-white dust, deeply embedded in its rough texture, its edges frayed and minutely torn. Suddenly, a powerful gust of wind sweeps in, lifting the flag uncontrollably upward, instantly dispersing the accumulated dust and forming a fleeting, murky cloud beneath the harsh, desolate sunlight. In this moment, the camera focuses on the fabric's microscopic details, clearly showing each individual fiber trembling and vibrating at high frequency, as if in a silent scream. The entire flag billows and twitches violently in the wind, while light pierces the dense dust, forming distinct beams of light, creating a dramatic atmosphere imbued with power, desolation, and a profound sense of historical vicissitudes.", + "seed": 15911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15911_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1459792480804026, + "vlm_logprob": -0.1459792480804026, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.1459792480804026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011383346281945705 + ], + "opposite_logprob_tokens": [ + -5.011383533477783 + ], + "logprob_reward": -0.011383346281945705, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011383346281945705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08605306595563889, + 0.04928814247250557, + 0.14066845178604126, + 0.07220552116632462, + 0.2799092233181, + 0.1184997409582138, + 0.03491224721074104, + 0.09216907620429993, + 0.09940247237682343, + 0.02689206227660179 + ], + "credit_map": [ + 0.0069544375874102116, + 0.008171620778739452, + 0.012229123152792454, + 0.010189059190452099, + 0.012131454423069954, + 0.011768479831516743, + 0.016840899363160133, + 0.007767985109239817, + 0.003883144585415721, + 0.005771668162196875, + 0.007677782792598009, + 0.008227335289120674, + 0.007061505690217018, + 0.005140671972185373, + 0.0060928273014724255, + 0.005433208774775267, + 0.011901096440851688, + 0.011317775584757328, + 0.021748360246419907, + 0.02061440423130989, + 0.0154346888884902, + 0.032085079699754715, + 0.013295959681272507, + 0.01427109632641077, + 0.004999377299100161, + 0.0063421837985515594, + 0.012764595448970795, + 0.017465444281697273, + 0.005891803186386824, + 0.005699259229004383, + 0.009614216163754463, + 0.009428643621504307, + 0.008161645382642746, + 0.02338998205959797, + 0.02200986258685589, + 0.02901601791381836, + 0.013519737869501114, + 0.013887880370020866, + 0.11582784354686737, + 0.05409625917673111, + 0.009320843033492565, + 0.01047181710600853, + 0.014259479008615017, + 0.009376732632517815, + 0.012051058933138847, + 0.018154211342334747, + 0.023365486413240433, + 0.021500108763575554, + 0.0025605792179703712, + 0.0043309954926371574, + 0.005091700702905655, + 0.005311319138854742, + 0.004707443993538618, + 0.0031791389919817448, + 0.004585484508424997, + 0.005145586095750332, + 0.006372097413986921, + 0.014653121121227741, + 0.01697523146867752, + 0.01168108731508255, + 0.011400166898965836, + 0.0075869145803153515, + 0.01316599827259779, + 0.010334462858736515, + 0.009257365018129349, + 0.004931453615427017, + 0.03246011957526207, + 0.026618342846632004, + 0.005926448851823807, + 0.004365288186818361, + 0.005776535719633102, + 0.01006691250950098, + 0.0025748403277248144, + 0.002408764325082302, + 0.0037934654392302036, + 0.003696031402796507, + 0.003225113498046994, + 0.006116773467510939, + 0.001971873454749584, + 0.0031052003614604473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.011383346281945705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.180171936750412 + ], + "opposite_logprob_tokens": [ + -1.8051719665527344 + ], + "logprob_reward": -0.180171936750412, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.180171936750412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052617765963077545, + 0.04341881722211838, + 0.015136998146772385, + 0.029452459886670113, + 0.6686329245567322, + 0.025704145431518555, + 0.11537794768810272, + 0.01274997927248478, + 0.02381390705704689, + 0.013095063157379627 + ], + "credit_map": [ + 0.003843961749225855, + 0.004199683666229248, + 0.0050528934225440025, + 0.006272055208683014, + 0.005964822135865688, + 0.008018936030566692, + 0.010409773327410221, + 0.008855639956891537, + 0.002798278583213687, + 0.004247126169502735, + 0.00575074041262269, + 0.00592164508998394, + 0.005826917476952076, + 0.0064455242827534676, + 0.007407010067254305, + 0.005021571647375822, + 0.001026731333695352, + 0.001362511538900435, + 0.001554492861032486, + 0.0023901488166302443, + 0.0016168527072295547, + 0.0023555709049105644, + 0.0028925947844982147, + 0.0019380953162908554, + 0.0028397233691066504, + 0.003716329811140895, + 0.0030506884213536978, + 0.005634917411953211, + 0.004016593564301729, + 0.0028413867112249136, + 0.00394643098115921, + 0.003406390780583024, + 0.01624419540166855, + 0.057794708758592606, + 0.051520127803087234, + 0.06482860445976257, + 0.031418733298778534, + 0.02516036666929722, + 0.283659428358078, + 0.1380067616701126, + 0.0017248140648007393, + 0.001661407994106412, + 0.002344500506296754, + 0.0020857735071331263, + 0.002790076658129692, + 0.004903850145637989, + 0.005607589613646269, + 0.004586131777614355, + 0.004310088697820902, + 0.030134666711091995, + 0.01786925457417965, + 0.04748468101024628, + 0.004896620288491249, + 0.0021833607461303473, + 0.0050743152387440205, + 0.003424959722906351, + 0.0011251085670664907, + 0.0019250285113230348, + 0.002424103207886219, + 0.0023476560600101948, + 0.0012030950747430325, + 0.001054260297678411, + 0.0013095230096951127, + 0.0013612053589895368, + 0.001982850255444646, + 0.0011670960811898112, + 0.008165672421455383, + 0.006147839594632387, + 0.001340447342954576, + 0.0011197051499038935, + 0.0017710138345137239, + 0.0021192796994000673, + 0.0008870902820490301, + 0.0012791177723556757, + 0.0023411926813423634, + 0.0016439359169453382, + 0.002976379357278347, + 0.0008578866836614907, + 0.0022829121444374323, + 0.0008265484939329326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.180171936750412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539356015622616 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539356015622616, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539356015622616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5214267373085022, + 0.059962574392557144, + 0.03788619115948677, + 0.024197684600949287, + 0.1045440062880516, + 0.08335187286138535, + 0.11135128885507584, + 0.02149859629571438, + 0.02383064478635788, + 0.011950376443564892 + ], + "credit_map": [ + 0.018216464668512344, + 0.01526031456887722, + 0.18568748235702515, + 0.06065155938267708, + 0.03056156635284424, + 0.04120386019349098, + 0.04235900938510895, + 0.12748649716377258, + 0.005510269198566675, + 0.008778378367424011, + 0.008423807099461555, + 0.0076214331202209, + 0.0077583203092217445, + 0.006722333375364542, + 0.008051391690969467, + 0.007096644025295973, + 0.0033843815326690674, + 0.003060877090319991, + 0.0049255904741585255, + 0.005699031054973602, + 0.003590319072827697, + 0.00808115303516388, + 0.00401002774015069, + 0.005134814418852329, + 0.0015118931187316775, + 0.002055393997579813, + 0.004000140819698572, + 0.00620475597679615, + 0.0020418744534254074, + 0.002141458448022604, + 0.003029878716915846, + 0.003212289186194539, + 0.00589719507843256, + 0.008809654973447323, + 0.011952532455325127, + 0.014450049959123135, + 0.007284459192305803, + 0.008341951295733452, + 0.031118197366595268, + 0.016689976677298546, + 0.006721883546561003, + 0.006919356528669596, + 0.008979734033346176, + 0.00813094712793827, + 0.009074801579117775, + 0.014600177295506, + 0.016059868037700653, + 0.012865109369158745, + 0.004170547239482403, + 0.027232954278588295, + 0.01785433664917946, + 0.04517485201358795, + 0.005340703297406435, + 0.0025515169836580753, + 0.0054399496875703335, + 0.003586436156183481, + 0.001872848835773766, + 0.004377458244562149, + 0.003153336700052023, + 0.003473423421382904, + 0.0025073690339922905, + 0.0014432875905185938, + 0.0023885066621005535, + 0.002282367553561926, + 0.0021400838159024715, + 0.0017748140962794423, + 0.006013353820890188, + 0.0021535945124924183, + 0.0027127342764288187, + 0.0014064518036320806, + 0.004001418128609657, + 0.0036281943321228027, + 0.0009869844652712345, + 0.0017875018529593945, + 0.002200027462095022, + 0.0010902106296271086, + 0.0026655762922018766, + 0.0010107186390087008, + 0.0011451225727796555, + 0.0010642351116985083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.012539356015622616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32538118958473206, + 0.049155302345752716, + 0.043578870594501495, + 0.09001684933900833, + 0.08469892293214798, + 0.09747961908578873, + 0.24355366826057434, + 0.03092123195528984, + 0.022240636870265007, + 0.012973755598068237 + ], + "credit_map": [ + 0.011133935302495956, + 0.010359255596995354, + 0.11545853316783905, + 0.04483423009514809, + 0.016726933419704437, + 0.017638172954320908, + 0.019929075613617897, + 0.08930103480815887, + 0.00432788347825408, + 0.007494128774851561, + 0.0064012729562819, + 0.004900929983705282, + 0.00437752204015851, + 0.006471860222518444, + 0.009882486425340176, + 0.005299214273691177, + 0.0036784352269023657, + 0.0034570512361824512, + 0.008325841277837753, + 0.005142247304320335, + 0.0045272549614310265, + 0.008212906308472157, + 0.005622821860015392, + 0.004612312186509371, + 0.007798629347234964, + 0.009383470751345158, + 0.010289003141224384, + 0.01909332349896431, + 0.012331651523709297, + 0.007694729138165712, + 0.012707089073956013, + 0.010718952864408493, + 0.004546744748950005, + 0.009307174012064934, + 0.010750377550721169, + 0.014990217052400112, + 0.0058145588263869286, + 0.006959599442780018, + 0.02274460345506668, + 0.009585647843778133, + 0.008550717495381832, + 0.008676082827150822, + 0.013056322000920773, + 0.01006417814642191, + 0.01304719690233469, + 0.013557166792452335, + 0.01632028818130493, + 0.014207661151885986, + 0.008919322863221169, + 0.062043920159339905, + 0.038643382489681244, + 0.10054467618465424, + 0.010563827119767666, + 0.004898971412330866, + 0.011237315833568573, + 0.006702248472720385, + 0.0026429772842675447, + 0.0055782063864171505, + 0.005412823986262083, + 0.004178926348686218, + 0.0031481946352869272, + 0.002735209185630083, + 0.0035293239634484053, + 0.003695570630952716, + 0.0027257371693849564, + 0.0014792870497331023, + 0.0026665523182600737, + 0.004911900497972965, + 0.0017103208228945732, + 0.0013407183578237891, + 0.004184046760201454, + 0.003222075290977955, + 0.0020075200591236353, + 0.0014696032740175724, + 0.00127784616779536, + 0.001950146397575736, + 0.0015704783145338297, + 0.0016765007749199867, + 0.0016543070087209344, + 0.0013673531357198954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0024877809919416904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233138203620911 + ], + "opposite_logprob_tokens": [ + -0.8983138203620911 + ], + "logprob_reward": -0.5233138203620911, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233138203620911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09322791546583176, + 0.04206415265798569, + 0.046413153409957886, + 0.06449200212955475, + 0.5638474225997925, + 0.08975376933813095, + 0.029469821602106094, + 0.02489345334470272, + 0.03640849143266678, + 0.00942988134920597 + ], + "credit_map": [ + 0.004400485660880804, + 0.004941561724990606, + 0.017278708517551422, + 0.010747941210865974, + 0.006810057442635298, + 0.010035740211606026, + 0.014997370541095734, + 0.02401605248451233, + 0.0034340887796133757, + 0.005669774487614632, + 0.0055078319273889065, + 0.0055918931029737, + 0.0037320104893296957, + 0.004610850941389799, + 0.008259625174105167, + 0.005258078221231699, + 0.003408943535760045, + 0.003036217764019966, + 0.009647145867347717, + 0.006130937021225691, + 0.004685464780777693, + 0.01094758603721857, + 0.004195665940642357, + 0.004361190367490053, + 0.005592735018581152, + 0.006537795066833496, + 0.008617819286882877, + 0.014081388711929321, + 0.008256576955318451, + 0.005522985477000475, + 0.00888169463723898, + 0.007001005113124847, + 0.011479811742901802, + 0.05116099491715431, + 0.038491662591695786, + 0.05404573306441307, + 0.021070042625069618, + 0.015852708369493484, + 0.25395092368125916, + 0.11779554933309555, + 0.007474889978766441, + 0.006981313228607178, + 0.011741024442017078, + 0.009234413504600525, + 0.009782894514501095, + 0.015688374638557434, + 0.014493248425424099, + 0.014357606880366802, + 0.00136156368535012, + 0.006895578000694513, + 0.004078174941241741, + 0.010149797424674034, + 0.0015542252222076058, + 0.0010494496673345566, + 0.002268613316118717, + 0.002112417947500944, + 0.0019542670343071222, + 0.003830583766102791, + 0.005409713834524155, + 0.0028869332745671272, + 0.0029533004853874445, + 0.0016145285917446017, + 0.0036096510011702776, + 0.002634473843500018, + 0.003561899531632662, + 0.00252356450073421, + 0.01159555371850729, + 0.008621894754469395, + 0.002873303834348917, + 0.0016325076576322317, + 0.0021793886553496122, + 0.0034203757531940937, + 0.00093043566448614, + 0.0007918900228105485, + 0.0018855445086956024, + 0.001092070946469903, + 0.0013150505255907774, + 0.001502356375567615, + 0.0007561714737676084, + 0.0011563620064407587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.5233138203620911 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera rapidly tilts upward, capturing a flag made of heavy, faded crimson cotton and linen fabric. Initially still and drooping, the flag is coated in a thick layer of grayish-white dust, deeply embedded in its rough texture, its edges frayed and minutely torn. Suddenly, a powerful gust of wind sweeps in, lifting the flag uncontrollably upward, instantly dispersing the accumulated dust and forming a fleeting, murky cloud beneath the harsh, desolate sunlight. In this moment, the camera focuses on the fabric's microscopic details, clearly showing each individual fiber trembling and vibrating at high frequency, as if in a silent scream. The entire flag billows and twitches violently in the wind, while light pierces the dense dust, forming distinct beams of light, creating a dramatic atmosphere imbued with power, desolation, and a profound sense of historical vicissitudes.", + "seed": 15912, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15912_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1809136433294043, + "vlm_logprob": -0.1809136433294043, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.1809136433294043, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019275493919849396 + ], + "opposite_logprob_tokens": [ + -5.019275665283203 + ], + "logprob_reward": -0.019275493919849396, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.019275493919849396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2749449610710144, + 0.06312733143568039, + 0.12102203816175461, + 0.10361351817846298, + 0.1009586900472641, + 0.1133541464805603, + 0.0504813976585865, + 0.07131973654031754, + 0.06854598969221115, + 0.032632190734148026 + ], + "credit_map": [ + 0.020715296268463135, + 0.03188696503639221, + 0.037247829139232635, + 0.043783579021692276, + 0.039016153663396835, + 0.0320984423160553, + 0.037219833582639694, + 0.03297687694430351, + 0.005269257817417383, + 0.006067774258553982, + 0.01199396327137947, + 0.014760268852114677, + 0.00529881939291954, + 0.0039990730583667755, + 0.006900209467858076, + 0.008837971836328506, + 0.013102305121719837, + 0.014253580011427402, + 0.01482488214969635, + 0.021230624988675117, + 0.021542824804782867, + 0.011818794533610344, + 0.013187943026423454, + 0.011061088182032108, + 0.008898871950805187, + 0.019960273057222366, + 0.015952838584780693, + 0.017973091453313828, + 0.014133892953395844, + 0.00551587762311101, + 0.013247090391814709, + 0.007931592874228954, + 0.007273497059941292, + 0.023129982873797417, + 0.010880714282393456, + 0.010309198871254921, + 0.005706836003810167, + 0.0053849234245717525, + 0.03196410834789276, + 0.006309427320957184, + 0.01415827963501215, + 0.014773133210837841, + 0.02387094870209694, + 0.011699346825480461, + 0.013488012365996838, + 0.007479461841285229, + 0.014745793305337429, + 0.013139177113771439, + 0.006661843042820692, + 0.007316017523407936, + 0.005868051666766405, + 0.006488002371042967, + 0.007330349646508694, + 0.005465315189212561, + 0.004269538912922144, + 0.007082283031195402, + 0.023700689896941185, + 0.026029394939541817, + 0.0051626181229949, + 0.0024206817615777254, + 0.005493894685059786, + 0.0028756153769791126, + 0.0022005268838256598, + 0.003436310915276408, + 0.014277080073952675, + 0.009545911103487015, + 0.007715403102338314, + 0.008898595348000526, + 0.006178839132189751, + 0.006472286768257618, + 0.009030221961438656, + 0.006427651736885309, + 0.0035339489113539457, + 0.0025964737869799137, + 0.004124493803828955, + 0.004711353685706854, + 0.00287288217805326, + 0.007335367146879435, + 0.003933257889002562, + 0.0035244161263108253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.019275493919849396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1801937371492386 + ], + "opposite_logprob_tokens": [ + -1.8051937818527222 + ], + "logprob_reward": -0.1801937371492386, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1801937371492386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17006029188632965, + 0.06987390667200089, + 0.058707959949970245, + 0.14929896593093872, + 0.043107494711875916, + 0.04599371179938316, + 0.1751241683959961, + 0.20561935007572174, + 0.06379752606153488, + 0.01841657981276512 + ], + "credit_map": [ + 0.008247727528214455, + 0.023303719237446785, + 0.033377110958099365, + 0.036118049174547195, + 0.01627298817038536, + 0.017403623089194298, + 0.020039677619934082, + 0.015297395177185535, + 0.007671051193028688, + 0.006486407481133938, + 0.015000714920461178, + 0.010114114731550217, + 0.006684248335659504, + 0.00469612842425704, + 0.009023312479257584, + 0.010197929106652737, + 0.005932799074798822, + 0.00812885258346796, + 0.008112730458378792, + 0.011811157688498497, + 0.0055700684897601604, + 0.0041697658598423, + 0.008434046059846878, + 0.006548543460667133, + 0.012726441025733948, + 0.02215701900422573, + 0.020163899287581444, + 0.021773213520646095, + 0.02726537361741066, + 0.011063328944146633, + 0.022557370364665985, + 0.01159230899065733, + 0.004426994360983372, + 0.009547071531414986, + 0.003896849462762475, + 0.005582693498581648, + 0.002695660572499037, + 0.0032273847609758377, + 0.00974366907030344, + 0.003987172618508339, + 0.0064801680855453014, + 0.011018235236406326, + 0.006428765133023262, + 0.004396884702146053, + 0.0037146294489502907, + 0.0025464785285294056, + 0.0060480800457298756, + 0.005360467359423637, + 0.02496458776295185, + 0.022295422852039337, + 0.027400342747569084, + 0.022204579785466194, + 0.025071218609809875, + 0.012026922777295113, + 0.02431410364806652, + 0.01684698276221752, + 0.07712102681398392, + 0.08731864392757416, + 0.011510335840284824, + 0.005168442148715258, + 0.009329925291240215, + 0.003894871100783348, + 0.004490610212087631, + 0.006785509642213583, + 0.010282871313393116, + 0.007966802455484867, + 0.0059894719161093235, + 0.007698629051446915, + 0.008690407499670982, + 0.005966463591903448, + 0.010334708727896214, + 0.006868166383355856, + 0.0011087736347690225, + 0.0013321484439074993, + 0.003216609824448824, + 0.0027145612984895706, + 0.002247157273814082, + 0.004050559364259243, + 0.0017517202068120241, + 0.001995051046833396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.1801937371492386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978813599795103 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.00978813599795103, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978813599795103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6367855072021484, + 0.06075810268521309, + 0.09033926576375961, + 0.0740034431219101, + 0.02891782857477665, + 0.01583573967218399, + 0.04446708783507347, + 0.025580715388059616, + 0.0164451003074646, + 0.006867255084216595 + ], + "credit_map": [ + 0.0327235646545887, + 0.06813870370388031, + 0.1383701115846634, + 0.13906115293502808, + 0.04951898753643036, + 0.053554292768239975, + 0.0862567275762558, + 0.06916186213493347, + 0.002914681565016508, + 0.004151197150349617, + 0.014172454364597797, + 0.020653625950217247, + 0.003775640157982707, + 0.002480266150087118, + 0.0075191776268184185, + 0.005091048311442137, + 0.00919314008206129, + 0.007181914057582617, + 0.012426084838807583, + 0.013241419568657875, + 0.007301293779164553, + 0.005596795119345188, + 0.025122519582509995, + 0.010276084765791893, + 0.006619587540626526, + 0.01278779935091734, + 0.011804924346506596, + 0.012491481378674507, + 0.010471636429429054, + 0.004880837164819241, + 0.00938886497169733, + 0.005558300763368607, + 0.0017304883804172277, + 0.010675016790628433, + 0.0018679631175473332, + 0.001632033265195787, + 0.0013288112822920084, + 0.0009498393628746271, + 0.009144739247858524, + 0.0015889297937974334, + 0.0016830087406560779, + 0.004420689307153225, + 0.0018386277370154858, + 0.0018258733907714486, + 0.001235266332514584, + 0.0007445178343914449, + 0.00235820934176445, + 0.001729544484987855, + 0.005824055057018995, + 0.006665342021733522, + 0.00444979639723897, + 0.005490005016326904, + 0.0070036184042692184, + 0.004751548171043396, + 0.003633577609434724, + 0.0066491374745965, + 0.008098432794213295, + 0.009155348874628544, + 0.0023135500960052013, + 0.0008814125903882086, + 0.0021181851625442505, + 0.0012384328292682767, + 0.0006902107270434499, + 0.0010851406259462237, + 0.003069961443543434, + 0.002223620656877756, + 0.002129272324964404, + 0.0018465984612703323, + 0.0018503781175240874, + 0.0017991053173318505, + 0.0022080282215029, + 0.0013181318063288927, + 0.0005917336093261838, + 0.0005822269595228136, + 0.0006417179247364402, + 0.0007996104541234672, + 0.0006600297638215125, + 0.0019188710721209645, + 0.0008194549009203911, + 0.00085360964294523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.00978813599795103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37197354435920715, + 0.14823280274868011, + 0.11695164442062378, + 0.09058304876089096, + 0.08208591490983963, + 0.03657142445445061, + 0.06781556457281113, + 0.04301753267645836, + 0.026341615244746208, + 0.016426932066679 + ], + "credit_map": [ + 0.023990392684936523, + 0.04331817105412483, + 0.06466609984636307, + 0.06850869208574295, + 0.03978802263736725, + 0.040420934557914734, + 0.05001450702548027, + 0.041266702115535736, + 0.006353114265948534, + 0.00847871694713831, + 0.044426579028367996, + 0.054612647742033005, + 0.006039243657141924, + 0.0049333940260112286, + 0.013792173005640507, + 0.009596943855285645, + 0.010658394545316696, + 0.009080281481146812, + 0.01564807817339897, + 0.018984558060765266, + 0.019484523683786392, + 0.012123823165893555, + 0.01668689399957657, + 0.014285089448094368, + 0.008655319921672344, + 0.01819618232548237, + 0.013461258262395859, + 0.011102940887212753, + 0.015171686187386513, + 0.005074539687484503, + 0.011795608326792717, + 0.007125507108867168, + 0.004029062110930681, + 0.024419564753770828, + 0.005209198221564293, + 0.006081335712224245, + 0.0039126756601035595, + 0.002546665957197547, + 0.03007715381681919, + 0.005810258444398642, + 0.0032078612130135298, + 0.010044872760772705, + 0.005798922386020422, + 0.003340713679790497, + 0.0034102860372513533, + 0.001770324888639152, + 0.00389657448977232, + 0.005101870279759169, + 0.011754874140024185, + 0.010517219081521034, + 0.009336937218904495, + 0.007804383989423513, + 0.007392787840217352, + 0.004261443857103586, + 0.007915589027106762, + 0.00883233267813921, + 0.013985677622258663, + 0.01574019528925419, + 0.003386122640222311, + 0.0016741567524150014, + 0.0034244318958371878, + 0.001744852401316166, + 0.0012241825461387634, + 0.0018379135290160775, + 0.00400206632912159, + 0.003490401664748788, + 0.0038037023041397333, + 0.002802171278744936, + 0.003217719728127122, + 0.0032013969030231237, + 0.0035212102811783552, + 0.002302946988493204, + 0.0017610149225220084, + 0.0016200027894228697, + 0.0025696875527501106, + 0.0020763669162988663, + 0.00167425989639014, + 0.0030922957230359316, + 0.0019536681938916445, + 0.001679634558968246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933678388595581 + ], + "opposite_logprob_tokens": [ + -0.6933678388595581 + ], + "logprob_reward": -0.6933678388595581, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933678388595581, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07682489603757858, + 0.08713903278112411, + 0.12225756049156189, + 0.032879605889320374, + 0.21611124277114868, + 0.15510304272174835, + 0.07326377183198929, + 0.17308776080608368, + 0.024954991415143013, + 0.03837815299630165 + ], + "credit_map": [ + 0.007535472512245178, + 0.008494055829942226, + 0.009432162158191204, + 0.011837559752166271, + 0.007661723531782627, + 0.010245992802083492, + 0.011832828633487225, + 0.009785096161067486, + 0.006537238135933876, + 0.007026872131973505, + 0.019320856779813766, + 0.02653542533516884, + 0.0051465812139213085, + 0.0035947049036622047, + 0.012173424474895, + 0.006803920958191156, + 0.012999720871448517, + 0.018555551767349243, + 0.016663076356053352, + 0.01936968043446541, + 0.013238449580967426, + 0.007818643003702164, + 0.0188034288585186, + 0.014808992855250835, + 0.002601293846964836, + 0.004251921083778143, + 0.0047704316675662994, + 0.0059448787942528725, + 0.004101197700947523, + 0.0027336920611560345, + 0.004929648712277412, + 0.003546537132933736, + 0.008325357921421528, + 0.07804404199123383, + 0.008646674454212189, + 0.01352350041270256, + 0.00810623075813055, + 0.004416014067828655, + 0.08521571755409241, + 0.009833697229623795, + 0.014264152385294437, + 0.056305594742298126, + 0.02180464379489422, + 0.015296366065740585, + 0.007857369258999825, + 0.005600633565336466, + 0.01046217791736126, + 0.02351209707558155, + 0.010438566096127033, + 0.009862100705504417, + 0.01078218687325716, + 0.010283008217811584, + 0.011624718084931374, + 0.0065751452930271626, + 0.0065880874171853065, + 0.0071099549531936646, + 0.06242156773805618, + 0.06992995738983154, + 0.008447518572211266, + 0.0045454795472323895, + 0.007748904172331095, + 0.005763336084783077, + 0.005423865746706724, + 0.008807118982076645, + 0.0037070002872496843, + 0.0032836899627000093, + 0.0020652241073548794, + 0.0027156402356922626, + 0.0029633932281285524, + 0.0032136260997503996, + 0.004193238914012909, + 0.0028131757862865925, + 0.0033967539202421904, + 0.0025363839231431484, + 0.0039083147421479225, + 0.004687219392508268, + 0.005426282528787851, + 0.009919422678649426, + 0.004114779178053141, + 0.004388992674648762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.6933678388595581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In an extreme close-up, the camera rapidly tilts upward, capturing a flag made of heavy, faded crimson cotton and linen fabric. Initially still and drooping, the flag is coated in a thick layer of grayish-white dust, deeply embedded in its rough texture, its edges frayed and minutely torn. Suddenly, a powerful gust of wind sweeps in, lifting the flag uncontrollably upward, instantly dispersing the accumulated dust and forming a fleeting, murky cloud beneath the harsh, desolate sunlight. In this moment, the camera focuses on the fabric's microscopic details, clearly showing each individual fiber trembling and vibrating at high frequency, as if in a silent scream. The entire flag billows and twitches violently in the wind, while light pierces the dense dust, forming distinct beams of light, creating a dramatic atmosphere imbued with power, desolation, and a profound sense of historical vicissitudes.", + "seed": 15913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15913_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1371525769820437, + "vlm_logprob": -0.1371525769820437, + "vlm_yes_no_margin": 3.799999988079071, + "vlm_token_credit": -0.1371525769820437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018424442037940025 + ], + "opposite_logprob_tokens": [ + -4.7684245109558105 + ], + "logprob_reward": -0.018424442037940025, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.018424442037940025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013610191643238068, + 0.019514963030815125, + 0.016747096553444862, + 0.010641001164913177, + 0.04365716129541397, + 0.03165951743721962, + 0.23332689702510834, + 0.2672397792339325, + 0.012619908899068832, + 0.3509834110736847 + ], + "credit_map": [ + 0.0023558761458843946, + 0.001827916712500155, + 0.0014058757806196809, + 0.0017504439456388354, + 0.001775982789695263, + 0.0013880300102755427, + 0.001385749434120953, + 0.001720318105071783, + 0.0017483734991401434, + 0.003690210869535804, + 0.002737157978117466, + 0.0031196894124150276, + 0.002428282517939806, + 0.0017454661428928375, + 0.0018316981149837375, + 0.0022140881046652794, + 0.0023128599859774113, + 0.0015756285283714533, + 0.002973281079903245, + 0.0020990106277167797, + 0.003494906472042203, + 0.001978108659386635, + 0.0008722921484149992, + 0.0014410113217309117, + 0.0016079021152108908, + 0.0012292993487790227, + 0.001091823447495699, + 0.001936394372023642, + 0.0010010559344664216, + 0.0010297202970832586, + 0.0009389478946104646, + 0.0018058588029816747, + 0.011635618284344673, + 0.0036130279768258333, + 0.012011738494038582, + 0.00587652949616313, + 0.0027890638448297977, + 0.00214117718860507, + 0.0030718212947249413, + 0.0025181882083415985, + 0.00609548157081008, + 0.005765988491475582, + 0.006800701841711998, + 0.0035559511743485928, + 0.0020125675946474075, + 0.002334506018087268, + 0.003093679202720523, + 0.002000645734369755, + 0.008091390132904053, + 0.09806374460458755, + 0.10375001281499863, + 0.007625045720487833, + 0.009439727291464806, + 0.0018652165308594704, + 0.0022635480854660273, + 0.0022282302379608154, + 0.010271959006786346, + 0.11329105496406555, + 0.1107567697763443, + 0.008453900925815105, + 0.013575072400271893, + 0.004082873929291964, + 0.00368965370580554, + 0.0031185259576886892, + 0.0015330135356634855, + 0.0015578120946884155, + 0.0028612229507416487, + 0.002116311574354768, + 0.0010210437467321754, + 0.0011065135477110744, + 0.0013999606017023325, + 0.0010240322444587946, + 0.06345821171998978, + 0.14815999567508698, + 0.015869418159127235, + 0.019871702417731285, + 0.040976349264383316, + 0.01841096580028534, + 0.029924536123871803, + 0.014312260784208775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.018424442037940025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31392890214920044 + ], + "opposite_logprob_tokens": [ + -1.3139288425445557 + ], + "logprob_reward": -0.31392890214920044, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31392890214920044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044194817543029785, + 0.06672331690788269, + 0.04629216343164444, + 0.01039730105549097, + 0.0723387748003006, + 0.0878884345293045, + 0.2147369533777237, + 0.14433221518993378, + 0.01006488036364317, + 0.30303114652633667 + ], + "credit_map": [ + 0.007303970400243998, + 0.005906233564019203, + 0.004175166133791208, + 0.005298633128404617, + 0.004218973685055971, + 0.0036442172713577747, + 0.005851393099874258, + 0.007796228863298893, + 0.0060531580820679665, + 0.009283671155571938, + 0.006279734894633293, + 0.011955770663917065, + 0.012325751595199108, + 0.009900984354317188, + 0.005669177509844303, + 0.005255068186670542, + 0.007017227355390787, + 0.003947018645703793, + 0.010500898584723473, + 0.0029574681539088488, + 0.012519312091171741, + 0.00543029373511672, + 0.001889116014353931, + 0.0020308259408921003, + 0.0013646002626046538, + 0.0011365171521902084, + 0.0010987760033458471, + 0.001529906177893281, + 0.001221233163960278, + 0.0013451825361698866, + 0.0012006909819319844, + 0.0015003945445641875, + 0.016394425183534622, + 0.011266822926700115, + 0.017011607065796852, + 0.010218439623713493, + 0.004955297335982323, + 0.00322751235216856, + 0.004712862893939018, + 0.004551807884126902, + 0.01822175830602646, + 0.018346885219216347, + 0.0197035800665617, + 0.009065886959433556, + 0.005740463268011808, + 0.005461275577545166, + 0.006720856297761202, + 0.004627719521522522, + 0.008652213029563427, + 0.08617886155843735, + 0.09116180986166, + 0.009842327795922756, + 0.008028823882341385, + 0.0028409725055098534, + 0.004460176452994347, + 0.003571751993149519, + 0.005583013407886028, + 0.0616987869143486, + 0.05980400741100311, + 0.004379583057016134, + 0.008092659525573254, + 0.0015739600639790297, + 0.0017651240341365337, + 0.001435071462765336, + 0.0010155219351872802, + 0.0013052105205133557, + 0.0022740827407687902, + 0.0016205357387661934, + 0.0009942010510712862, + 0.0008129262714646757, + 0.0012280161026865244, + 0.0008143839659169316, + 0.05538135766983032, + 0.1284291297197342, + 0.01390462089329958, + 0.015869243070483208, + 0.03564686328172684, + 0.016138598322868347, + 0.025218656286597252, + 0.012442638166248798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.31392890214920044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036204999778419733 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036204999778419733, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036204999778419733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02590956538915634, + 0.03178992122411728, + 0.06254477053880692, + 0.010040702298283577, + 0.027707278728485107, + 0.03145703673362732, + 0.5185642242431641, + 0.184565931558609, + 0.007333059329539537, + 0.10008745640516281 + ], + "credit_map": [ + 0.004979616031050682, + 0.0035124679561704397, + 0.0032566417939960957, + 0.002689699176698923, + 0.0027041814755648375, + 0.0025718288961797953, + 0.0029486583080142736, + 0.003246471518650651, + 0.002001395681872964, + 0.00520889088511467, + 0.0026521107647567987, + 0.003953020088374615, + 0.006424830295145512, + 0.005642037373036146, + 0.003433091565966606, + 0.002474543172866106, + 0.009629474021494389, + 0.004682351369410753, + 0.014546588063240051, + 0.003989357501268387, + 0.016802920028567314, + 0.007365796715021133, + 0.002599041908979416, + 0.002929240232333541, + 0.0016723430017009377, + 0.001271640881896019, + 0.0009902440942823887, + 0.0014351251302286983, + 0.0011754859006032348, + 0.0011455989442765713, + 0.0008202898898161948, + 0.0015299749793484807, + 0.0030401854310184717, + 0.0031809962820261717, + 0.0034849774092435837, + 0.00732630118727684, + 0.002692022593691945, + 0.0021278169006109238, + 0.0032627603504806757, + 0.0025922174099832773, + 0.00667768158018589, + 0.004462411627173424, + 0.0076124826446175575, + 0.00283208885230124, + 0.0021561014000326395, + 0.0025274879299104214, + 0.002843230264261365, + 0.002345555229112506, + 0.01837385818362236, + 0.21763432025909424, + 0.23043957352638245, + 0.017461078241467476, + 0.020598020404577255, + 0.004317722748965025, + 0.005106301046907902, + 0.004633392207324505, + 0.007095387205481529, + 0.0790516808629036, + 0.07736719399690628, + 0.005586913786828518, + 0.009233525954186916, + 0.0023418557830154896, + 0.0021783073898404837, + 0.001711058896034956, + 0.001005177036859095, + 0.0008090496412478387, + 0.0012892732629552484, + 0.0012608913239091635, + 0.0006175429443828762, + 0.0008706850348971784, + 0.0008173772948794067, + 0.0006630626157857478, + 0.018227238208055496, + 0.04227922856807709, + 0.004744669422507286, + 0.0053754947148263454, + 0.011826342903077602, + 0.005577751435339451, + 0.00802383478730917, + 0.004032895900309086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0036204999778419733 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011812621960416436 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011812621960416436, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011812621960416436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02124791219830513, + 0.018782207742333412, + 0.02386036328971386, + 0.0075491368770599365, + 0.04587029665708542, + 0.020711375400424004, + 0.48318424820899963, + 0.29876360297203064, + 0.008440660312771797, + 0.07159020006656647 + ], + "credit_map": [ + 0.0026483836118131876, + 0.0030839366372674704, + 0.002161612268537283, + 0.003236432559788227, + 0.002790174912661314, + 0.00204708194360137, + 0.002236239379271865, + 0.0030440506525337696, + 0.0011404273100197315, + 0.002787622855976224, + 0.001071099890395999, + 0.002645211759954691, + 0.004410906229168177, + 0.0036828401498496532, + 0.0015473831444978714, + 0.0014967146562412381, + 0.0038094979245215654, + 0.0017044651322066784, + 0.005378340370953083, + 0.0020012082532048225, + 0.005797325633466244, + 0.0026101877447217703, + 0.001092923921532929, + 0.0014664113987237215, + 0.001661563292145729, + 0.0008477561641484499, + 0.001041831448674202, + 0.0009845958556979895, + 0.0007206545560620725, + 0.0007181771798059344, + 0.0006627932889387012, + 0.0009117641602642834, + 0.01231812871992588, + 0.004842039197683334, + 0.013548417948186398, + 0.006555305328220129, + 0.0018890511710196733, + 0.0013138714712113142, + 0.0027766432613134384, + 0.002626838395372033, + 0.004458321258425713, + 0.004253224469721317, + 0.0045349737629294395, + 0.002067520748823881, + 0.0015534493140876293, + 0.0011090236948803067, + 0.0016810829984024167, + 0.0010537771740928292, + 0.017863670364022255, + 0.20218150317668915, + 0.21441525220870972, + 0.016438186168670654, + 0.01965653896331787, + 0.0037735954392701387, + 0.004597509279847145, + 0.004258013796061277, + 0.011749409139156342, + 0.1266305297613144, + 0.12365403771400452, + 0.009461945854127407, + 0.01503303274512291, + 0.004729699343442917, + 0.004253429360687733, + 0.003251495771110058, + 0.0009639953495934606, + 0.0009962361073121428, + 0.0017551399068906903, + 0.0015349917812272906, + 0.0008222464239224792, + 0.0006943918415345252, + 0.0009757158113643527, + 0.0006979428580962121, + 0.013057064265012741, + 0.030243201181292534, + 0.0034035458229482174, + 0.0037405965849757195, + 0.008359060622751713, + 0.004100635182112455, + 0.005759302526712418, + 0.0029267838690429926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0011812621960416436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34860777854919434 + ], + "opposite_logprob_tokens": [ + -1.2236077785491943 + ], + "logprob_reward": -0.34860777854919434, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34860777854919434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04534298926591873, + 0.0442737452685833, + 0.04939641058444977, + 0.0822955071926117, + 0.1454542726278305, + 0.1185610368847847, + 0.14819027483463287, + 0.06734960526227951, + 0.20806972682476044, + 0.09106635302305222 + ], + "credit_map": [ + 0.007459883112460375, + 0.006642931140959263, + 0.005252419039607048, + 0.005932270549237728, + 0.004797189496457577, + 0.004448604770004749, + 0.005795159377157688, + 0.005014535039663315, + 0.004485498648136854, + 0.0046340287663042545, + 0.0054083033464848995, + 0.008842072449624538, + 0.006045089568942785, + 0.005019249394536018, + 0.005099250935018063, + 0.004740254022181034, + 0.005885131191462278, + 0.006558633409440517, + 0.004347996320575476, + 0.00967327505350113, + 0.005326886661350727, + 0.006458823103457689, + 0.004492751322686672, + 0.006652919575572014, + 0.014285542070865631, + 0.010892393067479134, + 0.00907689705491066, + 0.011021151207387447, + 0.007958708330988884, + 0.008578687906265259, + 0.007018517702817917, + 0.01346361730247736, + 0.017346827313303947, + 0.03334497660398483, + 0.016801362857222557, + 0.030498461797833443, + 0.01305878534913063, + 0.008700359612703323, + 0.01369116548448801, + 0.012012342922389507, + 0.027755288407206535, + 0.024456121027469635, + 0.02537149377167225, + 0.010734938085079193, + 0.010457971133291721, + 0.004926593508571386, + 0.008268003351986408, + 0.006590640638023615, + 0.006986468564718962, + 0.05428940802812576, + 0.055696360766887665, + 0.008698646910488605, + 0.00942324660718441, + 0.00411518570035696, + 0.005020963493734598, + 0.003960018511861563, + 0.008211522363126278, + 0.018357448279857635, + 0.016243554651737213, + 0.005674222484230995, + 0.007336494978517294, + 0.005454507656395435, + 0.0036301796790212393, + 0.002441681921482086, + 0.022817401215434074, + 0.024341503158211708, + 0.048656631261110306, + 0.037682514637708664, + 0.019632166251540184, + 0.016585061326622963, + 0.021381031721830368, + 0.01697343774139881, + 0.013403178192675114, + 0.031092515215277672, + 0.006728317588567734, + 0.00953507050871849, + 0.010060346685349941, + 0.006610753946006298, + 0.007400850765407085, + 0.006235326640307903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.34860777854919434 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera rapidly tilts upward, capturing a flag made of heavy, faded crimson cotton and linen fabric. Initially still and drooping, the flag is coated in a thick layer of grayish-white dust, deeply embedded in its rough texture, its edges frayed and minutely torn. Suddenly, a powerful gust of wind sweeps in, lifting the flag uncontrollably upward, instantly dispersing the accumulated dust and forming a fleeting, murky cloud beneath the harsh, desolate sunlight. In this moment, the camera focuses on the fabric's microscopic details, clearly showing each individual fiber trembling and vibrating at high frequency, as if in a silent scream. The entire flag billows and twitches violently in the wind, while light pierces the dense dust, forming distinct beams of light, creating a dramatic atmosphere imbued with power, desolation, and a profound sense of historical vicissitudes.", + "seed": 15914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_7/index_8665_seed_15914_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21456568972207607, + "vlm_logprob": -0.21456568972207607, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.21456568972207607, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018701888620853424 + ], + "opposite_logprob_tokens": [ + -4.518702030181885 + ], + "logprob_reward": -0.018701888620853424, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.018701888620853424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09891805797815323, + 0.04678765684366226, + 0.08388536423444748, + 0.14141663908958435, + 0.052280958741903305, + 0.05495895445346832, + 0.03390174359083176, + 0.07310208678245544, + 0.18661528825759888, + 0.22813330590724945 + ], + "credit_map": [ + 0.008549409918487072, + 0.010310946963727474, + 0.010998635552823544, + 0.018422726541757584, + 0.011457572691142559, + 0.00938411895185709, + 0.016970155760645866, + 0.012824500910937786, + 0.005439257249236107, + 0.005043399520218372, + 0.004814082756638527, + 0.0069586788304150105, + 0.00530878733843565, + 0.005575011949986219, + 0.008029421791434288, + 0.005619017872959375, + 0.008985050953924656, + 0.008935275487601757, + 0.009119394235312939, + 0.0183158740401268, + 0.007584377191960812, + 0.009756063111126423, + 0.013732139952480793, + 0.007457192987203598, + 0.011931913904845715, + 0.01663140021264553, + 0.01357936765998602, + 0.02299833483994007, + 0.016279185190796852, + 0.01390869915485382, + 0.024767611175775528, + 0.021320128813385963, + 0.006191766355186701, + 0.007561566773802042, + 0.0060597858391702175, + 0.012864284217357635, + 0.005316767375916243, + 0.0026758122257888317, + 0.00667010759934783, + 0.004940870217978954, + 0.009503565728664398, + 0.012916757725179195, + 0.005474812351167202, + 0.008695024996995926, + 0.0038109866436570883, + 0.0037042531184852123, + 0.006627412047237158, + 0.004226143471896648, + 0.004663356579840183, + 0.004480098374187946, + 0.0031341772992163897, + 0.002820058725774288, + 0.004005797673016787, + 0.005212939344346523, + 0.005819165147840977, + 0.003766152309253812, + 0.007558612618595362, + 0.010722711682319641, + 0.005394230131059885, + 0.01946176402270794, + 0.008276134729385376, + 0.007019404321908951, + 0.010322137735784054, + 0.004347092937678099, + 0.01774330995976925, + 0.00937605183571577, + 0.05612238869071007, + 0.01516665332019329, + 0.018434938043355942, + 0.012183966115117073, + 0.038523782044649124, + 0.019064201042056084, + 0.0258223507553339, + 0.06010611727833748, + 0.02258114702999592, + 0.05869641527533531, + 0.016704706475138664, + 0.009510539472103119, + 0.017914259806275368, + 0.016797758638858795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.018701888620853424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3140784204006195 + ], + "opposite_logprob_tokens": [ + -1.314078450202942 + ], + "logprob_reward": -0.3140784204006195, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3140784204006195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08584434539079666, + 0.050385747104883194, + 0.0685657411813736, + 0.09294333308935165, + 0.04320529103279114, + 0.08865232020616531, + 0.022960813716053963, + 0.12902803719043732, + 0.37547871470451355, + 0.042935676872730255 + ], + "credit_map": [ + 0.004098472185432911, + 0.006263313349336386, + 0.008058502338826656, + 0.015782922506332397, + 0.013544587418437004, + 0.006988438777625561, + 0.020893830806016922, + 0.01021427009254694, + 0.004571439232677221, + 0.005836979020386934, + 0.006600839551538229, + 0.00811463501304388, + 0.004822501912713051, + 0.005741924978792667, + 0.008935824036598206, + 0.005761600565165281, + 0.004661095328629017, + 0.006479606032371521, + 0.006226828321814537, + 0.023584293201565742, + 0.005073664244264364, + 0.005706751719117165, + 0.010449987836182117, + 0.006383513566106558, + 0.010095048695802689, + 0.016104912385344505, + 0.009329285472631454, + 0.014179526828229427, + 0.012256626971065998, + 0.0074689206667244434, + 0.012608406133949757, + 0.010900605469942093, + 0.0051294430159032345, + 0.00616619735956192, + 0.004009300842881203, + 0.009078320115804672, + 0.00490522151812911, + 0.002461544703692198, + 0.007529410533607006, + 0.003925855737179518, + 0.019579017534852028, + 0.02172171138226986, + 0.007425845135003328, + 0.009816967882215977, + 0.00748438248410821, + 0.005907685030251741, + 0.00977832730859518, + 0.006938381586223841, + 0.004466362297534943, + 0.002859443426132202, + 0.0017865020781755447, + 0.0015306398272514343, + 0.002489016856998205, + 0.004173217341303825, + 0.0027507119812071323, + 0.0029049196746200323, + 0.01679549366235733, + 0.019929973408579826, + 0.03214484825730324, + 0.021374467760324478, + 0.01347341388463974, + 0.008176065050065517, + 0.013055982068181038, + 0.004077782854437828, + 0.02609512396156788, + 0.017525987699627876, + 0.1559729427099228, + 0.025084255263209343, + 0.02290763333439827, + 0.012426248751580715, + 0.10167507827281952, + 0.013791420497000217, + 0.004407141823321581, + 0.007839181460440159, + 0.006405688356608152, + 0.007060812786221504, + 0.005581068340688944, + 0.0024185965303331614, + 0.005559928715229034, + 0.003663254901766777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.3140784204006195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043072305619716644 + ], + "opposite_logprob_tokens": [ + -3.16807222366333 + ], + "logprob_reward": -0.043072305619716644, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043072305619716644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15338297188282013, + 0.12731127440929413, + 0.09709946811199188, + 0.33019164204597473, + 0.038665059953927994, + 0.054030515253543854, + 0.02737121842801571, + 0.08335475623607635, + 0.04333360120654106, + 0.04525942727923393 + ], + "credit_map": [ + 0.006539655849337578, + 0.01152514573186636, + 0.013377510942518711, + 0.026268936693668365, + 0.02239665389060974, + 0.012976988218724728, + 0.04140534996986389, + 0.018892740830779076, + 0.012560390867292881, + 0.015238975174725056, + 0.018689624965190887, + 0.019889209419488907, + 0.011102201417088509, + 0.01418311893939972, + 0.0204244926571846, + 0.015223265625536442, + 0.010519716888666153, + 0.009106184355914593, + 0.00881197303533554, + 0.023108959197998047, + 0.009516437537968159, + 0.011212005279958248, + 0.013053033500909805, + 0.011771164834499359, + 0.02269301936030388, + 0.04656268283724785, + 0.027361087501049042, + 0.0484197624027729, + 0.03301909938454628, + 0.036393746733665466, + 0.05851840227842331, + 0.0572238564491272, + 0.005748979281634092, + 0.004489005543291569, + 0.003360142931342125, + 0.00866913702338934, + 0.003942922223359346, + 0.0018986021168529987, + 0.005963428411632776, + 0.004592845216393471, + 0.012749598361551762, + 0.017739541828632355, + 0.0032913011964410543, + 0.0063736289739608765, + 0.003402582136914134, + 0.0025707159657031298, + 0.005076178349554539, + 0.0028269654139876366, + 0.004096985794603825, + 0.002976533491164446, + 0.0030136851128190756, + 0.0024753049947321415, + 0.003661759663373232, + 0.0040029436349868774, + 0.004426552914083004, + 0.002717451425269246, + 0.009165680035948753, + 0.015531663782894611, + 0.010488789528608322, + 0.014791253954172134, + 0.01158207282423973, + 0.00748617947101593, + 0.01115406770259142, + 0.0031550542917102575, + 0.0026591643691062927, + 0.0032017005141824484, + 0.009634682908654213, + 0.004392031580209732, + 0.003954948857426643, + 0.0029430135618895292, + 0.011445177718997002, + 0.005102882627397776, + 0.0030429274775087833, + 0.008636221289634705, + 0.0061727543361485004, + 0.008354402147233486, + 0.004855455830693245, + 0.0025195118505507708, + 0.005528383422642946, + 0.006149772088974714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.043072305619716644 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036171742249280214 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036171742249280214, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036171742249280214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07714471966028214, + 0.07963664829730988, + 0.12265514582395554, + 0.19517043232917786, + 0.08550243079662323, + 0.09479709714651108, + 0.029560476541519165, + 0.07987908273935318, + 0.1941189467906952, + 0.04153502359986305 + ], + "credit_map": [ + 0.0040643587708473206, + 0.005631802137941122, + 0.008928438648581505, + 0.012718337588012218, + 0.009471269324421883, + 0.005107387434691191, + 0.0214993916451931, + 0.00972373504191637, + 0.007424047216773033, + 0.008431337773799896, + 0.011523891240358353, + 0.014366504736244678, + 0.007043229881674051, + 0.00959279015660286, + 0.010521840304136276, + 0.010733005590736866, + 0.012136430479586124, + 0.011431340128183365, + 0.012496192008256912, + 0.02951209992170334, + 0.012514972127974033, + 0.015696726739406586, + 0.018841998651623726, + 0.01002538949251175, + 0.014902211725711823, + 0.02979557402431965, + 0.016177957877516747, + 0.02782164141535759, + 0.019965309649705887, + 0.02109876088798046, + 0.03220566362142563, + 0.033203303813934326, + 0.010700693354010582, + 0.013942766934633255, + 0.00876968540251255, + 0.02371262013912201, + 0.0069174692034721375, + 0.003969200886785984, + 0.009704327210783958, + 0.00778566487133503, + 0.025781339034438133, + 0.03515847772359848, + 0.004133003763854504, + 0.009187119081616402, + 0.006112396717071533, + 0.004130707122385502, + 0.006451992783695459, + 0.0038420564960688353, + 0.004192535299807787, + 0.002822312293574214, + 0.002787188161164522, + 0.0023841809015721083, + 0.004109278786927462, + 0.004915218334645033, + 0.004911419004201889, + 0.003438344458118081, + 0.010346714407205582, + 0.0163810383528471, + 0.01681511104106903, + 0.018472662195563316, + 0.008276475593447685, + 0.0029159891419112682, + 0.005020134150981903, + 0.0016509565757587552, + 0.012606631964445114, + 0.010466535575687885, + 0.08276420831680298, + 0.01320763397961855, + 0.010285218246281147, + 0.00786937028169632, + 0.048960309475660324, + 0.007959041744470596, + 0.003096314612776041, + 0.006924009881913662, + 0.005099043250083923, + 0.006151778623461723, + 0.006743854843080044, + 0.0038798220921307802, + 0.005571724846959114, + 0.004068476147949696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.0036171742249280214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up of a flag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flag made of heavy, faded crimson cotton and linen fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flag initially shown still, drooping, and coated in a thick layer of grayish-white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a powerful gust of wind lift the flag uncontrollably upward and disperse the dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933586597442627 + ], + "opposite_logprob_tokens": [ + -0.6933586597442627 + ], + "logprob_reward": -0.6933586597442627, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933586597442627, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14366930723190308, + 0.0698355883359909, + 0.08167025446891785, + 0.14557641744613647, + 0.08352397382259369, + 0.12896524369716644, + 0.05214857682585716, + 0.1215820387005806, + 0.11883906275033951, + 0.05418950691819191 + ], + "credit_map": [ + 0.007089713122695684, + 0.013553328812122345, + 0.014602607116103172, + 0.02251208946108818, + 0.02117958478629589, + 0.01204600464552641, + 0.035200439393520355, + 0.017485545948147774, + 0.005833356641232967, + 0.005560788791626692, + 0.01294083334505558, + 0.011341576464474201, + 0.00561545742675662, + 0.008954940363764763, + 0.009820888750255108, + 0.009767746552824974, + 0.010248507373034954, + 0.008635835722088814, + 0.008065495640039444, + 0.0175795741379261, + 0.008040245622396469, + 0.007825849577784538, + 0.014097923412919044, + 0.007176822982728481, + 0.014284223318099976, + 0.017484156414866447, + 0.01151455007493496, + 0.02554398588836193, + 0.02132256329059601, + 0.012805846519768238, + 0.0242497306317091, + 0.018371354788541794, + 0.01117319893091917, + 0.01001379732042551, + 0.01062717940658331, + 0.022137446328997612, + 0.007099909707903862, + 0.004420375917106867, + 0.012065554969012737, + 0.005986503325402737, + 0.02057638019323349, + 0.025840729475021362, + 0.01513782236725092, + 0.022008739411830902, + 0.011138339526951313, + 0.008383714593946934, + 0.0155940605327487, + 0.010285448282957077, + 0.006228102371096611, + 0.00498909130692482, + 0.004913415759801865, + 0.00507336063310504, + 0.0068912445567548275, + 0.008488950319588184, + 0.009194339625537395, + 0.00637007225304842, + 0.014143064618110657, + 0.025601990520954132, + 0.011564706452190876, + 0.022993270307779312, + 0.01626262255012989, + 0.010597514919936657, + 0.015849003568291664, + 0.004569869954138994, + 0.01299021765589714, + 0.008847412653267384, + 0.01820945367217064, + 0.01523572951555252, + 0.012256764806807041, + 0.011430527083575726, + 0.02448374591767788, + 0.015385212376713753, + 0.006423591170459986, + 0.01162162609398365, + 0.006380284205079079, + 0.00974526721984148, + 0.0052688829600811005, + 0.003705993527546525, + 0.0057318261824548244, + 0.005312034394592047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.6933586597442627 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_8/index_tensor([2759, 2759])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_8/index_tensor([2759, 2759])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cf9b624232950f15d8f96315550d36b8f06f9db7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_8/index_tensor([2759, 2759])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, side-on tracking shot, held close to the ground, captures the breathtaking moment of a magnificent tiger charging through the humid, dense rainforest. The tiger's vibrant orange-yellow fur is streaked with deep, irregular black stripes, and its muscles bulge, each one tense and rippled with explosive force. It lunges forward with breathtaking force, its body stretched to its limit, its spine like a fully drawn bow, its four powerful paws alternately striking the ground. Each stomp sends dirt and debris from the humus and dead leaves cascading violently backward, creating a fleeting, chaotic trail through the air. Dappled sunlight filters through the layers of treetops, casting rapidly shifting light and shadows on its body and the ground, further heightening the sense of speed. Its extreme speed stretches the massive ferns and tangled vines in the background into a motion-blur of deep greens and browns, dramatically amplifying the visual impact. Its eyes are wide open, its expression fierce and focused, and the whole picture is full of tension, wildness and unstoppable vitality.", + "seed": 21859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_8/index_2759_seed_21859_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03453755446244031, + "vlm_logprob": -0.03453755446244031, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.03453755446244031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07156369090080261 + ], + "opposite_logprob_tokens": [ + -3.446563720703125 + ], + "logprob_reward": -0.07156369090080261, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.07156369090080261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22834454476833344, + 0.265544593334198, + 0.16485357284545898, + 0.09159410744905472, + 0.032353777438402176, + 0.049057792872190475, + 0.023770486935973167, + 0.04057551920413971, + 0.058809831738471985, + 0.045095790177583694 + ], + "credit_map": [ + 0.05316014215350151, + 0.023775504902005196, + 0.03386584296822548, + 0.028109565377235413, + 0.017100738361477852, + 0.02052794024348259, + 0.01950904168188572, + 0.032295770943164825, + 0.01963871717453003, + 0.06889216601848602, + 0.08836141228675842, + 0.01929190196096897, + 0.011948910541832447, + 0.009295488707721233, + 0.014736685901880264, + 0.03337932005524635, + 0.013418660499155521, + 0.0443279892206192, + 0.058744631707668304, + 0.012090569362044334, + 0.010586759075522423, + 0.0050720167346298695, + 0.007929865270853043, + 0.012683087028563023, + 0.027325160801410675, + 0.006960054859519005, + 0.014860355295240879, + 0.00816800631582737, + 0.00740086892619729, + 0.006764511112123728, + 0.009093547239899635, + 0.011021594516932964, + 0.002715888898819685, + 0.00299710500985384, + 0.0034718778915703297, + 0.005470946431159973, + 0.00589568680152297, + 0.0036788478028029203, + 0.004763961769640446, + 0.003359463531523943, + 0.004285464994609356, + 0.0067971935495734215, + 0.006374203599989414, + 0.005482096690684557, + 0.008770477026700974, + 0.004469567444175482, + 0.006406831555068493, + 0.006471953354775906, + 0.002915829187259078, + 0.0033133761025965214, + 0.003002278273925185, + 0.003737888764590025, + 0.001774617936462164, + 0.002798201749101281, + 0.002725935308262706, + 0.0035023584496229887, + 0.004756129812449217, + 0.004297426901757717, + 0.006834850646555424, + 0.004064284265041351, + 0.0042015304788947105, + 0.00713446456938982, + 0.004633279517292976, + 0.004653552547097206, + 0.006562632508575916, + 0.005414193030446768, + 0.009424657560884953, + 0.009367039427161217, + 0.00609411858022213, + 0.006164887920022011, + 0.0095836091786623, + 0.006198696326464415, + 0.004544154740869999, + 0.0062651908956468105, + 0.005084158852696419, + 0.006046414375305176, + 0.005919518880546093, + 0.007046546787023544, + 0.005736448336392641, + 0.004453354515135288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07156369090080261 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904676884412766 + ], + "opposite_logprob_tokens": [ + -2.4640467166900635 + ], + "logprob_reward": -0.08904676884412766, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904676884412766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11892566829919815, + 0.2653360962867737, + 0.12484174966812134, + 0.1314234435558319, + 0.04748796671628952, + 0.05839059501886368, + 0.10785553604364395, + 0.0702773705124855, + 0.04777428135275841, + 0.02768726646900177 + ], + "credit_map": [ + 0.030553283169865608, + 0.016041219234466553, + 0.020626315847039223, + 0.016267696395516396, + 0.00798396673053503, + 0.00884238351136446, + 0.007542399223893881, + 0.011068412102758884, + 0.028656167909502983, + 0.02381775714457035, + 0.05315456539392471, + 0.04722781106829643, + 0.020193450152873993, + 0.017655272036790848, + 0.037306349724531174, + 0.037324707955121994, + 0.009720388799905777, + 0.033666715025901794, + 0.04281323775649071, + 0.010313344188034534, + 0.008457883261144161, + 0.005084057804197073, + 0.005947012919932604, + 0.008839116431772709, + 0.0115733677521348, + 0.012119855731725693, + 0.013855664990842342, + 0.019154759123921394, + 0.023323390632867813, + 0.015562870539724827, + 0.020308976992964745, + 0.015524564310908318, + 0.003986784722656012, + 0.0053744446486234665, + 0.007637839298695326, + 0.00532331969588995, + 0.005978151690214872, + 0.006405085325241089, + 0.005280820652842522, + 0.007501520682126284, + 0.006278343498706818, + 0.004603874403983355, + 0.007494721096009016, + 0.009653457440435886, + 0.005687836557626724, + 0.004624288063496351, + 0.01079829502850771, + 0.009249784052371979, + 0.010141766630113125, + 0.025597121566534042, + 0.0119178406894207, + 0.015179760754108429, + 0.008951767347753048, + 0.011137735098600388, + 0.012630912475287914, + 0.01229863241314888, + 0.006903482601046562, + 0.014521153643727303, + 0.012589586898684502, + 0.006930422503501177, + 0.007066818885505199, + 0.006624183617532253, + 0.007824565283954144, + 0.007817155681550503, + 0.005940845236182213, + 0.0066461567766964436, + 0.006839509587734938, + 0.006621746812015772, + 0.0058450838550925255, + 0.0055530862882733345, + 0.005608857609331608, + 0.004718996584415436, + 0.0026862192898988724, + 0.004519329406321049, + 0.00280126859433949, + 0.004139118827879429, + 0.0023957770317792892, + 0.004286836367100477, + 0.0030854593496769667, + 0.003773256903514266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.08904676884412766 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017254954436793923 + ], + "opposite_logprob_tokens": [ + -6.376725673675537 + ], + "logprob_reward": -0.0017254954436793923, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017254954436793923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1424274444580078, + 0.1892331838607788, + 0.22207899391651154, + 0.05697856843471527, + 0.05429653823375702, + 0.07886195927858353, + 0.08754726499319077, + 0.04246332868933678, + 0.027363331988453865, + 0.09874939918518066 + ], + "credit_map": [ + 0.017882511019706726, + 0.028731653466820717, + 0.020030664280056953, + 0.016857024282217026, + 0.011043179780244827, + 0.010667282156646252, + 0.01215909793972969, + 0.025056030601263046, + 0.020061878487467766, + 0.024140160530805588, + 0.049867257475852966, + 0.021756524220108986, + 0.009626179002225399, + 0.012597768567502499, + 0.016974279657006264, + 0.03420913964509964, + 0.017554614692926407, + 0.06578230112791061, + 0.07816198468208313, + 0.012757478281855583, + 0.01619374006986618, + 0.007151348982006311, + 0.00925171934068203, + 0.015225791372358799, + 0.009572722017765045, + 0.006181030999869108, + 0.006508410442620516, + 0.007620211690664291, + 0.007197321858257055, + 0.0049745552241802216, + 0.007962292991578579, + 0.006962020881474018, + 0.004018766339868307, + 0.0044486126862466335, + 0.006508651655167341, + 0.007804254535585642, + 0.01098241750150919, + 0.006297521758824587, + 0.008385835215449333, + 0.005850478075444698, + 0.007704471237957478, + 0.013771429657936096, + 0.010340040549635887, + 0.010037614963948727, + 0.010402487590909004, + 0.005364708136767149, + 0.011308492161333561, + 0.009932720102369785, + 0.006937857251614332, + 0.009551267139613628, + 0.010687081143260002, + 0.01298193633556366, + 0.007379166316241026, + 0.011789009906351566, + 0.016888557001948357, + 0.011332386173307896, + 0.003988758660852909, + 0.005390457808971405, + 0.013136996887624264, + 0.005574943497776985, + 0.00197348790243268, + 0.0021307310089468956, + 0.00599378626793623, + 0.004274166654795408, + 0.0032061031088232994, + 0.003934433218091726, + 0.003918536473065615, + 0.002976079238578677, + 0.003312828717753291, + 0.003686221083626151, + 0.002908846130594611, + 0.0034202835522592068, + 0.008421588689088821, + 0.007158861495554447, + 0.011445577256381512, + 0.019651150330901146, + 0.013290085829794407, + 0.013373148627579212, + 0.015655232593417168, + 0.00975375808775425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0017254954436793923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736899726092815 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736899726092815, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736899726092815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12602442502975464, + 0.1298350691795349, + 0.27742987871170044, + 0.09129176288843155, + 0.18917100131511688, + 0.024188445881009102, + 0.023514624685049057, + 0.0644238144159317, + 0.025719113647937775, + 0.04840191453695297 + ], + "credit_map": [ + 0.017513958737254143, + 0.017830222845077515, + 0.01896468922495842, + 0.01174414437264204, + 0.009109657257795334, + 0.011734643019735813, + 0.014220540411770344, + 0.024906577542424202, + 0.011086407117545605, + 0.022144418209791183, + 0.03548586741089821, + 0.017054060474038124, + 0.007736830972135067, + 0.005947261583060026, + 0.011557862162590027, + 0.018822351470589638, + 0.022299084812402725, + 0.0770425945520401, + 0.10052748024463654, + 0.018966373056173325, + 0.017894644290208817, + 0.007926366291940212, + 0.012022064067423344, + 0.02075127512216568, + 0.023340284824371338, + 0.011660544201731682, + 0.010511600412428379, + 0.010378144681453705, + 0.011103483848273754, + 0.006344615016132593, + 0.008284821175038815, + 0.009668268263339996, + 0.017168080434203148, + 0.015793202444911003, + 0.031197017058730125, + 0.020665431395173073, + 0.01789824292063713, + 0.03884236514568329, + 0.019069425761699677, + 0.02853723056614399, + 0.002388104796409607, + 0.003229954745620489, + 0.0029590565245598555, + 0.0030014761723577976, + 0.0030286868568509817, + 0.0034336012322455645, + 0.003145761787891388, + 0.0030018039979040623, + 0.0031478735618293285, + 0.0038590719923377037, + 0.0027450837660580873, + 0.0030513072852045298, + 0.002088345354422927, + 0.0025085024535655975, + 0.002983430866152048, + 0.003131007542833686, + 0.006384085398167372, + 0.009344564750790596, + 0.0241410993039608, + 0.006538639776408672, + 0.002086477819830179, + 0.00261558941565454, + 0.008737643249332905, + 0.004575713537633419, + 0.0028550326824188232, + 0.0030971637461334467, + 0.003186651272699237, + 0.003436755621805787, + 0.0032913086470216513, + 0.00301739270798862, + 0.003919912502169609, + 0.0029148973990231752, + 0.007117467001080513, + 0.0051177083514630795, + 0.005065243225544691, + 0.006847826298326254, + 0.005029064137488604, + 0.008768071420490742, + 0.005142903421074152, + 0.005313629284501076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.006736899726092815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036149173974990845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036149173974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20820198953151703, + 0.27727413177490234, + 0.14640024304389954, + 0.05178595706820488, + 0.059136874973773956, + 0.042114924639463425, + 0.05090855062007904, + 0.07691976428031921, + 0.045810528099536896, + 0.04144709184765816 + ], + "credit_map": [ + 0.021872950717806816, + 0.017849642783403397, + 0.030667897313833237, + 0.016844386234879494, + 0.015333245508372784, + 0.025087757036089897, + 0.023204650729894638, + 0.05734143778681755, + 0.020780542865395546, + 0.06698554009199142, + 0.09010907262563705, + 0.017137162387371063, + 0.013663342222571373, + 0.008821153081953526, + 0.02117009088397026, + 0.038607195019721985, + 0.012601112946867943, + 0.039737049490213394, + 0.04783545061945915, + 0.010342273861169815, + 0.010161926969885826, + 0.0063791340216994286, + 0.009179195389151573, + 0.01016409695148468, + 0.007083590142428875, + 0.004222011659294367, + 0.0069386037066578865, + 0.007410033605992794, + 0.005783667787909508, + 0.004656617529690266, + 0.008266529068350792, + 0.007424901705235243, + 0.006694613490253687, + 0.00507008982822299, + 0.007192108314484358, + 0.012419016100466251, + 0.007955707609653473, + 0.0042587630450725555, + 0.008345477283000946, + 0.007201092317700386, + 0.0050584133714437485, + 0.0053663174621760845, + 0.005791873205453157, + 0.0056878384202718735, + 0.006156824994832277, + 0.003318110015243292, + 0.0057498677633702755, + 0.004985674750059843, + 0.004937881138175726, + 0.0061112199909985065, + 0.00640297494828701, + 0.007422213442623615, + 0.0035952304024249315, + 0.005899328272789717, + 0.00948750413954258, + 0.007052188273519278, + 0.00626622699201107, + 0.009644173085689545, + 0.00760960578918457, + 0.009566361084580421, + 0.008263603784143925, + 0.013931330293416977, + 0.011037490330636501, + 0.010600967332720757, + 0.0050550298765301704, + 0.00520092761144042, + 0.006609877105802298, + 0.005659192334860563, + 0.006445943843573332, + 0.004806046839803457, + 0.005583988036960363, + 0.006449516396969557, + 0.004286102019250393, + 0.005797887686640024, + 0.0045815398916602135, + 0.0055691516026854515, + 0.004399473313242197, + 0.006381021812558174, + 0.00497512286528945, + 0.005456788931041956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0036149173974990845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, side-on tracking shot, held close to the ground, captures the breathtaking moment of a magnificent tiger charging through the humid, dense rainforest. The tiger's vibrant orange-yellow fur is streaked with deep, irregular black stripes, and its muscles bulge, each one tense and rippled with explosive force. It lunges forward with breathtaking force, its body stretched to its limit, its spine like a fully drawn bow, its four powerful paws alternately striking the ground. Each stomp sends dirt and debris from the humus and dead leaves cascading violently backward, creating a fleeting, chaotic trail through the air. Dappled sunlight filters through the layers of treetops, casting rapidly shifting light and shadows on its body and the ground, further heightening the sense of speed. Its extreme speed stretches the massive ferns and tangled vines in the background into a motion-blur of deep greens and browns, dramatically amplifying the visual impact. Its eyes are wide open, its expression fierce and focused, and the whole picture is full of tension, wildness and unstoppable vitality.", + "seed": 21860, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_8/index_2759_seed_21860_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03261776843573898, + "vlm_logprob": -0.03261776843573898, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.03261776843573898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11192241311073303 + ], + "opposite_logprob_tokens": [ + -2.4869225025177 + ], + "logprob_reward": -0.11192241311073303, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11192241311073303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062901072204113, + 0.034472789615392685, + 0.050461407750844955, + 0.03268790990114212, + 0.11808802932500839, + 0.06965065747499466, + 0.10389848798513412, + 0.39741483330726624, + 0.10489727556705475, + 0.025527531281113625 + ], + "credit_map": [ + 0.007900943048298359, + 0.006859866436570883, + 0.012809542007744312, + 0.007293245289474726, + 0.007597479037940502, + 0.007444854360073805, + 0.005042621865868568, + 0.007952519692480564, + 0.003586200764402747, + 0.004481385927647352, + 0.0037743765860795975, + 0.003406414994969964, + 0.005958057940006256, + 0.00406019389629364, + 0.005058537237346172, + 0.004147623665630817, + 0.0068513816222548485, + 0.006574776954948902, + 0.005555081646889448, + 0.004931661766022444, + 0.0061027174815535545, + 0.008297876454889774, + 0.0059284246526658535, + 0.006219492293894291, + 0.0033484601881355047, + 0.0022609829902648926, + 0.004965214990079403, + 0.0024492882657796144, + 0.002127809915691614, + 0.0055491989478468895, + 0.003560715587809682, + 0.008426240645349026, + 0.006341872736811638, + 0.007130286656320095, + 0.010400363244116306, + 0.0388372465968132, + 0.009448962286114693, + 0.010894468985497952, + 0.010825911536812782, + 0.0242089182138443, + 0.007185474503785372, + 0.005163860972970724, + 0.009957803413271904, + 0.008937516249716282, + 0.007968347519636154, + 0.011684012599289417, + 0.006180324591696262, + 0.012573318555951118, + 0.011556599289178848, + 0.019620342180132866, + 0.009633149951696396, + 0.013266244903206825, + 0.014133602380752563, + 0.007555536925792694, + 0.01656329073011875, + 0.011569716036319733, + 0.011265533976256847, + 0.017514703795313835, + 0.052038390189409256, + 0.017041834071278572, + 0.09583697468042374, + 0.02349478006362915, + 0.15371452271938324, + 0.02650812268257141, + 0.010261023417115211, + 0.015114838257431984, + 0.012834396213293076, + 0.013891112059354782, + 0.013974864967167377, + 0.01470077782869339, + 0.013808563351631165, + 0.01031169481575489, + 0.002376341260969639, + 0.0027635437436401844, + 0.0037085942458361387, + 0.003497926751151681, + 0.0028760754503309727, + 0.0030423225834965706, + 0.003205845830962062, + 0.004056880716234446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.11192241311073303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370113715529442 + ], + "opposite_logprob_tokens": [ + -3.651370048522949 + ], + "logprob_reward": -0.026370113715529442, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370113715529442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03126997500658035, + 0.052343808114528656, + 0.08473173528909683, + 0.03665312007069588, + 0.13673017919063568, + 0.10739760100841522, + 0.39954352378845215, + 0.08307284861803055, + 0.03445802628993988, + 0.03379916027188301 + ], + "credit_map": [ + 0.0034962445497512817, + 0.004037679173052311, + 0.005215817596763372, + 0.0032360353507101536, + 0.003586465958505869, + 0.004211007617413998, + 0.0033851247280836105, + 0.0041016023606061935, + 0.005179741885513067, + 0.007651929743587971, + 0.007308894768357277, + 0.007425408810377121, + 0.007139042019844055, + 0.005614087916910648, + 0.006937240716069937, + 0.005087465513497591, + 0.01010939385741949, + 0.013220292516052723, + 0.009115437045693398, + 0.008569964207708836, + 0.009221893735229969, + 0.0134038757532835, + 0.010320390574634075, + 0.010770496912300587, + 0.005725872237235308, + 0.0034989300183951855, + 0.0038279315922409296, + 0.0027245287783443928, + 0.002680747536942363, + 0.005961495451629162, + 0.003705862443894148, + 0.008527754805982113, + 0.007323501165956259, + 0.0076162246987223625, + 0.014544064179062843, + 0.0454062819480896, + 0.010638972744345665, + 0.011510047130286694, + 0.011533274315297604, + 0.028157832100987434, + 0.01694004237651825, + 0.010643243789672852, + 0.016663752496242523, + 0.016761241480708122, + 0.008810370229184628, + 0.015044944360852242, + 0.009011120535433292, + 0.01352289691567421, + 0.028878984972834587, + 0.03342510759830475, + 0.05833568051457405, + 0.0417950339615345, + 0.036403838545084, + 0.0438123382627964, + 0.08893217146396637, + 0.06796038150787354, + 0.003143982496112585, + 0.0035506878048181534, + 0.013218049891293049, + 0.004087014589458704, + 0.017262395471334457, + 0.00395832909271121, + 0.03235848248004913, + 0.0054939198307693005, + 0.0025102118961513042, + 0.005383337382227182, + 0.004088343121111393, + 0.004158731549978256, + 0.007380860857665539, + 0.005496779922395945, + 0.003077414119616151, + 0.002362349070608616, + 0.0029779227916151285, + 0.004487046971917152, + 0.0031585393007844687, + 0.004306148737668991, + 0.004769579507410526, + 0.004896049853414297, + 0.003698813496157527, + 0.005505061708390713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.026370113715529442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017256144201382995 + ], + "opposite_logprob_tokens": [ + -6.376725673675537 + ], + "logprob_reward": -0.0017256144201382995, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017256144201382995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03531370684504509, + 0.05022675171494484, + 0.04698842391371727, + 0.030875105410814285, + 0.24923375248908997, + 0.03354228287935257, + 0.07218629866838455, + 0.30997830629348755, + 0.14639733731746674, + 0.025258008390665054 + ], + "credit_map": [ + 0.0036505425814539194, + 0.003914766013622284, + 0.004593413323163986, + 0.0038951800670474768, + 0.005031058099120855, + 0.006645732559263706, + 0.0030536409467458725, + 0.004529374185949564, + 0.004892164375633001, + 0.0065844119526445866, + 0.005487298127263784, + 0.006370170041918755, + 0.008570787496864796, + 0.0050698393024504185, + 0.008593760430812836, + 0.004658321384340525, + 0.006315474398434162, + 0.007194804958999157, + 0.0052711740136146545, + 0.005149594973772764, + 0.004072561860084534, + 0.008372793905436993, + 0.004484051372855902, + 0.006127969827502966, + 0.006844097748398781, + 0.003888458712026477, + 0.0037324188742786646, + 0.002344219945371151, + 0.0024362190160900354, + 0.004611256066709757, + 0.002325885696336627, + 0.004692549351602793, + 0.01317811943590641, + 0.01148656103760004, + 0.018689371645450592, + 0.09317582845687866, + 0.01516581792384386, + 0.0158787090331316, + 0.017767461016774178, + 0.06389188021421432, + 0.004641368985176086, + 0.004004471004009247, + 0.004861111752688885, + 0.004903916269540787, + 0.003673224477097392, + 0.0037593538872897625, + 0.0028919638134539127, + 0.004806873854249716, + 0.00470814248546958, + 0.00803344789892435, + 0.01109276618808508, + 0.008238528855144978, + 0.006250734906643629, + 0.0060965875163674355, + 0.018169226124882698, + 0.009596867486834526, + 0.009621593169867992, + 0.012563462369143963, + 0.04393155872821808, + 0.01318374928086996, + 0.07273314148187637, + 0.016148796305060387, + 0.12247375398874283, + 0.01932222954928875, + 0.013742252252995968, + 0.023652013391256332, + 0.017272574827075005, + 0.018423937261104584, + 0.0231001116335392, + 0.018928933888673782, + 0.018520483747124672, + 0.012757042422890663, + 0.002505391836166382, + 0.0022182923275977373, + 0.0029658416751772165, + 0.0050708153285086155, + 0.0028004690539091825, + 0.0036108496133238077, + 0.002835480961948633, + 0.0032508689910173416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.0017256144201382995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581373944878578 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.020581373944878578, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581373944878578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1255740523338318, + 0.2058013528585434, + 0.09991463273763657, + 0.052414845675230026, + 0.13349145650863647, + 0.0895850732922554, + 0.10738755017518997, + 0.1355694830417633, + 0.02165207453072071, + 0.028609417378902435 + ], + "credit_map": [ + 0.016858164221048355, + 0.008916259743273258, + 0.021055098623037338, + 0.015634967014193535, + 0.01164756529033184, + 0.024056360125541687, + 0.011399814859032631, + 0.016005827113986015, + 0.022322308272123337, + 0.013459807261824608, + 0.01763225346803665, + 0.01331100519746542, + 0.042879268527030945, + 0.02876695618033409, + 0.03871545195579529, + 0.028714310377836227, + 0.012390532530844212, + 0.01214110478758812, + 0.009068244136869907, + 0.009631693363189697, + 0.018447821959853172, + 0.015603584237396717, + 0.012656616978347301, + 0.009975042194128036, + 0.00621198071166873, + 0.0055681257508695126, + 0.00656157499179244, + 0.0045813205651938915, + 0.005030414089560509, + 0.007797908503562212, + 0.005886543076485395, + 0.010776979848742485, + 0.005252838134765625, + 0.00576622411608696, + 0.008065973408520222, + 0.051973339170217514, + 0.006576562765985727, + 0.00928783044219017, + 0.009492943994700909, + 0.03707575052976608, + 0.012239189818501472, + 0.008481970988214016, + 0.014193116687238216, + 0.012379402294754982, + 0.009486928582191467, + 0.010716698132455349, + 0.00814391765743494, + 0.01394384540617466, + 0.007677180226892233, + 0.00845358893275261, + 0.014409014955163002, + 0.007778668310493231, + 0.011329524219036102, + 0.01601804420351982, + 0.02443060465157032, + 0.01729092188179493, + 0.003635744797065854, + 0.006806264631450176, + 0.01603911630809307, + 0.006203331518918276, + 0.03257700055837631, + 0.010602995753288269, + 0.04931310936808586, + 0.010391926392912865, + 0.002023412613198161, + 0.003344625933095813, + 0.0028617107309401035, + 0.0028654569759964943, + 0.0028654441703110933, + 0.0031526971142739058, + 0.0028114577289670706, + 0.0017272706609219313, + 0.002491554245352745, + 0.002786827040836215, + 0.0033237403258681297, + 0.006973341107368469, + 0.0021617503371089697, + 0.003973305691033602, + 0.003706003073602915, + 0.0031928953249007463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.020581373944878578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489326987415552 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489326987415552, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489326987415552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10154469311237335, + 0.0595100112259388, + 0.059465549886226654, + 0.016890686005353928, + 0.36300331354141235, + 0.039375972002744675, + 0.06837505847215652, + 0.18782611191272736, + 0.05349880829453468, + 0.05050976946949959 + ], + "credit_map": [ + 0.014008735306560993, + 0.008360608480870724, + 0.019849982112646103, + 0.011817087419331074, + 0.009960154071450233, + 0.01320677064359188, + 0.007476815953850746, + 0.016864551231265068, + 0.005363405216485262, + 0.005278684198856354, + 0.004759466741234064, + 0.005461397580802441, + 0.006821255665272474, + 0.0077608427964150906, + 0.010783565230667591, + 0.013281400315463543, + 0.0054648867808282375, + 0.008172680623829365, + 0.0059912316501140594, + 0.004931769799441099, + 0.008371810428798199, + 0.011822296306490898, + 0.007053667679429054, + 0.007657214533537626, + 0.0033497042022645473, + 0.0018163234926760197, + 0.0019774427637457848, + 0.001411108998581767, + 0.0013607562286779284, + 0.00248010759241879, + 0.0016165184788405895, + 0.0028787259943783283, + 0.010952560231089592, + 0.015543321147561073, + 0.018618501722812653, + 0.1361832320690155, + 0.01795061118900776, + 0.025003112852573395, + 0.02695254236459732, + 0.111799456179142, + 0.00311030144803226, + 0.0035462635569274426, + 0.005439198575913906, + 0.0034073535352945328, + 0.006260339170694351, + 0.006610469426959753, + 0.003937033005058765, + 0.007065016310662031, + 0.006159032229334116, + 0.009224732406437397, + 0.00895325280725956, + 0.007362727541476488, + 0.008575181476771832, + 0.00742163835093379, + 0.013225764967501163, + 0.007452739868313074, + 0.006050613708794117, + 0.010102551430463791, + 0.027554143220186234, + 0.008584407158195972, + 0.042829133570194244, + 0.007695161271840334, + 0.0718679279088974, + 0.013142184354364872, + 0.004107282962650061, + 0.007102977484464645, + 0.007003999315202236, + 0.006968583911657333, + 0.00991474837064743, + 0.007270848844200373, + 0.005988419987261295, + 0.005141953006386757, + 0.0036529891658574343, + 0.004594079684466124, + 0.006444850005209446, + 0.012389121577143669, + 0.005360520910471678, + 0.007913309149444103, + 0.00524562643840909, + 0.004909276030957699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.002489326987415552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, side-on tracking shot, held close to the ground, captures the breathtaking moment of a magnificent tiger charging through the humid, dense rainforest. The tiger's vibrant orange-yellow fur is streaked with deep, irregular black stripes, and its muscles bulge, each one tense and rippled with explosive force. It lunges forward with breathtaking force, its body stretched to its limit, its spine like a fully drawn bow, its four powerful paws alternately striking the ground. Each stomp sends dirt and debris from the humus and dead leaves cascading violently backward, creating a fleeting, chaotic trail through the air. Dappled sunlight filters through the layers of treetops, casting rapidly shifting light and shadows on its body and the ground, further heightening the sense of speed. Its extreme speed stretches the massive ferns and tangled vines in the background into a motion-blur of deep greens and browns, dramatically amplifying the visual impact. Its eyes are wide open, its expression fierce and focused, and the whole picture is full of tension, wildness and unstoppable vitality.", + "seed": 21861, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_8/index_2759_seed_21861_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06709712452720851, + "vlm_logprob": -0.06709712452720851, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.06709712452720851, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13306163251399994 + ], + "opposite_logprob_tokens": [ + -2.383061647415161 + ], + "logprob_reward": -0.13306163251399994, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.13306163251399994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3746144473552704, + 0.035109009593725204, + 0.03591926023364067, + 0.18371781706809998, + 0.03736412525177002, + 0.025820937007665634, + 0.07897850126028061, + 0.07185186445713043, + 0.05777079612016678, + 0.0988532230257988 + ], + "credit_map": [ + 0.0295904278755188, + 0.026437673717737198, + 0.021817345172166824, + 0.02688632532954216, + 0.02429838292300701, + 0.053660135716199875, + 0.10145870596170425, + 0.09046546369791031, + 0.003498762147501111, + 0.004240140784531832, + 0.004805356729775667, + 0.004141184035688639, + 0.004045379813760519, + 0.005705143325030804, + 0.0037644708063453436, + 0.004908572882413864, + 0.00317860278300941, + 0.0034991439897567034, + 0.0073544965125620365, + 0.0031587493140250444, + 0.010661263018846512, + 0.002914872020483017, + 0.002068880945444107, + 0.003083252115175128, + 0.007760681677609682, + 0.00828575063496828, + 0.034839700907468796, + 0.058431465178728104, + 0.005321841686964035, + 0.015033897012472153, + 0.009921461343765259, + 0.04412301257252693, + 0.003221262712031603, + 0.00383719801902771, + 0.0035340366885066032, + 0.0038698839489370584, + 0.0033242092467844486, + 0.004275368992239237, + 0.009980671107769012, + 0.0053214929066598415, + 0.002344745211303234, + 0.003038265975192189, + 0.003818206023424864, + 0.003767148358747363, + 0.002260028850287199, + 0.0032563323620706797, + 0.0026187601033598185, + 0.004717448726296425, + 0.009910671040415764, + 0.009378262795507908, + 0.009593930095434189, + 0.006960736587643623, + 0.006222348194569349, + 0.007733732461929321, + 0.01960662007331848, + 0.009572199545800686, + 0.008737551979720592, + 0.006277770269662142, + 0.003009573556482792, + 0.03187018260359764, + 0.0026030847802758217, + 0.004653273615986109, + 0.003462121356278658, + 0.01123830582946539, + 0.009375291876494884, + 0.004718402866274118, + 0.01326336432248354, + 0.007631539832800627, + 0.007313702255487442, + 0.006871345452964306, + 0.004183212295174599, + 0.004413933493196964, + 0.008244474418461323, + 0.00869505200535059, + 0.012631222605705261, + 0.0178313497453928, + 0.0073295519687235355, + 0.008020535111427307, + 0.01783854141831398, + 0.018262499943375587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.13306163251399994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038861870765686 + ], + "opposite_logprob_tokens": [ + -1.9103885889053345 + ], + "logprob_reward": -0.16038861870765686, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038861870765686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27946165204048157, + 0.06070054695010185, + 0.03892797604203224, + 0.08888953924179077, + 0.043369926512241364, + 0.09078102558851242, + 0.05759597569704056, + 0.19771923124790192, + 0.0774901881814003, + 0.06506393849849701 + ], + "credit_map": [ + 0.021048545837402344, + 0.025602804496884346, + 0.02371303178369999, + 0.028431780636310577, + 0.03195175155997276, + 0.034649647772312164, + 0.04789163917303085, + 0.06617244333028793, + 0.005917961709201336, + 0.009867919608950615, + 0.008580160327255726, + 0.0059655215591192245, + 0.008558545261621475, + 0.00818662904202938, + 0.008366413414478302, + 0.005257399287074804, + 0.004211365710943937, + 0.00597592256963253, + 0.006172527559101582, + 0.004507923033088446, + 0.008967505767941475, + 0.002896103309467435, + 0.0034954561851918697, + 0.0027011765632778406, + 0.0035712525714188814, + 0.00430330540984869, + 0.011487270705401897, + 0.014628663659095764, + 0.002936805598437786, + 0.023092100396752357, + 0.0041496935300529, + 0.0247204527258873, + 0.008503742516040802, + 0.004821535665541887, + 0.006035050377249718, + 0.0041426559910178185, + 0.003041746560484171, + 0.005159917287528515, + 0.006234270054847002, + 0.005431012250483036, + 0.008592459373176098, + 0.00967859011143446, + 0.011536696925759315, + 0.013165649026632309, + 0.010360674932599068, + 0.011720418930053711, + 0.013701698742806911, + 0.012024848721921444, + 0.0037709991447627544, + 0.005173140671104193, + 0.005913065746426582, + 0.004744512028992176, + 0.006329663097858429, + 0.007453990634530783, + 0.013802972622215748, + 0.010407635010778904, + 0.03984599560499191, + 0.01015231478959322, + 0.008261573500931263, + 0.08599461615085602, + 0.013501054607331753, + 0.015118722803890705, + 0.010068945586681366, + 0.014776019379496574, + 0.014495684765279293, + 0.006401316728442907, + 0.017156638205051422, + 0.006526486482471228, + 0.01376289315521717, + 0.00774949649348855, + 0.005076798610389233, + 0.0063208844512701035, + 0.007011646870523691, + 0.006943290587514639, + 0.008879575878381729, + 0.009270436130464077, + 0.007139946334064007, + 0.007766122464090586, + 0.009000450372695923, + 0.009052480570971966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.16038861870765686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002212697174400091 + ], + "opposite_logprob_tokens": [ + -6.1272125244140625 + ], + "logprob_reward": -0.002212697174400091, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002212697174400091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2765209376811981, + 0.03732311353087425, + 0.08295022696256638, + 0.2455900013446808, + 0.02999817579984665, + 0.025268511846661568, + 0.09898826479911804, + 0.12018487602472305, + 0.02684597298502922, + 0.056329913437366486 + ], + "credit_map": [ + 0.024272460490465164, + 0.01785527728497982, + 0.01666334457695484, + 0.017752472311258316, + 0.018649019300937653, + 0.040672753006219864, + 0.06505182385444641, + 0.07560379058122635, + 0.00396308396011591, + 0.004878484643995762, + 0.006217550951987505, + 0.004653542302548885, + 0.003779531689360738, + 0.004440159071236849, + 0.004325226414948702, + 0.00506553566083312, + 0.004951228853315115, + 0.004118278156965971, + 0.022851496934890747, + 0.004298544488847256, + 0.02974857948720455, + 0.007110226433724165, + 0.0047615570947527885, + 0.0051103136502206326, + 0.010966531932353973, + 0.011669541709125042, + 0.04877013340592384, + 0.08903924375772476, + 0.006482647266238928, + 0.020232636481523514, + 0.01383958850055933, + 0.04458966851234436, + 0.0052993991412222385, + 0.004396544303745031, + 0.0050594499334692955, + 0.002850596560165286, + 0.0035081685055047274, + 0.0028793420642614365, + 0.002876017242670059, + 0.0031286580488085747, + 0.0029062097892165184, + 0.002236875705420971, + 0.0050871046259999275, + 0.003414763603359461, + 0.002492128172889352, + 0.0035145243164151907, + 0.003136090701445937, + 0.0024808140005916357, + 0.00887485221028328, + 0.009720559231936932, + 0.008622991852462292, + 0.0060920617543160915, + 0.009087573736906052, + 0.007039854768663645, + 0.030580228194594383, + 0.018970148637890816, + 0.021186647936701775, + 0.006821633316576481, + 0.004561314824968576, + 0.05278458446264267, + 0.008182593621313572, + 0.008505589328706264, + 0.0059458366595208645, + 0.012196682393550873, + 0.0033306682016700506, + 0.0034556242171674967, + 0.003987294156104326, + 0.0029316660948097706, + 0.003950687125325203, + 0.0032495895866304636, + 0.003463307162746787, + 0.0024771359749138355, + 0.005974869709461927, + 0.006251694168895483, + 0.006579448468983173, + 0.007137128617614508, + 0.008099209517240524, + 0.00698627857491374, + 0.006212400272488594, + 0.009088883176445961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.002212697174400091 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810705617070198 + ], + "opposite_logprob_tokens": [ + -3.28810715675354 + ], + "logprob_reward": -0.03810705617070198, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810705617070198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40974095463752747, + 0.09230754524469376, + 0.1130712628364563, + 0.09931804984807968, + 0.06704846024513245, + 0.04761229455471039, + 0.0561240054666996, + 0.03412685915827751, + 0.051114123314619064, + 0.029536547139286995 + ], + "credit_map": [ + 0.03837687149643898, + 0.01891305483877659, + 0.03228228911757469, + 0.02866181917488575, + 0.02823043242096901, + 0.060052476823329926, + 0.09103985130786896, + 0.11218410730361938, + 0.01098804920911789, + 0.010105409659445286, + 0.012068943120539188, + 0.00815349817276001, + 0.009850912727415562, + 0.01271999254822731, + 0.01268037036061287, + 0.01574036292731762, + 0.006810235790908337, + 0.0053605264984071255, + 0.027897123247385025, + 0.00587488105520606, + 0.033958110958337784, + 0.013189558871090412, + 0.009114078246057034, + 0.010866732336580753, + 0.005057673901319504, + 0.005161956883966923, + 0.01964191533625126, + 0.03579684719443321, + 0.0027362932451069355, + 0.011132883839309216, + 0.0055453479290008545, + 0.014245113357901573, + 0.0036845363210886717, + 0.004052333068102598, + 0.0047488161362707615, + 0.00642600329592824, + 0.006116057746112347, + 0.010346898809075356, + 0.015009036287665367, + 0.016664765775203705, + 0.003827651496976614, + 0.005652023945003748, + 0.006876170169562101, + 0.005010918248444796, + 0.004510699305683374, + 0.006669397931545973, + 0.004679874051362276, + 0.010385549627244473, + 0.004885935224592686, + 0.00567980483174324, + 0.005785046145319939, + 0.004537772387266159, + 0.004883230663836002, + 0.005709097255021334, + 0.014685011468827724, + 0.009958096779882908, + 0.002440136391669512, + 0.003134737256914377, + 0.0020883074030280113, + 0.010436953976750374, + 0.0021388079039752483, + 0.003090228419750929, + 0.0015318405348807573, + 0.009265839122235775, + 0.007733547128736973, + 0.005948490463197231, + 0.011032572947442532, + 0.0061136228032410145, + 0.006743776611983776, + 0.00601772079244256, + 0.003600944532081485, + 0.00392344081774354, + 0.0024050751235336065, + 0.0022645657882094383, + 0.002875057514756918, + 0.00398329459130764, + 0.0025050947442650795, + 0.002674476709216833, + 0.0071706585586071014, + 0.005658319219946861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03810705617070198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017156180692836642 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.0017156180692836642, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017156180692836642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1987425982952118, + 0.033040888607501984, + 0.10665156692266464, + 0.18910977244377136, + 0.08978260308504105, + 0.05521291121840477, + 0.07752592116594315, + 0.11533522605895996, + 0.0920838713645935, + 0.042514607310295105 + ], + "credit_map": [ + 0.01596355251967907, + 0.012520085088908672, + 0.013604541309177876, + 0.013994710519909859, + 0.017631515860557556, + 0.02765379101037979, + 0.04786042124032974, + 0.04951399192214012, + 0.0025858948938548565, + 0.004507796373218298, + 0.004581360146403313, + 0.0037964668590575457, + 0.0031873818952590227, + 0.004363166633993387, + 0.004169850144535303, + 0.005848974455147982, + 0.006083452608436346, + 0.003344135358929634, + 0.031771305948495865, + 0.005154345650225878, + 0.04262174293398857, + 0.006647012662142515, + 0.005455370061099529, + 0.005574219860136509, + 0.005960531998425722, + 0.00558325182646513, + 0.021279608830809593, + 0.052820343524217606, + 0.004854267928749323, + 0.06580831110477448, + 0.009043430909514427, + 0.023760030046105385, + 0.005676421336829662, + 0.006630612071603537, + 0.006927430629730225, + 0.006898127496242523, + 0.008155351504683495, + 0.013985447585582733, + 0.01906489208340645, + 0.022444332018494606, + 0.005137874744832516, + 0.005069277714937925, + 0.010491181164979935, + 0.005038324277848005, + 0.00701874028891325, + 0.007203863933682442, + 0.008163358084857464, + 0.007090297993272543, + 0.009414744563400745, + 0.009546652436256409, + 0.012058207765221596, + 0.007188900373876095, + 0.006721549201756716, + 0.006353873759508133, + 0.016633400693535805, + 0.009608594700694084, + 0.02828318066895008, + 0.006811986211687326, + 0.005349594168365002, + 0.03217248246073723, + 0.011902223341166973, + 0.011367185041308403, + 0.006939292885363102, + 0.012509288266301155, + 0.01311538927257061, + 0.011569857597351074, + 0.017316823825240135, + 0.010363569483160973, + 0.011636526323854923, + 0.013066784478724003, + 0.006990272551774979, + 0.008024659939110279, + 0.0047568087466061115, + 0.0048265536315739155, + 0.005171482916921377, + 0.005194148514419794, + 0.00465501518920064, + 0.005542861297726631, + 0.006396292243152857, + 0.005971449427306652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0017156180692836642 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, side-on tracking shot, held close to the ground, captures the breathtaking moment of a magnificent tiger charging through the humid, dense rainforest. The tiger's vibrant orange-yellow fur is streaked with deep, irregular black stripes, and its muscles bulge, each one tense and rippled with explosive force. It lunges forward with breathtaking force, its body stretched to its limit, its spine like a fully drawn bow, its four powerful paws alternately striking the ground. Each stomp sends dirt and debris from the humus and dead leaves cascading violently backward, creating a fleeting, chaotic trail through the air. Dappled sunlight filters through the layers of treetops, casting rapidly shifting light and shadows on its body and the ground, further heightening the sense of speed. Its extreme speed stretches the massive ferns and tangled vines in the background into a motion-blur of deep greens and browns, dramatically amplifying the visual impact. Its eyes are wide open, its expression fierce and focused, and the whole picture is full of tension, wildness and unstoppable vitality.", + "seed": 21862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_8/index_2759_seed_21862_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03736917099449784, + "vlm_logprob": -0.03736917099449784, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.03736917099449784, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1384446769952774 + ], + "opposite_logprob_tokens": [ + -2.263444662094116 + ], + "logprob_reward": -0.1384446769952774, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1384446769952774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10598769783973694, + 0.06993336230516434, + 0.03584708273410797, + 0.05306432768702507, + 0.04799044877290726, + 0.02557501010596752, + 0.28895771503448486, + 0.04095299541950226, + 0.29324451088905334, + 0.03844683617353439 + ], + "credit_map": [ + 0.012427465058863163, + 0.012759984470903873, + 0.013006164692342281, + 0.013123324140906334, + 0.012840096838772297, + 0.012902689166367054, + 0.01474541611969471, + 0.014182564802467823, + 0.0057956380769610405, + 0.012228809297084808, + 0.012978115119040012, + 0.007535369601100683, + 0.007356287445873022, + 0.011408455669879913, + 0.004206758923828602, + 0.008423932828009129, + 0.004047017078846693, + 0.004205864854156971, + 0.0058804904110729694, + 0.004624963738024235, + 0.003790217451751232, + 0.005591489374637604, + 0.002362902043387294, + 0.005344140809029341, + 0.0060967677272856236, + 0.007729995530098677, + 0.008588731288909912, + 0.006311910692602396, + 0.0074271876364946365, + 0.007833353243768215, + 0.004624202381819487, + 0.004452183376997709, + 0.004834888968616724, + 0.0038350666873157024, + 0.011171409860253334, + 0.006362220738083124, + 0.004983426071703434, + 0.006840707268565893, + 0.004853647667914629, + 0.005109086167067289, + 0.002216707682237029, + 0.0028317102696746588, + 0.0036730989813804626, + 0.0025124431122094393, + 0.0031621430534869432, + 0.005052502267062664, + 0.003180814441293478, + 0.002945592626929283, + 0.011166450567543507, + 0.01589317061007023, + 0.01696779951453209, + 0.012523171491920948, + 0.022894738242030144, + 0.09333328157663345, + 0.0245493333786726, + 0.09162979573011398, + 0.003766220761463046, + 0.004932759329676628, + 0.00629976624622941, + 0.0036590618547052145, + 0.004102929029613733, + 0.007563482038676739, + 0.00391088193282485, + 0.0067178960889577866, + 0.05921684205532074, + 0.01641019992530346, + 0.020355254411697388, + 0.007096519693732262, + 0.010071001015603542, + 0.045264795422554016, + 0.012543327175080776, + 0.1222866103053093, + 0.0024544987827539444, + 0.002331957221031189, + 0.00897166971117258, + 0.002881237305700779, + 0.004243303090333939, + 0.0032448444981127977, + 0.00320745212957263, + 0.011111876927316189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.1384446769952774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812266141176224 + ], + "opposite_logprob_tokens": [ + -3.2881226539611816 + ], + "logprob_reward": -0.03812266141176224, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812266141176224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07367837429046631, + 0.05095120146870613, + 0.17139938473701477, + 0.05582241341471672, + 0.027300888672471046, + 0.05264471471309662, + 0.3475642502307892, + 0.03641526401042938, + 0.12286026030778885, + 0.061363205313682556 + ], + "credit_map": [ + 0.0064347488805651665, + 0.010327913798391819, + 0.01194678619503975, + 0.009469632059335709, + 0.00706123374402523, + 0.0076622795313596725, + 0.012752080336213112, + 0.008023707196116447, + 0.006048673763871193, + 0.007384409662336111, + 0.00695373211055994, + 0.007954678498208523, + 0.005352802574634552, + 0.0053265755996108055, + 0.006084697786718607, + 0.005845639389008284, + 0.017413238063454628, + 0.017721792683005333, + 0.02940238080918789, + 0.017396170645952225, + 0.021104367449879646, + 0.026157140731811523, + 0.019300319254398346, + 0.022903963923454285, + 0.006298045162111521, + 0.005481743719428778, + 0.006435101851820946, + 0.0052825529128313065, + 0.004543245770037174, + 0.007767938077449799, + 0.014296410605311394, + 0.005717382300645113, + 0.003477261168882251, + 0.0034890093374997377, + 0.004062386229634285, + 0.002925323089584708, + 0.0033593180123716593, + 0.003498943755403161, + 0.002463096287101507, + 0.004025553818792105, + 0.005828715395182371, + 0.007716030813753605, + 0.005736013408750296, + 0.008350440301001072, + 0.008256170898675919, + 0.006776218302547932, + 0.005076482426375151, + 0.004904651548713446, + 0.013324198313057423, + 0.017375506460666656, + 0.017192063853144646, + 0.013691514730453491, + 0.022290989756584167, + 0.11547191441059113, + 0.035308800637722015, + 0.11290928721427917, + 0.0044459691271185875, + 0.003221530932933092, + 0.0050256806425750256, + 0.0038190227933228016, + 0.004539722576737404, + 0.0038502353709191084, + 0.006783206481486559, + 0.0047299000434577465, + 0.02423521876335144, + 0.006983340252190828, + 0.008885177783668041, + 0.0036799439694732428, + 0.004563216120004654, + 0.018529582768678665, + 0.005890662781894207, + 0.05009312555193901, + 0.003718968015164137, + 0.0031760160345584154, + 0.006688327994197607, + 0.0027035053353756666, + 0.019028212875127792, + 0.0034787303302437067, + 0.003955420106649399, + 0.01861402951180935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.03812266141176224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013453251449391246 + ], + "opposite_logprob_tokens": [ + -6.626345157623291 + ], + "logprob_reward": -0.0013453251449391246, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013453251449391246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08408321440219879, + 0.0411594994366169, + 0.04213502258062363, + 0.07050240784883499, + 0.034705642610788345, + 0.0505375936627388, + 0.20400680601596832, + 0.0467965304851532, + 0.18669074773788452, + 0.2393825650215149 + ], + "credit_map": [ + 0.010317221283912659, + 0.0101897232234478, + 0.013175207190215588, + 0.01019322406500578, + 0.0056994082406163216, + 0.006848107557743788, + 0.016957156360149384, + 0.01070315670222044, + 0.004251780454069376, + 0.0051220618188381195, + 0.006111861206591129, + 0.005560533609241247, + 0.004790653940290213, + 0.006567827891558409, + 0.003970809280872345, + 0.004783971700817347, + 0.003728072624653578, + 0.0037946971133351326, + 0.005500922445207834, + 0.004617748782038689, + 0.0049202111549675465, + 0.008319755084812641, + 0.0030077083501964808, + 0.008245903998613358, + 0.010253407061100006, + 0.007275731302797794, + 0.009918594732880592, + 0.006955793127417564, + 0.006402680184692144, + 0.00867547933012247, + 0.011894763447344303, + 0.009125953540205956, + 0.0038666643667966127, + 0.005056510679423809, + 0.003815852338448167, + 0.004045269452035427, + 0.005541669670492411, + 0.005688882432878017, + 0.001973946113139391, + 0.004716845229268074, + 0.007042291108518839, + 0.01180964708328247, + 0.003916543908417225, + 0.005743816029280424, + 0.005531882867217064, + 0.006532010156661272, + 0.005276024341583252, + 0.004685372579842806, + 0.011355288326740265, + 0.010621726512908936, + 0.009526312351226807, + 0.012658056803047657, + 0.011372752487659454, + 0.06608667969703674, + 0.0197127778083086, + 0.0626731812953949, + 0.0049273231998085976, + 0.003129150252789259, + 0.010260148905217648, + 0.003545243525877595, + 0.004578066058456898, + 0.00790624599903822, + 0.003343908116221428, + 0.009106440469622612, + 0.03740343451499939, + 0.01378239318728447, + 0.01340414397418499, + 0.011712422594428062, + 0.014346792362630367, + 0.03298632800579071, + 0.01412620022892952, + 0.048929017037153244, + 0.01822645217180252, + 0.013223717920482159, + 0.01752450317144394, + 0.0116200540214777, + 0.08660481125116348, + 0.015036443248391151, + 0.017472993582487106, + 0.05967358499765396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.0013453251449391246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734886672347784 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734886672347784, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734886672347784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09326489269733429, + 0.03404950723052025, + 0.05464913696050644, + 0.03967026248574257, + 0.01919296383857727, + 0.013565837405622005, + 0.24182915687561035, + 0.0528082549571991, + 0.2648862898349762, + 0.18608374893665314 + ], + "credit_map": [ + 0.012684824876487255, + 0.007097790017724037, + 0.010020408779382706, + 0.009040419943630695, + 0.011434260755777359, + 0.015412045642733574, + 0.014272660948336124, + 0.013302478939294815, + 0.004089868161827326, + 0.004607208538800478, + 0.005203973967581987, + 0.003697984851896763, + 0.004388741217553616, + 0.005557399708777666, + 0.0023472171742469072, + 0.004157107323408127, + 0.005568380933254957, + 0.005061422009021044, + 0.007037260103970766, + 0.004659160505980253, + 0.006178421899676323, + 0.01028386689722538, + 0.005848603788763285, + 0.010012013837695122, + 0.004291387740522623, + 0.002817636588588357, + 0.005822944454848766, + 0.004024905152618885, + 0.004420297686010599, + 0.0049567436799407005, + 0.009134762920439243, + 0.004201579838991165, + 0.0017151649808511138, + 0.0023118872195482254, + 0.0032213919330388308, + 0.00273243454284966, + 0.0024874077644199133, + 0.002067954046651721, + 0.002013844670727849, + 0.0026428778655827045, + 0.001960244495421648, + 0.0022687476594001055, + 0.001702175009995699, + 0.0015823334688320756, + 0.0011402270756661892, + 0.0019911741837859154, + 0.001356015563942492, + 0.0015649194829165936, + 0.012297307141125202, + 0.013855954632163048, + 0.012190097011625767, + 0.016339853405952454, + 0.0156361386179924, + 0.07782802730798721, + 0.01886971853673458, + 0.07481203973293304, + 0.003941978793591261, + 0.006001378875225782, + 0.007506178226321936, + 0.005807321984320879, + 0.006298086605966091, + 0.009990246035158634, + 0.00599006237462163, + 0.007273002527654171, + 0.04820806160569191, + 0.019798507913947105, + 0.022769836708903313, + 0.017658397555351257, + 0.02181326597929001, + 0.04015563428401947, + 0.023897912353277206, + 0.07058463990688324, + 0.015550097450613976, + 0.00825162511318922, + 0.0612323135137558, + 0.010433959774672985, + 0.03468019515275955, + 0.01570810191333294, + 0.00874982587993145, + 0.03147761523723602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.006734886672347784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a side-on tracking shot of a tiger charging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange-yellow with deep, irregular black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a humid, dense rainforest with dappled sunlight filtering through treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the massive ferns and tangled vines in the background rendered as a motion-blur of deep greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021983047481626272 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.0021983047481626272, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021983047481626272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07636487483978271, + 0.047437362372875214, + 0.06215103343129158, + 0.043291542679071426, + 0.05421468988060951, + 0.02456471510231495, + 0.1636180430650711, + 0.03476981818675995, + 0.3194631040096283, + 0.1741247922182083 + ], + "credit_map": [ + 0.006153269670903683, + 0.005785329267382622, + 0.009453055448830128, + 0.007020395249128342, + 0.011971751227974892, + 0.010532190091907978, + 0.013821984641253948, + 0.011626911349594593, + 0.006138525437563658, + 0.0044029830023646355, + 0.009095153771340847, + 0.004949954338371754, + 0.0070319389924407005, + 0.008237083442509174, + 0.00285320938564837, + 0.004728521220386028, + 0.005108020268380642, + 0.00580774387344718, + 0.00613258546218276, + 0.005829323083162308, + 0.007534204050898552, + 0.01261279545724392, + 0.004750689957290888, + 0.01437568012624979, + 0.004451183136552572, + 0.004556489177048206, + 0.005099677015095949, + 0.005215903278440237, + 0.006284658797085285, + 0.005034076515585184, + 0.0078072021715343, + 0.0048423572443425655, + 0.005119939800351858, + 0.008363083004951477, + 0.008810342289507389, + 0.0074327620677649975, + 0.006369114387780428, + 0.008786712773144245, + 0.0037684987764805555, + 0.005564241204410791, + 0.00204634596593678, + 0.0035471650771796703, + 0.0035164072178304195, + 0.0029782650526612997, + 0.0024829250760376453, + 0.003943974617868662, + 0.003695536870509386, + 0.0023540984839200974, + 0.007885773666203022, + 0.01055101677775383, + 0.007576418574899435, + 0.01107032410800457, + 0.012538272887468338, + 0.04872400686144829, + 0.01855398342013359, + 0.04671826586127281, + 0.002818519249558449, + 0.00379253295250237, + 0.006099506746977568, + 0.003262399695813656, + 0.003671922953799367, + 0.0043592932634055614, + 0.004118700046092272, + 0.006646949332207441, + 0.05353793874382973, + 0.018016057088971138, + 0.0224745012819767, + 0.011263241060078144, + 0.012855535373091698, + 0.05222631245851517, + 0.01715525984764099, + 0.1319342851638794, + 0.010923412628471851, + 0.006901508197188377, + 0.0428687147796154, + 0.006701312027871609, + 0.026753442361950874, + 0.01190967857837677, + 0.009731190279126167, + 0.05833553895354271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.0021983047481626272 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_9/index_tensor([3799, 3799])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_9/index_tensor([3799, 3799])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..989c3102d8d82f009b3034f330fcdf253375653f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_9/index_tensor([3799, 3799])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic slow-motion close-up captures the action from a slightly elevated angle as a dark blue six-sided die falls into the frame. Made of polished, slightly translucent acrylic, the die impacts the warm, wood-textured surface of the kitchen table with a sharp impact. Under the soft, realistic overhead kitchen light, it bounces erratically, spinning in mid-air, causing the white, concave dots to blur into fleeting streaks. The camera remains static, allowing dynamic movement to fill the scene as the die loses momentum, gliding across the polished wood before swaying to a final, tense moment. It settles decisively, completely at rest, shifting the feel of the scene from chaotic energy to stillness. The final frame fixates on the perfectly still die, its top surface clearly displaying the number six. The warm light casts soft, fleeting shadows alongside it and creates subtle highlights along its smooth edges, creating a quiet, defined atmosphere. High resolution.", + "seed": 21612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_9/index_3799_seed_21612_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02195641864091158, + "vlm_logprob": -0.02195641864091158, + "vlm_yes_no_margin": 4.024999952316284, + "vlm_token_credit": -0.02195641864091158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0273251011967659 + ], + "opposite_logprob_tokens": [ + -3.90232515335083 + ], + "logprob_reward": -0.0273251011967659, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0273251011967659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10943728685379028, + 0.07490846514701843, + 0.04585462808609009, + 0.048944659531116486, + 0.07478931546211243, + 0.11018355935811996, + 0.10101915895938873, + 0.19383186101913452, + 0.06028718501329422, + 0.18074390292167664 + ], + "credit_map": [ + 0.011956211179494858, + 0.01868862845003605, + 0.008286678232252598, + 0.011724249459803104, + 0.014065823517739773, + 0.016793038696050644, + 0.013493653386831284, + 0.014429003931581974, + 0.007461121771484613, + 0.01063466165214777, + 0.009218361228704453, + 0.009003885090351105, + 0.009134180843830109, + 0.010126721113920212, + 0.009458363987505436, + 0.009871168062090874, + 0.00579819455742836, + 0.005886661354452372, + 0.007491367403417826, + 0.008600987493991852, + 0.0033915690146386623, + 0.0032002064399421215, + 0.0053510707803070545, + 0.006134574767202139, + 0.0032478293869644403, + 0.0037264470010995865, + 0.007878504693508148, + 0.008279531262814999, + 0.00493837334215641, + 0.005492270924150944, + 0.005908171180635691, + 0.009473533369600773, + 0.005716822110116482, + 0.005067647900432348, + 0.00991443358361721, + 0.010891885496675968, + 0.010991831310093403, + 0.010884263552725315, + 0.010239103808999062, + 0.011083324439823627, + 0.00813829991966486, + 0.010241095907986164, + 0.01493910700082779, + 0.00811704620718956, + 0.020577164366841316, + 0.015603671781718731, + 0.016612786799669266, + 0.015954377129673958, + 0.010646832175552845, + 0.007247413508594036, + 0.014081554487347603, + 0.01470231730490923, + 0.00998589489609003, + 0.015930363908410072, + 0.012976441532373428, + 0.015448340214788914, + 0.012184682302176952, + 0.0076036942191421986, + 0.057547181844711304, + 0.006476212292909622, + 0.009841647930443287, + 0.07795283943414688, + 0.00487022940069437, + 0.01735537312924862, + 0.004671107511967421, + 0.005056337453424931, + 0.012968449853360653, + 0.005579192657023668, + 0.005425882060080767, + 0.01438372302800417, + 0.005205290392041206, + 0.006997203920036554, + 0.00989783089607954, + 0.004147169645875692, + 0.05607305094599724, + 0.008328347466886044, + 0.012134590186178684, + 0.07474081218242645, + 0.004288301337510347, + 0.011133785359561443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.0273251011967659 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562665157020092 + ], + "opposite_logprob_tokens": [ + -4.3875627517700195 + ], + "logprob_reward": -0.012562665157020092, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562665157020092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12789735198020935, + 0.038894928991794586, + 0.057873230427503586, + 0.10401507467031479, + 0.046371251344680786, + 0.1214466392993927, + 0.048464950174093246, + 0.34802865982055664, + 0.06467213481664658, + 0.04233571141958237 + ], + "credit_map": [ + 0.01012367568910122, + 0.014538578689098358, + 0.019951792433857918, + 0.015260562300682068, + 0.023149406537413597, + 0.014139735139906406, + 0.014162218198180199, + 0.016571389511227608, + 0.0026780990883708, + 0.0051877461373806, + 0.004967155400663614, + 0.005408810917288065, + 0.004220266826450825, + 0.005161934066563845, + 0.006040074862539768, + 0.005230840761214495, + 0.005364392418414354, + 0.0041785635985434055, + 0.008717517368495464, + 0.008772579953074455, + 0.005810711998492479, + 0.0058007994666695595, + 0.009560116566717625, + 0.009668546728789806, + 0.007336982525885105, + 0.010602411814033985, + 0.01498282328248024, + 0.015013845637440681, + 0.011608162894845009, + 0.01485956646502018, + 0.01253530289977789, + 0.017075981944799423, + 0.004587519448250532, + 0.006196553353220224, + 0.006957666017115116, + 0.008189302869141102, + 0.004201951902359724, + 0.006078541744500399, + 0.0037901673931628466, + 0.006369545124471188, + 0.009458772838115692, + 0.009855124168097973, + 0.015292397700250149, + 0.010436839424073696, + 0.02456897310912609, + 0.01788102649152279, + 0.02017834782600403, + 0.01377516146749258, + 0.004950929898768663, + 0.0044439020566642284, + 0.006457951385527849, + 0.0081656314432621, + 0.0037234597839415073, + 0.008688969537615776, + 0.005145797971636057, + 0.006888309959322214, + 0.021382298320531845, + 0.012741459533572197, + 0.10806901007890701, + 0.015871554613113403, + 0.0170036181807518, + 0.14445804059505463, + 0.009576036594808102, + 0.018926644697785378, + 0.0035115336067974567, + 0.00193992059212178, + 0.01928539387881756, + 0.004318712279200554, + 0.003595224115997553, + 0.022756407037377357, + 0.0031358443666249514, + 0.006129096262156963, + 0.002681915881112218, + 0.0010660653933882713, + 0.012297200039029121, + 0.0017097663367167115, + 0.0027854880318045616, + 0.016849340870976448, + 0.0010788710787892342, + 0.0038670655339956284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.012562665157020092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864654965698719 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.00864654965698719, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864654965698719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16115562617778778, + 0.06309640407562256, + 0.08201338350772858, + 0.06085387244820595, + 0.08392543345689774, + 0.1685585230588913, + 0.169900581240654, + 0.08185135573148727, + 0.05798503756523132, + 0.07065972685813904 + ], + "credit_map": [ + 0.014276597648859024, + 0.02359343320131302, + 0.01491649728268385, + 0.022012270987033844, + 0.017953060567378998, + 0.01999955251812935, + 0.024177785962820053, + 0.024226443842053413, + 0.00871978048235178, + 0.00859471969306469, + 0.00714579364284873, + 0.0070738778449594975, + 0.004940313287079334, + 0.007755409926176071, + 0.00943758711218834, + 0.00942892674356699, + 0.0069641852751374245, + 0.006072260905057192, + 0.01167624443769455, + 0.010444837622344494, + 0.010709348134696484, + 0.00871066004037857, + 0.016270043328404427, + 0.011165815405547619, + 0.005113130901008844, + 0.00932219997048378, + 0.009209168143570423, + 0.011626833118498325, + 0.005084835924208164, + 0.005964741576462984, + 0.006433788686990738, + 0.008099177852272987, + 0.006158856209367514, + 0.006516058463603258, + 0.011682827025651932, + 0.012948956340551376, + 0.01212526299059391, + 0.01164278481155634, + 0.011255308985710144, + 0.011595388874411583, + 0.01222292985767126, + 0.012853733263909817, + 0.02047796919941902, + 0.014034549705684185, + 0.03603949025273323, + 0.023046990856528282, + 0.03256840631365776, + 0.017314467579126358, + 0.011227116920053959, + 0.009243862703442574, + 0.03583161532878876, + 0.017266640439629555, + 0.011390558443963528, + 0.040010929107666016, + 0.01134544238448143, + 0.03358442336320877, + 0.008322574198246002, + 0.004558552987873554, + 0.014765646308660507, + 0.0044637578539550304, + 0.005720030516386032, + 0.01801260933279991, + 0.005194351077079773, + 0.02081383764743805, + 0.003270120359957218, + 0.002438623458147049, + 0.015962930396199226, + 0.00467889616265893, + 0.003297323826700449, + 0.01917426660656929, + 0.0034856041893363, + 0.005677276290953159, + 0.004027186427265406, + 0.0016390252858400345, + 0.02119462750852108, + 0.0034659404773265123, + 0.0049130297265946865, + 0.02930580824613571, + 0.0018040781142190099, + 0.004310039337724447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.00864654965698719 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202990293502808 + ], + "opposite_logprob_tokens": [ + -4.018202781677246 + ], + "logprob_reward": -0.018202990293502808, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018202990293502808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1755027323961258, + 0.09228764474391937, + 0.04110851138830185, + 0.042358361184597015, + 0.058915168046951294, + 0.11973561346530914, + 0.0876668319106102, + 0.3068656027317047, + 0.04889094457030296, + 0.02666855789721012 + ], + "credit_map": [ + 0.0156742911785841, + 0.026964638382196426, + 0.01885964348912239, + 0.01866883412003517, + 0.023573579266667366, + 0.028629684820771217, + 0.025450529530644417, + 0.017681527882814407, + 0.007980470545589924, + 0.01309546921402216, + 0.012431016191840172, + 0.014044612646102905, + 0.008263777941465378, + 0.008836397901177406, + 0.011063095182180405, + 0.016572801396250725, + 0.0033330039586871862, + 0.004033063538372517, + 0.005946732126176357, + 0.0064392429776489735, + 0.004842505790293217, + 0.0038679144345223904, + 0.007953301072120667, + 0.004692743998020887, + 0.00368728069588542, + 0.003415742190554738, + 0.00614608358591795, + 0.006608365569263697, + 0.004384645260870457, + 0.0052202786318957806, + 0.005556951276957989, + 0.0073390137404203415, + 0.005731192883104086, + 0.005495349410921335, + 0.007457831408828497, + 0.009801018051803112, + 0.0073292856104671955, + 0.007426655851304531, + 0.007717891130596399, + 0.007955947890877724, + 0.010614932514727116, + 0.010026448406279087, + 0.01565261371433735, + 0.009877744130790234, + 0.02144630067050457, + 0.01913127489387989, + 0.02018703892827034, + 0.012799261137843132, + 0.009539623744785786, + 0.006321161054074764, + 0.012423558160662651, + 0.013514571823179722, + 0.008043617010116577, + 0.013951876200735569, + 0.00938399974256754, + 0.014488422311842442, + 0.02015054039657116, + 0.012644675560295582, + 0.0889044776558876, + 0.013956676237285137, + 0.015909595414996147, + 0.12127888947725296, + 0.009129423648118973, + 0.024891339242458344, + 0.003256079275161028, + 0.003941479604691267, + 0.010324660688638687, + 0.005673669744282961, + 0.004399755969643593, + 0.012658597901463509, + 0.003994202706962824, + 0.004642500542104244, + 0.002350427210330963, + 0.0014987954637035728, + 0.0059724668972194195, + 0.0016835985006764531, + 0.001585579477250576, + 0.006065484136343002, + 0.001710204523988068, + 0.00580200320109725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.018202990293502808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043044786900281906 + ], + "opposite_logprob_tokens": [ + -3.1680448055267334 + ], + "logprob_reward": -0.043044786900281906, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043044786900281906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13526417315006256, + 0.08180408924818039, + 0.048484835773706436, + 0.03555121645331383, + 0.036037035286426544, + 0.09441342204809189, + 0.05385734140872955, + 0.3786517083644867, + 0.024587856605648994, + 0.11134829372167587 + ], + "credit_map": [ + 0.011414624750614166, + 0.01759745180606842, + 0.01327530201524496, + 0.01908414624631405, + 0.01983223296701908, + 0.020514773204922676, + 0.013810333795845509, + 0.0197353083640337, + 0.008276041597127914, + 0.011826827190816402, + 0.01331060566008091, + 0.011257041245698929, + 0.005165873095393181, + 0.0060419547371566296, + 0.006597362458705902, + 0.01932837814092636, + 0.004406984895467758, + 0.0042312550358474255, + 0.0072588250041007996, + 0.0066642314195632935, + 0.006293380632996559, + 0.0048535168170928955, + 0.008826227858662605, + 0.005950414575636387, + 0.0034355768002569675, + 0.002978063188493252, + 0.00487494096159935, + 0.005316160619258881, + 0.004516315180808306, + 0.004247703589498997, + 0.0046289521269500256, + 0.005553507246077061, + 0.002623191336169839, + 0.0036436133086681366, + 0.0045823510736227036, + 0.005547522101551294, + 0.004611484240740538, + 0.0046527199447155, + 0.004240470007061958, + 0.006135683972388506, + 0.008873752318322659, + 0.00907359179109335, + 0.013817318715155125, + 0.009668760001659393, + 0.010012781247496605, + 0.01749807968735695, + 0.013194053433835506, + 0.012275083921849728, + 0.006101321429014206, + 0.003988771233707666, + 0.009932290762662888, + 0.0041669029742479324, + 0.004633399657905102, + 0.012833879329264164, + 0.004057534970343113, + 0.008143237791955471, + 0.025058167055249214, + 0.012382229790091515, + 0.09742897003889084, + 0.017959630116820335, + 0.021349437534809113, + 0.12912552058696747, + 0.014203593134880066, + 0.06114418804645538, + 0.0015809416072443128, + 0.0014908025041222572, + 0.00559633132070303, + 0.0031422951724380255, + 0.0018121517496183515, + 0.007297225296497345, + 0.0018251803703606129, + 0.001842930563725531, + 0.006288034841418266, + 0.00203821063041687, + 0.035038094967603683, + 0.003966709598898888, + 0.007314837537705898, + 0.0476684607565403, + 0.0018166383961215615, + 0.007217301055788994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.043044786900281906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic slow-motion close-up captures the action from a slightly elevated angle as a dark blue six-sided die falls into the frame. Made of polished, slightly translucent acrylic, the die impacts the warm, wood-textured surface of the kitchen table with a sharp impact. Under the soft, realistic overhead kitchen light, it bounces erratically, spinning in mid-air, causing the white, concave dots to blur into fleeting streaks. The camera remains static, allowing dynamic movement to fill the scene as the die loses momentum, gliding across the polished wood before swaying to a final, tense moment. It settles decisively, completely at rest, shifting the feel of the scene from chaotic energy to stillness. The final frame fixates on the perfectly still die, its top surface clearly displaying the number six. The warm light casts soft, fleeting shadows alongside it and creates subtle highlights along its smooth edges, creating a quiet, defined atmosphere. High resolution.", + "seed": 21613, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_9/index_3799_seed_21613_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03109834957867861, + "vlm_logprob": -0.03109834957867861, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.03109834957867861, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032292336225509644 + ], + "opposite_logprob_tokens": [ + -3.657292366027832 + ], + "logprob_reward": -0.032292336225509644, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.032292336225509644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10905937850475311, + 0.07797688990831375, + 0.06285494565963745, + 0.03713489696383476, + 0.14063729345798492, + 0.12439952790737152, + 0.15129294991493225, + 0.19117389619350433, + 0.07527842372655869, + 0.030191780999302864 + ], + "credit_map": [ + 0.011741632595658302, + 0.013593437150120735, + 0.022136332467198372, + 0.012678850442171097, + 0.01134501677006483, + 0.011313977651298046, + 0.012679318897426128, + 0.013570815324783325, + 0.005356530658900738, + 0.00809953361749649, + 0.009250530041754246, + 0.013146914541721344, + 0.012016084976494312, + 0.0064427549950778484, + 0.012606003321707249, + 0.01105854008346796, + 0.003642050549387932, + 0.008131549693644047, + 0.006736647803336382, + 0.005495226942002773, + 0.005314159672707319, + 0.014322089962661266, + 0.007039132993668318, + 0.012174088507890701, + 0.0033494140952825546, + 0.003863806603476405, + 0.00520274369046092, + 0.004033104982227087, + 0.004120236728340387, + 0.00573652284219861, + 0.006417371798306704, + 0.004411695525050163, + 0.012371928431093693, + 0.02301226556301117, + 0.028905482962727547, + 0.019667303189635277, + 0.014000793918967247, + 0.017928630113601685, + 0.013150349259376526, + 0.011600535362958908, + 0.009608106687664986, + 0.016491606831550598, + 0.013599350117146969, + 0.016976917162537575, + 0.01233822014182806, + 0.025636158883571625, + 0.018003851175308228, + 0.011745313182473183, + 0.012337462976574898, + 0.01573505997657776, + 0.033461906015872955, + 0.016120709478855133, + 0.01500377245247364, + 0.023164095357060432, + 0.013309174217283726, + 0.02216077782213688, + 0.018767954781651497, + 0.016181163489818573, + 0.025936389341950417, + 0.024252723902463913, + 0.023072415962815285, + 0.02779771015048027, + 0.03157617151737213, + 0.02358935959637165, + 0.007352378685027361, + 0.005232389084994793, + 0.009209951385855675, + 0.011095120571553707, + 0.008137239143252373, + 0.009071391075849533, + 0.013452483341097832, + 0.0117274709045887, + 0.004043704364448786, + 0.0023611080832779408, + 0.005099068861454725, + 0.0032037929631769657, + 0.002329186536371708, + 0.004518511239439249, + 0.003692137310281396, + 0.004944273270666599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.032292336225509644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026376847177743912 + ], + "opposite_logprob_tokens": [ + -3.651376962661743 + ], + "logprob_reward": -0.026376847177743912, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026376847177743912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1691674441099167, + 0.04787077382206917, + 0.06880100816488266, + 0.04394448921084404, + 0.09639506042003632, + 0.06523126363754272, + 0.2580536901950836, + 0.23360851407051086, + 0.008947200141847134, + 0.007980580441653728 + ], + "credit_map": [ + 0.013528760522603989, + 0.02515237219631672, + 0.023808537051081657, + 0.026752332225441933, + 0.018615813925862312, + 0.016997694969177246, + 0.023444822058081627, + 0.02086709998548031, + 0.005306038074195385, + 0.004100047517567873, + 0.009507056325674057, + 0.005502520594745874, + 0.006154260132461786, + 0.0048542930744588375, + 0.0058394926600158215, + 0.006607064511626959, + 0.005341251380741596, + 0.010307916440069675, + 0.013685375452041626, + 0.00954696536064148, + 0.005608564708381891, + 0.008925992995500565, + 0.0068881455808877945, + 0.008496803231537342, + 0.004163670353591442, + 0.008365324698388577, + 0.004742086865007877, + 0.004156035371124744, + 0.004487208090722561, + 0.011215224862098694, + 0.00341609213501215, + 0.003398846834897995, + 0.008351204916834831, + 0.016788462176918983, + 0.018413612619042397, + 0.012311327271163464, + 0.00957333855330944, + 0.01333083026111126, + 0.010101194493472576, + 0.007525090593844652, + 0.007394762244075537, + 0.007345751393586397, + 0.00942690297961235, + 0.00982159748673439, + 0.007307935506105423, + 0.008314955048263073, + 0.01031712256371975, + 0.005302241537719965, + 0.021530168130993843, + 0.027526330202817917, + 0.05126921460032463, + 0.02613730914890766, + 0.029750287532806396, + 0.04037497565150261, + 0.02439643070101738, + 0.03706898167729378, + 0.023954981938004494, + 0.01938071846961975, + 0.03072214312851429, + 0.031836967915296555, + 0.02531866356730461, + 0.03195986524224281, + 0.04142893850803375, + 0.02900623343884945, + 0.0013375055277720094, + 0.0008455600473098457, + 0.001501412014476955, + 0.0009481357992626727, + 0.00107031490188092, + 0.0009935421403497458, + 0.001106930896639824, + 0.0011437986977398396, + 0.0013825884088873863, + 0.0006094668060541153, + 0.0017891767201945186, + 0.0006212008302100003, + 0.0006666647968813777, + 0.0008432394824922085, + 0.0009107770165428519, + 0.0011574659729376435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.026376847177743912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202469028532505 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202469028532505, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202469028532505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15916754305362701, + 0.0914100632071495, + 0.11259336024522781, + 0.10754372924566269, + 0.12249865382909775, + 0.07516046613454819, + 0.07717642188072205, + 0.12071457505226135, + 0.03190986439585686, + 0.10182531923055649 + ], + "credit_map": [ + 0.013917366974055767, + 0.02361082285642624, + 0.022868381813168526, + 0.016666335985064507, + 0.020677918568253517, + 0.017809394747018814, + 0.02250535972416401, + 0.021111972630023956, + 0.0067451708018779755, + 0.00949084758758545, + 0.010001201182603836, + 0.013774984516203403, + 0.013668084517121315, + 0.008795191533863544, + 0.017050351947546005, + 0.011884241364896297, + 0.005724241957068443, + 0.010029763914644718, + 0.012824131175875664, + 0.010833850130438805, + 0.007907622493803501, + 0.028292670845985413, + 0.008917218074202538, + 0.028063874691724777, + 0.015057378448545933, + 0.011048402637243271, + 0.018180670216679573, + 0.011020216159522533, + 0.013295301236212254, + 0.013887947425246239, + 0.014333086088299751, + 0.010720730759203434, + 0.011741766706109047, + 0.013506428338587284, + 0.018617652356624603, + 0.013504465110599995, + 0.01587279513478279, + 0.016145238652825356, + 0.019114740192890167, + 0.013995575718581676, + 0.008625161834061146, + 0.009785490110516548, + 0.013213141821324825, + 0.01051592268049717, + 0.005733863916248083, + 0.01412320975214243, + 0.007751092314720154, + 0.00541258929297328, + 0.006946682929992676, + 0.006970248185098171, + 0.016411956399679184, + 0.009420818649232388, + 0.008518151007592678, + 0.008386540226638317, + 0.009548109024763107, + 0.010973919183015823, + 0.011562183499336243, + 0.009699348360300064, + 0.01647068001329899, + 0.016605179756879807, + 0.012952923774719238, + 0.015939801931381226, + 0.022154683247208595, + 0.015329774469137192, + 0.004146272782236338, + 0.002713414141908288, + 0.005944233853369951, + 0.003043035976588726, + 0.0032664071768522263, + 0.003181679639965296, + 0.004932632204145193, + 0.00468218931928277, + 0.01341506838798523, + 0.008614825084805489, + 0.016522403806447983, + 0.010767385363578796, + 0.008814116939902306, + 0.0151534304022789, + 0.013283757492899895, + 0.015254332683980465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.014202469028532505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591415464878082 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591415464878082, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591415464878082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26361286640167236, + 0.1164160966873169, + 0.06855630874633789, + 0.04571665823459625, + 0.09200365841388702, + 0.1022130474448204, + 0.11987088620662689, + 0.10649207979440689, + 0.05559789761900902, + 0.029520491138100624 + ], + "credit_map": [ + 0.0255522970110178, + 0.04107056185603142, + 0.037864841520786285, + 0.03958722949028015, + 0.029033076018095016, + 0.024549834430217743, + 0.03549977019429207, + 0.030455250293016434, + 0.012117414735257626, + 0.00914109405130148, + 0.025470567867159843, + 0.017820758745074272, + 0.011329306289553642, + 0.013801036402583122, + 0.010008118115365505, + 0.01672780141234398, + 0.004368886351585388, + 0.007868246175348759, + 0.0078652985394001, + 0.00529453856870532, + 0.00615380285307765, + 0.016447627916932106, + 0.0064141531474888325, + 0.014143754728138447, + 0.0030105437617748976, + 0.008813670836389065, + 0.00404747948050499, + 0.004085130523890257, + 0.003859623335301876, + 0.014061548747122288, + 0.005199648905545473, + 0.0026390119455754757, + 0.009267177432775497, + 0.012928834185004234, + 0.01531666424125433, + 0.009232199750840664, + 0.011609374545514584, + 0.015798252075910568, + 0.009193985722959042, + 0.008657174184918404, + 0.009593508206307888, + 0.014264444820582867, + 0.021995913237333298, + 0.012929658405482769, + 0.008219726383686066, + 0.012060800567269325, + 0.012207035906612873, + 0.010941964574158192, + 0.017816031351685524, + 0.011397965252399445, + 0.02084927447140217, + 0.015132863074541092, + 0.013073851354420185, + 0.011564225889742374, + 0.014396978542208672, + 0.015639690682291985, + 0.013895783573389053, + 0.009672282263636589, + 0.021005479618906975, + 0.01094009354710579, + 0.010025852359831333, + 0.01354781910777092, + 0.013117642141878605, + 0.014287120662629604, + 0.008571121841669083, + 0.004915217403322458, + 0.012281886301934719, + 0.005654486827552319, + 0.004229311365634203, + 0.007161132991313934, + 0.005590599961578846, + 0.007194141857326031, + 0.0037049672100692987, + 0.0018536730203777552, + 0.0073409900069236755, + 0.0025318614207208157, + 0.002418390940874815, + 0.002730096923187375, + 0.00468247989192605, + 0.004258031956851482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.020591415464878082 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620286799967289 + ], + "opposite_logprob_tokens": [ + -2.812028646469116 + ], + "logprob_reward": -0.0620286799967289, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620286799967289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11985981464385986, + 0.0724007710814476, + 0.05360323190689087, + 0.04633239284157753, + 0.21032503247261047, + 0.3549559712409973, + 0.055957335978746414, + 0.05970803648233414, + 0.01874871738255024, + 0.008108711801469326 + ], + "credit_map": [ + 0.006970755290240049, + 0.013561403378844261, + 0.014284156262874603, + 0.013075705617666245, + 0.012755125761032104, + 0.022994935512542725, + 0.010982790030539036, + 0.025234950706362724, + 0.00697383051738143, + 0.008952448144555092, + 0.011526725254952908, + 0.007884254679083824, + 0.011323044076561928, + 0.005947127938270569, + 0.0116360979154706, + 0.008157243952155113, + 0.003473102580755949, + 0.004599816165864468, + 0.005984903313219547, + 0.006251228973269463, + 0.005593856796622276, + 0.012052524834871292, + 0.0061919004656374454, + 0.009455897845327854, + 0.0031258456874638796, + 0.008539918810129166, + 0.0032968884333968163, + 0.0042976983822882175, + 0.004563871771097183, + 0.01500942837446928, + 0.0045232526026666164, + 0.0029754876159131527, + 0.014311833307147026, + 0.02928762510418892, + 0.03400713577866554, + 0.026661472395062447, + 0.027289632707834244, + 0.037288371473550797, + 0.02267778478562832, + 0.018801167607307434, + 0.03151140362024307, + 0.051469095051288605, + 0.05431566759943962, + 0.04365299269556999, + 0.027845920994877815, + 0.07636158913373947, + 0.041146330535411835, + 0.02865295298397541, + 0.00517182657495141, + 0.00625921506434679, + 0.01025676354765892, + 0.005305358674377203, + 0.006654071621596813, + 0.009218121878802776, + 0.005064226686954498, + 0.008027754724025726, + 0.006274704821407795, + 0.004948541056364775, + 0.007931185886263847, + 0.007683014962822199, + 0.006349424831569195, + 0.008967963978648186, + 0.010116222314536572, + 0.007436975836753845, + 0.001855370937846601, + 0.0013462117640301585, + 0.0030890724156051874, + 0.0019486949313431978, + 0.0017699283780530095, + 0.002987463027238846, + 0.0030105949845165014, + 0.0027413798961788416, + 0.0009544766508042812, + 0.0007589756278321147, + 0.0021459858398884535, + 0.0007336729322560132, + 0.0007438799948431551, + 0.0007896857568994164, + 0.0008825470576994121, + 0.0010994885815307498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0620286799967289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic slow-motion close-up captures the action from a slightly elevated angle as a dark blue six-sided die falls into the frame. Made of polished, slightly translucent acrylic, the die impacts the warm, wood-textured surface of the kitchen table with a sharp impact. Under the soft, realistic overhead kitchen light, it bounces erratically, spinning in mid-air, causing the white, concave dots to blur into fleeting streaks. The camera remains static, allowing dynamic movement to fill the scene as the die loses momentum, gliding across the polished wood before swaying to a final, tense moment. It settles decisively, completely at rest, shifting the feel of the scene from chaotic energy to stillness. The final frame fixates on the perfectly still die, its top surface clearly displaying the number six. The warm light casts soft, fleeting shadows alongside it and creates subtle highlights along its smooth edges, creating a quiet, defined atmosphere. High resolution.", + "seed": 21614, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_9/index_3799_seed_21614_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02575335744768381, + "vlm_logprob": -0.02575335744768381, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02575335744768381, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01871475763618946 + ], + "opposite_logprob_tokens": [ + -4.518714904785156 + ], + "logprob_reward": -0.01871475763618946, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01871475763618946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26692771911621094, + 0.06245080381631851, + 0.21042153239250183, + 0.10526696592569351, + 0.027157891541719437, + 0.12294387072324753, + 0.030307579785585403, + 0.048363205045461655, + 0.08468884229660034, + 0.041471585631370544 + ], + "credit_map": [ + 0.016313286498188972, + 0.008470806293189526, + 0.10514898598194122, + 0.010314247570931911, + 0.006918648723512888, + 0.1072981059551239, + 0.006304324604570866, + 0.0061592827551066875, + 0.007900320924818516, + 0.007520290557295084, + 0.015487782657146454, + 0.0058495053090155125, + 0.007237826939672232, + 0.007972268387675285, + 0.005645234137773514, + 0.0048375739715993404, + 0.018901627510786057, + 0.07319115847349167, + 0.07246380299329758, + 0.0171948354691267, + 0.007835078053176403, + 0.00428068358451128, + 0.006012459751218557, + 0.010541887022554874, + 0.011685377918183804, + 0.013484815135598183, + 0.013221760280430317, + 0.013849721290171146, + 0.0110658910125494, + 0.016060838475823402, + 0.011137309484183788, + 0.014761249534785748, + 0.0030888987239450216, + 0.0041512795723974705, + 0.005366859957575798, + 0.004606776405125856, + 0.002646440640091896, + 0.002292853780090809, + 0.002128954278305173, + 0.0028758288826793432, + 0.012740484438836575, + 0.020844312384724617, + 0.02543325163424015, + 0.022438012063503265, + 0.00947107095271349, + 0.008823228068649769, + 0.012964785099029541, + 0.010228724218904972, + 0.003502804087474942, + 0.009707021526992321, + 0.00949000008404255, + 0.002122970763593912, + 0.0014480951940640807, + 0.0017547316383570433, + 0.0010506637627258897, + 0.0012312948238104582, + 0.00588610302656889, + 0.002648956375196576, + 0.014733969233930111, + 0.0029004255775362253, + 0.0019015244906768203, + 0.014536377973854542, + 0.0029797889292240143, + 0.0027760572265833616, + 0.00601243507117033, + 0.0018025243189185858, + 0.03254299610853195, + 0.0036590404342859983, + 0.004130102228373289, + 0.031351517885923386, + 0.002791822189465165, + 0.0023984117433428764, + 0.004832741338759661, + 0.004311314783990383, + 0.010499040596187115, + 0.004293784499168396, + 0.0030133589170873165, + 0.007803775370121002, + 0.004253148566931486, + 0.002464422956109047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01871475763618946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982042171061039 + ], + "opposite_logprob_tokens": [ + -3.529820442199707 + ], + "logprob_reward": -0.02982042171061039, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982042171061039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24738375842571259, + 0.03621836006641388, + 0.07854534685611725, + 0.07142727822065353, + 0.03900511562824249, + 0.06456614285707474, + 0.30335313081741333, + 0.08262868225574493, + 0.018593987450003624, + 0.05827820301055908 + ], + "credit_map": [ + 0.013517497107386589, + 0.005401951726526022, + 0.10056453198194504, + 0.011720775626599789, + 0.006300835870206356, + 0.09384924173355103, + 0.008846416138112545, + 0.007182506378740072, + 0.004280114080756903, + 0.0034420795273035765, + 0.00700060511007905, + 0.005430595483630896, + 0.006675563752651215, + 0.0026786711532622576, + 0.003897113725543022, + 0.002813617931678891, + 0.006220417097210884, + 0.020138565450906754, + 0.022079722955822945, + 0.006950006820261478, + 0.004834676627069712, + 0.006315602920949459, + 0.0045796954073011875, + 0.0074266609735786915, + 0.008910312317311764, + 0.008955376222729683, + 0.009077755734324455, + 0.00996063556522131, + 0.007033458910882473, + 0.010171115398406982, + 0.007386409677565098, + 0.009932216256856918, + 0.004514369182288647, + 0.006785247474908829, + 0.0076918164268136024, + 0.007195177022367716, + 0.004144232254475355, + 0.00258180545642972, + 0.002799900248646736, + 0.003292568726465106, + 0.005694128572940826, + 0.010242338292300701, + 0.012942897155880928, + 0.012152284383773804, + 0.005558800417929888, + 0.005376351997256279, + 0.007699156645685434, + 0.004900180734694004, + 0.01585535705089569, + 0.1246376559138298, + 0.12450806796550751, + 0.013852479867637157, + 0.0050913323648273945, + 0.01163612212985754, + 0.004173042252659798, + 0.003599063027650118, + 0.014585685916244984, + 0.00452306168153882, + 0.023431917652487755, + 0.0029991809278726578, + 0.0030968915671110153, + 0.028103284537792206, + 0.0021019482519477606, + 0.003786711022257805, + 0.0015918145654723048, + 0.0009528283262625337, + 0.006183812394738197, + 0.0013501690700650215, + 0.001023639808408916, + 0.005516356322914362, + 0.0010630616452544928, + 0.0009123057243414223, + 0.00661424407735467, + 0.0030670599080622196, + 0.019621560350060463, + 0.0034699037205427885, + 0.002984958002343774, + 0.01722247526049614, + 0.0027718457859009504, + 0.0025261591654270887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.02982042171061039 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202351681888103 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014202351681888103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202351681888103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26087892055511475, + 0.05587223544716835, + 0.15258973836898804, + 0.110263392329216, + 0.064888596534729, + 0.07301764190196991, + 0.05754059553146362, + 0.03362026438117027, + 0.1076182872056961, + 0.08371035009622574 + ], + "credit_map": [ + 0.01559016015380621, + 0.009454101324081421, + 0.10041923820972443, + 0.01306602731347084, + 0.011197246611118317, + 0.09234436601400375, + 0.0104623232036829, + 0.008345454931259155, + 0.006691911723464727, + 0.006758596282452345, + 0.008958037942647934, + 0.011181057430803776, + 0.005414447281509638, + 0.005950056482106447, + 0.006453635171055794, + 0.00446448614820838, + 0.015688981860876083, + 0.04706653207540512, + 0.043343398720026016, + 0.013885379768908024, + 0.008230149745941162, + 0.008040924556553364, + 0.0054891533218324184, + 0.010845206677913666, + 0.011267529800534248, + 0.014509639702737331, + 0.01359823439270258, + 0.014552714303135872, + 0.011743396520614624, + 0.01745995692908764, + 0.011503176763653755, + 0.015628738328814507, + 0.007438189350068569, + 0.010185071267187595, + 0.011030717752873898, + 0.008819503709673882, + 0.007812153548002243, + 0.006235848646610975, + 0.005332659929990768, + 0.00803444441407919, + 0.006830333266407251, + 0.012737330049276352, + 0.01781444624066353, + 0.01423718873411417, + 0.003582998178899288, + 0.006386408116668463, + 0.0067847734317183495, + 0.004644164815545082, + 0.004057339392602444, + 0.020011847838759422, + 0.01951226405799389, + 0.0033274292945861816, + 0.002572619356215, + 0.0027633097488433123, + 0.0027179578319191933, + 0.0025778200943022966, + 0.004015079699456692, + 0.0020844326354563236, + 0.00924668088555336, + 0.0018266424303874373, + 0.0019367446657270193, + 0.010163634084165096, + 0.002319958293810487, + 0.0020270885434001684, + 0.007861494086682796, + 0.002633047988638282, + 0.0421091727912426, + 0.004050025716423988, + 0.003148416057229042, + 0.0413956344127655, + 0.0034073367714881897, + 0.0030131444800645113, + 0.01122246216982603, + 0.0070107076317071915, + 0.026158558204770088, + 0.006684330757707357, + 0.004362477455288172, + 0.018676258623600006, + 0.0044594984501600266, + 0.0051360526122152805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.014202351681888103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087710037827492 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011087710037827492, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087710037827492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3339501917362213, + 0.09392233192920685, + 0.1787133663892746, + 0.04354829713702202, + 0.02255941927433014, + 0.033736348152160645, + 0.13771352171897888, + 0.06170424818992615, + 0.04100189730525017, + 0.05315040051937103 + ], + "credit_map": [ + 0.020003844052553177, + 0.0071505410596728325, + 0.1355811208486557, + 0.012833877466619015, + 0.007307954132556915, + 0.13768364489078522, + 0.005819785874336958, + 0.00756942480802536, + 0.010149247013032436, + 0.011113333515822887, + 0.018228676170110703, + 0.011755900457501411, + 0.011445578187704086, + 0.013048364780843258, + 0.010700142942368984, + 0.0074810851365327835, + 0.014298435300588608, + 0.0651707723736763, + 0.06464366614818573, + 0.011937837116420269, + 0.005767913069576025, + 0.005975795444101095, + 0.004485935438424349, + 0.006433007773011923, + 0.004008859395980835, + 0.006293446756899357, + 0.008258134126663208, + 0.007200754247605801, + 0.0037221380043774843, + 0.005614916328340769, + 0.0036180943716317415, + 0.004831952974200249, + 0.002627417678013444, + 0.003947766963392496, + 0.0038996159564703703, + 0.003768575843423605, + 0.001865147496573627, + 0.0019295758102089167, + 0.0021796380169689655, + 0.002341682091355324, + 0.003706869902089238, + 0.005817243829369545, + 0.008705844171345234, + 0.007234520744532347, + 0.0017040962120518088, + 0.0027414446230977774, + 0.00207155360840261, + 0.0017547737807035446, + 0.008029771968722343, + 0.05431064963340759, + 0.05462043359875679, + 0.006577633786946535, + 0.0028550433926284313, + 0.005781055428087711, + 0.0022238451056182384, + 0.0033150948584079742, + 0.01225225254893303, + 0.004738502204418182, + 0.01361377164721489, + 0.002981500467285514, + 0.003076930996030569, + 0.01924622245132923, + 0.0023177461698651314, + 0.003477324265986681, + 0.0033140783198177814, + 0.0012166710803285241, + 0.01479752454906702, + 0.002071578288450837, + 0.002036664867773652, + 0.014301764778792858, + 0.0016783344326540828, + 0.001585280755534768, + 0.00609479658305645, + 0.002848150907084346, + 0.016286099329590797, + 0.00596762727946043, + 0.0034285984002053738, + 0.013185523450374603, + 0.002450599567964673, + 0.0028890068642795086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.011087710037827492 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494154617190361 + ], + "opposite_logprob_tokens": [ + -2.9299416542053223 + ], + "logprob_reward": -0.05494154617190361, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494154617190361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026570016518235207, + 0.021998604759573936, + 0.052613403648138046, + 0.022099873051047325, + 0.01669403165578842, + 0.05396762862801552, + 0.6804008483886719, + 0.08841689676046371, + 0.005655190907418728, + 0.031583480536937714 + ], + "credit_map": [ + 0.002595247235149145, + 0.0024524214677512646, + 0.007254117634147406, + 0.0022914153523743153, + 0.002336672740057111, + 0.006624915171414614, + 0.0018422295106574893, + 0.001172998920083046, + 0.0023298528976738453, + 0.0025527409743517637, + 0.0035231609363108873, + 0.004236720968037844, + 0.0019453142303973436, + 0.002819306217133999, + 0.002171615371480584, + 0.002419892232865095, + 0.0042792921885848045, + 0.019411230459809303, + 0.01960264891386032, + 0.003600704250857234, + 0.0016750299837440252, + 0.0011607129126787186, + 0.0013792506651952863, + 0.0015045326435938478, + 0.0021893614903092384, + 0.0032181076239794493, + 0.0029038512147963047, + 0.0026898160576820374, + 0.002486469456925988, + 0.0035756397992372513, + 0.0021225649397820234, + 0.002914064098149538, + 0.0019941984210163355, + 0.0031468039378523827, + 0.0035626895260065794, + 0.0032455050386488438, + 0.0010014022700488567, + 0.0011875757481902838, + 0.0011226485949009657, + 0.0014332098653540015, + 0.005757228471338749, + 0.00971483625471592, + 0.013876769691705704, + 0.011905581690371037, + 0.00253221089951694, + 0.0038017663173377514, + 0.002765457611531019, + 0.0036137779243290424, + 0.03708428516983986, + 0.2792849838733673, + 0.2771233320236206, + 0.03172671049833298, + 0.01094312034547329, + 0.025606997311115265, + 0.009190852753818035, + 0.009440581314265728, + 0.016655417159199715, + 0.004922070074826479, + 0.023678001016378403, + 0.0031160376965999603, + 0.0032989580649882555, + 0.030014971271157265, + 0.0022714489605277777, + 0.004459993913769722, + 0.0006616276223212481, + 0.0004922869266010821, + 0.0014144397573545575, + 0.0006441567093133926, + 0.0005106926546432078, + 0.0008480232208967209, + 0.0005215405253693461, + 0.0005624233745038509, + 0.003757667727768421, + 0.0018584165954962373, + 0.009381919167935848, + 0.0041162786073982716, + 0.0015083267353475094, + 0.006974126677960157, + 0.0014654900878667831, + 0.0025212534237653017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.05494154617190361 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic slow-motion close-up captures the action from a slightly elevated angle as a dark blue six-sided die falls into the frame. Made of polished, slightly translucent acrylic, the die impacts the warm, wood-textured surface of the kitchen table with a sharp impact. Under the soft, realistic overhead kitchen light, it bounces erratically, spinning in mid-air, causing the white, concave dots to blur into fleeting streaks. The camera remains static, allowing dynamic movement to fill the scene as the die loses momentum, gliding across the polished wood before swaying to a final, tense moment. It settles decisively, completely at rest, shifting the feel of the scene from chaotic energy to stillness. The final frame fixates on the perfectly still die, its top surface clearly displaying the number six. The warm light casts soft, fleeting shadows alongside it and creates subtle highlights along its smooth edges, creating a quiet, defined atmosphere. High resolution.", + "seed": 21615, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000020/group_9/index_3799_seed_21615_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020675426907837392, + "vlm_logprob": -0.020675426907837392, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.020675426907837392, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026844296604394913 + ], + "opposite_logprob_tokens": [ + -4.026844501495361 + ], + "logprob_reward": -0.026844296604394913, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.026844296604394913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07259350270032883, + 0.03966662660241127, + 0.10962721705436707, + 0.08375462144613266, + 0.05412472411990166, + 0.11882022768259048, + 0.23702655732631683, + 0.21552418172359467, + 0.06075382977724075, + 0.008108445443212986 + ], + "credit_map": [ + 0.005035324953496456, + 0.009524716064333916, + 0.009765488095581532, + 0.009119205176830292, + 0.00883095059543848, + 0.009403963573276997, + 0.01124156266450882, + 0.00967229064553976, + 0.003930925391614437, + 0.003563484875485301, + 0.004737940151244402, + 0.005920594558119774, + 0.0044082715176045895, + 0.005532427225261927, + 0.006617893930524588, + 0.004955090116709471, + 0.010658187791705132, + 0.009589732624590397, + 0.01291543710976839, + 0.012979943305253983, + 0.01052104216068983, + 0.015831492841243744, + 0.02154036983847618, + 0.01559100765734911, + 0.007900484837591648, + 0.014510628767311573, + 0.008649204857647419, + 0.008043263107538223, + 0.005812885705381632, + 0.021832764148712158, + 0.006883418653160334, + 0.010121963918209076, + 0.005067851394414902, + 0.005884069483727217, + 0.006857350002974272, + 0.008759940974414349, + 0.005995582789182663, + 0.008686047978699207, + 0.008137119933962822, + 0.004736760165542364, + 0.012648306787014008, + 0.007575939409434795, + 0.013842620886862278, + 0.020296568050980568, + 0.013763431459665298, + 0.018841756507754326, + 0.01621844805777073, + 0.01563316583633423, + 0.02499186433851719, + 0.018769262358546257, + 0.028805505484342575, + 0.035674065351486206, + 0.02226051688194275, + 0.042184460908174515, + 0.03499235957860947, + 0.029348522424697876, + 0.023103736340999603, + 0.015738284215331078, + 0.03041253238916397, + 0.024459242820739746, + 0.02034270204603672, + 0.04095359891653061, + 0.028133630752563477, + 0.03238046169281006, + 0.008021355606615543, + 0.004536783322691917, + 0.006455155089497566, + 0.008688344620168209, + 0.007510682102292776, + 0.007909353822469711, + 0.009258365258574486, + 0.008373789489269257, + 0.0008475047652609646, + 0.0006354170618578792, + 0.001018781797029078, + 0.0011640670709311962, + 0.000998009811155498, + 0.0008496764348819852, + 0.0009845305467024446, + 0.0016104583628475666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.026844296604394913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009808440692722797 + ], + "opposite_logprob_tokens": [ + -4.634808540344238 + ], + "logprob_reward": -0.009808440692722797, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009808440692722797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08443102240562439, + 0.09474489092826843, + 0.19523562490940094, + 0.125420480966568, + 0.06204359978437424, + 0.03929959982633591, + 0.1641150563955307, + 0.12397921085357666, + 0.10110318660736084, + 0.009627312421798706 + ], + "credit_map": [ + 0.008065575733780861, + 0.01143805030733347, + 0.009685560129582882, + 0.0073873684741556644, + 0.011678552255034447, + 0.01350414752960205, + 0.012022619135677814, + 0.010649153962731361, + 0.009267068468034267, + 0.008839171379804611, + 0.011167805641889572, + 0.011735145002603531, + 0.013421394862234592, + 0.010336051695048809, + 0.015724431723356247, + 0.014253819361329079, + 0.015339270234107971, + 0.01864267885684967, + 0.020202359184622765, + 0.02199261076748371, + 0.01786765456199646, + 0.032334696501493454, + 0.038413822650909424, + 0.03044254146516323, + 0.011092170141637325, + 0.012144297361373901, + 0.018859151750802994, + 0.022791394963860512, + 0.01140694972127676, + 0.012678021565079689, + 0.024037694558501244, + 0.01241080928593874, + 0.007605418562889099, + 0.006024011876434088, + 0.007566925138235092, + 0.008862782269716263, + 0.006627045106142759, + 0.010876611806452274, + 0.008072507567703724, + 0.006408293731510639, + 0.006560539826750755, + 0.0028156184125691652, + 0.004514311905950308, + 0.004594941157847643, + 0.004244795069098473, + 0.007452433463186026, + 0.004647290334105492, + 0.00446966802701354, + 0.018086684867739677, + 0.011930733919143677, + 0.019919883459806442, + 0.02950894646346569, + 0.014979034662246704, + 0.02650943398475647, + 0.024256683886051178, + 0.018923666328191757, + 0.019848302006721497, + 0.010700386017560959, + 0.014672168530523777, + 0.013860203325748444, + 0.01050365250557661, + 0.017659807577729225, + 0.013202190399169922, + 0.023532502353191376, + 0.016039375215768814, + 0.008714082650840282, + 0.011288506910204887, + 0.014992641285061836, + 0.00898807868361473, + 0.015750816091895103, + 0.010515475645661354, + 0.01481421198695898, + 0.0015830672346055508, + 0.0006913180113770068, + 0.0010655553778633475, + 0.0010297161061316729, + 0.0008880644454620779, + 0.0017602185253053904, + 0.0010741029400378466, + 0.0015352697810158134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.009808440692722797 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079746186733246 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079746186733246, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079746186733246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13663749396800995, + 0.18683817982673645, + 0.16386942565441132, + 0.10452023148536682, + 0.09918756037950516, + 0.10379431396722794, + 0.07806815207004547, + 0.06699877232313156, + 0.04047823324799538, + 0.019607678055763245 + ], + "credit_map": [ + 0.009615792892873287, + 0.017422322183847427, + 0.018132289871573448, + 0.015891941264271736, + 0.01229129359126091, + 0.012708988972008228, + 0.014089732430875301, + 0.03648513928055763, + 0.015663735568523407, + 0.01774541102349758, + 0.0241082850843668, + 0.02923564240336418, + 0.02222532220184803, + 0.01599949039518833, + 0.03571756184101105, + 0.02614273875951767, + 0.011779451742768288, + 0.015625888481736183, + 0.01726803556084633, + 0.01921570859849453, + 0.01681237667798996, + 0.0228141937404871, + 0.03538069874048233, + 0.024973079562187195, + 0.007606813684105873, + 0.01077819149941206, + 0.01440612506121397, + 0.02054784819483757, + 0.010289071127772331, + 0.012253798544406891, + 0.01893549971282482, + 0.009702886454761028, + 0.01293612364679575, + 0.007404129486531019, + 0.012216121889650822, + 0.011295430362224579, + 0.008627437055110931, + 0.019954359158873558, + 0.01291772723197937, + 0.01383622270077467, + 0.010769589804112911, + 0.007746763527393341, + 0.01361453253775835, + 0.017053427174687386, + 0.01228413451462984, + 0.012614306062459946, + 0.014871856197714806, + 0.014839708805084229, + 0.009360800497233868, + 0.006049068178981543, + 0.010345451533794403, + 0.012972994707524776, + 0.006573658902198076, + 0.015440020710229874, + 0.009082197211682796, + 0.008243966847658157, + 0.008567891083657742, + 0.0056905015371739864, + 0.007461564615368843, + 0.010906125418841839, + 0.005622078664600849, + 0.0098939910531044, + 0.009359030053019524, + 0.009497582912445068, + 0.006383194588124752, + 0.0032076314091682434, + 0.004840710200369358, + 0.005187677219510078, + 0.004163029137998819, + 0.005678859539330006, + 0.005027967970818281, + 0.005989162717014551, + 0.002402234822511673, + 0.0013505261158570647, + 0.0028119313064962626, + 0.0025567635893821716, + 0.0017648764187470078, + 0.003356716362759471, + 0.0018740220693871379, + 0.0034906058572232723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.016079746186733246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556072324514389 + ], + "opposite_logprob_tokens": [ + -4.387556076049805 + ], + "logprob_reward": -0.012556072324514389, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556072324514389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13773247599601746, + 0.10424475371837616, + 0.12674836814403534, + 0.11409036815166473, + 0.05887388810515404, + 0.07032053172588348, + 0.1170327439904213, + 0.12741385400295258, + 0.11381665617227554, + 0.029726430773735046 + ], + "credit_map": [ + 0.015446737408638, + 0.02200229838490486, + 0.013608582317829132, + 0.012920696288347244, + 0.016844699159264565, + 0.025654001161456108, + 0.014834036119282246, + 0.01642141304910183, + 0.00902829971164465, + 0.011493563652038574, + 0.0165900606662035, + 0.019531384110450745, + 0.00957477930933237, + 0.013922943733632565, + 0.014079378917813301, + 0.010024331510066986, + 0.007626436185091734, + 0.01298197265714407, + 0.014722393825650215, + 0.013764061033725739, + 0.011902101337909698, + 0.02091217041015625, + 0.023349709808826447, + 0.021489504724740982, + 0.009874124079942703, + 0.018656106665730476, + 0.01318044401705265, + 0.013581236824393272, + 0.007392533123493195, + 0.027257435023784637, + 0.011194148100912571, + 0.012954327277839184, + 0.004686372820287943, + 0.0058042253367602825, + 0.007357303984463215, + 0.00969130452722311, + 0.00562727265059948, + 0.007046011742204428, + 0.011021200567483902, + 0.007640187628567219, + 0.006470679305493832, + 0.005145466886460781, + 0.00945717841386795, + 0.011846944689750671, + 0.00706800352782011, + 0.009563014842569828, + 0.010390056297183037, + 0.010379189625382423, + 0.011166328564286232, + 0.009014695882797241, + 0.01585046574473381, + 0.01596788316965103, + 0.011947893537580967, + 0.017735036090016365, + 0.01571808010339737, + 0.019632350653409958, + 0.018193798139691353, + 0.009544354863464832, + 0.01513324398547411, + 0.0163982305675745, + 0.0131999421864748, + 0.018324101343750954, + 0.014488700777292252, + 0.022131457924842834, + 0.012045166455209255, + 0.008180147036910057, + 0.014680997468531132, + 0.017945801839232445, + 0.010155205614864826, + 0.020985782146453857, + 0.015310877934098244, + 0.014512656256556511, + 0.00398936215788126, + 0.002771195024251938, + 0.003999548032879829, + 0.0036704707890748978, + 0.0034531860146671534, + 0.0031467131339013577, + 0.003928739111870527, + 0.004767212085425854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.012556072324514389 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark blue six-sided die made of polished, slightly translucent acrylic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the die fall into the frame and impact a warm, wood-textured kitchen table surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the die bounce erratically, spin in mid-air, and glide across the table before coming to a complete rest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, realistic overhead kitchen light, casting soft shadows and highlights on the die?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot, capturing the die's movement from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808857873082161 + ], + "opposite_logprob_tokens": [ + -3.28808856010437 + ], + "logprob_reward": -0.03808857873082161, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808857873082161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07629024237394333, + 0.06953587383031845, + 0.1712881624698639, + 0.07300766557455063, + 0.06659666448831558, + 0.06288361549377441, + 0.2106378972530365, + 0.18395870923995972, + 0.07411735504865646, + 0.011683803983032703 + ], + "credit_map": [ + 0.0069357664324343204, + 0.010830859653651714, + 0.008818701840937138, + 0.008272986859083176, + 0.008149963803589344, + 0.011023860424757004, + 0.0052801985293626785, + 0.01697789877653122, + 0.006337348371744156, + 0.007412085309624672, + 0.010773503221571445, + 0.012726095505058765, + 0.007330806460231543, + 0.007993508130311966, + 0.009225817397236824, + 0.007736704312264919, + 0.014902440831065178, + 0.015474975109100342, + 0.019109003245830536, + 0.01984909176826477, + 0.015885649248957634, + 0.02772928774356842, + 0.032985277473926544, + 0.02535242773592472, + 0.006512773223221302, + 0.010704997926950455, + 0.009018992073833942, + 0.009339822456240654, + 0.005387983750551939, + 0.015436073765158653, + 0.008185174316167831, + 0.00842184666544199, + 0.004782239906489849, + 0.006448318250477314, + 0.009203355759382248, + 0.01301434077322483, + 0.006643411237746477, + 0.00869572814553976, + 0.010832836851477623, + 0.006976431235671043, + 0.006874597165733576, + 0.003697391599416733, + 0.006296392530202866, + 0.01073307916522026, + 0.0072435480542480946, + 0.011451194062829018, + 0.009154162369668484, + 0.00743324588984251, + 0.021838942542672157, + 0.01610834337770939, + 0.025516221299767494, + 0.0351380817592144, + 0.019604118540883064, + 0.036347512155771255, + 0.03095400147140026, + 0.02513067051768303, + 0.023124897852540016, + 0.012910331599414349, + 0.024695424363017082, + 0.02096501737833023, + 0.01713152416050434, + 0.03154410049319267, + 0.026944603770971298, + 0.026642808690667152, + 0.012084259651601315, + 0.006257692351937294, + 0.008350403979420662, + 0.009569807909429073, + 0.007909168489277363, + 0.010242036543786526, + 0.009777111932635307, + 0.00992687325924635, + 0.0014366194372996688, + 0.0008203188772313297, + 0.0013879397884011269, + 0.0011842858511954546, + 0.000950935180298984, + 0.002584770554676652, + 0.001321166753768921, + 0.0019977677147835493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.03808857873082161 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_0/index_tensor([7324, 7324])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_0/index_tensor([7324, 7324])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c650c49d33b1710d033d00530a3fecc9e8af39e4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_0/index_tensor([7324, 7324])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, first-person perspective shot looks directly down upon a massive architectural blueprint, spread across a dark, matte drafting table. The design's intricate, sharp lines represent complex structures and pathways, glowing a brilliant, inner cyan against the deep indigo paper. This perfect digital precision dissolves as the luminescent ink begins to accumulate, forming shimmering beads of liquid directly onto the glowing carrier lines. Then, slowly and relentlessly, these droplets of pure light begin to leak outward, smudged from their defined paths as the glowing liquid bleeds across the surface of the paper, smudged the clear geometric patterns into a soft, ethereal blur, and coalesced at intersections to create even brighter nodes of light before spilling out in fragile, uncontrolled tendrils. This self-luminous ink casts a cool, otherworldly glow, contrasting starkly with the warm, dim ambient light of the surrounding, unseen room, creating a focused, captivating atmosphere of magical transformation.", + "seed": 9609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_0/index_7324_seed_9609_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05693302601575852, + "vlm_logprob": -0.05693302601575852, + "vlm_yes_no_margin": 3.200000023841858, + "vlm_token_credit": -0.05693302601575852, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02204333432018757 + ], + "opposite_logprob_tokens": [ + -4.147043228149414 + ], + "logprob_reward": -0.02204333432018757, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02204333432018757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1894053816795349, + 0.03381539508700371, + 0.023316476494073868, + 0.021455850452184677, + 0.5293787717819214, + 0.02582155354321003, + 0.015554002486169338, + 0.011988949961960316, + 0.06144774705171585, + 0.08781586587429047 + ], + "credit_map": [ + 0.015591557137668133, + 0.03151895850896835, + 0.022581500932574272, + 0.03381675109267235, + 0.012868550606071949, + 0.010369513183832169, + 0.03015729784965515, + 0.03250125050544739, + 0.004247425124049187, + 0.005239398218691349, + 0.004138907417654991, + 0.004194903653115034, + 0.002375063020735979, + 0.0030879632104188204, + 0.005988508928567171, + 0.004543224815279245, + 0.0024490817449986935, + 0.0024990581441670656, + 0.0026117293164134026, + 0.00376619934104383, + 0.0012885044561699033, + 0.0019121520454064012, + 0.003575708717107773, + 0.005214042030274868, + 0.0022287273313850164, + 0.0039779250510036945, + 0.002486773068085313, + 0.003437540028244257, + 0.001858413452282548, + 0.0016892470885068178, + 0.0032865935936570168, + 0.0024906310718506575, + 0.01948549412190914, + 0.015907317399978638, + 0.23333360254764557, + 0.07917119562625885, + 0.007793179713189602, + 0.013039822690188885, + 0.026638781651854515, + 0.13400936126708984, + 0.00334466272033751, + 0.006668172311037779, + 0.0035803341306746006, + 0.0035056930501013994, + 0.002329546259716153, + 0.0018295948393642902, + 0.0024113969411700964, + 0.0021521546877920628, + 0.0016260565025731921, + 0.002667826134711504, + 0.0023170551285147667, + 0.002558079082518816, + 0.001586325466632843, + 0.0015061808517202735, + 0.001682245172560215, + 0.0016102339141070843, + 0.0015332202892750502, + 0.0021280685905367136, + 0.0017765972297638655, + 0.0008336787577718496, + 0.0013783444883301854, + 0.0018309296574443579, + 0.0010553710162639618, + 0.001452740398235619, + 0.00659656897187233, + 0.009731877595186234, + 0.004459324758499861, + 0.0063107009045779705, + 0.011946771293878555, + 0.009193514473736286, + 0.006219292059540749, + 0.006989699322730303, + 0.004906462039798498, + 0.007146965712308884, + 0.004811454098671675, + 0.007410619407892227, + 0.01072327047586441, + 0.008502244018018246, + 0.022349275648593903, + 0.021965578198432922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.02204333432018757 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016089366748929024 + ], + "opposite_logprob_tokens": [ + -4.14108943939209 + ], + "logprob_reward": -0.016089366748929024, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016089366748929024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19852185249328613, + 0.039093442261219025, + 0.027731547132134438, + 0.01990082487463951, + 0.5387564301490784, + 0.022213442251086235, + 0.018237510696053505, + 0.050628673285245895, + 0.04941786825656891, + 0.03549845516681671 + ], + "credit_map": [ + 0.01974523812532425, + 0.037896208465099335, + 0.02399837225675583, + 0.031148023903369904, + 0.015042977407574654, + 0.01495263073593378, + 0.03135644271969795, + 0.024381931871175766, + 0.003459834959357977, + 0.00600549578666687, + 0.00513180997222662, + 0.004951485898345709, + 0.0030100876465439796, + 0.0035728018265217543, + 0.006164692807942629, + 0.006797224748879671, + 0.0019193800399079919, + 0.004249054007232189, + 0.003337664296850562, + 0.003729032352566719, + 0.0021250310819596052, + 0.0024843954015523195, + 0.004928835667669773, + 0.004958153236657381, + 0.001967009389773011, + 0.002265677321702242, + 0.0027171557303518057, + 0.004029604606330395, + 0.0013955909525975585, + 0.001621308270841837, + 0.0028646928258240223, + 0.003039784263819456, + 0.019831104204058647, + 0.014908197335898876, + 0.239114448428154, + 0.08199075609445572, + 0.008097385056316853, + 0.012932040728628635, + 0.02991955541074276, + 0.13196294009685516, + 0.0027131952811032534, + 0.005291204433888197, + 0.003751893062144518, + 0.0026263443287461996, + 0.002279153559356928, + 0.0017872574971988797, + 0.002138752955943346, + 0.0016256371745839715, + 0.0023013383615761995, + 0.0035139776300638914, + 0.002230497542768717, + 0.0029928479343652725, + 0.001446178532205522, + 0.002085198415443301, + 0.0020810116548091173, + 0.0015864563174545765, + 0.003866947488859296, + 0.006769907660782337, + 0.008541575632989407, + 0.003989556338638067, + 0.008436020463705063, + 0.007091520819813013, + 0.003828703425824642, + 0.008104436099529266, + 0.005663429852575064, + 0.007389356382191181, + 0.0036800468806177378, + 0.006008145399391651, + 0.008998283185064793, + 0.006673869211226702, + 0.005281078163534403, + 0.005723653361201286, + 0.0025687769521027803, + 0.003575640032067895, + 0.0023284642957150936, + 0.002944523235782981, + 0.005696638487279415, + 0.004593353718519211, + 0.006450559012591839, + 0.007340490818023682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.016089366748929024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428166776895523 + ], + "opposite_logprob_tokens": [ + -2.0178167819976807 + ], + "logprob_reward": -0.1428166776895523, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428166776895523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048640910536050797, + 0.045989010483026505, + 0.026030931621789932, + 0.018014539033174515, + 0.7557886838912964, + 0.035685330629348755, + 0.020940763875842094, + 0.018916694447398186, + 0.006490248721092939, + 0.023502860218286514 + ], + "credit_map": [ + 0.007778072729706764, + 0.007671517785638571, + 0.005618126131594181, + 0.0075744567438960075, + 0.0040753991343081, + 0.004023054149001837, + 0.005998144391924143, + 0.005902138538658619, + 0.004484415985643864, + 0.010042555630207062, + 0.004814478103071451, + 0.0057664006017148495, + 0.0030211026314646006, + 0.0032721511088311672, + 0.006926802918314934, + 0.007661104667931795, + 0.003511951304972172, + 0.004774794913828373, + 0.002367588924244046, + 0.003156979102641344, + 0.002418479649350047, + 0.0020457536447793245, + 0.0033853943459689617, + 0.0043699913658201694, + 0.001516783726401627, + 0.002893280703574419, + 0.001958837267011404, + 0.0022112943697720766, + 0.0014682826586067677, + 0.001425332622602582, + 0.0031384651083499193, + 0.0034022629261016846, + 0.0282135047018528, + 0.020847825333476067, + 0.33400845527648926, + 0.11595801264047623, + 0.012027643620967865, + 0.01855255849659443, + 0.04021850600838661, + 0.18596220016479492, + 0.004669909831136465, + 0.009470769204199314, + 0.006095502059906721, + 0.004212858155369759, + 0.003135201521217823, + 0.0023931509349495173, + 0.002927070949226618, + 0.0027808721642941236, + 0.002212878782302141, + 0.004008880350738764, + 0.0027703698724508286, + 0.002643620828166604, + 0.0020013866014778614, + 0.002134061185643077, + 0.0025359566789120436, + 0.0026336105074733496, + 0.0022163798566907644, + 0.003931573126465082, + 0.0033981595188379288, + 0.001701423549093306, + 0.0017238399013876915, + 0.0018575701396912336, + 0.001755774486809969, + 0.0023319756146520376, + 0.0006051256787031889, + 0.0010676005622372031, + 0.0005383216193877161, + 0.0007288184715434909, + 0.0009520722087472677, + 0.0008123782463371754, + 0.0006947102956473827, + 0.0010912223951891065, + 0.0015332637121900916, + 0.003145799273625016, + 0.0015255501493811607, + 0.002063656924292445, + 0.0030858484096825123, + 0.0023421309888362885, + 0.005369897931814194, + 0.004436712712049484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.1428166776895523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499965697526932 + ], + "opposite_logprob_tokens": [ + -2.929999589920044 + ], + "logprob_reward": -0.05499965697526932, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499965697526932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055822085589170456, + 0.041825003921985626, + 0.029249906539916992, + 0.02327071875333786, + 0.755653440952301, + 0.02479996532201767, + 0.01634378358721733, + 0.021532008424401283, + 0.010522225871682167, + 0.020980948582291603 + ], + "credit_map": [ + 0.005423842463642359, + 0.009062372148036957, + 0.005683889612555504, + 0.006595343351364136, + 0.005716920364648104, + 0.007839006371796131, + 0.00801965780556202, + 0.007481047417968512, + 0.0037232362665235996, + 0.008667506277561188, + 0.004275804851204157, + 0.004500118549913168, + 0.003748327260836959, + 0.004953429102897644, + 0.004835805390030146, + 0.007120775990188122, + 0.004537960048764944, + 0.0035321100149303675, + 0.0028283351566642523, + 0.003912015818059444, + 0.0019765247125178576, + 0.002358830301091075, + 0.004427597392350435, + 0.005676534026861191, + 0.003969268407672644, + 0.0026078857481479645, + 0.002720253774896264, + 0.0039636059664189816, + 0.001841917634010315, + 0.0027656478341668844, + 0.0025963226798921824, + 0.002805817173793912, + 0.027943989261984825, + 0.020745432004332542, + 0.3355887532234192, + 0.11606039851903915, + 0.011603360064327717, + 0.018401814624667168, + 0.04102827236056328, + 0.18428140878677368, + 0.0033321441151201725, + 0.006542415823787451, + 0.003401012159883976, + 0.003050216007977724, + 0.0026257741264998913, + 0.0015965532511472702, + 0.002147363731637597, + 0.002104486571624875, + 0.00158780871424824, + 0.00289658410474658, + 0.002665364183485508, + 0.0018656185129657388, + 0.0015112581895664334, + 0.0016071487916633487, + 0.0017314258730039, + 0.00247857510112226, + 0.0024461401626467705, + 0.003026288468390703, + 0.002598096150904894, + 0.0018415403319522738, + 0.0038438832852989435, + 0.0028120751958340406, + 0.002091826405376196, + 0.002872159006074071, + 0.0013445391086861491, + 0.0018301488598808646, + 0.0011793321464210749, + 0.0013759455177932978, + 0.0009661443764343858, + 0.0014530908083543181, + 0.0013047854881733656, + 0.0010682395659387112, + 0.00239674374461174, + 0.0037118252366781235, + 0.0018935702973976731, + 0.0022799763828516006, + 0.002959438366815448, + 0.0024664378724992275, + 0.002971160924062133, + 0.002301797503605485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.05499965697526932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048716094344854355 + ], + "opposite_logprob_tokens": [ + -3.0487160682678223 + ], + "logprob_reward": -0.048716094344854355, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048716094344854355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22086483240127563, + 0.08512218296527863, + 0.041357412934303284, + 0.03706106171011925, + 0.49256432056427, + 0.02088387869298458, + 0.009588785469532013, + 0.012507646344602108, + 0.017466951161623, + 0.06258288770914078 + ], + "credit_map": [ + 0.02437940239906311, + 0.026023179292678833, + 0.017370956018567085, + 0.026672884821891785, + 0.02506239339709282, + 0.035134490579366684, + 0.03138267248868942, + 0.03483887016773224, + 0.009928902611136436, + 0.008170412853360176, + 0.00736643560230732, + 0.009224215522408485, + 0.009011573158204556, + 0.013261700980365276, + 0.013048231601715088, + 0.015110718086361885, + 0.004314027726650238, + 0.006484734360128641, + 0.003303660312667489, + 0.004523677751421928, + 0.005227984860539436, + 0.006543096620589495, + 0.005320601165294647, + 0.005639632698148489, + 0.003667864017188549, + 0.006254729349166155, + 0.003049192251637578, + 0.0035069556906819344, + 0.004253777675330639, + 0.005493059754371643, + 0.0050641330890357494, + 0.005771354306489229, + 0.018612094223499298, + 0.014156841672956944, + 0.21853211522102356, + 0.07612977921962738, + 0.00774286687374115, + 0.011895963922142982, + 0.028417911380529404, + 0.1170768067240715, + 0.0017213887767866254, + 0.0026972019113600254, + 0.006357943173497915, + 0.002467283047735691, + 0.0011921522673219442, + 0.0012575725559145212, + 0.0014040449168533087, + 0.003786293789744377, + 0.0009367857710458338, + 0.0010483134537935257, + 0.001730866264551878, + 0.001423651003278792, + 0.0008664667257107794, + 0.0011120610870420933, + 0.0014891254249960184, + 0.0009815163211897016, + 0.0012739767553284764, + 0.002309402683749795, + 0.0027358243241906166, + 0.0014814555179327726, + 0.0009341083350591362, + 0.000905188440810889, + 0.0013566307025030255, + 0.0015110602835193276, + 0.001408491050824523, + 0.0035907647106796503, + 0.0017758896574378014, + 0.0017559207044541836, + 0.0022279727272689342, + 0.002166285878047347, + 0.0016433349810540676, + 0.0028982926160097122, + 0.003717867424711585, + 0.0063356561586260796, + 0.0030608114320784807, + 0.005698340479284525, + 0.006561589427292347, + 0.005268312059342861, + 0.017138058319687843, + 0.014802252873778343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.048716094344854355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, first-person perspective shot looks directly down upon a massive architectural blueprint, spread across a dark, matte drafting table. The design's intricate, sharp lines represent complex structures and pathways, glowing a brilliant, inner cyan against the deep indigo paper. This perfect digital precision dissolves as the luminescent ink begins to accumulate, forming shimmering beads of liquid directly onto the glowing carrier lines. Then, slowly and relentlessly, these droplets of pure light begin to leak outward, smudged from their defined paths as the glowing liquid bleeds across the surface of the paper, smudged the clear geometric patterns into a soft, ethereal blur, and coalesced at intersections to create even brighter nodes of light before spilling out in fragile, uncontrolled tendrils. This self-luminous ink casts a cool, otherworldly glow, contrasting starkly with the warm, dim ambient light of the surrounding, unseen room, creating a focused, captivating atmosphere of magical transformation.", + "seed": 9610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_0/index_7324_seed_9610_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.051286513730883596, + "vlm_logprob": -0.051286513730883596, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.051286513730883596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041308097541332245 + ], + "opposite_logprob_tokens": [ + -3.4163081645965576 + ], + "logprob_reward": -0.041308097541332245, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.041308097541332245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07317239791154861, + 0.027727780863642693, + 0.047794654965400696, + 0.06872457265853882, + 0.24514570832252502, + 0.26963338255882263, + 0.04573769122362137, + 0.11665406823158264, + 0.07777772843837738, + 0.027632039040327072 + ], + "credit_map": [ + 0.00525629660114646, + 0.005582538433372974, + 0.008865578100085258, + 0.010233504697680473, + 0.01729143224656582, + 0.006719064433127642, + 0.009118977934122086, + 0.010105008259415627, + 0.004044528119266033, + 0.003576327580958605, + 0.003113223472610116, + 0.0031918224412947893, + 0.0028998020570725203, + 0.004861480090767145, + 0.0027383719570934772, + 0.0033022246789187193, + 0.0071716816164553165, + 0.00847617257386446, + 0.0062600187957286835, + 0.005103882402181625, + 0.0063368272967636585, + 0.006487122271209955, + 0.003946478944271803, + 0.0040124692022800446, + 0.00986070279031992, + 0.0070216781459748745, + 0.010871152393519878, + 0.007412448991090059, + 0.008899375796318054, + 0.009617466479539871, + 0.007524292916059494, + 0.007517451886087656, + 0.02550344169139862, + 0.026425400748848915, + 0.024492574855685234, + 0.019567478448152542, + 0.03347555175423622, + 0.06850491464138031, + 0.021086011081933975, + 0.0260903462767601, + 0.020369376987218857, + 0.022283188998699188, + 0.03185547888278961, + 0.042377643287181854, + 0.036462362855672836, + 0.027727559208869934, + 0.05365562438964844, + 0.03490215539932251, + 0.004438482224941254, + 0.005248906556516886, + 0.006942455191165209, + 0.007553529925644398, + 0.0041561187244951725, + 0.006894128862768412, + 0.004982547368854284, + 0.005521520972251892, + 0.02897028625011444, + 0.007922191172838211, + 0.027633052319288254, + 0.011521982960402966, + 0.011547544039785862, + 0.006254938896745443, + 0.01323770172894001, + 0.009566363878548145, + 0.010822712443768978, + 0.00931363832205534, + 0.0073951855301856995, + 0.009589153341948986, + 0.008806077763438225, + 0.009628260508179665, + 0.010487447492778301, + 0.01173525769263506, + 0.003919650334864855, + 0.0033738056663423777, + 0.0015364562859758735, + 0.001303178956732154, + 0.003970250021666288, + 0.007675213739275932, + 0.0019976526964455843, + 0.003855830989778042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.041308097541332245 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026367442682385445 + ], + "opposite_logprob_tokens": [ + -3.651367425918579 + ], + "logprob_reward": -0.026367442682385445, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026367442682385445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2504616677761078, + 0.036643821746110916, + 0.034590255469083786, + 0.11830825358629227, + 0.09123671799898148, + 0.20917311310768127, + 0.03721495717763901, + 0.13181374967098236, + 0.06659785658121109, + 0.023959597572684288 + ], + "credit_map": [ + 0.019334087148308754, + 0.019358910620212555, + 0.02792186476290226, + 0.03276460990309715, + 0.05239533632993698, + 0.023007169365882874, + 0.03780478611588478, + 0.037874892354011536, + 0.003656848333775997, + 0.00421480368822813, + 0.005115778651088476, + 0.005904531572014093, + 0.0044419062323868275, + 0.004098997917026281, + 0.004623192362487316, + 0.004587764386087656, + 0.004838329739868641, + 0.005344214383512735, + 0.003634885884821415, + 0.004894070792943239, + 0.005435050465166569, + 0.004322963301092386, + 0.003032510634511709, + 0.003088229801505804, + 0.01795242913067341, + 0.011454952880740166, + 0.019949261099100113, + 0.010309145785868168, + 0.014725161716341972, + 0.019608864560723305, + 0.012431396171450615, + 0.011877039447426796, + 0.00961229670792818, + 0.010723776184022427, + 0.009025245904922485, + 0.0063514686189591885, + 0.011882791295647621, + 0.02600688301026821, + 0.006888349540531635, + 0.010745907202363014, + 0.01375952921807766, + 0.01693541742861271, + 0.021898362785577774, + 0.029195662587881088, + 0.034774262458086014, + 0.030538376420736313, + 0.037319499999284744, + 0.024752004072070122, + 0.0030483293812721968, + 0.003239487763494253, + 0.008433184586465359, + 0.009384451434016228, + 0.0028244212735444307, + 0.0033249848056584597, + 0.004129244480282068, + 0.002830854384228587, + 0.028828302398324013, + 0.008321499451994896, + 0.030405327677726746, + 0.014333929866552353, + 0.01443614810705185, + 0.00919592846184969, + 0.014779089950025082, + 0.011513526551425457, + 0.007976101711392403, + 0.007319261785596609, + 0.007075084373354912, + 0.008991990238428116, + 0.007954441010951996, + 0.007826306857168674, + 0.009525788016617298, + 0.009928880259394646, + 0.004867251496762037, + 0.003170185023918748, + 0.0026608076877892017, + 0.0027922766748815775, + 0.0020722297485917807, + 0.0021355811040848494, + 0.0034747403115034103, + 0.0027865252923220396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.026367442682385445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002748548984528 + ], + "opposite_logprob_tokens": [ + -2.6950275897979736 + ], + "logprob_reward": -0.07002748548984528, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002748548984528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11232704669237137, + 0.06605461239814758, + 0.11034059524536133, + 0.05894243344664574, + 0.0415329709649086, + 0.1095975935459137, + 0.09265923500061035, + 0.24465808272361755, + 0.11290038377046585, + 0.05098702386021614 + ], + "credit_map": [ + 0.010751872323453426, + 0.014026585966348648, + 0.013005184940993786, + 0.01760771870613098, + 0.01717524044215679, + 0.011365320533514023, + 0.014351112768054008, + 0.014044014737010002, + 0.009419098496437073, + 0.007575801573693752, + 0.007546401582658291, + 0.012760847806930542, + 0.008834293112158775, + 0.00788851361721754, + 0.006518653128296137, + 0.005511005874723196, + 0.017539672553539276, + 0.014923960901796818, + 0.01295340433716774, + 0.009679686278104782, + 0.013195554725825787, + 0.02215844951570034, + 0.008997472003102303, + 0.010892407968640327, + 0.01169655192643404, + 0.008630783297121525, + 0.009412623941898346, + 0.005358152557164431, + 0.005136293359100819, + 0.007114031817764044, + 0.005443165078759193, + 0.006150838453322649, + 0.004181958734989166, + 0.005711071193218231, + 0.0042646946385502815, + 0.003979088272899389, + 0.0050354753620922565, + 0.008721667341887951, + 0.004111424088478088, + 0.0055275955237448215, + 0.00844135507941246, + 0.010452961549162865, + 0.009322283789515495, + 0.01228887028992176, + 0.022642092779278755, + 0.021804289892315865, + 0.01419390644878149, + 0.010451843962073326, + 0.017043137922883034, + 0.009573652409017086, + 0.01340764481574297, + 0.010884293355047703, + 0.006414782255887985, + 0.014147154055535793, + 0.011228254064917564, + 0.009960325434803963, + 0.05329233407974243, + 0.01758587546646595, + 0.0650131031870842, + 0.023627547547221184, + 0.0191304050385952, + 0.01824042946100235, + 0.02663220837712288, + 0.0211362112313509, + 0.013789664953947067, + 0.011922399513423443, + 0.010976859368383884, + 0.013442994095385075, + 0.012801111675798893, + 0.014408169314265251, + 0.014984130859375, + 0.020575059577822685, + 0.0063424245454370975, + 0.007245032582432032, + 0.006770579144358635, + 0.005917507689446211, + 0.004487455356866121, + 0.007024249993264675, + 0.006615666206926107, + 0.006584115792065859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.07002748548984528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700361579656601 + ], + "opposite_logprob_tokens": [ + -2.6950361728668213 + ], + "logprob_reward": -0.0700361579656601, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700361579656601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13602183759212494, + 0.051343489438295364, + 0.07520753890275955, + 0.057607535272836685, + 0.19053882360458374, + 0.1452789604663849, + 0.056642744690179825, + 0.18251587450504303, + 0.05676182731986046, + 0.04808133840560913 + ], + "credit_map": [ + 0.019197911024093628, + 0.012941500172019005, + 0.021848447620868683, + 0.015822634100914, + 0.015455749817192554, + 0.0126999756321311, + 0.01703367754817009, + 0.021021945402026176, + 0.007248181849718094, + 0.00482176523655653, + 0.0061185602098703384, + 0.0071555026806890965, + 0.0062739793211221695, + 0.008771948516368866, + 0.005859754513949156, + 0.005093797575682402, + 0.009369355626404285, + 0.0087366234511137, + 0.006865072529762983, + 0.008118191733956337, + 0.011528104543685913, + 0.016423694789409637, + 0.006513929460197687, + 0.007652571424841881, + 0.007419345434755087, + 0.006089466158300638, + 0.008439137600362301, + 0.00914844125509262, + 0.006632823962718248, + 0.006558889988809824, + 0.005948870442807674, + 0.007370558567345142, + 0.019706126302480698, + 0.023753272369503975, + 0.017491798847913742, + 0.013636702671647072, + 0.027176495641469955, + 0.054187674075365067, + 0.013766036368906498, + 0.020820731297135353, + 0.01022957731038332, + 0.011827726848423481, + 0.016263585537672043, + 0.023935528472065926, + 0.020818844437599182, + 0.018176307901740074, + 0.027092844247817993, + 0.01693454384803772, + 0.010766868479549885, + 0.004820357076823711, + 0.012167681939899921, + 0.00585837522521615, + 0.0048898980021476746, + 0.004316059872508049, + 0.00776375737041235, + 0.006059746257960796, + 0.030612517148256302, + 0.016185954213142395, + 0.03134249895811081, + 0.028351889923214912, + 0.02051875926554203, + 0.01465783640742302, + 0.02357184886932373, + 0.01727456785738468, + 0.0075227199122309685, + 0.005442386493086815, + 0.006152078043669462, + 0.006870224606245756, + 0.007522997446358204, + 0.00578351691365242, + 0.008002347312867641, + 0.009465556591749191, + 0.0072411419823765755, + 0.006673756986856461, + 0.005645096302032471, + 0.005591381806880236, + 0.005498433485627174, + 0.006292794831097126, + 0.005843294318765402, + 0.005295441020280123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.0700361579656601 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869338497519493 + ], + "opposite_logprob_tokens": [ + -3.0486934185028076 + ], + "logprob_reward": -0.04869338497519493, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869338497519493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1343446969985962, + 0.047280918806791306, + 0.08097232133150101, + 0.06261910498142242, + 0.10278552770614624, + 0.13771000504493713, + 0.05761098116636276, + 0.10724158585071564, + 0.18276762962341309, + 0.08666721731424332 + ], + "credit_map": [ + 0.019756179302930832, + 0.013335351832211018, + 0.016393454745411873, + 0.015431372448801994, + 0.018226828426122665, + 0.018549539148807526, + 0.016700828447937965, + 0.015951164066791534, + 0.006094710435718298, + 0.008862783201038837, + 0.006766623351722956, + 0.005503730848431587, + 0.004804104100912809, + 0.005842838902026415, + 0.004245416726917028, + 0.005160712171345949, + 0.012039173394441605, + 0.012880456633865833, + 0.010012582875788212, + 0.009769923985004425, + 0.008775030262768269, + 0.009299375116825104, + 0.00902971439063549, + 0.009166077710688114, + 0.011205310933291912, + 0.00624019606038928, + 0.011667385697364807, + 0.0044480799697339535, + 0.006476701702922583, + 0.011028004810214043, + 0.005616568494588137, + 0.005936861038208008, + 0.009505408816039562, + 0.013588023371994495, + 0.009003440849483013, + 0.00861336663365364, + 0.014032288454473019, + 0.027514437213540077, + 0.00937595684081316, + 0.011152613908052444, + 0.010712440125644207, + 0.010501018725335598, + 0.016391240060329437, + 0.02187490463256836, + 0.020660432055592537, + 0.014798830263316631, + 0.02599981427192688, + 0.01677132397890091, + 0.006148165091872215, + 0.007140720263123512, + 0.00853651948273182, + 0.010489356704056263, + 0.004004095681011677, + 0.006961142644286156, + 0.0075561124831438065, + 0.006774870678782463, + 0.014318724162876606, + 0.007714168634265661, + 0.010622691363096237, + 0.01484625693410635, + 0.013996575027704239, + 0.01347248163074255, + 0.014460287056863308, + 0.017810417339205742, + 0.020588651299476624, + 0.0176493301987648, + 0.026770712807774544, + 0.02582581713795662, + 0.01809115707874298, + 0.02116028405725956, + 0.025145389139652252, + 0.027536295354366302, + 0.014069455675780773, + 0.014149390161037445, + 0.009894623421132565, + 0.006442708894610405, + 0.009275349788367748, + 0.015519140288233757, + 0.007234379183501005, + 0.010082168504595757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.04869338497519493 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, first-person perspective shot looks directly down upon a massive architectural blueprint, spread across a dark, matte drafting table. The design's intricate, sharp lines represent complex structures and pathways, glowing a brilliant, inner cyan against the deep indigo paper. This perfect digital precision dissolves as the luminescent ink begins to accumulate, forming shimmering beads of liquid directly onto the glowing carrier lines. Then, slowly and relentlessly, these droplets of pure light begin to leak outward, smudged from their defined paths as the glowing liquid bleeds across the surface of the paper, smudged the clear geometric patterns into a soft, ethereal blur, and coalesced at intersections to create even brighter nodes of light before spilling out in fragile, uncontrolled tendrils. This self-luminous ink casts a cool, otherworldly glow, contrasting starkly with the warm, dim ambient light of the surrounding, unseen room, creating a focused, captivating atmosphere of magical transformation.", + "seed": 9611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_0/index_7324_seed_9611_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2074367344379425, + "vlm_logprob": -0.2074367344379425, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.2074367344379425, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0864500105381012 + ], + "opposite_logprob_tokens": [ + -2.5864500999450684 + ], + "logprob_reward": -0.0864500105381012, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0864500105381012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21066759526729584, + 0.032951224595308304, + 0.04084014892578125, + 0.3652181327342987, + 0.033515386283397675, + 0.09685095399618149, + 0.06782607734203339, + 0.04277148097753525, + 0.045109815895557404, + 0.06424913555383682 + ], + "credit_map": [ + 0.011501572094857693, + 0.018559623509645462, + 0.03849772736430168, + 0.022250263020396233, + 0.025127889588475227, + 0.019012408331036568, + 0.05014689639210701, + 0.025571241974830627, + 0.0019869208335876465, + 0.003954901359975338, + 0.003436419414356351, + 0.004496681038290262, + 0.003897987073287368, + 0.0028191322926431894, + 0.005934514105319977, + 0.006424667313694954, + 0.003452819073572755, + 0.006830587983131409, + 0.0038025083485990763, + 0.006788064260035753, + 0.00534460786730051, + 0.0032478650100529194, + 0.006196700967848301, + 0.005176997277885675, + 0.04596701264381409, + 0.03505387529730797, + 0.050783757120370865, + 0.037469103932380676, + 0.0455315001308918, + 0.053678859025239944, + 0.04321248084306717, + 0.05352157726883888, + 0.0034549166448414326, + 0.002528965473175049, + 0.0032385908998548985, + 0.0058622462674975395, + 0.003688626689836383, + 0.003984655253589153, + 0.004686771426349878, + 0.006070617586374283, + 0.015337052755057812, + 0.014346737414598465, + 0.013476342894136906, + 0.010783393867313862, + 0.009586005471646786, + 0.009558952413499355, + 0.012029372155666351, + 0.011733110062777996, + 0.0021974442061036825, + 0.00350356986746192, + 0.0044021280482411385, + 0.006458605639636517, + 0.0036498503759503365, + 0.004111751448363066, + 0.024731112644076347, + 0.018771616742014885, + 0.003664341988041997, + 0.0033523747697472572, + 0.01003717165440321, + 0.0027586589567363262, + 0.004532119259238243, + 0.0030092401430010796, + 0.011410093866288662, + 0.004007483366876841, + 0.0035014392342418432, + 0.00419090548530221, + 0.00511953653767705, + 0.0037626561243087053, + 0.00439664488658309, + 0.015291181392967701, + 0.0034306782763451338, + 0.0054167769849300385, + 0.0051754070445895195, + 0.006889841984957457, + 0.006591187324374914, + 0.01079305075109005, + 0.008723911829292774, + 0.008284362033009529, + 0.009419369511306286, + 0.008372007869184017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0864500105381012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908253908157349 + ], + "opposite_logprob_tokens": [ + -2.4640824794769287 + ], + "logprob_reward": -0.08908253908157349, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908253908157349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12398817390203476, + 0.05097300559282303, + 0.034931082278490067, + 0.4614781439304352, + 0.04366648569703102, + 0.033988289535045624, + 0.10371547192335129, + 0.027154577895998955, + 0.03460989147424698, + 0.08549486845731735 + ], + "credit_map": [ + 0.006802413612604141, + 0.012396380305290222, + 0.023048436269164085, + 0.013719303533434868, + 0.014906470663845539, + 0.012911184690892696, + 0.02651248686015606, + 0.013691497966647148, + 0.004163756035268307, + 0.004133973736315966, + 0.005429703276604414, + 0.005278581287711859, + 0.004990659654140472, + 0.003538657445460558, + 0.014065852388739586, + 0.009371825493872166, + 0.003269148524850607, + 0.00421650055795908, + 0.002957922639325261, + 0.0032042500097304583, + 0.0039587016217410564, + 0.002799244364723563, + 0.005830287467688322, + 0.008695028722286224, + 0.06058530509471893, + 0.044697266072034836, + 0.06287528574466705, + 0.04683540761470795, + 0.054465774446725845, + 0.07412011921405792, + 0.05483454093337059, + 0.06306444108486176, + 0.005651457235217094, + 0.005445562303066254, + 0.004870652221143246, + 0.004702620208263397, + 0.00796788465231657, + 0.00546983303502202, + 0.0051012830808758736, + 0.004457192495465279, + 0.006608220282942057, + 0.0043463860638439655, + 0.004844909068197012, + 0.0038837622851133347, + 0.0027370050083845854, + 0.0027720495127141476, + 0.0035755401477217674, + 0.005220416467636824, + 0.004150799475610256, + 0.004111508838832378, + 0.008097056299448013, + 0.01111595332622528, + 0.005627245642244816, + 0.004911792930215597, + 0.03842465206980705, + 0.02727646753191948, + 0.002075032563880086, + 0.002783755538985133, + 0.005025362595915794, + 0.0032124570570886135, + 0.002885931869968772, + 0.0022802366875112057, + 0.005494851619005203, + 0.0033969522919505835, + 0.0030328696593642235, + 0.002820921828970313, + 0.005087800789624453, + 0.0038674534298479557, + 0.0057153175584971905, + 0.0077154082246124744, + 0.003014965681359172, + 0.003355157095938921, + 0.008546593599021435, + 0.009984944947063923, + 0.009888329543173313, + 0.010841121897101402, + 0.012508565559983253, + 0.01206907257437706, + 0.012302793562412262, + 0.009353451430797577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.08908253908157349 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866965487599373 + ], + "opposite_logprob_tokens": [ + -3.0486695766448975 + ], + "logprob_reward": -0.04866965487599373, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866965487599373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17045481503009796, + 0.14269709587097168, + 0.05821561440825462, + 0.2627730071544647, + 0.029469996690750122, + 0.044926274567842484, + 0.0960983857512474, + 0.1100349873304367, + 0.0226745642721653, + 0.0626552551984787 + ], + "credit_map": [ + 0.007743270602077246, + 0.01119927130639553, + 0.05595315620303154, + 0.014729059301316738, + 0.013675892725586891, + 0.0077945636585354805, + 0.04617979750037193, + 0.013179798610508442, + 0.00939041655510664, + 0.011423508636653423, + 0.017306460067629814, + 0.013536762446165085, + 0.01227529812604189, + 0.007903649471700191, + 0.041303712874650955, + 0.02955729514360428, + 0.005042306147515774, + 0.007654478773474693, + 0.005554166156798601, + 0.005332666914910078, + 0.007335758302360773, + 0.003742859698832035, + 0.00955163687467575, + 0.014001740142703056, + 0.03330841660499573, + 0.024741437286138535, + 0.037084586918354034, + 0.028075864538550377, + 0.030779490247368813, + 0.0368463508784771, + 0.02957926318049431, + 0.04235760122537613, + 0.003209773451089859, + 0.0025270457845181227, + 0.003209092654287815, + 0.003816730109974742, + 0.004725004080682993, + 0.00414986303076148, + 0.002881958382204175, + 0.004950530361384153, + 0.004904152825474739, + 0.007042902056127787, + 0.0054681915789842606, + 0.005171396303921938, + 0.004732179921120405, + 0.00426629139110446, + 0.007132026832550764, + 0.006209137383848429, + 0.004340289626270533, + 0.00656229630112648, + 0.008278045803308487, + 0.010571290738880634, + 0.00550588546320796, + 0.0045905751176178455, + 0.03295641764998436, + 0.02329358458518982, + 0.006291623692959547, + 0.004731052555143833, + 0.03406591713428497, + 0.00872110202908516, + 0.005669621285051107, + 0.0038137768860906363, + 0.0391070730984211, + 0.007634822279214859, + 0.003157147439196706, + 0.0019630370661616325, + 0.0023695044219493866, + 0.001899620983749628, + 0.003480822080746293, + 0.005554639734327793, + 0.001925639109686017, + 0.0023241539020091295, + 0.008599378168582916, + 0.007598640397191048, + 0.009012258611619473, + 0.00881294347345829, + 0.006348713301122189, + 0.005006297491490841, + 0.009559319354593754, + 0.007717708591371775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.04866965487599373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996948689222336 + ], + "opposite_logprob_tokens": [ + -2.929996967315674 + ], + "logprob_reward": -0.054996948689222336, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996948689222336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14918243885040283, + 0.2194433957338333, + 0.042641542851924896, + 0.1463145613670349, + 0.05076202005147934, + 0.07712576538324356, + 0.09043273329734802, + 0.11426566541194916, + 0.04462546855211258, + 0.06520643830299377 + ], + "credit_map": [ + 0.009795880876481533, + 0.017675282433629036, + 0.03650061786174774, + 0.016330866143107414, + 0.011628271080553532, + 0.011661914177238941, + 0.028327377513051033, + 0.017262224107980728, + 0.017154518514871597, + 0.012156841345131397, + 0.027626052498817444, + 0.01588701643049717, + 0.014477714896202087, + 0.011177205480635166, + 0.0750194564461708, + 0.04594458267092705, + 0.002896849764510989, + 0.003817926160991192, + 0.003002335550263524, + 0.0047322423197329044, + 0.005951739381998777, + 0.0066173989325761795, + 0.00734051875770092, + 0.008282526396214962, + 0.01673329435288906, + 0.01377809140831232, + 0.02048775926232338, + 0.014535664580762386, + 0.018562637269496918, + 0.02130114659667015, + 0.017399601638317108, + 0.023516353219747543, + 0.0034939430188387632, + 0.004160317592322826, + 0.002748003462329507, + 0.01099439151585102, + 0.006970573216676712, + 0.006019207648932934, + 0.004387948662042618, + 0.011987632140517235, + 0.010046773590147495, + 0.009636085480451584, + 0.011294920928776264, + 0.010685743764042854, + 0.008613691665232182, + 0.008008761331439018, + 0.009331888519227505, + 0.009507890790700912, + 0.00405606534332037, + 0.008903800509870052, + 0.006939175073057413, + 0.009161788038909435, + 0.005293834488838911, + 0.005996492225676775, + 0.030773542821407318, + 0.019308028742671013, + 0.004507729317992926, + 0.0046076648868620396, + 0.03902004286646843, + 0.007467223331332207, + 0.006218180060386658, + 0.0033600488677620888, + 0.041704561561346054, + 0.007380204740911722, + 0.004238705150783062, + 0.005037417635321617, + 0.004871387965977192, + 0.005083170253783464, + 0.004065521992743015, + 0.011428771540522575, + 0.0048661245964467525, + 0.0050343675538897514, + 0.00745053356513381, + 0.006063244771212339, + 0.00698784738779068, + 0.00880563072860241, + 0.008368880487978458, + 0.006740143056958914, + 0.01348647940903902, + 0.0073036737740039825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.054996948689222336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579845190048218 + ], + "opposite_logprob_tokens": [ + -0.6329845190048218 + ], + "logprob_reward": -0.7579845190048218, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579845190048218, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08538196980953217, + 0.03965902328491211, + 0.02307610772550106, + 0.5956428647041321, + 0.024941403418779373, + 0.014500759541988373, + 0.01624060422182083, + 0.11108671128749847, + 0.032161954790353775, + 0.05730850249528885 + ], + "credit_map": [ + 0.005457019433379173, + 0.009176583029329777, + 0.014886271208524704, + 0.010772270150482655, + 0.011354582384228706, + 0.008570725098252296, + 0.01260745245963335, + 0.0125570734962821, + 0.002548233838751912, + 0.003191406838595867, + 0.0042884317226707935, + 0.003349506761878729, + 0.0032744070049375296, + 0.0026695579290390015, + 0.010892958380281925, + 0.009444523602724075, + 0.001993070123717189, + 0.002958624390885234, + 0.0021529674995690584, + 0.002213100902736187, + 0.0031845499761402607, + 0.0022943478543311357, + 0.003234775271266699, + 0.005044673569500446, + 0.07718490809202194, + 0.05747228488326073, + 0.08300141990184784, + 0.06117589399218559, + 0.07194362580776215, + 0.08870341628789902, + 0.07058822363615036, + 0.08557315915822983, + 0.003139183856546879, + 0.0028533865697681904, + 0.0028648595325648785, + 0.0024675652384757996, + 0.004771028179675341, + 0.002742473501712084, + 0.0030442308634519577, + 0.0030586763750761747, + 0.0018655082676559687, + 0.0021997173316776752, + 0.0019422749755904078, + 0.0017604525201022625, + 0.001397312618792057, + 0.0014101715059950948, + 0.001965669682249427, + 0.001959653338417411, + 0.0011965313460677862, + 0.001598019734956324, + 0.002473425352945924, + 0.0028005221392959356, + 0.0012014034437015653, + 0.0007806946523487568, + 0.0036725783720612526, + 0.002517429878935218, + 0.00537704536691308, + 0.0027999705635011196, + 0.040945567190647125, + 0.005958182737231255, + 0.004024954978376627, + 0.0023626249749213457, + 0.04229475185275078, + 0.007323623634874821, + 0.003634355729445815, + 0.003464870620518923, + 0.006614913698285818, + 0.003391942009329796, + 0.005099920090287924, + 0.0029622074216604233, + 0.0025094207376241684, + 0.004484325647354126, + 0.006614444311708212, + 0.006528552155941725, + 0.007583789527416229, + 0.010111833922564983, + 0.006748946383595467, + 0.0065073673613369465, + 0.006445004604756832, + 0.006768569350242615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.7579845190048218 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, first-person perspective shot looks directly down upon a massive architectural blueprint, spread across a dark, matte drafting table. The design's intricate, sharp lines represent complex structures and pathways, glowing a brilliant, inner cyan against the deep indigo paper. This perfect digital precision dissolves as the luminescent ink begins to accumulate, forming shimmering beads of liquid directly onto the glowing carrier lines. Then, slowly and relentlessly, these droplets of pure light begin to leak outward, smudged from their defined paths as the glowing liquid bleeds across the surface of the paper, smudged the clear geometric patterns into a soft, ethereal blur, and coalesced at intersections to create even brighter nodes of light before spilling out in fragile, uncontrolled tendrils. This self-luminous ink casts a cool, otherworldly glow, contrasting starkly with the warm, dim ambient light of the surrounding, unseen room, creating a focused, captivating atmosphere of magical transformation.", + "seed": 9612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_0/index_7324_seed_9612_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10560021549463272, + "vlm_logprob": -0.10560021549463272, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.10560021549463272, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16849200427532196 + ], + "opposite_logprob_tokens": [ + -1.9184919595718384 + ], + "logprob_reward": -0.16849200427532196, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16849200427532196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6616529822349548, + 0.183891162276268, + 0.01209373027086258, + 0.023168109357357025, + 0.012729593552649021, + 0.013952837325632572, + 0.029913656413555145, + 0.010896248742938042, + 0.01484726369380951, + 0.03685447573661804 + ], + "credit_map": [ + 0.017724180594086647, + 0.018385108560323715, + 0.033146362751722336, + 0.08584650605916977, + 0.02621510624885559, + 0.02040097303688526, + 0.2581208348274231, + 0.20181387662887573, + 0.024908145889639854, + 0.017450103536248207, + 0.031105147674679756, + 0.01592676155269146, + 0.014148817397654057, + 0.02388448640704155, + 0.029562901705503464, + 0.026904791593551636, + 0.0027827529702335596, + 0.0008809803985059261, + 0.0020975391380488873, + 0.0013936251634731889, + 0.0010408855741843581, + 0.001203737105242908, + 0.0014118836261332035, + 0.0012823264114558697, + 0.0027290547732263803, + 0.0031436823774129152, + 0.0023506558500230312, + 0.003325861878693104, + 0.002821215195581317, + 0.0032849141862243414, + 0.0027100450824946165, + 0.002802679780870676, + 0.0028236955404281616, + 0.002032515360042453, + 0.0014585829339921474, + 0.0009431059006601572, + 0.00122273329179734, + 0.0017942851409316063, + 0.0008898808155208826, + 0.0015647943364456296, + 0.001139185857027769, + 0.001210937974974513, + 0.0015886485343798995, + 0.0032024115789681673, + 0.001225300831720233, + 0.001654303283430636, + 0.0024334131740033627, + 0.0014986367896199226, + 0.004178307950496674, + 0.003867305815219879, + 0.002675263909623027, + 0.004334586206823587, + 0.0031345600727945566, + 0.003117717569693923, + 0.0048690298572182655, + 0.003736885031685233, + 0.0017755457665771246, + 0.0011023666011169553, + 0.00092587701510638, + 0.0010791266104206443, + 0.0011842395178973675, + 0.0017585230525583029, + 0.0013009256217628717, + 0.0017696439754217863, + 0.0013252322096377611, + 0.0013256260426715016, + 0.0012727859430015087, + 0.0015420510899275541, + 0.002563147572800517, + 0.003171083051711321, + 0.0018220110796391964, + 0.0018253264715895057, + 0.004568913020193577, + 0.0027514598332345486, + 0.00282495841383934, + 0.004216192290186882, + 0.00494728097692132, + 0.006067908834666014, + 0.004961100406944752, + 0.006516661029309034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.16849200427532196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428404450416565 + ], + "opposite_logprob_tokens": [ + -2.0178403854370117 + ], + "logprob_reward": -0.1428404450416565, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428404450416565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6994319558143616, + 0.13575448095798492, + 0.024206794798374176, + 0.029672108590602875, + 0.011424233205616474, + 0.012480687350034714, + 0.024462534114718437, + 0.010525439865887165, + 0.022071609273552895, + 0.029970165342092514 + ], + "credit_map": [ + 0.023133084177970886, + 0.022528886795043945, + 0.028235431760549545, + 0.08217548578977585, + 0.031243057921528816, + 0.029171118512749672, + 0.2724519968032837, + 0.21049290895462036, + 0.01892124116420746, + 0.013774414546787739, + 0.020407970994710922, + 0.012383125722408295, + 0.010454384610056877, + 0.018419193103909492, + 0.02261354960501194, + 0.018780609592795372, + 0.006208011414855719, + 0.0017122133867815137, + 0.0029636449180543423, + 0.003020227188244462, + 0.002013872843235731, + 0.002402315614745021, + 0.002825481118634343, + 0.0030610281974077225, + 0.00321570853702724, + 0.0045408932492136955, + 0.002826859476044774, + 0.005449963733553886, + 0.00279074814170599, + 0.0034293828066438437, + 0.004146330524235964, + 0.003272221190854907, + 0.001255428302101791, + 0.0012463389430195093, + 0.0024079380091279745, + 0.001659254776313901, + 0.0012250255094841123, + 0.001238707802258432, + 0.0010329891229048371, + 0.0013585499254986644, + 0.0013790145749226213, + 0.0014371366705745459, + 0.0014312140410766006, + 0.0022250062320381403, + 0.001111471327021718, + 0.0017477682558819652, + 0.0017069971654564142, + 0.0014420782681554556, + 0.0038576731458306313, + 0.003336816094815731, + 0.002311265328899026, + 0.0034071977715939283, + 0.00236929371021688, + 0.0027027735486626625, + 0.003698212094604969, + 0.0027793028857558966, + 0.0020293898414820433, + 0.001670560915954411, + 0.0008741205674596131, + 0.0009304698905907571, + 0.001399864093400538, + 0.0010928408009931445, + 0.0010727130575105548, + 0.0014554803492501378, + 0.0021361426915973425, + 0.0016545627731829882, + 0.0020891481544822454, + 0.002310577780008316, + 0.0032333489507436752, + 0.0052231429144740105, + 0.0027023358270525932, + 0.00272234994918108, + 0.005355044733732939, + 0.0030435696244239807, + 0.0026434806641191244, + 0.003369265701621771, + 0.0038897383492439985, + 0.0029664586763828993, + 0.0040448796935379505, + 0.004657728131860495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.1428404450416565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965805262327194 + ], + "opposite_logprob_tokens": [ + -2.9299657344818115 + ], + "logprob_reward": -0.054965805262327194, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965805262327194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5400803089141846, + 0.06110949069261551, + 0.045215629041194916, + 0.037026505917310715, + 0.03839697688817978, + 0.0634499117732048, + 0.09459102898836136, + 0.039191629737615585, + 0.035673972219228745, + 0.045264508575201035 + ], + "credit_map": [ + 0.022088827565312386, + 0.029015837237238884, + 0.045219067484140396, + 0.09363790601491928, + 0.026654934510588646, + 0.02999955043196678, + 0.17000962793827057, + 0.12345461547374725, + 0.00923944916576147, + 0.0070814099162817, + 0.009875210002064705, + 0.00668753357604146, + 0.005283416248857975, + 0.008628296665847301, + 0.005492216907441616, + 0.008821960538625717, + 0.005881607532501221, + 0.003789337119087577, + 0.007709990721195936, + 0.005366733763366938, + 0.005254573188722134, + 0.004619218874722719, + 0.005410752724856138, + 0.007183419074863195, + 0.004472800064831972, + 0.004266185685992241, + 0.005747776012867689, + 0.0057248640805482864, + 0.003537715645506978, + 0.003830504836514592, + 0.005540011450648308, + 0.003906650468707085, + 0.004744473844766617, + 0.004905304405838251, + 0.005241850391030312, + 0.004673117771744728, + 0.004990157205611467, + 0.005548403132706881, + 0.0036918623372912407, + 0.004601811058819294, + 0.005568847060203552, + 0.006716208066791296, + 0.008127162232995033, + 0.012421802617609501, + 0.006246998906135559, + 0.005545034073293209, + 0.01077310461550951, + 0.008050753735005856, + 0.011663857847452164, + 0.010433226823806763, + 0.011250816285610199, + 0.012189756147563457, + 0.009896543808281422, + 0.010746389627456665, + 0.017508480697870255, + 0.01090197078883648, + 0.004351743031293154, + 0.005222274921834469, + 0.0028655901551246643, + 0.003976201172918081, + 0.005949106533080339, + 0.005880295764654875, + 0.004475514404475689, + 0.006470908876508474, + 0.004017919301986694, + 0.0035448281560093164, + 0.0035793068818747997, + 0.0037870912346988916, + 0.004029444884508848, + 0.00923476554453373, + 0.0031013719271868467, + 0.0043792459182441235, + 0.006139843259006739, + 0.003829948604106903, + 0.004830844700336456, + 0.0034292927011847496, + 0.0077554709278047085, + 0.005154299084097147, + 0.006755563896149397, + 0.007369250059127808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.054965805262327194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867975786328316 + ], + "opposite_logprob_tokens": [ + -3.048679828643799 + ], + "logprob_reward": -0.04867975786328316, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867975786328316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6497987508773804, + 0.10454875230789185, + 0.04382239654660225, + 0.022034233435988426, + 0.027865026146173477, + 0.05200834572315216, + 0.033460330218076706, + 0.02408476173877716, + 0.011371133849024773, + 0.031006256118416786 + ], + "credit_map": [ + 0.014864186756312847, + 0.03236444666981697, + 0.04069903492927551, + 0.07332612574100494, + 0.024416569620370865, + 0.021863320842385292, + 0.24999216198921204, + 0.19227291643619537, + 0.013577018864452839, + 0.009787840768694878, + 0.016320692375302315, + 0.010943963192403316, + 0.008276491425931454, + 0.013296453282237053, + 0.017899081110954285, + 0.014447217807173729, + 0.01153611857444048, + 0.0030382450204342604, + 0.007074656430631876, + 0.005095692817121744, + 0.0033568183425813913, + 0.004706417676061392, + 0.004296634811908007, + 0.004717813339084387, + 0.0019981178920716047, + 0.0021564238704741, + 0.002399083925411105, + 0.004413713701069355, + 0.0020755715668201447, + 0.002071839990094304, + 0.003903623204678297, + 0.0030158592853695154, + 0.0039127292111516, + 0.0030929616186767817, + 0.0045101223513484, + 0.003568542655557394, + 0.002903400454670191, + 0.003971341997385025, + 0.0026533063501119614, + 0.003252618946135044, + 0.003391195321455598, + 0.0054356688633561134, + 0.006000660825520754, + 0.0101015605032444, + 0.005392666906118393, + 0.00600475026294589, + 0.009534487500786781, + 0.006147355306893587, + 0.004557505249977112, + 0.0035012022126466036, + 0.0044481512159109116, + 0.0037354512605816126, + 0.0037327269092202187, + 0.0036809267476201057, + 0.005833033937960863, + 0.003971331287175417, + 0.0024807394947856665, + 0.003367855679243803, + 0.001721699140034616, + 0.00209597684442997, + 0.0038883090019226074, + 0.0038881597574800253, + 0.0026048438157886267, + 0.004037178587168455, + 0.0009077118011191487, + 0.0009869657224044204, + 0.0015550231328234076, + 0.0015847113681957126, + 0.001518857548944652, + 0.0013953058514744043, + 0.001838106196373701, + 0.001584452809765935, + 0.004923302214592695, + 0.0032341990154236555, + 0.003150831675156951, + 0.0028254613280296326, + 0.004965749103575945, + 0.0037147821858525276, + 0.003832207527011633, + 0.004359721671789885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04867975786328316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, first-person perspective looking directly down at a massive architectural blueprint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the blueprint spread across a dark, matte drafting table with sharp lines glowing brilliant cyan against deep indigo paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glowing ink dissolve into shimmering beads of liquid that leak outward and smudge the geometric patterns into a soft blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the smudged lines coalesce at intersections to create brighter nodes of light before spilling out in fragile, uncontrolled tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the self-luminous ink cast a cool, otherworldly glow that contrasts with the warm, dim ambient light of the surrounding room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1130230650305748 + ], + "opposite_logprob_tokens": [ + -2.238023042678833 + ], + "logprob_reward": -0.1130230650305748, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1130230650305748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16024304926395416, + 0.19383254647254944, + 0.13423967361450195, + 0.07710552960634232, + 0.07876280695199966, + 0.06817085295915604, + 0.056120745837688446, + 0.05385100468993187, + 0.027816910296678543, + 0.14985689520835876 + ], + "credit_map": [ + 0.017625458538532257, + 0.01868915557861328, + 0.023196391761302948, + 0.020348578691482544, + 0.015021141618490219, + 0.03156571090221405, + 0.01557120867073536, + 0.018225396052002907, + 0.024517938494682312, + 0.020005684345960617, + 0.03000037558376789, + 0.022299762815237045, + 0.015957068651914597, + 0.02413434349000454, + 0.03270725533366203, + 0.024210108444094658, + 0.03342743590474129, + 0.012676496058702469, + 0.01826312579214573, + 0.014096160419285297, + 0.00985007919371128, + 0.012834020890295506, + 0.014685500413179398, + 0.01840684376657009, + 0.007539799436926842, + 0.018147502094507217, + 0.007172941230237484, + 0.010075727477669716, + 0.007907094433903694, + 0.011155552230775356, + 0.007085988763719797, + 0.008020921610295773, + 0.017380746081471443, + 0.009222546592354774, + 0.009854543954133987, + 0.005758463405072689, + 0.006471451837569475, + 0.011686005629599094, + 0.005526582710444927, + 0.012862457893788815, + 0.005713097285479307, + 0.00793689489364624, + 0.008120975457131863, + 0.0137217678129673, + 0.0068425689823925495, + 0.007286611013114452, + 0.010408916510641575, + 0.008140016347169876, + 0.008154640905559063, + 0.007089528255164623, + 0.006590057164430618, + 0.006461469456553459, + 0.006634000688791275, + 0.006702033802866936, + 0.008003154769539833, + 0.006485858466476202, + 0.006204427685588598, + 0.01025066152215004, + 0.0039605083875358105, + 0.003989425953477621, + 0.008378558792173862, + 0.009005826897919178, + 0.006096211727708578, + 0.005965381860733032, + 0.002606752095744014, + 0.002712582005187869, + 0.002787105506286025, + 0.0045671723783016205, + 0.004127673339098692, + 0.0028506461530923843, + 0.004296009428799152, + 0.003868967294692993, + 0.021416999399662018, + 0.016917360946536064, + 0.010274887084960938, + 0.015317820012569427, + 0.027915988117456436, + 0.02223505824804306, + 0.019716277718544006, + 0.016062486916780472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.1130230650305748 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_tensor([6560, 6560])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_tensor([6560, 6560])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0b98f177ffd17dec584e5f4a0c64497e8d4a4dfc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_tensor([6560, 6560])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle close-up captures a meticulously orchestrated scientific procedure on a sterile stainless steel workbench, beneath bright, even overhead fluorescent lighting. A pair of hands, clad in tight-fitting blue butyl gloves, manipulate with precision. The gloved right hand first firmly grasps a clear plastic syringe with black measurement markings and shakes it quickly and deliberately to mix the clear liquid within. Simultaneously, the gloved left hand steadily slides a clear, round Petri dish across the reflective steel surface and into the center of the frame. The right hand then moves to a small glass bottle, deftly unscrewing the plastic cap and carefully placing it on a plain white cardboard box nearby. After another quick shake of the syringe, the right hand guides the needle, decisively inserting it into the bottle's rubber stopper. The left hand then carefully lifts the lid of the Petri dish and secures it above the base. The right hand tilts the syringe over the open Petri dish, carefully pressing the plunger with its thumb, releasing a few drops of clear liquid onto the culture medium within, keeping it steady. The syringe is then smoothly lifted, and the left hand instantly and precisely closes the cap, sealing the Petri dish, creating a clinical and highly professional atmosphere throughout the entire process.", + "seed": 22396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_6560_seed_22396_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.044044694490730764, + "vlm_logprob": -0.044044694490730764, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.044044694490730764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044637568295001984 + ], + "opposite_logprob_tokens": [ + -3.419637680053711 + ], + "logprob_reward": -0.044637568295001984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.044637568295001984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387915462255478, + 0.09733057022094727, + 0.08219209313392639, + 0.051125332713127136, + 0.03820577636361122, + 0.08938643336296082, + 0.1662093847990036, + 0.13204972445964813, + 0.09192860871553421, + 0.11278043687343597 + ], + "credit_map": [ + 0.020729653537273407, + 0.023294445127248764, + 0.015386359766125679, + 0.015566794201731682, + 0.015082216821610928, + 0.02965734340250492, + 0.007470904849469662, + 0.011603837832808495, + 0.01323009468615055, + 0.010897830128669739, + 0.016268707811832428, + 0.010779938660562038, + 0.010120811872184277, + 0.016730118542909622, + 0.009670230560004711, + 0.009632844477891922, + 0.01401137001812458, + 0.012757047079503536, + 0.010941815562546253, + 0.00905599445104599, + 0.01716606691479683, + 0.008621261455118656, + 0.0050789411179721355, + 0.0045596035197377205, + 0.006954483687877655, + 0.008043364621698856, + 0.007661005016416311, + 0.006085429806262255, + 0.003102157963439822, + 0.0069631836377084255, + 0.007167584728449583, + 0.00514812720939517, + 0.004834061022847891, + 0.005679829977452755, + 0.005284543614834547, + 0.005954626947641373, + 0.0028818980790674686, + 0.0030540290754288435, + 0.005428976379334927, + 0.005087816156446934, + 0.01127349678426981, + 0.012499256059527397, + 0.008998743258416653, + 0.01337581779807806, + 0.00769016332924366, + 0.00994690228253603, + 0.009433696046471596, + 0.016168365254998207, + 0.022006481885910034, + 0.015825722366571426, + 0.020472312346100807, + 0.03436647355556488, + 0.016034280881285667, + 0.013740678317844868, + 0.030566899105906487, + 0.01319655030965805, + 0.019551584497094154, + 0.019408315420150757, + 0.01696845144033432, + 0.01751186139881611, + 0.013597318902611732, + 0.014094731770455837, + 0.017702383920550346, + 0.0132150799036026, + 0.01018431968986988, + 0.007737904321402311, + 0.006641844287514687, + 0.010376447811722755, + 0.015221583656966686, + 0.017561934888362885, + 0.011263903230428696, + 0.012940687127411366, + 0.016055192798376083, + 0.01286908145993948, + 0.016767529770731926, + 0.009482664056122303, + 0.008205565623939037, + 0.00992885697633028, + 0.01763414964079857, + 0.02183740772306919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.044637568295001984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422515232115984 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.01422515232115984, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422515232115984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12330755591392517, + 0.05845584347844124, + 0.03871452808380127, + 0.10369548946619034, + 0.12552465498447418, + 0.10393920540809631, + 0.21883757412433624, + 0.10937945544719696, + 0.05176106467843056, + 0.06638465076684952 + ], + "credit_map": [ + 0.023123016580939293, + 0.026732327416539192, + 0.012101569212973118, + 0.01763596571981907, + 0.009711742401123047, + 0.0159586351364851, + 0.007703171577304602, + 0.010341123677790165, + 0.0072201951406896114, + 0.009944393299520016, + 0.007220932748168707, + 0.007850725203752518, + 0.005925294011831284, + 0.008628752082586288, + 0.005475422367453575, + 0.006190128158777952, + 0.006755396258085966, + 0.004885223228484392, + 0.005496842786669731, + 0.0026379020418971777, + 0.007525049615651369, + 0.004377249628305435, + 0.0029903193935751915, + 0.0040465486235916615, + 0.01972595602273941, + 0.025126250460743904, + 0.015926815569400787, + 0.010481956414878368, + 0.00499881012365222, + 0.010123862884938717, + 0.009788896888494492, + 0.007522935047745705, + 0.021654080599546432, + 0.021316708996891975, + 0.01288010086864233, + 0.0180680975317955, + 0.009874633513391018, + 0.011645224876701832, + 0.015656594187021255, + 0.014429216273128986, + 0.012261188589036465, + 0.013211260549724102, + 0.010660005733370781, + 0.014217858202755451, + 0.0084524592384696, + 0.013747360557317734, + 0.012736125849187374, + 0.01865294575691223, + 0.02975141443312168, + 0.024509936571121216, + 0.025429220870137215, + 0.04965854436159134, + 0.022231966257095337, + 0.01720660924911499, + 0.0296781063079834, + 0.02037177048623562, + 0.010586650110781193, + 0.013931686989963055, + 0.014487690292298794, + 0.011442865245044231, + 0.010206588543951511, + 0.02268197201192379, + 0.01455261092633009, + 0.011489395052194595, + 0.00824874360114336, + 0.005980895832180977, + 0.004074367228895426, + 0.005098234862089157, + 0.005716250278055668, + 0.006709855515509844, + 0.007993154227733612, + 0.007939564064145088, + 0.007223252207040787, + 0.008535949513316154, + 0.008428010158240795, + 0.006386397872120142, + 0.007079811301082373, + 0.007425460033118725, + 0.008932357653975487, + 0.012373405508697033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.01422515232115984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371930867433548 + ], + "opposite_logprob_tokens": [ + -3.408719301223755 + ], + "logprob_reward": -0.03371930867433548, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371930867433548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05636347830295563, + 0.0550694577395916, + 0.034463632851839066, + 0.041755594313144684, + 0.07780436426401138, + 0.08450251817703247, + 0.4400339126586914, + 0.14164137840270996, + 0.040497634559869766, + 0.0278680007904768 + ], + "credit_map": [ + 0.01359939482063055, + 0.012496362440288067, + 0.0058527979999780655, + 0.008045865222811699, + 0.0037553247530013323, + 0.006488305050879717, + 0.002697244519367814, + 0.00342818396165967, + 0.010193209163844585, + 0.00957992859184742, + 0.007885660976171494, + 0.006480936426669359, + 0.005875025875866413, + 0.005610882770270109, + 0.004701577592641115, + 0.004742239136248827, + 0.004851604346185923, + 0.0047224415466189384, + 0.004559971392154694, + 0.002360749989748001, + 0.006851350888609886, + 0.00487694563344121, + 0.0029127434827387333, + 0.003327829297631979, + 0.005376768764108419, + 0.009677066467702389, + 0.005903040524572134, + 0.00510711595416069, + 0.0031185876578092575, + 0.003538887482136488, + 0.00475452933460474, + 0.004279597662389278, + 0.01268770918250084, + 0.014452815987169743, + 0.008028104901313782, + 0.010509598068892956, + 0.008389982394874096, + 0.005510946735739708, + 0.01082137692719698, + 0.007403825409710407, + 0.012377643957734108, + 0.008079955354332924, + 0.009430985897779465, + 0.010604767128825188, + 0.007856517098844051, + 0.014207028783857822, + 0.007983864285051823, + 0.013961758464574814, + 0.0622469000518322, + 0.045068319886922836, + 0.05195581167936325, + 0.0890793725848198, + 0.044603683054447174, + 0.03582421690225601, + 0.06857392936944962, + 0.04268169403076172, + 0.01379725057631731, + 0.017705490812659264, + 0.019533026963472366, + 0.016075775027275085, + 0.018689487129449844, + 0.024227315559983253, + 0.018523864448070526, + 0.013089167885482311, + 0.00449446402490139, + 0.004917202517390251, + 0.002726395847275853, + 0.004279425833374262, + 0.007463850546628237, + 0.006536101922392845, + 0.004472768399864435, + 0.005607425235211849, + 0.0036744449753314257, + 0.003702381392940879, + 0.003267789725214243, + 0.0041207075119018555, + 0.0027185857761651278, + 0.0037132413126528263, + 0.003852861002087593, + 0.0028179893270134926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.03371930867433548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899484038352966 + ], + "opposite_logprob_tokens": [ + -2.5789947509765625 + ], + "logprob_reward": -0.07899484038352966, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899484038352966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06897858530282974, + 0.07518772780895233, + 0.04622470587491989, + 0.09010786563158035, + 0.040129076689481735, + 0.15510417520999908, + 0.11048804223537445, + 0.3389871418476105, + 0.050006501376628876, + 0.024786178022623062 + ], + "credit_map": [ + 0.01274364348500967, + 0.015636134892702103, + 0.00854133628308773, + 0.007753125391900539, + 0.00598104577511549, + 0.008907853625714779, + 0.004368508234620094, + 0.005046936683356762, + 0.014288515783846378, + 0.013278036378324032, + 0.009576746262609959, + 0.008945805951952934, + 0.008189240470528603, + 0.009809551760554314, + 0.006224197801202536, + 0.004875633399933577, + 0.007078263908624649, + 0.006064604967832565, + 0.005658254027366638, + 0.0039792051538825035, + 0.008569728583097458, + 0.006135848816484213, + 0.004937932826578617, + 0.003800871316343546, + 0.014543174766004086, + 0.01835673302412033, + 0.012440957129001617, + 0.009507405571639538, + 0.0070165349170565605, + 0.008021638728678226, + 0.010893505997955799, + 0.009327911771833897, + 0.005801139399409294, + 0.004423605743795633, + 0.00463060662150383, + 0.007150216493755579, + 0.003747026203200221, + 0.0036654267460107803, + 0.00629060436040163, + 0.004420452751219273, + 0.025003930553793907, + 0.01685892790555954, + 0.017378654330968857, + 0.025548173114657402, + 0.010909494943916798, + 0.022187238559126854, + 0.01777053065598011, + 0.01944722980260849, + 0.018912237137556076, + 0.008832638151943684, + 0.014113488607108593, + 0.02634020335972309, + 0.009621749632060528, + 0.0071457126177847385, + 0.014407985843718052, + 0.011114032007753849, + 0.034719012677669525, + 0.036252111196517944, + 0.045150190591812134, + 0.039010897278785706, + 0.024597441777586937, + 0.0685557946562767, + 0.04819811135530472, + 0.04250359907746315, + 0.005870283115655184, + 0.005433748476207256, + 0.004073654301464558, + 0.004683999810367823, + 0.00658924225717783, + 0.00857880525290966, + 0.006947904359549284, + 0.007828864268958569, + 0.002860578242689371, + 0.003102983348071575, + 0.0032793167047202587, + 0.002115590963512659, + 0.002054997021332383, + 0.0025019750464707613, + 0.004135544877499342, + 0.004735194146633148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.07899484038352966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646602779626846 + ], + "opposite_logprob_tokens": [ + -3.0486466884613037 + ], + "logprob_reward": -0.048646602779626846, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646602779626846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0926324650645256, + 0.0446610152721405, + 0.029236283153295517, + 0.03337598219513893, + 0.030303752049803734, + 0.09517958015203476, + 0.21599061787128448, + 0.3840876817703247, + 0.048543985933065414, + 0.025988668203353882 + ], + "credit_map": [ + 0.018268998712301254, + 0.023013390600681305, + 0.010398216545581818, + 0.011583942919969559, + 0.008127342909574509, + 0.011570597067475319, + 0.004341250751167536, + 0.005328732077032328, + 0.007653824519366026, + 0.008928114548325539, + 0.006628792267292738, + 0.003901052288711071, + 0.005494480021297932, + 0.004256995860487223, + 0.004026499576866627, + 0.0037712559569627047, + 0.00488520972430706, + 0.0038502714596688747, + 0.003472640411928296, + 0.00296135270036757, + 0.004805884789675474, + 0.0034852568060159683, + 0.0023210477083921432, + 0.003454619087278843, + 0.004309951327741146, + 0.007834953255951405, + 0.006146383471786976, + 0.003750365227460861, + 0.002196183195337653, + 0.004411966539919376, + 0.0027103498578071594, + 0.002015829784795642, + 0.004663284868001938, + 0.00430563697591424, + 0.0032533081248402596, + 0.004978554788976908, + 0.00344631215557456, + 0.0024150714743882418, + 0.004073810297995806, + 0.0031677735969424248, + 0.012631281279027462, + 0.008511518128216267, + 0.01008172519505024, + 0.015184490010142326, + 0.009224362671375275, + 0.012748573906719685, + 0.011701100505888462, + 0.015096531249582767, + 0.028825443238019943, + 0.023611510172486305, + 0.024746792390942574, + 0.0437275655567646, + 0.023671552538871765, + 0.021446092054247856, + 0.027864614501595497, + 0.022097058594226837, + 0.052720919251441956, + 0.05057166516780853, + 0.04561556503176689, + 0.04934344068169594, + 0.03845500573515892, + 0.06190200522542, + 0.045987319201231, + 0.03949175029993057, + 0.00559128588065505, + 0.004717723000794649, + 0.003089744132012129, + 0.005627283360809088, + 0.007629729341715574, + 0.008812679909169674, + 0.008227979764342308, + 0.004847559612244368, + 0.003923586569726467, + 0.0030663793440908194, + 0.0031575544271618128, + 0.003277231939136982, + 0.0020139170810580254, + 0.0029475055634975433, + 0.003666934324428439, + 0.003935560118407011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.048646602779626846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a meticulously orchestrated scientific procedure on a sterile stainless steel workbench, beneath bright, even overhead fluorescent lighting. A pair of hands, clad in tight-fitting blue butyl gloves, manipulate with precision. The gloved right hand first firmly grasps a clear plastic syringe with black measurement markings and shakes it quickly and deliberately to mix the clear liquid within. Simultaneously, the gloved left hand steadily slides a clear, round Petri dish across the reflective steel surface and into the center of the frame. The right hand then moves to a small glass bottle, deftly unscrewing the plastic cap and carefully placing it on a plain white cardboard box nearby. After another quick shake of the syringe, the right hand guides the needle, decisively inserting it into the bottle's rubber stopper. The left hand then carefully lifts the lid of the Petri dish and secures it above the base. The right hand tilts the syringe over the open Petri dish, carefully pressing the plunger with its thumb, releasing a few drops of clear liquid onto the culture medium within, keeping it steady. The syringe is then smoothly lifted, and the left hand instantly and precisely closes the cap, sealing the Petri dish, creating a clinical and highly professional atmosphere throughout the entire process.", + "seed": 22397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_6560_seed_22397_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05048861727118492, + "vlm_logprob": -0.05048861727118492, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.05048861727118492, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10972068458795547 + ], + "opposite_logprob_tokens": [ + -2.3597207069396973 + ], + "logprob_reward": -0.10972068458795547, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10972068458795547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13021554052829742, + 0.036264680325984955, + 0.10849888622760773, + 0.22212031483650208, + 0.054468750953674316, + 0.057837411761283875, + 0.04302263259887695, + 0.05918059125542641, + 0.1799180805683136, + 0.10847314447164536 + ], + "credit_map": [ + 0.023467617109417915, + 0.010946229100227356, + 0.01680927351117134, + 0.016491267830133438, + 0.009793098084628582, + 0.00799697544425726, + 0.016421470791101456, + 0.028289606794714928, + 0.004837819375097752, + 0.004249502904713154, + 0.004282819107174873, + 0.005329952109605074, + 0.0035771974362432957, + 0.0033425698056817055, + 0.0049990457482635975, + 0.005645770579576492, + 0.014080946333706379, + 0.015927333384752274, + 0.016573401167988777, + 0.021786924451589584, + 0.006760717369616032, + 0.008193220011889935, + 0.016456063836812973, + 0.00872026663273573, + 0.038001928478479385, + 0.05122804269194603, + 0.023585280403494835, + 0.023964248597621918, + 0.016825448721647263, + 0.01758161373436451, + 0.027750289067626, + 0.023183437064290047, + 0.004120492842048407, + 0.010184204205870628, + 0.009003864601254463, + 0.007152003701776266, + 0.007991647347807884, + 0.005704822018742561, + 0.005854019429534674, + 0.00445769727230072, + 0.006566185969859362, + 0.007598142139613628, + 0.011176289059221745, + 0.0073437499813735485, + 0.006993464659899473, + 0.005960356444120407, + 0.007580234203487635, + 0.004618983715772629, + 0.004755884874612093, + 0.0056249825283885, + 0.005719348322600126, + 0.004802749026566744, + 0.006829618941992521, + 0.004914894234389067, + 0.005925876088440418, + 0.00444927578791976, + 0.005365211982280016, + 0.011541256681084633, + 0.011316698975861073, + 0.006277685519307852, + 0.007323374040424824, + 0.0050751687958836555, + 0.006621657870709896, + 0.005659529473632574, + 0.03138197213411331, + 0.0273080226033926, + 0.027722937986254692, + 0.029302168637514114, + 0.018349919468164444, + 0.016527587547898293, + 0.015562120825052261, + 0.013763339258730412, + 0.012147180736064911, + 0.015518981032073498, + 0.01531003974378109, + 0.012993949465453625, + 0.013372172601521015, + 0.014905505813658237, + 0.012366804294288158, + 0.011858497746288776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.10972068458795547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637336403131485 + ], + "opposite_logprob_tokens": [ + -3.6513733863830566 + ], + "logprob_reward": -0.02637336403131485, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637336403131485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03548163175582886, + 0.05371568351984024, + 0.20067837834358215, + 0.07033852487802505, + 0.03976740688085556, + 0.06977282464504242, + 0.08001622557640076, + 0.19139453768730164, + 0.14001020789146423, + 0.1188245639204979 + ], + "credit_map": [ + 0.004833206534385681, + 0.003393567632883787, + 0.004292651545256376, + 0.004571214783936739, + 0.004576313775032759, + 0.003483016276732087, + 0.005658288486301899, + 0.004673373885452747, + 0.00809711217880249, + 0.005440812092274427, + 0.006909651216119528, + 0.007677264511585236, + 0.005428246222436428, + 0.004947546869516373, + 0.006910715717822313, + 0.008304339833557606, + 0.03010767325758934, + 0.04828105866909027, + 0.02594609372317791, + 0.019802812486886978, + 0.016306698322296143, + 0.016175031661987305, + 0.02493244782090187, + 0.019126567989587784, + 0.009922655299305916, + 0.01577324979007244, + 0.00856605265289545, + 0.008538770489394665, + 0.006394804921001196, + 0.006543497554957867, + 0.007461799308657646, + 0.007137699518352747, + 0.004199650138616562, + 0.008542053401470184, + 0.006123295519500971, + 0.005750028882175684, + 0.0045806351117789745, + 0.003482344327494502, + 0.0043327282182872295, + 0.002756670117378235, + 0.009461026638746262, + 0.011606280691921711, + 0.013367915526032448, + 0.009509843774139881, + 0.008473727852106094, + 0.005847882013767958, + 0.006600027903914452, + 0.004906117916107178, + 0.00898418202996254, + 0.014366176910698414, + 0.011865969747304916, + 0.007331592030823231, + 0.007952047511935234, + 0.011346419341862202, + 0.008517921902239323, + 0.009651916101574898, + 0.017898941412568092, + 0.04787222295999527, + 0.03961314260959625, + 0.02024787850677967, + 0.021933116018772125, + 0.011898874305188656, + 0.017693813890218735, + 0.014236570335924625, + 0.028538726270198822, + 0.014350762590765953, + 0.03791174292564392, + 0.013884607702493668, + 0.01320788636803627, + 0.010559236630797386, + 0.009619129821658134, + 0.011938116513192654, + 0.012988729402422905, + 0.011422565206885338, + 0.020901784300804138, + 0.011556564830243587, + 0.009848985821008682, + 0.020260866731405258, + 0.01077948696911335, + 0.021065589040517807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.02637336403131485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301932960748672 + ], + "opposite_logprob_tokens": [ + -3.773301839828491 + ], + "logprob_reward": -0.023301932960748672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301932960748672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06707017868757248, + 0.0685887262225151, + 0.15005545318126678, + 0.08358930796384811, + 0.04723600670695305, + 0.08054158091545105, + 0.04270278662443161, + 0.06774987280368805, + 0.34984686970710754, + 0.04261920228600502 + ], + "credit_map": [ + 0.00971419271081686, + 0.00534688588231802, + 0.008471024222671986, + 0.007873505353927612, + 0.008108851499855518, + 0.005511854309588671, + 0.01057893130928278, + 0.01146493386477232, + 0.009494917467236519, + 0.0067940871231257915, + 0.010061059147119522, + 0.010231362655758858, + 0.006897749844938517, + 0.007370690815150738, + 0.010552107356488705, + 0.0071867541410028934, + 0.02001131884753704, + 0.0363626591861248, + 0.020894862711429596, + 0.019466565921902657, + 0.011719447560608387, + 0.009545067325234413, + 0.021149005740880966, + 0.010906516574323177, + 0.014544985257089138, + 0.017590897157788277, + 0.008538338355720043, + 0.008674437180161476, + 0.007428732234984636, + 0.0070512667298316956, + 0.011635174043476582, + 0.008125478401780128, + 0.005826306529343128, + 0.006698692683130503, + 0.012844860553741455, + 0.008070983923971653, + 0.003949237987399101, + 0.0032655864488333464, + 0.003728944342583418, + 0.002851393772289157, + 0.010957387275993824, + 0.010803007520735264, + 0.014692576602101326, + 0.010477565228939056, + 0.011346363462507725, + 0.008628462441265583, + 0.007439822889864445, + 0.006196397356688976, + 0.005351573694497347, + 0.00505151366814971, + 0.007936175912618637, + 0.00609163660556078, + 0.005230915732681751, + 0.004231759812682867, + 0.005245085805654526, + 0.0035641249269247055, + 0.0074073863215744495, + 0.019326822832226753, + 0.013999063521623611, + 0.006470540538430214, + 0.005888976156711578, + 0.004340050742030144, + 0.005146821960806847, + 0.0051702093333005905, + 0.09598827362060547, + 0.032348375767469406, + 0.10118014365434647, + 0.03600212186574936, + 0.023715903982520103, + 0.026293860748410225, + 0.01620745100080967, + 0.0181107260286808, + 0.004950175527483225, + 0.004966259002685547, + 0.005920876748859882, + 0.004945868626236916, + 0.004734831862151623, + 0.005373495165258646, + 0.005922445561736822, + 0.005805245600640774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.023301932960748672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495542660355568 + ], + "opposite_logprob_tokens": [ + -2.92995548248291 + ], + "logprob_reward": -0.05495542660355568, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495542660355568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06517670303583145, + 0.05856461450457573, + 0.08003215491771698, + 0.11382008343935013, + 0.06102897971868515, + 0.11201971769332886, + 0.07157904654741287, + 0.04623311758041382, + 0.2678849697113037, + 0.12366054952144623 + ], + "credit_map": [ + 0.010146279819309711, + 0.005953498650342226, + 0.00910030398517847, + 0.011665874160826206, + 0.004805475939065218, + 0.005677004344761372, + 0.008924243971705437, + 0.008904018439352512, + 0.008351719938218594, + 0.0052771358750760555, + 0.006924243178218603, + 0.00810911413282156, + 0.006835523527115583, + 0.0063466159626841545, + 0.007924170233309269, + 0.0087960883975029, + 0.010028399527072906, + 0.013350307010114193, + 0.010316956788301468, + 0.0077574667520821095, + 0.008265029639005661, + 0.0075421882793307304, + 0.013985629193484783, + 0.008786185644567013, + 0.019640721380710602, + 0.023374419659376144, + 0.013143456541001797, + 0.012154866941273212, + 0.009071284905076027, + 0.00880018062889576, + 0.015689680352807045, + 0.01194547489285469, + 0.006333055905997753, + 0.010801094584167004, + 0.015050665475428104, + 0.009556741453707218, + 0.0053274608217179775, + 0.004880732391029596, + 0.005293185822665691, + 0.0037860467564314604, + 0.013980493880808353, + 0.017106806859374046, + 0.017527565360069275, + 0.013514232821762562, + 0.01446638535708189, + 0.013294534757733345, + 0.013420539908111095, + 0.00870915874838829, + 0.007864989340305328, + 0.0094516072422266, + 0.013568558730185032, + 0.011047989130020142, + 0.008697319775819778, + 0.0068281772546470165, + 0.008302738890051842, + 0.00581766851246357, + 0.005028021056205034, + 0.010241852141916752, + 0.009843474254012108, + 0.006300001870840788, + 0.004577711224555969, + 0.004074797034263611, + 0.00317843584343791, + 0.0029888302087783813, + 0.0906912162899971, + 0.013020174577832222, + 0.10414335131645203, + 0.0270292479544878, + 0.010187098756432533, + 0.007887665182352066, + 0.007016794290393591, + 0.007909407839179039, + 0.015192209742963314, + 0.014215445145964622, + 0.01666700653731823, + 0.013586612418293953, + 0.015844255685806274, + 0.01780986785888672, + 0.014549888670444489, + 0.015795255079865456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.05495542660355568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809167817234993 + ], + "opposite_logprob_tokens": [ + -3.2880916595458984 + ], + "logprob_reward": -0.03809167817234993, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809167817234993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05342927575111389, + 0.032017264515161514, + 0.13603650033473969, + 0.07722053676843643, + 0.04070105031132698, + 0.1911957710981369, + 0.06260771304368973, + 0.045317187905311584, + 0.26307621598243713, + 0.09839854389429092 + ], + "credit_map": [ + 0.008615067228674889, + 0.004654145333915949, + 0.006154465489089489, + 0.007758902385830879, + 0.005236273165792227, + 0.004741949494928122, + 0.00791558064520359, + 0.008352886885404587, + 0.0047570993192493916, + 0.0029774033464491367, + 0.004058398772031069, + 0.005385594442486763, + 0.0031557432375848293, + 0.003090806771069765, + 0.005402722861617804, + 0.0031894941348582506, + 0.015892738476395607, + 0.02385306917130947, + 0.01888173073530197, + 0.019910156726837158, + 0.011712283827364445, + 0.00850031990557909, + 0.02103196643292904, + 0.01625422015786171, + 0.011123337782919407, + 0.018111417070031166, + 0.00849751103669405, + 0.008109468035399914, + 0.006252479739487171, + 0.006288744509220123, + 0.010447440668940544, + 0.008390138857066631, + 0.0038445377722382545, + 0.0070667508989572525, + 0.007063626777380705, + 0.005146154202520847, + 0.00497828982770443, + 0.005400435067713261, + 0.004224156029522419, + 0.002977092284709215, + 0.01868976652622223, + 0.02475610561668873, + 0.01958329975605011, + 0.020436564460396767, + 0.038110487163066864, + 0.02174542471766472, + 0.028158051893115044, + 0.019716057926416397, + 0.006279418710619211, + 0.00883472990244627, + 0.009098494425415993, + 0.006871609017252922, + 0.007330751046538353, + 0.007880267687141895, + 0.008667160756886005, + 0.007645272184163332, + 0.004652052186429501, + 0.005994015838950872, + 0.010589051060378551, + 0.007664533331990242, + 0.0056433286517858505, + 0.0023727647494524717, + 0.005501722451299429, + 0.0028997191693633795, + 0.07754314690828323, + 0.020627351477742195, + 0.07716113328933716, + 0.027221480384469032, + 0.014866655692458153, + 0.011204243637621403, + 0.01722317561507225, + 0.017229007557034492, + 0.01554414164274931, + 0.009566773660480976, + 0.018702127039432526, + 0.010452806949615479, + 0.007594629656523466, + 0.01283099316060543, + 0.010474633425474167, + 0.013232431374490261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.03809167817234993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a meticulously orchestrated scientific procedure on a sterile stainless steel workbench, beneath bright, even overhead fluorescent lighting. A pair of hands, clad in tight-fitting blue butyl gloves, manipulate with precision. The gloved right hand first firmly grasps a clear plastic syringe with black measurement markings and shakes it quickly and deliberately to mix the clear liquid within. Simultaneously, the gloved left hand steadily slides a clear, round Petri dish across the reflective steel surface and into the center of the frame. The right hand then moves to a small glass bottle, deftly unscrewing the plastic cap and carefully placing it on a plain white cardboard box nearby. After another quick shake of the syringe, the right hand guides the needle, decisively inserting it into the bottle's rubber stopper. The left hand then carefully lifts the lid of the Petri dish and secures it above the base. The right hand tilts the syringe over the open Petri dish, carefully pressing the plunger with its thumb, releasing a few drops of clear liquid onto the culture medium within, keeping it steady. The syringe is then smoothly lifted, and the left hand instantly and precisely closes the cap, sealing the Petri dish, creating a clinical and highly professional atmosphere throughout the entire process.", + "seed": 22398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_6560_seed_22398_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05861643552780151, + "vlm_logprob": -0.05861643552780151, + "vlm_yes_no_margin": 3.049999976158142, + "vlm_token_credit": -0.05861643552780151, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06102766469120979 + ], + "opposite_logprob_tokens": [ + -3.061027765274048 + ], + "logprob_reward": -0.06102766469120979, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06102766469120979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17845715582370758, + 0.19288618862628937, + 0.06996682286262512, + 0.07523897290229797, + 0.026678048074245453, + 0.024224458262324333, + 0.06582342088222504, + 0.10351280868053436, + 0.12318765372037888, + 0.14002445340156555 + ], + "credit_map": [ + 0.02127985656261444, + 0.018936563283205032, + 0.02206035517156124, + 0.011760747991502285, + 0.02056884765625, + 0.03824666142463684, + 0.016228148713707924, + 0.029375987127423286, + 0.014863370917737484, + 0.011725496500730515, + 0.0671466812491417, + 0.06828322261571884, + 0.00439242459833622, + 0.0057395691983401775, + 0.012215250171720982, + 0.00852019153535366, + 0.0129522904753685, + 0.0076980069279670715, + 0.015450346283614635, + 0.012433214113116264, + 0.0027206377126276493, + 0.004644790198653936, + 0.008651698008179665, + 0.005415845662355423, + 0.008487490937113762, + 0.012636554427444935, + 0.01451780553907156, + 0.01106122974306345, + 0.008402523584663868, + 0.006394690368324518, + 0.0068369763903319836, + 0.006901709362864494, + 0.004480751231312752, + 0.0034722250420600176, + 0.0031586112454533577, + 0.0039733340963721275, + 0.0022154664620757103, + 0.0034356052055954933, + 0.003173505188897252, + 0.0027685528621077538, + 0.004490363411605358, + 0.004371462855488062, + 0.004026818089187145, + 0.002331117633730173, + 0.0023897981736809015, + 0.001809862907975912, + 0.00201663700863719, + 0.002788400277495384, + 0.012039830908179283, + 0.011944049969315529, + 0.01248825341463089, + 0.00851985439658165, + 0.004270504228770733, + 0.003568403422832489, + 0.006401062943041325, + 0.006591469049453735, + 0.026632143184542656, + 0.014636024832725525, + 0.015583946369588375, + 0.013526790775358677, + 0.00911402516067028, + 0.005639670416712761, + 0.008557458408176899, + 0.009822756983339787, + 0.028577417135238647, + 0.025468505918979645, + 0.02132805809378624, + 0.01886409893631935, + 0.007146105170249939, + 0.008007640950381756, + 0.007931594736874104, + 0.005864245817065239, + 0.02310258522629738, + 0.019360912963747978, + 0.024324709549546242, + 0.03897339105606079, + 0.008861850947141647, + 0.009075448848307133, + 0.00845450721681118, + 0.007871042937040329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.06102766469120979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018214579671621323 + ], + "opposite_logprob_tokens": [ + -4.018214702606201 + ], + "logprob_reward": -0.018214579671621323, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018214579671621323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11948411911725998, + 0.08566737174987793, + 0.06007307395339012, + 0.1894836127758026, + 0.11586467176675797, + 0.027241433039307594, + 0.11467256397008896, + 0.040596578270196915, + 0.13519230484962463, + 0.11172427982091904 + ], + "credit_map": [ + 0.020584790036082268, + 0.01630220003426075, + 0.01791398972272873, + 0.010930236428976059, + 0.011659340932965279, + 0.013450649566948414, + 0.011767346411943436, + 0.016875572502613068, + 0.009620417840778828, + 0.010402027517557144, + 0.012391355820000172, + 0.012211748398840427, + 0.0073910970240831375, + 0.010360896587371826, + 0.012677622959017754, + 0.010612205602228642, + 0.011347023770213127, + 0.009338709525763988, + 0.008126035332679749, + 0.008910974487662315, + 0.005537149030715227, + 0.00418004859238863, + 0.00529952859506011, + 0.007333601824939251, + 0.020672952756285667, + 0.022828876972198486, + 0.04159810394048691, + 0.035040922462940216, + 0.017836490646004677, + 0.011471395380795002, + 0.018987057730555534, + 0.021047810092568398, + 0.01342232059687376, + 0.011894466355443, + 0.016973914578557014, + 0.012186244130134583, + 0.010194528847932816, + 0.011314881034195423, + 0.020426427945494652, + 0.019451888278126717, + 0.005035815294831991, + 0.0036447111051529646, + 0.0044814590364694595, + 0.002941854763776064, + 0.0030717107001692057, + 0.0018540449673309922, + 0.002051425166428089, + 0.0041604116559028625, + 0.015923017635941505, + 0.012957118451595306, + 0.021968506276607513, + 0.017545927315950394, + 0.00878226663917303, + 0.006447716616094112, + 0.009699237532913685, + 0.021348778158426285, + 0.00645361328497529, + 0.0065771955996751785, + 0.006535772234201431, + 0.00781610794365406, + 0.0033151796087622643, + 0.002688596723601222, + 0.00305636296980083, + 0.004153749439865351, + 0.015462197363376617, + 0.015888335183262825, + 0.018940605223178864, + 0.041248563677072525, + 0.00705775897949934, + 0.006790023762732744, + 0.01564093492925167, + 0.01416388526558876, + 0.018610212951898575, + 0.02231290377676487, + 0.019801082089543343, + 0.021975789219141006, + 0.007275303825736046, + 0.006176459603011608, + 0.008874197490513325, + 0.006698330864310265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.018214579671621323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038125988095998764 + ], + "opposite_logprob_tokens": [ + -3.288125991821289 + ], + "logprob_reward": -0.038125988095998764, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038125988095998764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12750700116157532, + 0.12657880783081055, + 0.13900256156921387, + 0.10820391029119492, + 0.03967690095305443, + 0.10081621259450912, + 0.07312066853046417, + 0.07547315955162048, + 0.13416898250579834, + 0.0754518136382103 + ], + "credit_map": [ + 0.013768436387181282, + 0.027345385402441025, + 0.012919207103550434, + 0.014502846635878086, + 0.010523411445319653, + 0.014011317864060402, + 0.015623889863491058, + 0.018812503665685654, + 0.023538382723927498, + 0.017341116443276405, + 0.026181351393461227, + 0.01647462509572506, + 0.009694254025816917, + 0.010198350995779037, + 0.011716552078723907, + 0.011434176005423069, + 0.019245987758040428, + 0.023006130009889603, + 0.028016528114676476, + 0.02127470076084137, + 0.007321270182728767, + 0.012639048509299755, + 0.014123254455626011, + 0.013375651091337204, + 0.0136820487678051, + 0.011146889068186283, + 0.02534586563706398, + 0.021770605817437172, + 0.012330221943557262, + 0.007807868532836437, + 0.00973346084356308, + 0.006386950146406889, + 0.006682912353426218, + 0.005199807230383158, + 0.003823376726359129, + 0.005506361834704876, + 0.004110273439437151, + 0.0034630282316356897, + 0.007292645983397961, + 0.0035984960850328207, + 0.018476154655218124, + 0.015581057406961918, + 0.016903942450881004, + 0.01459580473601818, + 0.010817653499543667, + 0.007666566409170628, + 0.006792943459004164, + 0.009982091374695301, + 0.017441842705011368, + 0.010055833496153355, + 0.013354256749153137, + 0.009334379807114601, + 0.004335618577897549, + 0.0031300708651542664, + 0.005348397884517908, + 0.010120273567736149, + 0.014144107699394226, + 0.011977873742580414, + 0.013736478053033352, + 0.015453863888978958, + 0.004257583059370518, + 0.0047689261846244335, + 0.003946065902709961, + 0.007188261486589909, + 0.02711738459765911, + 0.024136412888765335, + 0.020326701924204826, + 0.019771045073866844, + 0.009329832158982754, + 0.009356556460261345, + 0.013605683110654354, + 0.010525367222726345, + 0.010909244418144226, + 0.012423459440469742, + 0.013169152662158012, + 0.0168558731675148, + 0.006068022456020117, + 0.005131950136274099, + 0.005378865636885166, + 0.005515248980373144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.038125988095998764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706075608730316 + ], + "opposite_logprob_tokens": [ + -2.127060651779175 + ], + "logprob_reward": -0.12706075608730316, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706075608730316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1394188553094864, + 0.14647862315177917, + 0.19714733958244324, + 0.09067228436470032, + 0.0929207056760788, + 0.041398994624614716, + 0.06976980715990067, + 0.058800362050533295, + 0.11691907793283463, + 0.04647395759820938 + ], + "credit_map": [ + 0.02802032232284546, + 0.016341496258974075, + 0.020068349316716194, + 0.015321633778512478, + 0.011041424237191677, + 0.019019316881895065, + 0.013735326938331127, + 0.015870973467826843, + 0.010721331462264061, + 0.010364114306867123, + 0.03788391873240471, + 0.04001791775226593, + 0.008873634971678257, + 0.01145798247307539, + 0.017208779230713844, + 0.009950936771929264, + 0.031370293349027634, + 0.02897242270410061, + 0.03962825611233711, + 0.0332954078912735, + 0.013416854664683342, + 0.019991952925920486, + 0.017945872619748116, + 0.01252627931535244, + 0.010641563683748245, + 0.014256744645535946, + 0.024218223989009857, + 0.016475681215524673, + 0.0062810322269797325, + 0.005277419928461313, + 0.007975388318300247, + 0.005546228960156441, + 0.015658589079976082, + 0.014205467887222767, + 0.014514414593577385, + 0.014202390797436237, + 0.010421219281852245, + 0.007878611795604229, + 0.010811376385390759, + 0.005228639114648104, + 0.007612713612616062, + 0.006853010505437851, + 0.006959317252039909, + 0.004782915581017733, + 0.0034938945900648832, + 0.00335982465185225, + 0.0040132650174200535, + 0.004324054345488548, + 0.012680264189839363, + 0.008092681877315044, + 0.013587230816483498, + 0.014975138939917088, + 0.0036213304847478867, + 0.0029917170759290457, + 0.006215597502887249, + 0.007605849765241146, + 0.013291207142174244, + 0.009453736245632172, + 0.008434156887233257, + 0.010157654993236065, + 0.003336502006277442, + 0.0036675806622952223, + 0.0037950167898088694, + 0.006664505694061518, + 0.015137319453060627, + 0.01462156604975462, + 0.014611732214689255, + 0.03183463215827942, + 0.007607952691614628, + 0.006640619598329067, + 0.013909963890910149, + 0.012555285356938839, + 0.007975831627845764, + 0.007339939940720797, + 0.00758892297744751, + 0.007930507883429527, + 0.004387552849948406, + 0.003960254602134228, + 0.003525089006870985, + 0.0037658573128283024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.12706075608730316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865318909287453 + ], + "opposite_logprob_tokens": [ + -3.0486531257629395 + ], + "logprob_reward": -0.04865318909287453, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865318909287453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2177952378988266, + 0.25159740447998047, + 0.04607393220067024, + 0.05264580622315407, + 0.12261875718832016, + 0.06526467949151993, + 0.056421346962451935, + 0.07309224456548691, + 0.05084219574928284, + 0.06364836543798447 + ], + "credit_map": [ + 0.046401411294937134, + 0.024189269170165062, + 0.04285323992371559, + 0.015613089315593243, + 0.012846003286540508, + 0.0281353909522295, + 0.01871049404144287, + 0.029046358540654182, + 0.01376231200993061, + 0.020990040153265, + 0.089170902967453, + 0.08282171189785004, + 0.010657381266355515, + 0.010613814927637577, + 0.012298773974180222, + 0.011282481253147125, + 0.004945008084177971, + 0.005483440589159727, + 0.009001780301332474, + 0.006410067435353994, + 0.0028805884066969156, + 0.006235474720597267, + 0.005836356431245804, + 0.005281213670969009, + 0.005674389190971851, + 0.006039670202881098, + 0.010436894372105598, + 0.007519552484154701, + 0.004841152112931013, + 0.0038614978548139334, + 0.005790309514850378, + 0.00848233699798584, + 0.018555201590061188, + 0.017538519576191902, + 0.012543206103146076, + 0.014087124727666378, + 0.01371829118579626, + 0.013933624140918255, + 0.018773909658193588, + 0.013468876481056213, + 0.012171385809779167, + 0.01068036537617445, + 0.010998903773725033, + 0.006700349971652031, + 0.007064328994601965, + 0.005652956198900938, + 0.00407499261200428, + 0.007921400479972363, + 0.011055643670260906, + 0.010922967456281185, + 0.010600244626402855, + 0.00861523486673832, + 0.002201010473072529, + 0.002361743478104472, + 0.003806514898315072, + 0.006857985630631447, + 0.01277155987918377, + 0.015565506182610989, + 0.011790754273533821, + 0.0178393367677927, + 0.0031442525796592236, + 0.004019242245703936, + 0.0034956447780132294, + 0.0044659459963440895, + 0.009227472357451916, + 0.006491193547844887, + 0.008020507171750069, + 0.010015253908932209, + 0.003487513866275549, + 0.004003561567515135, + 0.004689906723797321, + 0.004906788934022188, + 0.01320401020348072, + 0.009971676394343376, + 0.010189720429480076, + 0.011947193183004856, + 0.004262602888047695, + 0.004820899106562138, + 0.005078025162220001, + 0.004174236673861742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.04865318909287453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a meticulously orchestrated scientific procedure on a sterile stainless steel workbench, beneath bright, even overhead fluorescent lighting. A pair of hands, clad in tight-fitting blue butyl gloves, manipulate with precision. The gloved right hand first firmly grasps a clear plastic syringe with black measurement markings and shakes it quickly and deliberately to mix the clear liquid within. Simultaneously, the gloved left hand steadily slides a clear, round Petri dish across the reflective steel surface and into the center of the frame. The right hand then moves to a small glass bottle, deftly unscrewing the plastic cap and carefully placing it on a plain white cardboard box nearby. After another quick shake of the syringe, the right hand guides the needle, decisively inserting it into the bottle's rubber stopper. The left hand then carefully lifts the lid of the Petri dish and secures it above the base. The right hand tilts the syringe over the open Petri dish, carefully pressing the plunger with its thumb, releasing a few drops of clear liquid onto the culture medium within, keeping it steady. The syringe is then smoothly lifted, and the left hand instantly and precisely closes the cap, sealing the Petri dish, creating a clinical and highly professional atmosphere throughout the entire process.", + "seed": 22399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_1/index_6560_seed_22399_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.050624997913837434, + "vlm_logprob": -0.050624997913837434, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.050624997913837434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043990034610033035 + ], + "opposite_logprob_tokens": [ + -3.293990135192871 + ], + "logprob_reward": -0.043990034610033035, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.043990034610033035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24650223553180695, + 0.11998865008354187, + 0.13071101903915405, + 0.10264231264591217, + 0.048613183200359344, + 0.038962848484516144, + 0.11152230203151703, + 0.08629027754068375, + 0.07294536381959915, + 0.041821788996458054 + ], + "credit_map": [ + 0.01434391550719738, + 0.01197726745158434, + 0.01639351062476635, + 0.021800799295306206, + 0.06311798095703125, + 0.04478897154331207, + 0.03464733436703682, + 0.03943243622779846, + 0.012042556889355183, + 0.009573579765856266, + 0.016738755628466606, + 0.021303769201040268, + 0.014996998012065887, + 0.017509272322058678, + 0.013416492380201817, + 0.014407219365239143, + 0.013470674864947796, + 0.020822608843445778, + 0.02009444124996662, + 0.015386799350380898, + 0.01589144952595234, + 0.01726764254271984, + 0.011804893612861633, + 0.015972500666975975, + 0.0053309230133891106, + 0.007530989125370979, + 0.012609113939106464, + 0.02086205594241619, + 0.0160362645983696, + 0.021763380616903305, + 0.00902184471487999, + 0.00948774628341198, + 0.005501202307641506, + 0.006450287066400051, + 0.007513110525906086, + 0.0034169626887887716, + 0.007401751819998026, + 0.008958670310676098, + 0.0036211328115314245, + 0.005750063341110945, + 0.00292246718890965, + 0.006204612087458372, + 0.0035902082454413176, + 0.0024175173602998257, + 0.007126032840460539, + 0.006468835286796093, + 0.004699244163930416, + 0.005533932242542505, + 0.009907395578920841, + 0.03056630678474903, + 0.009236854501068592, + 0.005604648031294346, + 0.007539966143667698, + 0.010072954930365086, + 0.011175896041095257, + 0.02741827815771103, + 0.007668092381209135, + 0.0073659177869558334, + 0.023125285282731056, + 0.007427508942782879, + 0.00674804113805294, + 0.0077952914871275425, + 0.005243997555226088, + 0.020916137844324112, + 0.011211304925382137, + 0.007184173911809921, + 0.007013315334916115, + 0.004549399949610233, + 0.010672850534319878, + 0.014460052363574505, + 0.010477418079972267, + 0.007376844063401222, + 0.004828400909900665, + 0.005444681271910667, + 0.0046262433752417564, + 0.005659902468323708, + 0.003968146163970232, + 0.005239156540483236, + 0.006356892175972462, + 0.005698366090655327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.043990034610033035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023316025733947754 + ], + "opposite_logprob_tokens": [ + -3.773315906524658 + ], + "logprob_reward": -0.023316025733947754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023316025733947754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13697035610675812, + 0.2337687909603119, + 0.13929259777069092, + 0.15164463222026825, + 0.05618596076965332, + 0.033516645431518555, + 0.025286255404353142, + 0.14170564711093903, + 0.04804942011833191, + 0.03357965499162674 + ], + "credit_map": [ + 0.017538035288453102, + 0.014993726275861263, + 0.014643818140029907, + 0.019729584455490112, + 0.01937611773610115, + 0.0170486681163311, + 0.01770809292793274, + 0.015932321548461914, + 0.024470752105116844, + 0.018092801794409752, + 0.0336427316069603, + 0.041518766433000565, + 0.025690989568829536, + 0.03300820291042328, + 0.028482133522629738, + 0.02886243909597397, + 0.018581729382276535, + 0.023218655958771706, + 0.017914006486535072, + 0.010623078793287277, + 0.019501907750964165, + 0.01547801773995161, + 0.017767932265996933, + 0.016207290813326836, + 0.012689519673585892, + 0.01749611645936966, + 0.021148771047592163, + 0.023839876055717468, + 0.017619619145989418, + 0.02674090676009655, + 0.014353048987686634, + 0.017756784334778786, + 0.005843760911375284, + 0.0077783395536243916, + 0.007890996523201466, + 0.006339463405311108, + 0.007464696653187275, + 0.009778928942978382, + 0.004600265063345432, + 0.006489516235888004, + 0.004024879075586796, + 0.0054027424193918705, + 0.003159677144140005, + 0.002429267857223749, + 0.005664190277457237, + 0.005007219035178423, + 0.003484599757939577, + 0.004344069864600897, + 0.0025997571647167206, + 0.004738784860819578, + 0.002417138312011957, + 0.0022786809131503105, + 0.003247757675126195, + 0.0035525320563465357, + 0.0021192452404648066, + 0.004332363605499268, + 0.011788611300289631, + 0.0093100406229496, + 0.039357610046863556, + 0.012406421825289726, + 0.009742937982082367, + 0.01281694695353508, + 0.009711289778351784, + 0.03657180443406105, + 0.008079488761723042, + 0.007231155410408974, + 0.004186117090284824, + 0.003928587306290865, + 0.005046415608376265, + 0.0071761771105229855, + 0.0062231384217739105, + 0.006178345065563917, + 0.003724395763128996, + 0.005691599566489458, + 0.004867302253842354, + 0.0031496109440922737, + 0.004011052194982767, + 0.004358807113021612, + 0.0033663352951407433, + 0.004410553723573685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.023316025733947754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029811974614858627 + ], + "opposite_logprob_tokens": [ + -3.5298118591308594 + ], + "logprob_reward": -0.029811974614858627, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811974614858627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06839392334222794, + 0.43383780121803284, + 0.07532676309347153, + 0.07732205092906952, + 0.059737950563430786, + 0.015684738755226135, + 0.04423629492521286, + 0.10858660191297531, + 0.0878552719950676, + 0.029018618166446686 + ], + "credit_map": [ + 0.00674691004678607, + 0.008611495606601238, + 0.00699334591627121, + 0.0076153697445988655, + 0.009160363115370274, + 0.011114891618490219, + 0.008083687163889408, + 0.01006785873323679, + 0.04249004274606705, + 0.034153953194618225, + 0.062057506293058395, + 0.07833798229694366, + 0.048393234610557556, + 0.062114037573337555, + 0.05231339856982231, + 0.053977612406015396, + 0.008658153004944324, + 0.013281098566949368, + 0.009908270090818405, + 0.005575144197791815, + 0.00946837104856968, + 0.010131759569048882, + 0.01025270577520132, + 0.008051259443163872, + 0.007189089432358742, + 0.007723658345639706, + 0.010588021017611027, + 0.008752867579460144, + 0.009163237176835537, + 0.01124527957290411, + 0.005264237988740206, + 0.017395658418536186, + 0.006668677553534508, + 0.007846453227102757, + 0.011571909300982952, + 0.007770889438688755, + 0.00824817642569542, + 0.009197442792356014, + 0.0035160360857844353, + 0.004918363410979509, + 0.001362979761324823, + 0.0020582901779562235, + 0.0016087752301245928, + 0.0009473968530073762, + 0.003012903267517686, + 0.0032188708428293467, + 0.0010117479832842946, + 0.0024637735914438963, + 0.0036368623841553926, + 0.012401612475514412, + 0.0036076901014894247, + 0.0025292213540524244, + 0.0032047873828560114, + 0.0049848295748233795, + 0.0029396379832178354, + 0.010931652039289474, + 0.009224890731275082, + 0.009189398027956486, + 0.02537214569747448, + 0.01197162363678217, + 0.007909024134278297, + 0.010477320291101933, + 0.009227429516613483, + 0.025214768946170807, + 0.014719715341925621, + 0.015244930982589722, + 0.008890505880117416, + 0.007704537361860275, + 0.010146440006792545, + 0.01142432726919651, + 0.009350108914077282, + 0.010374709032475948, + 0.005345238372683525, + 0.004746638238430023, + 0.0043333047069609165, + 0.0027684890665113926, + 0.002578843617811799, + 0.002575053833425045, + 0.0032672849483788013, + 0.003403766779229045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.029811974614858627 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129462942481041 + ], + "opposite_logprob_tokens": [ + -2.2379462718963623 + ], + "logprob_reward": -0.1129462942481041, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129462942481041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16553469002246857, + 0.17440937459468842, + 0.10660340636968613, + 0.0893329456448555, + 0.06095576658844948, + 0.04057734087109566, + 0.028092486783862114, + 0.07631334662437439, + 0.2275872677564621, + 0.03059336356818676 + ], + "credit_map": [ + 0.01051198597997427, + 0.011971070431172848, + 0.010557874105870724, + 0.015186556614935398, + 0.04072253406047821, + 0.026079032570123672, + 0.024377649649977684, + 0.026127979159355164, + 0.02018643356859684, + 0.01433894969522953, + 0.025678055360913277, + 0.02823689579963684, + 0.01908188872039318, + 0.023886043578386307, + 0.02048521488904953, + 0.022515885531902313, + 0.009447988122701645, + 0.02127179689705372, + 0.009467764757573605, + 0.00908021628856659, + 0.017816966399550438, + 0.015509354881942272, + 0.012770081870257854, + 0.011239243671298027, + 0.008706875145435333, + 0.009600606746971607, + 0.012015555053949356, + 0.012260757386684418, + 0.009868563152849674, + 0.010800452902913094, + 0.0068745301105082035, + 0.01920560747385025, + 0.006322046276181936, + 0.007645171135663986, + 0.009413359686732292, + 0.007072606589645147, + 0.008305066265165806, + 0.009353049099445343, + 0.004994845017790794, + 0.007849622517824173, + 0.003073224099352956, + 0.005809268914163113, + 0.006529681850224733, + 0.0038883488159626722, + 0.0063877832144498825, + 0.007595485541969538, + 0.0028541628271341324, + 0.004439385607838631, + 0.002469436964020133, + 0.0037990163546055555, + 0.0036146813072264194, + 0.0019810767844319344, + 0.0021184878423810005, + 0.0045523690059781075, + 0.004743831232190132, + 0.004813586827367544, + 0.006910253316164017, + 0.009642766788601875, + 0.01951475627720356, + 0.0058671776205301285, + 0.00852354709059, + 0.0069731962867081165, + 0.003468835959210992, + 0.01541281770914793, + 0.03311985731124878, + 0.03897267207503319, + 0.022695722058415413, + 0.018587736412882805, + 0.019534606486558914, + 0.035318925976753235, + 0.035423941910266876, + 0.023933809250593185, + 0.005189210642129183, + 0.0036707979161292315, + 0.003270010696724057, + 0.003857692703604698, + 0.003134560538455844, + 0.0032467416021972895, + 0.004494420252740383, + 0.0037299287505447865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.1129462942481041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static high-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands wearing tight-fitting blue butyl gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the right hand shake a clear plastic syringe with black measurement markings to mix liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the right hand insert the syringe needle into a small glass bottle with a rubber stopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the right hand release clear liquid from the syringe into a clear round Petri dish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306066036224365 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306066036224365, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306066036224365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16759276390075684, + 0.2084105759859085, + 0.04715057462453842, + 0.1003037542104721, + 0.02966775745153427, + 0.04096603021025658, + 0.12721188366413116, + 0.15756277740001678, + 0.06936942785978317, + 0.051764484494924545 + ], + "credit_map": [ + 0.009065482765436172, + 0.010540959425270557, + 0.012034773826599121, + 0.015002570115029812, + 0.03709021210670471, + 0.028583737090229988, + 0.024658765643835068, + 0.03061625361442566, + 0.016101012006402016, + 0.014470398426055908, + 0.02953951060771942, + 0.04059246927499771, + 0.022453583776950836, + 0.030476650223135948, + 0.02969808503985405, + 0.025078872218728065, + 0.007645746227353811, + 0.008580132387578487, + 0.004699439741671085, + 0.00401023356243968, + 0.005370819475501776, + 0.0049775405786931515, + 0.0065963189117610455, + 0.005270346067845821, + 0.011506009846925735, + 0.009166240692138672, + 0.011863816529512405, + 0.010210978798568249, + 0.012073718942701817, + 0.01464855670928955, + 0.005728443153202534, + 0.02510599046945572, + 0.003731240751221776, + 0.004048217087984085, + 0.004336829297244549, + 0.002711581764742732, + 0.00393562950193882, + 0.004402542486786842, + 0.0026245457120239735, + 0.00387717061676085, + 0.004188344348222017, + 0.005313631612807512, + 0.005266275256872177, + 0.0037874991539865732, + 0.006669638212770224, + 0.006928903982043266, + 0.0033196841832250357, + 0.00549205532297492, + 0.011007294990122318, + 0.03597402200102806, + 0.007351868785917759, + 0.006682716775685549, + 0.008543468080461025, + 0.010914133861660957, + 0.014046194963157177, + 0.032692182809114456, + 0.015201244503259659, + 0.02187945507466793, + 0.026506591588258743, + 0.014210781082510948, + 0.022957896813750267, + 0.01916179433465004, + 0.014834316447377205, + 0.022810695692896843, + 0.007738564629107714, + 0.010083737783133984, + 0.0072138686664402485, + 0.005755133926868439, + 0.009510759264230728, + 0.010406491346657276, + 0.011776668019592762, + 0.006884200964123011, + 0.008909854106605053, + 0.008214598521590233, + 0.007969182915985584, + 0.004965946543961763, + 0.005275635048747063, + 0.005373573862016201, + 0.0051857223734259605, + 0.005869971588253975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.04306066036224365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_tensor([2988, 2988])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_tensor([2988, 2988])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0e83c87e2e84756878b34af3137190c0daa72c16 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_tensor([2988, 2988])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro close-up, shot at an extremely slow speed, tilted upwards from below, focuses on a massive jeweled egg, carved from a material resembling moonstone or translucent crystal. Captured at a high frame rate, the scene appears as extreme slow motion. The eggshell, polished to a mirror-smooth finish, is meticulously engraved with intricate, ancient runes. From within, a warm, soft white light steadily emanates, illuminating the entire egg and making the grooves within the runes seem to flow with a faint energy. Several fine, natural cracks dot the eggshell, through which crystal-clear water slowly and steadily seeps. Each plump droplet slides carefully along the graceful curve of the eggshell, its surface tension maintaining its perfect spherical shape and clearly reflecting the detailed runes and the soft light filtering through. As it slides, the droplet leaves a fleeting, slightly damp mark on the smooth surface. Set against a pure black background, the entire image is completely illuminated by the inner light of the egg, presenting an extremely pure texture and an otherworldly luster, creating an atmosphere that is sacred, quiet and full of the miracle of life.", + "seed": 32816, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_2988_seed_32816_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2652937933802605, + "vlm_logprob": -0.2652937933802605, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.2652937933802605, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05749718099832535 + ], + "opposite_logprob_tokens": [ + -2.932497262954712 + ], + "logprob_reward": -0.05749718099832535, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05749718099832535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23185186088085175, + 0.08276430517435074, + 0.050394441932439804, + 0.07396570593118668, + 0.1259365975856781, + 0.050139620900154114, + 0.07957176864147186, + 0.027800776064395905, + 0.08260180056095123, + 0.19497321546077728 + ], + "credit_map": [ + 0.010038566775619984, + 0.01392182894051075, + 0.02465173415839672, + 0.019842995330691338, + 0.0245222020894289, + 0.01387710776180029, + 0.057538192719221115, + 0.06745921820402145, + 0.0048363772220909595, + 0.004970751237124205, + 0.013835939578711987, + 0.010517922230064869, + 0.015814175829291344, + 0.00480850413441658, + 0.01617630012333393, + 0.01180432178080082, + 0.0025151066947728395, + 0.002833849983289838, + 0.005909828934818506, + 0.009670188650488853, + 0.005528884474188089, + 0.0030281650833785534, + 0.008711747825145721, + 0.012196664698421955, + 0.0038099917583167553, + 0.006137082818895578, + 0.013609075918793678, + 0.010437854565680027, + 0.015655795112252235, + 0.003909251186996698, + 0.011686216108500957, + 0.00872043427079916, + 0.00902184471487999, + 0.009508899413049221, + 0.020592205226421356, + 0.02119552530348301, + 0.017218412831425667, + 0.008463205769658089, + 0.02301264926791191, + 0.01692385971546173, + 0.002430737018585205, + 0.004833003971725702, + 0.009848997928202152, + 0.00788445957005024, + 0.005299444310367107, + 0.0034929702524095774, + 0.009967342019081116, + 0.0063826581463217735, + 0.006958675105124712, + 0.008940384723246098, + 0.011273907497525215, + 0.01161948125809431, + 0.010286842472851276, + 0.005822749808430672, + 0.015645550563931465, + 0.00902416929602623, + 0.0018127845833078027, + 0.0029411788564175367, + 0.004878162406384945, + 0.0033626605290919542, + 0.0029215016402304173, + 0.002337503246963024, + 0.004983959719538689, + 0.004563022404909134, + 0.004022715613245964, + 0.008388453163206577, + 0.014864516444504261, + 0.010949152521789074, + 0.018193252384662628, + 0.005102128721773624, + 0.013581548817455769, + 0.007500024512410164, + 0.011774538084864616, + 0.015311624854803085, + 0.01608225516974926, + 0.04442337527871132, + 0.017361968755722046, + 0.012159422039985657, + 0.022444823756814003, + 0.055415183305740356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.05749718099832535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9744046926498413 + ], + "opposite_logprob_tokens": [ + -0.4744046926498413 + ], + "logprob_reward": -0.9744046926498413, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9744046926498413, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14271754026412964, + 0.06293173879384995, + 0.10890654474496841, + 0.2013763189315796, + 0.19529545307159424, + 0.02758358046412468, + 0.03404653072357178, + 0.033092062920331955, + 0.09696550667285919, + 0.09708468616008759 + ], + "credit_map": [ + 0.006862789858132601, + 0.016963478177785873, + 0.02558285742998123, + 0.014877512119710445, + 0.027259282767772675, + 0.011899873614311218, + 0.02033928409218788, + 0.018932469189167023, + 0.0035736560821533203, + 0.005411758087575436, + 0.014351565390825272, + 0.007920050993561745, + 0.010946573689579964, + 0.004346690606325865, + 0.008823773823678493, + 0.007557670585811138, + 0.003995646722614765, + 0.007839399389922619, + 0.014768682420253754, + 0.017786899581551552, + 0.022437041625380516, + 0.006393264513462782, + 0.020894132554531097, + 0.01479148119688034, + 0.007859169505536556, + 0.01694757305085659, + 0.032432667911052704, + 0.024635186418890953, + 0.04698033630847931, + 0.012019354850053787, + 0.03522542119026184, + 0.025276636704802513, + 0.007418634835630655, + 0.013630148954689503, + 0.03012882173061371, + 0.026540206745266914, + 0.042046528309583664, + 0.011385800316929817, + 0.037725962698459625, + 0.02641935832798481, + 0.001260167919099331, + 0.0020375894382596016, + 0.004933849908411503, + 0.0036784259136766195, + 0.0030489335767924786, + 0.00233853654935956, + 0.003886668710038066, + 0.006399409845471382, + 0.002462050411850214, + 0.002973719732835889, + 0.007116451393812895, + 0.005469790659844875, + 0.0036309855058789253, + 0.002187065314501524, + 0.00631677033379674, + 0.0038897027261555195, + 0.0022852541878819466, + 0.00353886722587049, + 0.0061821104027330875, + 0.004779470153152943, + 0.0033193903509527445, + 0.0023829657584428787, + 0.005030322819948196, + 0.005573686212301254, + 0.005376592744141817, + 0.005471527110785246, + 0.01308920606970787, + 0.019614772871136665, + 0.011401131749153137, + 0.003530556336045265, + 0.027485603466629982, + 0.010996131226420403, + 0.004810255486518145, + 0.007309476844966412, + 0.0077148014679551125, + 0.02165571041405201, + 0.012047777883708477, + 0.00565493106842041, + 0.010376734659075737, + 0.027514994144439697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.9744046926498413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16045604646205902 + ], + "opposite_logprob_tokens": [ + -1.9104560613632202 + ], + "logprob_reward": -0.16045604646205902, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16045604646205902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09075197577476501, + 0.06059735640883446, + 0.17472834885120392, + 0.05987846106290817, + 0.2986215651035309, + 0.023005640134215355, + 0.07174667716026306, + 0.024511640891432762, + 0.06175363063812256, + 0.13440465927124023 + ], + "credit_map": [ + 0.00499667227268219, + 0.007124507799744606, + 0.017361057922244072, + 0.011899341829121113, + 0.018596310168504715, + 0.005326738115400076, + 0.012355427257716656, + 0.0130919199436903, + 0.0033693676814436913, + 0.004621320869773626, + 0.010459230281412601, + 0.008379126898944378, + 0.010142210870981216, + 0.003655931679531932, + 0.012003729119896889, + 0.007966439239680767, + 0.00748111167922616, + 0.009954246692359447, + 0.02463693916797638, + 0.04537109285593033, + 0.021861325949430466, + 0.008639108389616013, + 0.019508985802531242, + 0.03727554529905319, + 0.005053970962762833, + 0.004283561371266842, + 0.00923524983227253, + 0.011243564076721668, + 0.013192887417972088, + 0.0030010028276592493, + 0.008446951396763325, + 0.00542127201333642, + 0.0151706263422966, + 0.023591015487909317, + 0.04852083697915077, + 0.03921777755022049, + 0.05311606451869011, + 0.018367527052760124, + 0.05806248262524605, + 0.042575228959321976, + 0.0007685772143304348, + 0.0020351491402834654, + 0.004513300023972988, + 0.003423742949962616, + 0.003103023860603571, + 0.0013025328516960144, + 0.004538880195468664, + 0.0033204350620508194, + 0.005371284205466509, + 0.00669526681303978, + 0.014354539103806019, + 0.010610065422952175, + 0.009153109975159168, + 0.004858165048062801, + 0.01254100538790226, + 0.008163240738213062, + 0.001719060936011374, + 0.0029643720481544733, + 0.0035703941248357296, + 0.003867312101647258, + 0.0027936187107115984, + 0.0011881034588441253, + 0.004002111032605171, + 0.004406669642776251, + 0.0034683675039559603, + 0.006199303083121777, + 0.011430887505412102, + 0.008455507457256317, + 0.00863710418343544, + 0.0034472919069230556, + 0.013423670083284378, + 0.006691500544548035, + 0.007943267934024334, + 0.00928429514169693, + 0.01054296176880598, + 0.03244347125291824, + 0.011915110982954502, + 0.007913215085864067, + 0.014269557781517506, + 0.04009278863668442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.16045604646205902 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003900095820427 + ], + "opposite_logprob_tokens": [ + -2.3503899574279785 + ], + "logprob_reward": -0.1003900095820427, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003900095820427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10060273110866547, + 0.09216092526912689, + 0.30514395236968994, + 0.060611266642808914, + 0.10937508940696716, + 0.030188128352165222, + 0.0724177435040474, + 0.0204660352319479, + 0.16829703748226166, + 0.0407370962202549 + ], + "credit_map": [ + 0.00532516511157155, + 0.009659034200012684, + 0.013245750218629837, + 0.009584315121173859, + 0.012429801747202873, + 0.00830000452697277, + 0.018120286986231804, + 0.023938369005918503, + 0.004628945607692003, + 0.00732054840773344, + 0.019759051501750946, + 0.012310588732361794, + 0.012089810334146023, + 0.005262557417154312, + 0.014292296953499317, + 0.016497133299708366, + 0.010588024742901325, + 0.019002655521035194, + 0.0407072938978672, + 0.061305370181798935, + 0.06155857816338539, + 0.01565372571349144, + 0.05191932991147041, + 0.04440898075699806, + 0.0034240693785250187, + 0.0050843083299696445, + 0.01107831858098507, + 0.008060946129262447, + 0.0107105178758502, + 0.0033046884927898645, + 0.010290242731571198, + 0.00865817628800869, + 0.00620014313608408, + 0.008236708119511604, + 0.018661536276340485, + 0.016472049057483673, + 0.018726972863078117, + 0.005789883900433779, + 0.020723514258861542, + 0.014564287848770618, + 0.0013362473109737039, + 0.0021610711701214314, + 0.004576662089675665, + 0.0063619352877140045, + 0.0032087229192256927, + 0.0015853282529860735, + 0.006383280735462904, + 0.004574878606945276, + 0.0052556530572474, + 0.004906246438622475, + 0.012628073804080486, + 0.01286002341657877, + 0.011986345052719116, + 0.0037949206307530403, + 0.01180180162191391, + 0.009184679947793484, + 0.0011146097676828504, + 0.002251756377518177, + 0.003717664396390319, + 0.003761539002880454, + 0.001996807986870408, + 0.0012844069860875607, + 0.0033291387371718884, + 0.0030101125594228506, + 0.008219724521040916, + 0.018071891739964485, + 0.030854765325784683, + 0.023044930770993233, + 0.03293261304497719, + 0.010704359039664268, + 0.02300111949443817, + 0.02146763913333416, + 0.0024063924793154, + 0.0026602146681398153, + 0.0038631849456578493, + 0.010299154557287693, + 0.0030924950260668993, + 0.002297700848430395, + 0.003949510864913464, + 0.012168445624411106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.1003900095820427 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033721037209033966 + ], + "opposite_logprob_tokens": [ + -3.4087209701538086 + ], + "logprob_reward": -0.033721037209033966, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033721037209033966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13798154890537262, + 0.05254446342587471, + 0.0929756835103035, + 0.17197036743164062, + 0.36460980772972107, + 0.033832039684057236, + 0.06586810946464539, + 0.025768669322133064, + 0.011333542875945568, + 0.04311573877930641 + ], + "credit_map": [ + 0.016759369522333145, + 0.013397252187132835, + 0.020312439650297165, + 0.016289161518216133, + 0.02299673855304718, + 0.008226500824093819, + 0.021279146894812584, + 0.01872093230485916, + 0.003962135408073664, + 0.00521805277094245, + 0.009938299655914307, + 0.005798115860670805, + 0.008867772296071053, + 0.003909225109964609, + 0.007788642309606075, + 0.007062217686325312, + 0.008623919449746609, + 0.005698380991816521, + 0.00941151287406683, + 0.008970662020146847, + 0.005505945533514023, + 0.005219204351305962, + 0.02220326103270054, + 0.027342796325683594, + 0.01538514532148838, + 0.015883637592196465, + 0.022002091631293297, + 0.02777947299182415, + 0.023134108632802963, + 0.014792492613196373, + 0.02572440169751644, + 0.027269016951322556, + 0.04494863376021385, + 0.031750891357660294, + 0.057725388556718826, + 0.05977635830640793, + 0.03757504001259804, + 0.02558549866080284, + 0.05728776752948761, + 0.04996024817228317, + 0.0029268276412039995, + 0.002387367421761155, + 0.0046138158068060875, + 0.0055945999920368195, + 0.004790003411471844, + 0.001727779977954924, + 0.005716539919376373, + 0.006075103767216206, + 0.006149404216557741, + 0.005641003604978323, + 0.009917672723531723, + 0.011419006623327732, + 0.007639504503458738, + 0.0037810138892382383, + 0.01145949400961399, + 0.009861010126769543, + 0.0015377035597339272, + 0.0023807573597878218, + 0.004555848892778158, + 0.004758947994560003, + 0.00268946448341012, + 0.0017473954940214753, + 0.003975024912506342, + 0.0041235266253352165, + 0.0006336382939480245, + 0.0008416391792707145, + 0.0017615009564906359, + 0.0015970024978742003, + 0.0018526247004047036, + 0.000721437914762646, + 0.0024528345093131065, + 0.0014728637179359794, + 0.0029376335442066193, + 0.0021379722747951746, + 0.003587951185181737, + 0.01138261891901493, + 0.0020662930328398943, + 0.0026553268544375896, + 0.004433267284184694, + 0.013914673589169979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.033721037209033966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, shot at an extremely slow speed, tilted upwards from below, focuses on a massive jeweled egg, carved from a material resembling moonstone or translucent crystal. Captured at a high frame rate, the scene appears as extreme slow motion. The eggshell, polished to a mirror-smooth finish, is meticulously engraved with intricate, ancient runes. From within, a warm, soft white light steadily emanates, illuminating the entire egg and making the grooves within the runes seem to flow with a faint energy. Several fine, natural cracks dot the eggshell, through which crystal-clear water slowly and steadily seeps. Each plump droplet slides carefully along the graceful curve of the eggshell, its surface tension maintaining its perfect spherical shape and clearly reflecting the detailed runes and the soft light filtering through. As it slides, the droplet leaves a fleeting, slightly damp mark on the smooth surface. Set against a pure black background, the entire image is completely illuminated by the inner light of the egg, presenting an extremely pure texture and an otherworldly luster, creating an atmosphere that is sacred, quiet and full of the miracle of life.", + "seed": 32817, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_2988_seed_32817_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2331987701356411, + "vlm_logprob": -0.2331987701356411, + "vlm_yes_no_margin": 1.85, + "vlm_token_credit": -0.2331987701356411, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05750708654522896 + ], + "opposite_logprob_tokens": [ + -2.932507038116455 + ], + "logprob_reward": -0.05750708654522896, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05750708654522896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.139273539185524, + 0.056898582726716995, + 0.05142040550708771, + 0.024131089448928833, + 0.05180582031607628, + 0.03290138021111488, + 0.039221517741680145, + 0.11794772744178772, + 0.04003968462347984, + 0.4463602602481842 + ], + "credit_map": [ + 0.007353090215474367, + 0.024667996913194656, + 0.009688768535852432, + 0.019774416461586952, + 0.028431931510567665, + 0.015641242265701294, + 0.018009858205914497, + 0.015706228092312813, + 0.003233330324292183, + 0.00874425284564495, + 0.0037282509729266167, + 0.008742225356400013, + 0.010777847841382027, + 0.005370212718844414, + 0.007469372358173132, + 0.008833089843392372, + 0.0017663498874753714, + 0.007023300044238567, + 0.002337388228625059, + 0.008293814957141876, + 0.01336820051074028, + 0.0054008145816624165, + 0.005428140051662922, + 0.007802395150065422, + 0.0010504028759896755, + 0.0036892800126224756, + 0.001502299215644598, + 0.003261619247496128, + 0.003896798938512802, + 0.0026521196123212576, + 0.0032207337208092213, + 0.004857837222516537, + 0.0037689690943807364, + 0.005309142172336578, + 0.004262519534677267, + 0.008774850517511368, + 0.010067366994917393, + 0.004773502703756094, + 0.00793883204460144, + 0.0069106388837099075, + 0.0014193160459399223, + 0.005463433917611837, + 0.001850963570177555, + 0.005598708521574736, + 0.0068016150034964085, + 0.00292311143130064, + 0.004853593651205301, + 0.003990636207163334, + 0.0018731789896264672, + 0.0036492508370429277, + 0.0017025175038725138, + 0.009038900025188923, + 0.005373172461986542, + 0.002595399273559451, + 0.011309681460261345, + 0.003679415676742792, + 0.0043830000795423985, + 0.014005637727677822, + 0.006081598810851574, + 0.023701202124357224, + 0.022458134219050407, + 0.00902964361011982, + 0.02347957342863083, + 0.014808934181928635, + 0.0019491268321871758, + 0.0037640798836946487, + 0.002215625951066613, + 0.005905301310122013, + 0.008938726037740707, + 0.00270988536067307, + 0.008309231139719486, + 0.006247710902243853, + 0.019038692116737366, + 0.0238514244556427, + 0.02165145054459572, + 0.15451118350028992, + 0.06445787847042084, + 0.01942082680761814, + 0.10367446392774582, + 0.039754319936037064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.05750708654522896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.758097231388092 + ], + "opposite_logprob_tokens": [ + -0.633097231388092 + ], + "logprob_reward": -0.758097231388092, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.758097231388092, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06515560299158096, + 0.05013245344161987, + 0.06596225500106812, + 0.014694706536829472, + 0.022685140371322632, + 0.014722784049808979, + 0.043130062520504, + 0.10725980997085571, + 0.037697792053222656, + 0.5785594582557678 + ], + "credit_map": [ + 0.003305587451905012, + 0.010558070614933968, + 0.0035074353218078613, + 0.009047823958098888, + 0.011586545035243034, + 0.007586352527141571, + 0.008208008483052254, + 0.01135577354580164, + 0.0027227068785578012, + 0.006078092381358147, + 0.0032648746855556965, + 0.008681830018758774, + 0.009621203877031803, + 0.004098987206816673, + 0.006709055043756962, + 0.008955704048275948, + 0.002558153122663498, + 0.008974703028798103, + 0.003286351915448904, + 0.01032939087599516, + 0.01600349135696888, + 0.00512562645599246, + 0.009063840843737125, + 0.010620692744851112, + 0.0008746628882363439, + 0.002410015556961298, + 0.0007658054819330573, + 0.0019783242605626583, + 0.0026487030554562807, + 0.0014474802883341908, + 0.0021481646690517664, + 0.0024215509183704853, + 0.001345346332527697, + 0.00323427258990705, + 0.001933336490765214, + 0.0034489163663238287, + 0.0035367512609809637, + 0.002072889357805252, + 0.004039646591991186, + 0.003073981963098049, + 0.0007779632578603923, + 0.002487855963408947, + 0.0011977175017818809, + 0.00204148655757308, + 0.002813598606735468, + 0.0013625783612951636, + 0.002431596163660288, + 0.001609987928532064, + 0.0017473107436671853, + 0.005514953751116991, + 0.0022993171587586403, + 0.008397447876632214, + 0.007519289385527372, + 0.00261375168338418, + 0.01125329453498125, + 0.003784696338698268, + 0.003637513145804405, + 0.012765203602612019, + 0.004858207423239946, + 0.01826792024075985, + 0.02846175618469715, + 0.0065619018860161304, + 0.02224176935851574, + 0.010465542785823345, + 0.0016703850124031305, + 0.0038871830329298973, + 0.001980886096134782, + 0.0072788214311003685, + 0.00852617621421814, + 0.0018497479613870382, + 0.007705173920840025, + 0.004799419082701206, + 0.022488078102469444, + 0.03156382218003273, + 0.024331312626600266, + 0.18506687879562378, + 0.0973580926656723, + 0.02446487359702587, + 0.14537465572357178, + 0.047911711037158966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.758097231388092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007994502782822 + ], + "opposite_logprob_tokens": [ + -2.695080041885376 + ], + "logprob_reward": -0.07007994502782822, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007994502782822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10150061547756195, + 0.08942507207393646, + 0.07871513068675995, + 0.013646100647747517, + 0.019237563014030457, + 0.032832179218530655, + 0.07872175425291061, + 0.128350168466568, + 0.038496263325214386, + 0.41907525062561035 + ], + "credit_map": [ + 0.005024500656872988, + 0.01614365726709366, + 0.006102858576923609, + 0.013656930066645145, + 0.017178967595100403, + 0.011083096265792847, + 0.011904405429959297, + 0.020406173542141914, + 0.004662870895117521, + 0.011707041412591934, + 0.006049684248864651, + 0.012449266389012337, + 0.016818754374980927, + 0.00791681744158268, + 0.011575480923056602, + 0.018245143815875053, + 0.0018791115144267678, + 0.010209865868091583, + 0.0028301780112087727, + 0.014649264514446259, + 0.023617422208189964, + 0.005808683577924967, + 0.009185832925140858, + 0.01053475309163332, + 0.0005333792068995535, + 0.0018941935850307345, + 0.0010053711012005806, + 0.0017206481425091624, + 0.0022531875874847174, + 0.0016746516339480877, + 0.001544271595776081, + 0.0030203950591385365, + 0.0010141583625227213, + 0.002662685001268983, + 0.0013015319127589464, + 0.0031020729802548885, + 0.004449997562915087, + 0.0017464925767853856, + 0.0025602553505450487, + 0.002400363562628627, + 0.0017100432887673378, + 0.0056882514618337154, + 0.002122972859069705, + 0.004677299410104752, + 0.005291699431836605, + 0.003949264530092478, + 0.00495290569961071, + 0.004439734388142824, + 0.004480809438973665, + 0.007189772557467222, + 0.0038211033679544926, + 0.022164970636367798, + 0.013095377013087273, + 0.0048659020103514194, + 0.014204071834683418, + 0.008899732492864132, + 0.0038227480836212635, + 0.014493314549326897, + 0.005738039966672659, + 0.02784796990454197, + 0.028763050213456154, + 0.009590374305844307, + 0.021926062181591988, + 0.016168592497706413, + 0.0015591143164783716, + 0.002597959479317069, + 0.002184545388445258, + 0.0076935747638344765, + 0.006883368361741304, + 0.0023141715209931135, + 0.007116351742297411, + 0.008147169835865498, + 0.017746862024068832, + 0.02401137165725231, + 0.020772123709321022, + 0.13146543502807617, + 0.07037434726953506, + 0.019539520144462585, + 0.09548846632242203, + 0.039677057415246964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.07007994502782822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10037900507450104 + ], + "opposite_logprob_tokens": [ + -2.35037899017334 + ], + "logprob_reward": -0.10037900507450104, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10037900507450104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06441348791122437, + 0.09515117108821869, + 0.0973716601729393, + 0.05125142261385918, + 0.04897955805063248, + 0.04686248302459717, + 0.07972924411296844, + 0.35744813084602356, + 0.045331671833992004, + 0.11346114426851273 + ], + "credit_map": [ + 0.004888809286057949, + 0.010341218672692776, + 0.004260194022208452, + 0.007707097567617893, + 0.010838338173925877, + 0.008989639580249786, + 0.00931040570139885, + 0.00807778537273407, + 0.005829227156937122, + 0.011968529783189297, + 0.005756002850830555, + 0.012825739569962025, + 0.020307159051299095, + 0.009214087389409542, + 0.011441889218986034, + 0.017808541655540466, + 0.004527238663285971, + 0.013974900357425213, + 0.005801455583423376, + 0.013758910819888115, + 0.022969180718064308, + 0.010202309116721153, + 0.011971402913331985, + 0.01416626013815403, + 0.00226507056504488, + 0.006794343702495098, + 0.0035179259721189737, + 0.006626034155488014, + 0.010464518330991268, + 0.005505938548594713, + 0.007133531849831343, + 0.008944055996835232, + 0.0024989282246679068, + 0.00712534599006176, + 0.003431099932640791, + 0.007086927071213722, + 0.0094913886860013, + 0.005229437258094549, + 0.0075632124207913876, + 0.006553217303007841, + 0.002313950564712286, + 0.01136683952063322, + 0.002796551678329706, + 0.0064329719170928, + 0.007693960797041655, + 0.003978106193244457, + 0.005973807070404291, + 0.00630629388615489, + 0.003458176041021943, + 0.007398115936666727, + 0.003346774261444807, + 0.01858595758676529, + 0.010250632651150227, + 0.005322492215782404, + 0.023213524371385574, + 0.008153575472533703, + 0.013926754705607891, + 0.03401963412761688, + 0.018763624131679535, + 0.08401202410459518, + 0.06068916618824005, + 0.024178635329008102, + 0.0739494115114212, + 0.047908857464790344, + 0.0020505120046436787, + 0.004094230011105537, + 0.0027728655841201544, + 0.010510467924177647, + 0.009197288192808628, + 0.0022587282583117485, + 0.008471156470477581, + 0.005976421758532524, + 0.006182100158184767, + 0.008143088780343533, + 0.006652966607362032, + 0.017123691737651825, + 0.026501310989260674, + 0.006365489214658737, + 0.028345027938485146, + 0.014147469773888588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.10037900507450104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993058264255524 + ], + "opposite_logprob_tokens": [ + -1.804930567741394 + ], + "logprob_reward": -0.17993058264255524, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993058264255524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2293800562620163, + 0.06141577288508415, + 0.15062075853347778, + 0.08962064236402512, + 0.11948087811470032, + 0.07316498458385468, + 0.030456949025392532, + 0.07440197467803955, + 0.044823311269283295, + 0.12663471698760986 + ], + "credit_map": [ + 0.015467538498342037, + 0.03791697695851326, + 0.0144553417339921, + 0.03682988882064819, + 0.048713814467191696, + 0.026463337242603302, + 0.02665950171649456, + 0.022873662412166595, + 0.004067640285938978, + 0.008504565805196762, + 0.004140661563724279, + 0.009931210428476334, + 0.009869596920907497, + 0.004982312209904194, + 0.008901834487915039, + 0.011017945595085621, + 0.005938026588410139, + 0.027567511424422264, + 0.009221309795975685, + 0.021329659968614578, + 0.026765482500195503, + 0.019142458215355873, + 0.016422705724835396, + 0.024233603850007057, + 0.004819652996957302, + 0.014073976315557957, + 0.005978491622954607, + 0.013351179659366608, + 0.016929777339100838, + 0.009039611555635929, + 0.010686623863875866, + 0.01474132388830185, + 0.007539695594459772, + 0.014416089281439781, + 0.008648869581520557, + 0.020489368587732315, + 0.027377350255846977, + 0.01198171079158783, + 0.014073456637561321, + 0.014954342506825924, + 0.004059058614075184, + 0.016647571697831154, + 0.004544609691947699, + 0.010655594058334827, + 0.010388853028416634, + 0.0061398521065711975, + 0.009249743074178696, + 0.011479697190225124, + 0.0015106966020539403, + 0.0028961135540157557, + 0.0015864541055634618, + 0.006016397383064032, + 0.006096207536756992, + 0.002059554448351264, + 0.006883538793772459, + 0.003407987067475915, + 0.002473394153639674, + 0.0067588211968541145, + 0.003746424335986376, + 0.018935516476631165, + 0.012453421950340271, + 0.005183154717087746, + 0.014922497794032097, + 0.00992874801158905, + 0.0018015705281868577, + 0.0035846128594130278, + 0.0031715885270386934, + 0.006978526245802641, + 0.009396690875291824, + 0.0028985810931771994, + 0.00731904711574316, + 0.009672694839537144, + 0.005786375142633915, + 0.010124331340193748, + 0.008666426874697208, + 0.02609381638467312, + 0.026099639013409615, + 0.007844379171729088, + 0.02683367393910885, + 0.015186074189841747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.17993058264255524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, shot at an extremely slow speed, tilted upwards from below, focuses on a massive jeweled egg, carved from a material resembling moonstone or translucent crystal. Captured at a high frame rate, the scene appears as extreme slow motion. The eggshell, polished to a mirror-smooth finish, is meticulously engraved with intricate, ancient runes. From within, a warm, soft white light steadily emanates, illuminating the entire egg and making the grooves within the runes seem to flow with a faint energy. Several fine, natural cracks dot the eggshell, through which crystal-clear water slowly and steadily seeps. Each plump droplet slides carefully along the graceful curve of the eggshell, its surface tension maintaining its perfect spherical shape and clearly reflecting the detailed runes and the soft light filtering through. As it slides, the droplet leaves a fleeting, slightly damp mark on the smooth surface. Set against a pure black background, the entire image is completely illuminated by the inner light of the egg, presenting an extremely pure texture and an otherworldly luster, creating an atmosphere that is sacred, quiet and full of the miracle of life.", + "seed": 32818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_2988_seed_32818_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12092207819223404, + "vlm_logprob": -0.12092207819223404, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.12092207819223404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026128582656383514 + ], + "opposite_logprob_tokens": [ + -3.7761285305023193 + ], + "logprob_reward": -0.026128582656383514, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026128582656383514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2815006971359253, + 0.11494268476963043, + 0.1107451543211937, + 0.034644439816474915, + 0.03229071572422981, + 0.057082291692495346, + 0.04099547490477562, + 0.04020768776535988, + 0.20038722455501556, + 0.08720360696315765 + ], + "credit_map": [ + 0.0066953687928617, + 0.04929918050765991, + 0.015430646017193794, + 0.1140589788556099, + 0.015587560832500458, + 0.004378811921924353, + 0.06305665522813797, + 0.012993484735488892, + 0.0046658762730658054, + 0.03724797070026398, + 0.006014578510075808, + 0.013465587981045246, + 0.00845941063016653, + 0.0038588412571698427, + 0.035214003175497055, + 0.00601642020046711, + 0.0038046787958592176, + 0.01145250629633665, + 0.010017937049269676, + 0.026453977450728416, + 0.020535120740532875, + 0.0032586418092250824, + 0.028919747099280357, + 0.006302551832050085, + 0.0022072079591453075, + 0.0036732975859194994, + 0.009625964798033237, + 0.006625253241509199, + 0.00505739776417613, + 0.001285678124986589, + 0.0030373355839401484, + 0.003132306970655918, + 0.002320557599887252, + 0.00601716386154294, + 0.006002240348607302, + 0.004420901648700237, + 0.005855053663253784, + 0.0015158391324803233, + 0.0035092781763523817, + 0.0026496821083128452, + 0.0033201263286173344, + 0.006546315737068653, + 0.008108466863632202, + 0.009672184474766254, + 0.01245398260653019, + 0.004055248573422432, + 0.005841077771037817, + 0.007084892131388187, + 0.002171458676457405, + 0.008018904365599155, + 0.006311435718089342, + 0.007810445036739111, + 0.00572897819802165, + 0.0028016222640872, + 0.004284640308469534, + 0.0038679938297718763, + 0.0028948651161044836, + 0.005718560423702002, + 0.004805689677596092, + 0.004000593442469835, + 0.00897305179387331, + 0.0038898512721061707, + 0.0056938002817332745, + 0.00423128018155694, + 0.008360895328223705, + 0.01652359589934349, + 0.022100383415818214, + 0.015601826831698418, + 0.03955546021461487, + 0.01352308876812458, + 0.05870353430509567, + 0.026018453761935234, + 0.00404330575838685, + 0.011785736307501793, + 0.00878229085355997, + 0.009062150493264198, + 0.013109741732478142, + 0.008041651919484138, + 0.015721557661890984, + 0.016657179221510887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.026128582656383514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22569651901721954 + ], + "opposite_logprob_tokens": [ + -1.6006965637207031 + ], + "logprob_reward": -0.22569651901721954, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22569651901721954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039984237402677536, + 0.010636577382683754, + 0.03195413947105408, + 0.01441939640790224, + 0.008279486559331417, + 0.01201740000396967, + 0.027104893699288368, + 0.0736355185508728, + 0.29995274543762207, + 0.48201560974121094 + ], + "credit_map": [ + 0.0015661217039451003, + 0.0067890714854002, + 0.00250211451202631, + 0.015826279297471046, + 0.002264725975692272, + 0.0008506276644766331, + 0.008571888320147991, + 0.001613409025594592, + 0.0005804842221550643, + 0.0025752615183591843, + 0.0011189274955540895, + 0.001551647437736392, + 0.0013908910332247615, + 0.0005224961787462234, + 0.002074957126751542, + 0.0008219117880798876, + 0.001008670311421156, + 0.0029962414409965277, + 0.00292053259909153, + 0.008887852542102337, + 0.004385456442832947, + 0.0011461114045232534, + 0.008607613854110241, + 0.0020016625057905912, + 0.0008108484908007085, + 0.00308057083748281, + 0.00205218861810863, + 0.0017187572084367275, + 0.002790173515677452, + 0.0008145414176397026, + 0.0018553308909758925, + 0.0012969853123649955, + 0.0004092850140295923, + 0.0016579091316089034, + 0.0010341998422518373, + 0.0009081230964511633, + 0.0017874246696010232, + 0.0005481485277414322, + 0.0010563533287495375, + 0.0008780428324826062, + 0.0007217360544018447, + 0.0019338568672537804, + 0.0015736279310658574, + 0.0016950168646872044, + 0.0024440698325634003, + 0.0009432372753508389, + 0.0016127810813486576, + 0.0010930740972980857, + 0.0012596298474818468, + 0.0036585081834346056, + 0.0037450348027050495, + 0.0030261592473834753, + 0.007003451697528362, + 0.0026244192849844694, + 0.0033852888736873865, + 0.002402401529252529, + 0.0018481684383004904, + 0.005348734091967344, + 0.014946157112717628, + 0.004951151087880135, + 0.008088299073278904, + 0.0028614194598048925, + 0.025185944512486458, + 0.010405649431049824, + 0.010735103860497475, + 0.025133579969406128, + 0.032503705471754074, + 0.023075856268405914, + 0.050637852400541306, + 0.015080510638654232, + 0.09677992016077042, + 0.0460062101483345, + 0.006431099958717823, + 0.014998420141637325, + 0.014647964388132095, + 0.03642543405294418, + 0.021440299227833748, + 0.01383461244404316, + 0.20419622957706451, + 0.1700415462255478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.22569651901721954 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890762135386467 + ], + "opposite_logprob_tokens": [ + -2.464076280593872 + ], + "logprob_reward": -0.0890762135386467, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890762135386467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.092631034553051, + 0.05672607198357582, + 0.04205050319433212, + 0.07907137274742126, + 0.03188476338982582, + 0.15335580706596375, + 0.028280461207032204, + 0.19426637887954712, + 0.24163837730884552, + 0.08009517937898636 + ], + "credit_map": [ + 0.002900860970839858, + 0.012655370868742466, + 0.006138421595096588, + 0.03450537472963333, + 0.005733533762395382, + 0.0029270178638398647, + 0.022174887359142303, + 0.005595565307885408, + 0.0026334202848374844, + 0.016727151349186897, + 0.0036298194900155067, + 0.007095780689269304, + 0.004935720004141331, + 0.0023715419229120016, + 0.015627577900886536, + 0.0037050603423267603, + 0.0015025604516267776, + 0.003735792590305209, + 0.0032655796967446804, + 0.013249504379928112, + 0.004480272531509399, + 0.0012307503493502736, + 0.01128267589956522, + 0.003303365781903267, + 0.004409872926771641, + 0.006167061626911163, + 0.02567918598651886, + 0.011574380099773407, + 0.010136188939213753, + 0.002359057543799281, + 0.013955974020063877, + 0.004789646249264479, + 0.0015924673061817884, + 0.0058895619586110115, + 0.00446134340018034, + 0.004592210985720158, + 0.0064679719507694244, + 0.001810212153941393, + 0.00435296818614006, + 0.0027180276811122894, + 0.008653505705296993, + 0.01858530566096306, + 0.020915204659104347, + 0.018973814323544502, + 0.04444099962711334, + 0.009721074253320694, + 0.014721756801009178, + 0.017344152554869652, + 0.0015204880619421601, + 0.0059404741041362286, + 0.004522891249507666, + 0.004382978193461895, + 0.004315850790590048, + 0.002107493579387665, + 0.002709507942199707, + 0.002780777169391513, + 0.005003317724913359, + 0.013136961497366428, + 0.03848648816347122, + 0.012930450029671192, + 0.020382404327392578, + 0.00752177694812417, + 0.06806407123804092, + 0.028740916401147842, + 0.008670877665281296, + 0.01949159801006317, + 0.027449751272797585, + 0.019292717799544334, + 0.04321686923503876, + 0.01332940999418497, + 0.07572238147258759, + 0.03446478024125099, + 0.003638894995674491, + 0.007746351882815361, + 0.008165085688233376, + 0.009098614566028118, + 0.01702709309756756, + 0.010086268186569214, + 0.014397429302334785, + 0.009935439564287663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.0890762135386467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811577707529068 + ], + "opposite_logprob_tokens": [ + -3.2881157398223877 + ], + "logprob_reward": -0.03811577707529068, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811577707529068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05307720601558685, + 0.01835126429796219, + 0.01581348478794098, + 0.03093133680522442, + 0.011429005302488804, + 0.11170072853565216, + 0.025572361424565315, + 0.20884887874126434, + 0.3414559066295624, + 0.1828199028968811 + ], + "credit_map": [ + 0.0018718461506068707, + 0.011359418742358685, + 0.0031525364611297846, + 0.021626882255077362, + 0.0032916655763983727, + 0.0013875003205612302, + 0.008228763937950134, + 0.0021585861686617136, + 0.0011368870036676526, + 0.0039965081959962845, + 0.0020579693373292685, + 0.0033245233353227377, + 0.0030608398374170065, + 0.0010299400892108679, + 0.002106189727783203, + 0.0016384058399125934, + 0.0008160233264788985, + 0.002728856634348631, + 0.002823065035045147, + 0.0022419027518481016, + 0.003486485220491886, + 0.0007664079312235117, + 0.001815598108805716, + 0.0011351443827152252, + 0.0016254232032224536, + 0.004211751278489828, + 0.007456481922417879, + 0.0030596046708524227, + 0.004022496752440929, + 0.0010917165782302618, + 0.0071567208506166935, + 0.0023071374744176865, + 0.0007746414630673826, + 0.0019541627261787653, + 0.0020851939916610718, + 0.0013579828664660454, + 0.0022239082027226686, + 0.0007610528264194727, + 0.0013612376060336828, + 0.0009108242229558527, + 0.006450379267334938, + 0.015039968304336071, + 0.014922658912837505, + 0.012505746446549892, + 0.034740135073661804, + 0.005549989640712738, + 0.012038969434797764, + 0.010452864691615105, + 0.0018313175532966852, + 0.0033670268021523952, + 0.0034439596347510815, + 0.0038730523083359003, + 0.006500737741589546, + 0.0019363361643627286, + 0.002569714793935418, + 0.002050213748589158, + 0.004628845956176519, + 0.014331269077956676, + 0.04215788096189499, + 0.013246317394077778, + 0.025863729417324066, + 0.007464960217475891, + 0.071368008852005, + 0.029787857085466385, + 0.010719136334955692, + 0.029907837510108948, + 0.03634871169924736, + 0.026788223534822464, + 0.05762886255979538, + 0.016272952780127525, + 0.10969192534685135, + 0.05409820377826691, + 0.003489796305075288, + 0.00713088596239686, + 0.007491010241210461, + 0.017621710896492004, + 0.017091602087020874, + 0.008087188936769962, + 0.07000818848609924, + 0.0518995076417923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.03811577707529068 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559329867362976 + ], + "opposite_logprob_tokens": [ + -1.6005933284759521 + ], + "logprob_reward": -0.22559329867362976, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559329867362976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18108095228672028, + 0.060724373906850815, + 0.13234221935272217, + 0.05473800003528595, + 0.06093635782599449, + 0.04045255854725838, + 0.07614786922931671, + 0.20532748103141785, + 0.10037039965391159, + 0.08787985891103745 + ], + "credit_map": [ + 0.005103104282170534, + 0.018680710345506668, + 0.008311830461025238, + 0.07066833972930908, + 0.01171380840241909, + 0.004531332291662693, + 0.05256422609090805, + 0.009507598355412483, + 0.004000906832516193, + 0.014589490368962288, + 0.0057358527556061745, + 0.009683145210146904, + 0.010214393027126789, + 0.003778677200898528, + 0.006689786445349455, + 0.006032123230397701, + 0.003365830983966589, + 0.011950422078371048, + 0.008373348973691463, + 0.049030426889657974, + 0.008121603168547153, + 0.0030383989214897156, + 0.041062477976083755, + 0.007399701047688723, + 0.002799291629344225, + 0.0053450968116521835, + 0.016474200412631035, + 0.011507723480463028, + 0.006081068888306618, + 0.0026432510931044817, + 0.005124759394675493, + 0.004762604366987944, + 0.004623936954885721, + 0.008571666665375233, + 0.008749261498451233, + 0.00941151287406683, + 0.01013084128499031, + 0.004819795489311218, + 0.0088956318795681, + 0.005733713507652283, + 0.002636946039274335, + 0.005554355680942535, + 0.005857025273144245, + 0.005418724380433559, + 0.007070330902934074, + 0.004399319179356098, + 0.004785800818353891, + 0.004730053246021271, + 0.0032399743795394897, + 0.014351298101246357, + 0.010523735545575619, + 0.011386734433472157, + 0.01348956860601902, + 0.005930454935878515, + 0.010241634212434292, + 0.006984470412135124, + 0.005873905960470438, + 0.020151950418949127, + 0.032753836363554, + 0.015683680772781372, + 0.020074764266610146, + 0.009986825287342072, + 0.06853007525205612, + 0.03227243945002556, + 0.003922338597476482, + 0.009830424562096596, + 0.011240927502512932, + 0.008040539920330048, + 0.02071111463010311, + 0.006742770783603191, + 0.02706269733607769, + 0.012819585390388966, + 0.0037585177924484015, + 0.005518039222806692, + 0.00804379303008318, + 0.006341693457216024, + 0.007874329574406147, + 0.005861002020537853, + 0.027417415753006935, + 0.023065069690346718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.22559329867362976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, shot at an extremely slow speed, tilted upwards from below, focuses on a massive jeweled egg, carved from a material resembling moonstone or translucent crystal. Captured at a high frame rate, the scene appears as extreme slow motion. The eggshell, polished to a mirror-smooth finish, is meticulously engraved with intricate, ancient runes. From within, a warm, soft white light steadily emanates, illuminating the entire egg and making the grooves within the runes seem to flow with a faint energy. Several fine, natural cracks dot the eggshell, through which crystal-clear water slowly and steadily seeps. Each plump droplet slides carefully along the graceful curve of the eggshell, its surface tension maintaining its perfect spherical shape and clearly reflecting the detailed runes and the soft light filtering through. As it slides, the droplet leaves a fleeting, slightly damp mark on the smooth surface. Set against a pure black background, the entire image is completely illuminated by the inner light of the egg, presenting an extremely pure texture and an otherworldly luster, creating an atmosphere that is sacred, quiet and full of the miracle of life.", + "seed": 32819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_10/index_2988_seed_32819_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10441030561923981, + "vlm_logprob": -0.10441030561923981, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.10441030561923981, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03833331912755966 + ], + "opposite_logprob_tokens": [ + -3.4133334159851074 + ], + "logprob_reward": -0.03833331912755966, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03833331912755966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28807616233825684, + 0.12532344460487366, + 0.07775530219078064, + 0.16598477959632874, + 0.08198823779821396, + 0.043228182941675186, + 0.08028969168663025, + 0.05669001489877701, + 0.04680945351719856, + 0.03385473042726517 + ], + "credit_map": [ + 0.03450527414679527, + 0.040961552411317825, + 0.02810036949813366, + 0.03001045249402523, + 0.05115048587322235, + 0.03930021449923515, + 0.04155097156763077, + 0.022496839985251427, + 0.019451644271612167, + 0.016547825187444687, + 0.018958084285259247, + 0.015804948285222054, + 0.019260965287685394, + 0.007244309410452843, + 0.02076510153710842, + 0.0072905574925243855, + 0.008550206199288368, + 0.010142497718334198, + 0.009544303640723228, + 0.008852947503328323, + 0.013618968427181244, + 0.00687954667955637, + 0.012782294303178787, + 0.007384536787867546, + 0.01719716377556324, + 0.025511955842375755, + 0.01885787770152092, + 0.01822354644536972, + 0.02842346578836441, + 0.014238961040973663, + 0.03239840269088745, + 0.011133399792015553, + 0.01367622334510088, + 0.01087132003158331, + 0.013734382577240467, + 0.009855528362095356, + 0.012700645253062248, + 0.0035451198928058147, + 0.013396496884524822, + 0.004208516329526901, + 0.0061625209636986256, + 0.005602410528808832, + 0.0071413954719901085, + 0.008453676477074623, + 0.004996027331799269, + 0.0015706283738836646, + 0.006792906206101179, + 0.002508618403226137, + 0.0090943593531847, + 0.010319072753190994, + 0.01024708617478609, + 0.009823564440011978, + 0.011996123939752579, + 0.0030300088692456484, + 0.020608006045222282, + 0.0051714698784053326, + 0.00521708931773901, + 0.007600904908031225, + 0.00917140580713749, + 0.007950312457978725, + 0.008536715991795063, + 0.002449854975566268, + 0.01259855180978775, + 0.0031651800964027643, + 0.0034743878059089184, + 0.007845811545848846, + 0.005301597062498331, + 0.008726839907467365, + 0.004394877701997757, + 0.0017472094623371959, + 0.012128966860473156, + 0.0031897625885903835, + 0.002974027069285512, + 0.005994659848511219, + 0.004023467656224966, + 0.0052665830589830875, + 0.003913058899343014, + 0.0010908475378528237, + 0.008368095383048058, + 0.0022239896934479475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.03833331912755966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31349891424179077 + ], + "opposite_logprob_tokens": [ + -1.3134989738464355 + ], + "logprob_reward": -0.31349891424179077, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31349891424179077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11170274764299393, + 0.12641581892967224, + 0.06325488537549973, + 0.2904752194881439, + 0.16421541571617126, + 0.02581595443189144, + 0.09241937100887299, + 0.09804506599903107, + 0.013898291625082493, + 0.01375722698867321 + ], + "credit_map": [ + 0.013375549577176571, + 0.016612917184829712, + 0.009641879238188267, + 0.012726744636893272, + 0.01763945445418358, + 0.009998563677072525, + 0.023266490548849106, + 0.008441142737865448, + 0.01929498463869095, + 0.01724829152226448, + 0.018779611214995384, + 0.016591647639870644, + 0.020002055913209915, + 0.00805036909878254, + 0.018939899280667305, + 0.007508974056690931, + 0.009086793288588524, + 0.008548806421458721, + 0.007976524531841278, + 0.005404464900493622, + 0.011056637391448021, + 0.004064000211656094, + 0.01155082881450653, + 0.005566829815506935, + 0.031101565808057785, + 0.04409874975681305, + 0.03531469777226448, + 0.03310803323984146, + 0.04758267104625702, + 0.02453841082751751, + 0.055286649614572525, + 0.01944444142282009, + 0.02228955179452896, + 0.02149876393377781, + 0.0334884375333786, + 0.021468205377459526, + 0.023840943351387978, + 0.007370044011622667, + 0.026333037763834, + 0.007926438003778458, + 0.0036202913615852594, + 0.0032556988298892975, + 0.0042989482171833515, + 0.004505496006458998, + 0.003639912698417902, + 0.0010894322767853737, + 0.0038538475055247545, + 0.0015523276524618268, + 0.007585915736854076, + 0.014190482906997204, + 0.007410822436213493, + 0.011203257367014885, + 0.014632659032940865, + 0.0029320595785975456, + 0.027726344764232635, + 0.0067378259263932705, + 0.00984883215278387, + 0.012965962290763855, + 0.012849267572164536, + 0.015116753987967968, + 0.014567558653652668, + 0.0028140826616436243, + 0.024619601666927338, + 0.00526300398632884, + 0.0010335115948691964, + 0.002329872688278556, + 0.0016235910588875413, + 0.0027271255385130644, + 0.0014695284189656377, + 0.0005011566099710763, + 0.0032529900781810284, + 0.0009605149971321225, + 0.0012390154879540205, + 0.0029548618476837873, + 0.0017397836782038212, + 0.002420229371637106, + 0.0015472575323656201, + 0.0005009803571738303, + 0.002444298006594181, + 0.0009108015219680965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.31349891424179077 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062071140855550766 + ], + "opposite_logprob_tokens": [ + -2.8120710849761963 + ], + "logprob_reward": -0.062071140855550766, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062071140855550766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18860983848571777, + 0.1015271246433258, + 0.0897364690899849, + 0.17192982137203217, + 0.21623243391513824, + 0.06638724356889725, + 0.06369426846504211, + 0.05062677711248398, + 0.02557072788476944, + 0.025685321539640427 + ], + "credit_map": [ + 0.02002924308180809, + 0.027858050540089607, + 0.018274009227752686, + 0.016950558871030807, + 0.03126168251037598, + 0.018875135108828545, + 0.04243459925055504, + 0.012926554307341576, + 0.017744198441505432, + 0.013482684269547462, + 0.014571204781532288, + 0.012953590601682663, + 0.015131726861000061, + 0.005738569889217615, + 0.016302065923810005, + 0.005603081081062555, + 0.01103086955845356, + 0.012192672118544579, + 0.012112561613321304, + 0.011255981400609016, + 0.013115703128278255, + 0.0066255866549909115, + 0.014527966268360615, + 0.008875131607055664, + 0.017685886472463608, + 0.025887228548526764, + 0.019421283155679703, + 0.01818276382982731, + 0.028045155107975006, + 0.014472569338977337, + 0.036513812839984894, + 0.011721132323145866, + 0.034065838903188705, + 0.029424909502267838, + 0.03999444097280502, + 0.02528221532702446, + 0.031791843473911285, + 0.009540502913296223, + 0.03561500087380409, + 0.01051767636090517, + 0.008779221214354038, + 0.00770634226500988, + 0.009429226629436016, + 0.009057409130036831, + 0.009694743901491165, + 0.0038780809845775366, + 0.01376461610198021, + 0.004077602177858353, + 0.007320433389395475, + 0.009645952843129635, + 0.005876242648810148, + 0.008457459509372711, + 0.010850486345589161, + 0.0018246032996103168, + 0.01586676575243473, + 0.003852320834994316, + 0.005145670846104622, + 0.006548184435814619, + 0.00924544408917427, + 0.007328240666538477, + 0.006870039273053408, + 0.0020199783612042665, + 0.010054797865450382, + 0.003414424601942301, + 0.0016968630952760577, + 0.006576725747436285, + 0.0024916091933846474, + 0.00624948600307107, + 0.0019731009379029274, + 0.0007443630020134151, + 0.004531209357082844, + 0.0013073697919026017, + 0.0021089643705636263, + 0.004479855764657259, + 0.003846922190859914, + 0.004105154424905777, + 0.0027328948490321636, + 0.0007025333470664918, + 0.006123065482825041, + 0.0015859314007684588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.062071140855550766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811221942305565 + ], + "opposite_logprob_tokens": [ + -3.288112163543701 + ], + "logprob_reward": -0.03811221942305565, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811221942305565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2536628544330597, + 0.04626119136810303, + 0.15987908840179443, + 0.057113707065582275, + 0.13576669991016388, + 0.09085749834775925, + 0.08765097707509995, + 0.09504354000091553, + 0.040433865040540695, + 0.033330585807561874 + ], + "credit_map": [ + 0.03396762162446976, + 0.038261983543634415, + 0.027194052934646606, + 0.02713744156062603, + 0.042858727276325226, + 0.02387169934809208, + 0.04212786629796028, + 0.018243450671434402, + 0.006599804386496544, + 0.005601954646408558, + 0.007977636530995369, + 0.005979218985885382, + 0.005242866463959217, + 0.003980738576501608, + 0.006682873237878084, + 0.004196097142994404, + 0.016810830682516098, + 0.018789827823638916, + 0.025138260796666145, + 0.01799660362303257, + 0.024150000885128975, + 0.012776228599250317, + 0.03153460472822189, + 0.01268273126333952, + 0.005935319233685732, + 0.008587698452174664, + 0.00668408814817667, + 0.006707688793540001, + 0.010725194588303566, + 0.0026453116443008184, + 0.012164540588855743, + 0.003663860959932208, + 0.015369633212685585, + 0.017079610377550125, + 0.024634355679154396, + 0.017308920621871948, + 0.0230462234467268, + 0.006379660684615374, + 0.024340610951185226, + 0.0076076798141002655, + 0.01144376676529646, + 0.011667505837976933, + 0.008222240023314953, + 0.010726225562393665, + 0.019776714965701103, + 0.005440042819827795, + 0.018223484978079796, + 0.005357515998184681, + 0.009810850024223328, + 0.013406770303845406, + 0.01004350371658802, + 0.016285819932818413, + 0.013941198587417603, + 0.0026044927071779966, + 0.015931444242596626, + 0.005626900587230921, + 0.010549521073698997, + 0.010688906535506248, + 0.012275147251784801, + 0.01010302733629942, + 0.017483914270997047, + 0.006790579296648502, + 0.018305165693163872, + 0.008847269229590893, + 0.004124285653233528, + 0.0057647922076284885, + 0.004812019877135754, + 0.006027325987815857, + 0.004630589857697487, + 0.0012155777076259255, + 0.010954992845654488, + 0.0029042824171483517, + 0.003397840540856123, + 0.00449364073574543, + 0.0037982873618602753, + 0.005013494286686182, + 0.003066826146095991, + 0.0010538940550759435, + 0.010179351083934307, + 0.0023272514808923006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.03811221942305565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a massive egg carved from a material resembling moonstone or translucent crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eggshell polished to a mirror-smooth finish and engraved with intricate, ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a warm, soft white light emanate from within the egg, illuminating the grooves in the runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, natural cracks visible on the eggshell with crystal-clear water seeping through them and forming spherical droplets that slide along the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background pure black while the scene is tilted upwards from below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003593444824219 + ], + "opposite_logprob_tokens": [ + -2.695035934448242 + ], + "logprob_reward": -0.07003593444824219, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003593444824219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2254229038953781, + 0.04733727499842644, + 0.04480154812335968, + 0.08375625312328339, + 0.08992566913366318, + 0.08242176473140717, + 0.21531420946121216, + 0.17004622519016266, + 0.011764892376959324, + 0.029209313914179802 + ], + "credit_map": [ + 0.029414914548397064, + 0.030280640348792076, + 0.023888595402240753, + 0.025608330965042114, + 0.03222564607858658, + 0.021235547959804535, + 0.04709985852241516, + 0.015669366344809532, + 0.008596695028245449, + 0.006415079347789288, + 0.005707846488803625, + 0.006010913290083408, + 0.006629297509789467, + 0.002922912361100316, + 0.007671676576137543, + 0.0033828529994934797, + 0.008043651469051838, + 0.005875201430171728, + 0.00786823220551014, + 0.004891624674201012, + 0.005979129113256931, + 0.002868322655558586, + 0.006485173013061285, + 0.002790208673104644, + 0.008881844580173492, + 0.012959737330675125, + 0.010180380195379257, + 0.011785158887505531, + 0.013281416147947311, + 0.007211202289909124, + 0.01383445505052805, + 0.005622058641165495, + 0.01829756796360016, + 0.010924180038273335, + 0.011990065686404705, + 0.011424856260418892, + 0.014936753548681736, + 0.004959251265972853, + 0.01033510360866785, + 0.0070578958839178085, + 0.010968361049890518, + 0.009146795608103275, + 0.012786944396793842, + 0.009508551098406315, + 0.012886653654277325, + 0.0050294725224375725, + 0.016193166375160217, + 0.00590181490406394, + 0.03448544815182686, + 0.030988529324531555, + 0.02364949882030487, + 0.03520408645272255, + 0.03037312440574169, + 0.010463407263159752, + 0.033139917999506, + 0.017010189592838287, + 0.022081511095166206, + 0.023348266258835793, + 0.019879871979355812, + 0.025193853303790092, + 0.03144565969705582, + 0.007734208833426237, + 0.028557566925883293, + 0.011805284768342972, + 0.0009425700409337878, + 0.0028038141317665577, + 0.0016836002469062805, + 0.0026258465368300676, + 0.0009835383389145136, + 0.0003715534112416208, + 0.0017674933187663555, + 0.0005864763516001403, + 0.003647826611995697, + 0.0049213264137506485, + 0.0055909291841089725, + 0.005712898913770914, + 0.0020825453102588654, + 0.0009011160582304001, + 0.004369834437966347, + 0.0019828365184366703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.07003593444824219 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_11/index_tensor([9007, 9007])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_11/index_tensor([9007, 9007])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0ba0488ad4c01495475ee1a68e87a0b7c5eafa4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_11/index_tensor([9007, 9007])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, using head-tracking technology, locked onto a cyborg speaking on a news commentary program. His serious, focused face dominated the frame. He had immaculately combed dark short hair, eagle-like eyes, a faint blue data halo glimmering in his pupils. A high nose bridge and a chiseled jawline revealed faint seams of silver metal that blended seamlessly into his skin. He wore a high-collared, slim-fitting charcoal gray shirt over a well-tailored black blazer. The fabric shone subtly under the cool studio lighting. As he calmly delivered his analysis, the camera smoothly adjusted with his slightest head movements, maintaining perfect composition. At the bottom edge of the frame, one of his hands rested on the sleek black glass tabletop. Its mechanical fingers, composed entirely of polished chrome and carbon fiber, tapped the surface with a steady, deliberate rhythm, producing a barely audible, crisp sound as their fingertips made contact. The background is a huge, slightly out-of-focus LED screen with abstract blue data flow graphics flowing on it. Precise top light and contour light outline his clear silhouette and reflect sharp highlights on the joints of his metal fingers. The whole scene creates an extremely calm, professional analytical atmosphere full of wisdom and futuristic technology.", + "seed": 4146, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_11/index_9007_seed_4146_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04191050799563527, + "vlm_logprob": -0.04191050799563527, + "vlm_yes_no_margin": 4.349999976158142, + "vlm_token_credit": -0.04191050799563527, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005461059045046568 + ], + "opposite_logprob_tokens": [ + -5.3804612159729 + ], + "logprob_reward": -0.005461059045046568, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005461059045046568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5671352744102478, + 0.07643337547779083, + 0.038439661264419556, + 0.029294170439243317, + 0.07958478480577469, + 0.06414247304201126, + 0.03966653719544411, + 0.02570280060172081, + 0.04588494822382927, + 0.033716022968292236 + ], + "credit_map": [ + 0.06716732680797577, + 0.046928152441978455, + 0.07457791268825531, + 0.09325146675109863, + 0.04987269639968872, + 0.0410144180059433, + 0.05879201740026474, + 0.13553127646446228, + 0.010340513661503792, + 0.007728178054094315, + 0.01255783811211586, + 0.013298249803483486, + 0.010794377885758877, + 0.006751759443432093, + 0.008331301622092724, + 0.00663115456700325, + 0.005284322891384363, + 0.00428298395127058, + 0.006824366748332977, + 0.004767850507050753, + 0.005853700917214155, + 0.002439384115859866, + 0.005276739597320557, + 0.003710312070325017, + 0.0034263823181390762, + 0.0031184540130198, + 0.00291263684630394, + 0.0036160429008305073, + 0.004749672953039408, + 0.0021356912329792976, + 0.005980806890875101, + 0.003354481654241681, + 0.011175911873579025, + 0.006644938141107559, + 0.012310297228395939, + 0.007725968025624752, + 0.008074636571109295, + 0.00575941801071167, + 0.01749289594590664, + 0.010400720871984959, + 0.008679118938744068, + 0.007618648931384087, + 0.010587876662611961, + 0.010089241899549961, + 0.008238464593887329, + 0.004628575406968594, + 0.008322457782924175, + 0.005978090222924948, + 0.0058062938041985035, + 0.0024077643174678087, + 0.007448699790984392, + 0.0053778402507305145, + 0.0034609308931976557, + 0.002734109293669462, + 0.008960211649537086, + 0.0034706895239651203, + 0.003046911209821701, + 0.0024015414528548717, + 0.0035569113679230213, + 0.002871886594220996, + 0.00465152133256197, + 0.003623757278546691, + 0.002880746964365244, + 0.0026695223059505224, + 0.00771996658295393, + 0.003909347113221884, + 0.006659382488578558, + 0.005857620853930712, + 0.00594952329993248, + 0.004714342299848795, + 0.006149868946522474, + 0.0049248929135501385, + 0.005323908757418394, + 0.003059027250856161, + 0.004040708765387535, + 0.004546997603029013, + 0.006147944834083319, + 0.0031200461089611053, + 0.003305343445390463, + 0.004172045271843672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005461059045046568 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110103473067284 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110103473067284, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110103473067284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3045848309993744, + 0.09907234460115433, + 0.08244007080793381, + 0.060256436467170715, + 0.07494903355836868, + 0.07273117452859879, + 0.04504896700382233, + 0.05900447443127632, + 0.14506712555885315, + 0.05684560909867287 + ], + "credit_map": [ + 0.035218495875597, + 0.02825782634317875, + 0.03186365216970444, + 0.04094616696238518, + 0.04297507181763649, + 0.027136316522955894, + 0.03275805711746216, + 0.0654292181134224, + 0.014286816120147705, + 0.010554620064795017, + 0.013260094448924065, + 0.017510879784822464, + 0.014044113457202911, + 0.007167550269514322, + 0.010750585235655308, + 0.011497680097818375, + 0.01292612124234438, + 0.009208171628415585, + 0.012744610197842121, + 0.011023367755115032, + 0.010660259984433651, + 0.005848639644682407, + 0.008935063146054745, + 0.011093823239207268, + 0.010376677848398685, + 0.0077486406080424786, + 0.007172812707722187, + 0.0070157963782548904, + 0.008090239949524403, + 0.004916203673928976, + 0.008214255794882774, + 0.006721805315464735, + 0.010153715498745441, + 0.006801237352192402, + 0.01176514383405447, + 0.00860739964991808, + 0.009115044958889484, + 0.0070129940286278725, + 0.014305436983704567, + 0.007188054732978344, + 0.009929836727678776, + 0.00570069020614028, + 0.01441970095038414, + 0.010195075534284115, + 0.009992176666855812, + 0.006395525299012661, + 0.008006220683455467, + 0.008091938681900501, + 0.007815510034561157, + 0.004523281008005142, + 0.005031509790569544, + 0.008402695879340172, + 0.00688827084377408, + 0.0034523210488259792, + 0.004807714838534594, + 0.004127658437937498, + 0.009075131267309189, + 0.005075523164123297, + 0.0064706518314778805, + 0.006855440326035023, + 0.010562200099229813, + 0.007228319998830557, + 0.008385468274354935, + 0.00535173062235117, + 0.02546083554625511, + 0.009568066336214542, + 0.027543149888515472, + 0.018381740897893906, + 0.019261790439486504, + 0.015841903164982796, + 0.01800546608865261, + 0.011004152707755566, + 0.008704529143869877, + 0.005200853571295738, + 0.00917457789182663, + 0.007277637720108032, + 0.008824795484542847, + 0.005433427169919014, + 0.005710568744689226, + 0.0065192170441150665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004110103473067284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499108135700226 + ], + "opposite_logprob_tokens": [ + -2.9299910068511963 + ], + "logprob_reward": -0.05499108135700226, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499108135700226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46061304211616516, + 0.0924549326300621, + 0.07789523899555206, + 0.03989137336611748, + 0.08367405086755753, + 0.08264941722154617, + 0.057880379259586334, + 0.03441723436117172, + 0.029588134959340096, + 0.040936220437288284 + ], + "credit_map": [ + 0.05687776207923889, + 0.040459685027599335, + 0.04530820623040199, + 0.07616797089576721, + 0.04244381561875343, + 0.030661897733807564, + 0.05276890844106674, + 0.11592477560043335, + 0.012145447544753551, + 0.00850851833820343, + 0.010608636774122715, + 0.008070845156908035, + 0.020133525133132935, + 0.008238258771598339, + 0.012389405630528927, + 0.012360291555523872, + 0.0109335882589221, + 0.008831536397337914, + 0.005996035877615213, + 0.01094741839915514, + 0.010206150822341442, + 0.006324063520878553, + 0.01649927906692028, + 0.008157174102962017, + 0.007059451192617416, + 0.004793549422174692, + 0.004359485115855932, + 0.0046274675987660885, + 0.00570703437551856, + 0.0034326473250985146, + 0.0055764224380254745, + 0.004335315898060799, + 0.00988624058663845, + 0.007578929420560598, + 0.010961178690195084, + 0.009578641504049301, + 0.014432164840400219, + 0.006633047480136156, + 0.01573835127055645, + 0.008865500800311565, + 0.009470432996749878, + 0.009258830919861794, + 0.013148072175681591, + 0.012500016950070858, + 0.012187855318188667, + 0.007965831086039543, + 0.009686192497611046, + 0.008432184346020222, + 0.00798216462135315, + 0.004501263611018658, + 0.013162625022232533, + 0.011909207329154015, + 0.004685587715357542, + 0.0043625724501907825, + 0.0065653324127197266, + 0.004711626097559929, + 0.004774404224008322, + 0.0030744632240384817, + 0.004157747607678175, + 0.0037232849281281233, + 0.00783378817141056, + 0.0040220534428954124, + 0.00411883182823658, + 0.002712660003453493, + 0.0041126967407763, + 0.002317792735993862, + 0.004485825076699257, + 0.005309010855853558, + 0.0041629415936768055, + 0.0034326124005019665, + 0.002604685490950942, + 0.003162568435072899, + 0.006083499174565077, + 0.003847584594041109, + 0.0068250736221671104, + 0.0068246410228312016, + 0.005282348953187466, + 0.003237747820094228, + 0.004264952149242163, + 0.004570377059280872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05499108135700226 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196044661104679 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002196044661104679, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196044661104679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2933303117752075, + 0.11420033872127533, + 0.13328097760677338, + 0.09736842662096024, + 0.08740399032831192, + 0.06776610761880875, + 0.06316517293453217, + 0.07009939849376678, + 0.04164053872227669, + 0.03174465894699097 + ], + "credit_map": [ + 0.03397412970662117, + 0.026875685900449753, + 0.037497133016586304, + 0.044378120452165604, + 0.027940426021814346, + 0.024157466366887093, + 0.03200612589716911, + 0.06650124490261078, + 0.015604067593812943, + 0.008995245210826397, + 0.025842737406492233, + 0.018606556579470634, + 0.014287917874753475, + 0.0072976648807525635, + 0.01210501417517662, + 0.01146115642040968, + 0.02509447932243347, + 0.010786348022520542, + 0.032512109726667404, + 0.018432490527629852, + 0.010449389927089214, + 0.008369052782654762, + 0.011836910620331764, + 0.015800217166543007, + 0.019090555608272552, + 0.00805709883570671, + 0.026296168565750122, + 0.009151427075266838, + 0.0079877944663167, + 0.006364400964230299, + 0.01030784286558628, + 0.01011315081268549, + 0.011224242858588696, + 0.008171862922608852, + 0.012731114402413368, + 0.016515420749783516, + 0.01050414890050888, + 0.007694204803556204, + 0.011652403511106968, + 0.008910605683922768, + 0.008940327912569046, + 0.006319891661405563, + 0.015506597235798836, + 0.009650567546486855, + 0.006073632277548313, + 0.005732180085033178, + 0.008772098459303379, + 0.006770823150873184, + 0.009699597023427486, + 0.005261512938886881, + 0.008219865150749683, + 0.012703719548881054, + 0.007024089340120554, + 0.004804594907909632, + 0.009717986918985844, + 0.005733818281441927, + 0.010807685554027557, + 0.007487485650926828, + 0.013443834148347378, + 0.011275356635451317, + 0.00753407459706068, + 0.005557909607887268, + 0.0061844587326049805, + 0.007808600086718798, + 0.007173388730734587, + 0.0035161597188562155, + 0.007214434444904327, + 0.007716490421444178, + 0.004731135442852974, + 0.004028742667287588, + 0.003025359008461237, + 0.004234835505485535, + 0.00475463317707181, + 0.002696976764127612, + 0.004170575179159641, + 0.0037368761841207743, + 0.006075477693229914, + 0.002411242574453354, + 0.003713551675900817, + 0.004185331054031849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002196044661104679 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279425144195557 + ], + "opposite_logprob_tokens": [ + -2.017794132232666 + ], + "logprob_reward": -0.14279425144195557, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14279425144195557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19789136946201324, + 0.1044512614607811, + 0.13594795763492584, + 0.05118769034743309, + 0.12590526044368744, + 0.1840677112340927, + 0.03552202880382538, + 0.02080175280570984, + 0.10802599787712097, + 0.036199018359184265 + ], + "credit_map": [ + 0.023665117099881172, + 0.019450953230261803, + 0.022299209609627724, + 0.027210963889956474, + 0.032861534506082535, + 0.01652129553258419, + 0.02405695617198944, + 0.031825318932533264, + 0.013733482919633389, + 0.01144846249371767, + 0.011625085957348347, + 0.011905786581337452, + 0.022269967943429947, + 0.00882373470813036, + 0.015378807671368122, + 0.00926592294126749, + 0.018088003620505333, + 0.013207195326685905, + 0.01799033395946026, + 0.023390239104628563, + 0.019800608977675438, + 0.010393225587904453, + 0.020483702421188354, + 0.012594630941748619, + 0.006738404743373394, + 0.005396328400820494, + 0.005529999267309904, + 0.008473830297589302, + 0.007517061196267605, + 0.0037572376895695925, + 0.009118816815316677, + 0.004656002391129732, + 0.014988571405410767, + 0.012267407961189747, + 0.013805434107780457, + 0.012066797353327274, + 0.01653413288295269, + 0.013809310272336006, + 0.02937152609229088, + 0.0130620701238513, + 0.025482026860117912, + 0.020380008965730667, + 0.031628575176000595, + 0.03285793215036392, + 0.021294504404067993, + 0.01531493104994297, + 0.019593529403209686, + 0.01751619018614292, + 0.004664805717766285, + 0.0030013707000762224, + 0.005505267530679703, + 0.007741072680801153, + 0.003682335838675499, + 0.0031846691854298115, + 0.004318075254559517, + 0.0034244267735630274, + 0.0028702320996671915, + 0.0021488203201442957, + 0.002404219703748822, + 0.0031716153025627136, + 0.0035768689122051, + 0.003209585091099143, + 0.0018773567862808704, + 0.0015430526109412313, + 0.01599091850221157, + 0.0072458600625395775, + 0.020174982026219368, + 0.015028704889118671, + 0.013917437754571438, + 0.009582838043570518, + 0.014122985303401947, + 0.011962258256971836, + 0.0047765630297362804, + 0.0037960573099553585, + 0.004217505920678377, + 0.00589801836758852, + 0.005955988075584173, + 0.0040139369666576385, + 0.003574658650904894, + 0.0039662872441112995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.14279425144195557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, using head-tracking technology, locked onto a cyborg speaking on a news commentary program. His serious, focused face dominated the frame. He had immaculately combed dark short hair, eagle-like eyes, a faint blue data halo glimmering in his pupils. A high nose bridge and a chiseled jawline revealed faint seams of silver metal that blended seamlessly into his skin. He wore a high-collared, slim-fitting charcoal gray shirt over a well-tailored black blazer. The fabric shone subtly under the cool studio lighting. As he calmly delivered his analysis, the camera smoothly adjusted with his slightest head movements, maintaining perfect composition. At the bottom edge of the frame, one of his hands rested on the sleek black glass tabletop. Its mechanical fingers, composed entirely of polished chrome and carbon fiber, tapped the surface with a steady, deliberate rhythm, producing a barely audible, crisp sound as their fingertips made contact. The background is a huge, slightly out-of-focus LED screen with abstract blue data flow graphics flowing on it. Precise top light and contour light outline his clear silhouette and reflect sharp highlights on the joints of his metal fingers. The whole scene creates an extremely calm, professional analytical atmosphere full of wisdom and futuristic technology.", + "seed": 4147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_11/index_9007_seed_4147_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03042817839886993, + "vlm_logprob": -0.03042817839886993, + "vlm_yes_no_margin": 4.6749999761581424, + "vlm_token_credit": -0.03042817839886993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008683658204972744 + ], + "opposite_logprob_tokens": [ + -4.883683681488037 + ], + "logprob_reward": -0.008683658204972744, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008683658204972744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21738409996032715, + 0.1718805432319641, + 0.12037227302789688, + 0.0913509801030159, + 0.04808620363473892, + 0.09009342640638351, + 0.0710720419883728, + 0.11531206220388412, + 0.04074050486087799, + 0.03370784595608711 + ], + "credit_map": [ + 0.028365161269903183, + 0.025638068094849586, + 0.022912824526429176, + 0.022934135049581528, + 0.036740973591804504, + 0.026384999975562096, + 0.03190197795629501, + 0.022505950182676315, + 0.017459966242313385, + 0.014637183398008347, + 0.016834966838359833, + 0.02114270068705082, + 0.0230744406580925, + 0.024814024567604065, + 0.02478780224919319, + 0.029129469767212868, + 0.012909791432321072, + 0.014678030274808407, + 0.01496335119009018, + 0.01510926615446806, + 0.013739763759076595, + 0.01260174810886383, + 0.01937507465481758, + 0.01699524000287056, + 0.011144465766847134, + 0.015720872208476067, + 0.006974521093070507, + 0.008549096062779427, + 0.016519762575626373, + 0.011645819991827011, + 0.011140948161482811, + 0.009655491448938847, + 0.0056043826043605804, + 0.00901462696492672, + 0.005439257249236107, + 0.006001266185194254, + 0.0058054751716554165, + 0.00808089878410101, + 0.004799744114279747, + 0.0033405516296625137, + 0.009770456701517105, + 0.013482307083904743, + 0.010621970519423485, + 0.00899581704288721, + 0.012669739313423634, + 0.010715826414525509, + 0.016119958832859993, + 0.007717355154454708, + 0.009794140234589577, + 0.010474195703864098, + 0.004675936419516802, + 0.008031905628740788, + 0.008403932675719261, + 0.009256442077457905, + 0.01201386284083128, + 0.008421629667282104, + 0.016471676528453827, + 0.013546505011618137, + 0.0132836252450943, + 0.010697205550968647, + 0.01490795612335205, + 0.016151193529367447, + 0.013920395635068417, + 0.016333507373929024, + 0.006491930689662695, + 0.007878370583057404, + 0.0030773323960602283, + 0.0032657962292432785, + 0.006402280181646347, + 0.006025558337569237, + 0.005118323024362326, + 0.0024809150490909815, + 0.008071537129580975, + 0.005365714430809021, + 0.002857318613678217, + 0.0025761141441762447, + 0.00479977298527956, + 0.0057425061240792274, + 0.0024771809112280607, + 0.0018177032470703125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008683658204972744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036320213694125414 + ], + "opposite_logprob_tokens": [ + -5.628632068634033 + ], + "logprob_reward": -0.0036320213694125414, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036320213694125414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22897885739803314, + 0.15644972026348114, + 0.10160250961780548, + 0.09816520661115646, + 0.07092317938804626, + 0.05463210120797157, + 0.07443707436323166, + 0.13387426733970642, + 0.03971202298998833, + 0.04122518375515938 + ], + "credit_map": [ + 0.0292068924754858, + 0.038679931312799454, + 0.020152762532234192, + 0.029548712074756622, + 0.04469481483101845, + 0.02874252013862133, + 0.02473917044699192, + 0.013214017264544964, + 0.015113250352442265, + 0.020577043294906616, + 0.012484238483011723, + 0.027122164145112038, + 0.02071707881987095, + 0.017776288092136383, + 0.026670506224036217, + 0.015989139676094055, + 0.016595508903265, + 0.009847406297922134, + 0.00881465245038271, + 0.01303018443286419, + 0.012011592276394367, + 0.012786168605089188, + 0.016769709065556526, + 0.011747278273105621, + 0.013327913358807564, + 0.017223278060555458, + 0.009106381796300411, + 0.008309464901685715, + 0.013077356852591038, + 0.013527201488614082, + 0.012150943279266357, + 0.011442654766142368, + 0.009353045374155045, + 0.014076505787670612, + 0.007117521483451128, + 0.00822711456567049, + 0.0084149781614542, + 0.011613359674811363, + 0.007485350128263235, + 0.004635299555957317, + 0.0056334733963012695, + 0.008379502221941948, + 0.005673656240105629, + 0.005333863198757172, + 0.010021844878792763, + 0.007014069240540266, + 0.007631712593138218, + 0.0049439724534749985, + 0.010838955640792847, + 0.011328823864459991, + 0.005406343378126621, + 0.005660868249833584, + 0.011102614924311638, + 0.009510962292551994, + 0.010423888452351093, + 0.010164606384932995, + 0.01978132128715515, + 0.020041758194565773, + 0.016031136736273766, + 0.010825525969266891, + 0.01760910451412201, + 0.018360501155257225, + 0.015276136808097363, + 0.015948759391903877, + 0.005944651551544666, + 0.007889732718467712, + 0.004499985370784998, + 0.003649587742984295, + 0.0051765539683401585, + 0.005007721018046141, + 0.0036067813634872437, + 0.00393700459972024, + 0.009508454240858555, + 0.005837568081915379, + 0.0036588141229003668, + 0.0029602618888020515, + 0.00605191383510828, + 0.006344611756503582, + 0.003545420477166772, + 0.003318137489259243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0036320213694125414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078514158725739 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078514158725739, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078514158725739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09926654398441315, + 0.31265968084335327, + 0.2244543433189392, + 0.04588502645492554, + 0.04849095270037651, + 0.04384337365627289, + 0.04751371219754219, + 0.03352339565753937, + 0.0676942989230156, + 0.07666860520839691 + ], + "credit_map": [ + 0.011922157369554043, + 0.013522367924451828, + 0.012503420934081078, + 0.010585799813270569, + 0.017156675457954407, + 0.012126106768846512, + 0.010552394203841686, + 0.010897618718445301, + 0.035373155027627945, + 0.02896019257605076, + 0.03075975924730301, + 0.05444250628352165, + 0.0330120325088501, + 0.03587237000465393, + 0.06088029965758324, + 0.03335937112569809, + 0.02062724530696869, + 0.022018736228346825, + 0.02400679513812065, + 0.03562267869710922, + 0.021322228014469147, + 0.021229632198810577, + 0.04905658960342407, + 0.03057044930756092, + 0.008262011222541332, + 0.007720555178821087, + 0.004404687322676182, + 0.0033333436585962772, + 0.006534544751048088, + 0.006197801325470209, + 0.005102129187434912, + 0.004329955670982599, + 0.006700546480715275, + 0.007514636032283306, + 0.0053340597078204155, + 0.005481267347931862, + 0.005756671540439129, + 0.006897287908941507, + 0.006205930840224028, + 0.004600553307682276, + 0.006317400839179754, + 0.008042250759899616, + 0.004423211328685284, + 0.004347685724496841, + 0.007746426854282618, + 0.006540311500430107, + 0.003422694280743599, + 0.003003392368555069, + 0.005350688938051462, + 0.006300772074609995, + 0.004430347587913275, + 0.0031461745966225863, + 0.00590172316879034, + 0.007309857755899429, + 0.0049500372260808945, + 0.010124110616743565, + 0.0032856559846550226, + 0.0043684192933142185, + 0.0037010409869253635, + 0.0027543040923774242, + 0.004948153626173735, + 0.004673533607274294, + 0.005492735654115677, + 0.004299551248550415, + 0.007925532758235931, + 0.012459073215723038, + 0.006963212508708239, + 0.006768406368792057, + 0.010782387107610703, + 0.008411332964897156, + 0.008375507779419422, + 0.006008842960000038, + 0.010667919181287289, + 0.016359683126211166, + 0.004562527872622013, + 0.009782074019312859, + 0.011784828267991543, + 0.009599000215530396, + 0.00880957767367363, + 0.005102998577058315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.011078514158725739 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017114529618993402 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017114529618993402, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017114529618993402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2157505601644516, + 0.08004742860794067, + 0.07568279653787613, + 0.15231455862522125, + 0.07117181271314621, + 0.05564749240875244, + 0.12114115059375763, + 0.1015438660979271, + 0.056573428213596344, + 0.07012691348791122 + ], + "credit_map": [ + 0.029845083132386208, + 0.02512820065021515, + 0.029202252626419067, + 0.03360404446721077, + 0.025474032387137413, + 0.026994438841938972, + 0.02826320379972458, + 0.017239311710000038, + 0.011816115118563175, + 0.011235460638999939, + 0.008221859112381935, + 0.00783112645149231, + 0.014874706044793129, + 0.01024719700217247, + 0.0071188719011843204, + 0.00870208628475666, + 0.014029928483068943, + 0.007132525555789471, + 0.008835293352603912, + 0.007397941779345274, + 0.005919473245739937, + 0.00947504211217165, + 0.012789334170520306, + 0.010103259235620499, + 0.035305947065353394, + 0.015926511958241463, + 0.016642233356833458, + 0.012106640264391899, + 0.012919548898935318, + 0.0233237873762846, + 0.01885969564318657, + 0.017230184748768806, + 0.011734338477253914, + 0.013730010017752647, + 0.007337289862334728, + 0.007173057645559311, + 0.010726592503488064, + 0.009417146444320679, + 0.006164779420942068, + 0.004888602998107672, + 0.005136724561452866, + 0.010029174387454987, + 0.0048851128667593, + 0.004997047130018473, + 0.009389062412083149, + 0.006994923111051321, + 0.00986403226852417, + 0.004351417534053326, + 0.014233658090233803, + 0.017803369089961052, + 0.008178265765309334, + 0.008317080326378345, + 0.022066034376621246, + 0.017270106822252274, + 0.021467335522174835, + 0.011805306188762188, + 0.016313739120960236, + 0.012722575105726719, + 0.014910606667399406, + 0.008586157113313675, + 0.011869666166603565, + 0.013277621008455753, + 0.008983125910162926, + 0.014880380593240261, + 0.00962922815233469, + 0.008526794612407684, + 0.007166521158069372, + 0.006426437757909298, + 0.00508547481149435, + 0.00786046963185072, + 0.005520198959857225, + 0.006358305923640728, + 0.013656111434102058, + 0.013459332287311554, + 0.006113606970757246, + 0.006424500606954098, + 0.009136509150266647, + 0.01001190859824419, + 0.006392709445208311, + 0.004932237323373556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0017114529618993402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270352452993393 + ], + "opposite_logprob_tokens": [ + -2.127035140991211 + ], + "logprob_reward": -0.1270352452993393, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270352452993393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2856080234050751, + 0.10185274481773376, + 0.07104457169771194, + 0.1104641854763031, + 0.04242941364645958, + 0.06290777027606964, + 0.12236607074737549, + 0.08127319067716599, + 0.06860075891017914, + 0.053453244268894196 + ], + "credit_map": [ + 0.04039686173200607, + 0.03621939197182655, + 0.03051314316689968, + 0.06063677743077278, + 0.032657381147146225, + 0.03273729234933853, + 0.03478742018342018, + 0.01765977218747139, + 0.019693933427333832, + 0.011027732864022255, + 0.00949034933000803, + 0.013770879246294498, + 0.011825169436633587, + 0.013940275646746159, + 0.012829716317355633, + 0.009274695068597794, + 0.010139644145965576, + 0.006636773701757193, + 0.00817867275327444, + 0.0070899673737585545, + 0.007374690379947424, + 0.009007632732391357, + 0.009213925339281559, + 0.01340327039361, + 0.017223570495843887, + 0.01773879863321781, + 0.01034888718277216, + 0.01090812124311924, + 0.014431240037083626, + 0.014807983301579952, + 0.010175986215472221, + 0.014829607680439949, + 0.008085541427135468, + 0.0066644879989326, + 0.004607465583831072, + 0.004034109879285097, + 0.004218091256916523, + 0.005570665001869202, + 0.005896945483982563, + 0.00335211050696671, + 0.007712807971984148, + 0.008374743163585663, + 0.009842950850725174, + 0.005983803421258926, + 0.0069625284522771835, + 0.008144708350300789, + 0.009351897053420544, + 0.006534337066113949, + 0.014762569218873978, + 0.016854964196681976, + 0.00860565435141325, + 0.009690879844129086, + 0.019505424425005913, + 0.0169724952429533, + 0.023778023198246956, + 0.012196066789329052, + 0.012458285316824913, + 0.012318669818341732, + 0.010771477594971657, + 0.007494681514799595, + 0.010132472962141037, + 0.010741227306425571, + 0.008182424120604992, + 0.009173968806862831, + 0.007131335791200399, + 0.014482522383332253, + 0.01040306780487299, + 0.0063906772993505, + 0.007739387918263674, + 0.006918365601450205, + 0.008481393568217754, + 0.007054012268781662, + 0.007097845431417227, + 0.009168541990220547, + 0.007323126308619976, + 0.00501254852861166, + 0.007787481881678104, + 0.006704398896545172, + 0.005423440597951412, + 0.004935865756124258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.1270352452993393 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, using head-tracking technology, locked onto a cyborg speaking on a news commentary program. His serious, focused face dominated the frame. He had immaculately combed dark short hair, eagle-like eyes, a faint blue data halo glimmering in his pupils. A high nose bridge and a chiseled jawline revealed faint seams of silver metal that blended seamlessly into his skin. He wore a high-collared, slim-fitting charcoal gray shirt over a well-tailored black blazer. The fabric shone subtly under the cool studio lighting. As he calmly delivered his analysis, the camera smoothly adjusted with his slightest head movements, maintaining perfect composition. At the bottom edge of the frame, one of his hands rested on the sleek black glass tabletop. Its mechanical fingers, composed entirely of polished chrome and carbon fiber, tapped the surface with a steady, deliberate rhythm, producing a barely audible, crisp sound as their fingertips made contact. The background is a huge, slightly out-of-focus LED screen with abstract blue data flow graphics flowing on it. Precise top light and contour light outline his clear silhouette and reflect sharp highlights on the joints of his metal fingers. The whole scene creates an extremely calm, professional analytical atmosphere full of wisdom and futuristic technology.", + "seed": 4148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_11/index_9007_seed_4148_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02163952700793743, + "vlm_logprob": -0.02163952700793743, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02163952700793743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013892029412090778 + ], + "opposite_logprob_tokens": [ + -4.38889217376709 + ], + "logprob_reward": -0.013892029412090778, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013892029412090778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08915763348340988, + 0.051787715405225754, + 0.02214192971587181, + 0.023370660841464996, + 0.05169769749045372, + 0.1772644817829132, + 0.3935141861438751, + 0.036801960319280624, + 0.10580598562955856, + 0.04845777899026871 + ], + "credit_map": [ + 0.015437480993568897, + 0.010274671949446201, + 0.01526959240436554, + 0.007733096834272146, + 0.009261362254619598, + 0.010065391659736633, + 0.008752463385462761, + 0.012363567017018795, + 0.006288967560976744, + 0.00438604224473238, + 0.005893912632018328, + 0.006755352485924959, + 0.007832461036741734, + 0.006021875888109207, + 0.009722618386149406, + 0.0048864781856536865, + 0.0037705381400883198, + 0.0018308914732187986, + 0.002904729451984167, + 0.003121187211945653, + 0.002159352879971266, + 0.002570651238784194, + 0.002905667992308736, + 0.0028789108619093895, + 0.003583888988941908, + 0.0018918898422271013, + 0.003418409265577793, + 0.0026892689056694508, + 0.002118953038007021, + 0.003333114320412278, + 0.0031520025804638863, + 0.003183132503181696, + 0.006713093724101782, + 0.004086092580109835, + 0.005118255037814379, + 0.007270229514688253, + 0.0066530886106193066, + 0.0071921320632100105, + 0.008692129515111446, + 0.0059726727195084095, + 0.04941987246274948, + 0.011812462471425533, + 0.011311748996376991, + 0.008037956431508064, + 0.010351062752306461, + 0.06354115158319473, + 0.012330214492976665, + 0.010459999553859234, + 0.13328610360622406, + 0.024492159485816956, + 0.03147237002849579, + 0.017308196052908897, + 0.013470890000462532, + 0.13529103994369507, + 0.0109843285754323, + 0.02720908634364605, + 0.007591530214995146, + 0.0028135112952440977, + 0.005845450796186924, + 0.003501768922433257, + 0.0025388463400304317, + 0.005706519819796085, + 0.00427722092717886, + 0.004527108743786812, + 0.035320062190294266, + 0.006609792821109295, + 0.03478556498885155, + 0.005144906230270863, + 0.0038861429784446955, + 0.008916362188756466, + 0.005079210735857487, + 0.006063946057111025, + 0.010625076480209827, + 0.003929515834897757, + 0.006095422897487879, + 0.004120905417948961, + 0.0031822295859456062, + 0.007803064770996571, + 0.004352862015366554, + 0.008348703384399414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.013892029412090778 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076443287543952465 + ], + "opposite_logprob_tokens": [ + -4.882644176483154 + ], + "logprob_reward": -0.0076443287543952465, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076443287543952465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07648801803588867, + 0.05256946012377739, + 0.017623914405703545, + 0.02118857391178608, + 0.11244668811559677, + 0.11669804900884628, + 0.10195513069629669, + 0.06274883449077606, + 0.3832172751426697, + 0.055064041167497635 + ], + "credit_map": [ + 0.015241824090480804, + 0.007556047756224871, + 0.01151357777416706, + 0.007296637631952763, + 0.00797976553440094, + 0.009656865149736404, + 0.008019790053367615, + 0.009223505854606628, + 0.010268068872392178, + 0.004789566155523062, + 0.005943671800196171, + 0.004910001531243324, + 0.0056928060948848724, + 0.009254219941794872, + 0.006910231895744801, + 0.00480089383199811, + 0.003077751724049449, + 0.001667071133852005, + 0.002170863561332226, + 0.0015855339588597417, + 0.0017940821126103401, + 0.003015759401023388, + 0.0020890068262815475, + 0.0022238458041101694, + 0.004173446446657181, + 0.0017246498027816415, + 0.0020921926479786634, + 0.002494648564606905, + 0.0015358065720647573, + 0.005154830869287252, + 0.002273037564009428, + 0.0017399612115696073, + 0.03137451410293579, + 0.007212171331048012, + 0.007159732282161713, + 0.006514564156532288, + 0.006481995340436697, + 0.04038029909133911, + 0.006215002853423357, + 0.007108405232429504, + 0.029827890917658806, + 0.00858627911657095, + 0.007772024720907211, + 0.007699282839894295, + 0.008447282016277313, + 0.038577377796173096, + 0.008242194540798664, + 0.007545723579823971, + 0.03220000118017197, + 0.006405305117368698, + 0.010481741279363632, + 0.004887535236775875, + 0.005810906179249287, + 0.029626762494444847, + 0.0051634726114571095, + 0.0073794047348201275, + 0.013837222009897232, + 0.005104579962790012, + 0.008517024107277393, + 0.005443072412163019, + 0.0057954927906394005, + 0.009593471884727478, + 0.006546884775161743, + 0.007911087945103645, + 0.15279732644557953, + 0.023217041045427322, + 0.09565151482820511, + 0.011798076331615448, + 0.008815775625407696, + 0.05681968480348587, + 0.010204137302935123, + 0.023913709446787834, + 0.008673076517879963, + 0.006454137619584799, + 0.0069577437825500965, + 0.005091584753245115, + 0.005883815232664347, + 0.005489703267812729, + 0.01145953219383955, + 0.005054446402937174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.0076443287543952465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328970469534397 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328970469534397, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328970469534397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10244124382734299, + 0.05631212890148163, + 0.03865678980946541, + 0.19328780472278595, + 0.21467812359333038, + 0.1395065188407898, + 0.0485992468893528, + 0.059102658182382584, + 0.1183280423283577, + 0.029087461531162262 + ], + "credit_map": [ + 0.01353161409497261, + 0.00938118901103735, + 0.013365613296627998, + 0.01100277341902256, + 0.008917700499296188, + 0.023242438212037086, + 0.009519187733530998, + 0.01348073035478592, + 0.00828694086521864, + 0.004666568245738745, + 0.00817271787673235, + 0.0071617476642131805, + 0.005507286172360182, + 0.006578573491424322, + 0.007075391244143248, + 0.008862907998263836, + 0.006262185052037239, + 0.00340613117441535, + 0.003919765818864107, + 0.004117534030228853, + 0.0043528638780117035, + 0.005708867218345404, + 0.004946363624185324, + 0.005943077616393566, + 0.049251195043325424, + 0.011816469021141529, + 0.020738933235406876, + 0.02017287164926529, + 0.01292529609054327, + 0.040243156254291534, + 0.020464548841118813, + 0.01767534576356411, + 0.06596646457910538, + 0.013670745305716991, + 0.012823659926652908, + 0.00932674016803503, + 0.011145872063934803, + 0.07952632755041122, + 0.008881471119821072, + 0.013336854055523872, + 0.03565371781587601, + 0.009380956180393696, + 0.009236034005880356, + 0.009008613415062428, + 0.00917607732117176, + 0.044625747948884964, + 0.012582509778439999, + 0.009842862375080585, + 0.005979435984045267, + 0.0034357879776507616, + 0.0091283293440938, + 0.0033626356162130833, + 0.0037007273640483618, + 0.013519282452762127, + 0.005015453789383173, + 0.0044575948268175125, + 0.00833911169320345, + 0.0037042300682514906, + 0.008900115266442299, + 0.008182698860764503, + 0.005662910640239716, + 0.0073215351440012455, + 0.007098873611539602, + 0.009893180802464485, + 0.02994447760283947, + 0.005218249745666981, + 0.041819993406534195, + 0.005823523737490177, + 0.003994655795395374, + 0.01911957748234272, + 0.0035571360494941473, + 0.008850427344441414, + 0.0033807610161602497, + 0.0022025734651833773, + 0.003695001592859626, + 0.0025337187107652426, + 0.0027371759060770273, + 0.004256263375282288, + 0.0062850890681147575, + 0.003996876068413258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.02328970469534397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013344916515052319 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.0013344916515052319, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013344916515052319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044942643493413925, + 0.04563470557332039, + 0.030168408527970314, + 0.02321496792137623, + 0.027502931654453278, + 0.1878879964351654, + 0.07882668823003769, + 0.06645449995994568, + 0.4731059670448303, + 0.022261198610067368 + ], + "credit_map": [ + 0.005945916287600994, + 0.004315383732318878, + 0.0060145677998661995, + 0.004160713870078325, + 0.003313514171168208, + 0.010805899277329445, + 0.004154182504862547, + 0.006232460495084524, + 0.005885724443942308, + 0.00405890354886651, + 0.005172192584723234, + 0.00613459711894393, + 0.005184310022741556, + 0.008258719928562641, + 0.005265239160507917, + 0.005675013177096844, + 0.007074984721839428, + 0.0025751441717147827, + 0.00391141464933753, + 0.002131083747372031, + 0.002132188295945525, + 0.005661300383508205, + 0.002855307189747691, + 0.0038269804790616035, + 0.0030066119506955147, + 0.002355349948629737, + 0.002765519777312875, + 0.0021260541398078203, + 0.0019184279954060912, + 0.004378945101052523, + 0.002617776859551668, + 0.004046279471367598, + 0.0059185391291975975, + 0.001803731545805931, + 0.003457040758803487, + 0.002675409894436598, + 0.0016664661234244704, + 0.006483323872089386, + 0.002509970683604479, + 0.0029884467367082834, + 0.05009862035512924, + 0.012721979059278965, + 0.012224633246660233, + 0.008735038340091705, + 0.011498266831040382, + 0.06726512312889099, + 0.01249210350215435, + 0.012852202169597149, + 0.026907125487923622, + 0.0050594378262758255, + 0.007469302974641323, + 0.003678339533507824, + 0.0033060761634260416, + 0.023683791980147362, + 0.0026448904536664486, + 0.00607771473005414, + 0.011567864567041397, + 0.00464541744440794, + 0.009689344093203545, + 0.006303806789219379, + 0.005259813275188208, + 0.014475341886281967, + 0.006211626343429089, + 0.008301278576254845, + 0.1890193372964859, + 0.02763105183839798, + 0.13565538823604584, + 0.0157321784645319, + 0.009928498417139053, + 0.05465998873114586, + 0.011154117994010448, + 0.02932536043226719, + 0.0037398801650851965, + 0.0017454397166147828, + 0.0034762464929372072, + 0.0016007611993700266, + 0.0016539997886866331, + 0.004813533741980791, + 0.0021841819398105145, + 0.0030471535865217447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.0013344916515052319 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203708052635193 + ], + "opposite_logprob_tokens": [ + -2.8120369911193848 + ], + "logprob_reward": -0.06203708052635193, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203708052635193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06810273975133896, + 0.02343050204217434, + 0.02526305429637432, + 0.036140039563179016, + 0.15401771664619446, + 0.4400080442428589, + 0.04453488066792488, + 0.044873256236314774, + 0.09930220246315002, + 0.06432761251926422 + ], + "credit_map": [ + 0.011616239324212074, + 0.006198992021381855, + 0.007299823686480522, + 0.009561177343130112, + 0.006505058147013187, + 0.013289748691022396, + 0.006797021254897118, + 0.006834669504314661, + 0.003089754143729806, + 0.0020316641312092543, + 0.0030706122051924467, + 0.00307083479128778, + 0.003165043191984296, + 0.002765984507277608, + 0.0032496503554284573, + 0.002986957784742117, + 0.003946960903704166, + 0.00256406725384295, + 0.00259813922457397, + 0.002749228850007057, + 0.00312239583581686, + 0.003019749652594328, + 0.004185330588370562, + 0.003077179891988635, + 0.008464070037007332, + 0.0030337017960846424, + 0.00420645996928215, + 0.0032358395401388407, + 0.002867369679734111, + 0.00569117208942771, + 0.003937108907848597, + 0.004704317543655634, + 0.044869888573884964, + 0.009053750894963741, + 0.010222618468105793, + 0.007580931298434734, + 0.006890130694955587, + 0.05978237837553024, + 0.005743572488427162, + 0.00987442396581173, + 0.12025320529937744, + 0.02927863411605358, + 0.029070023447275162, + 0.020375186577439308, + 0.02522587589919567, + 0.15933454036712646, + 0.02691907249391079, + 0.029551470652222633, + 0.012613235972821712, + 0.003062122268602252, + 0.0068372623063623905, + 0.003489581635221839, + 0.004008548334240913, + 0.006640256382524967, + 0.003631418338045478, + 0.004252448212355375, + 0.004469829611480236, + 0.0032919365912675858, + 0.005754008889198303, + 0.00562082976102829, + 0.005298393312841654, + 0.006986725609749556, + 0.007000476121902466, + 0.0064510502852499485, + 0.03906738758087158, + 0.0059384568594396114, + 0.026829475536942482, + 0.0033553566318005323, + 0.0024775355122983456, + 0.012664073146879673, + 0.002788458950817585, + 0.0061814552173018456, + 0.013097879476845264, + 0.004622539039701223, + 0.009027736261487007, + 0.0036390358582139015, + 0.004500864539295435, + 0.015210257843136787, + 0.00513711292296648, + 0.009092176333069801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.06203708052635193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, using head-tracking technology, locked onto a cyborg speaking on a news commentary program. His serious, focused face dominated the frame. He had immaculately combed dark short hair, eagle-like eyes, a faint blue data halo glimmering in his pupils. A high nose bridge and a chiseled jawline revealed faint seams of silver metal that blended seamlessly into his skin. He wore a high-collared, slim-fitting charcoal gray shirt over a well-tailored black blazer. The fabric shone subtly under the cool studio lighting. As he calmly delivered his analysis, the camera smoothly adjusted with his slightest head movements, maintaining perfect composition. At the bottom edge of the frame, one of his hands rested on the sleek black glass tabletop. Its mechanical fingers, composed entirely of polished chrome and carbon fiber, tapped the surface with a steady, deliberate rhythm, producing a barely audible, crisp sound as their fingertips made contact. The background is a huge, slightly out-of-focus LED screen with abstract blue data flow graphics flowing on it. Precise top light and contour light outline his clear silhouette and reflect sharp highlights on the joints of his metal fingers. The whole scene creates an extremely calm, professional analytical atmosphere full of wisdom and futuristic technology.", + "seed": 4149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_11/index_9007_seed_4149_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03424463432747871, + "vlm_logprob": -0.03424463432747871, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.03424463432747871, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024639787152409554 + ], + "opposite_logprob_tokens": [ + -3.774639844894409 + ], + "logprob_reward": -0.024639787152409554, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024639787152409554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26960596442222595, + 0.08629639446735382, + 0.055041950196027756, + 0.08454523980617523, + 0.08579665422439575, + 0.09919767081737518, + 0.05190921202301979, + 0.08591623604297638, + 0.10760481655597687, + 0.07408590614795685 + ], + "credit_map": [ + 0.03174573555588722, + 0.03953735530376434, + 0.018491020426154137, + 0.01962233893573284, + 0.03769813850522041, + 0.03405959531664848, + 0.051520250737667084, + 0.03693152219057083, + 0.015699738636612892, + 0.011352204717695713, + 0.0067182984203100204, + 0.005310770124197006, + 0.01190656516700983, + 0.009436685591936111, + 0.014432437717914581, + 0.011439689435064793, + 0.00808880478143692, + 0.007240472827106714, + 0.007325438782572746, + 0.005574686452746391, + 0.006590970326215029, + 0.004283060319721699, + 0.008443023078143597, + 0.007495491299778223, + 0.01161886379122734, + 0.010927963070571423, + 0.011471020989120007, + 0.008788193576037884, + 0.010194013826549053, + 0.011618616990745068, + 0.00999632477760315, + 0.009930241852998734, + 0.010122803039848804, + 0.01089850440621376, + 0.014866624027490616, + 0.008587552234530449, + 0.009628674946725368, + 0.009256355464458466, + 0.011635190807282925, + 0.01080094464123249, + 0.011576424352824688, + 0.012490423396229744, + 0.010409943759441376, + 0.007259706035256386, + 0.015891125425696373, + 0.014035976491868496, + 0.015898844227194786, + 0.01163522806018591, + 0.0055988444946706295, + 0.008652819320559502, + 0.003512284718453884, + 0.0035774442367255688, + 0.006318867206573486, + 0.007317071780562401, + 0.007789995055645704, + 0.009141881950199604, + 0.014849714003503323, + 0.009270815178751945, + 0.007706440053880215, + 0.008132914081215858, + 0.007412330247461796, + 0.00628996454179287, + 0.007476125378161669, + 0.024777935817837715, + 0.026338934898376465, + 0.03105311468243599, + 0.0062138657085597515, + 0.00898673478513956, + 0.008750995621085167, + 0.011683841235935688, + 0.006992198061197996, + 0.007585134822875261, + 0.008541937917470932, + 0.008969110436737537, + 0.009675717912614346, + 0.008298792876303196, + 0.009332036599516869, + 0.0075868465937674046, + 0.008771783672273159, + 0.012909679673612118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.024639787152409554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005968841258436441 + ], + "opposite_logprob_tokens": [ + -5.130969047546387 + ], + "logprob_reward": -0.005968841258436441, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005968841258436441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16484972834587097, + 0.08828642219305038, + 0.07898616790771484, + 0.07283447682857513, + 0.05164705961942673, + 0.04465381056070328, + 0.044045086950063705, + 0.10245778411626816, + 0.11280598491430283, + 0.23943349719047546 + ], + "credit_map": [ + 0.01961294747889042, + 0.0227731354534626, + 0.01387113332748413, + 0.016569945961236954, + 0.021671518683433533, + 0.01879776082932949, + 0.03112378902733326, + 0.02042950876057148, + 0.011349272914230824, + 0.011761613190174103, + 0.0047960649244487286, + 0.006653776858001947, + 0.01641828753054142, + 0.010774785652756691, + 0.018202831968665123, + 0.008329792879521847, + 0.010245075449347496, + 0.010473748669028282, + 0.010953579097986221, + 0.0117758484557271, + 0.00721316784620285, + 0.00612677913159132, + 0.008939744904637337, + 0.013258215971291065, + 0.00941718090325594, + 0.00941503793001175, + 0.010364131070673466, + 0.011989678256213665, + 0.008109106682240963, + 0.0060733165591955185, + 0.00962024461477995, + 0.007845786400139332, + 0.005341767333447933, + 0.00822359323501587, + 0.005446682684123516, + 0.005543982144445181, + 0.0059820860624313354, + 0.006423039361834526, + 0.006899705622345209, + 0.007786201778799295, + 0.004597807768732309, + 0.006479053758084774, + 0.0034396646078675985, + 0.004470126237720251, + 0.007783135864883661, + 0.004890911281108856, + 0.007758072577416897, + 0.005235037766396999, + 0.005723882466554642, + 0.006715384777635336, + 0.0033002735581249, + 0.0031476961448788643, + 0.005483656655997038, + 0.005176698789000511, + 0.005871625151485205, + 0.008625868707895279, + 0.01458625216037035, + 0.013309239409863949, + 0.012178000994026661, + 0.012076100334525108, + 0.008456241339445114, + 0.013156553730368614, + 0.0088178850710392, + 0.019877510145306587, + 0.02688399888575077, + 0.02899632416665554, + 0.008767000399529934, + 0.00758246798068285, + 0.010860991664230824, + 0.010762808844447136, + 0.010217161849141121, + 0.008735239505767822, + 0.03125154972076416, + 0.02637287974357605, + 0.026912055909633636, + 0.02576308138668537, + 0.04863917827606201, + 0.019782952964305878, + 0.03559071943163872, + 0.02512107975780964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005968841258436441 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352928951382637 + ], + "opposite_logprob_tokens": [ + -3.651352882385254 + ], + "logprob_reward": -0.026352928951382637, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352928951382637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17875568568706512, + 0.07031276077032089, + 0.0904705598950386, + 0.06045689061284065, + 0.10387958586215973, + 0.056230053305625916, + 0.18472230434417725, + 0.1429772973060608, + 0.05575378239154816, + 0.05644107609987259 + ], + "credit_map": [ + 0.023429494351148605, + 0.025465501472353935, + 0.015562404878437519, + 0.017530834302306175, + 0.024097416549921036, + 0.022497477009892464, + 0.028486555442214012, + 0.0216860119253397, + 0.010358335450291634, + 0.010265050455927849, + 0.006447472609579563, + 0.007079228293150663, + 0.01002590637654066, + 0.008830859325826168, + 0.010028460063040257, + 0.007277445401996374, + 0.014069936238229275, + 0.013904652558267117, + 0.011917348951101303, + 0.017253035679459572, + 0.00881911814212799, + 0.005763179622590542, + 0.00925698596984148, + 0.009486300870776176, + 0.007146039977669716, + 0.011779083870351315, + 0.005863960366696119, + 0.009097292087972164, + 0.005814098287373781, + 0.007505742367357016, + 0.00556485541164875, + 0.007685817778110504, + 0.014204772189259529, + 0.015336934477090836, + 0.010685889050364494, + 0.014300921931862831, + 0.013207141309976578, + 0.013455166481435299, + 0.012545726262032986, + 0.010143023915588856, + 0.007690240629017353, + 0.007412147708237171, + 0.0050818598829209805, + 0.0073836371302604675, + 0.008450808934867382, + 0.00574721721932292, + 0.008348026312887669, + 0.006116114091128111, + 0.025718364864587784, + 0.022650524973869324, + 0.02154899388551712, + 0.02173311449587345, + 0.021208010613918304, + 0.019759047776460648, + 0.021921314299106598, + 0.030182937160134315, + 0.01893191784620285, + 0.016626497730612755, + 0.01872745156288147, + 0.019399600103497505, + 0.01405210793018341, + 0.02163154073059559, + 0.013432683423161507, + 0.020175503566861153, + 0.009942922741174698, + 0.010056914761662483, + 0.006308460608124733, + 0.008574333973228931, + 0.005107801407575607, + 0.006057028658688068, + 0.004456170368939638, + 0.0052501484751701355, + 0.007200967520475388, + 0.006121940910816193, + 0.007445405703037977, + 0.0063865515403449535, + 0.007590028923004866, + 0.005777016747742891, + 0.006169924046844244, + 0.00974924024194479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.026352928951382637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335444045253098 + ], + "opposite_logprob_tokens": [ + -6.626335620880127 + ], + "logprob_reward": -0.001335444045253098, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335444045253098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16548262536525726, + 0.13679413497447968, + 0.1504049301147461, + 0.04253784939646721, + 0.054042037576436996, + 0.06710617244243622, + 0.06739442050457001, + 0.14562606811523438, + 0.07885503768920898, + 0.09175668656826019 + ], + "credit_map": [ + 0.022301852703094482, + 0.02641662023961544, + 0.015753885731101036, + 0.013942950405180454, + 0.021006330847740173, + 0.01879342831671238, + 0.020764505490660667, + 0.026503054425120354, + 0.026198439300060272, + 0.01822582073509693, + 0.015150009654462337, + 0.022841108962893486, + 0.012192429974675179, + 0.013713398016989231, + 0.015311411581933498, + 0.013161521404981613, + 0.022466840222477913, + 0.033384811133146286, + 0.014231255277991295, + 0.015760887414216995, + 0.01389106921851635, + 0.013489864766597748, + 0.016355853527784348, + 0.020824357867240906, + 0.006358677987009287, + 0.004984609317034483, + 0.0047956532798707485, + 0.005245962645858526, + 0.005210528615862131, + 0.005574302282184362, + 0.0038423442747443914, + 0.006525769364088774, + 0.00634385971352458, + 0.008195349015295506, + 0.0062190694734454155, + 0.00523696094751358, + 0.00602924358099699, + 0.005745565984398127, + 0.00810698326677084, + 0.008165002800524235, + 0.00862949714064598, + 0.009623087011277676, + 0.0059328447096049786, + 0.005423333961516619, + 0.008625488728284836, + 0.009912385605275631, + 0.008094646967947483, + 0.01086488924920559, + 0.008316506631672382, + 0.008726486004889011, + 0.007163507863879204, + 0.005734610371291637, + 0.008534325286746025, + 0.007233744021505117, + 0.010046481154859066, + 0.011638756841421127, + 0.022906623780727386, + 0.016868282109498978, + 0.015443310141563416, + 0.01759559102356434, + 0.012438995763659477, + 0.016948001459240913, + 0.010960319079458714, + 0.032464947551488876, + 0.019218487665057182, + 0.0199700016528368, + 0.008788099512457848, + 0.00612969696521759, + 0.006381121464073658, + 0.008403034880757332, + 0.004227088764309883, + 0.0057375081814825535, + 0.012072044424712658, + 0.009781507775187492, + 0.008537467569112778, + 0.009384889155626297, + 0.014936654828488827, + 0.01044872310012579, + 0.008822210133075714, + 0.017773190513253212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.001335444045253098 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on a cyborg's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cyborg have dark short hair, eagle-like eyes, and a faint blue data halo in his pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cyborg wearing a high-collared charcoal gray shirt under a black blazer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large, out-of-focus LED screen displaying abstract blue data flow graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly adjust to the cyborg's head movements to maintain composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129261702299118 + ], + "opposite_logprob_tokens": [ + -2.2379262447357178 + ], + "logprob_reward": -0.1129261702299118, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129261702299118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16773772239685059, + 0.11473348736763, + 0.08431965857744217, + 0.08505606651306152, + 0.10066696256399155, + 0.10613354295492172, + 0.07620420306921005, + 0.14201441407203674, + 0.04855949059128761, + 0.07457452267408371 + ], + "credit_map": [ + 0.022237107157707214, + 0.030649157240986824, + 0.01632346585392952, + 0.014205398969352245, + 0.020651264116168022, + 0.018111394718289375, + 0.024500751867890358, + 0.02105916664004326, + 0.023989573121070862, + 0.02100558951497078, + 0.010089567862451077, + 0.01249005738645792, + 0.010600184090435505, + 0.010631587356328964, + 0.011526091955602169, + 0.014400821179151535, + 0.011251930147409439, + 0.009279877878725529, + 0.012486028485000134, + 0.008039378561079502, + 0.008645380847156048, + 0.008114160038530827, + 0.009551556780934334, + 0.016951341181993484, + 0.011861458420753479, + 0.008462038822472095, + 0.014191203750669956, + 0.013739303685724735, + 0.010396244935691357, + 0.006795099470764399, + 0.00851682759821415, + 0.01109388004988432, + 0.01286204531788826, + 0.014738589525222778, + 0.01432798895984888, + 0.009699540212750435, + 0.010094803757965565, + 0.011353110894560814, + 0.014817673712968826, + 0.012773205526173115, + 0.017436927184462547, + 0.018508141860365868, + 0.010952392593026161, + 0.009666990488767624, + 0.012346042320132256, + 0.013249264098703861, + 0.011999184265732765, + 0.01197458803653717, + 0.009471329860389233, + 0.00963060837239027, + 0.007732844445854425, + 0.008144206367433071, + 0.009171892888844013, + 0.01108636800199747, + 0.013436476700007915, + 0.007530469447374344, + 0.02680247835814953, + 0.014799683354794979, + 0.012150319293141365, + 0.014004248194396496, + 0.015279777348041534, + 0.014688554219901562, + 0.011177697218954563, + 0.03311164304614067, + 0.011998577043414116, + 0.013546491973102093, + 0.0027240081690251827, + 0.0020740721374750137, + 0.00545164430513978, + 0.005514777731150389, + 0.0033128790091723204, + 0.003937034402042627, + 0.008848140016198158, + 0.010591430589556694, + 0.008469131775200367, + 0.011559277772903442, + 0.010912355035543442, + 0.0073828063905239105, + 0.00852409191429615, + 0.008287286385893822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.1129261702299118 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_12/index_tensor([2840, 2840])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_12/index_tensor([2840, 2840])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f93b1741bb7a0e6e753cfdc3cd7b9ce53d2c8f80 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_12/index_tensor([2840, 2840])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static shot captured from a medium close-up angle, using high-frame-rate slow-motion cinematography, focuses on a middle-aged Asian male chef. He wears a dark gray cotton headscarf, his face stained with grease and sweat from a hard day's work in the kitchen. He has just paused from his rapid chopping, one hand still tightly gripping a heavy kitchen knife, the blade hovering a few millimeters above the flour-stained wooden cutting board, the other resting on the stainless steel countertop. A strong overhead light streams down from above, sculpting the sharp lines of his face and highlighting the beads of sweat on his forehead and the oily sheen on his cheeks. His brow is slightly furrowed, his eyes sharp and focused, as if piercing the air before him, fixated on a single point beyond the camera. The slow-motion effect captures the countless tiny particles of flour dust and oil smoke floating in the air, dancing slowly within the beam of light like shimmering stardust. The background is a slightly blurred professional kitchen, and the stainless steel kitchenware reflects a cold luster. The whole scene creates an atmosphere of extreme concentration, tension and power in realistic and slightly warm colors.", + "seed": 14360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_12/index_2840_seed_14360_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033869693800806996, + "vlm_logprob": -0.033869693800806996, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.033869693800806996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04563586786389351 + ], + "opposite_logprob_tokens": [ + -3.295635938644409 + ], + "logprob_reward": -0.04563586786389351, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04563586786389351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25411921739578247, + 0.06238444894552231, + 0.09693235158920288, + 0.23300597071647644, + 0.04338134825229645, + 0.061914462596178055, + 0.11027448624372482, + 0.04368419945240021, + 0.06001908332109451, + 0.034284479916095734 + ], + "credit_map": [ + 0.02261926233768463, + 0.03609145060181618, + 0.03407255932688713, + 0.045470867305994034, + 0.02954266220331192, + 0.02974064275622368, + 0.03372564539313316, + 0.0228560958057642, + 0.00477175647392869, + 0.0076933009549975395, + 0.007297942880541086, + 0.005728668998926878, + 0.009558048099279404, + 0.009075372479856014, + 0.013544159010052681, + 0.004715193063020706, + 0.012432696297764778, + 0.009438145905733109, + 0.014912509359419346, + 0.010145404376089573, + 0.009286333806812763, + 0.014832533895969391, + 0.01280570775270462, + 0.01307900995016098, + 0.04610668867826462, + 0.031246209517121315, + 0.021002858877182007, + 0.021544627845287323, + 0.02906756103038788, + 0.05029352009296417, + 0.014445727691054344, + 0.01929876208305359, + 0.010015006177127361, + 0.004974768031388521, + 0.004891636781394482, + 0.004981603007763624, + 0.003975728526711464, + 0.008770542219281197, + 0.00264740944840014, + 0.003124651499092579, + 0.01158886682242155, + 0.007095322944223881, + 0.005704676732420921, + 0.006674407981336117, + 0.008135403506457806, + 0.012542870827019215, + 0.004887240473181009, + 0.005285665392875671, + 0.01855602115392685, + 0.013989080674946308, + 0.009943196550011635, + 0.010090556927025318, + 0.01692747324705124, + 0.01800239272415638, + 0.009310981258749962, + 0.013454769738018513, + 0.011233390308916569, + 0.006456418428570032, + 0.003956624772399664, + 0.0034815724939107895, + 0.004361533559858799, + 0.009470845572650433, + 0.0021917724516242743, + 0.002532037440687418, + 0.015018676407635212, + 0.007616474758833647, + 0.00370906968601048, + 0.007647097576409578, + 0.004398292861878872, + 0.013979543931782246, + 0.0035817602183669806, + 0.0040681627579033375, + 0.003702425165101886, + 0.004663285333663225, + 0.0038381563499569893, + 0.004843248054385185, + 0.005556133575737476, + 0.00432145269587636, + 0.0034926861990243196, + 0.0038670902140438557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.04563586786389351 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055013421922922134 + ], + "opposite_logprob_tokens": [ + -2.930013418197632 + ], + "logprob_reward": -0.055013421922922134, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055013421922922134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14515018463134766, + 0.07468084990978241, + 0.05301807075738907, + 0.1772785633802414, + 0.15585845708847046, + 0.10282360762357712, + 0.046503230929374695, + 0.11459684371948242, + 0.09872075170278549, + 0.03136949613690376 + ], + "credit_map": [ + 0.010782674886286259, + 0.025544896721839905, + 0.015419645234942436, + 0.01816484145820141, + 0.023150861263275146, + 0.017846819013357162, + 0.01975190080702305, + 0.014488539658486843, + 0.0059284004382789135, + 0.012656861916184425, + 0.008809864521026611, + 0.0075218225829303265, + 0.011180633679032326, + 0.010471601039171219, + 0.01084542740136385, + 0.007266237400472164, + 0.004418946802616119, + 0.005638071335852146, + 0.007665560115128756, + 0.008993173018097878, + 0.006893208250403404, + 0.006178738549351692, + 0.0077081178314983845, + 0.005522254388779402, + 0.03178995102643967, + 0.028953973203897476, + 0.015569204464554787, + 0.019868746399879456, + 0.021276436746120453, + 0.031390998512506485, + 0.010147887282073498, + 0.01828134059906006, + 0.04455392435193062, + 0.019633106887340546, + 0.009405595250427723, + 0.01150498352944851, + 0.012667758390307426, + 0.037872910499572754, + 0.007538426201790571, + 0.012681734748184681, + 0.02668444626033306, + 0.011699831113219261, + 0.005980059038847685, + 0.009074673056602478, + 0.008943545632064342, + 0.027253136038780212, + 0.005393848288804293, + 0.007794064935296774, + 0.005271459463983774, + 0.006374314893037081, + 0.004407104104757309, + 0.005021759774535894, + 0.008080406114459038, + 0.006450037006288767, + 0.004497373476624489, + 0.006400777027010918, + 0.034408025443553925, + 0.016464998945593834, + 0.007081587798893452, + 0.008979443460702896, + 0.0076064011082053185, + 0.02941708452999592, + 0.004242025315761566, + 0.006397268269211054, + 0.025178518146276474, + 0.012592569924890995, + 0.006486966740339994, + 0.01282862201333046, + 0.006786992773413658, + 0.023635029792785645, + 0.005063769873231649, + 0.006148274056613445, + 0.0038041621446609497, + 0.00424005463719368, + 0.003589803818613291, + 0.004579758271574974, + 0.004108533263206482, + 0.004170028492808342, + 0.0029368940740823746, + 0.003940261900424957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.055013421922922134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371515870094299 + ], + "opposite_logprob_tokens": [ + -3.40871524810791 + ], + "logprob_reward": -0.03371515870094299, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371515870094299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13346777856349945, + 0.058441489934921265, + 0.05337325856089592, + 0.35751014947891235, + 0.044989727437496185, + 0.1693219542503357, + 0.0591738261282444, + 0.08283617347478867, + 0.01388080045580864, + 0.0270049087703228 + ], + "credit_map": [ + 0.009779905900359154, + 0.015402297489345074, + 0.017143331468105316, + 0.014894586056470871, + 0.021231969818472862, + 0.018574466928839684, + 0.018661484122276306, + 0.017779730260372162, + 0.004833227954804897, + 0.006014290265738964, + 0.012213831767439842, + 0.005914157722145319, + 0.005724004469811916, + 0.007767783477902412, + 0.0052174897864460945, + 0.01075670588761568, + 0.006310334429144859, + 0.004990909248590469, + 0.005929779261350632, + 0.006319274194538593, + 0.005957688204944134, + 0.008117103017866611, + 0.007500362582504749, + 0.008247804827988148, + 0.09072217345237732, + 0.04237133264541626, + 0.02715384028851986, + 0.03395165130496025, + 0.03126247599720955, + 0.08853323757648468, + 0.015365407802164555, + 0.028150029480457306, + 0.007023834157735109, + 0.004978650249540806, + 0.007217178121209145, + 0.008988757617771626, + 0.004510899074375629, + 0.005568528547883034, + 0.0037024898920208216, + 0.0029993895441293716, + 0.04624757543206215, + 0.01994006149470806, + 0.009400295093655586, + 0.015624632127583027, + 0.012747595086693764, + 0.046301502734422684, + 0.00741376169025898, + 0.011646542698144913, + 0.007976454682648182, + 0.0077920774929225445, + 0.007648114580661058, + 0.0056482055224478245, + 0.009770532138645649, + 0.008110224269330502, + 0.0056249964982271194, + 0.00660322280600667, + 0.023298203945159912, + 0.012242945842444897, + 0.0051613436080515385, + 0.00738875288516283, + 0.0061429813504219055, + 0.020434726029634476, + 0.003704766510054469, + 0.004462455399334431, + 0.0012059310683980584, + 0.0015944059705361724, + 0.002072620438411832, + 0.0018965420313179493, + 0.0022078496403992176, + 0.001741744694299996, + 0.0014689918607473373, + 0.001692714635282755, + 0.004159429110586643, + 0.004739334341138601, + 0.002702274825423956, + 0.0034620105288922787, + 0.002899543382227421, + 0.003867194289341569, + 0.002179959323257208, + 0.002995162969455123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.03371515870094299 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351653039455414 + ], + "opposite_logprob_tokens": [ + -3.6513516902923584 + ], + "logprob_reward": -0.026351653039455414, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351653039455414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0877687856554985, + 0.03476070240139961, + 0.052799127995967865, + 0.4212489426136017, + 0.03140592202544212, + 0.2094642072916031, + 0.035029809921979904, + 0.01662091352045536, + 0.0622127503156662, + 0.04868878796696663 + ], + "credit_map": [ + 0.009204918518662453, + 0.011483617126941681, + 0.007971939630806446, + 0.009062724187970161, + 0.012025642208755016, + 0.011729990132153034, + 0.015260480344295502, + 0.011029469780623913, + 0.002989178989082575, + 0.004199910443276167, + 0.004625264089554548, + 0.002922507468611002, + 0.005769395735114813, + 0.004772083833813667, + 0.005038691218942404, + 0.004443668760359287, + 0.004545143339782953, + 0.00489405170083046, + 0.005255087744444609, + 0.006509209517389536, + 0.006940328981727362, + 0.008309073746204376, + 0.005956141743808985, + 0.010390090756118298, + 0.10702396929264069, + 0.04891258105635643, + 0.03135858103632927, + 0.03940214589238167, + 0.037145115435123444, + 0.10532981157302856, + 0.01884874515235424, + 0.033227983862161636, + 0.0035875244066119194, + 0.004409646615386009, + 0.004598632920533419, + 0.005961229559034109, + 0.0037414622493088245, + 0.004101520869880915, + 0.0026081025134772062, + 0.0023978031240403652, + 0.058129843324422836, + 0.024972833693027496, + 0.01212596707046032, + 0.018988098949193954, + 0.014476943761110306, + 0.058178920298814774, + 0.008313183672726154, + 0.014278415590524673, + 0.005653027445077896, + 0.004159433301538229, + 0.0045247687958180904, + 0.0026917902287095785, + 0.005767719820141792, + 0.005599189084023237, + 0.002991778077557683, + 0.003642104333266616, + 0.0018859519623219967, + 0.002880843821913004, + 0.0027532184030860662, + 0.0020608329214155674, + 0.0024809411261230707, + 0.001919227885082364, + 0.0013392684049904346, + 0.0013006288791075349, + 0.011659098789095879, + 0.007034826558083296, + 0.006881353911012411, + 0.009504350833594799, + 0.006288218777626753, + 0.01113717257976532, + 0.004401825368404388, + 0.005305905360728502, + 0.009121439419686794, + 0.00841040350496769, + 0.005116929765790701, + 0.005664447788149118, + 0.004391075111925602, + 0.008650360628962517, + 0.0031143701635301113, + 0.004219765774905682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.026351653039455414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632367476820946 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.008632367476820946, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632367476820946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2135756015777588, + 0.04465235769748688, + 0.0487256683409214, + 0.20602606236934662, + 0.08641061931848526, + 0.05400839075446129, + 0.0618363581597805, + 0.12137338519096375, + 0.03551480174064636, + 0.12787678837776184 + ], + "credit_map": [ + 0.020631058141589165, + 0.02974603697657585, + 0.03246689960360527, + 0.039837151765823364, + 0.024927997961640358, + 0.02341311238706112, + 0.02392863854765892, + 0.01862470991909504, + 0.0047733718529343605, + 0.005719996057450771, + 0.008662964217364788, + 0.004660917911678553, + 0.00597407715395093, + 0.004756726790219545, + 0.004795700777322054, + 0.005308604799211025, + 0.007384608034044504, + 0.005217608530074358, + 0.004319381434470415, + 0.009197354316711426, + 0.004423781763762236, + 0.00834875088185072, + 0.0037934216670691967, + 0.0060407621785998344, + 0.03583553805947304, + 0.0300089530646801, + 0.020696213468909264, + 0.01955719292163849, + 0.027043916285037994, + 0.03816855326294899, + 0.013569180853664875, + 0.021146506071090698, + 0.023763442412018776, + 0.01081695407629013, + 0.005158708896487951, + 0.005914990324527025, + 0.008905571885406971, + 0.020898325368762016, + 0.003916983958333731, + 0.00703564053401351, + 0.014278387650847435, + 0.006642517633736134, + 0.003043420147150755, + 0.006074285600334406, + 0.003716652747243643, + 0.014067953452467918, + 0.002823140937834978, + 0.0033620342146605253, + 0.012110556475818157, + 0.007102270144969225, + 0.006022678688168526, + 0.005468445830047131, + 0.00801089871674776, + 0.009737279266119003, + 0.006090130191296339, + 0.007294099777936935, + 0.03665297478437424, + 0.01798931322991848, + 0.005867247004061937, + 0.009257754310965538, + 0.008065109141170979, + 0.032049041241407394, + 0.0049362280406057835, + 0.0065557146444916725, + 0.008563169278204441, + 0.004319941624999046, + 0.002595579018816352, + 0.004559510853141546, + 0.0027441608253866434, + 0.008392319083213806, + 0.0020621144212782383, + 0.0022780047729611397, + 0.0318598598241806, + 0.01684476248919964, + 0.009195555932819843, + 0.01130704302340746, + 0.009984973818063736, + 0.032082222402095795, + 0.006468451116234064, + 0.010133911855518818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.008632367476820946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot captured from a medium close-up angle, using high-frame-rate slow-motion cinematography, focuses on a middle-aged Asian male chef. He wears a dark gray cotton headscarf, his face stained with grease and sweat from a hard day's work in the kitchen. He has just paused from his rapid chopping, one hand still tightly gripping a heavy kitchen knife, the blade hovering a few millimeters above the flour-stained wooden cutting board, the other resting on the stainless steel countertop. A strong overhead light streams down from above, sculpting the sharp lines of his face and highlighting the beads of sweat on his forehead and the oily sheen on his cheeks. His brow is slightly furrowed, his eyes sharp and focused, as if piercing the air before him, fixated on a single point beyond the camera. The slow-motion effect captures the countless tiny particles of flour dust and oil smoke floating in the air, dancing slowly within the beam of light like shimmering stardust. The background is a slightly blurred professional kitchen, and the stainless steel kitchenware reflects a cold luster. The whole scene creates an atmosphere of extreme concentration, tension and power in realistic and slightly warm colors.", + "seed": 14361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_12/index_2840_seed_14361_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0317212613299489, + "vlm_logprob": -0.0317212613299489, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.0317212613299489, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025489868596196175 + ], + "opposite_logprob_tokens": [ + -3.7754898071289062 + ], + "logprob_reward": -0.025489868596196175, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025489868596196175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29979193210601807, + 0.09713658690452576, + 0.06360730528831482, + 0.0988626703619957, + 0.09899138659238815, + 0.0745583102107048, + 0.09910416603088379, + 0.03440038859844208, + 0.03678606078028679, + 0.09676119685173035 + ], + "credit_map": [ + 0.0353689007461071, + 0.036617130041122437, + 0.05074299871921539, + 0.029038934037089348, + 0.052656140178442, + 0.02857905998826027, + 0.0404556579887867, + 0.026333104819059372, + 0.00927428063005209, + 0.010989655740559101, + 0.017523590475320816, + 0.01449565775692463, + 0.01557479240000248, + 0.008285907097160816, + 0.01239088550209999, + 0.008601810783147812, + 0.00623791990801692, + 0.012092150747776031, + 0.008583597838878632, + 0.008384348824620247, + 0.007280901074409485, + 0.0041929273866117, + 0.006521041505038738, + 0.01031441893428564, + 0.011668090708553791, + 0.015263555571436882, + 0.0127290403470397, + 0.019454864785075188, + 0.011087480001151562, + 0.006309633143246174, + 0.011610426008701324, + 0.010739577934145927, + 0.009262127801775932, + 0.009695804677903652, + 0.015010055154561996, + 0.007750269491225481, + 0.012836875393986702, + 0.009758472442626953, + 0.010684821754693985, + 0.023992963135242462, + 0.008504400961101055, + 0.010307885706424713, + 0.013327116146683693, + 0.008625784888863564, + 0.008232470601797104, + 0.004439054988324642, + 0.008545417338609695, + 0.012576174922287464, + 0.01033959724009037, + 0.014979361556470394, + 0.014709414914250374, + 0.01284155622124672, + 0.01260676421225071, + 0.008528425358235836, + 0.010548890568315983, + 0.01455016154795885, + 0.00416137557476759, + 0.004097622353583574, + 0.005565478466451168, + 0.004243901465088129, + 0.004227536730468273, + 0.0029461882077157497, + 0.0034781547728925943, + 0.005680131260305643, + 0.004352746531367302, + 0.007001073099672794, + 0.005603509489446878, + 0.00400462094694376, + 0.005136239342391491, + 0.003094025421887636, + 0.003344491356983781, + 0.004249355290085077, + 0.0096230898052454, + 0.009518329985439777, + 0.025991301983594894, + 0.017320087179541588, + 0.008809362538158894, + 0.008296908810734749, + 0.01044581737369299, + 0.006756295915693045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.025489868596196175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048703715205192566 + ], + "opposite_logprob_tokens": [ + -3.048703670501709 + ], + "logprob_reward": -0.048703715205192566, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048703715205192566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12197446078062057, + 0.06760860234498978, + 0.05115743726491928, + 0.048967454582452774, + 0.04659562557935715, + 0.1439111828804016, + 0.17746765911579132, + 0.24891598522663116, + 0.018652906641364098, + 0.0747487023472786 + ], + "credit_map": [ + 0.012950954027473927, + 0.015806732699275017, + 0.020358609035611153, + 0.01499144732952118, + 0.019130567088723183, + 0.00746356463059783, + 0.019099941477179527, + 0.01217263750731945, + 0.007128806319087744, + 0.008780935779213905, + 0.011909925378859043, + 0.008311493322253227, + 0.010540712624788284, + 0.005503020714968443, + 0.010302189737558365, + 0.005131518468260765, + 0.005557829048484564, + 0.009699730202555656, + 0.007128448691219091, + 0.007777675054967403, + 0.007064357865601778, + 0.0038238780107349157, + 0.004472882952541113, + 0.005632632412016392, + 0.005374192725867033, + 0.00749631505459547, + 0.007275902200490236, + 0.006874740589410067, + 0.00890304520726204, + 0.003093516919761896, + 0.004815473686903715, + 0.005134271457791328, + 0.004681383725255728, + 0.007274901960045099, + 0.007316981442272663, + 0.005032951012253761, + 0.00547507731243968, + 0.003908196929842234, + 0.0036475223023444414, + 0.009258612059056759, + 0.015934955328702927, + 0.019583120942115784, + 0.01950686238706112, + 0.01772123947739601, + 0.017216073349118233, + 0.011249852366745472, + 0.01525187585502863, + 0.027447201311588287, + 0.01660805754363537, + 0.017774006351828575, + 0.021954156458377838, + 0.028287189081311226, + 0.02466416545212269, + 0.025326045230031013, + 0.022874563932418823, + 0.019979465752840042, + 0.026401888579130173, + 0.03095267154276371, + 0.038591016083955765, + 0.048189908266067505, + 0.033181432634592056, + 0.024601168930530548, + 0.02604028396308422, + 0.020957622677087784, + 0.0018127212533727288, + 0.0019474891014397144, + 0.003729848423972726, + 0.0031388895586133003, + 0.0030922184232622385, + 0.0017266697250306606, + 0.0015079526929184794, + 0.0016971175791695714, + 0.007172990590333939, + 0.006829975172877312, + 0.020079363137483597, + 0.012743578292429447, + 0.007198241073638201, + 0.007913605310022831, + 0.007284295745193958, + 0.005526654422283173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.048703715205192566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043083496391773224 + ], + "opposite_logprob_tokens": [ + -3.168083429336548 + ], + "logprob_reward": -0.043083496391773224, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083496391773224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17270289361476898, + 0.08172933012247086, + 0.11413025110960007, + 0.11685675382614136, + 0.16988982260227203, + 0.11079301685094833, + 0.11932290345430374, + 0.045883744955062866, + 0.02492809109389782, + 0.043763190507888794 + ], + "credit_map": [ + 0.024481387808918953, + 0.02496112696826458, + 0.024020018056035042, + 0.02820787951350212, + 0.02461530826985836, + 0.009920565411448479, + 0.024537507444620132, + 0.011959106661379337, + 0.010598046705126762, + 0.0105755515396595, + 0.013126426376402378, + 0.014095896854996681, + 0.009183328598737717, + 0.00926636066287756, + 0.00905101653188467, + 0.005832710303366184, + 0.011102278716862202, + 0.015802742913365364, + 0.014531961642205715, + 0.02385883592069149, + 0.011215513572096825, + 0.007900513708591461, + 0.010889570228755474, + 0.018828844651579857, + 0.014128077775239944, + 0.02428877167403698, + 0.014074750244617462, + 0.01983593963086605, + 0.012818352319300175, + 0.007038778625428677, + 0.008769785054028034, + 0.01590230129659176, + 0.01833549700677395, + 0.02053852379322052, + 0.03340180590748787, + 0.020006904378533363, + 0.01965901255607605, + 0.011928478255867958, + 0.024070516228675842, + 0.02194909192621708, + 0.012214538641273975, + 0.014277546666562557, + 0.022805582731962204, + 0.01182646956294775, + 0.013183780014514923, + 0.009836231358349323, + 0.012003188021481037, + 0.014645680785179138, + 0.0129861980676651, + 0.018512513488531113, + 0.020472541451454163, + 0.015510841272771358, + 0.015914427116513252, + 0.008794600144028664, + 0.013953951187431812, + 0.01317784283310175, + 0.006034924183040857, + 0.005842410959303379, + 0.007224672008305788, + 0.00882430374622345, + 0.005650426261126995, + 0.003746502101421356, + 0.00439548771828413, + 0.00416502496227622, + 0.003278314834460616, + 0.00367423752322793, + 0.005346737802028656, + 0.0029419560451060534, + 0.002349653048440814, + 0.002012514276430011, + 0.0023795904126018286, + 0.002945088315755129, + 0.004089169204235077, + 0.005141828674823046, + 0.00696448003873229, + 0.008730131201446056, + 0.0064578247256577015, + 0.0043140980415046215, + 0.004793914500623941, + 0.0032717487774789333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.043083496391773224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033705245703458786 + ], + "opposite_logprob_tokens": [ + -3.408705234527588 + ], + "logprob_reward": -0.033705245703458786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033705245703458786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.202521413564682, + 0.12792342901229858, + 0.1045740619301796, + 0.14780673384666443, + 0.08928948640823364, + 0.061322420835494995, + 0.060443200170993805, + 0.06314653903245926, + 0.06268748641014099, + 0.08028529584407806 + ], + "credit_map": [ + 0.022857682779431343, + 0.030370021238923073, + 0.02753913216292858, + 0.03604709357023239, + 0.026352105662226677, + 0.016697002574801445, + 0.02554105967283249, + 0.017117304727435112, + 0.011740577407181263, + 0.017527462914586067, + 0.026664361357688904, + 0.022285662591457367, + 0.013354374095797539, + 0.009087084792554379, + 0.01755940355360508, + 0.009704504162073135, + 0.012077681720256805, + 0.014221115969121456, + 0.018419604748487473, + 0.015153942629694939, + 0.0142148956656456, + 0.007999674417078495, + 0.009405356831848621, + 0.013081791810691357, + 0.020146569237113, + 0.020835837349295616, + 0.019233843311667442, + 0.026777274906635284, + 0.017413616180419922, + 0.00884414091706276, + 0.014500745572149754, + 0.020054711028933525, + 0.01121033076196909, + 0.010953199118375778, + 0.014789339154958725, + 0.011982560157775879, + 0.012798127718269825, + 0.007105399388819933, + 0.008737181313335896, + 0.011713347397744656, + 0.006968171801418066, + 0.00810960028320551, + 0.012057642452418804, + 0.007516269572079182, + 0.007467067334800959, + 0.005245739594101906, + 0.005797759164124727, + 0.008160170167684555, + 0.005838575307279825, + 0.008841803297400475, + 0.009103539399802685, + 0.009385617449879646, + 0.0064363325946033, + 0.004953867755830288, + 0.008644401095807552, + 0.007239064201712608, + 0.0071707130409777164, + 0.0096292058005929, + 0.010209948755800724, + 0.007221943698823452, + 0.00626419298350811, + 0.005058879032731056, + 0.005167813040316105, + 0.012423835694789886, + 0.006773584987968206, + 0.00913239922374487, + 0.008210821077227592, + 0.006417407654225826, + 0.009647423401474953, + 0.007404860109090805, + 0.005841695237904787, + 0.009259295649826527, + 0.0076792072504758835, + 0.00892669428139925, + 0.017489293590188026, + 0.015390876680612564, + 0.00913151539862156, + 0.007640155032277107, + 0.007433904800564051, + 0.006593642756342888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.033705245703458786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762398075312376 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.00762398075312376, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762398075312376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23338159918785095, + 0.10549204796552658, + 0.05457047000527382, + 0.07111606001853943, + 0.05883878841996193, + 0.05622218921780586, + 0.09525220841169357, + 0.1966785341501236, + 0.03271641954779625, + 0.095731720328331 + ], + "credit_map": [ + 0.02963283658027649, + 0.040962040424346924, + 0.049911726266145706, + 0.03136375546455383, + 0.020933963358402252, + 0.015169769525527954, + 0.026308255270123482, + 0.019099263474345207, + 0.010441986843943596, + 0.013470008969306946, + 0.01584063097834587, + 0.02020995318889618, + 0.012494999915361404, + 0.010359343141317368, + 0.01118618156760931, + 0.011488945223391056, + 0.004586885217577219, + 0.006001662462949753, + 0.00782451219856739, + 0.010642546229064465, + 0.006673239171504974, + 0.003868447383865714, + 0.00534020783379674, + 0.00963296927511692, + 0.0068766875192523, + 0.0096150953322649, + 0.008328048512339592, + 0.014406302943825722, + 0.010219228453934193, + 0.005971514619886875, + 0.007195534650236368, + 0.008503643795847893, + 0.006549629848450422, + 0.0060818069614470005, + 0.009785627014935017, + 0.008624179288744926, + 0.008249111473560333, + 0.005073836538940668, + 0.005471510346978903, + 0.00900308508425951, + 0.005940255243331194, + 0.007764664478600025, + 0.007668651640415192, + 0.0055457791313529015, + 0.008159950375556946, + 0.005359270144253969, + 0.006169620435684919, + 0.009613996371626854, + 0.009292668662965298, + 0.008890480734407902, + 0.01122044213116169, + 0.015147843398153782, + 0.01497039757668972, + 0.015088964253664017, + 0.012133356183767319, + 0.008508051745593548, + 0.020338745787739754, + 0.02579120546579361, + 0.028768770396709442, + 0.0353773795068264, + 0.026432057842612267, + 0.02091442607343197, + 0.02230938710272312, + 0.016746556386351585, + 0.003473785473033786, + 0.0039273337461054325, + 0.0039220391772687435, + 0.003218273399397731, + 0.006259151268750429, + 0.005530113819986582, + 0.0034393775276839733, + 0.0029463451355695724, + 0.010811774991452694, + 0.010933415964245796, + 0.030241213738918304, + 0.011905230581760406, + 0.00950630847364664, + 0.006234397180378437, + 0.007876073941588402, + 0.008223299868404865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.00762398075312376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot captured from a medium close-up angle, using high-frame-rate slow-motion cinematography, focuses on a middle-aged Asian male chef. He wears a dark gray cotton headscarf, his face stained with grease and sweat from a hard day's work in the kitchen. He has just paused from his rapid chopping, one hand still tightly gripping a heavy kitchen knife, the blade hovering a few millimeters above the flour-stained wooden cutting board, the other resting on the stainless steel countertop. A strong overhead light streams down from above, sculpting the sharp lines of his face and highlighting the beads of sweat on his forehead and the oily sheen on his cheeks. His brow is slightly furrowed, his eyes sharp and focused, as if piercing the air before him, fixated on a single point beyond the camera. The slow-motion effect captures the countless tiny particles of flour dust and oil smoke floating in the air, dancing slowly within the beam of light like shimmering stardust. The background is a slightly blurred professional kitchen, and the stainless steel kitchenware reflects a cold luster. The whole scene creates an atmosphere of extreme concentration, tension and power in realistic and slightly warm colors.", + "seed": 14362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_12/index_2840_seed_14362_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04072345439344645, + "vlm_logprob": -0.04072345439344645, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04072345439344645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02610616758465767 + ], + "opposite_logprob_tokens": [ + -3.776106119155884 + ], + "logprob_reward": -0.02610616758465767, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02610616758465767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31981298327445984, + 0.2293628305196762, + 0.060316625982522964, + 0.09604859352111816, + 0.0551784485578537, + 0.0283048115670681, + 0.028776714578270912, + 0.06646253168582916, + 0.07178696244955063, + 0.043949492275714874 + ], + "credit_map": [ + 0.056180182844400406, + 0.03452169895172119, + 0.044190362095832825, + 0.025381432846188545, + 0.05643906444311142, + 0.02827480062842369, + 0.026536565274000168, + 0.04828888177871704, + 0.03520236536860466, + 0.019671374931931496, + 0.01786721870303154, + 0.023950429633259773, + 0.01738571748137474, + 0.02250448428094387, + 0.019551247358322144, + 0.07323001325130463, + 0.007433434948325157, + 0.009616744704544544, + 0.011027591302990913, + 0.008578796871006489, + 0.00789344310760498, + 0.005075682420283556, + 0.003923750948160887, + 0.006767182145267725, + 0.014295610599219799, + 0.019221169874072075, + 0.017275024205446243, + 0.011339528486132622, + 0.011879871599376202, + 0.005824707914143801, + 0.009743739850819111, + 0.006468946114182472, + 0.011436673812568188, + 0.014710860326886177, + 0.007247234228998423, + 0.004801287315785885, + 0.0042639607563614845, + 0.002826410811394453, + 0.006274597253650427, + 0.003617428708821535, + 0.0031393945682793856, + 0.003950884565711021, + 0.0045829820446670055, + 0.0025454629212617874, + 0.006215570028871298, + 0.0027174712158739567, + 0.0027066345792263746, + 0.00244641350582242, + 0.0035012923181056976, + 0.0038736299611628056, + 0.004302952438592911, + 0.004482051357626915, + 0.004415001720190048, + 0.002556656487286091, + 0.0027064215391874313, + 0.002938711317256093, + 0.0168059840798378, + 0.020114582031965256, + 0.004840624053031206, + 0.004158523865044117, + 0.007283403072506189, + 0.0041581932455301285, + 0.0045416452921926975, + 0.00455958629027009, + 0.018109692260622978, + 0.024086968973279, + 0.007320455275475979, + 0.004389457870274782, + 0.006229445803910494, + 0.004175394307821989, + 0.003688932629302144, + 0.0037866171915084124, + 0.010238710790872574, + 0.016541525721549988, + 0.0027939591091126204, + 0.0030066552571952343, + 0.004078171215951443, + 0.003106646705418825, + 0.0018848282052204013, + 0.002298998413607478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02610616758465767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871052876114845 + ], + "opposite_logprob_tokens": [ + -3.048710584640503 + ], + "logprob_reward": -0.04871052876114845, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871052876114845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.441849023103714, + 0.1965399831533432, + 0.06300494074821472, + 0.059980642050504684, + 0.06515605002641678, + 0.02590787597000599, + 0.03360987454652786, + 0.030255531892180443, + 0.041545379906892776, + 0.04215068742632866 + ], + "credit_map": [ + 0.05150208622217178, + 0.05048622190952301, + 0.05977695807814598, + 0.04676399379968643, + 0.07432384043931961, + 0.037784554064273834, + 0.03173529729247093, + 0.089476078748703, + 0.02922671101987362, + 0.01599024422466755, + 0.017631754279136658, + 0.02072514221072197, + 0.0207777488976717, + 0.01806217059493065, + 0.016286838799715042, + 0.05783938989043236, + 0.009258275851607323, + 0.0067286319099366665, + 0.014125979505479336, + 0.005143502727150917, + 0.009887930937111378, + 0.004166466183960438, + 0.00522382790222764, + 0.008470335975289345, + 0.005661460570991039, + 0.008126869797706604, + 0.013482760637998581, + 0.00787695124745369, + 0.008038097061216831, + 0.004066724330186844, + 0.007247745059430599, + 0.005480038467794657, + 0.011906342580914497, + 0.016041813418269157, + 0.00941076036542654, + 0.0077473497949540615, + 0.00480859586969018, + 0.003914718981832266, + 0.006803118158131838, + 0.004523356445133686, + 0.004310897551476955, + 0.0043141404166817665, + 0.003836266230791807, + 0.002293697092682123, + 0.004393122158944607, + 0.0022863950580358505, + 0.0022502809297293425, + 0.0022230776958167553, + 0.005162031855434179, + 0.006642412394285202, + 0.004605424124747515, + 0.005311326123774052, + 0.003760863794013858, + 0.001982593210414052, + 0.0031408665236085653, + 0.0030043574515730143, + 0.0032108996529132128, + 0.004048225935548544, + 0.003700202563777566, + 0.003658327041193843, + 0.006056519690901041, + 0.0032003161031752825, + 0.002668551867827773, + 0.0037124897353351116, + 0.006680040154606104, + 0.01052180863916874, + 0.003897738177329302, + 0.0032226615585386753, + 0.0077938660979270935, + 0.003690072102472186, + 0.002854143502190709, + 0.0028850550297647715, + 0.011469143442809582, + 0.012139969505369663, + 0.003987658303231001, + 0.0028065741062164307, + 0.004402091260999441, + 0.002567768096923828, + 0.002168788108974695, + 0.0026086976286023855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04871052876114845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900695502758026 + ], + "opposite_logprob_tokens": [ + -2.5790069103240967 + ], + "logprob_reward": -0.07900695502758026, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900695502758026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28249040246009827, + 0.20965434610843658, + 0.10776671022176743, + 0.12412559986114502, + 0.05873968452215195, + 0.03158348798751831, + 0.07514315843582153, + 0.06018916517496109, + 0.01569106988608837, + 0.0346163734793663 + ], + "credit_map": [ + 0.03923506662249565, + 0.03153696656227112, + 0.04353543743491173, + 0.023776814341545105, + 0.04445556551218033, + 0.02251489832997322, + 0.01931745745241642, + 0.058118220418691635, + 0.02806689590215683, + 0.015720615163445473, + 0.023553265258669853, + 0.022842280566692352, + 0.019021784886717796, + 0.019912954419851303, + 0.01668669655919075, + 0.06384985148906708, + 0.021719498559832573, + 0.012044640257954597, + 0.01779361255466938, + 0.014689854346215725, + 0.012028919532895088, + 0.009695406071841717, + 0.007667200174182653, + 0.012127580121159554, + 0.016661547124385834, + 0.023621615022420883, + 0.027918443083763123, + 0.01466098241508007, + 0.013742342591285706, + 0.008137683384120464, + 0.011220534332096577, + 0.008162444457411766, + 0.011058160103857517, + 0.01483287662267685, + 0.007700534537434578, + 0.006521974224597216, + 0.004124497063457966, + 0.0033843950368463993, + 0.006609255447983742, + 0.004507989622652531, + 0.004340616520494223, + 0.004550935234874487, + 0.005034963600337505, + 0.0047205789014697075, + 0.0035296494606882334, + 0.0025739071425050497, + 0.003747823415324092, + 0.0030850127805024385, + 0.022842148318886757, + 0.02804398722946644, + 0.00439481670036912, + 0.004399132449179888, + 0.004957912024110556, + 0.0025608011055737734, + 0.004106837324798107, + 0.00383752747438848, + 0.007914742454886436, + 0.006677559576928616, + 0.012705236673355103, + 0.007847997359931469, + 0.005964658223092556, + 0.006264713127166033, + 0.006261082366108894, + 0.00655317259952426, + 0.0023705391213297844, + 0.003782725427299738, + 0.0019236132502555847, + 0.0018705372931435704, + 0.0017197313718497753, + 0.001021908363327384, + 0.0016477697063237429, + 0.0013542457018047571, + 0.006498660426586866, + 0.010512898676097393, + 0.004775397013872862, + 0.002490938175469637, + 0.0040067629888653755, + 0.0021815479267388582, + 0.002003075322136283, + 0.002147095976397395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07900695502758026 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306145757436752 + ], + "opposite_logprob_tokens": [ + -3.1680614948272705 + ], + "logprob_reward": -0.04306145757436752, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306145757436752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4328382909297943, + 0.2325112521648407, + 0.09600608050823212, + 0.049336597323417664, + 0.05881929025053978, + 0.024667713791131973, + 0.01582346484065056, + 0.03019196353852749, + 0.017178984358906746, + 0.042626313865184784 + ], + "credit_map": [ + 0.042874451726675034, + 0.04555235803127289, + 0.05215240642428398, + 0.05009164661169052, + 0.08595523983240128, + 0.03160128742456436, + 0.02623945288360119, + 0.09837145358324051, + 0.041511040180921555, + 0.022149542346596718, + 0.017159074544906616, + 0.011815771460533142, + 0.016724079847335815, + 0.023695921525359154, + 0.01827421598136425, + 0.0811816081404686, + 0.02270788699388504, + 0.01603245548903942, + 0.020484240725636482, + 0.013024435378611088, + 0.006869528442621231, + 0.005433793645352125, + 0.004999686498194933, + 0.006454061716794968, + 0.007808972150087357, + 0.010981928557157516, + 0.007777079474180937, + 0.004687946755439043, + 0.007282142527401447, + 0.0028023794293403625, + 0.00453892583027482, + 0.003457226324826479, + 0.006551769562065601, + 0.010182799771428108, + 0.011471048928797245, + 0.008393664844334126, + 0.005307970102876425, + 0.00459721265360713, + 0.00701491255313158, + 0.005299918819218874, + 0.0032733341213315725, + 0.004530524369329214, + 0.0028966879472136497, + 0.0032816948369145393, + 0.0038684001192450523, + 0.0019134187605232, + 0.0028507302049547434, + 0.0020529243629425764, + 0.0018081278540194035, + 0.002085977466776967, + 0.0031346501782536507, + 0.0022025106009095907, + 0.0022685849107801914, + 0.0013803355395793915, + 0.0014259193558245897, + 0.0015173598658293486, + 0.003998558036983013, + 0.006173160392791033, + 0.005270580295473337, + 0.002487465273588896, + 0.0035955768544226885, + 0.0032102346885949373, + 0.0024593379348516464, + 0.0029970528557896614, + 0.0019332243828102946, + 0.0025440866593271494, + 0.0021326295100152493, + 0.0026355721056461334, + 0.0026948575396090746, + 0.0014719758182764053, + 0.0017068339511752129, + 0.00205980590544641, + 0.011968297883868217, + 0.017470909282565117, + 0.002447592793032527, + 0.001630262122489512, + 0.0033453309442847967, + 0.00233327760361135, + 0.0014435600023716688, + 0.001987090101465583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04306145757436752 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732163019478321 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732163019478321, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732163019478321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2806556820869446, + 0.16519974172115326, + 0.09737522900104523, + 0.12989380955696106, + 0.09314825385808945, + 0.05353229492902756, + 0.03695836290717125, + 0.09100470691919327, + 0.024913297966122627, + 0.027318663895130157 + ], + "credit_map": [ + 0.02422078140079975, + 0.02863204851746559, + 0.05213150009512901, + 0.04072938859462738, + 0.061409011483192444, + 0.02086680196225643, + 0.024171313270926476, + 0.028494851663708687, + 0.029763951897621155, + 0.016353189945220947, + 0.012001588940620422, + 0.010777046903967857, + 0.01257886178791523, + 0.016676658764481544, + 0.014524824917316437, + 0.05252361297607422, + 0.01892458088696003, + 0.008083446882665157, + 0.026538357138633728, + 0.01425299420952797, + 0.007883516140282154, + 0.005731153767555952, + 0.0077282339334487915, + 0.008232949301600456, + 0.02306796796619892, + 0.03046964667737484, + 0.027811143547296524, + 0.009835658594965935, + 0.011134468019008636, + 0.008703754283487797, + 0.010316764935851097, + 0.008554401807487011, + 0.01854630745947361, + 0.024116748943924904, + 0.010748803615570068, + 0.010061179287731647, + 0.008880283683538437, + 0.0057367864064872265, + 0.007715873420238495, + 0.007342261262238026, + 0.006990200839936733, + 0.007789064664393663, + 0.007964910008013248, + 0.007416492328047752, + 0.006998293101787567, + 0.00520365172997117, + 0.006694604177027941, + 0.004475078079849482, + 0.010265526361763477, + 0.011418038047850132, + 0.002857799408957362, + 0.0031100832857191563, + 0.0034116925671696663, + 0.0015758662484586239, + 0.0023418522905558348, + 0.001977503765374422, + 0.029663585126399994, + 0.03598325327038765, + 0.005882257129997015, + 0.003735621925443411, + 0.005129226017743349, + 0.0030876474920660257, + 0.004495871718972921, + 0.0030272495932877064, + 0.005758870858699083, + 0.007746345829218626, + 0.0018525490304455161, + 0.0017516011139377952, + 0.0033960926812142134, + 0.0014658874133601785, + 0.0016054996522143483, + 0.0013364519691094756, + 0.006066455505788326, + 0.006444927770644426, + 0.0023922708351165056, + 0.0022139796055853367, + 0.0033486001193523407, + 0.0026101747062057257, + 0.002249594312161207, + 0.0019926608074456453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.006732163019478321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot captured from a medium close-up angle, using high-frame-rate slow-motion cinematography, focuses on a middle-aged Asian male chef. He wears a dark gray cotton headscarf, his face stained with grease and sweat from a hard day's work in the kitchen. He has just paused from his rapid chopping, one hand still tightly gripping a heavy kitchen knife, the blade hovering a few millimeters above the flour-stained wooden cutting board, the other resting on the stainless steel countertop. A strong overhead light streams down from above, sculpting the sharp lines of his face and highlighting the beads of sweat on his forehead and the oily sheen on his cheeks. His brow is slightly furrowed, his eyes sharp and focused, as if piercing the air before him, fixated on a single point beyond the camera. The slow-motion effect captures the countless tiny particles of flour dust and oil smoke floating in the air, dancing slowly within the beam of light like shimmering stardust. The background is a slightly blurred professional kitchen, and the stainless steel kitchenware reflects a cold luster. The whole scene creates an atmosphere of extreme concentration, tension and power in realistic and slightly warm colors.", + "seed": 14363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_12/index_2840_seed_14363_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023960961541160942, + "vlm_logprob": -0.023960961541160942, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.023960961541160942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03261199966073036 + ], + "opposite_logprob_tokens": [ + -3.5326120853424072 + ], + "logprob_reward": -0.03261199966073036, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03261199966073036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14943164587020874, + 0.04192166402935982, + 0.44796112179756165, + 0.04870111867785454, + 0.07126566767692566, + 0.02003418281674385, + 0.08685921877622604, + 0.06249877065420151, + 0.027358287945389748, + 0.043968286365270615 + ], + "credit_map": [ + 0.015644704923033714, + 0.02182888425886631, + 0.018983162939548492, + 0.02219364419579506, + 0.019386693835258484, + 0.01802569068968296, + 0.019433598965406418, + 0.013935266993939877, + 0.004768777638673782, + 0.0038203629665076733, + 0.00569075858220458, + 0.006216744426637888, + 0.006070080678910017, + 0.005033623892813921, + 0.004959736950695515, + 0.005361583549529314, + 0.028288530185818672, + 0.03417918086051941, + 0.033985983580350876, + 0.05023763328790665, + 0.08962373435497284, + 0.07886555045843124, + 0.03960961475968361, + 0.0931708887219429, + 0.0033643313217908144, + 0.0023864225950092077, + 0.0033953189849853516, + 0.003459048457443714, + 0.010169742628932, + 0.009045714512467384, + 0.007680635433644056, + 0.009199908934533596, + 0.004672315903007984, + 0.006131071597337723, + 0.004816890694200993, + 0.007128067780286074, + 0.013875826261937618, + 0.01104191318154335, + 0.01174844242632389, + 0.011851146817207336, + 0.0021780456881970167, + 0.001962691778317094, + 0.0018666264368221164, + 0.0018049736972898245, + 0.003199897473677993, + 0.0027822174597531557, + 0.0026499731466174126, + 0.0035897581838071346, + 0.004694278817623854, + 0.005060391966253519, + 0.0047233388759195805, + 0.008841590955853462, + 0.012569587677717209, + 0.014955587685108185, + 0.025552639737725258, + 0.01046180259436369, + 0.003624530741944909, + 0.0033863272983580828, + 0.0032129990868270397, + 0.003886463586241007, + 0.014937302097678185, + 0.011631558649241924, + 0.009335304610431194, + 0.012484284117817879, + 0.0033033217769116163, + 0.003278267104178667, + 0.002379980171099305, + 0.002964058890938759, + 0.004145554732531309, + 0.004174896981567144, + 0.0027735040057450533, + 0.004338703118264675, + 0.0032807355746626854, + 0.003219539998099208, + 0.002313100267201662, + 0.00252538756467402, + 0.010317527689039707, + 0.008142910897731781, + 0.004999659024178982, + 0.009169425815343857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.03261199966073036 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311124235391617 + ], + "opposite_logprob_tokens": [ + -3.1681113243103027 + ], + "logprob_reward": -0.04311124235391617, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311124235391617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21019111573696136, + 0.16296613216400146, + 0.16043326258659363, + 0.06865193694829941, + 0.06609677523374557, + 0.034046780318021774, + 0.10209640115499496, + 0.08959223330020905, + 0.03848317265510559, + 0.06744223833084106 + ], + "credit_map": [ + 0.017889443784952164, + 0.0167233943939209, + 0.02270166575908661, + 0.024614164605736732, + 0.039158809930086136, + 0.032612379640340805, + 0.0237660501152277, + 0.03272518515586853, + 0.01650739088654518, + 0.01294107548892498, + 0.02680644765496254, + 0.02007448673248291, + 0.026276644319295883, + 0.02111651748418808, + 0.014890533871948719, + 0.024353018030524254, + 0.016644898802042007, + 0.013835974968969822, + 0.020596280694007874, + 0.022246262058615685, + 0.029375305399298668, + 0.020188061520457268, + 0.019659986719489098, + 0.017886491492390633, + 0.006648844573646784, + 0.008053150959312916, + 0.006924356799572706, + 0.0062019494362175465, + 0.012731673195958138, + 0.010002854280173779, + 0.0066842855885624886, + 0.011404813267290592, + 0.004776503425091505, + 0.008840386755764484, + 0.005398273933678865, + 0.0076190815307199955, + 0.010792705230414867, + 0.009939986281096935, + 0.009878954850137234, + 0.008850883692502975, + 0.003292741719633341, + 0.0028412879910320044, + 0.0044533563777804375, + 0.0037405923940241337, + 0.006377493031322956, + 0.005273357965052128, + 0.0031019991729408503, + 0.004965949337929487, + 0.005095993168652058, + 0.006332384422421455, + 0.005556646268814802, + 0.010402519255876541, + 0.01371034886687994, + 0.01786942034959793, + 0.03066065162420273, + 0.012468431144952774, + 0.007464535068720579, + 0.005361180752515793, + 0.00488635478541255, + 0.004266112577170134, + 0.022104058414697647, + 0.016443658620119095, + 0.011040850542485714, + 0.0180254727602005, + 0.004339428618550301, + 0.002838261891156435, + 0.003320613643154502, + 0.003015730530023575, + 0.006272720638662577, + 0.006257684901356697, + 0.006407399661839008, + 0.00603133300319314, + 0.004905648995190859, + 0.004653631243854761, + 0.004692491143941879, + 0.00420777453109622, + 0.014762545004487038, + 0.011506706476211548, + 0.009306964464485645, + 0.013406465761363506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04311124235391617 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029814520850777626 + ], + "opposite_logprob_tokens": [ + -3.5298144817352295 + ], + "logprob_reward": -0.029814520850777626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029814520850777626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12132340669631958, + 0.05511278659105301, + 0.1868455410003662, + 0.10497903823852539, + 0.14599518477916718, + 0.07043275982141495, + 0.09922319650650024, + 0.030690697953104973, + 0.13490581512451172, + 0.050491634756326675 + ], + "credit_map": [ + 0.014720180071890354, + 0.019621573388576508, + 0.015391708351671696, + 0.01789931021630764, + 0.01715436577796936, + 0.01434016041457653, + 0.011368297971785069, + 0.0108278077095747, + 0.003930370323359966, + 0.004742476157844067, + 0.008189788088202477, + 0.006799658294767141, + 0.009369362145662308, + 0.0071836733259260654, + 0.007601223886013031, + 0.00729623157531023, + 0.015020676888525486, + 0.029643557965755463, + 0.02225014567375183, + 0.026986895129084587, + 0.016948137432336807, + 0.022068167105317116, + 0.026639558374881744, + 0.027288397774100304, + 0.007799828425049782, + 0.012567291036248207, + 0.008864657022058964, + 0.015304799191653728, + 0.016629375517368317, + 0.014638502150774002, + 0.01292894035577774, + 0.016245637089014053, + 0.009392878971993923, + 0.009776903316378593, + 0.00955699011683464, + 0.015752069652080536, + 0.032057445496320724, + 0.023887798190116882, + 0.018960149958729744, + 0.026610948145389557, + 0.004794447682797909, + 0.005170272663235664, + 0.007993621751666069, + 0.007708529941737652, + 0.015322655439376831, + 0.01022711955010891, + 0.00858725793659687, + 0.010628858581185341, + 0.00505405617877841, + 0.005423280410468578, + 0.0060602291487157345, + 0.010198177769780159, + 0.012870796024799347, + 0.017222514376044273, + 0.029763586819171906, + 0.012630551122128963, + 0.0030478816479444504, + 0.001962261041626334, + 0.0024288094136863947, + 0.0025178310461342335, + 0.005043583922088146, + 0.005260549020022154, + 0.006445640232414007, + 0.0039841411635279655, + 0.008037402294576168, + 0.007754485588520765, + 0.010839121416211128, + 0.012440686114132404, + 0.03083861619234085, + 0.02366112358868122, + 0.012886187061667442, + 0.028448190540075302, + 0.0035511492751538754, + 0.003206277498975396, + 0.006859525106847286, + 0.0032865882385522127, + 0.010178150609135628, + 0.007965045049786568, + 0.006403539329767227, + 0.009041357785463333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.029814520850777626 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078160256147385 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078160256147385, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078160256147385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24526697397232056, + 0.06600067764520645, + 0.0803900733590126, + 0.1588241010904312, + 0.04533802717924118, + 0.09112103283405304, + 0.17417483031749725, + 0.06102735176682472, + 0.05207961052656174, + 0.025777382776141167 + ], + "credit_map": [ + 0.023254601284861565, + 0.02364315092563629, + 0.028187185525894165, + 0.029889296740293503, + 0.04230325669050217, + 0.03844572603702545, + 0.024682991206645966, + 0.03486078232526779, + 0.004736996255815029, + 0.004597155377268791, + 0.006120721343904734, + 0.00989443901926279, + 0.012442849576473236, + 0.009790265001356602, + 0.0059858704917132854, + 0.012432379648089409, + 0.008597955107688904, + 0.008603602647781372, + 0.007469167001545429, + 0.009213250130414963, + 0.017089154571294785, + 0.011301180347800255, + 0.008653121069073677, + 0.009462637826800346, + 0.017795350402593613, + 0.03078893944621086, + 0.01723286882042885, + 0.01878182776272297, + 0.0189763642847538, + 0.019741881638765335, + 0.018833640962839127, + 0.016673218458890915, + 0.0030818607192486525, + 0.005552332848310471, + 0.003310069441795349, + 0.006213825661689043, + 0.005874790251255035, + 0.006779145449399948, + 0.007934452034533024, + 0.006591551471501589, + 0.00579011021181941, + 0.00459453696385026, + 0.006732761859893799, + 0.007206721697002649, + 0.02019585482776165, + 0.016345176845788956, + 0.009445798583328724, + 0.020810073241591454, + 0.008237487636506557, + 0.009469983167946339, + 0.01061118021607399, + 0.017816083505749702, + 0.024540739133954048, + 0.030234163627028465, + 0.051247738301754, + 0.022017447277903557, + 0.003560945624485612, + 0.003213118528947234, + 0.003588078310713172, + 0.003210667986422777, + 0.0159166119992733, + 0.011117011308670044, + 0.00808478519320488, + 0.012336133047938347, + 0.005159942898899317, + 0.004043675493448973, + 0.006726844236254692, + 0.005956638604402542, + 0.00884795468300581, + 0.008042627014219761, + 0.005033867433667183, + 0.00826806016266346, + 0.0016116517363116145, + 0.0025562418159097433, + 0.0031854792032390833, + 0.0030917769763618708, + 0.004241019021719694, + 0.003940153867006302, + 0.0036769742146134377, + 0.0034740872215479612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.011078160256147385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged Asian male chef?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef wearing a dark gray cotton headscarf with grease and sweat stains on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef holding a heavy kitchen knife above a flour-stained wooden cutting board with one hand on a stainless steel countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot as a static medium close-up with high-frame-rate slow-motion capturing floating flour dust and oil smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a slightly blurred professional kitchen with stainless steel kitchenware reflecting a cold luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003188884584233165 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.003188884584233165, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003188884584233165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18011188507080078, + 0.2001509964466095, + 0.14676889777183533, + 0.07335980236530304, + 0.07856371253728867, + 0.025584660470485687, + 0.15553094446659088, + 0.06450553983449936, + 0.025916507467627525, + 0.0495070144534111 + ], + "credit_map": [ + 0.019081491976976395, + 0.0216806810349226, + 0.02948850765824318, + 0.03203628584742546, + 0.02690732292830944, + 0.020540837198495865, + 0.013892196118831635, + 0.01648457534611225, + 0.01744586043059826, + 0.013781099580228329, + 0.022548723965883255, + 0.0196104533970356, + 0.03858628123998642, + 0.031145086511969566, + 0.022093649953603745, + 0.034939851611852646, + 0.011018041521310806, + 0.014501862227916718, + 0.01779826544225216, + 0.011988108046352863, + 0.02399202063679695, + 0.02490812912583351, + 0.012338366359472275, + 0.030224112793803215, + 0.005748612340539694, + 0.006954239681363106, + 0.006569618824869394, + 0.00893857330083847, + 0.013977187685668468, + 0.011178226210176945, + 0.011563942767679691, + 0.008429397828876972, + 0.00731324777007103, + 0.010546390898525715, + 0.008633341640233994, + 0.009414209052920341, + 0.0120107876136899, + 0.010802950710058212, + 0.009646421298384666, + 0.010196365416049957, + 0.0016342306043952703, + 0.0028827099595218897, + 0.002705017337575555, + 0.003059182083234191, + 0.004211684688925743, + 0.003895514877513051, + 0.004184663761407137, + 0.0030116578564047813, + 0.016130926087498665, + 0.018030351027846336, + 0.029707401990890503, + 0.020980507135391235, + 0.01796318218111992, + 0.018520638346672058, + 0.017373232170939445, + 0.01682470738887787, + 0.006077465135604143, + 0.006535895634442568, + 0.004811289254575968, + 0.005422120448201895, + 0.01187288947403431, + 0.009564229287207127, + 0.009024612605571747, + 0.011197030544281006, + 0.002566928742453456, + 0.0019950270652770996, + 0.0038856547325849533, + 0.0025517747271806, + 0.003947761375457048, + 0.0038002256769686937, + 0.0034353614319115877, + 0.0037337741814553738, + 0.004845260642468929, + 0.004449991509318352, + 0.0050377920269966125, + 0.0043972874991595745, + 0.008952107280492783, + 0.007529026363044977, + 0.0063408855348825455, + 0.007954665459692478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.003188884584233165 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_tensor([7125, 7125])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_tensor([7125, 7125])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..49206f816e330358709bee8fe21a85e1a0fca6d1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_tensor([7125, 7125])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, static shot, shot from a horizontal angle, depicts a world constructed in a refined pixel art style, with sharp, blocky edges. At the center of the frame is an elderly man with a graying beard and hair, his sparse white hair pulled back in a small bun. His face is lined with deep wrinkles, created by pixelated blocks, and his eyes are focused and serene. He wears a traditional dark gray coarse Chinese robe, its folds rendered with varying shades of gray. He stands in a quaint Chinese courtyard, his ground paved with bluestone slabs, interspersed between the cracks of the slabs with a profusion of golden and crimson pixelated fallen leaves. The elderly man is slightly bent over, his hands tightly gripping a long-handled bamboo broom—the handle rendered as a long string of brown pixels, the end of the broom a cluster of scattered yellow pixels. He methodically and unhurriedly swept the fallen leaves to one side. Each sweep was slow and rhythmic. The tip of the bamboo broom gently brushed the ground, sending pixelated leaves fluttering briefly in the air before gathering into a small pile. A soft, diffuse light evenly bathed the courtyard, devoid of harsh highlights, casting only soft, jagged, blocky shadows around the figures and the corners of the walls. The scene, dominated by earth tones and grays, was punctuated by the vibrant, warm colors of the fallen leaves, creating a tranquil and focused atmosphere that blended the nostalgia of retro gaming with the essence of Eastern Zen.", + "seed": 23362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_7125_seed_23362_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02394847306422889, + "vlm_logprob": -0.02394847306422889, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.02394847306422889, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07053662836551666 + ], + "opposite_logprob_tokens": [ + -2.8205366134643555 + ], + "logprob_reward": -0.07053662836551666, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07053662836551666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08426698297262192, + 0.1973448395729065, + 0.031133275479078293, + 0.046486444771289825, + 0.2138834297657013, + 0.2793271839618683, + 0.056173406541347504, + 0.04746118560433388, + 0.021356191486120224, + 0.022567102685570717 + ], + "credit_map": [ + 0.02317608892917633, + 0.004777677357196808, + 0.006776100490242243, + 0.0034673467744141817, + 0.003917235881090164, + 0.029665183275938034, + 0.005050395615398884, + 0.007436948828399181, + 0.07584158331155777, + 0.007098884787410498, + 0.00812754686921835, + 0.005636897869408131, + 0.006478165742009878, + 0.07766387611627579, + 0.005629143211990595, + 0.010868746787309647, + 0.00512361153960228, + 0.003310608211904764, + 0.00514706876128912, + 0.0037568893749266863, + 0.005088428035378456, + 0.002824444090947509, + 0.0034590037539601326, + 0.002423218684270978, + 0.0072133783251047134, + 0.004846339114010334, + 0.006081848405301571, + 0.0048768180422484875, + 0.009759726002812386, + 0.0046477108262479305, + 0.0033202480990439653, + 0.005740368273109198, + 0.0822409838438034, + 0.008654720149934292, + 0.008838449604809284, + 0.004800237715244293, + 0.0057941279374063015, + 0.08446084707975388, + 0.006315687671303749, + 0.012778360396623611, + 0.1096598282456398, + 0.009662223979830742, + 0.010698922909796238, + 0.005289511755108833, + 0.008891926147043705, + 0.11170785874128342, + 0.008809193037450314, + 0.014607681892812252, + 0.014402803033590317, + 0.003939557820558548, + 0.0037468806840479374, + 0.004542100243270397, + 0.002805139636620879, + 0.019024331122636795, + 0.0036856592632830143, + 0.004026929382234812, + 0.012713571079075336, + 0.0034011737443506718, + 0.0044378419406712055, + 0.0022964216768741608, + 0.0030375749338418245, + 0.014620200730860233, + 0.003065873868763447, + 0.0038885201793164015, + 0.0028760400600731373, + 0.003932021092623472, + 0.0021130559034645557, + 0.0023653488606214523, + 0.002325677312910557, + 0.0022724997252225876, + 0.0027006713207811117, + 0.0027708744164556265, + 0.0026548742316663265, + 0.002000881126150489, + 0.003169455798342824, + 0.003493719734251499, + 0.0025445520877838135, + 0.0017175039974972606, + 0.002427377039566636, + 0.004558737389743328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.07053662836551666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067501612938940525 + ], + "opposite_logprob_tokens": [ + -5.006750106811523 + ], + "logprob_reward": -0.0067501612938940525, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067501612938940525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21249696612358093, + 0.10821875929832458, + 0.06777394562959671, + 0.07943369448184967, + 0.06483976542949677, + 0.22407203912734985, + 0.12083931267261505, + 0.01606537029147148, + 0.07910466939210892, + 0.02715550921857357 + ], + "credit_map": [ + 0.05643285810947418, + 0.019307393580675125, + 0.014295640401542187, + 0.012559822760522366, + 0.012234366498887539, + 0.06832180917263031, + 0.011331574060022831, + 0.01801351085305214, + 0.031235015019774437, + 0.00636259512975812, + 0.004998823162168264, + 0.011957733891904354, + 0.006346139125525951, + 0.03489396721124649, + 0.0061129191890358925, + 0.006311560980975628, + 0.011254576034843922, + 0.0057699428871273994, + 0.007745495997369289, + 0.009645783342421055, + 0.009763715788722038, + 0.010314707644283772, + 0.007789155468344688, + 0.005490571726113558, + 0.01928113028407097, + 0.007595962844789028, + 0.01068661455065012, + 0.005327623337507248, + 0.011379009112715721, + 0.010963788256049156, + 0.008060270920395851, + 0.006139293313026428, + 0.01447885949164629, + 0.0070992824621498585, + 0.004600713029503822, + 0.0052384426817297935, + 0.007154968101531267, + 0.01540534757077694, + 0.005623459815979004, + 0.005238698795437813, + 0.08776652812957764, + 0.007220901548862457, + 0.008346674963831902, + 0.004005781374871731, + 0.007451435085386038, + 0.08963926136493683, + 0.007228404749184847, + 0.012413023971021175, + 0.02760154753923416, + 0.00934064295142889, + 0.008327892981469631, + 0.008230119943618774, + 0.009202258661389351, + 0.037768419831991196, + 0.010159523226320744, + 0.01020890660583973, + 0.0030153533443808556, + 0.0020282079931348562, + 0.0011709339451044798, + 0.0017653751419857144, + 0.0020529129542410374, + 0.0030295199248939753, + 0.0016441787593066692, + 0.0013588879955932498, + 0.02794836461544037, + 0.004804361145943403, + 0.004048143979161978, + 0.004075154196470976, + 0.004448655527085066, + 0.02493234910070896, + 0.0032015019096434116, + 0.005646140314638615, + 0.004161699675023556, + 0.0022355737164616585, + 0.00458167027682066, + 0.0029396158643066883, + 0.004498901776969433, + 0.002624273067340255, + 0.0032726412173360586, + 0.0028411352541297674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0067501612938940525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028225842397660017 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.0028225842397660017, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028225842397660017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19599682092666626, + 0.06029251962900162, + 0.0449112169444561, + 0.04113418981432915, + 0.23971089720726013, + 0.1468663215637207, + 0.0344468355178833, + 0.06808841228485107, + 0.09614485502243042, + 0.07240800559520721 + ], + "credit_map": [ + 0.05727251246571541, + 0.009845311753451824, + 0.012884367257356644, + 0.007597798015922308, + 0.011837831698358059, + 0.07158214598894119, + 0.010963067412376404, + 0.014013770036399364, + 0.005863082595169544, + 0.006184497848153114, + 0.005627407692372799, + 0.01465493906289339, + 0.005741064436733723, + 0.0071341185830533504, + 0.007664933335036039, + 0.007422476075589657, + 0.006408055312931538, + 0.0060997758992016315, + 0.005225280299782753, + 0.004045493435114622, + 0.006325824186205864, + 0.004441170487552881, + 0.005720399785786867, + 0.006645215675234795, + 0.007812545634806156, + 0.004170569591224194, + 0.0064496672712266445, + 0.003969486337155104, + 0.004947707988321781, + 0.004663838539272547, + 0.003772847354412079, + 0.0053475224412977695, + 0.08953483402729034, + 0.010670596733689308, + 0.011108647100627422, + 0.005754019133746624, + 0.007761728018522263, + 0.09518412500619888, + 0.0067775920033454895, + 0.012919336557388306, + 0.0563039593398571, + 0.005419519264250994, + 0.006083193700760603, + 0.0028442207258194685, + 0.00523997750133276, + 0.05776000767946243, + 0.005352313164621592, + 0.007863126695156097, + 0.003733324585482478, + 0.005405514035373926, + 0.004548058845102787, + 0.004420233424752951, + 0.0035292357206344604, + 0.0032793977297842503, + 0.004511610139161348, + 0.005019461736083031, + 0.01753513514995575, + 0.004614479839801788, + 0.007927837781608105, + 0.0036379259545356035, + 0.003950728569179773, + 0.020097315311431885, + 0.004360288847237825, + 0.005964695475995541, + 0.02915235236287117, + 0.008257784880697727, + 0.005337280686944723, + 0.004621084779500961, + 0.006691805087029934, + 0.030033618211746216, + 0.005779558792710304, + 0.00627136928960681, + 0.014368794858455658, + 0.005979607347398996, + 0.008473647758364677, + 0.010762574151158333, + 0.00697017228230834, + 0.00884268432855606, + 0.005878137890249491, + 0.011132379993796349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0028225842397660017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594549672678113 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.00594549672678113, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594549672678113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1641414910554886, + 0.41777265071868896, + 0.030107973143458366, + 0.020120160654187202, + 0.0818609893321991, + 0.13324709236621857, + 0.039309680461883545, + 0.06278622150421143, + 0.027423322200775146, + 0.023230448365211487 + ], + "credit_map": [ + 0.048532430082559586, + 0.010197769850492477, + 0.009840158745646477, + 0.006821997929364443, + 0.007489548996090889, + 0.061217498034238815, + 0.008030489087104797, + 0.012011590413749218, + 0.16269074380397797, + 0.013949510641396046, + 0.01683511957526207, + 0.011366747319698334, + 0.013587002642452717, + 0.16480277478694916, + 0.010963059030473232, + 0.02357768639922142, + 0.004945690277963877, + 0.003612907836213708, + 0.002941614016890526, + 0.0034655178897082806, + 0.004865581169724464, + 0.003489024704322219, + 0.002695024013519287, + 0.004092613235116005, + 0.0025948642287403345, + 0.0024064050521701574, + 0.002734315348789096, + 0.002855257596820593, + 0.00303466385230422, + 0.0024449490010738373, + 0.0015699046198278666, + 0.0024798004887998104, + 0.02694106288254261, + 0.005186517722904682, + 0.004592423792928457, + 0.003957049921154976, + 0.005088404286652803, + 0.026594894006848335, + 0.004711829125881195, + 0.004788803867995739, + 0.048916857689619064, + 0.004744064528495073, + 0.006015103776007891, + 0.0038893024902790785, + 0.005165421403944492, + 0.04839995130896568, + 0.006239997688680887, + 0.00987639557570219, + 0.007601471617817879, + 0.0043968805111944675, + 0.00396398501470685, + 0.0038933989126235247, + 0.003561439225450158, + 0.00962808821350336, + 0.003353644162416458, + 0.0029107718728482723, + 0.019159117713570595, + 0.003247759770601988, + 0.004314003977924585, + 0.002293214201927185, + 0.003134624334052205, + 0.02257842943072319, + 0.003215826116502285, + 0.004843247588723898, + 0.009168758988380432, + 0.0018129285890609026, + 0.0016197626246139407, + 0.0020035787019878626, + 0.0017944255378097296, + 0.006952131167054176, + 0.0018216753378510475, + 0.002250062068924308, + 0.007043099962174892, + 0.0020091324113309383, + 0.001879714778624475, + 0.00206491956487298, + 0.0018236400792375207, + 0.004582047928124666, + 0.0017546711023896933, + 0.0020732213743031025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.00594549672678113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033687494695186615 + ], + "opposite_logprob_tokens": [ + -3.4086875915527344 + ], + "logprob_reward": -0.033687494695186615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687494695186615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06488362699747086, + 0.18301920592784882, + 0.03520561382174492, + 0.03245176002383232, + 0.15554699301719666, + 0.2588305175304413, + 0.07877586036920547, + 0.09852439910173416, + 0.051139894872903824, + 0.04162215068936348 + ], + "credit_map": [ + 0.013649455271661282, + 0.005939087364822626, + 0.004952478222548962, + 0.004638342652469873, + 0.006540818605571985, + 0.01754145324230194, + 0.00712020555511117, + 0.004501786082983017, + 0.06574741005897522, + 0.007665297947824001, + 0.00791550986468792, + 0.005505864508450031, + 0.00965289305895567, + 0.06638866662979126, + 0.008851378224790096, + 0.011292166076600552, + 0.006613656412810087, + 0.0026353425346314907, + 0.004473272245377302, + 0.0027245134115219116, + 0.004073743708431721, + 0.005252284929156303, + 0.002580120228230953, + 0.0068526798859238625, + 0.008021173998713493, + 0.003972878213971853, + 0.0030535145197063684, + 0.003293173387646675, + 0.00453140027821064, + 0.004180451855063438, + 0.002151173073798418, + 0.0032479942310601473, + 0.05948053300380707, + 0.006362559739500284, + 0.006768564227968454, + 0.003647606121376157, + 0.004564245697110891, + 0.06119861453771591, + 0.004812333267182112, + 0.008712545037269592, + 0.10085704177618027, + 0.009278452023863792, + 0.010261088609695435, + 0.005125124007463455, + 0.008715813048183918, + 0.10250432789325714, + 0.00871069636195898, + 0.013377990573644638, + 0.018107276409864426, + 0.008021153509616852, + 0.005198706407099962, + 0.005933831445872784, + 0.004951049108058214, + 0.025260210037231445, + 0.00539066968485713, + 0.005912966560572386, + 0.026756178587675095, + 0.007241424173116684, + 0.009210462681949139, + 0.0051318202167749405, + 0.0064771901816129684, + 0.029937004670500755, + 0.005489283706992865, + 0.008281029760837555, + 0.018458513543009758, + 0.0027106325142085552, + 0.0026400024071335793, + 0.003092181868851185, + 0.0023160125128924847, + 0.016614174470305443, + 0.002049891510978341, + 0.003258483950048685, + 0.005966537166386843, + 0.0034029134549200535, + 0.005940691567957401, + 0.0053266906179487705, + 0.004562345799058676, + 0.003598820883780718, + 0.0046042175963521, + 0.00821993499994278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.033687494695186615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a horizontal angle, depicts a world constructed in a refined pixel art style, with sharp, blocky edges. At the center of the frame is an elderly man with a graying beard and hair, his sparse white hair pulled back in a small bun. His face is lined with deep wrinkles, created by pixelated blocks, and his eyes are focused and serene. He wears a traditional dark gray coarse Chinese robe, its folds rendered with varying shades of gray. He stands in a quaint Chinese courtyard, his ground paved with bluestone slabs, interspersed between the cracks of the slabs with a profusion of golden and crimson pixelated fallen leaves. The elderly man is slightly bent over, his hands tightly gripping a long-handled bamboo broom—the handle rendered as a long string of brown pixels, the end of the broom a cluster of scattered yellow pixels. He methodically and unhurriedly swept the fallen leaves to one side. Each sweep was slow and rhythmic. The tip of the bamboo broom gently brushed the ground, sending pixelated leaves fluttering briefly in the air before gathering into a small pile. A soft, diffuse light evenly bathed the courtyard, devoid of harsh highlights, casting only soft, jagged, blocky shadows around the figures and the corners of the walls. The scene, dominated by earth tones and grays, was punctuated by the vibrant, warm colors of the fallen leaves, creating a tranquil and focused atmosphere that blended the nostalgia of retro gaming with the essence of Eastern Zen.", + "seed": 23363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_7125_seed_23363_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03402455030009151, + "vlm_logprob": -0.03402455030009151, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03402455030009151, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08558270335197449 + ], + "opposite_logprob_tokens": [ + -2.585582733154297 + ], + "logprob_reward": -0.08558270335197449, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08558270335197449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10400526225566864, + 0.02925686351954937, + 0.03019416704773903, + 0.42038974165916443, + 0.055669356137514114, + 0.10776124149560928, + 0.04874452203512192, + 0.16360527276992798, + 0.011090457439422607, + 0.02928309701383114 + ], + "credit_map": [ + 0.010362566448748112, + 0.015127964317798615, + 0.01478167437016964, + 0.011861938051879406, + 0.01697186380624771, + 0.010080894455313683, + 0.011857914738357067, + 0.012960436753928661, + 0.0023585890885442495, + 0.002594568068161607, + 0.00783961359411478, + 0.0019200381357222795, + 0.002255209954455495, + 0.0023436828050762415, + 0.003441974986344576, + 0.006503187119960785, + 0.002336443867534399, + 0.004181195981800556, + 0.005782689433544874, + 0.002760097151622176, + 0.004859124310314655, + 0.0030248891562223434, + 0.0030142772011458874, + 0.004235450644046068, + 0.017276082187891006, + 0.018689725548028946, + 0.16244812309741974, + 0.011172570288181305, + 0.01871398091316223, + 0.02374821901321411, + 0.014947064220905304, + 0.15339395403862, + 0.004595210310071707, + 0.0028879314195364714, + 0.017519483342766762, + 0.0017668867949396372, + 0.0024099654983729124, + 0.00502075906842947, + 0.002694708527997136, + 0.018774408847093582, + 0.005882460158318281, + 0.00640275701880455, + 0.03985636681318283, + 0.004813628736883402, + 0.007204766385257244, + 0.009635191410779953, + 0.0041199675761163235, + 0.029846101999282837, + 0.004356893245130777, + 0.005099068861454725, + 0.004957230761647224, + 0.005058078095316887, + 0.010392291471362114, + 0.0055417693220078945, + 0.005425977520644665, + 0.007913215085864067, + 0.01927720010280609, + 0.02810499630868435, + 0.02417834661900997, + 0.011709382757544518, + 0.04094971343874931, + 0.01533590629696846, + 0.012149411253631115, + 0.011900312267243862, + 0.0012699811486527324, + 0.0006016136612743139, + 0.003331599058583379, + 0.0006546899094246328, + 0.0007089503924362361, + 0.001876987749710679, + 0.0004896271857433021, + 0.0021570080425590277, + 0.00856045912951231, + 0.003428178606554866, + 0.003606105223298073, + 0.002778218360617757, + 0.003486114786937833, + 0.003477230668067932, + 0.0015867751790210605, + 0.002360016107559204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.08558270335197449 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982146292924881 + ], + "opposite_logprob_tokens": [ + -3.5298213958740234 + ], + "logprob_reward": -0.02982146292924881, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982146292924881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09070473164319992, + 0.022870458662509918, + 0.05466945469379425, + 0.6238429546356201, + 0.009392593055963516, + 0.03058783710002899, + 0.015425379388034344, + 0.1330711543560028, + 0.01112509984523058, + 0.008310317993164062 + ], + "credit_map": [ + 0.008401851169764996, + 0.010054612532258034, + 0.016139596700668335, + 0.009362990036606789, + 0.012679432518780231, + 0.00854411069303751, + 0.007121119182556868, + 0.018401026725769043, + 0.0020076422952115536, + 0.00402757478877902, + 0.0022623410914093256, + 0.0023216092959046364, + 0.0036156727001070976, + 0.002309798263013363, + 0.0034155843313783407, + 0.0029102384578436613, + 0.004517194349318743, + 0.005393383093178272, + 0.012961247004568577, + 0.0039104861207306385, + 0.006770967040210962, + 0.008861858397722244, + 0.005773703567683697, + 0.006480621639639139, + 0.02306170016527176, + 0.007769917603582144, + 0.26782941818237305, + 0.015599104575812817, + 0.010469000786542892, + 0.04281742498278618, + 0.019955171272158623, + 0.23634123802185059, + 0.0010845307260751724, + 0.0008807911071926355, + 0.0014164403546601534, + 0.0006616796599701047, + 0.001308010658249259, + 0.0015960122691467404, + 0.0013827107613906264, + 0.0010624181013554335, + 0.0034934389404952526, + 0.002207380486652255, + 0.009487872011959553, + 0.001762565691024065, + 0.0028780107386410236, + 0.0031525385566055775, + 0.0011950030457228422, + 0.006411031819880009, + 0.0017447412246838212, + 0.001466357964091003, + 0.0025071054697036743, + 0.0014904369600117207, + 0.0019201880786567926, + 0.0017383320955559611, + 0.0012771161273121834, + 0.0032811034470796585, + 0.021528998389840126, + 0.021748322993516922, + 0.015075969509780407, + 0.00999810453504324, + 0.02885282225906849, + 0.014409820549190044, + 0.011027902364730835, + 0.010429230518639088, + 0.0012039016000926495, + 0.0006059861625544727, + 0.0034750718623399734, + 0.0005535341333597898, + 0.0008976289536803961, + 0.0015948100481182337, + 0.0005179151194170117, + 0.0022762527223676443, + 0.0011169200297445059, + 0.000945711275562644, + 0.001454756362363696, + 0.0007756512495689094, + 0.0011586896143853664, + 0.0009111855761148036, + 0.0006903400644659996, + 0.0012570646358653903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.02982146292924881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742346566170454 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.006742346566170454, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742346566170454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07837294787168503, + 0.04434632137417793, + 0.040575698018074036, + 0.6540109515190125, + 0.047538287937641144, + 0.014219708740711212, + 0.03739193081855774, + 0.05605141818523407, + 0.01828278973698616, + 0.009209971874952316 + ], + "credit_map": [ + 0.0069160605780780315, + 0.013642956502735615, + 0.00953696295619011, + 0.00885588489472866, + 0.011194788850843906, + 0.009248415008187294, + 0.009428827092051506, + 0.009549051523208618, + 0.0031719987746328115, + 0.006721192970871925, + 0.007211725693196058, + 0.0045552970841526985, + 0.006239979062229395, + 0.00443601980805397, + 0.007242097519338131, + 0.004768010228872299, + 0.0035370795521885157, + 0.0032179877161979675, + 0.011062998324632645, + 0.002460182411596179, + 0.004709169734269381, + 0.005547548644244671, + 0.0029980360995978117, + 0.007042697165161371, + 0.025132253766059875, + 0.0053834132850170135, + 0.28418171405792236, + 0.015209128148853779, + 0.008443357422947884, + 0.04618535563349724, + 0.02024507150053978, + 0.24923063814640045, + 0.003401316935196519, + 0.002744590165093541, + 0.013230690732598305, + 0.0024355431087315083, + 0.0026061679236590862, + 0.004528904799371958, + 0.004806847311556339, + 0.013784226961433887, + 0.0016854240093380213, + 0.0024989256635308266, + 0.002175971632823348, + 0.0009116976871155202, + 0.0028107387479394674, + 0.0018317059148102999, + 0.0011254908749833703, + 0.0011797542683780193, + 0.0034479540772736073, + 0.0060715158469974995, + 0.0059787617065012455, + 0.002966273808851838, + 0.00925303716212511, + 0.0033841358963400126, + 0.002865248592570424, + 0.0034250037278980017, + 0.01067502610385418, + 0.007758655585348606, + 0.006080285180360079, + 0.004478437826037407, + 0.009888870641589165, + 0.00669608311727643, + 0.0049598035402596, + 0.005514258053153753, + 0.001917281886562705, + 0.000871760246809572, + 0.00587807921692729, + 0.0007623402052558959, + 0.001430931151844561, + 0.0028401713352650404, + 0.0007376974681392312, + 0.0038445270620286465, + 0.0010188259184360504, + 0.0011369718704372644, + 0.0011924629798159003, + 0.0009740397799760103, + 0.0017047355649992824, + 0.0009546068613417447, + 0.0009324325947090983, + 0.0012958967126905918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.006742346566170454 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186016008257866 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186016008257866, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186016008257866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062483713030815125, + 0.03391614556312561, + 0.022846940904855728, + 0.6946957111358643, + 0.020765138790011406, + 0.06857167929410934, + 0.036608804017305374, + 0.031483788043260574, + 0.013066524639725685, + 0.015561556443572044 + ], + "credit_map": [ + 0.008822700008749962, + 0.007842780090868473, + 0.007147195283323526, + 0.00666464027017355, + 0.008103900589048862, + 0.009015502408146858, + 0.007677836809307337, + 0.007209164090454578, + 0.00459658307954669, + 0.0035754407290369272, + 0.006299260072410107, + 0.0032554827630519867, + 0.0025459264870733023, + 0.003306567668914795, + 0.004090824164450169, + 0.006246062461286783, + 0.002302211942151189, + 0.002891691867262125, + 0.004538039211183786, + 0.001906630932353437, + 0.0024299027863889933, + 0.002132967347279191, + 0.0016535627655684948, + 0.0049919369630515575, + 0.025973305106163025, + 0.006524805910885334, + 0.30143478512763977, + 0.016326535493135452, + 0.009680907242000103, + 0.04888593405485153, + 0.022211959585547447, + 0.26365751028060913, + 0.002008760580793023, + 0.003008366096764803, + 0.003211941570043564, + 0.0011950558982789516, + 0.0023111540358513594, + 0.0022860486060380936, + 0.0021921752486377954, + 0.004551640246063471, + 0.0064477878622710705, + 0.003360586706548929, + 0.025402933359146118, + 0.0021328828297555447, + 0.003642695490270853, + 0.00583775294944644, + 0.0019350942457094789, + 0.019811948761343956, + 0.004059745464473963, + 0.004835738334804773, + 0.005993842612951994, + 0.0029342027846723795, + 0.0072311172261834145, + 0.004043444525450468, + 0.0028659659437835217, + 0.004644752014428377, + 0.0034747598692774773, + 0.004657173529267311, + 0.0035508207511156797, + 0.003075825748965144, + 0.006862078793346882, + 0.0035228957422077656, + 0.0028597416821867228, + 0.003480494488030672, + 0.001468811766244471, + 0.00115316326264292, + 0.0031806607730686665, + 0.0007555597694590688, + 0.0013650527689605951, + 0.0035145527217537165, + 0.0006373062497004867, + 0.0009914186084643006, + 0.0018154297722503543, + 0.001333613763563335, + 0.0026801086496561766, + 0.0011196320410817862, + 0.00219825841486454, + 0.0018241845536977053, + 0.0011135709937661886, + 0.003476759186014533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.018186016008257866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790222644805908 + ], + "opposite_logprob_tokens": [ + -3.529790163040161 + ], + "logprob_reward": -0.029790222644805908, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790222644805908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22646945714950562, + 0.057022273540496826, + 0.06085244566202164, + 0.2349342703819275, + 0.043745845556259155, + 0.06585901230573654, + 0.04430082067847252, + 0.1938582807779312, + 0.035861413925886154, + 0.03709619119763374 + ], + "credit_map": [ + 0.02645207941532135, + 0.03579537570476532, + 0.0307137593626976, + 0.02807428129017353, + 0.032254986464977264, + 0.02246163785457611, + 0.024572310969233513, + 0.026145050302147865, + 0.010512829758226871, + 0.006766550708562136, + 0.007095092907547951, + 0.005727382842451334, + 0.00811396911740303, + 0.006549302954226732, + 0.006149533670395613, + 0.006107611116021872, + 0.006684652995318174, + 0.0072535863146185875, + 0.011207381263375282, + 0.004694292787462473, + 0.010918878950178623, + 0.009435081854462624, + 0.004813385661691427, + 0.005845182575285435, + 0.00982157327234745, + 0.00855121947824955, + 0.09569139033555984, + 0.006563262548297644, + 0.007468389347195625, + 0.02127625234425068, + 0.007328846026211977, + 0.07823335379362106, + 0.006163408979773521, + 0.003159940242767334, + 0.007155042607337236, + 0.002729739062488079, + 0.0029521684627979994, + 0.006606758572161198, + 0.003636270062997937, + 0.011342517100274563, + 0.0059151300229132175, + 0.004579371307045221, + 0.020530417561531067, + 0.003143724286928773, + 0.005852839909493923, + 0.006253836210817099, + 0.004331427626311779, + 0.015252264216542244, + 0.008870215155184269, + 0.005113425198942423, + 0.00444081611931324, + 0.004968677181750536, + 0.006272866390645504, + 0.005707889795303345, + 0.0035878147464245558, + 0.005339117720723152, + 0.028888968750834465, + 0.02993195876479149, + 0.02575334534049034, + 0.013914154842495918, + 0.044324878603219986, + 0.01950753666460514, + 0.016009477898478508, + 0.015527955256402493, + 0.008482682518661022, + 0.0029469169676303864, + 0.0048486110754311085, + 0.002726102015003562, + 0.003715098137035966, + 0.0043632155284285545, + 0.002286657691001892, + 0.006492129527032375, + 0.0046247621066868305, + 0.004693260416388512, + 0.003978378605097532, + 0.0057654608972370625, + 0.004071294795721769, + 0.004645481240004301, + 0.005133919883519411, + 0.004183629993349314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.029790222644805908 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a horizontal angle, depicts a world constructed in a refined pixel art style, with sharp, blocky edges. At the center of the frame is an elderly man with a graying beard and hair, his sparse white hair pulled back in a small bun. His face is lined with deep wrinkles, created by pixelated blocks, and his eyes are focused and serene. He wears a traditional dark gray coarse Chinese robe, its folds rendered with varying shades of gray. He stands in a quaint Chinese courtyard, his ground paved with bluestone slabs, interspersed between the cracks of the slabs with a profusion of golden and crimson pixelated fallen leaves. The elderly man is slightly bent over, his hands tightly gripping a long-handled bamboo broom—the handle rendered as a long string of brown pixels, the end of the broom a cluster of scattered yellow pixels. He methodically and unhurriedly swept the fallen leaves to one side. Each sweep was slow and rhythmic. The tip of the bamboo broom gently brushed the ground, sending pixelated leaves fluttering briefly in the air before gathering into a small pile. A soft, diffuse light evenly bathed the courtyard, devoid of harsh highlights, casting only soft, jagged, blocky shadows around the figures and the corners of the walls. The scene, dominated by earth tones and grays, was punctuated by the vibrant, warm colors of the fallen leaves, creating a tranquil and focused atmosphere that blended the nostalgia of retro gaming with the essence of Eastern Zen.", + "seed": 23364, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_7125_seed_23364_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025431547313928604, + "vlm_logprob": -0.025431547313928604, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.025431547313928604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07666334509849548 + ], + "opposite_logprob_tokens": [ + -2.7016632556915283 + ], + "logprob_reward": -0.07666334509849548, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07666334509849548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3937545716762543, + 0.1414170116186142, + 0.03931887075304985, + 0.039872948080301285, + 0.048828866332769394, + 0.06038486957550049, + 0.10174360871315002, + 0.05810016393661499, + 0.04147611930966377, + 0.07510296255350113 + ], + "credit_map": [ + 0.04841560497879982, + 0.06879980117082596, + 0.028777770698070526, + 0.047681961208581924, + 0.05390537157654762, + 0.03973117843270302, + 0.07336670905351639, + 0.03307616338133812, + 0.020010152831673622, + 0.012916253879666328, + 0.009005282074213028, + 0.015266324393451214, + 0.022594250738620758, + 0.027916505932807922, + 0.018671536818146706, + 0.015036720782518387, + 0.004600157495588064, + 0.004870079457759857, + 0.0038811436388641596, + 0.004096432588994503, + 0.005923624616116285, + 0.005195844452828169, + 0.00685883779078722, + 0.003892746986821294, + 0.004471536260098219, + 0.0032537486404180527, + 0.002401885110884905, + 0.003610413521528244, + 0.006847776938229799, + 0.009050958789885044, + 0.005973664578050375, + 0.004262963309884071, + 0.005588561296463013, + 0.009992992505431175, + 0.0032792112324386835, + 0.007094154134392738, + 0.005778813268989325, + 0.005690793506801128, + 0.006687791086733341, + 0.00471655186265707, + 0.007359265349805355, + 0.011283599771559238, + 0.0038821741472929716, + 0.007728132884949446, + 0.007469826843589544, + 0.007751223165541887, + 0.009255051612854004, + 0.005655595567077398, + 0.01152525469660759, + 0.00849706307053566, + 0.0065204850398004055, + 0.006727506872266531, + 0.016667913645505905, + 0.027656447142362595, + 0.014796220697462559, + 0.009352720342576504, + 0.006599807180464268, + 0.008047825656831264, + 0.004088655114173889, + 0.006141240708529949, + 0.006946137174963951, + 0.00628542760387063, + 0.012926037423312664, + 0.007065027952194214, + 0.004114342853426933, + 0.007578573189675808, + 0.00312733743339777, + 0.005840410478413105, + 0.005602213088423014, + 0.0049904389306902885, + 0.006568692624568939, + 0.003654112806543708, + 0.009171607904136181, + 0.007977631874382496, + 0.005531200673431158, + 0.0061502098105847836, + 0.009313891641795635, + 0.012740946374833584, + 0.017992107197642326, + 0.00622536800801754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.07666334509849548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592467859387398 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592467859387398, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592467859387398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6079645752906799, + 0.1498132050037384, + 0.04445890709757805, + 0.021328609436750412, + 0.01535603404045105, + 0.014871330931782722, + 0.029926978051662445, + 0.022852346301078796, + 0.024233363568782806, + 0.06919466704130173 + ], + "credit_map": [ + 0.08203420788049698, + 0.07643384486436844, + 0.050122376531362534, + 0.04634464159607887, + 0.10339253395795822, + 0.13219185173511505, + 0.062102556228637695, + 0.05534256622195244, + 0.016028694808483124, + 0.009950626641511917, + 0.010896059684455395, + 0.009917253628373146, + 0.027495477348566055, + 0.03669301047921181, + 0.015196560882031918, + 0.023635532706975937, + 0.005812777206301689, + 0.005327627994120121, + 0.0034317101817578077, + 0.004325139336287975, + 0.006789525039494038, + 0.007115395739674568, + 0.007543203420937061, + 0.0041135274805128574, + 0.0024972534738481045, + 0.002105234656482935, + 0.0013722809962928295, + 0.0019731938373297453, + 0.0031485864892601967, + 0.0044082761742174625, + 0.00421945471316576, + 0.0016043278155848384, + 0.0019721076823771, + 0.0015134001150727272, + 0.0011589028872549534, + 0.0011465713614597917, + 0.0018612457206472754, + 0.0027699072379618883, + 0.0018062322633340955, + 0.003127665026113391, + 0.0020316538866609335, + 0.0019543569069355726, + 0.0009267033892683685, + 0.0019277858082205057, + 0.0021142621990293264, + 0.002906547160819173, + 0.0017788824625313282, + 0.0012311384780332446, + 0.003737869905307889, + 0.004056144971400499, + 0.0016911897109821439, + 0.002998007694259286, + 0.00495710177347064, + 0.00672624446451664, + 0.003514074021950364, + 0.0022463458590209484, + 0.0026268260553479195, + 0.004105385858565569, + 0.0023538742680102587, + 0.0026429477147758007, + 0.0021859980188310146, + 0.002449544845148921, + 0.003233359893783927, + 0.0032544077839702368, + 0.002784645650535822, + 0.0031718553509563208, + 0.0016198825323954225, + 0.0031042718328535557, + 0.0032072460744529963, + 0.0037765013985335827, + 0.0030743135139346123, + 0.00349464756436646, + 0.009255403652787209, + 0.006174031645059586, + 0.005172783508896828, + 0.006092596333473921, + 0.007589123677462339, + 0.01253232080489397, + 0.017033245414495468, + 0.005345162004232407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.020592467859387398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099062643945217 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004099062643945217, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099062643945217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2687707543373108, + 0.2492319941520691, + 0.06706484407186508, + 0.08886604011058807, + 0.060122910887002945, + 0.03412312641739845, + 0.051151569932699203, + 0.03608804568648338, + 0.0546596460044384, + 0.08992110937833786 + ], + "credit_map": [ + 0.03330956771969795, + 0.032673463225364685, + 0.028780965134501457, + 0.02543015591800213, + 0.0324004665017128, + 0.036838799715042114, + 0.042907845228910446, + 0.0364295057952404, + 0.02470223419368267, + 0.017801519483327866, + 0.015276134945452213, + 0.01605803146958351, + 0.046553488820791245, + 0.06303097307682037, + 0.022815916687250137, + 0.0429936908185482, + 0.009076233953237534, + 0.005966606549918652, + 0.0051320320926606655, + 0.005121856462210417, + 0.009190521202981472, + 0.014303312636911869, + 0.011637848801910877, + 0.006636434234678745, + 0.010737608186900616, + 0.011813772842288017, + 0.005763666704297066, + 0.009831934235990047, + 0.012386814691126347, + 0.01717638038098812, + 0.013491679914295673, + 0.007664186414331198, + 0.0058520641177892685, + 0.007745669223368168, + 0.004210040904581547, + 0.005946206394582987, + 0.010820047929883003, + 0.0074801272712647915, + 0.008386681787669659, + 0.009682074189186096, + 0.0045152450911700726, + 0.004587994422763586, + 0.003064660122618079, + 0.004770382773131132, + 0.003946136683225632, + 0.003978515509516001, + 0.005534876137971878, + 0.003725314512848854, + 0.006609050557017326, + 0.009897532872855663, + 0.0031539525371044874, + 0.006225196644663811, + 0.006800198927521706, + 0.008914096280932426, + 0.0059527019038796425, + 0.0035988413728773594, + 0.0047971331514418125, + 0.0059869312681257725, + 0.0029145849402993917, + 0.0037083595525473356, + 0.0038011271972209215, + 0.004984733648598194, + 0.0058168526738882065, + 0.004078318830579519, + 0.005870571359992027, + 0.00914006493985653, + 0.0037765351589769125, + 0.006990815047174692, + 0.00824535358697176, + 0.006798573769629002, + 0.0074645644053816795, + 0.006373168900609016, + 0.01339237205684185, + 0.011662696488201618, + 0.007175698410719633, + 0.006985814776271582, + 0.010834434069693089, + 0.016854701563715935, + 0.01659899204969406, + 0.006416400894522667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.004099062643945217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624453864991665 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624453864991665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624453864991665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2039225846529007, + 0.4657166302204132, + 0.0678233653306961, + 0.05366393178701401, + 0.046930309385061264, + 0.025097209960222244, + 0.020145731046795845, + 0.03228558599948883, + 0.03103093057870865, + 0.0533837266266346 + ], + "credit_map": [ + 0.026124561205506325, + 0.020445996895432472, + 0.019273122772574425, + 0.021993786096572876, + 0.02459738403558731, + 0.0358671173453331, + 0.02801969088613987, + 0.027600929141044617, + 0.043287523090839386, + 0.03228463977575302, + 0.027070846408605576, + 0.02875353954732418, + 0.08733066916465759, + 0.1252465397119522, + 0.039241522550582886, + 0.0825013667345047, + 0.007595399860292673, + 0.010721445083618164, + 0.005386986769735813, + 0.006610780023038387, + 0.009001058526337147, + 0.008520388044416904, + 0.009885315783321857, + 0.010101993568241596, + 0.0049852244555950165, + 0.009318484924733639, + 0.004292594734579325, + 0.005509571172297001, + 0.006979098543524742, + 0.007018421310931444, + 0.009784931316971779, + 0.00577560905367136, + 0.005502020008862019, + 0.00680176168680191, + 0.0041424850933253765, + 0.006231325212866068, + 0.0069666896015405655, + 0.005831404123455286, + 0.00610812334343791, + 0.005346498917788267, + 0.0028207467403262854, + 0.0023015395272523165, + 0.002690957859158516, + 0.0036069161724299192, + 0.0032163315918296576, + 0.0025710132904350758, + 0.004742116667330265, + 0.0031475869473069906, + 0.0019966226536780596, + 0.0020460595842450857, + 0.0019371508387848735, + 0.0018711291486397386, + 0.0033804383128881454, + 0.0028690402396023273, + 0.003042858559638262, + 0.0030024335719645023, + 0.0037200746592134237, + 0.0044738552533090115, + 0.002141433535143733, + 0.0035397280007600784, + 0.00451891403645277, + 0.003317898605018854, + 0.0072215888649225235, + 0.0033520960714668036, + 0.0032463253010064363, + 0.0031834938563406467, + 0.0023245096672326326, + 0.002035709796473384, + 0.005559803452342749, + 0.007328887470066547, + 0.00430653290823102, + 0.0030456697568297386, + 0.007939732633531094, + 0.005272856447845697, + 0.0039023756980895996, + 0.003806674387305975, + 0.0062195616774261, + 0.010834727436304092, + 0.011465451680123806, + 0.003942349459975958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007624453864991665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178407102823257 + ], + "opposite_logprob_tokens": [ + -4.018178462982178 + ], + "logprob_reward": -0.018178407102823257, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178407102823257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40683504939079285, + 0.10746124386787415, + 0.042831212282180786, + 0.16756035387516022, + 0.16885031759738922, + 0.01623893715441227, + 0.018954332917928696, + 0.023594709113240242, + 0.017909614369273186, + 0.029764242470264435 + ], + "credit_map": [ + 0.06067971885204315, + 0.033844269812107086, + 0.025959094986319542, + 0.026826562359929085, + 0.0768929123878479, + 0.1110440343618393, + 0.04649919271469116, + 0.025089256465435028, + 0.01278297696262598, + 0.010887193493545055, + 0.007925468496978283, + 0.008034803904592991, + 0.016950305551290512, + 0.02432095631957054, + 0.013205639086663723, + 0.013353895395994186, + 0.0064783114939928055, + 0.005176535341888666, + 0.0036651778500527143, + 0.0040547652170062065, + 0.005094262305647135, + 0.01036946102976799, + 0.003963134251534939, + 0.004029565490782261, + 0.017767634242773056, + 0.023416951298713684, + 0.011140147224068642, + 0.01816646382212639, + 0.02254415862262249, + 0.02873157523572445, + 0.026476753875613213, + 0.019316671416163445, + 0.01591578871011734, + 0.03087262623012066, + 0.01027572900056839, + 0.02000734582543373, + 0.02539890632033348, + 0.024074574932456017, + 0.02143053337931633, + 0.02087480202317238, + 0.0018374169012531638, + 0.0022074843291193247, + 0.001499223755672574, + 0.0022487249225378036, + 0.002155875088647008, + 0.002015021163970232, + 0.0026536318473517895, + 0.0016215596115216613, + 0.002312550088390708, + 0.0031194211915135384, + 0.0015100346645340323, + 0.001931191305629909, + 0.0031547776889055967, + 0.00315482122823596, + 0.0022746636532247066, + 0.0014968744944781065, + 0.003228025045245886, + 0.0022653338965028524, + 0.0015216120518743992, + 0.0013806564966216683, + 0.00415625236928463, + 0.006632914301007986, + 0.0026983085554093122, + 0.0017116060480475426, + 0.0016492960276082158, + 0.002579656895250082, + 0.0014434315962716937, + 0.0014959030086174607, + 0.003431799355894327, + 0.003170182229951024, + 0.0022567990235984325, + 0.0018825458828359842, + 0.003151330165565014, + 0.002534622559323907, + 0.002101818099617958, + 0.0027892536018043756, + 0.004097993951290846, + 0.0036308655980974436, + 0.0086832195520401, + 0.002775139408186078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.018178407102823257 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a horizontal angle, depicts a world constructed in a refined pixel art style, with sharp, blocky edges. At the center of the frame is an elderly man with a graying beard and hair, his sparse white hair pulled back in a small bun. His face is lined with deep wrinkles, created by pixelated blocks, and his eyes are focused and serene. He wears a traditional dark gray coarse Chinese robe, its folds rendered with varying shades of gray. He stands in a quaint Chinese courtyard, his ground paved with bluestone slabs, interspersed between the cracks of the slabs with a profusion of golden and crimson pixelated fallen leaves. The elderly man is slightly bent over, his hands tightly gripping a long-handled bamboo broom—the handle rendered as a long string of brown pixels, the end of the broom a cluster of scattered yellow pixels. He methodically and unhurriedly swept the fallen leaves to one side. Each sweep was slow and rhythmic. The tip of the bamboo broom gently brushed the ground, sending pixelated leaves fluttering briefly in the air before gathering into a small pile. A soft, diffuse light evenly bathed the courtyard, devoid of harsh highlights, casting only soft, jagged, blocky shadows around the figures and the corners of the walls. The scene, dominated by earth tones and grays, was punctuated by the vibrant, warm colors of the fallen leaves, creating a tranquil and focused atmosphere that blended the nostalgia of retro gaming with the essence of Eastern Zen.", + "seed": 23365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_13/index_7125_seed_23365_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04669993352144956, + "vlm_logprob": -0.04669993352144956, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.04669993352144956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18796195089817047 + ], + "opposite_logprob_tokens": [ + -1.8129619359970093 + ], + "logprob_reward": -0.18796195089817047, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18796195089817047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7743499279022217, + 0.05741467326879501, + 0.01926060952246189, + 0.017969563603401184, + 0.0203305222094059, + 0.015994245186448097, + 0.0131369112059474, + 0.03733742609620094, + 0.019450465217232704, + 0.02475561387836933 + ], + "credit_map": [ + 0.12077678740024567, + 0.14286161959171295, + 0.10162370651960373, + 0.08696727454662323, + 0.09147997945547104, + 0.07096660137176514, + 0.06663194298744202, + 0.0930420383810997, + 0.007031833287328482, + 0.010666735470294952, + 0.0069955927319824696, + 0.007090417202562094, + 0.007755108177661896, + 0.005310960114002228, + 0.00608970457687974, + 0.006474318448454142, + 0.0023737403098493814, + 0.00300692324526608, + 0.0026180418208241463, + 0.0025990833528339863, + 0.001999430824071169, + 0.0017991593340411782, + 0.0026608342304825783, + 0.00220339628867805, + 0.0023293665144592524, + 0.0024760644882917404, + 0.002846338087692857, + 0.002153802430257201, + 0.002380043501034379, + 0.0018532309914007783, + 0.0019861902110278606, + 0.001944524934515357, + 0.0025897191371768713, + 0.002850366523489356, + 0.002663274295628071, + 0.0025347298942506313, + 0.002119675511494279, + 0.002676428062841296, + 0.0017595385434105992, + 0.0031367894262075424, + 0.003990215715020895, + 0.0015085207996889949, + 0.0032743082847446203, + 0.0019021691987290978, + 0.0017587728798389435, + 0.0009491447126492858, + 0.001333414576947689, + 0.0012776997173205018, + 0.001403456786647439, + 0.001804633648134768, + 0.0017163380980491638, + 0.0015165643999353051, + 0.0019151561427861452, + 0.0012976955622434616, + 0.0014466902939602733, + 0.0020363761577755213, + 0.01506700087338686, + 0.00273825298063457, + 0.010332117788493633, + 0.0023265809286385775, + 0.0019194930791854858, + 0.0013988768914714456, + 0.002186625264585018, + 0.0013684744480997324, + 0.003549658227711916, + 0.0024497725535184145, + 0.002787412377074361, + 0.0026008347049355507, + 0.0031365060713142157, + 0.0013937964104115963, + 0.0018931126687675714, + 0.0016393704572692513, + 0.004680439364165068, + 0.0029475665651261806, + 0.0029423863161355257, + 0.00332614965736866, + 0.0033854078501462936, + 0.0019505479140207171, + 0.0032651915680617094, + 0.0022579210344702005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.18796195089817047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088770814239979 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088770814239979, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088770814239979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45386725664138794, + 0.13446767628192902, + 0.05882498621940613, + 0.05138583481311798, + 0.033166415989398956, + 0.032483525574207306, + 0.02918996475636959, + 0.09498144686222076, + 0.04067065939307213, + 0.07096228748559952 + ], + "credit_map": [ + 0.04579370096325874, + 0.051400166004896164, + 0.05086878687143326, + 0.06829576194286346, + 0.08794661611318588, + 0.04300229251384735, + 0.051162078976631165, + 0.05539782717823982, + 0.022183848544955254, + 0.01703728176653385, + 0.014988811686635017, + 0.018954724073410034, + 0.02588845044374466, + 0.00844082236289978, + 0.014294401742517948, + 0.012679332867264748, + 0.010507303290069103, + 0.0059630307368934155, + 0.009130077436566353, + 0.00788620300590992, + 0.006841413676738739, + 0.004939349368214607, + 0.0068778968416154385, + 0.006679703947156668, + 0.012279004789888859, + 0.00401666946709156, + 0.007435477804392576, + 0.006967201363295317, + 0.006547477096319199, + 0.0036164335906505585, + 0.005226528272032738, + 0.005297043826431036, + 0.010966816917061806, + 0.0024670185521245003, + 0.0051069059409201145, + 0.0032096016220748425, + 0.0031227809377014637, + 0.0019866719376295805, + 0.0030541110318154097, + 0.0032525062561035156, + 0.004120988305658102, + 0.004194386303424835, + 0.004584042355418205, + 0.005549176596105099, + 0.004801275674253702, + 0.0024776554200798273, + 0.003514582524076104, + 0.003241416299715638, + 0.008367734961211681, + 0.0030741391237825155, + 0.004129934590309858, + 0.002925693290308118, + 0.003936429042369127, + 0.0016668459866195917, + 0.0026650098152458668, + 0.002424175152555108, + 0.04008325934410095, + 0.008894344791769981, + 0.020762626081705093, + 0.0046953181736171246, + 0.004875896032899618, + 0.004018634092062712, + 0.003407321171835065, + 0.008244034834206104, + 0.012569365091621876, + 0.004010540433228016, + 0.005010325461626053, + 0.004272865131497383, + 0.0033398689702153206, + 0.0029245219193398952, + 0.0032786380033940077, + 0.005264533217996359, + 0.016908006742596626, + 0.006266534794121981, + 0.016823288053274155, + 0.008702403865754604, + 0.005501465871930122, + 0.0038674601819366217, + 0.0070822276175022125, + 0.005810893140733242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011088770814239979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099300131201744 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099300131201744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099300131201744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5176980495452881, + 0.06975020468235016, + 0.03685406222939491, + 0.03493720293045044, + 0.03139680251479149, + 0.05705761909484863, + 0.020942073315382004, + 0.12265875935554504, + 0.034523773938417435, + 0.07418143004179001 + ], + "credit_map": [ + 0.07441425323486328, + 0.0947730764746666, + 0.0669851079583168, + 0.06588342785835266, + 0.06304384768009186, + 0.0499274842441082, + 0.047285471111536026, + 0.055385369807481766, + 0.015757881104946136, + 0.005919512361288071, + 0.011630908586084843, + 0.010609408840537071, + 0.009080756455659866, + 0.004583725705742836, + 0.007029799744486809, + 0.005138217471539974, + 0.005548995919525623, + 0.004856437910348177, + 0.0042949216440320015, + 0.004593321587890387, + 0.004556260537356138, + 0.003526808228343725, + 0.005866031628102064, + 0.003611284773796797, + 0.006928995717316866, + 0.004649781621992588, + 0.004049322102218866, + 0.004793776664882898, + 0.0047187465243041515, + 0.003685151692479849, + 0.00279253046028316, + 0.0033188967499881983, + 0.005175435449928045, + 0.004148039501160383, + 0.0038022655062377453, + 0.0034263283014297485, + 0.0038294869009405375, + 0.003472589887678623, + 0.004386048298329115, + 0.003156607272103429, + 0.014861227944493294, + 0.004880449268966913, + 0.013699003495275974, + 0.005748976953327656, + 0.0043121566995978355, + 0.004751015920192003, + 0.0041243587620556355, + 0.004680430516600609, + 0.003596547059714794, + 0.002772021107375622, + 0.0030008521862328053, + 0.0026118468958884478, + 0.00209037191234529, + 0.002077023033052683, + 0.0028348411433398724, + 0.001958570210263133, + 0.05107247829437256, + 0.010416780598461628, + 0.02012469433248043, + 0.005844488739967346, + 0.0067177810706198215, + 0.006624104455113411, + 0.004464866127818823, + 0.017393557354807854, + 0.007668006233870983, + 0.0034152287989854813, + 0.0043219453655183315, + 0.0050890035927295685, + 0.0034158334601670504, + 0.002722162986174226, + 0.00343445991165936, + 0.004457134287804365, + 0.020410336554050446, + 0.004964533727616072, + 0.022429052740335464, + 0.007310440298169851, + 0.0046584224328398705, + 0.0030745395924896, + 0.0043074944987893105, + 0.00702660670503974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.004099300131201744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977633148431778 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.00977633148431778, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977633148431778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42921677231788635, + 0.12019447237253189, + 0.05408402159810066, + 0.051811788231134415, + 0.04209364578127861, + 0.041945479810237885, + 0.029284028336405754, + 0.07252596318721771, + 0.0752796158194542, + 0.08356417715549469 + ], + "credit_map": [ + 0.05069581791758537, + 0.06470608711242676, + 0.05763053521513939, + 0.06483927369117737, + 0.052349381148815155, + 0.04764265939593315, + 0.04998794570565224, + 0.04136507213115692, + 0.02230895683169365, + 0.011918565258383751, + 0.019846491515636444, + 0.01780811883509159, + 0.013598295859992504, + 0.009844020009040833, + 0.013653848320245743, + 0.0112161785364151, + 0.009830038994550705, + 0.005944305099546909, + 0.007757962681353092, + 0.008048255927860737, + 0.006258588284254074, + 0.004792361985892057, + 0.005896177142858505, + 0.005556333344429731, + 0.010571838356554508, + 0.007425801828503609, + 0.005662212148308754, + 0.007233536336570978, + 0.0065446943044662476, + 0.0051277694292366505, + 0.003896421752870083, + 0.005349516402930021, + 0.009427507407963276, + 0.0050598494708538055, + 0.005022192839533091, + 0.004626001697033644, + 0.005397903732955456, + 0.003805600805208087, + 0.004353160038590431, + 0.004401428624987602, + 0.006192890461534262, + 0.003721072571352124, + 0.009899280965328217, + 0.004464650992304087, + 0.004249860066920519, + 0.004163164179772139, + 0.003282021265476942, + 0.005972541403025389, + 0.005261654499918222, + 0.004075375385582447, + 0.003620976349338889, + 0.0038876731414347887, + 0.0037523808423429728, + 0.003153805620968342, + 0.003047306789085269, + 0.002484857104718685, + 0.02904420718550682, + 0.005827052053064108, + 0.012858131900429726, + 0.004777347669005394, + 0.004509165417402983, + 0.004163363017141819, + 0.0030731474980711937, + 0.00827354658395052, + 0.020410677418112755, + 0.007513900753110647, + 0.009896689094603062, + 0.008468013256788254, + 0.008540928363800049, + 0.004470225889235735, + 0.006119170226156712, + 0.009860006161034107, + 0.024835724383592606, + 0.007535753306001425, + 0.021441416814923286, + 0.009001855738461018, + 0.006914088036864996, + 0.004106235224753618, + 0.004983160179108381, + 0.004745947662740946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00977633148431778 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a refined pixel art style with sharp, blocky edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with a graying beard, sparse white hair in a bun, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a traditional dark gray coarse Chinese robe and holding a long-handled bamboo broom with yellow bristles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Chinese courtyard with bluestone slab paving and golden and crimson fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a horizontal angle, showing the man sweeping leaves methodically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573314279317856 + ], + "opposite_logprob_tokens": [ + -3.895573377609253 + ], + "logprob_reward": -0.020573314279317856, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573314279317856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30250057578086853, + 0.18304024636745453, + 0.1031411737203598, + 0.04865095391869545, + 0.05718204006552696, + 0.09556149691343307, + 0.044147178530693054, + 0.029727358371019363, + 0.03809240460395813, + 0.09795664250850677 + ], + "credit_map": [ + 0.04828986898064613, + 0.03519165888428688, + 0.04199628159403801, + 0.053196314722299576, + 0.03729633241891861, + 0.019027790054678917, + 0.04011700302362442, + 0.02738529071211815, + 0.027951907366514206, + 0.029801158234477043, + 0.02615252137184143, + 0.02427263930439949, + 0.0248044915497303, + 0.015130226500332355, + 0.015613025985658169, + 0.01931423880159855, + 0.015397865325212479, + 0.015652703121304512, + 0.016724977642297745, + 0.015051676891744137, + 0.013453340157866478, + 0.00942364614456892, + 0.008850428275763988, + 0.008586520329117775, + 0.008048022165894508, + 0.009932405315339565, + 0.005682682618498802, + 0.004629351198673248, + 0.0058150868862867355, + 0.005468674004077911, + 0.003590882755815983, + 0.00548383966088295, + 0.009719032794237137, + 0.008696649223566055, + 0.008178066462278366, + 0.007718583103269339, + 0.005867634899914265, + 0.007279610726982355, + 0.004797128029167652, + 0.004925319459289312, + 0.02697763219475746, + 0.007064618170261383, + 0.026898056268692017, + 0.007254742551594973, + 0.006255443673580885, + 0.005815534852445126, + 0.006478557828813791, + 0.008816896006464958, + 0.005433071870356798, + 0.006288766395300627, + 0.005533003713935614, + 0.0058107320219278336, + 0.004571958445012569, + 0.007253363728523254, + 0.004356072749942541, + 0.004900200292468071, + 0.005617222748696804, + 0.0033246527891606092, + 0.0037711020559072495, + 0.0034935353323817253, + 0.0030114848632365465, + 0.00285809813067317, + 0.003260793862864375, + 0.00439046323299408, + 0.006750263273715973, + 0.004141486715525389, + 0.004503527656197548, + 0.005455563776195049, + 0.005086417309939861, + 0.002640750491991639, + 0.005819848272949457, + 0.003694538725540042, + 0.022937873378396034, + 0.009456508792936802, + 0.02806994505226612, + 0.008692474104464054, + 0.005905838683247566, + 0.006612909957766533, + 0.006740680895745754, + 0.009540391154587269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020573314279317856 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_14/index_tensor([1522, 1522])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_14/index_tensor([1522, 1522])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..61c24f2e8b5e40ad15ab16d1918f893fa3f70a46 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_14/index_tensor([1522, 1522])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, macro-style shot remains perfectly static, focusing on the complex landscape of a human face. The frame is dominated by a gleaming eye with a crystal-like pale blue iris, its intricate fibrous pattern visible in ultra-fine detail. A single, cold-toned key light from a high side angle illuminates the scene, giving the cornea a damp sheen and revealing its delicate network of capillaries. From the delicate inner corner of the eye, a viscous black liquid begins to gush, accumulating erratically before steadily and deliberately seeping down to the lower lash line. This dark liquid, with its oily, reflective surface, captures the single light source, following a slow, uninterrupted path along pale skin, navigating the fine, almost invisible fuzz and delicate texture of pores. The stark contrast between the dark liquid, pale skin, and bright eye, combined with the subject's unblinking stillness, creates a deeply unsettling and surreal atmosphere, filled with unspoken tension.", + "seed": 11722, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_14/index_1522_seed_11722_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21574844988062977, + "vlm_logprob": -0.21574844988062977, + "vlm_yes_no_margin": 3.5249998807907104, + "vlm_token_credit": -0.21574844988062977, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031714164651930332 + ], + "opposite_logprob_tokens": [ + -8.00317096710205 + ], + "logprob_reward": -0.0031714164651930332, + "yes_no_margin_reward": 7.999999523162842, + "token_credit_reward": -0.0031714164651930332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0314364992082119, + 0.059916332364082336, + 0.022162364795804024, + 0.02601049654185772, + 0.044313639402389526, + 0.22127510607242584, + 0.4016342759132385, + 0.1442740559577942, + 0.024757226929068565, + 0.02422001212835312 + ], + "credit_map": [ + 0.005486306268721819, + 0.0033918507397174835, + 0.003467809408903122, + 0.003525440115481615, + 0.0031199574004858732, + 0.003812817158177495, + 0.004550591576844454, + 0.004081726539880037, + 0.0034050389658659697, + 0.004502573981881142, + 0.0027259797789156437, + 0.0034639991354197264, + 0.0033510976936668158, + 0.02049899287521839, + 0.0043066260404884815, + 0.01766202598810196, + 0.0022004577331244946, + 0.004184851888567209, + 0.0019293185323476791, + 0.0030539077706635, + 0.003438602201640606, + 0.0024406048469245434, + 0.0026016137562692165, + 0.0023130071349442005, + 0.0025573645252734423, + 0.0031933344434946775, + 0.0018805023282766342, + 0.003527329070493579, + 0.005439103115350008, + 0.0032298865262418985, + 0.003148831892758608, + 0.0030341448727995157, + 0.005536384880542755, + 0.00486741354689002, + 0.003982217516750097, + 0.005812458693981171, + 0.0040588900446891785, + 0.0074364738538861275, + 0.005569736007601023, + 0.007050064858049154, + 0.013577759265899658, + 0.0057108704932034016, + 0.008342165499925613, + 0.006113288924098015, + 0.011341632343828678, + 0.08547328412532806, + 0.014792020432651043, + 0.0759240910410881, + 0.022567765787243843, + 0.004056508652865887, + 0.009855695068836212, + 0.004201369360089302, + 0.014189778827130795, + 0.17794226109981537, + 0.01737942546606064, + 0.15144148468971252, + 0.008665869012475014, + 0.002993126632645726, + 0.004272629506886005, + 0.002869594842195511, + 0.006371034309267998, + 0.05979885160923004, + 0.006402245257049799, + 0.05290072038769722, + 0.0029975208453834057, + 0.002407610882073641, + 0.002447624923661351, + 0.0023293853737413883, + 0.00367776514030993, + 0.002933942247182131, + 0.004282092209905386, + 0.0036812867037951946, + 0.0019332712981849909, + 0.002603298518806696, + 0.0018497988348826766, + 0.00198118994012475, + 0.003638779977336526, + 0.004125386942178011, + 0.003454839810729027, + 0.004633448552340269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0031714164651930332 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221953883767128 + ], + "opposite_logprob_tokens": [ + -4.018221855163574 + ], + "logprob_reward": -0.018221953883767128, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221953883767128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04391442611813545, + 0.050141341984272, + 0.03872650861740112, + 0.056590549647808075, + 0.02516409009695053, + 0.10383923351764679, + 0.47132381796836853, + 0.12968750298023224, + 0.05431797727942467, + 0.02629457600414753 + ], + "credit_map": [ + 0.008254244923591614, + 0.0038311798125505447, + 0.004977118223905563, + 0.004628774244338274, + 0.005019516218453646, + 0.005176803562790155, + 0.005556973163038492, + 0.006469815969467163, + 0.005705390591174364, + 0.009296176955103874, + 0.004963854793459177, + 0.006282357964664698, + 0.007530940696597099, + 0.0054051741026341915, + 0.005668033845722675, + 0.005289411172270775, + 0.003922365605831146, + 0.004863062873482704, + 0.003546939231455326, + 0.004564183298498392, + 0.0055788136087358, + 0.005349600687623024, + 0.004180488642305136, + 0.006721058394759893, + 0.007893800735473633, + 0.005494551733136177, + 0.004229194950312376, + 0.005593548063188791, + 0.010599900037050247, + 0.007635956630110741, + 0.005901769734919071, + 0.009241829626262188, + 0.0038418788462877274, + 0.0028855970595031977, + 0.0023935569915920496, + 0.004069858696311712, + 0.002433834597468376, + 0.002766244113445282, + 0.003306205850094557, + 0.003466914175078273, + 0.006222342140972614, + 0.0037754594814032316, + 0.00308715901337564, + 0.0034956224262714386, + 0.004368620924651623, + 0.04204653948545456, + 0.005753280594944954, + 0.03509020805358887, + 0.026151005178689957, + 0.008046779781579971, + 0.011295921169221401, + 0.005587167572230101, + 0.020309098064899445, + 0.20536042749881744, + 0.021943725645542145, + 0.17262966930866241, + 0.006710925605148077, + 0.0035383906215429306, + 0.004090618342161179, + 0.0022292383946478367, + 0.005232942756265402, + 0.05242551490664482, + 0.005355750676244497, + 0.05010412260890007, + 0.003419595304876566, + 0.0022001140750944614, + 0.0026116641238331795, + 0.002572215162217617, + 0.0023649840150028467, + 0.019645018503069878, + 0.00235461862757802, + 0.019149769097566605, + 0.0032747737132012844, + 0.0026007506530731916, + 0.0025961678475141525, + 0.002802471397444606, + 0.003273187903687358, + 0.0049338866956532, + 0.003217281773686409, + 0.003596055321395397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.018221953883767128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14286835491657257 + ], + "opposite_logprob_tokens": [ + -2.0178682804107666 + ], + "logprob_reward": -0.14286835491657257, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14286835491657257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022914858534932137, + 0.07183670252561569, + 0.023812899366021156, + 0.00609531719237566, + 0.06935456395149231, + 0.0540287010371685, + 0.5502806901931763, + 0.12665772438049316, + 0.04279344528913498, + 0.03222523257136345 + ], + "credit_map": [ + 0.002926592482253909, + 0.0019219964742660522, + 0.0020394541788846254, + 0.002597214188426733, + 0.0022236634977161884, + 0.0038273537065833807, + 0.002392304129898548, + 0.004986279644072056, + 0.0040603941306471825, + 0.0050236256793141365, + 0.0030528476927429438, + 0.003584526479244232, + 0.0033821361139416695, + 0.027136970311403275, + 0.005602359306067228, + 0.019993841648101807, + 0.002031539101153612, + 0.0030112890526652336, + 0.002074320102110505, + 0.0025435511488467455, + 0.0032783173955976963, + 0.003465745598077774, + 0.0025808867067098618, + 0.004827249329537153, + 0.0008797875489108264, + 0.000745581288356334, + 0.0006316955550573766, + 0.0006625144160352647, + 0.0006842939183115959, + 0.0009797612437978387, + 0.0006306502618826926, + 0.0008810327854007483, + 0.004222256131470203, + 0.004192946013063192, + 0.004062095656991005, + 0.0038874323945492506, + 0.0031525783706456423, + 0.02232925221323967, + 0.00563785620033741, + 0.021870141848921776, + 0.00479589868336916, + 0.003758443519473076, + 0.0028625563718378544, + 0.004824990872293711, + 0.0030438988469541073, + 0.01718972437083721, + 0.005016746465116739, + 0.01253644097596407, + 0.031259577721357346, + 0.005876096431165934, + 0.013397221453487873, + 0.0063567934557795525, + 0.020865991711616516, + 0.2438437044620514, + 0.023840466514229774, + 0.2048407942056656, + 0.0065384856425225735, + 0.0022038985043764114, + 0.00447582732886076, + 0.002424788661301136, + 0.005277454853057861, + 0.05121432989835739, + 0.006002245470881462, + 0.04852070286870003, + 0.0020622434094548225, + 0.0012366496957838535, + 0.002046457026153803, + 0.0021103275939822197, + 0.0017086948500946164, + 0.01583200693130493, + 0.0024676083121448755, + 0.015329455956816673, + 0.002220326103270054, + 0.002385614439845085, + 0.00197717291302979, + 0.001947283511981368, + 0.003319907933473587, + 0.008352304808795452, + 0.003281357465311885, + 0.008741266094148159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.14286835491657257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607646234333515 + ], + "opposite_logprob_tokens": [ + -4.141076564788818 + ], + "logprob_reward": -0.01607646234333515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607646234333515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02447598986327648, + 0.028838789090514183, + 0.03356793150305748, + 0.025866547599434853, + 0.12403004616498947, + 0.07089708000421524, + 0.4059409201145172, + 0.264231413602829, + 0.00983433611690998, + 0.01231691986322403 + ], + "credit_map": [ + 0.0023578605614602566, + 0.001414638478308916, + 0.0022282670252025127, + 0.001896885223686695, + 0.0017447896534577012, + 0.00446213036775589, + 0.002941062208265066, + 0.007430355995893478, + 0.0023958967067301273, + 0.003985149320214987, + 0.002506157848984003, + 0.0034970208071172237, + 0.0024971445091068745, + 0.007142587099224329, + 0.002969314344227314, + 0.003845514729619026, + 0.0030333329923450947, + 0.006318260449916124, + 0.0030154685955494642, + 0.004756465088576078, + 0.004364276770502329, + 0.003559310920536518, + 0.0035992597695440054, + 0.004921554587781429, + 0.002842327579855919, + 0.001526329549960792, + 0.002109288237988949, + 0.0019834470003843307, + 0.0014398789498955011, + 0.006446459796279669, + 0.0015548435039818287, + 0.00796397216618061, + 0.009256544522941113, + 0.007908113300800323, + 0.005956939421594143, + 0.007785205263644457, + 0.007040579803287983, + 0.040510352700948715, + 0.010694047436118126, + 0.03487825766205788, + 0.0077699534595012665, + 0.004359615035355091, + 0.005712837912142277, + 0.004332107491791248, + 0.005640366580337286, + 0.01832439750432968, + 0.007635066751390696, + 0.017122728750109673, + 0.023106234148144722, + 0.004449660889804363, + 0.010113573633134365, + 0.004263861570507288, + 0.014010575599968433, + 0.17972546815872192, + 0.017307445406913757, + 0.15296408534049988, + 0.013748177327215672, + 0.00472393911331892, + 0.008235577493906021, + 0.00433039665222168, + 0.011102688498795033, + 0.10961885750293732, + 0.011594347655773163, + 0.10087738931179047, + 0.0009999023750424385, + 0.0010593869956210256, + 0.0013448019744828343, + 0.0012193771544843912, + 0.00096268265042454, + 0.0016929575940594077, + 0.0012405108427628875, + 0.0013147161807864904, + 0.0013247068272903562, + 0.0016584635013714433, + 0.0012181809870526195, + 0.0014747919049113989, + 0.0017302829073742032, + 0.0014141142601147294, + 0.0020352236460894346, + 0.001461155479773879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.01607646234333515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898404061794281 + ], + "opposite_logprob_tokens": [ + -0.523404061794281 + ], + "logprob_reward": -0.898404061794281, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898404061794281, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.008858912624418736, + 0.05116433650255203, + 0.027695653960108757, + 0.03409789502620697, + 0.04152892157435417, + 0.2514406442642212, + 0.3741758167743683, + 0.17730426788330078, + 0.021435629576444626, + 0.012297878041863441 + ], + "credit_map": [ + 0.001203453866764903, + 0.0007339160656556487, + 0.0009253103635273874, + 0.0009062729659490287, + 0.0008010591845959425, + 0.0012450311332941055, + 0.0013184304116293788, + 0.0017254386330023408, + 0.0030511722434312105, + 0.0029374901205301285, + 0.002211517421528697, + 0.001990474294871092, + 0.0025650118477642536, + 0.01905006170272827, + 0.0036612655967473984, + 0.015697341412305832, + 0.0024214296136051416, + 0.004855879582464695, + 0.0025030672550201416, + 0.0035573525819927454, + 0.0042234137654304504, + 0.003201841376721859, + 0.0029744403436779976, + 0.003958229441195726, + 0.0037849571090191603, + 0.002527683973312378, + 0.0025108810514211655, + 0.0031430041417479515, + 0.00386730395257473, + 0.006694139912724495, + 0.0029273678082972765, + 0.008642555214464664, + 0.0035030010621994734, + 0.004245180636644363, + 0.0026952058542519808, + 0.003383581293746829, + 0.0027207592502236366, + 0.010111888870596886, + 0.004122029524296522, + 0.010747273452579975, + 0.01410252321511507, + 0.002283655572682619, + 0.0064913309179246426, + 0.0030731556471437216, + 0.008882605470716953, + 0.10994712263345718, + 0.01167032029479742, + 0.09498994797468185, + 0.020961333066225052, + 0.007937316782772541, + 0.0094896936789155, + 0.006871757097542286, + 0.020215144380927086, + 0.1588113009929657, + 0.018963579088449478, + 0.13092568516731262, + 0.008995452895760536, + 0.004268012009561062, + 0.006593627855181694, + 0.004339799750596285, + 0.009213017299771309, + 0.06915922462940216, + 0.00928074773401022, + 0.0654543936252594, + 0.0013631118927150965, + 0.0011916097719222307, + 0.0009988914243876934, + 0.0007706551114097238, + 0.001138067222200334, + 0.007764966227114201, + 0.0010827021906152368, + 0.007125625386834145, + 0.0014544507721439004, + 0.0010366329224780202, + 0.000983276404440403, + 0.0010754881659522653, + 0.0011131888022646308, + 0.002798123052343726, + 0.001118563930504024, + 0.002718153176829219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.898404061794281 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, macro-style shot remains perfectly static, focusing on the complex landscape of a human face. The frame is dominated by a gleaming eye with a crystal-like pale blue iris, its intricate fibrous pattern visible in ultra-fine detail. A single, cold-toned key light from a high side angle illuminates the scene, giving the cornea a damp sheen and revealing its delicate network of capillaries. From the delicate inner corner of the eye, a viscous black liquid begins to gush, accumulating erratically before steadily and deliberately seeping down to the lower lash line. This dark liquid, with its oily, reflective surface, captures the single light source, following a slow, uninterrupted path along pale skin, navigating the fine, almost invisible fuzz and delicate texture of pores. The stark contrast between the dark liquid, pale skin, and bright eye, combined with the subject's unblinking stillness, creates a deeply unsettling and surreal atmosphere, filled with unspoken tension.", + "seed": 11723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_14/index_1522_seed_11723_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17388028209097683, + "vlm_logprob": -0.17388028209097683, + "vlm_yes_no_margin": 3.9000000238418577, + "vlm_token_credit": -0.17388028209097683, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00392037071287632 + ], + "opposite_logprob_tokens": [ + -8.128920555114746 + ], + "logprob_reward": -0.00392037071287632, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.00392037071287632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08062127232551575, + 0.04271414130926132, + 0.02988804318010807, + 0.028854096308350563, + 0.029464172199368477, + 0.04336036741733551, + 0.18382035195827484, + 0.1827285885810852, + 0.2723604440689087, + 0.10618849843740463 + ], + "credit_map": [ + 0.01310852263122797, + 0.005077974405139685, + 0.009601529687643051, + 0.005615211091935635, + 0.00840821024030447, + 0.0068594482727348804, + 0.011346966959536076, + 0.020603418350219727, + 0.007502029649913311, + 0.0033550127409398556, + 0.0051321908831596375, + 0.003217578399926424, + 0.004559141118079424, + 0.005034294445067644, + 0.0050454698503017426, + 0.008868427015841007, + 0.003893111366778612, + 0.002115303184837103, + 0.0032088684383779764, + 0.002409544540569186, + 0.003286749357357621, + 0.0033064913004636765, + 0.0038776136934757233, + 0.007790361996740103, + 0.0036432184278964996, + 0.0018624794902279973, + 0.0029434903990477324, + 0.0019815575797110796, + 0.004454857669770718, + 0.0030834481585770845, + 0.005843460094183683, + 0.005041586700826883, + 0.004254597704857588, + 0.001965396571904421, + 0.002699194708839059, + 0.002662291517481208, + 0.003536354983225465, + 0.0028018876910209656, + 0.00624614953994751, + 0.005298301111906767, + 0.004517079796642065, + 0.0031871232204139233, + 0.0036711629945784807, + 0.003987223841249943, + 0.00647954735904932, + 0.004241448827087879, + 0.009226451627910137, + 0.008050336502492428, + 0.022603457793593407, + 0.024944208562374115, + 0.01889427751302719, + 0.021183505654335022, + 0.02376580983400345, + 0.015424897894263268, + 0.031112078577280045, + 0.02589213103055954, + 0.016055598855018616, + 0.014341386966407299, + 0.019041690975427628, + 0.020615560933947563, + 0.0209756251424551, + 0.02007533423602581, + 0.03649690002202988, + 0.03512650355696678, + 0.030710110440850258, + 0.03266894444823265, + 0.028575953096151352, + 0.032171912491321564, + 0.03100898116827011, + 0.03129076957702637, + 0.04214631766080856, + 0.04378746822476387, + 0.009406758472323418, + 0.013256247155368328, + 0.01279655285179615, + 0.01297285221517086, + 0.016598787158727646, + 0.009945015422999859, + 0.01709573343396187, + 0.014116553589701653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.00392037071287632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637394517660141 + ], + "opposite_logprob_tokens": [ + -3.651373863220215 + ], + "logprob_reward": -0.02637394517660141, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637394517660141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16449914872646332, + 0.07828298956155777, + 0.11199944466352463, + 0.11426518112421036, + 0.07118424028158188, + 0.07945103943347931, + 0.14377054572105408, + 0.09661224484443665, + 0.08091937005519867, + 0.059015773236751556 + ], + "credit_map": [ + 0.01805277355015278, + 0.011083369143307209, + 0.01477143820375204, + 0.01406971924006939, + 0.025331493467092514, + 0.016525689512491226, + 0.03855758532881737, + 0.02610706351697445, + 0.010891621932387352, + 0.005496542435139418, + 0.007658395450562239, + 0.0061307866126298904, + 0.012130333110690117, + 0.009219042956829071, + 0.01464115735143423, + 0.012115114368498325, + 0.020612353459000587, + 0.005923897493630648, + 0.012930981814861298, + 0.008038613013923168, + 0.011207051575183868, + 0.013129754923284054, + 0.015383926220238209, + 0.02477286010980606, + 0.011950178071856499, + 0.005246116779744625, + 0.010526460595428944, + 0.006857037544250488, + 0.021707704290747643, + 0.010555131360888481, + 0.03131764382123947, + 0.016104906797409058, + 0.009028008207678795, + 0.0037265538703650236, + 0.008721227757632732, + 0.0060881138779222965, + 0.008391034789383411, + 0.009117374196648598, + 0.00806010514497757, + 0.01805182546377182, + 0.006667406763881445, + 0.004138166084885597, + 0.008318142965435982, + 0.006541060283780098, + 0.012436021119356155, + 0.006667871959507465, + 0.024171924218535423, + 0.010510445572435856, + 0.01128070056438446, + 0.016996974125504494, + 0.010302773676812649, + 0.015441489405930042, + 0.02122833952307701, + 0.011634166352450848, + 0.040854811668395996, + 0.01603129878640175, + 0.008086810819804668, + 0.008936147205531597, + 0.007630537264049053, + 0.014585128985345364, + 0.012696282006800175, + 0.009073163382709026, + 0.020980756729841232, + 0.014623417519032955, + 0.007296773139387369, + 0.009908165782690048, + 0.008516710251569748, + 0.013560515828430653, + 0.008566999807953835, + 0.006442359182983637, + 0.015945546329021454, + 0.010682297870516777, + 0.008658631704747677, + 0.006252750288695097, + 0.00640887999907136, + 0.007436835672706366, + 0.005653250031173229, + 0.006774656940251589, + 0.007910729385912418, + 0.009920042008161545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.02637394517660141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287082850933075 + ], + "opposite_logprob_tokens": [ + -2.0178709030151367 + ], + "logprob_reward": -0.14287082850933075, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14287082850933075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17941279709339142, + 0.09701430052518845, + 0.11964808404445648, + 0.08904900401830673, + 0.08411585539579391, + 0.10853163152933121, + 0.09778283536434174, + 0.0701519101858139, + 0.10280029475688934, + 0.051493316888809204 + ], + "credit_map": [ + 0.0304703488945961, + 0.012176896445453167, + 0.01939643733203411, + 0.014797229319810867, + 0.017460282891988754, + 0.018555335700511932, + 0.022150512784719467, + 0.044405750930309296, + 0.012982632964849472, + 0.006087462417781353, + 0.008931847289204597, + 0.00835881195962429, + 0.011235613375902176, + 0.007742292247712612, + 0.017038574442267418, + 0.024637067690491676, + 0.01635647937655449, + 0.007206049282103777, + 0.012075935490429401, + 0.010715903714299202, + 0.013024144805967808, + 0.0141068110242486, + 0.016227608546614647, + 0.02993515133857727, + 0.015520649962127209, + 0.004509604535996914, + 0.0110922921448946, + 0.004859736654907465, + 0.013737916015088558, + 0.009895382449030876, + 0.016142813488841057, + 0.013290607370436192, + 0.007843082770705223, + 0.005647348240017891, + 0.006346321199089289, + 0.0087567875161767, + 0.013134102337062359, + 0.005716474261134863, + 0.025276677682995796, + 0.011395060457289219, + 0.010468650609254837, + 0.005591411143541336, + 0.012216979637742043, + 0.010679533705115318, + 0.016277385875582695, + 0.008842683397233486, + 0.028687816113233566, + 0.01576717011630535, + 0.012993951328098774, + 0.013401933014392853, + 0.00962867308408022, + 0.00964038074016571, + 0.010068605653941631, + 0.011415733024477959, + 0.011661683209240437, + 0.018971871584653854, + 0.007878652773797512, + 0.005734972655773163, + 0.006088923197239637, + 0.013732880353927612, + 0.008186573162674904, + 0.006088356487452984, + 0.01518369372934103, + 0.007257855962961912, + 0.008874926716089249, + 0.006906712893396616, + 0.009379656054079533, + 0.012962005101144314, + 0.014327997341752052, + 0.00928320363163948, + 0.02808336541056633, + 0.012982415035367012, + 0.007856719195842743, + 0.004504564683884382, + 0.005767198279500008, + 0.00602311547845602, + 0.005663128569722176, + 0.005580496042966843, + 0.008541951887309551, + 0.007556139025837183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.14287082850933075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820087829604745 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820087829604745, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820087829604745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06653789430856705, + 0.043911952525377274, + 0.02145002782344818, + 0.034783072769641876, + 0.06328969448804855, + 0.0762488842010498, + 0.17120660841464996, + 0.05553143844008446, + 0.30660876631736755, + 0.16043168306350708 + ], + "credit_map": [ + 0.010761920362710953, + 0.004986388608813286, + 0.007459057029336691, + 0.00519236084073782, + 0.007591220084577799, + 0.006479682866483927, + 0.010299136862158775, + 0.013768130913376808, + 0.006703948602080345, + 0.002759253839030862, + 0.005126445088535547, + 0.003141382010653615, + 0.005164712201803923, + 0.004707318730652332, + 0.00685088150203228, + 0.009458011947572231, + 0.003001823788508773, + 0.0015557791339233518, + 0.0025501372292637825, + 0.0018377264495939016, + 0.0021544408518821, + 0.0023670506197959185, + 0.0033100582659244537, + 0.004673011600971222, + 0.006830403581261635, + 0.0018240142380818725, + 0.004343915265053511, + 0.0025311654899269342, + 0.0042274328880012035, + 0.0036848534364253283, + 0.004967005457729101, + 0.006374284625053406, + 0.006715340539813042, + 0.004603046458214521, + 0.006127667613327503, + 0.0059361509047448635, + 0.007488973904401064, + 0.006203800439834595, + 0.011306392960250378, + 0.014908325858414173, + 0.009626145474612713, + 0.004458251409232616, + 0.0068021854385733604, + 0.004293684381991625, + 0.012628358788788319, + 0.00781271606683731, + 0.016771551221609116, + 0.013855991885066032, + 0.021858010441064835, + 0.01480499655008316, + 0.01776125654578209, + 0.011762765236198902, + 0.01912877894937992, + 0.017324650660157204, + 0.038196299225091934, + 0.03036985732614994, + 0.006643199827522039, + 0.00554776331409812, + 0.005607037339359522, + 0.009913327172398567, + 0.006495324894785881, + 0.004894925281405449, + 0.009921453893184662, + 0.006508401595056057, + 0.03438838943839073, + 0.020441699773073196, + 0.02637430652976036, + 0.03364979103207588, + 0.04060758277773857, + 0.03138994798064232, + 0.07869426161050797, + 0.04106279835104942, + 0.018477223813533783, + 0.01511083822697401, + 0.013519302941858768, + 0.023649020120501518, + 0.02245773933827877, + 0.016972877085208893, + 0.02866719476878643, + 0.021577490493655205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.002820087829604745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693416178226471 + ], + "opposite_logprob_tokens": [ + -0.693416178226471 + ], + "logprob_reward": -0.693416178226471, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693416178226471, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17935095727443695, + 0.0409829206764698, + 0.0623399093747139, + 0.11576852202415466, + 0.108983114361763, + 0.0789981484413147, + 0.11481118202209473, + 0.125893697142601, + 0.12115985155105591, + 0.05171166732907295 + ], + "credit_map": [ + 0.02709193341434002, + 0.013492611236870289, + 0.017607903108000755, + 0.013685157522559166, + 0.022243928164243698, + 0.019187461584806442, + 0.029763786122202873, + 0.03627820312976837, + 0.004995640832930803, + 0.003953530918806791, + 0.0032300336752086878, + 0.0029487255960702896, + 0.0068980916403234005, + 0.005130869802087545, + 0.0076226480305194855, + 0.0062033855356276035, + 0.008823796175420284, + 0.004293051548302174, + 0.006804963573813438, + 0.005736515391618013, + 0.0065543875098228455, + 0.006798204500228167, + 0.008291691541671753, + 0.015037301927804947, + 0.013265072368085384, + 0.00584731251001358, + 0.01096996758133173, + 0.007882404141128063, + 0.017268238589167595, + 0.011790642514824867, + 0.02439393661916256, + 0.024350957944989204, + 0.010132872499525547, + 0.0061110868118703365, + 0.010361934080719948, + 0.00885712169110775, + 0.014928558841347694, + 0.010893695056438446, + 0.02269837260246277, + 0.02499949000775814, + 0.010518701747059822, + 0.004376010969281197, + 0.006574285216629505, + 0.00621319143101573, + 0.010918347164988518, + 0.008588004857301712, + 0.01879107765853405, + 0.0130185317248106, + 0.009079331532120705, + 0.009392148815095425, + 0.008935441263020039, + 0.017473405227065086, + 0.014030678197741508, + 0.010467649437487125, + 0.027695825323462486, + 0.017736714333295822, + 0.013115703128278255, + 0.009447213262319565, + 0.01061087753623724, + 0.023705970495939255, + 0.01511908695101738, + 0.012177806347608566, + 0.02457202970981598, + 0.01714502088725567, + 0.01019690278917551, + 0.009286253713071346, + 0.012790882028639317, + 0.012286156415939331, + 0.01853441260755062, + 0.011494149453938007, + 0.030785568058490753, + 0.01578553207218647, + 0.005223752930760384, + 0.004654479678720236, + 0.004618362989276648, + 0.006322451867163181, + 0.007132034283131361, + 0.006872135680168867, + 0.009977751411497593, + 0.006910704076290131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.693416178226471 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, macro-style shot remains perfectly static, focusing on the complex landscape of a human face. The frame is dominated by a gleaming eye with a crystal-like pale blue iris, its intricate fibrous pattern visible in ultra-fine detail. A single, cold-toned key light from a high side angle illuminates the scene, giving the cornea a damp sheen and revealing its delicate network of capillaries. From the delicate inner corner of the eye, a viscous black liquid begins to gush, accumulating erratically before steadily and deliberately seeping down to the lower lash line. This dark liquid, with its oily, reflective surface, captures the single light source, following a slow, uninterrupted path along pale skin, navigating the fine, almost invisible fuzz and delicate texture of pores. The stark contrast between the dark liquid, pale skin, and bright eye, combined with the subject's unblinking stillness, creates a deeply unsettling and surreal atmosphere, filled with unspoken tension.", + "seed": 11724, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_14/index_1522_seed_11724_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20807800544425845, + "vlm_logprob": -0.20807800544425845, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.20807800544425845, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004909721668809652 + ], + "opposite_logprob_tokens": [ + -8.25490951538086 + ], + "logprob_reward": -0.004909721668809652, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.004909721668809652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0365789458155632, + 0.019395064562559128, + 0.013809900730848312, + 0.01766122505068779, + 0.024232590571045876, + 0.03821364790201187, + 0.10652332752943039, + 0.2817714810371399, + 0.08470126986503601, + 0.37711259722709656 + ], + "credit_map": [ + 0.007954373955726624, + 0.0028060267213732004, + 0.006211753003299236, + 0.002952813170850277, + 0.002548783551901579, + 0.0036224827636033297, + 0.004917206708341837, + 0.00556550920009613, + 0.0032954418566077948, + 0.0013346774503588676, + 0.002750558312982321, + 0.0012025327887386084, + 0.001500390120781958, + 0.0024052925873547792, + 0.0023785901721566916, + 0.004527580458670855, + 0.002870732918381691, + 0.0010160672245547175, + 0.0025071720592677593, + 0.000747710233554244, + 0.0009371615597046912, + 0.0020587947219610214, + 0.0018006410682573915, + 0.0018716215854510665, + 0.003228007582947612, + 0.0011289797257632017, + 0.00251115788705647, + 0.0005721377674490213, + 0.0009901807643473148, + 0.003457498736679554, + 0.0017409310676157475, + 0.004032331984490156, + 0.0029685304034501314, + 0.0014678057050332427, + 0.0050416202284395695, + 0.0008969189948402345, + 0.0014327798271551728, + 0.003942383453249931, + 0.001501671620644629, + 0.0069808801636099815, + 0.00657985033467412, + 0.004178559873253107, + 0.005148658063262701, + 0.0014780061319470406, + 0.0018680922221392393, + 0.006449215579777956, + 0.003081447212025523, + 0.009429818019270897, + 0.014449349604547024, + 0.011022713035345078, + 0.016429753974080086, + 0.0071953777223825455, + 0.006853156257420778, + 0.01713753305375576, + 0.00948713906109333, + 0.023948298767209053, + 0.029447928071022034, + 0.017851658165454865, + 0.01578742451965809, + 0.006219762377440929, + 0.006139146164059639, + 0.09198134392499924, + 0.011001030914485455, + 0.1033431738615036, + 0.014788180589675903, + 0.008243335410952568, + 0.018967559561133385, + 0.0034268274903297424, + 0.004957562778145075, + 0.01212750468403101, + 0.003821553662419319, + 0.01836874708533287, + 0.03687462955713272, + 0.015844518318772316, + 0.021309487521648407, + 0.010172941721975803, + 0.008323030546307564, + 0.1288832724094391, + 0.010940399020910263, + 0.14476431906223297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.004909721668809652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0182094294577837 + ], + "opposite_logprob_tokens": [ + -4.018209457397461 + ], + "logprob_reward": -0.0182094294577837, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0182094294577837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07694501429796219, + 0.06014426052570343, + 0.048824358731508255, + 0.043985478579998016, + 0.039447203278541565, + 0.029407592490315437, + 0.17685802280902863, + 0.19725240767002106, + 0.07795698195695877, + 0.24917861819267273 + ], + "credit_map": [ + 0.011359541676938534, + 0.00588097283616662, + 0.008275782689452171, + 0.0060801636427640915, + 0.007897157222032547, + 0.012946810573339462, + 0.00850110687315464, + 0.016003485769033432, + 0.0109684644266963, + 0.003996945917606354, + 0.0070467242039740086, + 0.0030959981959313154, + 0.003622811520472169, + 0.008514828979969025, + 0.006604006979614496, + 0.016294484958052635, + 0.0054932222701609135, + 0.0030520125292241573, + 0.004429193679243326, + 0.0024454842787235975, + 0.003368870820850134, + 0.011483453214168549, + 0.005060549359768629, + 0.013491573743522167, + 0.005836084950715303, + 0.002222505398094654, + 0.004163251258432865, + 0.0011574242962524295, + 0.002840779023244977, + 0.011055673472583294, + 0.002068523084744811, + 0.014641241170465946, + 0.009149952791631222, + 0.00269905524328351, + 0.007429715245962143, + 0.0019614610355347395, + 0.0032474887557327747, + 0.004557284060865641, + 0.003120903857052326, + 0.0072813453152775764, + 0.003773766104131937, + 0.003893592394888401, + 0.005457236431539059, + 0.0019659181125462055, + 0.0035240454599261284, + 0.003198630642145872, + 0.0037363101728260517, + 0.0038580961991101503, + 0.021212179213762283, + 0.017169471830129623, + 0.026689892634749413, + 0.014094190672039986, + 0.013275306671857834, + 0.028392182663083076, + 0.017756622284650803, + 0.03826818987727165, + 0.02300853282213211, + 0.013358375988900661, + 0.017298396676778793, + 0.0046159359626472, + 0.008296546526253223, + 0.05914735794067383, + 0.00872002076357603, + 0.06280725449323654, + 0.011629913002252579, + 0.006743077654391527, + 0.013315809890627861, + 0.0029930558521300554, + 0.005897428374737501, + 0.014447446912527084, + 0.0044440519995987415, + 0.018486209213733673, + 0.02496062032878399, + 0.010122478939592838, + 0.01420681457966566, + 0.006688904482871294, + 0.005348729435354471, + 0.08488230407238007, + 0.007038872689008713, + 0.09592990577220917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.0182094294577837 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907435834407806 + ], + "opposite_logprob_tokens": [ + -2.4640743732452393 + ], + "logprob_reward": -0.08907435834407806, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907435834407806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06716547906398773, + 0.04749083146452904, + 0.03916085883975029, + 0.12820760905742645, + 0.05340125411748886, + 0.027997255325317383, + 0.16974207758903503, + 0.2567412257194519, + 0.05402320995926857, + 0.15607021749019623 + ], + "credit_map": [ + 0.012572934851050377, + 0.003651488572359085, + 0.014852062799036503, + 0.005221884232014418, + 0.0039002324920147657, + 0.009194834157824516, + 0.008620422333478928, + 0.009151616133749485, + 0.007102248724550009, + 0.0022667315788567066, + 0.004565846174955368, + 0.002128841821104288, + 0.0021190023981034756, + 0.00949074886739254, + 0.004505186807364225, + 0.015312223695218563, + 0.004420527722686529, + 0.0027399747632443905, + 0.005091932136565447, + 0.0016901347553357482, + 0.002920174039900303, + 0.007380054797977209, + 0.005571701098233461, + 0.009346360340714455, + 0.02149280533194542, + 0.006616351194679737, + 0.01295558176934719, + 0.0031491597183048725, + 0.00484504085034132, + 0.033002015203237534, + 0.00740804523229599, + 0.03873860836029053, + 0.006107229273766279, + 0.0024729985743761063, + 0.008562987670302391, + 0.0017248166259378195, + 0.002223244169726968, + 0.011235153302550316, + 0.002914857119321823, + 0.01815997064113617, + 0.003789646318182349, + 0.003540563164278865, + 0.0055582644417881966, + 0.0013992469757795334, + 0.0017597682308405638, + 0.003899332834407687, + 0.00271989475004375, + 0.005330538842827082, + 0.022994646802544594, + 0.01445626001805067, + 0.029965301975607872, + 0.011587935499846935, + 0.009152773767709732, + 0.027571912854909897, + 0.015384195372462273, + 0.03862905129790306, + 0.026671407744288445, + 0.01632045954465866, + 0.013825920410454273, + 0.0056623839773237705, + 0.005454455502331257, + 0.08449024707078934, + 0.0102510591968894, + 0.09406527876853943, + 0.0084488894790411, + 0.0046984730288386345, + 0.01040436141192913, + 0.0022391381207853556, + 0.0032930627930909395, + 0.009615072049200535, + 0.0030966508202254772, + 0.01222756877541542, + 0.015236400067806244, + 0.00616714870557189, + 0.008884597569704056, + 0.004246399272233248, + 0.0035706637427210808, + 0.053203705698251724, + 0.004320630803704262, + 0.06044067442417145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.08907435834407806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803065583109856 + ], + "opposite_logprob_tokens": [ + -3.5298030376434326 + ], + "logprob_reward": -0.029803065583109856, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803065583109856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049215011298656464, + 0.021609442308545113, + 0.04521910846233368, + 0.06233009696006775, + 0.03981463238596916, + 0.0801217332482338, + 0.42792579531669617, + 0.020444780588150024, + 0.08365540951490402, + 0.16966399550437927 + ], + "credit_map": [ + 0.007594346068799496, + 0.002983296290040016, + 0.006528479512780905, + 0.0038137563969939947, + 0.0036305643152445555, + 0.008340245112776756, + 0.005895004607737064, + 0.010429318994283676, + 0.002648850902915001, + 0.0011516865342855453, + 0.004013285972177982, + 0.0012037204578518867, + 0.0012334140483289957, + 0.003987605683505535, + 0.0023153237998485565, + 0.005055554676800966, + 0.004625278525054455, + 0.0019295229576528072, + 0.0034092850983142853, + 0.0014821719378232956, + 0.002577480860054493, + 0.014288946986198425, + 0.0037345096934586763, + 0.013171914033591747, + 0.007469854783266783, + 0.003158493898808956, + 0.0059527079574763775, + 0.001538167241960764, + 0.002211362821981311, + 0.01720612123608589, + 0.003127624047920108, + 0.021665766835212708, + 0.009916596114635468, + 0.003056098008528352, + 0.009310769848525524, + 0.0018765936838462949, + 0.002694027964025736, + 0.003838072530925274, + 0.0029017520137131214, + 0.006220724433660507, + 0.010623989626765251, + 0.005390813108533621, + 0.013817835599184036, + 0.0041380394250154495, + 0.005707607138901949, + 0.013183657079935074, + 0.009198970161378384, + 0.018060822039842606, + 0.05491666495800018, + 0.04153217747807503, + 0.07245790958404541, + 0.03615134581923485, + 0.026690155267715454, + 0.06304006278514862, + 0.043305374681949615, + 0.08983210474252701, + 0.002708286279812455, + 0.001672443002462387, + 0.0026174255181103945, + 0.0011454958003014326, + 0.00257405755110085, + 0.0037785309832543135, + 0.0016435040161013603, + 0.0043050372041761875, + 0.009720570407807827, + 0.008533000946044922, + 0.008171966299414635, + 0.004302633460611105, + 0.011444848030805588, + 0.016610722988843918, + 0.005646728910505772, + 0.01922493241727352, + 0.016783222556114197, + 0.00655832514166832, + 0.00959023553878069, + 0.004650827497243881, + 0.004038672894239426, + 0.05778315290808678, + 0.004771075211465359, + 0.06548848748207092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.029803065583109856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898393452167511 + ], + "opposite_logprob_tokens": [ + -0.523393452167511 + ], + "logprob_reward": -0.898393452167511, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898393452167511, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.023494521155953407, + 0.012161738239228725, + 0.05757156014442444, + 0.012820792384445667, + 0.02012527920305729, + 0.04805292189121246, + 0.10576987266540527, + 0.6394766569137573, + 0.0312033761292696, + 0.04932323843240738 + ], + "credit_map": [ + 0.004256768152117729, + 0.0019853273406624794, + 0.002952044364064932, + 0.0019818972796201706, + 0.002275282982736826, + 0.0035409238189458847, + 0.0030556435231119394, + 0.0034466360229998827, + 0.0016766691114753485, + 0.0010015678126364946, + 0.0022115251049399376, + 0.0010921269422397017, + 0.0010534694883972406, + 0.0014796812320128083, + 0.001683542737737298, + 0.0019631574396044016, + 0.008906455710530281, + 0.0026476033963263035, + 0.006653830409049988, + 0.0019510862184688449, + 0.002385699888691306, + 0.013391073793172836, + 0.0058456165716052055, + 0.01579020358622074, + 0.0026960691902786493, + 0.0009755194187164307, + 0.0024071475490927696, + 0.0004677929973695427, + 0.0009583166684024036, + 0.0018385410076007247, + 0.0014499528333544731, + 0.002027454087510705, + 0.005164796952158213, + 0.0020654306281358004, + 0.00412042997777462, + 0.0009686952107585967, + 0.0014521335251629353, + 0.001794920302927494, + 0.0014342048671096563, + 0.003124670824036002, + 0.010719293728470802, + 0.004397745709866285, + 0.008852806873619556, + 0.0018277611816301942, + 0.0023906289134174585, + 0.006969891022890806, + 0.003197359386831522, + 0.009697441011667252, + 0.013346127234399319, + 0.012197775766253471, + 0.01710866577923298, + 0.008282778784632683, + 0.007532253861427307, + 0.01327580027282238, + 0.01335613802075386, + 0.020670350641012192, + 0.07303903996944427, + 0.042579058557748795, + 0.050551850348711014, + 0.014290137216448784, + 0.019093377515673637, + 0.19914256036281586, + 0.026323605328798294, + 0.21445709466934204, + 0.0051260944455862045, + 0.003263481892645359, + 0.00673076743260026, + 0.001245796331204474, + 0.0025396738201379776, + 0.0044587706215679646, + 0.0016998026985675097, + 0.006138992495834827, + 0.00482527818530798, + 0.002274684375151992, + 0.0027003006543964148, + 0.0014429913135245442, + 0.0013997872592881322, + 0.016981935128569603, + 0.001507397973909974, + 0.018190870061516762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.898393452167511 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, macro-style shot remains perfectly static, focusing on the complex landscape of a human face. The frame is dominated by a gleaming eye with a crystal-like pale blue iris, its intricate fibrous pattern visible in ultra-fine detail. A single, cold-toned key light from a high side angle illuminates the scene, giving the cornea a damp sheen and revealing its delicate network of capillaries. From the delicate inner corner of the eye, a viscous black liquid begins to gush, accumulating erratically before steadily and deliberately seeping down to the lower lash line. This dark liquid, with its oily, reflective surface, captures the single light source, following a slow, uninterrupted path along pale skin, navigating the fine, almost invisible fuzz and delicate texture of pores. The stark contrast between the dark liquid, pale skin, and bright eye, combined with the subject's unblinking stillness, creates a deeply unsettling and surreal atmosphere, filled with unspoken tension.", + "seed": 11725, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_14/index_1522_seed_11725_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17120869350619614, + "vlm_logprob": -0.17120869350619614, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.17120869350619614, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0023636280093342066 + ], + "opposite_logprob_tokens": [ + -8.752363204956055 + ], + "logprob_reward": -0.0023636280093342066, + "yes_no_margin_reward": 8.75, + "token_credit_reward": -0.0023636280093342066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13944850862026215, + 0.052713360637426376, + 0.05378536134958267, + 0.089113749563694, + 0.10133053362369537, + 0.11790350824594498, + 0.18175289034843445, + 0.07478095591068268, + 0.14473983645439148, + 0.04443126916885376 + ], + "credit_map": [ + 0.01198939885944128, + 0.013377556577324867, + 0.023805392906069756, + 0.020107699558138847, + 0.016652537509799004, + 0.013983875513076782, + 0.022331375628709793, + 0.01720067299902439, + 0.010100306943058968, + 0.0052260952070355415, + 0.008992939256131649, + 0.006072923541069031, + 0.005678938236087561, + 0.003966040443629026, + 0.007036702241748571, + 0.0056394170969724655, + 0.007221728563308716, + 0.005029902793467045, + 0.012763802893459797, + 0.005767012946307659, + 0.00571511359885335, + 0.004875735845416784, + 0.005794997327029705, + 0.00661707017570734, + 0.00839547161012888, + 0.008503523655235767, + 0.009769503027200699, + 0.014020350761711597, + 0.010984168387949467, + 0.006932332646101713, + 0.021468885242938995, + 0.009039518423378468, + 0.010364546440541744, + 0.010865249671041965, + 0.014124966226518154, + 0.016595976427197456, + 0.010273296386003494, + 0.009375455789268017, + 0.019231094047427177, + 0.01049994956701994, + 0.014087335206568241, + 0.012053865939378738, + 0.020171446725726128, + 0.01523884292691946, + 0.013715707696974277, + 0.011664967983961105, + 0.01704912632703781, + 0.013922211714088917, + 0.016010375693440437, + 0.01921071670949459, + 0.02139267325401306, + 0.0294975396245718, + 0.024793580174446106, + 0.01782650128006935, + 0.034574951976537704, + 0.01844656839966774, + 0.007685861550271511, + 0.00781477615237236, + 0.010046251118183136, + 0.009614307433366776, + 0.011189674027264118, + 0.004872411023825407, + 0.015830663964152336, + 0.00772700784727931, + 0.02546333335340023, + 0.013714850880205631, + 0.026611313223838806, + 0.02630983479321003, + 0.01334954984486103, + 0.00875098630785942, + 0.015177113004028797, + 0.01536285225301981, + 0.005716749466955662, + 0.004519292619079351, + 0.0076911840587854385, + 0.006579766981303692, + 0.007174219004809856, + 0.0034249373711645603, + 0.0044388980604708195, + 0.004886219743639231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.0023636280093342066 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609283819794655 + ], + "opposite_logprob_tokens": [ + -3.8956093788146973 + ], + "logprob_reward": -0.020609283819794655, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609283819794655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1978590041399002, + 0.14263233542442322, + 0.07433509826660156, + 0.09980250149965286, + 0.06496918201446533, + 0.09174726158380508, + 0.09262438863515854, + 0.07781330496072769, + 0.1000981479883194, + 0.0581187978386879 + ], + "credit_map": [ + 0.018443966284394264, + 0.01882546953856945, + 0.03776348754763603, + 0.026463299989700317, + 0.028907228261232376, + 0.016521122306585312, + 0.02887476049363613, + 0.02205965481698513, + 0.014014998450875282, + 0.012998584657907486, + 0.018689490854740143, + 0.01726953126490116, + 0.025510970503091812, + 0.01290502306073904, + 0.02671693079173565, + 0.014526791870594025, + 0.007572028320282698, + 0.0049561127088963985, + 0.017251791432499886, + 0.006907665636390448, + 0.009565122425556183, + 0.0065384493209421635, + 0.011745549738407135, + 0.009798371233046055, + 0.012851315550506115, + 0.009570460766553879, + 0.011888133361935616, + 0.012440997175872326, + 0.012089486233890057, + 0.008742578327655792, + 0.022094350308179855, + 0.010125174187123775, + 0.00780474441125989, + 0.005797646939754486, + 0.010783777572214603, + 0.010418680496513844, + 0.005694138817489147, + 0.006883897352963686, + 0.010362266562879086, + 0.007224028930068016, + 0.0074682217091321945, + 0.009577935561537743, + 0.011665319092571735, + 0.01443478837609291, + 0.012821856886148453, + 0.00839980784803629, + 0.017519904300570488, + 0.009859412908554077, + 0.010120940394699574, + 0.009180190972983837, + 0.015943219885230064, + 0.014499318785965443, + 0.01286966074258089, + 0.007929484359920025, + 0.012480349279940128, + 0.009601219557225704, + 0.007397894281893969, + 0.008021735586225986, + 0.008399012498557568, + 0.011918146163225174, + 0.01051844097673893, + 0.005423910450190306, + 0.018381239846348763, + 0.007752919569611549, + 0.015512797981500626, + 0.010223979130387306, + 0.015277629718184471, + 0.015602074563503265, + 0.011472993530333042, + 0.007222285028547049, + 0.014981318265199661, + 0.009805064648389816, + 0.011561227031052113, + 0.0059669348411262035, + 0.010651263408362865, + 0.007793894037604332, + 0.005370998755097389, + 0.003784497966989875, + 0.006281101610511541, + 0.006708874367177486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.020609283819794655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710757553577423 + ], + "opposite_logprob_tokens": [ + -2.127107620239258 + ], + "logprob_reward": -0.12710757553577423, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710757553577423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0878254771232605, + 0.05576048418879509, + 0.03646073117852211, + 0.06362248957157135, + 0.05681364983320236, + 0.12187475711107254, + 0.34138041734695435, + 0.1489216685295105, + 0.0396210178732872, + 0.047719258815050125 + ], + "credit_map": [ + 0.015957355499267578, + 0.007056865841150284, + 0.019773008301854134, + 0.009338604286313057, + 0.007318838499486446, + 0.009039654396474361, + 0.00857527181506157, + 0.010765883140265942, + 0.006626216694712639, + 0.005152128171175718, + 0.012589375488460064, + 0.00614452688023448, + 0.005463229492306709, + 0.00494171679019928, + 0.006972983945161104, + 0.007870306260883808, + 0.0034820029977709055, + 0.003923076204955578, + 0.005012876354157925, + 0.004906273912638426, + 0.0056562586687505245, + 0.003588454332202673, + 0.00630615558475256, + 0.00358563382178545, + 0.005992445163428783, + 0.006915808189660311, + 0.00766530679538846, + 0.009243682958185673, + 0.009896283969283104, + 0.0055731842294335365, + 0.010815201327204704, + 0.007520577870309353, + 0.012120612896978855, + 0.00513504771515727, + 0.009962260723114014, + 0.006787791848182678, + 0.004251915495842695, + 0.0065657347440719604, + 0.004740351345390081, + 0.007249936927109957, + 0.009488481096923351, + 0.012758440338075161, + 0.02209317684173584, + 0.01989111490547657, + 0.015558951534330845, + 0.01011993270367384, + 0.018207672983407974, + 0.013756987638771534, + 0.03357299044728279, + 0.03724503889679909, + 0.032715607434511185, + 0.061511848121881485, + 0.040243878960609436, + 0.0319523811340332, + 0.07071533799171448, + 0.033423345535993576, + 0.01314788032323122, + 0.01583591103553772, + 0.01666327938437462, + 0.022507229819893837, + 0.02105589397251606, + 0.009107667952775955, + 0.03543977066874504, + 0.015164035372436047, + 0.006759260315448046, + 0.0035752099938690662, + 0.00473400903865695, + 0.005399059969931841, + 0.006194344721734524, + 0.0031849094666540623, + 0.005263045430183411, + 0.0045111775398254395, + 0.005318309646099806, + 0.004933829419314861, + 0.005661796312779188, + 0.0068091824650764465, + 0.008761816658079624, + 0.005388803314417601, + 0.006374051794409752, + 0.004471468739211559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.12710757553577423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544417753815651 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544417753815651, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544417753815651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08452137559652328, + 0.0843995213508606, + 0.05919049680233002, + 0.08741040527820587, + 0.17271585762500763, + 0.40064147114753723, + 0.03846190497279167, + 0.027017749845981598, + 0.01798279583454132, + 0.027658330276608467 + ], + "credit_map": [ + 0.008263098075985909, + 0.008008800446987152, + 0.022723475471138954, + 0.009378173388540745, + 0.007934097200632095, + 0.007848200388252735, + 0.00941422302275896, + 0.010951303876936436, + 0.007930017076432705, + 0.006881832145154476, + 0.020924026146531105, + 0.009133891202509403, + 0.0073525384068489075, + 0.00810349639505148, + 0.012531072832643986, + 0.011542638763785362, + 0.005902628879994154, + 0.005180736072361469, + 0.011516849510371685, + 0.006825463846325874, + 0.006987359374761581, + 0.0060592168010771275, + 0.009457343257963657, + 0.007260897662490606, + 0.009954696521162987, + 0.009239193052053452, + 0.011481760069727898, + 0.012211217544972897, + 0.010898102074861526, + 0.006738726980984211, + 0.01817481778562069, + 0.00871189497411251, + 0.015964215621352196, + 0.017977943643927574, + 0.017446059733629227, + 0.029456036165356636, + 0.02126123011112213, + 0.016927704215049744, + 0.03605947270989418, + 0.01762319542467594, + 0.03806113451719284, + 0.041703347116708755, + 0.05601475387811661, + 0.05875353887677193, + 0.06309565901756287, + 0.03505202755331993, + 0.06733977049589157, + 0.04062126204371452, + 0.004272335208952427, + 0.003142127301543951, + 0.004685545340180397, + 0.007346447091549635, + 0.004196210764348507, + 0.00262737600132823, + 0.008628020994365215, + 0.0035638422705233097, + 0.0023143019061535597, + 0.002426259219646454, + 0.004879512824118137, + 0.003939277492463589, + 0.00419269222766161, + 0.002022399799898267, + 0.004478953313082457, + 0.0027643523644655943, + 0.003439375665038824, + 0.0019290510099381208, + 0.0028916639275848866, + 0.0016383897745981812, + 0.0019785359036177397, + 0.0017343260115012527, + 0.0023684517946094275, + 0.0020030022133141756, + 0.0037275590002536774, + 0.002338431077077985, + 0.00430477038025856, + 0.005344554781913757, + 0.002438050927594304, + 0.0022429495584219694, + 0.004425558261573315, + 0.0028364553581923246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.012544417753815651 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, macro-style view of a human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris pale blue with a visible crystal-like fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single cold-toned key light from a high side angle illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous black liquid gush from the inner corner of the eye and seep down to the lower lash line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693418562412262 + ], + "opposite_logprob_tokens": [ + -0.693418562412262 + ], + "logprob_reward": -0.693418562412262, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693418562412262, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09977146983146667, + 0.07494333386421204, + 0.030079592019319534, + 0.09688782691955566, + 0.038553185760974884, + 0.26734185218811035, + 0.18434710800647736, + 0.13367925584316254, + 0.04157698154449463, + 0.032819438725709915 + ], + "credit_map": [ + 0.01923847571015358, + 0.009603142738342285, + 0.020390305668115616, + 0.009219421073794365, + 0.010190550237894058, + 0.010425594635307789, + 0.00890276487916708, + 0.01180120836943388, + 0.009704415686428547, + 0.00587554881349206, + 0.015116262249648571, + 0.009889744222164154, + 0.007299821823835373, + 0.008239611983299255, + 0.009176166728138924, + 0.009641760028898716, + 0.004773086402565241, + 0.0027655153535306454, + 0.00403987430036068, + 0.004880297929048538, + 0.003762573469430208, + 0.0031064453069120646, + 0.00389143917709589, + 0.0028603598475456238, + 0.008772513829171658, + 0.01022875402122736, + 0.013814018107950687, + 0.014400755055248737, + 0.013901795260608196, + 0.008308438584208488, + 0.018186915665864944, + 0.009274633601307869, + 0.005957192275673151, + 0.004004928283393383, + 0.006191523279994726, + 0.005156633444130421, + 0.0036995913833379745, + 0.004092520568519831, + 0.005532854702323675, + 0.00391794228926301, + 0.024003339931368828, + 0.028083408251404762, + 0.03996295854449272, + 0.04277341440320015, + 0.03844689577817917, + 0.025963859632611275, + 0.04161577671766281, + 0.02649218589067459, + 0.016032375395298004, + 0.0207834392786026, + 0.01511302962899208, + 0.030320456251502037, + 0.022429389879107475, + 0.018619177863001823, + 0.042341578751802444, + 0.018707649782299995, + 0.012710366398096085, + 0.014072079211473465, + 0.013936510309576988, + 0.019674858078360558, + 0.01969265751540661, + 0.008646907284855843, + 0.03157550469040871, + 0.013370365835726261, + 0.007410888094455004, + 0.004181951750069857, + 0.006048492155969143, + 0.004687381908297539, + 0.004666826222091913, + 0.003564129350706935, + 0.006592290010303259, + 0.004425025545060635, + 0.0035735839046537876, + 0.0035667777992784977, + 0.004219874739646912, + 0.005449994001537561, + 0.004889254458248615, + 0.00298386812210083, + 0.00460720295086503, + 0.003528882982209325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.693418562412262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_15/index_tensor([6060, 6060])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_15/index_tensor([6060, 6060])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..817fc4e6be0c4666677c666e66159ac09750541a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_15/index_tensor([6060, 6060])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "gentle downward tilt begins, following and capturing an elegant celadon wine jug as it spins slowly and gracefully in mid-air, seemingly defying gravity. A classic Longquan kiln creation, the jug boasts a long neck and rounded belly. Its glaze is a soft, jade-like pale green, its surface adorned with delicate, natural ice cracks, shimmering with a restrained luster under the soft, diffuse light filtering through the paper window. As the camera descends, the jug's trajectory is precise, until its base, with near-silent grace, touches the deep-grained, richly colored pearwood table, stirring up a barely perceptible puff of dust. In the background, an elegant Song Dynasty study emerges. Through a latticed window, a few sparse, slanting plum blossoms peek through the courtyard, permeating the space with a tranquil, Zen-like atmosphere. The entire picture strictly follows the style of Song Dynasty court painting, with precise composition, elegant colors, even and soft lighting, no strong contrast between light and dark, and emphasis on the outline and texture of objects, creating an ultimate classical aesthetic and a quiet atmosphere of forgetting oneself and the world.", + "seed": 15365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_15/index_6060_seed_15365_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010686351638287305, + "vlm_logprob": -0.010686351638287305, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.010686351638287305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004004911985248327 + ], + "opposite_logprob_tokens": [ + -6.504004955291748 + ], + "logprob_reward": -0.004004911985248327, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004004911985248327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16319356858730316, + 0.05424237996339798, + 0.0559631884098053, + 0.15745730698108673, + 0.18747901916503906, + 0.08447758853435516, + 0.05410057678818703, + 0.07314381003379822, + 0.09810096770524979, + 0.07184156030416489 + ], + "credit_map": [ + 0.017557218670845032, + 0.02132333815097809, + 0.01187230460345745, + 0.03063497133553028, + 0.021205278113484383, + 0.02182280272245407, + 0.028246847912669182, + 0.010530811734497547, + 0.006311497651040554, + 0.006175513379275799, + 0.008030685596168041, + 0.008938211016356945, + 0.005792834330350161, + 0.005744637921452522, + 0.008506144396960735, + 0.004742863122373819, + 0.006368787027895451, + 0.0052806842140853405, + 0.007279748097062111, + 0.008281271904706955, + 0.006690545007586479, + 0.005894633941352367, + 0.009920136071741581, + 0.006247387733310461, + 0.009785422123968601, + 0.007908923551440239, + 0.019214576110243797, + 0.007951261475682259, + 0.04608212783932686, + 0.036696240305900574, + 0.014704160392284393, + 0.015114594250917435, + 0.007525424472987652, + 0.006958026438951492, + 0.07825847715139389, + 0.01356560830026865, + 0.005252478178590536, + 0.006766447797417641, + 0.014808451756834984, + 0.054344117641448975, + 0.009095784276723862, + 0.014794763177633286, + 0.01050223782658577, + 0.012765971943736076, + 0.008744887076318264, + 0.0072522591799497604, + 0.010443164967000484, + 0.010878529399633408, + 0.005623324308544397, + 0.005947841331362724, + 0.007608947344124317, + 0.010030617006123066, + 0.004618771839886904, + 0.005396625027060509, + 0.008579165674746037, + 0.0062952907755970955, + 0.005912916734814644, + 0.00814108643680811, + 0.01514472160488367, + 0.013931347988545895, + 0.005659232847392559, + 0.004898209124803543, + 0.008749889209866524, + 0.010706412605941296, + 0.008866718038916588, + 0.01400749757885933, + 0.014499969780445099, + 0.017095431685447693, + 0.007497529499232769, + 0.012039723806083202, + 0.01574227586388588, + 0.0083518261089921, + 0.007013194728642702, + 0.007698370143771172, + 0.009907218627631664, + 0.011899256147444248, + 0.007259600795805454, + 0.007430246099829674, + 0.012844565324485302, + 0.007789114955812693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.004004911985248327 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074232757091522 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.016074232757091522, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074232757091522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21299263834953308, + 0.04458285868167877, + 0.039834070950746536, + 0.23373399674892426, + 0.08307138830423355, + 0.060149841010570526, + 0.06497304886579514, + 0.0554032176733017, + 0.08567005395889282, + 0.11958888918161392 + ], + "credit_map": [ + 0.012667074799537659, + 0.033787451684474945, + 0.01458414550870657, + 0.0439814068377018, + 0.029341669753193855, + 0.027800867334008217, + 0.03667644038796425, + 0.014153567142784595, + 0.005654926411807537, + 0.004716488532721996, + 0.003590809879824519, + 0.00697947246953845, + 0.00919555313885212, + 0.005304696504026651, + 0.006793609820306301, + 0.002347303321585059, + 0.004051098600029945, + 0.004714278504252434, + 0.005203097127377987, + 0.005829618312418461, + 0.004143423866480589, + 0.0029723751358687878, + 0.005897936411201954, + 0.007022243924438953, + 0.01642255112528801, + 0.010649295523762703, + 0.0487975999712944, + 0.0180674996227026, + 0.03501598536968231, + 0.05079455301165581, + 0.016859693452715874, + 0.0371268056333065, + 0.00365633238106966, + 0.003318779869005084, + 0.03387465700507164, + 0.005875215400010347, + 0.0034856463316828012, + 0.0035950299352407455, + 0.008020449429750443, + 0.021245280280709267, + 0.004982122220098972, + 0.009029241278767586, + 0.009710455313324928, + 0.00807652622461319, + 0.00733774621039629, + 0.005267463158816099, + 0.008329484611749649, + 0.007416804786771536, + 0.0075940582901239395, + 0.004864695481956005, + 0.011812039650976658, + 0.0072696516290307045, + 0.005601578392088413, + 0.007220451720058918, + 0.01077917031943798, + 0.009831408970057964, + 0.008373687975108624, + 0.004666952416300774, + 0.008904277347028255, + 0.00818960927426815, + 0.004128785338252783, + 0.005787648726254702, + 0.008980928920209408, + 0.006371328141540289, + 0.00585155189037323, + 0.012112277559936047, + 0.01770040951669216, + 0.012554123997688293, + 0.008022296242415905, + 0.003926546778529882, + 0.019038217142224312, + 0.006464625708758831, + 0.01039492804557085, + 0.01576954498887062, + 0.015529968775808811, + 0.016472715884447098, + 0.020666958764195442, + 0.007604614831507206, + 0.016373032703995705, + 0.016777122393250465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.016074232757091522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095382057130337 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095382057130337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095382057130337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20960848033428192, + 0.06000218167901039, + 0.05998513475060463, + 0.13115382194519043, + 0.06849264353513718, + 0.04378156736493111, + 0.09545004367828369, + 0.09956429898738861, + 0.08825013041496277, + 0.143711656332016 + ], + "credit_map": [ + 0.020497066900134087, + 0.02472814917564392, + 0.02084016241133213, + 0.03056425042450428, + 0.03930550813674927, + 0.0309562087059021, + 0.026262089610099792, + 0.01645505800843239, + 0.005803778301924467, + 0.004938086029142141, + 0.005799294449388981, + 0.008039696142077446, + 0.009213642217218876, + 0.007530308328568935, + 0.011287365108728409, + 0.007390008307993412, + 0.006778639741241932, + 0.004581511486321688, + 0.009516352787613869, + 0.007615118753165007, + 0.006569518707692623, + 0.005812934599816799, + 0.008928090333938599, + 0.01018296368420124, + 0.007949158549308777, + 0.0107351616024971, + 0.032969534397125244, + 0.012329877354204655, + 0.012344719842076302, + 0.018153658136725426, + 0.01061251387000084, + 0.02605920284986496, + 0.005277085583657026, + 0.006771511398255825, + 0.016888583078980446, + 0.007232731208205223, + 0.008026834577322006, + 0.005159958731383085, + 0.010095035657286644, + 0.009040904231369495, + 0.003845303086563945, + 0.007005404680967331, + 0.005422081798315048, + 0.00502962339669466, + 0.004959960002452135, + 0.006892996374517679, + 0.005358079914003611, + 0.005268121138215065, + 0.007769309915602207, + 0.0077016884461045265, + 0.020518707111477852, + 0.01726582832634449, + 0.005157215986400843, + 0.005670592654496431, + 0.017076708376407623, + 0.014289992861449718, + 0.008928315714001656, + 0.016381649300456047, + 0.009949550963938236, + 0.01943192072212696, + 0.011225562542676926, + 0.01120867021381855, + 0.016230223700404167, + 0.006208400242030621, + 0.006352652329951525, + 0.007895631715655327, + 0.024923287332057953, + 0.009205413982272148, + 0.005498984828591347, + 0.005221010651439428, + 0.023297008126974106, + 0.0058561344631016254, + 0.01239536888897419, + 0.02016516402363777, + 0.02140495367348194, + 0.018912797793745995, + 0.021221978589892387, + 0.009484543465077877, + 0.018552586436271667, + 0.021574264392256737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.004095382057130337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184494227170944 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184494227170944, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184494227170944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0779353603720665, + 0.2598043382167816, + 0.04089445248246193, + 0.1579519361257553, + 0.10498341172933578, + 0.02383803017437458, + 0.02447950281202793, + 0.07261938601732254, + 0.1752193719148636, + 0.062274269759655 + ], + "credit_map": [ + 0.008897284045815468, + 0.00999727938324213, + 0.007883057929575443, + 0.010142296552658081, + 0.011915726587176323, + 0.010559389367699623, + 0.011112594977021217, + 0.00742773013189435, + 0.03008074127137661, + 0.028271257877349854, + 0.03453639894723892, + 0.035760298371315, + 0.03874808922410011, + 0.027936480939388275, + 0.039781901985406876, + 0.02468918263912201, + 0.006606067530810833, + 0.00445418618619442, + 0.004480367060750723, + 0.004888469353318214, + 0.004905050154775381, + 0.004864932503551245, + 0.005860589444637299, + 0.004834788851439953, + 0.012220557779073715, + 0.0096803093329072, + 0.010495951399207115, + 0.008986453525722027, + 0.04635996371507645, + 0.04233884438872337, + 0.015875615179538727, + 0.011994238011538982, + 0.010574348270893097, + 0.0077174026519060135, + 0.023577412590384483, + 0.008364969864487648, + 0.009674563072621822, + 0.014016623608767986, + 0.011197608895599842, + 0.019860481843352318, + 0.0029054884798824787, + 0.002544577233493328, + 0.0036996176932007074, + 0.0026742229238152504, + 0.002605279441922903, + 0.0047464678063988686, + 0.0023720432072877884, + 0.002290332457050681, + 0.0019380878657102585, + 0.0027177997399121523, + 0.005163290072232485, + 0.006076654884964228, + 0.00277139525860548, + 0.001307719387114048, + 0.002631395822390914, + 0.001873158267699182, + 0.006891743279993534, + 0.006256367079913616, + 0.011006172746419907, + 0.009792580269277096, + 0.005305347498506308, + 0.008386524394154549, + 0.008531843312084675, + 0.016448812559247017, + 0.01051204465329647, + 0.009476291947066784, + 0.06576031446456909, + 0.009395226836204529, + 0.0061794426292181015, + 0.004406442865729332, + 0.06173701584339142, + 0.00775258569046855, + 0.006370670162141323, + 0.012940499000251293, + 0.010281424969434738, + 0.008984076790511608, + 0.009023811668157578, + 0.0032788184471428394, + 0.005706867203116417, + 0.0056880987249314785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.018184494227170944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072737164795399 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072737164795399, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072737164795399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15376295149326324, + 0.09648402035236359, + 0.034541670233011246, + 0.17115530371665955, + 0.12145598232746124, + 0.06812607496976852, + 0.08214086294174194, + 0.05661822855472565, + 0.044578272849321365, + 0.17113666236400604 + ], + "credit_map": [ + 0.016255894675850868, + 0.020824292674660683, + 0.017050309106707573, + 0.02131728082895279, + 0.023516995832324028, + 0.02125743217766285, + 0.019831325858831406, + 0.013709415681660175, + 0.009307406842708588, + 0.009269343689084053, + 0.007687922567129135, + 0.01597573608160019, + 0.016559036448597908, + 0.013802586123347282, + 0.017165012657642365, + 0.006716975010931492, + 0.0037232371978461742, + 0.003489595139399171, + 0.0035030017606914043, + 0.005313246510922909, + 0.005839427467435598, + 0.004043668508529663, + 0.005828201770782471, + 0.0028012944385409355, + 0.013493271544575691, + 0.006820250302553177, + 0.034376516938209534, + 0.011911706998944283, + 0.028218602761626244, + 0.039153438061475754, + 0.011440624482929707, + 0.025740887969732285, + 0.005609510000795126, + 0.010768861509859562, + 0.03902372345328331, + 0.011528457514941692, + 0.006354358047246933, + 0.005067267455160618, + 0.019880510866642, + 0.02322329767048359, + 0.00695160822942853, + 0.009574593044817448, + 0.009883711114525795, + 0.008210069499909878, + 0.007349142339080572, + 0.010913594625890255, + 0.00904439389705658, + 0.006198957562446594, + 0.006860424764454365, + 0.0072077251970767975, + 0.016905564814805984, + 0.015036946162581444, + 0.008021443150937557, + 0.006021447945386171, + 0.011181777343153954, + 0.010905534960329533, + 0.004535755142569542, + 0.008646713569760323, + 0.008144235238432884, + 0.011791076511144638, + 0.004886058624833822, + 0.005008436273783445, + 0.00958214420825243, + 0.0040238057263195515, + 0.0029111735057085752, + 0.007028725929558277, + 0.009806875139474869, + 0.006379132159054279, + 0.0034745705779641867, + 0.002503950148820877, + 0.009505021385848522, + 0.0029688228387385607, + 0.016414716839790344, + 0.030061136931180954, + 0.028561051934957504, + 0.021808745339512825, + 0.02303195931017399, + 0.009662923403084278, + 0.01901518553495407, + 0.022580942139029503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.011072737164795399 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "gentle downward tilt begins, following and capturing an elegant celadon wine jug as it spins slowly and gracefully in mid-air, seemingly defying gravity. A classic Longquan kiln creation, the jug boasts a long neck and rounded belly. Its glaze is a soft, jade-like pale green, its surface adorned with delicate, natural ice cracks, shimmering with a restrained luster under the soft, diffuse light filtering through the paper window. As the camera descends, the jug's trajectory is precise, until its base, with near-silent grace, touches the deep-grained, richly colored pearwood table, stirring up a barely perceptible puff of dust. In the background, an elegant Song Dynasty study emerges. Through a latticed window, a few sparse, slanting plum blossoms peek through the courtyard, permeating the space with a tranquil, Zen-like atmosphere. The entire picture strictly follows the style of Song Dynasty court painting, with precise composition, elegant colors, even and soft lighting, no strong contrast between light and dark, and emphasis on the outline and texture of objects, creating an ultimate classical aesthetic and a quiet atmosphere of forgetting oneself and the world.", + "seed": 15366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_15/index_6060_seed_15366_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028428797889500857, + "vlm_logprob": -0.028428797889500857, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.028428797889500857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004763564094901085 + ], + "opposite_logprob_tokens": [ + -5.879763603210449 + ], + "logprob_reward": -0.004763564094901085, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004763564094901085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09964358806610107, + 0.05404343083500862, + 0.05483430624008179, + 0.04282553866505623, + 0.13703975081443787, + 0.23987707495689392, + 0.18321377038955688, + 0.02652168460190296, + 0.08457981050014496, + 0.07742100208997726 + ], + "credit_map": [ + 0.007498473860323429, + 0.009558948688209057, + 0.011521613225340843, + 0.016601070761680603, + 0.01365041546523571, + 0.01270297821611166, + 0.018388593569397926, + 0.009721502661705017, + 0.00611080089583993, + 0.0069123744033277035, + 0.007068552076816559, + 0.011106351390480995, + 0.00650417897850275, + 0.004616638645529747, + 0.008376243524253368, + 0.0033482953440397978, + 0.005991909187287092, + 0.008928516879677773, + 0.004193349741399288, + 0.006283932831138372, + 0.010446111671626568, + 0.006127263884991407, + 0.007518661208450794, + 0.005344567354768515, + 0.005042386241257191, + 0.0049458956345915794, + 0.004172213841229677, + 0.004484380129724741, + 0.0097048906609416, + 0.0033505763858556747, + 0.007023785728961229, + 0.004101415630429983, + 0.01701958477497101, + 0.023617926985025406, + 0.011062400415539742, + 0.019654760137200356, + 0.014034806750714779, + 0.02102559432387352, + 0.017087869346141815, + 0.013536826707422733, + 0.018562231212854385, + 0.05057552829384804, + 0.014552758075296879, + 0.056077271699905396, + 0.020349659025669098, + 0.010744795203208923, + 0.02942703664302826, + 0.03958781808614731, + 0.03463984280824661, + 0.018102429807186127, + 0.015935208648443222, + 0.022898871451616287, + 0.017665859311819077, + 0.027359696105122566, + 0.02318175695836544, + 0.02343011647462845, + 0.0025802694726735353, + 0.002887668088078499, + 0.0036237549502402544, + 0.00420679384842515, + 0.002765234326943755, + 0.0024748919531702995, + 0.0036740938667207956, + 0.004308979958295822, + 0.00783038605004549, + 0.01355042029172182, + 0.007980799302458763, + 0.012201945297420025, + 0.010486980900168419, + 0.009340095333755016, + 0.010822328738868237, + 0.012366860173642635, + 0.00728596281260252, + 0.016722489148378372, + 0.005894100293517113, + 0.0075060040690004826, + 0.010369694791734219, + 0.0067201703786849976, + 0.010313799604773521, + 0.012608783319592476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.004763564094901085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805263504385948 + ], + "opposite_logprob_tokens": [ + -3.5298051834106445 + ], + "logprob_reward": -0.029805263504385948, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805263504385948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039365775883197784, + 0.016715094447135925, + 0.011850729584693909, + 0.010721319355070591, + 0.6220807433128357, + 0.09137815982103348, + 0.03948047757148743, + 0.022474907338619232, + 0.057646602392196655, + 0.0882861539721489 + ], + "credit_map": [ + 0.0030209077522158623, + 0.003424249356612563, + 0.003454364137724042, + 0.0039208801463246346, + 0.009459426626563072, + 0.005734867416322231, + 0.0073137045837938786, + 0.0030373777262866497, + 0.0016424329951405525, + 0.0018097509164363146, + 0.0012735602213069797, + 0.0043974146246910095, + 0.0019432207336649299, + 0.0019428940722718835, + 0.002474132226780057, + 0.0012316879583522677, + 0.0012560939649119973, + 0.0014108106261119246, + 0.0010771772358566523, + 0.001410788856446743, + 0.0020812044385820627, + 0.0017099695978686213, + 0.001431113458238542, + 0.0014735725708305836, + 0.0012503907782956958, + 0.001752666081301868, + 0.0018080975860357285, + 0.0011715422151610255, + 0.00148684938903898, + 0.000745913537684828, + 0.0013468828983604908, + 0.0011589773930609226, + 0.07884698361158371, + 0.11394896358251572, + 0.05053155496716499, + 0.08605401962995529, + 0.06200536713004112, + 0.09554911404848099, + 0.07322769612073898, + 0.061917033046483994, + 0.005859586410224438, + 0.01606556586921215, + 0.00599458534270525, + 0.024998635053634644, + 0.0056058745831251144, + 0.004088287707418203, + 0.00977628119289875, + 0.01898934133350849, + 0.004749669693410397, + 0.0038230905774980783, + 0.0032441657967865467, + 0.006493132095783949, + 0.003725296352058649, + 0.003968584816902876, + 0.007441957015544176, + 0.00603457959368825, + 0.001380715286359191, + 0.002458063652738929, + 0.0032671939115971327, + 0.0034520139452069998, + 0.0026800541672855616, + 0.002068591071292758, + 0.0033483945298939943, + 0.0038198810070753098, + 0.006245314609259367, + 0.015981560572981834, + 0.004499200265854597, + 0.004167709033936262, + 0.007159184664487839, + 0.005985000636428595, + 0.004725874867290258, + 0.00888275820761919, + 0.008542999625205994, + 0.025569312274456024, + 0.00820006150752306, + 0.008280220441520214, + 0.011042464524507523, + 0.007418734487146139, + 0.006544788368046284, + 0.012687569484114647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.029805263504385948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952369887381792 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952369887381792, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952369887381792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06800919771194458, + 0.036089248955249786, + 0.030777407810091972, + 0.04567711055278778, + 0.0564349927008152, + 0.06213488057255745, + 0.04339700937271118, + 0.04356526955962181, + 0.09358184039592743, + 0.5203330516815186 + ], + "credit_map": [ + 0.00683471979573369, + 0.011511022225022316, + 0.01348405797034502, + 0.008832825347781181, + 0.005131629761308432, + 0.007133053615689278, + 0.009313887916505337, + 0.00576799688860774, + 0.004570699296891689, + 0.0048385076224803925, + 0.006097360514104366, + 0.0047595505602657795, + 0.0037059190217405558, + 0.005236112978309393, + 0.004151501227170229, + 0.0027295954059809446, + 0.00333642540499568, + 0.0038463971577584743, + 0.0032156272791326046, + 0.0037738943938165903, + 0.005437607876956463, + 0.0027641740161925554, + 0.0035755608696490526, + 0.004827718250453472, + 0.006076646037399769, + 0.005120143294334412, + 0.0037610132712870836, + 0.00594872934743762, + 0.00974279548972845, + 0.0037401753943413496, + 0.006493477616459131, + 0.004794124513864517, + 0.006666599772870541, + 0.010468586347997189, + 0.004576290491968393, + 0.007835236378014088, + 0.006494654342532158, + 0.008170835673809052, + 0.007051051128655672, + 0.005171731114387512, + 0.0063234916888177395, + 0.011911639012396336, + 0.004614312667399645, + 0.00848049484193325, + 0.00807004701346159, + 0.004287415184080601, + 0.010898872278630733, + 0.0075486041605472565, + 0.00545503431931138, + 0.0044165407307446, + 0.003631938248872757, + 0.00544762285426259, + 0.004797654692083597, + 0.006803075782954693, + 0.005929837469011545, + 0.006915301084518433, + 0.0031638285145163536, + 0.004997014068067074, + 0.004389399196952581, + 0.006533170584589243, + 0.005563219543546438, + 0.00341336359269917, + 0.008300983346998692, + 0.007204288616776466, + 0.008334346115589142, + 0.02265555039048195, + 0.005775912664830685, + 0.010264664888381958, + 0.011030583642423153, + 0.007920668460428715, + 0.00961779523640871, + 0.01798231154680252, + 0.052599452435970306, + 0.1633836328983307, + 0.0482386089861393, + 0.034894052892923355, + 0.05303521826863289, + 0.04348394274711609, + 0.04093809053301811, + 0.08376006036996841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.005952369887381792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908308297395706 + ], + "opposite_logprob_tokens": [ + -2.464083194732666 + ], + "logprob_reward": -0.08908308297395706, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908308297395706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11951673775911331, + 0.06421110779047012, + 0.053374823182821274, + 0.059625692665576935, + 0.2553514540195465, + 0.06344924122095108, + 0.11285994201898575, + 0.049230437725782394, + 0.032199859619140625, + 0.1901807188987732 + ], + "credit_map": [ + 0.011295288801193237, + 0.014731785282492638, + 0.01534329541027546, + 0.015561440959572792, + 0.01738724112510681, + 0.01541191153228283, + 0.016757700592279434, + 0.013028069399297237, + 0.008769231848418713, + 0.008899320848286152, + 0.01126361545175314, + 0.00942307524383068, + 0.00690059969201684, + 0.007075953762978315, + 0.00729436706751585, + 0.004584945272654295, + 0.005558369215577841, + 0.007439649663865566, + 0.004896480590105057, + 0.006063381675630808, + 0.009134123101830482, + 0.00521551538258791, + 0.007580655626952648, + 0.007486650720238686, + 0.007728675380349159, + 0.00703025795519352, + 0.006231781095266342, + 0.005670176353305578, + 0.012176492251455784, + 0.006819970440119505, + 0.007264578249305487, + 0.006703759077936411, + 0.03418955206871033, + 0.044222138822078705, + 0.02157614193856716, + 0.03563155606389046, + 0.024936281144618988, + 0.04107598215341568, + 0.02790520153939724, + 0.025814590975642204, + 0.003287526546046138, + 0.007368180435150862, + 0.004657939542084932, + 0.017132652923464775, + 0.00565787497907877, + 0.0030602130573242903, + 0.009078568778932095, + 0.01320628635585308, + 0.020642194896936417, + 0.00753329461440444, + 0.009690291248261929, + 0.012363341636955738, + 0.013826432637870312, + 0.018479496240615845, + 0.015030286274850368, + 0.015294606797397137, + 0.0037196993362158537, + 0.004639363847672939, + 0.0064923400059342384, + 0.007024068385362625, + 0.005789291113615036, + 0.006679263431578875, + 0.008706541731953621, + 0.006179869640618563, + 0.0028907908126711845, + 0.003108361503109336, + 0.004101164638996124, + 0.004449510481208563, + 0.004077273420989513, + 0.003570455824956298, + 0.004138079471886158, + 0.0058642225340008736, + 0.019598474726080894, + 0.0600690022110939, + 0.017773931846022606, + 0.012777848169207573, + 0.018398914486169815, + 0.015659073367714882, + 0.015588228590786457, + 0.03031524270772934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.08908308297395706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539708986878395 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539708986878395, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539708986878395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12097567319869995, + 0.035992227494716644, + 0.013699076138436794, + 0.05073019489645958, + 0.03391587361693382, + 0.14728040993213654, + 0.03509581834077835, + 0.041742414236068726, + 0.08405549079179764, + 0.4365127980709076 + ], + "credit_map": [ + 0.007194232195615768, + 0.014399769715964794, + 0.0118464520201087, + 0.01846734993159771, + 0.024207018315792084, + 0.015607916750013828, + 0.022307541221380234, + 0.006945401895791292, + 0.0074790650978684425, + 0.005773847922682762, + 0.0058105201460421085, + 0.003993551712483168, + 0.002706058556213975, + 0.004527073819190264, + 0.003364959731698036, + 0.002337152138352394, + 0.0014747347449883819, + 0.001263732323423028, + 0.0010343166068196297, + 0.002214631764218211, + 0.0030825596768409014, + 0.0015113449189811945, + 0.0017995252273976803, + 0.0013182308757677674, + 0.0053067454136908054, + 0.005687932949513197, + 0.004929131828248501, + 0.007133091799914837, + 0.010347585193812847, + 0.0033630202524363995, + 0.007664224598556757, + 0.00629846379160881, + 0.004547137767076492, + 0.005181856919080019, + 0.0025898138992488384, + 0.004247395787388086, + 0.004483402241021395, + 0.003690257202833891, + 0.0051128072664141655, + 0.0040632025338709354, + 0.014182394370436668, + 0.03255104646086693, + 0.010487526655197144, + 0.020647605881094933, + 0.018587058410048485, + 0.008520323783159256, + 0.016780609264969826, + 0.025523841381072998, + 0.0052191284485161304, + 0.0026256106793880463, + 0.0038128583692014217, + 0.004333137068897486, + 0.0037832511588931084, + 0.003598499344661832, + 0.005145803093910217, + 0.006577526219189167, + 0.002558484673500061, + 0.005493163131177425, + 0.004761073738336563, + 0.007599801290780306, + 0.005087795201689005, + 0.0037465498317033052, + 0.0071557494811713696, + 0.005339795723557472, + 0.006228311453014612, + 0.019799163565039635, + 0.005508267320692539, + 0.009813875891268253, + 0.00922606885433197, + 0.00803644023835659, + 0.007963702082633972, + 0.017479658126831055, + 0.044733379036188126, + 0.1371249109506607, + 0.04111476242542267, + 0.030599946156144142, + 0.04302598163485527, + 0.03653683885931969, + 0.0343509204685688, + 0.06902605295181274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.012539708986878395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "gentle downward tilt begins, following and capturing an elegant celadon wine jug as it spins slowly and gracefully in mid-air, seemingly defying gravity. A classic Longquan kiln creation, the jug boasts a long neck and rounded belly. Its glaze is a soft, jade-like pale green, its surface adorned with delicate, natural ice cracks, shimmering with a restrained luster under the soft, diffuse light filtering through the paper window. As the camera descends, the jug's trajectory is precise, until its base, with near-silent grace, touches the deep-grained, richly colored pearwood table, stirring up a barely perceptible puff of dust. In the background, an elegant Song Dynasty study emerges. Through a latticed window, a few sparse, slanting plum blossoms peek through the courtyard, permeating the space with a tranquil, Zen-like atmosphere. The entire picture strictly follows the style of Song Dynasty court painting, with precise composition, elegant colors, even and soft lighting, no strong contrast between light and dark, and emphasis on the outline and texture of objects, creating an ultimate classical aesthetic and a quiet atmosphere of forgetting oneself and the world.", + "seed": 15367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_15/index_6060_seed_15367_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.034694074653089046, + "vlm_logprob": -0.034694074653089046, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.034694074653089046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0051378123462200165 + ], + "opposite_logprob_tokens": [ + -5.755137920379639 + ], + "logprob_reward": -0.0051378123462200165, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0051378123462200165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08054564893245697, + 0.13027890026569366, + 0.0429675318300724, + 0.12146065384149551, + 0.03681352734565735, + 0.23598657548427582, + 0.05070294439792633, + 0.14216455817222595, + 0.042149562388658524, + 0.1169300451874733 + ], + "credit_map": [ + 0.007825516164302826, + 0.013980978168547153, + 0.010562868788838387, + 0.008974660187959671, + 0.008560912683606148, + 0.01260567456483841, + 0.008064880035817623, + 0.009970166720449924, + 0.017679067328572273, + 0.01919880509376526, + 0.0107810627669096, + 0.020266355946660042, + 0.021383963525295258, + 0.011293701827526093, + 0.022026920691132545, + 0.007649038918316364, + 0.00566024100407958, + 0.005409781355410814, + 0.006049059797078371, + 0.007951930165290833, + 0.0035231562796980143, + 0.003986469469964504, + 0.007111487910151482, + 0.003275409108027816, + 0.007461552042514086, + 0.011497964151203632, + 0.005865505896508694, + 0.025468099862337112, + 0.01823994144797325, + 0.009898071177303791, + 0.02985701523721218, + 0.013172524981200695, + 0.003667548531666398, + 0.003409468336030841, + 0.004153699614107609, + 0.00529382424429059, + 0.0038861853536218405, + 0.008425634354352951, + 0.005100918933749199, + 0.002876253565773368, + 0.03513999283313751, + 0.08999338746070862, + 0.00969797931611538, + 0.018554195761680603, + 0.03370681777596474, + 0.00905713438987732, + 0.021761877462267876, + 0.018075212836265564, + 0.003910416271537542, + 0.006632640957832336, + 0.006714513059705496, + 0.008889065124094486, + 0.005796489305794239, + 0.0033723560627549887, + 0.008218453265726566, + 0.007169016171246767, + 0.007632827386260033, + 0.01916411705315113, + 0.02579253539443016, + 0.01851709373295307, + 0.011859117075800896, + 0.007658747490495443, + 0.017842164263129234, + 0.03369796276092529, + 0.003421633504331112, + 0.006197984330356121, + 0.005662867333739996, + 0.005221055820584297, + 0.005633658263832331, + 0.003217834047973156, + 0.007279794197529554, + 0.005514740478247404, + 0.011030782014131546, + 0.015181979164481163, + 0.01556587964296341, + 0.015996331349015236, + 0.01531488448381424, + 0.008112109266221523, + 0.014721579849720001, + 0.021006502211093903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.0051378123462200165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003659754991531 + ], + "opposite_logprob_tokens": [ + -2.6950366497039795 + ], + "logprob_reward": -0.07003659754991531, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003659754991531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09912068396806717, + 0.059331342577934265, + 0.025974810123443604, + 0.0914606973528862, + 0.035017721354961395, + 0.0527169406414032, + 0.10670997947454453, + 0.09420698881149292, + 0.0347076952457428, + 0.40075308084487915 + ], + "credit_map": [ + 0.010922348126769066, + 0.012477543205022812, + 0.011370212770998478, + 0.013320058584213257, + 0.01711198315024376, + 0.014679034240543842, + 0.010567083023488522, + 0.008672425523400307, + 0.010949373245239258, + 0.006079140584915876, + 0.008547486737370491, + 0.006118283607065678, + 0.009732875041663647, + 0.006033917423337698, + 0.0071871294640004635, + 0.004683133214712143, + 0.003762908047065139, + 0.003701647976413369, + 0.004285850562155247, + 0.00373983196914196, + 0.002094465307891369, + 0.0043048313818871975, + 0.0022405055351555347, + 0.0018447688780725002, + 0.006187619175761938, + 0.009980461560189724, + 0.006085658445954323, + 0.018222618848085403, + 0.011338158510625362, + 0.007497366052120924, + 0.021679500117897987, + 0.010469318367540836, + 0.003566285129636526, + 0.00394763657823205, + 0.004765485879033804, + 0.004300127737224102, + 0.004738478921353817, + 0.0072103035636246204, + 0.003959260415285826, + 0.002530141733586788, + 0.004795330576598644, + 0.010472114197909832, + 0.004350985400378704, + 0.007829135283827782, + 0.008653713390231133, + 0.0045985993929207325, + 0.006470768712460995, + 0.00554629135876894, + 0.0060055931098759174, + 0.012895052321255207, + 0.01056505274027586, + 0.017808442935347557, + 0.008995848707854748, + 0.006092955358326435, + 0.024744434282183647, + 0.019602596759796143, + 0.00462889950722456, + 0.014842170290648937, + 0.015686986967921257, + 0.01113983616232872, + 0.01012665405869484, + 0.004905538633465767, + 0.011153916828334332, + 0.021722985431551933, + 0.0032250373624265194, + 0.004674149677157402, + 0.00466583576053381, + 0.004258644767105579, + 0.005161916837096214, + 0.003088908502832055, + 0.0050505162216722965, + 0.004582685884088278, + 0.04083652421832085, + 0.05978836119174957, + 0.04715681076049805, + 0.05196911469101906, + 0.04790044203400612, + 0.022296104580163956, + 0.05766291543841362, + 0.07314281910657883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.07003659754991531 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738675758242607 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738675758242607, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738675758242607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10094106942415237, + 0.2647486925125122, + 0.04611124098300934, + 0.12912170588970184, + 0.05063546076416969, + 0.08086108416318893, + 0.06305506080389023, + 0.10998503863811493, + 0.06368935853242874, + 0.0908513218164444 + ], + "credit_map": [ + 0.01154186762869358, + 0.01206483319401741, + 0.01180438045412302, + 0.014307265169918537, + 0.012010923586785793, + 0.017116962000727654, + 0.012225444428622723, + 0.009869392029941082, + 0.04367077350616455, + 0.031142406165599823, + 0.031143810600042343, + 0.030963685363531113, + 0.045132242143154144, + 0.03236095607280731, + 0.035107195377349854, + 0.015227610245347023, + 0.006084761116653681, + 0.005863674916327, + 0.005465459544211626, + 0.007698179222643375, + 0.004441300872713327, + 0.005925905425101519, + 0.006447847466915846, + 0.004184110090136528, + 0.011313525028526783, + 0.010597470216453075, + 0.007818569429218769, + 0.02632642164826393, + 0.016415437683463097, + 0.015399249270558357, + 0.027625329792499542, + 0.013625706546008587, + 0.0062593258917331696, + 0.00602420698851347, + 0.00568222114816308, + 0.00759870233014226, + 0.004465332254767418, + 0.009646537713706493, + 0.005927445366978645, + 0.005031687673181295, + 0.009837380610406399, + 0.027739910408854485, + 0.004554540850222111, + 0.007891060784459114, + 0.01213005930185318, + 0.004412778187543154, + 0.008662865497171879, + 0.005632487591356039, + 0.003587489016354084, + 0.004809372592717409, + 0.010810700245201588, + 0.012583010829985142, + 0.006545356474816799, + 0.004654483404010534, + 0.009685625322163105, + 0.010379022918641567, + 0.007041479926556349, + 0.01419698353856802, + 0.019702883437275887, + 0.01305911224335432, + 0.009308062493801117, + 0.00714512262493372, + 0.013679061084985733, + 0.025852328166365623, + 0.003907257225364447, + 0.0075123123824596405, + 0.007039594929665327, + 0.00843118503689766, + 0.011062584817409515, + 0.003862544195726514, + 0.013391317799687386, + 0.008482559584081173, + 0.005044156685471535, + 0.008357536979019642, + 0.019605042412877083, + 0.006834772415459156, + 0.009008877910673618, + 0.005047092214226723, + 0.02512597106397152, + 0.011827865615487099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.006738675758242607 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901687175035477 + ], + "opposite_logprob_tokens": [ + -2.579016923904419 + ], + "logprob_reward": -0.07901687175035477, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901687175035477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13221962749958038, + 0.31036075949668884, + 0.03696196898818016, + 0.052016645669937134, + 0.04540756717324257, + 0.023966915905475616, + 0.04664679989218712, + 0.1413586437702179, + 0.06348889321088791, + 0.14757224917411804 + ], + "credit_map": [ + 0.013334368355572224, + 0.032249074429273605, + 0.01581614278256893, + 0.012791100889444351, + 0.013929079286754131, + 0.018015332520008087, + 0.011351392604410648, + 0.01473313383758068, + 0.05596603453159332, + 0.03377523645758629, + 0.03428231552243233, + 0.03285883367061615, + 0.05996685475111008, + 0.037353381514549255, + 0.0396726056933403, + 0.016485469415783882, + 0.005021979101002216, + 0.005112063139677048, + 0.004114228300750256, + 0.005591544788330793, + 0.004028955474495888, + 0.004740097094327211, + 0.004723743069916964, + 0.0036293535958975554, + 0.005683105438947678, + 0.004838624969124794, + 0.007881170138716698, + 0.008881804533302784, + 0.00498030474409461, + 0.0069977599196136, + 0.006125523708760738, + 0.006628342904150486, + 0.008493675850331783, + 0.005203235428780317, + 0.008758949115872383, + 0.0058877128176391125, + 0.0031023691408336163, + 0.0042786612175405025, + 0.0060536800883710384, + 0.003629276528954506, + 0.003126918338239193, + 0.00471279164776206, + 0.0028757592663168907, + 0.003342957468703389, + 0.0022338975686579943, + 0.0019185496494174004, + 0.003214762546122074, + 0.002541277091950178, + 0.0036678584292531013, + 0.003055453533306718, + 0.006534477695822716, + 0.00810755230486393, + 0.003195184748619795, + 0.0025751737412065268, + 0.009726370684802532, + 0.009784726426005363, + 0.008956961333751678, + 0.020710209384560585, + 0.021353978663682938, + 0.016457397490739822, + 0.01594621129333973, + 0.009628892876207829, + 0.02064366824924946, + 0.02766129933297634, + 0.007902225479483604, + 0.007025437895208597, + 0.010463486425578594, + 0.0067797270603477955, + 0.00828773993998766, + 0.00418882304802537, + 0.008558675646781921, + 0.010282768867909908, + 0.007565995678305626, + 0.0055585154332220554, + 0.038915060460567474, + 0.007860993966460228, + 0.012693064287304878, + 0.006483378820121288, + 0.05293547734618187, + 0.015559745021164417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.07901687175035477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540415860712528 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.012540415860712528, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540415860712528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10490100830793381, + 0.35383155941963196, + 0.03390975669026375, + 0.07384788990020752, + 0.0442015677690506, + 0.1254417896270752, + 0.03740082308650017, + 0.08389553427696228, + 0.052705686539411545, + 0.08986440300941467 + ], + "credit_map": [ + 0.010489800944924355, + 0.022836480289697647, + 0.011121692135930061, + 0.014123432338237762, + 0.011965017765760422, + 0.014568795450031757, + 0.011557688936591148, + 0.008238100446760654, + 0.043345652520656586, + 0.05152304843068123, + 0.03493070229887962, + 0.05434669554233551, + 0.052059952169656754, + 0.036971334367990494, + 0.059364691376686096, + 0.021289492025971413, + 0.005062753334641457, + 0.0038361139595508575, + 0.00525604048743844, + 0.0030962240416556597, + 0.003712611272931099, + 0.005370764527469873, + 0.003612447762861848, + 0.003962798044085503, + 0.013172517530620098, + 0.005505411420017481, + 0.006181762553751469, + 0.006396444980055094, + 0.0037673725746572018, + 0.009823096916079521, + 0.008502661250531673, + 0.020498620346188545, + 0.005567505955696106, + 0.00434131408110261, + 0.004122598096728325, + 0.006503491662442684, + 0.005622306372970343, + 0.007848951034247875, + 0.006042627617716789, + 0.004152772948145866, + 0.016952401027083397, + 0.045711491256952286, + 0.006050396710634232, + 0.011668024584650993, + 0.019230036064982414, + 0.005709879100322723, + 0.011648313142359257, + 0.00847125705331564, + 0.002619135659188032, + 0.00306421029381454, + 0.0048598963767290115, + 0.007786192931234837, + 0.005218395497649908, + 0.0023447542916983366, + 0.007049191277474165, + 0.004459044896066189, + 0.006321705412119627, + 0.011146710254251957, + 0.007970761507749557, + 0.011350514367222786, + 0.013233667239546776, + 0.006101418752223253, + 0.015172742307186127, + 0.012598006054759026, + 0.0032646465115249157, + 0.00613360432907939, + 0.007402083836495876, + 0.005990182980895042, + 0.007604226004332304, + 0.0032163767609745264, + 0.009552148170769215, + 0.009542417712509632, + 0.008243480697274208, + 0.010104581713676453, + 0.011559014208614826, + 0.012886988930404186, + 0.013159758411347866, + 0.006256141699850559, + 0.014012753963470459, + 0.013641687110066414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.012540415860712528 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "gentle downward tilt begins, following and capturing an elegant celadon wine jug as it spins slowly and gracefully in mid-air, seemingly defying gravity. A classic Longquan kiln creation, the jug boasts a long neck and rounded belly. Its glaze is a soft, jade-like pale green, its surface adorned with delicate, natural ice cracks, shimmering with a restrained luster under the soft, diffuse light filtering through the paper window. As the camera descends, the jug's trajectory is precise, until its base, with near-silent grace, touches the deep-grained, richly colored pearwood table, stirring up a barely perceptible puff of dust. In the background, an elegant Song Dynasty study emerges. Through a latticed window, a few sparse, slanting plum blossoms peek through the courtyard, permeating the space with a tranquil, Zen-like atmosphere. The entire picture strictly follows the style of Song Dynasty court painting, with precise composition, elegant colors, even and soft lighting, no strong contrast between light and dark, and emphasis on the outline and texture of objects, creating an ultimate classical aesthetic and a quiet atmosphere of forgetting oneself and the world.", + "seed": 15368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_15/index_6060_seed_15368_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014202888729050756, + "vlm_logprob": -0.014202888729050756, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.014202888729050756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004432852379977703 + ], + "opposite_logprob_tokens": [ + -6.254432678222656 + ], + "logprob_reward": -0.004432852379977703, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004432852379977703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07349470257759094, + 0.06295610964298248, + 0.03014625608921051, + 0.17525842785835266, + 0.18562157452106476, + 0.18345296382904053, + 0.08977624028921127, + 0.08508053421974182, + 0.07348782569169998, + 0.040725402534008026 + ], + "credit_map": [ + 0.00726340664550662, + 0.009034296497702599, + 0.006167371291667223, + 0.012152709066867828, + 0.015103290788829327, + 0.0069139995612204075, + 0.01152605377137661, + 0.005333562847226858, + 0.009345971047878265, + 0.007204513065516949, + 0.004640464670956135, + 0.008742027916014194, + 0.007637913338840008, + 0.006816445849835873, + 0.009593173861503601, + 0.008975591510534286, + 0.0024185243528336287, + 0.005725285038352013, + 0.003987159114331007, + 0.0027118190191686153, + 0.005752007942646742, + 0.0025263011921197176, + 0.00515148788690567, + 0.0018736679339781404, + 0.01113180909305811, + 0.01892073079943657, + 0.009304937906563282, + 0.030350372195243835, + 0.03152954578399658, + 0.011651516892015934, + 0.04380772262811661, + 0.018561767414212227, + 0.011082868091762066, + 0.009414102882146835, + 0.01496043335646391, + 0.010374837554991245, + 0.04680948704481125, + 0.008213896304368973, + 0.06117413192987442, + 0.023591795936226845, + 0.008020224049687386, + 0.0064185443334281445, + 0.028172003105282784, + 0.008180311881005764, + 0.050690874457359314, + 0.007817551493644714, + 0.058896664530038834, + 0.015256762504577637, + 0.005590684711933136, + 0.010569698177278042, + 0.014260859228670597, + 0.009465809911489487, + 0.014648993499577045, + 0.0038642953149974346, + 0.019865093752741814, + 0.011510795913636684, + 0.007761988788843155, + 0.0078203696757555, + 0.013329435139894485, + 0.008495889604091644, + 0.014625188894569874, + 0.004063894040882587, + 0.016227571293711662, + 0.012756187468767166, + 0.0049507287330925465, + 0.009427815675735474, + 0.008321321569383144, + 0.006625011097639799, + 0.013372994028031826, + 0.003034917637705803, + 0.018341967836022377, + 0.00941306259483099, + 0.002824053168296814, + 0.004937850404530764, + 0.005326222162693739, + 0.0035519094672054052, + 0.00644319923594594, + 0.0028056371957063675, + 0.00917165819555521, + 0.00566486082971096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.004432852379977703 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193624913692474 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193624913692474, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193624913692474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055081360042095184, + 0.10675597190856934, + 0.04605691879987717, + 0.08895975351333618, + 0.2619624733924866, + 0.13954029977321625, + 0.16724488139152527, + 0.07347296178340912, + 0.04253925010561943, + 0.01838614046573639 + ], + "credit_map": [ + 0.004606370814144611, + 0.006059105973690748, + 0.004820127505809069, + 0.0072097270749509335, + 0.011453979648649693, + 0.006356323603540659, + 0.00966673158109188, + 0.004908996634185314, + 0.0069861579686403275, + 0.014636315405368805, + 0.00739416154101491, + 0.022634977474808693, + 0.017131801694631577, + 0.0104800621047616, + 0.019375264644622803, + 0.008117236196994781, + 0.0022114755120128393, + 0.006318497471511364, + 0.0022287212777882814, + 0.006607202347368002, + 0.011474289000034332, + 0.00385228986851871, + 0.010060584172606468, + 0.0033038603141903877, + 0.008410309441387653, + 0.012374191544950008, + 0.006917314138263464, + 0.01236108597368002, + 0.012566769495606422, + 0.008464572951197624, + 0.014707065187394619, + 0.013158442452549934, + 0.018081191927194595, + 0.015624837018549442, + 0.03595851734280586, + 0.017265653237700462, + 0.04983216151595116, + 0.013037054799497128, + 0.0689597874879837, + 0.04320325702428818, + 0.007266901899129152, + 0.00872159842401743, + 0.026374706998467445, + 0.010556564666330814, + 0.03404035046696663, + 0.010519581846892834, + 0.02818458154797554, + 0.013876022771000862, + 0.01960698887705803, + 0.019205689430236816, + 0.020891517400741577, + 0.022686149924993515, + 0.02690570242702961, + 0.00900644063949585, + 0.03365134447813034, + 0.01529105007648468, + 0.007419322617352009, + 0.006869954988360405, + 0.012889332138001919, + 0.00621615769341588, + 0.013101012445986271, + 0.003986425697803497, + 0.010019741952419281, + 0.012971016578376293, + 0.002491458784788847, + 0.0048811729066073895, + 0.00666562095284462, + 0.004623760469257832, + 0.006499805022031069, + 0.0020358453039079905, + 0.007966340519487858, + 0.007375244051218033, + 0.001081110560335219, + 0.0018413447542116046, + 0.0029913713224232197, + 0.001705401111394167, + 0.002817961387336254, + 0.0011544957524165511, + 0.00402815779671073, + 0.0027662983629852533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.018193624913692474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616105066612363 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.003616105066612363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616105066612363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06508838385343552, + 0.049441851675510406, + 0.06034604087471962, + 0.12275268137454987, + 0.38469821214675903, + 0.15741859376430511, + 0.06317312270402908, + 0.03734232857823372, + 0.036427080631256104, + 0.023311715573072433 + ], + "credit_map": [ + 0.005287500564008951, + 0.007973764091730118, + 0.004300269298255444, + 0.012278652749955654, + 0.013463662937283516, + 0.005336924456059933, + 0.012040457688272, + 0.0044071488082408905, + 0.0032947715371847153, + 0.006545080803334713, + 0.0027869686018675566, + 0.00880456157028675, + 0.006937634665518999, + 0.006990823429077864, + 0.009028120897710323, + 0.005053890869021416, + 0.0030408098828047514, + 0.007837209850549698, + 0.005720611661672592, + 0.005603787023574114, + 0.011727506294846535, + 0.011286386288702488, + 0.01080559566617012, + 0.004324135836213827, + 0.005087544675916433, + 0.017068620771169662, + 0.005349964369088411, + 0.01891833357512951, + 0.037642110139131546, + 0.005882991477847099, + 0.024794386699795723, + 0.00800873152911663, + 0.023124761879444122, + 0.01962718553841114, + 0.02789931185543537, + 0.021807890385389328, + 0.09703265130519867, + 0.013402066193521023, + 0.1311609148979187, + 0.05064341425895691, + 0.005142916459590197, + 0.006710731890052557, + 0.018452627584338188, + 0.005956034641712904, + 0.04913055896759033, + 0.004967827349901199, + 0.0521557480096817, + 0.014902148395776749, + 0.006549052894115448, + 0.0071288165636360645, + 0.007131732534617186, + 0.008907264098525047, + 0.01034687552601099, + 0.0032313410192728043, + 0.013853676617145538, + 0.006024361588060856, + 0.003198150312528014, + 0.005056777969002724, + 0.004062652587890625, + 0.0033776552882045507, + 0.0068399254232645035, + 0.002689473098143935, + 0.0087368069216609, + 0.0033808881416916847, + 0.003475163597613573, + 0.0036616104189306498, + 0.003765692235901952, + 0.004105688538402319, + 0.005990501027554274, + 0.0022706410381942987, + 0.009217542596161366, + 0.00394024234265089, + 0.00166686053853482, + 0.0024748535361140966, + 0.0029036561027169228, + 0.002060161903500557, + 0.003885725513100624, + 0.0015256335027515888, + 0.005942595191299915, + 0.0028522294014692307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.003616105066612363 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369971364736557 + ], + "opposite_logprob_tokens": [ + -3.4086997509002686 + ], + "logprob_reward": -0.03369971364736557, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369971364736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13105475902557373, + 0.05842011794447899, + 0.08511585742235184, + 0.11036758869886398, + 0.18627697229385376, + 0.2267274111509323, + 0.061629705131053925, + 0.026406293734908104, + 0.019741613417863846, + 0.09425967186689377 + ], + "credit_map": [ + 0.012705330736935139, + 0.016411826014518738, + 0.00957795511931181, + 0.02342311479151249, + 0.019994201138615608, + 0.01617729850113392, + 0.02238214761018753, + 0.0103828813880682, + 0.006245575845241547, + 0.009763257578015327, + 0.005988761782646179, + 0.007069947198033333, + 0.006741091143339872, + 0.009758259169757366, + 0.007473721168935299, + 0.005379499401897192, + 0.005860757548362017, + 0.010883403941988945, + 0.010166612453758717, + 0.007551996503025293, + 0.01622733660042286, + 0.014085170812904835, + 0.014808960258960724, + 0.00553162069991231, + 0.008489311672747135, + 0.01563839614391327, + 0.006457000505179167, + 0.013542830012738705, + 0.01321986224502325, + 0.006000501103699207, + 0.03097871132194996, + 0.016040978953242302, + 0.015151452273130417, + 0.009993904270231724, + 0.030196532607078552, + 0.01051382627338171, + 0.03115730918943882, + 0.010057495906949043, + 0.044543031603097916, + 0.034663405269384384, + 0.0070752170868217945, + 0.008207589387893677, + 0.03207056596875191, + 0.008819579146802425, + 0.06569357216358185, + 0.00845749955624342, + 0.07000681757926941, + 0.02639657072722912, + 0.007149023935198784, + 0.005551367532461882, + 0.007020739838480949, + 0.007746003568172455, + 0.009606433100998402, + 0.005926024168729782, + 0.012384874746203423, + 0.006245233118534088, + 0.0022950454149395227, + 0.002033878117799759, + 0.004479455761611462, + 0.002222453709691763, + 0.004385102074593306, + 0.002570754149928689, + 0.004516432993113995, + 0.003903171746060252, + 0.0019329189090058208, + 0.001561379642225802, + 0.0037532481364905834, + 0.0015460038557648659, + 0.003077475819736719, + 0.0014553613727912307, + 0.0026338703464716673, + 0.0037813547533005476, + 0.01390809752047062, + 0.010438409633934498, + 0.013261528685688972, + 0.007866896688938141, + 0.012761566787958145, + 0.008302210830152035, + 0.015510737895965576, + 0.012210225686430931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.03369971364736557 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a celadon wine jug with a long neck and rounded belly visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jug spinning slowly in mid-air before touching a pearwood table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jug feature a soft jade-like pale green glaze with delicate ice cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an elegant Song Dynasty study with a latticed window showing plum blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Song Dynasty court painting, featuring soft diffuse light and no strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072147637605667 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072147637605667, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072147637605667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12498980760574341, + 0.057356469333171844, + 0.0699942335486412, + 0.052296318113803864, + 0.45239126682281494, + 0.04903194680809975, + 0.07797316461801529, + 0.04850999638438225, + 0.029927924275398254, + 0.03752894327044487 + ], + "credit_map": [ + 0.011689689010381699, + 0.015771087259054184, + 0.0073544918559491634, + 0.023460399359464645, + 0.024446791037917137, + 0.010144012980163097, + 0.02272077091038227, + 0.009402565658092499, + 0.00541291106492281, + 0.011169820092618465, + 0.004162691067904234, + 0.0057239048182964325, + 0.007746437564492226, + 0.010377050377428532, + 0.009080030024051666, + 0.003683624556288123, + 0.006072180811315775, + 0.006596859078854322, + 0.004051053896546364, + 0.009898359887301922, + 0.01493956707417965, + 0.0071953595615923405, + 0.014047364704310894, + 0.007193492725491524, + 0.004888423718512058, + 0.005935142748057842, + 0.003754941513761878, + 0.008059331215918064, + 0.007717076223343611, + 0.002673041308298707, + 0.012591680511832237, + 0.006676679011434317, + 0.02633313089609146, + 0.022306839004158974, + 0.03217938914895058, + 0.02678992971777916, + 0.11670937389135361, + 0.015607691369950771, + 0.15461981296539307, + 0.0578450933098793, + 0.00273215607739985, + 0.0029209950007498264, + 0.00904682744294405, + 0.0026394850574433804, + 0.009333268739283085, + 0.00235912692733109, + 0.011268120259046555, + 0.008731968700885773, + 0.009585101157426834, + 0.0079768942669034, + 0.009574230760335922, + 0.01000859122723341, + 0.012206712737679482, + 0.005804950837045908, + 0.014058186672627926, + 0.008758495561778545, + 0.0052965921349823475, + 0.006869554985314608, + 0.007700871676206589, + 0.003665399504825473, + 0.00790055189281702, + 0.004156875424087048, + 0.007231733296066523, + 0.005688420031219721, + 0.002529639983549714, + 0.0021106558851897717, + 0.00457761762663722, + 0.002502265153452754, + 0.0048217615112662315, + 0.002205067314207554, + 0.00628275703638792, + 0.004898160230368376, + 0.004290826618671417, + 0.0033281417563557625, + 0.005169783253222704, + 0.003597063012421131, + 0.005270787049084902, + 0.0031934105791151524, + 0.00764311570674181, + 0.005035814829170704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.011072147637605667 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_tensor([6957, 6957])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_tensor([6957, 6957])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..24a85fc4c725af8905a6a7ef61d6d4bb4e1107ad --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_tensor([6957, 6957])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angle close-up captures a modern, minimalist gooseneck table lamp resting on a lightly textured wooden surface. Finished in a uniformly smooth matte black, the lamp features a slender, flexible neck that gracefully transitions from a heavy, round base to a simple, disc-shaped LED head. When on, the lamp casts a focused, circular beam of bright, cool white light onto the table, highlighting the subtle grain of the wood. Following a faint, invisible surface vibration, the lamp's head and upper neck begin to vibrate gently from side to side. The movement is fluid and subtle, a brief, silent dance that rapidly diminishes as the amplitude of the vibration steadily decreases, until, in an instant, all movement ceases and the lamp returns to a perfect state of stillness, its elegant form a stark, silent silhouette against the softly lit, out-of-focus background of the quiet study. The atmosphere is one of focused tranquility and minimalist calm, the fleeting movements only emphasizing the scene's profound stillness.", + "seed": 16210, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_6957_seed_16210_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040898833563551305, + "vlm_logprob": -0.040898833563551305, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.040898833563551305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00914787407964468 + ], + "opposite_logprob_tokens": [ + -6.509147644042969 + ], + "logprob_reward": -0.00914787407964468, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00914787407964468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11961202323436737, + 0.09405866265296936, + 0.0761740654706955, + 0.04722458869218826, + 0.09565858542919159, + 0.26750755310058594, + 0.13707056641578674, + 0.10111125558614731, + 0.04456473886966705, + 0.017017971724271774 + ], + "credit_map": [ + 0.023859219625592232, + 0.00707707554101944, + 0.02981242723762989, + 0.010614021681249142, + 0.00849380623549223, + 0.021146826446056366, + 0.00819548312574625, + 0.010413164272904396, + 0.013737094588577747, + 0.008712022565305233, + 0.020611625164747238, + 0.009726411662995815, + 0.006103808060288429, + 0.015485226176679134, + 0.006007440388202667, + 0.013675033114850521, + 0.014734824188053608, + 0.007010486908257008, + 0.01818246580660343, + 0.0065167793072760105, + 0.00603587506338954, + 0.010098136030137539, + 0.005612769164144993, + 0.00798272155225277, + 0.01522282138466835, + 0.0033332977909594774, + 0.006529765669256449, + 0.002619961742311716, + 0.0019905180670320988, + 0.003995584324002266, + 0.002199497539550066, + 0.011333142407238483, + 0.01571844518184662, + 0.007624237332493067, + 0.017939168959856033, + 0.006609669886529446, + 0.006831618491560221, + 0.022760571911931038, + 0.005358709488064051, + 0.01281615998595953, + 0.059651535004377365, + 0.016260452568531036, + 0.04035048186779022, + 0.014627278782427311, + 0.012970895506441593, + 0.04406601935625076, + 0.012245512567460537, + 0.06733540445566177, + 0.04505130648612976, + 0.006317794322967529, + 0.017885953187942505, + 0.004224318079650402, + 0.003616398898884654, + 0.012318617664277554, + 0.005097812507301569, + 0.04255836457014084, + 0.036316607147455215, + 0.004270208068192005, + 0.012830612249672413, + 0.0024592617992311716, + 0.0021620807237923145, + 0.009507394395768642, + 0.003074166364967823, + 0.030490931123495102, + 0.006345248315483332, + 0.0037599296774715185, + 0.0073061599396169186, + 0.0044240765273571014, + 0.0033732068259269, + 0.007076486945152283, + 0.00391607778146863, + 0.008363551460206509, + 0.002452484332025051, + 0.0015563301276415586, + 0.002776277717202902, + 0.001038836780935526, + 0.0017890633316710591, + 0.003687764285132289, + 0.0014023430412635207, + 0.00231487350538373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.00914787407964468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647494949400425 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647494949400425, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647494949400425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14264750480651855, + 0.03952736780047417, + 0.01950482837855816, + 0.010075797326862812, + 0.07450135052204132, + 0.022171927616000175, + 0.13425032794475555, + 0.49334266781806946, + 0.015718771144747734, + 0.048259422183036804 + ], + "credit_map": [ + 0.021795958280563354, + 0.01030413806438446, + 0.03175852820277214, + 0.010123503394424915, + 0.009850339964032173, + 0.028984583914279938, + 0.00850051548331976, + 0.021329939365386963, + 0.0068436674773693085, + 0.002488625468686223, + 0.008211496286094189, + 0.0031184873078018427, + 0.0018700980581343174, + 0.005827472545206547, + 0.0027400748804211617, + 0.008427450433373451, + 0.005813912022858858, + 0.0012145215878263116, + 0.004415432456880808, + 0.00123489904217422, + 0.0007857718737795949, + 0.0018775302451103926, + 0.0011518842075020075, + 0.003010879037901759, + 0.0024687121622264385, + 0.0006139784236438572, + 0.0012924126349389553, + 0.0007531522423960268, + 0.00046462248428724706, + 0.0009584875078871846, + 0.0006566413794644177, + 0.002867792034521699, + 0.01781640760600567, + 0.004585811868309975, + 0.016899457201361656, + 0.004911010153591633, + 0.003688995959237218, + 0.010958070866763592, + 0.0038911947049200535, + 0.011750409379601479, + 0.005861442070454359, + 0.0010285390308126807, + 0.0034608172718435526, + 0.0007843063212931156, + 0.0007620758260600269, + 0.0030048058833926916, + 0.0008507011225447059, + 0.006419242825359106, + 0.04387465864419937, + 0.006133733317255974, + 0.016557279974222183, + 0.005278669763356447, + 0.004083578009158373, + 0.011219289153814316, + 0.005897287745028734, + 0.041205838322639465, + 0.18255512416362762, + 0.021508632227778435, + 0.06311028450727463, + 0.012551571242511272, + 0.009423894807696342, + 0.03789214789867401, + 0.015195507556200027, + 0.15110549330711365, + 0.0031532878056168556, + 0.0007802567561157048, + 0.0033367271535098553, + 0.0009207159746438265, + 0.0006443441961891949, + 0.0017318599857389927, + 0.0009474277612753212, + 0.00420415261760354, + 0.00867943000048399, + 0.002767282770946622, + 0.009672753512859344, + 0.002942649181932211, + 0.0023666389752179384, + 0.006098765414208174, + 0.0034512511920183897, + 0.012280660681426525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.008647494949400425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203025320544839 + ], + "opposite_logprob_tokens": [ + -5.75320291519165 + ], + "logprob_reward": -0.003203025320544839, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203025320544839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06352845579385757, + 0.07853726297616959, + 0.21976812183856964, + 0.10285675525665283, + 0.15196795761585236, + 0.08202993869781494, + 0.04973980784416199, + 0.11655957251787186, + 0.03543860465288162, + 0.099573515355587 + ], + "credit_map": [ + 0.014182751066982746, + 0.004356035962700844, + 0.015637662261724472, + 0.004649174865335226, + 0.0037637008354067802, + 0.010468553751707077, + 0.004106238018721342, + 0.0063643380999565125, + 0.01386204082518816, + 0.005636724177747965, + 0.016116756945848465, + 0.006773932371288538, + 0.004933769349008799, + 0.015720048919320107, + 0.007967730984091759, + 0.007526256609708071, + 0.06264794617891312, + 0.014574447646737099, + 0.04639384150505066, + 0.013989460654556751, + 0.008404538035392761, + 0.02410816214978695, + 0.01123070903122425, + 0.03841901570558548, + 0.031198769807815552, + 0.005338774062693119, + 0.01787620037794113, + 0.005140152294188738, + 0.003018365241587162, + 0.010161664336919785, + 0.005337403155863285, + 0.024785421788692474, + 0.03284544497728348, + 0.009712129831314087, + 0.033903930336236954, + 0.009337764233350754, + 0.008205763064324856, + 0.029533561319112778, + 0.008235742338001728, + 0.020193617790937424, + 0.017102226614952087, + 0.005082882009446621, + 0.013133361004292965, + 0.0047246916219592094, + 0.004309551324695349, + 0.014800108037889004, + 0.0037256244104355574, + 0.01915149763226509, + 0.01625313051044941, + 0.002183883683755994, + 0.005806455388665199, + 0.0021976009011268616, + 0.0018530430970713496, + 0.004737436305731535, + 0.00247667427174747, + 0.01423158310353756, + 0.04356560483574867, + 0.0048979101702570915, + 0.01484688464552164, + 0.002746257930994034, + 0.0020219506695866585, + 0.009163402952253819, + 0.003483803477138281, + 0.03583375737071037, + 0.0044968691654503345, + 0.002645599190145731, + 0.008610352873802185, + 0.002508594421669841, + 0.0019442630000412464, + 0.004593881778419018, + 0.002603633562102914, + 0.008035412058234215, + 0.016072073951363564, + 0.006809796206653118, + 0.020328499376773834, + 0.006814694497734308, + 0.006042883265763521, + 0.01437161210924387, + 0.0068621342070400715, + 0.022271817550063133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.003203025320544839 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988049983978271 + ], + "opposite_logprob_tokens": [ + -1.8048804998397827 + ], + "logprob_reward": -0.17988049983978271, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988049983978271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11214646697044373, + 0.07303345203399658, + 0.014399577863514423, + 0.15515096485614777, + 0.07535991072654724, + 0.042882394045591354, + 0.32505273818969727, + 0.17293672263622284, + 0.01275950949639082, + 0.016278238967061043 + ], + "credit_map": [ + 0.02083633467555046, + 0.007835413329303265, + 0.02022663503885269, + 0.008173109963536263, + 0.007989257574081421, + 0.021889083087444305, + 0.0071234689094126225, + 0.01807316578924656, + 0.021906452253460884, + 0.004085906781256199, + 0.01298128254711628, + 0.004814370535314083, + 0.0032396498136222363, + 0.00871849711984396, + 0.0045260777696967125, + 0.01276121661067009, + 0.002336829202249646, + 0.0013817192520946264, + 0.0029905179981142282, + 0.0009548740345053375, + 0.0010076200123876333, + 0.002343609696254134, + 0.0012914539547637105, + 0.0020929542370140553, + 0.04605308175086975, + 0.00942966528236866, + 0.028470102697610855, + 0.010102443397045135, + 0.0054861679673194885, + 0.0133403604850173, + 0.007269491441547871, + 0.034999649971723557, + 0.019376276060938835, + 0.00448651472106576, + 0.016380231827497482, + 0.004285469651222229, + 0.004001609981060028, + 0.01040224265307188, + 0.004335279110819101, + 0.01209228578954935, + 0.00954064354300499, + 0.003127525094896555, + 0.007802380248904228, + 0.00273521919734776, + 0.0026111439801752567, + 0.006274822633713484, + 0.0027701843064278364, + 0.008020474575459957, + 0.10803984105587006, + 0.014439116232097149, + 0.039263539016246796, + 0.012223680503666401, + 0.009245835244655609, + 0.027261817827820778, + 0.014075648970901966, + 0.10050326585769653, + 0.0650496780872345, + 0.007260279729962349, + 0.020766373723745346, + 0.004175929818302393, + 0.003583339275792241, + 0.015834014862775803, + 0.005354418884962797, + 0.050912681967020035, + 0.0020539439283311367, + 0.0010135185439139605, + 0.0021532834507524967, + 0.0007590852910652757, + 0.0010357139399275184, + 0.0025920423213392496, + 0.0014251614920794964, + 0.0017267607618123293, + 0.0036488622426986694, + 0.0011565330205485225, + 0.0025714244693517685, + 0.0009647739352658391, + 0.0008972507202997804, + 0.0025151041336357594, + 0.0013261995045468211, + 0.0031980916392058134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.17988049983978271 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615273628383875 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615273628383875, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615273628383875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12997667491436005, + 0.0735597088932991, + 0.057730186730623245, + 0.05601406842470169, + 0.12597084045410156, + 0.03553532063961029, + 0.35095274448394775, + 0.12526953220367432, + 0.016608763486146927, + 0.028382087126374245 + ], + "credit_map": [ + 0.01976032741367817, + 0.008795579895377159, + 0.02706805244088173, + 0.007389511447399855, + 0.008091907016932964, + 0.028068330138921738, + 0.007333195302635431, + 0.02346978150308132, + 0.020715219900012016, + 0.004372536204755306, + 0.011696324683725834, + 0.005619647912681103, + 0.0035828864201903343, + 0.008278202265501022, + 0.005928108002990484, + 0.013366784900426865, + 0.010258261114358902, + 0.0050970930606126785, + 0.01509249396622181, + 0.006412812974303961, + 0.003235156647861004, + 0.00882469117641449, + 0.003868683474138379, + 0.004940997809171677, + 0.01430097408592701, + 0.003580502700060606, + 0.007661197334527969, + 0.004635739140212536, + 0.002673500217497349, + 0.005498817190527916, + 0.003597427159547806, + 0.014065911993384361, + 0.03696234151721001, + 0.00657543633133173, + 0.023555899038910866, + 0.00684801721945405, + 0.00497299013659358, + 0.011555836535990238, + 0.006599280517548323, + 0.02890104241669178, + 0.00467611663043499, + 0.0034474043641239405, + 0.007946306839585304, + 0.0035510114394128323, + 0.0034142553340643644, + 0.006022751796990633, + 0.003405373077839613, + 0.003072102088481188, + 0.11674880981445312, + 0.01652044616639614, + 0.042941056191921234, + 0.012797437608242035, + 0.00954076275229454, + 0.029100002720952034, + 0.013575251214206219, + 0.10972901433706284, + 0.04770674183964729, + 0.0052560921758413315, + 0.015413239598274231, + 0.003038200084120035, + 0.0023412134032696486, + 0.010172763839364052, + 0.0034538726322352886, + 0.03788742050528526, + 0.003844005987048149, + 0.0014851228334009647, + 0.002447145991027355, + 0.0009064472396858037, + 0.0013110693544149399, + 0.0029147693421691656, + 0.0012716013006865978, + 0.0024286031257361174, + 0.004717991687357426, + 0.0016220235265791416, + 0.005677336826920509, + 0.0018844607984647155, + 0.0018304475815966725, + 0.004101198632270098, + 0.0027848633471876383, + 0.005763766821473837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.003615273628383875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a modern, minimalist gooseneck table lamp resting on a lightly textured wooden surface. Finished in a uniformly smooth matte black, the lamp features a slender, flexible neck that gracefully transitions from a heavy, round base to a simple, disc-shaped LED head. When on, the lamp casts a focused, circular beam of bright, cool white light onto the table, highlighting the subtle grain of the wood. Following a faint, invisible surface vibration, the lamp's head and upper neck begin to vibrate gently from side to side. The movement is fluid and subtle, a brief, silent dance that rapidly diminishes as the amplitude of the vibration steadily decreases, until, in an instant, all movement ceases and the lamp returns to a perfect state of stillness, its elegant form a stark, silent silhouette against the softly lit, out-of-focus background of the quiet study. The atmosphere is one of focused tranquility and minimalist calm, the fleeting movements only emphasizing the scene's profound stillness.", + "seed": 16211, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_6957_seed_16211_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03692807792685926, + "vlm_logprob": -0.03692807792685926, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.03692807792685926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006111979018896818 + ], + "opposite_logprob_tokens": [ + -6.006112098693848 + ], + "logprob_reward": -0.006111979018896818, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.006111979018896818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2840907871723175, + 0.1140999123454094, + 0.056532274931669235, + 0.08562706410884857, + 0.048327650874853134, + 0.0851159319281578, + 0.1424405425786972, + 0.062118612229824066, + 0.07990800589323044, + 0.04173921421170235 + ], + "credit_map": [ + 0.02856813743710518, + 0.03825851529836655, + 0.06153801456093788, + 0.021886367350816727, + 0.044582299888134, + 0.04240470379590988, + 0.02859710529446602, + 0.018255626782774925, + 0.011150995269417763, + 0.020288558676838875, + 0.015931719914078712, + 0.0076240310445427895, + 0.024103617295622826, + 0.015354695729911327, + 0.013006984256207943, + 0.00663930457085371, + 0.006145010236650705, + 0.009927080012857914, + 0.008349359035491943, + 0.0033646938391029835, + 0.008528782054781914, + 0.008679525926709175, + 0.0075940401293337345, + 0.003943784162402153, + 0.009594987146556377, + 0.013488433323800564, + 0.01748817041516304, + 0.006601647008210421, + 0.009533184580504894, + 0.01251305639743805, + 0.006094713695347309, + 0.010312872938811779, + 0.005267839413136244, + 0.007137998007237911, + 0.011443219147622585, + 0.0028128072153776884, + 0.006762431934475899, + 0.007173980586230755, + 0.00442203413695097, + 0.0033073395024985075, + 0.009888820350170135, + 0.009570359252393246, + 0.018133917823433876, + 0.0073459213599562645, + 0.01225664559751749, + 0.012432051822543144, + 0.0085158571600914, + 0.006972355768084526, + 0.013583563268184662, + 0.019940758123993874, + 0.022356748580932617, + 0.010024392046034336, + 0.026108909398317337, + 0.02354041300714016, + 0.016243048012256622, + 0.010642718523740768, + 0.0062483735382556915, + 0.00897842738777399, + 0.008956766687333584, + 0.004192847292870283, + 0.012794045731425285, + 0.008266628719866276, + 0.0072801318019628525, + 0.005401384551078081, + 0.008874648250639439, + 0.011273249052464962, + 0.01235717162489891, + 0.005275113508105278, + 0.012160999700427055, + 0.012661582790315151, + 0.005886795464903116, + 0.011418438516557217, + 0.004058670252561569, + 0.005958331283181906, + 0.006189865991473198, + 0.0029930986929684877, + 0.008231176994740963, + 0.00646330788731575, + 0.0035691754892468452, + 0.0042755878530442715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.006111979018896818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028237728402018547 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.0028237728402018547, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028237728402018547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4404962956905365, + 0.21055136620998383, + 0.051618631929159164, + 0.0465703159570694, + 0.02488463930785656, + 0.0395766943693161, + 0.041551373898983, + 0.06557714194059372, + 0.0449645034968853, + 0.034209005534648895 + ], + "credit_map": [ + 0.05075947195291519, + 0.08535230159759521, + 0.08478161692619324, + 0.03614288195967674, + 0.03668244928121567, + 0.05074254050850868, + 0.02421438880264759, + 0.07182063907384872, + 0.02285519801080227, + 0.03769920393824577, + 0.037357110530138016, + 0.015275980345904827, + 0.0231325700879097, + 0.027441924437880516, + 0.01093076542019844, + 0.0358586348593235, + 0.00569509994238615, + 0.008740955032408237, + 0.008266492746770382, + 0.004052013158798218, + 0.004275625105947256, + 0.006595297250896692, + 0.003519895952194929, + 0.010473255068063736, + 0.004336428362876177, + 0.006478205323219299, + 0.008813435211777687, + 0.003204136388376355, + 0.007008017972111702, + 0.007302854210138321, + 0.003652516519650817, + 0.005774726625531912, + 0.002651296090334654, + 0.0036714007146656513, + 0.005448593758046627, + 0.001748118200339377, + 0.003207628382369876, + 0.0033802574034780264, + 0.001905077719129622, + 0.0028722682036459446, + 0.004899133928120136, + 0.005519323516637087, + 0.010169655084609985, + 0.003540676785632968, + 0.0033638759050518274, + 0.005850130692124367, + 0.002314613200724125, + 0.003919288050383329, + 0.003932930529117584, + 0.005636070389300585, + 0.006183878518640995, + 0.0026787552051246166, + 0.007768876850605011, + 0.005170729942619801, + 0.004431628622114658, + 0.005748505238443613, + 0.006506627891212702, + 0.010106872767210007, + 0.010506090708076954, + 0.005042866338044405, + 0.009879556484520435, + 0.008378891274333, + 0.005999451503157616, + 0.009156795218586922, + 0.005381747614592314, + 0.006152431946247816, + 0.011121115647256374, + 0.0032164668664336205, + 0.005420559551566839, + 0.0072397529147565365, + 0.0023708390071988106, + 0.004061592742800713, + 0.003576324786990881, + 0.004831300582736731, + 0.006142801139503717, + 0.003165595466271043, + 0.004747092258185148, + 0.005307255312800407, + 0.0026621604338288307, + 0.00377647764980793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0028237728402018547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545830570161343 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545830570161343, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545830570161343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38830652832984924, + 0.10833540558815002, + 0.05192193388938904, + 0.05702843889594078, + 0.08240234851837158, + 0.08212772011756897, + 0.050153158605098724, + 0.05264930799603462, + 0.09552983939647675, + 0.03154527768492699 + ], + "credit_map": [ + 0.03973716124892235, + 0.05464964732527733, + 0.06144271790981293, + 0.03025289811193943, + 0.06015041843056679, + 0.0603359080851078, + 0.03095262311398983, + 0.05078520625829697, + 0.010840334929525852, + 0.017872467637062073, + 0.01947527751326561, + 0.008138254284858704, + 0.013159051537513733, + 0.01451919600367546, + 0.008522368036210537, + 0.0158084686845541, + 0.005664012860506773, + 0.008155922405421734, + 0.010725369676947594, + 0.004223998636007309, + 0.004851270001381636, + 0.008437982760369778, + 0.003262836020439863, + 0.006600548047572374, + 0.005251189228147268, + 0.007331575732678175, + 0.009733621962368488, + 0.0037020882591605186, + 0.013664810918271542, + 0.008105195127427578, + 0.005946822464466095, + 0.0032931421883404255, + 0.009095223620533943, + 0.012174723669886589, + 0.018070634454488754, + 0.007204427383840084, + 0.00830770656466484, + 0.01390976645052433, + 0.0047630611807107925, + 0.008876817300915718, + 0.008393536321818829, + 0.00967343244701624, + 0.01153477095067501, + 0.00698713306337595, + 0.014675906859338284, + 0.014357076026499271, + 0.008577042259275913, + 0.007928835228085518, + 0.006469447631388903, + 0.007988440804183483, + 0.011394739151000977, + 0.0038016040343791246, + 0.0051426696591079235, + 0.0071532344445586205, + 0.002942177001386881, + 0.005260855425149202, + 0.005036680959165096, + 0.007068227045238018, + 0.0074879880994558334, + 0.0037054787389934063, + 0.011211507022380829, + 0.007872259244322777, + 0.00579756498336792, + 0.004469609819352627, + 0.01022378634661436, + 0.01307649165391922, + 0.01240755058825016, + 0.004850273951888084, + 0.023256106302142143, + 0.01275811530649662, + 0.012573819607496262, + 0.006383707746863365, + 0.003945833072066307, + 0.004830248653888702, + 0.0060602459125220776, + 0.0029400577768683434, + 0.004483079072088003, + 0.004616308957338333, + 0.0021140463650226593, + 0.0025554599706083536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.012545830570161343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034038364887238 + ], + "opposite_logprob_tokens": [ + -1.910340428352356 + ], + "logprob_reward": -0.16034038364887238, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034038364887238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2991545498371124, + 0.06995806843042374, + 0.05148075520992279, + 0.06112641096115112, + 0.21902038156986237, + 0.09183380007743835, + 0.06411127001047134, + 0.046857770532369614, + 0.05831350013613701, + 0.038143452256917953 + ], + "credit_map": [ + 0.03416214510798454, + 0.04306738078594208, + 0.05455118045210838, + 0.020963164046406746, + 0.04851331561803818, + 0.035598866641521454, + 0.029754165560007095, + 0.032544322311878204, + 0.007826970890164375, + 0.010909055359661579, + 0.012481090612709522, + 0.004452865570783615, + 0.0122068477794528, + 0.009267599321901798, + 0.0055421022698283195, + 0.0072715370915830135, + 0.00538786081597209, + 0.006581269670277834, + 0.00796393770724535, + 0.004319577943533659, + 0.008744852617383003, + 0.006717129610478878, + 0.0035708313807845116, + 0.008195293135941029, + 0.006624877452850342, + 0.008758907206356525, + 0.012473395094275475, + 0.004033155273646116, + 0.010107687674462795, + 0.00966661423444748, + 0.004748922307044268, + 0.004712848924100399, + 0.02219451405107975, + 0.03407616913318634, + 0.026351310312747955, + 0.014045911841094494, + 0.04804994538426399, + 0.03663484379649162, + 0.01918141543865204, + 0.018486274406313896, + 0.010352964513003826, + 0.012602381408214569, + 0.014818339608609676, + 0.006917939987033606, + 0.016008982434868813, + 0.014200755394995213, + 0.010037788189947605, + 0.00689464807510376, + 0.007019014563411474, + 0.009721285663545132, + 0.007648809812963009, + 0.003503041109070182, + 0.015785181894898415, + 0.009910174645483494, + 0.006873071193695068, + 0.0036506913602352142, + 0.005588787142187357, + 0.005861495155841112, + 0.010158129036426544, + 0.003899963339790702, + 0.00672728568315506, + 0.0065999808721244335, + 0.004513060674071312, + 0.003509068163111806, + 0.006453860085457563, + 0.009598172269761562, + 0.007744263857603073, + 0.004378819838166237, + 0.01080416515469551, + 0.009131291881203651, + 0.005903251469135284, + 0.004299672320485115, + 0.004443637561053038, + 0.006523094139993191, + 0.006590661127120256, + 0.002606552792713046, + 0.00658287201076746, + 0.0054932269267737865, + 0.0032703380566090345, + 0.0026330705732107162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.16034038364887238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818423556163907 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818423556163907, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818423556163907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25839290022850037, + 0.09815875440835953, + 0.04820482060313225, + 0.039861779659986496, + 0.12814010679721832, + 0.06148437038064003, + 0.09686093777418137, + 0.08317657560110092, + 0.14362843334674835, + 0.04209136590361595 + ], + "credit_map": [ + 0.028561506420373917, + 0.03806572034955025, + 0.041221898049116135, + 0.017735792323946953, + 0.042079858481884, + 0.03801121190190315, + 0.028443187475204468, + 0.024273721501231194, + 0.010553054511547089, + 0.014989360235631466, + 0.018204448744654655, + 0.00789419375360012, + 0.014482944272458553, + 0.013505936600267887, + 0.008942866697907448, + 0.00958594586700201, + 0.004990969784557819, + 0.007997742854058743, + 0.009024903178215027, + 0.003357328474521637, + 0.0073540182784199715, + 0.0067804306745529175, + 0.004265408962965012, + 0.004434018861502409, + 0.00460738455876708, + 0.0052227783016860485, + 0.006123860366642475, + 0.0028025528881698847, + 0.007492462173104286, + 0.005215365439653397, + 0.004085463471710682, + 0.004311912693083286, + 0.012448114342987537, + 0.018263069912791252, + 0.02371959388256073, + 0.009682153351604939, + 0.020741555839776993, + 0.019800761714577675, + 0.011153615079820156, + 0.012331229634582996, + 0.006024811416864395, + 0.008005023002624512, + 0.012684092856943607, + 0.004364751745015383, + 0.010249601677060127, + 0.008360839448869228, + 0.005663848947733641, + 0.0061314040794968605, + 0.011259130202233791, + 0.0149996317923069, + 0.0203469917178154, + 0.006474434398114681, + 0.008240937255322933, + 0.015058213844895363, + 0.006318686995655298, + 0.014162907376885414, + 0.010188965126872063, + 0.01132993120700121, + 0.021436430513858795, + 0.006733761169016361, + 0.006484290584921837, + 0.013098749332129955, + 0.005191569682210684, + 0.008712878450751305, + 0.018006423488259315, + 0.022559134289622307, + 0.03247538208961487, + 0.011260406114161015, + 0.011315838433802128, + 0.020452413707971573, + 0.008861170150339603, + 0.01869766227900982, + 0.004223912954330444, + 0.00579312676563859, + 0.006857638712972403, + 0.0031082304194569588, + 0.008238581009209156, + 0.005994219332933426, + 0.004115635063499212, + 0.0037600232753902674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.002818423556163907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a modern, minimalist gooseneck table lamp resting on a lightly textured wooden surface. Finished in a uniformly smooth matte black, the lamp features a slender, flexible neck that gracefully transitions from a heavy, round base to a simple, disc-shaped LED head. When on, the lamp casts a focused, circular beam of bright, cool white light onto the table, highlighting the subtle grain of the wood. Following a faint, invisible surface vibration, the lamp's head and upper neck begin to vibrate gently from side to side. The movement is fluid and subtle, a brief, silent dance that rapidly diminishes as the amplitude of the vibration steadily decreases, until, in an instant, all movement ceases and the lamp returns to a perfect state of stillness, its elegant form a stark, silent silhouette against the softly lit, out-of-focus background of the quiet study. The atmosphere is one of focused tranquility and minimalist calm, the fleeting movements only emphasizing the scene's profound stillness.", + "seed": 16212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_6957_seed_16212_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07418909026309847, + "vlm_logprob": -0.07418909026309847, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.07418909026309847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005823194514960051 + ], + "opposite_logprob_tokens": [ + -6.130823135375977 + ], + "logprob_reward": -0.005823194514960051, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005823194514960051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20575159788131714, + 0.2062123715877533, + 0.06406224519014359, + 0.10407551378011703, + 0.05339740589261055, + 0.10649717599153519, + 0.07798846065998077, + 0.06216795742511749, + 0.08249184489250183, + 0.03735540062189102 + ], + "credit_map": [ + 0.02145947515964508, + 0.02331644482910633, + 0.029495518654584885, + 0.01921869069337845, + 0.0360964871942997, + 0.020713871344923973, + 0.019968362525105476, + 0.03548276424407959, + 0.024015476927161217, + 0.018163790926337242, + 0.030590075999498367, + 0.014593584463000298, + 0.016367468982934952, + 0.0227960254997015, + 0.02225462533533573, + 0.05743132904171944, + 0.007791243959218264, + 0.0077039883472025394, + 0.007467362564057112, + 0.006015835329890251, + 0.011446951888501644, + 0.010895955376327038, + 0.00466942461207509, + 0.008071486838161945, + 0.014435560442507267, + 0.015736950561404228, + 0.009717878885567188, + 0.009662492200732231, + 0.012285789474844933, + 0.014542309567332268, + 0.010238840244710445, + 0.01745569333434105, + 0.007331108208745718, + 0.006342762615531683, + 0.007059893105179071, + 0.00643201544880867, + 0.004757188260555267, + 0.006797004956752062, + 0.004554215352982283, + 0.010123215615749359, + 0.01135735958814621, + 0.006254703737795353, + 0.012832450680434704, + 0.007924106903374195, + 0.011493239551782608, + 0.013700824230909348, + 0.012628003023564816, + 0.03030649572610855, + 0.008408349007368088, + 0.008344302885234356, + 0.007485353387892246, + 0.006286365911364555, + 0.011023333296179771, + 0.008951413445174694, + 0.007292329333722591, + 0.02019701525568962, + 0.007005960214883089, + 0.012503257021307945, + 0.0059837051667273045, + 0.006564540322870016, + 0.008478446863591671, + 0.008075802586972713, + 0.006179709918797016, + 0.007376536726951599, + 0.012315685860812664, + 0.009897192940115929, + 0.008640834130346775, + 0.00797534640878439, + 0.014795573428273201, + 0.010016631335020065, + 0.008150594308972359, + 0.01069999486207962, + 0.005222318693995476, + 0.006347353104501963, + 0.0067322575487196445, + 0.004068468231707811, + 0.003878758056089282, + 0.004526271019130945, + 0.0025956809986382723, + 0.0039842925034463406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005823194514960051 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032007675617933273 + ], + "opposite_logprob_tokens": [ + -5.753200531005859 + ], + "logprob_reward": -0.0032007675617933273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032007675617933273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25649330019950867, + 0.06170310080051422, + 0.0218037161976099, + 0.04440843686461449, + 0.07880841940641403, + 0.0545605830848217, + 0.03260505199432373, + 0.3517639636993408, + 0.04178769141435623, + 0.05606578662991524 + ], + "credit_map": [ + 0.02652726322412491, + 0.016012780368328094, + 0.04663096368312836, + 0.016497399657964706, + 0.01339976117014885, + 0.026848021894693375, + 0.031275615096092224, + 0.07930147647857666, + 0.005515230353921652, + 0.00540871312841773, + 0.009206542745232582, + 0.0049665141850709915, + 0.004821333568543196, + 0.00608041463419795, + 0.007473664823919535, + 0.018230687826871872, + 0.0023774344008415937, + 0.0023423563688993454, + 0.0035176610108464956, + 0.001764854183420539, + 0.0017312123673036695, + 0.0026134063955396414, + 0.0022937837056815624, + 0.005163007881492376, + 0.0043665640987455845, + 0.0034515862353146076, + 0.006201891228556633, + 0.003452012548223138, + 0.003371080383658409, + 0.004705172963440418, + 0.004818669520318508, + 0.014041458256542683, + 0.008790059946477413, + 0.005107360426336527, + 0.010894589126110077, + 0.005041863769292831, + 0.004378617741167545, + 0.008315887302160263, + 0.009523065760731697, + 0.026756975799798965, + 0.00596360769122839, + 0.0035635342355817556, + 0.007425679359585047, + 0.0036476068198680878, + 0.004235853906720877, + 0.005931082181632519, + 0.007243713829666376, + 0.016549505293369293, + 0.0031436302233487368, + 0.002976377494633198, + 0.004479503259062767, + 0.002045947825536132, + 0.0034691146574914455, + 0.0033775921911001205, + 0.003081684233620763, + 0.010031202808022499, + 0.03687145560979843, + 0.0230746790766716, + 0.04734105244278908, + 0.020467134192585945, + 0.024868449196219444, + 0.037192147225141525, + 0.04139265418052673, + 0.12055634707212448, + 0.004082608502358198, + 0.004790303763002157, + 0.004197775386273861, + 0.0030781393870711327, + 0.00618836609646678, + 0.004391820169985294, + 0.003617864567786455, + 0.011440812610089779, + 0.0060814968310296535, + 0.002810846082866192, + 0.008122391998767853, + 0.004055391065776348, + 0.0023876240011304617, + 0.006349700037389994, + 0.007284370716661215, + 0.01897396706044674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.0032007675617933273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651750162243843 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.008651750162243843, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651750162243843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17455437779426575, + 0.0929664894938469, + 0.03719639405608177, + 0.3633146584033966, + 0.06147714704275131, + 0.051555875688791275, + 0.06181173026561737, + 0.11065845936536789, + 0.02358417958021164, + 0.022880729287862778 + ], + "credit_map": [ + 0.021110234782099724, + 0.0129920719191432, + 0.02997780591249466, + 0.018559319898486137, + 0.013858266174793243, + 0.017355674877762794, + 0.018876422196626663, + 0.0418245755136013, + 0.009904943406581879, + 0.007113860920071602, + 0.01071528997272253, + 0.009817306883633137, + 0.009288633242249489, + 0.009340559132397175, + 0.009846349246799946, + 0.026939542964100838, + 0.004849278833717108, + 0.003353839972987771, + 0.007366647012531757, + 0.0031637893989682198, + 0.003563197562471032, + 0.004219191148877144, + 0.004020745866000652, + 0.006659699138253927, + 0.037963010370731354, + 0.037200067192316055, + 0.049351222813129425, + 0.027285296469926834, + 0.035553939640522, + 0.037911251187324524, + 0.03415021300315857, + 0.10389965027570724, + 0.00705913919955492, + 0.005143404006958008, + 0.007354221772402525, + 0.004242017399519682, + 0.005619984120130539, + 0.0075102499686181545, + 0.00610997574403882, + 0.018438151106238365, + 0.006443339865654707, + 0.006140385754406452, + 0.005700135603547096, + 0.005243851337581873, + 0.005978529341518879, + 0.005261654499918222, + 0.004426030442118645, + 0.012361946515738964, + 0.006285055074840784, + 0.005320197902619839, + 0.007521797902882099, + 0.004535397980362177, + 0.00422526802867651, + 0.006153373047709465, + 0.006586963776499033, + 0.021183673292398453, + 0.012910530902445316, + 0.0137990303337574, + 0.011078610084950924, + 0.007965244352817535, + 0.015621677041053772, + 0.011931382119655609, + 0.009415178559720516, + 0.02793680503964424, + 0.0022646584548056126, + 0.003109934041276574, + 0.0035283458419144154, + 0.0023925388231873512, + 0.003105918178334832, + 0.0022023252677172422, + 0.0019074395531788468, + 0.0050730169750750065, + 0.002872464479878545, + 0.0016947070835158229, + 0.0028585316613316536, + 0.0019538146443665028, + 0.0016034391010180116, + 0.0025220459792762995, + 0.002834153361618519, + 0.0065415711142122746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.008651750162243843 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34863099455833435 + ], + "opposite_logprob_tokens": [ + -1.2236310243606567 + ], + "logprob_reward": -0.34863099455833435, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34863099455833435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20160827040672302, + 0.15720953047275543, + 0.045893345028162, + 0.1411038339138031, + 0.08835320919752121, + 0.06423285603523254, + 0.08121192455291748, + 0.09596316516399384, + 0.1010947898030281, + 0.023329077288508415 + ], + "credit_map": [ + 0.024374494329094887, + 0.02570464462041855, + 0.03472813591361046, + 0.017329175025224686, + 0.022530382499098778, + 0.024793464690446854, + 0.016631321981549263, + 0.0355166532099247, + 0.01894664764404297, + 0.014646378345787525, + 0.02306578867137432, + 0.012181992642581463, + 0.01096763089299202, + 0.015984756872057915, + 0.016494110226631165, + 0.04492221772670746, + 0.0066405292600393295, + 0.0064688436686992645, + 0.005759215913712978, + 0.00643582921475172, + 0.005755320657044649, + 0.005270056892186403, + 0.003185227746143937, + 0.006378323305398226, + 0.017787130549550056, + 0.015640275552868843, + 0.019306927919387817, + 0.008928162045776844, + 0.01428209524601698, + 0.014678982086479664, + 0.013034376315772533, + 0.03744589164853096, + 0.011200714856386185, + 0.011466121301054955, + 0.009814123623073101, + 0.006028275471180677, + 0.00744473934173584, + 0.009556279517710209, + 0.008529312908649445, + 0.02431364357471466, + 0.006956818979233503, + 0.009290272369980812, + 0.006746106781065464, + 0.006157325115054846, + 0.009099634364247322, + 0.006819761823862791, + 0.0073532783426344395, + 0.011809653602540493, + 0.009410503320395947, + 0.008906972594559193, + 0.011077520437538624, + 0.0073416042141616344, + 0.008994671516120434, + 0.008463844656944275, + 0.008290053345263004, + 0.01872675120830536, + 0.01177745871245861, + 0.019969725981354713, + 0.008343341760337353, + 0.008849444799125195, + 0.019286060705780983, + 0.012135176919400692, + 0.007254316471517086, + 0.008347638882696629, + 0.010639706626534462, + 0.008374381810426712, + 0.013346604071557522, + 0.006959717720746994, + 0.011523996479809284, + 0.009983540512621403, + 0.010205973871052265, + 0.030060864984989166, + 0.003016280010342598, + 0.0038074334152042866, + 0.0021996551658958197, + 0.0019548849668353796, + 0.0031895479187369347, + 0.0032006814144551754, + 0.0019513830775395036, + 0.004009210038930178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.34863099455833435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16248106956481934, + 0.13661004602909088, + 0.04035176336765289, + 0.1826375126838684, + 0.07120942324399948, + 0.10318411886692047, + 0.10307947546243668, + 0.05044211819767952, + 0.11918721348047256, + 0.030817296355962753 + ], + "credit_map": [ + 0.018506577238440514, + 0.018074287101626396, + 0.018992923200130463, + 0.012977655977010727, + 0.016450095921754837, + 0.02573905512690544, + 0.016647985205054283, + 0.035092487931251526, + 0.016313793137669563, + 0.013663978315889835, + 0.01705656200647354, + 0.008933189325034618, + 0.010761828161776066, + 0.01755339466035366, + 0.01639365963637829, + 0.03593363240361214, + 0.005078981164842844, + 0.0066033825278282166, + 0.0036285254172980785, + 0.00398891419172287, + 0.0033463179133832455, + 0.00579911470413208, + 0.004069688264280558, + 0.007836841978132725, + 0.019802827388048172, + 0.01978364586830139, + 0.024682406336069107, + 0.01428981963545084, + 0.016101766377687454, + 0.018152499571442604, + 0.018201416358351707, + 0.05162312835454941, + 0.008983243256807327, + 0.009430044330656528, + 0.009887587279081345, + 0.007953937165439129, + 0.00750353280454874, + 0.008098040707409382, + 0.005628257058560848, + 0.013724779710173607, + 0.011709540151059628, + 0.008166643790900707, + 0.014561145566403866, + 0.008015726692974567, + 0.0105816051363945, + 0.013754813931882381, + 0.012709438800811768, + 0.023685209453105927, + 0.010889573954045773, + 0.010585455223917961, + 0.013037101365625858, + 0.006725965067744255, + 0.012684023007750511, + 0.011055824346840382, + 0.012061068788170815, + 0.026040462777018547, + 0.005192297510802746, + 0.008318483829498291, + 0.004976820666342974, + 0.004761152900755405, + 0.00667578587308526, + 0.004733155481517315, + 0.0042813243344426155, + 0.011503095738589764, + 0.011760643683373928, + 0.013392476364970207, + 0.011895856820046902, + 0.009946330450475216, + 0.014194690622389317, + 0.014270564541220665, + 0.009521839208900928, + 0.034204814583063126, + 0.004895190708339214, + 0.003134314203634858, + 0.0034374415408819914, + 0.002556155901402235, + 0.0028986467514187098, + 0.0039436412043869495, + 0.004160582087934017, + 0.005791323259472847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00463874451816082 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a modern, minimalist gooseneck table lamp resting on a lightly textured wooden surface. Finished in a uniformly smooth matte black, the lamp features a slender, flexible neck that gracefully transitions from a heavy, round base to a simple, disc-shaped LED head. When on, the lamp casts a focused, circular beam of bright, cool white light onto the table, highlighting the subtle grain of the wood. Following a faint, invisible surface vibration, the lamp's head and upper neck begin to vibrate gently from side to side. The movement is fluid and subtle, a brief, silent dance that rapidly diminishes as the amplitude of the vibration steadily decreases, until, in an instant, all movement ceases and the lamp returns to a perfect state of stillness, its elegant form a stark, silent silhouette against the softly lit, out-of-focus background of the quiet study. The atmosphere is one of focused tranquility and minimalist calm, the fleeting movements only emphasizing the scene's profound stillness.", + "seed": 16213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_2/index_6957_seed_16213_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04719406245276332, + "vlm_logprob": -0.04719406245276332, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.04719406245276332, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007242022547870874 + ], + "opposite_logprob_tokens": [ + -5.632242202758789 + ], + "logprob_reward": -0.007242022547870874, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007242022547870874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36443185806274414, + 0.10213224589824677, + 0.08556617051362991, + 0.06352141499519348, + 0.056626517325639725, + 0.0643688514828682, + 0.04336518794298172, + 0.05119572579860687, + 0.0898716077208519, + 0.07892033457756042 + ], + "credit_map": [ + 0.03575129061937332, + 0.030544249340891838, + 0.055737193673849106, + 0.05749407410621643, + 0.0356370247900486, + 0.0600818432867527, + 0.03476959839463234, + 0.05441658943891525, + 0.014913040213286877, + 0.009805282577872276, + 0.015866607427597046, + 0.009177911095321178, + 0.01312992163002491, + 0.016206523403525352, + 0.01230017002671957, + 0.010732791386544704, + 0.010114024393260479, + 0.008648662827908993, + 0.01281716488301754, + 0.012903816066682339, + 0.007624731864780188, + 0.010375681333243847, + 0.009259592741727829, + 0.01382249966263771, + 0.011107545346021652, + 0.0066127167083323, + 0.010616417042911053, + 0.005631070584058762, + 0.005621918942779303, + 0.011215622536838055, + 0.005611381493508816, + 0.007104748394340277, + 0.011335211805999279, + 0.005375238601118326, + 0.007478202227503061, + 0.004412483889609575, + 0.005543812643736601, + 0.009615267626941204, + 0.00553935719653964, + 0.007326944265514612, + 0.009645464830100536, + 0.007363434880971909, + 0.007674143183976412, + 0.0076932888478040695, + 0.006977392360568047, + 0.009039085358381271, + 0.008551993407309055, + 0.007424050942063332, + 0.00740671344101429, + 0.005207933485507965, + 0.005422200076282024, + 0.005690366495400667, + 0.005060363560914993, + 0.005325638223439455, + 0.005017171613872051, + 0.00423479825258255, + 0.010143527761101723, + 0.003888868261128664, + 0.0067453449591994286, + 0.0056689707562327385, + 0.0032309433445334435, + 0.01031250599771738, + 0.005417040083557367, + 0.005788526497781277, + 0.013235609047114849, + 0.008864514529705048, + 0.01336919516324997, + 0.010586290620267391, + 0.009455952793359756, + 0.011447671800851822, + 0.010893096216022968, + 0.012019277550280094, + 0.009719613008201122, + 0.006831086706370115, + 0.011024954728782177, + 0.0105335284024477, + 0.006078030914068222, + 0.015770789235830307, + 0.007214054930955172, + 0.011748280376195908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.007242022547870874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37073490023612976, + 0.05239706113934517, + 0.08016851544380188, + 0.05988733470439911, + 0.11252865940332413, + 0.03860080987215042, + 0.013911035843193531, + 0.06837321817874908, + 0.03300084173679352, + 0.1703975945711136 + ], + "credit_map": [ + 0.049610886722803116, + 0.044386520981788635, + 0.05102616921067238, + 0.05282575264573097, + 0.03549764305353165, + 0.0591806061565876, + 0.02734294906258583, + 0.050864361226558685, + 0.00984720978885889, + 0.004473046399652958, + 0.007293347734957933, + 0.0065403347834944725, + 0.005026713013648987, + 0.00917719304561615, + 0.003270077286288142, + 0.006769139319658279, + 0.013231180608272552, + 0.007070017512887716, + 0.010963885113596916, + 0.011930584907531738, + 0.005395975429564714, + 0.01411276776343584, + 0.005861838813871145, + 0.011602267622947693, + 0.008253338746726513, + 0.004737575072795153, + 0.009385740384459496, + 0.007887332700192928, + 0.0034936207812279463, + 0.01009510736912489, + 0.005012947600334883, + 0.011021670885384083, + 0.02335646003484726, + 0.00941711850464344, + 0.015246925875544548, + 0.010429196059703827, + 0.007189754396677017, + 0.021289825439453125, + 0.009103711694478989, + 0.016495652496814728, + 0.007877449505031109, + 0.0033675527665764093, + 0.005644301883876324, + 0.004862066823989153, + 0.0023542759008705616, + 0.0075618308037519455, + 0.002765439450740814, + 0.004167891573160887, + 0.002859744941815734, + 0.0011456827633082867, + 0.001835886505432427, + 0.0017251733224838972, + 0.000789626850746572, + 0.002489596139639616, + 0.0010715584503486753, + 0.0019937665201723576, + 0.015940846875309944, + 0.006314569618552923, + 0.009277633391320705, + 0.005867215804755688, + 0.003800397040322423, + 0.01583671011030674, + 0.004349271766841412, + 0.006986574735492468, + 0.006694398354738951, + 0.0031215709168463945, + 0.0045277271419763565, + 0.0025151472073048353, + 0.0018919064896181226, + 0.007469917181879282, + 0.00199014856480062, + 0.004790022969245911, + 0.024819817394018173, + 0.014285667799413204, + 0.024464787915349007, + 0.020917033776640892, + 0.008993270806968212, + 0.03121667169034481, + 0.015228738076984882, + 0.030471598729491234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.009783769026398659 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202704653143883 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202704653143883, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202704653143883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4593164920806885, + 0.22644193470478058, + 0.09273930639028549, + 0.025079257786273956, + 0.029298977926373482, + 0.01698237657546997, + 0.020280878990888596, + 0.02965467981994152, + 0.048872675746679306, + 0.05133338272571564 + ], + "credit_map": [ + 0.07139356434345245, + 0.04416697472333908, + 0.07034450024366379, + 0.06123163551092148, + 0.032027095556259155, + 0.0729430690407753, + 0.038787592202425, + 0.06842207908630371, + 0.03886403888463974, + 0.020506486296653748, + 0.0329003781080246, + 0.02330663427710533, + 0.01734010875225067, + 0.03761199116706848, + 0.025091441348195076, + 0.030820870772004128, + 0.013208974152803421, + 0.009923379868268967, + 0.01224441733211279, + 0.012838849797844887, + 0.0070040421560406685, + 0.014462814666330814, + 0.009760911576449871, + 0.013295922428369522, + 0.004106434062123299, + 0.002392204711213708, + 0.003843525657430291, + 0.0023536630906164646, + 0.0027985661290585995, + 0.004488087259232998, + 0.0019595432095229626, + 0.003137233667075634, + 0.0055566043592989445, + 0.002793419174849987, + 0.003963189199566841, + 0.0019182610558345914, + 0.002550852485001087, + 0.006673905998468399, + 0.0023689628578722477, + 0.0034737857058644295, + 0.0024234075099229813, + 0.0018852688372135162, + 0.002432382432743907, + 0.0021765651181340218, + 0.0018900095019489527, + 0.002531882608309388, + 0.001540581346489489, + 0.0021022821310907602, + 0.002891301643103361, + 0.0021578615996986628, + 0.002615109784528613, + 0.0023680320009589195, + 0.002508817007765174, + 0.003304167417809367, + 0.0024442889261990786, + 0.0019913024734705687, + 0.005057378672063351, + 0.0029570856131613255, + 0.004097434692084789, + 0.002879000036045909, + 0.003487729700282216, + 0.004803492687642574, + 0.0033609268721193075, + 0.003011631313711405, + 0.006267630960792303, + 0.004781028255820274, + 0.0067846523597836494, + 0.0064916289411485195, + 0.00581661332398653, + 0.005999553482979536, + 0.005791761446744204, + 0.00693981209769845, + 0.009189914911985397, + 0.006199803669005632, + 0.006169280037283897, + 0.004120159428566694, + 0.007076844107359648, + 0.009405330754816532, + 0.004427776206284761, + 0.0047442736104130745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014202704653143883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015494853258133 + ], + "opposite_logprob_tokens": [ + -1.7015495300292969 + ], + "logprob_reward": -0.2015494853258133, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015494853258133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33342427015304565, + 0.11551537364721298, + 0.08768905699253082, + 0.06254507601261139, + 0.05828835815191269, + 0.07500800490379333, + 0.08262596279382706, + 0.07539426535367966, + 0.03573618084192276, + 0.07377343624830246 + ], + "credit_map": [ + 0.05387578904628754, + 0.034338776022195816, + 0.0443100780248642, + 0.0376901775598526, + 0.028653282672166824, + 0.06354363262653351, + 0.02672099694609642, + 0.04429153352975845, + 0.020505618304014206, + 0.013305715285241604, + 0.01754213124513626, + 0.008616434410214424, + 0.009736807085573673, + 0.023832568898797035, + 0.009421351365745068, + 0.012554759159684181, + 0.016125621274113655, + 0.010731445625424385, + 0.011984416283667088, + 0.008650465868413448, + 0.006760809570550919, + 0.018126649782061577, + 0.005847252439707518, + 0.009462411515414715, + 0.011917885392904282, + 0.0075355819426476955, + 0.007597274612635374, + 0.006073849741369486, + 0.008437584154307842, + 0.011091522872447968, + 0.005619326140731573, + 0.00427205441519618, + 0.006244742311537266, + 0.0044760857708752155, + 0.007971210405230522, + 0.006456943228840828, + 0.0065347496420145035, + 0.006564535200595856, + 0.01079679373651743, + 0.009243304841220379, + 0.008417665958404541, + 0.006993682589381933, + 0.010297217406332493, + 0.00941890012472868, + 0.008548993617296219, + 0.010129468515515327, + 0.011407921090722084, + 0.009794159792363644, + 0.010995462536811829, + 0.008389377035200596, + 0.010687264613807201, + 0.009463655762374401, + 0.00737879890948534, + 0.009570710361003876, + 0.012592527084052563, + 0.013548175804316998, + 0.012346041388809681, + 0.00785426702350378, + 0.011085034348070621, + 0.00796737801283598, + 0.0066657112911343575, + 0.012369546107947826, + 0.006361585576087236, + 0.010744714178144932, + 0.005286070983856916, + 0.0030146799981594086, + 0.005121335387229919, + 0.0052176122553646564, + 0.003320936346426606, + 0.006179892458021641, + 0.003062230534851551, + 0.004533425904810429, + 0.008559918962419033, + 0.010093554854393005, + 0.009407630190253258, + 0.0075815701857209206, + 0.014210866764187813, + 0.008129543624818325, + 0.009295015595853329, + 0.006495343986898661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2015494853258133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern, minimalist gooseneck table lamp with a matte black finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp resting on a lightly textured wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp cast a focused, circular beam of bright, cool white light onto the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lamp's head and upper neck vibrate gently from side to side before coming to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031923307105898857 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031923307105898857, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031923307105898857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25449830293655396, + 0.12622924149036407, + 0.13537710905075073, + 0.12704266607761383, + 0.07543512433767319, + 0.05082250013947487, + 0.0850895494222641, + 0.029383569955825806, + 0.05296100303530693, + 0.06316090375185013 + ], + "credit_map": [ + 0.03431156277656555, + 0.02677440084517002, + 0.03258237615227699, + 0.027735624462366104, + 0.030211130157113075, + 0.04457695782184601, + 0.0236666277050972, + 0.0346396304666996, + 0.022596975788474083, + 0.013043833896517754, + 0.017862165346741676, + 0.014095847494900227, + 0.010826232843101025, + 0.024831006303429604, + 0.008663921616971493, + 0.014309258200228214, + 0.02926674857735634, + 0.01222813781350851, + 0.017667138949036598, + 0.01297328807413578, + 0.008357087150216103, + 0.02675984799861908, + 0.009727923199534416, + 0.018396932631731033, + 0.022144567221403122, + 0.011719560250639915, + 0.016263838857412338, + 0.01377314142882824, + 0.01281790528446436, + 0.02067667804658413, + 0.01233226340264082, + 0.01731470599770546, + 0.009890139102935791, + 0.007828759960830212, + 0.012666512280702591, + 0.008221512660384178, + 0.0070185880176723, + 0.01018320582807064, + 0.00768137164413929, + 0.011945035308599472, + 0.007688137236982584, + 0.00555743183940649, + 0.006364434026181698, + 0.006034455262124538, + 0.004799973219633102, + 0.008547996170818806, + 0.005887428764253855, + 0.005942645017057657, + 0.012181295081973076, + 0.006916701328009367, + 0.011558976955711842, + 0.014413531869649887, + 0.007580606732517481, + 0.011078493669629097, + 0.009869562461972237, + 0.011490380391478539, + 0.004143813159316778, + 0.003241807222366333, + 0.004203799180686474, + 0.0038525043055415154, + 0.0029758706223219633, + 0.00425617815926671, + 0.003031498286873102, + 0.0036781018134206533, + 0.010481453500688076, + 0.0046710972674191, + 0.0063460515812039375, + 0.005427581258118153, + 0.0048618740402162075, + 0.009866313077509403, + 0.004249215591698885, + 0.007057414390146732, + 0.01078285463154316, + 0.004836339503526688, + 0.008213581517338753, + 0.008984032087028027, + 0.0050615184009075165, + 0.010229090228676796, + 0.006611425429582596, + 0.00844205915927887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0031923307105898857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_tensor([9871, 9871])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_tensor([9871, 9871])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0301c76a7ca5d57e5a1d5cfe25248d832cc15b63 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_tensor([9871, 9871])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, horizontal, medium close-up shot focuses on an elderly craftsman sitting in a dimly lit, yet orderly, restoration studio. His graying hair is meticulously combed back, his face deeply wrinkled. A pair of old-fashioned round magnifying glasses rest on his nose, his gaze focused and serene. He wears a faded indigo cotton work shirt and a brown leather apron stained with white dust and the marks of time. Before him, fixed to his workbench, is a dusty marble bust of a classical woman. The statue's face is elegant and peaceful, but a tiny crack mars its integrity. A beam of warm afternoon sunlight, streaming from a high window to the side, penetrates the fine dust particles floating in the air, forming distinct beams (a Tyndall effect), softly illuminating the craftsman's profile and the marble surface, revealing its veins and flaws. The craftsman holds the statue's shoulder steadily with one hand, while the other carefully grips a delicate scraper. With a near-meditative focus, he gently scrapes away the old, solidified filling from the cracks with the tip of the scraper. Fine white powder slowly flakes off as he moves. The dominant tones of the painting are warm brown, off-white, and vintage blue, creating a soft, retro aesthetic and a tranquil, focused, and timeless artistic atmosphere.", + "seed": 13576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_9871_seed_13576_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.048230660147964954, + "vlm_logprob": -0.048230660147964954, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.048230660147964954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014371342957019806 + ], + "opposite_logprob_tokens": [ + -5.014371395111084 + ], + "logprob_reward": -0.014371342957019806, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.014371342957019806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2585762143135071, + 0.06314276158809662, + 0.08504725247621536, + 0.109487384557724, + 0.08588265627622604, + 0.1110101118683815, + 0.059329744428396225, + 0.044935908168554306, + 0.08899711072444916, + 0.0935908779501915 + ], + "credit_map": [ + 0.05266564339399338, + 0.04789194092154503, + 0.028878573328256607, + 0.024172255769371986, + 0.0332178957760334, + 0.018286800011992455, + 0.03523094579577446, + 0.01823217235505581, + 0.009116354398429394, + 0.009555712342262268, + 0.007300901226699352, + 0.010054610669612885, + 0.005446022842079401, + 0.005083829630166292, + 0.010277009569108486, + 0.006308318581432104, + 0.012526562437415123, + 0.011079283431172371, + 0.010124195367097855, + 0.010016250424087048, + 0.00835293997079134, + 0.007754125632345676, + 0.016275234520435333, + 0.008918658830225468, + 0.01577111892402172, + 0.01270954217761755, + 0.012760106474161148, + 0.009866469539701939, + 0.01834758371114731, + 0.01246910635381937, + 0.01321304403245449, + 0.014350414276123047, + 0.012262629345059395, + 0.007451795507222414, + 0.01097834948450327, + 0.009460479021072388, + 0.007534217555075884, + 0.013526178896427155, + 0.01190927717834711, + 0.012759728357195854, + 0.0182704608887434, + 0.010027787648141384, + 0.029944684356451035, + 0.014876820147037506, + 0.00891323946416378, + 0.0073761665262281895, + 0.01361480075865984, + 0.00798614602535963, + 0.00998654030263424, + 0.0069433581084012985, + 0.010220072232186794, + 0.0055661736987531185, + 0.011367441155016422, + 0.005001431796699762, + 0.004758153110742569, + 0.005486571229994297, + 0.009105078876018524, + 0.005108469165861607, + 0.0045851171016693115, + 0.004589139483869076, + 0.010117408819496632, + 0.004035905469208956, + 0.00380209693685174, + 0.0035926930140703917, + 0.01655278168618679, + 0.00818698015064001, + 0.007736970204859972, + 0.008016477338969707, + 0.020186230540275574, + 0.009247807785868645, + 0.009234455414116383, + 0.009835413657128811, + 0.015684910118579865, + 0.010624109767377377, + 0.022259818390011787, + 0.0099498201161623, + 0.009658494032919407, + 0.007442929316312075, + 0.010505945421755314, + 0.007464852649718523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.014371342957019806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206089336425066 + ], + "opposite_logprob_tokens": [ + -3.895608901977539 + ], + "logprob_reward": -0.0206089336425066, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206089336425066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26314106583595276, + 0.13352546095848083, + 0.08750584721565247, + 0.08739118278026581, + 0.074942946434021, + 0.05126764252781868, + 0.06559694558382034, + 0.12194045633077621, + 0.05583465099334717, + 0.05885380506515503 + ], + "credit_map": [ + 0.049416083842515945, + 0.04354654625058174, + 0.027569111436605453, + 0.026605667546391487, + 0.028644943609833717, + 0.024212343618273735, + 0.041334789246320724, + 0.02181156538426876, + 0.026333894580602646, + 0.016129590570926666, + 0.01357151847332716, + 0.020844245329499245, + 0.02225606143474579, + 0.010256964713335037, + 0.012697240337729454, + 0.011435936205089092, + 0.014484420418739319, + 0.008581708185374737, + 0.016186682507395744, + 0.012071611359715462, + 0.01276092417538166, + 0.006242044270038605, + 0.009515062905848026, + 0.007663391064852476, + 0.012438171543180943, + 0.011726882308721542, + 0.012061531655490398, + 0.009274374693632126, + 0.011630097404122353, + 0.007120196707546711, + 0.014529102481901646, + 0.008610821329057217, + 0.011123095639050007, + 0.009730498306453228, + 0.01066094171255827, + 0.007832531817257404, + 0.006441303063184023, + 0.010905093513429165, + 0.009671676903963089, + 0.008577804081141949, + 0.008034136146306992, + 0.006810929160565138, + 0.011312955990433693, + 0.00534517876803875, + 0.004475340247154236, + 0.005864412058144808, + 0.0049976771697402, + 0.004427017644047737, + 0.010384946130216122, + 0.007624112535268068, + 0.01008006650954485, + 0.006266606040298939, + 0.01190190576016903, + 0.0048659369349479675, + 0.009084075689315796, + 0.005389295518398285, + 0.028631851077079773, + 0.007807696238160133, + 0.012015508487820625, + 0.007332728710025549, + 0.035651255398988724, + 0.00948404148221016, + 0.01130798738449812, + 0.009709394536912441, + 0.011279787868261337, + 0.004610556177794933, + 0.005906857084482908, + 0.003988112788647413, + 0.013080400414764881, + 0.005419733002781868, + 0.00532372435554862, + 0.006225478369742632, + 0.010162098333239555, + 0.0072278836742043495, + 0.009742396883666515, + 0.0057237460277974606, + 0.009439055807888508, + 0.005201226100325584, + 0.005507451016455889, + 0.005849950946867466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0206089336425066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200001023709774 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200001023709774, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200001023709774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20905928313732147, + 0.24399949610233307, + 0.10687672346830368, + 0.1223648339509964, + 0.06863021850585938, + 0.051740702241659164, + 0.037868432700634, + 0.0421014167368412, + 0.08045291155576706, + 0.03690594807267189 + ], + "credit_map": [ + 0.03454096242785454, + 0.03968886286020279, + 0.026484543457627296, + 0.02230013906955719, + 0.018743550404906273, + 0.018818240612745285, + 0.03312242031097412, + 0.015360568650066853, + 0.06210615113377571, + 0.02692100591957569, + 0.02424153871834278, + 0.02406897023320198, + 0.05305727198719978, + 0.017789514735341072, + 0.022839227691292763, + 0.012975823134183884, + 0.014286443591117859, + 0.013846210204064846, + 0.01419273391366005, + 0.014100030064582825, + 0.011640235781669617, + 0.014142205007374287, + 0.01622793637216091, + 0.008440930396318436, + 0.01731340028345585, + 0.01435215212404728, + 0.016367899253964424, + 0.015200371854007244, + 0.017957087606191635, + 0.009857365861535072, + 0.018252892419695854, + 0.013063664548099041, + 0.008479347452521324, + 0.008964618667960167, + 0.010714353993535042, + 0.008043034002184868, + 0.006336871534585953, + 0.010604657232761383, + 0.008274437859654427, + 0.007212898228317499, + 0.007466891314834356, + 0.004624627064913511, + 0.007733636535704136, + 0.005330484826117754, + 0.010430294089019299, + 0.00571048678830266, + 0.005896707996726036, + 0.004547575488686562, + 0.0055846828036010265, + 0.005232872441411018, + 0.007321871817111969, + 0.004479436669498682, + 0.002957375952973962, + 0.003560638288035989, + 0.0053268373012542725, + 0.0034047174267470837, + 0.006916233338415623, + 0.004267712123692036, + 0.005329801235347986, + 0.005462150555104017, + 0.007102031260728836, + 0.0038616242818534374, + 0.0054878611117601395, + 0.003674002131447196, + 0.01656220108270645, + 0.006658515427261591, + 0.008842278271913528, + 0.00588833075016737, + 0.018592307344079018, + 0.007728953845798969, + 0.007210640702396631, + 0.008969686925411224, + 0.006294887978583574, + 0.0045742676593363285, + 0.005824250169098377, + 0.003826052648946643, + 0.006519600749015808, + 0.0028293312061578035, + 0.0032972812186926603, + 0.003740275278687477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.014200001023709774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901235669851303 + ], + "opposite_logprob_tokens": [ + -2.579012393951416 + ], + "logprob_reward": -0.07901235669851303, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901235669851303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12311239540576935, + 0.08718250691890717, + 0.0745106190443039, + 0.0614628903567791, + 0.08920909464359283, + 0.05277680605649948, + 0.07976634055376053, + 0.3393428325653076, + 0.071280837059021, + 0.02135571464896202 + ], + "credit_map": [ + 0.02019771933555603, + 0.01678437925875187, + 0.01787642017006874, + 0.014654922299087048, + 0.015279592014849186, + 0.010253136046230793, + 0.02032426744699478, + 0.0077419583685696125, + 0.01977822370827198, + 0.010078235529363155, + 0.007765351794660091, + 0.010989954695105553, + 0.01840449869632721, + 0.0060807084664702415, + 0.007057832553982735, + 0.0070277014747262, + 0.010750953108072281, + 0.009767412208020687, + 0.007532730232924223, + 0.007933521643280983, + 0.01088416576385498, + 0.006948638707399368, + 0.012038453482091427, + 0.008654745295643806, + 0.008514758199453354, + 0.008886548690497875, + 0.006815522909164429, + 0.006928958464413881, + 0.006078769452869892, + 0.005200203508138657, + 0.011676657944917679, + 0.00736146979033947, + 0.011964589357376099, + 0.009927448816597462, + 0.013140765018761158, + 0.01208541076630354, + 0.007371306885033846, + 0.008296751417219639, + 0.017018377780914307, + 0.009404442273080349, + 0.00988922268152237, + 0.005710198078304529, + 0.007466283161193132, + 0.0051125576719641685, + 0.010222385637462139, + 0.004409503191709518, + 0.006080598570406437, + 0.0038860556669533253, + 0.011688459664583206, + 0.011420335620641708, + 0.011964472010731697, + 0.009053558111190796, + 0.006701576057821512, + 0.009199206717312336, + 0.010984210297465324, + 0.008754529058933258, + 0.0789090022444725, + 0.02315879426896572, + 0.028899431228637695, + 0.018863119184970856, + 0.10409866273403168, + 0.03075781650841236, + 0.02820216678082943, + 0.026453832164406776, + 0.014479845762252808, + 0.005702643189579248, + 0.007776536978781223, + 0.005371260456740856, + 0.016080699861049652, + 0.00738952960819006, + 0.00643191346898675, + 0.008048410527408123, + 0.003688280936330557, + 0.002404205035418272, + 0.0034175016917288303, + 0.0018464019522070885, + 0.003682836890220642, + 0.0017868184950202703, + 0.0023913979530334473, + 0.002138270763680339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.07901235669851303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296066641807556 + ], + "opposite_logprob_tokens": [ + -2.2379605770111084 + ], + "logprob_reward": -0.11296066641807556, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296066641807556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2840479612350464, + 0.0950627475976944, + 0.07057370245456696, + 0.11410917341709137, + 0.12457220256328583, + 0.07546039670705795, + 0.07397512346506119, + 0.1103440374135971, + 0.026644332334399223, + 0.025210315361618996 + ], + "credit_map": [ + 0.06053183600306511, + 0.02887708880007267, + 0.025542348623275757, + 0.024789968505501747, + 0.05082115903496742, + 0.02297893911600113, + 0.04762326553463936, + 0.022883398458361626, + 0.015119006857275963, + 0.01298903301358223, + 0.013193953782320023, + 0.01128675788640976, + 0.010248739272356033, + 0.011442244984209538, + 0.01089471485465765, + 0.009888291358947754, + 0.010397527366876602, + 0.013586168177425861, + 0.007503158878535032, + 0.00886464025825262, + 0.008171360939741135, + 0.006952742580324411, + 0.007633876521140337, + 0.007464228197932243, + 0.01606825925409794, + 0.017119504511356354, + 0.014195723459124565, + 0.01214680541306734, + 0.012064595706760883, + 0.011234551668167114, + 0.015300614759325981, + 0.01597912609577179, + 0.01714407280087471, + 0.012271102517843246, + 0.01183199044317007, + 0.014528272673487663, + 0.010785648599267006, + 0.018275108188390732, + 0.02076444961130619, + 0.018971562385559082, + 0.011588208377361298, + 0.00634092278778553, + 0.010536395944654942, + 0.008241209201514721, + 0.012272438034415245, + 0.009826652705669403, + 0.00881286896765232, + 0.007841699756681919, + 0.010555536486208439, + 0.009572194889187813, + 0.01316569559276104, + 0.009814306162297726, + 0.0050107198767364025, + 0.009960556402802467, + 0.008893080987036228, + 0.00700303865596652, + 0.023134473711252213, + 0.00965487863868475, + 0.011380256153643131, + 0.008912166580557823, + 0.028815871104598045, + 0.008840873837471008, + 0.010111336596310139, + 0.00949418731033802, + 0.004076026845723391, + 0.0025837691500782967, + 0.0036660279147326946, + 0.0035208677873015404, + 0.003191846888512373, + 0.0033664461225271225, + 0.0034233781043440104, + 0.00281597301363945, + 0.004371278453618288, + 0.0027443335857242346, + 0.003115281229838729, + 0.0031033349223434925, + 0.004095497541129589, + 0.0017430748557671905, + 0.0030430748593062162, + 0.0029944423586130142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.11296066641807556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium close-up shot focuses on an elderly craftsman sitting in a dimly lit, yet orderly, restoration studio. His graying hair is meticulously combed back, his face deeply wrinkled. A pair of old-fashioned round magnifying glasses rest on his nose, his gaze focused and serene. He wears a faded indigo cotton work shirt and a brown leather apron stained with white dust and the marks of time. Before him, fixed to his workbench, is a dusty marble bust of a classical woman. The statue's face is elegant and peaceful, but a tiny crack mars its integrity. A beam of warm afternoon sunlight, streaming from a high window to the side, penetrates the fine dust particles floating in the air, forming distinct beams (a Tyndall effect), softly illuminating the craftsman's profile and the marble surface, revealing its veins and flaws. The craftsman holds the statue's shoulder steadily with one hand, while the other carefully grips a delicate scraper. With a near-meditative focus, he gently scrapes away the old, solidified filling from the cracks with the tip of the scraper. Fine white powder slowly flakes off as he moves. The dominant tones of the painting are warm brown, off-white, and vintage blue, creating a soft, retro aesthetic and a tranquil, focused, and timeless artistic atmosphere.", + "seed": 13577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_9871_seed_13577_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.030368633940815924, + "vlm_logprob": -0.030368633940815924, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.030368633940815924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009151536040008068 + ], + "opposite_logprob_tokens": [ + -5.759151458740234 + ], + "logprob_reward": -0.009151536040008068, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009151536040008068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2462266981601715, + 0.1290617436170578, + 0.08972334861755371, + 0.05595714598894119, + 0.07477792352437973, + 0.10303079336881638, + 0.07987483590841293, + 0.06769464910030365, + 0.06550777703523636, + 0.08814502507448196 + ], + "credit_map": [ + 0.017252184450626373, + 0.028536567464470863, + 0.03696895018219948, + 0.03359811007976532, + 0.02222733199596405, + 0.03357711434364319, + 0.036812882870435715, + 0.037253569811582565, + 0.009533560834825039, + 0.01458735577762127, + 0.015297098085284233, + 0.03703559190034866, + 0.010227341204881668, + 0.013492683880031109, + 0.011950468644499779, + 0.01693764515221119, + 0.007115449756383896, + 0.010809122584760189, + 0.011226270347833633, + 0.02428019978106022, + 0.00826601404696703, + 0.010100060142576694, + 0.008423528634011745, + 0.009502707980573177, + 0.00495403166860342, + 0.007178249768912792, + 0.008658722043037415, + 0.009739642962813377, + 0.005813331808894873, + 0.00853454414755106, + 0.005358406808227301, + 0.00572021771222353, + 0.008548559620976448, + 0.00782649964094162, + 0.008776456117630005, + 0.014127579517662525, + 0.00918355118483305, + 0.011075708083808422, + 0.006009465083479881, + 0.00923011265695095, + 0.009425560012459755, + 0.01591009646654129, + 0.01083437167108059, + 0.021231980994343758, + 0.013318914920091629, + 0.017628125846385956, + 0.006319137290120125, + 0.008362611755728722, + 0.010535470210015774, + 0.011591434478759766, + 0.011812972836196423, + 0.013378428295254707, + 0.00864829495549202, + 0.010437916964292526, + 0.00527306878939271, + 0.008197244256734848, + 0.00932164117693901, + 0.0067557464353740215, + 0.015438910573720932, + 0.011818349361419678, + 0.005725472699850798, + 0.0075583080761134624, + 0.0039077987894415855, + 0.007168425712734461, + 0.006714155897498131, + 0.007580569013953209, + 0.011333159171044827, + 0.012252897955477238, + 0.007363589014858007, + 0.008196947164833546, + 0.005367433186620474, + 0.006699024233967066, + 0.010944021865725517, + 0.008822761476039886, + 0.020055588334798813, + 0.012690993025898933, + 0.011494474485516548, + 0.010894445702433586, + 0.006900507025420666, + 0.0063422308303415775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.009151536040008068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233156755566597 + ], + "opposite_logprob_tokens": [ + -3.773315668106079 + ], + "logprob_reward": -0.0233156755566597, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233156755566597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1746305525302887, + 0.07657267898321152, + 0.2281302958726883, + 0.07633417844772339, + 0.09194248169660568, + 0.07247445732355118, + 0.07538846880197525, + 0.0583692230284214, + 0.07242463529109955, + 0.07373291254043579 + ], + "credit_map": [ + 0.01690499857068062, + 0.01797470822930336, + 0.02388455905020237, + 0.031259797513484955, + 0.018637798726558685, + 0.023071754723787308, + 0.022036219015717506, + 0.020860733464360237, + 0.00835405196994543, + 0.008727365173399448, + 0.010988717898726463, + 0.016109002754092216, + 0.0056159235537052155, + 0.007807568646967411, + 0.00702296057716012, + 0.011947098188102245, + 0.04339710995554924, + 0.07871698588132858, + 0.025074651464819908, + 0.023990536108613014, + 0.015953421592712402, + 0.016618262976408005, + 0.013949085026979446, + 0.01043027639389038, + 0.006549129728227854, + 0.007703016512095928, + 0.008287817239761353, + 0.007097679655998945, + 0.013941898010671139, + 0.0187651626765728, + 0.005142475012689829, + 0.00884700845927, + 0.00962089840322733, + 0.008603743277490139, + 0.009512112475931644, + 0.010583274997770786, + 0.021590519696474075, + 0.012564499862492085, + 0.0070329089649021626, + 0.01243453100323677, + 0.006257403641939163, + 0.011209441348910332, + 0.007715546526014805, + 0.012497675605118275, + 0.010372576303780079, + 0.012570678256452084, + 0.004736628849059343, + 0.007114512845873833, + 0.008854598738253117, + 0.009093075059354305, + 0.010101680643856525, + 0.011158779263496399, + 0.012146384455263615, + 0.0102129727602005, + 0.005804555024951696, + 0.008016427047550678, + 0.007502749562263489, + 0.0066188033670187, + 0.005765069741755724, + 0.008576806634664536, + 0.010858584195375443, + 0.007056687492877245, + 0.005722647067159414, + 0.006267883814871311, + 0.007781504653394222, + 0.0075606717728078365, + 0.009702224284410477, + 0.008678669109940529, + 0.009968205355107784, + 0.009253802709281445, + 0.007484393659979105, + 0.01199517771601677, + 0.009817181155085564, + 0.00930140819400549, + 0.008295513689517975, + 0.00936121866106987, + 0.013656464405357838, + 0.00964217260479927, + 0.0058619617484509945, + 0.007797010242938995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0233156755566597 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644777350127697 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644777350127697, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644777350127697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15506185591220856, + 0.10401642322540283, + 0.14964544773101807, + 0.1540856808423996, + 0.08977263420820236, + 0.11593383550643921, + 0.09462860226631165, + 0.0408870093524456, + 0.04863722249865532, + 0.047331299632787704 + ], + "credit_map": [ + 0.013149996288120747, + 0.017429785802960396, + 0.028755486011505127, + 0.023095211014151573, + 0.014129242859780788, + 0.02693280391395092, + 0.020555488765239716, + 0.011013831943273544, + 0.009319649077951908, + 0.01563699170947075, + 0.014026488177478313, + 0.021779043599963188, + 0.010268950834870338, + 0.014673231169581413, + 0.011206773109734058, + 0.007105294149369001, + 0.023581577464938164, + 0.042575038969516754, + 0.01961425133049488, + 0.02282932586967945, + 0.012038144282996655, + 0.010896419174969196, + 0.010589059442281723, + 0.007521623745560646, + 0.01805867627263069, + 0.013683859258890152, + 0.01852048933506012, + 0.03246026486158371, + 0.014231456443667412, + 0.019851764664053917, + 0.023898886516690254, + 0.013380272313952446, + 0.011254395358264446, + 0.009694079868495464, + 0.01313052698969841, + 0.017753995954990387, + 0.011290510185062885, + 0.014171922579407692, + 0.00519655691459775, + 0.0072806477546691895, + 0.011255851946771145, + 0.017399122938513756, + 0.01508883386850357, + 0.02479011006653309, + 0.009590424597263336, + 0.018433140590786934, + 0.006951075047254562, + 0.012425259687006474, + 0.011398480273783207, + 0.012100918218493462, + 0.012586639262735844, + 0.01434695441275835, + 0.012113709934055805, + 0.01200675219297409, + 0.008044248446822166, + 0.012030897662043571, + 0.0057008652947843075, + 0.005315459333360195, + 0.009073219262063503, + 0.006945389788597822, + 0.004556148312985897, + 0.004455919843167067, + 0.002133946167305112, + 0.0027060569263994694, + 0.005864445120096207, + 0.007534550037235022, + 0.007910051383078098, + 0.008371359668672085, + 0.004937630146741867, + 0.00578237185254693, + 0.003361338283866644, + 0.004875475075095892, + 0.0072517539374530315, + 0.009596399962902069, + 0.006575003266334534, + 0.0058694081380963326, + 0.005769026931375265, + 0.0063327886164188385, + 0.0027683961670845747, + 0.003168517490848899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008644777350127697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206833943724632 + ], + "opposite_logprob_tokens": [ + -2.812068223953247 + ], + "logprob_reward": -0.06206833943724632, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206833943724632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12922555208206177, + 0.06316347420215607, + 0.36362555623054504, + 0.04514751210808754, + 0.05869472399353981, + 0.13517753779888153, + 0.10152901709079742, + 0.03163474425673485, + 0.04326345771551132, + 0.028538396582007408 + ], + "credit_map": [ + 0.009241831488907337, + 0.015221212059259415, + 0.013511684723198414, + 0.025347188115119934, + 0.013096149079501629, + 0.02624385431408882, + 0.015499787405133247, + 0.011063851416110992, + 0.005410320591181517, + 0.006910646799951792, + 0.006741034798324108, + 0.010966950096189976, + 0.004595247097313404, + 0.009623470716178417, + 0.007155838888138533, + 0.011759965680539608, + 0.09038067609071732, + 0.15651783347129822, + 0.024340970441699028, + 0.015198071487247944, + 0.01845915988087654, + 0.02205043099820614, + 0.02424030192196369, + 0.012438124977052212, + 0.005743334069848061, + 0.004583359695971012, + 0.0062102205120027065, + 0.007997186854481697, + 0.005281807389110327, + 0.007057195995002985, + 0.004438559990376234, + 0.0038358482997864485, + 0.006557390559464693, + 0.006322410423308611, + 0.00787392444908619, + 0.007230374030768871, + 0.008133724331855774, + 0.011216454207897186, + 0.004261160269379616, + 0.007099286653101444, + 0.01265626959502697, + 0.023383116349577904, + 0.015580938197672367, + 0.029157428070902824, + 0.010734203271567822, + 0.02677064575254917, + 0.008533657528460026, + 0.008361263200640678, + 0.008812622167170048, + 0.015534088015556335, + 0.009507028385996819, + 0.018489200621843338, + 0.010061472654342651, + 0.014621381647884846, + 0.010133502073585987, + 0.014369716867804527, + 0.002698550932109356, + 0.0037249657325446606, + 0.004618890583515167, + 0.004346771631389856, + 0.005296566057950258, + 0.003735373727977276, + 0.003949036356061697, + 0.0032645901665091515, + 0.0038378769531846046, + 0.004974778275936842, + 0.005200439132750034, + 0.007639661896973848, + 0.0061494940891861916, + 0.006095875054597855, + 0.004776057321578264, + 0.00458927359431982, + 0.004003420472145081, + 0.0042365845292806625, + 0.004937265999615192, + 0.003570705419406295, + 0.003857908071950078, + 0.003857896663248539, + 0.0018870345083996654, + 0.002187583362683654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.06206833943724632 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866284132003784 + ], + "opposite_logprob_tokens": [ + -3.0486629009246826 + ], + "logprob_reward": -0.04866284132003784, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866284132003784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40376514196395874, + 0.03724999353289604, + 0.09888095408678055, + 0.045063287019729614, + 0.11781033128499985, + 0.17108352482318878, + 0.06035559996962547, + 0.0194822009652853, + 0.01984109915792942, + 0.026467835530638695 + ], + "credit_map": [ + 0.03890527784824371, + 0.054410286247730255, + 0.06593617796897888, + 0.073964424431324, + 0.03811349719762802, + 0.04695908725261688, + 0.044382184743881226, + 0.041094232350587845, + 0.003159792860969901, + 0.0042933146469295025, + 0.0045392210595309734, + 0.008303849957883358, + 0.003886363236233592, + 0.004235921893268824, + 0.004385864362120628, + 0.004445664118975401, + 0.01582152396440506, + 0.024103669449687004, + 0.011853955686092377, + 0.016577597707509995, + 0.008211574517190456, + 0.00740097276866436, + 0.007011663634330034, + 0.007899994030594826, + 0.004052597098052502, + 0.003842972218990326, + 0.006155863404273987, + 0.008968627080321312, + 0.005989416968077421, + 0.0060958703979849815, + 0.004890694748610258, + 0.005067245569080114, + 0.010164083912968636, + 0.01330257672816515, + 0.017820606008172035, + 0.017799746245145798, + 0.022042114287614822, + 0.016518322750926018, + 0.009393577463924885, + 0.010769306682050228, + 0.016826434060931206, + 0.027350716292858124, + 0.023442881181836128, + 0.029464134946465492, + 0.019009124487638474, + 0.026431899517774582, + 0.013188274577260017, + 0.015370038338005543, + 0.007064794190227985, + 0.01103967148810625, + 0.0044801863841712475, + 0.013936503790318966, + 0.006046508438885212, + 0.009281259030103683, + 0.0036570685915648937, + 0.004849607590585947, + 0.0026774078141897917, + 0.0024883421137928963, + 0.004163060337305069, + 0.002382308477535844, + 0.002756433794274926, + 0.002364877611398697, + 0.0014076469233259559, + 0.0012421230785548687, + 0.0031384523026645184, + 0.0029329906683415174, + 0.0023604617454111576, + 0.002690875669941306, + 0.002398506738245487, + 0.0026390566490590572, + 0.0018946132622659206, + 0.0017861437518149614, + 0.004400928504765034, + 0.005164842586964369, + 0.0035043659154325724, + 0.0033856602385640144, + 0.00323973735794425, + 0.0030709379352629185, + 0.001925593358464539, + 0.0017757684690877795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.04866284132003784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium close-up shot focuses on an elderly craftsman sitting in a dimly lit, yet orderly, restoration studio. His graying hair is meticulously combed back, his face deeply wrinkled. A pair of old-fashioned round magnifying glasses rest on his nose, his gaze focused and serene. He wears a faded indigo cotton work shirt and a brown leather apron stained with white dust and the marks of time. Before him, fixed to his workbench, is a dusty marble bust of a classical woman. The statue's face is elegant and peaceful, but a tiny crack mars its integrity. A beam of warm afternoon sunlight, streaming from a high window to the side, penetrates the fine dust particles floating in the air, forming distinct beams (a Tyndall effect), softly illuminating the craftsman's profile and the marble surface, revealing its veins and flaws. The craftsman holds the statue's shoulder steadily with one hand, while the other carefully grips a delicate scraper. With a near-meditative focus, he gently scrapes away the old, solidified filling from the cracks with the tip of the scraper. Fine white powder slowly flakes off as he moves. The dominant tones of the painting are warm brown, off-white, and vintage blue, creating a soft, retro aesthetic and a tranquil, focused, and timeless artistic atmosphere.", + "seed": 13578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_9871_seed_13578_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031208811420947315, + "vlm_logprob": -0.031208811420947315, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.031208811420947315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008195106871426105 + ], + "opposite_logprob_tokens": [ + -5.508194923400879 + ], + "logprob_reward": -0.008195106871426105, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008195106871426105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03794926032423973, + 0.011946936137974262, + 0.010308018885552883, + 0.5407222509384155, + 0.017912713810801506, + 0.0377756804227829, + 0.17807213962078094, + 0.043701376765966415, + 0.11281096935272217, + 0.008800656534731388 + ], + "credit_map": [ + 0.008122737519443035, + 0.006916081998497248, + 0.0030503084417432547, + 0.003390053752809763, + 0.004625501576811075, + 0.0031554580200463533, + 0.0032597575336694717, + 0.005429361946880817, + 0.0015634094597771764, + 0.002134074456989765, + 0.0011486371513456106, + 0.0010892024729400873, + 0.0013398746959865093, + 0.0011328303953632712, + 0.0014758991310372949, + 0.002063007326796651, + 0.001460172818042338, + 0.0015926887281239033, + 0.0007595114875584841, + 0.0005515934317372739, + 0.0012402513530105352, + 0.0016625968273729086, + 0.0010334387188777328, + 0.002007765229791403, + 0.21324951946735382, + 0.21783284842967987, + 0.03532998263835907, + 0.014217379502952099, + 0.023395992815494537, + 0.012619896791875362, + 0.011580548249185085, + 0.012496094219386578, + 0.004624516237527132, + 0.004470868967473507, + 0.0011080465046688914, + 0.000829310913104564, + 0.0010398119920864701, + 0.0013530818978324533, + 0.0012146522058174014, + 0.003272423753514886, + 0.009621248580515385, + 0.009460711851716042, + 0.0019993549212813377, + 0.001616858527995646, + 0.0025374768301844597, + 0.002932824194431305, + 0.002664907369762659, + 0.006942296866327524, + 0.060093797743320465, + 0.061831049621105194, + 0.010393002070486546, + 0.005953623913228512, + 0.008664537221193314, + 0.00840019341558218, + 0.006087239366024733, + 0.016648689284920692, + 0.010743406601250172, + 0.010598978959023952, + 0.0033012989442795515, + 0.0024424628354609013, + 0.0032896988559514284, + 0.004754426423460245, + 0.0028769290074706078, + 0.005694174207746983, + 0.04301401972770691, + 0.04631056636571884, + 0.0069959089159965515, + 0.003529765410348773, + 0.004987654741853476, + 0.0026764688082039356, + 0.0024074800312519073, + 0.002889103488996625, + 0.0014048963785171509, + 0.001266130362637341, + 0.0010847254889085889, + 0.0010304204188287258, + 0.0009071959648281336, + 0.0009053212124854326, + 0.0011945100268349051, + 0.001007456099614501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.008195106871426105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086315736174583 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086315736174583, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086315736174583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02891889400780201, + 0.01716219261288643, + 0.008137797936797142, + 0.5406298041343689, + 0.017490070313215256, + 0.03812773525714874, + 0.10261989384889603, + 0.027478229254484177, + 0.19076865911483765, + 0.02866673469543457 + ], + "credit_map": [ + 0.004759680014103651, + 0.005106640048325062, + 0.0026383670046925545, + 0.002960293786600232, + 0.002644225722178817, + 0.003162805689498782, + 0.002685027662664652, + 0.004961855243891478, + 0.0037771393544971943, + 0.004288291558623314, + 0.0012464113533496857, + 0.000972019275650382, + 0.0012076746206730604, + 0.0014398725470528007, + 0.001681572524830699, + 0.002549211960285902, + 0.0012018129928037524, + 0.0013897968456149101, + 0.0009052385576069355, + 0.0007044706726446748, + 0.001117854262702167, + 0.0009285798878408968, + 0.0007175052305683494, + 0.0011725395452231169, + 0.2133946567773819, + 0.21747246384620667, + 0.0352310873568058, + 0.014550742693245411, + 0.023181194439530373, + 0.012664523907005787, + 0.011626774445176125, + 0.012508376501500607, + 0.004493059124797583, + 0.00444336561486125, + 0.0010152338072657585, + 0.0007635606452822685, + 0.0010557244531810284, + 0.0012974049896001816, + 0.0012257145717740059, + 0.0031960082706063986, + 0.009442328475415707, + 0.00954074040055275, + 0.0021317133214324713, + 0.0020208447240293026, + 0.0025942069478332996, + 0.002850125776603818, + 0.002777351066470146, + 0.0067704240791499615, + 0.032826922833919525, + 0.03336796909570694, + 0.005954429041594267, + 0.003410806180909276, + 0.005200677085667849, + 0.006096947472542524, + 0.004163700621575117, + 0.011598451994359493, + 0.004625231027603149, + 0.005351055413484573, + 0.002194539178162813, + 0.0018814251525327563, + 0.002814909443259239, + 0.0033792227040976286, + 0.0022856127470731735, + 0.004946233704686165, + 0.07328205555677414, + 0.07875313609838486, + 0.011780356988310814, + 0.005946635734289885, + 0.008393085561692715, + 0.004444032907485962, + 0.0037875554990023375, + 0.004381806124001741, + 0.006809052545577288, + 0.007627906743437052, + 0.0020519443787634373, + 0.0016560208750888705, + 0.001984277041628957, + 0.002140183700248599, + 0.0032448212150484324, + 0.003152530873194337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.016086315736174583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631670217961073 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631670217961073, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631670217961073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04298510029911995, + 0.04850935935974121, + 0.02623780630528927, + 0.4562830328941345, + 0.05878308415412903, + 0.04080764949321747, + 0.13646018505096436, + 0.049946412444114685, + 0.10532540082931519, + 0.03466196730732918 + ], + "credit_map": [ + 0.007131871301680803, + 0.007840478792786598, + 0.004926471039652824, + 0.004166560247540474, + 0.0033431611955165863, + 0.00502034230157733, + 0.004263573791831732, + 0.00629263324663043, + 0.007429078221321106, + 0.008662059903144836, + 0.0024626506492495537, + 0.0031368050258606672, + 0.004022024106234312, + 0.006378086283802986, + 0.005795191042125225, + 0.010623457841575146, + 0.0035642858128994703, + 0.00398407643660903, + 0.0022118103224784136, + 0.0019385189516469836, + 0.0023842838127166033, + 0.0044820657931268215, + 0.002523212693631649, + 0.005149549804627895, + 0.1787441372871399, + 0.18225036561489105, + 0.029629016295075417, + 0.012214604765176773, + 0.019554078578948975, + 0.011389356106519699, + 0.010262612253427505, + 0.012238813564181328, + 0.012592971324920654, + 0.01276881992816925, + 0.00281426589936018, + 0.002373785013332963, + 0.0042927460744977, + 0.007034835871309042, + 0.004704225808382034, + 0.012201431207358837, + 0.009721193462610245, + 0.009622115641832352, + 0.0022580837830901146, + 0.0018333669286221266, + 0.00279886438511312, + 0.00328929559327662, + 0.0032141937408596277, + 0.00807053130120039, + 0.04368383437395096, + 0.04469453543424606, + 0.007689132820814848, + 0.004853745456784964, + 0.0070551251992583275, + 0.007750426884740591, + 0.0054268017411231995, + 0.015306573361158371, + 0.012019123882055283, + 0.011953232809901237, + 0.004161498509347439, + 0.0031052434351295233, + 0.003529729787260294, + 0.005223983433097601, + 0.0035705564077943563, + 0.006383040454238653, + 0.03684492036700249, + 0.03756232559680939, + 0.006330814212560654, + 0.005753722507506609, + 0.005137462634593248, + 0.003994825761765242, + 0.004033071920275688, + 0.005668247118592262, + 0.009653693996369839, + 0.010346874594688416, + 0.002627486130222678, + 0.001516096293926239, + 0.002044658176600933, + 0.0019039191538468003, + 0.0035371212288737297, + 0.0030321141239255667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.007631670217961073 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207001954317093 + ], + "opposite_logprob_tokens": [ + -2.81207013130188 + ], + "logprob_reward": -0.06207001954317093, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207001954317093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11320697516202927, + 0.03233571723103523, + 0.020285142585635185, + 0.3360632359981537, + 0.03090897761285305, + 0.02590394951403141, + 0.013505830429494381, + 0.02967997081577778, + 0.3642209470272064, + 0.033889297395944595 + ], + "credit_map": [ + 0.025075383484363556, + 0.02208288013935089, + 0.00966346450150013, + 0.013773207552731037, + 0.009007451124489307, + 0.009287801571190357, + 0.010925732553005219, + 0.013391057029366493, + 0.005106660537421703, + 0.00646177027374506, + 0.0020495764911174774, + 0.002342653227970004, + 0.0026939604431390762, + 0.0035717468708753586, + 0.0040516420267522335, + 0.006057702004909515, + 0.0029390286654233932, + 0.0026901063974946737, + 0.0016743199666962028, + 0.0014085525181144476, + 0.0019104700768366456, + 0.003356280503794551, + 0.0020876736380159855, + 0.004218708258122206, + 0.12913022935390472, + 0.13276241719722748, + 0.02161005325615406, + 0.010203949175775051, + 0.014242068864405155, + 0.010495120659470558, + 0.007949070073664188, + 0.009670321829617023, + 0.006378070451319218, + 0.006652506534010172, + 0.0016152231255546212, + 0.0012751554604619741, + 0.002281207824125886, + 0.0037227303255349398, + 0.0023943765554577112, + 0.006589705590158701, + 0.0032172584906220436, + 0.003007668536156416, + 0.0016781798331066966, + 0.00202420842833817, + 0.002326027024537325, + 0.004041668493300676, + 0.003161117434501648, + 0.006447820458561182, + 0.002371298847720027, + 0.0023032778408378363, + 0.0009433642844669521, + 0.0008918163366615772, + 0.001076537650078535, + 0.0017953872447833419, + 0.001376382657326758, + 0.0027477650437504053, + 0.00914562027901411, + 0.009947801008820534, + 0.0018751068273559213, + 0.0013991894666105509, + 0.0015221708454191685, + 0.0015844665467739105, + 0.0016465879743918777, + 0.0025590236764401197, + 0.14070728421211243, + 0.1497056484222412, + 0.02205934189260006, + 0.011054160073399544, + 0.015403739176690578, + 0.008541136048734188, + 0.007053186185657978, + 0.009696400724351406, + 0.009406432509422302, + 0.01038471981883049, + 0.0021977645810693502, + 0.0018932610983029008, + 0.0020957784727215767, + 0.0026476182974874973, + 0.0027421072591096163, + 0.0025216147769242525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.06207001954317093 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062060944736003876 + ], + "opposite_logprob_tokens": [ + -2.812060832977295 + ], + "logprob_reward": -0.062060944736003876, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062060944736003876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025802478194236755, + 0.0069632744416594505, + 0.006924903020262718, + 0.41608673334121704, + 0.0336175262928009, + 0.07754533737897873, + 0.22136405110359192, + 0.06546254456043243, + 0.09798919409513474, + 0.04824390634894371 + ], + "credit_map": [ + 0.004015850368887186, + 0.004205432720482349, + 0.002300173742696643, + 0.002591256517916918, + 0.0030765687115490437, + 0.0027614880818873644, + 0.0026797379832714796, + 0.004171969369053841, + 0.0009748878655955195, + 0.0013235316146165133, + 0.00050078344065696, + 0.00047061199438758194, + 0.0007014285074546933, + 0.0008996946853585541, + 0.0008060343097895384, + 0.0012863022275269032, + 0.0014981450513005257, + 0.0017224678304046392, + 0.0004835925064980984, + 0.0005540586425922811, + 0.0005545092863030732, + 0.0007106589619070292, + 0.0006854425882920623, + 0.0007160283857956529, + 0.16314060986042023, + 0.16708803176879883, + 0.027032645419239998, + 0.010871575213968754, + 0.018129901960492134, + 0.010462774895131588, + 0.009448233991861343, + 0.009912961162626743, + 0.005583552643656731, + 0.005837995558977127, + 0.0016250023618340492, + 0.0015687653794884682, + 0.0024319251533597708, + 0.00419765105471015, + 0.003203212982043624, + 0.009169422090053558, + 0.01753576472401619, + 0.018024802207946777, + 0.004114591982215643, + 0.0034957630559802055, + 0.005651683080941439, + 0.006882346235215664, + 0.006297520361840725, + 0.015542873181402683, + 0.07292637228965759, + 0.0746525302529335, + 0.012735447846353054, + 0.00742889940738678, + 0.010939893312752247, + 0.0116910170763731, + 0.00832580216228962, + 0.022664090618491173, + 0.016486147418618202, + 0.016662558540701866, + 0.004961268976330757, + 0.004134373739361763, + 0.0045713866129517555, + 0.0061348238959908485, + 0.004356526304036379, + 0.00815545953810215, + 0.03775475546717644, + 0.040412772446870804, + 0.006026430055499077, + 0.003001732286065817, + 0.004197739064693451, + 0.0022585466504096985, + 0.001951466896571219, + 0.0023857560008764267, + 0.011217537336051464, + 0.01121353916823864, + 0.004291504621505737, + 0.0024860079865902662, + 0.003490820527076721, + 0.0036603084299713373, + 0.005557754077017307, + 0.0063264332711696625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.062060944736003876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium close-up shot focuses on an elderly craftsman sitting in a dimly lit, yet orderly, restoration studio. His graying hair is meticulously combed back, his face deeply wrinkled. A pair of old-fashioned round magnifying glasses rest on his nose, his gaze focused and serene. He wears a faded indigo cotton work shirt and a brown leather apron stained with white dust and the marks of time. Before him, fixed to his workbench, is a dusty marble bust of a classical woman. The statue's face is elegant and peaceful, but a tiny crack mars its integrity. A beam of warm afternoon sunlight, streaming from a high window to the side, penetrates the fine dust particles floating in the air, forming distinct beams (a Tyndall effect), softly illuminating the craftsman's profile and the marble surface, revealing its veins and flaws. The craftsman holds the statue's shoulder steadily with one hand, while the other carefully grips a delicate scraper. With a near-meditative focus, he gently scrapes away the old, solidified filling from the cracks with the tip of the scraper. Fine white powder slowly flakes off as he moves. The dominant tones of the painting are warm brown, off-white, and vintage blue, creating a soft, retro aesthetic and a tranquil, focused, and timeless artistic atmosphere.", + "seed": 13579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_3/index_9871_seed_13579_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05518079660832882, + "vlm_logprob": -0.05518079660832882, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05518079660832882, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01036805659532547 + ], + "opposite_logprob_tokens": [ + -5.6353678703308105 + ], + "logprob_reward": -0.01036805659532547, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.01036805659532547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31888547539711, + 0.10982701927423477, + 0.07207068055868149, + 0.06313524395227432, + 0.07129190862178802, + 0.08135104924440384, + 0.056143417954444885, + 0.05765529349446297, + 0.07361012697219849, + 0.09602975100278854 + ], + "credit_map": [ + 0.038876596838235855, + 0.049010079354047775, + 0.03292201831936836, + 0.02397191897034645, + 0.05589016526937485, + 0.0532638318836689, + 0.025619585067033768, + 0.03933129459619522, + 0.01374314446002245, + 0.0184120275080204, + 0.009986392222344875, + 0.009196671657264233, + 0.014368775300681591, + 0.017675578594207764, + 0.011244233697652817, + 0.015200205147266388, + 0.010286829434335232, + 0.009949037805199623, + 0.006650299299508333, + 0.0042791059240698814, + 0.007927785627543926, + 0.01507748942822218, + 0.007072692271322012, + 0.010827447287738323, + 0.008439507335424423, + 0.011374368332326412, + 0.006218391004949808, + 0.004929464776068926, + 0.006986369378864765, + 0.007993362843990326, + 0.010762078687548637, + 0.006431711837649345, + 0.008658049628138542, + 0.009601515717804432, + 0.006415906827896833, + 0.007563289720565081, + 0.009530722163617611, + 0.01482689380645752, + 0.007753834594041109, + 0.006941698491573334, + 0.011056373827159405, + 0.008918671868741512, + 0.006809787359088659, + 0.004953410476446152, + 0.009792023338377476, + 0.017420824617147446, + 0.009394010528922081, + 0.013005958870053291, + 0.0077232858166098595, + 0.005910371895879507, + 0.005912084598094225, + 0.005258909892290831, + 0.00795386079698801, + 0.008767524734139442, + 0.006139423232525587, + 0.008477961644530296, + 0.007464247290045023, + 0.0076665799133479595, + 0.005812690127640963, + 0.004613349214196205, + 0.00725335069000721, + 0.009674971923232079, + 0.004961193539202213, + 0.010208912193775177, + 0.008376962505280972, + 0.00659124506637454, + 0.006902436725795269, + 0.005589799955487251, + 0.009617106057703495, + 0.014414907433092594, + 0.0068544805981218815, + 0.015263191424310207, + 0.011705071665346622, + 0.009462741203606129, + 0.009538902901113033, + 0.009537918493151665, + 0.011731057427823544, + 0.01731545478105545, + 0.008606111630797386, + 0.018132498487830162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.01036805659532547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562312185764313 + ], + "opposite_logprob_tokens": [ + -4.387562274932861 + ], + "logprob_reward": -0.012562312185764313, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562312185764313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36595121026039124, + 0.12838828563690186, + 0.07943426817655563, + 0.0873030424118042, + 0.07101263850927353, + 0.06033049151301384, + 0.03773441165685654, + 0.04250086471438408, + 0.0383802205324173, + 0.0889645591378212 + ], + "credit_map": [ + 0.05043162778019905, + 0.058162663131952286, + 0.02959412895143032, + 0.02264244295656681, + 0.04462985321879387, + 0.06596273928880692, + 0.0317925289273262, + 0.06273523718118668, + 0.02089976891875267, + 0.018207000568509102, + 0.014106838032603264, + 0.010637493804097176, + 0.016247177496552467, + 0.01770014315843582, + 0.011439274996519089, + 0.01915058307349682, + 0.014222291298210621, + 0.014948062598705292, + 0.009216931648552418, + 0.0068717943504452705, + 0.008929256349802017, + 0.009099658578634262, + 0.0072008040733635426, + 0.00894546415656805, + 0.013091923668980598, + 0.012246859259903431, + 0.008075103163719177, + 0.00906650722026825, + 0.011049403809010983, + 0.01291812863200903, + 0.009823711588978767, + 0.01103140227496624, + 0.009405568242073059, + 0.009127472527325153, + 0.0062219793908298016, + 0.005476966500282288, + 0.009357737377285957, + 0.01728789694607258, + 0.004941646009683609, + 0.00919337384402752, + 0.008067498914897442, + 0.008957305923104286, + 0.006987493485212326, + 0.0063865529373288155, + 0.007325952872633934, + 0.007975002750754356, + 0.00622969213873148, + 0.008400991559028625, + 0.005578816868364811, + 0.005303400568664074, + 0.0034929693210870028, + 0.0038215918466448784, + 0.005733894649893045, + 0.005870295688509941, + 0.003479217877611518, + 0.0044542280957102776, + 0.007021118421107531, + 0.008585087954998016, + 0.005683980416506529, + 0.0047404151409864426, + 0.005520702339708805, + 0.00388223840855062, + 0.0029795716982334852, + 0.004087747074663639, + 0.004839474800974131, + 0.004083959851413965, + 0.004422105383127928, + 0.00412529893219471, + 0.004848857875913382, + 0.007987234741449356, + 0.0034243452828377485, + 0.0046489425003528595, + 0.013333267532289028, + 0.01481458730995655, + 0.013115311041474342, + 0.007523630745708942, + 0.013762105256319046, + 0.010290372185409069, + 0.006281269248574972, + 0.009844012558460236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012562312185764313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978943519294262 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.00978943519294262, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978943519294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32310056686401367, + 0.1658995896577835, + 0.07953235507011414, + 0.08859048038721085, + 0.03567732125520706, + 0.07399982213973999, + 0.059888292104005814, + 0.0814051628112793, + 0.04514559358358383, + 0.046760786324739456 + ], + "credit_map": [ + 0.040220025926828384, + 0.04233406111598015, + 0.03259929269552231, + 0.023960068821907043, + 0.04442577809095383, + 0.050909142941236496, + 0.024961622431874275, + 0.06369057297706604, + 0.020525064319372177, + 0.01867726631462574, + 0.014174423180520535, + 0.0110938074067235, + 0.019181804731488228, + 0.03113548271358013, + 0.01578338071703911, + 0.03532835841178894, + 0.010704653337597847, + 0.014441468752920628, + 0.0069937752559781075, + 0.005813693627715111, + 0.01047816127538681, + 0.014359861612319946, + 0.0068387119099497795, + 0.009902025572955608, + 0.011473439633846283, + 0.00948731042444706, + 0.008955893106758595, + 0.00803576409816742, + 0.012045323848724365, + 0.012674955651164055, + 0.009827048517763615, + 0.016090743243694305, + 0.0051747583784163, + 0.004951880779117346, + 0.003559060860425234, + 0.002398526994511485, + 0.00523769948631525, + 0.005335772410035133, + 0.00422228779643774, + 0.004797334782779217, + 0.008625500835478306, + 0.008880320005118847, + 0.006577115040272474, + 0.0032713990658521652, + 0.010925737209618092, + 0.01558009535074234, + 0.007218529935926199, + 0.012921120040118694, + 0.009289047680795193, + 0.006186218000948429, + 0.006374855525791645, + 0.003684280440211296, + 0.008316081948578358, + 0.009098337963223457, + 0.005812849383801222, + 0.011126618832349777, + 0.010737157426774502, + 0.009971436113119125, + 0.00862114131450653, + 0.005108204670250416, + 0.010564344003796577, + 0.01300199143588543, + 0.008461827412247658, + 0.01493906881660223, + 0.006454160436987877, + 0.005601168144494295, + 0.003942918032407761, + 0.002679030643776059, + 0.006351057905703783, + 0.007585232611745596, + 0.00509747676551342, + 0.007434550207108259, + 0.005713032092899084, + 0.00561714731156826, + 0.0060665966011583805, + 0.0028131259605288506, + 0.0066810110583901405, + 0.008136254735291004, + 0.0039658998139202595, + 0.0077677201479673386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00978943519294262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035334527492523 + ], + "opposite_logprob_tokens": [ + -2.350353240966797 + ], + "logprob_reward": -0.10035334527492523, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035334527492523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35828298330307007, + 0.1409986913204193, + 0.06796540319919586, + 0.07320187985897064, + 0.06364018470048904, + 0.04584925249218941, + 0.045701585710048676, + 0.05503537133336067, + 0.07253246754407883, + 0.07679221779108047 + ], + "credit_map": [ + 0.047050319612026215, + 0.05701659247279167, + 0.025737645104527473, + 0.026461981236934662, + 0.04138068109750748, + 0.07693466544151306, + 0.03232269361615181, + 0.05137842521071434, + 0.017380831763148308, + 0.01749386638402939, + 0.01086198166012764, + 0.00969015620648861, + 0.017284346744418144, + 0.030541272833943367, + 0.009969162754714489, + 0.027777064591646194, + 0.009327307343482971, + 0.008470505475997925, + 0.006703735329210758, + 0.005104399751871824, + 0.007674242835491896, + 0.015092725865542889, + 0.005636989139020443, + 0.009955495595932007, + 0.009112166240811348, + 0.00784127227962017, + 0.005583648104220629, + 0.0038907909765839577, + 0.008967372588813305, + 0.01572587341070175, + 0.007269518915563822, + 0.014811237342655659, + 0.008117462508380413, + 0.007045299746096134, + 0.004016145598143339, + 0.004164121579378843, + 0.008113442920148373, + 0.016032664105296135, + 0.005209980066865683, + 0.01094107050448656, + 0.005895227659493685, + 0.005258776247501373, + 0.004510906990617514, + 0.003121362067759037, + 0.006455711089074612, + 0.006749466992914677, + 0.0056759146973490715, + 0.008181886747479439, + 0.00552291888743639, + 0.004328225273638964, + 0.004308505915105343, + 0.0027459017001092434, + 0.006101213861256838, + 0.01149748731404543, + 0.003566961968317628, + 0.0076303682290017605, + 0.007724250666797161, + 0.00686024222522974, + 0.0058785066939890385, + 0.0026139034889638424, + 0.0059414515271782875, + 0.010180281475186348, + 0.005606420338153839, + 0.010230312123894691, + 0.008375552482903004, + 0.008865044452250004, + 0.0073567721992731094, + 0.005718541797250509, + 0.00931770633906126, + 0.013455932959914207, + 0.007132574450224638, + 0.01231034379452467, + 0.00884653627872467, + 0.006443001329898834, + 0.007759598549455404, + 0.0045639825984835625, + 0.009927201084792614, + 0.018739929422736168, + 0.005941583774983883, + 0.014570387080311775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.10035334527492523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on an elderly craftsman with graying hair combed back and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded indigo cotton work shirt and a brown leather apron stained with white dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dusty marble bust of a classical woman fixed to the workbench in front of the craftsman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a delicate scraper to remove old filling from a tiny crack in the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a beam of warm afternoon sunlight streaming from a high window, illuminating dust particles and the statue's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283083379268646 + ], + "opposite_logprob_tokens": [ + -2.0178308486938477 + ], + "logprob_reward": -0.14283083379268646, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283083379268646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29781055450439453, + 0.1012132316827774, + 0.05585850030183792, + 0.13959746062755585, + 0.12920242547988892, + 0.10528360307216644, + 0.03490642458200455, + 0.03687787428498268, + 0.06268846988677979, + 0.03656153008341789 + ], + "credit_map": [ + 0.03706713765859604, + 0.03797976300120354, + 0.027431970462203026, + 0.021972181275486946, + 0.05193902924656868, + 0.04623216763138771, + 0.023862523958086967, + 0.05132576823234558, + 0.014195079915225506, + 0.02012709714472294, + 0.008814365603029728, + 0.0055478597059845924, + 0.01126498356461525, + 0.014635555446147919, + 0.00959252379834652, + 0.017035754397511482, + 0.00886598415672779, + 0.012831877917051315, + 0.006384553853422403, + 0.004916520789265633, + 0.007131462916731834, + 0.005881104152649641, + 0.004323503002524376, + 0.0055234916508197784, + 0.01691567897796631, + 0.015507661737501621, + 0.013161133043467999, + 0.008212529122829437, + 0.015750283375382423, + 0.025965357199311256, + 0.02452167496085167, + 0.01956312544643879, + 0.01639782451093197, + 0.016351960599422455, + 0.010994481854140759, + 0.012233594432473183, + 0.017652682960033417, + 0.02946849726140499, + 0.01172299962490797, + 0.014380368404090405, + 0.015014849603176117, + 0.009498256258666515, + 0.007366123143583536, + 0.006412302143871784, + 0.012834576889872551, + 0.019447272643446922, + 0.013074589893221855, + 0.0216356310993433, + 0.004724235739558935, + 0.003811470465734601, + 0.002831053454428911, + 0.0028372786473482847, + 0.005674069281667471, + 0.007332570385187864, + 0.0028054900467395782, + 0.004890254233032465, + 0.006206865888088942, + 0.0049020228907465935, + 0.0036999473813921213, + 0.0018960732268169522, + 0.006541301496326923, + 0.0040653119795024395, + 0.0028174840845167637, + 0.0067488644272089005, + 0.00777416629716754, + 0.006934680975973606, + 0.004867204464972019, + 0.003992053214460611, + 0.008659783750772476, + 0.013704212382435799, + 0.005685725715011358, + 0.011070641689002514, + 0.0053938827477395535, + 0.0047140284441411495, + 0.0036128812935203314, + 0.0022469961550086737, + 0.006975421216338873, + 0.005486311390995979, + 0.003171201329678297, + 0.004960805643349886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.14283083379268646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_4/index_tensor([5565, 5565])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_4/index_tensor([5565, 5565])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f6becb4b878cb9cfeb3562ec5f09327695c9c70a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_4/index_tensor([5565, 5565])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This stunning, cinematic macro shot begins in extreme close-up, depicting the intricate, sun-bleached pattern of a shell necklace against the warm, sun-kissed skin of a teenage Pacific Islander girl's collarbone. Each small, off-white shell, varying in shape and texture, is strung together to create a striking, matte-like effect of natural imperfections. As she confidently and steadily walks along a city sidewalk, the camera slowly and steadily pans left, maintaining a shallow depth of field. This deliberate movement gradually reveals her profile, then her full face, defined by high cheekbones and a resolute expression; her dark, focused eyes gaze straight ahead, her jaw quietly firm. Her long, wavy black hair, streaked with sun-lit strands, lifts slightly in the city breeze, framing her face. She wears a simple coral cotton tank top, providing a vibrant pop of color. The background fades into a vibrant, out-of-focus bokeh of bustling city life—the blurred figures of pedestrians, the blurred glow of storefronts, and the fleeting flashes of passing cars in the bright afternoon sun. This warm light casts a golden glow on her skin and creates a soft rim light around her hair, enhancing the rich, cinematic quality of the scene and creating a powerful, focused atmosphere amidst the chaos of the city.", + "seed": 30180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_4/index_5565_seed_30180_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026681663934141397, + "vlm_logprob": -0.026681663934141397, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.026681663934141397, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06164464354515076 + ], + "opposite_logprob_tokens": [ + -2.9366445541381836 + ], + "logprob_reward": -0.06164464354515076, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06164464354515076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09843068569898605, + 0.010557468049228191, + 0.004710194654762745, + 0.024301500990986824, + 0.007459090556949377, + 0.020732926204800606, + 0.008697756566107273, + 0.03161963075399399, + 0.6239846348762512, + 0.16950611770153046 + ], + "credit_map": [ + 0.00844012014567852, + 0.009344026446342468, + 0.0072230324149131775, + 0.012991366907954216, + 0.016087615862488747, + 0.018486468121409416, + 0.012716230005025864, + 0.013141822069883347, + 0.0006865110481157899, + 0.0012551923282444477, + 0.0011976482346653938, + 0.001193914213217795, + 0.0014747204259037971, + 0.001922538154758513, + 0.001525669009424746, + 0.0013012748677283525, + 0.0003221580118406564, + 0.00044306309428066015, + 0.0009329927270300686, + 0.0005157993291504681, + 0.0006390053313225508, + 0.0004315997357480228, + 0.0005097279790788889, + 0.000915848242584616, + 0.0019053247524425387, + 0.0027415864169597626, + 0.002625103574246168, + 0.0042748888954520226, + 0.003264544066041708, + 0.0021892341319471598, + 0.004374056123197079, + 0.0029267636127769947, + 0.0007910779677331448, + 0.0008044701535254717, + 0.0008062844863161445, + 0.0009024948813021183, + 0.0007669856422580779, + 0.0012903899187222123, + 0.0012061592424288392, + 0.000891228555701673, + 0.004156893119215965, + 0.003592661116272211, + 0.0021189849358052015, + 0.0021427851170301437, + 0.00124304776545614, + 0.002830816898494959, + 0.0020850615110248327, + 0.0025626765564084053, + 0.001147494767792523, + 0.0010338559513911605, + 0.0012825696030631661, + 0.0012077286373823881, + 0.0007907277322374284, + 0.0013372123939916492, + 0.0008462049881927669, + 0.001051962492056191, + 0.005160072352737188, + 0.0037380883004516363, + 0.003334740875288844, + 0.0042298841290175915, + 0.0029217114206403494, + 0.004760516341775656, + 0.0028211474418640137, + 0.004653471056371927, + 0.02247406728565693, + 0.023564307019114494, + 0.019358433783054352, + 0.04987029358744621, + 0.25794947147369385, + 0.028583895415067673, + 0.057957615703344345, + 0.16422656178474426, + 0.006288166623562574, + 0.06860093027353287, + 0.0043404665775597095, + 0.06745726615190506, + 0.004969785921275616, + 0.003995492123067379, + 0.010595561005175114, + 0.0032584548462182283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.06164464354515076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309982433915138 + ], + "opposite_logprob_tokens": [ + -3.168099880218506 + ], + "logprob_reward": -0.04309982433915138, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309982433915138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18762630224227905, + 0.06373248249292374, + 0.013090608641505241, + 0.019916722550988197, + 0.017385080456733704, + 0.06045857444405556, + 0.019616348668932915, + 0.03202441707253456, + 0.44028639793395996, + 0.14586301147937775 + ], + "credit_map": [ + 0.025457581505179405, + 0.014646985568106174, + 0.026924453675746918, + 0.023723339661955833, + 0.02218533121049404, + 0.021210921928286552, + 0.023541167378425598, + 0.029936520382761955, + 0.004267184995114803, + 0.006211825180798769, + 0.009576489217579365, + 0.008163820952177048, + 0.007527827750891447, + 0.012337585911154747, + 0.006490760482847691, + 0.009156997315585613, + 0.001004946418106556, + 0.0011312926653772593, + 0.0031126863323152065, + 0.0016979437787085772, + 0.0021051804069429636, + 0.0010725645115599036, + 0.0016124241519719362, + 0.0013535715406760573, + 0.0016997258644551039, + 0.0016970449360087514, + 0.004222574178129435, + 0.0019014886347576976, + 0.0016820385353639722, + 0.0035115720238536596, + 0.002202721545472741, + 0.0029995576478540897, + 0.001638562767766416, + 0.0018725665286183357, + 0.0022700189147144556, + 0.002155026188120246, + 0.00143102474976331, + 0.002625667955726385, + 0.002658140379935503, + 0.0027340734377503395, + 0.004380485042929649, + 0.005880603101104498, + 0.007387019693851471, + 0.009361845441162586, + 0.004426661878824234, + 0.00693847332149744, + 0.01353220734745264, + 0.008551282808184624, + 0.0035761985927820206, + 0.0031111561693251133, + 0.0019293908262625337, + 0.0034377770498394966, + 0.001610597944818437, + 0.0017782810609787703, + 0.001622806186787784, + 0.0025501414202153683, + 0.0036798431538045406, + 0.00332035799510777, + 0.0038998136296868324, + 0.004006318747997284, + 0.003111598314717412, + 0.005356843117624521, + 0.0028560971841216087, + 0.005793544463813305, + 0.015257253311574459, + 0.01761598326265812, + 0.011427297256886959, + 0.01571550779044628, + 0.1755714863538742, + 0.02490311861038208, + 0.038195520639419556, + 0.14160025119781494, + 0.005358753725886345, + 0.057707611471414566, + 0.005098887253552675, + 0.05718449875712395, + 0.004714086186140776, + 0.0036645547952502966, + 0.00923091359436512, + 0.002903718501329422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.04309982433915138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052612656727433205 + ], + "opposite_logprob_tokens": [ + -5.255261421203613 + ], + "logprob_reward": -0.0052612656727433205, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052612656727433205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.136850968003273, + 0.17379681766033173, + 0.019131606444716454, + 0.05327283591032028, + 0.04113417863845825, + 0.018660306930541992, + 0.05304966866970062, + 0.04038994386792183, + 0.43705040216445923, + 0.026663323864340782 + ], + "credit_map": [ + 0.01609264500439167, + 0.013884368352591991, + 0.01495007611811161, + 0.017453202977776527, + 0.016303522512316704, + 0.01908309943974018, + 0.016276979818940163, + 0.022807080298662186, + 0.011669172905385494, + 0.018219532445073128, + 0.03760238364338875, + 0.021669190376996994, + 0.023607952520251274, + 0.02124987356364727, + 0.02075958251953125, + 0.01901913434267044, + 0.0015519632725045085, + 0.0018968352815136313, + 0.0023710497189313173, + 0.003087762277573347, + 0.0031702436972409487, + 0.0017931925831362605, + 0.0024644008371979, + 0.0027961591258645058, + 0.0028727506287395954, + 0.004245805088430643, + 0.005466976203024387, + 0.009270021691918373, + 0.006794838234782219, + 0.005639241077005863, + 0.01098431646823883, + 0.007998882792890072, + 0.003474883269518614, + 0.005209630820900202, + 0.004592659883201122, + 0.008116603828966618, + 0.003828116226941347, + 0.007413580548018217, + 0.00435347855091095, + 0.004145229700952768, + 0.0033920989371836185, + 0.002139277756214142, + 0.0017506007570773363, + 0.002317618578672409, + 0.0014587664045393467, + 0.0019797328859567642, + 0.0033506331965327263, + 0.002271577250212431, + 0.008251924999058247, + 0.007612953893840313, + 0.006583864800632, + 0.011445621959865093, + 0.0037222097162157297, + 0.004300636705011129, + 0.005323734134435654, + 0.005808720830827951, + 0.006683623418211937, + 0.006404863204807043, + 0.0044589173048734665, + 0.006778300274163485, + 0.0025008725933730602, + 0.005052190274000168, + 0.0029320192988961935, + 0.005579160992056131, + 0.014569531194865704, + 0.017207374796271324, + 0.01623368449509144, + 0.05195392668247223, + 0.11760841310024261, + 0.030076241120696068, + 0.024594061076641083, + 0.16480717062950134, + 0.0034352727234363556, + 0.004760683514177799, + 0.0039114439859986305, + 0.003706760471686721, + 0.0027154956478625536, + 0.0028316068928688765, + 0.003090625861659646, + 0.002211435232311487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0052612656727433205 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584875717759132 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.020584875717759132, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584875717759132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18375761806964874, + 0.029800988733768463, + 0.008815280161798, + 0.04674161598086357, + 0.019057001918554306, + 0.0465925857424736, + 0.10250520706176758, + 0.0512220524251461, + 0.31763163208961487, + 0.19387610256671906 + ], + "credit_map": [ + 0.018112236633896828, + 0.017489679157733917, + 0.019563836976885796, + 0.022299932315945625, + 0.026194535195827484, + 0.022989097982645035, + 0.02641342021524906, + 0.030694877728819847, + 0.0020290601532906294, + 0.0031659926753491163, + 0.003625117940828204, + 0.004193430300801992, + 0.0031921446789056063, + 0.005333172157406807, + 0.0034249890595674515, + 0.004837082698941231, + 0.000719510659109801, + 0.0010084324749186635, + 0.001165088382549584, + 0.0012838402763009071, + 0.0012778632808476686, + 0.0008502010605297983, + 0.0011720893671736121, + 0.0013382541947066784, + 0.0034449512604624033, + 0.004598753526806831, + 0.009298119693994522, + 0.006085141096264124, + 0.00534630473703146, + 0.007369804661720991, + 0.005704537499696016, + 0.004894002806395292, + 0.001985257724300027, + 0.0021860399283468723, + 0.002237086184322834, + 0.0037132767029106617, + 0.0021220226772129536, + 0.0028488743118941784, + 0.0020068311132490635, + 0.0019576128106564283, + 0.0046187699772417545, + 0.005258042365312576, + 0.006148321554064751, + 0.005928969010710716, + 0.002072978066280484, + 0.0052459449507296085, + 0.011084779165685177, + 0.006234781350940466, + 0.009802707470953465, + 0.013883997686207294, + 0.012000192888081074, + 0.023205455392599106, + 0.007226275745779276, + 0.008992180228233337, + 0.01492907665669918, + 0.012465320527553558, + 0.010013746097683907, + 0.006674695294350386, + 0.00486348569393158, + 0.006598959676921368, + 0.0039022115524858236, + 0.007638637442141771, + 0.003731748554855585, + 0.0077985674142837524, + 0.014948129653930664, + 0.01291046291589737, + 0.022823359817266464, + 0.08904672414064407, + 0.10191816836595535, + 0.018067140132188797, + 0.02812289260327816, + 0.029794717207551003, + 0.009401666931807995, + 0.07543753832578659, + 0.009710414335131645, + 0.07032454013824463, + 0.007021846249699593, + 0.006289037875831127, + 0.011244772002100945, + 0.004446282051503658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020584875717759132 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028177103959023952 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028177103959023952, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028177103959023952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0708245262503624, + 0.01276480033993721, + 0.0034214050974696875, + 0.01856003701686859, + 0.013018473982810974, + 0.024981243535876274, + 0.01646941713988781, + 0.05147943273186684, + 0.2822102904319763, + 0.5062702894210815 + ], + "credit_map": [ + 0.006552752573043108, + 0.007155093364417553, + 0.0065430509857833385, + 0.009090017527341843, + 0.01122275460511446, + 0.012069135904312134, + 0.009504308924078941, + 0.008687419816851616, + 0.0011621330631896853, + 0.0009764641872607172, + 0.0029802126809954643, + 0.0021786412689834833, + 0.0014349876437336206, + 0.0008763328660279512, + 0.001312227570451796, + 0.001843802398070693, + 0.00023711408721283078, + 0.0004853040736634284, + 0.00037841356243006885, + 0.0004682107246480882, + 0.0006611812859773636, + 0.0003533805429469794, + 0.0003064281481783837, + 0.0005313726142048836, + 0.0013310224749147892, + 0.0023507659789174795, + 0.0020630122162401676, + 0.003005041740834713, + 0.002237084088847041, + 0.002076792297884822, + 0.003158591454848647, + 0.0023377262987196445, + 0.0013369991211220622, + 0.0020653854589909315, + 0.0016099525382742286, + 0.0033038328401744366, + 0.0010152071481570601, + 0.0014158299891278148, + 0.0013626356376335025, + 0.0009086327045224607, + 0.004573123063892126, + 0.004747510887682438, + 0.002340311650186777, + 0.0031087035313248634, + 0.0018715830519795418, + 0.0028438398148864508, + 0.0030040720012038946, + 0.0024921002332121134, + 0.0012813258217647672, + 0.002138013020157814, + 0.0020908848382532597, + 0.004977424629032612, + 0.0011800582287833095, + 0.0014015500200912356, + 0.0017456019995734096, + 0.0016545599792152643, + 0.007139790803194046, + 0.007484794594347477, + 0.006386386230587959, + 0.00712035084143281, + 0.005394896026700735, + 0.006841205060482025, + 0.00451139360666275, + 0.006600619293749332, + 0.014654666185379028, + 0.01135653629899025, + 0.01981208845973015, + 0.06934872269630432, + 0.10794644802808762, + 0.01511092483997345, + 0.028619840741157532, + 0.01536109484732151, + 0.02695414610207081, + 0.18656796216964722, + 0.027816884219646454, + 0.19191603362560272, + 0.015051279217004776, + 0.013034646399319172, + 0.034072961658239365, + 0.010856421664357185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0028177103959023952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This stunning, cinematic macro shot begins in extreme close-up, depicting the intricate, sun-bleached pattern of a shell necklace against the warm, sun-kissed skin of a teenage Pacific Islander girl's collarbone. Each small, off-white shell, varying in shape and texture, is strung together to create a striking, matte-like effect of natural imperfections. As she confidently and steadily walks along a city sidewalk, the camera slowly and steadily pans left, maintaining a shallow depth of field. This deliberate movement gradually reveals her profile, then her full face, defined by high cheekbones and a resolute expression; her dark, focused eyes gaze straight ahead, her jaw quietly firm. Her long, wavy black hair, streaked with sun-lit strands, lifts slightly in the city breeze, framing her face. She wears a simple coral cotton tank top, providing a vibrant pop of color. The background fades into a vibrant, out-of-focus bokeh of bustling city life—the blurred figures of pedestrians, the blurred glow of storefronts, and the fleeting flashes of passing cars in the bright afternoon sun. This warm light casts a golden glow on her skin and creates a soft rim light around her hair, enhancing the rich, cinematic quality of the scene and creating a powerful, focused atmosphere amidst the chaos of the city.", + "seed": 30181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_4/index_5565_seed_30181_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.033205113094300034, + "vlm_logprob": -0.033205113094300034, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.033205113094300034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023087849840521812 + ], + "opposite_logprob_tokens": [ + -3.898087739944458 + ], + "logprob_reward": -0.023087849840521812, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023087849840521812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10077289491891861, + 0.12844955921173096, + 0.12175807356834412, + 0.028518132865428925, + 0.05299859121441841, + 0.10877872258424759, + 0.1957167536020279, + 0.08668576925992966, + 0.11607536673545837, + 0.06024608761072159 + ], + "credit_map": [ + 0.018282759934663773, + 0.01366292592138052, + 0.011547391302883625, + 0.00946112535893917, + 0.01229608803987503, + 0.01304639969021082, + 0.012036115862429142, + 0.01044008880853653, + 0.012256396003067493, + 0.012376435101032257, + 0.018842177465558052, + 0.014659135602414608, + 0.019673960283398628, + 0.012144580483436584, + 0.021320849657058716, + 0.017176032066345215, + 0.008434932678937912, + 0.019775474444031715, + 0.008734548464417458, + 0.018577562645077705, + 0.026240777224302292, + 0.012828304432332516, + 0.017559321597218513, + 0.009607149288058281, + 0.0022736431565135717, + 0.002546662464737892, + 0.0025313717778772116, + 0.0032036351040005684, + 0.005901495460420847, + 0.004066383466124535, + 0.003649843856692314, + 0.004345097113400698, + 0.00510754669085145, + 0.007191985845565796, + 0.007670190185308456, + 0.004184601362794638, + 0.011058934964239597, + 0.005953511688858271, + 0.005312396679073572, + 0.006519427057355642, + 0.01048724353313446, + 0.009104670956730843, + 0.012780467048287392, + 0.02174968272447586, + 0.0158349871635437, + 0.008975023403763771, + 0.01611778698861599, + 0.013728857971727848, + 0.012611334212124348, + 0.006743073929101229, + 0.048126671463251114, + 0.010504493489861488, + 0.023797757923603058, + 0.013830420561134815, + 0.06819076836109161, + 0.011912237852811813, + 0.01218885462731123, + 0.007140851113945246, + 0.0091590266674757, + 0.006171541288495064, + 0.018766576424241066, + 0.006619968917220831, + 0.013944965787231922, + 0.012693986296653748, + 0.015170853585004807, + 0.014867383055388927, + 0.01681484654545784, + 0.008056593127548695, + 0.022964302450418472, + 0.009265275672078133, + 0.01954818144440651, + 0.009387930855154991, + 0.010846557095646858, + 0.007001131772994995, + 0.008851786144077778, + 0.004935181699693203, + 0.010596590116620064, + 0.006415335461497307, + 0.005470465403050184, + 0.006129037588834763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.023087849840521812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005304843187332 + ], + "opposite_logprob_tokens": [ + -2.6950531005859375 + ], + "logprob_reward": -0.07005304843187332, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005304843187332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12099820375442505, + 0.5212714076042175, + 0.06340298056602478, + 0.10523899644613266, + 0.02803714945912361, + 0.04554534703493118, + 0.055228885263204575, + 0.021753044798970222, + 0.02432931587100029, + 0.014194673858582973 + ], + "credit_map": [ + 0.013020935468375683, + 0.014436354860663414, + 0.011781008914113045, + 0.010323882102966309, + 0.015885302796959877, + 0.024940483272075653, + 0.016828998923301697, + 0.013781231828033924, + 0.04856768250465393, + 0.06033310294151306, + 0.07504832744598389, + 0.06840725243091583, + 0.0760079100728035, + 0.03860719874501228, + 0.09149804711341858, + 0.06280186772346497, + 0.004280705936253071, + 0.010514737106859684, + 0.006584173534065485, + 0.0075642699375748634, + 0.009534096345305443, + 0.0072946189902722836, + 0.011041765101253986, + 0.006588613614439964, + 0.008064146153628826, + 0.008359654806554317, + 0.009853934869170189, + 0.01549622230231762, + 0.01698130927979946, + 0.013428319245576859, + 0.01889377273619175, + 0.014161628670990467, + 0.003493035677820444, + 0.002553027356043458, + 0.0048278807662427425, + 0.0029942344408482313, + 0.004565520212054253, + 0.0031819683499634266, + 0.0040496033616364, + 0.002371878130361438, + 0.00373639864847064, + 0.00340446294285357, + 0.0044095986522734165, + 0.009973721578717232, + 0.006985359359532595, + 0.004742683842778206, + 0.005482194945216179, + 0.006810924969613552, + 0.004079975187778473, + 0.0015277581987902522, + 0.01277098897844553, + 0.002785458229482174, + 0.006666587200015783, + 0.0031729473266750574, + 0.020628295838832855, + 0.0035968665033578873, + 0.003230515168979764, + 0.0019696278031915426, + 0.002031563315540552, + 0.001189615111798048, + 0.0046662078239023685, + 0.002857876941561699, + 0.003408282296732068, + 0.002399355173110962, + 0.0034892912954092026, + 0.0020940133836120367, + 0.002703330246731639, + 0.0021572711411863565, + 0.003711084369570017, + 0.004030567593872547, + 0.004000286106020212, + 0.0021434698719531298, + 0.002190627623349428, + 0.00185636174865067, + 0.002078291028738022, + 0.0014931948389858007, + 0.0023344114888459444, + 0.0011471707839518785, + 0.0019312456715852022, + 0.0011633704416453838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.07005304843187332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102861508727074 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.004102861508727074, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102861508727074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24809522926807404, + 0.10092700272798538, + 0.16739283502101898, + 0.17695100605487823, + 0.04488668590784073, + 0.0690016970038414, + 0.07642290741205215, + 0.04360925033688545, + 0.04657188430428505, + 0.026141470298171043 + ], + "credit_map": [ + 0.029511721804738045, + 0.02698943018913269, + 0.02238561026751995, + 0.023131396621465683, + 0.026234833523631096, + 0.046258680522441864, + 0.04447224736213684, + 0.029111292213201523, + 0.010831957682967186, + 0.013783585280179977, + 0.012422291561961174, + 0.009718368761241436, + 0.01476033590734005, + 0.00906718522310257, + 0.01808934286236763, + 0.01225393544882536, + 0.009783336892724037, + 0.026427708566188812, + 0.011846916750073433, + 0.024620383977890015, + 0.03304038941860199, + 0.017001984640955925, + 0.0296782236546278, + 0.014993882738053799, + 0.009221581742167473, + 0.02536756731569767, + 0.015587433241307735, + 0.02643338218331337, + 0.02603583037853241, + 0.01626584492623806, + 0.04168759658932686, + 0.01635175757110119, + 0.004244199022650719, + 0.004747504834085703, + 0.004571851808577776, + 0.00491482624784112, + 0.0061051687225699425, + 0.004157036542892456, + 0.009815115481615067, + 0.006330983713269234, + 0.004618760664016008, + 0.00775762926787138, + 0.007811117451637983, + 0.01603994332253933, + 0.009960518218576908, + 0.003881655167788267, + 0.011850577779114246, + 0.007081500720232725, + 0.0056846123188734055, + 0.005609387997537851, + 0.0071891979314386845, + 0.00850817933678627, + 0.017621293663978577, + 0.013699078001081944, + 0.009510193020105362, + 0.008600965142250061, + 0.007979819551110268, + 0.004194512497633696, + 0.004429432097822428, + 0.0022857554722577333, + 0.009037299081683159, + 0.0035272210370749235, + 0.007080510258674622, + 0.005074700806289911, + 0.007983789779245853, + 0.004018818493932486, + 0.004281037487089634, + 0.004383210092782974, + 0.007280881516635418, + 0.005518998019397259, + 0.009211051277816296, + 0.0038940981030464172, + 0.003632548963651061, + 0.003265306819230318, + 0.003996358253061771, + 0.003074402455240488, + 0.004150158725678921, + 0.0025448312517255545, + 0.0031125738751143217, + 0.002365290420129895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.004102861508727074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204526126384735 + ], + "opposite_logprob_tokens": [ + -2.8120453357696533 + ], + "logprob_reward": -0.06204526126384735, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204526126384735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07552793622016907, + 0.151402086019516, + 0.09504109621047974, + 0.05720068886876106, + 0.129531592130661, + 0.08335191011428833, + 0.2987631559371948, + 0.03862060233950615, + 0.03073650784790516, + 0.039824388921260834 + ], + "credit_map": [ + 0.009684440679848194, + 0.008764717727899551, + 0.006017493084073067, + 0.00831189751625061, + 0.009886342100799084, + 0.014253399334847927, + 0.01003960520029068, + 0.008570040576159954, + 0.013838373124599457, + 0.015261330641806126, + 0.019725507125258446, + 0.016091281548142433, + 0.023729095235466957, + 0.014372944831848145, + 0.02644389495253563, + 0.021939652040600777, + 0.009396499954164028, + 0.012250146828591824, + 0.008631511591374874, + 0.015396525152027607, + 0.018661072477698326, + 0.011412243358790874, + 0.011510328389704227, + 0.007782759610563517, + 0.005229300353676081, + 0.005926542915403843, + 0.006077636498957872, + 0.006846948992460966, + 0.0104823661968112, + 0.006612665485590696, + 0.00961148552596569, + 0.006413740571588278, + 0.010988330468535423, + 0.015804806724190712, + 0.02318417839705944, + 0.012485967017710209, + 0.02366875484585762, + 0.015235795639455318, + 0.01429952122271061, + 0.013864235952496529, + 0.006241658702492714, + 0.008007451891899109, + 0.00786706991493702, + 0.012071117758750916, + 0.014446012675762177, + 0.009266158565878868, + 0.013131068088114262, + 0.012321376241743565, + 0.01926993764936924, + 0.007593504153192043, + 0.08612080663442612, + 0.014969319105148315, + 0.03105046972632408, + 0.011927738785743713, + 0.1096646785736084, + 0.01816672272980213, + 0.005348145496100187, + 0.0027453156653791666, + 0.0056050559505820274, + 0.0022843542974442244, + 0.00602054875344038, + 0.006357710808515549, + 0.005455922335386276, + 0.00480355229228735, + 0.004741252399981022, + 0.004106974694877863, + 0.002861748682335019, + 0.00249437615275383, + 0.00523685198277235, + 0.004299025051295757, + 0.004149693064391613, + 0.002846585353836417, + 0.008808222599327564, + 0.002700639655813575, + 0.00784379430115223, + 0.0037714149802923203, + 0.003479823237285018, + 0.0038121140096336603, + 0.004902935121208429, + 0.004505445249378681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.06204526126384735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067365444265306 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.0067365444265306, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067365444265306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0905187651515007, + 0.09252392500638962, + 0.05032320320606232, + 0.03806066885590553, + 0.056890442967414856, + 0.10127443820238113, + 0.18760338425636292, + 0.04511037468910217, + 0.23054194450378418, + 0.10715276747941971 + ], + "credit_map": [ + 0.013355151750147343, + 0.010895995423197746, + 0.009154644794762135, + 0.010294411331415176, + 0.009225835092365742, + 0.01618112064898014, + 0.01101108267903328, + 0.010400531813502312, + 0.008376561105251312, + 0.010244209319353104, + 0.012425022199749947, + 0.010338077321648598, + 0.014897207729518414, + 0.00805285107344389, + 0.019342178478837013, + 0.008847825229167938, + 0.005393593106418848, + 0.006217254791408777, + 0.005877841729670763, + 0.00569685734808445, + 0.005357870366424322, + 0.007724426221102476, + 0.007656730245798826, + 0.006398631725460291, + 0.0029019115027040243, + 0.003697626292705536, + 0.0030775347258895636, + 0.004828049335628748, + 0.005760887172073126, + 0.007196884136646986, + 0.005705665331333876, + 0.0048921117559075356, + 0.0062049394473433495, + 0.007964199408888817, + 0.0062242308631539345, + 0.004662226419895887, + 0.00978067796677351, + 0.007515314966440201, + 0.006943808402866125, + 0.00759505107998848, + 0.01687910594046116, + 0.008185038343071938, + 0.014848629012703896, + 0.011769689619541168, + 0.015925556421279907, + 0.006257866974920034, + 0.01797940395772457, + 0.009429163299500942, + 0.013682734221220016, + 0.009755461476743221, + 0.03524619713425636, + 0.01130448654294014, + 0.017730552703142166, + 0.048282407224178314, + 0.030063200742006302, + 0.02153835818171501, + 0.007596560753881931, + 0.0034492528066039085, + 0.006260768976062536, + 0.0021595158614218235, + 0.00815922673791647, + 0.00453962804749608, + 0.007593606598675251, + 0.005351819097995758, + 0.055877525359392166, + 0.01534392312169075, + 0.027422040700912476, + 0.016698535531759262, + 0.024159180000424385, + 0.04465211555361748, + 0.019874241203069687, + 0.026514410972595215, + 0.022511078044772148, + 0.011214328929781914, + 0.017069416120648384, + 0.00833407137542963, + 0.012718666344881058, + 0.013009071350097656, + 0.0108156967908144, + 0.011480451561510563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0067365444265306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This stunning, cinematic macro shot begins in extreme close-up, depicting the intricate, sun-bleached pattern of a shell necklace against the warm, sun-kissed skin of a teenage Pacific Islander girl's collarbone. Each small, off-white shell, varying in shape and texture, is strung together to create a striking, matte-like effect of natural imperfections. As she confidently and steadily walks along a city sidewalk, the camera slowly and steadily pans left, maintaining a shallow depth of field. This deliberate movement gradually reveals her profile, then her full face, defined by high cheekbones and a resolute expression; her dark, focused eyes gaze straight ahead, her jaw quietly firm. Her long, wavy black hair, streaked with sun-lit strands, lifts slightly in the city breeze, framing her face. She wears a simple coral cotton tank top, providing a vibrant pop of color. The background fades into a vibrant, out-of-focus bokeh of bustling city life—the blurred figures of pedestrians, the blurred glow of storefronts, and the fleeting flashes of passing cars in the bright afternoon sun. This warm light casts a golden glow on her skin and creates a soft rim light around her hair, enhancing the rich, cinematic quality of the scene and creating a powerful, focused atmosphere amidst the chaos of the city.", + "seed": 30182, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_4/index_5565_seed_30182_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02345373681746423, + "vlm_logprob": -0.02345373681746423, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.02345373681746423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025242777541279793 + ], + "opposite_logprob_tokens": [ + -3.775242805480957 + ], + "logprob_reward": -0.025242777541279793, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025242777541279793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2771126627922058, + 0.034820783883333206, + 0.08464360237121582, + 0.039898380637168884, + 0.07306130975484848, + 0.07012318074703217, + 0.07196477800607681, + 0.14388033747673035, + 0.10373726487159729, + 0.1007576510310173 + ], + "credit_map": [ + 0.06266019493341446, + 0.055255550891160965, + 0.029983950778841972, + 0.029769666492938995, + 0.026467230170965195, + 0.014615483582019806, + 0.03654414042830467, + 0.02181645669043064, + 0.002636341145262122, + 0.0034666615538299084, + 0.0039342246018350124, + 0.004848471377044916, + 0.004368973895907402, + 0.0032852846197783947, + 0.006184845697134733, + 0.006095979828387499, + 0.006244950462132692, + 0.005398824345320463, + 0.032100796699523926, + 0.009547201916575432, + 0.010882358998060226, + 0.003542544785887003, + 0.007725670002400875, + 0.009201263077557087, + 0.005028336308896542, + 0.0046860221773386, + 0.004426429979503155, + 0.007443526294082403, + 0.0028527136892080307, + 0.0028382812160998583, + 0.005647334735840559, + 0.006975735537707806, + 0.010012935847043991, + 0.015316539444029331, + 0.010137385688722134, + 0.008657229132950306, + 0.009733644314110279, + 0.0069242739118635654, + 0.006983320694416761, + 0.005295983050018549, + 0.009334194473922253, + 0.010226737707853317, + 0.01416417770087719, + 0.010055435821413994, + 0.010299489833414555, + 0.004747265484184027, + 0.007943051867187023, + 0.0033528294879943132, + 0.007146439980715513, + 0.010916459374129772, + 0.01104956679046154, + 0.008069916628301144, + 0.009754685685038567, + 0.004291278310120106, + 0.014926127158105373, + 0.005810304544866085, + 0.020307881757616997, + 0.011655834503471851, + 0.016476431861519814, + 0.010906721465289593, + 0.018384354189038277, + 0.007759404834359884, + 0.04924175515770912, + 0.009147952310740948, + 0.015369869768619537, + 0.03434549272060394, + 0.012359409593045712, + 0.01028132438659668, + 0.00911355298012495, + 0.004772896412760019, + 0.012069136835634708, + 0.005425583571195602, + 0.017088547348976135, + 0.009134052321314812, + 0.00948647316545248, + 0.00985602755099535, + 0.019013091921806335, + 0.01134489569813013, + 0.01708422601222992, + 0.007750328164547682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.025242777541279793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043096285313367844 + ], + "opposite_logprob_tokens": [ + -3.1680963039398193 + ], + "logprob_reward": -0.043096285313367844, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043096285313367844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22762002050876617, + 0.0782051756978035, + 0.36326542496681213, + 0.03881954401731491, + 0.03216857463121414, + 0.0722985789179802, + 0.02940196357667446, + 0.0704287588596344, + 0.04231832176446915, + 0.0454736202955246 + ], + "credit_map": [ + 0.02567065879702568, + 0.03998064249753952, + 0.02939549833536148, + 0.02240748517215252, + 0.02202226035296917, + 0.024739263579249382, + 0.041717905551195145, + 0.021686295047402382, + 0.005346179008483887, + 0.010620411485433578, + 0.010633600875735283, + 0.009639423340559006, + 0.007935275323688984, + 0.0062295738607645035, + 0.015863047912716866, + 0.011937662027776241, + 0.020300887525081635, + 0.05113264545798302, + 0.15118473768234253, + 0.04090775176882744, + 0.011267464607954025, + 0.009810148738324642, + 0.02946295030415058, + 0.049198828637599945, + 0.005542973522096872, + 0.0052431849762797356, + 0.0028623042162507772, + 0.0077482424676418304, + 0.002873690566048026, + 0.004633182194083929, + 0.005414165556430817, + 0.004501798655837774, + 0.005503168795257807, + 0.004912675824016333, + 0.004777889233082533, + 0.004526473581790924, + 0.003923037089407444, + 0.0023377700708806515, + 0.003857395378872752, + 0.002330165356397629, + 0.0041472879238426685, + 0.006796219386160374, + 0.008191998116672039, + 0.007600302807986736, + 0.026758071035146713, + 0.005467872601002455, + 0.010759511031210423, + 0.0025773104280233383, + 0.0025257430970668793, + 0.007017703261226416, + 0.004838235210627317, + 0.004007371608167887, + 0.003545651910826564, + 0.0013457867316901684, + 0.004038349259644747, + 0.0020831229630857706, + 0.014370517805218697, + 0.007407030090689659, + 0.007916771806776524, + 0.004950143396854401, + 0.007930639199912548, + 0.004562032409012318, + 0.018088920041918755, + 0.0052027106285095215, + 0.006004307884722948, + 0.006847004871815443, + 0.006127981469035149, + 0.005282052326947451, + 0.004482409916818142, + 0.0034866086207330227, + 0.004238387569785118, + 0.005849569104611874, + 0.006139840465039015, + 0.0047992561012506485, + 0.007711330894380808, + 0.0053000678308308125, + 0.006494603585451841, + 0.004770859610289335, + 0.00694747269153595, + 0.003310187254101038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.043096285313367844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003199935657903552 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003199935657903552, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003199935657903552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15463215112686157, + 0.10953547060489655, + 0.34710854291915894, + 0.10471981018781662, + 0.030367130413651466, + 0.02672920562326908, + 0.03183427080512047, + 0.1372699737548828, + 0.03950287029147148, + 0.018300607800483704 + ], + "credit_map": [ + 0.026057902723550797, + 0.028844494372606277, + 0.016209129244089127, + 0.0162724107503891, + 0.01528427004814148, + 0.011569560505449772, + 0.021608691662549973, + 0.01878567785024643, + 0.007183464244008064, + 0.011824991554021835, + 0.012552172876894474, + 0.018339093774557114, + 0.011421553790569305, + 0.010407274588942528, + 0.018820596858859062, + 0.01898632012307644, + 0.01730775646865368, + 0.06905349344015121, + 0.13579697906970978, + 0.030623847618699074, + 0.016282308846712112, + 0.008980153128504753, + 0.028603458777070045, + 0.040460508316755295, + 0.004429067485034466, + 0.007739518769085407, + 0.0038732716348022223, + 0.043216172605752945, + 0.0034862435422837734, + 0.01679924502968788, + 0.02163810469210148, + 0.0035381847992539406, + 0.004838143475353718, + 0.006424545310437679, + 0.004831570666283369, + 0.0036133029498159885, + 0.0028462097980082035, + 0.002557589439675212, + 0.0027824509888887405, + 0.002473316853865981, + 0.0021690146531909704, + 0.002890769625082612, + 0.004195474553853273, + 0.0056574102491140366, + 0.004219012800604105, + 0.00308994110673666, + 0.0024508412461727858, + 0.002056739991530776, + 0.0032952423207461834, + 0.007110847160220146, + 0.005057783797383308, + 0.004651891998946667, + 0.0033036326058208942, + 0.002221113769337535, + 0.003726780414581299, + 0.002466976409777999, + 0.02522347867488861, + 0.010685452260077, + 0.014256490394473076, + 0.009083444252610207, + 0.016961580142378807, + 0.007279483135789633, + 0.04427112638950348, + 0.009508916176855564, + 0.005238776095211506, + 0.008214137516915798, + 0.0039429129101336, + 0.004958933684974909, + 0.003866568673402071, + 0.003221285529434681, + 0.005340528208762407, + 0.00471972580999136, + 0.0020079459063708782, + 0.00294910604134202, + 0.002170585561543703, + 0.0023311376571655273, + 0.0025649345479905605, + 0.0018314113840460777, + 0.002746416023001075, + 0.0016990697477012873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.003199935657903552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810097277164459 + ], + "opposite_logprob_tokens": [ + -3.2881009578704834 + ], + "logprob_reward": -0.03810097277164459, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810097277164459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16680164635181427, + 0.05750136822462082, + 0.24584169685840607, + 0.06165693700313568, + 0.05473613366484642, + 0.14181214570999146, + 0.060032714158296585, + 0.13101641833782196, + 0.044951051473617554, + 0.03564982861280441 + ], + "credit_map": [ + 0.02767120487987995, + 0.03590194135904312, + 0.015519270673394203, + 0.014239943586289883, + 0.015217541716992855, + 0.01436927355825901, + 0.026660041883587837, + 0.017222467809915543, + 0.00450525525957346, + 0.008634866215288639, + 0.005375318694859743, + 0.010734070092439651, + 0.008892759680747986, + 0.004666794091463089, + 0.008860275149345398, + 0.0058320374228060246, + 0.012284684926271439, + 0.04207448661327362, + 0.09664785861968994, + 0.02404516376554966, + 0.011769169941544533, + 0.007321331650018692, + 0.019955499097704887, + 0.031743526458740234, + 0.007516221143305302, + 0.007967119105160236, + 0.0055585275404155254, + 0.015835804864764214, + 0.003366141114383936, + 0.006144117563962936, + 0.011640319600701332, + 0.0036287002731114626, + 0.007393942680209875, + 0.006328432820737362, + 0.00873566884547472, + 0.006450179498642683, + 0.010028811171650887, + 0.0051871491596102715, + 0.004514796659350395, + 0.006097155623137951, + 0.018405918031930923, + 0.02237125299870968, + 0.022193655371665955, + 0.02177342399954796, + 0.026879599317908287, + 0.008484709076583385, + 0.014091924764215946, + 0.0076116761192679405, + 0.008446523919701576, + 0.008137761615216732, + 0.014735853299498558, + 0.006688877008855343, + 0.0062185730785131454, + 0.002748129889369011, + 0.008756197988986969, + 0.00430080434307456, + 0.02281743660569191, + 0.012400750070810318, + 0.018168212845921516, + 0.010346445254981518, + 0.016371222212910652, + 0.007351280655711889, + 0.033241309225559235, + 0.010319792665541172, + 0.006004770286381245, + 0.009431831538677216, + 0.004833663813769817, + 0.003961197566241026, + 0.006847021635621786, + 0.0029010928701609373, + 0.007070448715239763, + 0.003901031333953142, + 0.005405120085924864, + 0.007527383975684643, + 0.0077940840274095535, + 0.004355099983513355, + 0.003144959220662713, + 0.00199318490922451, + 0.003160553751513362, + 0.0022694470826536417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.03810097277164459 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076287128031253815 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.0076287128031253815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076287128031253815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041292693465948105, + 0.024263450875878334, + 0.03768331930041313, + 0.056915175169706345, + 0.057045161724090576, + 0.05953533947467804, + 0.05328204110264778, + 0.49879390001296997, + 0.1304377168416977, + 0.040751248598098755 + ], + "credit_map": [ + 0.0038541103713214397, + 0.0034668773878365755, + 0.004422055557370186, + 0.004331821575760841, + 0.003528582165017724, + 0.005602202378213406, + 0.004901889245957136, + 0.01118515431880951, + 0.001451021060347557, + 0.002632620744407177, + 0.002814731327816844, + 0.003528458531945944, + 0.0018818887183442712, + 0.0026658503338694572, + 0.003644559532403946, + 0.005644321441650391, + 0.002989477012306452, + 0.005538996309041977, + 0.008689879439771175, + 0.007147922180593014, + 0.003375724656507373, + 0.0020687563810497522, + 0.0035184938460588455, + 0.004354072734713554, + 0.003096712986007333, + 0.0045763649977743626, + 0.002377411350607872, + 0.023354584351181984, + 0.0019063837826251984, + 0.0068392520770430565, + 0.01243047509342432, + 0.0023339928593486547, + 0.008331980556249619, + 0.01347501389682293, + 0.008607135154306889, + 0.006317345891147852, + 0.006007632706314325, + 0.005575955845415592, + 0.005040457472205162, + 0.003689639503136277, + 0.0031028175726532936, + 0.005310709122568369, + 0.004564852919429541, + 0.00880750548094511, + 0.02042287029325962, + 0.0056589702144265175, + 0.00927148386836052, + 0.0023961274418979883, + 0.007439094595611095, + 0.007821660488843918, + 0.012332171201705933, + 0.007237262558192015, + 0.005931473337113857, + 0.0026192183140665293, + 0.006203650496900082, + 0.0036975103430449963, + 0.09172111749649048, + 0.034837085753679276, + 0.05452801659703255, + 0.03062562644481659, + 0.0638309121131897, + 0.024843042716383934, + 0.1644786149263382, + 0.0339294970035553, + 0.01869063451886177, + 0.04879984259605408, + 0.015487871132791042, + 0.014100326225161552, + 0.008836161345243454, + 0.004900341387838125, + 0.012891112826764584, + 0.006731440778821707, + 0.006694938987493515, + 0.005959162022918463, + 0.011336266063153744, + 0.005432541016489267, + 0.00320514221675694, + 0.0020250696688890457, + 0.0037141023203730583, + 0.0023840281646698713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0076287128031253815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This stunning, cinematic macro shot begins in extreme close-up, depicting the intricate, sun-bleached pattern of a shell necklace against the warm, sun-kissed skin of a teenage Pacific Islander girl's collarbone. Each small, off-white shell, varying in shape and texture, is strung together to create a striking, matte-like effect of natural imperfections. As she confidently and steadily walks along a city sidewalk, the camera slowly and steadily pans left, maintaining a shallow depth of field. This deliberate movement gradually reveals her profile, then her full face, defined by high cheekbones and a resolute expression; her dark, focused eyes gaze straight ahead, her jaw quietly firm. Her long, wavy black hair, streaked with sun-lit strands, lifts slightly in the city breeze, framing her face. She wears a simple coral cotton tank top, providing a vibrant pop of color. The background fades into a vibrant, out-of-focus bokeh of bustling city life—the blurred figures of pedestrians, the blurred glow of storefronts, and the fleeting flashes of passing cars in the bright afternoon sun. This warm light casts a golden glow on her skin and creates a soft rim light around her hair, enhancing the rich, cinematic quality of the scene and creating a powerful, focused atmosphere amidst the chaos of the city.", + "seed": 30183, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_4/index_5565_seed_30183_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0631343905813992, + "vlm_logprob": -0.0631343905813992, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0631343905813992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.264884889125824 + ], + "opposite_logprob_tokens": [ + -1.5148849487304688 + ], + "logprob_reward": -0.264884889125824, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.264884889125824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08212877064943314, + 0.06893002986907959, + 0.03670622408390045, + 0.08683452755212784, + 0.14248915016651154, + 0.20045430958271027, + 0.07832358032464981, + 0.08736100792884827, + 0.1751444786787033, + 0.04162788763642311 + ], + "credit_map": [ + 0.012115842662751675, + 0.015331084839999676, + 0.010315751656889915, + 0.010577693581581116, + 0.008572499267756939, + 0.0062989951111376286, + 0.009656185284256935, + 0.009260720573365688, + 0.006883848924189806, + 0.010450526140630245, + 0.00516326492652297, + 0.010554471984505653, + 0.007641533389687538, + 0.010065850801765919, + 0.009238586761057377, + 0.008931949734687805, + 0.004001880995929241, + 0.005642385222017765, + 0.00455123046413064, + 0.00465345149859786, + 0.0038848838303238153, + 0.004333900287747383, + 0.004531717859208584, + 0.005106775090098381, + 0.011897395364940166, + 0.008094937540590763, + 0.011147542856633663, + 0.014390736818313599, + 0.009436240419745445, + 0.00805280078202486, + 0.01409096084535122, + 0.009723911993205547, + 0.006490800529718399, + 0.013559672050178051, + 0.016648586839437485, + 0.03895070403814316, + 0.01863286644220352, + 0.010130684822797775, + 0.02487960085272789, + 0.013196243904531002, + 0.021765580400824547, + 0.016575030982494354, + 0.03008325584232807, + 0.031802840530872345, + 0.018051426857709885, + 0.02682613395154476, + 0.029463956132531166, + 0.025886090472340584, + 0.00829868484288454, + 0.009512456133961678, + 0.006681670900434256, + 0.012116760015487671, + 0.008806527592241764, + 0.009617029689252377, + 0.012544677592813969, + 0.010745772160589695, + 0.007095261011272669, + 0.00796340312808752, + 0.010968592017889023, + 0.015409485436975956, + 0.01413396093994379, + 0.012544086202979088, + 0.0094070415943861, + 0.009839178994297981, + 0.00990158412605524, + 0.019735092297196388, + 0.019583353772759438, + 0.017079226672649384, + 0.03999669477343559, + 0.026215234771370888, + 0.029790280386805534, + 0.012843004427850246, + 0.0037859980948269367, + 0.004669826477766037, + 0.00427357479929924, + 0.0036951093934476376, + 0.004393635783344507, + 0.005696265026926994, + 0.00621981592848897, + 0.008893661201000214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.264884889125824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016114704310894012 + ], + "opposite_logprob_tokens": [ + -4.141114711761475 + ], + "logprob_reward": -0.016114704310894012, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016114704310894012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22105880081653595, + 0.0768643319606781, + 0.12815842032432556, + 0.09819305688142776, + 0.0850539430975914, + 0.05788327753543854, + 0.11588072776794434, + 0.0824701339006424, + 0.07016491889953613, + 0.06427238136529922 + ], + "credit_map": [ + 0.030443580821156502, + 0.022937258705496788, + 0.03340991586446762, + 0.028107767924666405, + 0.040278565138578415, + 0.02172940969467163, + 0.027014248073101044, + 0.017138035967946053, + 0.00806156825274229, + 0.010122661478817463, + 0.008202644996345043, + 0.012893767096102238, + 0.006581699009984732, + 0.01214515883475542, + 0.00950746238231659, + 0.009349365718662739, + 0.017011526972055435, + 0.01573232002556324, + 0.015537263825535774, + 0.017382005229592323, + 0.012658697552978992, + 0.01188105158507824, + 0.02054709754884243, + 0.01740846037864685, + 0.01082666963338852, + 0.011646359227597713, + 0.013232444413006306, + 0.018386008217930794, + 0.011609895154833794, + 0.006566253956407309, + 0.01481721643358469, + 0.011108210310339928, + 0.007848923094570637, + 0.012540126219391823, + 0.012734218500554562, + 0.010006611235439777, + 0.011652952060103416, + 0.008879045955836773, + 0.014049533754587173, + 0.0073425257578492165, + 0.004271895624697208, + 0.0036194908898323774, + 0.005512196570634842, + 0.01048713456839323, + 0.005355570465326309, + 0.009119962342083454, + 0.010778790339827538, + 0.008738238364458084, + 0.009163693524897099, + 0.013348760083317757, + 0.007863194681704044, + 0.007831014692783356, + 0.015521853230893612, + 0.03263668343424797, + 0.011334561742842197, + 0.01818097196519375, + 0.007802372332662344, + 0.012752121314406395, + 0.00700916163623333, + 0.007773345336318016, + 0.011412328109145164, + 0.014504100196063519, + 0.010033508762717247, + 0.011183197610080242, + 0.006201384123414755, + 0.008318755775690079, + 0.004565115552395582, + 0.004760191310197115, + 0.015282922424376011, + 0.009703841991722584, + 0.011452087201178074, + 0.009880618192255497, + 0.006873424630612135, + 0.008103668689727783, + 0.006438090931624174, + 0.007874754257500172, + 0.009825939312577248, + 0.006860928609967232, + 0.011599521152675152, + 0.006696050986647606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.016114704310894012 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747082807123661 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747082807123661, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747082807123661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19915622472763062, + 0.11147370934486389, + 0.14905719459056854, + 0.10983730852603912, + 0.11890755593776703, + 0.0317147821187973, + 0.10799845308065414, + 0.037978142499923706, + 0.05716920644044876, + 0.07670743763446808 + ], + "credit_map": [ + 0.032186344265937805, + 0.018709301948547363, + 0.04059768468141556, + 0.03843081369996071, + 0.0200472604483366, + 0.012137072160840034, + 0.023191779851913452, + 0.013855995610356331, + 0.013766899704933167, + 0.015013204887509346, + 0.014236767776310444, + 0.020435433834791183, + 0.011032790876924992, + 0.010170361958444118, + 0.0162561796605587, + 0.01056208461523056, + 0.03277863189578056, + 0.018339039757847786, + 0.015283321030437946, + 0.017050882801413536, + 0.021077176555991173, + 0.012240713462233543, + 0.019979076460003853, + 0.012308360077440739, + 0.00902475230395794, + 0.011224946938455105, + 0.013395891524851322, + 0.018785560503602028, + 0.011672166176140308, + 0.00925357360392809, + 0.020872976630926132, + 0.015607448294758797, + 0.007434002123773098, + 0.011695457622408867, + 0.01151216123253107, + 0.0290116798132658, + 0.014263873919844627, + 0.011742942035198212, + 0.01638653129339218, + 0.01686091534793377, + 0.0027331640012562275, + 0.0023034317418932915, + 0.0037909189704805613, + 0.005395603831857443, + 0.0032695329282432795, + 0.004816337954252958, + 0.004610938485711813, + 0.004794853739440441, + 0.012178732082247734, + 0.01580432616174221, + 0.01030690036714077, + 0.010026175528764725, + 0.012949001044034958, + 0.018950888887047768, + 0.010089692659676075, + 0.01769273541867733, + 0.0038549399469047785, + 0.003968759439885616, + 0.00489101093262434, + 0.004325856454670429, + 0.005086516495794058, + 0.006163882091641426, + 0.0055640642531216145, + 0.004123116377741098, + 0.0046727098524570465, + 0.0056458874605596066, + 0.008397677913308144, + 0.007274770177900791, + 0.006564687006175518, + 0.00839090533554554, + 0.008833696134388447, + 0.007388876751065254, + 0.007400102913379669, + 0.00827572587877512, + 0.005632851272821426, + 0.007253963500261307, + 0.009995092637836933, + 0.009252315387129784, + 0.014210822992026806, + 0.014686569571495056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.006747082807123661 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288074880838394 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023288074880838394, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288074880838394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03985327109694481, + 0.038338541984558105, + 0.059170108288526535, + 0.049075983464717865, + 0.08915339410305023, + 0.053447872400283813, + 0.14438261091709137, + 0.16342537105083466, + 0.1968565434217453, + 0.16629625856876373 + ], + "credit_map": [ + 0.006469511892646551, + 0.005473315715789795, + 0.006826432887464762, + 0.005813934374600649, + 0.002767522819340229, + 0.0036349871661514044, + 0.004310211166739464, + 0.0045573594979941845, + 0.004378653131425381, + 0.005009257700294256, + 0.004356393590569496, + 0.007615272421389818, + 0.004901268519461155, + 0.0026466010604053736, + 0.005649368744343519, + 0.0037817282136529684, + 0.014454267919063568, + 0.007478024810552597, + 0.006616109516471624, + 0.006450239568948746, + 0.007765879854559898, + 0.00392424501478672, + 0.006928082089871168, + 0.005553268361836672, + 0.005550738889724016, + 0.0057157487608492374, + 0.007813358679413795, + 0.008709282614290714, + 0.004689755383878946, + 0.0031270566396415234, + 0.0062401629984378815, + 0.007229886949062347, + 0.006229009944945574, + 0.010868790559470654, + 0.009704574942588806, + 0.009675832465291023, + 0.011010373011231422, + 0.011261415667831898, + 0.018759645521640778, + 0.011643760837614536, + 0.00540428701788187, + 0.004336160607635975, + 0.005863436963409185, + 0.007103501353412867, + 0.005286675412207842, + 0.008678643964231014, + 0.0077268765307962894, + 0.009048295207321644, + 0.012117949314415455, + 0.017870860174298286, + 0.011163001880049706, + 0.010253272019326687, + 0.024020962417125702, + 0.03569437563419342, + 0.011339259333908558, + 0.021922945976257324, + 0.017243338748812675, + 0.020128708332777023, + 0.020119130611419678, + 0.02124534547328949, + 0.022506386041641235, + 0.026538273319602013, + 0.019017856568098068, + 0.016626350581645966, + 0.019483881071209908, + 0.03044828400015831, + 0.025315633043646812, + 0.019321877509355545, + 0.03316182643175125, + 0.02258305810391903, + 0.02990569733083248, + 0.016636308282613754, + 0.01778857223689556, + 0.014959280379116535, + 0.014193166978657246, + 0.01618446595966816, + 0.03532816469669342, + 0.01889076456427574, + 0.023181861266493797, + 0.025769997388124466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.023288074880838394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with high cheekbones and a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a simple coral cotton tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a shell necklace with small, off-white shells strung against her collarbone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan left while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vibrant, out-of-focus bokeh of a city sidewalk with blurred pedestrians and storefronts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046372017823159695 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.0046372017823159695, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046372017823159695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03883049264550209, + 0.020807046443223953, + 0.027326112613081932, + 0.04627382010221481, + 0.041566625237464905, + 0.03466833010315895, + 0.08738142997026443, + 0.12916985154151917, + 0.191194549202919, + 0.3827817440032959 + ], + "credit_map": [ + 0.00580689636990428, + 0.005563323386013508, + 0.004319265019148588, + 0.007990067824721336, + 0.0034803268499672413, + 0.003542287740856409, + 0.0046659777872264385, + 0.003462349995970726, + 0.0022094890009611845, + 0.0028106458485126495, + 0.0019367788918316364, + 0.002001935848966241, + 0.002190475584939122, + 0.004481134470552206, + 0.0023601779248565435, + 0.0028164086397737265, + 0.0048736995086073875, + 0.0033953702077269554, + 0.0033617201261222363, + 0.0038618824910372496, + 0.0027173778507858515, + 0.0026674724649637938, + 0.0032469595316797495, + 0.0032016292680054903, + 0.0050306362099945545, + 0.004868270363658667, + 0.0043903980404138565, + 0.009947494603693485, + 0.004594465252012014, + 0.0059760925360023975, + 0.006305110175162554, + 0.005161353852599859, + 0.003430766286328435, + 0.005688546225428581, + 0.004255096893757582, + 0.007349704392254353, + 0.005192715208977461, + 0.003874737536534667, + 0.007193158846348524, + 0.004581902641803026, + 0.002828203374519944, + 0.0024183294735848904, + 0.0037111162673681974, + 0.006465031765401363, + 0.0037580339703708887, + 0.005352074280381203, + 0.004777850583195686, + 0.005357689689844847, + 0.010630669072270393, + 0.010030757635831833, + 0.008150370791554451, + 0.00814933143556118, + 0.011460362933576107, + 0.019027408212423325, + 0.008039669133722782, + 0.011892861686646938, + 0.013869582675397396, + 0.015046440996229649, + 0.021507062017917633, + 0.019528977572917938, + 0.016724929213523865, + 0.015879515558481216, + 0.016284627839922905, + 0.010328717529773712, + 0.013242602348327637, + 0.021569188684225082, + 0.021648051217198372, + 0.019028866663575172, + 0.04123768210411072, + 0.028167126700282097, + 0.033474452793598175, + 0.01282657589763403, + 0.04823371022939682, + 0.06703995913267136, + 0.043810244649648666, + 0.04765799641609192, + 0.06149058789014816, + 0.03662412241101265, + 0.039806053042411804, + 0.03811909630894661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0046372017823159695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_5/index_tensor([435, 435])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_5/index_tensor([435, 435])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a4d89ee5dd42b33abbc5878110aa5f684e8888ac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_5/index_tensor([435, 435])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static close-up captures a confident teenage African-American woman of average build, her black hair slicked back, wet and clinging to her scalp, droplets of water tracking along her strong jawline. She sits in an indoor swimming pool, treading water so vigorously that her shoulders and head remain elevated and steady above the shimmering surface, a testament to her athleticism, allowing small, concentric ripples to spread around her. In her right hand, she expertly holds a bright yellow, textured water ball, her fingers gripping its dimpled surface. The entire scene is bathed in the intense, warm golden glow of the overhead natatorium lights, which shimmer on her skin and create brilliant highlights on the disturbed water. She makes a slight, deliberate motion to lift the ball, a subtle but controlled movement, her head tilted steadily, her focused eyes scanning the length of the pool with strategic intent. The interplay of the warm light, the deep blue water, and her determined expression creates a dynamic, high-contrast visual that encapsulates the atmosphere of peak athletic focus and poised power.", + "seed": 18080, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_5/index_435_seed_18080_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04447463410906494, + "vlm_logprob": -0.04447463410906494, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.04447463410906494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08743169903755188 + ], + "opposite_logprob_tokens": [ + -2.5874316692352295 + ], + "logprob_reward": -0.08743169903755188, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08743169903755188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1905307173728943, + 0.06385242938995361, + 0.04934822395443916, + 0.07273562252521515, + 0.06541425734758377, + 0.044133517891168594, + 0.09935562312602997, + 0.2034522444009781, + 0.0656818151473999, + 0.14549556374549866 + ], + "credit_map": [ + 0.03030599281191826, + 0.01720459759235382, + 0.03158705309033394, + 0.01772751286625862, + 0.027019865810871124, + 0.02050473913550377, + 0.01916077360510826, + 0.027020161971449852, + 0.006790624465793371, + 0.005401648581027985, + 0.0069371857680380344, + 0.006764395162463188, + 0.01313982903957367, + 0.006899063941091299, + 0.008988865651190281, + 0.008930808864533901, + 0.007655392400920391, + 0.0031911469995975494, + 0.006361579522490501, + 0.0035440672654658556, + 0.007982797920703888, + 0.00639924593269825, + 0.006151045206934214, + 0.008062942884862423, + 0.008055057376623154, + 0.007659032940864563, + 0.008491055108606815, + 0.008651147596538067, + 0.011563923209905624, + 0.010806458070874214, + 0.010013449005782604, + 0.007495488040149212, + 0.006859235465526581, + 0.0057518924586474895, + 0.006238546222448349, + 0.006602061912417412, + 0.015255059115588665, + 0.007402090821415186, + 0.013557583093643188, + 0.003747781505808234, + 0.007033003494143486, + 0.004373228643089533, + 0.0035890284925699234, + 0.004925508052110672, + 0.006642024964094162, + 0.005984433460980654, + 0.007267187349498272, + 0.004319098778069019, + 0.007168095093220472, + 0.006777611095458269, + 0.006399419158697128, + 0.01146375760436058, + 0.02690086141228676, + 0.010377553291618824, + 0.02530629001557827, + 0.004962030332535505, + 0.030238928273320198, + 0.007634187117218971, + 0.04932273179292679, + 0.020275544375181198, + 0.03572985157370567, + 0.01660463586449623, + 0.029793281108140945, + 0.013853071257472038, + 0.006650074385106564, + 0.006389637477695942, + 0.00977393426001072, + 0.009209505282342434, + 0.009183446876704693, + 0.006570682395249605, + 0.013321475125849247, + 0.0045830546878278255, + 0.01544554065912962, + 0.012890772894024849, + 0.014242666773498058, + 0.020965470001101494, + 0.021828249096870422, + 0.015594924800097942, + 0.021962130442261696, + 0.022565782070159912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.08743169903755188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206710636615753 + ], + "opposite_logprob_tokens": [ + -2.8120670318603516 + ], + "logprob_reward": -0.06206710636615753, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206710636615753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13331539928913116, + 0.07809236645698547, + 0.0616571307182312, + 0.07267549633979797, + 0.03257128968834877, + 0.05331791564822197, + 0.08612208813428879, + 0.3097635507583618, + 0.05935526639223099, + 0.11312951892614365 + ], + "credit_map": [ + 0.013194833882153034, + 0.012816719710826874, + 0.015520271845161915, + 0.018322331830859184, + 0.022032642737030983, + 0.014783852733671665, + 0.020207826048135757, + 0.016436908394098282, + 0.008176692761480808, + 0.005540892481803894, + 0.011299393139779568, + 0.007942956872284412, + 0.010146685875952244, + 0.008226701989769936, + 0.013518973253667355, + 0.013240069150924683, + 0.006535940803587437, + 0.004721760284155607, + 0.005035719368606806, + 0.006354072596877813, + 0.009285736829042435, + 0.010537009686231613, + 0.010171724483370781, + 0.009015165269374847, + 0.006648743059486151, + 0.006461454555392265, + 0.009719017893075943, + 0.012701950967311859, + 0.009171323850750923, + 0.01110148523002863, + 0.008453008718788624, + 0.008418506942689419, + 0.003501832950860262, + 0.0025277792010456324, + 0.002742479555308819, + 0.005095067899674177, + 0.005531429313123226, + 0.004055144265294075, + 0.006915406323969364, + 0.00220215180888772, + 0.0058138552121818066, + 0.0037035176064819098, + 0.006765071302652359, + 0.0067407540045678616, + 0.006646177731454372, + 0.006665973924100399, + 0.009772436693310738, + 0.007210128009319305, + 0.008886156603693962, + 0.005944411735981703, + 0.012130692601203918, + 0.011184859089553356, + 0.01302268821746111, + 0.011204955168068409, + 0.015760425478219986, + 0.007987907156348228, + 0.037749532610177994, + 0.011484960094094276, + 0.07345619797706604, + 0.04140334948897362, + 0.04757673293352127, + 0.03080051764845848, + 0.04322478547692299, + 0.024067459627985954, + 0.006509417202323675, + 0.005131830926984549, + 0.005573653150349855, + 0.006375640630722046, + 0.012980747036635876, + 0.005433966871351004, + 0.012110095471143723, + 0.005239913240075111, + 0.012634549289941788, + 0.009674652479588985, + 0.013424355536699295, + 0.01194088626652956, + 0.017768584191799164, + 0.010318819433450699, + 0.01895115152001381, + 0.018416516482830048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.06206710636615753 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641016039997339 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007641016039997339, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641016039997339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1372535526752472, + 0.36166471242904663, + 0.04247758165001869, + 0.04725191742181778, + 0.052779097110033035, + 0.042937517166137695, + 0.040481239557266235, + 0.11599922925233841, + 0.07574734836816788, + 0.08340781182050705 + ], + "credit_map": [ + 0.01523174624890089, + 0.010682953521609306, + 0.02566768415272236, + 0.013776634819805622, + 0.013766077347099781, + 0.015112903900444508, + 0.013838457874953747, + 0.02917708456516266, + 0.034162718802690506, + 0.03091951087117195, + 0.04356386512517929, + 0.07103372365236282, + 0.04072164371609688, + 0.03429502621293068, + 0.04500720277428627, + 0.06196099892258644, + 0.006019809748977423, + 0.003640613052994013, + 0.0056489272974431515, + 0.006191179621964693, + 0.005389800760895014, + 0.005365158896893263, + 0.0047498466446995735, + 0.005472245626151562, + 0.004178074654191732, + 0.005218858364969492, + 0.005282366182655096, + 0.008837533183395863, + 0.007393863517791033, + 0.00589248351752758, + 0.006436835043132305, + 0.004011901561170816, + 0.004500129725784063, + 0.005048684775829315, + 0.003620404051616788, + 0.004075518809258938, + 0.013913076370954514, + 0.00573206739500165, + 0.012654407881200314, + 0.0032348085660487413, + 0.0032088442239910364, + 0.0032175106462091208, + 0.0037879536394029856, + 0.0049784863367676735, + 0.00873524323105812, + 0.004604237154126167, + 0.009994307532906532, + 0.004410932771861553, + 0.0042762598022818565, + 0.0025524809025228024, + 0.004051460884511471, + 0.006026019342243671, + 0.0055325510911643505, + 0.005623871926218271, + 0.007896412163972855, + 0.0045221829786896706, + 0.010278522036969662, + 0.005064154975116253, + 0.021988917142152786, + 0.0170611422508955, + 0.01802193932235241, + 0.015805883333086967, + 0.01917123608291149, + 0.008607429452240467, + 0.00556572899222374, + 0.0069189672358334064, + 0.007353318855166435, + 0.007397206034511328, + 0.014674225822091103, + 0.004169892985373735, + 0.02436133660376072, + 0.005306669510900974, + 0.008046113885939121, + 0.006312097422778606, + 0.012868219055235386, + 0.009494374506175518, + 0.013248054310679436, + 0.007300346624106169, + 0.012134694494307041, + 0.014003911055624485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.007641016039997339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031945884693413973 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031945884693413973, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031945884693413973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10642047226428986, + 0.07788663357496262, + 0.11144351959228516, + 0.2784571051597595, + 0.04997226223349571, + 0.09235825389623642, + 0.022410264238715172, + 0.17304639518260956, + 0.02834312990307808, + 0.05966196209192276 + ], + "credit_map": [ + 0.019309746101498604, + 0.011283784173429012, + 0.014376966282725334, + 0.012177626602351665, + 0.014950854703783989, + 0.010092414915561676, + 0.010898459702730179, + 0.013330611400306225, + 0.01012633927166462, + 0.006898646242916584, + 0.012341137044131756, + 0.00993426889181137, + 0.00852530263364315, + 0.008172539994120598, + 0.010289336554706097, + 0.011599059216678143, + 0.013510224409401417, + 0.0063746003434062, + 0.013558642938733101, + 0.013290400616824627, + 0.01122883427888155, + 0.01892636902630329, + 0.016222208738327026, + 0.018332239240407944, + 0.042149607092142105, + 0.028980674222111702, + 0.029267894104123116, + 0.0440157875418663, + 0.030210714787244797, + 0.04725341126322746, + 0.03146006166934967, + 0.02511894330382347, + 0.004864602815359831, + 0.004655680153518915, + 0.008091205731034279, + 0.006614817306399345, + 0.006806608289480209, + 0.005982196424156427, + 0.008065987378358841, + 0.004891165532171726, + 0.01597495563328266, + 0.008486420847475529, + 0.010230022482573986, + 0.009433615021407604, + 0.012217343784868717, + 0.013132210820913315, + 0.012819497846066952, + 0.010064189322292805, + 0.00360860675573349, + 0.002635740675032139, + 0.0019427320221439004, + 0.0027148800436407328, + 0.003641925984993577, + 0.0025377448182553053, + 0.003821972291916609, + 0.0015066618798300624, + 0.019427767023444176, + 0.006717807147651911, + 0.039621952921152115, + 0.022349577397108078, + 0.026774385944008827, + 0.01940395124256611, + 0.02520020306110382, + 0.013550755567848682, + 0.002870089141651988, + 0.0023596561513841152, + 0.0034994676243513823, + 0.0047866408713161945, + 0.004379963967949152, + 0.0031151887960731983, + 0.005089837126433849, + 0.0022422848269343376, + 0.006586253177374601, + 0.004684740211814642, + 0.00615933071821928, + 0.010420646518468857, + 0.006129469256848097, + 0.0057331956923007965, + 0.007405598647892475, + 0.012542733922600746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.0031945884693413973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062038760632276535 + ], + "opposite_logprob_tokens": [ + -2.8120386600494385 + ], + "logprob_reward": -0.062038760632276535, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062038760632276535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09534827619791031, + 0.04300785809755325, + 0.06763417273759842, + 0.050929322838783264, + 0.05208471044898033, + 0.06120951473712921, + 0.0865260437130928, + 0.13602976500988007, + 0.333482563495636, + 0.07374780625104904 + ], + "credit_map": [ + 0.009199514053761959, + 0.010184122249484062, + 0.01241397950798273, + 0.010892199352383614, + 0.017302261665463448, + 0.008897269144654274, + 0.01013234443962574, + 0.016326582059264183, + 0.005205866880714893, + 0.004385655280202627, + 0.005781233310699463, + 0.006039817817509174, + 0.0076940301805734634, + 0.004774376284331083, + 0.004838970024138689, + 0.004287907388061285, + 0.004834257066249847, + 0.0051838248036801815, + 0.009586072526872158, + 0.00920899212360382, + 0.010714911855757236, + 0.006228251848369837, + 0.012543756514787674, + 0.009334107860922813, + 0.00414530374109745, + 0.004806363023817539, + 0.006588498130440712, + 0.006501303054392338, + 0.009373943321406841, + 0.006007141899317503, + 0.007928294129669666, + 0.0055784741416573524, + 0.005111320409923792, + 0.003937628120183945, + 0.006124350242316723, + 0.006513641215860844, + 0.009103680029511452, + 0.0055138529278337955, + 0.011266508139669895, + 0.004513729829341173, + 0.007657986134290695, + 0.0035063212271779776, + 0.0057832966558635235, + 0.006776468362659216, + 0.012491290457546711, + 0.00518661318346858, + 0.016332687810063362, + 0.003474852070212364, + 0.012068265117704868, + 0.005565265193581581, + 0.010606818832457066, + 0.011511805467307568, + 0.016089195385575294, + 0.010591855272650719, + 0.012609373778104782, + 0.007483466994017363, + 0.013539575040340424, + 0.00470897788181901, + 0.033380262553691864, + 0.0169623214751482, + 0.019132817164063454, + 0.01976175233721733, + 0.020601622760295868, + 0.007942434400320053, + 0.03138657659292221, + 0.03041459619998932, + 0.03256581723690033, + 0.028316836804151535, + 0.05975043773651123, + 0.02181154675781727, + 0.10205870866775513, + 0.02717803791165352, + 0.009586844593286514, + 0.006103111896663904, + 0.009233619086444378, + 0.008013254031538963, + 0.01107910554856062, + 0.00864246767014265, + 0.008896827697753906, + 0.012192572467029095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.062038760632276535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up captures a confident teenage African-American woman of average build, her black hair slicked back, wet and clinging to her scalp, droplets of water tracking along her strong jawline. She sits in an indoor swimming pool, treading water so vigorously that her shoulders and head remain elevated and steady above the shimmering surface, a testament to her athleticism, allowing small, concentric ripples to spread around her. In her right hand, she expertly holds a bright yellow, textured water ball, her fingers gripping its dimpled surface. The entire scene is bathed in the intense, warm golden glow of the overhead natatorium lights, which shimmer on her skin and create brilliant highlights on the disturbed water. She makes a slight, deliberate motion to lift the ball, a subtle but controlled movement, her head tilted steadily, her focused eyes scanning the length of the pool with strategic intent. The interplay of the warm light, the deep blue water, and her determined expression creates a dynamic, high-contrast visual that encapsulates the atmosphere of peak athletic focus and poised power.", + "seed": 18081, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_5/index_435_seed_18081_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.038888022676110266, + "vlm_logprob": -0.038888022676110266, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.038888022676110266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02585156261920929 + ], + "opposite_logprob_tokens": [ + -3.9008514881134033 + ], + "logprob_reward": -0.02585156261920929, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02585156261920929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17934679985046387, + 0.2823221683502197, + 0.09138146042823792, + 0.06391265988349915, + 0.09690424054861069, + 0.11030308902263641, + 0.0685204565525055, + 0.06509210914373398, + 0.01597249135375023, + 0.02624453604221344 + ], + "credit_map": [ + 0.016469541937112808, + 0.017282430082559586, + 0.014205388724803925, + 0.02516176924109459, + 0.03181597217917442, + 0.020177086815238, + 0.03767566755414009, + 0.01655893214046955, + 0.022771501913666725, + 0.10302023589611053, + 0.029295237734913826, + 0.012498493306338787, + 0.03959914669394493, + 0.034068793058395386, + 0.025788210332393646, + 0.015280550345778465, + 0.008059833198785782, + 0.012012902647256851, + 0.0076534622348845005, + 0.009452997706830502, + 0.0164264477789402, + 0.012673363089561462, + 0.014369387179613113, + 0.010733064264059067, + 0.007046089973300695, + 0.0071168988943099976, + 0.007255870383232832, + 0.005219549871981144, + 0.011749735102057457, + 0.008118848316371441, + 0.009282927960157394, + 0.008122737519443035, + 0.008916039951145649, + 0.01187044009566307, + 0.01400639396160841, + 0.016774717718362808, + 0.009596149437129498, + 0.010968083515763283, + 0.013267390429973602, + 0.011505020782351494, + 0.009472469799220562, + 0.026548856869339943, + 0.018799282610416412, + 0.012367099523544312, + 0.008380426093935966, + 0.017533227801322937, + 0.008166437037289143, + 0.009035291150212288, + 0.007543503772467375, + 0.007928438484668732, + 0.01707504689693451, + 0.012563174590468407, + 0.004706934094429016, + 0.005700814072042704, + 0.007151855155825615, + 0.005850684829056263, + 0.0087426183745265, + 0.00980934128165245, + 0.009184902533888817, + 0.007604195270687342, + 0.007929619401693344, + 0.0074432105757296085, + 0.008118540048599243, + 0.006259686313569546, + 0.001679450273513794, + 0.0018235736060887575, + 0.0034043879713863134, + 0.0023992538917809725, + 0.0014343255897983909, + 0.0015508591895923018, + 0.0019554526079446077, + 0.0017251879908144474, + 0.007471044082194567, + 0.0020415890030562878, + 0.001977583160623908, + 0.003871977562084794, + 0.002266157418489456, + 0.002016597194597125, + 0.004525984171777964, + 0.002073603216558695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.02585156261920929 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900739461183548 + ], + "opposite_logprob_tokens": [ + -2.579007387161255 + ], + "logprob_reward": -0.07900739461183548, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900739461183548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1733729988336563, + 0.14840272068977356, + 0.10075294971466064, + 0.09482129663228989, + 0.20696642994880676, + 0.028218653053045273, + 0.06541436165571213, + 0.08289128541946411, + 0.045640960335731506, + 0.053518325090408325 + ], + "credit_map": [ + 0.01618768647313118, + 0.024354932829737663, + 0.014397522434592247, + 0.021973736584186554, + 0.03312556445598602, + 0.01905612275004387, + 0.025123128667473793, + 0.019154299050569534, + 0.010346097871661186, + 0.05819370225071907, + 0.017300814390182495, + 0.0073307109996676445, + 0.013979745097458363, + 0.026161015033721924, + 0.008418183773756027, + 0.006672448478639126, + 0.013942647725343704, + 0.008992090821266174, + 0.008313877508044243, + 0.01235500443726778, + 0.01094350777566433, + 0.015975965186953545, + 0.011816276237368584, + 0.018413584679365158, + 0.007914613001048565, + 0.00815371610224247, + 0.008016837760806084, + 0.006304221693426371, + 0.019617605954408646, + 0.013978248462080956, + 0.01429644413292408, + 0.016539618372917175, + 0.02028888650238514, + 0.0206622201949358, + 0.04253964126110077, + 0.029833316802978516, + 0.019694678485393524, + 0.022059908136725426, + 0.02523617073893547, + 0.026651613414287567, + 0.002364126034080982, + 0.005314207170158625, + 0.0033254441805183887, + 0.0029196003451943398, + 0.0034977360628545284, + 0.0034150523133575916, + 0.004394258372485638, + 0.002988229040056467, + 0.006164629478007555, + 0.006141130346804857, + 0.012403745204210281, + 0.0071821860037744045, + 0.007055320776998997, + 0.0064385440200567245, + 0.009575961157679558, + 0.010452844202518463, + 0.0068655190989375114, + 0.008732420392334461, + 0.005502438172698021, + 0.010547640733420849, + 0.008693256415426731, + 0.008603894151747227, + 0.02185196615755558, + 0.012094146572053432, + 0.005023798439651728, + 0.0053910319693386555, + 0.009649444371461868, + 0.005982405040413141, + 0.006168271414935589, + 0.004030992742627859, + 0.005204436369240284, + 0.004190576262772083, + 0.008693734183907509, + 0.004305319860577583, + 0.005200599320232868, + 0.007448206655681133, + 0.005267265718430281, + 0.004058436490595341, + 0.013787658885121346, + 0.0047570993192493916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.07900739461183548 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791677817702293 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009791677817702293, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791677817702293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14813008904457092, + 0.1087755486369133, + 0.12299375236034393, + 0.11049216985702515, + 0.2560284435749054, + 0.048621874302625656, + 0.04528288543224335, + 0.05860842019319534, + 0.069425567984581, + 0.03164124861359596 + ], + "credit_map": [ + 0.013109700754284859, + 0.014671526849269867, + 0.01629156805574894, + 0.021858615800738335, + 0.028671808540821075, + 0.012454413808882236, + 0.021569272503256798, + 0.01950317621231079, + 0.011273529380559921, + 0.02448873408138752, + 0.012438558042049408, + 0.007484506815671921, + 0.016175350174307823, + 0.01567641645669937, + 0.011302335187792778, + 0.009936132468283176, + 0.0126657048240304, + 0.01533114816993475, + 0.007887637242674828, + 0.011532791890203953, + 0.021640634164214134, + 0.02218518778681755, + 0.012708565220236778, + 0.01904209330677986, + 0.013548538088798523, + 0.012752112001180649, + 0.018817132338881493, + 0.010074903257191181, + 0.015069611370563507, + 0.014739287085831165, + 0.011970105580985546, + 0.013520483858883381, + 0.0286470428109169, + 0.025588728487491608, + 0.08177608996629715, + 0.03436416760087013, + 0.021781116724014282, + 0.02205200120806694, + 0.021512480452656746, + 0.020306838676333427, + 0.004797640722244978, + 0.011934359557926655, + 0.004936543758958578, + 0.00473976181820035, + 0.0035007272381335497, + 0.008359435945749283, + 0.005718882195651531, + 0.004634527489542961, + 0.004146936349570751, + 0.006156097631901503, + 0.00970851443707943, + 0.006332840770483017, + 0.0046078599989414215, + 0.0037450427189469337, + 0.006480355281382799, + 0.004105241969227791, + 0.008061526343226433, + 0.009470337070524693, + 0.013372655026614666, + 0.008134860545396805, + 0.005091968923807144, + 0.00471827108412981, + 0.004781947936862707, + 0.00497685931622982, + 0.010645539499819279, + 0.010699416510760784, + 0.012276771478354931, + 0.009742445312440395, + 0.005291702691465616, + 0.004694152623414993, + 0.0108597781509161, + 0.005215764511376619, + 0.01113254576921463, + 0.004121878650039434, + 0.003167497692629695, + 0.004444792866706848, + 0.002054680837318301, + 0.0024579674936830997, + 0.0025028486270457506, + 0.0017590393545106053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.009791677817702293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978199765086174 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.00978199765086174, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978199765086174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35616615414619446, + 0.07011030614376068, + 0.10498373955488205, + 0.0900762602686882, + 0.11199437826871872, + 0.05690864473581314, + 0.11358623206615448, + 0.03672036901116371, + 0.029659533873200417, + 0.029794353991746902 + ], + "credit_map": [ + 0.03702656552195549, + 0.04292263835668564, + 0.03325924277305603, + 0.03861977159976959, + 0.060865044593811035, + 0.04865437373518944, + 0.047038450837135315, + 0.047780077904462814, + 0.00611483259126544, + 0.014109943993389606, + 0.005418889690190554, + 0.004809259437024593, + 0.01319740992039442, + 0.008539238013327122, + 0.01217931229621172, + 0.005741414148360491, + 0.016581472009420395, + 0.013476417399942875, + 0.01215919479727745, + 0.013313427567481995, + 0.014779130928218365, + 0.012254372239112854, + 0.01084405928850174, + 0.011575656943023205, + 0.01570441946387291, + 0.012103213928639889, + 0.015252722427248955, + 0.01129163708537817, + 0.010446926578879356, + 0.00817814003676176, + 0.008972893469035625, + 0.008126307278871536, + 0.013134075328707695, + 0.011870762333273888, + 0.03022126480937004, + 0.017666855826973915, + 0.009240265004336834, + 0.008156361058354378, + 0.012255656532943249, + 0.009449129924178123, + 0.00845969095826149, + 0.007971026003360748, + 0.01325030904263258, + 0.006966634653508663, + 0.003382570343092084, + 0.00694203469902277, + 0.005733231082558632, + 0.004203152377158403, + 0.01112930104136467, + 0.024398326873779297, + 0.02230915240943432, + 0.015072742477059364, + 0.006737532559782267, + 0.009541728533804417, + 0.00966715719550848, + 0.014730297029018402, + 0.0040347641333937645, + 0.005500027909874916, + 0.007533332798629999, + 0.004273498430848122, + 0.0030015013180673122, + 0.0029958535451442003, + 0.006463904399424791, + 0.002917483914643526, + 0.0036782615352422, + 0.0044963667169213295, + 0.004536529071629047, + 0.0034453964326530695, + 0.00329088163562119, + 0.002586695598438382, + 0.00503443693742156, + 0.0025909689720720053, + 0.0038719321601092815, + 0.003601427422836423, + 0.005214993376284838, + 0.005984024610370398, + 0.0030510840006172657, + 0.0018748410511761904, + 0.003615729045122862, + 0.0025803251191973686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.00978199765086174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000748068094254 + ], + "opposite_logprob_tokens": [ + -2.695007562637329 + ], + "logprob_reward": -0.07000748068094254, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000748068094254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26980218291282654, + 0.27871450781822205, + 0.04202893003821373, + 0.07793982326984406, + 0.06762132048606873, + 0.03878647834062576, + 0.06811632215976715, + 0.05207061395049095, + 0.03918612375855446, + 0.06573373079299927 + ], + "credit_map": [ + 0.03339149430394173, + 0.02250831387937069, + 0.023874228820204735, + 0.02838151901960373, + 0.054901476949453354, + 0.04657137021422386, + 0.032703153789043427, + 0.02747059240937233, + 0.02213572897017002, + 0.11477455496788025, + 0.04100509360432625, + 0.013094575144350529, + 0.022690633311867714, + 0.038386762142181396, + 0.013031798414885998, + 0.01359532494097948, + 0.004158785566687584, + 0.0053680771961808205, + 0.003366281511262059, + 0.005561233032494783, + 0.007567735388875008, + 0.005998786073178053, + 0.005386807955801487, + 0.004621224012225866, + 0.008836510591208935, + 0.007624164689332247, + 0.012479661963880062, + 0.007897265255451202, + 0.01200303714722395, + 0.008079241029918194, + 0.01143767312169075, + 0.00958226714283228, + 0.009565652348101139, + 0.0065081422217190266, + 0.010007880628108978, + 0.00934569165110588, + 0.008237016387283802, + 0.006927968934178352, + 0.008176743984222412, + 0.008852217346429825, + 0.0047997403889894485, + 0.003667552024126053, + 0.009001289494335651, + 0.005671259947121143, + 0.003295500762760639, + 0.0033844676800072193, + 0.004354298580437899, + 0.004612367134541273, + 0.006932821590453386, + 0.00520391296595335, + 0.0076676602475345135, + 0.006708630360662937, + 0.010984184220433235, + 0.008983840234577656, + 0.012028982862830162, + 0.00960628129541874, + 0.007978410460054874, + 0.007607588078826666, + 0.00850166380405426, + 0.006954115349799395, + 0.003914268687367439, + 0.004272444639354944, + 0.008463675156235695, + 0.004378444980829954, + 0.0066909389570355415, + 0.004650673363357782, + 0.007472349796444178, + 0.004778318107128143, + 0.0035977880470454693, + 0.0030435442458838224, + 0.005486959591507912, + 0.0034655469935387373, + 0.02113170735538006, + 0.014315924607217312, + 0.0054858229123055935, + 0.0064299944788217545, + 0.004413430579006672, + 0.003906998783349991, + 0.006313737481832504, + 0.0037361157592386007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.07000748068094254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up captures a confident teenage African-American woman of average build, her black hair slicked back, wet and clinging to her scalp, droplets of water tracking along her strong jawline. She sits in an indoor swimming pool, treading water so vigorously that her shoulders and head remain elevated and steady above the shimmering surface, a testament to her athleticism, allowing small, concentric ripples to spread around her. In her right hand, she expertly holds a bright yellow, textured water ball, her fingers gripping its dimpled surface. The entire scene is bathed in the intense, warm golden glow of the overhead natatorium lights, which shimmer on her skin and create brilliant highlights on the disturbed water. She makes a slight, deliberate motion to lift the ball, a subtle but controlled movement, her head tilted steadily, her focused eyes scanning the length of the pool with strategic intent. The interplay of the warm light, the deep blue water, and her determined expression creates a dynamic, high-contrast visual that encapsulates the atmosphere of peak athletic focus and poised power.", + "seed": 18082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_5/index_435_seed_18082_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05584029546007514, + "vlm_logprob": -0.05584029546007514, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05584029546007514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09851255267858505 + ], + "opposite_logprob_tokens": [ + -2.473512649536133 + ], + "logprob_reward": -0.09851255267858505, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09851255267858505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13656198978424072, + 0.04595688730478287, + 0.04261840507388115, + 0.07102344930171967, + 0.05400055646896362, + 0.09071473032236099, + 0.07195144146680832, + 0.35547372698783875, + 0.11379688233137131, + 0.017901895567774773 + ], + "credit_map": [ + 0.0116199292242527, + 0.01375850010663271, + 0.012661737389862537, + 0.012902120128273964, + 0.030735865235328674, + 0.019206441938877106, + 0.020774194970726967, + 0.014903194271028042, + 0.006705470383167267, + 0.011388053186237812, + 0.003906663507223129, + 0.0023329684045165777, + 0.007171350996941328, + 0.005131295882165432, + 0.0053843301720917225, + 0.003936756402254105, + 0.004824225325137377, + 0.004611221607774496, + 0.0060431999154388905, + 0.003633334767073393, + 0.005247224122285843, + 0.00770751154050231, + 0.005147088784724474, + 0.00540459668263793, + 0.008530545979738235, + 0.008898208849132061, + 0.00639837421476841, + 0.005994011182337999, + 0.012804551981389523, + 0.008696814067661762, + 0.009370180778205395, + 0.010330768302083015, + 0.006787923630326986, + 0.008621684275567532, + 0.004847664386034012, + 0.004906822927296162, + 0.007392173167318106, + 0.00910857506096363, + 0.0065300967544317245, + 0.00580561812967062, + 0.03017766959965229, + 0.005989764351397753, + 0.004574082791805267, + 0.003495615441352129, + 0.0057076746597886086, + 0.03156796470284462, + 0.004997778218239546, + 0.004204185679554939, + 0.01722128503024578, + 0.008406168781220913, + 0.003025335492566228, + 0.0033462238498032093, + 0.004387899301946163, + 0.02675885520875454, + 0.003990168683230877, + 0.004815502557903528, + 0.1253277063369751, + 0.027812128886580467, + 0.009297734126448631, + 0.0052414764650166035, + 0.023633494973182678, + 0.13568726181983948, + 0.015176218934357166, + 0.013297703117132187, + 0.03298058733344078, + 0.00946265272796154, + 0.00777012761682272, + 0.008387716487050056, + 0.012885103933513165, + 0.011062665842473507, + 0.011658597737550735, + 0.01958942785859108, + 0.0035961747635155916, + 0.001592346583493054, + 0.0019160552183166146, + 0.0013723671436309814, + 0.0017211023950949311, + 0.001622332725673914, + 0.001969977281987667, + 0.004111538175493479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.09851255267858505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208436191082001 + ], + "opposite_logprob_tokens": [ + -2.812084436416626 + ], + "logprob_reward": -0.06208436191082001, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208436191082001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08207447081804276, + 0.058480117470026016, + 0.11764951795339584, + 0.0771855041384697, + 0.02667083404958248, + 0.044302381575107574, + 0.08296047151088715, + 0.34359878301620483, + 0.09286180138587952, + 0.07421614229679108 + ], + "credit_map": [ + 0.006831361446529627, + 0.008439881727099419, + 0.009143847040832043, + 0.006426261737942696, + 0.01696927472949028, + 0.011052146553993225, + 0.01408722996711731, + 0.00912446528673172, + 0.0049589890986680984, + 0.010531634092330933, + 0.004634777549654245, + 0.003686173353344202, + 0.00662885932251811, + 0.016012299805879593, + 0.007256753277033567, + 0.004770629107952118, + 0.014997454360127449, + 0.010210876353085041, + 0.014423596672713757, + 0.010953840799629688, + 0.014523940160870552, + 0.02207469753921032, + 0.014532610774040222, + 0.01593250222504139, + 0.010975480079650879, + 0.015395883470773697, + 0.006632713600993156, + 0.006601348053663969, + 0.010370073840022087, + 0.008424839936196804, + 0.009982354938983917, + 0.008802808821201324, + 0.004301123321056366, + 0.003270065179094672, + 0.004201406612992287, + 0.002353413263335824, + 0.0027295108884572983, + 0.004337838850915432, + 0.0028872548136860132, + 0.0025902206543833017, + 0.01306463684886694, + 0.003367986297234893, + 0.0025724321603775024, + 0.0022693658247590065, + 0.004240554757416248, + 0.012974346056580544, + 0.003799822647124529, + 0.002013237215578556, + 0.014507081359624863, + 0.01033787988126278, + 0.005463839042931795, + 0.005305327475070953, + 0.007123867981135845, + 0.025746725499629974, + 0.007779499050229788, + 0.006696243770420551, + 0.11636991798877716, + 0.028702113777399063, + 0.009293956682085991, + 0.005462408997118473, + 0.02244238741695881, + 0.13213923573493958, + 0.01484137773513794, + 0.014347375370562077, + 0.02804489992558956, + 0.0063187479972839355, + 0.008110255934298038, + 0.005901565309613943, + 0.009421642869710922, + 0.01627924293279648, + 0.00937977246940136, + 0.009405669756233692, + 0.01972484402358532, + 0.006655153352767229, + 0.005412451457232237, + 0.004536140710115433, + 0.005523405037820339, + 0.006224500481039286, + 0.007322521880269051, + 0.018817124888300896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.06208436191082001 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209873043000698 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209873043000698, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209873043000698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22219619154930115, + 0.06780356168746948, + 0.05768432095646858, + 0.06470208615064621, + 0.021854771301150322, + 0.05637325718998909, + 0.02606256864964962, + 0.4058457911014557, + 0.0421014204621315, + 0.03537600859999657 + ], + "credit_map": [ + 0.018797999247908592, + 0.032689426094293594, + 0.022093981504440308, + 0.0235512126237154, + 0.0421995148062706, + 0.030298875644803047, + 0.029423099011182785, + 0.02314208634197712, + 0.009275753982365131, + 0.023864228278398514, + 0.0036540438886731863, + 0.0027322918176651, + 0.00543230539187789, + 0.013873416930437088, + 0.004531495273113251, + 0.004440030083060265, + 0.007834207266569138, + 0.004289875738322735, + 0.007843849249184132, + 0.005339552648365498, + 0.005960698239505291, + 0.012222697958350182, + 0.006579893175512552, + 0.007613546214997768, + 0.008047698065638542, + 0.01338758785277605, + 0.006361024454236031, + 0.0050954632461071014, + 0.008120964281260967, + 0.007214441895484924, + 0.00830098707228899, + 0.008173922076821327, + 0.0027984273619949818, + 0.0034430299419909716, + 0.003550335532054305, + 0.0018616010202094913, + 0.0022724035661667585, + 0.0028722130227833986, + 0.002736240392550826, + 0.0023205226752907038, + 0.019116614013910294, + 0.003266823710873723, + 0.0030961474403738976, + 0.0019720522686839104, + 0.0036697285249829292, + 0.019024305045604706, + 0.0034641448874026537, + 0.002763444557785988, + 0.003613351611420512, + 0.003491569310426712, + 0.0036013047210872173, + 0.002170368330553174, + 0.00343607971444726, + 0.003507541259750724, + 0.0034312079660594463, + 0.002811148762702942, + 0.14128583669662476, + 0.03218358755111694, + 0.010819850489497185, + 0.00823229644447565, + 0.02685546875, + 0.153899148106575, + 0.017323046922683716, + 0.01524659525603056, + 0.00798468291759491, + 0.004309683572500944, + 0.006287259515374899, + 0.005307520739734173, + 0.004293839447200298, + 0.006584991700947285, + 0.0044499109499156475, + 0.0028835339471697807, + 0.01024566125124693, + 0.0029878539498895407, + 0.002143226796761155, + 0.002389349276199937, + 0.0025601827073842287, + 0.002936271484941244, + 0.003833336755633354, + 0.008280128240585327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.014209873043000698 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098706413060427 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004098706413060427, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098706413060427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18340663611888885, + 0.1061711385846138, + 0.3328503668308258, + 0.11337264627218246, + 0.021094096824526787, + 0.019888928160071373, + 0.10069696605205536, + 0.07732182741165161, + 0.026392322033643723, + 0.01880507543683052 + ], + "credit_map": [ + 0.015043945983052254, + 0.018505321815609932, + 0.01671481877565384, + 0.021770717576146126, + 0.03159330412745476, + 0.03344164416193962, + 0.028275301679968834, + 0.018061593174934387, + 0.008961119689047337, + 0.02267734706401825, + 0.008829875849187374, + 0.005627832375466824, + 0.010831749998033047, + 0.028592310845851898, + 0.012234742753207684, + 0.008416158147156239, + 0.04059800133109093, + 0.03511234372854233, + 0.04951797425746918, + 0.040768224745988846, + 0.029901426285505295, + 0.05803529545664787, + 0.03924274072051048, + 0.039674341678619385, + 0.015927257016301155, + 0.012958801351487637, + 0.013507919386029243, + 0.013401268050074577, + 0.015274416655302048, + 0.015673033893108368, + 0.013757232576608658, + 0.012872717343270779, + 0.002260420238599181, + 0.0026226963382214308, + 0.0022414831910282373, + 0.0028277586679905653, + 0.002494805259630084, + 0.0025271084159612656, + 0.0029120591934770346, + 0.0032077650539577007, + 0.005353386979550123, + 0.0019108726410195231, + 0.0015031684888526797, + 0.0009537687292322516, + 0.0016192587791010737, + 0.005771198309957981, + 0.0014029481681063771, + 0.0013743271119892597, + 0.023300036787986755, + 0.010263357311487198, + 0.008904931135475636, + 0.00388705893419683, + 0.006500865798443556, + 0.03472784161567688, + 0.006490055005997419, + 0.0066228266805410385, + 0.02537950687110424, + 0.005526474677026272, + 0.004736305680125952, + 0.002447087550535798, + 0.005565289873629808, + 0.026030752807855606, + 0.0038494339678436518, + 0.003786979243159294, + 0.0053545031696558, + 0.003344570519402623, + 0.0023720255121588707, + 0.0029019739013165236, + 0.002799017820507288, + 0.005107192788273096, + 0.0030055579263716936, + 0.001507478766143322, + 0.004018053412437439, + 0.0025676924269646406, + 0.001836678246036172, + 0.0014968038303777575, + 0.0019772006198763847, + 0.0014891797909513116, + 0.0020584010053426027, + 0.0033610665705055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.004098706413060427 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029598325490952 + ], + "opposite_logprob_tokens": [ + -2.3502960205078125 + ], + "logprob_reward": -0.10029598325490952, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029598325490952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1670926958322525, + 0.16500629484653473, + 0.12330549955368042, + 0.049372900277376175, + 0.06302741914987564, + 0.03004869446158409, + 0.07486055046319962, + 0.1971472203731537, + 0.09231064468622208, + 0.037828050553798676 + ], + "credit_map": [ + 0.014002836309373379, + 0.022324614226818085, + 0.014308298006653786, + 0.014291911385953426, + 0.03161889687180519, + 0.026846153661608696, + 0.024975920096039772, + 0.018724068999290466, + 0.012459865771234035, + 0.04272007942199707, + 0.011756345629692078, + 0.010018152184784412, + 0.01894790679216385, + 0.042527198791503906, + 0.014760160818696022, + 0.011816591955721378, + 0.01820986159145832, + 0.008806978352367878, + 0.013465872034430504, + 0.01005635131150484, + 0.012814010493457317, + 0.028864258900284767, + 0.013561871834099293, + 0.017526304349303246, + 0.008241496048867702, + 0.008013349026441574, + 0.003128400770947337, + 0.003868355881422758, + 0.005768980365246534, + 0.008492219261825085, + 0.005545713473111391, + 0.006314385216683149, + 0.008412542752921581, + 0.011710512451827526, + 0.008045337162911892, + 0.0054633645340800285, + 0.007827522233128548, + 0.009569419547915459, + 0.006996958050876856, + 0.0050017642788589, + 0.008845271542668343, + 0.0021881558932363987, + 0.001603874727152288, + 0.001536361058242619, + 0.002210089936852455, + 0.009650005958974361, + 0.0022484574001282454, + 0.0017664802726358175, + 0.019243624061346054, + 0.007486162707209587, + 0.004750219639390707, + 0.004034081939607859, + 0.004989911802113056, + 0.026104364544153214, + 0.003768709721043706, + 0.004483481869101524, + 0.05323731526732445, + 0.019502773880958557, + 0.015301891602575779, + 0.008902259171009064, + 0.012178703211247921, + 0.06706462055444717, + 0.008857139386236668, + 0.012102534994482994, + 0.03113291598856449, + 0.004838082008063793, + 0.006456772796809673, + 0.005043479613959789, + 0.008904258720576763, + 0.017846936360001564, + 0.007705157622694969, + 0.010383046232163906, + 0.009223748929798603, + 0.004135056398808956, + 0.0032863507512956858, + 0.0029841323848813772, + 0.0028987873811274767, + 0.003087815595790744, + 0.004104678053408861, + 0.008107482455670834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.10029598325490952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up captures a confident teenage African-American woman of average build, her black hair slicked back, wet and clinging to her scalp, droplets of water tracking along her strong jawline. She sits in an indoor swimming pool, treading water so vigorously that her shoulders and head remain elevated and steady above the shimmering surface, a testament to her athleticism, allowing small, concentric ripples to spread around her. In her right hand, she expertly holds a bright yellow, textured water ball, her fingers gripping its dimpled surface. The entire scene is bathed in the intense, warm golden glow of the overhead natatorium lights, which shimmer on her skin and create brilliant highlights on the disturbed water. She makes a slight, deliberate motion to lift the ball, a subtle but controlled movement, her head tilted steadily, her focused eyes scanning the length of the pool with strategic intent. The interplay of the warm light, the deep blue water, and her determined expression creates a dynamic, high-contrast visual that encapsulates the atmosphere of peak athletic focus and poised power.", + "seed": 18083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_5/index_435_seed_18083_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08190029952675104, + "vlm_logprob": -0.08190029952675104, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.08190029952675104, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23433764278888702 + ], + "opposite_logprob_tokens": [ + -1.6093376874923706 + ], + "logprob_reward": -0.23433764278888702, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23433764278888702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17472031712532043, + 0.060313597321510315, + 0.08029350638389587, + 0.06103738397359848, + 0.05090110003948212, + 0.37635892629623413, + 0.11897537857294083, + 0.02091396041214466, + 0.02642437256872654, + 0.030061503872275352 + ], + "credit_map": [ + 0.02508554421365261, + 0.028491096571087837, + 0.008262606337666512, + 0.00942887645214796, + 0.037295110523700714, + 0.031128238886594772, + 0.025636648759245872, + 0.009392187930643559, + 0.011293559335172176, + 0.01655609719455242, + 0.00468432204797864, + 0.002935783239081502, + 0.007307486142963171, + 0.008325460366904736, + 0.004873912315815687, + 0.004336978308856487, + 0.009899388067424297, + 0.010845158249139786, + 0.00989451166242361, + 0.004545485135167837, + 0.01345934346318245, + 0.010201173834502697, + 0.015104915015399456, + 0.006343530490994453, + 0.008699034340679646, + 0.012476664036512375, + 0.003184645902365446, + 0.005657895002514124, + 0.010380921885371208, + 0.0053838337771594524, + 0.009981068782508373, + 0.0052733211778104305, + 0.006158799864351749, + 0.011508466675877571, + 0.0036070523783564568, + 0.0032182035502046347, + 0.010394543409347534, + 0.006983439903706312, + 0.005349074024707079, + 0.003681517904624343, + 0.14274616539478302, + 0.021993819624185562, + 0.01029999740421772, + 0.005094047170132399, + 0.015431306324899197, + 0.16490937769412994, + 0.005084834061563015, + 0.010799375362694263, + 0.017052754759788513, + 0.022624263539910316, + 0.012400751002132893, + 0.007006362080574036, + 0.014833586290478706, + 0.022670892998576164, + 0.010846483521163464, + 0.011540288105607033, + 0.0027640615589916706, + 0.004449150525033474, + 0.0013493065489456058, + 0.001959713874384761, + 0.002715261420235038, + 0.002554016886278987, + 0.0026592605281621218, + 0.0024631915148347616, + 0.003348337020725012, + 0.004991590976715088, + 0.0020247003994882107, + 0.005512331612408161, + 0.003439143532887101, + 0.0027938862331211567, + 0.002293484518304467, + 0.0020208985079079866, + 0.0034795980900526047, + 0.005368457175791264, + 0.004584945272654295, + 0.004983771126717329, + 0.0040636188350617886, + 0.0033254267182201147, + 0.0020072211045771837, + 0.002248466480523348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.23433764278888702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906181901693344 + ], + "opposite_logprob_tokens": [ + -2.464061737060547 + ], + "logprob_reward": -0.08906181901693344, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906181901693344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09194021672010422, + 0.057354193180799484, + 0.04814297705888748, + 0.0899004265666008, + 0.026755737140774727, + 0.5568851232528687, + 0.032616619020700455, + 0.03583063930273056, + 0.038468461483716965, + 0.022105539217591286 + ], + "credit_map": [ + 0.013677149079740047, + 0.017989570274949074, + 0.008049985393881798, + 0.004404861479997635, + 0.012596448883414268, + 0.021853219717741013, + 0.008557150140404701, + 0.0048118396662175655, + 0.009613226167857647, + 0.014441986568272114, + 0.006557869724929333, + 0.0034154485911130905, + 0.00549085158854723, + 0.008774361573159695, + 0.004718593321740627, + 0.0043418630957603455, + 0.00720693776383996, + 0.00746502261608839, + 0.004860060755163431, + 0.0034244840499013662, + 0.007349053863435984, + 0.0077910879626870155, + 0.004818148910999298, + 0.005228186491876841, + 0.012152626179158688, + 0.01998826675117016, + 0.005615935195237398, + 0.009184163995087147, + 0.0113321952521801, + 0.01040751300752163, + 0.014401339925825596, + 0.006818399298936129, + 0.004209219478070736, + 0.0047119795344769955, + 0.002914134180173278, + 0.0016568839782848954, + 0.004590568132698536, + 0.0032273565884679556, + 0.003298077266663313, + 0.002147521823644638, + 0.21132692694664001, + 0.033086858689785004, + 0.013736451044678688, + 0.008321864530444145, + 0.022561537101864815, + 0.2439315915107727, + 0.007846087217330933, + 0.01607384718954563, + 0.005846802145242691, + 0.0052191950380802155, + 0.002903055865317583, + 0.002233243780210614, + 0.0032272732350975275, + 0.006659857463091612, + 0.002826876938343048, + 0.0037003171164542437, + 0.0049684070982038975, + 0.008991057053208351, + 0.002681209472939372, + 0.004176731687039137, + 0.004147444851696491, + 0.0036270185373723507, + 0.0035340385511517525, + 0.0037047388032078743, + 0.004890082869678736, + 0.008754152804613113, + 0.0026790499687194824, + 0.006226435769349337, + 0.0050394730642437935, + 0.005455489736050367, + 0.002399744698777795, + 0.0030240395572036505, + 0.0037306330632418394, + 0.0035629086196422577, + 0.002232570433989167, + 0.0020838002674281597, + 0.0026472541503608227, + 0.0025587081909179688, + 0.00298503995873034, + 0.002304625930264592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.08906181901693344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646339945495129 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646339945495129, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646339945495129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10764565318822861, + 0.11533576250076294, + 0.08841254562139511, + 0.04564814642071724, + 0.03433993086218834, + 0.318968802690506, + 0.11894938349723816, + 0.051842961460351944, + 0.04922015964984894, + 0.06963662058115005 + ], + "credit_map": [ + 0.015394722111523151, + 0.02172992005944252, + 0.013542106375098228, + 0.007571808993816376, + 0.0151229752227664, + 0.013079703785479069, + 0.013901167549192905, + 0.007303244434297085, + 0.030633186921477318, + 0.03958900272846222, + 0.007695599924772978, + 0.003692035563290119, + 0.010798431001603603, + 0.012497699819505215, + 0.005449394695460796, + 0.004980419296771288, + 0.01963975466787815, + 0.009702341631054878, + 0.007603420410305262, + 0.004258630331605673, + 0.010291253216564655, + 0.02437777630984783, + 0.007348071318119764, + 0.005191299133002758, + 0.00633809482678771, + 0.008062941953539848, + 0.003845584113150835, + 0.002483720425516367, + 0.007953260093927383, + 0.005413148086518049, + 0.007028738968074322, + 0.004522660747170448, + 0.005473425146192312, + 0.005659244954586029, + 0.0032255479600280523, + 0.003429555334150791, + 0.00546537758782506, + 0.004431314766407013, + 0.004109290894120932, + 0.0025461737532168627, + 0.11711220443248749, + 0.021581241860985756, + 0.007860353216528893, + 0.005371865350753069, + 0.015962732955813408, + 0.13484351336956024, + 0.006340010557323694, + 0.009896903298795223, + 0.016289541497826576, + 0.017841899767518044, + 0.011007538065314293, + 0.010117217898368835, + 0.014256823807954788, + 0.018607448786497116, + 0.011516452766954899, + 0.019312461838126183, + 0.008166661486029625, + 0.008438349701464176, + 0.0043321880511939526, + 0.006159894168376923, + 0.0053366366773843765, + 0.008269652724266052, + 0.005583098158240318, + 0.005556478165090084, + 0.00705261854454875, + 0.008711506612598896, + 0.004494389519095421, + 0.008407416753470898, + 0.007411432918161154, + 0.006031053606420755, + 0.0034587127156555653, + 0.0036530315410345793, + 0.011436819098889828, + 0.010370219126343727, + 0.004409434739500284, + 0.006251530256122351, + 0.011488877236843109, + 0.013033386319875717, + 0.00796034187078476, + 0.004686015658080578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.007646339945495129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048658184707164764 + ], + "opposite_logprob_tokens": [ + -3.0486581325531006 + ], + "logprob_reward": -0.048658184707164764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048658184707164764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32245635986328125, + 0.10180115699768066, + 0.05929941684007645, + 0.08867863565683365, + 0.05333961918950081, + 0.22280368208885193, + 0.02336394600570202, + 0.01968909054994583, + 0.03566046804189682, + 0.07290762662887573 + ], + "credit_map": [ + 0.06870293617248535, + 0.06209089607000351, + 0.020748553797602654, + 0.026893269270658493, + 0.03457874432206154, + 0.059300683438777924, + 0.03274239972233772, + 0.017398888245224953, + 0.020685594528913498, + 0.023768465965986252, + 0.01382683590054512, + 0.006079571321606636, + 0.010201551951467991, + 0.010276230052113533, + 0.009884105995297432, + 0.007078804541379213, + 0.011986654251813889, + 0.01158990990370512, + 0.0064875115640461445, + 0.0027233180589973927, + 0.009303676895797253, + 0.008048763498663902, + 0.0059855603612959385, + 0.0031740230042487383, + 0.015078718774020672, + 0.014032658189535141, + 0.008848769590258598, + 0.00891442596912384, + 0.011167047545313835, + 0.012596364133059978, + 0.010422619991004467, + 0.00761802913621068, + 0.008556093089282513, + 0.0098070427775383, + 0.005998225882649422, + 0.0043892282992601395, + 0.00838380679488182, + 0.007830984890460968, + 0.004468090366572142, + 0.003906145691871643, + 0.0819883793592453, + 0.01491820253431797, + 0.00616165017709136, + 0.004123439081013203, + 0.008929531089961529, + 0.09523096680641174, + 0.004170565865933895, + 0.007280959747731686, + 0.003871653228998184, + 0.0037116932217031717, + 0.0022337259724736214, + 0.0014170564245432615, + 0.0034159738570451736, + 0.004090502392500639, + 0.0018771450268104672, + 0.0027461969293653965, + 0.002630516653880477, + 0.0038341106846928596, + 0.001701985253021121, + 0.0016568938735872507, + 0.0036249267868697643, + 0.0022732641082257032, + 0.0016915189335122705, + 0.0022758746054023504, + 0.004952722694724798, + 0.005716701038181782, + 0.0035070041194558144, + 0.006556936074048281, + 0.005221654195338488, + 0.004130511544644833, + 0.0024751380551606417, + 0.003099802415817976, + 0.014225183986127377, + 0.015351149253547192, + 0.005309586878865957, + 0.00951799750328064, + 0.0082170395180583, + 0.011455735191702843, + 0.0048029376193881035, + 0.004027999006211758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.048658184707164764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a confident teenage African-American woman with black hair slicked back and wet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting in an indoor swimming pool and treading water vigorously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a bright yellow, textured water ball in her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense warm golden overhead lighting creating highlights on the water and skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman make a slight, deliberate motion to lift the water ball while her head remains tilted steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979751117527485 + ], + "opposite_logprob_tokens": [ + -3.5297975540161133 + ], + "logprob_reward": -0.02979751117527485, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979751117527485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12514476478099823, + 0.07347836345434189, + 0.08132390677928925, + 0.11463145166635513, + 0.039840031415224075, + 0.24791838228702545, + 0.07325813174247742, + 0.12548702955245972, + 0.06652563065290451, + 0.05239231139421463 + ], + "credit_map": [ + 0.01571986824274063, + 0.02257676236331463, + 0.00958948489278555, + 0.007499044761061668, + 0.029833883047103882, + 0.013209037482738495, + 0.02031796984374523, + 0.006398702040314674, + 0.010363076813519001, + 0.012542251497507095, + 0.01053517498075962, + 0.0057697235606610775, + 0.009851275011897087, + 0.009949089027941227, + 0.007955585606396198, + 0.006512182764708996, + 0.01663431152701378, + 0.011604883708059788, + 0.004583324305713177, + 0.004004983231425285, + 0.009755588136613369, + 0.019526537507772446, + 0.0108328340575099, + 0.0043814522214233875, + 0.015768466517329216, + 0.020591257140040398, + 0.0067898244597017765, + 0.008491717278957367, + 0.01704000122845173, + 0.010277344845235348, + 0.02428996004164219, + 0.0113828731700778, + 0.005158030427992344, + 0.005486049223691225, + 0.003426677081733942, + 0.0035834177397191525, + 0.00733915762975812, + 0.004014643840491772, + 0.00825525727123022, + 0.0025767977349460125, + 0.09328881651163101, + 0.014366869814693928, + 0.0059079453349113464, + 0.003757086116820574, + 0.010291596874594688, + 0.10693320631980896, + 0.005091676954180002, + 0.008281184360384941, + 0.010210201144218445, + 0.015043803490698338, + 0.011113159358501434, + 0.005875570233911276, + 0.00970240868628025, + 0.010045967996120453, + 0.005308741703629494, + 0.00595827866345644, + 0.02227412909269333, + 0.033329445868730545, + 0.007881640456616879, + 0.013651032000780106, + 0.016100071370601654, + 0.01646876521408558, + 0.007912972941994667, + 0.007868992164731026, + 0.007544708903878927, + 0.011310545727610588, + 0.00904312077909708, + 0.010694337077438831, + 0.007275017909705639, + 0.007975478656589985, + 0.005418822169303894, + 0.007263597566634417, + 0.008786100894212723, + 0.005996368359774351, + 0.0031739347614347935, + 0.005147597752511501, + 0.005068338010460138, + 0.009994118474423885, + 0.006999386940151453, + 0.0072264643386006355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.02979751117527485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_tensor([846, 846])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_tensor([846, 846])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a09a5550274a46e15877b3007ff1b677206d4b4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_tensor([846, 846])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful low-angle tracking shot follows a heavy steel freight train, its charcoal-black front covered in riveted steel and mottled with rust, as it hurtles forward with unstoppable momentum. The lone round headlight casts a glaring beam through the gathering twilight, illuminating the abandoned, overgrown tracks ahead. The massive steel wheels relentlessly grind against the rusted track, emitting a dull metallic thud, shaking the splintered railroad ties and the surrounding gravel roadbed slightly and violently pushing the overgrown weeds to the sides. The low-slanting sunset casts long, twisting shadows across the barren landscape, gilding the weathered sides of the train with a fleeting warm golden hue. Overhead, dark clouds gather, foreshadowing a storm. The dramatic interplay of light and shadow imbues the entire scene with a heavy, resolute, and desolate atmosphere.", + "seed": 23252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_846_seed_23252_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19199804328382014, + "vlm_logprob": -0.19199804328382014, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.19199804328382014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24209122359752655 + ], + "opposite_logprob_tokens": [ + -1.617091178894043 + ], + "logprob_reward": -0.24209122359752655, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24209122359752655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17610131204128265, + 0.035628266632556915, + 0.040005188435316086, + 0.3266759514808655, + 0.02650666981935501, + 0.12164786458015442, + 0.04368743300437927, + 0.0362013578414917, + 0.03989967703819275, + 0.15364624559879303 + ], + "credit_map": [ + 0.012323739007115364, + 0.06022047996520996, + 0.013994974084198475, + 0.01816125027835369, + 0.02169831469655037, + 0.021967759355902672, + 0.010755719617009163, + 0.016979089006781578, + 0.0030436471570283175, + 0.007837883196771145, + 0.003125872928649187, + 0.00395278399810195, + 0.004954817704856396, + 0.004559568595141172, + 0.004926970694214106, + 0.0032267235219478607, + 0.003384549170732498, + 0.006975645199418068, + 0.0035811741836369038, + 0.0043168338015675545, + 0.008422650396823883, + 0.005361427087336779, + 0.003617028007283807, + 0.004345880821347237, + 0.020046047866344452, + 0.021563421934843063, + 0.021542692556977272, + 0.0358562245965004, + 0.022676393389701843, + 0.0805349349975586, + 0.10370061546564102, + 0.02075560949742794, + 0.0016623310511931777, + 0.003111822297796607, + 0.0033908458426594734, + 0.0030534067191183567, + 0.0036169670056551695, + 0.004449636209756136, + 0.0030261389911174774, + 0.004195521119982004, + 0.005746110342442989, + 0.011796403676271439, + 0.0093555161729455, + 0.016778824850916862, + 0.01977492682635784, + 0.018704984337091446, + 0.02059515193104744, + 0.018895944580435753, + 0.0023611485958099365, + 0.0034682981204241514, + 0.006894341669976711, + 0.005306906532496214, + 0.008699650876224041, + 0.004422094207257032, + 0.006202682852745056, + 0.00633231271058321, + 0.0019310148200020194, + 0.0024606711231172085, + 0.0035887903068214655, + 0.006112789735198021, + 0.005252372939139605, + 0.005038382951170206, + 0.007016947027295828, + 0.004800386726856232, + 0.003394964151084423, + 0.003257161006331444, + 0.008491839282214642, + 0.005135641898959875, + 0.006076643243432045, + 0.006371255032718182, + 0.003685403149574995, + 0.003486769972369075, + 0.010982878506183624, + 0.00869296956807375, + 0.012724664993584156, + 0.02214796096086502, + 0.025078415870666504, + 0.027937179431319237, + 0.024593666195869446, + 0.021488510072231293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.24209122359752655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328296065330505 + ], + "opposite_logprob_tokens": [ + -0.7578296065330505 + ], + "logprob_reward": -0.6328296065330505, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328296065330505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05230625346302986, + 0.036017339676618576, + 0.027957858517766, + 0.5708874464035034, + 0.03363363817334175, + 0.06781679391860962, + 0.055019136518239975, + 0.09871452301740646, + 0.015196164138615131, + 0.04245082288980484 + ], + "credit_map": [ + 0.005522485356777906, + 0.015738531947135925, + 0.0035111578181385994, + 0.004980172496289015, + 0.005750955548137426, + 0.008004160597920418, + 0.004031999036669731, + 0.004766792990267277, + 0.003405072260648012, + 0.007811316754668951, + 0.005341139622032642, + 0.0037013848777860403, + 0.003343805903568864, + 0.005466506350785494, + 0.0035749187227338552, + 0.0033731956500560045, + 0.00270766019821167, + 0.004200897179543972, + 0.002315906109288335, + 0.002834385260939598, + 0.006086552981287241, + 0.005045743193477392, + 0.002365939086303115, + 0.0024007735773921013, + 0.03233986720442772, + 0.04060586169362068, + 0.04088815674185753, + 0.059023477137088776, + 0.04409072548151016, + 0.1508614718914032, + 0.18167005479335785, + 0.02140786685049534, + 0.002141492674127221, + 0.002424427308142185, + 0.003778217127546668, + 0.00253273150883615, + 0.008130076341331005, + 0.0037842532619833946, + 0.003186250803992152, + 0.007656186819076538, + 0.0024705303367227316, + 0.0050633312202990055, + 0.004667573142796755, + 0.008771711960434914, + 0.009419441223144531, + 0.015145843848586082, + 0.009771130047738552, + 0.01250723097473383, + 0.0032406183890998363, + 0.0042707244865596294, + 0.00785752758383751, + 0.005632971413433552, + 0.011658423580229282, + 0.006715239491313696, + 0.006775196176022291, + 0.008868439123034477, + 0.0062978221103549, + 0.0058604017831385136, + 0.010389882139861584, + 0.015178716741502285, + 0.010690237395465374, + 0.011769120581448078, + 0.0247577503323555, + 0.01377058494836092, + 0.001299907686188817, + 0.0007783457986079156, + 0.002481134608387947, + 0.001780560240149498, + 0.0016835472779348493, + 0.002383011393249035, + 0.0017465975834056735, + 0.003043058793991804, + 0.0027483359444886446, + 0.0030204253271222115, + 0.0036954516544938087, + 0.006433341186493635, + 0.0077780853025615215, + 0.00848264992237091, + 0.007001344580203295, + 0.003291188506409526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.6328296065330505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371855288743973 + ], + "opposite_logprob_tokens": [ + -3.651371955871582 + ], + "logprob_reward": -0.026371855288743973, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371855288743973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24390439689159393, + 0.08842063695192337, + 0.057941824197769165, + 0.33566054701805115, + 0.029856175184249878, + 0.06564822793006897, + 0.05488719791173935, + 0.02805044874548912, + 0.03571585565805435, + 0.05991477146744728 + ], + "credit_map": [ + 0.015693949535489082, + 0.07536502927541733, + 0.02042197249829769, + 0.028557900339365005, + 0.029536113142967224, + 0.02883552573621273, + 0.01959289237856865, + 0.025901010259985924, + 0.008848950266838074, + 0.014317359775304794, + 0.007055300287902355, + 0.00756870349869132, + 0.013262360356748104, + 0.017945604398846626, + 0.011592650786042213, + 0.00782970804721117, + 0.005629526451230049, + 0.012797949835658073, + 0.004413872491568327, + 0.006546286400407553, + 0.011102184653282166, + 0.008259451016783714, + 0.004534801468253136, + 0.0046577490866184235, + 0.02077125385403633, + 0.024302057921886444, + 0.022667989134788513, + 0.036447133868932724, + 0.028270110487937927, + 0.07919187843799591, + 0.10290414839982986, + 0.021105969324707985, + 0.002025154884904623, + 0.004385279957205057, + 0.0036436289083212614, + 0.002812882186844945, + 0.004493690095841885, + 0.005366485565900803, + 0.0034703630954027176, + 0.003658688860014081, + 0.003282304620370269, + 0.005942533724009991, + 0.0066358414478600025, + 0.007334953639656305, + 0.011396876536309719, + 0.011640897952020168, + 0.009211571887135506, + 0.01020325068384409, + 0.004939500708132982, + 0.0042942422442138195, + 0.008678952232003212, + 0.006412121467292309, + 0.011342082172632217, + 0.005478852894157171, + 0.005801465827971697, + 0.007939977571368217, + 0.0025314996019005775, + 0.002225721487775445, + 0.0043505653738975525, + 0.0049540507607162, + 0.002856943989172578, + 0.0030215715523809195, + 0.00435118330642581, + 0.003758911741897464, + 0.0026174760423600674, + 0.0031354816164821386, + 0.00637383246794343, + 0.004073846619576216, + 0.005885849706828594, + 0.005292851943522692, + 0.004460252821445465, + 0.0038762653712183237, + 0.004352041985839605, + 0.0032098672818392515, + 0.004142068792134523, + 0.008244985714554787, + 0.009896742179989815, + 0.010599897243082523, + 0.010085215792059898, + 0.009383950382471085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.026371855288743973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811405599117279 + ], + "opposite_logprob_tokens": [ + -3.288114070892334 + ], + "logprob_reward": -0.03811405599117279, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811405599117279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19555266201496124, + 0.04763060808181763, + 0.06557852774858475, + 0.2550070881843567, + 0.06880602985620499, + 0.12420301139354706, + 0.04377787187695503, + 0.08044459670782089, + 0.045315176248550415, + 0.07368436455726624 + ], + "credit_map": [ + 0.01255098171532154, + 0.06011504307389259, + 0.015599900856614113, + 0.019063150510191917, + 0.02330007404088974, + 0.025546250864863396, + 0.015180841088294983, + 0.024196423590183258, + 0.004468837752938271, + 0.00970937218517065, + 0.004203410353511572, + 0.004780913237482309, + 0.0066213407553732395, + 0.0068693459033966064, + 0.006848047953099012, + 0.00412934273481369, + 0.005785994231700897, + 0.008113594725728035, + 0.006518357899039984, + 0.007470059674233198, + 0.008840953931212425, + 0.009398226626217365, + 0.007327686529606581, + 0.012123662047088146, + 0.014218373224139214, + 0.017942266538739204, + 0.01696605421602726, + 0.02945581264793873, + 0.019837936386466026, + 0.05946891009807587, + 0.08108242601156235, + 0.016035335138440132, + 0.004524515941739082, + 0.008738977834582329, + 0.006120066624134779, + 0.007398543879389763, + 0.009708896279335022, + 0.01250150054693222, + 0.010147368535399437, + 0.009666171856224537, + 0.006574824918061495, + 0.01018358301371336, + 0.010272545740008354, + 0.018848218023777008, + 0.01861480623483658, + 0.017888445407152176, + 0.023220723494887352, + 0.01859988085925579, + 0.00351316062733531, + 0.004173917695879936, + 0.004119198303669691, + 0.004678965080529451, + 0.010499865747988224, + 0.005349821876734495, + 0.005384017247706652, + 0.006058926694095135, + 0.006051782052963972, + 0.0057151904329657555, + 0.009517841972410679, + 0.015184225514531136, + 0.007022981997579336, + 0.009003257378935814, + 0.01752595603466034, + 0.010423365980386734, + 0.003503479529172182, + 0.0035718893632292747, + 0.008821175433695316, + 0.0054293423891067505, + 0.007001150399446487, + 0.006333004683256149, + 0.005984614137560129, + 0.004670523572713137, + 0.004780471790581942, + 0.005006616003811359, + 0.003927474841475487, + 0.010235285386443138, + 0.014418546110391617, + 0.01526684034615755, + 0.012582755647599697, + 0.007466380018740892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.03811405599117279 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3396685719490051, + 0.16287297010421753, + 0.024504633620381355, + 0.2675396203994751, + 0.026798242703080177, + 0.034411486238241196, + 0.02566542848944664, + 0.07441544532775879, + 0.021786419674754143, + 0.022337138652801514 + ], + "credit_map": [ + 0.020798396319150925, + 0.08112508058547974, + 0.024150582030415535, + 0.032216496765613556, + 0.041655197739601135, + 0.05790261924266815, + 0.03328191116452217, + 0.04853830486536026, + 0.01347802672535181, + 0.024365441873669624, + 0.010376779362559319, + 0.01291039027273655, + 0.022261474281549454, + 0.03494258224964142, + 0.031297773122787476, + 0.013240509666502476, + 0.0019331284565851092, + 0.00474747596308589, + 0.002193786669522524, + 0.0025114293675869703, + 0.003743499517440796, + 0.0036540785804390907, + 0.0034996378235518932, + 0.0022215996868908405, + 0.015756765380501747, + 0.012004079297184944, + 0.016570301726460457, + 0.028691833838820457, + 0.01401561964303255, + 0.07471566647291183, + 0.09241829812526703, + 0.013367045670747757, + 0.0015648090047761798, + 0.0042393263429403305, + 0.002726803533732891, + 0.003446100279688835, + 0.0030909853521734476, + 0.003798224264755845, + 0.0050050485879182816, + 0.0029269454535096884, + 0.0015999894822016358, + 0.0035399007610976696, + 0.0020452183671295643, + 0.005623529199510813, + 0.0036542010493576527, + 0.00516819953918457, + 0.007511962670832872, + 0.005268485751003027, + 0.0024930608924478292, + 0.0017483500996604562, + 0.003950225654989481, + 0.002657139440998435, + 0.00434254901483655, + 0.00337739335373044, + 0.003102084156125784, + 0.003994627855718136, + 0.004274492152035236, + 0.004440880846232176, + 0.007088880054652691, + 0.0110640162602067, + 0.009220948442816734, + 0.009344798512756824, + 0.01864941045641899, + 0.010332026518881321, + 0.001597715774551034, + 0.0017013452015817165, + 0.0040646810084581375, + 0.0026509789749979973, + 0.00269807456061244, + 0.003259263001382351, + 0.0027478907722979784, + 0.0030664701480418444, + 0.0019307106267660856, + 0.0013660038821399212, + 0.002660646801814437, + 0.003400103421881795, + 0.0038245057221502066, + 0.0038074625190347433, + 0.003814345458522439, + 0.0015333605697378516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02058347500860691 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful low-angle tracking shot follows a heavy steel freight train, its charcoal-black front covered in riveted steel and mottled with rust, as it hurtles forward with unstoppable momentum. The lone round headlight casts a glaring beam through the gathering twilight, illuminating the abandoned, overgrown tracks ahead. The massive steel wheels relentlessly grind against the rusted track, emitting a dull metallic thud, shaking the splintered railroad ties and the surrounding gravel roadbed slightly and violently pushing the overgrown weeds to the sides. The low-slanting sunset casts long, twisting shadows across the barren landscape, gilding the weathered sides of the train with a fleeting warm golden hue. Overhead, dark clouds gather, foreshadowing a storm. The dramatic interplay of light and shadow imbues the entire scene with a heavy, resolute, and desolate atmosphere.", + "seed": 23253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_846_seed_23253_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24835051102563738, + "vlm_logprob": -0.24835051102563738, + "vlm_yes_no_margin": 2.7749999880790712, + "vlm_token_credit": -0.24835051102563738, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038940150290727615 + ], + "opposite_logprob_tokens": [ + -3.413940191268921 + ], + "logprob_reward": -0.038940150290727615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.038940150290727615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0854106917977333, + 0.026181019842624664, + 0.02991507574915886, + 0.11505556106567383, + 0.07420483231544495, + 0.15514661371707916, + 0.11896686255931854, + 0.05901780351996422, + 0.10673771053552628, + 0.2293638437986374 + ], + "credit_map": [ + 0.011176399886608124, + 0.012652790173888206, + 0.010661039501428604, + 0.010418067686259747, + 0.011256751604378223, + 0.010068397969007492, + 0.007412695791572332, + 0.01176455244421959, + 0.004500101786106825, + 0.003358517773449421, + 0.0031516056042164564, + 0.0037416648119688034, + 0.0033884451258927584, + 0.0024851947091519833, + 0.0029871519654989243, + 0.0025683376006782055, + 0.004359698388725519, + 0.0035326615907251835, + 0.003244072897359729, + 0.003730854718014598, + 0.0049780914559960365, + 0.002353473100811243, + 0.004673883318901062, + 0.0030423409771174192, + 0.011982478201389313, + 0.013607789762318134, + 0.013487685471773148, + 0.009531931951642036, + 0.031895387917757034, + 0.010256465524435043, + 0.01501814741641283, + 0.009275674819946289, + 0.011423799209296703, + 0.00881093367934227, + 0.010993174277245998, + 0.005942274816334248, + 0.008425487205386162, + 0.011989463120698929, + 0.009566579945385456, + 0.007053117733448744, + 0.014372826553881168, + 0.012387845665216446, + 0.017662985250353813, + 0.015541085042059422, + 0.032736923545598984, + 0.016263287514448166, + 0.027620350942015648, + 0.018561309203505516, + 0.021204469725489616, + 0.011236399412155151, + 0.011620404198765755, + 0.009026588872075081, + 0.01841539517045021, + 0.022461166605353355, + 0.013912631198763847, + 0.011089809238910675, + 0.00595081876963377, + 0.0039436607621610165, + 0.005639619659632444, + 0.0058474852703511715, + 0.007736261934041977, + 0.009568881243467331, + 0.01097757089883089, + 0.00935350451618433, + 0.008178831078112125, + 0.006834868341684341, + 0.00950463954359293, + 0.008222549222409725, + 0.015271713957190514, + 0.01834951899945736, + 0.023632729426026344, + 0.01674286089837551, + 0.0137868020683527, + 0.017462147399783134, + 0.014424211345613003, + 0.015062873251736164, + 0.046554554253816605, + 0.035954639315605164, + 0.04511662572622299, + 0.041001979261636734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.038940150290727615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538957118988037 + ], + "opposite_logprob_tokens": [ + -0.4288957715034485 + ], + "logprob_reward": -1.0538957118988037, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0538957118988037, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.049286242574453354, + 0.03185250610113144, + 0.03142884001135826, + 0.1733611822128296, + 0.08853572607040405, + 0.10603073984384537, + 0.22270935773849487, + 0.11099152266979218, + 0.06787838786840439, + 0.11792550981044769 + ], + "credit_map": [ + 0.007092077285051346, + 0.006405549123883247, + 0.004987846128642559, + 0.006218737456947565, + 0.008120153099298477, + 0.006492984015494585, + 0.004970446694642305, + 0.004998443182557821, + 0.005143975373357534, + 0.004570393823087215, + 0.0032456640619784594, + 0.0030126760248094797, + 0.00650301668792963, + 0.003311741165816784, + 0.0034418206196278334, + 0.0026232183445245028, + 0.004035087767988443, + 0.0034672380425035954, + 0.0039748018607497215, + 0.003765692003071308, + 0.006102273240685463, + 0.002560567343607545, + 0.00498319324105978, + 0.0025399886071681976, + 0.02080887369811535, + 0.022419512271881104, + 0.015639491379261017, + 0.013924075290560722, + 0.04004542529582977, + 0.020365381613373756, + 0.02365056425333023, + 0.016507865861058235, + 0.013040862046182156, + 0.009877599775791168, + 0.009762894362211227, + 0.00715956836938858, + 0.013239705935120583, + 0.017930682748556137, + 0.008287898264825344, + 0.00923650711774826, + 0.012624246068298817, + 0.010043978691101074, + 0.013425475917756557, + 0.008561343885958195, + 0.026385722681879997, + 0.011572902090847492, + 0.014865233562886715, + 0.008551841601729393, + 0.04903482645750046, + 0.010637854225933552, + 0.01464046724140644, + 0.014759634621441364, + 0.023410091176629066, + 0.062752865254879, + 0.02785375714302063, + 0.019619883969426155, + 0.013161562383174896, + 0.0075326766818761826, + 0.010266016237437725, + 0.008031046018004417, + 0.01787777803838253, + 0.018190480768680573, + 0.017225541174411774, + 0.01870642602443695, + 0.006259841378778219, + 0.0036882739514112473, + 0.006752369925379753, + 0.0046178544871509075, + 0.010549215599894524, + 0.012177986092865467, + 0.012516429647803307, + 0.011316419579088688, + 0.0060398462228477, + 0.00496967788785696, + 0.006591728422790766, + 0.004823371302336454, + 0.016157280653715134, + 0.023443741723895073, + 0.024696793407201767, + 0.031203074380755424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -1.0538957118988037 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212341979146004 + ], + "opposite_logprob_tokens": [ + -4.264212131500244 + ], + "logprob_reward": -0.014212341979146004, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212341979146004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15211300551891327, + 0.03595041483640671, + 0.0519220307469368, + 0.2245161533355713, + 0.054668426513671875, + 0.161545991897583, + 0.056978076696395874, + 0.10401464253664017, + 0.060778334736824036, + 0.09751297533512115 + ], + "credit_map": [ + 0.027724003419280052, + 0.01784246787428856, + 0.018581582233309746, + 0.01697039045393467, + 0.018978528678417206, + 0.01807224564254284, + 0.01272489782422781, + 0.021218866109848022, + 0.005013828165829182, + 0.005745814181864262, + 0.003717228304594755, + 0.004106283187866211, + 0.0076853688806295395, + 0.003660243470221758, + 0.003341076895594597, + 0.002680564997717738, + 0.006775912828743458, + 0.005814289674162865, + 0.006234403699636459, + 0.006035730708390474, + 0.008385594934225082, + 0.004489782731980085, + 0.008885174058377743, + 0.005301142111420631, + 0.028678959235548973, + 0.029218433424830437, + 0.01648183912038803, + 0.01911233365535736, + 0.05222173035144806, + 0.016114603728055954, + 0.0400724858045578, + 0.022615747526288033, + 0.007502881810069084, + 0.005985886789858341, + 0.006710386835038662, + 0.0039011116605252028, + 0.00881862360984087, + 0.009395943023264408, + 0.00681905634701252, + 0.005534532945603132, + 0.015271099284291267, + 0.014147450216114521, + 0.01823967695236206, + 0.015216709114611149, + 0.03323730453848839, + 0.017199313268065453, + 0.028421884402632713, + 0.019812537357211113, + 0.007003750652074814, + 0.00307068950496614, + 0.006281376350671053, + 0.005304490216076374, + 0.011428960599005222, + 0.010928424075245857, + 0.007624433375895023, + 0.00533594936132431, + 0.00856821984052658, + 0.00907852128148079, + 0.007840326055884361, + 0.0069129266776144505, + 0.018113993108272552, + 0.018828008323907852, + 0.02075693942606449, + 0.01391570270061493, + 0.004845122806727886, + 0.0037068245001137257, + 0.004914130549877882, + 0.004941346123814583, + 0.008315171115100384, + 0.010501942597329617, + 0.01524238009005785, + 0.008311409503221512, + 0.005781663116067648, + 0.005946204997599125, + 0.007429997902363539, + 0.0055436729453504086, + 0.014901149086654186, + 0.01569226011633873, + 0.02164537087082863, + 0.020572656765580177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.014212341979146004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707315385341644 + ], + "opposite_logprob_tokens": [ + -2.127073049545288 + ], + "logprob_reward": -0.12707315385341644, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12707315385341644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09541492909193039, + 0.058606378734111786, + 0.0736355260014534, + 0.07192685455083847, + 0.22456730902194977, + 0.1119757816195488, + 0.07500346750020981, + 0.09146305918693542, + 0.06434634327888489, + 0.13306032121181488 + ], + "credit_map": [ + 0.0149246696382761, + 0.013591655530035496, + 0.010404777713119984, + 0.012523823417723179, + 0.013383232057094574, + 0.010794314555823803, + 0.009263797663152218, + 0.010528667829930782, + 0.010867903009057045, + 0.006839442066848278, + 0.006899026222527027, + 0.006244716700166464, + 0.00874674879014492, + 0.006369935814291239, + 0.005502419546246529, + 0.007136190310120583, + 0.011575768701732159, + 0.009326811879873276, + 0.006714547984302044, + 0.00940160546451807, + 0.012883094139397144, + 0.005243406631052494, + 0.011633875779807568, + 0.006856421008706093, + 0.009806330315768719, + 0.009390883147716522, + 0.007167892996221781, + 0.007426092401146889, + 0.01229309756308794, + 0.005482763517647982, + 0.013394368812441826, + 0.006965433247387409, + 0.03456934168934822, + 0.01899886690080166, + 0.04355520382523537, + 0.024348994717001915, + 0.021294962614774704, + 0.03131674602627754, + 0.027385443449020386, + 0.023097774013876915, + 0.01393096148967743, + 0.008216498419642448, + 0.01907830685377121, + 0.008568248711526394, + 0.014242591336369514, + 0.01228751614689827, + 0.015391228720545769, + 0.020260436460375786, + 0.01599368080496788, + 0.009713627398014069, + 0.0055425432510674, + 0.004467816557735205, + 0.011318805627524853, + 0.014338142238557339, + 0.007458413951098919, + 0.006170447915792465, + 0.006995823699980974, + 0.005551726091653109, + 0.009533267468214035, + 0.007621225900948048, + 0.014477645047008991, + 0.015884781256318092, + 0.0183017048984766, + 0.013096891343593597, + 0.005662698298692703, + 0.003258055541664362, + 0.006577180698513985, + 0.0035581840202212334, + 0.006941178813576698, + 0.012499113567173481, + 0.009942766278982162, + 0.015907173976302147, + 0.007146420888602734, + 0.006257643923163414, + 0.0070439232513308525, + 0.005437060724943876, + 0.012198631651699543, + 0.026348523795604706, + 0.027631647884845734, + 0.040996477007865906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.12707315385341644 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631197106093168 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631197106093168, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631197106093168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1264457255601883, + 0.04098382964730263, + 0.04643065482378006, + 0.12215618044137955, + 0.11146420240402222, + 0.17318353056907654, + 0.12414617836475372, + 0.11453315615653992, + 0.06987253576517105, + 0.07078403234481812 + ], + "credit_map": [ + 0.021938860416412354, + 0.015369786880910397, + 0.0182432122528553, + 0.012598089873790741, + 0.020266924053430557, + 0.014066861942410469, + 0.008326591923832893, + 0.015635374933481216, + 0.007433538790792227, + 0.005823367740958929, + 0.0054968735203146935, + 0.003876845818012953, + 0.005910252220928669, + 0.004008176736533642, + 0.003688855329528451, + 0.004745913669466972, + 0.006367848254740238, + 0.005160727072507143, + 0.006500461138784885, + 0.00482627097517252, + 0.008544118143618107, + 0.0035006923135370016, + 0.007471193093806505, + 0.004059338476508856, + 0.01367608830332756, + 0.016300497576594353, + 0.011520720086991787, + 0.011244534514844418, + 0.02434396557509899, + 0.013195347972214222, + 0.019159825518727303, + 0.012715190649032593, + 0.01614413410425186, + 0.01832401566207409, + 0.011448383331298828, + 0.009066474623978138, + 0.012612112797796726, + 0.024629270657896996, + 0.008759500458836555, + 0.010480297729372978, + 0.020310085266828537, + 0.012967702001333237, + 0.02169874683022499, + 0.014842210337519646, + 0.04201779142022133, + 0.019097069278359413, + 0.024448929354548454, + 0.01780097559094429, + 0.024924634024500847, + 0.011895515955984592, + 0.010547122918069363, + 0.008353911340236664, + 0.012709325179457664, + 0.031709931790828705, + 0.015054618008434772, + 0.008951103314757347, + 0.012011486105620861, + 0.008083103224635124, + 0.016120070591568947, + 0.006799234077334404, + 0.0167582705616951, + 0.020309310406446457, + 0.018720127642154694, + 0.01573154143989086, + 0.0071128131821751595, + 0.003335167421028018, + 0.006440440192818642, + 0.005004659295082092, + 0.010768061503767967, + 0.01272627618163824, + 0.014351828023791313, + 0.010133284144103527, + 0.004768603015691042, + 0.0048109800554811954, + 0.005930168554186821, + 0.004608028568327427, + 0.013870900496840477, + 0.012411851435899734, + 0.012202429585158825, + 0.012181064113974571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.007631197106093168 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful low-angle tracking shot follows a heavy steel freight train, its charcoal-black front covered in riveted steel and mottled with rust, as it hurtles forward with unstoppable momentum. The lone round headlight casts a glaring beam through the gathering twilight, illuminating the abandoned, overgrown tracks ahead. The massive steel wheels relentlessly grind against the rusted track, emitting a dull metallic thud, shaking the splintered railroad ties and the surrounding gravel roadbed slightly and violently pushing the overgrown weeds to the sides. The low-slanting sunset casts long, twisting shadows across the barren landscape, gilding the weathered sides of the train with a fleeting warm golden hue. Overhead, dark clouds gather, foreshadowing a storm. The dramatic interplay of light and shadow imbues the entire scene with a heavy, resolute, and desolate atmosphere.", + "seed": 23254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_846_seed_23254_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2550742167979479, + "vlm_logprob": -0.2550742167979479, + "vlm_yes_no_margin": 1.774999988079071, + "vlm_token_credit": -0.2550742167979479, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10684261471033096 + ], + "opposite_logprob_tokens": [ + -2.356842517852783 + ], + "logprob_reward": -0.10684261471033096, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10684261471033096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11388026177883148, + 0.04621458426117897, + 0.08954896032810211, + 0.2506265938282013, + 0.10278484970331192, + 0.05188075080513954, + 0.09766742587089539, + 0.1000177264213562, + 0.07437179237604141, + 0.07300706207752228 + ], + "credit_map": [ + 0.011637318879365921, + 0.011857372708618641, + 0.020091600716114044, + 0.01937008462846279, + 0.013347678817808628, + 0.016766024753451347, + 0.011362044140696526, + 0.009448135271668434, + 0.005910911131650209, + 0.007946847938001156, + 0.005344685632735491, + 0.006682229693979025, + 0.0069249714724719524, + 0.005500449799001217, + 0.004481601994484663, + 0.0034228863660246134, + 0.009715946391224861, + 0.012292197905480862, + 0.009638153947889805, + 0.010863786563277245, + 0.010521347634494305, + 0.011538006365299225, + 0.015823878347873688, + 0.009155641309916973, + 0.02747698500752449, + 0.02214496023952961, + 0.026254044845700264, + 0.026750389486551285, + 0.04487612843513489, + 0.025554178282618523, + 0.034690339118242264, + 0.04287956655025482, + 0.011163442395627499, + 0.011489996686577797, + 0.010389399714767933, + 0.010889534838497639, + 0.012963475659489632, + 0.0176159106194973, + 0.011876204051077366, + 0.01639687642455101, + 0.005592151079326868, + 0.007037293165922165, + 0.005179125349968672, + 0.0040865372866392136, + 0.0081256702542305, + 0.008402508683502674, + 0.006232522428035736, + 0.0072249467484653, + 0.007407242897897959, + 0.006547839380800724, + 0.013058997690677643, + 0.00569369038566947, + 0.02125885896384716, + 0.01816604472696781, + 0.008183974772691727, + 0.01735076867043972, + 0.008128116838634014, + 0.007617517840117216, + 0.01678839884698391, + 0.01038400735706091, + 0.02020408771932125, + 0.0166380163282156, + 0.010110270231962204, + 0.010147311724722385, + 0.007423246745020151, + 0.007449890952557325, + 0.012894256971776485, + 0.00704526249319315, + 0.015269801951944828, + 0.0095292953774333, + 0.007836472243070602, + 0.006923562381416559, + 0.005447497591376305, + 0.006399990990757942, + 0.012768944725394249, + 0.00703981751576066, + 0.01465456373989582, + 0.009660659357905388, + 0.010216460563242435, + 0.006819127127528191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.10684261471033096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576180636882782 + ], + "opposite_logprob_tokens": [ + -0.826180636882782 + ], + "logprob_reward": -0.576180636882782, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576180636882782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049603257328271866, + 0.053604789078235626, + 0.040263719856739044, + 0.18377982079982758, + 0.07981414347887039, + 0.10710052400827408, + 0.19573278725147247, + 0.07724989950656891, + 0.1747254580259323, + 0.0381256565451622 + ], + "credit_map": [ + 0.004980685189366341, + 0.004346190486103296, + 0.007635687943547964, + 0.00728705944493413, + 0.006626774091273546, + 0.007201198022812605, + 0.00554527435451746, + 0.005980384070426226, + 0.005641746334731579, + 0.010234986431896687, + 0.007077022455632687, + 0.009112918749451637, + 0.005437795538455248, + 0.006742001511156559, + 0.0048006330616772175, + 0.0045576817356050014, + 0.002820835215970874, + 0.004763805773109198, + 0.003420416731387377, + 0.004317558370530605, + 0.0068101948127150536, + 0.006019366439431906, + 0.008367015980184078, + 0.0037445304915308952, + 0.01632123999297619, + 0.013762249611318111, + 0.023027528077363968, + 0.02195170894265175, + 0.032698821276426315, + 0.018176045268774033, + 0.028968380764126778, + 0.02887384034693241, + 0.006020646542310715, + 0.008326097391545773, + 0.006917412392795086, + 0.008556507527828217, + 0.01246473379433155, + 0.014716777950525284, + 0.009480873122811317, + 0.013331092894077301, + 0.009737161919474602, + 0.014002732932567596, + 0.013783207163214684, + 0.01194376964122057, + 0.010800282470881939, + 0.018221670761704445, + 0.014678207226097584, + 0.013933487236499786, + 0.019681306555867195, + 0.015203738585114479, + 0.033250100910663605, + 0.012960732914507389, + 0.030018175020813942, + 0.04901326820254326, + 0.01480073295533657, + 0.020804721862077713, + 0.008533963933587074, + 0.007431536912918091, + 0.009199312888085842, + 0.009323921985924244, + 0.011862017214298248, + 0.01393430121243, + 0.00745543884113431, + 0.009509408846497536, + 0.01557998452335596, + 0.017854822799563408, + 0.03952408581972122, + 0.019833626225590706, + 0.027131300419569016, + 0.01886974833905697, + 0.02136407606303692, + 0.014567800797522068, + 0.003347956109791994, + 0.004389454610645771, + 0.007288026623427868, + 0.004537152126431465, + 0.005731396377086639, + 0.004607957322150469, + 0.0042305076494812965, + 0.003993204329162836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.576180636882782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282546937465668 + ], + "opposite_logprob_tokens": [ + -2.0178253650665283 + ], + "logprob_reward": -0.14282546937465668, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14282546937465668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15534928441047668, + 0.06340203434228897, + 0.05195832625031471, + 0.17284324765205383, + 0.13862524926662445, + 0.09531988203525543, + 0.08622745424509048, + 0.06796538084745407, + 0.07184027135372162, + 0.09646884351968765 + ], + "credit_map": [ + 0.015263800509274006, + 0.017568795010447502, + 0.029627608135342598, + 0.027171647176146507, + 0.01378680020570755, + 0.022644009441137314, + 0.014801114797592163, + 0.014485502615571022, + 0.005847644060850143, + 0.013756310567259789, + 0.006837815977632999, + 0.011368362233042717, + 0.006550367455929518, + 0.008587323129177094, + 0.005578221753239632, + 0.00487598916515708, + 0.005219015292823315, + 0.0049736895598471165, + 0.00410564336925745, + 0.005307725630700588, + 0.007265143096446991, + 0.007268564309924841, + 0.010487375780940056, + 0.007331166882067919, + 0.018094133585691452, + 0.015435499139130116, + 0.018899038434028625, + 0.025283852592110634, + 0.03188440203666687, + 0.018330493941903114, + 0.02290617860853672, + 0.02200964279472828, + 0.018648093566298485, + 0.01454512681812048, + 0.012710788287222385, + 0.013656605035066605, + 0.01816960610449314, + 0.025278838351368904, + 0.012113554403185844, + 0.023502646014094353, + 0.014594083651900291, + 0.014959747903048992, + 0.012266771867871284, + 0.009384525008499622, + 0.011765995994210243, + 0.01316201500594616, + 0.009871878661215305, + 0.009314868599176407, + 0.01105549931526184, + 0.008137608878314495, + 0.011424797587096691, + 0.004659310914576054, + 0.013042107224464417, + 0.012687569484114647, + 0.006947964429855347, + 0.01827259734272957, + 0.010194403119385242, + 0.004951481241732836, + 0.009660202078521252, + 0.005849598906934261, + 0.013602770864963531, + 0.01191919669508934, + 0.004544222727417946, + 0.007243509404361248, + 0.008914614096283913, + 0.0062429350800812244, + 0.011218182742595673, + 0.006943820044398308, + 0.01522369310259819, + 0.008984208106994629, + 0.007978181354701519, + 0.006334643345326185, + 0.009647400118410587, + 0.010603496804833412, + 0.01695738174021244, + 0.008956545032560825, + 0.013230695389211178, + 0.013770218938589096, + 0.012189144268631935, + 0.011113964952528477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.14282546937465668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289371371269226 + ], + "opposite_logprob_tokens": [ + -1.0539371967315674 + ], + "logprob_reward": -0.4289371371269226, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4289371371269226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1667531579732895, + 0.08732668310403824, + 0.11394083499908447, + 0.09619085490703583, + 0.08036015182733536, + 0.04879819601774216, + 0.08698255568742752, + 0.04203103110194206, + 0.03710178658366203, + 0.24051474034786224 + ], + "credit_map": [ + 0.016986418515443802, + 0.01649104803800583, + 0.02672746405005455, + 0.031279392540454865, + 0.016309374943375587, + 0.0243512112647295, + 0.017457973212003708, + 0.01715027540922165, + 0.008472239598631859, + 0.018836405128240585, + 0.010325456038117409, + 0.01620081625878811, + 0.008128861896693707, + 0.012386462651193142, + 0.007085448130965233, + 0.005890996195375919, + 0.010037084110081196, + 0.008752275258302689, + 0.011093409731984138, + 0.012747813016176224, + 0.011714172549545765, + 0.015881642699241638, + 0.025861328467726707, + 0.017853114753961563, + 0.010382849723100662, + 0.008012566715478897, + 0.011127963662147522, + 0.012528783641755581, + 0.020938947796821594, + 0.011017419397830963, + 0.013697054237127304, + 0.008485260419547558, + 0.008799226023256779, + 0.009271502494812012, + 0.007544301450252533, + 0.011057037860155106, + 0.01371031068265438, + 0.011143216863274574, + 0.010992542840540409, + 0.007842015475034714, + 0.004129003267735243, + 0.004709105007350445, + 0.006822744384407997, + 0.006905063521116972, + 0.005800250452011824, + 0.007458285894244909, + 0.0061813825741410255, + 0.006792357657104731, + 0.0088497968390584, + 0.006779346615076065, + 0.016589131206274033, + 0.0052598160691559315, + 0.014231822453439236, + 0.017960749566555023, + 0.005650491453707218, + 0.011661396361887455, + 0.0037737353704869747, + 0.0045370664447546005, + 0.0066000125370919704, + 0.0033009860198944807, + 0.010307750664651394, + 0.005845820996910334, + 0.0031902713235467672, + 0.004475388675928116, + 0.0042631300166249275, + 0.0048957099206745625, + 0.005465388298034668, + 0.0038287625648081303, + 0.006299453787505627, + 0.0042862556874752045, + 0.00470573827624321, + 0.0033573471009731293, + 0.026271339505910873, + 0.027340145781636238, + 0.04470719024538994, + 0.025823751464486122, + 0.03743770718574524, + 0.0303829163312912, + 0.027810899540781975, + 0.020740780979394913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.4289371371269226 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585225895047188 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.020585225895047188, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585225895047188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1432497352361679, + 0.06610651314258575, + 0.08328013122081757, + 0.1928534209728241, + 0.06844338774681091, + 0.0922597199678421, + 0.14798936247825623, + 0.10876382887363434, + 0.025421328842639923, + 0.07163266092538834 + ], + "credit_map": [ + 0.013870364986360073, + 0.012408290058374405, + 0.014909574761986732, + 0.015153036452829838, + 0.027958692982792854, + 0.021768344566226006, + 0.01848601922392845, + 0.018695395439863205, + 0.0065887318924069405, + 0.008133316412568092, + 0.006546928081661463, + 0.010039412416517735, + 0.010558819398283958, + 0.009561605751514435, + 0.008482127450406551, + 0.006195561029016972, + 0.008480210788547993, + 0.0074959103949368, + 0.005929941311478615, + 0.009807290509343147, + 0.011431047692894936, + 0.012195880524814129, + 0.016164595261216164, + 0.01177524495869875, + 0.023801762610673904, + 0.015709688887000084, + 0.021696550771594048, + 0.024348529055714607, + 0.036710310727357864, + 0.01999935321509838, + 0.024652892723679543, + 0.025934327393770218, + 0.011521887965500355, + 0.0070942179299890995, + 0.005474034696817398, + 0.007173503749072552, + 0.012573257088661194, + 0.008704462088644505, + 0.009446708485484123, + 0.00645530316978693, + 0.01695103384554386, + 0.014124763198196888, + 0.0096044996753335, + 0.006797850131988525, + 0.015985766425728798, + 0.010935243219137192, + 0.007010176777839661, + 0.010850377380847931, + 0.015681659802794456, + 0.010930081829428673, + 0.023949114605784416, + 0.00983076449483633, + 0.029191015288233757, + 0.02711009792983532, + 0.012380796484649181, + 0.018915819004178047, + 0.010264944285154343, + 0.008573491126298904, + 0.017574865370988846, + 0.010442167520523071, + 0.017793618142604828, + 0.023965010419487953, + 0.006291453260928392, + 0.013858272694051266, + 0.002873346908017993, + 0.002623542444780469, + 0.004679376725107431, + 0.0027580871246755123, + 0.003958544693887234, + 0.0029816606547683477, + 0.0028908306267112494, + 0.002655936870723963, + 0.0052322689443826675, + 0.008696201257407665, + 0.010258953087031841, + 0.007157652173191309, + 0.016132734715938568, + 0.010313187725841999, + 0.006749969441443682, + 0.007091694977134466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.020585225895047188 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful low-angle tracking shot follows a heavy steel freight train, its charcoal-black front covered in riveted steel and mottled with rust, as it hurtles forward with unstoppable momentum. The lone round headlight casts a glaring beam through the gathering twilight, illuminating the abandoned, overgrown tracks ahead. The massive steel wheels relentlessly grind against the rusted track, emitting a dull metallic thud, shaking the splintered railroad ties and the surrounding gravel roadbed slightly and violently pushing the overgrown weeds to the sides. The low-slanting sunset casts long, twisting shadows across the barren landscape, gilding the weathered sides of the train with a fleeting warm golden hue. Overhead, dark clouds gather, foreshadowing a storm. The dramatic interplay of light and shadow imbues the entire scene with a heavy, resolute, and desolate atmosphere.", + "seed": 23255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_6/index_846_seed_23255_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1579155908897519, + "vlm_logprob": -0.1579155908897519, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.1579155908897519, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07961999624967575 + ], + "opposite_logprob_tokens": [ + -2.704619884490967 + ], + "logprob_reward": -0.07961999624967575, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07961999624967575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06525937467813492, + 0.024115033447742462, + 0.008680629543960094, + 0.19663800299167633, + 0.40780889987945557, + 0.07813315093517303, + 0.04322921112179756, + 0.10014765709638596, + 0.03727683052420616, + 0.038711193948984146 + ], + "credit_map": [ + 0.006825328804552555, + 0.007838349789381027, + 0.005111840087920427, + 0.009956375695765018, + 0.007885602302849293, + 0.00896016601473093, + 0.007823789492249489, + 0.01085792575031519, + 0.0029492133762687445, + 0.003588159568607807, + 0.0022750364150851965, + 0.004362720064818859, + 0.0033312635496258736, + 0.0021675850730389357, + 0.0031246887519955635, + 0.0023163696750998497, + 0.000990005792118609, + 0.0010805358178913593, + 0.000840288819745183, + 0.001573187648318708, + 0.001160053419880569, + 0.0009448667988181114, + 0.0009961487958207726, + 0.001095543266274035, + 0.045444026589393616, + 0.007546392735093832, + 0.00711715966463089, + 0.015522667206823826, + 0.012781141325831413, + 0.06911788135766983, + 0.02800159901380539, + 0.011107155121862888, + 0.024632517248392105, + 0.044030994176864624, + 0.028417790308594704, + 0.06667304784059525, + 0.054608363658189774, + 0.0644451379776001, + 0.05756755918264389, + 0.06743351370096207, + 0.004195943474769592, + 0.006937368772923946, + 0.007055792957544327, + 0.00896994024515152, + 0.013499565422534943, + 0.013996653258800507, + 0.013183350674808025, + 0.010294534265995026, + 0.004105149302631617, + 0.0021159399766474962, + 0.0029050814919173717, + 0.004108464810997248, + 0.006770419888198376, + 0.007856149226427078, + 0.004837915766984224, + 0.010530093684792519, + 0.007513930089771748, + 0.008081842213869095, + 0.007162377703934908, + 0.01248469389975071, + 0.01678977906703949, + 0.015244298614561558, + 0.010709133930504322, + 0.02216160111129284, + 0.003135278122499585, + 0.0053334604017436504, + 0.0020669803488999605, + 0.004547561518847942, + 0.003906456753611565, + 0.005380035378038883, + 0.005799734033644199, + 0.007107328157871962, + 0.003921667579561472, + 0.0062790922820568085, + 0.0029947482980787754, + 0.006643794476985931, + 0.0034649961162358522, + 0.0043974146246910095, + 0.004764437675476074, + 0.006245047319680452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.07961999624967575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328422427177429 + ], + "opposite_logprob_tokens": [ + -0.7578422427177429 + ], + "logprob_reward": -0.6328422427177429, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328422427177429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10350288450717926, + 0.026750674471259117, + 0.03415960818529129, + 0.3913956880569458, + 0.13024404644966125, + 0.03872724995017052, + 0.017335418611764908, + 0.1639692783355713, + 0.040029820054769516, + 0.05388539284467697 + ], + "credit_map": [ + 0.009233944118022919, + 0.014728305861353874, + 0.009016294963657856, + 0.025088144466280937, + 0.008272897452116013, + 0.012040826492011547, + 0.01314057968556881, + 0.011981888674199581, + 0.0031920168548822403, + 0.0036819244269281626, + 0.0028947300743311644, + 0.004888082388788462, + 0.002619198290631175, + 0.0027085868641734123, + 0.0032445210963487625, + 0.0035216123796999454, + 0.003796396544203162, + 0.004891924560070038, + 0.0024535413831472397, + 0.0072733936831355095, + 0.0042301504872739315, + 0.004049633629620075, + 0.004232341423630714, + 0.003232228336855769, + 0.0890614241361618, + 0.015249209478497505, + 0.010569757781922817, + 0.03007967211306095, + 0.02403012290596962, + 0.13801394402980804, + 0.05989569425582886, + 0.024495847523212433, + 0.012213705107569695, + 0.009484095498919487, + 0.010051336139440536, + 0.01700109802186489, + 0.01660793647170067, + 0.01699475757777691, + 0.022344259545207024, + 0.025546859949827194, + 0.0025456782896071672, + 0.002442492637783289, + 0.0023791915737092495, + 0.00503110745921731, + 0.007427228149026632, + 0.00637945719063282, + 0.006787047255784273, + 0.005735044367611408, + 0.0015426224563270807, + 0.0009843098232522607, + 0.0019895138684660196, + 0.0018063305178657174, + 0.0032501397654414177, + 0.0026292181573808193, + 0.0021747262217104435, + 0.0029585587326437235, + 0.012228958308696747, + 0.013560432009398937, + 0.012987381778657436, + 0.0220614206045866, + 0.030158046633005142, + 0.028707420453429222, + 0.01912025548517704, + 0.025145364925265312, + 0.0036614483688026667, + 0.00734208058565855, + 0.0026275739073753357, + 0.0055513228289783, + 0.004638205748051405, + 0.0048101856373250484, + 0.00449535483494401, + 0.006903648842126131, + 0.006129146553575993, + 0.012666658498346806, + 0.004122389946132898, + 0.007284750230610371, + 0.005753001198172569, + 0.004961194936186075, + 0.006973545998334885, + 0.005994703155010939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.6328422427177429 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979887880384922 + ], + "opposite_logprob_tokens": [ + -4.634799003601074 + ], + "logprob_reward": -0.00979887880384922, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979887880384922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26727694272994995, + 0.08093936741352081, + 0.023342479020357132, + 0.08263693749904633, + 0.2769912779331207, + 0.055401720106601715, + 0.039262983947992325, + 0.035852961242198944, + 0.05651458725333214, + 0.08178078383207321 + ], + "credit_map": [ + 0.020794706419110298, + 0.03755248337984085, + 0.025765204802155495, + 0.066138356924057, + 0.022341445088386536, + 0.030135104432702065, + 0.03266449645161629, + 0.03188515082001686, + 0.011002126149833202, + 0.012009824626147747, + 0.005781239364296198, + 0.01389458030462265, + 0.011743564158678055, + 0.010052776895463467, + 0.010677887126803398, + 0.005777365993708372, + 0.002722316887229681, + 0.0036247032694518566, + 0.001846725819632411, + 0.004176461137831211, + 0.002691937843337655, + 0.0027507964987307787, + 0.002661456586793065, + 0.0028680807445198298, + 0.018486404791474342, + 0.0057452707551419735, + 0.004891721997410059, + 0.007098622154444456, + 0.007764844223856926, + 0.02626555785536766, + 0.0065582855604588985, + 0.005826230626553297, + 0.01739266887307167, + 0.026059867814183235, + 0.019549250602722168, + 0.04559403285384178, + 0.03544493019580841, + 0.04258423298597336, + 0.03787030279636383, + 0.05249597877264023, + 0.0033852756023406982, + 0.004242242313921452, + 0.004183996468782425, + 0.00678755808621645, + 0.010340890847146511, + 0.00865020789206028, + 0.009087826125323772, + 0.008723723702132702, + 0.003650807775557041, + 0.0033366845455020666, + 0.0042027924209833145, + 0.004194873850792646, + 0.00678933784365654, + 0.006662366911768913, + 0.004481379874050617, + 0.005944738630205393, + 0.002701054560020566, + 0.0036030772607773542, + 0.0029570823535323143, + 0.004973448812961578, + 0.005366894416511059, + 0.005654351320117712, + 0.004097505938261747, + 0.0064995442517101765, + 0.00527911726385355, + 0.009408622980117798, + 0.003409133292734623, + 0.007231228053569794, + 0.005792583338916302, + 0.00691297696903348, + 0.007968952879309654, + 0.010511975735425949, + 0.007853363640606403, + 0.012134874239563942, + 0.006429871544241905, + 0.017912911251187325, + 0.005844983272254467, + 0.006933967582881451, + 0.013424107804894447, + 0.011246699839830399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.00979887880384922 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620616190135479 + ], + "opposite_logprob_tokens": [ + -2.8120615482330322 + ], + "logprob_reward": -0.0620616190135479, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620616190135479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.263685405254364, + 0.04843825101852417, + 0.044636670500040054, + 0.14180327951908112, + 0.22082418203353882, + 0.05557309091091156, + 0.035857904702425, + 0.03805755078792572, + 0.0553087517619133, + 0.09581483900547028 + ], + "credit_map": [ + 0.024097200483083725, + 0.019913338124752045, + 0.02156083472073078, + 0.03994937613606453, + 0.026399865746498108, + 0.03613058477640152, + 0.034809861332178116, + 0.06082436814904213, + 0.006339742802083492, + 0.005651114974170923, + 0.0039935181848704815, + 0.008843342773616314, + 0.004959040321409702, + 0.0054813106544315815, + 0.007980676367878914, + 0.005189512390643358, + 0.004844334442168474, + 0.005961315706372261, + 0.0037060317117720842, + 0.008808345533907413, + 0.006857705768197775, + 0.00491527933627367, + 0.0056712995283305645, + 0.0038723587058484554, + 0.03070860542356968, + 0.006095560733228922, + 0.005606424994766712, + 0.012427521869540215, + 0.010250923223793507, + 0.04653748869895935, + 0.02111072652041912, + 0.009066027589142323, + 0.014186995103955269, + 0.023897549137473106, + 0.01455787941813469, + 0.03987007215619087, + 0.02921600639820099, + 0.029638366773724556, + 0.030925532802939415, + 0.038531795144081116, + 0.0034301942214369774, + 0.004178416449576616, + 0.005479080136865377, + 0.007969134487211704, + 0.006667761132121086, + 0.008379878476262093, + 0.010019516572356224, + 0.009449114091694355, + 0.0032237162813544273, + 0.0023627765476703644, + 0.0035101522225886583, + 0.00360603048466146, + 0.004790154285728931, + 0.0057207378558814526, + 0.004397697281092405, + 0.008246644400060177, + 0.002998419338837266, + 0.0035747650545090437, + 0.002632624702528119, + 0.005319472402334213, + 0.006693228147923946, + 0.00651558069512248, + 0.004391044843941927, + 0.005932419560849667, + 0.004620129708200693, + 0.010209374129772186, + 0.004538179840892553, + 0.006211065221577883, + 0.005866410676389933, + 0.00714016892015934, + 0.006270421668887138, + 0.010452999733388424, + 0.007179250940680504, + 0.010096794925630093, + 0.007446467876434326, + 0.0267795380204916, + 0.010329659096896648, + 0.0062578474171459675, + 0.01502661406993866, + 0.012698673643171787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0620616190135479 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy steel freight train with a charcoal-black front covered in riveted steel and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving forward with high speed and unstoppable momentum?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone round headlight visible casting a glaring beam through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on abandoned, overgrown tracks with splintered railroad ties and gravel roadbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low angle with a low-slanting sunset casting long, twisting shadows and a warm golden hue on the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052552176639437675 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.0052552176639437675, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052552176639437675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11673682183027267, + 0.037927016615867615, + 0.050091348588466644, + 0.22877483069896698, + 0.2583285868167877, + 0.04858558252453804, + 0.028155343607068062, + 0.03824639692902565, + 0.02946729212999344, + 0.1636868119239807 + ], + "credit_map": [ + 0.013655601069331169, + 0.016242926940321922, + 0.010421695187687874, + 0.023244548588991165, + 0.010266966186463833, + 0.01628195121884346, + 0.013838340528309345, + 0.012784785591065884, + 0.0071639930829405785, + 0.004172370303422213, + 0.003151807002723217, + 0.004919897299259901, + 0.004826910328119993, + 0.004371011629700661, + 0.005303341895341873, + 0.0040176790207624435, + 0.00586724653840065, + 0.006858367007225752, + 0.0038142232224345207, + 0.009487579576671124, + 0.005953670013695955, + 0.006968429312109947, + 0.00560466293245554, + 0.005537167191505432, + 0.051725756376981735, + 0.007675538305193186, + 0.006938092410564423, + 0.01538663450628519, + 0.012332268059253693, + 0.07994813472032547, + 0.04000338539481163, + 0.014764996245503426, + 0.016713431105017662, + 0.02572050131857395, + 0.019523056223988533, + 0.034395795315504074, + 0.034928251057863235, + 0.04286671429872513, + 0.03900851309299469, + 0.045172303915023804, + 0.004141210112720728, + 0.004565258510410786, + 0.0036663904320448637, + 0.006295781582593918, + 0.006672389805316925, + 0.006431968417018652, + 0.010715984739363194, + 0.006096599157899618, + 0.0020432709716260433, + 0.0022107160184532404, + 0.002447487087920308, + 0.003434019396081567, + 0.00440241489559412, + 0.00531961303204298, + 0.0031871541868895292, + 0.005110668484121561, + 0.003554545110091567, + 0.004509957041591406, + 0.0061705647967755795, + 0.004243892151862383, + 0.006244057789444923, + 0.005638212896883488, + 0.0037284137215465307, + 0.004156751092523336, + 0.003352144733071327, + 0.004583393689244986, + 0.0029588702600449324, + 0.00395572604611516, + 0.003221635241061449, + 0.0037366636097431183, + 0.0031059104949235916, + 0.004552945960313082, + 0.013560086488723755, + 0.02499307692050934, + 0.014701662585139275, + 0.03797142952680588, + 0.014259058982133865, + 0.012778230011463165, + 0.025430869311094284, + 0.019992398098111153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.0052552176639437675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_tensor([9633, 9633])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_tensor([9633, 9633])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6832d65fd7ffb83630ea64d5bc94c187f210fe8d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_tensor([9633, 9633])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, sweeping camera shot arcs around a fearsome warrior as he meticulously scales a sheer, treacherous cliff face. Rendered in a detailed, heroic fantasy illustration reminiscent of classic Dungeons & Dragons art, the scene is a study in tension and determination. The warrior, a man of heroic stature, has long, black braids blown by the high-altitude wind. His face is a mask of intense concentration, his jaw set, his eyes fixed on his next handhold. His muscular body is clad in worn dark leather armor and iron-studded pauldrons that have seen many battles. Strapped to his back is a massive, double-headed battle axe, its polished steel edge gleaming in the light, its heavy wooden handle intricately carved with patterns. His gloved hands dig into the jagged, unforgiving rock, his body moving with a steady and deliberate rhythm that demonstrates his immense strength. As the camera continues to sweep, it captures the dramatic interplay of shadow and light cast by the low-angle sun, highlighting the rough texture of the granite and the climber's strained muscles. The camera zooms out slightly to reveal the abyss below, a vertiginous chasm filled with swirling mist that obscures the bottom, creating a powerful and heroic atmosphere of a lonely struggle against an epic, unforgiving landscape.", + "seed": 23391, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23391_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021286090649664403, + "vlm_logprob": -0.021286090649664403, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.021286090649664403, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011736392974853516 + ], + "opposite_logprob_tokens": [ + -4.8867363929748535 + ], + "logprob_reward": -0.011736392974853516, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011736392974853516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12489370256662369, + 0.37147751450538635, + 0.06782716512680054, + 0.0652252733707428, + 0.09666859358549118, + 0.044721975922584534, + 0.06826023757457733, + 0.033684588968753815, + 0.06182354688644409, + 0.0654173269867897 + ], + "credit_map": [ + 0.011546117253601551, + 0.015963494777679443, + 0.011976262554526329, + 0.021062549203634262, + 0.00878457073122263, + 0.016855698078870773, + 0.014958557672798634, + 0.023746466264128685, + 0.01734369434416294, + 0.012014961801469326, + 0.019673924893140793, + 0.020696258172392845, + 0.008840436115860939, + 0.14802983403205872, + 0.010286682285368443, + 0.134591743350029, + 0.008394734002649784, + 0.007500867825001478, + 0.008479214273393154, + 0.0077817500568926334, + 0.008161232806742191, + 0.0073351929895579815, + 0.00914519652724266, + 0.011028977110981941, + 0.010161319747567177, + 0.007222979329526424, + 0.008339437656104565, + 0.00559264374896884, + 0.007016059011220932, + 0.007149231620132923, + 0.009339412674307823, + 0.010404187254607677, + 0.02211865410208702, + 0.011370851658284664, + 0.02382754534482956, + 0.007719116751104593, + 0.007563532330095768, + 0.010671006515622139, + 0.005598811432719231, + 0.007799078244715929, + 0.007299443706870079, + 0.007067757658660412, + 0.007205409463495016, + 0.004183284938335419, + 0.004772967658936977, + 0.003882854711264372, + 0.004877988714724779, + 0.005432270932942629, + 0.008954238146543503, + 0.006236731540411711, + 0.013416184112429619, + 0.00791112333536148, + 0.007154833525419235, + 0.00837001297622919, + 0.0076378099620342255, + 0.008579310029745102, + 0.005307858344167471, + 0.00402535917237401, + 0.004145653452724218, + 0.0035321437753736973, + 0.003997716121375561, + 0.0033067320473492146, + 0.005226829554885626, + 0.004142296500504017, + 0.010977251455187798, + 0.004477722104638815, + 0.007147538010030985, + 0.01028232742100954, + 0.005969776771962643, + 0.008517096750438213, + 0.007522918283939362, + 0.006928921677172184, + 0.009469744749367237, + 0.004522974602878094, + 0.009920059703290462, + 0.006290198769420385, + 0.008759189397096634, + 0.009531009942293167, + 0.008886962197721004, + 0.008037184365093708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.011736392974853516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076110303401947 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016076110303401947, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076110303401947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1049867793917656, + 0.23367206752300262, + 0.0903749167919159, + 0.1270150989294052, + 0.08962349593639374, + 0.038724664598703384, + 0.09962498396635056, + 0.05263888090848923, + 0.136485293507576, + 0.026853859424591064 + ], + "credit_map": [ + 0.010067942552268505, + 0.010146256536245346, + 0.010057217441499233, + 0.01901109516620636, + 0.006521420553326607, + 0.0170572642236948, + 0.01782165840268135, + 0.014303931035101414, + 0.010438403114676476, + 0.0076265581883490086, + 0.013901659287512302, + 0.012766199186444283, + 0.005563326645642519, + 0.0944768637418747, + 0.007830630987882614, + 0.08106840401887894, + 0.010273227468132973, + 0.004635828081518412, + 0.011281353421509266, + 0.00954804290086031, + 0.005464308895170689, + 0.026957359164953232, + 0.010697671212255955, + 0.011517125181853771, + 0.016053080558776855, + 0.011130533181130886, + 0.011136802844703197, + 0.013663867488503456, + 0.019308587536215782, + 0.017884133383631706, + 0.019336653873324394, + 0.018501440063118935, + 0.014053295366466045, + 0.011296723037958145, + 0.014445848762989044, + 0.009917868301272392, + 0.007306510116904974, + 0.012575598433613777, + 0.009135052561759949, + 0.010892603546380997, + 0.004537372384220362, + 0.006287397816777229, + 0.004756160546094179, + 0.004381898324936628, + 0.005100287962704897, + 0.003912132233381271, + 0.004854549188166857, + 0.0048948670737445354, + 0.011524194851517677, + 0.008575765416026115, + 0.019706806167960167, + 0.012603006325662136, + 0.010379898361861706, + 0.01046952698379755, + 0.012204624712467194, + 0.014161165803670883, + 0.00637155631557107, + 0.004868177231401205, + 0.005761501844972372, + 0.006113178096711636, + 0.005818595644086599, + 0.005412050988525152, + 0.01018696092069149, + 0.008106858469545841, + 0.012290211394429207, + 0.013092394918203354, + 0.009477181360125542, + 0.017692765220999718, + 0.028572294861078262, + 0.01969730295240879, + 0.01870759017765522, + 0.016955548897385597, + 0.0030300470534712076, + 0.0020487289875745773, + 0.0040459115989506245, + 0.0049278875812888145, + 0.0023538703098893166, + 0.0033675977028906345, + 0.0039058395195752382, + 0.0031739783007651567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.016076110303401947 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809374198317528 + ], + "opposite_logprob_tokens": [ + -3.2880938053131104 + ], + "logprob_reward": -0.03809374198317528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809374198317528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09380141645669937, + 0.26755934953689575, + 0.047348503023386, + 0.2497345358133316, + 0.04296965152025223, + 0.04015734791755676, + 0.057941701263189316, + 0.10317426174879074, + 0.05820593237876892, + 0.039107296615839005 + ], + "credit_map": [ + 0.009424840100109577, + 0.01313700806349516, + 0.011008125729858875, + 0.016648266464471817, + 0.007388023193925619, + 0.013188319280743599, + 0.0108198756352067, + 0.012186951003968716, + 0.012730536051094532, + 0.008982514962553978, + 0.01450866088271141, + 0.01666727103292942, + 0.007277808152139187, + 0.10153202712535858, + 0.010715283453464508, + 0.09514525532722473, + 0.005656465422362089, + 0.004315964877605438, + 0.006477372255176306, + 0.005824742838740349, + 0.0040686288848519325, + 0.009998280555009842, + 0.004961569793522358, + 0.0060454802587628365, + 0.0359332375228405, + 0.031025560572743416, + 0.03564174845814705, + 0.03370143100619316, + 0.02068692445755005, + 0.020497597754001617, + 0.029918640851974487, + 0.042329393327236176, + 0.007524642162024975, + 0.004983119200915098, + 0.006554438732564449, + 0.005668964236974716, + 0.0035597565583884716, + 0.004876237828284502, + 0.0047384463250637054, + 0.005064044147729874, + 0.0057290708646178246, + 0.005531159229576588, + 0.0052492134273052216, + 0.005326943006366491, + 0.00503286300227046, + 0.003719016909599304, + 0.0047874958254396915, + 0.004781586118042469, + 0.007576778996735811, + 0.004977590404450893, + 0.013833974488079548, + 0.005845615640282631, + 0.0062754894606769085, + 0.006134212017059326, + 0.006307878065854311, + 0.006990162190049887, + 0.018262209370732307, + 0.009836175478994846, + 0.01766677014529705, + 0.011067011393606663, + 0.012485429644584656, + 0.00976371020078659, + 0.011684838682413101, + 0.012408117763698101, + 0.0057932850904762745, + 0.004888759460300207, + 0.0034388352651149035, + 0.013012515380978584, + 0.008661008439958096, + 0.007930061779916286, + 0.007814685814082623, + 0.006666786037385464, + 0.005780650302767754, + 0.003277689218521118, + 0.006142999045550823, + 0.00445074075832963, + 0.003988584503531456, + 0.006435073912143707, + 0.004693529102951288, + 0.004338029772043228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.03809374198317528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633737027645111 + ], + "opposite_logprob_tokens": [ + -3.6513373851776123 + ], + "logprob_reward": -0.02633737027645111, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633737027645111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0712471753358841, + 0.420758455991745, + 0.03422519564628601, + 0.02618269808590412, + 0.11810533702373505, + 0.05131031945347786, + 0.16890104115009308, + 0.02886217087507248, + 0.06677237153053284, + 0.01363526564091444 + ], + "credit_map": [ + 0.007343876175582409, + 0.011778132990002632, + 0.0063744718208909035, + 0.011664699763059616, + 0.0045826612040400505, + 0.01158012542873621, + 0.008583301678299904, + 0.009339908137917519, + 0.019740846008062363, + 0.012568349950015545, + 0.022968845441937447, + 0.022831860929727554, + 0.009984090924263, + 0.1683870255947113, + 0.012252489104866982, + 0.15202490985393524, + 0.005336061120033264, + 0.0030595266725867987, + 0.003148823045194149, + 0.0034093852154910564, + 0.0030704799573868513, + 0.004884663503617048, + 0.00446699233725667, + 0.00684926426038146, + 0.0025742908474057913, + 0.00201943377032876, + 0.002250920282676816, + 0.002068008529022336, + 0.003993089776486158, + 0.005846097599714994, + 0.004500813782215118, + 0.002930041868239641, + 0.019353903830051422, + 0.014799121767282486, + 0.02278764359652996, + 0.012956326827406883, + 0.009199573658406734, + 0.014204435981810093, + 0.011787491850554943, + 0.013016832992434502, + 0.010451994836330414, + 0.007405445910990238, + 0.00883031077682972, + 0.0056860302574932575, + 0.005576810333877802, + 0.0034497182350605726, + 0.004971892107278109, + 0.0049381195567548275, + 0.018198132514953613, + 0.019760766997933388, + 0.028391951695084572, + 0.024301862344145775, + 0.02465507946908474, + 0.016189558431506157, + 0.0165821872651577, + 0.020821504294872284, + 0.003824278712272644, + 0.002657178323715925, + 0.003891113679856062, + 0.0030719805508852005, + 0.004836468491703272, + 0.00331909884698689, + 0.003776217345148325, + 0.003485835390165448, + 0.0052383216097950935, + 0.0063057211227715015, + 0.003893317421898246, + 0.011469638906419277, + 0.014797842130064964, + 0.007855379022657871, + 0.008631408214569092, + 0.008580741472542286, + 0.0019621467217803, + 0.0012720068916678429, + 0.001679866574704647, + 0.0012886648764833808, + 0.0010311946971341968, + 0.0030371739994734526, + 0.0014344488736242056, + 0.0019297630060464144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.02633737027645111 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186837710440159 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014186837710440159, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186837710440159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16222776472568512, + 0.2737789452075958, + 0.03663943335413933, + 0.06359300762414932, + 0.06226750835776329, + 0.0482923723757267, + 0.11247242242097855, + 0.12693490087985992, + 0.06894204020500183, + 0.04485158994793892 + ], + "credit_map": [ + 0.0242456067353487, + 0.018751289695501328, + 0.02411847561597824, + 0.022633105516433716, + 0.013999155722558498, + 0.02012905664741993, + 0.015168050304055214, + 0.023183029145002365, + 0.012984603643417358, + 0.009927339851856232, + 0.01638907752931118, + 0.01438919547945261, + 0.00790497288107872, + 0.10919451713562012, + 0.009374886751174927, + 0.09361434727907181, + 0.003774132113903761, + 0.0031581781804561615, + 0.004170052707195282, + 0.004404720850288868, + 0.0028974979650229216, + 0.008164465427398682, + 0.005817570723593235, + 0.004252814222127199, + 0.00595417246222496, + 0.004844778683036566, + 0.007989740930497646, + 0.010182349011301994, + 0.009153611026704311, + 0.009215934202075005, + 0.00915062427520752, + 0.007101801689714193, + 0.011584829539060593, + 0.00811079703271389, + 0.008947071619331837, + 0.009126116521656513, + 0.004632262513041496, + 0.00687051797285676, + 0.0063528139144182205, + 0.006643094588071108, + 0.005505353678017855, + 0.006443467456847429, + 0.006460242439061403, + 0.006919463165104389, + 0.0050384532660245895, + 0.004407082684338093, + 0.00701102614402771, + 0.006507286801934242, + 0.01613268069922924, + 0.010190817527472973, + 0.021348442882299423, + 0.013462010771036148, + 0.011955584399402142, + 0.01196777168661356, + 0.012676670216023922, + 0.014738447032868862, + 0.02343420870602131, + 0.015164639800786972, + 0.018037807196378708, + 0.013139182701706886, + 0.015117855742573738, + 0.014042096212506294, + 0.01272696815431118, + 0.01527214515954256, + 0.00928753986954689, + 0.005540394689887762, + 0.009829611517488956, + 0.006821535062044859, + 0.010496964678168297, + 0.00950036570429802, + 0.008173570968210697, + 0.009292054921388626, + 0.007643654942512512, + 0.004895922727882862, + 0.005513670854270458, + 0.0056665693409740925, + 0.004056410863995552, + 0.0055657727643847466, + 0.0061333575285971165, + 0.0053762346506118774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.014186837710440159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, sweeping camera shot arcs around a fearsome warrior as he meticulously scales a sheer, treacherous cliff face. Rendered in a detailed, heroic fantasy illustration reminiscent of classic Dungeons & Dragons art, the scene is a study in tension and determination. The warrior, a man of heroic stature, has long, black braids blown by the high-altitude wind. His face is a mask of intense concentration, his jaw set, his eyes fixed on his next handhold. His muscular body is clad in worn dark leather armor and iron-studded pauldrons that have seen many battles. Strapped to his back is a massive, double-headed battle axe, its polished steel edge gleaming in the light, its heavy wooden handle intricately carved with patterns. His gloved hands dig into the jagged, unforgiving rock, his body moving with a steady and deliberate rhythm that demonstrates his immense strength. As the camera continues to sweep, it captures the dramatic interplay of shadow and light cast by the low-angle sun, highlighting the rough texture of the granite and the climber's strained muscles. The camera zooms out slightly to reveal the abyss below, a vertiginous chasm filled with swirling mist that obscures the bottom, creating a powerful and heroic atmosphere of a lonely struggle against an epic, unforgiving landscape.", + "seed": 23392, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23392_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025130323879420758, + "vlm_logprob": -0.025130323879420758, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.025130323879420758, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011730972677469254 + ], + "opposite_logprob_tokens": [ + -4.636731147766113 + ], + "logprob_reward": -0.011730972677469254, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011730972677469254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.613004744052887, + 0.025039395317435265, + 0.02821795828640461, + 0.04068724438548088, + 0.02856404520571232, + 0.06440423429012299, + 0.024295177310705185, + 0.026765508577227592, + 0.08039560168981552, + 0.06862609833478928 + ], + "credit_map": [ + 0.02171584963798523, + 0.029723387211561203, + 0.040438439697027206, + 0.023772258311510086, + 0.24373377859592438, + 0.04890979453921318, + 0.06353585422039032, + 0.14117538928985596, + 0.0021158556919544935, + 0.002051105024293065, + 0.005008446518331766, + 0.0037572477012872696, + 0.0029556432273238897, + 0.0032725371420383453, + 0.003902348456904292, + 0.001976210391148925, + 0.003254414303228259, + 0.0028610308654606342, + 0.0028622797690331936, + 0.002885670866817236, + 0.004914429038763046, + 0.0026237706188112497, + 0.0058730244636535645, + 0.0029433395247906446, + 0.003211419330909848, + 0.0014921167166903615, + 0.010536868125200272, + 0.005031001288443804, + 0.002501358510926366, + 0.010223781690001488, + 0.0020546582527458668, + 0.005636037793010473, + 0.003228241577744484, + 0.003776717698201537, + 0.0038021968211978674, + 0.0033287564292550087, + 0.002904534572735429, + 0.0034489037934690714, + 0.0055896565318107605, + 0.002485036849975586, + 0.006596884690225124, + 0.0065071070566773415, + 0.011136976070702076, + 0.006532008294016123, + 0.005914183799177408, + 0.01094904262572527, + 0.007280364166945219, + 0.009487668983638287, + 0.0032866375986486673, + 0.0037860500160604715, + 0.003642770228907466, + 0.0034069158136844635, + 0.003594267414882779, + 0.002353833056986332, + 0.0026285359635949135, + 0.001596167916432023, + 0.002389661269262433, + 0.004802974406629801, + 0.0037755495868623257, + 0.003217377932742238, + 0.003175681922584772, + 0.002668440341949463, + 0.0030749610159546137, + 0.00366086233407259, + 0.014583289623260498, + 0.0076419152319431305, + 0.01280593778938055, + 0.012390044517815113, + 0.0036688896361738443, + 0.009726799093186855, + 0.006199742667376995, + 0.013378984294831753, + 0.008502247743308544, + 0.006255093961954117, + 0.016784489154815674, + 0.005913705565035343, + 0.004071488045156002, + 0.00874741654843092, + 0.006253769621253014, + 0.01209788303822279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.011730972677469254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085587553679943 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085587553679943, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085587553679943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5252657532691956, + 0.03207913041114807, + 0.02683044970035553, + 0.1071222722530365, + 0.04642214626073837, + 0.05507737025618553, + 0.05012381821870804, + 0.02003616653382778, + 0.04793684557080269, + 0.08910605311393738 + ], + "credit_map": [ + 0.018921159207820892, + 0.025461990386247635, + 0.023385968059301376, + 0.01878972537815571, + 0.20996075868606567, + 0.04290389642119408, + 0.050960391759872437, + 0.134881854057312, + 0.0023523883428424597, + 0.0028717194218188524, + 0.0034270822070538998, + 0.0051530939526855946, + 0.0054003517143428326, + 0.004117731470614672, + 0.004421693738549948, + 0.004335067700594664, + 0.003263851860538125, + 0.0021462023723870516, + 0.002309631323441863, + 0.0028295302763581276, + 0.00445124227553606, + 0.0036603088956326246, + 0.003326445585116744, + 0.004843234550207853, + 0.008217042312026024, + 0.0036885428708046675, + 0.027050623670220375, + 0.013448132202029228, + 0.004601156339049339, + 0.029896531254053116, + 0.004391415975987911, + 0.01582881063222885, + 0.005471914075314999, + 0.007396599277853966, + 0.005872501991689205, + 0.005695442669093609, + 0.007272852584719658, + 0.005197213031351566, + 0.00517634442076087, + 0.004339275881648064, + 0.005863131955265999, + 0.005962283816188574, + 0.0062338910065591335, + 0.006590103730559349, + 0.005888682324439287, + 0.011303182691335678, + 0.0074048093520104885, + 0.005831281654536724, + 0.005778160411864519, + 0.009622273966670036, + 0.008170921355485916, + 0.00795246846973896, + 0.004958910867571831, + 0.004231300670653582, + 0.005444677080959082, + 0.003965101204812527, + 0.0021694416645914316, + 0.0038002983201295137, + 0.003262232057750225, + 0.0021682779770344496, + 0.002335568191483617, + 0.0019390167435631156, + 0.0022989704739302397, + 0.002062358893454075, + 0.0075728981755673885, + 0.004750433377921581, + 0.005983349867165089, + 0.0071441479958593845, + 0.0028602946549654007, + 0.006879675202071667, + 0.004349291790276766, + 0.008396750316023827, + 0.011776833795011044, + 0.011422766372561455, + 0.013510040007531643, + 0.01304539106786251, + 0.005907054059207439, + 0.011376348324120045, + 0.009627807885408401, + 0.012439803220331669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.011085587553679943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495677888393402 + ], + "opposite_logprob_tokens": [ + -2.9299566745758057 + ], + "logprob_reward": -0.05495677888393402, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495677888393402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5576052069664001, + 0.024583909660577774, + 0.02198748104274273, + 0.04038350656628609, + 0.03073747642338276, + 0.042300090193748474, + 0.034856293350458145, + 0.0844537541270256, + 0.041331905871629715, + 0.12176038324832916 + ], + "credit_map": [ + 0.020646315068006516, + 0.029608504846692085, + 0.033110737800598145, + 0.01729034073650837, + 0.21740055084228516, + 0.04444427788257599, + 0.05423816666007042, + 0.14086630940437317, + 0.0017716269940137863, + 0.0025846457574516535, + 0.0024581667967140675, + 0.003690616926178336, + 0.003757848870009184, + 0.0032916658092290163, + 0.0039039403200149536, + 0.0031253991182893515, + 0.0022100561764091253, + 0.002263708272948861, + 0.002555225742980838, + 0.0018438516417518258, + 0.003960881847888231, + 0.002063591266050935, + 0.0040628365240991116, + 0.0030273303855210543, + 0.003143543377518654, + 0.002824964001774788, + 0.009806700050830841, + 0.005872929003089666, + 0.0026315401773899794, + 0.008804261684417725, + 0.0027798672672361135, + 0.004519700072705746, + 0.0033099930733442307, + 0.004212571308016777, + 0.003957157023251057, + 0.0038429712876677513, + 0.004668813664466143, + 0.0027206314262002707, + 0.005286300089210272, + 0.002739035990089178, + 0.0047641173005104065, + 0.004877940285950899, + 0.004324011970311403, + 0.003719376167282462, + 0.003062534611672163, + 0.006396360695362091, + 0.0043751560151577, + 0.010780593380331993, + 0.005019271746277809, + 0.003997601103037596, + 0.005643055774271488, + 0.004436557646840811, + 0.005676334723830223, + 0.0038410299457609653, + 0.0033851657062768936, + 0.0028572746086865664, + 0.010960215702652931, + 0.013063021004199982, + 0.014183119870722294, + 0.019162384793162346, + 0.005472402088344097, + 0.004228034522384405, + 0.010061261244118214, + 0.007323319558054209, + 0.008394578471779823, + 0.0037402065936475992, + 0.0066053494811058044, + 0.007518725004047155, + 0.002083418658003211, + 0.004297301638871431, + 0.002757288981229067, + 0.005935031920671463, + 0.013825672678649426, + 0.01310733798891306, + 0.027242328971624374, + 0.01488020084798336, + 0.006418684497475624, + 0.012051939964294434, + 0.010725888423621655, + 0.023508327081799507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.05495677888393402 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369003161787987 + ], + "opposite_logprob_tokens": [ + -3.4086899757385254 + ], + "logprob_reward": -0.03369003161787987, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369003161787987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0596243180334568, + 0.0691186860203743, + 0.028698788955807686, + 0.0687146857380867, + 0.080707848072052, + 0.07425938546657562, + 0.17328883707523346, + 0.04751388356089592, + 0.15131935477256775, + 0.24675413966178894 + ], + "credit_map": [ + 0.003500631544739008, + 0.008035531267523766, + 0.003477726597338915, + 0.005302970763295889, + 0.014264161698520184, + 0.007436507847160101, + 0.007866009138524532, + 0.009740782901644707, + 0.005920762661844492, + 0.009012894704937935, + 0.009935081005096436, + 0.008439337834715843, + 0.006777163129299879, + 0.011806250549852848, + 0.01027060579508543, + 0.006956599187105894, + 0.0034819887951016426, + 0.004029759671539068, + 0.002899510320276022, + 0.00399094307795167, + 0.004128139466047287, + 0.0028578233905136585, + 0.003912953194230795, + 0.0033976752310991287, + 0.006098249927163124, + 0.003495266893878579, + 0.017910445109009743, + 0.008850221522152424, + 0.0028896471485495567, + 0.016304705291986465, + 0.004023502115160227, + 0.00914265401661396, + 0.010527192614972591, + 0.00835546012967825, + 0.010313532315194607, + 0.007751444820314646, + 0.011024387553334236, + 0.014155535958707333, + 0.00813025888055563, + 0.010450048372149467, + 0.007613092195242643, + 0.005983863491564989, + 0.007232639938592911, + 0.006015714257955551, + 0.008161811158061028, + 0.01563350111246109, + 0.01071248110383749, + 0.012906283140182495, + 0.016088884323835373, + 0.020125510171055794, + 0.02576545998454094, + 0.018957750871777534, + 0.033680956810712814, + 0.02444588392972946, + 0.017122896388173103, + 0.017101513221859932, + 0.00798365194350481, + 0.00633970694616437, + 0.006780811119824648, + 0.005852918606251478, + 0.006987045053392649, + 0.004702682141214609, + 0.0038750022649765015, + 0.00499206967651844, + 0.03486642614006996, + 0.014734181575477123, + 0.026480335742235184, + 0.024850619956851006, + 0.005819618236273527, + 0.012910687364637852, + 0.011905327439308167, + 0.01975216716527939, + 0.0324723981320858, + 0.027108820155262947, + 0.06461065262556076, + 0.03292227163910866, + 0.01252791192382574, + 0.022461747750639915, + 0.01760111376643181, + 0.03704923018813133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.03369003161787987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188248664140701 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188248664140701, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188248664140701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5182598233222961, + 0.03219320997595787, + 0.023754166439175606, + 0.023166289553046227, + 0.10383862257003784, + 0.047677334398031235, + 0.034524355083703995, + 0.026141101494431496, + 0.09483109414577484, + 0.09561392664909363 + ], + "credit_map": [ + 0.018477031961083412, + 0.025411007925868034, + 0.03048882633447647, + 0.02011149749159813, + 0.2059338092803955, + 0.04141369089484215, + 0.050837136805057526, + 0.1255868673324585, + 0.0028196563944220543, + 0.0035166670568287373, + 0.003511079354211688, + 0.005480894353240728, + 0.004642108455300331, + 0.0038065230473876, + 0.005013744346797466, + 0.0034025416243821383, + 0.0031744639854878187, + 0.002260778797790408, + 0.0021792540792375803, + 0.0020031463354825974, + 0.0026089900638908148, + 0.002112759044393897, + 0.006303905509412289, + 0.003110869089141488, + 0.0021977166179567575, + 0.0019229855388402939, + 0.0033018263056874275, + 0.0029072375036776066, + 0.002926864894106984, + 0.003456764155998826, + 0.0038308880757540464, + 0.0026220090221613646, + 0.012346668168902397, + 0.01298236008733511, + 0.016440384089946747, + 0.012904838658869267, + 0.009656655602157116, + 0.016781581565737724, + 0.012456720694899559, + 0.010269428603351116, + 0.006306929979473352, + 0.005903332494199276, + 0.006129757035523653, + 0.00517237326130271, + 0.003507608314976096, + 0.00592529634013772, + 0.0049506668001413345, + 0.009781372733414173, + 0.004430785309523344, + 0.005282491445541382, + 0.0058795977383852005, + 0.006457467097789049, + 0.003995444159954786, + 0.0024171266704797745, + 0.003436147468164563, + 0.002625297522172332, + 0.0034307537134736776, + 0.00320193893276155, + 0.003721672110259533, + 0.004747284110635519, + 0.0034534609876573086, + 0.0019857471343129873, + 0.0025489493273198605, + 0.0030512982048094273, + 0.01772308722138405, + 0.010662462562322617, + 0.01436975784599781, + 0.016076793894171715, + 0.00522040668874979, + 0.01094034593552351, + 0.006616788916289806, + 0.013221455737948418, + 0.013554391451179981, + 0.007262132596224546, + 0.03209344670176506, + 0.009474832564592361, + 0.0037160394713282585, + 0.007122904062271118, + 0.004981564357876778, + 0.01740863174200058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.014188248664140701 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, sweeping camera shot arcs around a fearsome warrior as he meticulously scales a sheer, treacherous cliff face. Rendered in a detailed, heroic fantasy illustration reminiscent of classic Dungeons & Dragons art, the scene is a study in tension and determination. The warrior, a man of heroic stature, has long, black braids blown by the high-altitude wind. His face is a mask of intense concentration, his jaw set, his eyes fixed on his next handhold. His muscular body is clad in worn dark leather armor and iron-studded pauldrons that have seen many battles. Strapped to his back is a massive, double-headed battle axe, its polished steel edge gleaming in the light, its heavy wooden handle intricately carved with patterns. His gloved hands dig into the jagged, unforgiving rock, his body moving with a steady and deliberate rhythm that demonstrates his immense strength. As the camera continues to sweep, it captures the dramatic interplay of shadow and light cast by the low-angle sun, highlighting the rough texture of the granite and the climber's strained muscles. The camera zooms out slightly to reveal the abyss below, a vertiginous chasm filled with swirling mist that obscures the bottom, creating a powerful and heroic atmosphere of a lonely struggle against an epic, unforgiving landscape.", + "seed": 23393, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23393_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01840332169085741, + "vlm_logprob": -0.01840332169085741, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.01840332169085741, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018859826028347015 + ], + "opposite_logprob_tokens": [ + -4.26885986328125 + ], + "logprob_reward": -0.018859826028347015, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018859826028347015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22374874353408813, + 0.08129280805587769, + 0.06477892398834229, + 0.06418512761592865, + 0.06491401791572571, + 0.0568089559674263, + 0.07578439265489578, + 0.06022649630904198, + 0.23380818963050842, + 0.07445236295461655 + ], + "credit_map": [ + 0.03071032464504242, + 0.03530575707554817, + 0.03469723090529442, + 0.026936490088701248, + 0.02334195002913475, + 0.023350410163402557, + 0.022571342065930367, + 0.026835236698389053, + 0.01160738617181778, + 0.013583979569375515, + 0.006875178776681423, + 0.00753778638318181, + 0.01041970681399107, + 0.014737652614712715, + 0.007621549069881439, + 0.008909570053219795, + 0.008593685925006866, + 0.013433498330414295, + 0.007066862192004919, + 0.009985102340579033, + 0.0065162042155861855, + 0.007687046658247709, + 0.0047720312140882015, + 0.006724494509398937, + 0.01156223751604557, + 0.008020466193556786, + 0.007954580709338188, + 0.010285487398505211, + 0.004824519157409668, + 0.008977249264717102, + 0.00635925680398941, + 0.0062013352289795876, + 0.0073026311583817005, + 0.01299350056797266, + 0.00568661792203784, + 0.007837647572159767, + 0.009808010421693325, + 0.007915466092526913, + 0.006438582669943571, + 0.006931557320058346, + 0.0051506683230400085, + 0.008440469391644001, + 0.007527833338826895, + 0.003548351116478443, + 0.007808741647750139, + 0.01218697614967823, + 0.0054864767007529736, + 0.006659441627562046, + 0.009305448271334171, + 0.009267151355743408, + 0.011158685199916363, + 0.00807365495711565, + 0.010466196574270725, + 0.013438736088573933, + 0.006022593472152948, + 0.008051923476159573, + 0.006767961662262678, + 0.009337367489933968, + 0.012246984988451004, + 0.006473316811025143, + 0.0070723011158406734, + 0.005601199343800545, + 0.0054231975227594376, + 0.007304168771952391, + 0.05048602074384689, + 0.01324744988232851, + 0.09107059240341187, + 0.012108173221349716, + 0.009901040233671665, + 0.01287611573934555, + 0.00603735214099288, + 0.03808145225048065, + 0.011826126836240292, + 0.008194921538233757, + 0.006525860168039799, + 0.006763812154531479, + 0.012855712324380875, + 0.008952995762228966, + 0.004798920825123787, + 0.014534014277160168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.018859826028347015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796753525733948 + ], + "opposite_logprob_tokens": [ + -4.634796619415283 + ], + "logprob_reward": -0.009796753525733948, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796753525733948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2591116726398468, + 0.05120057985186577, + 0.03769369795918465, + 0.04926953464746475, + 0.03032028302550316, + 0.1572161465883255, + 0.075795017182827, + 0.05698300153017044, + 0.16639064252376556, + 0.11601938307285309 + ], + "credit_map": [ + 0.0400034599006176, + 0.0343228280544281, + 0.03336095064878464, + 0.03002222254872322, + 0.029129810631275177, + 0.03397541865706444, + 0.030313637107610703, + 0.027983352541923523, + 0.0067371497862041, + 0.006591151934117079, + 0.0041627720929682255, + 0.006169838830828667, + 0.007351282052695751, + 0.008023872971534729, + 0.006452065426856279, + 0.005712447687983513, + 0.004585996735841036, + 0.006851312704384327, + 0.0032960514072328806, + 0.004236565437167883, + 0.0049736700020730495, + 0.004730327054858208, + 0.0031866279896348715, + 0.00583315035328269, + 0.006918464321643114, + 0.005271523259580135, + 0.004374405834823847, + 0.00421128561720252, + 0.006233342923223972, + 0.011323143728077412, + 0.005015519913285971, + 0.005921850446611643, + 0.004368027672171593, + 0.00394008494913578, + 0.0038713710382580757, + 0.0025140747893601656, + 0.0036943398881703615, + 0.005065690726041794, + 0.003163957269862294, + 0.003702739253640175, + 0.016045967116951942, + 0.02254837192595005, + 0.01328892633318901, + 0.012911028228700161, + 0.034887101501226425, + 0.02389952540397644, + 0.014706204645335674, + 0.018929030746221542, + 0.010105710476636887, + 0.011011774651706219, + 0.016675550490617752, + 0.006225787103176117, + 0.006013641599565744, + 0.009927813895046711, + 0.006644957233220339, + 0.0091897863894701, + 0.007286611944437027, + 0.0058746603317558765, + 0.013599736616015434, + 0.006974189076572657, + 0.006062410306185484, + 0.005708192940801382, + 0.00385845685377717, + 0.007618745788931847, + 0.02556564472615719, + 0.020069943740963936, + 0.04667024314403534, + 0.013392245396971703, + 0.019702665507793427, + 0.00968069490045309, + 0.009358521550893784, + 0.02195069007575512, + 0.014147509820759296, + 0.012984110973775387, + 0.013162502087652683, + 0.011591188609600067, + 0.02207387425005436, + 0.013133869506418705, + 0.011075012385845184, + 0.017851321026682854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.009796753525733948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809775784611702 + ], + "opposite_logprob_tokens": [ + -3.288097858428955 + ], + "logprob_reward": -0.03809775784611702, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809775784611702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18116526305675507, + 0.045803628861904144, + 0.09341150522232056, + 0.08233339339494705, + 0.07435011863708496, + 0.13442476093769073, + 0.15193301439285278, + 0.11923690140247345, + 0.06933126598596573, + 0.048010218888521194 + ], + "credit_map": [ + 0.02938801236450672, + 0.02167689800262451, + 0.018885618075728416, + 0.012787815183401108, + 0.02040480449795723, + 0.0329103022813797, + 0.02256746217608452, + 0.022544335573911667, + 0.006268960423767567, + 0.006376262288540602, + 0.0068892198614776134, + 0.0036718128249049187, + 0.005470091477036476, + 0.007675684988498688, + 0.005091227125376463, + 0.004360366612672806, + 0.022977065294981003, + 0.009470805525779724, + 0.008710645139217377, + 0.008569211699068546, + 0.0082634836435318, + 0.013163300231099129, + 0.006343770772218704, + 0.015913212671875954, + 0.019408510997891426, + 0.007735954597592354, + 0.009329589083790779, + 0.008360748179256916, + 0.004949292168021202, + 0.014435382559895515, + 0.008734204806387424, + 0.009379694238305092, + 0.010388254188001156, + 0.008737965486943722, + 0.008116984739899635, + 0.006916313897818327, + 0.007585496176034212, + 0.012492024339735508, + 0.006697326432913542, + 0.013415743596851826, + 0.01312479842454195, + 0.01826418936252594, + 0.012758840806782246, + 0.010668259114027023, + 0.02506966143846512, + 0.027364304289221764, + 0.010429401881992817, + 0.01674528419971466, + 0.02337346225976944, + 0.018375322222709656, + 0.041276585310697556, + 0.013150540180504322, + 0.012406374327838421, + 0.01850508153438568, + 0.009375214576721191, + 0.0154704125598073, + 0.01820502057671547, + 0.01638997346162796, + 0.02283884584903717, + 0.012401682324707508, + 0.012164703570306301, + 0.012572473846375942, + 0.008997243829071522, + 0.015666941180825233, + 0.011404219083487988, + 0.011576595716178417, + 0.019840210676193237, + 0.005404582712799311, + 0.004771205596625805, + 0.0044723860919475555, + 0.00288753816857934, + 0.008974513970315456, + 0.006591080222278833, + 0.00758162047713995, + 0.005481142550706863, + 0.004838635213673115, + 0.007760246284306049, + 0.006291735451668501, + 0.004983227699995041, + 0.0044825212098658085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.03809775784611702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13046850264072418, + 0.04417916759848595, + 0.04477573186159134, + 0.053092844784259796, + 0.04873383790254593, + 0.07586484402418137, + 0.21184779703617096, + 0.03487958386540413, + 0.31992781162261963, + 0.03622991591691971 + ], + "credit_map": [ + 0.020481225103139877, + 0.01748831570148468, + 0.01995961368083954, + 0.015656061470508575, + 0.011866734363138676, + 0.015797065570950508, + 0.014384206384420395, + 0.014835276640951633, + 0.0052118320018053055, + 0.005745213478803635, + 0.005709764547646046, + 0.0050939409993588924, + 0.00406326400116086, + 0.008157886564731598, + 0.005157312378287315, + 0.005039957817643881, + 0.007978811860084534, + 0.007195570040494204, + 0.00477733975276351, + 0.0035990700125694275, + 0.00407622242346406, + 0.005192385986447334, + 0.004589727148413658, + 0.00736660324037075, + 0.010933665558695793, + 0.005302295554429293, + 0.004966565873473883, + 0.0040652393363416195, + 0.006118339952081442, + 0.012119513936340809, + 0.003436771919950843, + 0.006150454748421907, + 0.006071618292480707, + 0.0050367992371320724, + 0.006537301931530237, + 0.0032847702968865633, + 0.007770877331495285, + 0.008777476847171783, + 0.004036865197122097, + 0.007218129932880402, + 0.007584827020764351, + 0.010533005930483341, + 0.010665801353752613, + 0.00642730388790369, + 0.008508292958140373, + 0.01635645143687725, + 0.006779143121093512, + 0.009010015986859798, + 0.030160831287503242, + 0.033136870712041855, + 0.055553507059812546, + 0.01439990010112524, + 0.015139438211917877, + 0.02494201995432377, + 0.01449376717209816, + 0.024021463468670845, + 0.004658656194806099, + 0.003146145958453417, + 0.007693364750593901, + 0.0042934962548315525, + 0.003339502727612853, + 0.0036465178709477186, + 0.0033330915030092, + 0.0047688111662864685, + 0.06466018408536911, + 0.022198081016540527, + 0.11949091404676437, + 0.01768561452627182, + 0.01857958920300007, + 0.01726076751947403, + 0.009681039489805698, + 0.050371620804071426, + 0.0062561421655118465, + 0.004610391333699226, + 0.005970189813524485, + 0.003946250304579735, + 0.0036236420273780823, + 0.0052999649196863174, + 0.0025675888173282146, + 0.00395574513822794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.01419165637344122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107061468064785 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.01107061468064785, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107061468064785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20155812799930573, + 0.04478110373020172, + 0.06858652085065842, + 0.053160328418016434, + 0.033858563750982285, + 0.07498761266469955, + 0.2303728461265564, + 0.11657662689685822, + 0.05126520246267319, + 0.12485314160585403 + ], + "credit_map": [ + 0.02044222503900528, + 0.02772957645356655, + 0.019195973873138428, + 0.028320863842964172, + 0.033525120466947556, + 0.031847428530454636, + 0.01860368438065052, + 0.02189325913786888, + 0.006338137201964855, + 0.00447826087474823, + 0.0036507025361061096, + 0.004536010324954987, + 0.0056311446242034435, + 0.00876561738550663, + 0.005445755552500486, + 0.005935477092862129, + 0.012993060052394867, + 0.010018382221460342, + 0.0078965425491333, + 0.00525936484336853, + 0.008565986528992653, + 0.00840972550213337, + 0.00482810428366065, + 0.010615351609885693, + 0.008619849570095539, + 0.006949650589376688, + 0.005566511768847704, + 0.008281586691737175, + 0.004041966050863266, + 0.00780563335865736, + 0.005867592059075832, + 0.006027538795024157, + 0.0041946410201489925, + 0.004059616010636091, + 0.003907608799636364, + 0.003915258217602968, + 0.0031498647294938564, + 0.006926111876964569, + 0.0037352556828409433, + 0.003970205318182707, + 0.01009769644588232, + 0.014543399214744568, + 0.008392060175538063, + 0.0071237944066524506, + 0.010630897246301174, + 0.010019760578870773, + 0.006269044708460569, + 0.007910956628620625, + 0.0357491672039032, + 0.0297491904348135, + 0.0643298327922821, + 0.0163390152156353, + 0.01659424975514412, + 0.026628203690052032, + 0.015357984229922295, + 0.025625189766287804, + 0.012225953862071037, + 0.016765229403972626, + 0.026632020249962807, + 0.013719804584980011, + 0.01169844064861536, + 0.012304398231208324, + 0.00828047189861536, + 0.014950308948755264, + 0.008649465627968311, + 0.006124895066022873, + 0.009864608757197857, + 0.00610937038436532, + 0.004124882631003857, + 0.006370446179062128, + 0.004208200145512819, + 0.0058133359998464584, + 0.026394415646791458, + 0.012132133357226849, + 0.013927476480603218, + 0.009513854049146175, + 0.00694353599101305, + 0.012653324753046036, + 0.0059702033177018166, + 0.03731819614768028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.01107061468064785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, sweeping camera shot arcs around a fearsome warrior as he meticulously scales a sheer, treacherous cliff face. Rendered in a detailed, heroic fantasy illustration reminiscent of classic Dungeons & Dragons art, the scene is a study in tension and determination. The warrior, a man of heroic stature, has long, black braids blown by the high-altitude wind. His face is a mask of intense concentration, his jaw set, his eyes fixed on his next handhold. His muscular body is clad in worn dark leather armor and iron-studded pauldrons that have seen many battles. Strapped to his back is a massive, double-headed battle axe, its polished steel edge gleaming in the light, its heavy wooden handle intricately carved with patterns. His gloved hands dig into the jagged, unforgiving rock, his body moving with a steady and deliberate rhythm that demonstrates his immense strength. As the camera continues to sweep, it captures the dramatic interplay of shadow and light cast by the low-angle sun, highlighting the rough texture of the granite and the climber's strained muscles. The camera zooms out slightly to reveal the abyss below, a vertiginous chasm filled with swirling mist that obscures the bottom, creating a powerful and heroic atmosphere of a lonely struggle against an epic, unforgiving landscape.", + "seed": 23394, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_7/index_9633_seed_23394_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.035486551374197005, + "vlm_logprob": -0.035486551374197005, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.035486551374197005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01357361301779747 + ], + "opposite_logprob_tokens": [ + -4.51357364654541 + ], + "logprob_reward": -0.01357361301779747, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01357361301779747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1850469708442688, + 0.07767286896705627, + 0.08903337270021439, + 0.06400418281555176, + 0.056791920214891434, + 0.07158567756414413, + 0.24612343311309814, + 0.10860845446586609, + 0.06043253466486931, + 0.040700607001781464 + ], + "credit_map": [ + 0.019378257915377617, + 0.019484933465719223, + 0.02966427616775036, + 0.019944339990615845, + 0.03531516715884209, + 0.027702581137418747, + 0.016069037839770317, + 0.01748836599290371, + 0.011623304337263107, + 0.011389640159904957, + 0.0074392869137227535, + 0.009119413793087006, + 0.01202124822884798, + 0.009877527132630348, + 0.009480160661041737, + 0.006722288206219673, + 0.011817612685263157, + 0.006877968553453684, + 0.007234314922243357, + 0.0060878414660692215, + 0.01578279212117195, + 0.01356741413474083, + 0.008184649050235748, + 0.01948077790439129, + 0.007559929974377155, + 0.005402767565101385, + 0.006831386126577854, + 0.0036250422708690166, + 0.011606846004724503, + 0.009740054607391357, + 0.006091912277042866, + 0.013146239332854748, + 0.006092825438827276, + 0.0072282589972019196, + 0.006518007256090641, + 0.005840009078383446, + 0.007037213537842035, + 0.007555113174021244, + 0.006444311700761318, + 0.010076174512505531, + 0.015100670047104359, + 0.006763080134987831, + 0.013837289065122604, + 0.004594635218381882, + 0.0064743272960186005, + 0.008430740796029568, + 0.0062576537020504475, + 0.01012727152556181, + 0.046110909432172775, + 0.019522076472640038, + 0.0434957891702652, + 0.017298664897680283, + 0.02698364295065403, + 0.03056519664824009, + 0.027721021324396133, + 0.034426115453243256, + 0.009384222328662872, + 0.009885468520224094, + 0.011446637101471424, + 0.01061223354190588, + 0.016353558748960495, + 0.017483597621321678, + 0.01980608142912388, + 0.013636640273034573, + 0.01280087698251009, + 0.004110304638743401, + 0.015380378812551498, + 0.0039195953868329525, + 0.004387666936963797, + 0.00674035307019949, + 0.0054093278013169765, + 0.007684020791202784, + 0.007485520094633102, + 0.0032836697064340115, + 0.005863993428647518, + 0.002855908125638962, + 0.004489310551434755, + 0.005742088425904512, + 0.007166135590523481, + 0.0038139799144119024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.01357361301779747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636094018816948 + ], + "opposite_logprob_tokens": [ + -3.6513609886169434 + ], + "logprob_reward": -0.02636094018816948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636094018816948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11475960165262222, + 0.07717017829418182, + 0.07928413897752762, + 0.03247550129890442, + 0.20419549942016602, + 0.03538156673312187, + 0.24622344970703125, + 0.09010133892297745, + 0.08850105106830597, + 0.03190765529870987 + ], + "credit_map": [ + 0.012395903468132019, + 0.00892302393913269, + 0.014216041192412376, + 0.016729267314076424, + 0.026182837784290314, + 0.01701531372964382, + 0.008924616500735283, + 0.010372606106102467, + 0.010621920228004456, + 0.01298022735863924, + 0.006819709669798613, + 0.012400765903294086, + 0.011504646390676498, + 0.008076581172645092, + 0.00822990108281374, + 0.006536433473229408, + 0.008187096565961838, + 0.0063957832753658295, + 0.007871254347264767, + 0.006513136904686689, + 0.011812751181423664, + 0.011625055223703384, + 0.009309527464210987, + 0.017569540068507195, + 0.0045169987715780735, + 0.0025044362992048264, + 0.003524630796164274, + 0.00179710378870368, + 0.004226208198815584, + 0.00544782355427742, + 0.0028834932018071413, + 0.007574810180813074, + 0.02242830954492092, + 0.03453867509961128, + 0.023170461878180504, + 0.01894601806998253, + 0.025548305362462997, + 0.02538093365728855, + 0.024374712258577347, + 0.029808087274432182, + 0.006060771178454161, + 0.0036194496788084507, + 0.004407565109431744, + 0.0021785073913633823, + 0.00460132397711277, + 0.005237148609012365, + 0.004639503080397844, + 0.0046372986398637295, + 0.043803609907627106, + 0.024227634072303772, + 0.045961640775203705, + 0.016785800457000732, + 0.02487211860716343, + 0.026736333966255188, + 0.029608452692627907, + 0.0342278927564621, + 0.0075295595452189445, + 0.006528226658701897, + 0.006820105016231537, + 0.00854349136352539, + 0.017999673262238503, + 0.015147330239415169, + 0.01275612786412239, + 0.014776838012039661, + 0.011085638776421547, + 0.010466930456459522, + 0.013017035089433193, + 0.006008714437484741, + 0.012121093459427357, + 0.012720298022031784, + 0.01379964966326952, + 0.009281697683036327, + 0.004568410571664572, + 0.004116641823202372, + 0.005781887099146843, + 0.001594285829924047, + 0.003167494898661971, + 0.005690369755029678, + 0.003911982290446758, + 0.003076584078371525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.02636094018816948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495350807905197 + ], + "opposite_logprob_tokens": [ + -2.9299535751342773 + ], + "logprob_reward": -0.05495350807905197, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495350807905197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15391822159290314, + 0.08567926287651062, + 0.08109048008918762, + 0.07226686924695969, + 0.05456879734992981, + 0.08470424264669418, + 0.22915634512901306, + 0.12123245745897293, + 0.06365169584751129, + 0.053731586784124374 + ], + "credit_map": [ + 0.020587192848324776, + 0.014504121616482735, + 0.017359012737870216, + 0.018428975716233253, + 0.033669933676719666, + 0.025767289102077484, + 0.010933807119727135, + 0.012667888775467873, + 0.013528810814023018, + 0.00966765359044075, + 0.007362496107816696, + 0.006650518160313368, + 0.012691065669059753, + 0.010261178016662598, + 0.009834041818976402, + 0.015683505684137344, + 0.01705768145620823, + 0.006417927797883749, + 0.008037502877414227, + 0.004240947309881449, + 0.01308339461684227, + 0.011519674211740494, + 0.00787369068711996, + 0.01285966020077467, + 0.008261334151029587, + 0.007284488994628191, + 0.009036973118782043, + 0.00607095193117857, + 0.011666149832308292, + 0.009590676985681057, + 0.008418157696723938, + 0.011938130483031273, + 0.0064000096172094345, + 0.0053041051141917706, + 0.011493884027004242, + 0.004835308063775301, + 0.005072300788015127, + 0.007643742021173239, + 0.00744972750544548, + 0.0063697220757603645, + 0.013348943553864956, + 0.007475375197827816, + 0.012420494109392166, + 0.004283626563847065, + 0.010668091475963593, + 0.011209040880203247, + 0.011113577522337437, + 0.01418509241193533, + 0.02554921992123127, + 0.012987964786589146, + 0.02961624041199684, + 0.014058192260563374, + 0.033378470689058304, + 0.05556432902812958, + 0.020186517387628555, + 0.03781542554497719, + 0.014247026294469833, + 0.009998951107263565, + 0.016947325319051743, + 0.011569956317543983, + 0.016728945076465607, + 0.0181625597178936, + 0.018199028447270393, + 0.015378667041659355, + 0.013970568776130676, + 0.003846873063594103, + 0.014294995926320553, + 0.004156504757702351, + 0.005666783545166254, + 0.007021613419055939, + 0.00861525721848011, + 0.0060791014693677425, + 0.007682040799409151, + 0.005218026228249073, + 0.008566208183765411, + 0.003083444433286786, + 0.007363923825323582, + 0.009539213962852955, + 0.0070096454583108425, + 0.0052690827287733555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.05495350807905197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000981271266937 + ], + "opposite_logprob_tokens": [ + -2.695009708404541 + ], + "logprob_reward": -0.07000981271266937, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000981271266937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2044454663991928, + 0.04844583570957184, + 0.11089407652616501, + 0.09084869921207428, + 0.3010542392730713, + 0.060326412320137024, + 0.087052121758461, + 0.05329787731170654, + 0.020705299451947212, + 0.02292989194393158 + ], + "credit_map": [ + 0.022180533036589622, + 0.01718408614397049, + 0.03606407716870308, + 0.02211698144674301, + 0.06395845860242844, + 0.018762879073619843, + 0.011888401582837105, + 0.012290057726204395, + 0.005520250648260117, + 0.0048054782673716545, + 0.00495719350874424, + 0.00533438753336668, + 0.0074512106366455555, + 0.007343480829149485, + 0.004744917154312134, + 0.008288916200399399, + 0.01142637524753809, + 0.007764230947941542, + 0.010106631554663181, + 0.009195653721690178, + 0.01587083376944065, + 0.015653662383556366, + 0.0145207354798913, + 0.026355959475040436, + 0.00849156454205513, + 0.005723730660974979, + 0.010350469499826431, + 0.0035535250790417194, + 0.012400344014167786, + 0.016515744850039482, + 0.008377439342439175, + 0.025435879826545715, + 0.035059425979852676, + 0.046909742057323456, + 0.035878609865903854, + 0.024665970355272293, + 0.03707199916243553, + 0.03853681683540344, + 0.036789048463106155, + 0.04614263400435448, + 0.0057974765077233315, + 0.0053358315490186214, + 0.0062106819823384285, + 0.0050372472032904625, + 0.009919681586325169, + 0.007938754744827747, + 0.007348009385168552, + 0.012738724239170551, + 0.011128445155918598, + 0.008470553904771805, + 0.012110783718526363, + 0.007791406940668821, + 0.011706586927175522, + 0.012916434556245804, + 0.010353339836001396, + 0.0125745739787817, + 0.003172138938680291, + 0.005674116313457489, + 0.004372903611510992, + 0.005881220102310181, + 0.010362472385168076, + 0.009286188520491123, + 0.008193970657885075, + 0.006354866083711386, + 0.0041091302409768105, + 0.0016158826183527708, + 0.004246428143233061, + 0.0013738104607909918, + 0.0021022134460508823, + 0.0024607337545603514, + 0.001759944949299097, + 0.0030371572356671095, + 0.004254729021340609, + 0.0026792241260409355, + 0.003952403552830219, + 0.0012670735595747828, + 0.0018842756981030107, + 0.0038128721062093973, + 0.003002770710736513, + 0.002076544566079974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07000981271266937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man wearing worn dark leather armor and iron-studded pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior scaling a sheer cliff face with his gloved hands digging into jagged rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive double-headed battle axe strapped to the warrior's back with a polished steel edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera arc around the warrior and zoom out to reveal a vertiginous chasm filled with swirling mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a detailed heroic fantasy illustration reminiscent of classic Dungeons & Dragons art?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534882873296738 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.012534882873296738, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534882873296738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10753651708364487, + 0.0699995756149292, + 0.09939935058355331, + 0.040065936744213104, + 0.061771195381879807, + 0.18043792247772217, + 0.1882655918598175, + 0.1648782193660736, + 0.03348960354924202, + 0.05415603518486023 + ], + "credit_map": [ + 0.011941731907427311, + 0.01151341199874878, + 0.010057375766336918, + 0.011193983256816864, + 0.023812564089894295, + 0.015851125121116638, + 0.00889702420681715, + 0.014269295148551464, + 0.008191950619220734, + 0.007938439957797527, + 0.007702661678195, + 0.007311146706342697, + 0.012600768357515335, + 0.009899559430778027, + 0.009761115536093712, + 0.006593933328986168, + 0.022741587832570076, + 0.009000733494758606, + 0.009129400365054607, + 0.005436658393591642, + 0.013130069710314274, + 0.015581289306282997, + 0.009533502161502838, + 0.014846106991171837, + 0.007213488686829805, + 0.003426157170906663, + 0.005714870989322662, + 0.0023863916285336018, + 0.005668531637638807, + 0.0056537846103310585, + 0.004161378834396601, + 0.005841333419084549, + 0.006107860244810581, + 0.009756897576153278, + 0.007758675143122673, + 0.00630682660266757, + 0.0077901543118059635, + 0.00831999909132719, + 0.006639896892011166, + 0.009090887382626534, + 0.024716056883335114, + 0.018747087568044662, + 0.02488541789352894, + 0.01344601996243, + 0.026954391971230507, + 0.024683574214577675, + 0.020272333174943924, + 0.026733040809631348, + 0.026058310642838478, + 0.010838362388312817, + 0.033434439450502396, + 0.010985583066940308, + 0.02094571478664875, + 0.03295186907052994, + 0.020491788163781166, + 0.03255951404571533, + 0.025409294292330742, + 0.011395305395126343, + 0.021202953532338142, + 0.012924314476549625, + 0.02581392601132393, + 0.02600405178964138, + 0.01882188953459263, + 0.02330648899078369, + 0.00853507686406374, + 0.0020672951359301805, + 0.008631717413663864, + 0.0018196441233158112, + 0.002675228053703904, + 0.0033922709990292788, + 0.002680229488760233, + 0.0036881447304040194, + 0.007566820830106735, + 0.005961294751614332, + 0.011469762772321701, + 0.003213230287656188, + 0.006521924864500761, + 0.0077702817507088184, + 0.006575684528797865, + 0.005077039822936058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.012534882873296738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_8/index_tensor([409, 409])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_8/index_tensor([409, 409])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fe62ea77ff5dd9c2a444414bd45bfacca95ad19d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_8/index_tensor([409, 409])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, mid-range shot from a horizontal angle focuses firmly on the tuning dial of an old radio. The radio, with its dark brown wooden case, bears subtle scratches and a patina of age. It sits on a wooden tabletop, its background softly blurred, the outlines of a bookshelf barely visible. The space is dimly lit. The dial is a rectangular glass panel, with a yellowed frequency scale and city names printed behind it. A slender red hand motionlessly points to a specific frequency. A soft, warm amber light filters through from within the dial, gradually pulsating from a faint glow to a bright, then slowly fading in a steady, soothing rhythm. Each time the light illuminates, it not only clearly illuminates the details on the dial but also casts dynamic shadows on the wooden case. This rhythmic light, combined with the texture of the old object, creates a profoundly tranquil, warm, and nostalgic atmosphere of contemplation.", + "seed": 32477, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_8/index_409_seed_32477_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03584291690494865, + "vlm_logprob": -0.03584291690494865, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03584291690494865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09853891283273697 + ], + "opposite_logprob_tokens": [ + -2.473538875579834 + ], + "logprob_reward": -0.09853891283273697, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09853891283273697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4536726474761963, + 0.10350553691387177, + 0.059285473078489304, + 0.07873896509408951, + 0.047632958739995956, + 0.04050637036561966, + 0.07163088768720627, + 0.04380186274647713, + 0.06536974012851715, + 0.03585564345121384 + ], + "credit_map": [ + 0.04530477151274681, + 0.04694121330976486, + 0.07084420323371887, + 0.06795921176671982, + 0.05372009798884392, + 0.05620167404413223, + 0.05730991065502167, + 0.05539151281118393, + 0.00964544340968132, + 0.008477292023599148, + 0.016399921849370003, + 0.014481722377240658, + 0.013731150887906551, + 0.014022478833794594, + 0.011238397099077702, + 0.015509111806750298, + 0.006382638122886419, + 0.00477982172742486, + 0.008977980352938175, + 0.007265252526849508, + 0.006820144131779671, + 0.007300063967704773, + 0.006479123141616583, + 0.011280439794063568, + 0.006906855385750532, + 0.010316172614693642, + 0.011626257561147213, + 0.009216394275426865, + 0.01098096463829279, + 0.01023936364799738, + 0.008893889375030994, + 0.01055905781686306, + 0.00457228347659111, + 0.004627615679055452, + 0.00732668861746788, + 0.006766323000192642, + 0.006548498757183552, + 0.004752850625663996, + 0.006356888450682163, + 0.006681803148239851, + 0.003165758214890957, + 0.004499122034758329, + 0.0060857185162603855, + 0.005573732312768698, + 0.005847200285643339, + 0.0051215034909546375, + 0.00509653752669692, + 0.0051167914643883705, + 0.006842584349215031, + 0.006502852775156498, + 0.01111839059740305, + 0.009226417168974876, + 0.011769344098865986, + 0.00882069393992424, + 0.0069404952228069305, + 0.01041010208427906, + 0.004325482528656721, + 0.004269986413419247, + 0.006707665044814348, + 0.005305729806423187, + 0.005993135739117861, + 0.005867987405508757, + 0.005998009350150824, + 0.005333862267434597, + 0.005480983294546604, + 0.00647599995136261, + 0.009788507595658302, + 0.006657915655523539, + 0.008721664547920227, + 0.01037693303078413, + 0.00661497563123703, + 0.011252752505242825, + 0.003809940768405795, + 0.004079884849488735, + 0.005500636994838715, + 0.0043932609260082245, + 0.004047894384711981, + 0.0036223987117409706, + 0.004223253577947617, + 0.006178371608257294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.09853891283273697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221626333892345 + ], + "opposite_logprob_tokens": [ + -4.264221668243408 + ], + "logprob_reward": -0.014221626333892345, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221626333892345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3629800081253052, + 0.15692803263664246, + 0.1128813847899437, + 0.04998528212308884, + 0.049270711839199066, + 0.038094840943813324, + 0.05035407096147537, + 0.06825299561023712, + 0.05585239455103874, + 0.05540020391345024 + ], + "credit_map": [ + 0.05140896141529083, + 0.05198247730731964, + 0.053779684007167816, + 0.043605051934719086, + 0.038052450865507126, + 0.038647785782814026, + 0.034674182534217834, + 0.05082940310239792, + 0.02190503664314747, + 0.029284682124853134, + 0.02450004406273365, + 0.01707707718014717, + 0.01837281510233879, + 0.012171234004199505, + 0.01308755949139595, + 0.020529592409729958, + 0.01584594137966633, + 0.020666155964136124, + 0.017110425978899002, + 0.013855050317943096, + 0.011248858645558357, + 0.011348924599587917, + 0.009029163978993893, + 0.013776866719126701, + 0.005172242410480976, + 0.007742880843579769, + 0.007461422123014927, + 0.005826236680150032, + 0.007491408381611109, + 0.004799774847924709, + 0.004211627412587404, + 0.007279688958078623, + 0.006719834171235561, + 0.007548417896032333, + 0.007446505129337311, + 0.005333181004971266, + 0.005633189808577299, + 0.004385187290608883, + 0.0038263669703155756, + 0.008378027006983757, + 0.004160204436630011, + 0.0039916448295116425, + 0.005828517023473978, + 0.0050480663776397705, + 0.00483787851408124, + 0.003618682036176324, + 0.00384928728453815, + 0.006760560441762209, + 0.0053563774563372135, + 0.0044747754000127316, + 0.0076941680163145065, + 0.007113093044608831, + 0.006655702833086252, + 0.004285189788788557, + 0.004802363459020853, + 0.009972402825951576, + 0.009218188002705574, + 0.009397575631737709, + 0.010971380397677422, + 0.00715242326259613, + 0.008272279053926468, + 0.007602793630212545, + 0.006921618245542049, + 0.008716738782823086, + 0.008138157427310944, + 0.0058618709444999695, + 0.00925123319029808, + 0.008816166780889034, + 0.008634586818516254, + 0.004562932997941971, + 0.0052033280953764915, + 0.005384119693189859, + 0.006773761939257383, + 0.008327780291438103, + 0.008767781779170036, + 0.006303583737462759, + 0.00652838172391057, + 0.004847263917326927, + 0.00527700362727046, + 0.008574648760259151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014221626333892345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019496734021231532 + ], + "opposite_logprob_tokens": [ + -6.251949787139893 + ], + "logprob_reward": -0.0019496734021231532, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019496734021231532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37655460834503174, + 0.13424581289291382, + 0.0832514613866806, + 0.0641934871673584, + 0.05403808876872063, + 0.042464111000299454, + 0.08404918760061264, + 0.048157479614019394, + 0.058878470212221146, + 0.05416730418801308 + ], + "credit_map": [ + 0.04550790414214134, + 0.03660183772444725, + 0.05814608186483383, + 0.04556746408343315, + 0.05063670128583908, + 0.046012260019779205, + 0.04515747353434563, + 0.04892481863498688, + 0.01452920027077198, + 0.013936731033027172, + 0.019620181992650032, + 0.017110349610447884, + 0.02447568252682686, + 0.015429341234266758, + 0.013425629585981369, + 0.015718689188361168, + 0.009459096938371658, + 0.008978231810033321, + 0.012460757978260517, + 0.012248559854924679, + 0.00986659899353981, + 0.007495408412069082, + 0.00914554949849844, + 0.013597247190773487, + 0.007809762842953205, + 0.007898783311247826, + 0.009334062226116657, + 0.006678110919892788, + 0.009669995866715908, + 0.008139683865010738, + 0.005256015341728926, + 0.009407068602740765, + 0.005751355551183224, + 0.005849235691130161, + 0.009164949879050255, + 0.006553030572831631, + 0.005519787780940533, + 0.007240437902510166, + 0.004793060477823019, + 0.009166223928332329, + 0.004117059521377087, + 0.004596540238708258, + 0.006747180595993996, + 0.005400358699262142, + 0.004897823557257652, + 0.0048316465690732, + 0.0043724398128688335, + 0.007501058746129274, + 0.00858510285615921, + 0.00739565072581172, + 0.013089524582028389, + 0.00875014066696167, + 0.013862930238246918, + 0.010621030814945698, + 0.006051959935575724, + 0.015692846849560738, + 0.004179039504379034, + 0.004154168535023928, + 0.008290516212582588, + 0.005214212462306023, + 0.007308948785066605, + 0.0046531702391803265, + 0.00441652862355113, + 0.009940888732671738, + 0.007380675058811903, + 0.006299258675426245, + 0.009447276592254639, + 0.008210585452616215, + 0.008966987021267414, + 0.0062993550673127174, + 0.0065652430057525635, + 0.005709085147827864, + 0.005295894108712673, + 0.0061181820929050446, + 0.00864182598888874, + 0.007515303324908018, + 0.007824084721505642, + 0.005447370000183582, + 0.006921121384948492, + 0.006403519306331873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0019496734021231532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814319893717766 + ], + "opposite_logprob_tokens": [ + -3.2881431579589844 + ], + "logprob_reward": -0.03814319893717766, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814319893717766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2980009615421295, + 0.1085209921002388, + 0.07204648852348328, + 0.07007253915071487, + 0.08555254340171814, + 0.0568537712097168, + 0.04972120746970177, + 0.11970045417547226, + 0.08317629992961884, + 0.05635476112365723 + ], + "credit_map": [ + 0.047485895454883575, + 0.03344297409057617, + 0.053220510482788086, + 0.03814607113599777, + 0.033885106444358826, + 0.029964424669742584, + 0.03032541275024414, + 0.031530559062957764, + 0.020703379064798355, + 0.013579006306827068, + 0.016697751358151436, + 0.009765710681676865, + 0.012198817916214466, + 0.012184900231659412, + 0.00867174006998539, + 0.014719683676958084, + 0.011617463082075119, + 0.00876088161021471, + 0.01116419117897749, + 0.008476059883832932, + 0.008346102200448513, + 0.006576168350875378, + 0.0062002502381801605, + 0.010905379429459572, + 0.0073188431560993195, + 0.008223709650337696, + 0.01170773059129715, + 0.007847225293517113, + 0.013690956868231297, + 0.005193803459405899, + 0.00875288899987936, + 0.007337379734963179, + 0.01053010206669569, + 0.007864118553698063, + 0.012355105020105839, + 0.015361757017672062, + 0.009936362504959106, + 0.008570433594286442, + 0.008839534595608711, + 0.012095130048692226, + 0.007612281944602728, + 0.006298352964222431, + 0.009441209957003593, + 0.007579296827316284, + 0.008206712082028389, + 0.005916311405599117, + 0.0071708387695252895, + 0.004628767725080252, + 0.0062063164077699184, + 0.005795522592961788, + 0.007443251553922892, + 0.007123332004994154, + 0.005145165603607893, + 0.005073457956314087, + 0.004783134441822767, + 0.008151027373969555, + 0.012128505855798721, + 0.011141168884932995, + 0.01836247183382511, + 0.016847234219312668, + 0.016304099932312965, + 0.01635853201150894, + 0.01680520549416542, + 0.011753236874938011, + 0.008100045844912529, + 0.008942642249166965, + 0.012238089926540852, + 0.011424963362514973, + 0.01265487726777792, + 0.010234005749225616, + 0.008098226971924305, + 0.011483447626233101, + 0.007210998795926571, + 0.005084268283098936, + 0.008463054895401001, + 0.008637727238237858, + 0.007774772122502327, + 0.0050421967171132565, + 0.007214605808258057, + 0.00692713912576437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03814319893717766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361173018813133 + ], + "opposite_logprob_tokens": [ + -3.6513612270355225 + ], + "logprob_reward": -0.026361173018813133, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361173018813133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3477220833301544, + 0.11936567723751068, + 0.07484018802642822, + 0.07944683730602264, + 0.0688900575041771, + 0.07686594128608704, + 0.04585310444235802, + 0.04616222903132439, + 0.10231689363718033, + 0.03853699192404747 + ], + "credit_map": [ + 0.0422416627407074, + 0.039665140211582184, + 0.0523473396897316, + 0.051302921026945114, + 0.043657086789608, + 0.03792432323098183, + 0.0372406467795372, + 0.04334297776222229, + 0.013249377720057964, + 0.013191294856369495, + 0.01686589978635311, + 0.016077350825071335, + 0.018886147066950798, + 0.011917958036065102, + 0.014295145869255066, + 0.014882504008710384, + 0.007389893755316734, + 0.00857947114855051, + 0.0110301049426198, + 0.010326755233108997, + 0.010698567144572735, + 0.007992969825863838, + 0.008322242647409439, + 0.010500180535018444, + 0.009083667770028114, + 0.009851659648120403, + 0.012665480375289917, + 0.009699135087430477, + 0.008717145770788193, + 0.008766462095081806, + 0.007363503333181143, + 0.013299781829118729, + 0.00824244599789381, + 0.007532100658863783, + 0.01000634953379631, + 0.01087963581085205, + 0.007768719922751188, + 0.009393949061632156, + 0.007321823388338089, + 0.007745029870420694, + 0.00795056764036417, + 0.009927939623594284, + 0.012252098880708218, + 0.012027798220515251, + 0.00877473782747984, + 0.008804187178611755, + 0.009810288436710835, + 0.007318323943763971, + 0.007705322466790676, + 0.005074487533420324, + 0.006931471638381481, + 0.0057680741883814335, + 0.005251959431916475, + 0.004025368019938469, + 0.004568925593048334, + 0.006527496036142111, + 0.00577619019895792, + 0.004984579514712095, + 0.0067884051240980625, + 0.007374108303338289, + 0.0052496218122541904, + 0.0050744242034852505, + 0.005503314547240734, + 0.005411586258560419, + 0.009451738558709621, + 0.010564512573182583, + 0.016802068799734116, + 0.015293626114726067, + 0.01322846207767725, + 0.013283517211675644, + 0.013369263149797916, + 0.010323706082999706, + 0.005847290623933077, + 0.0041136047802865505, + 0.005955898202955723, + 0.0049008880741894245, + 0.0036285589449107647, + 0.0030074468813836575, + 0.00463656522333622, + 0.006446740590035915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.026361173018813133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-range shot from a horizontal angle focuses firmly on the tuning dial of an old radio. The radio, with its dark brown wooden case, bears subtle scratches and a patina of age. It sits on a wooden tabletop, its background softly blurred, the outlines of a bookshelf barely visible. The space is dimly lit. The dial is a rectangular glass panel, with a yellowed frequency scale and city names printed behind it. A slender red hand motionlessly points to a specific frequency. A soft, warm amber light filters through from within the dial, gradually pulsating from a faint glow to a bright, then slowly fading in a steady, soothing rhythm. Each time the light illuminates, it not only clearly illuminates the details on the dial but also casts dynamic shadows on the wooden case. This rhythmic light, combined with the texture of the old object, creates a profoundly tranquil, warm, and nostalgic atmosphere of contemplation.", + "seed": 32478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_8/index_409_seed_32478_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03432943420484662, + "vlm_logprob": -0.03432943420484662, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03432943420484662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06953976303339005 + ], + "opposite_logprob_tokens": [ + -2.819539785385132 + ], + "logprob_reward": -0.06953976303339005, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06953976303339005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27947521209716797, + 0.15560641884803772, + 0.11779910326004028, + 0.0662534162402153, + 0.08116452395915985, + 0.06777717918157578, + 0.07763213664293289, + 0.07064498960971832, + 0.049135033041238785, + 0.03451196104288101 + ], + "credit_map": [ + 0.021974263712763786, + 0.049340952187776566, + 0.02949240617454052, + 0.018691612407565117, + 0.052909303456544876, + 0.05393107607960701, + 0.013658113777637482, + 0.039477478712797165, + 0.01125540491193533, + 0.034765925258398056, + 0.015433249063789845, + 0.00923958234488964, + 0.0261851754039526, + 0.030242955312132835, + 0.004639494698494673, + 0.02384462207555771, + 0.007904835976660252, + 0.026721462607383728, + 0.011387652717530727, + 0.0057342457585036755, + 0.019249068573117256, + 0.024061623960733414, + 0.003723110305145383, + 0.01901709847152233, + 0.004989061504602432, + 0.014839048497378826, + 0.007386200595647097, + 0.0038928361609578133, + 0.016632970422506332, + 0.007884322665631771, + 0.002599040511995554, + 0.008029937744140625, + 0.005521745886653662, + 0.013943766243755817, + 0.0070863268338143826, + 0.0025075054727494717, + 0.007307341787964106, + 0.023356692865490913, + 0.0022884563077241182, + 0.019152691587805748, + 0.0046212272718548775, + 0.012697847560048103, + 0.005853236187249422, + 0.002831768011674285, + 0.006487631238996983, + 0.01595865562558174, + 0.0036251258570700884, + 0.015701690688729286, + 0.006094175390899181, + 0.017055638134479523, + 0.006141562946140766, + 0.003806617809459567, + 0.006590571720153093, + 0.01782785728573799, + 0.0041655488312244415, + 0.0159501601010561, + 0.004180487245321274, + 0.015770139172673225, + 0.005566261243075132, + 0.0021619799081236124, + 0.0065260762348771095, + 0.018814563751220703, + 0.0019376151030883193, + 0.01568787358701229, + 0.003019982483237982, + 0.007886121980845928, + 0.00409098993986845, + 0.0020071452017873526, + 0.005059271585196257, + 0.014207465574145317, + 0.002313322154805064, + 0.010550737380981445, + 0.0023300559259951115, + 0.006074345204979181, + 0.0031930108088999987, + 0.0012142243795096874, + 0.003817955730482936, + 0.009102296084165573, + 0.0013265055604279041, + 0.007453566882759333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06953976303339005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016098398715257645 + ], + "opposite_logprob_tokens": [ + -4.141098499298096 + ], + "logprob_reward": -0.016098398715257645, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016098398715257645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1985449343919754, + 0.1630367487668991, + 0.06733827292919159, + 0.05794990062713623, + 0.0929282084107399, + 0.06924336403608322, + 0.09893596172332764, + 0.1238780990242958, + 0.04056257754564285, + 0.08758192509412766 + ], + "credit_map": [ + 0.020179279148578644, + 0.042379144579172134, + 0.026619669049978256, + 0.01415919978171587, + 0.027753572911024094, + 0.02810857817530632, + 0.013813983649015427, + 0.025531500577926636, + 0.013606051914393902, + 0.038580529391765594, + 0.017503876239061356, + 0.008314579725265503, + 0.020668594166636467, + 0.03299197554588318, + 0.0072673955000936985, + 0.02410375140607357, + 0.0059579056687653065, + 0.017457658424973488, + 0.006922620814293623, + 0.004024328663945198, + 0.008381742052733898, + 0.011913237161934376, + 0.0031871008686721325, + 0.009493677876889706, + 0.004944708663970232, + 0.01647537760436535, + 0.005706179421395063, + 0.0031673687044531107, + 0.008967659436166286, + 0.008582823909819126, + 0.0025156608317047358, + 0.007590123917907476, + 0.007815900258719921, + 0.01767197623848915, + 0.009961686097085476, + 0.005676139146089554, + 0.008738149888813496, + 0.0217275470495224, + 0.004188028164207935, + 0.017148779705166817, + 0.00594932958483696, + 0.016043975949287415, + 0.008618301711976528, + 0.004449415486305952, + 0.006638198159635067, + 0.011452986858785152, + 0.004140571225434542, + 0.01195058785378933, + 0.007058985065668821, + 0.023579131811857224, + 0.009447549469769001, + 0.003952484577894211, + 0.010164516977965832, + 0.02193428948521614, + 0.0037928116507828236, + 0.019006196409463882, + 0.008594627492129803, + 0.029959436506032944, + 0.012263355776667595, + 0.007066164165735245, + 0.012118976563215256, + 0.027250096201896667, + 0.004512917250394821, + 0.022112516686320305, + 0.0051312875002622604, + 0.010259590111672878, + 0.0037811296060681343, + 0.002342043910175562, + 0.006704986561089754, + 0.004866988863795996, + 0.0027171627152711153, + 0.004759387578815222, + 0.007968494668602943, + 0.025395182892680168, + 0.007792861666530371, + 0.004394311457872391, + 0.011124507524073124, + 0.014324724674224854, + 0.0032032181043177843, + 0.013378629460930824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016098398715257645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644558299332857 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644558299332857, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644558299332857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4311588406562805, + 0.19376753270626068, + 0.10564767569303513, + 0.056684281677007675, + 0.03830542042851448, + 0.03642091155052185, + 0.04328000918030739, + 0.031911443918943405, + 0.030210351571440697, + 0.032613541930913925 + ], + "credit_map": [ + 0.02980257384479046, + 0.07576802372932434, + 0.04969056323170662, + 0.02417639084160328, + 0.03772282600402832, + 0.11007776856422424, + 0.024918189272284508, + 0.0790025144815445, + 0.014654322527348995, + 0.03826258331537247, + 0.018918372690677643, + 0.010668751783668995, + 0.01749274879693985, + 0.04784112796187401, + 0.007786007132381201, + 0.03814362734556198, + 0.007213344797492027, + 0.019705547019839287, + 0.009881841018795967, + 0.004945856053382158, + 0.013045767322182655, + 0.02625616081058979, + 0.004250731319189072, + 0.020348427817225456, + 0.0038559220265597105, + 0.010219537653028965, + 0.00510292686522007, + 0.0026682009920477867, + 0.005705536808818579, + 0.014958715997636318, + 0.001931241131387651, + 0.012242197059094906, + 0.003337293863296509, + 0.009534618817269802, + 0.0037589974235743284, + 0.0017869217554107308, + 0.004378485027700663, + 0.007513025309890509, + 0.0019256980158388615, + 0.006070379633456469, + 0.0033477137330919504, + 0.006446895655244589, + 0.004017239436507225, + 0.002622711006551981, + 0.003431775141507387, + 0.00739953201264143, + 0.0038587613962590694, + 0.005296280607581139, + 0.0025111455470323563, + 0.008633604273200035, + 0.0030279315542429686, + 0.0019009499810636044, + 0.003542596474289894, + 0.01169760711491108, + 0.0019042445346713066, + 0.010061930865049362, + 0.0019708622712641954, + 0.005882714409381151, + 0.0025326840113848448, + 0.0016321629518643022, + 0.0034276240039616823, + 0.008544743061065674, + 0.0011144160525873303, + 0.006806239020079374, + 0.003094604006037116, + 0.0034444485791027546, + 0.0032102521508932114, + 0.0018046066397801042, + 0.0033212509006261826, + 0.007213412318378687, + 0.002319533843547106, + 0.005802245810627937, + 0.0030947464983910322, + 0.009313234128057957, + 0.0031209797598421574, + 0.001605476369149983, + 0.004536032676696777, + 0.005061809439212084, + 0.0021420365665107965, + 0.0037392256781458855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004644558299332857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500687658786774 + ], + "opposite_logprob_tokens": [ + -2.930006980895996 + ], + "logprob_reward": -0.05500687658786774, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500687658786774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4919717311859131, + 0.06365470588207245, + 0.07050621509552002, + 0.05321364104747772, + 0.03617270290851593, + 0.06160520762205124, + 0.1232207790017128, + 0.02819693088531494, + 0.04223953187465668, + 0.029218491166830063 + ], + "credit_map": [ + 0.04486521705985069, + 0.14910389482975006, + 0.057829055935144424, + 0.026675686240196228, + 0.05952530726790428, + 0.07577390968799591, + 0.015060340985655785, + 0.06313832849264145, + 0.004537197761237621, + 0.01540224626660347, + 0.006861857138574123, + 0.0035866601392626762, + 0.008442992344498634, + 0.010699907317757607, + 0.004226734396070242, + 0.009897110983729362, + 0.004870333708822727, + 0.021324610337615013, + 0.00800730474293232, + 0.0033489572815597057, + 0.009602338075637817, + 0.011932543478906155, + 0.0023536637891083956, + 0.009066460654139519, + 0.0042105200700461864, + 0.015188470482826233, + 0.005820204969495535, + 0.0031363072339445353, + 0.00931360200047493, + 0.0071821375750005245, + 0.0025248981546610594, + 0.0058374954387545586, + 0.003223617561161518, + 0.006075489334762096, + 0.004250087775290012, + 0.002263256348669529, + 0.0036996055860072374, + 0.008214980363845825, + 0.002135814866051078, + 0.006309852469712496, + 0.0038486463017761707, + 0.01171878445893526, + 0.004658758174628019, + 0.002760240575298667, + 0.005704833660274744, + 0.016436893492937088, + 0.002941999351605773, + 0.013535053469240665, + 0.006716518197208643, + 0.02309073880314827, + 0.009709488600492477, + 0.003260186407715082, + 0.009804042056202888, + 0.0366157665848732, + 0.003235860262066126, + 0.030788172036409378, + 0.002247414318844676, + 0.007138956803828478, + 0.0034914750140160322, + 0.0011983398580923676, + 0.003773770062252879, + 0.004514440428465605, + 0.001834323862567544, + 0.003998209722340107, + 0.004517769441008568, + 0.008382641710340977, + 0.005040948279201984, + 0.002688077511265874, + 0.0041261035948991776, + 0.008294341154396534, + 0.0022514592856168747, + 0.00693818973377347, + 0.0028555660974234343, + 0.003551863133907318, + 0.0022510667331516743, + 0.001729178475216031, + 0.003028458682820201, + 0.007523481268435717, + 0.001969935605302453, + 0.0063089425675570965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.05500687658786774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635757438838482 + ], + "opposite_logprob_tokens": [ + -3.651357650756836 + ], + "logprob_reward": -0.02635757438838482, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635757438838482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40680861473083496, + 0.13403913378715515, + 0.0924200564622879, + 0.045219019055366516, + 0.020435409620404243, + 0.033081091940402985, + 0.035030703991651535, + 0.04079359397292137, + 0.0869692862033844, + 0.10520310699939728 + ], + "credit_map": [ + 0.03230111300945282, + 0.09514790028333664, + 0.04502540081739426, + 0.019001325592398643, + 0.05886932834982872, + 0.08056436479091644, + 0.01648571528494358, + 0.05941342934966087, + 0.00980798527598381, + 0.029647964984178543, + 0.013660921715199947, + 0.007029610686004162, + 0.018702644854784012, + 0.02836688607931137, + 0.005659635178744793, + 0.02116348221898079, + 0.006034861318767071, + 0.01931086741387844, + 0.009070334956049919, + 0.004295776132494211, + 0.013135078363120556, + 0.021127335727214813, + 0.0035392125137150288, + 0.01590658910572529, + 0.003099419642239809, + 0.010262899100780487, + 0.004669835790991783, + 0.0026927387807518244, + 0.005776500795036554, + 0.009439324960112572, + 0.002105897758156061, + 0.007172400131821632, + 0.001795488758943975, + 0.0028786459006369114, + 0.003219668986275792, + 0.0015532482648268342, + 0.0021787090227007866, + 0.0035835709422826767, + 0.001647112425416708, + 0.0035789639223366976, + 0.0026539857499301434, + 0.0036410351749509573, + 0.003585752099752426, + 0.0015625340165570378, + 0.0037941746413707733, + 0.007397344335913658, + 0.0025851221289485693, + 0.007861142046749592, + 0.0031919158063828945, + 0.012104430235922337, + 0.003939532209187746, + 0.0022350831422954798, + 0.003909618128091097, + 0.003933310974389315, + 0.0018150652758777142, + 0.0039017461240291595, + 0.002100394805893302, + 0.005370773375034332, + 0.003039189148694277, + 0.0017788885161280632, + 0.0027582873590290546, + 0.013713252730667591, + 0.0015032561495900154, + 0.010529550723731518, + 0.005315025802701712, + 0.011584239080548286, + 0.008370103314518929, + 0.004372209310531616, + 0.010491720400750637, + 0.02533661760389805, + 0.004503460600972176, + 0.016995903104543686, + 0.006393650081008673, + 0.021648086607456207, + 0.00920153595507145, + 0.004032669123262167, + 0.007964540272951126, + 0.027629774063825607, + 0.004530468489974737, + 0.02380237728357315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02635757438838482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-range shot from a horizontal angle focuses firmly on the tuning dial of an old radio. The radio, with its dark brown wooden case, bears subtle scratches and a patina of age. It sits on a wooden tabletop, its background softly blurred, the outlines of a bookshelf barely visible. The space is dimly lit. The dial is a rectangular glass panel, with a yellowed frequency scale and city names printed behind it. A slender red hand motionlessly points to a specific frequency. A soft, warm amber light filters through from within the dial, gradually pulsating from a faint glow to a bright, then slowly fading in a steady, soothing rhythm. Each time the light illuminates, it not only clearly illuminates the details on the dial but also casts dynamic shadows on the wooden case. This rhythmic light, combined with the texture of the old object, creates a profoundly tranquil, warm, and nostalgic atmosphere of contemplation.", + "seed": 32479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_8/index_409_seed_32479_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021658957260660826, + "vlm_logprob": -0.021658957260660826, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.021658957260660826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041685499250888824 + ], + "opposite_logprob_tokens": [ + -3.2916855812072754 + ], + "logprob_reward": -0.041685499250888824, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041685499250888824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5308367609977722, + 0.08990906178951263, + 0.04546961560845375, + 0.0470694974064827, + 0.05915224552154541, + 0.031223304569721222, + 0.0827227383852005, + 0.03523612394928932, + 0.05044347420334816, + 0.02793717011809349 + ], + "credit_map": [ + 0.055971041321754456, + 0.04140635207295418, + 0.05717795714735985, + 0.09188597649335861, + 0.09642195701599121, + 0.03239765390753746, + 0.10689235478639603, + 0.04868347570300102, + 0.013301401399075985, + 0.015096508897840977, + 0.01035105437040329, + 0.011844526045024395, + 0.01239257212728262, + 0.005844962317496538, + 0.012181777507066727, + 0.008896264247596264, + 0.007742722984403372, + 0.006471370812505484, + 0.010593865998089314, + 0.004835969768464565, + 0.0037764960434287786, + 0.003173618810251355, + 0.004364336375147104, + 0.00451123621314764, + 0.006662043277174234, + 0.004437030293047428, + 0.005299725569784641, + 0.011007271707057953, + 0.004266875796020031, + 0.004378868266940117, + 0.00580937135964632, + 0.005208311136811972, + 0.007270468398928642, + 0.0062954602763056755, + 0.010221895761787891, + 0.009761876426637173, + 0.00742425536736846, + 0.004602834116667509, + 0.006643909960985184, + 0.006931552197784185, + 0.0041344054043293, + 0.004466237500309944, + 0.004050026647746563, + 0.005123571027070284, + 0.0033578015863895416, + 0.0034251525066792965, + 0.0032646493054926395, + 0.003401463618502021, + 0.010399386286735535, + 0.008374652825295925, + 0.013247187249362469, + 0.010620919987559319, + 0.009792255237698555, + 0.009726322256028652, + 0.010394290089607239, + 0.010167721658945084, + 0.004620423074811697, + 0.00382660492323339, + 0.004584669601172209, + 0.0071710930205881596, + 0.003381884889677167, + 0.0032880189828574657, + 0.004157782532274723, + 0.004205650184303522, + 0.007027591113001108, + 0.006161749362945557, + 0.011115122586488724, + 0.007575828116387129, + 0.005446718540042639, + 0.0032776924781501293, + 0.0055629354901611805, + 0.004275844432413578, + 0.0042419894598424435, + 0.0036988460924476385, + 0.00482725165784359, + 0.0025516278110444546, + 0.0024697589688003063, + 0.002360852900892496, + 0.003405297175049782, + 0.004381546750664711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.041685499250888824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076436190865933895 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.0076436190865933895, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076436190865933895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4680323600769043, + 0.149145245552063, + 0.05145243927836418, + 0.038070905953645706, + 0.075380839407444, + 0.03674834594130516, + 0.0673656091094017, + 0.04915199801325798, + 0.0275657270103693, + 0.03708651661872864 + ], + "credit_map": [ + 0.07166343182325363, + 0.04634130001068115, + 0.1188124418258667, + 0.05472960323095322, + 0.05285894125699997, + 0.024870799854397774, + 0.05039658769965172, + 0.04835924506187439, + 0.023591214790940285, + 0.015692701563239098, + 0.036299046128988266, + 0.018478181213140488, + 0.014471747912466526, + 0.008451303467154503, + 0.015863707289099693, + 0.016297345981001854, + 0.009154221042990685, + 0.006272666156291962, + 0.008843663148581982, + 0.006622419226914644, + 0.005341206677258015, + 0.00368282082490623, + 0.0056893653236329556, + 0.005846076179295778, + 0.007398833520710468, + 0.004500307142734528, + 0.006291178986430168, + 0.005475436337292194, + 0.0038807850796729326, + 0.0026293564587831497, + 0.0037635567132383585, + 0.004131450317800045, + 0.009496175684034824, + 0.006760289426892996, + 0.008778007701039314, + 0.012964649125933647, + 0.008854703046381474, + 0.007451137062162161, + 0.012001159600913525, + 0.009074719622731209, + 0.006181615870445967, + 0.007472159806638956, + 0.005114493891596794, + 0.00381063437089324, + 0.0038145985454320908, + 0.002564158523455262, + 0.004053044132888317, + 0.0037376410327851772, + 0.009047812782227993, + 0.007185536436736584, + 0.013885699212551117, + 0.00958152487874031, + 0.008525765500962734, + 0.004789732396602631, + 0.008423865772783756, + 0.005925673991441727, + 0.0074320645071566105, + 0.006073426455259323, + 0.007078222930431366, + 0.0059771169908344746, + 0.006252916995435953, + 0.003265044651925564, + 0.0063058617524802685, + 0.00676734559237957, + 0.003736890619620681, + 0.004986233543604612, + 0.0040486035868525505, + 0.003543832339346409, + 0.002499534050002694, + 0.003097481792792678, + 0.002576722064986825, + 0.003076429944485426, + 0.005729635711759329, + 0.0056459433399140835, + 0.005795719102025032, + 0.00339917722158134, + 0.003678297158330679, + 0.0031333051156252623, + 0.0033565277699381113, + 0.00634791050106287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0076436190865933895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3718995153903961, + 0.13171574473381042, + 0.061465729027986526, + 0.06227508559823036, + 0.11404702812433243, + 0.059417616575956345, + 0.06233223155140877, + 0.03562978282570839, + 0.03465801477432251, + 0.06655921787023544 + ], + "credit_map": [ + 0.04762224107980728, + 0.03528258204460144, + 0.059507813304662704, + 0.05384580418467522, + 0.057711876928806305, + 0.021724609658122063, + 0.054190538823604584, + 0.042014043778181076, + 0.018801560625433922, + 0.01515210885554552, + 0.012857582420110703, + 0.01910136640071869, + 0.01688782311975956, + 0.011048552580177784, + 0.023249657824635506, + 0.014617101289331913, + 0.008680742233991623, + 0.01161937415599823, + 0.010266164317727089, + 0.006865890230983496, + 0.005685308016836643, + 0.004314086865633726, + 0.007842740043997765, + 0.006191421765834093, + 0.009203110821545124, + 0.007813545875251293, + 0.014402470551431179, + 0.006823363713920116, + 0.007353551220148802, + 0.003539671190083027, + 0.006346681155264378, + 0.006792688276618719, + 0.015733731910586357, + 0.008899250067770481, + 0.01333143375813961, + 0.020718494430184364, + 0.012648462317883968, + 0.010282862931489944, + 0.017995258793234825, + 0.014437531121075153, + 0.009241319261491299, + 0.011966357007622719, + 0.01109185628592968, + 0.007462090812623501, + 0.005340464413166046, + 0.0036112761590629816, + 0.005188543815165758, + 0.005515714175999165, + 0.008370710536837578, + 0.006035387050360441, + 0.007186040747910738, + 0.010309235192835331, + 0.007237312849611044, + 0.006162542849779129, + 0.007817848585546017, + 0.009213155135512352, + 0.005528295878320932, + 0.0044835531152784824, + 0.006783954333513975, + 0.005602966994047165, + 0.003209722228348255, + 0.0027449943590909243, + 0.0027654990553855896, + 0.004510796628892422, + 0.004759983159601688, + 0.00544240977615118, + 0.0048344735987484455, + 0.0066605741158127785, + 0.002664020983502269, + 0.003588058054447174, + 0.0034050641115754843, + 0.0033034307416528463, + 0.009867554530501366, + 0.00950139481574297, + 0.009442885406315327, + 0.007783497683703899, + 0.00767850037664175, + 0.004748906474560499, + 0.009128843434154987, + 0.008407633751630783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033728644251823425 + ], + "opposite_logprob_tokens": [ + -3.40872859954834 + ], + "logprob_reward": -0.033728644251823425, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033728644251823425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3740188181400299, + 0.10176888853311539, + 0.04798628017306328, + 0.07416083663702011, + 0.13168160617351532, + 0.055006496608257294, + 0.03779224678874016, + 0.06995722651481628, + 0.049644943326711655, + 0.0579826682806015 + ], + "credit_map": [ + 0.051052775233983994, + 0.031340617686510086, + 0.07068334519863129, + 0.04594641551375389, + 0.05593537166714668, + 0.020453158766031265, + 0.06114652007818222, + 0.037460606545209885, + 0.014051933772861958, + 0.01336081326007843, + 0.010173913091421127, + 0.016788503155112267, + 0.0075636254623532295, + 0.010800975374877453, + 0.015114367939531803, + 0.013914764858782291, + 0.007963594049215317, + 0.007485734764486551, + 0.004804037977010012, + 0.006961153820157051, + 0.004328599199652672, + 0.004316053353250027, + 0.005050997715443373, + 0.007076110690832138, + 0.009559374302625656, + 0.008852417580783367, + 0.009911713190376759, + 0.01635129377245903, + 0.007291747722774744, + 0.00506417965516448, + 0.009868298657238483, + 0.0072618089616298676, + 0.017398670315742493, + 0.012760793790221214, + 0.018496809527277946, + 0.020870832726359367, + 0.014763899147510529, + 0.014725223183631897, + 0.015985630452632904, + 0.016679758206009865, + 0.0063860658556222916, + 0.00541164493188262, + 0.006928334478288889, + 0.010750509798526764, + 0.007202111184597015, + 0.0038612675853073597, + 0.008304351940751076, + 0.006162208504974842, + 0.0063055590726435184, + 0.006527341902256012, + 0.0040910267271101475, + 0.006316883489489555, + 0.0033216492738574743, + 0.002275118138641119, + 0.0035156747326254845, + 0.005438992753624916, + 0.010521476157009602, + 0.009879819117486477, + 0.00807464960962534, + 0.010260755196213722, + 0.011113710701465607, + 0.004451035521924496, + 0.008997390978038311, + 0.00665839109569788, + 0.006406941916793585, + 0.0053854756988584995, + 0.01100930292159319, + 0.008664190769195557, + 0.0039044010918587446, + 0.0029316460713744164, + 0.004713509231805801, + 0.0066294739954173565, + 0.008085121400654316, + 0.0057700080797076225, + 0.009071215987205505, + 0.005830506794154644, + 0.009419064037501812, + 0.005465702153742313, + 0.009015492163598537, + 0.005325556732714176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.033728644251823425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294012993574142 + ], + "opposite_logprob_tokens": [ + -3.773293972015381 + ], + "logprob_reward": -0.023294012993574142, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294012993574142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3832172751426697, + 0.08301665633916855, + 0.038686685264110565, + 0.07872312515974045, + 0.19855889678001404, + 0.04339722916483879, + 0.02992042899131775, + 0.07236987352371216, + 0.02101362869143486, + 0.05109626427292824 + ], + "credit_map": [ + 0.055601999163627625, + 0.038163769990205765, + 0.10849422216415405, + 0.04704052209854126, + 0.031264085322618484, + 0.03186473622918129, + 0.031772080808877945, + 0.03901582956314087, + 0.01183721050620079, + 0.009449654258787632, + 0.02123301848769188, + 0.012020285241305828, + 0.006226609461009502, + 0.005792247597128153, + 0.00936878751963377, + 0.007088834885507822, + 0.0052653043530881405, + 0.004634147975593805, + 0.0061854287050664425, + 0.0037584779784083366, + 0.004174343775957823, + 0.004440564662218094, + 0.003965690266340971, + 0.0062627182342112064, + 0.010573236271739006, + 0.010343882255256176, + 0.01146172359585762, + 0.01535138487815857, + 0.007969807833433151, + 0.005979517009109259, + 0.00909468624740839, + 0.007948878221213818, + 0.027754252776503563, + 0.016316194087266922, + 0.023513473570346832, + 0.035058606415987015, + 0.02475445717573166, + 0.017535585910081863, + 0.029362231492996216, + 0.024264072999358177, + 0.005727002397179604, + 0.005296797025948763, + 0.0077425590716302395, + 0.006411777343600988, + 0.005279168486595154, + 0.0033629126846790314, + 0.006014405284076929, + 0.0035626054741442204, + 0.004052502103149891, + 0.004233235493302345, + 0.004303030204027891, + 0.00326776341535151, + 0.005568128079175949, + 0.0018614697037264705, + 0.0037264146376401186, + 0.002907881746068597, + 0.008896009996533394, + 0.008209899067878723, + 0.008540586568415165, + 0.01355898380279541, + 0.012119315564632416, + 0.00360541557893157, + 0.010000047273933887, + 0.007439604494720697, + 0.003313209628686309, + 0.003056366927921772, + 0.0023874014150351286, + 0.0023052706383168697, + 0.0028656369540840387, + 0.002061892068013549, + 0.0028590471483767033, + 0.0021648018155246973, + 0.005987522192299366, + 0.004934967029839754, + 0.00903943832963705, + 0.0070264143869280815, + 0.007658509071916342, + 0.003315212670713663, + 0.009212162345647812, + 0.003922030329704285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.023294012993574142 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-range shot from a horizontal angle focuses firmly on the tuning dial of an old radio. The radio, with its dark brown wooden case, bears subtle scratches and a patina of age. It sits on a wooden tabletop, its background softly blurred, the outlines of a bookshelf barely visible. The space is dimly lit. The dial is a rectangular glass panel, with a yellowed frequency scale and city names printed behind it. A slender red hand motionlessly points to a specific frequency. A soft, warm amber light filters through from within the dial, gradually pulsating from a faint glow to a bright, then slowly fading in a steady, soothing rhythm. Each time the light illuminates, it not only clearly illuminates the details on the dial but also casts dynamic shadows on the wooden case. This rhythmic light, combined with the texture of the old object, creates a profoundly tranquil, warm, and nostalgic atmosphere of contemplation.", + "seed": 32480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_8/index_409_seed_32480_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15606967387720944, + "vlm_logprob": -0.15606967387720944, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.15606967387720944, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6411290168762207 + ], + "opposite_logprob_tokens": [ + -0.7661290168762207 + ], + "logprob_reward": -0.6411290168762207, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6411290168762207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39814725518226624, + 0.15532678365707397, + 0.0621156319975853, + 0.039256852120161057, + 0.05224139243364334, + 0.09053387492895126, + 0.05788891017436981, + 0.06336095184087753, + 0.04894978925585747, + 0.03217850998044014 + ], + "credit_map": [ + 0.03188696876168251, + 0.05665642395615578, + 0.06055482476949692, + 0.029999179765582085, + 0.052778374403715134, + 0.07928566634654999, + 0.0314413420855999, + 0.05554449185729027, + 0.012564516626298428, + 0.023348761722445488, + 0.024477360770106316, + 0.011600902304053307, + 0.013535024598240852, + 0.03501737862825394, + 0.010605381801724434, + 0.024177471175789833, + 0.008238722570240498, + 0.007352072279900312, + 0.009799969382584095, + 0.0040071383118629456, + 0.006753391586244106, + 0.010590857826173306, + 0.006534579209983349, + 0.008838905952870846, + 0.004143420606851578, + 0.003563367063179612, + 0.00630944361910224, + 0.003706842428073287, + 0.0036790966987609863, + 0.007506291847676039, + 0.004943490494042635, + 0.00540490448474884, + 0.004921845160424709, + 0.004243944771587849, + 0.007848849520087242, + 0.0049550761468708515, + 0.012149847112596035, + 0.006475814618170261, + 0.004906881600618362, + 0.006739138159900904, + 0.00619683600962162, + 0.007302530109882355, + 0.010507599450647831, + 0.009644757956266403, + 0.02107558585703373, + 0.0131099633872509, + 0.011108191683888435, + 0.01158841885626316, + 0.0058599780313670635, + 0.005086004734039307, + 0.006719064433127642, + 0.006248053628951311, + 0.014361580833792686, + 0.006877930834889412, + 0.0056387740187346935, + 0.007097526453435421, + 0.005959944799542427, + 0.005919999442994595, + 0.00796817522495985, + 0.006217959336936474, + 0.013427138328552246, + 0.01013151090592146, + 0.00659638736397028, + 0.007139844354242086, + 0.005684956442564726, + 0.004143961705267429, + 0.007269704714417458, + 0.004624201916158199, + 0.008040959015488625, + 0.006809723563492298, + 0.004994530230760574, + 0.007381755858659744, + 0.0034415137488394976, + 0.002852837787941098, + 0.005467204377055168, + 0.0024076111149042845, + 0.004318687133491039, + 0.0049195000901818275, + 0.0037113402504473925, + 0.005059819668531418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.6411290168762207 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043120261281728745 + ], + "opposite_logprob_tokens": [ + -3.1681201457977295 + ], + "logprob_reward": -0.043120261281728745, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043120261281728745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32119351625442505, + 0.14447090029716492, + 0.08795982599258423, + 0.062365688383579254, + 0.14545167982578278, + 0.03538460284471512, + 0.031365931034088135, + 0.0684157982468605, + 0.05494361370801926, + 0.04844847321510315 + ], + "credit_map": [ + 0.026819149032235146, + 0.03273138031363487, + 0.05694260075688362, + 0.026283210143446922, + 0.050420474261045456, + 0.05725647881627083, + 0.02275698259472847, + 0.04798322170972824, + 0.012743117287755013, + 0.021084774285554886, + 0.024488206952810287, + 0.011450880207121372, + 0.012783443555235863, + 0.02991114743053913, + 0.008969435468316078, + 0.023039886727929115, + 0.010378070175647736, + 0.009297225624322891, + 0.013863476924598217, + 0.00717466976493597, + 0.014244653284549713, + 0.012617683038115501, + 0.009023981168866158, + 0.011360058560967445, + 0.0059032863937318325, + 0.005666463635861874, + 0.00829403754323721, + 0.0065260292030870914, + 0.013026089407503605, + 0.008679728023707867, + 0.0061561050824820995, + 0.008113952353596687, + 0.01124337688088417, + 0.011551334522664547, + 0.019177483394742012, + 0.017603924497961998, + 0.034403543919324875, + 0.01594480685889721, + 0.018439285457134247, + 0.017087921500205994, + 0.003614692948758602, + 0.004535981919616461, + 0.006517883390188217, + 0.003253200324252248, + 0.003584415651857853, + 0.005480663850903511, + 0.003155038459226489, + 0.005242727696895599, + 0.0030388410668820143, + 0.003494664328172803, + 0.004250447731465101, + 0.002030101604759693, + 0.006683081854134798, + 0.004092058632522821, + 0.0033981662709265947, + 0.004378571640700102, + 0.005259009078145027, + 0.006802771706134081, + 0.00767151964828372, + 0.006472035776823759, + 0.017673809081315994, + 0.00898604467511177, + 0.008025280199944973, + 0.007525325752794743, + 0.008201987482607365, + 0.004504806362092495, + 0.00819421373307705, + 0.005640351679176092, + 0.007091461215168238, + 0.005969132762402296, + 0.007418622728437185, + 0.00792304240167141, + 0.004381448961794376, + 0.0038925933185964823, + 0.008405850268900394, + 0.00620806822553277, + 0.007891513407230377, + 0.004952618386596441, + 0.005853043403476477, + 0.006863338407129049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.043120261281728745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639005314558744 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007639005314558744, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639005314558744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2694578468799591, + 0.18472951650619507, + 0.1357647180557251, + 0.09462114423513412, + 0.12064716219902039, + 0.04122684523463249, + 0.01861105114221573, + 0.05612790584564209, + 0.03561614081263542, + 0.043197717517614365 + ], + "credit_map": [ + 0.02085954137146473, + 0.04292809218168259, + 0.04303921386599541, + 0.02127351239323616, + 0.029888227581977844, + 0.049748729914426804, + 0.022283686324954033, + 0.03943685069680214, + 0.01429110486060381, + 0.02491622045636177, + 0.030034251511096954, + 0.012099606916308403, + 0.022047385573387146, + 0.03817405924201012, + 0.013406849466264248, + 0.029760031029582024, + 0.012947636656463146, + 0.01589907705783844, + 0.020124012604355812, + 0.008388698101043701, + 0.02426217682659626, + 0.02163104899227619, + 0.011058413423597813, + 0.021453645080327988, + 0.006949822884052992, + 0.00876929797232151, + 0.01228366605937481, + 0.008074563927948475, + 0.02040027268230915, + 0.014519750140607357, + 0.011036565527319908, + 0.012587197124958038, + 0.00978959258645773, + 0.008987956680357456, + 0.015457934699952602, + 0.014572171494364738, + 0.027473755180835724, + 0.014438757672905922, + 0.016134947538375854, + 0.013792043551802635, + 0.0026866926345974207, + 0.003321928670629859, + 0.004980407655239105, + 0.003481226274743676, + 0.009881410747766495, + 0.006567209493368864, + 0.004674133844673634, + 0.005633837077766657, + 0.0017868287395685911, + 0.002264190698042512, + 0.0032043701503425837, + 0.002133241156116128, + 0.0022056675516068935, + 0.002596695441752672, + 0.0015886171022430062, + 0.0028314394876360893, + 0.004801566246896982, + 0.004967497196048498, + 0.007543770596385002, + 0.00430351821705699, + 0.013707328587770462, + 0.007850118912756443, + 0.004590934608131647, + 0.008363170549273491, + 0.005226162262260914, + 0.0029604455921798944, + 0.004330928903073072, + 0.003816270735114813, + 0.004980276804417372, + 0.0041463542729616165, + 0.005206303205341101, + 0.0049494002014398575, + 0.003237067023292184, + 0.0037964859511703253, + 0.007205917499959469, + 0.0028685168363153934, + 0.008150633424520493, + 0.007052089553326368, + 0.0040100738406181335, + 0.006876930128782988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.007639005314558744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210239604115486 + ], + "opposite_logprob_tokens": [ + -2.8121023178100586 + ], + "logprob_reward": -0.06210239604115486, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210239604115486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22953768074512482, + 0.08923603594303131, + 0.06478307396173477, + 0.061832115054130554, + 0.055442530661821365, + 0.04649308696389198, + 0.08977480977773666, + 0.1510930359363556, + 0.1528569906949997, + 0.05895064026117325 + ], + "credit_map": [ + 0.025425756350159645, + 0.025356994941830635, + 0.028640158474445343, + 0.02381814830005169, + 0.047065384685993195, + 0.028069835156202316, + 0.022277837619185448, + 0.028883574530482292, + 0.01036654319614172, + 0.013789242133498192, + 0.012931494042277336, + 0.007115093991160393, + 0.012219605967402458, + 0.011766230687499046, + 0.009040584787726402, + 0.012007245793938637, + 0.004885400645434856, + 0.005729366093873978, + 0.008487311191856861, + 0.004691243637353182, + 0.01786045730113983, + 0.008015941828489304, + 0.0070030284114181995, + 0.008110322058200836, + 0.00747675122693181, + 0.005519963800907135, + 0.008225128054618835, + 0.0045959483832120895, + 0.010615449398756027, + 0.009728414937853813, + 0.006753609050065279, + 0.008916853927075863, + 0.004356112331151962, + 0.005959921982139349, + 0.006696375086903572, + 0.00491288723424077, + 0.010783625766634941, + 0.00849882885813713, + 0.007878844626247883, + 0.006355935707688332, + 0.0054355282336473465, + 0.004269538912922144, + 0.006734144873917103, + 0.005548418499529362, + 0.0059922668151557446, + 0.005933028645813465, + 0.006677832920104265, + 0.005902324803173542, + 0.014810090884566307, + 0.007170235272496939, + 0.013785157352685928, + 0.006547499913722277, + 0.011560550890862942, + 0.015452847816050053, + 0.00760674150660634, + 0.012841683812439442, + 0.01487867720425129, + 0.014246544800698757, + 0.01847023516893387, + 0.015990346670150757, + 0.031505558639764786, + 0.022560741752386093, + 0.017081132158637047, + 0.01635979674756527, + 0.013135932385921478, + 0.013112618587911129, + 0.019077599048614502, + 0.017099648714065552, + 0.03483780100941658, + 0.018920205533504486, + 0.019352393224835396, + 0.017320791259407997, + 0.00508755212649703, + 0.008657663129270077, + 0.010228321887552738, + 0.005306582897901535, + 0.007946130819618702, + 0.007397817447781563, + 0.00466864462941885, + 0.009657928720116615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.06210239604115486 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of the tuning dial of an old radio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the radio case dark brown wood with visible scratches and signs of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dial a rectangular glass panel with a yellowed frequency scale and city names?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a slender red hand point motionlessly to a specific frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a soft warm amber light inside the dial pulse rhythmically from faint to bright and then fade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635768987238407 + ], + "opposite_logprob_tokens": [ + -3.651357650756836 + ], + "logprob_reward": -0.02635768987238407, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635768987238407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2202032208442688, + 0.08840187638998032, + 0.11074259132146835, + 0.06369239091873169, + 0.04792243242263794, + 0.08597100526094437, + 0.06083713844418526, + 0.0805993601679802, + 0.18470728397369385, + 0.056922730058431625 + ], + "credit_map": [ + 0.017644179984927177, + 0.021636145189404488, + 0.028333887457847595, + 0.01937878504395485, + 0.04579140990972519, + 0.0423407107591629, + 0.01545944157987833, + 0.02961866743862629, + 0.007589332759380341, + 0.014199970290064812, + 0.013331758789718151, + 0.005711676552891731, + 0.012869558297097683, + 0.017382122576236725, + 0.005422372370958328, + 0.01189508382230997, + 0.008961028419435024, + 0.009036029689013958, + 0.01408443320542574, + 0.009136435575783253, + 0.028993630781769753, + 0.015305643901228905, + 0.011609762907028198, + 0.013615625910460949, + 0.006229359656572342, + 0.005352675449103117, + 0.009521818719804287, + 0.006953597534447908, + 0.010262088850140572, + 0.007738160435110331, + 0.00857219286262989, + 0.009062495082616806, + 0.003864027326926589, + 0.005824629217386246, + 0.0062734829261898994, + 0.004921876825392246, + 0.00783773884177208, + 0.007072060368955135, + 0.004661635495722294, + 0.0074669793248176575, + 0.006824595853686333, + 0.007687736768275499, + 0.010251052677631378, + 0.011283688247203827, + 0.01737602800130844, + 0.01159830205142498, + 0.01077396422624588, + 0.010175635106861591, + 0.004874869715422392, + 0.004808831959962845, + 0.007799304556101561, + 0.00560313044115901, + 0.014289778657257557, + 0.00789260771125555, + 0.007406304590404034, + 0.008162311278283596, + 0.005763971246778965, + 0.006922921631485224, + 0.010287673212587833, + 0.00710919126868248, + 0.019636517390608788, + 0.010751639492809772, + 0.010000314563512802, + 0.01012712623924017, + 0.0128854438662529, + 0.015371736139059067, + 0.02351006306707859, + 0.0159287266433239, + 0.04497137665748596, + 0.026135355234146118, + 0.022613754495978355, + 0.023290827870368958, + 0.004444707650691271, + 0.005132071208208799, + 0.006504699122160673, + 0.008980502374470234, + 0.011563395150005817, + 0.005584187805652618, + 0.00805073045194149, + 0.006662433501332998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.02635768987238407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_9/index_tensor([3097, 3097])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_9/index_tensor([3097, 3097])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb1e67e001a2f9c207466938bdd9a97ff321a929 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_9/index_tensor([3097, 3097])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, wide-angle shot captures the surreal and expansive transformation of Millennium City Park, rendered entirely in the distinctive artistic style of a David Hockney painting. The scene is defined by clean lines, planes of vibrant, saturated color, and a slightly distorted, flattened perspective. A traditional Chinese pagoda, with its ornate terracotta tile roof and an ornate pavilion with imperial yellow accents, stands majestically above the surface of a vast, incredibly clear lake that fills the park's basin. Beneath the tranquil, azure waters, the underwater forms of tranquil gardens, intricate stone paths, and manicured trees are clearly visible. The central element is the gentle, rhythmic pulsation of the water itself; the water level rises steadily and smoothly, clinging to the pavilion's wooden latticework, then gracefully recedes, allowing the building's surfaces to shimmer in bright, shadowless sunlight against the azure sky. This constant, hypnotic fluctuation creates an atmosphere of dreamlike tranquility, a beautiful yet melancholic natural landscape restored to its man-made splendor in brilliant, serene colors.", + "seed": 27246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_9/index_3097_seed_27246_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07841506376862525, + "vlm_logprob": -0.07841506376862525, + "vlm_yes_no_margin": 3.3749999523162844, + "vlm_token_credit": -0.07841506376862525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023788632825016975 + ], + "opposite_logprob_tokens": [ + -3.8987886905670166 + ], + "logprob_reward": -0.023788632825016975, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023788632825016975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27242136001586914, + 0.08703571557998657, + 0.05461554601788521, + 0.049723412841558456, + 0.1143111139535904, + 0.07581217586994171, + 0.2137613296508789, + 0.060104336589574814, + 0.02571449801325798, + 0.04650050029158592 + ], + "credit_map": [ + 0.02829955518245697, + 0.0659845620393753, + 0.051576148718595505, + 0.02259339950978756, + 0.020548824220895767, + 0.033059027045965195, + 0.02017485536634922, + 0.030185000970959663, + 0.007301343139261007, + 0.024729352444410324, + 0.022420652210712433, + 0.004775607958436012, + 0.005147400312125683, + 0.0077437604777514935, + 0.004959688521921635, + 0.009957914240658283, + 0.006103658117353916, + 0.007594243157655001, + 0.00642229663208127, + 0.005855613853782415, + 0.008471254259347916, + 0.008527085185050964, + 0.005887696053832769, + 0.005753697827458382, + 0.006701774429529905, + 0.005873825866729021, + 0.009748533368110657, + 0.005427927244454622, + 0.0036114125978201628, + 0.0061103967018425465, + 0.004352600779384375, + 0.007896943017840385, + 0.015130637213587761, + 0.018127596005797386, + 0.016860686242580414, + 0.01247992180287838, + 0.007798532955348492, + 0.01297911535948515, + 0.009402929805219173, + 0.02153169736266136, + 0.00803385954350233, + 0.016424676403403282, + 0.0077991606667637825, + 0.011210277676582336, + 0.0063948798924684525, + 0.008923441171646118, + 0.007346062455326319, + 0.00967982318252325, + 0.025739435106515884, + 0.024041224271059036, + 0.03456054627895355, + 0.03039063699543476, + 0.015427635982632637, + 0.01977342925965786, + 0.02078976295888424, + 0.04303866624832153, + 0.004036413040012121, + 0.015379827469587326, + 0.004380076192319393, + 0.019035518169403076, + 0.0030281643848866224, + 0.003908480983227491, + 0.0027982499450445175, + 0.007537608966231346, + 0.0027922617737203836, + 0.002838158281520009, + 0.0022266660816967487, + 0.006877015344798565, + 0.001464083674363792, + 0.0024085131008177996, + 0.001801700214855373, + 0.00530609954148531, + 0.005256752949208021, + 0.012216486036777496, + 0.003984205890446901, + 0.005765232257544994, + 0.0035616292152553797, + 0.005446175578981638, + 0.004088338930159807, + 0.006181680131703615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.023788632825016975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599357783794403 + ], + "opposite_logprob_tokens": [ + -3.895599365234375 + ], + "logprob_reward": -0.020599357783794403, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599357783794403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28864073753356934, + 0.09575867652893066, + 0.11204267293214798, + 0.0635577142238617, + 0.0646882876753807, + 0.053409650921821594, + 0.07912229746580124, + 0.05745336785912514, + 0.04798303544521332, + 0.137343630194664 + ], + "credit_map": [ + 0.028670212253928185, + 0.06138346344232559, + 0.046564407646656036, + 0.023806259036064148, + 0.03110705129802227, + 0.03549407422542572, + 0.026888156309723854, + 0.034727104008197784, + 0.010174261406064034, + 0.013974932953715324, + 0.010993888601660728, + 0.009606159292161465, + 0.01391301117837429, + 0.012028768658638, + 0.010700859129428864, + 0.014366772957146168, + 0.013580329716205597, + 0.01295398734509945, + 0.011678844690322876, + 0.010644864290952682, + 0.020743412896990776, + 0.01897469162940979, + 0.012002435512840748, + 0.011464090086519718, + 0.006689902860671282, + 0.009433274157345295, + 0.009087925776839256, + 0.005436131730675697, + 0.00863607320934534, + 0.009584423154592514, + 0.004906916990876198, + 0.00978305283933878, + 0.00613038893789053, + 0.01708977296948433, + 0.0041146669536828995, + 0.005448364187031984, + 0.007841043174266815, + 0.009477161802351475, + 0.005532810930162668, + 0.009054074995219707, + 0.004938007332384586, + 0.01296901609748602, + 0.005033727735280991, + 0.005867087747901678, + 0.005931952502578497, + 0.005827324464917183, + 0.004605778027325869, + 0.008236751891672611, + 0.009531832300126553, + 0.011623675003647804, + 0.011495410464704037, + 0.01089763268828392, + 0.005940473172813654, + 0.006943644490092993, + 0.005890123546123505, + 0.01679949276149273, + 0.003334494074806571, + 0.010361320339143276, + 0.00476636178791523, + 0.01805015280842781, + 0.004603833891451359, + 0.00418988848105073, + 0.004224094562232494, + 0.007923212833702564, + 0.0044708289206027985, + 0.00456588389351964, + 0.003237077733501792, + 0.010430500842630863, + 0.006355663295835257, + 0.006606180686503649, + 0.0039630914106965065, + 0.008353802375495434, + 0.010912662371993065, + 0.029998797923326492, + 0.008171364665031433, + 0.03432789072394371, + 0.012126635760068893, + 0.011791547760367393, + 0.010002522729337215, + 0.02001218870282173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.020599357783794403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189646303653717 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189646303653717, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018189646303653717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20811717212200165, + 0.07679753750562668, + 0.07655224204063416, + 0.07111982256174088, + 0.06473146378993988, + 0.14051462709903717, + 0.18724502623081207, + 0.04414663463830948, + 0.017788074910640717, + 0.11298743635416031 + ], + "credit_map": [ + 0.02548515982925892, + 0.029587728902697563, + 0.02993251010775566, + 0.021949920803308487, + 0.021971551701426506, + 0.030219057574868202, + 0.01905248872935772, + 0.029918765649199486, + 0.007434711791574955, + 0.013911287300288677, + 0.021272676065564156, + 0.00598675524815917, + 0.004305408801883459, + 0.005574258975684643, + 0.0054184612818062305, + 0.012893971055746078, + 0.007989432662725449, + 0.011452948674559593, + 0.012150559574365616, + 0.008500820957124233, + 0.009797405451536179, + 0.009165526367723942, + 0.008432324044406414, + 0.00906322617083788, + 0.00990336388349533, + 0.010393992997705936, + 0.014051019214093685, + 0.007365437690168619, + 0.006085637025535107, + 0.008637893944978714, + 0.005657565779983997, + 0.00902490597218275, + 0.006244217045605183, + 0.021430257707834244, + 0.005302255507558584, + 0.005142465233802795, + 0.003933139145374298, + 0.0064504435285925865, + 0.0034702306147664785, + 0.012758464552462101, + 0.013878594152629375, + 0.04519517719745636, + 0.013220701366662979, + 0.016138171777129173, + 0.008750955574214458, + 0.014002426527440548, + 0.009391174651682377, + 0.0199374258518219, + 0.020817479118704796, + 0.036612290889024734, + 0.024513479322195053, + 0.02245185151696205, + 0.012849063612520695, + 0.01845596358180046, + 0.01384876947849989, + 0.037696126848459244, + 0.00430120388045907, + 0.009888367727398872, + 0.004171264357864857, + 0.01082694623619318, + 0.002818414941430092, + 0.004025888163596392, + 0.002997249597683549, + 0.005117298569530249, + 0.0029839107301086187, + 0.0029314851853996515, + 0.0017764773219823837, + 0.001875618938356638, + 0.0017491275211796165, + 0.0024076616391539574, + 0.0018684524111449718, + 0.0021953419782221317, + 0.008036601357161999, + 0.02221183106303215, + 0.00707222567871213, + 0.03322836756706238, + 0.0060649472288787365, + 0.00787363387644291, + 0.00583385955542326, + 0.022665970027446747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.018189646303653717 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134286105632782 + ], + "opposite_logprob_tokens": [ + -1.3134286403656006 + ], + "logprob_reward": -0.3134286105632782, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134286105632782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12019506096839905, + 0.07276596128940582, + 0.09878627955913544, + 0.06936193257570267, + 0.11514101922512054, + 0.23773764073848724, + 0.12554797530174255, + 0.04624514654278755, + 0.04397140443325043, + 0.07024763524532318 + ], + "credit_map": [ + 0.014089670032262802, + 0.015711551532149315, + 0.014558689668774605, + 0.015776624903082848, + 0.016140121966600418, + 0.01797664165496826, + 0.010999237187206745, + 0.014942512847483158, + 0.006741237360984087, + 0.0076722390949726105, + 0.015783339738845825, + 0.007145590148866177, + 0.00790228508412838, + 0.008977835066616535, + 0.0072202179580926895, + 0.01132320798933506, + 0.01231701672077179, + 0.02426174096763134, + 0.011232859455049038, + 0.011675257235765457, + 0.007666397374123335, + 0.012975282035768032, + 0.00729139382019639, + 0.011366333812475204, + 0.006320094224065542, + 0.012273336760699749, + 0.012448602356016636, + 0.008384821936488152, + 0.0086888512596488, + 0.00779341347515583, + 0.005987238604575396, + 0.007465575821697712, + 0.011535321362316608, + 0.037276484072208405, + 0.007914142683148384, + 0.006845727097243071, + 0.010250133462250233, + 0.013876646757125854, + 0.008009920828044415, + 0.019432641565799713, + 0.024441801011562347, + 0.08251671493053436, + 0.021425114944577217, + 0.020708704367280006, + 0.01669609732925892, + 0.023068979382514954, + 0.015681130811572075, + 0.03319908678531647, + 0.015269797295331955, + 0.012209600768983364, + 0.018361534923315048, + 0.019676508381962776, + 0.012709544040262699, + 0.016050908714532852, + 0.012617970816791058, + 0.01865210011601448, + 0.003371255239471793, + 0.007434373255819082, + 0.005375780630856752, + 0.005432841833680868, + 0.009129608981311321, + 0.004812278319150209, + 0.006264607422053814, + 0.004424398764967918, + 0.005603476893156767, + 0.006375195924192667, + 0.004198553040623665, + 0.01053044106811285, + 0.0024787746369838715, + 0.005083237309008837, + 0.0029099530074745417, + 0.006791766732931137, + 0.006452922709286213, + 0.009040791541337967, + 0.005887683480978012, + 0.018644221127033234, + 0.004451444372534752, + 0.006678207777440548, + 0.004958757199347019, + 0.014133602380752563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.3134286105632782 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069071367383003 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.016069071367383003, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069071367383003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1574207842350006, + 0.06542404741048813, + 0.11778516322374344, + 0.04483247920870781, + 0.031036831438541412, + 0.04110264778137207, + 0.059732381254434586, + 0.3735233247280121, + 0.05284249782562256, + 0.05629979446530342 + ], + "credit_map": [ + 0.02454822137951851, + 0.02295564115047455, + 0.027492769062519073, + 0.013740558177232742, + 0.011494494043290615, + 0.022742729634046555, + 0.014345974661409855, + 0.020100390538573265, + 0.006242783740162849, + 0.012424714863300323, + 0.015645509585738182, + 0.007062003016471863, + 0.004266888368874788, + 0.005817627999931574, + 0.004888160619884729, + 0.009076358750462532, + 0.013293915428221226, + 0.03778264299035072, + 0.01369602233171463, + 0.009421154856681824, + 0.008289643563330173, + 0.01293993555009365, + 0.008553345687687397, + 0.013808499090373516, + 0.004161624703556299, + 0.008496015332639217, + 0.007689120713621378, + 0.003833228023722768, + 0.003469643648713827, + 0.006560126785188913, + 0.004081187304109335, + 0.006541533395648003, + 0.003099749330431223, + 0.008501807227730751, + 0.0037038668524473906, + 0.002928153146058321, + 0.002223048359155655, + 0.003122220980003476, + 0.0019985416438430548, + 0.005459444131702185, + 0.003534104209393263, + 0.0066563403233885765, + 0.0049587432295084, + 0.006534065119922161, + 0.004973395727574825, + 0.005353992339223623, + 0.004312776494771242, + 0.004779231734573841, + 0.006933202967047691, + 0.005756020080298185, + 0.011758450418710709, + 0.009090971201658249, + 0.0038155964575707912, + 0.006652791053056717, + 0.004753719549626112, + 0.010971629992127419, + 0.019710185006260872, + 0.0797903835773468, + 0.03264018893241882, + 0.1384715586900711, + 0.012013568542897701, + 0.014944123104214668, + 0.015394111163914204, + 0.06055919826030731, + 0.003617068286985159, + 0.0042418222874403, + 0.006099130026996136, + 0.01691848412156105, + 0.0025967187248170376, + 0.003289248328655958, + 0.002402404323220253, + 0.013677619397640228, + 0.00475322175770998, + 0.013387276791036129, + 0.004158936906605959, + 0.017092114314436913, + 0.0023274337872862816, + 0.0036686428356915712, + 0.0030847350135445595, + 0.007827428169548512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.016069071367383003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot captures the surreal and expansive transformation of Millennium City Park, rendered entirely in the distinctive artistic style of a David Hockney painting. The scene is defined by clean lines, planes of vibrant, saturated color, and a slightly distorted, flattened perspective. A traditional Chinese pagoda, with its ornate terracotta tile roof and an ornate pavilion with imperial yellow accents, stands majestically above the surface of a vast, incredibly clear lake that fills the park's basin. Beneath the tranquil, azure waters, the underwater forms of tranquil gardens, intricate stone paths, and manicured trees are clearly visible. The central element is the gentle, rhythmic pulsation of the water itself; the water level rises steadily and smoothly, clinging to the pavilion's wooden latticework, then gracefully recedes, allowing the building's surfaces to shimmer in bright, shadowless sunlight against the azure sky. This constant, hypnotic fluctuation creates an atmosphere of dreamlike tranquility, a beautiful yet melancholic natural landscape restored to its man-made splendor in brilliant, serene colors.", + "seed": 27247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_9/index_3097_seed_27247_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025824072398245336, + "vlm_logprob": -0.025824072398245336, + "vlm_yes_no_margin": 4.299999952316284, + "vlm_token_credit": -0.025824072398245336, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006155224516987801 + ], + "opposite_logprob_tokens": [ + -5.381155014038086 + ], + "logprob_reward": -0.006155224516987801, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006155224516987801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9243955612182617, + 0.014582817442715168, + 0.005996066145598888, + 0.010802999138832092, + 0.0092551838606596, + 0.01595095545053482, + 0.006384237203747034, + 0.004986325744539499, + 0.0033285163808614016, + 0.004317351616919041 + ], + "credit_map": [ + 0.02544366754591465, + 0.02999509871006012, + 0.0025919920299202204, + 0.0034926552325487137, + 0.44342508912086487, + 0.40747860074043274, + 0.006419493816792965, + 0.0055488464422523975, + 0.0006473054527305067, + 0.004040701314806938, + 0.0005732927820645273, + 0.0004885264206677675, + 0.006087624467909336, + 0.002004097681492567, + 0.00043141740025021136, + 0.00030985032208263874, + 0.0004652927163988352, + 0.0018153010169044137, + 0.000467656587716192, + 0.0004335828125476837, + 0.0014947158051654696, + 0.0005509370821528137, + 0.0003733912017196417, + 0.0003951884573325515, + 0.001998298801481724, + 0.0012406653258949518, + 0.0012102012988179922, + 0.0009613896836526692, + 0.0012956871651113033, + 0.002534399041905999, + 0.0008231670944951475, + 0.0007391893304884434, + 0.0010111830197274685, + 0.0011459981324151158, + 0.0008115287637338042, + 0.0013546687550842762, + 0.0011775316670536995, + 0.0013845606008544564, + 0.0013742082519456744, + 0.0009955036221072078, + 0.001623977324925363, + 0.001687571289949119, + 0.002306457841768861, + 0.0025134508032351732, + 0.0020567281171679497, + 0.0015859509585425258, + 0.002463878132402897, + 0.0017129391198977828, + 0.000586146954447031, + 0.00045227250666357577, + 0.0008588184136897326, + 0.001012407592497766, + 0.0005637353169731796, + 0.0007232265779748559, + 0.0010116149205714464, + 0.0011760141933336854, + 0.00043807391193695366, + 0.0006162344361655414, + 0.0005527900066226721, + 0.0006541242473758757, + 0.0010053168516606092, + 0.0006270723533816636, + 0.0006129755638539791, + 0.0004797378496732563, + 0.00022921930940356106, + 0.000282876193523407, + 0.0005256686126813293, + 0.0005267690867185593, + 0.00039384327828884125, + 0.0004245519812684506, + 0.00043219889630563557, + 0.0005133886006660759, + 0.0004222616262268275, + 0.00045964098535478115, + 0.0005962030845694244, + 0.0006892802775837481, + 0.0003739372477866709, + 0.0004569966986309737, + 0.0006834528758190572, + 0.0006355787045322359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.006155224516987801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798170067369938 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009798170067369938, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798170067369938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9375216364860535, + 0.020222272723913193, + 0.008068963885307312, + 0.0067513734102249146, + 0.0035162887070327997, + 0.005356364883482456, + 0.007563257124274969, + 0.006278303451836109, + 0.0022985837422311306, + 0.002422898542135954 + ], + "credit_map": [ + 0.023329060524702072, + 0.03961522504687309, + 0.003362206509336829, + 0.004846230149269104, + 0.45326894521713257, + 0.4001445472240448, + 0.006805311422795057, + 0.006150181870907545, + 0.0008971320348791778, + 0.007627403363585472, + 0.0008381033549085259, + 0.0014107703464105725, + 0.005050841253250837, + 0.002721829107031226, + 0.0009311405592598021, + 0.0007450521807186306, + 0.00024071114603430033, + 0.00288850418291986, + 0.00039612496038898826, + 0.000418781244661659, + 0.003068124409765005, + 0.0003964848001487553, + 0.0003685109259095043, + 0.0002917217498179525, + 0.0008343678782694042, + 0.000748654652852565, + 0.0007172693149186671, + 0.0006754540372639894, + 0.001315419445745647, + 0.0014380301581695676, + 0.0005673397099599242, + 0.0004548383003566414, + 0.0004439380718395114, + 0.00042922570719383657, + 0.00033896631794050336, + 0.0005274104187265038, + 0.00043238693615421653, + 0.00045542497537098825, + 0.0005659366142936051, + 0.0003229996364098042, + 0.0006040846928954124, + 0.0007448327378369868, + 0.0006741195684298873, + 0.0005933266947977245, + 0.0008007344440557063, + 0.0007121397648006678, + 0.0006256020278669894, + 0.0006015243707224727, + 0.000698244315572083, + 0.000967857544310391, + 0.0007868378306739032, + 0.0010797426803037524, + 0.001588254002854228, + 0.0008464285056106746, + 0.0007498592603951693, + 0.0008460328099317849, + 0.0005790836876258254, + 0.0009929685620591044, + 0.0006664155516773462, + 0.0007054570596665144, + 0.0011556316167116165, + 0.000673255359288305, + 0.0007608412997797132, + 0.0007446500821970403, + 0.00019216824148315936, + 0.0002528762852307409, + 0.0003165197267662734, + 0.0003836129035335034, + 0.00023632000375073403, + 0.00030524199246428907, + 0.0002700621262192726, + 0.00034178222995251417, + 0.0001697375118965283, + 0.00026541034458205104, + 0.0004035394231323153, + 0.0003071256505791098, + 0.00039262796053662896, + 0.00023148288892116398, + 0.00035756686702370644, + 0.0002954080409836024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009798170067369938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953673273324966 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.005953673273324966, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953673273324966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6901697516441345, + 0.06531218439340591, + 0.022604623809456825, + 0.02904619835317135, + 0.02821364626288414, + 0.019266242161393166, + 0.02775881625711918, + 0.06544402241706848, + 0.0438668467104435, + 0.008317704312503338 + ], + "credit_map": [ + 0.017555253580212593, + 0.021541349589824677, + 0.007007600273936987, + 0.009298414923250675, + 0.31887930631637573, + 0.2943243086338043, + 0.009530472569167614, + 0.012033131904900074, + 0.0029278879519551992, + 0.022558683529496193, + 0.0043018897995352745, + 0.0030125374905765057, + 0.015080044977366924, + 0.010493401437997818, + 0.003238318022340536, + 0.0036994207184761763, + 0.001967125339433551, + 0.005493941716849804, + 0.0019692089408636093, + 0.001738322083838284, + 0.005626726895570755, + 0.0019325243774801493, + 0.0025261545088142157, + 0.0013506190152838826, + 0.005241021513938904, + 0.002990567358210683, + 0.0030998121947050095, + 0.0027551602106541395, + 0.004333615303039551, + 0.0061579737812280655, + 0.0020301425829529762, + 0.0024379054084420204, + 0.0028676758520305157, + 0.0031266731675714254, + 0.002695912029594183, + 0.003230469999834895, + 0.006283886265009642, + 0.0037921913899481297, + 0.003739702282473445, + 0.002477135043591261, + 0.0022013611160218716, + 0.0028323789592832327, + 0.002416930627077818, + 0.002048910828307271, + 0.002367512322962284, + 0.0026554837822914124, + 0.0019512663129717112, + 0.0027923991438001394, + 0.002564399503171444, + 0.00370811577886343, + 0.0030168641824275255, + 0.005269414745271206, + 0.0028395550325512886, + 0.00272613437846303, + 0.004129257518798113, + 0.003505075117573142, + 0.006064583081752062, + 0.010238280519843102, + 0.008023574016988277, + 0.008025256916880608, + 0.010327503085136414, + 0.006113029085099697, + 0.008603325113654137, + 0.008048470132052898, + 0.003134733298793435, + 0.0044843279756605625, + 0.006187813822180033, + 0.008757062256336212, + 0.004818248562514782, + 0.004692199174314737, + 0.005758219864219427, + 0.006034242920577526, + 0.0008390998700633645, + 0.0009414033265784383, + 0.001248909393325448, + 0.0008431168389506638, + 0.0009161994094029069, + 0.0011261559557169676, + 0.0009024993050843477, + 0.001500320970080793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.005953673273324966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902844786643982 + ], + "opposite_logprob_tokens": [ + -2.4640283584594727 + ], + "logprob_reward": -0.08902844786643982, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902844786643982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8499746322631836, + 0.057887520641088486, + 0.01913009025156498, + 0.018754465505480766, + 0.014898847788572311, + 0.011739044450223446, + 0.008981107734143734, + 0.009809128940105438, + 0.005142138805240393, + 0.003682954702526331 + ], + "credit_map": [ + 0.023156458511948586, + 0.03003680147230625, + 0.0036526662297546864, + 0.004728076979517937, + 0.40723225474357605, + 0.3681330680847168, + 0.0060762157663702965, + 0.006959051359444857, + 0.0015884782187640667, + 0.024852881208062172, + 0.0012853513471782207, + 0.0016352856764569879, + 0.02342517301440239, + 0.002783863339573145, + 0.0011790135176852345, + 0.0011374737368896604, + 0.0007169721066020429, + 0.007029786240309477, + 0.0006512990221381187, + 0.000864258618094027, + 0.007578067481517792, + 0.0010657026432454586, + 0.0005792395095340908, + 0.0006447656196542084, + 0.003324708202853799, + 0.0018551526591181755, + 0.002289149444550276, + 0.0018973839469254017, + 0.0025548639241605997, + 0.004364838358014822, + 0.0010903557995334268, + 0.001378011773340404, + 0.0014834403991699219, + 0.002222958020865917, + 0.00193193054292351, + 0.002065831795334816, + 0.0023784805089235306, + 0.0012798634124919772, + 0.0019120336510241032, + 0.001624309690669179, + 0.001312850508838892, + 0.0011228462681174278, + 0.0021773416083306074, + 0.0015772065380588174, + 0.0013535430189222097, + 0.0015276859048753977, + 0.0012794408248737454, + 0.0013881297782063484, + 0.0008501726551912725, + 0.0009925166377797723, + 0.0010893021244555712, + 0.0011799747589975595, + 0.0015812538331374526, + 0.00122225028462708, + 0.0011141145369037986, + 0.0009515230194665492, + 0.0009716332424432039, + 0.0016759643331170082, + 0.0010197876254096627, + 0.0010784713085740805, + 0.001811364432796836, + 0.0011434174375608563, + 0.00088978506391868, + 0.0012187062529847026, + 0.00045006838627159595, + 0.0004383520281407982, + 0.0007500961655750871, + 0.0010297003900632262, + 0.00043713970808312297, + 0.0005570369539782405, + 0.0007106844568625093, + 0.0007690608035773039, + 0.00028311205096542835, + 0.0004884840454906225, + 0.0005768404807895422, + 0.0003988273674622178, + 0.0004870733537245542, + 0.0004934497992508113, + 0.00041056956979446113, + 0.0005445982678793371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08902844786643982 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818484626710415 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.01818484626710415, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818484626710415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8866877555847168, + 0.03599599748849869, + 0.015566980466246605, + 0.009429649449884892, + 0.01092150341719389, + 0.01393958367407322, + 0.009983758442103863, + 0.006574880797415972, + 0.008170978166162968, + 0.002728869905695319 + ], + "credit_map": [ + 0.025290947407484055, + 0.029279861599206924, + 0.0030899718403816223, + 0.0037439430598169565, + 0.42413437366485596, + 0.3901468813419342, + 0.006180489901453257, + 0.00482126185670495, + 0.0019004108617082238, + 0.011801331304013729, + 0.0022618521470576525, + 0.0030376261565834284, + 0.007927941158413887, + 0.004715784452855587, + 0.00239562033675611, + 0.0019554300233721733, + 0.0009356861701235175, + 0.004101642407476902, + 0.0010282519506290555, + 0.0016663516871631145, + 0.004059514030814171, + 0.0010169475572183728, + 0.0017286406364291906, + 0.0010299446294084191, + 0.0015606849920004606, + 0.0012729294830933213, + 0.0012308481382206082, + 0.0011640797602012753, + 0.000862369779497385, + 0.001912661362439394, + 0.0006258400389924645, + 0.0008002351969480515, + 0.0011777083855122328, + 0.0014318424509838223, + 0.0010209849569946527, + 0.0015906787011772394, + 0.001634727930650115, + 0.0015747490106150508, + 0.0012365931179374456, + 0.0012542190961539745, + 0.0015707035781815648, + 0.0018063721945509315, + 0.002142765559256077, + 0.0018123522168025374, + 0.001321682007983327, + 0.002009125892072916, + 0.0016577003989368677, + 0.0016188820591196418, + 0.0007966263801790774, + 0.0009423253359273076, + 0.001297086477279663, + 0.0016236254014074802, + 0.0014238107251003385, + 0.0010752385715022683, + 0.0013827525544911623, + 0.0014422924723476171, + 0.0004762089520227164, + 0.000696093775331974, + 0.000739414943382144, + 0.0009983888594433665, + 0.00111497868783772, + 0.000749873579479754, + 0.0009502712055109441, + 0.0008496508235111833, + 0.0006038199062459171, + 0.001031166291795671, + 0.001032856642268598, + 0.0013259951956570148, + 0.0010369600495323539, + 0.0010201351251453161, + 0.001063089701347053, + 0.001056955079548061, + 0.00041203261935152113, + 0.0002976910036522895, + 0.0003993113059550524, + 0.0003069248341489583, + 0.00025295239174738526, + 0.0003754091449081898, + 0.0003172784636262804, + 0.00036727014230564237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01818484626710415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot captures the surreal and expansive transformation of Millennium City Park, rendered entirely in the distinctive artistic style of a David Hockney painting. The scene is defined by clean lines, planes of vibrant, saturated color, and a slightly distorted, flattened perspective. A traditional Chinese pagoda, with its ornate terracotta tile roof and an ornate pavilion with imperial yellow accents, stands majestically above the surface of a vast, incredibly clear lake that fills the park's basin. Beneath the tranquil, azure waters, the underwater forms of tranquil gardens, intricate stone paths, and manicured trees are clearly visible. The central element is the gentle, rhythmic pulsation of the water itself; the water level rises steadily and smoothly, clinging to the pavilion's wooden latticework, then gracefully recedes, allowing the building's surfaces to shimmer in bright, shadowless sunlight against the azure sky. This constant, hypnotic fluctuation creates an atmosphere of dreamlike tranquility, a beautiful yet melancholic natural landscape restored to its man-made splendor in brilliant, serene colors.", + "seed": 27248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_9/index_3097_seed_27248_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06437991783022881, + "vlm_logprob": -0.06437991783022881, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.06437991783022881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027388200163841248 + ], + "opposite_logprob_tokens": [ + -3.777388095855713 + ], + "logprob_reward": -0.027388200163841248, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027388200163841248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3474317789077759, + 0.19071198999881744, + 0.10881051421165466, + 0.10281788557767868, + 0.05001397058367729, + 0.04023198410868645, + 0.03141988068819046, + 0.031195661053061485, + 0.03816966712474823, + 0.05919662490487099 + ], + "credit_map": [ + 0.03430844098329544, + 0.04280799254775047, + 0.050333984196186066, + 0.050075046718120575, + 0.03316670283675194, + 0.04246802255511284, + 0.04827331379055977, + 0.04599827155470848, + 0.01905287615954876, + 0.031139645725488663, + 0.03423532471060753, + 0.02904808335006237, + 0.029615774750709534, + 0.012947050854563713, + 0.02137879468500614, + 0.013294436037540436, + 0.012245970778167248, + 0.01006153505295515, + 0.01824190281331539, + 0.019573809579014778, + 0.012233168818056583, + 0.013185453601181507, + 0.010891161859035492, + 0.012377514503896236, + 0.012332267127931118, + 0.008130845613777637, + 0.01410883478820324, + 0.020498672500252724, + 0.007055636495351791, + 0.010591158643364906, + 0.011253300122916698, + 0.018847167491912842, + 0.005442720837891102, + 0.003912514075636864, + 0.0078995106741786, + 0.006421332247555256, + 0.004929171409457922, + 0.005330976564437151, + 0.005313865374773741, + 0.010763880796730518, + 0.0035527371801435947, + 0.004290055017918348, + 0.0057440283708274364, + 0.006975916679948568, + 0.005051456857472658, + 0.00443545775488019, + 0.004663689061999321, + 0.005518642254173756, + 0.003023158758878708, + 0.0022091774735599756, + 0.005209540482610464, + 0.003997754771262407, + 0.002668769331648946, + 0.002993279369547963, + 0.004772443789988756, + 0.006545756943523884, + 0.003085886826738715, + 0.002435634145513177, + 0.004831655416637659, + 0.006542593706399202, + 0.0024620576296001673, + 0.002467081183567643, + 0.005276976153254509, + 0.004093772731721401, + 0.0051102288998663425, + 0.002886479487642646, + 0.004286069422960281, + 0.0043967049568891525, + 0.0036715602036565542, + 0.005125229246914387, + 0.0046050515957176685, + 0.008088343776762486, + 0.004282757174223661, + 0.00455462746322155, + 0.012017996981739998, + 0.010805019177496433, + 0.0052667525596916676, + 0.004959233570843935, + 0.006459845695644617, + 0.010850390419363976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.027388200163841248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797697886824608 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797697886824608, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797697886824608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30757611989974976, + 0.21753087639808655, + 0.09922554343938828, + 0.08918292075395584, + 0.09393617510795593, + 0.022449851036071777, + 0.04568220674991608, + 0.04306328669190407, + 0.019172322005033493, + 0.06218070909380913 + ], + "credit_map": [ + 0.0343233160674572, + 0.0750337615609169, + 0.03629862144589424, + 0.02360413782298565, + 0.06466858088970184, + 0.024219825863838196, + 0.024185849353671074, + 0.025242002680897713, + 0.017820997163653374, + 0.02252410165965557, + 0.056700367480516434, + 0.041803669184446335, + 0.03113836795091629, + 0.016080019995570183, + 0.01837752014398575, + 0.013085813261568546, + 0.011695962399244308, + 0.014459120109677315, + 0.012016492895781994, + 0.012174547649919987, + 0.017930490896105766, + 0.009592593647539616, + 0.011080042459070683, + 0.010276279412209988, + 0.009893910959362984, + 0.00712458323687315, + 0.011850631795823574, + 0.01621874049305916, + 0.0068227569572627544, + 0.010456490330398083, + 0.010902405716478825, + 0.01591339148581028, + 0.011051289737224579, + 0.0070377835072577, + 0.014017339795827866, + 0.012456499971449375, + 0.008006326854228973, + 0.01083599403500557, + 0.009682435542345047, + 0.020848488435149193, + 0.0020096800290048122, + 0.0017780526541173458, + 0.004753944464027882, + 0.005126481410115957, + 0.0025102647487074137, + 0.0019962508231401443, + 0.002009820658713579, + 0.0022653548512607813, + 0.0038388539105653763, + 0.005085427779704332, + 0.005561282392591238, + 0.0067931972444057465, + 0.006944018881767988, + 0.004414764698594809, + 0.0062421951442956924, + 0.006802465301007032, + 0.003863106481730938, + 0.00599854439496994, + 0.006396373733878136, + 0.00623658113181591, + 0.00679432787001133, + 0.005592034198343754, + 0.0037470159586519003, + 0.004435297567397356, + 0.0020289889071136713, + 0.0020635616965591908, + 0.0028080707415938377, + 0.0029482650570571423, + 0.0024412127677351236, + 0.0018225894309580326, + 0.0019986433908343315, + 0.0030609876848757267, + 0.0062833186239004135, + 0.00828973762691021, + 0.005760692525655031, + 0.007331690285354853, + 0.01220446452498436, + 0.0074513801373541355, + 0.007440690882503986, + 0.007418729830533266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009797697886824608 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307607561349869 + ], + "opposite_logprob_tokens": [ + -3.1680760383605957 + ], + "logprob_reward": -0.04307607561349869, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307607561349869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22695064544677734, + 0.14522038400173187, + 0.16226935386657715, + 0.0972808375954628, + 0.10518407821655273, + 0.055146269500255585, + 0.05977705121040344, + 0.08667983114719391, + 0.01832377165555954, + 0.04316772520542145 + ], + "credit_map": [ + 0.022878563031554222, + 0.031632307916879654, + 0.03206091746687889, + 0.0350201390683651, + 0.027318280190229416, + 0.023115063086152077, + 0.033499736338853836, + 0.021425655111670494, + 0.017480596899986267, + 0.027982326224446297, + 0.01657547429203987, + 0.01211705431342125, + 0.025702668353915215, + 0.00950217992067337, + 0.020388852804899216, + 0.015471242368221283, + 0.015272993594408035, + 0.009789964184165001, + 0.03754859045147896, + 0.047428205609321594, + 0.014942674897611141, + 0.011579364538192749, + 0.014835345558822155, + 0.010872209444642067, + 0.01226672064512968, + 0.007831291295588017, + 0.012202312238514423, + 0.015433787368237972, + 0.007073740474879742, + 0.010992860421538353, + 0.012143991887569427, + 0.019336139783263206, + 0.010983134619891644, + 0.0055792201310396194, + 0.019791940227150917, + 0.013264444656670094, + 0.006508305203169584, + 0.010409481823444366, + 0.011497456580400467, + 0.027150101959705353, + 0.00701898755505681, + 0.00365533959120512, + 0.008649944327771664, + 0.006508031859993935, + 0.004665513057261705, + 0.0061377231031656265, + 0.0054437145590782166, + 0.013067015446722507, + 0.00678656343370676, + 0.0038881665095686913, + 0.008575095795094967, + 0.009967084042727947, + 0.0045318095944821835, + 0.005882020574063063, + 0.007143145427107811, + 0.01300316583365202, + 0.00925971008837223, + 0.005684149917215109, + 0.014676726423203945, + 0.016383856534957886, + 0.006450679153203964, + 0.009824278764426708, + 0.012749428860843182, + 0.01165100373327732, + 0.0014203606406226754, + 0.0022409718949347734, + 0.0024628632236272097, + 0.0030226309318095446, + 0.003004449652507901, + 0.001718689571134746, + 0.002414601854979992, + 0.0020392034202814102, + 0.005183954257518053, + 0.0029362502973526716, + 0.006575943902134895, + 0.006463493220508099, + 0.00398636469617486, + 0.006362751591950655, + 0.005733348894864321, + 0.005925615783780813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04307607561349869 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255701869726181 + ], + "opposite_logprob_tokens": [ + -1.6005702018737793 + ], + "logprob_reward": -0.2255701869726181, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255701869726181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16924570500850677, + 0.21902765333652496, + 0.1745234876871109, + 0.05969875678420067, + 0.05138091370463371, + 0.05693141743540764, + 0.0454670786857605, + 0.11008350551128387, + 0.06863397359848022, + 0.04500746354460716 + ], + "credit_map": [ + 0.020424967631697655, + 0.01972828060388565, + 0.02011011354625225, + 0.02592063695192337, + 0.015083109959959984, + 0.01771290972828865, + 0.025536729022860527, + 0.024728965014219284, + 0.02514134719967842, + 0.040901727974414825, + 0.03139102831482887, + 0.030968008562922478, + 0.03606932982802391, + 0.015604387037456036, + 0.021901626139879227, + 0.01705019734799862, + 0.014259198680520058, + 0.012165137566626072, + 0.03757275268435478, + 0.055461518466472626, + 0.014606746844947338, + 0.010335222817957401, + 0.014834056608378887, + 0.015288859605789185, + 0.0073778219521045685, + 0.004951811395585537, + 0.007500005420297384, + 0.00798323005437851, + 0.005552776623517275, + 0.0072875781916081905, + 0.007842001505196095, + 0.011203531175851822, + 0.005007608328014612, + 0.003819499397650361, + 0.009384676814079285, + 0.010830149054527283, + 0.004758050665259361, + 0.005408541765064001, + 0.005443530157208443, + 0.006728856358677149, + 0.00398080563172698, + 0.004616497550159693, + 0.012180636636912823, + 0.009723520837724209, + 0.005070878658443689, + 0.005860051140189171, + 0.00747060077264905, + 0.008028429001569748, + 0.0032119667157530785, + 0.004065683111548424, + 0.009638061746954918, + 0.0058837151154875755, + 0.005186432972550392, + 0.0036059969570487738, + 0.0051805851981043816, + 0.008694633841514587, + 0.012837751768529415, + 0.015017584897577763, + 0.010794583708047867, + 0.014366859570145607, + 0.018822500482201576, + 0.012571054510772228, + 0.011019882746040821, + 0.014653287827968597, + 0.005240836180746555, + 0.007542170584201813, + 0.00943795032799244, + 0.010415607132017612, + 0.00997142307460308, + 0.006872907746583223, + 0.00826637540012598, + 0.010886707343161106, + 0.004938777070492506, + 0.0042294664308428764, + 0.005158926360309124, + 0.005575728602707386, + 0.006250333972275257, + 0.004973255563527346, + 0.005791223142296076, + 0.008089751936495304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.2255701869726181 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606742851436138 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606742851436138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606742851436138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30732592940330505, + 0.1592252105474472, + 0.11024507135152817, + 0.11617522686719894, + 0.05686374753713608, + 0.03823855146765709, + 0.1026139184832573, + 0.058685146272182465, + 0.01948939822614193, + 0.031137855723500252 + ], + "credit_map": [ + 0.030150840058922768, + 0.04082578420639038, + 0.046236585825681686, + 0.03539141267538071, + 0.04768044874072075, + 0.032043732702732086, + 0.042851950973272324, + 0.0321451760828495, + 0.021488510072231293, + 0.035147491842508316, + 0.016674892976880074, + 0.011063629761338234, + 0.028500571846961975, + 0.014587273821234703, + 0.0183893833309412, + 0.013373463414609432, + 0.008464204147458076, + 0.009557360783219337, + 0.025026511400938034, + 0.020932044833898544, + 0.011830960400402546, + 0.008795533329248428, + 0.010131736285984516, + 0.015506716445088387, + 0.014950455166399479, + 0.008884749375283718, + 0.016559556126594543, + 0.01953495293855667, + 0.009915824048221111, + 0.01139769796282053, + 0.013632144778966904, + 0.02129983901977539, + 0.008056625723838806, + 0.004545697011053562, + 0.007525632623583078, + 0.010677809827029705, + 0.005625573452562094, + 0.005663108546286821, + 0.006549535319209099, + 0.00821976363658905, + 0.00572317885234952, + 0.0038594650104641914, + 0.004866980016231537, + 0.007998126558959484, + 0.0034769673366099596, + 0.003260453464463353, + 0.004738797899335623, + 0.0043145837262272835, + 0.012257705442607403, + 0.006969945505261421, + 0.01400667428970337, + 0.013144743628799915, + 0.008703459985554218, + 0.010112442076206207, + 0.014232280664145947, + 0.023186668753623962, + 0.0066607133485376835, + 0.004855765495449305, + 0.009912152774631977, + 0.006800222210586071, + 0.006077282130718231, + 0.005701090209186077, + 0.0056106531992554665, + 0.013067272491753101, + 0.001751884468831122, + 0.0020606524776667356, + 0.002892427844926715, + 0.0024958692956715822, + 0.0026758951134979725, + 0.0026780879124999046, + 0.0019372932147234678, + 0.00299728917889297, + 0.003231198526918888, + 0.0025740093551576138, + 0.006055922247469425, + 0.00458812341094017, + 0.004030454438179731, + 0.002770536346361041, + 0.0037868639919906855, + 0.00410075020045042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01606742851436138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot captures the surreal and expansive transformation of Millennium City Park, rendered entirely in the distinctive artistic style of a David Hockney painting. The scene is defined by clean lines, planes of vibrant, saturated color, and a slightly distorted, flattened perspective. A traditional Chinese pagoda, with its ornate terracotta tile roof and an ornate pavilion with imperial yellow accents, stands majestically above the surface of a vast, incredibly clear lake that fills the park's basin. Beneath the tranquil, azure waters, the underwater forms of tranquil gardens, intricate stone paths, and manicured trees are clearly visible. The central element is the gentle, rhythmic pulsation of the water itself; the water level rises steadily and smoothly, clinging to the pavilion's wooden latticework, then gracefully recedes, allowing the building's surfaces to shimmer in bright, shadowless sunlight against the azure sky. This constant, hypnotic fluctuation creates an atmosphere of dreamlike tranquility, a beautiful yet melancholic natural landscape restored to its man-made splendor in brilliant, serene colors.", + "seed": 27249, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000057/group_9/index_3097_seed_27249_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06072502359747887, + "vlm_logprob": -0.06072502359747887, + "vlm_yes_no_margin": 3.974999952316284, + "vlm_token_credit": -0.06072502359747887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006875073071569204 + ], + "opposite_logprob_tokens": [ + -5.131875038146973 + ], + "logprob_reward": -0.006875073071569204, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006875073071569204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5763419270515442, + 0.14446404576301575, + 0.008426481857895851, + 0.023476872593164444, + 0.031419139355421066, + 0.056183136999607086, + 0.027663704007864, + 0.08413298428058624, + 0.036231521517038345, + 0.01166017260402441 + ], + "credit_map": [ + 0.015402776189148426, + 0.011838715523481369, + 0.008854878135025501, + 0.0052407607436180115, + 0.25050872564315796, + 0.25882548093795776, + 0.009458489716053009, + 0.016212087124586105, + 0.014349376782774925, + 0.02162075974047184, + 0.014073370024561882, + 0.010554309003055096, + 0.0253724604845047, + 0.025782696902751923, + 0.013157812878489494, + 0.019553259015083313, + 0.0007340079755522311, + 0.0015049788635224104, + 0.0008143881568685174, + 0.0007917515467852354, + 0.0009289292502216995, + 0.0016827769577503204, + 0.0008217546273954213, + 0.0011478946544229984, + 0.002512881299480796, + 0.0025278336834162474, + 0.0026282912585884333, + 0.0018271013395860791, + 0.0038104457780718803, + 0.0039817458018660545, + 0.0027481201104819775, + 0.003440454136580229, + 0.004092107061296701, + 0.0023268619552254677, + 0.003950218670070171, + 0.0030317511409521103, + 0.003513380652293563, + 0.004876294638961554, + 0.005050841718912125, + 0.004577684216201305, + 0.0026492977049201727, + 0.0017836636397987604, + 0.002750360406935215, + 0.0025394721888005733, + 0.0032957650255411863, + 0.012092065997421741, + 0.01605795882642269, + 0.015014548785984516, + 0.003243600483983755, + 0.0024339035153388977, + 0.002784227952361107, + 0.0026842893566936255, + 0.002599547151476145, + 0.003953342791646719, + 0.004804434720426798, + 0.005160355940461159, + 0.0025758142583072186, + 0.0034247797448188066, + 0.0029152040369808674, + 0.0036368623841553926, + 0.003899283241480589, + 0.004017101600766182, + 0.031842853873968124, + 0.03182107210159302, + 0.0058672153390944, + 0.004170526750385761, + 0.0036311671137809753, + 0.004504873882979155, + 0.003335562301799655, + 0.004248977173119783, + 0.004632779862731695, + 0.005840420722961426, + 0.0010896429885178804, + 0.0010195499053224921, + 0.0012811667984351516, + 0.0010107994312420487, + 0.0014007362769916654, + 0.0017143669538199902, + 0.002263799076899886, + 0.001880111638456583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.006875073071569204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003202312160283327 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.003202312160283327, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003202312160283327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38652268052101135, + 0.05224312096834183, + 0.010029765777289867, + 0.019575482234358788, + 0.03923614323139191, + 0.024072950705885887, + 0.01516991388052702, + 0.4190415143966675, + 0.02397654391825199, + 0.010131868533790112 + ], + "credit_map": [ + 0.012624015100300312, + 0.018092118203639984, + 0.008885784074664116, + 0.007235233671963215, + 0.15970277786254883, + 0.1559741199016571, + 0.011998663656413555, + 0.012009983882308006, + 0.005078985821455717, + 0.007872362621128559, + 0.00501105235889554, + 0.004123453050851822, + 0.008620155975222588, + 0.008831772021949291, + 0.006199143826961517, + 0.006506192963570356, + 0.0008408281719312072, + 0.0009537985897623003, + 0.0013496794272214174, + 0.001075970707461238, + 0.0015844178851693869, + 0.001996451523154974, + 0.0009619027841836214, + 0.0012667163973674178, + 0.0021030863281339407, + 0.00200269790366292, + 0.0024656786117702723, + 0.00175363605376333, + 0.002782784402370453, + 0.002629769267514348, + 0.0034871601965278387, + 0.0023506705183535814, + 0.005407346878200769, + 0.0032790983095765114, + 0.004710998851805925, + 0.003780209459364414, + 0.003927037585526705, + 0.00578919006511569, + 0.006555322092026472, + 0.005786942783743143, + 0.001566259772516787, + 0.0014575024833902717, + 0.0022728159092366695, + 0.0017880210652947426, + 0.0020660492591559887, + 0.0035497737117111683, + 0.006586926057934761, + 0.004785601980984211, + 0.0018678870983421803, + 0.0018240706995129585, + 0.0015992306871339679, + 0.0015614238800480962, + 0.0019625802524387836, + 0.001773672760464251, + 0.002266071503981948, + 0.002314976416528225, + 0.00995678175240755, + 0.014392484910786152, + 0.012800270691514015, + 0.012301160022616386, + 0.017580602318048477, + 0.01674767956137657, + 0.16898967325687408, + 0.16627287864685059, + 0.00373401353135705, + 0.0029222622979432344, + 0.00240797596052289, + 0.001958441687747836, + 0.002462844830006361, + 0.0033995024859905243, + 0.00339279742911458, + 0.0036987054627388716, + 0.0009591050329618156, + 0.0009181132190860808, + 0.0011678804876282811, + 0.0011513211065903306, + 0.001253051101230085, + 0.0014774749288335443, + 0.0015932030510157347, + 0.0016117201885208488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.003202312160283327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192455172538757 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192455172538757, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018192455172538757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.541625440120697, + 0.09377612918615341, + 0.005708519835025072, + 0.0202115960419178, + 0.00904457550495863, + 0.0486108772456646, + 0.040162280201911926, + 0.19811969995498657, + 0.026852017268538475, + 0.015888899564743042 + ], + "credit_map": [ + 0.014063031412661076, + 0.012974402867257595, + 0.008465378545224667, + 0.004949474707245827, + 0.23692774772644043, + 0.24034519493579865, + 0.009269405156373978, + 0.014630830846726894, + 0.00969640351831913, + 0.013853924348950386, + 0.009383555501699448, + 0.007219879422336817, + 0.01643579825758934, + 0.014577077701687813, + 0.01019983645528555, + 0.012409654445946217, + 0.0006200819625519216, + 0.0006080649327486753, + 0.0006305897841230035, + 0.0004495080793276429, + 0.0007064317469485104, + 0.001151145901530981, + 0.0005180005100555718, + 0.0010246969759464264, + 0.001995200291275978, + 0.0024130758829414845, + 0.0023602121509611607, + 0.001857183175161481, + 0.0030183426570147276, + 0.002980925841256976, + 0.003222454572096467, + 0.0023642019368708134, + 0.0010959317442029715, + 0.0010823957854881883, + 0.000886020774487406, + 0.0007866037194617093, + 0.0009643500088714063, + 0.0016739681595936418, + 0.001192503608763218, + 0.0013628019951283932, + 0.0025185649283230305, + 0.0016782389720901847, + 0.002420949749648571, + 0.002208239398896694, + 0.0025092733558267355, + 0.009180080145597458, + 0.014255048707127571, + 0.01384048257023096, + 0.0041786786168813705, + 0.004318707622587681, + 0.00399449747055769, + 0.0036648933310061693, + 0.003929902799427509, + 0.005914187990128994, + 0.00704567413777113, + 0.0071157412603497505, + 0.004519991111010313, + 0.006994728464633226, + 0.006053198594599962, + 0.005924970842897892, + 0.008379091508686543, + 0.008079194463789463, + 0.07948049902915955, + 0.07868802547454834, + 0.003699314082041383, + 0.0031772132497280836, + 0.0024990790989249945, + 0.002415707567706704, + 0.002801814815029502, + 0.004150084685534239, + 0.003546338528394699, + 0.0045624650083482265, + 0.0019612072501331568, + 0.002797583816573024, + 0.0020176314283162355, + 0.0015541169559583068, + 0.0014665561029687524, + 0.0016356128035113215, + 0.002466262085363269, + 0.0019899304024875164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.018192455172538757 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252070814371109 + ], + "opposite_logprob_tokens": [ + -1.5020707845687866 + ], + "logprob_reward": -0.252070814371109, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252070814371109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43186870217323303, + 0.05067001283168793, + 0.021956512704491615, + 0.03566093370318413, + 0.00930347852408886, + 0.09019898623228073, + 0.0227920264005661, + 0.28784096240997314, + 0.032776840031147, + 0.016931504011154175 + ], + "credit_map": [ + 0.011674589477479458, + 0.010991554707288742, + 0.007559418212622404, + 0.004174825269728899, + 0.1883402019739151, + 0.18880341947078705, + 0.008280421607196331, + 0.012044301256537437, + 0.005887035746127367, + 0.007506649941205978, + 0.005353343673050404, + 0.003843480022624135, + 0.008290795609354973, + 0.008741660043597221, + 0.005745432805269957, + 0.005301615223288536, + 0.001405030838213861, + 0.004583545029163361, + 0.0013347840867936611, + 0.0016355839325115085, + 0.002837738487869501, + 0.003936010412871838, + 0.001724323839880526, + 0.00449949549511075, + 0.004259895998984575, + 0.0038807743694633245, + 0.004431427456438541, + 0.0030883599538356066, + 0.004141113720834255, + 0.004532525781542063, + 0.006396749056875706, + 0.004930085968226194, + 0.0010260723065584898, + 0.0010083435336127877, + 0.001053925952874124, + 0.0010616523213684559, + 0.0008660583407618105, + 0.001148456009104848, + 0.0018722896929830313, + 0.0012666805414482951, + 0.004884618334472179, + 0.003074965439736843, + 0.005193705204874277, + 0.0046096970327198505, + 0.0055129555985331535, + 0.019443584606051445, + 0.023666653782129288, + 0.023812804371118546, + 0.002594664925709367, + 0.0038869171403348446, + 0.0023482569959014654, + 0.0020477930083870888, + 0.0024950632359832525, + 0.002464073710143566, + 0.0026589573826640844, + 0.00429630046710372, + 0.007230452261865139, + 0.009787492454051971, + 0.008968526497483253, + 0.009471824392676353, + 0.012120054103434086, + 0.011518520303070545, + 0.11537995934486389, + 0.11336413025856018, + 0.0032879465725272894, + 0.005267939995974302, + 0.0031000429298728704, + 0.004955586511641741, + 0.0033273829612880945, + 0.004821703303605318, + 0.0037910586688667536, + 0.004225179087370634, + 0.002448044717311859, + 0.0033576253335922956, + 0.0016944161616265774, + 0.0012779562966898084, + 0.0017475944478064775, + 0.0018706611590459943, + 0.001926440978422761, + 0.0026087637525051832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.252070814371109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the distinctive artistic style of a David Hockney painting with clean lines and vibrant, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a traditional Chinese pagoda with an ornate terracotta tile roof and imperial yellow accents visible above the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are underwater forms of gardens, stone paths, and trees clearly visible beneath the tranquil, azure waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the water level rise steadily and smoothly to cling to the pavilion's wooden latticework before receding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, high-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284463211894035 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284463211894035, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284463211894035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6863219738006592, + 0.0624706894159317, + 0.015503592789173126, + 0.011852232739329338, + 0.007209391333162785, + 0.029479028657078743, + 0.008997415192425251, + 0.15875235199928284, + 0.01129437517374754, + 0.008119062520563602 + ], + "credit_map": [ + 0.017062749713659286, + 0.012792946770787239, + 0.010559788905084133, + 0.005604865960776806, + 0.30241864919662476, + 0.3086118996143341, + 0.010983947664499283, + 0.018287014216184616, + 0.006798686925321817, + 0.008871939964592457, + 0.006664978340268135, + 0.00476393336430192, + 0.010791999287903309, + 0.0093681076541543, + 0.007037929259240627, + 0.00817310344427824, + 0.0012073421385139227, + 0.004005438648164272, + 0.0011211043456569314, + 0.0013009936083108187, + 0.0015650003915652633, + 0.0025194634217768908, + 0.0010176284704357386, + 0.0027666203677654266, + 0.0012454175157472491, + 0.001465335488319397, + 0.001478566788136959, + 0.0010026836534962058, + 0.0015808296157047153, + 0.0014488919405266643, + 0.001688827178440988, + 0.0019416797440499067, + 0.0007570409798063338, + 0.0006707181455567479, + 0.0007614888600073755, + 0.000657985161524266, + 0.0008095979574136436, + 0.0014817400369793177, + 0.0011607016203925014, + 0.0009101175237447023, + 0.0017253467813134193, + 0.0010731357615441084, + 0.00185937259811908, + 0.0015404000878334045, + 0.0018861296121031046, + 0.006925453897565603, + 0.007211368530988693, + 0.007257816847413778, + 0.0008990450296550989, + 0.0011002031387761235, + 0.00097976706456393, + 0.0009876200929284096, + 0.0007881300989538431, + 0.0012295568594709039, + 0.0013510228600353003, + 0.0016620690003037453, + 0.0042720334604382515, + 0.005740951746702194, + 0.005028617102652788, + 0.004646486137062311, + 0.006985560525208712, + 0.00672193942591548, + 0.06345172226428986, + 0.06190503388643265, + 0.002129795728251338, + 0.0014876594068482518, + 0.0011434119660407305, + 0.0012362846173346043, + 0.0009856538381427526, + 0.0010517697082832456, + 0.0015148469246923923, + 0.0017449521692469716, + 0.0007990744197741151, + 0.0012226583203300834, + 0.0009026931365951896, + 0.000836435123346746, + 0.0008458569063805044, + 0.0008918607491068542, + 0.001637750887311995, + 0.0009827320463955402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.023284463211894035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_0/index_tensor([6482, 6482])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_0/index_tensor([6482, 6482])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b1da72a1bd3c92f881b33bb9e834e3a5c037a4cd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_0/index_tensor([6482, 6482])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a slow-motion shot captured at a high frame rate, the camera begins a majestic spiral climb, gliding upwards along the skeleton of a massive, unfinished ship. The evening sun shone with a warm golden glow, penetrating a dense forest of metal scaffolding. Casting ever-shifting, elongated shadows across the cold, dark-gray steel ribs of the ship, a dynamic chessboard of light and shadow. As the camera gracefully swirls, light plays across rough welds and exposed rivets, evoking the raw texture of industrial construction. Deep within the hull structure, welding torches operated by unseen workers occasionally ignite with fleeting, brilliant sparks. These fiery orange sparks, like stars in the night sky, shimmer briefly in the twilight, momentarily illuminating a small patch of rusted steel before fading into deep shadow. This imposing yet tranquil industrial landscape imbued with a vibrant sense of movement, imbuing the image with the grand, poetic, and majestic beauty of industrial construction.", + "seed": 17570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_0/index_6482_seed_17570_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15322363264858724, + "vlm_logprob": -0.15322363264858724, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.15322363264858724, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03305071219801903 + ], + "opposite_logprob_tokens": [ + -3.658050775527954 + ], + "logprob_reward": -0.03305071219801903, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03305071219801903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2970227599143982, + 0.07896280288696289, + 0.16300810873508453, + 0.03195507079362869, + 0.04393716901540756, + 0.03698711097240448, + 0.10649905353784561, + 0.11139044910669327, + 0.04085461050271988, + 0.0893828347325325 + ], + "credit_map": [ + 0.029408015310764313, + 0.03908531367778778, + 0.044780172407627106, + 0.04265926778316498, + 0.034135833382606506, + 0.03264269977807999, + 0.03485163301229477, + 0.03945983946323395, + 0.01072560716420412, + 0.00989062525331974, + 0.010260013863444328, + 0.012389859184622765, + 0.007441444788128138, + 0.008618514984846115, + 0.01046623196452856, + 0.00917050801217556, + 0.014604492112994194, + 0.019050870090723038, + 0.021992310881614685, + 0.021947069093585014, + 0.016769511625170708, + 0.02487150765955448, + 0.017031937837600708, + 0.026740428060293198, + 0.0026834860909730196, + 0.0025838653091341257, + 0.0044491467997431755, + 0.003556949319317937, + 0.0030799629166722298, + 0.003072973806411028, + 0.007306886836886406, + 0.00522179901599884, + 0.00683949189260602, + 0.003663254203274846, + 0.008297240361571312, + 0.004011954180896282, + 0.003602883080020547, + 0.004386786837130785, + 0.006819047033786774, + 0.0063165160827338696, + 0.002799844602122903, + 0.0031036166474223137, + 0.005068936385214329, + 0.0062223561108112335, + 0.003738259430974722, + 0.003707197029143572, + 0.006456403061747551, + 0.005890500731766224, + 0.008389544673264027, + 0.023298637941479683, + 0.011072024703025818, + 0.012931433506309986, + 0.01164594292640686, + 0.007493602577596903, + 0.01721869222819805, + 0.014449181035161018, + 0.008916034363210201, + 0.009082556702196598, + 0.015189672820270061, + 0.017361382022500038, + 0.011558623984456062, + 0.014640520326793194, + 0.0183340422809124, + 0.016307631507515907, + 0.002093034563586116, + 0.004005322698503733, + 0.004441040568053722, + 0.005179841537028551, + 0.002964604878798127, + 0.008768163621425629, + 0.006695600692182779, + 0.00670700054615736, + 0.004108387045562267, + 0.003928080666810274, + 0.004385203588753939, + 0.006239910144358873, + 0.008026001043617725, + 0.02886960655450821, + 0.005147095303982496, + 0.02867855876684189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.03305071219801903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867158457636833 + ], + "opposite_logprob_tokens": [ + -3.0486714839935303 + ], + "logprob_reward": -0.04867158457636833, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867158457636833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2087569385766983, + 0.2711234986782074, + 0.05525670945644379, + 0.043203528970479965, + 0.028739511966705322, + 0.08813432604074478, + 0.1375521719455719, + 0.10895230621099472, + 0.024662962183356285, + 0.03361806273460388 + ], + "credit_map": [ + 0.020721152424812317, + 0.02514907717704773, + 0.02937726490199566, + 0.030161332339048386, + 0.020365238189697266, + 0.03809942305088043, + 0.022536735981702805, + 0.022346684709191322, + 0.023001957684755325, + 0.03307272866368294, + 0.03748411312699318, + 0.0419803187251091, + 0.03173385560512543, + 0.034482691437006, + 0.03825642168521881, + 0.03111140802502632, + 0.0045724837109446526, + 0.006501832976937294, + 0.011479652486741543, + 0.0051014539785683155, + 0.00536822946742177, + 0.0058156391605734825, + 0.005783425644040108, + 0.010633988305926323, + 0.002974096452817321, + 0.004285008180886507, + 0.00575823150575161, + 0.008055051788687706, + 0.003046096069738269, + 0.0035119226668030024, + 0.009609990753233433, + 0.005963131785392761, + 0.0035865583922713995, + 0.002640373073518276, + 0.004970056936144829, + 0.0038743193726986647, + 0.0022942875511944294, + 0.002712187124416232, + 0.004529278725385666, + 0.0041324482299387455, + 0.008001233451068401, + 0.006858943495899439, + 0.009148758836090565, + 0.010494065470993519, + 0.010115845128893852, + 0.012274635024368763, + 0.01913839764893055, + 0.012102444656193256, + 0.007502557244151831, + 0.011445403099060059, + 0.009116139262914658, + 0.012171541340649128, + 0.029118597507476807, + 0.02402542158961296, + 0.023450421169400215, + 0.020722083747386932, + 0.005134256090968847, + 0.008046623319387436, + 0.015117023140192032, + 0.011045697145164013, + 0.01068291999399662, + 0.015505662187933922, + 0.024577874690294266, + 0.018842237070202827, + 0.0021572569385170937, + 0.002584391040727496, + 0.003057441208511591, + 0.0031475729774683714, + 0.002073382493108511, + 0.0041567422449588776, + 0.003743401262909174, + 0.0037427719216793776, + 0.003752783639356494, + 0.00233545177616179, + 0.0023810050915926695, + 0.00288118957541883, + 0.004197061527520418, + 0.007655679248273373, + 0.002901775063946843, + 0.007513116113841534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04867158457636833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282061159610748 + ], + "opposite_logprob_tokens": [ + -2.0178205966949463 + ], + "logprob_reward": -0.14282061159610748, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282061159610748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31138044595718384, + 0.08847219496965408, + 0.08650515228509903, + 0.07520847022533417, + 0.14292332530021667, + 0.03167542442679405, + 0.038618117570877075, + 0.07413046061992645, + 0.02592327445745468, + 0.12516316771507263 + ], + "credit_map": [ + 0.03172805905342102, + 0.033012647181749344, + 0.039227262139320374, + 0.041486646980047226, + 0.029461784288287163, + 0.053594112396240234, + 0.04304083436727524, + 0.0398290641605854, + 0.008637450635433197, + 0.010747687891125679, + 0.011400598101317883, + 0.013634739443659782, + 0.010134764015674591, + 0.0129088731482625, + 0.012237345799803734, + 0.008770724758505821, + 0.005342240910977125, + 0.00533028831705451, + 0.023777326568961143, + 0.008361407555639744, + 0.004678416531533003, + 0.0073761651292443275, + 0.009072083048522472, + 0.02256721630692482, + 0.005167081952095032, + 0.006867437623441219, + 0.009825359098613262, + 0.01195277739316225, + 0.00649998290464282, + 0.005886494182050228, + 0.01876351609826088, + 0.010245809331536293, + 0.020426500588655472, + 0.015222185291349888, + 0.02532362937927246, + 0.014233416877686977, + 0.012684800662100315, + 0.014717631042003632, + 0.021279694512486458, + 0.01903546042740345, + 0.0024983531329780817, + 0.0037639643996953964, + 0.00336548313498497, + 0.004795303102582693, + 0.0017773358849808574, + 0.0023580272682011127, + 0.007490626070648432, + 0.005626330152153969, + 0.005049529019743204, + 0.005255310330539942, + 0.005288722459226847, + 0.004918931517750025, + 0.0039152055978775024, + 0.0027711139991879463, + 0.006266179494559765, + 0.005153123755007982, + 0.004907894413918257, + 0.005059135612100363, + 0.010045773349702358, + 0.007723064161837101, + 0.007406209129840136, + 0.010764295235276222, + 0.015858957543969154, + 0.01236511766910553, + 0.001360042137093842, + 0.0019583208486437798, + 0.001851018168963492, + 0.0022161765955388546, + 0.002878478728234768, + 0.007022325415164232, + 0.0038691486697643995, + 0.004767762031406164, + 0.006296367384493351, + 0.004624416120350361, + 0.005647845566272736, + 0.006875060964375734, + 0.005264031235128641, + 0.044215064495801926, + 0.007770467549562454, + 0.044469915330410004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.14282061159610748 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185080960392952 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185080960392952, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185080960392952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09822544455528259, + 0.059774868190288544, + 0.1327052265405655, + 0.02736610919237137, + 0.061208512634038925, + 0.03845035284757614, + 0.07414470613002777, + 0.180683895945549, + 0.07998152822256088, + 0.24745933711528778 + ], + "credit_map": [ + 0.012122792191803455, + 0.010551869869232178, + 0.01395510882139206, + 0.011555402539670467, + 0.010534476488828659, + 0.015061174519360065, + 0.0097762830555439, + 0.014668339863419533, + 0.004754181485623121, + 0.008396980352699757, + 0.00802696030586958, + 0.009171001613140106, + 0.007242911960929632, + 0.007724246475845575, + 0.007479947991669178, + 0.006978638470172882, + 0.007130713667720556, + 0.010467885062098503, + 0.03514557704329491, + 0.013460930436849594, + 0.009620933793485165, + 0.010209466330707073, + 0.012317522428929806, + 0.03435221314430237, + 0.002268589800223708, + 0.003392902435734868, + 0.004480226431041956, + 0.0031896280124783516, + 0.0016179801896214485, + 0.0028001803439110518, + 0.003690629731863737, + 0.0059259748086333275, + 0.007977660745382309, + 0.006095259450376034, + 0.010732335969805717, + 0.007542122155427933, + 0.005928479600697756, + 0.008718537166714668, + 0.007031346671283245, + 0.007182769943028688, + 0.0028734265360981226, + 0.002614469500258565, + 0.005395499058067799, + 0.0061807273887097836, + 0.0027374704368412495, + 0.0031933393329381943, + 0.009106611832976341, + 0.006348808761686087, + 0.008359427563846111, + 0.00806631613522768, + 0.006437875330448151, + 0.009462837129831314, + 0.008194224908947945, + 0.006341366097331047, + 0.015626611188054085, + 0.011656047776341438, + 0.01066217478364706, + 0.013822218403220177, + 0.02746046520769596, + 0.02347726933658123, + 0.015522561967372894, + 0.013827592134475708, + 0.046072252094745636, + 0.029839377850294113, + 0.007193018216639757, + 0.010417941957712173, + 0.010304603725671768, + 0.010084555484354496, + 0.006430432666093111, + 0.01129146572202444, + 0.013197730295360088, + 0.01106178481131792, + 0.019199809059500694, + 0.010488555766642094, + 0.01650167442858219, + 0.021533755585551262, + 0.021842168644070625, + 0.06259822100400925, + 0.026633266359567642, + 0.0686618760228157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.018185080960392952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233901739120483 + ], + "opposite_logprob_tokens": [ + -0.8983901739120483 + ], + "logprob_reward": -0.5233901739120483, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233901739120483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13622237741947174, + 0.06809995323419571, + 0.07091911882162094, + 0.06086971238255501, + 0.029507268220186234, + 0.07129006087779999, + 0.07751935720443726, + 0.17267955839633942, + 0.11499575525522232, + 0.1978968232870102 + ], + "credit_map": [ + 0.016436658799648285, + 0.014949847012758255, + 0.018329676240682602, + 0.020406795665621758, + 0.018486956134438515, + 0.016939181834459305, + 0.009775431826710701, + 0.020897844806313515, + 0.00820082612335682, + 0.009248589165508747, + 0.010207840241491795, + 0.009024867787957191, + 0.007458400912582874, + 0.00654204934835434, + 0.00797333288937807, + 0.009444044902920723, + 0.0057352399453520775, + 0.010560883209109306, + 0.012432592920958996, + 0.0074478923343122005, + 0.0049736276268959045, + 0.00841448176652193, + 0.007424157578498125, + 0.013930246233940125, + 0.004949883092194796, + 0.006367342546582222, + 0.009463133290410042, + 0.011060896329581738, + 0.005353592801839113, + 0.004512447398155928, + 0.011840160936117172, + 0.007322254125028849, + 0.004291248973459005, + 0.002673316514119506, + 0.005163308698683977, + 0.002522966591641307, + 0.00329927122220397, + 0.002554158214479685, + 0.0046215904876589775, + 0.004381408449262381, + 0.00512964092195034, + 0.0076842582784593105, + 0.010643139481544495, + 0.011849969625473022, + 0.004466686863452196, + 0.0052958461456000805, + 0.014073316007852554, + 0.012147205881774426, + 0.007816164754331112, + 0.007208495866507292, + 0.007454435806721449, + 0.011127528734505177, + 0.010255496017634869, + 0.01011921837925911, + 0.012808487750589848, + 0.010729529894888401, + 0.007994999177753925, + 0.011954185552895069, + 0.022237509489059448, + 0.02705126442015171, + 0.015190542675554752, + 0.0220076534897089, + 0.03869037702679634, + 0.027553021907806396, + 0.008909061551094055, + 0.012016135267913342, + 0.015136498026549816, + 0.018705647438764572, + 0.010310451500117779, + 0.018072163686156273, + 0.015469059348106384, + 0.016376737505197525, + 0.014797607436776161, + 0.008176030591130257, + 0.011687850579619408, + 0.011713515967130661, + 0.013481167145073414, + 0.0628293827176094, + 0.015367628075182438, + 0.05984364449977875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.5233901739120483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow-motion shot captured at a high frame rate, the camera begins a majestic spiral climb, gliding upwards along the skeleton of a massive, unfinished ship. The evening sun shone with a warm golden glow, penetrating a dense forest of metal scaffolding. Casting ever-shifting, elongated shadows across the cold, dark-gray steel ribs of the ship, a dynamic chessboard of light and shadow. As the camera gracefully swirls, light plays across rough welds and exposed rivets, evoking the raw texture of industrial construction. Deep within the hull structure, welding torches operated by unseen workers occasionally ignite with fleeting, brilliant sparks. These fiery orange sparks, like stars in the night sky, shimmer briefly in the twilight, momentarily illuminating a small patch of rusted steel before fading into deep shadow. This imposing yet tranquil industrial landscape imbued with a vibrant sense of movement, imbuing the image with the grand, poetic, and majestic beauty of industrial construction.", + "seed": 17571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_0/index_6482_seed_17571_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2081398706883192, + "vlm_logprob": -0.2081398706883192, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.2081398706883192, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024920864030718803 + ], + "opposite_logprob_tokens": [ + -4.02492094039917 + ], + "logprob_reward": -0.024920864030718803, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024920864030718803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13966712355613708, + 0.040765613317489624, + 0.0820205882191658, + 0.06013568490743637, + 0.41603755950927734, + 0.05495940521359444, + 0.0908028855919838, + 0.04932069405913353, + 0.04018056020140648, + 0.026109851896762848 + ], + "credit_map": [ + 0.013156714849174023, + 0.015180648304522038, + 0.026316305622458458, + 0.013191615231335163, + 0.010492155328392982, + 0.010045798495411873, + 0.027985306456685066, + 0.023298578336834908, + 0.004933109041303396, + 0.004769733175635338, + 0.006390016060322523, + 0.005943287629634142, + 0.00267872866243124, + 0.0035305237397551537, + 0.005782183725386858, + 0.0067380317486822605, + 0.008879799395799637, + 0.008944123983383179, + 0.015145241282880306, + 0.016385560855269432, + 0.005333754234015942, + 0.005723459646105766, + 0.009760168381035328, + 0.011848481371998787, + 0.007455367129296064, + 0.005487009882926941, + 0.0116371288895607, + 0.008498390205204487, + 0.005164570640772581, + 0.004399820230901241, + 0.008568276651203632, + 0.00892512034624815, + 0.028027543798089027, + 0.010276786983013153, + 0.15285906195640564, + 0.024665391072630882, + 0.012150980532169342, + 0.017632007598876953, + 0.01139223575592041, + 0.15903355181217194, + 0.0065913936123251915, + 0.005135879851877689, + 0.007722142152488232, + 0.01002295222133398, + 0.0059197633527219296, + 0.004992994014173746, + 0.007275798358023167, + 0.007298481650650501, + 0.008936560712754726, + 0.0073990593664348125, + 0.010751526802778244, + 0.010766117833554745, + 0.017666127532720566, + 0.011146165430545807, + 0.014758840203285217, + 0.00937848724424839, + 0.004777069203555584, + 0.0038989537861198187, + 0.008163119666278362, + 0.007061907555907965, + 0.003455197438597679, + 0.004738271702080965, + 0.00487051997333765, + 0.012355657294392586, + 0.004943164996802807, + 0.002854466438293457, + 0.005681521724909544, + 0.006284753326326609, + 0.00501483678817749, + 0.005778048187494278, + 0.004463807679712772, + 0.0051599652506411076, + 0.00269994605332613, + 0.002083859173581004, + 0.0035697235725820065, + 0.004748978186398745, + 0.003426413284614682, + 0.003388764103874564, + 0.003123329719528556, + 0.003068840131163597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.024920864030718803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033718500286340714 + ], + "opposite_logprob_tokens": [ + -3.4087185859680176 + ], + "logprob_reward": -0.033718500286340714, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033718500286340714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14078207314014435, + 0.055960215628147125, + 0.05032644048333168, + 0.2719239294528961, + 0.025389816612005234, + 0.044407039880752563, + 0.06075073033571243, + 0.17831142246723175, + 0.13926495611667633, + 0.032883357256650925 + ], + "credit_map": [ + 0.0164524856954813, + 0.017667260020971298, + 0.03036646917462349, + 0.018270766362547874, + 0.008941201493144035, + 0.010446819476783276, + 0.017720317468047142, + 0.020916737616062164, + 0.005397205241024494, + 0.009337062947452068, + 0.007551153656095266, + 0.006217930465936661, + 0.00432296097278595, + 0.005619075149297714, + 0.009664904326200485, + 0.007849923335015774, + 0.005780922248959541, + 0.0051488615572452545, + 0.009893021546304226, + 0.008563579991459846, + 0.003844451392069459, + 0.004111868795007467, + 0.004927822854369879, + 0.00805591233074665, + 0.05043861269950867, + 0.0326024666428566, + 0.04613671079277992, + 0.03777306526899338, + 0.018561124801635742, + 0.02400343492627144, + 0.023792650550603867, + 0.03861584886908531, + 0.0031351293437182903, + 0.003460473846644163, + 0.003885011188685894, + 0.002779304049909115, + 0.0033328647259622812, + 0.002555648796260357, + 0.002678444841876626, + 0.003562940750271082, + 0.0052278474904596806, + 0.0037752066273242235, + 0.007664968725293875, + 0.004121342673897743, + 0.0038102746475487947, + 0.005122361239045858, + 0.005454969592392445, + 0.009230069816112518, + 0.009176675230264664, + 0.0044238874688744545, + 0.008678499609231949, + 0.007118030451238155, + 0.009733064100146294, + 0.0048369658179581165, + 0.007202315144240856, + 0.00958129484206438, + 0.01578526943922043, + 0.010282199829816818, + 0.030913900583982468, + 0.029814427718520164, + 0.010008911602199078, + 0.019247572869062424, + 0.019133782014250755, + 0.04312534257769585, + 0.025805532932281494, + 0.008361590094864368, + 0.018575415015220642, + 0.021075937896966934, + 0.01101658120751381, + 0.021430427208542824, + 0.013784079812467098, + 0.019215401262044907, + 0.002914626384153962, + 0.002270661061629653, + 0.004613943863660097, + 0.005853721871972084, + 0.005286436527967453, + 0.0034782537259161472, + 0.0040810792706906796, + 0.0043846312910318375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.033718500286340714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42888131737709045 + ], + "opposite_logprob_tokens": [ + -1.053881287574768 + ], + "logprob_reward": -0.42888131737709045, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42888131737709045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29568496346473694, + 0.09172429144382477, + 0.05383164808154106, + 0.10422317683696747, + 0.11408507078886032, + 0.08197241276502609, + 0.07790502905845642, + 0.05654309689998627, + 0.05792175978422165, + 0.06610862165689468 + ], + "credit_map": [ + 0.03445572033524513, + 0.03583711013197899, + 0.059448160231113434, + 0.02925642766058445, + 0.02421514317393303, + 0.02214350923895836, + 0.04356478527188301, + 0.04676409065723419, + 0.008332788944244385, + 0.01090233214199543, + 0.016025584191083908, + 0.013268554583191872, + 0.007995190098881721, + 0.010420416481792927, + 0.011999484151601791, + 0.012779931537806988, + 0.005755848716944456, + 0.006931454874575138, + 0.009652635082602501, + 0.00785902515053749, + 0.003832438960671425, + 0.005560065619647503, + 0.005263180006295443, + 0.008976996876299381, + 0.019101839512586594, + 0.013833576813340187, + 0.019957784563302994, + 0.011754448525607586, + 0.008882619440555573, + 0.00893379282206297, + 0.008800632320344448, + 0.0129584651440382, + 0.008602210320532322, + 0.0052457875572144985, + 0.03745045140385628, + 0.007237718906253576, + 0.004757975693792105, + 0.0061615160666406155, + 0.006321609485894442, + 0.0383077971637249, + 0.009195945225656033, + 0.004991746507585049, + 0.013879378326237202, + 0.009802991524338722, + 0.007629326079040766, + 0.00887236837297678, + 0.009354734793305397, + 0.018245918676257133, + 0.013273806311190128, + 0.0058197868056595325, + 0.01111218798905611, + 0.010357480496168137, + 0.00845011230558157, + 0.006715118419378996, + 0.010138124227523804, + 0.01203840970993042, + 0.005698308814316988, + 0.0032074437476694584, + 0.009736985899508, + 0.009600641205906868, + 0.00332363392226398, + 0.006012581288814545, + 0.005985540337860584, + 0.012977958656847477, + 0.007010283414274454, + 0.004098251927644014, + 0.008114496245980263, + 0.009397764690220356, + 0.006658659316599369, + 0.00804099254310131, + 0.0063638752326369286, + 0.008237432688474655, + 0.0047989520244300365, + 0.003706427989527583, + 0.011495820246636868, + 0.012903823517262936, + 0.00994743313640356, + 0.005802168045192957, + 0.0075357346795499325, + 0.009918262250721455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.42888131737709045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979600802063942 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.02979600802063942, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979600802063942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0804026871919632, + 0.05045256391167641, + 0.0850905105471611, + 0.06358238309621811, + 0.05141673609614372, + 0.10893938690423965, + 0.09577221423387527, + 0.23506829142570496, + 0.1851479411125183, + 0.044127337634563446 + ], + "credit_map": [ + 0.010696398094296455, + 0.011326252482831478, + 0.011667807586491108, + 0.008781402371823788, + 0.007328834384679794, + 0.006326556205749512, + 0.012249942868947983, + 0.012025479227304459, + 0.005313637666404247, + 0.005916330963373184, + 0.007725325413048267, + 0.0066236574202775955, + 0.004168080165982246, + 0.0053241983987390995, + 0.007082561030983925, + 0.008298765867948532, + 0.010631008073687553, + 0.010437404736876488, + 0.012740171514451504, + 0.014148855581879616, + 0.006417557131499052, + 0.006797252222895622, + 0.010661979205906391, + 0.013256274163722992, + 0.005898936651647091, + 0.010782935656607151, + 0.009249486029148102, + 0.006816215347498655, + 0.009357305243611336, + 0.006936617661267519, + 0.005616212263703346, + 0.008924667723476887, + 0.006289336830377579, + 0.006863479968160391, + 0.00637228274717927, + 0.006385421846061945, + 0.0064901201985776424, + 0.006519612390547991, + 0.00565409054979682, + 0.006842381786555052, + 0.012982022948563099, + 0.009215694852173328, + 0.019302990287542343, + 0.012180188670754433, + 0.010592889040708542, + 0.01103915087878704, + 0.012592297047376633, + 0.021034136414527893, + 0.012053346261382103, + 0.006962502375245094, + 0.012823810800909996, + 0.011721862480044365, + 0.01370044331997633, + 0.008471486158668995, + 0.016190653666853905, + 0.013848097063601017, + 0.01700197532773018, + 0.012959258630871773, + 0.044317517429590225, + 0.0437302403151989, + 0.013069051317870617, + 0.024959227070212364, + 0.021291228011250496, + 0.05773976445198059, + 0.022793829441070557, + 0.012094967998564243, + 0.026978764683008194, + 0.03162316977977753, + 0.020074741914868355, + 0.022220712155103683, + 0.0238198135048151, + 0.02554192952811718, + 0.0047457460314035416, + 0.0029528404120355844, + 0.005899947136640549, + 0.007327544502913952, + 0.007354182656854391, + 0.004653202369809151, + 0.005489218980073929, + 0.005704651586711407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.02979600802063942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233826637268066 + ], + "opposite_logprob_tokens": [ + -0.8983826637268066 + ], + "logprob_reward": -0.5233826637268066, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233826637268066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07785236090421677, + 0.029523201286792755, + 0.02388875186443329, + 0.046549346297979355, + 0.5359811186790466, + 0.036360546946525574, + 0.043533384799957275, + 0.11543039232492447, + 0.06353650987148285, + 0.02734445594251156 + ], + "credit_map": [ + 0.009094968438148499, + 0.009599046781659126, + 0.015841657295823097, + 0.008096275851130486, + 0.0055947634391486645, + 0.0051496922969818115, + 0.011912325397133827, + 0.012563633732497692, + 0.0027786260470747948, + 0.00357514969073236, + 0.005044862162321806, + 0.004326133523136377, + 0.003072521183639765, + 0.0035846063401550055, + 0.002795364009216428, + 0.0043459380976855755, + 0.0020107347518205643, + 0.0030006009619683027, + 0.004924776963889599, + 0.002918840851634741, + 0.0018930057995021343, + 0.0023049484007060528, + 0.0031001099850982428, + 0.003735733451321721, + 0.007100021466612816, + 0.004959899000823498, + 0.007287048734724522, + 0.005803624168038368, + 0.00404060073196888, + 0.004295762162655592, + 0.006271554622799158, + 0.006790837273001671, + 0.036416057497262955, + 0.012480036355555058, + 0.19782432913780212, + 0.030532551929354668, + 0.01508249994367361, + 0.02250274084508419, + 0.014058681204915047, + 0.20708417892456055, + 0.004853828810155392, + 0.00305926357395947, + 0.005585412960499525, + 0.004406854975968599, + 0.004262357484549284, + 0.004088085610419512, + 0.005196526646614075, + 0.004908217582851648, + 0.0043034665286540985, + 0.0028214554768055677, + 0.006319468840956688, + 0.005832367110997438, + 0.0061331759206950665, + 0.0031602573581039906, + 0.006762745324522257, + 0.008200451731681824, + 0.008272247388958931, + 0.0059027597308158875, + 0.021929049864411354, + 0.021376019343733788, + 0.005588086787611246, + 0.01316732820123434, + 0.010883805342018604, + 0.028311103582382202, + 0.006698459852486849, + 0.004418843891471624, + 0.009550574235618114, + 0.011767774820327759, + 0.006453045643866062, + 0.007318816613405943, + 0.008391749113798141, + 0.008937245234847069, + 0.0026222479064017534, + 0.002339980099350214, + 0.004109879024326801, + 0.004415580071508884, + 0.005200183019042015, + 0.0031086355447769165, + 0.002675095573067665, + 0.0028728547040373087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.5233826637268066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow-motion shot captured at a high frame rate, the camera begins a majestic spiral climb, gliding upwards along the skeleton of a massive, unfinished ship. The evening sun shone with a warm golden glow, penetrating a dense forest of metal scaffolding. Casting ever-shifting, elongated shadows across the cold, dark-gray steel ribs of the ship, a dynamic chessboard of light and shadow. As the camera gracefully swirls, light plays across rough welds and exposed rivets, evoking the raw texture of industrial construction. Deep within the hull structure, welding torches operated by unseen workers occasionally ignite with fleeting, brilliant sparks. These fiery orange sparks, like stars in the night sky, shimmer briefly in the twilight, momentarily illuminating a small patch of rusted steel before fading into deep shadow. This imposing yet tranquil industrial landscape imbued with a vibrant sense of movement, imbuing the image with the grand, poetic, and majestic beauty of industrial construction.", + "seed": 17572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_0/index_6482_seed_17572_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.22073881775140763, + "vlm_logprob": -0.22073881775140763, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.22073881775140763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04471098259091377 + ], + "opposite_logprob_tokens": [ + -3.419710874557495 + ], + "logprob_reward": -0.04471098259091377, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04471098259091377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38552185893058777, + 0.16964773833751678, + 0.06360818445682526, + 0.038430728018283844, + 0.14240437746047974, + 0.06622173637151718, + 0.04366173595190048, + 0.036433666944503784, + 0.02418963797390461, + 0.029880309477448463 + ], + "credit_map": [ + 0.04632578045129776, + 0.08181442320346832, + 0.041539423167705536, + 0.08320052921772003, + 0.049064721912145615, + 0.02449745498597622, + 0.028024690225720406, + 0.031054873019456863, + 0.019910549744963646, + 0.02275807596743107, + 0.025539850816130638, + 0.021374942734837532, + 0.01799842156469822, + 0.015336550772190094, + 0.026543451473116875, + 0.020185908302664757, + 0.009780262596905231, + 0.012885447591543198, + 0.009150191210210323, + 0.005850742105394602, + 0.006353573873639107, + 0.006494976580142975, + 0.0047647785395383835, + 0.008328219875693321, + 0.004707308020442724, + 0.0030999479349702597, + 0.005731607787311077, + 0.006391949485987425, + 0.005591464228928089, + 0.003568575717508793, + 0.0058536166325211525, + 0.003486264729872346, + 0.01885218918323517, + 0.021058470010757446, + 0.020747065544128418, + 0.016867995262145996, + 0.02289535105228424, + 0.014791138470172882, + 0.015836015343666077, + 0.011356170289218426, + 0.007785381283611059, + 0.0060253809206187725, + 0.010272950865328312, + 0.010500038042664528, + 0.00914708897471428, + 0.005539964884519577, + 0.008735804818570614, + 0.00821513868868351, + 0.008541925810277462, + 0.005169653799384832, + 0.007266258355230093, + 0.004369761794805527, + 0.00442215520888567, + 0.004108226392418146, + 0.0060299355536699295, + 0.003753828117623925, + 0.008626818656921387, + 0.004217342007905245, + 0.006526056211441755, + 0.003250754438340664, + 0.0032821153290569782, + 0.0039889574982225895, + 0.003384954994544387, + 0.0031566719990223646, + 0.003075595945119858, + 0.003218819387257099, + 0.0031639114022254944, + 0.0030273250304162502, + 0.002076819073408842, + 0.0027364676352590322, + 0.0027463289443403482, + 0.004144374746829271, + 0.003527143970131874, + 0.004916144534945488, + 0.0043724942952394485, + 0.003632499370723963, + 0.0033628600649535656, + 0.0037904076743870974, + 0.0037105181254446507, + 0.002568245166912675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04471098259091377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309137538075447 + ], + "opposite_logprob_tokens": [ + -3.168091297149658 + ], + "logprob_reward": -0.04309137538075447, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309137538075447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1934232860803604, + 0.12494109570980072, + 0.16855038702487946, + 0.05372031033039093, + 0.1343240588903427, + 0.03656768798828125, + 0.06463485211133957, + 0.04887550324201584, + 0.07673437148332596, + 0.09822840988636017 + ], + "credit_map": [ + 0.03651880845427513, + 0.03273480758070946, + 0.020553672686219215, + 0.026894191280007362, + 0.028054123744368553, + 0.015317333862185478, + 0.01380319893360138, + 0.019547149538993835, + 0.017285499721765518, + 0.017249761149287224, + 0.015740126371383667, + 0.015251167118549347, + 0.014742228202521801, + 0.015041587874293327, + 0.01504472829401493, + 0.01458599790930748, + 0.01728004775941372, + 0.039344459772109985, + 0.029280290007591248, + 0.015474509447813034, + 0.015042567625641823, + 0.014512441121041775, + 0.014424072578549385, + 0.02319200150668621, + 0.007873337715864182, + 0.005015726666897535, + 0.008969107642769814, + 0.006734102964401245, + 0.008536393754184246, + 0.005158454645425081, + 0.005499731749296188, + 0.005933454260230064, + 0.019424384459853172, + 0.014575393870472908, + 0.019067984074354172, + 0.01957702450454235, + 0.015783166512846947, + 0.0172931719571352, + 0.016838377341628075, + 0.011764557100832462, + 0.008604126051068306, + 0.0029654251411557198, + 0.006017295178025961, + 0.004771851468831301, + 0.003773342352360487, + 0.002679796190932393, + 0.004562408197671175, + 0.003193446435034275, + 0.011559221893548965, + 0.006785433739423752, + 0.010561187751591206, + 0.006050302181392908, + 0.006905755493789911, + 0.005977105814963579, + 0.009163472801446915, + 0.007632375694811344, + 0.005944410804659128, + 0.006079987622797489, + 0.006746080704033375, + 0.0054115415550768375, + 0.0077774664387106895, + 0.005400398280471563, + 0.006032734178006649, + 0.005482881795614958, + 0.015298580750823021, + 0.008028574287891388, + 0.010358023457229137, + 0.009414141066372395, + 0.006505447905510664, + 0.010026546195149422, + 0.009153451770544052, + 0.00794960092753172, + 0.015150888822972775, + 0.012218399904668331, + 0.014394542202353477, + 0.01710173487663269, + 0.007132814731448889, + 0.0148349329829216, + 0.008768146857619286, + 0.008626953698694706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.04309137538075447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233217477798462 + ], + "opposite_logprob_tokens": [ + -0.8983217477798462 + ], + "logprob_reward": -0.5233217477798462, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233217477798462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2953634560108185, + 0.17601658403873444, + 0.21845166385173798, + 0.05326872318983078, + 0.07662861794233322, + 0.04197252541780472, + 0.015720728784799576, + 0.05104733258485794, + 0.047441769391298294, + 0.024088626727461815 + ], + "credit_map": [ + 0.05193604528903961, + 0.05752056464552879, + 0.030112450942397118, + 0.04487933591008186, + 0.04251468926668167, + 0.02089211530983448, + 0.015956144779920578, + 0.03155212104320526, + 0.018702546134591103, + 0.023416122421622276, + 0.019847895950078964, + 0.017331518232822418, + 0.031791020184755325, + 0.023612862452864647, + 0.020351067185401917, + 0.020963560789823532, + 0.021213026717305183, + 0.054651886224746704, + 0.0456705242395401, + 0.019414138048887253, + 0.020671697333455086, + 0.018659787252545357, + 0.012881015427410603, + 0.025289569050073624, + 0.007704796735197306, + 0.006045654881745577, + 0.009112843312323093, + 0.010932717472314835, + 0.006204376462846994, + 0.003331404412165284, + 0.0057870917953550816, + 0.00414983369410038, + 0.011726480908691883, + 0.008117244578897953, + 0.010462836362421513, + 0.008579659275710583, + 0.011266701854765415, + 0.009797263890504837, + 0.010071292519569397, + 0.006607137620449066, + 0.007882080040872097, + 0.004991447553038597, + 0.006540194619446993, + 0.00381078221835196, + 0.006071618292480707, + 0.0038800507318228483, + 0.0033225365914404392, + 0.0054738144390285015, + 0.0029708854854106903, + 0.0017584628658369184, + 0.0026203980669379234, + 0.001413697493262589, + 0.0019824167247861624, + 0.001553672133013606, + 0.001755932462401688, + 0.0016652642516419291, + 0.009308027103543282, + 0.004829455632716417, + 0.008532158099114895, + 0.004734594374895096, + 0.007833564653992653, + 0.00568497134372592, + 0.004917367361485958, + 0.00520719401538372, + 0.005450573284178972, + 0.006446774583309889, + 0.0066045732237398624, + 0.005655120126903057, + 0.0064158085733652115, + 0.006592758931219578, + 0.004765808116644621, + 0.005510353483259678, + 0.0034283308777958155, + 0.003473518881946802, + 0.00370158813893795, + 0.0027053572703152895, + 0.0016091899015009403, + 0.003113033017143607, + 0.0031044723000377417, + 0.002953137271106243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.5233217477798462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818835735321045 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818835735321045, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818835735321045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1748904287815094, + 0.09181758761405945, + 0.32187315821647644, + 0.042090024799108505, + 0.08253247290849686, + 0.10738848894834518, + 0.04453020542860031, + 0.057339321821928024, + 0.04064224660396576, + 0.03689613193273544 + ], + "credit_map": [ + 0.023116420954465866, + 0.03449023514986038, + 0.017815308645367622, + 0.037658821791410446, + 0.023258700966835022, + 0.009461017325520515, + 0.00940126832574606, + 0.01968863233923912, + 0.012161615304648876, + 0.014194021001458168, + 0.01766718551516533, + 0.012225210666656494, + 0.009983312338590622, + 0.008264221251010895, + 0.008802838623523712, + 0.008519181981682777, + 0.025227326899766922, + 0.09407352656126022, + 0.07966164499521255, + 0.022413667291402817, + 0.0248614139854908, + 0.02025202289223671, + 0.01972583495080471, + 0.035657692700624466, + 0.007787832524627447, + 0.005473965313285589, + 0.005547506734728813, + 0.007449164055287838, + 0.00539217097684741, + 0.002694963477551937, + 0.0044378540478646755, + 0.0033065644092857838, + 0.012213400565087795, + 0.008398554287850857, + 0.01125643402338028, + 0.01313873752951622, + 0.010685993358492851, + 0.010975481010973454, + 0.008521522395312786, + 0.007342334371060133, + 0.02003871090710163, + 0.01109330914914608, + 0.01655169390141964, + 0.014105554670095444, + 0.011905833147466183, + 0.011759227141737938, + 0.009747044183313847, + 0.012187110260128975, + 0.007494641933590174, + 0.004817810375243425, + 0.007268799468874931, + 0.003932781517505646, + 0.004287770949304104, + 0.004120697733014822, + 0.007580542005598545, + 0.005027159117162228, + 0.011565291322767735, + 0.005974056199193001, + 0.010084382258355618, + 0.005127450916916132, + 0.005515154451131821, + 0.005888338200747967, + 0.006502735894173384, + 0.00668190885335207, + 0.0046195681206882, + 0.0070586781948804855, + 0.006183837540447712, + 0.006003059446811676, + 0.00462677888572216, + 0.004306313581764698, + 0.004886024631559849, + 0.00295798247680068, + 0.005024999380111694, + 0.003705594688653946, + 0.005262667313218117, + 0.006075041368603706, + 0.003593635279685259, + 0.004894542973488569, + 0.004619854036718607, + 0.003719794563949108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.01818835735321045 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47438162565231323 + ], + "opposite_logprob_tokens": [ + -0.9743816256523132 + ], + "logprob_reward": -0.47438162565231323, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47438162565231323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1753542572259903, + 0.07685346156358719, + 0.17875148355960846, + 0.046507179737091064, + 0.04102937504649162, + 0.1790466457605362, + 0.11545134335756302, + 0.07853026688098907, + 0.06457410007715225, + 0.04390188306570053 + ], + "credit_map": [ + 0.031118521466851234, + 0.028138121590018272, + 0.021252796053886414, + 0.026457345113158226, + 0.022893140092492104, + 0.014097313396632671, + 0.012303773313760757, + 0.019093234091997147, + 0.011859484948217869, + 0.011502395384013653, + 0.012015768326818943, + 0.010257747024297714, + 0.008650035597383976, + 0.008798090741038322, + 0.006767261307686567, + 0.007002679631114006, + 0.013668730854988098, + 0.052922748029232025, + 0.04556817561388016, + 0.01366166491061449, + 0.013224015943706036, + 0.010536182671785355, + 0.010852603241801262, + 0.018317367881536484, + 0.006105631589889526, + 0.0068160961382091045, + 0.007658520247787237, + 0.005688458681106567, + 0.007745148614048958, + 0.003478247206658125, + 0.005160841625183821, + 0.003854239359498024, + 0.006030166056007147, + 0.004523970186710358, + 0.005449043121188879, + 0.004585179965943098, + 0.006129059474915266, + 0.006190447136759758, + 0.004537392407655716, + 0.0035841146018356085, + 0.03661583736538887, + 0.01820426806807518, + 0.028439881280064583, + 0.019793709740042686, + 0.020803524181246758, + 0.018432026728987694, + 0.015733052045106888, + 0.021024351939558983, + 0.02143792062997818, + 0.012749020010232925, + 0.01973903737962246, + 0.01073429360985756, + 0.012595356442034245, + 0.0108236875385046, + 0.016618072986602783, + 0.010753956623375416, + 0.014246820472180843, + 0.011072247289121151, + 0.011810755357146263, + 0.007576718460768461, + 0.010967847891151905, + 0.008604221977293491, + 0.006953764706850052, + 0.007297887932509184, + 0.010042508132755756, + 0.008706220425665379, + 0.009634751826524734, + 0.0070330314338207245, + 0.005697730462998152, + 0.007527660112828016, + 0.006440367549657822, + 0.009491823613643646, + 0.007831453345716, + 0.006522688083350658, + 0.00609830766916275, + 0.006553697399795055, + 0.003372286446392536, + 0.005901183933019638, + 0.00358096556738019, + 0.004041299223899841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.47438162565231323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow-motion shot captured at a high frame rate, the camera begins a majestic spiral climb, gliding upwards along the skeleton of a massive, unfinished ship. The evening sun shone with a warm golden glow, penetrating a dense forest of metal scaffolding. Casting ever-shifting, elongated shadows across the cold, dark-gray steel ribs of the ship, a dynamic chessboard of light and shadow. As the camera gracefully swirls, light plays across rough welds and exposed rivets, evoking the raw texture of industrial construction. Deep within the hull structure, welding torches operated by unseen workers occasionally ignite with fleeting, brilliant sparks. These fiery orange sparks, like stars in the night sky, shimmer briefly in the twilight, momentarily illuminating a small patch of rusted steel before fading into deep shadow. This imposing yet tranquil industrial landscape imbued with a vibrant sense of movement, imbuing the image with the grand, poetic, and majestic beauty of industrial construction.", + "seed": 17573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_0/index_6482_seed_17573_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15435565039515495, + "vlm_logprob": -0.15435565039515495, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.15435565039515495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20948293805122375 + ], + "opposite_logprob_tokens": [ + -1.7094829082489014 + ], + "logprob_reward": -0.20948293805122375, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20948293805122375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09330874681472778, + 0.15939396619796753, + 0.05617538467049599, + 0.0954204797744751, + 0.11881769448518753, + 0.044753991067409515, + 0.05320439115166664, + 0.05979916453361511, + 0.19527754187583923, + 0.12384860217571259 + ], + "credit_map": [ + 0.011098991148173809, + 0.010690762661397457, + 0.012133157812058926, + 0.013995317742228508, + 0.013655396178364754, + 0.016033172607421875, + 0.009126538410782814, + 0.0065754069946706295, + 0.022541841492056847, + 0.010184354148805141, + 0.024714551866054535, + 0.021950170397758484, + 0.01356915570795536, + 0.03065056912600994, + 0.019571393728256226, + 0.016211941838264465, + 0.006156567484140396, + 0.009955642744898796, + 0.007752156350761652, + 0.00999915786087513, + 0.006935618352144957, + 0.0041437153704464436, + 0.0068201026879251, + 0.0044124252162873745, + 0.008737253025174141, + 0.011681226082146168, + 0.012169093824923038, + 0.015751073136925697, + 0.016669828444719315, + 0.009355612099170685, + 0.01020672544836998, + 0.0108496667817235, + 0.00845244713127613, + 0.010474122129380703, + 0.013473612256348133, + 0.04082115739583969, + 0.010636954568326473, + 0.009254967793822289, + 0.010514810681343079, + 0.01518961600959301, + 0.006172372028231621, + 0.0066072954796254635, + 0.005397360306233168, + 0.00522055895999074, + 0.004978686571121216, + 0.0063180262222886086, + 0.004277354571968317, + 0.0057823387905955315, + 0.013692469336092472, + 0.0052358414977788925, + 0.0038645174354314804, + 0.002784744603559375, + 0.0027805303689092398, + 0.002450101776048541, + 0.009251140989363194, + 0.013145039789378643, + 0.005337982438504696, + 0.00701857777312398, + 0.01617616042494774, + 0.006078470963984728, + 0.0050056371837854385, + 0.003027914557605982, + 0.006689152214676142, + 0.010465266183018684, + 0.014966120943427086, + 0.01833631470799446, + 0.05886983498930931, + 0.02357139252126217, + 0.015132151544094086, + 0.014138548634946346, + 0.01259677391499281, + 0.03766641765832901, + 0.01191523764282465, + 0.031233886256814003, + 0.01882190816104412, + 0.010935487225651741, + 0.010547423735260963, + 0.005570946726948023, + 0.006795274093747139, + 0.028028422966599464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.20948293805122375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020610567182302475 + ], + "opposite_logprob_tokens": [ + -3.8956105709075928 + ], + "logprob_reward": -0.020610567182302475, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020610567182302475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08601148426532745, + 0.13070698082447052, + 0.06497833877801895, + 0.029620805755257607, + 0.04723232612013817, + 0.05049225687980652, + 0.180311381816864, + 0.10092464089393616, + 0.05387014150619507, + 0.255851686000824 + ], + "credit_map": [ + 0.009927090257406235, + 0.013464336283504963, + 0.009882505983114243, + 0.018894154578447342, + 0.007712367456406355, + 0.008605444803833961, + 0.010193990543484688, + 0.007331592962145805, + 0.014937370084226131, + 0.012230467982590199, + 0.020390553399920464, + 0.022183677181601524, + 0.011698191054165363, + 0.02092721499502659, + 0.015485212206840515, + 0.012854304164648056, + 0.008166125044226646, + 0.012124183587729931, + 0.0075347647070884705, + 0.010231728665530682, + 0.007649676408618689, + 0.00688818609341979, + 0.008630634285509586, + 0.0037530367262661457, + 0.002872027922421694, + 0.002860297914594412, + 0.0030695784371346235, + 0.0059556919150054455, + 0.00497245229780674, + 0.002392343943938613, + 0.005052067339420319, + 0.0024463478475809097, + 0.003234642557799816, + 0.004452839028090239, + 0.0060706413350999355, + 0.0035929589066654444, + 0.007830488495528698, + 0.004983025602996349, + 0.008006582036614418, + 0.009061147458851337, + 0.01415980327874422, + 0.019680282101035118, + 0.002725783037021756, + 0.002746580634266138, + 0.004407400265336037, + 0.0029278197325766087, + 0.0019285924499854445, + 0.0019159965449944139, + 0.008791028521955013, + 0.006686690263450146, + 0.009609120897948742, + 0.00831698626279831, + 0.011577135883271694, + 0.0075850277207791805, + 0.08334161341190338, + 0.04440378397703171, + 0.009042955935001373, + 0.0156362596899271, + 0.02427201345562935, + 0.014661637134850025, + 0.011214515194296837, + 0.005683614406734705, + 0.012328475713729858, + 0.008085168898105621, + 0.004874615464359522, + 0.007575292140245438, + 0.014089411124587059, + 0.007520490791648626, + 0.004978285636752844, + 0.003675815649330616, + 0.005982815753668547, + 0.005173420999199152, + 0.022924352437257767, + 0.07967853546142578, + 0.02291637472808361, + 0.015779128298163414, + 0.025614861398935318, + 0.010484057478606701, + 0.010797352530062199, + 0.06765702366828918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.020610567182302475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900442183017731 + ], + "opposite_logprob_tokens": [ + -2.5790045261383057 + ], + "logprob_reward": -0.07900442183017731, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900442183017731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04775363206863403, + 0.11750249564647675, + 0.019108157604932785, + 0.028718965128064156, + 0.04396890476346016, + 0.14290417730808258, + 0.1399150937795639, + 0.3711591958999634, + 0.02658078446984291, + 0.06238855421543121 + ], + "credit_map": [ + 0.005824211519211531, + 0.006387957837432623, + 0.005425771698355675, + 0.010137256234884262, + 0.005380813963711262, + 0.004386105574667454, + 0.005767837166786194, + 0.0044436766766011715, + 0.018728626891970634, + 0.008620725013315678, + 0.018879346549510956, + 0.018539629876613617, + 0.009874221868813038, + 0.018216749653220177, + 0.01218903623521328, + 0.012454154901206493, + 0.0020885758567601442, + 0.002352682175114751, + 0.0027474919334053993, + 0.0023043849505484104, + 0.0019893005955964327, + 0.0024788680020719767, + 0.002773172687739134, + 0.002373681403696537, + 0.003054422326385975, + 0.003236085409298539, + 0.0042477198876440525, + 0.0026176790706813335, + 0.004399375058710575, + 0.004047555383294821, + 0.004646926186978817, + 0.0024692018050700426, + 0.0025818180292844772, + 0.0036892062053084373, + 0.005713299382477999, + 0.015009748749434948, + 0.00534904794767499, + 0.003667345503345132, + 0.0037678086664527655, + 0.00419063214212656, + 0.037144679576158524, + 0.057091787457466125, + 0.005138433538377285, + 0.007832529954612255, + 0.01240250188857317, + 0.007803722750395536, + 0.0054635279811918736, + 0.010026995092630386, + 0.010917318984866142, + 0.006947023328393698, + 0.006739875301718712, + 0.005711779929697514, + 0.00829655397683382, + 0.005924183409661055, + 0.05685082823038101, + 0.03852752596139908, + 0.02964683435857296, + 0.058154504746198654, + 0.06467954814434052, + 0.05953506380319595, + 0.054222408682107925, + 0.02006261982023716, + 0.05829882249236107, + 0.02655937895178795, + 0.002810038859024644, + 0.003909894730895758, + 0.0034063300117850304, + 0.003298899857327342, + 0.005915387999266386, + 0.001969974022358656, + 0.003475819481536746, + 0.0017944391584023833, + 0.007962203584611416, + 0.022837627679109573, + 0.009145830757915974, + 0.004128213506191969, + 0.009228400886058807, + 0.0030708685517311096, + 0.0033248630352318287, + 0.002690550172701478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.07900442183017731 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370685875415802 + ], + "opposite_logprob_tokens": [ + -3.4087069034576416 + ], + "logprob_reward": -0.03370685875415802, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370685875415802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09615525603294373, + 0.21110308170318604, + 0.04656032845377922, + 0.05962004512548447, + 0.10206284373998642, + 0.11935046315193176, + 0.12496895343065262, + 0.028402643278241158, + 0.041086774319410324, + 0.17068959772586823 + ], + "credit_map": [ + 0.01283360831439495, + 0.01460317987948656, + 0.012215728871524334, + 0.012350082397460938, + 0.01459219679236412, + 0.01763082481920719, + 0.006353279110044241, + 0.005576355382800102, + 0.023546608164906502, + 0.01867034286260605, + 0.03278929367661476, + 0.032457463443279266, + 0.02000896818935871, + 0.03733885660767555, + 0.02501625195145607, + 0.021275311708450317, + 0.005036054644733667, + 0.0059961737133562565, + 0.005164526402950287, + 0.007518196478486061, + 0.006957031786441803, + 0.006538817659020424, + 0.004875079728662968, + 0.004474451299756765, + 0.008108637295663357, + 0.006115585099905729, + 0.007946138270199299, + 0.007421093061566353, + 0.010353735648095608, + 0.007488133851438761, + 0.007313164882361889, + 0.004873556084930897, + 0.006694619078189135, + 0.013741402886807919, + 0.01120852306485176, + 0.009460929781198502, + 0.02092668041586876, + 0.010720483027398586, + 0.011653265915811062, + 0.017656946554780006, + 0.03259214386343956, + 0.042061060667037964, + 0.006856678519397974, + 0.0074723088182508945, + 0.009197771549224854, + 0.008403103798627853, + 0.005715230479836464, + 0.007052171044051647, + 0.008069765754044056, + 0.006492391228675842, + 0.0064883180893957615, + 0.005433227401226759, + 0.008151267655193806, + 0.005427168682217598, + 0.05235747992992401, + 0.03254934027791023, + 0.003342007054015994, + 0.006365934852510691, + 0.003555822651833296, + 0.0032599144615232944, + 0.002352959243580699, + 0.0028396747075021267, + 0.0028681885451078415, + 0.0038181457202881575, + 0.0040344614535570145, + 0.005242775660008192, + 0.010610545985400677, + 0.005151084158569574, + 0.004683679901063442, + 0.00304599036462605, + 0.004251483362168074, + 0.004066759254783392, + 0.02013021521270275, + 0.061022549867630005, + 0.012551728636026382, + 0.008354539051651955, + 0.030102523043751717, + 0.008626123890280724, + 0.008967838250100613, + 0.020934103056788445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.03370685875415802 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, unfinished ship structure made of metal scaffolding?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by an evening sun with a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there visible welding torches igniting with fiery orange sparks inside the hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a spiral climb gliding upwards along the ship's skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion with a high frame rate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289734661579132 + ], + "opposite_logprob_tokens": [ + -1.0539734363555908 + ], + "logprob_reward": -0.4289734661579132, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289734661579132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10422493517398834, + 0.1357816755771637, + 0.05519777908921242, + 0.059781547635793686, + 0.11062696576118469, + 0.09400409460067749, + 0.19961664080619812, + 0.12769955396652222, + 0.044024959206581116, + 0.06904184818267822 + ], + "credit_map": [ + 0.013970887288451195, + 0.017774680629372597, + 0.011671519838273525, + 0.014110089279711246, + 0.016677383333444595, + 0.014989731833338737, + 0.008701478131115437, + 0.006329167168587446, + 0.013971925713121891, + 0.013996278867125511, + 0.020845498889684677, + 0.02290779910981655, + 0.013889108784496784, + 0.020450808107852936, + 0.016273433342576027, + 0.013446819968521595, + 0.006184048485010862, + 0.006129057612270117, + 0.0085188839584589, + 0.006853311788290739, + 0.007674157619476318, + 0.004309393931180239, + 0.008813644759356976, + 0.006715277675539255, + 0.00474927294999361, + 0.005793335847556591, + 0.006364110391587019, + 0.009035521186888218, + 0.010272212326526642, + 0.007246613502502441, + 0.007333623245358467, + 0.008986861445009708, + 0.007457856088876724, + 0.00826459564268589, + 0.010790573433041573, + 0.03603120520710945, + 0.01048253569751978, + 0.008564628660678864, + 0.011868513189256191, + 0.017167050391435623, + 0.013476085849106312, + 0.03049761988222599, + 0.005120472516864538, + 0.005501721985638142, + 0.009226878173649311, + 0.010959810577332973, + 0.007203368470072746, + 0.012018132023513317, + 0.021008430048823357, + 0.006165037397295237, + 0.01035253144800663, + 0.009456180967390537, + 0.011756934225559235, + 0.00836724042892456, + 0.09320343285799026, + 0.03930686041712761, + 0.010161404497921467, + 0.020503373816609383, + 0.028118999674916267, + 0.019276604056358337, + 0.019366638734936714, + 0.007358897011727095, + 0.01535609271377325, + 0.007557551376521587, + 0.0043695103377103806, + 0.0067279282957315445, + 0.007885461673140526, + 0.005333147943019867, + 0.007877818308770657, + 0.003323683748021722, + 0.0047579300589859486, + 0.0037494797725230455, + 0.007862863130867481, + 0.009469092823565006, + 0.011151602491736412, + 0.0052461265586316586, + 0.009033831767737865, + 0.006056390702724457, + 0.007926895283162594, + 0.012295041233301163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.4289734661579132 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_tensor([8381, 8381])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_tensor([8381, 8381])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a39b583ba51c5116bf368a5c485674c020dd7369 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_tensor([8381, 8381])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic time-lapse sequence, shot from a low angle and kept largely static to emphasize the epic scale, documents the silent, slow-motion battle between two antagonistic trees. In the foreground, a gnarled, ancient oak tree with dark, deeply furrowed bark covered in patches of green moss aggressively extends its thick, muscular limbs. Beside it, a taller, thinner birch tree, its pale bark peeling, twists its branches in a sinuous, relentless pursuit of any available gap in the canopy. The camera captures their gradual intertwining growth over many seasons; branches jostling, encircling, and strategically blocking each other in a relentless struggle for sunlight. The background reveals a dense, out-of-focus forest, while the forest floor below rapidly blurs from a carpet of summer ferns to autumnal leaves and a winter blanket of snow and ice. Golden light penetrates the treetops, moving swiftly as the sun arcs across the sky, creating a high-contrast, dramatic play of light and shadow, highlighting the texture of the bark and the slow, deliberate dance of the branches. The overall atmosphere is one of fierce, silent and primal competition, a beautiful yet brutal natural ballet of growth and survival.", + "seed": 13650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_8381_seed_13650_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15287680393084885, + "vlm_logprob": -0.15287680393084885, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.15287680393084885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35257720947265625 + ], + "opposite_logprob_tokens": [ + -1.7275772094726562 + ], + "logprob_reward": -0.35257720947265625, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.35257720947265625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1121005117893219, + 0.049759432673454285, + 0.06343033164739609, + 0.10587579756975174, + 0.19205635786056519, + 0.042657509446144104, + 0.16264215111732483, + 0.12510673701763153, + 0.06510995328426361, + 0.08126120269298553 + ], + "credit_map": [ + 0.009722962975502014, + 0.011000625789165497, + 0.019381975755095482, + 0.01746549643576145, + 0.00867494847625494, + 0.0075004794634878635, + 0.020034225657582283, + 0.01831979677081108, + 0.004201108124107122, + 0.004217464942485094, + 0.008053363300859928, + 0.007990886457264423, + 0.003567075589671731, + 0.0029671834781765938, + 0.013262149877846241, + 0.005500203464180231, + 0.005552289541810751, + 0.007627458311617374, + 0.01592349074780941, + 0.010639962740242481, + 0.003704866860061884, + 0.0035386409144848585, + 0.0067677246406674385, + 0.009675898589193821, + 0.007243865169584751, + 0.033525120466947556, + 0.03119766153395176, + 0.011091476306319237, + 0.005205216351896524, + 0.0038587504532188177, + 0.006336947903037071, + 0.007416755426675081, + 0.013990219682455063, + 0.02037668600678444, + 0.025885649025440216, + 0.02299407869577408, + 0.035131052136421204, + 0.012227604165673256, + 0.039211440831422806, + 0.022239629179239273, + 0.0033555161207914352, + 0.005167905706912279, + 0.005058765411376953, + 0.00560775538906455, + 0.005006546154618263, + 0.0034867851063609123, + 0.00656040757894516, + 0.008413825184106827, + 0.011918172240257263, + 0.01766047812998295, + 0.025141064077615738, + 0.024836037307977676, + 0.01675938442349434, + 0.012164807878434658, + 0.03263133391737938, + 0.021530870348215103, + 0.011857499368488789, + 0.010566866025328636, + 0.028466522693634033, + 0.023048337548971176, + 0.008808223530650139, + 0.007568477652966976, + 0.015874139964580536, + 0.01891668140888214, + 0.006303914822638035, + 0.007976801134645939, + 0.011714800260961056, + 0.01177071314305067, + 0.006433798465877771, + 0.007355190347880125, + 0.006269190926104784, + 0.007285543251782656, + 0.00979791209101677, + 0.01380656287074089, + 0.017648829147219658, + 0.01604047603905201, + 0.0050881230272352695, + 0.00630413880571723, + 0.005457230377942324, + 0.007117930334061384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.35257720947265625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256576120853424 + ], + "opposite_logprob_tokens": [ + -1.60065758228302 + ], + "logprob_reward": -0.2256576120853424, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256576120853424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0769275426864624, + 0.03747346252202988, + 0.029981398954987526, + 0.10641738772392273, + 0.06829210370779037, + 0.021825319156050682, + 0.2368043214082718, + 0.11831769347190857, + 0.18680888414382935, + 0.11715192347764969 + ], + "credit_map": [ + 0.004587944597005844, + 0.006278913468122482, + 0.011345293372869492, + 0.011170695535838604, + 0.005557205528020859, + 0.0034778823610395193, + 0.02255859225988388, + 0.011951007880270481, + 0.0021662029903382063, + 0.002847568830475211, + 0.005849134176969528, + 0.0060004182159900665, + 0.0025085618253797293, + 0.0011130067287012935, + 0.012144879437983036, + 0.004843686707317829, + 0.0023901115637272596, + 0.005494331941008568, + 0.005548895802348852, + 0.005580688826739788, + 0.0018952045356854796, + 0.001740411389619112, + 0.0030017744284123182, + 0.004329979419708252, + 0.00712022976949811, + 0.043231889605522156, + 0.021409165114164352, + 0.015335097908973694, + 0.004755849950015545, + 0.003922323230654001, + 0.006363581866025925, + 0.004279240965843201, + 0.004369741305708885, + 0.007771984674036503, + 0.00859313178807497, + 0.008428503759205341, + 0.012702620588243008, + 0.0040991478599607944, + 0.014047183096408844, + 0.008279779925942421, + 0.002589552430436015, + 0.0024989531375467777, + 0.0032805672381073236, + 0.004206902347505093, + 0.002562266308814287, + 0.0014851901214569807, + 0.0026959720999002457, + 0.002505912445485592, + 0.024074463173747063, + 0.023796552792191505, + 0.05519366264343262, + 0.04070565849542618, + 0.021586397662758827, + 0.01725180447101593, + 0.023639293387532234, + 0.030556470155715942, + 0.011884985491633415, + 0.009587788954377174, + 0.01951613649725914, + 0.019582221284508705, + 0.013083224184811115, + 0.008183589205145836, + 0.020666232332587242, + 0.015813494101166725, + 0.018849676474928856, + 0.017008613795042038, + 0.03919897973537445, + 0.03626781702041626, + 0.022162213921546936, + 0.011181753128767014, + 0.023300087079405785, + 0.018839724361896515, + 0.008988650515675545, + 0.01057581789791584, + 0.03187950700521469, + 0.03623194247484207, + 0.005368920974433422, + 0.005266007501631975, + 0.010741121135652065, + 0.008099940605461597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.2256576120853424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025215516798198223 + ], + "opposite_logprob_tokens": [ + -6.002521514892578 + ], + "logprob_reward": -0.0025215516798198223, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025215516798198223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10757658630609512, + 0.06906447559595108, + 0.05641906335949898, + 0.23111335933208466, + 0.03918445110321045, + 0.03227005526423454, + 0.17940951883792877, + 0.112169049680233, + 0.09425382316112518, + 0.07853960990905762 + ], + "credit_map": [ + 0.012617465108633041, + 0.011629701592028141, + 0.017117656767368317, + 0.016173090785741806, + 0.006603761576116085, + 0.005891790613532066, + 0.021463681012392044, + 0.01607944443821907, + 0.005672842264175415, + 0.004319603089243174, + 0.009422346018254757, + 0.009838152676820755, + 0.004805880133062601, + 0.0033901098649948835, + 0.023205609992146492, + 0.00840992946177721, + 0.004067603964358568, + 0.006302074063569307, + 0.014048530720174313, + 0.010858514346182346, + 0.0041788117960095406, + 0.0032654365058988333, + 0.007499717175960541, + 0.006198377814143896, + 0.013144693337380886, + 0.07692018151283264, + 0.08070231974124908, + 0.028999757021665573, + 0.009442951530218124, + 0.006107083056122065, + 0.009607513435184956, + 0.006188871804624796, + 0.00320909614674747, + 0.0031246929429471493, + 0.008473196066915989, + 0.008649605326354504, + 0.004591039381921291, + 0.002001872519031167, + 0.004349220544099808, + 0.0047857328318059444, + 0.0023984520230442286, + 0.0033637071028351784, + 0.0053012790158391, + 0.0057088700123131275, + 0.003732152981683612, + 0.0021387115120887756, + 0.004330497235059738, + 0.00529638584703207, + 0.018679723143577576, + 0.02267715148627758, + 0.05181144177913666, + 0.02730230987071991, + 0.009892602451145649, + 0.012022354640066624, + 0.020458798855543137, + 0.01656513847410679, + 0.008510829880833626, + 0.009745774790644646, + 0.026775794103741646, + 0.019239526242017746, + 0.008555551990866661, + 0.0052746678702533245, + 0.01608138345181942, + 0.017985528334975243, + 0.00850693415850401, + 0.009162899106740952, + 0.02221374213695526, + 0.01960018463432789, + 0.00906326249241829, + 0.007613480556756258, + 0.010031962767243385, + 0.008061359636485577, + 0.006179256830364466, + 0.008854947052896023, + 0.0133778415620327, + 0.01881556212902069, + 0.006253985222429037, + 0.005425822455435991, + 0.01153088454157114, + 0.008101317100226879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.0025215516798198223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280154913663864 + ], + "opposite_logprob_tokens": [ + -3.773280143737793 + ], + "logprob_reward": -0.023280154913663864, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280154913663864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1438201367855072, + 0.10715239495038986, + 0.0836312547326088, + 0.053464822471141815, + 0.1438412219285965, + 0.02696469984948635, + 0.1358337104320526, + 0.15749460458755493, + 0.06333326548337936, + 0.08446390926837921 + ], + "credit_map": [ + 0.01666387915611267, + 0.016855955123901367, + 0.022548291832208633, + 0.022009333595633507, + 0.010836114175617695, + 0.008408179506659508, + 0.02719934657216072, + 0.019299035891890526, + 0.0074308449402451515, + 0.011434475891292095, + 0.015995465219020844, + 0.0191238634288311, + 0.007111512124538422, + 0.00696235615760088, + 0.027503833174705505, + 0.011590039357542992, + 0.006537807639688253, + 0.010312583297491074, + 0.020916622132062912, + 0.014551651664078236, + 0.005317783914506435, + 0.004314654041081667, + 0.011157840490341187, + 0.01052230503410101, + 0.00346311810426414, + 0.008166766725480556, + 0.013122639618813992, + 0.011300851590931416, + 0.0037292735651135445, + 0.0027093160897493362, + 0.005582268349826336, + 0.005390590988099575, + 0.012078522704541683, + 0.011971067637205124, + 0.03344643488526344, + 0.03859582170844078, + 0.010793473571538925, + 0.0069736819714307785, + 0.013692437671124935, + 0.01628977619111538, + 0.0020038436632603407, + 0.0029942835681140423, + 0.005111861974000931, + 0.004465138539671898, + 0.002510619815438986, + 0.0014048240846022964, + 0.003748982911929488, + 0.004725146107375622, + 0.011482303030788898, + 0.016660170629620552, + 0.02898283116519451, + 0.0185414869338274, + 0.010200063697993755, + 0.008850044570863247, + 0.028790542855858803, + 0.012326267547905445, + 0.014782315120100975, + 0.019951045513153076, + 0.039348386228084564, + 0.022967146709561348, + 0.008112797513604164, + 0.010646428912878036, + 0.01470640767365694, + 0.0269800815731287, + 0.005951069761067629, + 0.005157812032848597, + 0.016068778932094574, + 0.013556997291743755, + 0.003509719157591462, + 0.004835470113903284, + 0.007542349398136139, + 0.006711073685437441, + 0.00774384755641222, + 0.008204801939427853, + 0.014795994386076927, + 0.021179035305976868, + 0.006524881813675165, + 0.005853812675923109, + 0.012028341181576252, + 0.008133196271955967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.023280154913663864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034749150276184 + ], + "opposite_logprob_tokens": [ + -1.9103474617004395 + ], + "logprob_reward": -0.16034749150276184, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034749150276184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12283039093017578, + 0.057699527591466904, + 0.05490817874670029, + 0.0951632708311081, + 0.044402703642845154, + 0.05944184213876724, + 0.19685059785842896, + 0.11081746965646744, + 0.15037447214126587, + 0.10751157253980637 + ], + "credit_map": [ + 0.01468174159526825, + 0.009276090189814568, + 0.01573409140110016, + 0.015307541936635971, + 0.008999503217637539, + 0.007287145592272282, + 0.03518379479646683, + 0.016360478475689888, + 0.004105415660887957, + 0.004475289955735207, + 0.007935072295367718, + 0.008875437080860138, + 0.003407340496778488, + 0.0019172391621395946, + 0.02098354697227478, + 0.006000180263072252, + 0.004015279933810234, + 0.011867189779877663, + 0.010397888720035553, + 0.010489272885024548, + 0.0037976556923240423, + 0.0030014123767614365, + 0.005512853153049946, + 0.005826618056744337, + 0.005910955369472504, + 0.037487272173166275, + 0.022264186292886734, + 0.01085711270570755, + 0.004505888558924198, + 0.003436222905293107, + 0.006774551700800657, + 0.003927074838429689, + 0.0030565750785171986, + 0.004426891449838877, + 0.006160857621580362, + 0.005778834689408541, + 0.008079425431787968, + 0.0026109046302735806, + 0.009132924489676952, + 0.00515628419816494, + 0.0050598857924342155, + 0.004979205317795277, + 0.006696745287626982, + 0.007141391281038523, + 0.007999664172530174, + 0.004684551153331995, + 0.011967538855969906, + 0.010912854224443436, + 0.022191490978002548, + 0.017050139605998993, + 0.05177746340632439, + 0.03139663115143776, + 0.014904908835887909, + 0.013071989640593529, + 0.029128341004252434, + 0.017329612746834755, + 0.007581540383398533, + 0.009040775708854198, + 0.02424635738134384, + 0.019637662917375565, + 0.006359742023050785, + 0.005131889600306749, + 0.01425953023135662, + 0.02455996349453926, + 0.0156459491699934, + 0.013050824403762817, + 0.03373131528496742, + 0.03111163154244423, + 0.009980883449316025, + 0.0101480046287179, + 0.018613191321492195, + 0.018092667683959007, + 0.01254338026046753, + 0.011485003866255283, + 0.021588174626231194, + 0.018374912440776825, + 0.009709758684039116, + 0.008440503850579262, + 0.015593587420880795, + 0.009776243008673191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.16034749150276184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic time-lapse sequence, shot from a low angle and kept largely static to emphasize the epic scale, documents the silent, slow-motion battle between two antagonistic trees. In the foreground, a gnarled, ancient oak tree with dark, deeply furrowed bark covered in patches of green moss aggressively extends its thick, muscular limbs. Beside it, a taller, thinner birch tree, its pale bark peeling, twists its branches in a sinuous, relentless pursuit of any available gap in the canopy. The camera captures their gradual intertwining growth over many seasons; branches jostling, encircling, and strategically blocking each other in a relentless struggle for sunlight. The background reveals a dense, out-of-focus forest, while the forest floor below rapidly blurs from a carpet of summer ferns to autumnal leaves and a winter blanket of snow and ice. Golden light penetrates the treetops, moving swiftly as the sun arcs across the sky, creating a high-contrast, dramatic play of light and shadow, highlighting the texture of the bark and the slow, deliberate dance of the branches. The overall atmosphere is one of fierce, silent and primal competition, a beautiful yet brutal natural ballet of growth and survival.", + "seed": 13651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_8381_seed_13651_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19957087524235248, + "vlm_logprob": -0.19957087524235248, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.19957087524235248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.519294261932373 + ], + "opposite_logprob_tokens": [ + -1.269294261932373 + ], + "logprob_reward": -0.519294261932373, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.519294261932373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5085045695304871, + 0.04305804520845413, + 0.031078187748789787, + 0.04428248479962349, + 0.047370657324790955, + 0.061269938945770264, + 0.032883498817682266, + 0.031628645956516266, + 0.06947988271713257, + 0.13044406473636627 + ], + "credit_map": [ + 0.03974856063723564, + 0.04075383394956589, + 0.0841652974486351, + 0.09891439974308014, + 0.04102952405810356, + 0.07197531312704086, + 0.06612658500671387, + 0.06579108536243439, + 0.0023671728558838367, + 0.0029900295194238424, + 0.006242766510695219, + 0.0052097635343670845, + 0.0037006933707743883, + 0.007874621078372002, + 0.0077349222265183926, + 0.006938077975064516, + 0.0027181035839021206, + 0.0024274466559290886, + 0.0044815558940172195, + 0.0042560207657516, + 0.0035738633014261723, + 0.00613526813685894, + 0.004051137249916792, + 0.0034347933251410723, + 0.0033634151332080364, + 0.0037122536450624466, + 0.005949388723820448, + 0.007611872628331184, + 0.0032871963921934366, + 0.007588453125208616, + 0.005344050470739603, + 0.007425851654261351, + 0.0036016067024320364, + 0.005550878122448921, + 0.007375065237283707, + 0.010620342567563057, + 0.004103022161871195, + 0.0051420810632407665, + 0.006717141717672348, + 0.004260517656803131, + 0.00546761741861701, + 0.005520916078239679, + 0.008566212840378284, + 0.010036041028797626, + 0.005870908033102751, + 0.006875940598547459, + 0.010629178956151009, + 0.008303121663630009, + 0.0029670868534594774, + 0.002287916373461485, + 0.004399093799293041, + 0.0051452601328492165, + 0.003024191129952669, + 0.003624353324994445, + 0.006264250259846449, + 0.005171346012502909, + 0.0047631580382585526, + 0.0025230238679796457, + 0.0050572678446769714, + 0.00455477787181735, + 0.0029009885620325804, + 0.003615419380366802, + 0.005182131193578243, + 0.003031876403838396, + 0.0060389540158212185, + 0.00876727793365717, + 0.010746256448328495, + 0.01895926147699356, + 0.0044145588763058186, + 0.00638966728001833, + 0.00801941379904747, + 0.006144494283944368, + 0.011243625544011593, + 0.010804972611367702, + 0.01894996128976345, + 0.02941269241273403, + 0.008502312935888767, + 0.018363740295171738, + 0.017243998125195503, + 0.015922769904136658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.519294261932373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16044700145721436 + ], + "opposite_logprob_tokens": [ + -1.9104470014572144 + ], + "logprob_reward": -0.16044700145721436, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16044700145721436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19845329225063324, + 0.13715402781963348, + 0.06580425798892975, + 0.07779508084058762, + 0.08601190894842148, + 0.09358981996774673, + 0.07474517077207565, + 0.11013295501470566, + 0.05942294001579285, + 0.09689048677682877 + ], + "credit_map": [ + 0.020031388849020004, + 0.018546054139733315, + 0.03319411352276802, + 0.0468100905418396, + 0.013331635855138302, + 0.02393692545592785, + 0.027629153802990913, + 0.014973935671150684, + 0.008329886011779308, + 0.011614673770964146, + 0.021366050466895103, + 0.024056216701865196, + 0.013662715442478657, + 0.025857655331492424, + 0.014063083566725254, + 0.018203765153884888, + 0.0055544148199260235, + 0.005224564112722874, + 0.00967456679791212, + 0.008737975731492043, + 0.006834468338638544, + 0.009016546420753002, + 0.012628281489014626, + 0.008133447729051113, + 0.004177910275757313, + 0.004805763252079487, + 0.012854181230068207, + 0.017367713153362274, + 0.005514041054993868, + 0.008891180157661438, + 0.01564911939203739, + 0.008535174652934074, + 0.004941401071846485, + 0.009312269277870655, + 0.012686602771282196, + 0.022450892254710197, + 0.010520859621465206, + 0.007948829792439938, + 0.009204699657857418, + 0.00894636008888483, + 0.010122637264430523, + 0.00964223314076662, + 0.011277319863438606, + 0.015570060350000858, + 0.010782892815768719, + 0.009250995703041553, + 0.01426252257078886, + 0.012681175954639912, + 0.006749140564352274, + 0.005010741297155619, + 0.010778404772281647, + 0.013053464703261852, + 0.005089279264211655, + 0.007048850879073143, + 0.0173153318464756, + 0.009699969552457333, + 0.013246812857687473, + 0.006511497776955366, + 0.017009329050779343, + 0.01821187511086464, + 0.012796089053153992, + 0.014738630503416061, + 0.015380529686808586, + 0.012238210998475552, + 0.004460685420781374, + 0.006495879031717777, + 0.009214546531438828, + 0.012006258592009544, + 0.005507581867277622, + 0.007928396575152874, + 0.007513854652643204, + 0.006295742001384497, + 0.009546013548970222, + 0.008551906794309616, + 0.01378593035042286, + 0.0166030116379261, + 0.008010759018361568, + 0.013352630659937859, + 0.013073202222585678, + 0.013967049308121204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.16044700145721436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054972801357507706 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.054972801357507706, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054972801357507706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3679923713207245, + 0.06992565840482712, + 0.05689265578985214, + 0.12304763495922089, + 0.059334032237529755, + 0.07374262809753418, + 0.038509078323841095, + 0.11301499605178833, + 0.044654425233602524, + 0.05288654938340187 + ], + "credit_map": [ + 0.03738837689161301, + 0.02678866870701313, + 0.05807792395353317, + 0.03630513697862625, + 0.026670748367905617, + 0.0732811689376831, + 0.04862683638930321, + 0.060853492468595505, + 0.0072446162812411785, + 0.005812524352222681, + 0.009521890431642532, + 0.008342188782989979, + 0.009476586245000362, + 0.010617855004966259, + 0.00920406449586153, + 0.009705930016934872, + 0.0038965896237641573, + 0.0031248226296156645, + 0.009245025925338268, + 0.01087556779384613, + 0.005471828393638134, + 0.007703728508204222, + 0.009394580498337746, + 0.007180511485785246, + 0.008231780491769314, + 0.008632992394268513, + 0.01932312548160553, + 0.022400876507163048, + 0.00823979265987873, + 0.017650263383984566, + 0.024347864091396332, + 0.014220929704606533, + 0.006833883933722973, + 0.00446157855913043, + 0.008102781139314175, + 0.010144004598259926, + 0.006386370398104191, + 0.0057284231297671795, + 0.008514927700161934, + 0.009162059053778648, + 0.005788848269730806, + 0.006518795154988766, + 0.010819931514561176, + 0.009864486753940582, + 0.007619471289217472, + 0.0116715207695961, + 0.009457306936383247, + 0.012002261355519295, + 0.0034411174710839987, + 0.004209400620311499, + 0.00489024119451642, + 0.006935820449143648, + 0.0032664272002875805, + 0.004869802389293909, + 0.005937027279287577, + 0.004959241021424532, + 0.00767475925385952, + 0.010377631522715092, + 0.017574327066540718, + 0.016594253480434418, + 0.012510723434388638, + 0.016308456659317017, + 0.014208484441041946, + 0.017766354605555534, + 0.0044050016440451145, + 0.006863907910883427, + 0.00589551217854023, + 0.006834662519395351, + 0.0032643689773976803, + 0.005305575206875801, + 0.007453131023794413, + 0.004632262978702784, + 0.004825493320822716, + 0.0034266049042344093, + 0.007315704599022865, + 0.009141521528363228, + 0.003805931191891432, + 0.009575752541422844, + 0.008320330642163754, + 0.006475209258496761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.054972801357507706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069318279623985 + ], + "opposite_logprob_tokens": [ + -4.511069297790527 + ], + "logprob_reward": -0.011069318279623985, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069318279623985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25524792075157166, + 0.0547303669154644, + 0.03417431563138962, + 0.15259115397930145, + 0.08333417773246765, + 0.07953090220689774, + 0.036831989884376526, + 0.2460782825946808, + 0.02729487046599388, + 0.0301860012114048 + ], + "credit_map": [ + 0.019122593104839325, + 0.013089915737509727, + 0.04078410193324089, + 0.04351115599274635, + 0.020549003034830093, + 0.05319756269454956, + 0.03513520583510399, + 0.02985839545726776, + 0.006181970238685608, + 0.0051704291254282, + 0.00897578801959753, + 0.009620211087167263, + 0.0038859883788973093, + 0.007236068602651358, + 0.007025477010756731, + 0.006634433288127184, + 0.0024421452544629574, + 0.0022100345231592655, + 0.005410024430602789, + 0.008237170986831188, + 0.0030815305653959513, + 0.004032420460134745, + 0.0050034066662192345, + 0.003757581114768982, + 0.010725289583206177, + 0.01020165253430605, + 0.022465495392680168, + 0.03047228418290615, + 0.010718384757637978, + 0.02247639186680317, + 0.017374228686094284, + 0.02815743163228035, + 0.011586837470531464, + 0.006521262694150209, + 0.010335731320083141, + 0.010414894670248032, + 0.013348284177482128, + 0.010212273336946964, + 0.009912095963954926, + 0.0110027976334095, + 0.006647113244980574, + 0.005649335216730833, + 0.011518357321619987, + 0.010122342966496944, + 0.00816788338124752, + 0.010010202415287495, + 0.01132506225258112, + 0.01609060913324356, + 0.0044224937446415424, + 0.00409808661788702, + 0.004745435435324907, + 0.005808364599943161, + 0.0031685507856309414, + 0.0041083768010139465, + 0.005159664899110794, + 0.005321020260453224, + 0.02715824544429779, + 0.017537903040647507, + 0.03692704066634178, + 0.03521983325481415, + 0.03117281384766102, + 0.04311142861843109, + 0.025778112933039665, + 0.029172901064157486, + 0.002828980330377817, + 0.003985986113548279, + 0.0036317207850515842, + 0.005028135143220425, + 0.001965037314221263, + 0.003201008541509509, + 0.004272553138434887, + 0.0023814495652914047, + 0.0021231791470199823, + 0.002126371953636408, + 0.00481279194355011, + 0.00776631198823452, + 0.0018366440199315548, + 0.0037701509427279234, + 0.004358718637377024, + 0.0033918314147740602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.011069318279623985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207099318504333 + ], + "opposite_logprob_tokens": [ + -1.5020710229873657 + ], + "logprob_reward": -0.25207099318504333, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207099318504333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2628772556781769, + 0.09137234836816788, + 0.07114154100418091, + 0.07919678837060928, + 0.1122274100780487, + 0.11010438948869705, + 0.039018239825963974, + 0.07457274198532104, + 0.05480526387691498, + 0.10468394309282303 + ], + "credit_map": [ + 0.027498966082930565, + 0.016847403720021248, + 0.04065834358334541, + 0.03324602544307709, + 0.023595279082655907, + 0.049500804394483566, + 0.03308703750371933, + 0.038443416357040405, + 0.00923874694854021, + 0.00673654954880476, + 0.013065733015537262, + 0.010409433394670486, + 0.012603268958628178, + 0.013685831800103188, + 0.013597395271062851, + 0.012035395950078964, + 0.006471329368650913, + 0.004799684509634972, + 0.010995501652359962, + 0.010052436962723732, + 0.007431072648614645, + 0.013349834829568863, + 0.010427678003907204, + 0.007614011876285076, + 0.006157916039228439, + 0.005426995921880007, + 0.01299582701176405, + 0.014493224211037159, + 0.006620767991989851, + 0.008927609771490097, + 0.014423607848584652, + 0.0101508479565382, + 0.013227654621005058, + 0.011597491800785065, + 0.015428651124238968, + 0.021970951929688454, + 0.009998923167586327, + 0.01197435799986124, + 0.014857684262096882, + 0.013171704486012459, + 0.008090072311460972, + 0.007841752842068672, + 0.016991062089800835, + 0.015385123901069164, + 0.014871313236653805, + 0.014038985595107079, + 0.016363082453608513, + 0.016523007303476334, + 0.0028568438719958067, + 0.0031495194416493177, + 0.006153360474854708, + 0.00749894417822361, + 0.003125059651210904, + 0.004168919287621975, + 0.006960088852792978, + 0.0051055047661066055, + 0.010522672906517982, + 0.0068266610614955425, + 0.01007906999439001, + 0.011985559947788715, + 0.006540000904351473, + 0.00956419575959444, + 0.010462358593940735, + 0.008592220954596996, + 0.003735884791240096, + 0.010378777980804443, + 0.0068727219477295876, + 0.00796443223953247, + 0.006515495013445616, + 0.006754845380783081, + 0.008847100660204887, + 0.003736009355634451, + 0.006996756419539452, + 0.007628875318914652, + 0.01741079054772854, + 0.02224833332002163, + 0.005514591466635466, + 0.014324945397675037, + 0.019455542787909508, + 0.01110412273555994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.25207099318504333 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic time-lapse sequence, shot from a low angle and kept largely static to emphasize the epic scale, documents the silent, slow-motion battle between two antagonistic trees. In the foreground, a gnarled, ancient oak tree with dark, deeply furrowed bark covered in patches of green moss aggressively extends its thick, muscular limbs. Beside it, a taller, thinner birch tree, its pale bark peeling, twists its branches in a sinuous, relentless pursuit of any available gap in the canopy. The camera captures their gradual intertwining growth over many seasons; branches jostling, encircling, and strategically blocking each other in a relentless struggle for sunlight. The background reveals a dense, out-of-focus forest, while the forest floor below rapidly blurs from a carpet of summer ferns to autumnal leaves and a winter blanket of snow and ice. Golden light penetrates the treetops, moving swiftly as the sun arcs across the sky, creating a high-contrast, dramatic play of light and shadow, highlighting the texture of the bark and the slow, deliberate dance of the branches. The overall atmosphere is one of fierce, silent and primal competition, a beautiful yet brutal natural ballet of growth and survival.", + "seed": 13652, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_8381_seed_13652_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20078470334410667, + "vlm_logprob": -0.20078470334410667, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.20078470334410667, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16009825468063354 + ], + "opposite_logprob_tokens": [ + -2.4100983142852783 + ], + "logprob_reward": -0.16009825468063354, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.16009825468063354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10124251246452332, + 0.06506847590208054, + 0.26909586787223816, + 0.1349049210548401, + 0.01706748455762863, + 0.22551147639751434, + 0.02833586372435093, + 0.019916413351893425, + 0.025459477677941322, + 0.11339748650789261 + ], + "credit_map": [ + 0.009235299192368984, + 0.009606080129742622, + 0.016389377415180206, + 0.012476762756705284, + 0.014068376272916794, + 0.009003710001707077, + 0.015364092774689198, + 0.015098817646503448, + 0.0029458217322826385, + 0.0021939657162874937, + 0.019403133541345596, + 0.004818954039365053, + 0.004332419950515032, + 0.003036882495507598, + 0.005965225864201784, + 0.02237207256257534, + 0.012297539971768856, + 0.006336212158203125, + 0.09221765398979187, + 0.014851626940071583, + 0.01137517485767603, + 0.010541846044361591, + 0.013926257379353046, + 0.10754957050085068, + 0.0072751170955598354, + 0.006474444642663002, + 0.024037383496761322, + 0.02393907681107521, + 0.01547911949455738, + 0.009945454075932503, + 0.030024588108062744, + 0.0177297405898571, + 0.0013083727099001408, + 0.0014379636850208044, + 0.0018255185568705201, + 0.0030548141803592443, + 0.003465535817667842, + 0.0020173436496406794, + 0.002118324162438512, + 0.00183961377479136, + 0.01113287452608347, + 0.0094339270144701, + 0.07729164510965347, + 0.007873945869505405, + 0.01273388136178255, + 0.009271202608942986, + 0.009722903370857239, + 0.08805111050605774, + 0.001574688940308988, + 0.0019703637808561325, + 0.005910856649279594, + 0.0028134107124060392, + 0.003397104563191533, + 0.004530471749603748, + 0.002806913573294878, + 0.005332056898623705, + 0.0010393104748800397, + 0.0011342540383338928, + 0.004729453008621931, + 0.0025799202267080545, + 0.0015348095912486315, + 0.002136862138286233, + 0.002052153693512082, + 0.0047096507623791695, + 0.0021140901371836662, + 0.0035750465467572212, + 0.004643330350518227, + 0.002357171382755041, + 0.002453374210745096, + 0.0032863870728760958, + 0.0026989199686795473, + 0.0043311575427651405, + 0.005547044333070517, + 0.008085695095360279, + 0.03167282044887543, + 0.01238584890961647, + 0.0060135116800665855, + 0.004620810505002737, + 0.010991891846060753, + 0.03407986834645271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.16009825468063354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129833459854126 + ], + "opposite_logprob_tokens": [ + -2.237983226776123 + ], + "logprob_reward": -0.1129833459854126, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129833459854126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05971785634756088, + 0.09657981246709824, + 0.03630129247903824, + 0.027803238481283188, + 0.03912721574306488, + 0.5398620367050171, + 0.01351463608443737, + 0.044634245336055756, + 0.024102330207824707, + 0.11835730075836182 + ], + "credit_map": [ + 0.0038828582037240267, + 0.0030775461345911026, + 0.011713940650224686, + 0.005746739450842142, + 0.008791233412921429, + 0.006612386554479599, + 0.011977076530456543, + 0.007916079834103584, + 0.0070484476163983345, + 0.0064965649507939816, + 0.020770883187651634, + 0.011388366110622883, + 0.010721225291490555, + 0.00625329464673996, + 0.007848360575735569, + 0.026052672415971756, + 0.0024228503461927176, + 0.002157733077183366, + 0.0095699243247509, + 0.002566857263445854, + 0.0036319943610578775, + 0.0028733776416629553, + 0.0033111923839896917, + 0.009767362847924232, + 0.002414579736068845, + 0.0030314549803733826, + 0.004128644708544016, + 0.002715839771553874, + 0.004768012557178736, + 0.004384012892842293, + 0.0025150699075311422, + 0.003845624392852187, + 0.001956711756065488, + 0.001808619243092835, + 0.009759794920682907, + 0.004728519357740879, + 0.0029870858415961266, + 0.0022490601986646652, + 0.005013762507587671, + 0.010623662732541561, + 0.026223203167319298, + 0.027648145332932472, + 0.17975710332393646, + 0.023969123139977455, + 0.028691932559013367, + 0.023663992062211037, + 0.025180602446198463, + 0.2047279328107834, + 0.0009539314196445048, + 0.000781528651714325, + 0.0033631769474595785, + 0.0015956675633788109, + 0.0009783878922462463, + 0.00138071458786726, + 0.0013394460547715425, + 0.0031217820942401886, + 0.0021747094579041004, + 0.002667518798261881, + 0.01243018638342619, + 0.004918831866234541, + 0.0024091931991279125, + 0.003410547273233533, + 0.0029230983927845955, + 0.013700158335268497, + 0.0014810366556048393, + 0.00284281256608665, + 0.006355722434818745, + 0.0025574511382728815, + 0.0023213322274386883, + 0.001576900132931769, + 0.0019324426539242268, + 0.005034634843468666, + 0.00596289336681366, + 0.004415108356624842, + 0.03852670639753342, + 0.010028544813394547, + 0.004712413530796766, + 0.006655014585703611, + 0.005828877445310354, + 0.04222773760557175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.1129833459854126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219863340258598 + ], + "opposite_logprob_tokens": [ + -4.264219760894775 + ], + "logprob_reward": -0.014219863340258598, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219863340258598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09151028096675873, + 0.11252137273550034, + 0.0766114592552185, + 0.042870670557022095, + 0.05021972954273224, + 0.41005563735961914, + 0.02449287846684456, + 0.01799296960234642, + 0.027236927300691605, + 0.14648808538913727 + ], + "credit_map": [ + 0.0074558197520673275, + 0.007204074878245592, + 0.013916526921093464, + 0.011515622958540916, + 0.014692987315356731, + 0.011923708952963352, + 0.013361439108848572, + 0.011440101079642773, + 0.004926749505102634, + 0.0033451984636485577, + 0.0346120186150074, + 0.01128369476646185, + 0.007618198171257973, + 0.0049991970881819725, + 0.00634861271828413, + 0.03938771039247513, + 0.003445790149271488, + 0.0021722048986703157, + 0.02649718150496483, + 0.0035208857152611017, + 0.004064616747200489, + 0.004041064064949751, + 0.004339200910180807, + 0.028530512005090714, + 0.0033385129645466805, + 0.004344726912677288, + 0.00820477120578289, + 0.004771193489432335, + 0.006326890084892511, + 0.006654749624431133, + 0.0044204252772033215, + 0.004809400998055935, + 0.002600474050268531, + 0.002204607240855694, + 0.014508920721709728, + 0.004452956840395927, + 0.0029270555824041367, + 0.00322744925506413, + 0.004610448610037565, + 0.015687815845012665, + 0.02012714557349682, + 0.01675652712583542, + 0.13983623683452606, + 0.014172462746500969, + 0.02306155301630497, + 0.016982462257146835, + 0.01874586194753647, + 0.16037338972091675, + 0.0015442540170624852, + 0.0013679645489901304, + 0.004930234979838133, + 0.002397971460595727, + 0.00219197035767138, + 0.003507401328533888, + 0.003386380150914192, + 0.005166702903807163, + 0.001561732729896903, + 0.002508560661226511, + 0.0038044636603444815, + 0.0014188153436407447, + 0.0013580652885138988, + 0.0014276155270636082, + 0.0018497644923627377, + 0.004063952248543501, + 0.0017457375070080161, + 0.0014893952757120132, + 0.006844536866992712, + 0.0025800636503845453, + 0.002029445255175233, + 0.0034578570630401373, + 0.00295055634342134, + 0.006139334291219711, + 0.007185876835137606, + 0.004493171814829111, + 0.039189741015434265, + 0.018865112215280533, + 0.005151723977178335, + 0.007286805659532547, + 0.01413514744490385, + 0.05018050968647003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.014219863340258598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327817492187023 + ], + "opposite_logprob_tokens": [ + -3.77327823638916 + ], + "logprob_reward": -0.02327817492187023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327817492187023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13526253402233124, + 0.11963704228401184, + 0.21930643916130066, + 0.021015413105487823, + 0.14208467304706573, + 0.12440706789493561, + 0.024109799414873123, + 0.012898897752165794, + 0.01882375031709671, + 0.18245433270931244 + ], + "credit_map": [ + 0.012520452961325645, + 0.008816772140562534, + 0.01722172275185585, + 0.009726124815642834, + 0.03127910941839218, + 0.0231512188911438, + 0.01619933359324932, + 0.01634780690073967, + 0.005237166304141283, + 0.0032478845678269863, + 0.03759780526161194, + 0.009169618599116802, + 0.00883516389876604, + 0.005044910591095686, + 0.006712055765092373, + 0.043792445212602615, + 0.009733101353049278, + 0.005353971850126982, + 0.07706930488348007, + 0.010610489174723625, + 0.00983973778784275, + 0.008669031783938408, + 0.010221959091722965, + 0.08780883997678757, + 0.0011236518621444702, + 0.0015459159621968865, + 0.003178286598995328, + 0.003381701884791255, + 0.0021407962776720524, + 0.002776613226160407, + 0.00416641915217042, + 0.0027020315174013376, + 0.00631313119083643, + 0.006721477955579758, + 0.026512881740927696, + 0.02603789232671261, + 0.015409517101943493, + 0.007056412752717733, + 0.024618975818157196, + 0.029414400458335876, + 0.005853049457073212, + 0.006765245459973812, + 0.03778740018606186, + 0.005688247736543417, + 0.007593438494950533, + 0.007015479728579521, + 0.009850088506937027, + 0.043854132294654846, + 0.0010338707361370325, + 0.0014238589210435748, + 0.005964847281575203, + 0.0016546504339203238, + 0.0021914984099566936, + 0.0030548868235200644, + 0.002519143745303154, + 0.00626704515889287, + 0.0011485930299386382, + 0.0012745085405185819, + 0.002369961002841592, + 0.0011697242734953761, + 0.0011688927188515663, + 0.0019752010703086853, + 0.0019917706958949566, + 0.0018002467695623636, + 0.0015463099116459489, + 0.0022787675261497498, + 0.0035946699790656567, + 0.0019226586446166039, + 0.001854232046753168, + 0.0027902855072170496, + 0.0018977680010721087, + 0.00293906033039093, + 0.008371079340577126, + 0.006545267533510923, + 0.061709899455308914, + 0.013886156491935253, + 0.0056934175081551075, + 0.006647196132689714, + 0.008836176246404648, + 0.07076513022184372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.02327817492187023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933438777923584 + ], + "opposite_logprob_tokens": [ + -0.6933438777923584 + ], + "logprob_reward": -0.6933438777923584, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933438777923584, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08392306417226791, + 0.1120009496808052, + 0.030583059415221214, + 0.0259584728628397, + 0.019599275663495064, + 0.6089832782745361, + 0.019563725218176842, + 0.028686953708529472, + 0.0258604995906353, + 0.04484071582555771 + ], + "credit_map": [ + 0.005465837195515633, + 0.004811582621186972, + 0.017038144171237946, + 0.010102854110300541, + 0.008872437290847301, + 0.009439361281692982, + 0.01560246292501688, + 0.012590380385518074, + 0.00671647721901536, + 0.005102820228785276, + 0.026406949386000633, + 0.013309167698025703, + 0.009997925721108913, + 0.006018020212650299, + 0.011296306736767292, + 0.033153288066387177, + 0.0025737567339092493, + 0.0032509013544768095, + 0.005885412450879812, + 0.002505303593352437, + 0.0064489166252315044, + 0.003696731524541974, + 0.002324202563613653, + 0.003897834336385131, + 0.002406191313639283, + 0.00323337409645319, + 0.0035827509127557278, + 0.0026778988540172577, + 0.005349399521946907, + 0.0028712411876767874, + 0.0026484748814255, + 0.0031891418620944023, + 0.0011450075544416904, + 0.0010689677437767386, + 0.0051267812959849834, + 0.001574366120621562, + 0.0017311115516349673, + 0.0013903881190344691, + 0.0016133890021592379, + 0.005949263460934162, + 0.029648466035723686, + 0.027838217094540596, + 0.20677079260349274, + 0.02266233041882515, + 0.033357828855514526, + 0.0241045244038105, + 0.028608085587620735, + 0.23599299788475037, + 0.0009713732288219035, + 0.0008614041726104915, + 0.0057061901316046715, + 0.001728056580759585, + 0.0017695127753540874, + 0.0013053924776613712, + 0.0016915968153625727, + 0.005530199501663446, + 0.0013979332288727164, + 0.0023564088623970747, + 0.006974621210247278, + 0.0035403617657721043, + 0.0021188536193221807, + 0.0026150341145694256, + 0.0027026019524782896, + 0.00698113813996315, + 0.0012036742409691215, + 0.0017125640297308564, + 0.005245742853730917, + 0.003390838857740164, + 0.0014046578435227275, + 0.002355341101065278, + 0.005068772006779909, + 0.005478910636156797, + 0.0036416694056242704, + 0.00404316233471036, + 0.007182964589446783, + 0.006185153033584356, + 0.004070960450917482, + 0.005824128165841103, + 0.005600850097835064, + 0.008291823789477348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.6933438777923584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "dramatic time-lapse sequence, shot from a low angle and kept largely static to emphasize the epic scale, documents the silent, slow-motion battle between two antagonistic trees. In the foreground, a gnarled, ancient oak tree with dark, deeply furrowed bark covered in patches of green moss aggressively extends its thick, muscular limbs. Beside it, a taller, thinner birch tree, its pale bark peeling, twists its branches in a sinuous, relentless pursuit of any available gap in the canopy. The camera captures their gradual intertwining growth over many seasons; branches jostling, encircling, and strategically blocking each other in a relentless struggle for sunlight. The background reveals a dense, out-of-focus forest, while the forest floor below rapidly blurs from a carpet of summer ferns to autumnal leaves and a winter blanket of snow and ice. Golden light penetrates the treetops, moving swiftly as the sun arcs across the sky, creating a high-contrast, dramatic play of light and shadow, highlighting the texture of the bark and the slow, deliberate dance of the branches. The overall atmosphere is one of fierce, silent and primal competition, a beautiful yet brutal natural ballet of growth and survival.", + "seed": 13653, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_1/index_8381_seed_13653_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1517211988568306, + "vlm_logprob": -0.1517211988568306, + "vlm_yes_no_margin": 2.7249999761581423, + "vlm_token_credit": -0.1517211988568306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4048946499824524 + ], + "opposite_logprob_tokens": [ + -1.5298945903778076 + ], + "logprob_reward": -0.4048946499824524, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.4048946499824524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22470682859420776, + 0.16833634674549103, + 0.11507067829370499, + 0.0822257325053215, + 0.08882927894592285, + 0.1512080579996109, + 0.06410475075244904, + 0.036500126123428345, + 0.030789058655500412, + 0.03822916001081467 + ], + "credit_map": [ + 0.029228394851088524, + 0.015326364897191525, + 0.04313024878501892, + 0.03069157712161541, + 0.016181986778974533, + 0.019914239645004272, + 0.04575895890593529, + 0.02447507716715336, + 0.024740666151046753, + 0.016736077144742012, + 0.022954149171710014, + 0.025245849043130875, + 0.013177375309169292, + 0.011998985894024372, + 0.03648551180958748, + 0.016997748985886574, + 0.011715760454535484, + 0.00893657561391592, + 0.019437920302152634, + 0.014130067080259323, + 0.005349365063011646, + 0.01541135273873806, + 0.015166898258030415, + 0.024922741577029228, + 0.009212127886712551, + 0.006849105004221201, + 0.014367339201271534, + 0.01093149185180664, + 0.005280700512230396, + 0.006410498637706041, + 0.020362162962555885, + 0.008812310174107552, + 0.012053148820996284, + 0.008856016211211681, + 0.015157478861510754, + 0.013060824945569038, + 0.005758923944085836, + 0.008396380580961704, + 0.010062448680400848, + 0.01548406295478344, + 0.026906412094831467, + 0.017174946144223213, + 0.020991627126932144, + 0.02079513482749462, + 0.008071651682257652, + 0.018949784338474274, + 0.023672832176089287, + 0.014645667746663094, + 0.006878044921904802, + 0.004715814255177975, + 0.008755862712860107, + 0.009453478269279003, + 0.004129251465201378, + 0.0070228879339993, + 0.014251258224248886, + 0.00889816228300333, + 0.0036197721492499113, + 0.002009124029427767, + 0.005660850089043379, + 0.006316036451607943, + 0.001564985141158104, + 0.0039395522326231, + 0.008777913637459278, + 0.004611895885318518, + 0.003263116115704179, + 0.00201489613391459, + 0.00536592211574316, + 0.00381508469581604, + 0.0014255926944315434, + 0.003834604984149337, + 0.005352129228413105, + 0.005717715714126825, + 0.004056614823639393, + 0.0028939100448042154, + 0.008473558351397514, + 0.005194276571273804, + 0.0028608269058167934, + 0.0066074482165277, + 0.004135308787226677, + 0.004007219802588224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.4048946499824524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284498989582062 + ], + "opposite_logprob_tokens": [ + -2.0178449153900146 + ], + "logprob_reward": -0.14284498989582062, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14284498989582062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2892749607563019, + 0.07822269201278687, + 0.10055233538150787, + 0.07205566018819809, + 0.072967529296875, + 0.073556087911129, + 0.0656948983669281, + 0.0848090648651123, + 0.13540400564670563, + 0.027462763711810112 + ], + "credit_map": [ + 0.05352718010544777, + 0.034135982394218445, + 0.0464651845395565, + 0.033196404576301575, + 0.024328971281647682, + 0.028694214299321175, + 0.0430467464029789, + 0.025880279019474983, + 0.00772260082885623, + 0.008558027446269989, + 0.011432954110205173, + 0.011083347722887993, + 0.007769683375954628, + 0.007839853875339031, + 0.013651208020746708, + 0.0101650170981884, + 0.01649024337530136, + 0.010862593539059162, + 0.015105576254427433, + 0.012982824817299843, + 0.005928392056375742, + 0.009776473976671696, + 0.01556918304413557, + 0.013837048783898354, + 0.010208097286522388, + 0.007154272403568029, + 0.011767776682972908, + 0.013201323337852955, + 0.004580788780003786, + 0.008215759880840778, + 0.009212484583258629, + 0.007715160492807627, + 0.010214290581643581, + 0.005028202664107084, + 0.012139790691435337, + 0.010995815508067608, + 0.003743197536095977, + 0.01033928245306015, + 0.008391882292926311, + 0.012115071527659893, + 0.005668642930686474, + 0.004253185819834471, + 0.012285888195037842, + 0.011923916637897491, + 0.003917153924703598, + 0.00507184537127614, + 0.021150432527065277, + 0.009285020641982555, + 0.008452068082988262, + 0.005666009150445461, + 0.009528061375021935, + 0.011496656574308872, + 0.0036012406926602125, + 0.006933808792382479, + 0.008160894736647606, + 0.011856158263981342, + 0.015273555181920528, + 0.006977010518312454, + 0.013700766488909721, + 0.015632042661309242, + 0.004224169068038464, + 0.010333213023841381, + 0.009163265116512775, + 0.009505043737590313, + 0.028573177754878998, + 0.012815261259675026, + 0.020399637520313263, + 0.013291827403008938, + 0.007598978467285633, + 0.016099724918603897, + 0.012979738414287567, + 0.023645656183362007, + 0.0031997421756386757, + 0.0029571750201284885, + 0.004549655131995678, + 0.0031176793854683638, + 0.0024938713759183884, + 0.004027930554002523, + 0.0030535550322383642, + 0.00406315503641963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.14284498989582062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004671018570661545 + ], + "opposite_logprob_tokens": [ + -5.379671096801758 + ], + "logprob_reward": -0.004671018570661545, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004671018570661545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2645080089569092, + 0.17487363517284393, + 0.15153171122074127, + 0.09458567202091217, + 0.04668547213077545, + 0.09851723909378052, + 0.029895322397351265, + 0.024806689471006393, + 0.033109333366155624, + 0.08148688822984695 + ], + "credit_map": [ + 0.03473927825689316, + 0.03360610827803612, + 0.0415814146399498, + 0.03586498275399208, + 0.02936583198606968, + 0.024567555636167526, + 0.04208894819021225, + 0.022693892940878868, + 0.030411172658205032, + 0.01780407503247261, + 0.026314646005630493, + 0.024153266102075577, + 0.009951772168278694, + 0.013867219910025597, + 0.028774132952094078, + 0.023597344756126404, + 0.019329100847244263, + 0.011371099390089512, + 0.027088865637779236, + 0.017903532832860947, + 0.006483498960733414, + 0.014529065228998661, + 0.019527023658156395, + 0.035299528390169144, + 0.012083977460861206, + 0.008572770282626152, + 0.013232802040874958, + 0.011071027256548405, + 0.006736782379448414, + 0.010240118950605392, + 0.020470844581723213, + 0.012177341617643833, + 0.00566064054146409, + 0.004317908082157373, + 0.0069445837289094925, + 0.00731914397329092, + 0.0029910386074334383, + 0.006470915395766497, + 0.0054483916610479355, + 0.007532850839197636, + 0.008916974999010563, + 0.009482132270932198, + 0.014997857622802258, + 0.016905182972550392, + 0.004445204045623541, + 0.00910721905529499, + 0.01961297169327736, + 0.015049691312015057, + 0.0032474403269588947, + 0.002320758532732725, + 0.004801577422767878, + 0.006285510491579771, + 0.0013309955829754472, + 0.0025449679233133793, + 0.005599042400717735, + 0.003765030996873975, + 0.004974325653165579, + 0.001843727775849402, + 0.003846335457637906, + 0.003404795192182064, + 0.0012846951140090823, + 0.0020495790522545576, + 0.005297825671732426, + 0.002105406951159239, + 0.0036119427531957626, + 0.0023809110280126333, + 0.005538958124816418, + 0.005801004357635975, + 0.0015809497563168406, + 0.0036425222642719746, + 0.005118389613926411, + 0.005434657912701368, + 0.010516725480556488, + 0.005328907165676355, + 0.020390165969729424, + 0.009169629774987698, + 0.004336656536906958, + 0.017246384173631668, + 0.008127851411700249, + 0.006370572838932276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004671018570661545 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335861533880234 + ], + "opposite_logprob_tokens": [ + -3.6513359546661377 + ], + "logprob_reward": -0.026335861533880234, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335861533880234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1872483640909195, + 0.2038625031709671, + 0.1636650413274765, + 0.1485941857099533, + 0.07665134221315384, + 0.0962623879313469, + 0.023718956857919693, + 0.030760353431105614, + 0.023450320586562157, + 0.045786574482917786 + ], + "credit_map": [ + 0.02499561570584774, + 0.023403193801641464, + 0.026930347084999084, + 0.03385297954082489, + 0.01992676965892315, + 0.017268382012844086, + 0.024593228474259377, + 0.016277840360999107, + 0.035401199012994766, + 0.019414862617850304, + 0.029711296781897545, + 0.028455134481191635, + 0.01240090373903513, + 0.018513154238462448, + 0.03341005742549896, + 0.026555901393294334, + 0.01589304953813553, + 0.010959452018141747, + 0.02815796062350273, + 0.021104946732521057, + 0.00589565047994256, + 0.016191722825169563, + 0.029557080939412117, + 0.035905178636312485, + 0.013024400919675827, + 0.010075671598315239, + 0.020668908953666687, + 0.020946675911545753, + 0.006611999589949846, + 0.012955086305737495, + 0.044151779264211655, + 0.020159652456641197, + 0.009748789481818676, + 0.006568345241248608, + 0.011069094762206078, + 0.013343896716833115, + 0.004390782210975885, + 0.010895421728491783, + 0.007587416563183069, + 0.013047602027654648, + 0.010188557207584381, + 0.005592137575149536, + 0.015105286613106728, + 0.019770486280322075, + 0.0029594418592751026, + 0.00871732272207737, + 0.02016231045126915, + 0.013766842894256115, + 0.0023557550739496946, + 0.001778221339918673, + 0.003831704379990697, + 0.004824896343052387, + 0.0011112663196399808, + 0.002383930841460824, + 0.003963463939726353, + 0.0034697181545197964, + 0.003696582280099392, + 0.0029495181515812874, + 0.0052054752595722675, + 0.0049279481172561646, + 0.0016214341158047318, + 0.0026784995570778847, + 0.005214142147451639, + 0.004466753453016281, + 0.002923649735748768, + 0.0018106203060597181, + 0.0040165726095438, + 0.0036157260183244944, + 0.0012432840885594487, + 0.0029492960311472416, + 0.0035216771066188812, + 0.0033694931771606207, + 0.005371211562305689, + 0.002886065747588873, + 0.009477933868765831, + 0.005155687220394611, + 0.0027871273923665285, + 0.008309467695653439, + 0.006905816495418549, + 0.004893265664577484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.026335861533880234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera largely static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gnarled ancient oak tree with dark, deeply furrowed bark and green moss visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a taller, thinner birch tree with pale, peeling bark visible beside the oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show the forest floor changing from summer ferns to autumn leaves and then to a winter blanket of snow and ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798594743013382 + ], + "opposite_logprob_tokens": [ + -1.8048595190048218 + ], + "logprob_reward": -0.1798594743013382, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798594743013382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13741439580917358, + 0.08135204762220383, + 0.1253671795129776, + 0.13764235377311707, + 0.19332042336463928, + 0.04522189125418663, + 0.07404764741659164, + 0.03248210623860359, + 0.04913357272744179, + 0.12401840090751648 + ], + "credit_map": [ + 0.020231768488883972, + 0.013208470307290554, + 0.023943349719047546, + 0.018955502659082413, + 0.009788010269403458, + 0.01214626431465149, + 0.02702277898788452, + 0.012118250131607056, + 0.01052269246429205, + 0.007164048030972481, + 0.013076812028884888, + 0.011631344445049763, + 0.006353224162012339, + 0.009355857037007809, + 0.013591820374131203, + 0.009656233713030815, + 0.01515602134168148, + 0.009648622944951057, + 0.020884744822978973, + 0.01422928087413311, + 0.004652983974665403, + 0.014207416214048862, + 0.02310052327811718, + 0.023487582802772522, + 0.016453232616186142, + 0.011872594244778156, + 0.019895348697900772, + 0.015387022867798805, + 0.00899519957602024, + 0.015760214999318123, + 0.03075697086751461, + 0.018521748483181, + 0.028105061501264572, + 0.013757211156189442, + 0.02906503900885582, + 0.031767163425683975, + 0.01115398108959198, + 0.025162694975733757, + 0.023594345897436142, + 0.030714912340044975, + 0.003958543296903372, + 0.003427241463214159, + 0.0068788425996899605, + 0.008039762265980244, + 0.0017693329136818647, + 0.004760114010423422, + 0.011153996922075748, + 0.005234058480709791, + 0.007888847030699253, + 0.005338974762707949, + 0.010725844651460648, + 0.014558137394487858, + 0.0035094404593110085, + 0.006371807307004929, + 0.014683345332741737, + 0.010971243493258953, + 0.004898698069155216, + 0.0029805335216224194, + 0.005219617392867804, + 0.003635275876149535, + 0.0020119056571274996, + 0.0035546510480344296, + 0.005552154034376144, + 0.004629269242286682, + 0.006474874448031187, + 0.0030753815080970526, + 0.007860972546041012, + 0.007561492267996073, + 0.0019522354705259204, + 0.00553904427215457, + 0.007299164775758982, + 0.009370407089591026, + 0.012988314032554626, + 0.005508383736014366, + 0.029819490388035774, + 0.014683174900710583, + 0.006709216628223658, + 0.027141541242599487, + 0.013300514779984951, + 0.013867758214473724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.1798594743013382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_tensor([4405, 4405])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_tensor([4405, 4405])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3705c9ef02b183656216d4d4b3893eeb92cb7338 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_tensor([4405, 4405])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle tracking shot slowly pans right across the icy landscape of a remote Arctic research station. A golden polar sun hangs high above the horizon, its slanting rays casting long blue shadows across the vast, pristine expanse of ice, outlining the texture of each snowdrift with remarkable clarity. In the foreground, a scientist, clad in a thick, bright orange thermal suit, a fully enclosed hood, and goggles that reflect the golden sunlight, stands out. Half-kneeling in the snow, he methodically adjusts the angle of a large solar panel, its deep blue photovoltaic surface coated with a thin layer of frost. One thickly gloved hand steadying the panel's cool metal frame, the other carefully tightening a connection point on the bracket with a tool, his movements steady and focused. Fine snowflakes swirl in the biting wind, some landing gently on the scientist's dark goggles and the black surface of the solar panel before being blown away. In the background, the research station’s silver geodesic dome and several communication antennas gleam with metallic luster in the golden afterglow. The whole scene forms a sharp contrast between the majestic nature and the tenacious technology of mankind, creating a tranquil atmosphere that is both lonely and full of hope.", + "seed": 25280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_4405_seed_25280_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05726142590865493, + "vlm_logprob": -0.05726142590865493, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.05726142590865493, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23863834142684937 + ], + "opposite_logprob_tokens": [ + -1.6136384010314941 + ], + "logprob_reward": -0.23863834142684937, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23863834142684937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2902618646621704, + 0.028440263122320175, + 0.1726662963628769, + 0.04401329532265663, + 0.14151127636432648, + 0.064860038459301, + 0.13156719505786896, + 0.03627205267548561, + 0.07169631868600845, + 0.018711432814598083 + ], + "credit_map": [ + 0.049313582479953766, + 0.03242243826389313, + 0.05293350666761398, + 0.02445068582892418, + 0.026462193578481674, + 0.0607433058321476, + 0.019866911694407463, + 0.024069204926490784, + 0.004131740890443325, + 0.0031795050017535686, + 0.00512432586401701, + 0.0021875067614018917, + 0.0022775838151574135, + 0.0042228312231600285, + 0.0019117941847071052, + 0.0054049743339419365, + 0.020648157224059105, + 0.016882849857211113, + 0.019153140485286713, + 0.0093997772783041, + 0.013135524466633797, + 0.03114091232419014, + 0.008899565786123276, + 0.05340636521577835, + 0.005880164913833141, + 0.006017834879457951, + 0.0053915842436254025, + 0.005164716392755508, + 0.007231078576296568, + 0.0056505645625293255, + 0.003724446753039956, + 0.00495290569961071, + 0.01866013929247856, + 0.014774782583117485, + 0.024363335222005844, + 0.017136087641119957, + 0.011108429171144962, + 0.018005231395363808, + 0.01698175072669983, + 0.020481521263718605, + 0.007557450328022242, + 0.007768047507852316, + 0.01007088553160429, + 0.008971977047622204, + 0.0054895710200071335, + 0.006924340967088938, + 0.00704864040017128, + 0.011029120534658432, + 0.016920624300837517, + 0.013905868865549564, + 0.020565684884786606, + 0.012739477679133415, + 0.015321459621191025, + 0.014708664268255234, + 0.019030282273888588, + 0.01837512105703354, + 0.004290976095944643, + 0.003550909226760268, + 0.004191589076071978, + 0.005302005913108587, + 0.003963623661547899, + 0.006398978177458048, + 0.0046950457617640495, + 0.003878924297168851, + 0.009774832986295223, + 0.009118449874222279, + 0.011073265224695206, + 0.015539680607616901, + 0.005585194565355778, + 0.006154677364975214, + 0.0061569842509925365, + 0.00829322449862957, + 0.002600355539470911, + 0.0022883217316120863, + 0.002730935113504529, + 0.0027882116846740246, + 0.0022014668211340904, + 0.0020160055719316006, + 0.001712936325930059, + 0.0023731982801109552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.23863834142684937 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964101292192936 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964101292192936, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964101292192936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2797919511795044, + 0.08736907690763474, + 0.11655093729496002, + 0.1975863128900528, + 0.06507086753845215, + 0.025022491812705994, + 0.15148383378982544, + 0.025137726217508316, + 0.03375088796019554, + 0.018235908821225166 + ], + "credit_map": [ + 0.0402597077190876, + 0.04272802919149399, + 0.03541567549109459, + 0.034489866346120834, + 0.024422822520136833, + 0.03392092511057854, + 0.026856359094381332, + 0.04169856011867523, + 0.01253352127969265, + 0.01316804625093937, + 0.011290911585092545, + 0.005787520669400692, + 0.008796815760433674, + 0.012598684057593346, + 0.005063450429588556, + 0.01813013106584549, + 0.013755608350038528, + 0.014950734563171864, + 0.015520439483225346, + 0.009052835404872894, + 0.010876251384615898, + 0.016498684883117676, + 0.008239803835749626, + 0.02765658125281334, + 0.02556931972503662, + 0.01999031752347946, + 0.019641783088445663, + 0.01691155508160591, + 0.044867776334285736, + 0.03195880725979805, + 0.016535082831978798, + 0.022111667320132256, + 0.008281346410512924, + 0.006587965413928032, + 0.006188704166561365, + 0.004368531983345747, + 0.0058336746878921986, + 0.010730160400271416, + 0.006547927390784025, + 0.01653255894780159, + 0.003288175677880645, + 0.002366993110626936, + 0.004253212362527847, + 0.0038750958628952503, + 0.0018703412497416139, + 0.0024308350402861834, + 0.0026684305630624294, + 0.004269407596439123, + 0.018704332411289215, + 0.022187722846865654, + 0.02136947400867939, + 0.029008736833930016, + 0.019205501303076744, + 0.01333880890160799, + 0.01397526916116476, + 0.013693996705114841, + 0.0033070575445890427, + 0.0022090638522058725, + 0.00348721188493073, + 0.0022399385925382376, + 0.0023344780784100294, + 0.0040544169023633, + 0.0032345280051231384, + 0.004271032754331827, + 0.004501687828451395, + 0.003950220532715321, + 0.004702683072537184, + 0.006355983670800924, + 0.0037018156144768, + 0.003239296842366457, + 0.003106796881183982, + 0.004192400258034468, + 0.003106408752501011, + 0.0033061562571674585, + 0.0020329849794507027, + 0.00293907942250371, + 0.0023194109089672565, + 0.0018863806035369635, + 0.0010948240524157882, + 0.0015506620984524488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.005964101292192936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738202180713415 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738202180713415, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738202180713415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13996770977973938, + 0.054411064833402634, + 0.06307537853717804, + 0.04002515971660614, + 0.30001822113990784, + 0.07005886733531952, + 0.13120564818382263, + 0.06401187926530838, + 0.1010449156165123, + 0.03618113324046135 + ], + "credit_map": [ + 0.022722793743014336, + 0.015471813268959522, + 0.03513237088918686, + 0.016790315508842468, + 0.008765253238379955, + 0.01585654355585575, + 0.010798705741763115, + 0.0144299091771245, + 0.008126823231577873, + 0.006638675928115845, + 0.008138740435242653, + 0.004917321726679802, + 0.004545941483229399, + 0.009324008598923683, + 0.004156394395977259, + 0.008563161827623844, + 0.007622583769261837, + 0.0073568993248045444, + 0.008049207739531994, + 0.0044707139022648335, + 0.005122287664562464, + 0.010604556649923325, + 0.003693485399708152, + 0.01615564338862896, + 0.00581892766058445, + 0.004923155996948481, + 0.006341175176203251, + 0.004226210061460733, + 0.004099105950444937, + 0.006481689866632223, + 0.002912321127951145, + 0.005222573410719633, + 0.03402461111545563, + 0.019712340086698532, + 0.051603663712739944, + 0.019421208649873734, + 0.0190028827637434, + 0.04538759961724281, + 0.027451571077108383, + 0.08341436833143234, + 0.009385102428495884, + 0.007796264253556728, + 0.010632854886353016, + 0.007298002019524574, + 0.00648699514567852, + 0.007365942467004061, + 0.008859911933541298, + 0.0122337955981493, + 0.019479388371109962, + 0.012057139538228512, + 0.028559964150190353, + 0.020732540637254715, + 0.00909256562590599, + 0.007902363315224648, + 0.021700045093894005, + 0.0116816321387887, + 0.007981423288583755, + 0.007288600318133831, + 0.00687487842515111, + 0.008248320780694485, + 0.00664531672373414, + 0.007426074240356684, + 0.009044022299349308, + 0.010503248311579227, + 0.012024728581309319, + 0.007702965289354324, + 0.011216088198125362, + 0.008665290661156178, + 0.010459146462380886, + 0.013740243390202522, + 0.01478049997240305, + 0.022455953061580658, + 0.006016854662448168, + 0.005843033082783222, + 0.004757747985422611, + 0.005499552469700575, + 0.004618375562131405, + 0.0038859103806316853, + 0.0024366709403693676, + 0.0031229897867888212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.006738202180713415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863224919885397 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.00863224919885397, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863224919885397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21874642372131348, + 0.11560305953025818, + 0.12345679849386215, + 0.16492865979671478, + 0.1297215074300766, + 0.0354914516210556, + 0.06155451387166977, + 0.04235357791185379, + 0.06446383893489838, + 0.043680235743522644 + ], + "credit_map": [ + 0.029392126947641373, + 0.026270486414432526, + 0.0269832294434309, + 0.01814352534711361, + 0.013937205076217651, + 0.034723132848739624, + 0.01576773077249527, + 0.053528983145952225, + 0.016119759529829025, + 0.01664404198527336, + 0.015277363359928131, + 0.007132816594094038, + 0.00849970243871212, + 0.020393848419189453, + 0.005129859317094088, + 0.02640566974878311, + 0.01590326800942421, + 0.01613404043018818, + 0.018151266500353813, + 0.008102119900286198, + 0.010284364223480225, + 0.02234490215778351, + 0.005729785654693842, + 0.02680705115199089, + 0.02174416184425354, + 0.013528715819120407, + 0.027311325073242188, + 0.014345203526318073, + 0.027034487575292587, + 0.029382269829511642, + 0.01423435565084219, + 0.017348146066069603, + 0.01895769126713276, + 0.013886257074773312, + 0.036267805844545364, + 0.00942308735102415, + 0.009794676676392555, + 0.02107703499495983, + 0.008539085276424885, + 0.011775879189372063, + 0.004429872613400221, + 0.0043487842194736, + 0.006524969357997179, + 0.004044856410473585, + 0.0031992194708436728, + 0.0036672239657491446, + 0.00463289488106966, + 0.004643633496016264, + 0.00812234915792942, + 0.0061354790814220905, + 0.007677681744098663, + 0.007751392666250467, + 0.0054916683584451675, + 0.008823481388390064, + 0.007700188551098108, + 0.009852267801761627, + 0.005758471321314573, + 0.004400753416121006, + 0.00669045839458704, + 0.004857571795582771, + 0.005000256933271885, + 0.005185871385037899, + 0.004567811265587807, + 0.005892383400350809, + 0.007589375600218773, + 0.004851029720157385, + 0.00775387417525053, + 0.005332200787961483, + 0.007724631577730179, + 0.009631487540900707, + 0.007950136438012123, + 0.013631101697683334, + 0.005325305741280317, + 0.005139034707099199, + 0.004912199452519417, + 0.004920090548694134, + 0.003961151000112295, + 0.005096545908600092, + 0.004803337622433901, + 0.009522565640509129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.00863224919885397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334235444664955 + ], + "opposite_logprob_tokens": [ + -3.651334285736084 + ], + "logprob_reward": -0.026334235444664955, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334235444664955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40946537256240845, + 0.07024187594652176, + 0.09036590903997421, + 0.05434272438287735, + 0.08689515292644501, + 0.07641958445310593, + 0.09624402225017548, + 0.05037183314561844, + 0.02151842974126339, + 0.04413502290844917 + ], + "credit_map": [ + 0.06630758196115494, + 0.05204644799232483, + 0.07224619388580322, + 0.03536903113126755, + 0.03634060546755791, + 0.08325120806694031, + 0.023761307820677757, + 0.04014300927519798, + 0.010440956801176071, + 0.011801455169916153, + 0.011391302570700645, + 0.005681955721229315, + 0.010911814868450165, + 0.00888168253004551, + 0.004894787445664406, + 0.00623793201521039, + 0.012833990156650543, + 0.015333968214690685, + 0.013337461277842522, + 0.0075864531099796295, + 0.010157053358852863, + 0.009477369487285614, + 0.006769148167222738, + 0.014870469458401203, + 0.007774574216455221, + 0.00852931011468172, + 0.007019299082458019, + 0.0051527670584619045, + 0.0070364815182983875, + 0.007185582537204027, + 0.004097501747310162, + 0.0075472090393304825, + 0.012837927788496017, + 0.011746427975594997, + 0.015468724071979523, + 0.010675097815692425, + 0.008462341502308846, + 0.010953899472951889, + 0.010468555614352226, + 0.006282186601310968, + 0.010516060516238213, + 0.008605134673416615, + 0.017378097400069237, + 0.01177287194877863, + 0.005182509310543537, + 0.008506580255925655, + 0.007639751303941011, + 0.006818588823080063, + 0.011317203752696514, + 0.010392778553068638, + 0.010575314983725548, + 0.01204556692391634, + 0.010965623892843723, + 0.012138831429183483, + 0.017081087455153465, + 0.011727621778845787, + 0.006945536006242037, + 0.006567409262061119, + 0.006358670070767403, + 0.010617553256452084, + 0.00355032947845757, + 0.005572405643761158, + 0.004847231786698103, + 0.005912698805332184, + 0.0030562502797693014, + 0.002395958174020052, + 0.003971122670918703, + 0.0039650313556194305, + 0.001419025706127286, + 0.0023255504202097654, + 0.0018408525502309203, + 0.0025446403305977583, + 0.004947358276695013, + 0.004977287724614143, + 0.004328317474573851, + 0.005365314427763224, + 0.0043138861656188965, + 0.004851294215768576, + 0.00548202870413661, + 0.009869539178907871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.026334235444664955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle tracking shot slowly pans right across the icy landscape of a remote Arctic research station. A golden polar sun hangs high above the horizon, its slanting rays casting long blue shadows across the vast, pristine expanse of ice, outlining the texture of each snowdrift with remarkable clarity. In the foreground, a scientist, clad in a thick, bright orange thermal suit, a fully enclosed hood, and goggles that reflect the golden sunlight, stands out. Half-kneeling in the snow, he methodically adjusts the angle of a large solar panel, its deep blue photovoltaic surface coated with a thin layer of frost. One thickly gloved hand steadying the panel's cool metal frame, the other carefully tightening a connection point on the bracket with a tool, his movements steady and focused. Fine snowflakes swirl in the biting wind, some landing gently on the scientist's dark goggles and the black surface of the solar panel before being blown away. In the background, the research station’s silver geodesic dome and several communication antennas gleam with metallic luster in the golden afterglow. The whole scene forms a sharp contrast between the majestic nature and the tenacious technology of mankind, creating a tranquil atmosphere that is both lonely and full of hope.", + "seed": 25281, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_4405_seed_25281_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21915739430114628, + "vlm_logprob": -0.21915739430114628, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.21915739430114628, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.988124430179596 + ], + "opposite_logprob_tokens": [ + -0.48812443017959595 + ], + "logprob_reward": -0.988124430179596, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.988124430179596, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.057296786457300186, + 0.027049493044614792, + 0.017661454156041145, + 0.026707978919148445, + 0.0669371485710144, + 0.016005633398890495, + 0.01656440645456314, + 0.014640397392213345, + 0.7399479150772095, + 0.017188796773552895 + ], + "credit_map": [ + 0.0074713751673698425, + 0.008358893916010857, + 0.0071518500335514545, + 0.005252277012914419, + 0.00412982190027833, + 0.010168478824198246, + 0.004398164339363575, + 0.010365924797952175, + 0.00352755980566144, + 0.0029479707591235638, + 0.003437020815908909, + 0.003689310746267438, + 0.0017192292725667357, + 0.004106719046831131, + 0.0027784588746726513, + 0.004843222443014383, + 0.0030906288884580135, + 0.0033882695715874434, + 0.0022107157856225967, + 0.0027128467336297035, + 0.001642647897824645, + 0.00227039004676044, + 0.0012208245461806655, + 0.0011251303367316723, + 0.0037856660783290863, + 0.003156944876536727, + 0.004227685276418924, + 0.004259961657226086, + 0.002086313208565116, + 0.003520514350384474, + 0.0029392295982688665, + 0.0027316631749272346, + 0.010363838635385036, + 0.015470079146325588, + 0.006936713121831417, + 0.008757867850363255, + 0.006094134412705898, + 0.0074840690940618515, + 0.00641182716935873, + 0.00541861355304718, + 0.0018017477123066783, + 0.0014395557809621096, + 0.002100807847455144, + 0.001240910030901432, + 0.0013671612832695246, + 0.0017637931741774082, + 0.0032691489905118942, + 0.00302250892855227, + 0.0015861536376178265, + 0.0021016160026192665, + 0.001845208229497075, + 0.0023369607515633106, + 0.0018607706297188997, + 0.0015676298644393682, + 0.002484928583726287, + 0.0027811392210423946, + 0.001108342781662941, + 0.0013693846995010972, + 0.0015853309305384755, + 0.0013948812847957015, + 0.0016272771172225475, + 0.0013115848414599895, + 0.0033572737593203783, + 0.002886321162804961, + 0.023629337549209595, + 0.009662697091698647, + 0.007748583797365427, + 0.020426228642463684, + 0.027180485427379608, + 0.021183080971240997, + 0.3204399347305298, + 0.3096776306629181, + 0.0016557954950258136, + 0.002116444753482938, + 0.0027965966146439314, + 0.001898996066302061, + 0.0016196616925299168, + 0.001353653846308589, + 0.0028128589037805796, + 0.0029347895178943872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.988124430179596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005969078280031681 + ], + "opposite_logprob_tokens": [ + -5.130969047546387 + ], + "logprob_reward": -0.005969078280031681, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005969078280031681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17432741820812225, + 0.08039391785860062, + 0.034655630588531494, + 0.045423321425914764, + 0.0314648374915123, + 0.05233476310968399, + 0.07223052531480789, + 0.1797132045030594, + 0.26068755984306335, + 0.06876882910728455 + ], + "credit_map": [ + 0.02142530493438244, + 0.019901785999536514, + 0.02333197183907032, + 0.020496517419815063, + 0.018351873382925987, + 0.04047297313809395, + 0.01412736251950264, + 0.01621962897479534, + 0.009869212284684181, + 0.008577273227274418, + 0.010221903212368488, + 0.006894667167216539, + 0.005530229303985834, + 0.011295451782643795, + 0.008806062862277031, + 0.019199112430214882, + 0.005652569700032473, + 0.0033794010523706675, + 0.004749649204313755, + 0.0053558568470180035, + 0.003511548275128007, + 0.005917429458349943, + 0.002841136883944273, + 0.0032480373047292233, + 0.006016841158270836, + 0.008634332567453384, + 0.006755529437214136, + 0.00640072813257575, + 0.005098658613860607, + 0.00579780712723732, + 0.003413981758058071, + 0.003305439604446292, + 0.0039947982877492905, + 0.004794442560523748, + 0.004510532598942518, + 0.005250250920653343, + 0.0021121022291481495, + 0.0030632701236754656, + 0.004015682730823755, + 0.0037237561773508787, + 0.007621166296303272, + 0.0059738000854849815, + 0.00841100700199604, + 0.0040095350705087185, + 0.007549741771072149, + 0.006475677713751793, + 0.006803567986935377, + 0.005490265320986509, + 0.006866921670734882, + 0.014320498332381248, + 0.009004762396216393, + 0.01152295246720314, + 0.006736882962286472, + 0.0063265105709433556, + 0.009095304645597935, + 0.00835668109357357, + 0.01682853326201439, + 0.021375320851802826, + 0.017025014385581017, + 0.01479765959084034, + 0.01070336066186428, + 0.00795014202594757, + 0.05106233432888985, + 0.03997083008289337, + 0.018472524359822273, + 0.007699890062212944, + 0.00511120306327939, + 0.010105527937412262, + 0.013616538606584072, + 0.009649036452174187, + 0.1035708636045456, + 0.0924619659781456, + 0.006103597115725279, + 0.00785995926707983, + 0.00702914223074913, + 0.007953980006277561, + 0.005531957373023033, + 0.005413153674453497, + 0.011719577014446259, + 0.017157459631562233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.005969078280031681 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954266060143709 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954266060143709, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954266060143709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03496923670172691, + 0.06063098832964897, + 0.020701831206679344, + 0.016914503648877144, + 0.020153062418103218, + 0.015803148970007896, + 0.057392142713069916, + 0.05227232724428177, + 0.6967252492904663, + 0.02443745546042919 + ], + "credit_map": [ + 0.005726621486246586, + 0.0052268183790147305, + 0.005735862068831921, + 0.004283057525753975, + 0.004487484227865934, + 0.0037073183339089155, + 0.0029675234109163284, + 0.002834548708051443, + 0.007524043321609497, + 0.006614641286432743, + 0.008418605662882328, + 0.005123474635183811, + 0.0037285799626260996, + 0.009739706292748451, + 0.006099984049797058, + 0.013381951488554478, + 0.0026255622506141663, + 0.0020461324602365494, + 0.002525837393477559, + 0.002759306924417615, + 0.0022275636438280344, + 0.004075181670486927, + 0.001866478705778718, + 0.0025757679250091314, + 0.002498394111171365, + 0.0030816439539194107, + 0.0018506721826270223, + 0.0023238544818013906, + 0.0014742572093382478, + 0.0019292272627353668, + 0.0020422027446329594, + 0.0017142519354820251, + 0.003454059362411499, + 0.0044528767466545105, + 0.0019446638179942966, + 0.0028354038950055838, + 0.0017695670248940587, + 0.001965549075976014, + 0.0019306465983390808, + 0.0018002955475822091, + 0.001865069498308003, + 0.0015802463749423623, + 0.0019295994425192475, + 0.0018075078260153532, + 0.0020373899023979902, + 0.0019324218155816197, + 0.002548812422901392, + 0.0021021028514951468, + 0.006796304602175951, + 0.007245076820254326, + 0.009505059570074081, + 0.0071945772506296635, + 0.005356733221560717, + 0.006226425524801016, + 0.008275911211967468, + 0.006792054977267981, + 0.004202761687338352, + 0.005182584747672081, + 0.006519137416034937, + 0.004815936088562012, + 0.004371303599327803, + 0.003940603230148554, + 0.012770939618349075, + 0.01046905666589737, + 0.022278182208538055, + 0.008916454389691353, + 0.0072240340523421764, + 0.019050605595111847, + 0.025707092136144638, + 0.020178772509098053, + 0.3018271327018738, + 0.29154297709465027, + 0.0022647769656032324, + 0.0025747327599674463, + 0.003525643376633525, + 0.0022953476291149855, + 0.002161209937185049, + 0.0014269181992858648, + 0.004094482399523258, + 0.006094345357269049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.005954266060143709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369348868727684 + ], + "opposite_logprob_tokens": [ + -3.408693552017212 + ], + "logprob_reward": -0.03369348868727684, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369348868727684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11057896167039871, + 0.05406125262379646, + 0.0674498900771141, + 0.04179862141609192, + 0.06507358700037003, + 0.0895184576511383, + 0.06803148984909058, + 0.05566124618053436, + 0.38786569237709045, + 0.0599607415497303 + ], + "credit_map": [ + 0.013700155541300774, + 0.011818334460258484, + 0.01162791159003973, + 0.012480917386710644, + 0.015339275822043419, + 0.01789422333240509, + 0.00910219457000494, + 0.018615957349538803, + 0.006296616513282061, + 0.0049479249864816666, + 0.006119956728070974, + 0.006589309778064489, + 0.006166765000671148, + 0.011211040429770947, + 0.0055433353409171104, + 0.007186307106167078, + 0.014351483434438705, + 0.01055937074124813, + 0.011124900542199612, + 0.006572749000042677, + 0.007091752719134092, + 0.0076725841499865055, + 0.005414173472672701, + 0.004662883002310991, + 0.006493946071714163, + 0.006739526521414518, + 0.00399535009637475, + 0.005525978282094002, + 0.004230951424688101, + 0.004622846841812134, + 0.005558385048061609, + 0.004631640389561653, + 0.010251706466078758, + 0.013499222695827484, + 0.007963230833411217, + 0.007842657156288624, + 0.006411486305296421, + 0.007222807966172695, + 0.006870676297694445, + 0.005011802073568106, + 0.01222345232963562, + 0.006763850338757038, + 0.013664009980857372, + 0.006204227916896343, + 0.012709429487586021, + 0.012521862052381039, + 0.014459686353802681, + 0.010971949435770512, + 0.00805571861565113, + 0.003961104899644852, + 0.012893789447844028, + 0.004993172828108072, + 0.0092175817117095, + 0.00884135439991951, + 0.01007626112550497, + 0.009992516599595547, + 0.005326082929968834, + 0.0038749012164771557, + 0.00610037986189127, + 0.004861111752688885, + 0.009394577704370022, + 0.007154393009841442, + 0.009923925623297691, + 0.009025877341628075, + 0.03876561298966408, + 0.01797259971499443, + 0.019350353628396988, + 0.022045904770493507, + 0.041411008685827255, + 0.044258758425712585, + 0.09810278564691544, + 0.10595869272947311, + 0.005577987991273403, + 0.004566087853163481, + 0.004949321039021015, + 0.00474180793389678, + 0.004387914668768644, + 0.003597984788939357, + 0.015782615169882774, + 0.016357026994228363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.03369348868727684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045708298683167 + ], + "opposite_logprob_tokens": [ + -2.8120458126068115 + ], + "logprob_reward": -0.062045708298683167, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045708298683167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11544115096330643, + 0.036820415407419205, + 0.00960599072277546, + 0.010477479547262192, + 0.03911833465099335, + 0.007451055571436882, + 0.01168036088347435, + 0.01973005011677742, + 0.7174081206321716, + 0.032267093658447266 + ], + "credit_map": [ + 0.015943368896842003, + 0.01605655811727047, + 0.012227502651512623, + 0.010304087772965431, + 0.008373597636818886, + 0.01920982450246811, + 0.009927828796207905, + 0.023398369550704956, + 0.0044516162015497684, + 0.003983248025178909, + 0.004678420722484589, + 0.0030306752305477858, + 0.0024182498455047607, + 0.005749731324613094, + 0.004087828099727631, + 0.008420641534030437, + 0.0014954200014472008, + 0.0018102776957675815, + 0.0012472057715058327, + 0.0014964203583076596, + 0.0010379509767517447, + 0.0009252015152014792, + 0.000866036512888968, + 0.000727477774489671, + 0.001656046137213707, + 0.002391767455264926, + 0.0013506785035133362, + 0.0013437598245218396, + 0.0012065935879945755, + 0.0010901095811277628, + 0.0008023061091080308, + 0.0006362174754031003, + 0.00671269791200757, + 0.009730024263262749, + 0.004260721616446972, + 0.005135968327522278, + 0.003530059475451708, + 0.00406141672283411, + 0.0034058564342558384, + 0.0022815880365669727, + 0.0008080666302703321, + 0.0006913481629453599, + 0.0008446851279586554, + 0.0005949182668700814, + 0.0010092556476593018, + 0.0008922719862312078, + 0.0013474177103489637, + 0.0012630916899070144, + 0.0011624470353126526, + 0.001548907719552517, + 0.0015901443548500538, + 0.0014802124351263046, + 0.0012443255400285125, + 0.0012863528681918979, + 0.00180635426659137, + 0.0015616160817444324, + 0.0016943304799497128, + 0.0025365417823195457, + 0.002610395196825266, + 0.0021407846361398697, + 0.0022748971823602915, + 0.0018222039798274636, + 0.0035770758986473083, + 0.0030738182831555605, + 0.024465098977088928, + 0.009437850676476955, + 0.007194029167294502, + 0.02003733441233635, + 0.02671028859913349, + 0.021062931045889854, + 0.31025707721710205, + 0.2982434332370758, + 0.0031754509545862675, + 0.0019672478083521128, + 0.0022017767187207937, + 0.0025452713016420603, + 0.002094028051942587, + 0.0018227689433842897, + 0.009006589651107788, + 0.00945395790040493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.062045708298683167 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle tracking shot slowly pans right across the icy landscape of a remote Arctic research station. A golden polar sun hangs high above the horizon, its slanting rays casting long blue shadows across the vast, pristine expanse of ice, outlining the texture of each snowdrift with remarkable clarity. In the foreground, a scientist, clad in a thick, bright orange thermal suit, a fully enclosed hood, and goggles that reflect the golden sunlight, stands out. Half-kneeling in the snow, he methodically adjusts the angle of a large solar panel, its deep blue photovoltaic surface coated with a thin layer of frost. One thickly gloved hand steadying the panel's cool metal frame, the other carefully tightening a connection point on the bracket with a tool, his movements steady and focused. Fine snowflakes swirl in the biting wind, some landing gently on the scientist's dark goggles and the black surface of the solar panel before being blown away. In the background, the research station’s silver geodesic dome and several communication antennas gleam with metallic luster in the golden afterglow. The whole scene forms a sharp contrast between the majestic nature and the tenacious technology of mankind, creating a tranquil atmosphere that is both lonely and full of hope.", + "seed": 25282, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_4405_seed_25282_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18157405527308584, + "vlm_logprob": -0.18157405527308584, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.18157405527308584, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7677204608917236 + ], + "opposite_logprob_tokens": [ + -0.6427204608917236 + ], + "logprob_reward": -0.7677204608917236, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7677204608917236, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06798576563596725, + 0.05384132266044617, + 0.05597609654068947, + 0.04795319214463234, + 0.026931913569569588, + 0.5249987840652466, + 0.0437605082988739, + 0.050500109791755676, + 0.08263412863016129, + 0.04541816562414169 + ], + "credit_map": [ + 0.007272906601428986, + 0.0057991365902125835, + 0.0061593311838805676, + 0.005213172174990177, + 0.011512325145304203, + 0.010573840700089931, + 0.011782573536038399, + 0.009672482497990131, + 0.0036074649542570114, + 0.003166208742186427, + 0.0029390817508101463, + 0.0031795399263501167, + 0.012851854786276817, + 0.010399672202765942, + 0.010525277815759182, + 0.007172221317887306, + 0.007362002041190863, + 0.0030000500846654177, + 0.0034025697968900204, + 0.005362592171877623, + 0.004568676929920912, + 0.005360757932066917, + 0.015219382010400295, + 0.011700070463120937, + 0.005629519000649452, + 0.006878366693854332, + 0.0043479325249791145, + 0.005421127658337355, + 0.005819226149469614, + 0.005425442010164261, + 0.008795871399343014, + 0.005635708570480347, + 0.0034477112349122763, + 0.004201633855700493, + 0.002317683305591345, + 0.0024868114851415157, + 0.0031312538776546717, + 0.0038757307920604944, + 0.004017692059278488, + 0.0034533990547060966, + 0.015329158864915371, + 0.010612784884870052, + 0.010626482777297497, + 0.00621301494538784, + 0.23175030946731567, + 0.2263408899307251, + 0.011406173929572105, + 0.012720010243356228, + 0.004746753256767988, + 0.004472560714930296, + 0.003003610298037529, + 0.002957703312858939, + 0.006189940497279167, + 0.0065192426554858685, + 0.008558252826333046, + 0.007312444970011711, + 0.004264078103005886, + 0.008432691916823387, + 0.005221846047788858, + 0.005515932571142912, + 0.006622446235269308, + 0.005472390912473202, + 0.006294204853475094, + 0.008676517754793167, + 0.011827977374196053, + 0.010810906998813152, + 0.00774976285174489, + 0.011347546242177486, + 0.011062723584473133, + 0.007012865971773863, + 0.01126490905880928, + 0.011557439342141151, + 0.006296943873167038, + 0.0032786796800792217, + 0.0029536900110542774, + 0.0038321800529956818, + 0.005401880946010351, + 0.004095322452485561, + 0.010651892051100731, + 0.008907578885555267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.7677204608917236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110459703952074 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110459703952074, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110459703952074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0599193312227726, + 0.10458756983280182, + 0.06459454447031021, + 0.016775522381067276, + 0.5499017834663391, + 0.09619124978780746, + 0.026143863797187805, + 0.013253201730549335, + 0.03809037804603577, + 0.030542535707354546 + ], + "credit_map": [ + 0.006378219462931156, + 0.004596271086484194, + 0.004234802909195423, + 0.005457939580082893, + 0.008211681619286537, + 0.007493141572922468, + 0.01406231988221407, + 0.00948495976626873, + 0.008715423755347729, + 0.006045743357390165, + 0.004098417237401009, + 0.005986332893371582, + 0.019077090546488762, + 0.0182512067258358, + 0.026074901223182678, + 0.016338463872671127, + 0.007564560975879431, + 0.0058160945773124695, + 0.0032089019659906626, + 0.004687098320573568, + 0.006718731019645929, + 0.007166485767811537, + 0.017137626186013222, + 0.012295051477849483, + 0.0021057387348264456, + 0.002169740851968527, + 0.0012950045056641102, + 0.0018125249771401286, + 0.0014096831437200308, + 0.001403382862918079, + 0.0039971210062503815, + 0.00258232606574893, + 0.016671283170580864, + 0.010850400663912296, + 0.007590888999402523, + 0.006637013051658869, + 0.2494787871837616, + 0.23766861855983734, + 0.01086006872355938, + 0.010144724510610104, + 0.0029705853667110205, + 0.0032548061572015285, + 0.0024561977479606867, + 0.0017871648306027055, + 0.040926381945610046, + 0.04003272205591202, + 0.00239138538017869, + 0.0023720180615782738, + 0.0027608489617705345, + 0.005367226433008909, + 0.002857989864423871, + 0.0026581217534840107, + 0.00337467179633677, + 0.0033186255022883415, + 0.002964962041005492, + 0.002841418841853738, + 0.0013826036592945457, + 0.0022400538437068462, + 0.0013850784162059426, + 0.0011925013968721032, + 0.0017432223539799452, + 0.001262328471057117, + 0.0016922326758503914, + 0.0023551820777356625, + 0.006669803988188505, + 0.002878891071304679, + 0.002096962183713913, + 0.0025569526478648186, + 0.005114318802952766, + 0.004378407262265682, + 0.0075832330621778965, + 0.006811813451349735, + 0.004345579072833061, + 0.0016543697565793991, + 0.001829679124057293, + 0.0017764851218089461, + 0.0045536113902926445, + 0.002538887783885002, + 0.007880029268562794, + 0.005963895469903946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.004110459703952074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077335104346275 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077335104346275, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077335104346275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03058869019150734, + 0.3188340961933136, + 0.04258345067501068, + 0.01980559341609478, + 0.015852663666009903, + 0.367258220911026, + 0.013299193233251572, + 0.021953240036964417, + 0.10626783967018127, + 0.0635569766163826 + ], + "credit_map": [ + 0.003388149430975318, + 0.002371647162362933, + 0.0028752360958606005, + 0.0024572850670665503, + 0.0045837960205972195, + 0.004982725251466036, + 0.005829032976180315, + 0.004100820515304804, + 0.02224052883684635, + 0.013043588027358055, + 0.008334288373589516, + 0.0153681430965662, + 0.07867752015590668, + 0.06630506366491318, + 0.06677970290184021, + 0.0480852834880352, + 0.005575301591306925, + 0.0021995543502271175, + 0.0020296310540288687, + 0.003324887715280056, + 0.0034941735211759806, + 0.0039530303329229355, + 0.012456836178898811, + 0.00955003872513771, + 0.002314358949661255, + 0.0027027863543480635, + 0.0020011477172374725, + 0.0026223468594253063, + 0.0015403109136968851, + 0.0019245145376771688, + 0.004022036213427782, + 0.002678093034774065, + 0.0010051833232864738, + 0.0011452801991254091, + 0.0006439489661715925, + 0.0006959308520890772, + 0.004914025776088238, + 0.004982009530067444, + 0.0011836866615340114, + 0.001282600685954094, + 0.010907086543738842, + 0.005835642572492361, + 0.0070771933533251286, + 0.003996843472123146, + 0.16395717859268188, + 0.1599600613117218, + 0.007323022000491619, + 0.008201210759580135, + 0.0017172356601804495, + 0.001633236650377512, + 0.0011242955224588513, + 0.001060206675902009, + 0.001756132347509265, + 0.0018629322294145823, + 0.0023659542202949524, + 0.0017792010912671685, + 0.0022858320735394955, + 0.002588673261925578, + 0.0017969895852729678, + 0.0016883971402421594, + 0.0031188849825412035, + 0.002656640252098441, + 0.003376459004357457, + 0.0044413660652935505, + 0.019103532657027245, + 0.0072285765781998634, + 0.0060935732908546925, + 0.00786104891449213, + 0.013771258294582367, + 0.013343146070837975, + 0.01955723948776722, + 0.01930946484208107, + 0.009100250899791718, + 0.003706266637891531, + 0.0036037913523614407, + 0.003469025483354926, + 0.007988808676600456, + 0.005455667618662119, + 0.016924014315009117, + 0.013309154659509659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.011077335104346275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494843050837517 + ], + "opposite_logprob_tokens": [ + -2.929948329925537 + ], + "logprob_reward": -0.05494843050837517, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494843050837517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19282414019107819, + 0.12794867157936096, + 0.3650711476802826, + 0.059207819402217865, + 0.041977763175964355, + 0.11823613196611404, + 0.026968322694301605, + 0.027530314400792122, + 0.030702082440257072, + 0.009533627890050411 + ], + "credit_map": [ + 0.017704758793115616, + 0.012908300384879112, + 0.010517315939068794, + 0.012689057737588882, + 0.03319339081645012, + 0.029885176569223404, + 0.04529493674635887, + 0.030631206929683685, + 0.006538253277540207, + 0.00596799049526453, + 0.0037750841584056616, + 0.004708993714302778, + 0.04272536188364029, + 0.03414696082472801, + 0.012208046391606331, + 0.017877981066703796, + 0.04575450345873833, + 0.015984920784831047, + 0.013998770155012608, + 0.028446970507502556, + 0.0325191468000412, + 0.03511808440089226, + 0.10994721949100494, + 0.08330153673887253, + 0.008238309994339943, + 0.0031221145763993263, + 0.0028680574614554644, + 0.004592708311975002, + 0.005591641645878553, + 0.007770979311317205, + 0.014583520591259003, + 0.012440484017133713, + 0.004730252083390951, + 0.002897563623264432, + 0.0021682842634618282, + 0.0021113345865160227, + 0.00825183093547821, + 0.00737486407160759, + 0.007170899771153927, + 0.007272737566381693, + 0.005604921840131283, + 0.004222484305500984, + 0.003807857632637024, + 0.0025329729542136192, + 0.0436575785279274, + 0.04177035763859749, + 0.00882714707404375, + 0.007812819443643093, + 0.0020134614314883947, + 0.003642202354967594, + 0.002077636541798711, + 0.0018572730477899313, + 0.005269910674542189, + 0.005631040781736374, + 0.0035316809080541134, + 0.0029451188165694475, + 0.0034132932778447866, + 0.003436095779761672, + 0.003169911215081811, + 0.002703477628529072, + 0.0035159687977284193, + 0.00394332455471158, + 0.003376473905518651, + 0.003971768077462912, + 0.003222979139536619, + 0.003347361460328102, + 0.003159400774165988, + 0.0029406417161226273, + 0.0042172991670668125, + 0.003087659366428852, + 0.0048411195166409016, + 0.005885622464120388, + 0.0015181121416389942, + 0.0011050580069422722, + 0.0006581720081157982, + 0.0008082800195552409, + 0.0013689687475562096, + 0.0010190738830715418, + 0.0012103329645469785, + 0.001845630002208054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.05494843050837517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001359015703201 + ], + "opposite_logprob_tokens": [ + -2.6950135231018066 + ], + "logprob_reward": -0.07001359015703201, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001359015703201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08201751112937927, + 0.11936824023723602, + 0.025685736909508705, + 0.04952745512127876, + 0.2423345148563385, + 0.1489294469356537, + 0.04936018958687782, + 0.07953003793954849, + 0.1131662130355835, + 0.09008067101240158 + ], + "credit_map": [ + 0.008357718586921692, + 0.009752433747053146, + 0.007374464068561792, + 0.008188780397176743, + 0.011893133632838726, + 0.014375337399542332, + 0.012703144922852516, + 0.009372499771416187, + 0.012658595107495785, + 0.009427684359252453, + 0.006173178553581238, + 0.008916615508496761, + 0.012770939618349075, + 0.018715592101216316, + 0.030155416578054428, + 0.020550217479467392, + 0.0033457449171692133, + 0.003309470834210515, + 0.0026076282374560833, + 0.002126120263710618, + 0.002528857672587037, + 0.0033272875007241964, + 0.005047688726335764, + 0.003392938757315278, + 0.005766474176198244, + 0.0029037497006356716, + 0.002687803702428937, + 0.0034725377336144447, + 0.006234943401068449, + 0.008350690826773643, + 0.011351007968187332, + 0.008760250173509121, + 0.00853120069950819, + 0.005348385311663151, + 0.00341089628636837, + 0.003709787270054221, + 0.10594059526920319, + 0.10120183974504471, + 0.00749256880953908, + 0.006699242629110813, + 0.012444819323718548, + 0.008881691843271255, + 0.007860572077333927, + 0.005390745587646961, + 0.042286042124032974, + 0.04044043645262718, + 0.01610989309847355, + 0.01551524642854929, + 0.004882433917373419, + 0.00440170569345355, + 0.0038102027028799057, + 0.0029390635900199413, + 0.008271473459899426, + 0.007982481271028519, + 0.00832012016326189, + 0.008752705529332161, + 0.009843830019235611, + 0.005927803460508585, + 0.004575374070554972, + 0.003802425228059292, + 0.012597832828760147, + 0.013540479354560375, + 0.014597095549106598, + 0.014645196497440338, + 0.016602642834186554, + 0.013914532959461212, + 0.007685792166739702, + 0.014806169085204601, + 0.012940388172864914, + 0.011170264333486557, + 0.016682298853993416, + 0.01936413161456585, + 0.013525609858334064, + 0.008254783228039742, + 0.006632549222558737, + 0.007299517281353474, + 0.011492905206978321, + 0.007604213897138834, + 0.016723405569791794, + 0.01854768767952919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.07001359015703201 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle tracking shot slowly pans right across the icy landscape of a remote Arctic research station. A golden polar sun hangs high above the horizon, its slanting rays casting long blue shadows across the vast, pristine expanse of ice, outlining the texture of each snowdrift with remarkable clarity. In the foreground, a scientist, clad in a thick, bright orange thermal suit, a fully enclosed hood, and goggles that reflect the golden sunlight, stands out. Half-kneeling in the snow, he methodically adjusts the angle of a large solar panel, its deep blue photovoltaic surface coated with a thin layer of frost. One thickly gloved hand steadying the panel's cool metal frame, the other carefully tightening a connection point on the bracket with a tool, his movements steady and focused. Fine snowflakes swirl in the biting wind, some landing gently on the scientist's dark goggles and the black surface of the solar panel before being blown away. In the background, the research station’s silver geodesic dome and several communication antennas gleam with metallic luster in the golden afterglow. The whole scene forms a sharp contrast between the majestic nature and the tenacious technology of mankind, creating a tranquil atmosphere that is both lonely and full of hope.", + "seed": 25283, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_10/index_4405_seed_25283_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.131517810607329, + "vlm_logprob": -0.131517810607329, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.131517810607329, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.588939368724823 + ], + "opposite_logprob_tokens": [ + -0.838939368724823 + ], + "logprob_reward": -0.588939368724823, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.588939368724823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22562552988529205, + 0.07901624590158463, + 0.09605535864830017, + 0.1101083904504776, + 0.06321531534194946, + 0.08565998077392578, + 0.07846520841121674, + 0.07869155704975128, + 0.11748544871807098, + 0.0656769797205925 + ], + "credit_map": [ + 0.0345008447766304, + 0.013697641901671886, + 0.03490278124809265, + 0.012884141877293587, + 0.022558780387043953, + 0.05215590447187424, + 0.015798965469002724, + 0.039126455783843994, + 0.01343648973852396, + 0.0049880193546414375, + 0.012590726837515831, + 0.0048195356503129005, + 0.007869482971727848, + 0.01988343335688114, + 0.0045462241396307945, + 0.01088232733309269, + 0.013268020935356617, + 0.0058631813153624535, + 0.018085263669490814, + 0.005367244128137827, + 0.010260806418955326, + 0.022409645840525627, + 0.005415224935859442, + 0.015385964885354042, + 0.014487349428236485, + 0.009596533142030239, + 0.016559280455112457, + 0.013367823325097561, + 0.010642293840646744, + 0.01966894418001175, + 0.009628028608858585, + 0.01615813747048378, + 0.008726987056434155, + 0.00611674040555954, + 0.00811091810464859, + 0.0051251971162855625, + 0.003972819074988365, + 0.0132674565538764, + 0.004261726047843695, + 0.013633464463055134, + 0.012899299152195454, + 0.006433695089071989, + 0.012528619728982449, + 0.006374881137162447, + 0.009993108920753002, + 0.01661745458841324, + 0.005980394780635834, + 0.014832527376711369, + 0.010909263044595718, + 0.006268346682190895, + 0.011009450070559978, + 0.007087478414177895, + 0.007592260371893644, + 0.01732722669839859, + 0.007925095036625862, + 0.010346083901822567, + 0.012843121774494648, + 0.0064446995966136456, + 0.009705142118036747, + 0.004989929031580687, + 0.009823587723076344, + 0.022222917526960373, + 0.004953611176460981, + 0.007708549499511719, + 0.022607212886214256, + 0.008776843547821045, + 0.01105727069079876, + 0.010033716447651386, + 0.007663997355848551, + 0.020751431584358215, + 0.01417661551386118, + 0.02241836115717888, + 0.00936878565698862, + 0.007862536236643791, + 0.005895939655601978, + 0.006949655711650848, + 0.006607662886381149, + 0.015289408154785633, + 0.007164903916418552, + 0.006538085173815489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.588939368724823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028363734018057585 + ], + "opposite_logprob_tokens": [ + -5.877836227416992 + ], + "logprob_reward": -0.0028363734018057585, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028363734018057585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14375479519367218, + 0.17186379432678223, + 0.1988522708415985, + 0.0692436695098877, + 0.039219219237565994, + 0.04984183609485626, + 0.040869418531656265, + 0.16656357049942017, + 0.06022419035434723, + 0.05956723541021347 + ], + "credit_map": [ + 0.021238934248685837, + 0.011746039614081383, + 0.019300054758787155, + 0.0148523123934865, + 0.01255202479660511, + 0.030299058184027672, + 0.0181773379445076, + 0.015589041635394096, + 0.02228705585002899, + 0.023126404732465744, + 0.027301348745822906, + 0.018825847655534744, + 0.01733734831213951, + 0.018400685861706734, + 0.017258983105421066, + 0.02732611447572708, + 0.026683248579502106, + 0.018431248143315315, + 0.029050754383206367, + 0.017734745517373085, + 0.020508551970124245, + 0.03614387288689613, + 0.018968500196933746, + 0.03133136406540871, + 0.009318426251411438, + 0.006701110862195492, + 0.011149059049785137, + 0.006911627482622862, + 0.00598875479772687, + 0.011529783718287945, + 0.005436578765511513, + 0.012208330444991589, + 0.0053087384440004826, + 0.003984539303928614, + 0.005148154217749834, + 0.0031262284610420465, + 0.003547649597749114, + 0.006871585268527269, + 0.0028097874019294977, + 0.008422533981502056, + 0.006495983339846134, + 0.005717191379517317, + 0.006343000568449497, + 0.004664561711251736, + 0.00638627028092742, + 0.00799550861120224, + 0.004939199425280094, + 0.007300121244043112, + 0.006888360716402531, + 0.004766033496707678, + 0.004858565982431173, + 0.003279614495113492, + 0.005577278323471546, + 0.007718556094914675, + 0.003149250987917185, + 0.004631758201867342, + 0.02584579400718212, + 0.017243657261133194, + 0.019047211855649948, + 0.01657669059932232, + 0.01071805227547884, + 0.04907945170998573, + 0.010753693990409374, + 0.017299026250839233, + 0.011113016866147518, + 0.0035948066506534815, + 0.008418777957558632, + 0.0045720478519797325, + 0.00430563697591424, + 0.012524276040494442, + 0.007021085359156132, + 0.008674541488289833, + 0.00882929377257824, + 0.004345909226685762, + 0.0055358316749334335, + 0.0036664907820522785, + 0.00966542400419712, + 0.008756027556955814, + 0.01073376089334488, + 0.00803449284285307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.0028363734018057585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640405531972647 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640405531972647, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640405531972647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16984562575817108, + 0.19335071742534637, + 0.03559364378452301, + 0.0334773063659668, + 0.0419132336974144, + 0.04044013097882271, + 0.08037526905536652, + 0.16357757151126862, + 0.046267565339803696, + 0.195158913731575 + ], + "credit_map": [ + 0.027881987392902374, + 0.010401998646557331, + 0.021081117913126945, + 0.014966992661356926, + 0.023181157186627388, + 0.03580376133322716, + 0.01370034459978342, + 0.022828267887234688, + 0.028383595868945122, + 0.00889596901834011, + 0.03687513992190361, + 0.0124704260379076, + 0.015826115384697914, + 0.029999805614352226, + 0.008759862743318081, + 0.052139800041913986, + 0.004801678936928511, + 0.0038632680661976337, + 0.004731039050966501, + 0.003437970532104373, + 0.0040794615633785725, + 0.006515148561447859, + 0.0028753967490047216, + 0.005289680790156126, + 0.004830922000110149, + 0.002871852368116379, + 0.005507605150341988, + 0.0033785910345613956, + 0.0033101278822869062, + 0.006203646305948496, + 0.003083230694755912, + 0.004291333258152008, + 0.005392283201217651, + 0.005592449102550745, + 0.0048793042078614235, + 0.005825722124427557, + 0.0035240836441516876, + 0.006053019780665636, + 0.0033798927906900644, + 0.007266476284712553, + 0.004939362406730652, + 0.004140847362577915, + 0.006170143838971853, + 0.004587413277477026, + 0.004480171948671341, + 0.006870806682854891, + 0.0036328518763184547, + 0.005618532188236713, + 0.012093039229512215, + 0.007783510256558657, + 0.009425753727555275, + 0.006189027801156044, + 0.013162651099264622, + 0.01591845229268074, + 0.006877612788230181, + 0.008925220929086208, + 0.024465778842568398, + 0.01715001091361046, + 0.02030029520392418, + 0.018453383818268776, + 0.018909621983766556, + 0.03455676510930061, + 0.011570543050765991, + 0.018171176314353943, + 0.009472054429352283, + 0.0024246377870440483, + 0.0035353226121515036, + 0.002311201300472021, + 0.002756665926426649, + 0.014821349643170834, + 0.0026082948315888643, + 0.008338036015629768, + 0.02739221602678299, + 0.012086157687008381, + 0.015319726429879665, + 0.013279694132506847, + 0.02529917098581791, + 0.035307951271533966, + 0.04213517904281616, + 0.024338828399777412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.004640405531972647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533116154372692 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533116154372692, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533116154372692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19891805946826935, + 0.18599627912044525, + 0.15859316289424896, + 0.08117423951625824, + 0.03032313659787178, + 0.06573808193206787, + 0.07825317233800888, + 0.06081583723425865, + 0.08223274350166321, + 0.057955287396907806 + ], + "credit_map": [ + 0.03212854638695717, + 0.012625453062355518, + 0.027979839593172073, + 0.013502702116966248, + 0.026058750227093697, + 0.055526863783597946, + 0.01329940464347601, + 0.01779649779200554, + 0.02906819060444832, + 0.010119495913386345, + 0.03491470217704773, + 0.01015172153711319, + 0.019722893834114075, + 0.038056742399930954, + 0.010270152240991592, + 0.0336923785507679, + 0.026282282546162605, + 0.009536480531096458, + 0.022621721029281616, + 0.008109305053949356, + 0.015729928389191628, + 0.049001727253198624, + 0.009927581064403057, + 0.017384130507707596, + 0.01314999908208847, + 0.0066951182670891285, + 0.010139256715774536, + 0.006483450531959534, + 0.009062539786100388, + 0.02067320980131626, + 0.005026394966989756, + 0.009944276884198189, + 0.004476155620068312, + 0.0039522540755569935, + 0.004304810892790556, + 0.0036774170584976673, + 0.0027317609637975693, + 0.004834793042391539, + 0.0019542661029845476, + 0.00439167907461524, + 0.010277435183525085, + 0.003873781766742468, + 0.008605913259088993, + 0.004656115546822548, + 0.007584027945995331, + 0.014544266276061535, + 0.0033414424397051334, + 0.01285509392619133, + 0.013426495715975761, + 0.006307787727564573, + 0.009386311285197735, + 0.004711883142590523, + 0.008612575009465218, + 0.022000005468726158, + 0.005308401305228472, + 0.008499713614583015, + 0.009308867156505585, + 0.005508657544851303, + 0.014668446034193039, + 0.003468071110546589, + 0.006657837890088558, + 0.010156506672501564, + 0.004090289119631052, + 0.006957160774618387, + 0.014026064425706863, + 0.004907573573291302, + 0.005276205483824015, + 0.005523748695850372, + 0.008565050549805164, + 0.01711721159517765, + 0.0123570766299963, + 0.014459808357059956, + 0.007141107693314552, + 0.004443306475877762, + 0.00337039353325963, + 0.003907658159732819, + 0.005925965029746294, + 0.01019599661231041, + 0.014996522106230259, + 0.007974335923790932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012533116154372692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle tracking shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a scientist wearing a thick bright orange thermal suit with a fully enclosed hood and goggles visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scientist half-kneeling in the snow while adjusting a large solar panel with a deep blue photovoltaic surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background featuring a silver geodesic dome and communication antennas gleaming in the golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fine snowflakes visible swirling in the wind and landing on the scientist's goggles and the solar panel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863978922367096 + ], + "opposite_logprob_tokens": [ + -3.0486397743225098 + ], + "logprob_reward": -0.04863978922367096, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863978922367096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2657005786895752, + 0.09447374939918518, + 0.08425693213939667, + 0.07211904972791672, + 0.030311180278658867, + 0.08645373582839966, + 0.07622803747653961, + 0.1519526243209839, + 0.05179743468761444, + 0.08670663833618164 + ], + "credit_map": [ + 0.04095030575990677, + 0.016619384288787842, + 0.03779760003089905, + 0.01738416962325573, + 0.02902008220553398, + 0.0697636753320694, + 0.020762134343385696, + 0.03340325132012367, + 0.01772848144173622, + 0.005403307266533375, + 0.01584351621568203, + 0.006922910921275616, + 0.009638798423111439, + 0.02278667502105236, + 0.003994803875684738, + 0.012155251577496529, + 0.01351091731339693, + 0.0041197226382792, + 0.01243117917329073, + 0.005489160306751728, + 0.006354470271617174, + 0.029543589800596237, + 0.004792114719748497, + 0.008015780709683895, + 0.01105518452823162, + 0.004956051241606474, + 0.012285982258617878, + 0.005565581377595663, + 0.00832256767898798, + 0.01523568108677864, + 0.005697418469935656, + 0.009000577963888645, + 0.004795736633241177, + 0.0029676109552383423, + 0.004120228812098503, + 0.002855942351743579, + 0.0026967953890562057, + 0.004435328301042318, + 0.0021175050642341375, + 0.006322032306343317, + 0.01127170491963625, + 0.006505394820123911, + 0.01098032109439373, + 0.006310153286904097, + 0.015190310776233673, + 0.0166405588388443, + 0.006735580507665873, + 0.012819705531001091, + 0.01341647282242775, + 0.008301869966089725, + 0.00968877226114273, + 0.007221544627100229, + 0.008384205400943756, + 0.01744903065264225, + 0.005318404175341129, + 0.0064477319829165936, + 0.02154613845050335, + 0.016771966591477394, + 0.03365325182676315, + 0.01137315109372139, + 0.014496980234980583, + 0.023371918126940727, + 0.012451154179871082, + 0.01828806847333908, + 0.0103075560182333, + 0.0030583820771425962, + 0.0032866790425032377, + 0.002729977248236537, + 0.005487932823598385, + 0.018170570954680443, + 0.0033025038428604603, + 0.005453833378851414, + 0.00973986741155386, + 0.00665097264572978, + 0.006441937293857336, + 0.005492238327860832, + 0.011419066227972507, + 0.008356817997992039, + 0.025168370455503464, + 0.013437367975711823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.04863978922367096 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_11/index_tensor([212, 212])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_11/index_tensor([212, 212])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aca1488fb4f3d111f0fb36e534e304b242d88679 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_11/index_tensor([212, 212])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Cinematic, high-speed macro footage, shot in ultra-slow motion, remains static in extreme close-up, perfectly aligned with the point of violent impact. A massive gauntlet, constructed of dark, riveted steel and thick, battle-worn leather, encases a powerfully clenched fist. The knuckles are fitted with heavy, tarnished brass dust collectors, their surface bearing the marks of countless altercations. As the fist collides with the rough, dark granite wall, the moment of impact is deconstructed with startling clarity. The brass gauntlet visibly compresses for a microsecond before releasing its force, causing a spiderweb of crystalline cracks to run across the stone surface. Simultaneously, a dazzling shower of white-hot sparks erupts from the friction, arcing gracefully through the air. A dense, billowing cloud of fine dust and grit erupts from the point of impact, its swirling particles caught in a single, harsh keylight, casting deep shadows and brightly illuminating the chaotic ballet of destruction. The atmosphere is tense and visceral, dominated by a cool, desaturated palette of dark stone and metal, with only brief, fiery sparks of raw power creating a surreal and awe-inspiring spectacle.", + "seed": 16283, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_11/index_212_seed_16283_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21812157109379768, + "vlm_logprob": -0.21812157109379768, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.21812157109379768, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11873720586299896 + ], + "opposite_logprob_tokens": [ + -2.24373722076416 + ], + "logprob_reward": -0.11873720586299896, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11873720586299896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.080856092274189, + 0.06155814230442047, + 0.024579636752605438, + 0.6292703151702881, + 0.01950550079345703, + 0.024318832904100418, + 0.025070445612072945, + 0.041190069168806076, + 0.08043759316205978, + 0.01321336254477501 + ], + "credit_map": [ + 0.0066841160878539085, + 0.004675817210227251, + 0.008995860815048218, + 0.006714488845318556, + 0.007741132751107216, + 0.017909806221723557, + 0.008506049402058125, + 0.019628815352916718, + 0.005296461284160614, + 0.005019515287131071, + 0.006990335416048765, + 0.006697431206703186, + 0.006586863659322262, + 0.015539981424808502, + 0.00884724035859108, + 0.0065803127363324165, + 0.0030061022844165564, + 0.0014507588930428028, + 0.0018587057711556554, + 0.0021248969715088606, + 0.0031371652148663998, + 0.0064671519212424755, + 0.0031559430062770844, + 0.003378910245373845, + 0.03368823230266571, + 0.008353388868272305, + 0.013257171958684921, + 0.009279844351112843, + 0.009341954253613949, + 0.2875123620033264, + 0.00968203041702509, + 0.25815534591674805, + 0.0033294151071459055, + 0.0017839637584984303, + 0.0017354023875668645, + 0.0018838170217350125, + 0.00230729510076344, + 0.0026195331010967493, + 0.002828140277415514, + 0.003017934039235115, + 0.003602582961320877, + 0.004366517066955566, + 0.0026282831095159054, + 0.002043981570750475, + 0.0030218418687582016, + 0.0024984823539853096, + 0.0032025454565882683, + 0.0029545992147177458, + 0.0023788081016391516, + 0.0033223209902644157, + 0.002749023260548711, + 0.006310387048870325, + 0.002518666675314307, + 0.0022107968106865883, + 0.0023441037628799677, + 0.0032363389618694782, + 0.004795250482857227, + 0.005862408317625523, + 0.0032635037787258625, + 0.007395503576844931, + 0.00455488171428442, + 0.004449540749192238, + 0.0051400549709796906, + 0.005728927440941334, + 0.009076551534235477, + 0.013703831471502781, + 0.010502767749130726, + 0.011812214739620686, + 0.007436268962919712, + 0.009885807521641254, + 0.006602359935641289, + 0.011417794041335583, + 0.0015923044411465526, + 0.0009719087393023074, + 0.002560157561674714, + 0.0017750890692695975, + 0.0012932195095345378, + 0.0016072146827355027, + 0.001624791999347508, + 0.0017886767163872719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.11873720586299896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07011306285858154 + ], + "opposite_logprob_tokens": [ + -2.695113182067871 + ], + "logprob_reward": -0.07011306285858154, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07011306285858154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25273576378822327, + 0.08897428214550018, + 0.032517120242118835, + 0.40584027767181396, + 0.026926245540380478, + 0.043792933225631714, + 0.018488513305783272, + 0.06617787480354309, + 0.03295142948627472, + 0.031595632433891296 + ], + "credit_map": [ + 0.013462395407259464, + 0.017086368054151535, + 0.013626372441649437, + 0.01568114571273327, + 0.023504050448536873, + 0.06912948936223984, + 0.026246165856719017, + 0.07399976998567581, + 0.006683931220322847, + 0.0034675435163080692, + 0.009096398949623108, + 0.004695649724453688, + 0.004410123452544212, + 0.02913626655936241, + 0.00954359769821167, + 0.021940777078270912, + 0.002906583482399583, + 0.0019981474615633488, + 0.002560712629929185, + 0.0017370260320603848, + 0.002457273658365011, + 0.010119285434484482, + 0.0035680956207215786, + 0.007169999182224274, + 0.022511247545480728, + 0.0057191005907952785, + 0.00876015517860651, + 0.00696156732738018, + 0.007223998662084341, + 0.182825967669487, + 0.006523824762552977, + 0.1653144210577011, + 0.004668934270739555, + 0.0022450427059084177, + 0.0029661741573363543, + 0.002477418864145875, + 0.0038141091354191303, + 0.003089110367000103, + 0.0037888705264776945, + 0.0038765850476920605, + 0.005566197447478771, + 0.006495361216366291, + 0.004551331512629986, + 0.005058293230831623, + 0.006079344544559717, + 0.004922837484627962, + 0.005978716537356377, + 0.00514084892347455, + 0.0020711026154458523, + 0.0023123929277062416, + 0.0017710161628201604, + 0.002318732440471649, + 0.0026896942872554064, + 0.0022830793168395758, + 0.0024396509397774935, + 0.0026028442662209272, + 0.006927073933184147, + 0.005399236921221018, + 0.005502886604517698, + 0.010396886616945267, + 0.008031425066292286, + 0.00887844618409872, + 0.011878998018801212, + 0.009162917733192444, + 0.004136284347623587, + 0.004979446064680815, + 0.003420136636123061, + 0.004578437190502882, + 0.0037704878486692905, + 0.004493864718824625, + 0.0031064427457749844, + 0.004466329701244831, + 0.0037726801820099354, + 0.0029614479281008244, + 0.004055657424032688, + 0.003773061092942953, + 0.0033391453325748444, + 0.004012883175164461, + 0.005455353297293186, + 0.0042254021391272545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.07011306285858154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262050151824951 + ], + "opposite_logprob_tokens": [ + -0.5762050151824951 + ], + "logprob_reward": -0.8262050151824951, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262050151824951, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.323382169008255, + 0.05695898458361626, + 0.05033063143491745, + 0.09529301524162292, + 0.03615669906139374, + 0.0580359622836113, + 0.03149143606424332, + 0.1579495221376419, + 0.13377489149570465, + 0.05662666633725166 + ], + "credit_map": [ + 0.021377919241786003, + 0.009681429713964462, + 0.020151007920503616, + 0.010174565017223358, + 0.013166521675884724, + 0.11950762569904327, + 0.02593364380300045, + 0.10338947176933289, + 0.006108197383582592, + 0.003806409193202853, + 0.00701405992731452, + 0.004323720000684261, + 0.004906887654215097, + 0.014257669448852539, + 0.006896935403347015, + 0.00964510627090931, + 0.003356596687808633, + 0.002620062790811062, + 0.002655520336702466, + 0.0025281894486397505, + 0.003522764192894101, + 0.018336378037929535, + 0.005772311706095934, + 0.011538808234035969, + 0.006014284212142229, + 0.001925528747960925, + 0.002304995432496071, + 0.002224830212071538, + 0.0034631751477718353, + 0.03946608304977417, + 0.0029258914291858673, + 0.03696822375059128, + 0.005328298546373844, + 0.0036011647898703814, + 0.003055840963497758, + 0.0026011483278125525, + 0.006263891234993935, + 0.005728908348828554, + 0.003936585038900375, + 0.0056408606469631195, + 0.008309750817716122, + 0.007094980217516422, + 0.008179355412721634, + 0.007878722622990608, + 0.0071321940049529076, + 0.0058377329260110855, + 0.006358950864523649, + 0.007244272157549858, + 0.0032732482068240643, + 0.005362158641219139, + 0.002121425699442625, + 0.004920199979096651, + 0.004558376967906952, + 0.0032847104594111443, + 0.003839778481051326, + 0.004131540190428495, + 0.017182858660817146, + 0.021233636885881424, + 0.01040171179920435, + 0.028242938220500946, + 0.019958006218075752, + 0.016858072951436043, + 0.022077757865190506, + 0.02199454791843891, + 0.013935744762420654, + 0.022217396646738052, + 0.01707913726568222, + 0.01903592236340046, + 0.01480554137378931, + 0.01685967482626438, + 0.010631988756358624, + 0.0192094836384058, + 0.004400273784995079, + 0.0041537038050591946, + 0.006063913926482201, + 0.00597061263397336, + 0.005551328882575035, + 0.007333524990826845, + 0.015607061795890331, + 0.007546244654804468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.8262050151824951 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058732882142067 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058732882142067, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058732882142067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04663455858826637, + 0.10913330316543579, + 0.05594198778271675, + 0.4206671118736267, + 0.047371622174978256, + 0.05728946998715401, + 0.05191028490662575, + 0.08548711985349655, + 0.08276962488889694, + 0.04279494658112526 + ], + "credit_map": [ + 0.007118845824152231, + 0.00558603648096323, + 0.008744762279093266, + 0.005303012672811747, + 0.004754277411848307, + 0.005450863391160965, + 0.005348789505660534, + 0.00432797335088253, + 0.00898459181189537, + 0.004263776820152998, + 0.00954105332493782, + 0.005530589260160923, + 0.006381923798471689, + 0.03957317769527435, + 0.012896447442471981, + 0.021961744874715805, + 0.004916021134704351, + 0.002386578358709812, + 0.003590472973883152, + 0.0027724860701709986, + 0.0031551725696772337, + 0.020635487511754036, + 0.007158080581575632, + 0.011327691376209259, + 0.022764073684811592, + 0.0054785446263849735, + 0.008484672755002975, + 0.0063667516224086285, + 0.007351835258305073, + 0.19179484248161316, + 0.006549317389726639, + 0.17187707126140594, + 0.006754604168236256, + 0.00489649036899209, + 0.004512814339250326, + 0.004130855202674866, + 0.005432276055216789, + 0.008948860689997673, + 0.005702919326722622, + 0.006992802023887634, + 0.005846921354532242, + 0.0066235302947461605, + 0.005017492454499006, + 0.0050609833560884, + 0.010793634690344334, + 0.0074292016215622425, + 0.009461404755711555, + 0.007056304253637791, + 0.005739310756325722, + 0.005425536073744297, + 0.004911895841360092, + 0.008633213117718697, + 0.006222986616194248, + 0.005812866147607565, + 0.008220971561968327, + 0.006943505257368088, + 0.00878768227994442, + 0.007830874063074589, + 0.00917951762676239, + 0.015203715302050114, + 0.009561944752931595, + 0.011456087231636047, + 0.011084750294685364, + 0.01238254178315401, + 0.010244613513350487, + 0.016501130536198616, + 0.008643447421491146, + 0.011755470186471939, + 0.007723815739154816, + 0.010023617185652256, + 0.006523365154862404, + 0.011354165151715279, + 0.004417601507157087, + 0.0036017715465277433, + 0.006118809338659048, + 0.00564249511808157, + 0.004925803747028112, + 0.0060539995320141315, + 0.005531757138669491, + 0.006502707488834858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.02058732882142067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965242743492126 + ], + "opposite_logprob_tokens": [ + -2.9299652576446533 + ], + "logprob_reward": -0.054965242743492126, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965242743492126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03416535630822182, + 0.1405591517686844, + 0.01585785113275051, + 0.6376708149909973, + 0.018761521205306053, + 0.010917959734797478, + 0.029177673161029816, + 0.029689667746424675, + 0.0660853311419487, + 0.01711469329893589 + ], + "credit_map": [ + 0.0038581108674407005, + 0.0023991416674107313, + 0.004238959401845932, + 0.003102228045463562, + 0.0032628958579152822, + 0.005608230363577604, + 0.0038159526884555817, + 0.007879836484789848, + 0.008519521914422512, + 0.006695201620459557, + 0.010812998749315739, + 0.005981796886771917, + 0.005160033702850342, + 0.05553736537694931, + 0.01670362800359726, + 0.031148601323366165, + 0.002543399343267083, + 0.001440937165170908, + 0.003346950514242053, + 0.001950094010680914, + 0.001350678619928658, + 0.001923048635944724, + 0.0018117574509233236, + 0.0014909851597622037, + 0.03394652158021927, + 0.008506235666573048, + 0.013200453482568264, + 0.009864688850939274, + 0.009856809861958027, + 0.29046180844306946, + 0.009977791458368301, + 0.26185643672943115, + 0.0028064700309187174, + 0.0018294693436473608, + 0.0025684491265565157, + 0.0017262864857912064, + 0.002170737599954009, + 0.0028168517164885998, + 0.0017890810267999768, + 0.0030541750602424145, + 0.001015310175716877, + 0.0012275722110643983, + 0.0012267037527635694, + 0.0013935562456026673, + 0.001801850856281817, + 0.0012716044439002872, + 0.0016848427476361394, + 0.0012965183705091476, + 0.002553333528339863, + 0.0037660005036741495, + 0.0028898660093545914, + 0.006164008751511574, + 0.0044549982994794846, + 0.0029198580887168646, + 0.0029601198621094227, + 0.0034694846253842115, + 0.003934259992092848, + 0.003718805033713579, + 0.004679768346250057, + 0.005701955407857895, + 0.002575877821072936, + 0.0029124123975634575, + 0.0022064719814807177, + 0.0039601135067641735, + 0.007301583886146545, + 0.012951595708727837, + 0.007144042756408453, + 0.0098187904804945, + 0.006582347676157951, + 0.008431539870798588, + 0.005137844942510128, + 0.008717586286365986, + 0.0023820858914405107, + 0.0016376961721107364, + 0.002235913183540106, + 0.0020505026914179325, + 0.0013831232208758593, + 0.002552059246227145, + 0.002301172586157918, + 0.002572139725089073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.054965242743492126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "Cinematic, high-speed macro footage, shot in ultra-slow motion, remains static in extreme close-up, perfectly aligned with the point of violent impact. A massive gauntlet, constructed of dark, riveted steel and thick, battle-worn leather, encases a powerfully clenched fist. The knuckles are fitted with heavy, tarnished brass dust collectors, their surface bearing the marks of countless altercations. As the fist collides with the rough, dark granite wall, the moment of impact is deconstructed with startling clarity. The brass gauntlet visibly compresses for a microsecond before releasing its force, causing a spiderweb of crystalline cracks to run across the stone surface. Simultaneously, a dazzling shower of white-hot sparks erupts from the friction, arcing gracefully through the air. A dense, billowing cloud of fine dust and grit erupts from the point of impact, its swirling particles caught in a single, harsh keylight, casting deep shadows and brightly illuminating the chaotic ballet of destruction. The atmosphere is tense and visceral, dominated by a cool, desaturated palette of dark stone and metal, with only brief, fiery sparks of raw power creating a surreal and awe-inspiring spectacle.", + "seed": 16284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_11/index_212_seed_16284_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1811800990253687, + "vlm_logprob": -0.1811800990253687, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.1811800990253687, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11958258599042892 + ], + "opposite_logprob_tokens": [ + -2.2445826530456543 + ], + "logprob_reward": -0.11958258599042892, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11958258599042892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14106647670269012, + 0.30356907844543457, + 0.06558041274547577, + 0.053421709686517715, + 0.03837142139673233, + 0.056904349476099014, + 0.03096620738506317, + 0.019392352551221848, + 0.12960197031497955, + 0.16112597286701202 + ], + "credit_map": [ + 0.014502551406621933, + 0.011841853149235249, + 0.02148938551545143, + 0.017449401319026947, + 0.011220596730709076, + 0.024350322782993317, + 0.019979534670710564, + 0.020232824608683586, + 0.011150559410452843, + 0.0041831363923847675, + 0.013412881642580032, + 0.008524717763066292, + 0.008988246321678162, + 0.11988591402769089, + 0.017028626054525375, + 0.1203950047492981, + 0.00775110675022006, + 0.0037018293514847755, + 0.012834941036999226, + 0.007905947975814342, + 0.004710678942501545, + 0.009822056628763676, + 0.007963590323925018, + 0.010890263132750988, + 0.006107722409069538, + 0.00295760459266603, + 0.008233003318309784, + 0.004566813353449106, + 0.003303463803604245, + 0.011917171068489552, + 0.0038202530704438686, + 0.012515680864453316, + 0.004288268741220236, + 0.00554866436868906, + 0.005425006151199341, + 0.005350457038730383, + 0.00356510141864419, + 0.004531915299594402, + 0.005073491483926773, + 0.004588516429066658, + 0.005572954658418894, + 0.006066572852432728, + 0.008023522794246674, + 0.003908822312951088, + 0.008240445517003536, + 0.00812717154622078, + 0.0077176159247756, + 0.009247244335711002, + 0.004326805006712675, + 0.00317183299921453, + 0.004249730613082647, + 0.0030320114456117153, + 0.003453335724771023, + 0.0038841343484818935, + 0.004075279459357262, + 0.00477307615801692, + 0.0018317141802981496, + 0.0033253428991883993, + 0.0017835145117715001, + 0.0037366636097431183, + 0.002109109191223979, + 0.002448529936373234, + 0.0021234622690826654, + 0.0020340164192020893, + 0.01078763697296381, + 0.013409950770437717, + 0.023173213005065918, + 0.02995758317410946, + 0.009297631680965424, + 0.017529098317027092, + 0.015624677762389183, + 0.009822186082601547, + 0.016189832240343094, + 0.018485670909285545, + 0.021204126998782158, + 0.03211572766304016, + 0.016396336257457733, + 0.020770443603396416, + 0.021976588293910027, + 0.013987252488732338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.11958258599042892 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043116722255945206 + ], + "opposite_logprob_tokens": [ + -3.168116807937622 + ], + "logprob_reward": -0.043116722255945206, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043116722255945206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1384400874376297, + 0.22302554547786713, + 0.07467154413461685, + 0.08001010864973068, + 0.016217151656746864, + 0.027997206896543503, + 0.054025158286094666, + 0.1798323392868042, + 0.12194565683603287, + 0.08383513242006302 + ], + "credit_map": [ + 0.012184947729110718, + 0.01205759309232235, + 0.019600339233875275, + 0.021100180223584175, + 0.009449848905205727, + 0.027295785024762154, + 0.015769895166158676, + 0.02098149247467518, + 0.006867174059152603, + 0.003649970283731818, + 0.00722909951582551, + 0.004834574181586504, + 0.006934404373168945, + 0.08884686231613159, + 0.013127611018717289, + 0.09153585135936737, + 0.010119298473000526, + 0.00526869110763073, + 0.01279165968298912, + 0.007612800691276789, + 0.006322680506855249, + 0.00956692174077034, + 0.010918017476797104, + 0.012071480043232441, + 0.008756028488278389, + 0.008408865891397, + 0.012082384899258614, + 0.009298671968281269, + 0.00711815943941474, + 0.01397857815027237, + 0.008388478308916092, + 0.011978939175605774, + 0.001995404250919819, + 0.0013215370709076524, + 0.00207506213337183, + 0.0019135333131998777, + 0.002055822405964136, + 0.0021465381141752005, + 0.002296395367011428, + 0.00241285958327353, + 0.0023299423046410084, + 0.0025746463797986507, + 0.004035260062664747, + 0.005611727945506573, + 0.003199918195605278, + 0.0031170176807790995, + 0.0028541686479002237, + 0.004274526610970497, + 0.005096136592328548, + 0.005825636442750692, + 0.004749832674860954, + 0.005650828592479229, + 0.007046863436698914, + 0.007361589930951595, + 0.0075895157642662525, + 0.010704757645726204, + 0.018295960500836372, + 0.02960795722901821, + 0.015045180916786194, + 0.03727210313081741, + 0.01745443418622017, + 0.023393524810671806, + 0.02345128543674946, + 0.015311903320252895, + 0.010820374824106693, + 0.013371454551815987, + 0.01907014101743698, + 0.030768783763051033, + 0.009701558388769627, + 0.016679806634783745, + 0.012194942682981491, + 0.00933859869837761, + 0.008215588517487049, + 0.008154536597430706, + 0.010409441776573658, + 0.01680099032819271, + 0.011555218137800694, + 0.01001706626266241, + 0.011784407310187817, + 0.00689788069576025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.043116722255945206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934189200401306 + ], + "opposite_logprob_tokens": [ + -0.6934189200401306 + ], + "logprob_reward": -0.6934189200401306, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934189200401306, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11633430421352386, + 0.1742352992296219, + 0.035223547369241714, + 0.06295116990804672, + 0.026322683319449425, + 0.026509473100304604, + 0.023236829787492752, + 0.1464024782180786, + 0.2862515151500702, + 0.1025327816605568 + ], + "credit_map": [ + 0.00984996184706688, + 0.01190516259521246, + 0.018204573541879654, + 0.020274529233574867, + 0.009828954003751278, + 0.016006257385015488, + 0.012929455377161503, + 0.017335403710603714, + 0.005309004336595535, + 0.002942651743069291, + 0.006066576112061739, + 0.0041774422861635685, + 0.006206294987350702, + 0.06742111593484879, + 0.010717812925577164, + 0.07139439135789871, + 0.003735943930223584, + 0.0028386912308633327, + 0.005228525493294001, + 0.0052062696777284145, + 0.003716768929734826, + 0.003412823658436537, + 0.005617185495793819, + 0.005467336159199476, + 0.005442644935101271, + 0.004291537217795849, + 0.005794834345579147, + 0.007716592866927385, + 0.0049509392119944096, + 0.015081427060067654, + 0.004727643448859453, + 0.014945545233786106, + 0.003121600253507495, + 0.003213755087926984, + 0.0034310074988752604, + 0.0032905314583331347, + 0.0031912834383547306, + 0.003830828471109271, + 0.0032469569705426693, + 0.0029967178124934435, + 0.003204147331416607, + 0.0027214123401790857, + 0.004147358238697052, + 0.0026239955332130194, + 0.003044197103008628, + 0.003890750464051962, + 0.002901565283536911, + 0.003976043779402971, + 0.0026557256933301687, + 0.0032469993457198143, + 0.0025179563090205193, + 0.0025897708255797625, + 0.0022663273848593235, + 0.0028642325196415186, + 0.0029280518647283316, + 0.0041677639819681644, + 0.014316439628601074, + 0.024421190842986107, + 0.012297399342060089, + 0.031213615089654922, + 0.013194501399993896, + 0.019448306411504745, + 0.018380502238869667, + 0.013130506500601768, + 0.02594185620546341, + 0.03143998607993126, + 0.04904600605368614, + 0.06573733687400818, + 0.019474122673273087, + 0.03916347026824951, + 0.032112523913383484, + 0.023336173966526985, + 0.008403530344367027, + 0.010230950079858303, + 0.01690678298473358, + 0.018084069713950157, + 0.012017415836453438, + 0.01489593181759119, + 0.012261198833584785, + 0.009732883423566818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.6934189200401306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607552357017994 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.01607552357017994, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607552357017994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05176997557282448, + 0.3518672585487366, + 0.02212110534310341, + 0.012643169611692429, + 0.02229742333292961, + 0.09808239340782166, + 0.04232247546315193, + 0.26719778776168823, + 0.0740860104560852, + 0.05761242285370827 + ], + "credit_map": [ + 0.004374436102807522, + 0.004594236146658659, + 0.008486645296216011, + 0.008764112368226051, + 0.0029475209303200245, + 0.009184825234115124, + 0.003869728185236454, + 0.009548468515276909, + 0.00997643731534481, + 0.0044253915548324585, + 0.008936782367527485, + 0.004902505315840244, + 0.010601182468235493, + 0.1445065587759018, + 0.01959296129643917, + 0.14892543852329254, + 0.0027893390506505966, + 0.0011704149655997753, + 0.00544118694961071, + 0.002488247584551573, + 0.0016027549281716347, + 0.0028930746484547853, + 0.002541658002883196, + 0.0031944280490279198, + 0.0017398296622559428, + 0.0010629297466948628, + 0.0025634148623794317, + 0.0014995824312791228, + 0.0008755761082284153, + 0.0018787635490298271, + 0.0009358508395962417, + 0.0020872221793979406, + 0.0019296613754704595, + 0.0016756899422034621, + 0.003353761974722147, + 0.0022160469088703394, + 0.004028976429253817, + 0.002480674535036087, + 0.0026544034481048584, + 0.003958209417760372, + 0.00881081074476242, + 0.010410530492663383, + 0.014139179140329361, + 0.010383851826190948, + 0.012190357781946659, + 0.014011400751769543, + 0.011714808642864227, + 0.016421448439359665, + 0.005337619222700596, + 0.003535655327141285, + 0.004057306330651045, + 0.002952143084257841, + 0.004105955362319946, + 0.006847499869763851, + 0.00524499174207449, + 0.010241304524242878, + 0.02768510766327381, + 0.042293254286050797, + 0.023488769307732582, + 0.05210646986961365, + 0.027621915563941002, + 0.03308425098657608, + 0.03714212775230408, + 0.02377590350806713, + 0.006599258165806532, + 0.006152637302875519, + 0.007755621336400509, + 0.01632543094456196, + 0.008424852974712849, + 0.009780360385775566, + 0.011133801192045212, + 0.007914043962955475, + 0.006703219376504421, + 0.006154550705105066, + 0.006790558807551861, + 0.011127491481602192, + 0.0069992998614907265, + 0.0077785346657037735, + 0.006404594052582979, + 0.005654172506183386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.01607552357017994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370674327015877 + ], + "opposite_logprob_tokens": [ + -3.4087066650390625 + ], + "logprob_reward": -0.03370674327015877, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370674327015877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16054300963878632, + 0.038945719599723816, + 0.027972958981990814, + 0.0221762303262949, + 0.10333247482776642, + 0.07299206405878067, + 0.10324078798294067, + 0.2031836211681366, + 0.19148367643356323, + 0.07612938433885574 + ], + "credit_map": [ + 0.014384589157998562, + 0.010653652250766754, + 0.02111423760652542, + 0.031864602118730545, + 0.00868003349751234, + 0.022497102618217468, + 0.017156362533569336, + 0.034192442893981934, + 0.0025036362931132317, + 0.0016515875468030572, + 0.004241950809955597, + 0.002191707259044051, + 0.0017401134828105569, + 0.013295878656208515, + 0.0036991185043007135, + 0.00962173193693161, + 0.004156715236604214, + 0.001863184035755694, + 0.005282036028802395, + 0.0028388022910803556, + 0.001643268740735948, + 0.004584872629493475, + 0.0026979465037584305, + 0.004906134679913521, + 0.002440114039927721, + 0.0017256297869607806, + 0.0022256101947277784, + 0.0028810095973312855, + 0.002225753152742982, + 0.0037454983685165644, + 0.002656412310898304, + 0.004276204854249954, + 0.00970496330410242, + 0.008238487876951694, + 0.015972187742590904, + 0.00843360647559166, + 0.015754949301481247, + 0.01437345054000616, + 0.013820271007716656, + 0.017034562304615974, + 0.005961972288787365, + 0.005542273633182049, + 0.0072115384973585606, + 0.0036023438442498446, + 0.01032442320138216, + 0.01453371811658144, + 0.008965881541371346, + 0.016849927604198456, + 0.012408454902470112, + 0.013236615806818008, + 0.009741789661347866, + 0.009014817886054516, + 0.010781041346490383, + 0.012461643666028976, + 0.016308460384607315, + 0.019287973642349243, + 0.019832542166113853, + 0.03281579911708832, + 0.020373636856675148, + 0.03845649212598801, + 0.0211168359965086, + 0.025161270052194595, + 0.025424789637327194, + 0.02000225894153118, + 0.021212760359048843, + 0.0167120099067688, + 0.02606000006198883, + 0.042653631418943405, + 0.01982988603413105, + 0.02826542779803276, + 0.019173843786120415, + 0.017576133832335472, + 0.005552005488425493, + 0.00810131337493658, + 0.01040232926607132, + 0.016523350030183792, + 0.006651683244854212, + 0.010943528264760971, + 0.010206654667854309, + 0.007748528849333525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.03370674327015877 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "Cinematic, high-speed macro footage, shot in ultra-slow motion, remains static in extreme close-up, perfectly aligned with the point of violent impact. A massive gauntlet, constructed of dark, riveted steel and thick, battle-worn leather, encases a powerfully clenched fist. The knuckles are fitted with heavy, tarnished brass dust collectors, their surface bearing the marks of countless altercations. As the fist collides with the rough, dark granite wall, the moment of impact is deconstructed with startling clarity. The brass gauntlet visibly compresses for a microsecond before releasing its force, causing a spiderweb of crystalline cracks to run across the stone surface. Simultaneously, a dazzling shower of white-hot sparks erupts from the friction, arcing gracefully through the air. A dense, billowing cloud of fine dust and grit erupts from the point of impact, its swirling particles caught in a single, harsh keylight, casting deep shadows and brightly illuminating the chaotic ballet of destruction. The atmosphere is tense and visceral, dominated by a cool, desaturated palette of dark stone and metal, with only brief, fiery sparks of raw power creating a surreal and awe-inspiring spectacle.", + "seed": 16285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_11/index_212_seed_16285_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20388015136122703, + "vlm_logprob": -0.20388015136122703, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.20388015136122703, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08565700799226761 + ], + "opposite_logprob_tokens": [ + -2.5856571197509766 + ], + "logprob_reward": -0.08565700799226761, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08565700799226761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08295752853155136, + 0.4589410126209259, + 0.025549015030264854, + 0.041690949350595474, + 0.042751673609018326, + 0.024575605988502502, + 0.05913173407316208, + 0.03907819464802742, + 0.17956796288490295, + 0.04575631022453308 + ], + "credit_map": [ + 0.00813282374292612, + 0.006929093971848488, + 0.010714900679886341, + 0.007632977794855833, + 0.009352474473416805, + 0.010771731846034527, + 0.014676379039883614, + 0.014747160486876965, + 0.024196557700634003, + 0.015519754961133003, + 0.015287132933735847, + 0.015750283375382423, + 0.020772123709321022, + 0.16831475496292114, + 0.02710605412721634, + 0.17199435830116272, + 0.004168809857219458, + 0.00424570357427001, + 0.003203079104423523, + 0.0023758357856422663, + 0.0034484181087464094, + 0.003187691792845726, + 0.0020148667972534895, + 0.002904611872509122, + 0.005738322157412767, + 0.003959968686103821, + 0.0056508989073336124, + 0.004387252498418093, + 0.006501621566712856, + 0.006395015865564346, + 0.0034698834642767906, + 0.0055879889987409115, + 0.004691916052252054, + 0.006028228905051947, + 0.004600811284035444, + 0.005243017803877592, + 0.004902261309325695, + 0.005568506196141243, + 0.005705593153834343, + 0.0060113403014838696, + 0.002843211404979229, + 0.0033548143692314625, + 0.002656572265550494, + 0.00394293200224638, + 0.0026986224111169577, + 0.00297673256136477, + 0.002676511649042368, + 0.0034262125845998526, + 0.008888473734259605, + 0.00578446127474308, + 0.00565074710175395, + 0.006690914276987314, + 0.00438983179628849, + 0.010045374743640423, + 0.008283798582851887, + 0.009398137219250202, + 0.008593643084168434, + 0.005999895744025707, + 0.0032569810282438993, + 0.003107023425400257, + 0.0032952961046248674, + 0.005827473476529121, + 0.002913673873990774, + 0.006084212567657232, + 0.014726509340107441, + 0.018366007134318352, + 0.03180475905537605, + 0.014498667791485786, + 0.01088110264390707, + 0.024128321558237076, + 0.01829247921705246, + 0.04687012732028961, + 0.004105250351130962, + 0.0067693921737372875, + 0.004957116208970547, + 0.00683202687650919, + 0.0036985166370868683, + 0.00573412049561739, + 0.007638133596628904, + 0.00602175947278738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.08565700799226761 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813252970576286 + ], + "opposite_logprob_tokens": [ + -3.288132429122925 + ], + "logprob_reward": -0.03813252970576286, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813252970576286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0603005588054657, + 0.6397539973258972, + 0.030132822692394257, + 0.05043124035000801, + 0.025587253272533417, + 0.030144404619932175, + 0.07262851297855377, + 0.025089081376791, + 0.03625479340553284, + 0.029677310958504677 + ], + "credit_map": [ + 0.008664602413773537, + 0.005257094744592905, + 0.00672939233481884, + 0.005867701023817062, + 0.005647714249789715, + 0.007888260297477245, + 0.008247052319347858, + 0.011998741887509823, + 0.014538172632455826, + 0.009875104762613773, + 0.009844393469393253, + 0.01402299478650093, + 0.012352519668638706, + 0.2785521149635315, + 0.018309002742171288, + 0.2822597026824951, + 0.0032449306454509497, + 0.004478149581700563, + 0.002779844682663679, + 0.0029305790085345507, + 0.004187286365777254, + 0.004439826123416424, + 0.003637012792751193, + 0.004435193724930286, + 0.006953331641852856, + 0.006326750386506319, + 0.006334581412374973, + 0.005615280009806156, + 0.00765587342903018, + 0.006837338209152222, + 0.004383498337119818, + 0.006324589252471924, + 0.0029771998524665833, + 0.0033592411782592535, + 0.0026123528368771076, + 0.0031210388988256454, + 0.0030514386016875505, + 0.003173672826960683, + 0.0038748111110180616, + 0.003417498432099819, + 0.004263204988092184, + 0.0032173688523471355, + 0.0030900000128895044, + 0.002580797765403986, + 0.003765380010008812, + 0.004490593448281288, + 0.0035117925144732, + 0.005225266795605421, + 0.006357045378535986, + 0.0039915116503834724, + 0.004788669291883707, + 0.00625478895381093, + 0.00411228509619832, + 0.01993689127266407, + 0.008542168885469437, + 0.018645158037543297, + 0.004669283516705036, + 0.004199502523988485, + 0.0020104774739593267, + 0.0021564136259257793, + 0.0024645000230520964, + 0.0035557702649384737, + 0.0025348104536533356, + 0.0034983244258910418, + 0.0035802568309009075, + 0.0047791278921067715, + 0.006199650000780821, + 0.0030179114546626806, + 0.0024690411519259214, + 0.004550221376121044, + 0.0030683765653520823, + 0.008590209297835827, + 0.002352488227188587, + 0.003214901313185692, + 0.0032907831482589245, + 0.006379330065101385, + 0.0021859121043235064, + 0.003390462836250663, + 0.004690499510616064, + 0.00417293468490243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.03813252970576286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261933326721191 + ], + "opposite_logprob_tokens": [ + -0.5761933326721191 + ], + "logprob_reward": -0.8261933326721191, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261933326721191, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12803803384304047, + 0.42900899052619934, + 0.06433703750371933, + 0.0634351298213005, + 0.07369013875722885, + 0.05766456946730614, + 0.033966004848480225, + 0.028808623552322388, + 0.07453297823667526, + 0.0465184710919857 + ], + "credit_map": [ + 0.012968776747584343, + 0.009645613841712475, + 0.016756199300289154, + 0.012279791757464409, + 0.014996051788330078, + 0.016496632248163223, + 0.01869370974600315, + 0.026201261207461357, + 0.02242763713002205, + 0.014893733896315098, + 0.014124387875199318, + 0.017492614686489105, + 0.01724771223962307, + 0.15788576006889343, + 0.027385111898183823, + 0.15755201876163483, + 0.007936075329780579, + 0.01151304692029953, + 0.005789199844002724, + 0.007434047758579254, + 0.010325614362955093, + 0.007321257144212723, + 0.006832306273281574, + 0.007185487542301416, + 0.008921048603951931, + 0.006546776741743088, + 0.00723819388076663, + 0.0071209692396223545, + 0.010693524032831192, + 0.008770110085606575, + 0.006034814286977053, + 0.008109692484140396, + 0.015033318661153316, + 0.009473126381635666, + 0.00812903605401516, + 0.008404993452131748, + 0.00706911226734519, + 0.00862002745270729, + 0.007449647411704063, + 0.009510873816907406, + 0.007841882295906544, + 0.008498198352754116, + 0.005818601697683334, + 0.005713135469704866, + 0.005376251880079508, + 0.007694198749959469, + 0.008063732646405697, + 0.008658569306135178, + 0.0034545492380857468, + 0.0019599846564233303, + 0.0036983806639909744, + 0.004393346142023802, + 0.0024791129399091005, + 0.00762215768918395, + 0.004680975805968046, + 0.005677497014403343, + 0.0030593043193221092, + 0.003727728733792901, + 0.002616302575916052, + 0.0029880560468882322, + 0.003434650832787156, + 0.0047262879088521, + 0.003896889742463827, + 0.004359402693808079, + 0.006337420549243689, + 0.00931218359619379, + 0.012401759624481201, + 0.010420513339340687, + 0.005422818474471569, + 0.009286398068070412, + 0.00709229800850153, + 0.014259586110711098, + 0.005628606770187616, + 0.006679171696305275, + 0.005242175422608852, + 0.007314287591725588, + 0.0032941072713583708, + 0.0057870326563715935, + 0.006129296496510506, + 0.006443793419748545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.8261933326721191 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635037526488304 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635037526488304, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635037526488304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03888504579663277, + 0.16191908717155457, + 0.023418813943862915, + 0.13574400544166565, + 0.048246778547763824, + 0.016934413462877274, + 0.0170152485370636, + 0.09063972532749176, + 0.44337406754493713, + 0.023822762072086334 + ], + "credit_map": [ + 0.0053267101757228374, + 0.004538037348538637, + 0.003934870474040508, + 0.003340024035423994, + 0.005223877727985382, + 0.005432900041341782, + 0.004887227434664965, + 0.006201405078172684, + 0.0064173596911132336, + 0.004908150061964989, + 0.003977667074650526, + 0.0049265301786363125, + 0.005266416352242231, + 0.06394127011299133, + 0.0069200266152620316, + 0.0655616894364357, + 0.003375895321369171, + 0.004134363029152155, + 0.0024099836591631174, + 0.0023311139084398746, + 0.002561726840212941, + 0.0033357630018144846, + 0.0019709512125700712, + 0.003299017669633031, + 0.020567284896969795, + 0.017142102122306824, + 0.022466793656349182, + 0.014032957144081593, + 0.015805725008249283, + 0.018627142533659935, + 0.010487151332199574, + 0.01661485992372036, + 0.009600082412362099, + 0.0083095608279109, + 0.004623467102646828, + 0.004982492420822382, + 0.004843654111027718, + 0.005263988394290209, + 0.007082780823111534, + 0.003540752222761512, + 0.002331169554963708, + 0.0023394012823700905, + 0.0015629922272637486, + 0.002262145048007369, + 0.0017103967256844044, + 0.0023332317359745502, + 0.0017467305297031999, + 0.0026483472902327776, + 0.0019460200564935803, + 0.0010631335899233818, + 0.0010393349220976233, + 0.001691419049166143, + 0.0014595035463571548, + 0.004052495583891869, + 0.002281473483890295, + 0.003481869585812092, + 0.014303389005362988, + 0.01127325464040041, + 0.007939675822854042, + 0.010365994647145271, + 0.008450956083834171, + 0.014055708423256874, + 0.009456543251872063, + 0.014794218353927135, + 0.035075895488262177, + 0.03497028723359108, + 0.09615067392587662, + 0.029903991147875786, + 0.02428308129310608, + 0.06035066023468971, + 0.03846477344632149, + 0.12417474389076233, + 0.002369439462199807, + 0.004208074416965246, + 0.002474679844453931, + 0.004081027116626501, + 0.0015916326083242893, + 0.002709779189899564, + 0.003184419823810458, + 0.003203712636604905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.02635037526488304 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067511171102524 + ], + "opposite_logprob_tokens": [ + -3.168067455291748 + ], + "logprob_reward": -0.043067511171102524, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067511171102524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08553017675876617, + 0.189634308218956, + 0.035735100507736206, + 0.0999959260225296, + 0.04952200874686241, + 0.04246744513511658, + 0.04968573525547981, + 0.10190697759389877, + 0.19949844479560852, + 0.14602386951446533 + ], + "credit_map": [ + 0.013819973915815353, + 0.006620379630476236, + 0.00641076173633337, + 0.004729823675006628, + 0.01079207006841898, + 0.013953902758657932, + 0.008766050450503826, + 0.0204372126609087, + 0.006205244455486536, + 0.003919650334864855, + 0.0032430975697934628, + 0.004012021701782942, + 0.003650846891105175, + 0.08126144856214523, + 0.0039843120612204075, + 0.08335768431425095, + 0.004280077293515205, + 0.0050241765566170216, + 0.003081877250224352, + 0.00415373407304287, + 0.005303790792822838, + 0.00606224499642849, + 0.0034912258852273226, + 0.004337972030043602, + 0.007078035734593868, + 0.010903852060437202, + 0.009677499532699585, + 0.010127045214176178, + 0.024018732830882072, + 0.01242496445775032, + 0.012063232250511646, + 0.013702563010156155, + 0.010494678281247616, + 0.007198303006589413, + 0.005421148147433996, + 0.005818309262394905, + 0.0053311544470489025, + 0.005335766822099686, + 0.005701327696442604, + 0.004221322946250439, + 0.0064060864970088005, + 0.006652459502220154, + 0.003502058098092675, + 0.003692599479109049, + 0.0037795472890138626, + 0.006009306292980909, + 0.004334671888500452, + 0.008090714924037457, + 0.004465118981897831, + 0.002755995374172926, + 0.0030546702910214663, + 0.004542459733784199, + 0.0027404152788221836, + 0.014706841669976711, + 0.005938194692134857, + 0.011482035741209984, + 0.017106885090470314, + 0.012054326944053173, + 0.00843134056776762, + 0.01122868899255991, + 0.008866707794368267, + 0.01666036807000637, + 0.010352429002523422, + 0.017206234857439995, + 0.015982387587428093, + 0.015578274615108967, + 0.042500805109739304, + 0.017846887931227684, + 0.015165505930781364, + 0.026314113289117813, + 0.016184654086828232, + 0.04992580786347389, + 0.012968157418072224, + 0.015395707450807095, + 0.018089357763528824, + 0.03373177349567413, + 0.011643209494650364, + 0.014271099120378494, + 0.02261568233370781, + 0.017308877781033516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.043067511171102524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "Cinematic, high-speed macro footage, shot in ultra-slow motion, remains static in extreme close-up, perfectly aligned with the point of violent impact. A massive gauntlet, constructed of dark, riveted steel and thick, battle-worn leather, encases a powerfully clenched fist. The knuckles are fitted with heavy, tarnished brass dust collectors, their surface bearing the marks of countless altercations. As the fist collides with the rough, dark granite wall, the moment of impact is deconstructed with startling clarity. The brass gauntlet visibly compresses for a microsecond before releasing its force, causing a spiderweb of crystalline cracks to run across the stone surface. Simultaneously, a dazzling shower of white-hot sparks erupts from the friction, arcing gracefully through the air. A dense, billowing cloud of fine dust and grit erupts from the point of impact, its swirling particles caught in a single, harsh keylight, casting deep shadows and brightly illuminating the chaotic ballet of destruction. The atmosphere is tense and visceral, dominated by a cool, desaturated palette of dark stone and metal, with only brief, fiery sparks of raw power creating a surreal and awe-inspiring spectacle.", + "seed": 16286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_11/index_212_seed_16286_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2617938555777073, + "vlm_logprob": -0.2617938555777073, + "vlm_yes_no_margin": 2.174999988079071, + "vlm_token_credit": -0.2617938555777073, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07465270906686783 + ], + "opposite_logprob_tokens": [ + -2.699652671813965 + ], + "logprob_reward": -0.07465270906686783, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07465270906686783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.215483620762825, + 0.10230327397584915, + 0.1733333319425583, + 0.05186684802174568, + 0.0389854870736599, + 0.09542851150035858, + 0.11592170596122742, + 0.0536443293094635, + 0.08337730914354324, + 0.06965560466051102 + ], + "credit_map": [ + 0.044196732342243195, + 0.030553743243217468, + 0.01632443629205227, + 0.018012428656220436, + 0.025491587817668915, + 0.02508300170302391, + 0.03126155212521553, + 0.024560140445828438, + 0.011020619422197342, + 0.009007089771330357, + 0.010319295339286327, + 0.00563611276447773, + 0.01073835976421833, + 0.02777130901813507, + 0.007087206933647394, + 0.02072327770292759, + 0.014518499374389648, + 0.007444398477673531, + 0.05384945496916771, + 0.010964114218950272, + 0.00753707904368639, + 0.0541394017636776, + 0.009952856227755547, + 0.014927518554031849, + 0.005435227882117033, + 0.0051569887436926365, + 0.004620583727955818, + 0.0042274328880012035, + 0.007401580922305584, + 0.010367912240326405, + 0.005379415582865477, + 0.00927770510315895, + 0.005647245794534683, + 0.0043205805122852325, + 0.002790976082906127, + 0.005627255886793137, + 0.006189445964992046, + 0.00519690802320838, + 0.003745423164218664, + 0.005467652343213558, + 0.01297362707555294, + 0.010294279083609581, + 0.01521748024970293, + 0.01147259958088398, + 0.011470554396510124, + 0.0130188362672925, + 0.009311568923294544, + 0.011669576168060303, + 0.02119390480220318, + 0.009026417508721352, + 0.01752609759569168, + 0.007179109379649162, + 0.009128419682383537, + 0.015792421996593475, + 0.009753680787980556, + 0.026321642100811005, + 0.005672386847436428, + 0.006204537581652403, + 0.004593957681208849, + 0.005839689169079065, + 0.0059981755912303925, + 0.006411662325263023, + 0.01236196979880333, + 0.0065619489178061485, + 0.013377360068261623, + 0.010466229170560837, + 0.014616290107369423, + 0.006936453282833099, + 0.008394883014261723, + 0.01083056628704071, + 0.010173222050070763, + 0.008582307025790215, + 0.009781824424862862, + 0.011055277660489082, + 0.010027890093624592, + 0.006649107206612825, + 0.006746073719114065, + 0.00815283041447401, + 0.00933022703975439, + 0.00791237410157919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.07465270906686783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373267874121666 + ], + "opposite_logprob_tokens": [ + -3.4087326526641846 + ], + "logprob_reward": -0.03373267874121666, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373267874121666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1985698640346527, + 0.15695181488990784, + 0.07785316556692123, + 0.11959835141897202, + 0.05762706696987152, + 0.0824577584862709, + 0.13775980472564697, + 0.06964165717363358, + 0.05802270025014877, + 0.04151780530810356 + ], + "credit_map": [ + 0.031597450375556946, + 0.020905273035168648, + 0.011852060444653034, + 0.01909114047884941, + 0.026663441210985184, + 0.02463747002184391, + 0.02255735732614994, + 0.04126567766070366, + 0.014364056289196014, + 0.014833054505288601, + 0.011922451667487621, + 0.015836184844374657, + 0.01759140007197857, + 0.03873012214899063, + 0.014853541739284992, + 0.028821000829339027, + 0.01155858114361763, + 0.004370759706944227, + 0.01266323123127222, + 0.005761235021054745, + 0.005363765172660351, + 0.023039544001221657, + 0.00552334962412715, + 0.009572695940732956, + 0.0038799364119768143, + 0.0033755828626453876, + 0.005566189531236887, + 0.004469876177608967, + 0.004156831651926041, + 0.04876597598195076, + 0.004457286559045315, + 0.044926680624485016, + 0.0073153357952833176, + 0.006379732862114906, + 0.004501677583903074, + 0.006244870834052563, + 0.010442445054650307, + 0.0076998090371489525, + 0.007937559857964516, + 0.007105640135705471, + 0.010768537409603596, + 0.006866636220365763, + 0.01070773508399725, + 0.006795702967792749, + 0.009554845280945301, + 0.011879842728376389, + 0.01191755197942257, + 0.013966912403702736, + 0.02512197755277157, + 0.009892826899886131, + 0.019979458302259445, + 0.008975926786661148, + 0.00857371836900711, + 0.01857147552073002, + 0.014592853374779224, + 0.03205156698822975, + 0.012182125821709633, + 0.007514779455959797, + 0.011969849467277527, + 0.004634816199541092, + 0.006574948318302631, + 0.00886173639446497, + 0.009985735639929771, + 0.007917673327028751, + 0.007282051257789135, + 0.009144037030637264, + 0.007650306914001703, + 0.005679144058376551, + 0.006243102252483368, + 0.007505709305405617, + 0.008270835503935814, + 0.006247515324503183, + 0.0052473340183496475, + 0.007277944125235081, + 0.006871505174785852, + 0.0031515450682491064, + 0.004727229475975037, + 0.0045919944532215595, + 0.0051495349034667015, + 0.004500722046941519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.03373267874121666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539436340332031 + ], + "opposite_logprob_tokens": [ + -0.42894357442855835 + ], + "logprob_reward": -1.0539436340332031, + "yes_no_margin_reward": -0.6250000596046448, + "token_credit_reward": -1.0539436340332031, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07868771255016327, + 0.08421152085065842, + 0.05359060689806938, + 0.5526719689369202, + 0.01518416590988636, + 0.05989846587181091, + 0.03917329013347626, + 0.03409900888800621, + 0.025206085294485092, + 0.0572771281003952 + ], + "credit_map": [ + 0.01260103564709425, + 0.007677125744521618, + 0.005370902828872204, + 0.009915781207382679, + 0.011964729987084866, + 0.008005419746041298, + 0.011263727210462093, + 0.011888989247381687, + 0.010534791275858879, + 0.007528641726821661, + 0.010366900824010372, + 0.007032215595245361, + 0.011192120611667633, + 0.01632661372423172, + 0.007789707276970148, + 0.013440527021884918, + 0.0056533850729465485, + 0.004123874474316835, + 0.008853604085743427, + 0.003619296010583639, + 0.0028051927220076323, + 0.016884703189134598, + 0.004384648986160755, + 0.007265906780958176, + 0.01054207980632782, + 0.008255554363131523, + 0.010051118209958076, + 0.012356259860098362, + 0.006345739588141441, + 0.2566431760787964, + 0.0074479770846664906, + 0.24103006720542908, + 0.002100322861224413, + 0.0017043587286025286, + 0.0017019748920574784, + 0.0017628654604777694, + 0.002191301202401519, + 0.0020725505892187357, + 0.0018342507537454367, + 0.0018165415385738015, + 0.008825945667922497, + 0.006808355450630188, + 0.00944378599524498, + 0.00449031125754118, + 0.006904008332639933, + 0.008195683360099792, + 0.008979223668575287, + 0.006251151207834482, + 0.005673838779330254, + 0.004268063697963953, + 0.005673630628734827, + 0.004157746210694313, + 0.0037580719217658043, + 0.004916310776025057, + 0.0047751013189554214, + 0.005950529593974352, + 0.00511186383664608, + 0.003403640817850828, + 0.004203918389976025, + 0.002396320691332221, + 0.004756580106914043, + 0.004198788199573755, + 0.004864880815148354, + 0.005163018591701984, + 0.0044114249758422375, + 0.002842473331838846, + 0.004650318529456854, + 0.0017689791275188327, + 0.0024258301127701998, + 0.0031748961191624403, + 0.002946041524410248, + 0.0029861219227313995, + 0.00820758193731308, + 0.007849961519241333, + 0.007768634706735611, + 0.005930447019636631, + 0.005686365067958832, + 0.007064885459840298, + 0.008099350146949291, + 0.006669903174042702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -1.0539436340332031 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370386362075806 + ], + "opposite_logprob_tokens": [ + -3.4087038040161133 + ], + "logprob_reward": -0.03370386362075806, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370386362075806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0611235648393631, + 0.05064792558550835, + 0.14929644763469696, + 0.5847492218017578, + 0.04269370809197426, + 0.048401616513729095, + 0.012024651281535625, + 0.015568777918815613, + 0.01240161620080471, + 0.023092463612556458 + ], + "credit_map": [ + 0.009455333463847637, + 0.00931556150317192, + 0.004961281083524227, + 0.006574749480932951, + 0.008248826488852501, + 0.006837205961346626, + 0.009707905352115631, + 0.006022705230861902, + 0.005434071645140648, + 0.004159749951213598, + 0.00503665953874588, + 0.00366392033174634, + 0.004692546557635069, + 0.01411727350205183, + 0.003705918090417981, + 0.00983778852969408, + 0.011313794180750847, + 0.007617058698087931, + 0.04546995460987091, + 0.007719718851149082, + 0.00779708893969655, + 0.05143727734684944, + 0.006784158293157816, + 0.011157399043440819, + 0.011203012429177761, + 0.008901355788111687, + 0.01082905475050211, + 0.013757639564573765, + 0.007687514182180166, + 0.2707210183143616, + 0.007586160209029913, + 0.2540634572505951, + 0.00661816680803895, + 0.005390331614762545, + 0.004115115385502577, + 0.004162585362792015, + 0.004294121637940407, + 0.005781948566436768, + 0.005263408645987511, + 0.007068032864481211, + 0.007697728928178549, + 0.0051350900903344154, + 0.008715536445379257, + 0.0035888496786355972, + 0.005223534069955349, + 0.008075756020843983, + 0.0055723595432937145, + 0.004392760340124369, + 0.0020796710159629583, + 0.000874944613315165, + 0.0017460781382396817, + 0.001054297899827361, + 0.0010357528226450086, + 0.0015117530710995197, + 0.001380575937218964, + 0.0023415787145495415, + 0.0027168949600309134, + 0.0016524563543498516, + 0.002708560787141323, + 0.0010673052165657282, + 0.002481699688360095, + 0.0018255515024065971, + 0.0015444313175976276, + 0.001571879256516695, + 0.0016911770217120647, + 0.001420664950273931, + 0.0021649033296853304, + 0.000948958215303719, + 0.0013479513581842184, + 0.0016040762420743704, + 0.0021317845676094294, + 0.0010921003995463252, + 0.0035293048713356256, + 0.00409039156511426, + 0.003603002056479454, + 0.0018690444994717836, + 0.0020170919597148895, + 0.0026585126761347055, + 0.0026612880174070597, + 0.00266383052803576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03370386362075806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a fist colliding with a wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gauntlet constructed of dark riveted steel and thick battle-worn leather?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knuckles fitted with heavy tarnished brass dust collectors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause a spiderweb of crystalline cracks on the dark granite wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a harsh keylight casting deep shadows on a billowing cloud of dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293639242649078 + ], + "opposite_logprob_tokens": [ + -2.237936496734619 + ], + "logprob_reward": -0.11293639242649078, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293639242649078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05399034172296524, + 0.0356295108795166, + 0.0614890418946743, + 0.5194662809371948, + 0.054892439395189285, + 0.1434696763753891, + 0.0359983928501606, + 0.015194275416433811, + 0.05411025881767273, + 0.025759823620319366 + ], + "credit_map": [ + 0.007654265966266394, + 0.00571552524343133, + 0.0031279108952730894, + 0.005044769961386919, + 0.01091929990798235, + 0.007395484950393438, + 0.008207891136407852, + 0.005925192032009363, + 0.003111151745542884, + 0.0033782008104026318, + 0.002766095334663987, + 0.0036162708420306444, + 0.0033640791662037373, + 0.00815904513001442, + 0.005195300560444593, + 0.0060393656603991985, + 0.007714689709246159, + 0.0036306041292846203, + 0.014546209014952183, + 0.004247277043759823, + 0.003145835129544139, + 0.01366431824862957, + 0.005341175477951765, + 0.009198932908475399, + 0.010210545733571053, + 0.008888831362128258, + 0.010122453793883324, + 0.012760791927576065, + 0.008283508010208607, + 0.23772495985031128, + 0.008142471313476562, + 0.2233327180147171, + 0.00495161535218358, + 0.006249910686165094, + 0.0052818795666098595, + 0.005481076426804066, + 0.010131453163921833, + 0.00781237194314599, + 0.008569958619773388, + 0.006414174567908049, + 0.022313542664051056, + 0.018604710698127747, + 0.028055904433131218, + 0.009838570840656757, + 0.01625671610236168, + 0.020805908367037773, + 0.01156340166926384, + 0.01603090576827526, + 0.005400550086051226, + 0.004493251442909241, + 0.005530929658561945, + 0.0035959482192993164, + 0.0037138164043426514, + 0.004457792732864618, + 0.0045957560651004314, + 0.0042103491723537445, + 0.0025837456341832876, + 0.0011495630024001002, + 0.0024605621583759785, + 0.0010496050817891955, + 0.0017223218455910683, + 0.0020359314512461424, + 0.0025463271886110306, + 0.0016462196363136172, + 0.007605845108628273, + 0.005806154105812311, + 0.013983124867081642, + 0.002987960120663047, + 0.005527498200535774, + 0.007134308572858572, + 0.006522939074784517, + 0.004542427137494087, + 0.0039358423091471195, + 0.004118895623832941, + 0.0043327538296580315, + 0.0016816813731566072, + 0.0024365384597331285, + 0.003040655981749296, + 0.0029901706147938967, + 0.0032232855446636677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.11293639242649078 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_12/index_tensor([4800, 4800])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_12/index_tensor([4800, 4800])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ab9f76b2e1330bdcfe9e4ea66ac77f942bef54d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_12/index_tensor([4800, 4800])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins in extreme close-up, focusing on the long, flowing sleeves of a gorgeous Jin Opera costume, crafted from bright pink silk. The cuffs are delicately embroidered with intricate peony and phoenix motifs in gold thread, resting quietly on the actress's lap. Then, with an incredibly slow and smooth rhythm, the camera tilts upward, sweeping across the layers of vibrantly embroidered bodice, finally settling firmly on the face of the Hua Dan actress. She wears an exquisitely crafted phoenix crown inlaid with kingfisher feathers and gemstones, but our gaze is drawn to her face, painted with delicate opera makeup: beneath a snow-white foundation, a deep sense of fatigue can't be concealed. Her dramatically painted eyes are slightly downcast, their gaze vacant, devoid of their stage presence. A beam of direct light from a side-mounted vanity mirror sharply outlines the contours of her cheeks and makes the strands of pearl tassels dangling limply from either side of her phoenix coronet appear especially distinct. They rest motionlessly against her temples, no longer swaying with her movements, as if they, along with their host, have fallen into the silence and ennui that follows the performance. The entire scene takes place in a dimly lit, slightly cluttered backstage corner, with faint traces of hangers and prop boxes visible in the background. This creates an intimate and sentimental atmosphere, a fusion of glamour and reality, a stark contrast between on-stage splendor and backstage exhaustion.", + "seed": 7822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_12/index_4800_seed_7822_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05521318051032722, + "vlm_logprob": -0.05521318051032722, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.05521318051032722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004337427206337452 + ], + "opposite_logprob_tokens": [ + -5.879337310791016 + ], + "logprob_reward": -0.004337427206337452, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004337427206337452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030467238277196884, + 0.7539739012718201, + 0.018334710970520973, + 0.019046857953071594, + 0.023142222315073013, + 0.010436267592012882, + 0.033221255987882614, + 0.04322566092014313, + 0.041903845965862274, + 0.02624806948006153 + ], + "credit_map": [ + 0.0024921183940023184, + 0.004282466601580381, + 0.003008065512403846, + 0.004647369030863047, + 0.004598456434905529, + 0.0038018005434423685, + 0.003893522545695305, + 0.003743440145626664, + 0.02073587477207184, + 0.010856756940484047, + 0.030395686626434326, + 0.0167095810174942, + 0.01349182054400444, + 0.3414803743362427, + 0.023781729862093925, + 0.2965220808982849, + 0.0019046987872570753, + 0.0019695013761520386, + 0.001452179509215057, + 0.0032245698384940624, + 0.0016861994517967105, + 0.0022057276219129562, + 0.0037774399388581514, + 0.002114395145326853, + 0.0025626097340136766, + 0.002341418992727995, + 0.0029501684475690126, + 0.0024728390853852034, + 0.001456535654142499, + 0.0034248821903020144, + 0.0016138842329382896, + 0.0022245203144848347, + 0.0033267089165747166, + 0.002427465282380581, + 0.003414271166548133, + 0.0033640898764133453, + 0.003299474949017167, + 0.0018970235250890255, + 0.0033444457221776247, + 0.0020687428768724203, + 0.0011265756329521537, + 0.001189824310131371, + 0.0016416226280853152, + 0.0011468566954135895, + 0.0013310397043824196, + 0.0012980606406927109, + 0.001393043203279376, + 0.0013092445442453027, + 0.0021890311036258936, + 0.003919287584722042, + 0.0028637913055717945, + 0.002362753264605999, + 0.00952619407325983, + 0.0050531779415905476, + 0.0047710249200463295, + 0.002535996725782752, + 0.0029368617106229067, + 0.005029106512665749, + 0.004990755580365658, + 0.003764715511351824, + 0.011524293571710587, + 0.005301208235323429, + 0.006375937256962061, + 0.0033027802128344774, + 0.0034418825525790453, + 0.004754831548780203, + 0.0037938901223242283, + 0.004385031294077635, + 0.008203835226595402, + 0.004722477402538061, + 0.007081836462020874, + 0.005520063452422619, + 0.0019183293916285038, + 0.002572219353169203, + 0.003144602756947279, + 0.0030695819295942783, + 0.0024306438863277435, + 0.004263273440301418, + 0.005076513160020113, + 0.0037729048635810614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.004337427206337452 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02986508421599865 + ], + "opposite_logprob_tokens": [ + -3.529865026473999 + ], + "logprob_reward": -0.02986508421599865, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02986508421599865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18829359114170074, + 0.45726627111434937, + 0.0171271450817585, + 0.039923571050167084, + 0.02716813236474991, + 0.056114453822374344, + 0.09090935438871384, + 0.0346711091697216, + 0.04946453496813774, + 0.03906187787652016 + ], + "credit_map": [ + 0.014129623770713806, + 0.026549391448497772, + 0.01365643460303545, + 0.02434832602739334, + 0.03837912157177925, + 0.0236799456179142, + 0.022681385278701782, + 0.024869369342923164, + 0.012350802309811115, + 0.007186594419181347, + 0.018351178616285324, + 0.01140128169208765, + 0.008252857252955437, + 0.20586666464805603, + 0.014070446603000164, + 0.17978641390800476, + 0.0017849502619355917, + 0.0032683697063475847, + 0.0014500346733257174, + 0.002951731439679861, + 0.0014089691685512662, + 0.002215609885752201, + 0.0018117327708750963, + 0.0022357453126460314, + 0.009287039749324322, + 0.0031927970703691244, + 0.0061151692643761635, + 0.0029004011303186417, + 0.0035796931479126215, + 0.005304154008626938, + 0.003378526074811816, + 0.006165789440274239, + 0.003227285807952285, + 0.0031272131018340588, + 0.005261507350951433, + 0.003421556204557419, + 0.003914650063961744, + 0.002570616314187646, + 0.002453765831887722, + 0.003191537456586957, + 0.004420938901603222, + 0.005819950718432665, + 0.006846883334219456, + 0.007938939146697521, + 0.009931241162121296, + 0.008320964872837067, + 0.0065795849077403545, + 0.006255950778722763, + 0.004220332019031048, + 0.010008483193814754, + 0.007343449629843235, + 0.0066567775793373585, + 0.029654178768396378, + 0.014395885169506073, + 0.012241542339324951, + 0.00638870894908905, + 0.0038622457068413496, + 0.003760767402127385, + 0.010223185643553734, + 0.002523132599890232, + 0.002772978972643614, + 0.005201635882258415, + 0.0033634358551353216, + 0.002963725244626403, + 0.005523874890059233, + 0.005800964310765266, + 0.0054288762621581554, + 0.005387853365391493, + 0.009169398806989193, + 0.004081268794834614, + 0.007910789921879768, + 0.006161509547382593, + 0.002665209583938122, + 0.00305533385835588, + 0.005057196598500013, + 0.0030827114824205637, + 0.00427718460559845, + 0.006761620752513409, + 0.00826039258390665, + 0.005902230739593506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.02986508421599865 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022003266494721174 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0022003266494721174, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022003266494721174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03476833179593086, + 0.7606819868087769, + 0.011800285428762436, + 0.009891332127153873, + 0.018220776692032814, + 0.007286165375262499, + 0.040411852300167084, + 0.028798876330256462, + 0.031634025275707245, + 0.056506358087062836 + ], + "credit_map": [ + 0.003082041395828128, + 0.003254205221310258, + 0.0027292976155877113, + 0.003703428665176034, + 0.005435091443359852, + 0.004842997994273901, + 0.005540916230529547, + 0.006180350668728352, + 0.0205527413636446, + 0.00975183304399252, + 0.029562760144472122, + 0.016955649480223656, + 0.013455248437821865, + 0.34432655572891235, + 0.02309020422399044, + 0.3029870390892029, + 0.0014654722763225436, + 0.0017911844188347459, + 0.0010841250186786056, + 0.0017197284614667296, + 0.001230037771165371, + 0.001565463375300169, + 0.0012253695167601109, + 0.0017189054051414132, + 0.0011956251692026854, + 0.001240373938344419, + 0.0013188575394451618, + 0.0010403384221717715, + 0.001094228238798678, + 0.0013608827721327543, + 0.0014563477598130703, + 0.0011846780544146895, + 0.0027901555877178907, + 0.0014786527026444674, + 0.0023635888937860727, + 0.0014246833743527532, + 0.003785294946283102, + 0.001227102242410183, + 0.0030100250151008368, + 0.0021412742789834738, + 0.0006304227281361818, + 0.0007007861277088523, + 0.0013609940651804209, + 0.0005937036476098001, + 0.0011328276013955474, + 0.001041683368384838, + 0.0012224779929965734, + 0.0006032702513039112, + 0.0024154512211680412, + 0.004495222587138414, + 0.00300078303553164, + 0.0028124211821705103, + 0.013696881011128426, + 0.006331368815153837, + 0.004704909399151802, + 0.002954813651740551, + 0.003047363366931677, + 0.0034117461182177067, + 0.0066065541468560696, + 0.0021667226683348417, + 0.004234607797116041, + 0.003220467595383525, + 0.003979507368057966, + 0.0021319095976650715, + 0.0028135247994214296, + 0.00403020391240716, + 0.0030314449686557055, + 0.003548800479620695, + 0.006362686865031719, + 0.0029112708289176226, + 0.005023316480219364, + 0.003912776708602905, + 0.0070221624337136745, + 0.006922244094312191, + 0.008468340151011944, + 0.004575869534164667, + 0.007095648907124996, + 0.007415351457893848, + 0.009141041897237301, + 0.005865699145942926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.0022003266494721174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038112908601760864 + ], + "opposite_logprob_tokens": [ + -3.2881128787994385 + ], + "logprob_reward": -0.038112908601760864, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038112908601760864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07688587158918381, + 0.6214439272880554, + 0.01103305909782648, + 0.009948519989848137, + 0.011372857727110386, + 0.00901902187615633, + 0.061666734516620636, + 0.0643148273229599, + 0.0717390924692154, + 0.06257611513137817 + ], + "credit_map": [ + 0.005900473799556494, + 0.00977169070392847, + 0.006054717116057873, + 0.008269439451396465, + 0.01331564411520958, + 0.01001920085400343, + 0.012321624904870987, + 0.011233078315854073, + 0.017188485711812973, + 0.008643830195069313, + 0.024360092356801033, + 0.013681809417903423, + 0.011011607944965363, + 0.28146493434906006, + 0.0190441757440567, + 0.24604888260364532, + 0.0011588437482714653, + 0.0019746441394090652, + 0.0008112338837236166, + 0.0019421088509261608, + 0.0009460890432819724, + 0.0013681589625775814, + 0.0014707980444654822, + 0.0013611825415864587, + 0.0015548976371064782, + 0.00115957111120224, + 0.0012603302020579576, + 0.0011593293165788054, + 0.0010100480867549777, + 0.001361342379823327, + 0.0011774433078244328, + 0.0012655581813305616, + 0.0012892979430034757, + 0.0011881761020049453, + 0.0013019520556554198, + 0.0014573736116290092, + 0.0016563583631068468, + 0.001197273493744433, + 0.0018215960590168834, + 0.00146082928404212, + 0.00095470214728266, + 0.0009184739901684225, + 0.0008017216459847987, + 0.0014867952559143305, + 0.0013417912414297462, + 0.0012627028627321124, + 0.0012691376032307744, + 0.000983696198090911, + 0.0032720149029046297, + 0.007059828843921423, + 0.004773348104208708, + 0.00494946027174592, + 0.01888958550989628, + 0.009129972197115421, + 0.008598752319812775, + 0.004993767011910677, + 0.0035627391189336777, + 0.005743158981204033, + 0.010243714787065983, + 0.004718791227787733, + 0.0160246342420578, + 0.008408008143305779, + 0.008310033939778805, + 0.007303744088858366, + 0.005419989116489887, + 0.007580129895359278, + 0.006191104184836149, + 0.009531505405902863, + 0.012621579691767693, + 0.008421486243605614, + 0.012206094339489937, + 0.009767203591763973, + 0.00659751333296299, + 0.006980174221098423, + 0.008053348399698734, + 0.004648690111935139, + 0.008376412093639374, + 0.011499532498419285, + 0.008697825483977795, + 0.007722609210759401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.038112908601760864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155015587806702 + ], + "opposite_logprob_tokens": [ + -1.7015501260757446 + ], + "logprob_reward": -0.20155015587806702, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155015587806702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033846840262413025, + 0.41506093740463257, + 0.06735121458768845, + 0.033332157880067825, + 0.041459646075963974, + 0.014249213971197605, + 0.05855768918991089, + 0.05037577450275421, + 0.1742212474346161, + 0.11154519766569138 + ], + "credit_map": [ + 0.004266984295099974, + 0.004865862429141998, + 0.0027354308404028416, + 0.004518914967775345, + 0.004464676603674889, + 0.003957072272896767, + 0.00488792872056365, + 0.004149974323809147, + 0.012235080823302269, + 0.010524103417992592, + 0.016460949555039406, + 0.011770669370889664, + 0.0107519356533885, + 0.18143734335899353, + 0.012997117824852467, + 0.15888376533985138, + 0.005930457264184952, + 0.01022600568830967, + 0.005569100379943848, + 0.01078111119568348, + 0.005824307445436716, + 0.008528796955943108, + 0.010553009808063507, + 0.009938436560332775, + 0.008938577026128769, + 0.004136953037232161, + 0.0032731257379055023, + 0.0036671848502010107, + 0.0022007550578564405, + 0.0037462853360921144, + 0.0028023358900099993, + 0.004566946066915989, + 0.0032766324002295732, + 0.0023458076175302267, + 0.0038975144270807505, + 0.0032885470427572727, + 0.012638195417821407, + 0.0027471312787383795, + 0.008279474452137947, + 0.0049863457679748535, + 0.0015871673822402954, + 0.0014458245132118464, + 0.0017077433876693249, + 0.0023848521523177624, + 0.0017795980675145984, + 0.0019453524146229029, + 0.0019168497528880835, + 0.0014818278141319752, + 0.002877733903005719, + 0.0068944902159273624, + 0.004136157687753439, + 0.004842770751565695, + 0.017691096290946007, + 0.008943131193518639, + 0.008179721422493458, + 0.004992594011127949, + 0.003928145859390497, + 0.005082797724753618, + 0.009638063609600067, + 0.0038616948295384645, + 0.009551753289997578, + 0.006433655507862568, + 0.008167671971023083, + 0.0037119982298463583, + 0.017055664211511612, + 0.016773492097854614, + 0.01822509989142418, + 0.027838030830025673, + 0.021774372085928917, + 0.025202130898833275, + 0.023830287158489227, + 0.023522192612290382, + 0.01426888257265091, + 0.016054747626185417, + 0.015448804013431072, + 0.011939414776861668, + 0.011974465101957321, + 0.015329047106206417, + 0.013366397470235825, + 0.013163452036678791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.20155015587806702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in extreme close-up, focusing on the long, flowing sleeves of a gorgeous Jin Opera costume, crafted from bright pink silk. The cuffs are delicately embroidered with intricate peony and phoenix motifs in gold thread, resting quietly on the actress's lap. Then, with an incredibly slow and smooth rhythm, the camera tilts upward, sweeping across the layers of vibrantly embroidered bodice, finally settling firmly on the face of the Hua Dan actress. She wears an exquisitely crafted phoenix crown inlaid with kingfisher feathers and gemstones, but our gaze is drawn to her face, painted with delicate opera makeup: beneath a snow-white foundation, a deep sense of fatigue can't be concealed. Her dramatically painted eyes are slightly downcast, their gaze vacant, devoid of their stage presence. A beam of direct light from a side-mounted vanity mirror sharply outlines the contours of her cheeks and makes the strands of pearl tassels dangling limply from either side of her phoenix coronet appear especially distinct. They rest motionlessly against her temples, no longer swaying with her movements, as if they, along with their host, have fallen into the silence and ennui that follows the performance. The entire scene takes place in a dimly lit, slightly cluttered backstage corner, with faint traces of hangers and prop boxes visible in the background. This creates an intimate and sentimental atmosphere, a fusion of glamour and reality, a stark contrast between on-stage splendor and backstage exhaustion.", + "seed": 7823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_12/index_4800_seed_7823_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04213684289716184, + "vlm_logprob": -0.04213684289716184, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04213684289716184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00969995092600584 + ], + "opposite_logprob_tokens": [ + -4.759699821472168 + ], + "logprob_reward": -0.00969995092600584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00969995092600584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.527228832244873, + 0.11963430792093277, + 0.04703875258564949, + 0.031332217156887054, + 0.016675883904099464, + 0.027275336906313896, + 0.08388306200504303, + 0.04389384761452675, + 0.08415815979242325, + 0.01887957938015461 + ], + "credit_map": [ + 0.06288152933120728, + 0.04277234151959419, + 0.05797649174928665, + 0.05124807357788086, + 0.15422043204307556, + 0.056000564247369766, + 0.062057893723249435, + 0.04007155820727348, + 0.013374675996601582, + 0.017537657171487808, + 0.013809599913656712, + 0.012536686845123768, + 0.01066415011882782, + 0.0284448079764843, + 0.014076871797442436, + 0.009189867414534092, + 0.005079091060906649, + 0.0037108897231519222, + 0.008521431125700474, + 0.005941714625805616, + 0.00652518542483449, + 0.0074012367986142635, + 0.004921554122120142, + 0.0049376534298062325, + 0.002916449448093772, + 0.003908355254679918, + 0.002476206747815013, + 0.00275669083930552, + 0.0038931318558752537, + 0.004236515611410141, + 0.006728814914822578, + 0.004416054580360651, + 0.002039432991296053, + 0.0027089861687272787, + 0.002772917738184333, + 0.0016962051158770919, + 0.0018042461015284061, + 0.0022583012469112873, + 0.0018905317410826683, + 0.0015052625676617026, + 0.00589826749637723, + 0.003979564178735018, + 0.003446379443630576, + 0.0025000658351927996, + 0.003943650051951408, + 0.0018469466594979167, + 0.0018599712057039142, + 0.0038004934322088957, + 0.015732107684016228, + 0.012906789779663086, + 0.01082875207066536, + 0.0061143673956394196, + 0.01248707715421915, + 0.012262970209121704, + 0.006941798608750105, + 0.006609207019209862, + 0.008403205312788486, + 0.004555391613394022, + 0.007219997234642506, + 0.004123762249946594, + 0.007136782631278038, + 0.004291676916182041, + 0.004373339004814625, + 0.0037896924186497927, + 0.012477384880185127, + 0.00670125987380743, + 0.010572683997452259, + 0.006001180503517389, + 0.011704601347446442, + 0.02386913076043129, + 0.0055360798723995686, + 0.00729584414511919, + 0.0028009237721562386, + 0.002280859975144267, + 0.0019512630533427, + 0.0019391804235056043, + 0.0030249322298914194, + 0.0022682328708469868, + 0.002569415606558323, + 0.002044773194938898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.00969995092600584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05509984493255615 + ], + "opposite_logprob_tokens": [ + -2.9300999641418457 + ], + "logprob_reward": -0.05509984493255615, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05509984493255615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4650982916355133, + 0.12849201261997223, + 0.03310171142220497, + 0.10087232291698456, + 0.08478348702192307, + 0.027792543172836304, + 0.024094369262456894, + 0.030791964381933212, + 0.07427558302879333, + 0.030697723850607872 + ], + "credit_map": [ + 0.05824393033981323, + 0.05003920570015907, + 0.05734868720173836, + 0.06197856366634369, + 0.0977502167224884, + 0.04638435319066048, + 0.06112145632505417, + 0.032231878489255905, + 0.016206413507461548, + 0.02429816871881485, + 0.01888902299106121, + 0.0174811203032732, + 0.012890812940895557, + 0.013328167609870434, + 0.015514787286520004, + 0.009883515536785126, + 0.002976152114570141, + 0.003392812330275774, + 0.00389260146766901, + 0.00480229826644063, + 0.004194512031972408, + 0.0043464284390211105, + 0.004293980076909065, + 0.005202928092330694, + 0.011223332956433296, + 0.007960374467074871, + 0.008437465876340866, + 0.008516292087733746, + 0.015316909179091454, + 0.013373327441513538, + 0.01528610847890377, + 0.02075851336121559, + 0.014485834166407585, + 0.01018452923744917, + 0.01892261579632759, + 0.006233683321624994, + 0.009427491575479507, + 0.007080982904881239, + 0.008618365973234177, + 0.009829988703131676, + 0.004339168313890696, + 0.0030335015617311, + 0.0033455458469688892, + 0.003139972686767578, + 0.003817863529548049, + 0.0030328365974128246, + 0.0037034666165709496, + 0.0033801875542849302, + 0.004127338994294405, + 0.002566736890003085, + 0.0032969433814287186, + 0.00151528709102422, + 0.004259980749338865, + 0.003141196444630623, + 0.0035249425563961267, + 0.001661942107602954, + 0.004050571005791426, + 0.003902920288965106, + 0.003905365476384759, + 0.0033309506252408028, + 0.004353927448391914, + 0.004530949983745813, + 0.002920117462053895, + 0.0037971632555127144, + 0.012560989707708359, + 0.008161121979355812, + 0.010098393075168133, + 0.005483122076839209, + 0.00858239270746708, + 0.016736363992094994, + 0.00628709327429533, + 0.006366106681525707, + 0.0046354844234883785, + 0.002962418831884861, + 0.003933236002922058, + 0.002914097625762224, + 0.003256336087360978, + 0.003469826653599739, + 0.004316990729421377, + 0.005209334660321474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.05509984493255615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024925374891608953 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024925374891608953, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024925374891608953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15316344797611237, + 0.17445388436317444, + 0.060880377888679504, + 0.08676783740520477, + 0.06121787801384926, + 0.07130078226327896, + 0.06040719524025917, + 0.11742623150348663, + 0.11796373128890991, + 0.09641861915588379 + ], + "credit_map": [ + 0.020829014480113983, + 0.014225275255739689, + 0.029150672256946564, + 0.016194120049476624, + 0.025211824104189873, + 0.015010212548077106, + 0.01963256672024727, + 0.012909774668514729, + 0.022031579166650772, + 0.02294890396296978, + 0.022879406809806824, + 0.017715461552143097, + 0.015131630003452301, + 0.043088801205158234, + 0.020608076825737953, + 0.010050027631223202, + 0.005344447214156389, + 0.007657038047909737, + 0.006605624686926603, + 0.007010871544480324, + 0.0082602733746171, + 0.008747830986976624, + 0.007795488461852074, + 0.009458803571760654, + 0.006878376938402653, + 0.009143341332674026, + 0.006031523458659649, + 0.009988147765398026, + 0.014705654233694077, + 0.012317129410803318, + 0.014998993836343288, + 0.01270467322319746, + 0.005124450195580721, + 0.007735977414995432, + 0.006743594538420439, + 0.007707501761615276, + 0.009210843592882156, + 0.010381930507719517, + 0.007826514542102814, + 0.006487064994871616, + 0.011505061760544777, + 0.006395665463060141, + 0.01908792555332184, + 0.0048414673656225204, + 0.008114087395370007, + 0.007247930858284235, + 0.0038402907084673643, + 0.010268342681229115, + 0.008141613565385342, + 0.010474205948412418, + 0.008460629731416702, + 0.004773756023496389, + 0.008038593456149101, + 0.007821060717105865, + 0.007380919996649027, + 0.005316417198628187, + 0.02464097924530506, + 0.012879934161901474, + 0.02896789461374283, + 0.007670605555176735, + 0.015228155069053173, + 0.013110407628118992, + 0.00867016427218914, + 0.006258099339902401, + 0.016262397170066833, + 0.009674500674009323, + 0.014046899974346161, + 0.007951932959258556, + 0.0183172058314085, + 0.030945802107453346, + 0.009954766370356083, + 0.010810227133333683, + 0.01589064486324787, + 0.01037980243563652, + 0.011259798891842365, + 0.010516619309782982, + 0.013774353079497814, + 0.011651149019598961, + 0.013597609475255013, + 0.009348643943667412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0024925374891608953 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043083611875772476 + ], + "opposite_logprob_tokens": [ + -3.168083667755127 + ], + "logprob_reward": -0.043083611875772476, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083611875772476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2152673304080963, + 0.16850581765174866, + 0.0648559182882309, + 0.08233339339494705, + 0.04139167442917824, + 0.04586438089609146, + 0.03961716964840889, + 0.05798812210559845, + 0.12847194075584412, + 0.15570427477359772 + ], + "credit_map": [ + 0.02287537232041359, + 0.02678365632891655, + 0.03661779314279556, + 0.026989035308361053, + 0.047216739505529404, + 0.018844814971089363, + 0.022625893354415894, + 0.013314023613929749, + 0.019476046785712242, + 0.0238762516528368, + 0.018782014027237892, + 0.018042191863059998, + 0.015789514407515526, + 0.041355304419994354, + 0.020042793825268745, + 0.011141698807477951, + 0.007505984511226416, + 0.004700831603258848, + 0.008761031553149223, + 0.007052337285131216, + 0.01177725475281477, + 0.00919023435562849, + 0.0073739998042583466, + 0.008494241163134575, + 0.007508002687245607, + 0.006934022065252066, + 0.00957933533936739, + 0.011227146722376347, + 0.0167272686958313, + 0.010779567994177341, + 0.010678280144929886, + 0.008899766020476818, + 0.0043081981129944324, + 0.005377302411943674, + 0.0044203856959939, + 0.004175391048192978, + 0.007966008968651295, + 0.004808667115867138, + 0.005927814170718193, + 0.004407909698784351, + 0.0069282944314181805, + 0.004891982767730951, + 0.005933792795985937, + 0.0038200642447918653, + 0.00788016989827156, + 0.005193565506488085, + 0.006333823781460524, + 0.004882688168436289, + 0.005413900129497051, + 0.004810125567018986, + 0.005316176917403936, + 0.0034218523651361465, + 0.00564044713973999, + 0.007815102115273476, + 0.004488778300583363, + 0.0027107882779091597, + 0.008922175504267216, + 0.005821943748742342, + 0.009181446395814419, + 0.007665286771953106, + 0.007024683058261871, + 0.005662605632096529, + 0.0073245069943368435, + 0.006385473068803549, + 0.017371222376823425, + 0.015230560675263405, + 0.015314287506043911, + 0.010507320053875446, + 0.016604365780949593, + 0.023768192157149315, + 0.013672907836735249, + 0.016003089025616646, + 0.0205979123711586, + 0.015160691924393177, + 0.02053855173289776, + 0.01581750065088272, + 0.02115805074572563, + 0.017148008570075035, + 0.017112912610173225, + 0.028170648962259293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.043083611875772476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030826926231384 + ], + "opposite_logprob_tokens": [ + -2.3503081798553467 + ], + "logprob_reward": -0.10030826926231384, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030826926231384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11865893751382828, + 0.35731029510498047, + 0.1520126312971115, + 0.04287264496088028, + 0.023775825276970863, + 0.022413192316889763, + 0.04448806121945381, + 0.057081133127212524, + 0.048336341977119446, + 0.13305096328258514 + ], + "credit_map": [ + 0.013430939055979252, + 0.01224397961050272, + 0.023798368871212006, + 0.010486166924238205, + 0.02899884060025215, + 0.010132134892046452, + 0.012184282764792442, + 0.0073842271231114864, + 0.04650005325675011, + 0.049047406762838364, + 0.035770483314991, + 0.031564779579639435, + 0.04078960418701172, + 0.08538370579481125, + 0.04374927654862404, + 0.024504978209733963, + 0.019503485411405563, + 0.011098689399659634, + 0.019303863868117332, + 0.016287293285131454, + 0.029181454330682755, + 0.022498706355690956, + 0.014515158720314503, + 0.019623981788754463, + 0.003615377005189657, + 0.003728665644302964, + 0.004355350509285927, + 0.003365017706528306, + 0.005811265204101801, + 0.006337902508676052, + 0.004810533951967955, + 0.010848534293472767, + 0.002181354444473982, + 0.0027637104503810406, + 0.003086511977016926, + 0.0026515561621636152, + 0.003599633928388357, + 0.0038174467626959085, + 0.003071489045396447, + 0.0026041229721158743, + 0.004369536880403757, + 0.002505751559510827, + 0.003785377135500312, + 0.002143458928912878, + 0.0034408627543598413, + 0.0023305665235966444, + 0.002101602964103222, + 0.0017360355705022812, + 0.007002145983278751, + 0.006670522503554821, + 0.005873648915439844, + 0.0036864005960524082, + 0.005701255984604359, + 0.00677719758823514, + 0.004567828960716724, + 0.004209062084555626, + 0.009127013385295868, + 0.006586372386664152, + 0.007014139089733362, + 0.00796559825539589, + 0.006123526953160763, + 0.006280266214162111, + 0.0055522918701171875, + 0.008431924507021904, + 0.008952931500971317, + 0.004908150527626276, + 0.006713357754051685, + 0.005204220302402973, + 0.004826307762414217, + 0.008091477677226067, + 0.004229491576552391, + 0.005410401616245508, + 0.019539572298526764, + 0.012469017878174782, + 0.015560180880129337, + 0.012907737866044044, + 0.009493235498666763, + 0.02282189019024372, + 0.011686108075082302, + 0.02857322245836258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.10030826926231384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in extreme close-up, focusing on the long, flowing sleeves of a gorgeous Jin Opera costume, crafted from bright pink silk. The cuffs are delicately embroidered with intricate peony and phoenix motifs in gold thread, resting quietly on the actress's lap. Then, with an incredibly slow and smooth rhythm, the camera tilts upward, sweeping across the layers of vibrantly embroidered bodice, finally settling firmly on the face of the Hua Dan actress. She wears an exquisitely crafted phoenix crown inlaid with kingfisher feathers and gemstones, but our gaze is drawn to her face, painted with delicate opera makeup: beneath a snow-white foundation, a deep sense of fatigue can't be concealed. Her dramatically painted eyes are slightly downcast, their gaze vacant, devoid of their stage presence. A beam of direct light from a side-mounted vanity mirror sharply outlines the contours of her cheeks and makes the strands of pearl tassels dangling limply from either side of her phoenix coronet appear especially distinct. They rest motionlessly against her temples, no longer swaying with her movements, as if they, along with their host, have fallen into the silence and ennui that follows the performance. The entire scene takes place in a dimly lit, slightly cluttered backstage corner, with faint traces of hangers and prop boxes visible in the background. This creates an intimate and sentimental atmosphere, a fusion of glamour and reality, a stark contrast between on-stage splendor and backstage exhaustion.", + "seed": 7824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_12/index_4800_seed_7824_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03297994621098042, + "vlm_logprob": -0.03297994621098042, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03297994621098042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0061727589927613735 + ], + "opposite_logprob_tokens": [ + -5.3811726570129395 + ], + "logprob_reward": -0.0061727589927613735, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0061727589927613735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15233398973941803, + 0.0961424931883812, + 0.2611238956451416, + 0.05196952819824219, + 0.025636520236730576, + 0.09071614593267441, + 0.09694673120975494, + 0.07072295248508453, + 0.11850594729185104, + 0.03590179234743118 + ], + "credit_map": [ + 0.012626108713448048, + 0.01531156525015831, + 0.029573988169431686, + 0.013702319003641605, + 0.02232501655817032, + 0.011232124641537666, + 0.025122124701738358, + 0.022440742701292038, + 0.013255172409117222, + 0.00885305367410183, + 0.008138271979987621, + 0.012229152023792267, + 0.010752061381936073, + 0.009879245422780514, + 0.022442558780312538, + 0.01059297937899828, + 0.020193971693515778, + 0.014672862365841866, + 0.008972814306616783, + 0.011854125186800957, + 0.019204307347536087, + 0.06027961149811745, + 0.10530604422092438, + 0.020640121772885323, + 0.0076550934463739395, + 0.009010516107082367, + 0.006447854917496443, + 0.0040466152131557465, + 0.006048312410712242, + 0.006701032165437937, + 0.007661147508770227, + 0.004398955963551998, + 0.007155645173043013, + 0.0031281698029488325, + 0.003745208727195859, + 0.0015742636751383543, + 0.0021486273035407066, + 0.004395229741930962, + 0.001505384803749621, + 0.0019839894957840443, + 0.0164801012724638, + 0.007204034831374884, + 0.01575300097465515, + 0.006927193142473698, + 0.006734075024724007, + 0.01712832599878311, + 0.008626717142760754, + 0.011862699873745441, + 0.010520632378757, + 0.016801118850708008, + 0.009743897244334221, + 0.004871361888945103, + 0.006923791486769915, + 0.014358896762132645, + 0.004787014797329903, + 0.028940023854374886, + 0.013760508969426155, + 0.008634600788354874, + 0.009134341962635517, + 0.006351983640342951, + 0.008922250010073185, + 0.008630544878542423, + 0.006161939352750778, + 0.009126787073910236, + 0.01583031937479973, + 0.018294228240847588, + 0.027442945167422295, + 0.01279149204492569, + 0.01433053333312273, + 0.011852983385324478, + 0.010471604764461517, + 0.00749184237793088, + 0.005582678597420454, + 0.00453554093837738, + 0.004007347859442234, + 0.004300618544220924, + 0.005600647069513798, + 0.004295483231544495, + 0.003934485837817192, + 0.0036449902690947056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0061727589927613735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07913549989461899 + ], + "opposite_logprob_tokens": [ + -2.5791354179382324 + ], + "logprob_reward": -0.07913549989461899, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07913549989461899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22242870926856995, + 0.06345246732234955, + 0.2812533676624298, + 0.0742977112531662, + 0.15529890358448029, + 0.07173248380422592, + 0.042706198990345, + 0.03460780531167984, + 0.023058613762259483, + 0.031163766980171204 + ], + "credit_map": [ + 0.015959177166223526, + 0.018120603635907173, + 0.03753504902124405, + 0.02674570120871067, + 0.030019324272871017, + 0.020132016390562057, + 0.0441504567861557, + 0.029766378924250603, + 0.008258968591690063, + 0.005186019465327263, + 0.00559531943872571, + 0.00928861740976572, + 0.007015140727162361, + 0.0072808656841516495, + 0.014146984554827213, + 0.006680550053715706, + 0.01893303543329239, + 0.011945249512791634, + 0.007479596883058548, + 0.01046809647232294, + 0.015259437263011932, + 0.09719522297382355, + 0.10029098391532898, + 0.019681723788380623, + 0.009296823292970657, + 0.0075239259749650955, + 0.011539623141288757, + 0.009454806335270405, + 0.007679619360715151, + 0.010960998944938183, + 0.01040954701602459, + 0.0074323671869933605, + 0.029239317402243614, + 0.009477641433477402, + 0.017897341400384903, + 0.007086328696459532, + 0.014504431746900082, + 0.0507395938038826, + 0.008656127378344536, + 0.01769810914993286, + 0.015747103840112686, + 0.006735424045473337, + 0.0054603698663413525, + 0.005095987115055323, + 0.005569687113165855, + 0.01964745856821537, + 0.003645126009359956, + 0.009831320494413376, + 0.0069320667535066605, + 0.0037424974143505096, + 0.0060623399913311005, + 0.00258026085793972, + 0.004166695289313793, + 0.004861167632043362, + 0.004145447630435228, + 0.010215722024440765, + 0.005455737933516502, + 0.007119113579392433, + 0.004103620070964098, + 0.003141806460916996, + 0.005500578321516514, + 0.0028082109056413174, + 0.0029683178290724754, + 0.003510416252538562, + 0.003582644509151578, + 0.002261448185890913, + 0.002918887184932828, + 0.0028287414461374283, + 0.003696648869663477, + 0.002826135605573654, + 0.0026246788911521435, + 0.0023194286040961742, + 0.004752279259264469, + 0.0030016375239938498, + 0.0029959380626678467, + 0.003016920993104577, + 0.005206973757594824, + 0.005502449348568916, + 0.003478762460872531, + 0.0032088065054267645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.07913549989461899 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028232974000275135 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.0028232974000275135, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028232974000275135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15386322140693665, + 0.033459052443504333, + 0.17807628214359283, + 0.06442847847938538, + 0.10865715146064758, + 0.10951423645019531, + 0.13092081248760223, + 0.08457234501838684, + 0.07695640623569489, + 0.059552013874053955 + ], + "credit_map": [ + 0.012452498078346252, + 0.013386057689785957, + 0.04045061394572258, + 0.01699608750641346, + 0.02729014679789543, + 0.012607424519956112, + 0.0165648702532053, + 0.014115510508418083, + 0.004368592984974384, + 0.002332228934392333, + 0.003939393442124128, + 0.003302109194919467, + 0.005364939104765654, + 0.003132498124614358, + 0.0061219907365739346, + 0.004897297825664282, + 0.012694467790424824, + 0.006749676074832678, + 0.00449852691963315, + 0.0058144270442426205, + 0.009802866727113724, + 0.06763327121734619, + 0.05766895040869713, + 0.013214096426963806, + 0.013063370250165462, + 0.00595381623134017, + 0.016680223867297173, + 0.0038995195645838976, + 0.004494919441640377, + 0.007505379151552916, + 0.006617010571062565, + 0.0062142410315573215, + 0.016029175370931625, + 0.006234894506633282, + 0.01442513894289732, + 0.005832899361848831, + 0.010350968688726425, + 0.03605420142412186, + 0.006542191840708256, + 0.013187685050070286, + 0.02607993222773075, + 0.008500342257320881, + 0.008937988430261612, + 0.0076874904334545135, + 0.008619139902293682, + 0.03074944019317627, + 0.005587063264101744, + 0.013352836482226849, + 0.017285700887441635, + 0.03210526332259178, + 0.00705766212195158, + 0.004418112337589264, + 0.00969680491834879, + 0.016048643738031387, + 0.006908188574016094, + 0.03740043193101883, + 0.014170869253575802, + 0.016248242929577827, + 0.007704082410782576, + 0.008651796728372574, + 0.01444641500711441, + 0.007765027694404125, + 0.0074710906483232975, + 0.00811481662094593, + 0.010749028995633125, + 0.008361339569091797, + 0.010371052660048008, + 0.010001937858760357, + 0.007830468937754631, + 0.012738681398332119, + 0.01004698220640421, + 0.00685691274702549, + 0.007651684805750847, + 0.0071186767891049385, + 0.007440520450472832, + 0.008403563871979713, + 0.006656127516180277, + 0.007577083073556423, + 0.008342952467501163, + 0.006361405365169048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0028232974000275135 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033715736120939255 + ], + "opposite_logprob_tokens": [ + -3.4087157249450684 + ], + "logprob_reward": -0.033715736120939255, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033715736120939255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0554744116961956, + 0.05159996077418327, + 0.17719174921512604, + 0.05180501192808151, + 0.06341732293367386, + 0.23484839498996735, + 0.07403388619422913, + 0.09729747474193573, + 0.07760104537010193, + 0.1167306900024414 + ], + "credit_map": [ + 0.005435090512037277, + 0.005562144331634045, + 0.006922953762114048, + 0.006231228355318308, + 0.009550914168357849, + 0.006683898624032736, + 0.007208299823105335, + 0.007879888638854027, + 0.005895962007343769, + 0.00396706210449338, + 0.005220348015427589, + 0.007203997112810612, + 0.005074042361229658, + 0.003604106605052948, + 0.014428005553781986, + 0.0062064435333013535, + 0.019384747371077538, + 0.005919094197452068, + 0.0065737394616007805, + 0.0059175388887524605, + 0.011563129723072052, + 0.07060838490724564, + 0.04260827228426933, + 0.014616858214139938, + 0.009368740022182465, + 0.003482055850327015, + 0.01678035780787468, + 0.0032157478854060173, + 0.002908045891672373, + 0.006615904625505209, + 0.003285088809207082, + 0.006149070803076029, + 0.009200519882142544, + 0.004864519461989403, + 0.007292718160897493, + 0.006155780516564846, + 0.006016730330884457, + 0.017170662060379982, + 0.005666395649313927, + 0.007050001993775368, + 0.09156911075115204, + 0.022813018411397934, + 0.03958745300769806, + 0.010300738736987114, + 0.012225084938108921, + 0.035060491412878036, + 0.01089613325893879, + 0.012396391481161118, + 0.009551596827805042, + 0.016514208167791367, + 0.0041730087250471115, + 0.003920047078281641, + 0.006517449859529734, + 0.008924457244575024, + 0.004972263239324093, + 0.019460860639810562, + 0.014683840796351433, + 0.018992220982909203, + 0.011095980182290077, + 0.008258954621851444, + 0.013834668323397636, + 0.011547224596142769, + 0.008101376704871655, + 0.010783214122056961, + 0.009265600703656673, + 0.008105449378490448, + 0.016602426767349243, + 0.008540324866771698, + 0.009844139218330383, + 0.01095672883093357, + 0.008237653411924839, + 0.0060487231239676476, + 0.018450142815709114, + 0.011578511446714401, + 0.010646509006619453, + 0.014982898719608784, + 0.02234203740954399, + 0.014342904090881348, + 0.014091040007770061, + 0.010296649299561977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.033715736120939255 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305243864655495 + ], + "opposite_logprob_tokens": [ + -3.1680524349212646 + ], + "logprob_reward": -0.04305243864655495, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305243864655495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06466900557279587, + 0.033372003585100174, + 0.04389466717839241, + 0.04664023220539093, + 0.08098331838846207, + 0.06563421338796616, + 0.19915497303009033, + 0.05802825838327408, + 0.33291563391685486, + 0.07470769435167313 + ], + "credit_map": [ + 0.0062201074324548244, + 0.006295191589742899, + 0.009477660059928894, + 0.008639630861580372, + 0.00938376598060131, + 0.008266310207545757, + 0.010218084789812565, + 0.006168256513774395, + 0.004024648107588291, + 0.0031015691347420216, + 0.003315585432574153, + 0.004128813277930021, + 0.003950026817619801, + 0.00317278946749866, + 0.007469876669347286, + 0.004208697471767664, + 0.0042605893686413765, + 0.0022052268031984568, + 0.0022057457827031612, + 0.003042252268642187, + 0.004511456470936537, + 0.01748286373913288, + 0.0052644903771579266, + 0.004922042600810528, + 0.009568248875439167, + 0.0036423918791115284, + 0.01169813796877861, + 0.0027646466623991728, + 0.003325364086776972, + 0.004668933339416981, + 0.0049914163537323475, + 0.0059810904785990715, + 0.011974318884313107, + 0.004574153572320938, + 0.010896534658968449, + 0.003907542210072279, + 0.007710018195211887, + 0.027115950360894203, + 0.004926746711134911, + 0.009878051467239857, + 0.024638595059514046, + 0.006188366562128067, + 0.012603644281625748, + 0.0031605097465217113, + 0.0034745733719319105, + 0.009095024317502975, + 0.0026854926254600286, + 0.003788009053096175, + 0.028997711837291718, + 0.03879989683628082, + 0.013429108075797558, + 0.010407621040940285, + 0.012460327707231045, + 0.025505835190415382, + 0.010563226416707039, + 0.05899125337600708, + 0.008440013974905014, + 0.010604383423924446, + 0.00516223581507802, + 0.007466822396963835, + 0.007967477664351463, + 0.005079067777842283, + 0.005769506562501192, + 0.0075387535616755486, + 0.04642361029982567, + 0.0361206941306591, + 0.0863543376326561, + 0.034539368003606796, + 0.030615640804171562, + 0.04985656589269638, + 0.02769174426794052, + 0.02131367102265358, + 0.010070606134831905, + 0.007976175285875797, + 0.010700822807848454, + 0.008141310885548592, + 0.008680816739797592, + 0.012949557043612003, + 0.007604096084833145, + 0.00858430564403534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.04305243864655495 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in extreme close-up, focusing on the long, flowing sleeves of a gorgeous Jin Opera costume, crafted from bright pink silk. The cuffs are delicately embroidered with intricate peony and phoenix motifs in gold thread, resting quietly on the actress's lap. Then, with an incredibly slow and smooth rhythm, the camera tilts upward, sweeping across the layers of vibrantly embroidered bodice, finally settling firmly on the face of the Hua Dan actress. She wears an exquisitely crafted phoenix crown inlaid with kingfisher feathers and gemstones, but our gaze is drawn to her face, painted with delicate opera makeup: beneath a snow-white foundation, a deep sense of fatigue can't be concealed. Her dramatically painted eyes are slightly downcast, their gaze vacant, devoid of their stage presence. A beam of direct light from a side-mounted vanity mirror sharply outlines the contours of her cheeks and makes the strands of pearl tassels dangling limply from either side of her phoenix coronet appear especially distinct. They rest motionlessly against her temples, no longer swaying with her movements, as if they, along with their host, have fallen into the silence and ennui that follows the performance. The entire scene takes place in a dimly lit, slightly cluttered backstage corner, with faint traces of hangers and prop boxes visible in the background. This creates an intimate and sentimental atmosphere, a fusion of glamour and reality, a stark contrast between on-stage splendor and backstage exhaustion.", + "seed": 7825, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_12/index_4800_seed_7825_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05148761961609125, + "vlm_logprob": -0.05148761961609125, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.05148761961609125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03523838892579079 + ], + "opposite_logprob_tokens": [ + -3.41023850440979 + ], + "logprob_reward": -0.03523838892579079, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03523838892579079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2012144923210144, + 0.14259061217308044, + 0.14797425270080566, + 0.07449644804000854, + 0.18058516085147858, + 0.05894499644637108, + 0.07658985257148743, + 0.03184126690030098, + 0.03255198895931244, + 0.05321095883846283 + ], + "credit_map": [ + 0.023147836327552795, + 0.03405965119600296, + 0.018931003287434578, + 0.024321895092725754, + 0.026180990040302277, + 0.021401364356279373, + 0.027696050703525543, + 0.02547568641602993, + 0.014537378214299679, + 0.025371231138706207, + 0.01614864356815815, + 0.02426018752157688, + 0.014843189157545567, + 0.024619823321700096, + 0.011561651714146137, + 0.01124850194901228, + 0.017297131940722466, + 0.018041742965579033, + 0.014820285141468048, + 0.018410688266158104, + 0.0181198101490736, + 0.01714428886771202, + 0.025251654908061028, + 0.018888650462031364, + 0.010710191912949085, + 0.013813252560794353, + 0.0101496996358037, + 0.0165519081056118, + 0.0051386915147304535, + 0.005911780055612326, + 0.005299663171172142, + 0.0069212596863508224, + 0.018782058730721474, + 0.03682950511574745, + 0.03215208277106285, + 0.05273881182074547, + 0.01290342677384615, + 0.010178245604038239, + 0.008922004140913486, + 0.008079026825726032, + 0.00797954946756363, + 0.008796269074082375, + 0.01196285616606474, + 0.009869837202131748, + 0.0043748305179178715, + 0.005748986266553402, + 0.006384704727679491, + 0.0038279639557003975, + 0.010042066685855389, + 0.011174093000590801, + 0.009018944576382637, + 0.010595393367111683, + 0.0093289315700531, + 0.01396594475954771, + 0.008504430763423443, + 0.003960052039474249, + 0.004626539535820484, + 0.0041790492832660675, + 0.004226686432957649, + 0.003918473608791828, + 0.003992719575762749, + 0.004248199053108692, + 0.004140797536820173, + 0.0025088004767894745, + 0.00417809234932065, + 0.004119539167732, + 0.004189192317426205, + 0.006514551118016243, + 0.003934395965188742, + 0.0025049110408872366, + 0.0052215405739843845, + 0.0018897677073255181, + 0.006959139835089445, + 0.006629087496548891, + 0.008265392854809761, + 0.005663600750267506, + 0.008855435065925121, + 0.007368070539087057, + 0.005201683379709721, + 0.004268549382686615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.03523838892579079 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1005532369017601 + ], + "opposite_logprob_tokens": [ + -2.350553274154663 + ], + "logprob_reward": -0.1005532369017601, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1005532369017601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1595989614725113, + 0.12147680670022964, + 0.08447352796792984, + 0.07036299258470535, + 0.05421239510178566, + 0.0855310782790184, + 0.04099714756011963, + 0.1446024626493454, + 0.09870392084121704, + 0.14004071056842804 + ], + "credit_map": [ + 0.015804622322320938, + 0.016383035108447075, + 0.022152207791805267, + 0.021349750459194183, + 0.01629556156694889, + 0.02324070781469345, + 0.01934429071843624, + 0.025028781965374947, + 0.01650128699839115, + 0.01085997186601162, + 0.013042970560491085, + 0.022496048361063004, + 0.009586929343640804, + 0.026317592710256577, + 0.010407896712422371, + 0.012264109216630459, + 0.016640126705169678, + 0.012821086682379246, + 0.015274016186594963, + 0.014343497343361378, + 0.007409264799207449, + 0.004811912775039673, + 0.006813261192291975, + 0.006360375322401524, + 0.010043160989880562, + 0.011618816293776035, + 0.011082212440669537, + 0.013449265621602535, + 0.003984545357525349, + 0.006085197441279888, + 0.0045425742864608765, + 0.009557225741446018, + 0.010001038201153278, + 0.007687981240451336, + 0.009537397883832455, + 0.007825355045497417, + 0.0056967041455209255, + 0.0051920185796916485, + 0.0054032993502914906, + 0.002868599956855178, + 0.013242950662970543, + 0.008622405119240284, + 0.019632676616311073, + 0.013911186717450619, + 0.009439999237656593, + 0.007172251585870981, + 0.007832332514226437, + 0.005677274893969297, + 0.006294101942330599, + 0.00436573289334774, + 0.0055894795805215836, + 0.005093539599329233, + 0.006506896112114191, + 0.00580369308590889, + 0.0037439933512359858, + 0.003599709365516901, + 0.019366463646292686, + 0.012771829962730408, + 0.017442770302295685, + 0.015916813164949417, + 0.028834206983447075, + 0.018401164561510086, + 0.0184406116604805, + 0.013428602367639542, + 0.010952042415738106, + 0.009138310328125954, + 0.014084739610552788, + 0.01165898609906435, + 0.012382262386381626, + 0.008004083298146725, + 0.021871725097298622, + 0.010611769743263721, + 0.01839528977870941, + 0.01916557177901268, + 0.019666939973831177, + 0.0176664087921381, + 0.020802924409508705, + 0.015269915573298931, + 0.017974266782402992, + 0.01109938882291317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.1005532369017601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642066545784473 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004642066545784473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642066545784473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10280996561050415, + 0.0853530690073967, + 0.0701594427227974, + 0.10693687945604324, + 0.05356841906905174, + 0.07019790261983871, + 0.10454855859279633, + 0.05377969890832901, + 0.19040131568908691, + 0.16224481165409088 + ], + "credit_map": [ + 0.012420304119586945, + 0.01426409836858511, + 0.014185122214257717, + 0.012198464013636112, + 0.012694589793682098, + 0.008036243729293346, + 0.013638557866215706, + 0.01537257432937622, + 0.012313900515437126, + 0.022768378257751465, + 0.007664511911571026, + 0.008696337230503559, + 0.007472645491361618, + 0.014484097249805927, + 0.0059271627105772495, + 0.006026025861501694, + 0.01042785495519638, + 0.010350842960178852, + 0.01188112422823906, + 0.0094252098351717, + 0.007066067308187485, + 0.004887678194791079, + 0.007221255451440811, + 0.008899408392608166, + 0.013818821869790554, + 0.019225651398301125, + 0.019088754430413246, + 0.028816893696784973, + 0.005328696221113205, + 0.006391594652086496, + 0.005528704263269901, + 0.00873775314539671, + 0.009024801664054394, + 0.012930820696055889, + 0.006241712719202042, + 0.008742738515138626, + 0.005754095036536455, + 0.004324480891227722, + 0.003535259747877717, + 0.003014502814039588, + 0.0065817879512906075, + 0.008184412494301796, + 0.013252760283648968, + 0.012394454330205917, + 0.010447283275425434, + 0.006772874854505062, + 0.007763294968754053, + 0.0048010279424488544, + 0.015172293409705162, + 0.013365603983402252, + 0.01164938509464264, + 0.013557430356740952, + 0.011876245029270649, + 0.022197300568223, + 0.009044789709150791, + 0.007685503456741571, + 0.007787884213030338, + 0.005788289941847324, + 0.00553011754527688, + 0.007282385136932135, + 0.008908526971936226, + 0.008240829221904278, + 0.0062227570451796055, + 0.004018903709948063, + 0.020553570240736008, + 0.018788693472743034, + 0.026367243379354477, + 0.023039057850837708, + 0.02414729632437229, + 0.01913583278656006, + 0.038568172603845596, + 0.019801434129476547, + 0.027180800214409828, + 0.02330157533288002, + 0.02333434298634529, + 0.019841019064188004, + 0.018360374495387077, + 0.01439150981605053, + 0.02261931635439396, + 0.013215851970016956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.004642066545784473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206128001213074 + ], + "opposite_logprob_tokens": [ + -2.812061309814453 + ], + "logprob_reward": -0.06206128001213074, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206128001213074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09272301942110062, + 0.16513727605342865, + 0.10576092451810837, + 0.08237546682357788, + 0.044028252363204956, + 0.08854377269744873, + 0.11559269577264786, + 0.061017654836177826, + 0.037648458033800125, + 0.2071724832057953 + ], + "credit_map": [ + 0.01189570501446724, + 0.010836591944098473, + 0.010003996081650257, + 0.013700652867555618, + 0.010404889471828938, + 0.008333941921591759, + 0.014668907038867474, + 0.01287833508104086, + 0.018825365230441093, + 0.019245222210884094, + 0.013791431672871113, + 0.022867515683174133, + 0.02697654813528061, + 0.027188992127776146, + 0.017548678442835808, + 0.018693525344133377, + 0.014497967436909676, + 0.014112098142504692, + 0.018574900925159454, + 0.023205874487757683, + 0.00861153844743967, + 0.0077443234622478485, + 0.009344051592051983, + 0.009670170024037361, + 0.01143103837966919, + 0.014180492609739304, + 0.013902922160923481, + 0.021826239302754402, + 0.004320142790675163, + 0.005193138960748911, + 0.004762403666973114, + 0.00675909174606204, + 0.006938825827091932, + 0.0064865113236010075, + 0.007989197038114071, + 0.008484171703457832, + 0.004313190001994371, + 0.003424228634685278, + 0.0040366146713495255, + 0.0023555136285722256, + 0.01208505779504776, + 0.013197358697652817, + 0.018632065504789352, + 0.01135355792939663, + 0.010454369708895683, + 0.007795294281095266, + 0.00821763277053833, + 0.006808445323258638, + 0.01704416610300541, + 0.012707916088402271, + 0.013825087808072567, + 0.011249043047428131, + 0.018386725336313248, + 0.024571441113948822, + 0.010716204531490803, + 0.007092118263244629, + 0.01081918366253376, + 0.0075494954362511635, + 0.007731411140412092, + 0.008317815139889717, + 0.006899167317897081, + 0.007712273858487606, + 0.007531868293881416, + 0.004456439986824989, + 0.005148657597601414, + 0.0056390054523944855, + 0.005279500503093004, + 0.007400708273053169, + 0.0034470693208277225, + 0.0036211046390235424, + 0.003909716382622719, + 0.003202693536877632, + 0.026636116206645966, + 0.026245631277561188, + 0.03224495053291321, + 0.021737942472100258, + 0.03315876051783562, + 0.02778846211731434, + 0.02567325532436371, + 0.013687366619706154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.06206128001213074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of long, flowing pink silk sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sleeves embroidered with gold thread peony and phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera tilt upward from the sleeves to reveal a Hua Dan actress wearing a phoenix crown with kingfisher feathers and gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the actress's face painted with white foundation makeup, downcast eyes, and pearl tassels hanging limply from her crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a dimly lit, cluttered backstage corner with visible hangers and prop boxes in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494312569499016 + ], + "opposite_logprob_tokens": [ + -2.929943084716797 + ], + "logprob_reward": -0.05494312569499016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494312569499016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06890608370304108, + 0.08794467896223068, + 0.17992033064365387, + 0.21303153038024902, + 0.10545565187931061, + 0.09238806366920471, + 0.07483956962823868, + 0.038392018526792526, + 0.04950892552733421, + 0.08961313217878342 + ], + "credit_map": [ + 0.006560137961059809, + 0.006844210904091597, + 0.006716820411384106, + 0.006238971371203661, + 0.008209123276174068, + 0.009262904524803162, + 0.01170878205448389, + 0.013365128077566624, + 0.012242325581610203, + 0.023773835971951485, + 0.006567050237208605, + 0.008826876990497112, + 0.007905494421720505, + 0.0171688050031662, + 0.005821962840855122, + 0.005638326983898878, + 0.03128652274608612, + 0.03248897194862366, + 0.03209029510617256, + 0.03336014598608017, + 0.010430470108985901, + 0.013438157737255096, + 0.012131367810070515, + 0.01469438336789608, + 0.027336355298757553, + 0.04778791591525078, + 0.02297634817659855, + 0.04372619464993477, + 0.010260660201311111, + 0.018111884593963623, + 0.00755142280831933, + 0.035280756652355194, + 0.03101380355656147, + 0.01277167908847332, + 0.010327196680009365, + 0.02059454843401909, + 0.009508827701210976, + 0.007146055344492197, + 0.006068462505936623, + 0.00802507996559143, + 0.013983563520014286, + 0.01337785366922617, + 0.02254582941532135, + 0.01541423425078392, + 0.009144742973148823, + 0.005257888697087765, + 0.007006355095654726, + 0.005657597910612822, + 0.014432135969400406, + 0.01475747674703598, + 0.011324419640004635, + 0.012912401929497719, + 0.005746382288634777, + 0.00318757607601583, + 0.009070092812180519, + 0.003409086726605892, + 0.005511642899364233, + 0.006761352531611919, + 0.004319340921938419, + 0.005527573637664318, + 0.005083537194877863, + 0.004160955082625151, + 0.00446311617270112, + 0.0025645012501627207, + 0.006455031223595142, + 0.005279019474983215, + 0.006030268967151642, + 0.007463081739842892, + 0.006102824583649635, + 0.005353368353098631, + 0.007805536966770887, + 0.005019797012209892, + 0.01869911700487137, + 0.018204068765044212, + 0.012936953455209732, + 0.017903374508023262, + 0.00621882826089859, + 0.00496618403121829, + 0.006626808550208807, + 0.004057803191244602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.05494312569499016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_tensor([3919, 3919])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_tensor([3919, 3919])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e28714b7c4858a3362a6bc5eb7002b5a8e055878 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_tensor([3919, 3919])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning high-angle drone shot, gliding smoothly forward and slowly panning to the right, reveals a breathtaking spectacle. A dense, serpentine stream of thousands of runners flows like a vibrant river along the scenic coastal highway, a long human procession stretching for miles until it disappears around a distant, sun-bleached cape. The collective is a moving mosaic of color, the athletes clad in a dazzling array of high-performance gear: neon yellow vests, electric blue leggings, and hot pink shorts create a stark and beautiful contrast against the dark gray asphalt. To the right, the azure ocean gleams in the bright morning sun, gentle waves breaking into white foam on the rocky shoreline. The runners move with a unified, rhythmic pulse, their arms pumping and their feet tapping the pavement in a steady, determined rhythm. Their diverse faces blend with intense focus and shared struggle. The entire scene is bathed in crisp, clear light, casting long, dynamic shadows, creating an epic, vibrant atmosphere of human endurance and collective will.", + "seed": 699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_3919_seed_699_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2120330136269331, + "vlm_logprob": -0.2120330136269331, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.2120330136269331, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1422227919101715 + ], + "opposite_logprob_tokens": [ + -2.1422228813171387 + ], + "logprob_reward": -0.1422227919101715, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1422227919101715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11893061548471451, + 0.03383675962686539, + 0.059326108545064926, + 0.04164109751582146, + 0.025907345116138458, + 0.072252556681633, + 0.02384309656918049, + 0.45099055767059326, + 0.11878363788127899, + 0.054488230496644974 + ], + "credit_map": [ + 0.015013555996119976, + 0.01330519374459982, + 0.011411950923502445, + 0.014771547168493271, + 0.012336061336100101, + 0.021839888766407967, + 0.016900615766644478, + 0.013351812027394772, + 0.003207501256838441, + 0.003049166640266776, + 0.0034813075326383114, + 0.004033118952065706, + 0.004334155935794115, + 0.006555026862770319, + 0.0056086694821715355, + 0.0035678143613040447, + 0.007519885897636414, + 0.005338065791875124, + 0.005474549252539873, + 0.006521070376038551, + 0.0074188970029354095, + 0.010840353555977345, + 0.009495098143815994, + 0.006718190386891365, + 0.005342127289623022, + 0.002987627172842622, + 0.006600651890039444, + 0.004652811214327812, + 0.0028822761960327625, + 0.00487936194986105, + 0.006239024456590414, + 0.008057222701609135, + 0.0023021858651190996, + 0.001759712235070765, + 0.0027430569753050804, + 0.0025794561952352524, + 0.002525001298636198, + 0.005868434440344572, + 0.0052668023854494095, + 0.0028626995626837015, + 0.015078078955411911, + 0.005157147068530321, + 0.008417122066020966, + 0.004975742194801569, + 0.008765211328864098, + 0.011980387382209301, + 0.009265203028917313, + 0.008613667450845242, + 0.0043050977401435375, + 0.0023758080787956715, + 0.0026906069833785295, + 0.0031884934287518263, + 0.0020231218077242374, + 0.0028552303556352854, + 0.00292163691483438, + 0.003483101725578308, + 0.025062302127480507, + 0.012270585633814335, + 0.015382896177470684, + 0.02716558612883091, + 0.022693676874041557, + 0.030912138521671295, + 0.13313937187194824, + 0.18436400592327118, + 0.016878299415111542, + 0.007057692389935255, + 0.013059153221547604, + 0.015073719434440136, + 0.008815651759505272, + 0.017794450744986534, + 0.014533375389873981, + 0.0255713053047657, + 0.0056238980032503605, + 0.0022277343086898327, + 0.0029760447796434164, + 0.0027277180925011635, + 0.0034545573871582747, + 0.005350994877517223, + 0.012015058659017086, + 0.020112229511141777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.1422227919101715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996270686388016 + ], + "opposite_logprob_tokens": [ + -2.9299962520599365 + ], + "logprob_reward": -0.054996270686388016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996270686388016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12856636941432953, + 0.05701492354273796, + 0.04399768263101578, + 0.054886966943740845, + 0.02149459347128868, + 0.041681330651044846, + 0.033004045486450195, + 0.46207720041275024, + 0.12422763556241989, + 0.033049240708351135 + ], + "credit_map": [ + 0.006898189894855022, + 0.00658581405878067, + 0.011823714710772038, + 0.03013797290623188, + 0.014305253513157368, + 0.02342911809682846, + 0.02195768803358078, + 0.013428607024252415, + 0.007122428622096777, + 0.0032910953741520643, + 0.005503566935658455, + 0.010338501073420048, + 0.007153010461479425, + 0.006356911733746529, + 0.01047036238014698, + 0.006779038347303867, + 0.007404231000691652, + 0.005212603136897087, + 0.00360751710832119, + 0.004111980553716421, + 0.0047065094113349915, + 0.007384622935205698, + 0.006963582243770361, + 0.004606636706739664, + 0.00410386873409152, + 0.0040076314471662045, + 0.006473169662058353, + 0.0068038348108530045, + 0.008200841955840588, + 0.009182380512356758, + 0.008890356868505478, + 0.007224877364933491, + 0.004254578147083521, + 0.0029716654680669308, + 0.0022703055292367935, + 0.002526967553421855, + 0.0019839478190988302, + 0.0025496294256299734, + 0.002424391685053706, + 0.002513106679543853, + 0.005230448674410582, + 0.002767866710200906, + 0.004544262774288654, + 0.004349608905613422, + 0.004804563242942095, + 0.008816372603178024, + 0.005890684202313423, + 0.005277520976960659, + 0.00391134899109602, + 0.001979675842449069, + 0.0038835753221064806, + 0.004144126083701849, + 0.0035351174883544445, + 0.0032201516442000866, + 0.006357511971145868, + 0.005972534883767366, + 0.026401812210679054, + 0.012501083314418793, + 0.01529092900454998, + 0.0266305860131979, + 0.023524584248661995, + 0.03218318521976471, + 0.13740435242652893, + 0.1881406605243683, + 0.0198487751185894, + 0.00891937781125307, + 0.008278877474367619, + 0.009974172338843346, + 0.008172223344445229, + 0.02035735361278057, + 0.03223486989736557, + 0.01644197478890419, + 0.0033529242500662804, + 0.0024348495062440634, + 0.0025537603069096804, + 0.0025036116130650043, + 0.0023465133272111416, + 0.003352176630869508, + 0.0065160589292645454, + 0.009989341720938683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.054996270686388016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064027324318886 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064027324318886, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064027324318886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09987922757863998, + 0.05485190823674202, + 0.02429124526679516, + 0.03821582719683647, + 0.033129870891571045, + 0.04744264855980873, + 0.021971914917230606, + 0.5633178949356079, + 0.07644420862197876, + 0.04045528918504715 + ], + "credit_map": [ + 0.010736090131103992, + 0.008796671405434608, + 0.010456492193043232, + 0.017174653708934784, + 0.010112200863659382, + 0.013506283983588219, + 0.017950642853975296, + 0.011146193370223045, + 0.009573137387633324, + 0.005676622036844492, + 0.005671944469213486, + 0.0048255654983222485, + 0.008741166442632675, + 0.007821924984455109, + 0.00677567720413208, + 0.005765870679169893, + 0.00469758827239275, + 0.003559091128408909, + 0.002263121772557497, + 0.0026431262958794832, + 0.0023799226619303226, + 0.002663311781361699, + 0.0032553679775446653, + 0.0028297167737036943, + 0.003808606881648302, + 0.002580871805548668, + 0.006819171365350485, + 0.0043071904219686985, + 0.004885930567979813, + 0.0054458980448544025, + 0.005542883649468422, + 0.004825272597372532, + 0.004535783547908068, + 0.0026619965210556984, + 0.0036667983513325453, + 0.0034593993332237005, + 0.003595635760575533, + 0.004375495947897434, + 0.006120534613728523, + 0.004714226350188255, + 0.009362753480672836, + 0.003770385403186083, + 0.004936865530908108, + 0.0042520202696323395, + 0.0055084386840462685, + 0.007079411763697863, + 0.006049076095223427, + 0.006483697798103094, + 0.002333074575290084, + 0.001245893887244165, + 0.002716143848374486, + 0.0035230012144893408, + 0.0020108441822230816, + 0.00267393933609128, + 0.004423439037054777, + 0.0030455784872174263, + 0.03132239356637001, + 0.014461120590567589, + 0.018390769138932228, + 0.032028067857027054, + 0.028026951476931572, + 0.039249226450920105, + 0.16731329262256622, + 0.23252609372138977, + 0.007438286207616329, + 0.003023154102265835, + 0.0035489534493535757, + 0.005977334454655647, + 0.0037905576173216105, + 0.00860699824988842, + 0.02505660429596901, + 0.019002316519618034, + 0.0033626637887209654, + 0.0016507034888491035, + 0.002358678262680769, + 0.0032043501269072294, + 0.0030860176775604486, + 0.003899794537574053, + 0.00927069503813982, + 0.013622385449707508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.016064027324318886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578755617141724 + ], + "opposite_logprob_tokens": [ + -0.6328755617141724 + ], + "logprob_reward": -0.7578755617141724, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578755617141724, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12887834012508392, + 0.047294050455093384, + 0.05604265630245209, + 0.1978771984577179, + 0.07834134250879288, + 0.08165396749973297, + 0.11773601174354553, + 0.05139981210231781, + 0.15335379540920258, + 0.08742281794548035 + ], + "credit_map": [ + 0.03131987154483795, + 0.018003715202212334, + 0.010977906174957752, + 0.013919533230364323, + 0.009824859909713268, + 0.012601767666637897, + 0.014943988993763924, + 0.017286689952015877, + 0.005934521555900574, + 0.004876445047557354, + 0.005702266003936529, + 0.005856830161064863, + 0.0049119750037789345, + 0.007777176797389984, + 0.007049937732517719, + 0.005184900015592575, + 0.014995702542364597, + 0.009026396088302135, + 0.0033148517832159996, + 0.003975772764533758, + 0.006054675672203302, + 0.005785336252301931, + 0.006559227593243122, + 0.006330693606287241, + 0.019314495846629143, + 0.013996141962707043, + 0.030436484143137932, + 0.029067708179354668, + 0.018512820824980736, + 0.023979758843779564, + 0.03426603972911835, + 0.028303755447268486, + 0.009264592081308365, + 0.0062287128530442715, + 0.007237997837364674, + 0.008863614872097969, + 0.008348257280886173, + 0.012769083492457867, + 0.014679639600217342, + 0.010949444957077503, + 0.011832634918391705, + 0.005670617800205946, + 0.008985019288957119, + 0.006751570384949446, + 0.011082402430474758, + 0.01705777272582054, + 0.010770647786557674, + 0.009503297507762909, + 0.007076996378600597, + 0.005199873354285955, + 0.0123207438737154, + 0.019854595884680748, + 0.011149496771395206, + 0.01571938954293728, + 0.030592910945415497, + 0.01582200452685356, + 0.003921097610145807, + 0.0020294494461268187, + 0.00241044326685369, + 0.004627547226846218, + 0.002758560236543417, + 0.003887638682499528, + 0.013356570154428482, + 0.018408505246043205, + 0.016891323029994965, + 0.0060475049540400505, + 0.006816027220338583, + 0.011527265422046185, + 0.007464769762009382, + 0.01814792864024639, + 0.05147790536284447, + 0.034981075674295425, + 0.006930049508810043, + 0.003251874353736639, + 0.004550084006041288, + 0.005891298875212669, + 0.004963252693414688, + 0.007569923996925354, + 0.023365728557109833, + 0.030900603160262108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.7578755617141724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900641649961472 + ], + "opposite_logprob_tokens": [ + -2.4640064239501953 + ], + "logprob_reward": -0.08900641649961472, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900641649961472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13908334076404572, + 0.12453324347734451, + 0.05791323632001877, + 0.09825137257575989, + 0.021947892382740974, + 0.014519821852445602, + 0.01701209507882595, + 0.31430283188819885, + 0.18478968739509583, + 0.02764645218849182 + ], + "credit_map": [ + 0.022874606773257256, + 0.01505354419350624, + 0.013022217899560928, + 0.019285572692751884, + 0.016478730365633965, + 0.020889416337013245, + 0.015444506891071796, + 0.016034770756959915, + 0.02432369627058506, + 0.01755836233496666, + 0.010756881907582283, + 0.011020739562809467, + 0.017650820314884186, + 0.015929358080029488, + 0.013492029160261154, + 0.013801363296806812, + 0.006526163313537836, + 0.004989321809262037, + 0.006999558303505182, + 0.006860441528260708, + 0.007245114538818598, + 0.007756104692816734, + 0.010075454600155354, + 0.007461079861968756, + 0.010309165343642235, + 0.007189885247498751, + 0.009551028721034527, + 0.008930583484470844, + 0.008599049411714077, + 0.01912495493888855, + 0.018251854926347733, + 0.01629485934972763, + 0.0033472063951194286, + 0.002108969958499074, + 0.0020860345102846622, + 0.0024551909882575274, + 0.002639800775796175, + 0.0026626416947692633, + 0.003732240991666913, + 0.002915810327976942, + 0.0015741349197924137, + 0.0009986533550545573, + 0.001299970201216638, + 0.0010872854618355632, + 0.0020420150831341743, + 0.0028208300936967134, + 0.0025638488586992025, + 0.0021330860909074545, + 0.002016194397583604, + 0.001333695137873292, + 0.001218646066263318, + 0.0020445873960852623, + 0.0016485988162457943, + 0.0030014135409146547, + 0.003300709882751107, + 0.0024482510052621365, + 0.02463667094707489, + 0.012816522270441055, + 0.013174613006412983, + 0.01830967701971531, + 0.01727229915559292, + 0.020155202597379684, + 0.08541339635848999, + 0.12252448499202728, + 0.017160428687930107, + 0.007028838153928518, + 0.007854548282921314, + 0.013060279190540314, + 0.009464473463594913, + 0.020646417513489723, + 0.0629768967628479, + 0.04659781605005264, + 0.0028816252015531063, + 0.002261609537526965, + 0.0019298199331387877, + 0.0014756264863535762, + 0.0024929067585617304, + 0.004119617864489555, + 0.004481132607907057, + 0.008004115894436836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.08900641649961472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "stunning high-angle drone shot, gliding smoothly forward and slowly panning to the right, reveals a breathtaking spectacle. A dense, serpentine stream of thousands of runners flows like a vibrant river along the scenic coastal highway, a long human procession stretching for miles until it disappears around a distant, sun-bleached cape. The collective is a moving mosaic of color, the athletes clad in a dazzling array of high-performance gear: neon yellow vests, electric blue leggings, and hot pink shorts create a stark and beautiful contrast against the dark gray asphalt. To the right, the azure ocean gleams in the bright morning sun, gentle waves breaking into white foam on the rocky shoreline. The runners move with a unified, rhythmic pulse, their arms pumping and their feet tapping the pavement in a steady, determined rhythm. Their diverse faces blend with intense focus and shared struggle. The entire scene is bathed in crisp, clear light, casting long, dynamic shadows, creating an epic, vibrant atmosphere of human endurance and collective will.", + "seed": 700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_3919_seed_700_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.175861132517457, + "vlm_logprob": -0.175861132517457, + "vlm_yes_no_margin": 2.474999952316284, + "vlm_token_credit": -0.175861132517457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13248133659362793 + ], + "opposite_logprob_tokens": [ + -2.257481336593628 + ], + "logprob_reward": -0.13248133659362793, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13248133659362793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30456098914146423, + 0.04715840891003609, + 0.15961042046546936, + 0.12612475454807281, + 0.1564771980047226, + 0.0813646987080574, + 0.01665313169360161, + 0.017806662246584892, + 0.04208166524767876, + 0.048162102699279785 + ], + "credit_map": [ + 0.03810857981443405, + 0.032026540488004684, + 0.013568238355219364, + 0.08849979937076569, + 0.03617379814386368, + 0.03821342810988426, + 0.03372162580490112, + 0.024248965084552765, + 0.006988899782299995, + 0.007674145977944136, + 0.006239501293748617, + 0.005789461079984903, + 0.006474134512245655, + 0.0038010498974472284, + 0.005488299764692783, + 0.004702911712229252, + 0.019722221419215202, + 0.023051921278238297, + 0.010593621991574764, + 0.03956042230129242, + 0.01580437831580639, + 0.014147618785500526, + 0.026334986090660095, + 0.010395249351859093, + 0.017302902415394783, + 0.013868358917534351, + 0.00799514539539814, + 0.011419111862778664, + 0.009472516365349293, + 0.015134995803236961, + 0.03849012777209282, + 0.012441584840416908, + 0.01686604879796505, + 0.011021796613931656, + 0.01099560409784317, + 0.008868389762938023, + 0.03487299382686615, + 0.05052638426423073, + 0.010929150506854057, + 0.012396832928061485, + 0.007301296107470989, + 0.006796001456677914, + 0.005019853822886944, + 0.013702398166060448, + 0.005734513979405165, + 0.00748762022703886, + 0.028053585439920425, + 0.007269425317645073, + 0.0019768872298300266, + 0.0025907724630087614, + 0.002389098284766078, + 0.0018963062902912498, + 0.0017357220640406013, + 0.001821689773350954, + 0.0022301673889160156, + 0.002012487966567278, + 0.001762968022376299, + 0.0034888717345893383, + 0.00151406682562083, + 0.002036899095401168, + 0.0021256094332784414, + 0.0021471858490258455, + 0.0027526479680091143, + 0.0019784134346991777, + 0.006990932859480381, + 0.004920376930385828, + 0.0036246280651539564, + 0.003327558282762766, + 0.007229863200336695, + 0.004960262216627598, + 0.006780749186873436, + 0.004247294273227453, + 0.006429252214729786, + 0.0039373235777020454, + 0.003331574844196439, + 0.003449651412665844, + 0.007015020586550236, + 0.004664717707782984, + 0.01300154346972704, + 0.006333017721772194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.13248133659362793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062088727951049805 + ], + "opposite_logprob_tokens": [ + -2.81208872795105 + ], + "logprob_reward": -0.062088727951049805, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062088727951049805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5429313778877258, + 0.054284386336803436, + 0.057569462805986404, + 0.061313696205616, + 0.05951326712965965, + 0.0335545614361763, + 0.07313649356365204, + 0.0293829794973135, + 0.03079098090529442, + 0.05752282589673996 + ], + "credit_map": [ + 0.07040295004844666, + 0.028555307537317276, + 0.02550012245774269, + 0.18251903355121613, + 0.07017438858747482, + 0.07028312981128693, + 0.053403306752443314, + 0.04209311306476593, + 0.00794420950114727, + 0.007542043458670378, + 0.006695571355521679, + 0.00533534400165081, + 0.008929828181862831, + 0.005282560829073191, + 0.007327013183385134, + 0.005227812100201845, + 0.0077695418149232864, + 0.00766319315880537, + 0.004032306373119354, + 0.006451715249568224, + 0.011418956331908703, + 0.007520077284425497, + 0.008545753546059132, + 0.00416791345924139, + 0.007917401380836964, + 0.005223663058131933, + 0.0036315899342298508, + 0.004771000239998102, + 0.007644034456461668, + 0.009364718571305275, + 0.016982953995466232, + 0.005778332706540823, + 0.009409157559275627, + 0.003722202265635133, + 0.002990968991070986, + 0.01155192218720913, + 0.008831430226564407, + 0.012574495747685432, + 0.005453776568174362, + 0.004979304503649473, + 0.005022265017032623, + 0.0032720824237912893, + 0.0030846167355775833, + 0.005096740555018187, + 0.004614105448126793, + 0.0028001933824270964, + 0.0058294422924518585, + 0.0038351125549525023, + 0.00871283933520317, + 0.015286210924386978, + 0.008270500227808952, + 0.0076860529370605946, + 0.009424007497727871, + 0.007058208342641592, + 0.008680184371769428, + 0.008018492721021175, + 0.004921664483845234, + 0.0023862640373408794, + 0.003891788190230727, + 0.0031748146284371614, + 0.0027649824041873217, + 0.002845285926014185, + 0.0049056424759328365, + 0.004492535721510649, + 0.005781067069619894, + 0.002699409844353795, + 0.002800870221108198, + 0.004063772968947887, + 0.004682105500251055, + 0.0023568409960716963, + 0.00481627881526947, + 0.003590630367398262, + 0.007221422623842955, + 0.0045518032275140285, + 0.003831971436738968, + 0.00448831869289279, + 0.008607318624854088, + 0.004858050961047411, + 0.016453824937343597, + 0.007510107476264238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.062088727951049805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336914137005806 + ], + "opposite_logprob_tokens": [ + -3.40869140625 + ], + "logprob_reward": -0.0336914137005806, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336914137005806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7630236744880676, + 0.01989552192389965, + 0.01954474486410618, + 0.036576226353645325, + 0.06170433387160301, + 0.01809726282954216, + 0.031098511070013046, + 0.004929297138005495, + 0.03719937801361084, + 0.007931055501103401 + ], + "credit_map": [ + 0.09803860634565353, + 0.036023713648319244, + 0.036372408270835876, + 0.2553829848766327, + 0.10125512629747391, + 0.10196980834007263, + 0.07103246450424194, + 0.06294852495193481, + 0.0027935707475990057, + 0.0039021018892526627, + 0.001638947520405054, + 0.002176759298890829, + 0.0024901588913053274, + 0.0021954928524792194, + 0.002582271583378315, + 0.0021162156481295824, + 0.0025193365290760994, + 0.0026275734417140484, + 0.0012085060589015484, + 0.002911279909312725, + 0.0031080003827810287, + 0.0022023722995072603, + 0.003223823383450508, + 0.001743850763887167, + 0.005304683931171894, + 0.0028968341648578644, + 0.0022163772955536842, + 0.0027361996471881866, + 0.004064450040459633, + 0.00633508013561368, + 0.01004851795732975, + 0.0029740810859948397, + 0.00831675436347723, + 0.003686492098495364, + 0.005848705302923918, + 0.021413510665297508, + 0.005938921123743057, + 0.003724604146555066, + 0.0074499426409602165, + 0.005325398873537779, + 0.0019537173211574554, + 0.002072911011055112, + 0.001351997721940279, + 0.004662694875150919, + 0.001063046045601368, + 0.0010155346244573593, + 0.004343738313764334, + 0.0016336216358467937, + 0.003544930135831237, + 0.008675099350512028, + 0.003608307335525751, + 0.0033618260640650988, + 0.003341088769957423, + 0.002482396550476551, + 0.003728028852492571, + 0.0023568335454910994, + 0.0006577146705240011, + 0.0006797733367420733, + 0.00047186718438751996, + 0.0006207539699971676, + 0.0007111324812285602, + 0.0005688329110853374, + 0.0007266978500410914, + 0.000492524413857609, + 0.004341701045632362, + 0.0028910478577017784, + 0.00256349123083055, + 0.003274590475484729, + 0.006294302176684141, + 0.003886075224727392, + 0.00903295073658228, + 0.004915216937661171, + 0.001257567317225039, + 0.0005299199838191271, + 0.0004927348927594721, + 0.0007368635851889849, + 0.0013152703177183867, + 0.0006654910976067185, + 0.0010897355386987329, + 0.0018434720113873482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0336914137005806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328717470169067 + ], + "opposite_logprob_tokens": [ + -0.7578717470169067 + ], + "logprob_reward": -0.6328717470169067, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328717470169067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5668454170227051, + 0.018362753093242645, + 0.0709056407213211, + 0.03284309804439545, + 0.038531288504600525, + 0.048059262335300446, + 0.06709231436252594, + 0.014817755669355392, + 0.04169346019625664, + 0.10084900259971619 + ], + "credit_map": [ + 0.07657400518655777, + 0.033283673226833344, + 0.027624068781733513, + 0.19076980650424957, + 0.07283305376768112, + 0.06821204721927643, + 0.05237418785691261, + 0.04517453908920288, + 0.0028444258496165276, + 0.004296815488487482, + 0.0012077669380232692, + 0.0018186423694714904, + 0.002790441969409585, + 0.001676278654485941, + 0.001891524181701243, + 0.0018368574092164636, + 0.013307273387908936, + 0.009241186082363129, + 0.005021636839956045, + 0.010114452801644802, + 0.009751955047249794, + 0.007440345827490091, + 0.008432413451373577, + 0.0075963824056088924, + 0.0035693703684955835, + 0.0025104465894401073, + 0.0020786873064935207, + 0.0025207148864865303, + 0.004052290227264166, + 0.005834807176142931, + 0.009598048403859138, + 0.0026787323877215385, + 0.007237495854496956, + 0.0031094099394977093, + 0.002837774809449911, + 0.002889457857236266, + 0.006782270036637783, + 0.005322191398590803, + 0.004013619385659695, + 0.006339069921523333, + 0.0044446540996432304, + 0.005678740795701742, + 0.0027351640164852142, + 0.008991964161396027, + 0.003725233254954219, + 0.0037788504268974066, + 0.014680342748761177, + 0.004024313762784004, + 0.00837503932416439, + 0.017070669680833817, + 0.009437819011509418, + 0.007629959844052792, + 0.006591953337192535, + 0.005158216692507267, + 0.006958723068237305, + 0.005869934801012278, + 0.0018600292969495058, + 0.0016136839985847473, + 0.001594302710145712, + 0.0016307099722325802, + 0.0020956823136657476, + 0.001301631098613143, + 0.0023208577185869217, + 0.002400859259068966, + 0.004596428945660591, + 0.003918681759387255, + 0.004367005079984665, + 0.004529967904090881, + 0.004997933283448219, + 0.003900931915268302, + 0.010374794714152813, + 0.005007716361433268, + 0.01127968542277813, + 0.00717886071652174, + 0.006870453245937824, + 0.007977904751896858, + 0.014083636924624443, + 0.009494715370237827, + 0.03148287907242775, + 0.012480868957936764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.6328717470169067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018172437325119972 + ], + "opposite_logprob_tokens": [ + -4.018172264099121 + ], + "logprob_reward": -0.018172437325119972, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018172437325119972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3881424069404602, + 0.04994569346308708, + 0.026915820315480232, + 0.0861661359667778, + 0.024627063423395157, + 0.32826486229896545, + 0.031549375504255295, + 0.009065326303243637, + 0.02522258460521698, + 0.030100731179118156 + ], + "credit_map": [ + 0.05061300843954086, + 0.022989235818386078, + 0.018109066411852837, + 0.12913256883621216, + 0.05077332258224487, + 0.04948818311095238, + 0.03598834574222565, + 0.031048700213432312, + 0.008685414679348469, + 0.00862777978181839, + 0.004042574670165777, + 0.004071309231221676, + 0.008690726943314075, + 0.004686971660703421, + 0.0064435615204274654, + 0.00469735823571682, + 0.003753596916794777, + 0.004104595631361008, + 0.002408724743872881, + 0.0039569721557199955, + 0.004112420603632927, + 0.0027857047971338034, + 0.003732529003173113, + 0.0020612748339772224, + 0.009199919179081917, + 0.005746827460825443, + 0.005700225476175547, + 0.006999824661761522, + 0.008799589239060879, + 0.01400692854076624, + 0.026380697265267372, + 0.009332124143838882, + 0.004001465160399675, + 0.0029410228598862886, + 0.0023252239916473627, + 0.0025827016215771437, + 0.004123913124203682, + 0.0035967945586889982, + 0.0024894471280276775, + 0.0025664945133030415, + 0.035840198397636414, + 0.03078615665435791, + 0.01992984116077423, + 0.025858046486973763, + 0.034012340009212494, + 0.04425031319260597, + 0.09970999509096146, + 0.03787796199321747, + 0.004518744070082903, + 0.005743679124861956, + 0.0025029166135936975, + 0.003817654913291335, + 0.00372308655641973, + 0.002854001708328724, + 0.004245844203978777, + 0.004143448080867529, + 0.0011191293597221375, + 0.0008292535203509033, + 0.000891612668056041, + 0.0006794802029617131, + 0.0009741563117131591, + 0.0013942958321422338, + 0.001997738378122449, + 0.0011796598555520177, + 0.0035203828010708094, + 0.002753194887191057, + 0.002236118307337165, + 0.00295509141869843, + 0.003103644121438265, + 0.0021391904447227716, + 0.004931286443024874, + 0.0035836773458868265, + 0.0031546736136078835, + 0.0025297561660408974, + 0.0029257116839289665, + 0.0028066812083125114, + 0.005412144586443901, + 0.0025721131823956966, + 0.006924033630639315, + 0.0037756161764264107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.018172437325119972 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle drone shot, gliding smoothly forward and slowly panning to the right, reveals a breathtaking spectacle. A dense, serpentine stream of thousands of runners flows like a vibrant river along the scenic coastal highway, a long human procession stretching for miles until it disappears around a distant, sun-bleached cape. The collective is a moving mosaic of color, the athletes clad in a dazzling array of high-performance gear: neon yellow vests, electric blue leggings, and hot pink shorts create a stark and beautiful contrast against the dark gray asphalt. To the right, the azure ocean gleams in the bright morning sun, gentle waves breaking into white foam on the rocky shoreline. The runners move with a unified, rhythmic pulse, their arms pumping and their feet tapping the pavement in a steady, determined rhythm. Their diverse faces blend with intense focus and shared struggle. The entire scene is bathed in crisp, clear light, casting long, dynamic shadows, creating an epic, vibrant atmosphere of human endurance and collective will.", + "seed": 701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_3919_seed_701_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1550087271258235, + "vlm_logprob": -0.1550087271258235, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.1550087271258235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0838075801730156 + ], + "opposite_logprob_tokens": [ + -2.7088074684143066 + ], + "logprob_reward": -0.0838075801730156, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0838075801730156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12987478077411652, + 0.05836138874292374, + 0.057997435331344604, + 0.10545630007982254, + 0.15465416014194489, + 0.13869759440422058, + 0.05710786208510399, + 0.06313908100128174, + 0.2056588977575302, + 0.029052508994936943 + ], + "credit_map": [ + 0.01663241535425186, + 0.01029424462467432, + 0.022316470742225647, + 0.017157604917883873, + 0.012346401810646057, + 0.0098867891356349, + 0.0250875111669302, + 0.016153346747159958, + 0.00825646985322237, + 0.003962154500186443, + 0.015067599713802338, + 0.004998119547963142, + 0.004817420616745949, + 0.004132786300033331, + 0.012921284884214401, + 0.004205555655062199, + 0.005441766697913408, + 0.006746451836079359, + 0.012192048132419586, + 0.009404659271240234, + 0.009625574573874474, + 0.006369324401021004, + 0.004167912993580103, + 0.004049698822200298, + 0.0165582075715065, + 0.00896476861089468, + 0.018545959144830704, + 0.01570582389831543, + 0.010137489065527916, + 0.011535437777638435, + 0.012127849273383617, + 0.011880752630531788, + 0.012298007495701313, + 0.012523704208433628, + 0.025547539815306664, + 0.045113757252693176, + 0.010350917465984821, + 0.02506326511502266, + 0.014425690285861492, + 0.00933128409087658, + 0.022839656099677086, + 0.006731439847499132, + 0.04357224702835083, + 0.0158836767077446, + 0.011525873094797134, + 0.020530229434370995, + 0.010610976256430149, + 0.007003497797995806, + 0.006457495037466288, + 0.007703201845288277, + 0.0068344539031386375, + 0.011158337816596031, + 0.0076714204624295235, + 0.005048444494605064, + 0.006975758355110884, + 0.0052587492391467094, + 0.006310173310339451, + 0.006326335482299328, + 0.009165511466562748, + 0.01817876100540161, + 0.004564633592963219, + 0.004945090971887112, + 0.006108251865953207, + 0.00754032377153635, + 0.029744146391749382, + 0.006524608004838228, + 0.07425680011510849, + 0.023907296359539032, + 0.015220856294035912, + 0.027555381879210472, + 0.011055552400648594, + 0.017394250258803368, + 0.005148204043507576, + 0.0036428524181246758, + 0.004136959556490183, + 0.004282075446099043, + 0.0025008819065988064, + 0.0022684598807245493, + 0.004661291837692261, + 0.0024117843713611364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.0838075801730156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337250716984272 + ], + "opposite_logprob_tokens": [ + -3.4087250232696533 + ], + "logprob_reward": -0.0337250716984272, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337250716984272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10629772394895554, + 0.0370759516954422, + 0.06949316710233688, + 0.04986689239740372, + 0.1473228633403778, + 0.10114032030105591, + 0.03257077932357788, + 0.03842514753341675, + 0.38788902759552, + 0.029918242245912552 + ], + "credit_map": [ + 0.013715721666812897, + 0.008188081905245781, + 0.013994651846587658, + 0.02015645243227482, + 0.012378661893308163, + 0.012166813015937805, + 0.013395058922469616, + 0.012302268296480179, + 0.007699309382587671, + 0.003443544963374734, + 0.008086640387773514, + 0.004465204197913408, + 0.0036352567840367556, + 0.003024469595402479, + 0.00395249854773283, + 0.0027690285351127386, + 0.008245494216680527, + 0.00795406848192215, + 0.010658209212124348, + 0.012431944720447063, + 0.0067437076941132545, + 0.008782359771430492, + 0.008278273977339268, + 0.006399098318070173, + 0.00553704472258687, + 0.004770203493535519, + 0.009294996038079262, + 0.0076189469546079636, + 0.0045229229144752026, + 0.00691461656242609, + 0.006658680737018585, + 0.004549476318061352, + 0.010474291630089283, + 0.009041761979460716, + 0.02190137654542923, + 0.05065319687128067, + 0.00938606821000576, + 0.02269943803548813, + 0.012606465257704258, + 0.01056024618446827, + 0.01800459437072277, + 0.009850885719060898, + 0.023955032229423523, + 0.011047505773603916, + 0.008137657307088375, + 0.009093355387449265, + 0.01104175578802824, + 0.01000952161848545, + 0.004860205575823784, + 0.0051949117332696915, + 0.004797241650521755, + 0.0046532368287444115, + 0.003231662791222334, + 0.002295296872034669, + 0.0039867558516561985, + 0.00355146499350667, + 0.00455334410071373, + 0.004803003277629614, + 0.00548796821385622, + 0.00897961389273405, + 0.003032525535672903, + 0.0025861395988613367, + 0.0034559613559395075, + 0.005526591092348099, + 0.05586567893624306, + 0.008269374258816242, + 0.14278675615787506, + 0.04206184670329094, + 0.031205445528030396, + 0.06262049823999405, + 0.01617569848895073, + 0.028903665021061897, + 0.004298779647797346, + 0.0030387393198907375, + 0.0038781356997787952, + 0.004742548801004887, + 0.0033403283450752497, + 0.0025613203179091215, + 0.004099883604794741, + 0.003958502784371376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.0337250716984272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575299859046936 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575299859046936, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575299859046936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08981768041849136, + 0.04041203856468201, + 0.23895515501499176, + 0.059421174228191376, + 0.053055018186569214, + 0.06671477109193802, + 0.014042350463569164, + 0.023859785869717598, + 0.40072575211524963, + 0.012996249832212925 + ], + "credit_map": [ + 0.020009636878967285, + 0.008213034830987453, + 0.01109166257083416, + 0.011123927310109138, + 0.008338410407304764, + 0.007252483628690243, + 0.014189240522682667, + 0.009599289856851101, + 0.009593719616532326, + 0.0051960148848593235, + 0.005238322075456381, + 0.00483061233535409, + 0.004625589586794376, + 0.0033556558191776276, + 0.004777453374117613, + 0.0027946732006967068, + 0.02187742479145527, + 0.032022010535001755, + 0.033129557967185974, + 0.04386185109615326, + 0.043127238750457764, + 0.025756116956472397, + 0.024025404825806618, + 0.015155559405684471, + 0.008300760760903358, + 0.006801325362175703, + 0.009073491208255291, + 0.010258974507451057, + 0.0052795265801250935, + 0.008946380577981472, + 0.006809064652770758, + 0.003951656632125378, + 0.004132518079131842, + 0.003964496776461601, + 0.008059794083237648, + 0.016359569504857063, + 0.0037520197220146656, + 0.0077655925415456295, + 0.005251126829534769, + 0.0037699053063988686, + 0.009455664083361626, + 0.006101177539676428, + 0.014553921297192574, + 0.008595947176218033, + 0.006463696714490652, + 0.005667516496032476, + 0.008869425393640995, + 0.007007427513599396, + 0.002512669190764427, + 0.0018288780702278018, + 0.001962163718417287, + 0.0022068643011152744, + 0.001043658354319632, + 0.001043759984895587, + 0.0017122504068538547, + 0.0017321070190519094, + 0.00256603816524148, + 0.0023672953248023987, + 0.003524268977344036, + 0.003688505617901683, + 0.0033710929565131664, + 0.002994992770254612, + 0.002184278564527631, + 0.0031633165199309587, + 0.058008771389722824, + 0.008786983788013458, + 0.14793887734413147, + 0.043450403958559036, + 0.0313984639942646, + 0.06548569351434708, + 0.017664406448602676, + 0.027992166578769684, + 0.0014637602725997567, + 0.0013725378084927797, + 0.0016668308526277542, + 0.002046386245638132, + 0.001443002955056727, + 0.0017533369828015566, + 0.0016407616203650832, + 0.00160963402595371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.020575299859046936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328465938568115 + ], + "opposite_logprob_tokens": [ + -0.7578465938568115 + ], + "logprob_reward": -0.6328465938568115, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328465938568115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24708020687103271, + 0.12910690903663635, + 0.08939331769943237, + 0.05773454159498215, + 0.03306462615728378, + 0.07033687829971313, + 0.09848020225763321, + 0.04393843188881874, + 0.2155260592699051, + 0.015338904224336147 + ], + "credit_map": [ + 0.048235733062028885, + 0.01627206802368164, + 0.0690799430012703, + 0.024917956441640854, + 0.013037767261266708, + 0.015291500836610794, + 0.03589661046862602, + 0.02434861660003662, + 0.019703123718500137, + 0.014807396568357944, + 0.025177868083119392, + 0.017791936174035072, + 0.00933084450662136, + 0.007504566572606564, + 0.022308774292469025, + 0.01248238980770111, + 0.010579603724181652, + 0.009439629502594471, + 0.014074388891458511, + 0.014099564403295517, + 0.014928244054317474, + 0.008710958994925022, + 0.008991855196654797, + 0.008569076657295227, + 0.009188841097056866, + 0.00814310647547245, + 0.010017597116529942, + 0.010227455757558346, + 0.004593484103679657, + 0.005269632209092379, + 0.005795496050268412, + 0.004498927854001522, + 0.0036981473676860332, + 0.0035616792738437653, + 0.004338672384619713, + 0.005883586592972279, + 0.004663089290261269, + 0.003293682588264346, + 0.00410856818780303, + 0.003517200704663992, + 0.008697166107594967, + 0.007332290057092905, + 0.012749589048326015, + 0.010086427442729473, + 0.0077019645832479, + 0.0083932401612401, + 0.007971269078552723, + 0.007404933217912912, + 0.016645515337586403, + 0.007185366936028004, + 0.023567168042063713, + 0.013036925345659256, + 0.00785867776721716, + 0.006123275030404329, + 0.014241103082895279, + 0.00982216838747263, + 0.005408552009612322, + 0.003995319828391075, + 0.007909915409982204, + 0.006963124964386225, + 0.0051098973490297794, + 0.0062911370769143105, + 0.003781045088544488, + 0.004479442723095417, + 0.0323995016515255, + 0.005494436249136925, + 0.07952789962291718, + 0.02448042295873165, + 0.015553855337202549, + 0.03068212792277336, + 0.009929501451551914, + 0.01745830848813057, + 0.0016965122194960713, + 0.0013831036631017923, + 0.001811814378015697, + 0.0023887930437922478, + 0.0017436520429328084, + 0.0019076167372986674, + 0.002449204446747899, + 0.0019582060631364584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.6328465938568115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004089090041816235 + ], + "opposite_logprob_tokens": [ + -5.504088878631592 + ], + "logprob_reward": -0.004089090041816235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004089090041816235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05222383514046669, + 0.012920775450766087, + 0.018967123702168465, + 0.016615888103842735, + 0.17678120732307434, + 0.08565262705087662, + 0.01864388771355152, + 0.026045434176921844, + 0.5790449976921082, + 0.013104172423481941 + ], + "credit_map": [ + 0.009590841829776764, + 0.0032090668100863695, + 0.01734771393239498, + 0.006498188711702824, + 0.0023869210854172707, + 0.004089082591235638, + 0.005026942119002342, + 0.0040750824846327305, + 0.002330781426280737, + 0.0010620680404826999, + 0.00367173389531672, + 0.0019409334054216743, + 0.0005413435865193605, + 0.0006512427935376763, + 0.0018284559482708573, + 0.0008942175772972405, + 0.0027948692440986633, + 0.0013607090804725885, + 0.003590703010559082, + 0.0034583094529807568, + 0.0013656455557793379, + 0.001928801299072802, + 0.0027537825517356396, + 0.0017143049044534564, + 0.0023079721722751856, + 0.0015300383092835546, + 0.0021634986624121666, + 0.002588131697848439, + 0.0029808711260557175, + 0.0018275505863130093, + 0.0017423474928364158, + 0.001475479337386787, + 0.011611877009272575, + 0.009684695862233639, + 0.026080965995788574, + 0.06475499272346497, + 0.010533415712416172, + 0.028663646429777145, + 0.01448407955467701, + 0.010967545211315155, + 0.018491534516215324, + 0.005347344093024731, + 0.02559150941669941, + 0.009401029907166958, + 0.006103635765612125, + 0.007678474765270948, + 0.007465799804776907, + 0.005573307164013386, + 0.0034119165502488613, + 0.002393313217908144, + 0.003098965622484684, + 0.002726638223975897, + 0.0021401136182248592, + 0.0016705975867807865, + 0.001827907282859087, + 0.0013744377065449953, + 0.00269952486269176, + 0.0019803133327513933, + 0.004068762995302677, + 0.0028161469381302595, + 0.004671802744269371, + 0.00334983691573143, + 0.0024612341076135635, + 0.003997813444584608, + 0.08448906242847443, + 0.012316704727709293, + 0.21494650840759277, + 0.06274165213108063, + 0.045964475721120834, + 0.09155569225549698, + 0.025347348302602768, + 0.041683629155159, + 0.0014415516052395105, + 0.0013383292825892568, + 0.0017094797221943736, + 0.002027781680226326, + 0.0014582712901756167, + 0.0016284555895254016, + 0.002091825008392334, + 0.0014084792928770185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.004089090041816235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle drone shot, gliding smoothly forward and slowly panning to the right, reveals a breathtaking spectacle. A dense, serpentine stream of thousands of runners flows like a vibrant river along the scenic coastal highway, a long human procession stretching for miles until it disappears around a distant, sun-bleached cape. The collective is a moving mosaic of color, the athletes clad in a dazzling array of high-performance gear: neon yellow vests, electric blue leggings, and hot pink shorts create a stark and beautiful contrast against the dark gray asphalt. To the right, the azure ocean gleams in the bright morning sun, gentle waves breaking into white foam on the rocky shoreline. The runners move with a unified, rhythmic pulse, their arms pumping and their feet tapping the pavement in a steady, determined rhythm. Their diverse faces blend with intense focus and shared struggle. The entire scene is bathed in crisp, clear light, casting long, dynamic shadows, creating an epic, vibrant atmosphere of human endurance and collective will.", + "seed": 702, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_13/index_3919_seed_702_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1526834571734071, + "vlm_logprob": -0.1526834571734071, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.1526834571734071, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1176810935139656 + ], + "opposite_logprob_tokens": [ + -2.3676810264587402 + ], + "logprob_reward": -0.1176810935139656, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1176810935139656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05844837427139282, + 0.0359438955783844, + 0.03472510352730751, + 0.038215819746255875, + 0.09721283614635468, + 0.04456595703959465, + 0.0669313371181488, + 0.044365979731082916, + 0.1246490329504013, + 0.45494166016578674 + ], + "credit_map": [ + 0.007289101369678974, + 0.01184066478163004, + 0.007585818413645029, + 0.006992105394601822, + 0.005726449657231569, + 0.0058017089031636715, + 0.007130097597837448, + 0.0060824258252978325, + 0.005256211385130882, + 0.003845278639346361, + 0.004417204763740301, + 0.0042188516817986965, + 0.003840170567855239, + 0.0033730517607182264, + 0.006181703880429268, + 0.004811421502381563, + 0.0032036835327744484, + 0.004673277959227562, + 0.005211309064179659, + 0.002724711550399661, + 0.006970668211579323, + 0.0042679933831095695, + 0.004929832182824612, + 0.002743622288107872, + 0.0051883612759411335, + 0.005993007682263851, + 0.006868560332804918, + 0.0046687680296599865, + 0.003474252065643668, + 0.004848144482821226, + 0.004298392683267593, + 0.0028763313312083483, + 0.011622917838394642, + 0.013260156847536564, + 0.013449140824377537, + 0.01753147877752781, + 0.019181376323103905, + 0.005718739703297615, + 0.009368778206408024, + 0.007080232258886099, + 0.006404672283679247, + 0.006066647358238697, + 0.004350241739302874, + 0.006445649079978466, + 0.005510337185114622, + 0.004176267888396978, + 0.005341698881238699, + 0.006270441226661205, + 0.007908369414508343, + 0.01381750125437975, + 0.007903662510216236, + 0.008319380693137646, + 0.007574337534606457, + 0.006219035480171442, + 0.008291847072541714, + 0.006897196173667908, + 0.006006758194416761, + 0.011703575029969215, + 0.006494081113487482, + 0.004752223379909992, + 0.006560642272233963, + 0.0028819802682846785, + 0.0031079750042408705, + 0.002858740510419011, + 0.015794282779097557, + 0.020916007459163666, + 0.014054952189326286, + 0.017690910026431084, + 0.019833430647850037, + 0.009468551725149155, + 0.011775013990700245, + 0.015115872956812382, + 0.07362528890371323, + 0.07171116769313812, + 0.05095590278506279, + 0.07465441524982452, + 0.04851757735013962, + 0.0388842448592186, + 0.05045423284173012, + 0.04613879695534706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.1176810935139656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700753852725029 + ], + "opposite_logprob_tokens": [ + -2.695075273513794 + ], + "logprob_reward": -0.0700753852725029, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700753852725029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20493820309638977, + 0.09228940308094025, + 0.11697448045015335, + 0.05142650008201599, + 0.12217841297388077, + 0.10300184041261673, + 0.07093542069196701, + 0.06832121312618256, + 0.1099923700094223, + 0.05994214117527008 + ], + "credit_map": [ + 0.019960060715675354, + 0.04255852848291397, + 0.022211598232388496, + 0.019858337938785553, + 0.02273428626358509, + 0.019292138516902924, + 0.033443011343479156, + 0.024880247190594673, + 0.0117673110216856, + 0.012129798531532288, + 0.011239188723266125, + 0.008503983728587627, + 0.010094461031258106, + 0.014080957509577274, + 0.013700509443879128, + 0.010773192159831524, + 0.008920345455408096, + 0.014105517417192459, + 0.016939111053943634, + 0.00826834887266159, + 0.03206951171159744, + 0.012383288703858852, + 0.014530419372022152, + 0.0097579350695014, + 0.005171224009245634, + 0.007581500802189112, + 0.010656560771167278, + 0.007606134749948978, + 0.0046418108977377415, + 0.006606896407902241, + 0.005693323444575071, + 0.003469049697741866, + 0.015564371831715107, + 0.01594384014606476, + 0.019043905660510063, + 0.018626512959599495, + 0.023688487708568573, + 0.00719129154458642, + 0.012939227744936943, + 0.009180774912238121, + 0.012881861999630928, + 0.02068287879228592, + 0.01391263771802187, + 0.013363895006477833, + 0.008849290199577808, + 0.01171094924211502, + 0.009329164400696754, + 0.01227116584777832, + 0.0068620434030890465, + 0.008146745152771473, + 0.009100181981921196, + 0.011012149974703789, + 0.007595673203468323, + 0.005399379879236221, + 0.014234118163585663, + 0.008585131727159023, + 0.010720057412981987, + 0.017288686707615852, + 0.008689253591001034, + 0.006960786879062653, + 0.010759023949503899, + 0.0042912340722978115, + 0.004915919154882431, + 0.004696258343756199, + 0.01157332118600607, + 0.014583165757358074, + 0.01172599010169506, + 0.023785598576068878, + 0.010537845082581043, + 0.007568628527224064, + 0.02106296829879284, + 0.009154858998954296, + 0.005830077920109034, + 0.006862891372293234, + 0.006917786784470081, + 0.009577691555023193, + 0.007017086260020733, + 0.007502634543925524, + 0.008007203228771687, + 0.008226767182350159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0700753852725029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808995708823204 + ], + "opposite_logprob_tokens": [ + -3.2880899906158447 + ], + "logprob_reward": -0.03808995708823204, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808995708823204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16776923835277557, + 0.13423146307468414, + 0.1683300882577896, + 0.025219127535820007, + 0.15411141514778137, + 0.1745222806930542, + 0.050626300275325775, + 0.02866065502166748, + 0.06286908686161041, + 0.03366038203239441 + ], + "credit_map": [ + 0.017893098294734955, + 0.027027497068047523, + 0.01919529400765896, + 0.019779173657298088, + 0.02213555946946144, + 0.016029801219701767, + 0.02432483807206154, + 0.021383967250585556, + 0.01955220103263855, + 0.020267223939299583, + 0.016320616006851196, + 0.014243125915527344, + 0.012003224343061447, + 0.012172657996416092, + 0.022860297933220863, + 0.01681211031973362, + 0.012319325469434261, + 0.021973440423607826, + 0.02357454039156437, + 0.01297780591994524, + 0.04940202459692955, + 0.01636311039328575, + 0.019497625529766083, + 0.012222224846482277, + 0.003218705765902996, + 0.003799814498052001, + 0.0037904817145317793, + 0.0027560279704630375, + 0.0039073084481060505, + 0.002337861806154251, + 0.003009199630469084, + 0.0023997274693101645, + 0.015526182949543, + 0.01753443293273449, + 0.017134912312030792, + 0.024808252230286598, + 0.03556249663233757, + 0.010747618041932583, + 0.018072005361318588, + 0.014725509099662304, + 0.028728431090712547, + 0.036656033247709274, + 0.017864378169178963, + 0.02603565715253353, + 0.014540519565343857, + 0.013579572550952435, + 0.018134789541363716, + 0.018982894718647003, + 0.006288925185799599, + 0.0060370503924787045, + 0.006295930128544569, + 0.0067194742150604725, + 0.0050375754944980145, + 0.0058412919752299786, + 0.0077886683866381645, + 0.0066173807717859745, + 0.0026846567634493113, + 0.004182528704404831, + 0.0040472266264259815, + 0.005166746210306883, + 0.003622753545641899, + 0.002575863851234317, + 0.0029926132410764694, + 0.003388265846297145, + 0.007895460352301598, + 0.012996017932891846, + 0.007022644858807325, + 0.00789935328066349, + 0.005829465109854937, + 0.006609045434743166, + 0.005592534318566322, + 0.009024566039443016, + 0.0047765374183654785, + 0.005572102498263121, + 0.0036764240358024836, + 0.005820007529109716, + 0.003561971243470907, + 0.002652133582159877, + 0.004006166011095047, + 0.0035950387828052044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.03808995708823204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233865976333618 + ], + "opposite_logprob_tokens": [ + -0.8983865976333618 + ], + "logprob_reward": -0.5233865976333618, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233865976333618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24144141376018524, + 0.10246460139751434, + 0.09080372750759125, + 0.0481913685798645, + 0.05556271970272064, + 0.17748314142227173, + 0.05642543360590935, + 0.13149239122867584, + 0.07671181857585907, + 0.019423427060246468 + ], + "credit_map": [ + 0.026775531470775604, + 0.04715864732861519, + 0.02701455168426037, + 0.03681061789393425, + 0.027131052687764168, + 0.01621614582836628, + 0.03537599742412567, + 0.024958858266472816, + 0.0110434265807271, + 0.02003595232963562, + 0.012217061594128609, + 0.01095577608793974, + 0.008931786753237247, + 0.011683698743581772, + 0.018107444047927856, + 0.009489456191658974, + 0.006540766451507807, + 0.010198247618973255, + 0.012889806181192398, + 0.007116974331438541, + 0.027590831741690636, + 0.00985714141279459, + 0.010614968836307526, + 0.005994988605380058, + 0.007469650357961655, + 0.008136257529258728, + 0.009992675855755806, + 0.005512978881597519, + 0.004758375696837902, + 0.0030561897438019514, + 0.006429605185985565, + 0.0028356341645121574, + 0.006724341306835413, + 0.008138900622725487, + 0.008069516159594059, + 0.012587162666022778, + 0.0076631056144833565, + 0.0035520121455192566, + 0.004951051436364651, + 0.0038766306824982166, + 0.01257083099335432, + 0.018097203224897385, + 0.01821187511086464, + 0.02754761092364788, + 0.02002577856183052, + 0.01895214058458805, + 0.03149630129337311, + 0.030581390485167503, + 0.0071408990770578384, + 0.007809972856193781, + 0.006435710936784744, + 0.007240752689540386, + 0.00457292515784502, + 0.005825450178235769, + 0.010544802993535995, + 0.006854920648038387, + 0.019726134836673737, + 0.0346422903239727, + 0.02465537004172802, + 0.017199888825416565, + 0.009922470897436142, + 0.006000515539199114, + 0.010075878351926804, + 0.00926983542740345, + 0.012228172272443771, + 0.020578885450959206, + 0.012575121596455574, + 0.007331490982323885, + 0.00713319331407547, + 0.0053354487754404545, + 0.005814346019178629, + 0.005715157836675644, + 0.0031016471330076456, + 0.0028896385338157415, + 0.0022211738396435976, + 0.00347212259657681, + 0.0016461533959954977, + 0.0012194670271128416, + 0.0026710103265941143, + 0.002202213043347001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.5233865976333618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera glide smoothly forward and slowly pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense stream of thousands of runners flowing along a scenic coastal highway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the runners wearing neon yellow vests, electric blue leggings, and hot pink shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the azure ocean visible to the right with gentle waves breaking into white foam on a rocky shoreline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014184252358973026 + ], + "opposite_logprob_tokens": [ + -4.264184474945068 + ], + "logprob_reward": -0.014184252358973026, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014184252358973026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2538965940475464, + 0.1307513415813446, + 0.09508777409791946, + 0.0661601722240448, + 0.048446010798215866, + 0.08934684097766876, + 0.07631538063287735, + 0.04995054379105568, + 0.12534302473068237, + 0.0647023469209671 + ], + "credit_map": [ + 0.027506690472364426, + 0.04533281549811363, + 0.025629106909036636, + 0.026370326057076454, + 0.04039238765835762, + 0.03396517038345337, + 0.030700188130140305, + 0.0239998921751976, + 0.02101239189505577, + 0.028955550864338875, + 0.016598060727119446, + 0.013372537679970264, + 0.011228339746594429, + 0.010484457947313786, + 0.0165749229490757, + 0.012525070458650589, + 0.007478869054466486, + 0.01258870679885149, + 0.013239186257123947, + 0.0069719841703772545, + 0.026557479053735733, + 0.009753872640430927, + 0.011486065573990345, + 0.00701160216704011, + 0.011192652396857738, + 0.011147206649184227, + 0.0099166389554739, + 0.007659338880330324, + 0.007281898055225611, + 0.00628670072183013, + 0.006186802871525288, + 0.006488929037004709, + 0.005547268316149712, + 0.009493515826761723, + 0.008079010993242264, + 0.009134604595601559, + 0.005503470078110695, + 0.003233165480196476, + 0.004388916306197643, + 0.0030660557094961405, + 0.013937165960669518, + 0.01708712987601757, + 0.011720532551407814, + 0.009554635733366013, + 0.011650641448795795, + 0.008584609255194664, + 0.009564761072397232, + 0.00724736088886857, + 0.007870541885495186, + 0.008244943805038929, + 0.007023245561867952, + 0.009067071601748466, + 0.0069861640222370625, + 0.0160265751183033, + 0.007635451387614012, + 0.013461383990943432, + 0.0053847539238631725, + 0.008929464034736156, + 0.0064550829119980335, + 0.0058953785337507725, + 0.007842606864869595, + 0.005495802033692598, + 0.005073817912489176, + 0.0048736343160271645, + 0.015688765794038773, + 0.01601725071668625, + 0.014169023372232914, + 0.015257914550602436, + 0.013135057874023914, + 0.01782805472612381, + 0.01419106125831604, + 0.019055897369980812, + 0.006306791212409735, + 0.0070614865981042385, + 0.007834190502762794, + 0.011059298180043697, + 0.00854974053800106, + 0.011273980140686035, + 0.007269838359206915, + 0.005347013007849455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014184252358973026 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_14/index_tensor([6594, 6594])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_14/index_tensor([6594, 6594])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ffdae51e9330a6bb76ec71d0cf5632d413d49196 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_14/index_tensor([6594, 6594])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, aerial drone shot captures a wide, viscous stream of molten rock, cascading from fissures atop a rugged, black volcano, against the backdrop of deep night. This stream of molten rock cascades from a dazzling white-hot center to a deep, blood-red rim, its surface roiling and bubbling, its intense heat sending ripples of distortion through the air. With a slow, inexorable flow, it initially diverges into intertwining tributaries that cling to the dark mountain like vast, luminous roots, before gradually converging into a broad, main stream. As the camera slowly pulls out, the image becomes dominated by this magnificent, flaming river, carving a sinuous lifeline through the inky landscape. The magma itself becomes the sole source of light, tinting the thick ash plume an ominous orange-red, creating an atmosphere both awe-inspiring and imbued with a primal, destructive force.", + "seed": 12856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_14/index_6594_seed_12856_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09493950139731169, + "vlm_logprob": -0.09493950139731169, + "vlm_yes_no_margin": 3.3500000119209288, + "vlm_token_credit": -0.09493950139731169, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015391147695481777 + ], + "opposite_logprob_tokens": [ + -4.7653913497924805 + ], + "logprob_reward": -0.015391147695481777, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015391147695481777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07781826704740524, + 0.07366924732923508, + 0.22229962050914764, + 0.08802352100610733, + 0.1014508605003357, + 0.10080688446760178, + 0.11280649900436401, + 0.1464250534772873, + 0.03237309679389, + 0.04432699456810951 + ], + "credit_map": [ + 0.011534160934388638, + 0.0063497028313577175, + 0.008335710503160954, + 0.009639187715947628, + 0.009984915144741535, + 0.012950192205607891, + 0.007419861853122711, + 0.011604535393416882, + 0.008758238516747952, + 0.005672710947692394, + 0.008175325579941273, + 0.011985254473984241, + 0.004675030708312988, + 0.010818952694535255, + 0.008281487971544266, + 0.015302245505154133, + 0.032302435487508774, + 0.026379026472568512, + 0.025331798940896988, + 0.02972220443189144, + 0.01833157427608967, + 0.03000575490295887, + 0.02542211301624775, + 0.03480470925569534, + 0.012708183377981186, + 0.009360331110656261, + 0.009170718491077423, + 0.010244274511933327, + 0.010727600194513798, + 0.011578289791941643, + 0.007480508182197809, + 0.01675361953675747, + 0.016711847856640816, + 0.011971816420555115, + 0.008395210839807987, + 0.009159024804830551, + 0.010466692969202995, + 0.01333246473222971, + 0.01074626948684454, + 0.020667526870965958, + 0.01850985921919346, + 0.010829501785337925, + 0.00963890552520752, + 0.010027843527495861, + 0.013281653635203838, + 0.013683058321475983, + 0.010540188290178776, + 0.014295874163508415, + 0.020662156865000725, + 0.011417052708566189, + 0.009614885784685612, + 0.011829561553895473, + 0.015769438818097115, + 0.01585189253091812, + 0.012469720095396042, + 0.015191787853837013, + 0.01399197243154049, + 0.017455553635954857, + 0.015109601430594921, + 0.01795414835214615, + 0.020259317010641098, + 0.018603332340717316, + 0.01862417347729206, + 0.024426940828561783, + 0.004092681687325239, + 0.0029744307976216078, + 0.003094888525083661, + 0.004871703684329987, + 0.0031973072327673435, + 0.004767351318150759, + 0.003575737588107586, + 0.005798995029181242, + 0.0041784136556088924, + 0.003461879678070545, + 0.004114937968552113, + 0.007708038669079542, + 0.005249463953077793, + 0.007118071895092726, + 0.005009783897548914, + 0.007486405782401562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.015391147695481777 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016091477125883102 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.016091477125883102, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091477125883102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0840492993593216, + 0.054326802492141724, + 0.11303170025348663, + 0.13885395228862762, + 0.08477455377578735, + 0.21209953725337982, + 0.1481785774230957, + 0.09433450549840927, + 0.02081226371228695, + 0.04953887686133385 + ], + "credit_map": [ + 0.0156118618324399, + 0.0058105746284127235, + 0.007970246486365795, + 0.007671043276786804, + 0.006281428970396519, + 0.01519289892166853, + 0.009616971015930176, + 0.015894271433353424, + 0.008730323053896427, + 0.005307118408381939, + 0.005498799961060286, + 0.006931790150702, + 0.004811435006558895, + 0.00808341708034277, + 0.005408461205661297, + 0.0095554543659091, + 0.01962435059249401, + 0.009077063761651516, + 0.011038947850465775, + 0.01437343005090952, + 0.005939672235399485, + 0.018442856147885323, + 0.011332400143146515, + 0.02320297434926033, + 0.022800972685217857, + 0.014608013443648815, + 0.01520313136279583, + 0.017508842051029205, + 0.010408615693449974, + 0.019481929019093513, + 0.012115703895688057, + 0.026726730167865753, + 0.012884723953902721, + 0.013381823897361755, + 0.007695014588534832, + 0.008254438638687134, + 0.010802321135997772, + 0.011130603030323982, + 0.008205423131585121, + 0.012420202605426311, + 0.03944908827543259, + 0.021115196868777275, + 0.01948700286448002, + 0.022005802020430565, + 0.0332077220082283, + 0.03169567137956619, + 0.018204525113105774, + 0.02693452127277851, + 0.02619165927171707, + 0.013709639199078083, + 0.014205771498382092, + 0.016100045293569565, + 0.02462349645793438, + 0.021445617079734802, + 0.013972328044474125, + 0.01793000102043152, + 0.010409989394247532, + 0.01557883620262146, + 0.007933939807116985, + 0.011028986424207687, + 0.011692148633301258, + 0.010693510062992573, + 0.010881610214710236, + 0.016115471720695496, + 0.0027158488519489765, + 0.002373390132561326, + 0.0017054579220712185, + 0.003440971253439784, + 0.0025858585722744465, + 0.002857283689081669, + 0.0019310892093926668, + 0.0032023636158555746, + 0.007134727202355862, + 0.006291973404586315, + 0.0034334261436015368, + 0.004994452930986881, + 0.008140061981976032, + 0.006642392836511135, + 0.0032187916804105043, + 0.009683044627308846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.016091477125883102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981162816286087 + ], + "opposite_logprob_tokens": [ + -3.5298116207122803 + ], + "logprob_reward": -0.02981162816286087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981162816286087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12008035182952881, + 0.06850747019052505, + 0.07793483883142471, + 0.10153636336326599, + 0.17442922294139862, + 0.07760834693908691, + 0.12690922617912292, + 0.1794256865978241, + 0.013861442916095257, + 0.059707023203372955 + ], + "credit_map": [ + 0.022006981074810028, + 0.010045040398836136, + 0.009504844434559345, + 0.012218046002089977, + 0.014715145342051983, + 0.019981952384114265, + 0.016380105167627335, + 0.01522823516279459, + 0.010832726024091244, + 0.006457527633756399, + 0.007970951497554779, + 0.01114599034190178, + 0.006584322080016136, + 0.00922123808413744, + 0.006124787498265505, + 0.010169925168156624, + 0.011128171347081661, + 0.007264275103807449, + 0.008397082798182964, + 0.011377068236470222, + 0.006485310383141041, + 0.01230926625430584, + 0.008257296867668629, + 0.012716366909444332, + 0.015102402307093143, + 0.011633350513875484, + 0.01050269789993763, + 0.01221622247248888, + 0.011686382815241814, + 0.014092671684920788, + 0.01040355022996664, + 0.01589909754693508, + 0.018268698826432228, + 0.02116701379418373, + 0.01786036230623722, + 0.020411966368556023, + 0.026039091870188713, + 0.020569127053022385, + 0.030396291986107826, + 0.01971668377518654, + 0.015341943129897118, + 0.007295229006558657, + 0.006219915114343166, + 0.0125587098300457, + 0.008376387879252434, + 0.010190318338572979, + 0.005156001076102257, + 0.012469851411879063, + 0.013930496759712696, + 0.01322206761687994, + 0.011942156590521336, + 0.014125270769000053, + 0.016637850552797318, + 0.016102759167551994, + 0.019557900726795197, + 0.021390719339251518, + 0.018724238499999046, + 0.02632407657802105, + 0.01711779274046421, + 0.02132299542427063, + 0.025126894935965538, + 0.02251431532204151, + 0.023777971044182777, + 0.024517416954040527, + 0.0022134806495159864, + 0.0013846325455233455, + 0.0009795376099646091, + 0.0020169042982161045, + 0.001493106479756534, + 0.002182232914492488, + 0.0013587099965661764, + 0.002232837723568082, + 0.005483889486640692, + 0.0056725749745965, + 0.003417491214349866, + 0.009020465426146984, + 0.005911176558583975, + 0.009755080565810204, + 0.003432618919759989, + 0.017013730481266975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.02981162816286087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635316178202629 + ], + "opposite_logprob_tokens": [ + -3.651353120803833 + ], + "logprob_reward": -0.02635316178202629, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635316178202629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09863774478435516, + 0.15257862210273743, + 0.05843111872673035, + 0.08306346088647842, + 0.07855657488107681, + 0.17447532713413239, + 0.20154836773872375, + 0.1169954240322113, + 0.017559582367539406, + 0.018153775483369827 + ], + "credit_map": [ + 0.021295521408319473, + 0.008456985466182232, + 0.010150158777832985, + 0.009873390197753906, + 0.008491020649671555, + 0.01384985912591219, + 0.012799522839486599, + 0.013721281662583351, + 0.019882092252373695, + 0.01703297719359398, + 0.016108695417642593, + 0.025444723665714264, + 0.017416227608919144, + 0.019686156883835793, + 0.021098066121339798, + 0.015909673646092415, + 0.007678453344851732, + 0.008672181516885757, + 0.005508571397513151, + 0.008849975652992725, + 0.007450646720826626, + 0.005707262549549341, + 0.007109830621629953, + 0.007454193662852049, + 0.012219726108014584, + 0.010762562043964863, + 0.008779223076999187, + 0.014444959349930286, + 0.007910220883786678, + 0.010729526169598103, + 0.006590910255908966, + 0.01162633951753378, + 0.013860251754522324, + 0.010361663065850735, + 0.007213914766907692, + 0.007774531375616789, + 0.010179086588323116, + 0.010390227660536766, + 0.008208180777728558, + 0.010568717494606972, + 0.020399898290634155, + 0.024095185101032257, + 0.014234146103262901, + 0.025235289707779884, + 0.02584999054670334, + 0.01879510097205639, + 0.020715540274977684, + 0.025150174275040627, + 0.02486768551170826, + 0.028586043044924736, + 0.018015746027231216, + 0.03366294875741005, + 0.02449849806725979, + 0.024678286164999008, + 0.02230328880250454, + 0.02493586391210556, + 0.016593972221016884, + 0.014380211941897869, + 0.013227637857198715, + 0.016040368005633354, + 0.015502801164984703, + 0.0151669355109334, + 0.01117180660367012, + 0.014911687932908535, + 0.0022140361834317446, + 0.002105164108797908, + 0.0019341910956427455, + 0.0030614356510341167, + 0.0018489286303520203, + 0.002487005665898323, + 0.0015488649951294065, + 0.0023599565029144287, + 0.0016154005425050855, + 0.00269931647926569, + 0.001644508563913405, + 0.0024068921338766813, + 0.002127616899088025, + 0.0024135084822773933, + 0.001735052326694131, + 0.003511479590088129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.02635316178202629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870500922203064 + ], + "opposite_logprob_tokens": [ + -1.1370501518249512 + ], + "logprob_reward": -0.3870500922203064, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870500922203064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05308490991592407, + 0.08118758350610733, + 0.05608981475234032, + 0.0845930427312851, + 0.25656652450561523, + 0.15149317681789398, + 0.11881624907255173, + 0.13370057940483093, + 0.02626682072877884, + 0.03820131719112396 + ], + "credit_map": [ + 0.00713268481194973, + 0.005127469077706337, + 0.004165423568338156, + 0.005406723823398352, + 0.006585482507944107, + 0.009761268272995949, + 0.005643051583319902, + 0.009262802079319954, + 0.008837409317493439, + 0.008781655691564083, + 0.006838504690676928, + 0.012948425486683846, + 0.007840028032660484, + 0.011964092031121254, + 0.00930197723209858, + 0.014675493352115154, + 0.007313636131584644, + 0.0044577354565262794, + 0.005455502774566412, + 0.0067421672865748405, + 0.005306672304868698, + 0.009522727690637112, + 0.005925290286540985, + 0.011366082355380058, + 0.012133614160120487, + 0.01097027212381363, + 0.008939920924603939, + 0.011136776767671108, + 0.0095553919672966, + 0.012124073691666126, + 0.008656870573759079, + 0.0110761234536767, + 0.03957311064004898, + 0.023389408364892006, + 0.02503323368728161, + 0.024878520518541336, + 0.03691478073596954, + 0.03464870899915695, + 0.03731948509812355, + 0.03480929136276245, + 0.01775488071143627, + 0.019911807030439377, + 0.015957333147525787, + 0.01768815517425537, + 0.02232353948056698, + 0.01659219153225422, + 0.02348027378320694, + 0.017785005271434784, + 0.011049267835915089, + 0.011843272484838963, + 0.012051022611558437, + 0.012492607347667217, + 0.017899932339787483, + 0.013877242803573608, + 0.01815655082464218, + 0.0214463509619236, + 0.017970280721783638, + 0.022340910509228706, + 0.010405694134533405, + 0.019503643736243248, + 0.01597091183066368, + 0.016311530023813248, + 0.009806965477764606, + 0.02139064110815525, + 0.003919723443686962, + 0.00270827766507864, + 0.002613934688270092, + 0.004293244332075119, + 0.0029184867162257433, + 0.0035094742197543383, + 0.002111417707055807, + 0.004192260093986988, + 0.004414156079292297, + 0.003831711132079363, + 0.003526286920532584, + 0.0037087248638272285, + 0.006045287009328604, + 0.005700577516108751, + 0.0036717026960104704, + 0.007302870973944664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.3870500922203064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, aerial drone shot captures a wide, viscous stream of molten rock, cascading from fissures atop a rugged, black volcano, against the backdrop of deep night. This stream of molten rock cascades from a dazzling white-hot center to a deep, blood-red rim, its surface roiling and bubbling, its intense heat sending ripples of distortion through the air. With a slow, inexorable flow, it initially diverges into intertwining tributaries that cling to the dark mountain like vast, luminous roots, before gradually converging into a broad, main stream. As the camera slowly pulls out, the image becomes dominated by this magnificent, flaming river, carving a sinuous lifeline through the inky landscape. The magma itself becomes the sole source of light, tinting the thick ash plume an ominous orange-red, creating an atmosphere both awe-inspiring and imbued with a primal, destructive force.", + "seed": 12857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_14/index_6594_seed_12857_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10650672139599919, + "vlm_logprob": -0.10650672139599919, + "vlm_yes_no_margin": 3.7749999523162843, + "vlm_token_credit": -0.10650672139599919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012741466984152794 + ], + "opposite_logprob_tokens": [ + -5.512741565704346 + ], + "logprob_reward": -0.012741466984152794, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.012741466984152794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1332264244556427, + 0.10123833268880844, + 0.06878858804702759, + 0.17776280641555786, + 0.10794609040021896, + 0.08273138105869293, + 0.07810546457767487, + 0.03721662238240242, + 0.09509442746639252, + 0.1178898960351944 + ], + "credit_map": [ + 0.008953602984547615, + 0.01062077097594738, + 0.011848528869450092, + 0.010578515008091927, + 0.01160674449056387, + 0.03512570634484291, + 0.01171221025288105, + 0.03278034180402756, + 0.009299538098275661, + 0.006363284774124622, + 0.011821409687399864, + 0.00515716290101409, + 0.004852485843002796, + 0.025637255981564522, + 0.013882734812796116, + 0.02422446385025978, + 0.008131589740514755, + 0.00617548543959856, + 0.013355765491724014, + 0.0069684721529483795, + 0.006620848085731268, + 0.008499139919877052, + 0.011700150556862354, + 0.007337136659771204, + 0.02868310734629631, + 0.012213503941893578, + 0.02881564572453499, + 0.018753115087747574, + 0.013747664168477058, + 0.026196260005235672, + 0.01832583360373974, + 0.03102768212556839, + 0.012821701355278492, + 0.013257223181426525, + 0.014670736156404018, + 0.012900101952254772, + 0.01864027790725231, + 0.012015367858111858, + 0.011813033372163773, + 0.01182765793055296, + 0.018198098987340927, + 0.008890903554856777, + 0.010615319013595581, + 0.007906227372586727, + 0.008202100172638893, + 0.012999555096030235, + 0.008228284306824207, + 0.00769089488312602, + 0.012196237221360207, + 0.009645836427807808, + 0.009937957860529423, + 0.010394690558314323, + 0.007478659972548485, + 0.009486570954322815, + 0.00891138706356287, + 0.010054121725261211, + 0.004795590415596962, + 0.0054315305314958096, + 0.006052367389202118, + 0.004784856457263231, + 0.003404987044632435, + 0.005149529781192541, + 0.003936721011996269, + 0.0036610416136682034, + 0.014610354788601398, + 0.008102692663669586, + 0.01764174923300743, + 0.013395167887210846, + 0.007072465028613806, + 0.010509124025702477, + 0.014209093526005745, + 0.009553773328661919, + 0.011374693363904953, + 0.011013061739504337, + 0.004104620777070522, + 0.006544994655996561, + 0.009202064014971256, + 0.03228573128581047, + 0.00944969430565834, + 0.03391503542661667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.012741466984152794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00675205560401082 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.00675205560401082, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00675205560401082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1074444130063057, + 0.14592234790325165, + 0.08110101521015167, + 0.038670510053634644, + 0.2710287272930145, + 0.05589083582162857, + 0.055188167840242386, + 0.07081473618745804, + 0.08003929257392883, + 0.09389994293451309 + ], + "credit_map": [ + 0.009371574968099594, + 0.011308816261589527, + 0.017056912183761597, + 0.008653395809233189, + 0.008168629370629787, + 0.021578006446361542, + 0.011244804598391056, + 0.0200622770935297, + 0.011776200495660305, + 0.00882773194462061, + 0.020397182554006577, + 0.010594571940600872, + 0.00756843714043498, + 0.035773735493421555, + 0.01650054194033146, + 0.03448394313454628, + 0.014362331479787827, + 0.00783830787986517, + 0.011690538376569748, + 0.007901300676167011, + 0.006184130907058716, + 0.010431155562400818, + 0.009966189041733742, + 0.012727062217891216, + 0.004925261251628399, + 0.004654034040868282, + 0.005499676335602999, + 0.003415757091715932, + 0.0040185838006436825, + 0.00547573110088706, + 0.005506896413862705, + 0.005174568854272366, + 0.028800619766116142, + 0.036337509751319885, + 0.03213585540652275, + 0.028430260717868805, + 0.04281315952539444, + 0.03146976977586746, + 0.038581203669309616, + 0.03246036544442177, + 0.009303227998316288, + 0.010804991237819195, + 0.005230159033089876, + 0.0046798354014754295, + 0.009017937816679478, + 0.007013330701738596, + 0.0049413819797337055, + 0.004899975378066301, + 0.006592623423784971, + 0.0044091809540987015, + 0.00956097524613142, + 0.008496764115989208, + 0.005820948630571365, + 0.007299947552382946, + 0.006085705943405628, + 0.00692202290520072, + 0.009303092956542969, + 0.006296860054135323, + 0.010379127226769924, + 0.0110198725014925, + 0.0043915375135838985, + 0.008495766669511795, + 0.011824247427284718, + 0.009104231372475624, + 0.004967928864061832, + 0.00244973530061543, + 0.002833620412275195, + 0.002347140107303858, + 0.0025246022269129753, + 0.030378812924027443, + 0.0032412793952971697, + 0.031296174973249435, + 0.0067172786220908165, + 0.007723935414105654, + 0.004706896375864744, + 0.00751290749758482, + 0.007949640043079853, + 0.024538999423384666, + 0.00962661113590002, + 0.025123680010437965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.00675205560401082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587211474776268 + ], + "opposite_logprob_tokens": [ + -3.895587205886841 + ], + "logprob_reward": -0.020587211474776268, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587211474776268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.383517861366272, + 0.08279009908437729, + 0.06061404570937157, + 0.0649951621890068, + 0.10152627527713776, + 0.06747231632471085, + 0.07697797566652298, + 0.06963910162448883, + 0.01829742081463337, + 0.07416970282793045 + ], + "credit_map": [ + 0.021655336022377014, + 0.03207797929644585, + 0.041233327239751816, + 0.025759877637028694, + 0.03384425491094589, + 0.09737619012594223, + 0.036665331572294235, + 0.09490560740232468, + 0.010700849816203117, + 0.00865884218364954, + 0.015190653502941132, + 0.006515560206025839, + 0.005145567934960127, + 0.011108574457466602, + 0.013447847217321396, + 0.012022205628454685, + 0.01232755184173584, + 0.006456505972892046, + 0.00783107802271843, + 0.004706619307398796, + 0.005252516362816095, + 0.007614973466843367, + 0.00787353701889515, + 0.00855126604437828, + 0.011387042701244354, + 0.0071710702031850815, + 0.00824142899364233, + 0.005436273757368326, + 0.006446667015552521, + 0.00849803164601326, + 0.007480214349925518, + 0.010334446094930172, + 0.01504741981625557, + 0.009828874841332436, + 0.01524216029793024, + 0.011555894277989864, + 0.012589487247169018, + 0.0130041828379035, + 0.015205850824713707, + 0.00905242096632719, + 0.011370436288416386, + 0.01070184726268053, + 0.006154099479317665, + 0.007173093035817146, + 0.010788248851895332, + 0.00869457982480526, + 0.005364584736526012, + 0.0072254310362041, + 0.0113153588026762, + 0.010395551100373268, + 0.012652603909373283, + 0.009032493457198143, + 0.008648676797747612, + 0.010127289220690727, + 0.007466500625014305, + 0.007339506410062313, + 0.010788586921989918, + 0.006416593212634325, + 0.009757531806826591, + 0.009538802318274975, + 0.00509061012417078, + 0.009308435954153538, + 0.009242524392902851, + 0.00949601735919714, + 0.0018034974345937371, + 0.00250726449303329, + 0.0013987334677949548, + 0.0017551786731928587, + 0.0018949200166389346, + 0.0032680637668818235, + 0.002113673370331526, + 0.0035560906399041414, + 0.005645993631333113, + 0.008315734565258026, + 0.0054018241353333, + 0.006293947342783213, + 0.007550413254648447, + 0.016693206503987312, + 0.0071711488999426365, + 0.01709744520485401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.020587211474776268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1215951219201088, + 0.1285053789615631, + 0.058161791414022446, + 0.07747481763362885, + 0.09121762216091156, + 0.09950239956378937, + 0.20854341983795166, + 0.039279840886592865, + 0.10318610817193985, + 0.0725335106253624 + ], + "credit_map": [ + 0.009964648634195328, + 0.022128112614154816, + 0.01570780761539936, + 0.017883148044347763, + 0.015933247283101082, + 0.011387784034013748, + 0.017269566655158997, + 0.011320799589157104, + 0.012167141772806644, + 0.008121911436319351, + 0.01720295660197735, + 0.011514812707901001, + 0.00650548143312335, + 0.02755025401711464, + 0.019421657547354698, + 0.026021162047982216, + 0.008244527503848076, + 0.006296715699136257, + 0.009491614066064358, + 0.005315098911523819, + 0.0048443968407809734, + 0.008190052583813667, + 0.00826986413449049, + 0.007509519346058369, + 0.011303857900202274, + 0.007051077671349049, + 0.011664380319416523, + 0.00934583880007267, + 0.007641843985766172, + 0.010371125303208828, + 0.011058662086725235, + 0.009038024581968784, + 0.013524672016501427, + 0.007420637179166079, + 0.012610851787030697, + 0.010867382399737835, + 0.012162470258772373, + 0.012394899502396584, + 0.010272040031850338, + 0.011964665725827217, + 0.013797249644994736, + 0.014244243502616882, + 0.013552136719226837, + 0.008995354175567627, + 0.016878262162208557, + 0.012311750091612339, + 0.008072221651673317, + 0.0116511769592762, + 0.03473904728889465, + 0.025693582370877266, + 0.03148204833269119, + 0.025650102645158768, + 0.01870906911790371, + 0.02624557353556156, + 0.02560749463737011, + 0.020416514948010445, + 0.004483198747038841, + 0.004766650032252073, + 0.004994608461856842, + 0.004721653647720814, + 0.004204364959150553, + 0.005147950723767281, + 0.005565411411225796, + 0.005396001506596804, + 0.011723068542778492, + 0.005957783665508032, + 0.010278420522809029, + 0.008533728308975697, + 0.004435147158801556, + 0.026810012757778168, + 0.008464372716844082, + 0.02698357403278351, + 0.009280909784138203, + 0.008292614482343197, + 0.003008222673088312, + 0.004021816421300173, + 0.008587246760725975, + 0.015436532907187939, + 0.006681988015770912, + 0.017224179580807686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.01818765513598919 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742652177810669 + ], + "opposite_logprob_tokens": [ + -0.9742652177810669 + ], + "logprob_reward": -0.4742652177810669, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742652177810669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13365308940410614, + 0.13808295130729675, + 0.09121239185333252, + 0.047971874475479126, + 0.1393071413040161, + 0.0729183629155159, + 0.06586628407239914, + 0.10051657259464264, + 0.12742497026920319, + 0.0830463096499443 + ], + "credit_map": [ + 0.006955946329981089, + 0.012095479294657707, + 0.013892200775444508, + 0.007185061927884817, + 0.017974412068724632, + 0.033468157052993774, + 0.011244352906942368, + 0.03083747997879982, + 0.006970101036131382, + 0.009838106110692024, + 0.012705573812127113, + 0.005737918894737959, + 0.010413610376417637, + 0.04087194800376892, + 0.012429656460881233, + 0.03911605104804039, + 0.013262752443552017, + 0.006404180079698563, + 0.015821870416402817, + 0.011686458252370358, + 0.0053040822967886925, + 0.011548656970262527, + 0.016054604202508926, + 0.011129790917038918, + 0.007978885434567928, + 0.0038032822776585817, + 0.006621934939175844, + 0.004095103591680527, + 0.004133053123950958, + 0.007228954695165157, + 0.006775141227990389, + 0.0073355198837816715, + 0.019388994202017784, + 0.016252823173999786, + 0.020725825801491737, + 0.012770689092576504, + 0.01785779371857643, + 0.017378179356455803, + 0.018480965867638588, + 0.01645188219845295, + 0.010820822790265083, + 0.009335312061011791, + 0.009770294651389122, + 0.009540125727653503, + 0.007217296399176121, + 0.009704453870654106, + 0.008047771640121937, + 0.008482301607728004, + 0.008757100440561771, + 0.008934871293604374, + 0.01149111706763506, + 0.008631506934762001, + 0.0066178953275084496, + 0.00864753220230341, + 0.0074259755201637745, + 0.005360291339457035, + 0.012154296040534973, + 0.010366964153945446, + 0.015424768440425396, + 0.016199296340346336, + 0.00535951042547822, + 0.01277463510632515, + 0.015805678442120552, + 0.012431428767740726, + 0.015330087393522263, + 0.009698009118437767, + 0.019566290080547333, + 0.013278083875775337, + 0.0065836310386657715, + 0.025446254760026932, + 0.012326906435191631, + 0.025195717811584473, + 0.009659120813012123, + 0.008611205965280533, + 0.007668452803045511, + 0.008831348270177841, + 0.009218744933605194, + 0.014333032071590424, + 0.00964151881635189, + 0.01508288737386465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.4742652177810669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, aerial drone shot captures a wide, viscous stream of molten rock, cascading from fissures atop a rugged, black volcano, against the backdrop of deep night. This stream of molten rock cascades from a dazzling white-hot center to a deep, blood-red rim, its surface roiling and bubbling, its intense heat sending ripples of distortion through the air. With a slow, inexorable flow, it initially diverges into intertwining tributaries that cling to the dark mountain like vast, luminous roots, before gradually converging into a broad, main stream. As the camera slowly pulls out, the image becomes dominated by this magnificent, flaming river, carving a sinuous lifeline through the inky landscape. The magma itself becomes the sole source of light, tinting the thick ash plume an ominous orange-red, creating an atmosphere both awe-inspiring and imbued with a primal, destructive force.", + "seed": 12858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_14/index_6594_seed_12858_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07655271794646978, + "vlm_logprob": -0.07655271794646978, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.07655271794646978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01718687079846859 + ], + "opposite_logprob_tokens": [ + -4.392186641693115 + ], + "logprob_reward": -0.01718687079846859, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01718687079846859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10279405862092972, + 0.04942351207137108, + 0.051167573779821396, + 0.07732363790273666, + 0.048099637031555176, + 0.2742069661617279, + 0.12099635601043701, + 0.07082176208496094, + 0.14359979331493378, + 0.06156673654913902 + ], + "credit_map": [ + 0.019263043999671936, + 0.011141691356897354, + 0.015901552513241768, + 0.011854859068989754, + 0.012884780764579773, + 0.009364629164338112, + 0.008142293430864811, + 0.014241205528378487, + 0.006351910997182131, + 0.005463462322950363, + 0.007230549585074186, + 0.005418866407126188, + 0.007233442738652229, + 0.004525280557572842, + 0.005691111087799072, + 0.007508889771997929, + 0.0059475963935256, + 0.006653403863310814, + 0.0071602147072553635, + 0.0041750334203243256, + 0.008962293155491352, + 0.00530315563082695, + 0.004881421569734812, + 0.008084459230303764, + 0.008291900157928467, + 0.011442432180047035, + 0.011066152714192867, + 0.006671214941889048, + 0.010579023510217667, + 0.012112024240195751, + 0.006138155236840248, + 0.011022738181054592, + 0.006851774174720049, + 0.006240147165954113, + 0.005923572462052107, + 0.005164614412933588, + 0.0050151050090789795, + 0.00809188187122345, + 0.00502644432708621, + 0.005786094348877668, + 0.03702489659190178, + 0.027304712682962418, + 0.0489545539021492, + 0.02027324214577675, + 0.023785214871168137, + 0.0508817657828331, + 0.019519580528140068, + 0.04646298289299011, + 0.016403909772634506, + 0.019781308248639107, + 0.015385706909000874, + 0.012722400948405266, + 0.014005660079419613, + 0.016601184383034706, + 0.01069364883005619, + 0.015402532182633877, + 0.008968723937869072, + 0.010274319909512997, + 0.009025569073855877, + 0.007304361555725336, + 0.008689417503774166, + 0.009946237318217754, + 0.006237024441361427, + 0.010376108810305595, + 0.012452168390154839, + 0.02209932915866375, + 0.016403064131736755, + 0.024713018909096718, + 0.018731815740466118, + 0.015560092404484749, + 0.014135429635643959, + 0.019504882395267487, + 0.006893203128129244, + 0.008410428650677204, + 0.0067700110375881195, + 0.008379889652132988, + 0.006133977323770523, + 0.008901986293494701, + 0.006936396472156048, + 0.009140844456851482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.01718687079846859 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801830165088177 + ], + "opposite_logprob_tokens": [ + -4.634801864624023 + ], + "logprob_reward": -0.009801830165088177, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801830165088177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16767600178718567, + 0.1195453405380249, + 0.09022313356399536, + 0.06403454393148422, + 0.04274476319551468, + 0.06710351258516312, + 0.1565559357404709, + 0.157096266746521, + 0.0675232857465744, + 0.06749723106622696 + ], + "credit_map": [ + 0.026660550385713577, + 0.0250515379011631, + 0.022237788885831833, + 0.0165188480168581, + 0.027079414576292038, + 0.017006292939186096, + 0.0158415287733078, + 0.01728004403412342, + 0.015916459262371063, + 0.013088173232972622, + 0.017634041607379913, + 0.015753187239170074, + 0.015521666035056114, + 0.011657734401524067, + 0.013608569279313087, + 0.016365500167012215, + 0.012675580568611622, + 0.009518752805888653, + 0.013146803714334965, + 0.005628503393381834, + 0.016097234562039375, + 0.01014005672186613, + 0.007763050962239504, + 0.015253150835633278, + 0.006931225303560495, + 0.01124480925500393, + 0.009989555925130844, + 0.0050484430976212025, + 0.0075184665620327, + 0.008755027316510677, + 0.005062183830887079, + 0.009484834969043732, + 0.006502863485366106, + 0.005224329419434071, + 0.0067292251624166965, + 0.004256071988493204, + 0.004451748915016651, + 0.005399725865572691, + 0.003981478046625853, + 0.006199319381266832, + 0.00802528578788042, + 0.008507821708917618, + 0.008662881329655647, + 0.006307151634246111, + 0.010390407405793667, + 0.01036470104008913, + 0.0055183838121593, + 0.009326878935098648, + 0.01815645582973957, + 0.022837677970528603, + 0.025650829076766968, + 0.016641007736325264, + 0.014292343519628048, + 0.01978912204504013, + 0.012829850427806377, + 0.02635863982141018, + 0.021597815677523613, + 0.021050283685326576, + 0.021019907668232918, + 0.01830759085714817, + 0.015899086371064186, + 0.021464278921484947, + 0.014618727378547192, + 0.023138580843806267, + 0.00755818746984005, + 0.008071869611740112, + 0.009664410725235939, + 0.009144315496087074, + 0.008728143759071827, + 0.008067884482443333, + 0.006897668354213238, + 0.009390804916620255, + 0.007027583196759224, + 0.01023904699832201, + 0.007807526737451553, + 0.009520758874714375, + 0.006617303937673569, + 0.009120493195950985, + 0.007165692280977964, + 0.009998828172683716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.009801830165088177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643949404358864 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008643949404358864, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643949404358864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387808918952942, + 0.07526714354753494, + 0.15660350024700165, + 0.05926748737692833, + 0.059163641184568405, + 0.10787571966648102, + 0.18203188478946686, + 0.07950015366077423, + 0.09344126284122467, + 0.04806828871369362 + ], + "credit_map": [ + 0.025709357112646103, + 0.0161086767911911, + 0.023666540160775185, + 0.01463761180639267, + 0.01695224642753601, + 0.011793045327067375, + 0.012698924168944359, + 0.017214488238096237, + 0.010478249751031399, + 0.011339525692164898, + 0.011067952960729599, + 0.007451421115547419, + 0.010211654007434845, + 0.008949520997703075, + 0.005432290490716696, + 0.010336528532207012, + 0.020803434774279594, + 0.021246325224637985, + 0.022970858961343765, + 0.012887392193078995, + 0.026175018399953842, + 0.01718689687550068, + 0.013231144286692142, + 0.02210243232548237, + 0.010322852991521358, + 0.0053654140792787075, + 0.009458531625568867, + 0.0032242632005363703, + 0.00808904692530632, + 0.007036335300654173, + 0.0038983849808573723, + 0.01187265757471323, + 0.01131163164973259, + 0.009254684671759605, + 0.008846109732985497, + 0.005618620198220015, + 0.005053127650171518, + 0.007564382627606392, + 0.004454073961824179, + 0.007061011157929897, + 0.015553608536720276, + 0.013858874328434467, + 0.01610383577644825, + 0.010465790517628193, + 0.013001271523535252, + 0.0168662890791893, + 0.011201230809092522, + 0.010824819095432758, + 0.025649400427937508, + 0.031192343682050705, + 0.029472507536411285, + 0.01886453479528427, + 0.013483612798154354, + 0.018154870718717575, + 0.014734309166669846, + 0.030480314046144485, + 0.009309017099440098, + 0.010273854248225689, + 0.008113767020404339, + 0.010300234891474247, + 0.011232567019760609, + 0.011264005675911903, + 0.008090885356068611, + 0.010915827006101608, + 0.008743125945329666, + 0.011462847702205181, + 0.012321218848228455, + 0.011477084830403328, + 0.013458676636219025, + 0.013301907107234001, + 0.01024660561233759, + 0.01242979895323515, + 0.00562505004927516, + 0.006617382634431124, + 0.005538150202482939, + 0.005478919018059969, + 0.004323998466134071, + 0.007321515120565891, + 0.005867311730980873, + 0.0072959596291184425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.008643949404358864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033705130219459534 + ], + "opposite_logprob_tokens": [ + -3.408705234527588 + ], + "logprob_reward": -0.033705130219459534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033705130219459534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07875941693782806, + 0.03757361322641373, + 0.059532586485147476, + 0.09834963828325272, + 0.06823523342609406, + 0.057643547654151917, + 0.11014531552791595, + 0.12110632658004761, + 0.07132074236869812, + 0.2973335385322571 + ], + "credit_map": [ + 0.016096804291009903, + 0.00957875233143568, + 0.013090434484183788, + 0.008260050788521767, + 0.007960979826748371, + 0.006193904671818018, + 0.006719968281686306, + 0.01085851714015007, + 0.004557245410978794, + 0.0037602772936224937, + 0.00518999295309186, + 0.006639858242124319, + 0.004361570347100496, + 0.003944776486605406, + 0.003922745585441589, + 0.005197145044803619, + 0.009849593043327332, + 0.006159421056509018, + 0.009919008240103722, + 0.005079344380646944, + 0.008191079832613468, + 0.0065790642984211445, + 0.004951637703925371, + 0.008803441189229488, + 0.01116642914712429, + 0.014147474430501461, + 0.014968842267990112, + 0.006574318278580904, + 0.01630123145878315, + 0.016481628641486168, + 0.007951932959258556, + 0.010757780633866787, + 0.009668935090303421, + 0.006539324298501015, + 0.012139503844082355, + 0.005578217562288046, + 0.005968746263533831, + 0.01053950097411871, + 0.00583928357809782, + 0.011961723677814007, + 0.00804171059280634, + 0.007795519195497036, + 0.00976885762065649, + 0.004478242248296738, + 0.0051796999759972095, + 0.00906510278582573, + 0.005059025250375271, + 0.008255389519035816, + 0.017387552186846733, + 0.015781955793499947, + 0.015702031552791595, + 0.009614718146622181, + 0.011022426187992096, + 0.014633089303970337, + 0.009433175437152386, + 0.016570372506976128, + 0.011585977859795094, + 0.018821921199560165, + 0.011650455184280872, + 0.01718469336628914, + 0.012996421195566654, + 0.02063547447323799, + 0.01123536191880703, + 0.01699601300060749, + 0.007361085619777441, + 0.008615109138190746, + 0.010831393301486969, + 0.008083255030214787, + 0.009170426987111568, + 0.008486645296216011, + 0.008966515772044659, + 0.009806315414607525, + 0.03246411308646202, + 0.045071110129356384, + 0.03147338330745697, + 0.04437132924795151, + 0.0260554701089859, + 0.04033359885215759, + 0.030744129791855812, + 0.04682042449712753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.033705130219459534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342580914497375 + ], + "opposite_logprob_tokens": [ + -1.3134257793426514 + ], + "logprob_reward": -0.31342580914497375, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342580914497375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1833689659833908, + 0.06165555864572525, + 0.05582014471292496, + 0.07703303545713425, + 0.05255628004670143, + 0.08752671629190445, + 0.09143522381782532, + 0.17646844685077667, + 0.05513475462794304, + 0.15900085866451263 + ], + "credit_map": [ + 0.03151316940784454, + 0.020096534863114357, + 0.025813216343522072, + 0.018552904948592186, + 0.027963995933532715, + 0.016865404322743416, + 0.020567234605550766, + 0.021996498107910156, + 0.007587269879877567, + 0.007495655212551355, + 0.00974863301962614, + 0.007637599483132362, + 0.007591745816171169, + 0.006590087432414293, + 0.007859623990952969, + 0.0071449424140155315, + 0.006581383757293224, + 0.0071366773918271065, + 0.007990377955138683, + 0.00487686088308692, + 0.007999368943274021, + 0.006292382720857859, + 0.007102146279066801, + 0.007840949110686779, + 0.010169452987611294, + 0.00928855687379837, + 0.01124646794050932, + 0.006463867612183094, + 0.007528956048190594, + 0.008499456569552422, + 0.00664177443832159, + 0.01719450019299984, + 0.009941534139215946, + 0.006946555804461241, + 0.009373076260089874, + 0.004383644554764032, + 0.00393395172432065, + 0.005374209955334663, + 0.0040372987277805805, + 0.008566007949411869, + 0.008933181874454021, + 0.009353214874863625, + 0.013126906007528305, + 0.007660661358386278, + 0.014601623639464378, + 0.012086285278201103, + 0.006978503428399563, + 0.014786336570978165, + 0.014748961664736271, + 0.012203785590827465, + 0.015008766204118729, + 0.00797196850180626, + 0.009779984131455421, + 0.00836368277668953, + 0.006298865657299757, + 0.017059216275811195, + 0.016574930399656296, + 0.027951553463935852, + 0.01623297668993473, + 0.026508420705795288, + 0.023221036419272423, + 0.02365170046687126, + 0.016579819843173027, + 0.025748006999492645, + 0.005918761715292931, + 0.007452978286892176, + 0.006673561409115791, + 0.006768514867872, + 0.009479734115302563, + 0.006044568028301001, + 0.00586641114205122, + 0.006930226925760508, + 0.016996461898088455, + 0.022390110418200493, + 0.015927307307720184, + 0.025981927290558815, + 0.013012613169848919, + 0.02087102085351944, + 0.01801210828125477, + 0.025809314101934433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.31342580914497375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, aerial drone shot captures a wide, viscous stream of molten rock, cascading from fissures atop a rugged, black volcano, against the backdrop of deep night. This stream of molten rock cascades from a dazzling white-hot center to a deep, blood-red rim, its surface roiling and bubbling, its intense heat sending ripples of distortion through the air. With a slow, inexorable flow, it initially diverges into intertwining tributaries that cling to the dark mountain like vast, luminous roots, before gradually converging into a broad, main stream. As the camera slowly pulls out, the image becomes dominated by this magnificent, flaming river, carving a sinuous lifeline through the inky landscape. The magma itself becomes the sole source of light, tinting the thick ash plume an ominous orange-red, creating an atmosphere both awe-inspiring and imbued with a primal, destructive force.", + "seed": 12859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_14/index_6594_seed_12859_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09432923514395952, + "vlm_logprob": -0.09432923514395952, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.09432923514395952, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014375808648765087 + ], + "opposite_logprob_tokens": [ + -5.014375686645508 + ], + "logprob_reward": -0.014375808648765087, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.014375808648765087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26450589299201965, + 0.08711035549640656, + 0.10154590755701065, + 0.06404323130846024, + 0.041544072329998016, + 0.13400301337242126, + 0.13571953773498535, + 0.04240816831588745, + 0.0609644316136837, + 0.06815534085035324 + ], + "credit_map": [ + 0.045268237590789795, + 0.01722317934036255, + 0.041615549474954605, + 0.020281711593270302, + 0.0112826619297266, + 0.05193759500980377, + 0.01896052062511444, + 0.05793643370270729, + 0.015142560936510563, + 0.006681086961179972, + 0.013159132562577724, + 0.009398777969181538, + 0.010751466266810894, + 0.007426884490996599, + 0.010875117965042591, + 0.013675328344106674, + 0.01672985963523388, + 0.006851317826658487, + 0.01858155056834221, + 0.01069582998752594, + 0.011391870677471161, + 0.00910199899226427, + 0.012118164449930191, + 0.016075314953923225, + 0.007627409882843494, + 0.00682327663525939, + 0.007984479889273643, + 0.007151819299906492, + 0.007919049821794033, + 0.008959386497735977, + 0.00869565550237894, + 0.008882155641913414, + 0.00488505931571126, + 0.003979973495006561, + 0.006801048293709755, + 0.004340809769928455, + 0.005015823990106583, + 0.004131433553993702, + 0.00644792802631855, + 0.005941996816545725, + 0.013431199826300144, + 0.014340881258249283, + 0.020140601322054863, + 0.013693276792764664, + 0.021459558978676796, + 0.01537279225885868, + 0.01846826635301113, + 0.017096439376473427, + 0.015606417320668697, + 0.0134891951456666, + 0.012463691644370556, + 0.01026190910488367, + 0.013876388780772686, + 0.013213646598160267, + 0.03718192130327225, + 0.019626375287771225, + 0.0049133035354316235, + 0.007807169109582901, + 0.0034941653721034527, + 0.003664223477244377, + 0.005732744932174683, + 0.005875902716070414, + 0.005333510227501392, + 0.005587146617472172, + 0.007393423933535814, + 0.0062628332525491714, + 0.003853779286146164, + 0.0039819516241550446, + 0.008167039602994919, + 0.007314321584999561, + 0.014032367616891861, + 0.00995871052145958, + 0.009428819641470909, + 0.009206773713231087, + 0.007430722936987877, + 0.006925850175321102, + 0.008188665844500065, + 0.0072599439881742, + 0.008280448615550995, + 0.011434116400778294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.014375808648765087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807850234210491 + ], + "opposite_logprob_tokens": [ + -4.63480806350708 + ], + "logprob_reward": -0.009807850234210491, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807850234210491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33442962169647217, + 0.24273507297039032, + 0.05455872416496277, + 0.10646044462919235, + 0.05364273115992546, + 0.03902051970362663, + 0.05892925336956978, + 0.03688390180468559, + 0.039256755262613297, + 0.03408302366733551 + ], + "credit_map": [ + 0.08109046518802643, + 0.019809262827038765, + 0.034950241446495056, + 0.01743895746767521, + 0.014321285299956799, + 0.1054069772362709, + 0.0251602865755558, + 0.036252137273550034, + 0.03095296584069729, + 0.01858954317867756, + 0.02155272290110588, + 0.019473854452371597, + 0.019687646999955177, + 0.0670316070318222, + 0.03334624320268631, + 0.03210047632455826, + 0.006960191298276186, + 0.0032525600399821997, + 0.010663400404155254, + 0.005259005818516016, + 0.005641867406666279, + 0.005897813942283392, + 0.006152724381536245, + 0.010731156915426254, + 0.012417692691087723, + 0.01303696259856224, + 0.012239016592502594, + 0.008740807883441448, + 0.017208324745297432, + 0.017365161329507828, + 0.013210376724600792, + 0.012242105789482594, + 0.004999219439923763, + 0.006522003095597029, + 0.005689235869795084, + 0.004332450684159994, + 0.012412342242896557, + 0.007471338380128145, + 0.006117223761975765, + 0.0060989162884652615, + 0.004680403973907232, + 0.004566878080368042, + 0.005367844365537167, + 0.0038150642067193985, + 0.005373584106564522, + 0.004470085725188255, + 0.005243275314569473, + 0.005503383930772543, + 0.0072455499321222305, + 0.0038161389529705048, + 0.004175595473498106, + 0.0026019387878477573, + 0.004022635985165834, + 0.006228233687579632, + 0.02155671827495098, + 0.009282444603741169, + 0.0037168071139603853, + 0.005257850978523493, + 0.0034078252501785755, + 0.0035839362535625696, + 0.0040943254716694355, + 0.0065757883712649345, + 0.005809654016047716, + 0.004437715280801058, + 0.005311725661158562, + 0.0038040513172745705, + 0.003076462307944894, + 0.003064533229917288, + 0.004329320043325424, + 0.0057189273647964, + 0.008537952788174152, + 0.0054137855768203735, + 0.004984825383871794, + 0.004671227652579546, + 0.002796295564621687, + 0.00321504520252347, + 0.0032504196278750896, + 0.003636470064520836, + 0.00640915846452117, + 0.0051195817068219185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009807850234210491 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979274045675993 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.00979274045675993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979274045675993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2654341161251068, + 0.23611031472682953, + 0.1347571462392807, + 0.08483420312404633, + 0.058333612978458405, + 0.03375415876507759, + 0.06785798072814941, + 0.03312697261571884, + 0.033054105937480927, + 0.05273738130927086 + ], + "credit_map": [ + 0.055978622287511826, + 0.019723398610949516, + 0.02277437224984169, + 0.014655380509793758, + 0.021917348727583885, + 0.07865714281797409, + 0.02684256248176098, + 0.024885259568691254, + 0.028572021052241325, + 0.026966700330376625, + 0.02227811887860298, + 0.022822873666882515, + 0.04830431938171387, + 0.03773702308535576, + 0.030221322551369667, + 0.019207919016480446, + 0.01785658486187458, + 0.011880186386406422, + 0.023081732913851738, + 0.01719714142382145, + 0.015504010021686554, + 0.012128882110118866, + 0.015427800826728344, + 0.021680809557437897, + 0.008898640982806683, + 0.009738058783113956, + 0.015575990080833435, + 0.008111618459224701, + 0.01072163786739111, + 0.009365656413137913, + 0.01128684263676405, + 0.011135748587548733, + 0.00665842741727829, + 0.006241292227059603, + 0.009711273945868015, + 0.006816963665187359, + 0.00594767089933157, + 0.005502322223037481, + 0.009108314290642738, + 0.008347340859472752, + 0.0035887344274669886, + 0.003331034444272518, + 0.00498677883297205, + 0.0036266897805035114, + 0.005071678664535284, + 0.004293678794056177, + 0.00418037548661232, + 0.004675186704844236, + 0.007207972928881645, + 0.007187411654740572, + 0.008240263909101486, + 0.0034996650647372007, + 0.00825553759932518, + 0.007715845480561256, + 0.01566586270928383, + 0.01008541602641344, + 0.0038060590159147978, + 0.004815323278307915, + 0.0029240529984235764, + 0.0024495983961969614, + 0.004905736539512873, + 0.0052682189270854, + 0.00413551228120923, + 0.004822467919439077, + 0.004511994309723377, + 0.0038910331204533577, + 0.0028548752889037132, + 0.0029961909167468548, + 0.003021922893822193, + 0.005021981429308653, + 0.005510286428034306, + 0.005245819687843323, + 0.006284363567829132, + 0.006685449741780758, + 0.0065775783732533455, + 0.005129591096192598, + 0.0071034422144293785, + 0.005678873509168625, + 0.007625481113791466, + 0.007652599364519119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00979274045675993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904164284467697 + ], + "opposite_logprob_tokens": [ + -2.4640417098999023 + ], + "logprob_reward": -0.08904164284467697, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904164284467697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17209023237228394, + 0.14040425419807434, + 0.07451849430799484, + 0.09461353719234467, + 0.06072530895471573, + 0.04663462936878204, + 0.12778693437576294, + 0.08576926589012146, + 0.13886579871177673, + 0.05859166383743286 + ], + "credit_map": [ + 0.04319993034005165, + 0.013111654669046402, + 0.029393209144473076, + 0.01861109584569931, + 0.01590365543961525, + 0.016888363286852837, + 0.01664741337299347, + 0.018334893509745598, + 0.021885890513658524, + 0.011172187514603138, + 0.014148426242172718, + 0.011344537138938904, + 0.017293773591518402, + 0.034338343888521194, + 0.017018474638462067, + 0.013202602043747902, + 0.011788702569901943, + 0.006115177646279335, + 0.007272963412106037, + 0.007593229413032532, + 0.011422574520111084, + 0.0076037864200770855, + 0.011247871443629265, + 0.011474179103970528, + 0.009934403002262115, + 0.0107671357691288, + 0.013991164974868298, + 0.010446779429912567, + 0.012509092688560486, + 0.011249620467424393, + 0.014617875218391418, + 0.011097455397248268, + 0.006280848756432533, + 0.005505161825567484, + 0.009126664139330387, + 0.0068567064590752125, + 0.006576823070645332, + 0.0069400956854224205, + 0.010319068096578121, + 0.009119933471083641, + 0.0051740482449531555, + 0.005342084914445877, + 0.00820931326597929, + 0.0035041936207562685, + 0.007704128045588732, + 0.0046371533535420895, + 0.005708898883312941, + 0.0063548036850988865, + 0.01694026030600071, + 0.008868341334164143, + 0.011126978322863579, + 0.007469863630831242, + 0.011284434236586094, + 0.017359398305416107, + 0.03792420029640198, + 0.01681344024837017, + 0.00962032936513424, + 0.011602199636399746, + 0.006632629781961441, + 0.006600914988666773, + 0.011099846102297306, + 0.011258448474109173, + 0.017668120563030243, + 0.011286762543022633, + 0.020168820396065712, + 0.012185398489236832, + 0.010319598950445652, + 0.010233337990939617, + 0.009746585972607136, + 0.016185445711016655, + 0.037140507251024246, + 0.022886089980602264, + 0.007147890515625477, + 0.006914432626217604, + 0.006309487856924534, + 0.006008121185004711, + 0.007956665940582752, + 0.006186033133417368, + 0.009394651278853416, + 0.008674376644194126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.08904164284467697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on a rugged black volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wide viscous stream of molten rock cascading from fissures on the volcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the molten rock stream transition from a white-hot center to a deep blood-red rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pull out to reveal the magma stream as the dominant light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the magma stream shown initially diverging into tributaries before converging into a broad main stream?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862813353538513 + ], + "opposite_logprob_tokens": [ + -1.2236281633377075 + ], + "logprob_reward": -0.34862813353538513, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34862813353538513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.260427325963974, + 0.2644032835960388, + 0.05601517856121063, + 0.10095541924238205, + 0.03357943147420883, + 0.056157875806093216, + 0.05787672474980354, + 0.05770014226436615, + 0.07275758683681488, + 0.04012700542807579 + ], + "credit_map": [ + 0.08615140616893768, + 0.014207849279046059, + 0.0397246815264225, + 0.012951317243278027, + 0.010628425516188145, + 0.04724704846739769, + 0.018194708973169327, + 0.03132190927863121, + 0.03169555962085724, + 0.02096221037209034, + 0.04104398190975189, + 0.03034408763051033, + 0.02652040496468544, + 0.057737525552511215, + 0.02720910683274269, + 0.02889041043817997, + 0.005710405297577381, + 0.006443096790462732, + 0.008122616447508335, + 0.0074688163585960865, + 0.006577777210623026, + 0.0058432347141206264, + 0.00827729981392622, + 0.007571932394057512, + 0.00870272982865572, + 0.011971383355557919, + 0.01423606276512146, + 0.012065237388014793, + 0.015564003959298134, + 0.012469936162233353, + 0.013162057846784592, + 0.012784007005393505, + 0.0031868061050772667, + 0.0033811372704803944, + 0.005349316168576479, + 0.0030919101554900408, + 0.004503481090068817, + 0.004231894854456186, + 0.005851421505212784, + 0.00398346409201622, + 0.006159430835396051, + 0.00658249668776989, + 0.008045993745326996, + 0.0040649776346981525, + 0.010196504183113575, + 0.0059655471704900265, + 0.007712949998676777, + 0.007429972290992737, + 0.007080295588821173, + 0.005220477934926748, + 0.005234090145677328, + 0.003911709878593683, + 0.004956194199621677, + 0.006331050302833319, + 0.01642688550055027, + 0.008716021664440632, + 0.006157008931040764, + 0.008204656653106213, + 0.006450135726481676, + 0.005767818074673414, + 0.004965933505445719, + 0.008896942250430584, + 0.010462712496519089, + 0.006794932764023542, + 0.008463514037430286, + 0.006623525638133287, + 0.006668528076261282, + 0.0065222508274018764, + 0.010726714506745338, + 0.008433857932686806, + 0.015806088224053383, + 0.00951310619711876, + 0.0053387912921607494, + 0.004808822181075811, + 0.003437499748542905, + 0.003730600466951728, + 0.004695560317486525, + 0.004318404942750931, + 0.00909303780645132, + 0.004704288672655821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.34862813353538513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_15/index_tensor([1061, 1061])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_15/index_tensor([1061, 1061])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bca99a1f545257937ce5d54c47b6068686e46c52 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_15/index_tensor([1061, 1061])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, dynamic tracking shot, filmed from the side, follows an ancient, massive bronze chariot, ablaze with ghostly blue flames, as it soars at breathtaking speed across a vast, bottomless chasm. Shrouded in dusk, thin mist lingers across the canyon. The chariot's body, intricately carved with war and mythological totems, is engulfed in a shimmering blue flame, emitting no smoke and stretching backward in a fiery wake as it soars. At that moment, the chariot's massive bronze wheels rub violently against a sharp rock protruding from the canyon's edge, instantly unleashing a series of dazzling orange-yellow sparks. These sparks, like lightning, pierce the darkness, clearly illuminating the rough texture of the rock and the ancient inscriptions on the chariot's wheel hubs. The striking contrast between the cool blue halo and the fiery orange-yellow sparks creates an atmosphere of epic mythical grandeur, power, and danger.", + "seed": 10633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_15/index_1061_seed_10633_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14920291770249605, + "vlm_logprob": -0.14920291770249605, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.14920291770249605, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6350304484367371 + ], + "opposite_logprob_tokens": [ + -0.7600304484367371 + ], + "logprob_reward": -0.6350304484367371, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6350304484367371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18200087547302246, + 0.07317039370536804, + 0.18855351209640503, + 0.14155231416225433, + 0.13129115104675293, + 0.07193545252084732, + 0.0786217674612999, + 0.043208930641412735, + 0.02085460163652897, + 0.06881102919578552 + ], + "credit_map": [ + 0.0186405461281538, + 0.01351827010512352, + 0.026020746678113937, + 0.02714233100414276, + 0.021395692601799965, + 0.02023543231189251, + 0.02848130837082863, + 0.02656654082238674, + 0.007226387970149517, + 0.007265913300216198, + 0.010327199473977089, + 0.009552774019539356, + 0.007681819144636393, + 0.012389003299176693, + 0.010522819124162197, + 0.008204479701817036, + 0.023013750091195107, + 0.02911248989403248, + 0.030371027067303658, + 0.017565114423632622, + 0.017315613105893135, + 0.019741665571928024, + 0.023220758885145187, + 0.028213104233145714, + 0.012016157619655132, + 0.012178345583379269, + 0.020892955362796783, + 0.015427504666149616, + 0.02309873141348362, + 0.01249835267663002, + 0.022961214184761047, + 0.02247905544936657, + 0.010734978131949902, + 0.012542610988020897, + 0.016811488196253777, + 0.021978776901960373, + 0.011699793860316277, + 0.019427623599767685, + 0.020093047991394997, + 0.018002832308411598, + 0.0075340066105127335, + 0.0074171265587210655, + 0.009386186487972736, + 0.01044280081987381, + 0.006114516872912645, + 0.007178995758295059, + 0.014111821539700031, + 0.009749998338520527, + 0.006840222515165806, + 0.012311768718063831, + 0.006244075950235128, + 0.011126134544610977, + 0.013677661307156086, + 0.009657295420765877, + 0.008643139153718948, + 0.010121469385921955, + 0.0039328597486019135, + 0.0030736064072698355, + 0.008386220782995224, + 0.004038838669657707, + 0.004066415596753359, + 0.006066569127142429, + 0.005985661875456572, + 0.007658760063350201, + 0.003720267442986369, + 0.001936781802214682, + 0.0031250605825334787, + 0.0017435422632843256, + 0.0016809470253065228, + 0.002335486002266407, + 0.002494457410648465, + 0.003818058641627431, + 0.007192249409854412, + 0.005798023659735918, + 0.008025292307138443, + 0.0056710741482675076, + 0.004915760830044746, + 0.009657628834247589, + 0.009567570872604847, + 0.01798342727124691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.6350304484367371 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033725760877132416 + ], + "opposite_logprob_tokens": [ + -3.4087257385253906 + ], + "logprob_reward": -0.033725760877132416, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033725760877132416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11998102813959122, + 0.07400672137737274, + 0.04880554974079132, + 0.0667005181312561, + 0.10621106624603271, + 0.11924005299806595, + 0.05691562592983246, + 0.13455794751644135, + 0.05614549294114113, + 0.2174360156059265 + ], + "credit_map": [ + 0.011961206793785095, + 0.010312854312360287, + 0.023154597729444504, + 0.01743939332664013, + 0.016735123470425606, + 0.009320014156401157, + 0.015001552179455757, + 0.016056284308433533, + 0.0077948118560016155, + 0.006810960825532675, + 0.011185714974999428, + 0.007956222631037235, + 0.01136130653321743, + 0.00957748107612133, + 0.008085043169558048, + 0.011235183104872704, + 0.0064370520412921906, + 0.005723480135202408, + 0.0073378137312829494, + 0.004274176433682442, + 0.007013053633272648, + 0.005454563535749912, + 0.0064908345229923725, + 0.006074576172977686, + 0.005399092100560665, + 0.006027779076248407, + 0.016007451340556145, + 0.005655548069626093, + 0.0073906127363443375, + 0.004758450202643871, + 0.005791930481791496, + 0.015669656917452812, + 0.009824027307331562, + 0.009287097491323948, + 0.009966728277504444, + 0.019040290266275406, + 0.007579907309263945, + 0.013246498070657253, + 0.022133616730570793, + 0.01513290498405695, + 0.010993964038789272, + 0.01148860715329647, + 0.017018573358654976, + 0.021606434136629105, + 0.008058678358793259, + 0.015462886542081833, + 0.020074816420674324, + 0.014536102302372456, + 0.004101661499589682, + 0.006125143263489008, + 0.005968836601823568, + 0.006516017485409975, + 0.007983846589922905, + 0.011067706160247326, + 0.007647004909813404, + 0.007505408022552729, + 0.014898089691996574, + 0.010153776034712791, + 0.026007167994976044, + 0.013451182283461094, + 0.013233228586614132, + 0.017677363008260727, + 0.021456627175211906, + 0.017680516466498375, + 0.00502003962174058, + 0.003421869594603777, + 0.01067402958869934, + 0.0037932051345705986, + 0.005126100033521652, + 0.008665681816637516, + 0.006420514080673456, + 0.013024051673710346, + 0.02159879356622696, + 0.018461525440216064, + 0.021488483995199203, + 0.0186341293156147, + 0.025928661227226257, + 0.03116350993514061, + 0.029482534155249596, + 0.050678372383117676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.033725760877132416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969754070043564 + ], + "opposite_logprob_tokens": [ + -2.9299697875976562 + ], + "logprob_reward": -0.054969754070043564, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969754070043564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0835169330239296, + 0.15706606209278107, + 0.11104337871074677, + 0.05880219116806984, + 0.13604003190994263, + 0.10197574645280838, + 0.03162206709384918, + 0.10032255202531815, + 0.05494621768593788, + 0.1646648347377777 + ], + "credit_map": [ + 0.009440955705940723, + 0.006752020679414272, + 0.010383864864706993, + 0.012627660296857357, + 0.013107655569911003, + 0.00943734124302864, + 0.01047612726688385, + 0.01129130832850933, + 0.017704827710986137, + 0.012473209761083126, + 0.025207888334989548, + 0.02332075498998165, + 0.02089102193713188, + 0.020112603902816772, + 0.0214382391422987, + 0.015917513519525528, + 0.012027991935610771, + 0.01492683868855238, + 0.015291185118258, + 0.010082724504172802, + 0.011223645880818367, + 0.010284002870321274, + 0.019041279330849648, + 0.01816570572555065, + 0.0050938925705850124, + 0.004136126022785902, + 0.007939453236758709, + 0.007421278394758701, + 0.0068396166898310184, + 0.0047811539843678474, + 0.012468390166759491, + 0.010122277773916721, + 0.008899454027414322, + 0.012508770450949669, + 0.01160738617181778, + 0.02355966530740261, + 0.009509511291980743, + 0.024318818002939224, + 0.025954298675060272, + 0.019682105630636215, + 0.008653104305267334, + 0.006796893198043108, + 0.013813511468470097, + 0.02464139834046364, + 0.006387761328369379, + 0.008437376469373703, + 0.01762588508427143, + 0.015619813464581966, + 0.0034060056786984205, + 0.00439715851098299, + 0.0031332068610936403, + 0.005151662975549698, + 0.002963313600048423, + 0.004964872729033232, + 0.0031737482640892267, + 0.00443209707736969, + 0.015814047306776047, + 0.008742570877075195, + 0.016705207526683807, + 0.011452913284301758, + 0.007047795224934816, + 0.014449349604547024, + 0.010470868088304996, + 0.01563979685306549, + 0.007825438864529133, + 0.005182043183594942, + 0.009609187953174114, + 0.00472934590652585, + 0.006172624416649342, + 0.006018626969307661, + 0.008149770088493824, + 0.00725918123498559, + 0.018522638827562332, + 0.011842474341392517, + 0.01756330206990242, + 0.01592331752181053, + 0.011466999538242817, + 0.021859977394342422, + 0.02672743983566761, + 0.04075866937637329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.054969754070043564 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181954987347126 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.0181954987347126, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181954987347126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07630744576454163, + 0.0884523093700409, + 0.2231835424900055, + 0.10344576835632324, + 0.1060495376586914, + 0.19153377413749695, + 0.04411255940794945, + 0.07438937574625015, + 0.04316636919975281, + 0.049359291791915894 + ], + "credit_map": [ + 0.008182844147086143, + 0.0050502498634159565, + 0.011628636159002781, + 0.009573101066052914, + 0.008919619023799896, + 0.010571980848908424, + 0.0132957948371768, + 0.0090852165594697, + 0.008653881028294563, + 0.008641325868666172, + 0.01295748632401228, + 0.010694080963730812, + 0.010769321583211422, + 0.01610584929585457, + 0.009786194190382957, + 0.010844164527952671, + 0.02603922411799431, + 0.031029993668198586, + 0.03623658046126366, + 0.022267473861575127, + 0.022309666499495506, + 0.024694278836250305, + 0.02802472747862339, + 0.032581593841314316, + 0.009201645851135254, + 0.016349393874406815, + 0.012095226906239986, + 0.013408347964286804, + 0.009087014943361282, + 0.01268905308097601, + 0.01799064129590988, + 0.012624441646039486, + 0.010275228880345821, + 0.007608100306242704, + 0.010552211664617062, + 0.0179995559155941, + 0.006917894352227449, + 0.018330220133066177, + 0.019468579441308975, + 0.01489774789661169, + 0.01494517270475626, + 0.008408572524785995, + 0.05168784409761429, + 0.03036622889339924, + 0.010143604129552841, + 0.023477790877223015, + 0.02305855043232441, + 0.029446007683873177, + 0.0031462323386222124, + 0.005717834923416376, + 0.004940703045576811, + 0.0064174034632742405, + 0.006504859775304794, + 0.007138156797736883, + 0.004152802284806967, + 0.006094568409025669, + 0.00601884163916111, + 0.004679625853896141, + 0.014170549809932709, + 0.007306914776563644, + 0.0069274925626814365, + 0.0111025869846344, + 0.009810890071094036, + 0.014372478239238262, + 0.005061674863100052, + 0.0027386872097849846, + 0.006626579444855452, + 0.0031104467343539, + 0.004231947474181652, + 0.006143579725176096, + 0.006384947337210178, + 0.008868507109582424, + 0.0059614055790007114, + 0.00709955021739006, + 0.007816914469003677, + 0.00554908812046051, + 0.004275358282029629, + 0.005838166456669569, + 0.006537182256579399, + 0.006281629204750061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.0181954987347126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093126393854618 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093126393854618, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093126393854618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387423425912857, + 0.09221040457487106, + 0.16944624483585358, + 0.0939590334892273, + 0.10016264766454697, + 0.082121841609478, + 0.04532023146748543, + 0.0626048892736435, + 0.08075115084648132, + 0.13468113541603088 + ], + "credit_map": [ + 0.014414935372769833, + 0.0102921724319458, + 0.0211113840341568, + 0.017622368410229683, + 0.01724592223763466, + 0.013313799165189266, + 0.018888046965003014, + 0.025853728875517845, + 0.010069867596030235, + 0.009629885666072369, + 0.01144564151763916, + 0.013929787091910839, + 0.010942306369543076, + 0.011798144318163395, + 0.013945670798420906, + 0.01044910866767168, + 0.019278699532151222, + 0.027189617976546288, + 0.029787791892886162, + 0.01871386729180813, + 0.016052309423685074, + 0.014878160320222378, + 0.020099740475416183, + 0.02344607189297676, + 0.008858753368258476, + 0.007518322207033634, + 0.012903328984975815, + 0.0102549958974123, + 0.011079361662268639, + 0.00988321378827095, + 0.017958473414182663, + 0.015502595342695713, + 0.009687550365924835, + 0.010515640489757061, + 0.00870158988982439, + 0.015728192403912544, + 0.00911620631814003, + 0.013620549812912941, + 0.0190897099673748, + 0.013703216798603535, + 0.008775170892477036, + 0.005599218886345625, + 0.013916874304413795, + 0.01378463301807642, + 0.005729008466005325, + 0.012764779850840569, + 0.010389808565378189, + 0.011162353679537773, + 0.0035362245980650187, + 0.005646843928843737, + 0.0052147237583994865, + 0.006711320485919714, + 0.004252925515174866, + 0.008792019449174404, + 0.005239239428192377, + 0.00592693779617548, + 0.010327568277716637, + 0.00438695726916194, + 0.009650967083871365, + 0.005258915945887566, + 0.004691614303737879, + 0.008798551745712757, + 0.007842302322387695, + 0.011648016981780529, + 0.010344418697059155, + 0.007402613293379545, + 0.013450167141854763, + 0.00740937190130353, + 0.0074198585934937, + 0.009203372523188591, + 0.010085580870509148, + 0.015435776673257351, + 0.01988578401505947, + 0.03151252120733261, + 0.013330942951142788, + 0.013583551160991192, + 0.014093280769884586, + 0.013951304368674755, + 0.014762700535356998, + 0.013561060652136803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.004093126393854618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot, filmed from the side, follows an ancient, massive bronze chariot, ablaze with ghostly blue flames, as it soars at breathtaking speed across a vast, bottomless chasm. Shrouded in dusk, thin mist lingers across the canyon. The chariot's body, intricately carved with war and mythological totems, is engulfed in a shimmering blue flame, emitting no smoke and stretching backward in a fiery wake as it soars. At that moment, the chariot's massive bronze wheels rub violently against a sharp rock protruding from the canyon's edge, instantly unleashing a series of dazzling orange-yellow sparks. These sparks, like lightning, pierce the darkness, clearly illuminating the rough texture of the rock and the ancient inscriptions on the chariot's wheel hubs. The striking contrast between the cool blue halo and the fiery orange-yellow sparks creates an atmosphere of epic mythical grandeur, power, and danger.", + "seed": 10634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_15/index_1061_seed_10634_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.030336236860603095, + "vlm_logprob": -0.030336236860603095, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.030336236860603095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08113560825586319 + ], + "opposite_logprob_tokens": [ + -2.5811355113983154 + ], + "logprob_reward": -0.08113560825586319, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08113560825586319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4122876524925232, + 0.04159775748848915, + 0.041505202651023865, + 0.06590201705694199, + 0.03413731977343559, + 0.06255203485488892, + 0.152584046125412, + 0.05660170316696167, + 0.03099900111556053, + 0.10183332860469818 + ], + "credit_map": [ + 0.032672107219696045, + 0.034406255930662155, + 0.0952930748462677, + 0.07827260345220566, + 0.029276302084326744, + 0.0823177918791771, + 0.029935477301478386, + 0.03011401928961277, + 0.005534524098038673, + 0.004246114753186703, + 0.005610904656350613, + 0.005406288895756006, + 0.0051708705723285675, + 0.004556695464998484, + 0.006333640776574612, + 0.004738716408610344, + 0.004427800886332989, + 0.005451268516480923, + 0.0047386279329657555, + 0.0058437129482626915, + 0.00807343889027834, + 0.0037181065417826176, + 0.005904199089854956, + 0.00334804505109787, + 0.00923170242458582, + 0.004999618511646986, + 0.007405966520309448, + 0.006677266210317612, + 0.007160940207540989, + 0.008928933180868626, + 0.011404840275645256, + 0.010092743672430515, + 0.005055542103946209, + 0.003581173485144973, + 0.0040786718018352985, + 0.007054758258163929, + 0.0030732490122318268, + 0.002724991412833333, + 0.005530547816306353, + 0.0030383849516510963, + 0.00810397882014513, + 0.006010842509567738, + 0.009131486527621746, + 0.011444348841905594, + 0.005650998093187809, + 0.009203354828059673, + 0.005839057732373476, + 0.007167965639382601, + 0.01455763541162014, + 0.011515422724187374, + 0.034082744270563126, + 0.015685582533478737, + 0.015189685858786106, + 0.028994424268603325, + 0.012845033779740334, + 0.01971350982785225, + 0.004285251256078482, + 0.0069040958769619465, + 0.012298877350986004, + 0.008770888671278954, + 0.008007703348994255, + 0.005920754745602608, + 0.005473482888191938, + 0.004940647166222334, + 0.002649080939590931, + 0.004648360889405012, + 0.003198472084477544, + 0.0037620337679982185, + 0.0039021125994622707, + 0.004166199825704098, + 0.00386299891397357, + 0.004809742793440819, + 0.01237267255783081, + 0.015885241329669952, + 0.011005928739905357, + 0.009198346175253391, + 0.014655891805887222, + 0.017785653471946716, + 0.011612297967076302, + 0.009317299351096153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.08113560825586319 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038121629506349564 + ], + "opposite_logprob_tokens": [ + -3.2881217002868652 + ], + "logprob_reward": -0.038121629506349564, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038121629506349564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5179401636123657, + 0.06721299141645432, + 0.0904904454946518, + 0.08610064536333084, + 0.03774968907237053, + 0.030637193471193314, + 0.02856171689927578, + 0.026616234332323074, + 0.02314894087612629, + 0.09154188632965088 + ], + "credit_map": [ + 0.030736485496163368, + 0.03531759977340698, + 0.12495602667331696, + 0.10156117379665375, + 0.03516515716910362, + 0.11378995329141617, + 0.033052511513233185, + 0.04336134344339371, + 0.010518738999962807, + 0.008862253278493881, + 0.009800744242966175, + 0.009208840318024158, + 0.007953654043376446, + 0.0059900605119764805, + 0.00921123567968607, + 0.005667474586516619, + 0.007770287338644266, + 0.01124155055731535, + 0.010173162445425987, + 0.009222653694450855, + 0.022845443338155746, + 0.010157129727303982, + 0.010899662971496582, + 0.008180568926036358, + 0.011130266822874546, + 0.0060740066692233086, + 0.01038057915866375, + 0.008570726960897446, + 0.008777707815170288, + 0.015070749446749687, + 0.011022336781024933, + 0.0150742894038558, + 0.005504317115992308, + 0.0029839163180440664, + 0.005086050368845463, + 0.006051099859178066, + 0.004405748099088669, + 0.0055021545849740505, + 0.004673844203352928, + 0.0035425624810159206, + 0.0028076819144189358, + 0.0029957129154354334, + 0.004390229936689138, + 0.0049547916278243065, + 0.0034423386678099632, + 0.0045666443184018135, + 0.0038290382362902164, + 0.0036507609765976667, + 0.0027267043478786945, + 0.0021591661497950554, + 0.004656369332224131, + 0.002909816335886717, + 0.003180011408403516, + 0.005511630792170763, + 0.003510852577164769, + 0.003907172009348869, + 0.002327765105292201, + 0.0031601516529917717, + 0.0029269736260175705, + 0.0031914745923131704, + 0.003400063142180443, + 0.004304392263293266, + 0.004265031311661005, + 0.0030403879936784506, + 0.0026793256402015686, + 0.0031338082626461983, + 0.0020351004786789417, + 0.0024193688295781612, + 0.003049880964681506, + 0.004928968381136656, + 0.0025880367029458284, + 0.0023144567385315895, + 0.009170682169497013, + 0.016377093270421028, + 0.010540592484176159, + 0.009717065840959549, + 0.012609385885298252, + 0.011192651465535164, + 0.012356146238744259, + 0.0095782820135355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.038121629506349564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608150638639927 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.01608150638639927, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608150638639927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33164140582084656, + 0.08738528192043304, + 0.24001671373844147, + 0.03644118830561638, + 0.026113023981451988, + 0.032040584832429886, + 0.025651561096310616, + 0.05564926564693451, + 0.06298050284385681, + 0.10208041220903397 + ], + "credit_map": [ + 0.017975185066461563, + 0.017614681273698807, + 0.08912625908851624, + 0.07109176367521286, + 0.019075272604823112, + 0.0759957879781723, + 0.01671275682747364, + 0.024049745872616768, + 0.017252163961529732, + 0.010061511769890785, + 0.01482458971440792, + 0.01582331396639347, + 0.007586691994220018, + 0.007646095473319292, + 0.008062898181378841, + 0.006128033623099327, + 0.02095465548336506, + 0.031515493988990784, + 0.026474032551050186, + 0.022759152576327324, + 0.062184855341911316, + 0.025708910077810287, + 0.028715750202536583, + 0.021703874692320824, + 0.00443411897867918, + 0.0028773006051778793, + 0.0054841917008161545, + 0.0043641068041324615, + 0.004807118326425552, + 0.006308491341769695, + 0.00414949469268322, + 0.004016369115561247, + 0.0033804266713559628, + 0.002786237047985196, + 0.003885740414261818, + 0.003912368789315224, + 0.0025976980105042458, + 0.0026839557103812695, + 0.003619173774495721, + 0.0032474249601364136, + 0.004120796453207731, + 0.0031787229236215353, + 0.004581590183079243, + 0.004073765594512224, + 0.00351439299993217, + 0.00573060754686594, + 0.0037074305582791567, + 0.0031332792714238167, + 0.0024839888792485, + 0.00231760716997087, + 0.005540065933018923, + 0.0029186084866523743, + 0.002743080025538802, + 0.0039374190382659435, + 0.0023672364186495543, + 0.003343557007610798, + 0.004858263302594423, + 0.006772980093955994, + 0.008153029717504978, + 0.0068209427408874035, + 0.0072408635169267654, + 0.007845539599657059, + 0.007437265478074551, + 0.006520385388284922, + 0.005179065279662609, + 0.008011949248611927, + 0.006261293776333332, + 0.007691932842135429, + 0.008533420972526073, + 0.012999841943383217, + 0.006529751233756542, + 0.007773251738399267, + 0.010728654451668262, + 0.014053590595722198, + 0.011013181880116463, + 0.007347367703914642, + 0.013112514279782772, + 0.023509889841079712, + 0.00839092768728733, + 0.01392428670078516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.01608150638639927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089949868619442 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089949868619442, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089949868619442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39416375756263733, + 0.06325256079435349, + 0.13901780545711517, + 0.032020267099142075, + 0.031114783138036728, + 0.09371799975633621, + 0.04166494309902191, + 0.09384504705667496, + 0.06004106625914574, + 0.051161736249923706 + ], + "credit_map": [ + 0.03169865533709526, + 0.0337655283510685, + 0.09069446474313736, + 0.08834472298622131, + 0.028150327503681183, + 0.0703015848994255, + 0.025871379300951958, + 0.02533709444105625, + 0.012722810730338097, + 0.007229225244373083, + 0.009785030968487263, + 0.0076628695242106915, + 0.006836132146418095, + 0.0050643812865018845, + 0.007892217487096786, + 0.00605989433825016, + 0.013663865625858307, + 0.019023442640900612, + 0.014721191488206387, + 0.01233876496553421, + 0.03614051640033722, + 0.013450573198497295, + 0.016499202698469162, + 0.01318025216460228, + 0.0042470283806324005, + 0.0027568466030061245, + 0.0040268381126224995, + 0.0037876872811466455, + 0.003411043668165803, + 0.003937583416700363, + 0.004877221770584583, + 0.004976020194590092, + 0.004938225261867046, + 0.003494290402159095, + 0.004768339451402426, + 0.005034772679209709, + 0.0029496990609914064, + 0.003190089715644717, + 0.0031729170586913824, + 0.003566451370716095, + 0.00931151770055294, + 0.010790071450173855, + 0.011921203695237637, + 0.019682059064507484, + 0.009851712733507156, + 0.011954658664762974, + 0.01038332562893629, + 0.009823452681303024, + 0.004735629074275494, + 0.003113236278295517, + 0.007528319954872131, + 0.0037766108289361, + 0.0046704839915037155, + 0.008051054552197456, + 0.0045674447901546955, + 0.005222163628786802, + 0.008149492554366589, + 0.009854610078036785, + 0.02117442712187767, + 0.010338339023292065, + 0.013579991646111012, + 0.012549635022878647, + 0.00929561909288168, + 0.008902939967811108, + 0.006267550401389599, + 0.010450014844536781, + 0.007087749429047108, + 0.006754798348993063, + 0.00935202743858099, + 0.008871078491210938, + 0.005465509369969368, + 0.005792340263724327, + 0.006529233418405056, + 0.010782436467707157, + 0.004742311779409647, + 0.00365334446541965, + 0.009816677309572697, + 0.007949035614728928, + 0.003928780555725098, + 0.003759917337447405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.011089949868619442 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3594486713409424, + 0.07336758822202682, + 0.10358379036188126, + 0.037931762635707855, + 0.035341933369636536, + 0.042817119508981705, + 0.1689966768026352, + 0.05490269884467125, + 0.03827196732163429, + 0.08533777296543121 + ], + "credit_map": [ + 0.030949514359235764, + 0.03095167689025402, + 0.07484642416238785, + 0.12328958511352539, + 0.02294698916375637, + 0.025384388864040375, + 0.02756081335246563, + 0.023519281297922134, + 0.012512916699051857, + 0.009240546263754368, + 0.012254560366272926, + 0.011278669349849224, + 0.006423356011509895, + 0.00684500252828002, + 0.007335432339459658, + 0.007477105129510164, + 0.012188848108053207, + 0.013236427679657936, + 0.011731714010238647, + 0.010051315650343895, + 0.02250346913933754, + 0.012043490074574947, + 0.013376655988395214, + 0.008451865054666996, + 0.004029037430882454, + 0.003926226403564215, + 0.0033208681270480156, + 0.004109795670956373, + 0.007129443809390068, + 0.004766370169818401, + 0.006798723712563515, + 0.003851299174129963, + 0.00605802284553647, + 0.002743435325101018, + 0.005190112628042698, + 0.005932276137173176, + 0.003377671120688319, + 0.004209224134683609, + 0.004152250941842794, + 0.0036789372097700834, + 0.005079213529825211, + 0.005719153210520744, + 0.006129930727183819, + 0.007601223420351744, + 0.004913285840302706, + 0.004565291106700897, + 0.00395837239921093, + 0.004850648809224367, + 0.01194220595061779, + 0.011156830936670303, + 0.029388006776571274, + 0.019272757694125175, + 0.020821502432227135, + 0.031666677445173264, + 0.020070290192961693, + 0.024678384885191917, + 0.00535823218524456, + 0.006201361771672964, + 0.009212893433868885, + 0.00944927241653204, + 0.008126756176352501, + 0.0070681520737707615, + 0.005090927705168724, + 0.0043951072730124, + 0.004091830924153328, + 0.005980170797556639, + 0.0033659529872238636, + 0.005110473372042179, + 0.00599656393751502, + 0.005885259713977575, + 0.003277248004451394, + 0.004564465954899788, + 0.011896681040525436, + 0.012467284686863422, + 0.008480286225676537, + 0.006727728992700577, + 0.010946456342935562, + 0.0171232670545578, + 0.008957303129136562, + 0.008738766424357891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.005252490285784006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot, filmed from the side, follows an ancient, massive bronze chariot, ablaze with ghostly blue flames, as it soars at breathtaking speed across a vast, bottomless chasm. Shrouded in dusk, thin mist lingers across the canyon. The chariot's body, intricately carved with war and mythological totems, is engulfed in a shimmering blue flame, emitting no smoke and stretching backward in a fiery wake as it soars. At that moment, the chariot's massive bronze wheels rub violently against a sharp rock protruding from the canyon's edge, instantly unleashing a series of dazzling orange-yellow sparks. These sparks, like lightning, pierce the darkness, clearly illuminating the rough texture of the rock and the ancient inscriptions on the chariot's wheel hubs. The striking contrast between the cool blue halo and the fiery orange-yellow sparks creates an atmosphere of epic mythical grandeur, power, and danger.", + "seed": 10635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_15/index_1061_seed_10635_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011917308811098338, + "vlm_logprob": -0.011917308811098338, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.011917308811098338, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035634998232126236 + ], + "opposite_logprob_tokens": [ + -3.410634994506836 + ], + "logprob_reward": -0.035634998232126236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.035634998232126236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2920399606227875, + 0.07734613120555878, + 0.05535854026675224, + 0.03931606560945511, + 0.01813737116754055, + 0.06862290948629379, + 0.07749532163143158, + 0.09353701770305634, + 0.21332867443561554, + 0.06481802463531494 + ], + "credit_map": [ + 0.022692926228046417, + 0.04158083721995354, + 0.04774056002497673, + 0.06482124328613281, + 0.030074847862124443, + 0.01886301301419735, + 0.0340033583343029, + 0.03226318955421448, + 0.007600598968565464, + 0.016349483281373978, + 0.009770289063453674, + 0.014821884222328663, + 0.006021083332598209, + 0.007036596070975065, + 0.006320216227322817, + 0.00942598469555378, + 0.009235925041139126, + 0.007458304986357689, + 0.0061074052937328815, + 0.006734779570251703, + 0.0048353830352425575, + 0.007501153275370598, + 0.005065832287073135, + 0.00841976236552, + 0.0049375263042747974, + 0.004542258568108082, + 0.004806548357009888, + 0.004489085171371698, + 0.004682363010942936, + 0.004566902760416269, + 0.005269042681902647, + 0.006022342946380377, + 0.002012992510572076, + 0.0020003002136945724, + 0.002192254876717925, + 0.0017448905855417252, + 0.00160791608504951, + 0.0027077691629529, + 0.003457731334492564, + 0.002413517329841852, + 0.004457601811736822, + 0.0044897557236254215, + 0.005505541805177927, + 0.00817171111702919, + 0.006777889095246792, + 0.009673280641436577, + 0.01763489842414856, + 0.011912244372069836, + 0.006195543333888054, + 0.007953313179314137, + 0.007300809025764465, + 0.009010945446789265, + 0.008954138495028019, + 0.015428624115884304, + 0.01012236438691616, + 0.012529588304460049, + 0.009299121797084808, + 0.01186328660696745, + 0.014774894341826439, + 0.011990622617304325, + 0.009611357934772968, + 0.011203229427337646, + 0.011322308331727982, + 0.013472205959260464, + 0.015056908130645752, + 0.013985968194901943, + 0.04548110067844391, + 0.038875602185726166, + 0.016578318551182747, + 0.0308685302734375, + 0.025725528597831726, + 0.026756737381219864, + 0.007658748887479305, + 0.010146219283342361, + 0.008248414844274521, + 0.010488414205610752, + 0.007706817705184221, + 0.007178599946200848, + 0.007263998035341501, + 0.006126822903752327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.035634998232126236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081741079688072 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.016081741079688072, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081741079688072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3836677074432373, + 0.06820948421955109, + 0.0748252421617508, + 0.024579869583249092, + 0.02237587608397007, + 0.052805908024311066, + 0.02942381054162979, + 0.06668214499950409, + 0.19831839203834534, + 0.07911160588264465 + ], + "credit_map": [ + 0.04022860899567604, + 0.0853210911154747, + 0.11770075559616089, + 0.07066081464290619, + 0.015494050458073616, + 0.014909514226019382, + 0.02009570226073265, + 0.01925715245306492, + 0.005403764545917511, + 0.018369892612099648, + 0.011355728842318058, + 0.019089631736278534, + 0.0034544640220701694, + 0.003142555709928274, + 0.00297371088527143, + 0.004419735167175531, + 0.015069584362208843, + 0.010817327536642551, + 0.010927075520157814, + 0.008675480261445045, + 0.005349378567188978, + 0.008187972009181976, + 0.005580289289355278, + 0.010218129493296146, + 0.0032150980550795794, + 0.0030837038066238165, + 0.0032385915983468294, + 0.0032707073260098696, + 0.002179856179282069, + 0.0024380057584494352, + 0.0029966887086629868, + 0.004157218616455793, + 0.0019591774325817823, + 0.0015839299885556102, + 0.0021300597582012415, + 0.002359159290790558, + 0.002355118514969945, + 0.002450980944558978, + 0.00635722978040576, + 0.0031802200246602297, + 0.0027198679745197296, + 0.002803829265758395, + 0.002885503461584449, + 0.004314331337809563, + 0.0037469924427568913, + 0.013613734394311905, + 0.008440425619482994, + 0.014281224459409714, + 0.0029395977035164833, + 0.0030643269419670105, + 0.002636266639456153, + 0.003864245256409049, + 0.0037883126642555, + 0.00440450431779027, + 0.004708598367869854, + 0.004017957486212254, + 0.007313730660825968, + 0.008665312081575394, + 0.008181173354387283, + 0.006365629844367504, + 0.008354668505489826, + 0.006804128643125296, + 0.011412190273404121, + 0.0095853041857481, + 0.011874846182763577, + 0.011250381357967854, + 0.05093074217438698, + 0.039496563374996185, + 0.012045592069625854, + 0.026369718834757805, + 0.023296309635043144, + 0.02305423840880394, + 0.010811529122292995, + 0.011811071075499058, + 0.00948124285787344, + 0.009931117296218872, + 0.010542468167841434, + 0.011182229034602642, + 0.004495391622185707, + 0.010856554843485355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.016081741079688072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001719426247291267 + ], + "opposite_logprob_tokens": [ + -6.3767194747924805 + ], + "logprob_reward": -0.001719426247291267, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001719426247291267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09392829239368439, + 0.04638853669166565, + 0.021707763895392418, + 0.02087942510843277, + 0.011486008763313293, + 0.02905416674911976, + 0.0436234176158905, + 0.02412928268313408, + 0.6569172739982605, + 0.05188586190342903 + ], + "credit_map": [ + 0.009821918793022633, + 0.015966931357979774, + 0.02448960393667221, + 0.017718080431222916, + 0.0063796634785830975, + 0.0049639237113296986, + 0.007893296889960766, + 0.0066948700696229935, + 0.005276719108223915, + 0.010982411913573742, + 0.009310392662882805, + 0.007737398147583008, + 0.003392522456124425, + 0.0031670243479311466, + 0.002354671712964773, + 0.004167395643889904, + 0.0031993782613426447, + 0.004217932466417551, + 0.0022694419603794813, + 0.003052776213735342, + 0.0024273921735584736, + 0.0025470261462032795, + 0.0017606690526008606, + 0.0022331455256789923, + 0.0035666190087795258, + 0.0024187795352190733, + 0.002710146363824606, + 0.0021594727877527475, + 0.001572367618791759, + 0.003070262959226966, + 0.0016401351895183325, + 0.003741641528904438, + 0.0010901918867602944, + 0.0009800653206184506, + 0.0015353785129263997, + 0.000995053444057703, + 0.0011510939802974463, + 0.0013655361253768206, + 0.0027227005921304226, + 0.0016459891339764, + 0.001420511631295085, + 0.0017381050856783986, + 0.001675211708061397, + 0.002262693829834461, + 0.0024701652582734823, + 0.0066161984577775, + 0.005100481677800417, + 0.007770799100399017, + 0.004685900639742613, + 0.0052310023456811905, + 0.0036708766128867865, + 0.004280820954591036, + 0.0044698528945446014, + 0.007474284619092941, + 0.007711883168667555, + 0.006098797079175711, + 0.00260526523925364, + 0.0025493456050753593, + 0.002628491260111332, + 0.0028797828126698732, + 0.0028463953640311956, + 0.003340673167258501, + 0.003699628636240959, + 0.0035796985030174255, + 0.037866733968257904, + 0.04670392721891403, + 0.1444990038871765, + 0.12298373132944107, + 0.05501065403223038, + 0.09419490396976471, + 0.07781372219324112, + 0.07784460484981537, + 0.006955002434551716, + 0.008651607669889927, + 0.006846032105386257, + 0.007005716674029827, + 0.0055024391040205956, + 0.003906785044819117, + 0.007112364750355482, + 0.005905916914343834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.001719426247291267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001515312702395022 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.001515312702395022, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001515312702395022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20037150382995605, + 0.07109954953193665, + 0.05167364701628685, + 0.07205551117658615, + 0.04258153587579727, + 0.02181055024266243, + 0.09563729912042618, + 0.049272093921899796, + 0.32915470004081726, + 0.06634353846311569 + ], + "credit_map": [ + 0.015629781410098076, + 0.02841176837682724, + 0.029170149937272072, + 0.04340243712067604, + 0.020135462284088135, + 0.01600324735045433, + 0.021207567304372787, + 0.026411112397909164, + 0.006193587090820074, + 0.013167527504265308, + 0.0136047238484025, + 0.011776771396398544, + 0.004044904373586178, + 0.004694937262684107, + 0.005353064276278019, + 0.012264038436114788, + 0.008274601772427559, + 0.008002684451639652, + 0.0067106448113918304, + 0.008438760414719582, + 0.00521497568115592, + 0.004933987278491259, + 0.004963352810591459, + 0.005134647712111473, + 0.009639140218496323, + 0.009861912578344345, + 0.009208925068378448, + 0.013227337040007114, + 0.005934697110205889, + 0.007602882105857134, + 0.00676392950117588, + 0.009816699661314487, + 0.003626825287938118, + 0.003225908847525716, + 0.00324293808080256, + 0.004449422936886549, + 0.003973815590143204, + 0.005289508495479822, + 0.012632299214601517, + 0.006140820682048798, + 0.001695988466963172, + 0.0015746690332889557, + 0.0014139458071440458, + 0.0025118605699390173, + 0.002235732041299343, + 0.0037110652774572372, + 0.004503550939261913, + 0.004163741134107113, + 0.008818881586194038, + 0.01012908760458231, + 0.013179023750126362, + 0.014094052836298943, + 0.009689751081168652, + 0.011958400718867779, + 0.013969570398330688, + 0.013798541389405727, + 0.005607079714536667, + 0.006777530536055565, + 0.0061632185243070126, + 0.005501183215528727, + 0.00433308957144618, + 0.007066863588988781, + 0.006669765338301659, + 0.007153370417654514, + 0.01897173747420311, + 0.02559550292789936, + 0.0710432231426239, + 0.0595768466591835, + 0.03198975697159767, + 0.046104926615953445, + 0.039096929132938385, + 0.03677580505609512, + 0.009855936281383038, + 0.010442372411489487, + 0.007413655519485474, + 0.007566938642412424, + 0.005718804430216551, + 0.006138679105788469, + 0.007490601856261492, + 0.011716566048562527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.001515312702395022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22392632067203522, + 0.12848976254463196, + 0.05754503607749939, + 0.0954950749874115, + 0.05408596247434616, + 0.06549347192049026, + 0.059236422181129456, + 0.06300562620162964, + 0.2252710908651352, + 0.02745124138891697 + ], + "credit_map": [ + 0.02352575585246086, + 0.039126332849264145, + 0.04866151884198189, + 0.041376348584890366, + 0.017938410863280296, + 0.01624208316206932, + 0.0180222075432539, + 0.01903367042541504, + 0.007987314835190773, + 0.031825508922338486, + 0.015261122025549412, + 0.03694641962647438, + 0.007168784737586975, + 0.00909519288688898, + 0.008171305060386658, + 0.012034113518893719, + 0.010858266614377499, + 0.006884344387799501, + 0.007907608523964882, + 0.008540060371160507, + 0.0052414024248719215, + 0.004582279361784458, + 0.005545977968722582, + 0.007985098287463188, + 0.012868259102106094, + 0.010770819149911404, + 0.011529195122420788, + 0.016007134690880775, + 0.008832911029458046, + 0.010925307869911194, + 0.011142920702695847, + 0.013418527320027351, + 0.006628383882343769, + 0.004802749026566744, + 0.005398083012551069, + 0.005504942499101162, + 0.005438411608338356, + 0.006277584470808506, + 0.012836539186537266, + 0.007199267856776714, + 0.003301723627373576, + 0.002388901775702834, + 0.0034656920470297337, + 0.002981768222525716, + 0.0027511660009622574, + 0.02278035879135132, + 0.0044179209508001804, + 0.02340594306588173, + 0.004963806364685297, + 0.005531858652830124, + 0.005134031642228365, + 0.006468848790973425, + 0.006221711169928312, + 0.011674655601382256, + 0.00891206692904234, + 0.010329446755349636, + 0.007114002015441656, + 0.007329072803258896, + 0.007584108505398035, + 0.007210708688944578, + 0.007137325592339039, + 0.006402504164725542, + 0.011033344082534313, + 0.009194553829729557, + 0.012451786547899246, + 0.015276911668479443, + 0.050770338624715805, + 0.04235507920384407, + 0.016851305961608887, + 0.03232638165354729, + 0.028174851089715958, + 0.027064427733421326, + 0.003467381466180086, + 0.005205010063946247, + 0.004086410626769066, + 0.00269793882034719, + 0.0026604202575981617, + 0.002087386790663004, + 0.0025023026391863823, + 0.0047443886287510395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.004635065793991089 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot, filmed from the side, follows an ancient, massive bronze chariot, ablaze with ghostly blue flames, as it soars at breathtaking speed across a vast, bottomless chasm. Shrouded in dusk, thin mist lingers across the canyon. The chariot's body, intricately carved with war and mythological totems, is engulfed in a shimmering blue flame, emitting no smoke and stretching backward in a fiery wake as it soars. At that moment, the chariot's massive bronze wheels rub violently against a sharp rock protruding from the canyon's edge, instantly unleashing a series of dazzling orange-yellow sparks. These sparks, like lightning, pierce the darkness, clearly illuminating the rough texture of the rock and the ancient inscriptions on the chariot's wheel hubs. The striking contrast between the cool blue halo and the fiery orange-yellow sparks creates an atmosphere of epic mythical grandeur, power, and danger.", + "seed": 10636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_15/index_1061_seed_10636_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04399398239329457, + "vlm_logprob": -0.04399398239329457, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04399398239329457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16421397030353546 + ], + "opposite_logprob_tokens": [ + -1.914214015007019 + ], + "logprob_reward": -0.16421397030353546, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16421397030353546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19502949714660645, + 0.0643976703286171, + 0.05934145301580429, + 0.13504768908023834, + 0.05984771251678467, + 0.1145205870270729, + 0.04285155236721039, + 0.01930147595703602, + 0.13182638585567474, + 0.17783598601818085 + ], + "credit_map": [ + 0.018652157858014107, + 0.02396601065993309, + 0.027188047766685486, + 0.025768933817744255, + 0.025341739878058434, + 0.024393079802393913, + 0.019966132938861847, + 0.029753388836979866, + 0.010290440171957016, + 0.008238933980464935, + 0.008415882475674152, + 0.007513119373470545, + 0.006734634283930063, + 0.004945334978401661, + 0.010630816221237183, + 0.0076285069808363914, + 0.0073469290509819984, + 0.005645802244544029, + 0.006966397166252136, + 0.009295212104916573, + 0.007741772569715977, + 0.005682302638888359, + 0.008203990757465363, + 0.00845904927700758, + 0.01613927260041237, + 0.008834007196128368, + 0.013051951304078102, + 0.017433583736419678, + 0.011680129915475845, + 0.009871399961411953, + 0.03476715832948685, + 0.023270193487405777, + 0.005124746356159449, + 0.0060486095026135445, + 0.008264150470495224, + 0.010773452930152416, + 0.007302018813788891, + 0.007930275052785873, + 0.009213133715093136, + 0.005191326607018709, + 0.008931946009397507, + 0.010637897066771984, + 0.014031829312443733, + 0.019250094890594482, + 0.01542110089212656, + 0.018744923174381256, + 0.016722826287150383, + 0.0107799731194973, + 0.003846314735710621, + 0.004138678777962923, + 0.006195767782628536, + 0.004597598221153021, + 0.006796783767640591, + 0.008773901499807835, + 0.004594217054545879, + 0.0039082919247448444, + 0.0017191896913573146, + 0.0024528687354177237, + 0.0027374280616641045, + 0.002366652013733983, + 0.0034612591844052076, + 0.0014087521703913808, + 0.0021173646673560143, + 0.003037960734218359, + 0.017965173348784447, + 0.016140848398208618, + 0.016604233533143997, + 0.014286825433373451, + 0.025658177211880684, + 0.018406445160508156, + 0.011356555856764317, + 0.011408133432269096, + 0.014901743270456791, + 0.04118562489748001, + 0.04213656857609749, + 0.02211490273475647, + 0.02151620201766491, + 0.011497939005494118, + 0.013007084839046001, + 0.011475925333797932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.16421397030353546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033734407275915146 + ], + "opposite_logprob_tokens": [ + -3.4087343215942383 + ], + "logprob_reward": -0.033734407275915146, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033734407275915146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16413052380084991, + 0.07632942497730255, + 0.15176443755626678, + 0.05445452407002449, + 0.030473927035927773, + 0.05013351887464523, + 0.026128508150577545, + 0.061465784907341, + 0.11501941084861755, + 0.270099937915802 + ], + "credit_map": [ + 0.011550034396350384, + 0.016777457669377327, + 0.016416436061263084, + 0.015382843092083931, + 0.0220765620470047, + 0.019069423899054527, + 0.025440113618969917, + 0.03741765767335892, + 0.0072098481468856335, + 0.0071823312900960445, + 0.012500095181167126, + 0.007874666713178158, + 0.008591599762439728, + 0.011365949176251888, + 0.01155884936451912, + 0.010046089068055153, + 0.015735775232315063, + 0.03372904285788536, + 0.013653484173119068, + 0.021219683811068535, + 0.013263128697872162, + 0.0141072953119874, + 0.016821086406707764, + 0.023234950378537178, + 0.006850932724773884, + 0.006356461439281702, + 0.0066762096248567104, + 0.005004498641937971, + 0.006002053618431091, + 0.004748210776597261, + 0.0065046148374676704, + 0.012311547994613647, + 0.0024777508806437254, + 0.003805433865636587, + 0.003707021242007613, + 0.003430264536291361, + 0.005390985403209925, + 0.0032175013329833746, + 0.006070407573133707, + 0.0023745631333440542, + 0.004817259963601828, + 0.005160235334187746, + 0.006261148024350405, + 0.007595748640596867, + 0.0075951372273266315, + 0.00562490988522768, + 0.007912318222224712, + 0.005166760645806789, + 0.0030014878138899803, + 0.002391721121966839, + 0.0038399642799049616, + 0.0033571766689419746, + 0.004891121760010719, + 0.004132484085857868, + 0.0020023563411086798, + 0.0025121967773884535, + 0.005671069025993347, + 0.007633064407855272, + 0.009114371612668037, + 0.007398487534373999, + 0.010514412075281143, + 0.005260096862912178, + 0.007870489731431007, + 0.008003794588148594, + 0.015396819449961185, + 0.013491561636328697, + 0.018158122897148132, + 0.023312022909522057, + 0.01143124420195818, + 0.009394608438014984, + 0.011156649328768253, + 0.01267838291823864, + 0.022835085168480873, + 0.03362101688981056, + 0.044403016567230225, + 0.05241117998957634, + 0.02971249632537365, + 0.02854304574429989, + 0.03553580492734909, + 0.02303829975426197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.033734407275915146 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007644092198461294 + ], + "opposite_logprob_tokens": [ + -4.882644176483154 + ], + "logprob_reward": -0.007644092198461294, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007644092198461294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0649174153804779, + 0.0658818706870079, + 0.021852226927876472, + 0.059922754764556885, + 0.03442695364356041, + 0.05968679115176201, + 0.020171944051980972, + 0.05574579909443855, + 0.12299945205450058, + 0.49439480900764465 + ], + "credit_map": [ + 0.0039043580181896687, + 0.007706626784056425, + 0.006085412111133337, + 0.006339632906019688, + 0.008520837873220444, + 0.010267004370689392, + 0.010454163886606693, + 0.011639381758868694, + 0.007554219104349613, + 0.00696986448019743, + 0.009649706073105335, + 0.006573955528438091, + 0.0067352671176195145, + 0.007411852478981018, + 0.008421053178608418, + 0.012565954588353634, + 0.002331622177734971, + 0.004487277008593082, + 0.002090164925903082, + 0.0023694790434092283, + 0.0024589160457253456, + 0.002521029906347394, + 0.0024647328536957502, + 0.003129007061943412, + 0.00798087753355503, + 0.005327204708009958, + 0.0066171493381261826, + 0.008164671249687672, + 0.005833112169057131, + 0.004150308668613434, + 0.010928940959274769, + 0.010920489206910133, + 0.003163931192830205, + 0.004087191075086594, + 0.004504145588725805, + 0.004705528728663921, + 0.005040484014898539, + 0.004297640174627304, + 0.005764612462371588, + 0.002863421104848385, + 0.004595635458827019, + 0.005953658372163773, + 0.007671009283512831, + 0.01272247638553381, + 0.008084381930530071, + 0.008641023188829422, + 0.006709934212267399, + 0.005308672785758972, + 0.0018251144792884588, + 0.0020686162170022726, + 0.0033763160463422537, + 0.0028518273029476404, + 0.0033867366146296263, + 0.00323427258990705, + 0.0015881946310400963, + 0.0018408669857308269, + 0.010263780131936073, + 0.007966711185872555, + 0.011588876135647297, + 0.006977564189583063, + 0.0048281471244990826, + 0.004067995585501194, + 0.00566819915547967, + 0.004384523723274469, + 0.01650402508676052, + 0.01564599759876728, + 0.014709938317537308, + 0.02240154519677162, + 0.01080318447202444, + 0.014781618490815163, + 0.013594676740467548, + 0.014558462426066399, + 0.04253639280796051, + 0.1602347195148468, + 0.1669645756483078, + 0.04543548449873924, + 0.021967481821775436, + 0.022912003099918365, + 0.016648132354021072, + 0.017696019262075424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.007644092198461294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641253061592579 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641253061592579, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641253061592579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10946270823478699, + 0.06838241219520569, + 0.09348704665899277, + 0.12450613081455231, + 0.028556814417243004, + 0.04699759557843208, + 0.0440230593085289, + 0.04050062596797943, + 0.11873593181371689, + 0.32534775137901306 + ], + "credit_map": [ + 0.010601622983813286, + 0.011843547224998474, + 0.009931539185345173, + 0.01566021516919136, + 0.01231640949845314, + 0.013691394589841366, + 0.017235610634088516, + 0.01818237453699112, + 0.00874155480414629, + 0.007138840854167938, + 0.007016669027507305, + 0.0075085838325321674, + 0.008633257821202278, + 0.005941362585872412, + 0.014256644994020462, + 0.009145497344434261, + 0.008809911087155342, + 0.022021623328328133, + 0.012378005310893059, + 0.010264125652611256, + 0.01126262079924345, + 0.007342653814703226, + 0.009406463243067265, + 0.01200164295732975, + 0.014297460205852985, + 0.023627137765288353, + 0.01819443702697754, + 0.013387039303779602, + 0.01274491660296917, + 0.010969864204525948, + 0.016559094190597534, + 0.014726181514561176, + 0.002235210267826915, + 0.0033982277382165194, + 0.0034856863785535097, + 0.004492585547268391, + 0.004375425633043051, + 0.002671258756890893, + 0.004919201135635376, + 0.002979218028485775, + 0.004004096612334251, + 0.004026614595204592, + 0.005169841926544905, + 0.007241366431117058, + 0.007780291605740786, + 0.0060124145820736885, + 0.008480367250740528, + 0.0042826030403375626, + 0.004452465567737818, + 0.005080859642475843, + 0.006145257968455553, + 0.0040769013576209545, + 0.00842891726642847, + 0.007270631846040487, + 0.004274777602404356, + 0.004293248988687992, + 0.004027300514280796, + 0.006505346857011318, + 0.00489196227863431, + 0.00512784393504262, + 0.007106905337423086, + 0.0035836126189678907, + 0.005755603779107332, + 0.0035020513460040092, + 0.017183946445584297, + 0.012115897610783577, + 0.015466706827282906, + 0.020188791677355766, + 0.017517635598778725, + 0.012941966764628887, + 0.013792060315608978, + 0.009528920985758305, + 0.0267269816249609, + 0.06480289995670319, + 0.06714935600757599, + 0.060823142528533936, + 0.02510903775691986, + 0.027104413136839867, + 0.030980154871940613, + 0.022651776671409607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.007641253061592579 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, massive bronze chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot engulfed in ghostly blue flames that emit no smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chariot rub against a sharp rock, creating orange-yellow sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks illuminating rough rock texture and ancient inscriptions on the wheel hubs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, dynamic tracking perspective from the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736189126968384 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736189126968384, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736189126968384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2236667275428772, + 0.06277608871459961, + 0.11550091207027435, + 0.07641129940748215, + 0.04608519747853279, + 0.16814620792865753, + 0.037860121577978134, + 0.06714412569999695, + 0.07469039410352707, + 0.1277189552783966 + ], + "credit_map": [ + 0.01826319471001625, + 0.02396627515554428, + 0.02808069810271263, + 0.039756178855895996, + 0.029000598937273026, + 0.0237469170242548, + 0.02899971976876259, + 0.03185315057635307, + 0.0055309077724814415, + 0.005421308800578117, + 0.007421361282467842, + 0.007075016852468252, + 0.008451089262962341, + 0.006678096950054169, + 0.012050477787852287, + 0.01014783140271902, + 0.01287575252354145, + 0.023032912984490395, + 0.011987545527517796, + 0.01950247772037983, + 0.01214497722685337, + 0.008504864759743214, + 0.013476009480655193, + 0.013976368121802807, + 0.006598002277314663, + 0.0061163874343037605, + 0.007490343414247036, + 0.011334261856973171, + 0.0066800182685256, + 0.007226388901472092, + 0.019551949575543404, + 0.01141394954174757, + 0.0044827209785580635, + 0.005376760847866535, + 0.004888840485364199, + 0.005825012922286987, + 0.006102351006120443, + 0.0069395992904901505, + 0.007278122939169407, + 0.005191788077354431, + 0.015043964609503746, + 0.020530493929982185, + 0.02423679083585739, + 0.03668813034892082, + 0.019009916111826897, + 0.022128397598862648, + 0.01583102159202099, + 0.014677504077553749, + 0.003743828274309635, + 0.004081920720636845, + 0.005882862024009228, + 0.0043463800102472305, + 0.0069490401074290276, + 0.0059083569794893265, + 0.0031168870627880096, + 0.003830846631899476, + 0.006749193649739027, + 0.00916103646159172, + 0.01100679486989975, + 0.008985784836113453, + 0.007840163074433804, + 0.005624915007501841, + 0.007357410620898008, + 0.010418820194900036, + 0.01093675009906292, + 0.00910601019859314, + 0.009602388367056847, + 0.010801414959132671, + 0.008201953023672104, + 0.007702968083322048, + 0.00910628866404295, + 0.009232619777321815, + 0.013995864428579807, + 0.025792887434363365, + 0.025340473279356956, + 0.017977850511670113, + 0.010625717230141163, + 0.01011270098388195, + 0.010137055069208145, + 0.013736397959291935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.006736189126968384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_tensor([6516, 6516])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_tensor([6516, 6516])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..04d8c2051ed5761584acb3d3ab959284aea452df --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_tensor([6516, 6516])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide, static shot, shot from a level angle, captures the center of an empty circus tent. A single spotlight shines vertically from the soaring canvas dome, casting a bright circle of light on the sawdust-strewn floor, illuminating the surrounding rows of empty wooden benches and deep shadows. In the center of this luminous light, a performer in classic clown makeup—a large, exaggerated red smile framed by a pale base, with blue teardrop patterns adorning the corners of his eyes—wears a slightly askew, curly orange wig and pedals a unicycle that is comically small and contrasts sharply with his body. He wears a saturated, incongruous outfit: a loose, bright yellow silk shirt tied with a large blue and white striped bow at the collar, and baggy trousers held up by red suspenders, one leg adorned with white dots on a red background, the other with yellow stripes on a green background. He leans forward, his brows furrowed beneath the paint, beads of sweat oozing from his forehead. His legs pedal ferociously and clumsily, his entire body swaying violently with every jolt of the unicycle, his arms stretched out to the sides to maintain his precarious balance. Although the smile on his face remains constant, his focused and strained posture contrasts sharply with the vast emptiness around him, creating a surreal atmosphere that blends the comical, the poignant, and the obsessive.", + "seed": 25872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_6516_seed_25872_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07249410040676593, + "vlm_logprob": -0.07249410040676593, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.07249410040676593, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051448240876197815 + ], + "opposite_logprob_tokens": [ + -3.051448345184326 + ], + "logprob_reward": -0.051448240876197815, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051448240876197815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10285323858261108, + 0.06653246283531189, + 0.060580652207136154, + 0.03653419017791748, + 0.2561979293823242, + 0.1655639261007309, + 0.04010465741157532, + 0.03542381897568703, + 0.0499015748500824, + 0.18630754947662354 + ], + "credit_map": [ + 0.013658161275088787, + 0.0124712735414505, + 0.014161394909024239, + 0.017205215990543365, + 0.006772065535187721, + 0.012332655489444733, + 0.010479222983121872, + 0.015773244202136993, + 0.007469817064702511, + 0.008795923553407192, + 0.008000924251973629, + 0.008286532014608383, + 0.008746667765080929, + 0.008590913377702236, + 0.006403051782399416, + 0.010238629765808582, + 0.009677954018115997, + 0.005637435708194971, + 0.008787370286881924, + 0.00577832106500864, + 0.005893934518098831, + 0.009859545156359673, + 0.006595904007554054, + 0.008350184187293053, + 0.006309793330729008, + 0.0028465886134654284, + 0.00569620355963707, + 0.002457978669553995, + 0.003975569736212492, + 0.006307231727987528, + 0.004196221008896828, + 0.004744606092572212, + 0.07404696196317673, + 0.015871744602918625, + 0.06333918869495392, + 0.01662609353661537, + 0.014745058491826057, + 0.027753405272960663, + 0.022359242662787437, + 0.021456247195601463, + 0.03238492086529732, + 0.014907373115420341, + 0.04185572639107704, + 0.009441975504159927, + 0.015118849463760853, + 0.02137817069888115, + 0.011645637452602386, + 0.018831269815564156, + 0.01111513003706932, + 0.002626248635351658, + 0.003821606980636716, + 0.0026030188892036676, + 0.003521302482113242, + 0.011207083240151405, + 0.0027382574044167995, + 0.002472007879987359, + 0.004564002156257629, + 0.00328264944255352, + 0.005152026191353798, + 0.0031511192210018635, + 0.004551894962787628, + 0.006498148664832115, + 0.005587290972471237, + 0.002636686898767948, + 0.006922084838151932, + 0.005327990278601646, + 0.012029021047055721, + 0.0036096670664846897, + 0.004058300517499447, + 0.00592946819961071, + 0.005177479237318039, + 0.006847565062344074, + 0.03988959267735481, + 0.02442578226327896, + 0.025374429300427437, + 0.010984564200043678, + 0.014087448827922344, + 0.026106638833880424, + 0.022040020674467087, + 0.023399081081151962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.051448240876197815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031517595052719 + ], + "opposite_logprob_tokens": [ + -2.3503150939941406 + ], + "logprob_reward": -0.10031517595052719, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031517595052719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30610963702201843, + 0.11088481545448303, + 0.07956503331661224, + 0.07087365537881851, + 0.0970526933670044, + 0.06679723411798477, + 0.02730041928589344, + 0.043585944920778275, + 0.10191529989242554, + 0.09591525048017502 + ], + "credit_map": [ + 0.039993129670619965, + 0.065621018409729, + 0.04901920258998871, + 0.03545507416129112, + 0.031816910952329636, + 0.027755245566368103, + 0.024173563346266747, + 0.032275501638650894, + 0.015125961042940617, + 0.014119320549070835, + 0.01528269611299038, + 0.00955985113978386, + 0.018337927758693695, + 0.013471005484461784, + 0.010595247149467468, + 0.01439280342310667, + 0.01262584887444973, + 0.009580247104167938, + 0.01190072763711214, + 0.00738002872094512, + 0.006580955348908901, + 0.012814530171453953, + 0.00834766123443842, + 0.010335037484765053, + 0.013771996833384037, + 0.004454766400158405, + 0.014009678736329079, + 0.004694649949669838, + 0.004717546049505472, + 0.011500844731926918, + 0.005670404527336359, + 0.012053768150508404, + 0.021792303770780563, + 0.008825897239148617, + 0.015279214829206467, + 0.007091124542057514, + 0.010441869497299194, + 0.015166930854320526, + 0.00978684239089489, + 0.008668511174619198, + 0.014702972024679184, + 0.0055500841699540615, + 0.01998736336827278, + 0.003970062360167503, + 0.0051025706343352795, + 0.00789344310760498, + 0.003931182902306318, + 0.0056595574133098125, + 0.0034562142100185156, + 0.0021844555158168077, + 0.0038588964380323887, + 0.0036566760390996933, + 0.0035992173943668604, + 0.003129676217213273, + 0.004195515997707844, + 0.0032197667751461267, + 0.007899616844952106, + 0.0031096080783754587, + 0.005652608349919319, + 0.004595463629812002, + 0.006089847534894943, + 0.0049600000493228436, + 0.008226076140999794, + 0.0030527266208082438, + 0.027123326435685158, + 0.01111688744276762, + 0.03405948728322983, + 0.005810118746012449, + 0.004474820103496313, + 0.007358135189861059, + 0.004858838859945536, + 0.007113693282008171, + 0.02095704898238182, + 0.010103990323841572, + 0.013445453718304634, + 0.005816846154630184, + 0.0065659647807478905, + 0.013890921138226986, + 0.013168048113584518, + 0.011966974474489689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.10031517595052719 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792767018079758 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.029792767018079758, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792767018079758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08141949027776718, + 0.051522910594940186, + 0.06261925399303436, + 0.021186714991927147, + 0.2691430151462555, + 0.44281983375549316, + 0.01629694551229477, + 0.019964376464486122, + 0.02354690246284008, + 0.011480513960123062 + ], + "credit_map": [ + 0.00783824734389782, + 0.02120819315314293, + 0.012926371768116951, + 0.013791757635772228, + 0.006420999299734831, + 0.006152214016765356, + 0.006141379475593567, + 0.0069403378292918205, + 0.011706000193953514, + 0.00676771579310298, + 0.008776424452662468, + 0.003514511277899146, + 0.0058558289892971516, + 0.007409262005239725, + 0.003825223771855235, + 0.0036679490003734827, + 0.00918255839496851, + 0.013004247099161148, + 0.01022701058536768, + 0.005872697569429874, + 0.006112930364906788, + 0.00836879201233387, + 0.004534533713012934, + 0.005316489841789007, + 0.003337845904752612, + 0.0015598149038851261, + 0.0038261713925749063, + 0.0013495221501216292, + 0.0016306900652125478, + 0.0030392457265406847, + 0.0018889441853389144, + 0.004554481245577335, + 0.06818452477455139, + 0.02297102101147175, + 0.04905073344707489, + 0.01643470861017704, + 0.02841777540743351, + 0.037652913480997086, + 0.02175120823085308, + 0.024680178612470627, + 0.09726247191429138, + 0.04719410091638565, + 0.12977923452854156, + 0.027426745742559433, + 0.03894103318452835, + 0.04608404263854027, + 0.025263430550694466, + 0.030868785455822945, + 0.003285796381533146, + 0.0014442384708672762, + 0.00177542504388839, + 0.0013221504632383585, + 0.0028083999641239643, + 0.002180196112021804, + 0.001514884876087308, + 0.0019658554811030626, + 0.004015245474874973, + 0.001640671631321311, + 0.004665979649871588, + 0.0017409776337444782, + 0.002975074341520667, + 0.0015756245702505112, + 0.0015398840187117457, + 0.0018109200755134225, + 0.006123396102339029, + 0.002293145516887307, + 0.006609131582081318, + 0.0016065880190581083, + 0.0016226915176957846, + 0.0020595474634319544, + 0.0014954281505197287, + 0.0017369763227179646, + 0.0017239183653146029, + 0.0009105663630180061, + 0.0027673097793012857, + 0.0009257387719117105, + 0.0012027887860313058, + 0.0017713886918500066, + 0.0010716320248320699, + 0.0011071719927713275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.029792767018079758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585110411047935 + ], + "opposite_logprob_tokens": [ + -3.895585060119629 + ], + "logprob_reward": -0.020585110411047935, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585110411047935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22719670832157135, + 0.06485462933778763, + 0.14910484850406647, + 0.0917092114686966, + 0.10076457262039185, + 0.07034554332494736, + 0.04933147504925728, + 0.07308927178382874, + 0.09677630662918091, + 0.07682748883962631 + ], + "credit_map": [ + 0.027752064168453217, + 0.04627121984958649, + 0.041928909718990326, + 0.02678234688937664, + 0.020058486610651016, + 0.01918792724609375, + 0.017232319340109825, + 0.027983440086245537, + 0.006616941187530756, + 0.006622596178203821, + 0.008925310336053371, + 0.0069033014588057995, + 0.012221934273838997, + 0.006040556821972132, + 0.0083150714635849, + 0.009208915755152702, + 0.02249022200703621, + 0.029026394709944725, + 0.024765964597463608, + 0.013481163419783115, + 0.012866752222180367, + 0.020716262981295586, + 0.011036697775125504, + 0.014721393585205078, + 0.01394970528781414, + 0.0064321341924369335, + 0.014171084389090538, + 0.0068741291761398315, + 0.008690285496413708, + 0.012630505487322807, + 0.00887535884976387, + 0.020086005330085754, + 0.023728633299469948, + 0.005678480491042137, + 0.033920105546712875, + 0.007902770303189754, + 0.007481394801288843, + 0.00891205109655857, + 0.007585694547742605, + 0.005555439740419388, + 0.012454257346689701, + 0.007979470305144787, + 0.014539401978254318, + 0.006456480827182531, + 0.006386756896972656, + 0.01021728664636612, + 0.007058960851281881, + 0.005252929404377937, + 0.006343476939946413, + 0.0042143589816987514, + 0.004193361848592758, + 0.005186478141695261, + 0.004270385485142469, + 0.013266842812299728, + 0.004849789664149284, + 0.007006782107055187, + 0.01284046284854412, + 0.0073634241707623005, + 0.016040807589888573, + 0.008569224737584591, + 0.007200051564723253, + 0.008022050373256207, + 0.007072498556226492, + 0.005980750545859337, + 0.023813359439373016, + 0.010338013991713524, + 0.025019215419888496, + 0.006350451149046421, + 0.006874565500766039, + 0.008581277914345264, + 0.006195745896548033, + 0.009603672660887241, + 0.01438418310135603, + 0.006306859198957682, + 0.010685262270271778, + 0.005508518777787685, + 0.00722225196659565, + 0.011184867471456528, + 0.01247340627014637, + 0.00906214490532875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.020585110411047935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160329207777977 + ], + "opposite_logprob_tokens": [ + -1.9103292226791382 + ], + "logprob_reward": -0.160329207777977, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160329207777977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15218405425548553, + 0.04836539551615715, + 0.09048838168382645, + 0.05558101832866669, + 0.06573376059532166, + 0.2240508794784546, + 0.0890391543507576, + 0.15582984685897827, + 0.05151541158556938, + 0.06721211969852448 + ], + "credit_map": [ + 0.013888204470276833, + 0.02730586752295494, + 0.028006652370095253, + 0.02087263949215412, + 0.011455999687314034, + 0.015528670512139797, + 0.011944572441279888, + 0.02318144217133522, + 0.006519756745547056, + 0.005685896147042513, + 0.0065761227160692215, + 0.004475543275475502, + 0.008518933318555355, + 0.005742749664932489, + 0.004893227480351925, + 0.00595316756516695, + 0.0140431709587574, + 0.01817985624074936, + 0.013864031061530113, + 0.01113842148333788, + 0.008093460462987423, + 0.008322220295667648, + 0.008952728472650051, + 0.007894491776823997, + 0.008411431685090065, + 0.006010076962411404, + 0.0074598235078155994, + 0.004875641316175461, + 0.005970457103103399, + 0.00763131445273757, + 0.007090178783982992, + 0.008132095448672771, + 0.015406051650643349, + 0.004264239687472582, + 0.021422715857625008, + 0.005092454142868519, + 0.0062065161764621735, + 0.004702475853264332, + 0.004892934113740921, + 0.0037463740445673466, + 0.0436365120112896, + 0.02583063393831253, + 0.06558898836374283, + 0.01565214991569519, + 0.020677154883742332, + 0.023408161476254463, + 0.011896281503140926, + 0.017360981553792953, + 0.02121409773826599, + 0.008527928031980991, + 0.012016536667943, + 0.007084933575242758, + 0.008477028459310532, + 0.013203090988099575, + 0.007435052189975977, + 0.011080486699938774, + 0.027430834248661995, + 0.01851973682641983, + 0.038278669118881226, + 0.016491694375872612, + 0.012521794997155666, + 0.015559962950646877, + 0.012839550152420998, + 0.01418760884553194, + 0.00800460297614336, + 0.009136485867202282, + 0.008120114915072918, + 0.004935212433338165, + 0.004951932467520237, + 0.005304382648319006, + 0.004418037366122007, + 0.00664464570581913, + 0.01532432809472084, + 0.007356745656579733, + 0.011641918681561947, + 0.003966385964304209, + 0.004246619064360857, + 0.00877879373729229, + 0.00827961228787899, + 0.007617715280503035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.160329207777977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, static shot, shot from a level angle, captures the center of an empty circus tent. A single spotlight shines vertically from the soaring canvas dome, casting a bright circle of light on the sawdust-strewn floor, illuminating the surrounding rows of empty wooden benches and deep shadows. In the center of this luminous light, a performer in classic clown makeup—a large, exaggerated red smile framed by a pale base, with blue teardrop patterns adorning the corners of his eyes—wears a slightly askew, curly orange wig and pedals a unicycle that is comically small and contrasts sharply with his body. He wears a saturated, incongruous outfit: a loose, bright yellow silk shirt tied with a large blue and white striped bow at the collar, and baggy trousers held up by red suspenders, one leg adorned with white dots on a red background, the other with yellow stripes on a green background. He leans forward, his brows furrowed beneath the paint, beads of sweat oozing from his forehead. His legs pedal ferociously and clumsily, his entire body swaying violently with every jolt of the unicycle, his arms stretched out to the sides to maintain his precarious balance. Although the smile on his face remains constant, his focused and strained posture contrasts sharply with the vast emptiness around him, creating a surreal atmosphere that blends the comical, the poignant, and the obsessive.", + "seed": 25873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_6516_seed_25873_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03549715168774128, + "vlm_logprob": -0.03549715168774128, + "vlm_yes_no_margin": 3.3749999523162844, + "vlm_token_credit": -0.03549715168774128, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04004986956715584 + ], + "opposite_logprob_tokens": [ + -3.2900497913360596 + ], + "logprob_reward": -0.04004986956715584, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04004986956715584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1512988656759262, + 0.03784485161304474, + 0.1214556097984314, + 0.1533636897802353, + 0.1374204158782959, + 0.14063440263271332, + 0.10817893594503403, + 0.06029297783970833, + 0.04198215529322624, + 0.04752805456519127 + ], + "credit_map": [ + 0.014806535094976425, + 0.022586537525057793, + 0.02406514808535576, + 0.01974761113524437, + 0.021148545667529106, + 0.024966565892100334, + 0.009431084617972374, + 0.014546854421496391, + 0.003828482935205102, + 0.005638901609927416, + 0.006706724874675274, + 0.004376688506454229, + 0.003820171346887946, + 0.006673216354101896, + 0.002872995100915432, + 0.003927673678845167, + 0.01108994334936142, + 0.01849176734685898, + 0.014671629294753075, + 0.01152847707271576, + 0.020931167528033257, + 0.016672499477863312, + 0.015443791635334492, + 0.012626337818801403, + 0.027211306616663933, + 0.019179923459887505, + 0.02058892883360386, + 0.014995282515883446, + 0.018760012462735176, + 0.019205566495656967, + 0.011511802673339844, + 0.02191087417304516, + 0.018303409218788147, + 0.015616606920957565, + 0.03532549366354942, + 0.010023877955973148, + 0.011555623263120651, + 0.021508777514100075, + 0.007326523307710886, + 0.017760109156370163, + 0.024780208244919777, + 0.012248433195054531, + 0.024956434965133667, + 0.009389353916049004, + 0.012738415971398354, + 0.03125378489494324, + 0.007236423436552286, + 0.018031366169452667, + 0.013566075824201107, + 0.006815711501985788, + 0.026350243017077446, + 0.007157273590564728, + 0.008387383073568344, + 0.03373647853732109, + 0.006391334813088179, + 0.005774452816694975, + 0.008350149728357792, + 0.008015306666493416, + 0.010046410374343395, + 0.0060273720882833, + 0.006595863960683346, + 0.007870246656239033, + 0.006654849275946617, + 0.006732782814651728, + 0.0047095962800085545, + 0.004383713472634554, + 0.006915230769664049, + 0.0036845144350081682, + 0.005402939859777689, + 0.007427604869008064, + 0.004555691033601761, + 0.004902869462966919, + 0.006223463453352451, + 0.0046020131558179855, + 0.00809430330991745, + 0.00313177821226418, + 0.004787835292518139, + 0.007348774932324886, + 0.0030054841190576553, + 0.010334406048059464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.04004986956715584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038107167929410934 + ], + "opposite_logprob_tokens": [ + -3.28810715675354 + ], + "logprob_reward": -0.038107167929410934, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038107167929410934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20607204735279083, + 0.11429592967033386, + 0.24475739896297455, + 0.05800763890147209, + 0.06947165727615356, + 0.03858526423573494, + 0.07011783868074417, + 0.056674931198358536, + 0.1052597239613533, + 0.03675755113363266 + ], + "credit_map": [ + 0.020983301103115082, + 0.027025246992707253, + 0.04263168200850487, + 0.020679106935858727, + 0.019672634080052376, + 0.043040819466114044, + 0.011459438130259514, + 0.02057981863617897, + 0.012863107956945896, + 0.015077748335897923, + 0.021496031433343887, + 0.010715851560235023, + 0.009036410599946976, + 0.02410953864455223, + 0.006683059502393007, + 0.014314183034002781, + 0.025995926931500435, + 0.04517209902405739, + 0.034065283834934235, + 0.02280629426240921, + 0.0274260975420475, + 0.03882971778512001, + 0.023593543097376823, + 0.02686844766139984, + 0.0096873939037323, + 0.006677539553493261, + 0.011108987033367157, + 0.004785832483321428, + 0.006892114412039518, + 0.0067590391263365746, + 0.005502240266650915, + 0.006594494450837374, + 0.010847171768546104, + 0.007788063492625952, + 0.01885135844349861, + 0.004640953149646521, + 0.005394789855927229, + 0.010944121517241001, + 0.004212657455354929, + 0.0067925467155873775, + 0.0054266187362372875, + 0.0029521770775318146, + 0.008044631220400333, + 0.002938629826530814, + 0.003973356448113918, + 0.008525343611836433, + 0.0030530989170074463, + 0.003671409096568823, + 0.01194270234555006, + 0.006287539377808571, + 0.012774264439940453, + 0.006245503202080727, + 0.00614326074719429, + 0.015599185600876808, + 0.005909558851271868, + 0.00521582318469882, + 0.007261249236762524, + 0.007233732379972935, + 0.008833176456391811, + 0.004977733828127384, + 0.006714196875691414, + 0.006849590688943863, + 0.005152300465852022, + 0.00965295173227787, + 0.009591534733772278, + 0.012751159258186817, + 0.01642215996980667, + 0.010012978687882423, + 0.014394700527191162, + 0.015034055337309837, + 0.013454889878630638, + 0.013598247431218624, + 0.0050946795381605625, + 0.004424192942678928, + 0.0058272345922887325, + 0.0029173640068620443, + 0.004944957327097654, + 0.005588826257735491, + 0.002675962634384632, + 0.005284333601593971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.038107167929410934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809535130858421 + ], + "opposite_logprob_tokens": [ + -3.288095235824585 + ], + "logprob_reward": -0.03809535130858421, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809535130858421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15848812460899353, + 0.05265084281563759, + 0.12472020834684372, + 0.08296383172273636, + 0.0781300812959671, + 0.03597446531057358, + 0.125674307346344, + 0.09796441346406937, + 0.20517414808273315, + 0.03825952857732773 + ], + "credit_map": [ + 0.016045555472373962, + 0.015833601355552673, + 0.021829182282090187, + 0.015930430963635445, + 0.039374399930238724, + 0.022601258009672165, + 0.01336434856057167, + 0.013509366661310196, + 0.007071151398122311, + 0.008430300280451775, + 0.00783163495361805, + 0.004024264868348837, + 0.005117034073919058, + 0.012129541486501694, + 0.003431167220696807, + 0.00461575435474515, + 0.013393210247159004, + 0.017261352390050888, + 0.018044380471110344, + 0.013067850843071938, + 0.015810970216989517, + 0.01824222505092621, + 0.014893352054059505, + 0.01400688011199236, + 0.01133399736136198, + 0.011172561906278133, + 0.013035658746957779, + 0.008644459769129753, + 0.010784223675727844, + 0.008874811232089996, + 0.011909514665603638, + 0.007208612747490406, + 0.00920400582253933, + 0.010633164085447788, + 0.010059974156320095, + 0.005728422664105892, + 0.011027802713215351, + 0.0100723747164011, + 0.011150816455483437, + 0.010253526270389557, + 0.004549538716673851, + 0.00362388719804585, + 0.007562607992440462, + 0.003264800412580371, + 0.0038548344746232033, + 0.006311839912086725, + 0.0034917856100946665, + 0.0033151740208268166, + 0.02149961143732071, + 0.011860102415084839, + 0.02449641190469265, + 0.011194221675395966, + 0.01281161978840828, + 0.022117028012871742, + 0.011252804659307003, + 0.010442521423101425, + 0.015476218424737453, + 0.009627814404666424, + 0.013932695612311363, + 0.008282358758151531, + 0.011012978851795197, + 0.011182539165019989, + 0.008314634673297405, + 0.0201351810246706, + 0.018010953441262245, + 0.027141837403178215, + 0.031346723437309265, + 0.019524134695529938, + 0.029616160318255424, + 0.02852589637041092, + 0.02719971351325512, + 0.023808732628822327, + 0.005242376588284969, + 0.004070068243891001, + 0.0065549504943192005, + 0.0030627381056547165, + 0.004841992631554604, + 0.0069288890808820724, + 0.0029639455024152994, + 0.004594569094479084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.03809535130858421 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188124522566795 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018188124522566795, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188124522566795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10749314725399017, + 0.037519391626119614, + 0.06501258164644241, + 0.05073961243033409, + 0.08136562258005142, + 0.05798434093594551, + 0.1965426653623581, + 0.13347697257995605, + 0.23535631597042084, + 0.034509338438510895 + ], + "credit_map": [ + 0.011415599845349789, + 0.012838182970881462, + 0.020651288330554962, + 0.014718735590577126, + 0.015032793395221233, + 0.015495490282773972, + 0.007620608434081078, + 0.009720448404550552, + 0.004886912647634745, + 0.006992064416408539, + 0.005622399505227804, + 0.003542245365679264, + 0.0040388647466897964, + 0.005904259625822306, + 0.002698443830013275, + 0.0038342007901519537, + 0.008384655229747295, + 0.0101081021130085, + 0.010100265964865685, + 0.005943734664469957, + 0.007560793776065111, + 0.010715856216847897, + 0.0051303100772202015, + 0.00706886500120163, + 0.008922452107071877, + 0.006788049358874559, + 0.007275513373315334, + 0.004765918944031, + 0.006412179674953222, + 0.0048337592743337154, + 0.004667927976697683, + 0.007073814515024424, + 0.011358078569173813, + 0.012504626996815205, + 0.01060344185680151, + 0.006318311206996441, + 0.009867881424725056, + 0.010895660147070885, + 0.007908734492957592, + 0.011908886022865772, + 0.009629313834011555, + 0.005338125396519899, + 0.01119634136557579, + 0.004000154789537191, + 0.006050202529877424, + 0.012010225094854832, + 0.003919800743460655, + 0.0058401781134307384, + 0.033937372267246246, + 0.018051451072096825, + 0.04051181301474571, + 0.017183028161525726, + 0.018314428627490997, + 0.03627096861600876, + 0.015693683177232742, + 0.016579922288656235, + 0.018177950754761696, + 0.01380105223506689, + 0.017535138875246048, + 0.011811998672783375, + 0.01750926859676838, + 0.016892146319150925, + 0.013571765273809433, + 0.024177638813853264, + 0.02090301923453808, + 0.02894190512597561, + 0.03686869889497757, + 0.022301187738776207, + 0.03456868976354599, + 0.032323796302080154, + 0.030284741893410683, + 0.029164284467697144, + 0.004839158616960049, + 0.004190043080598116, + 0.0045778341591358185, + 0.0027423652354627848, + 0.0060412888415157795, + 0.004955517128109932, + 0.002744306344538927, + 0.004418822471052408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.018188124522566795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304524511098862 + ], + "opposite_logprob_tokens": [ + -3.1680452823638916 + ], + "logprob_reward": -0.04304524511098862, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304524511098862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16629280149936676, + 0.035490456968545914, + 0.10602615773677826, + 0.17454342544078827, + 0.09949686378240585, + 0.035559725016355515, + 0.05503212660551071, + 0.19428904354572296, + 0.09430377185344696, + 0.03896569460630417 + ], + "credit_map": [ + 0.01792575977742672, + 0.017475245520472527, + 0.03411594405770302, + 0.01633608713746071, + 0.013987681828439236, + 0.03819149732589722, + 0.014419334940612316, + 0.013841231353580952, + 0.004480689764022827, + 0.005336860194802284, + 0.004246237222105265, + 0.0032185001764446497, + 0.004623427055776119, + 0.0058399406261742115, + 0.0038302429020404816, + 0.003914554603397846, + 0.012815399095416069, + 0.01003908459097147, + 0.019486751407384872, + 0.009145306423306465, + 0.009005503728985786, + 0.026102295145392418, + 0.011610474437475204, + 0.007821337319910526, + 0.02035275474190712, + 0.01911376416683197, + 0.029435796663165092, + 0.015603596344590187, + 0.027089109644293785, + 0.02280106581747532, + 0.025204775854945183, + 0.014942537993192673, + 0.012849303893744946, + 0.008451004512608051, + 0.02340555004775524, + 0.006923291832208633, + 0.011168603785336018, + 0.015546193346381187, + 0.01047067902982235, + 0.010682232677936554, + 0.0042469697073102, + 0.003879133379086852, + 0.006626759190112352, + 0.0029300940223038197, + 0.004531633574515581, + 0.005771353840827942, + 0.004161105956882238, + 0.003412672784179449, + 0.0089707737788558, + 0.006814758758991957, + 0.01132971327751875, + 0.004171543288975954, + 0.006675919983536005, + 0.007437264081090689, + 0.0046510896645486355, + 0.004981060046702623, + 0.026997612789273262, + 0.013092942535877228, + 0.048214107751846313, + 0.019435623660683632, + 0.01747586391866207, + 0.03128279745578766, + 0.01496165432035923, + 0.02282842993736267, + 0.009737701155245304, + 0.009974369779229164, + 0.01668584533035755, + 0.00800344068557024, + 0.012700304388999939, + 0.01426424365490675, + 0.012421829625964165, + 0.01051602978259325, + 0.006895241793245077, + 0.0028549297712743282, + 0.00612846901640296, + 0.0030055036768317223, + 0.004483652766793966, + 0.006654222495853901, + 0.0038714047987014055, + 0.005072267260402441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.04304524511098862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, static shot, shot from a level angle, captures the center of an empty circus tent. A single spotlight shines vertically from the soaring canvas dome, casting a bright circle of light on the sawdust-strewn floor, illuminating the surrounding rows of empty wooden benches and deep shadows. In the center of this luminous light, a performer in classic clown makeup—a large, exaggerated red smile framed by a pale base, with blue teardrop patterns adorning the corners of his eyes—wears a slightly askew, curly orange wig and pedals a unicycle that is comically small and contrasts sharply with his body. He wears a saturated, incongruous outfit: a loose, bright yellow silk shirt tied with a large blue and white striped bow at the collar, and baggy trousers held up by red suspenders, one leg adorned with white dots on a red background, the other with yellow stripes on a green background. He leans forward, his brows furrowed beneath the paint, beads of sweat oozing from his forehead. His legs pedal ferociously and clumsily, his entire body swaying violently with every jolt of the unicycle, his arms stretched out to the sides to maintain his precarious balance. Although the smile on his face remains constant, his focused and strained posture contrasts sharply with the vast emptiness around him, creating a surreal atmosphere that blends the comical, the poignant, and the obsessive.", + "seed": 25874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_6516_seed_25874_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10199105702340602, + "vlm_logprob": -0.10199105702340602, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.10199105702340602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.161458820104599 + ], + "opposite_logprob_tokens": [ + -1.9114588499069214 + ], + "logprob_reward": -0.161458820104599, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.161458820104599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20542870461940765, + 0.24948252737522125, + 0.05019768327474594, + 0.09793175756931305, + 0.06691070646047592, + 0.10437049716711044, + 0.08913031965494156, + 0.029222015291452408, + 0.0900428518652916, + 0.017283014953136444 + ], + "credit_map": [ + 0.025951389223337173, + 0.019472582265734673, + 0.0477510504424572, + 0.018526824191212654, + 0.013210933655500412, + 0.03606454282999039, + 0.021208839491009712, + 0.023242546245455742, + 0.028440332040190697, + 0.021004900336265564, + 0.05051000043749809, + 0.013038765639066696, + 0.028035076335072517, + 0.0618426576256752, + 0.02249288372695446, + 0.02411791868507862, + 0.007024894002825022, + 0.004847978241741657, + 0.012653240002691746, + 0.002931135706603527, + 0.003249221481382847, + 0.010706616565585136, + 0.0035120435059070587, + 0.005272556096315384, + 0.020557159557938576, + 0.00757864722982049, + 0.01585443876683712, + 0.004482741467654705, + 0.01157496590167284, + 0.01410094927996397, + 0.007767093367874622, + 0.016015756875276566, + 0.010732020251452923, + 0.005281063728034496, + 0.01159692369401455, + 0.004168329294770956, + 0.00616732332855463, + 0.011942612938582897, + 0.005987811367958784, + 0.011034620925784111, + 0.017790786921977997, + 0.012119279243052006, + 0.016923099756240845, + 0.005861729849129915, + 0.010751685127615929, + 0.019224995747208595, + 0.010380228981375694, + 0.011318695731461048, + 0.016927748918533325, + 0.005610608030110598, + 0.014760695397853851, + 0.006090363021939993, + 0.007703294046223164, + 0.015487302094697952, + 0.008098750375211239, + 0.014451556839048862, + 0.003829728811979294, + 0.0021045077592134476, + 0.005253512877970934, + 0.0018991512479260564, + 0.00369372870773077, + 0.004372790455818176, + 0.003929248545318842, + 0.004139346536248922, + 0.018500208854675293, + 0.005144047550857067, + 0.022234715521335602, + 0.004648113157600164, + 0.009430361911654472, + 0.015116226859390736, + 0.005859869532287121, + 0.009109308943152428, + 0.0038697202689945698, + 0.0019197955261915922, + 0.0041087595745921135, + 0.0011481990804895759, + 0.001297815702855587, + 0.002148979576304555, + 0.0015211189165711403, + 0.001268626656383276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.161458820104599 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812912166118622 + ], + "opposite_logprob_tokens": [ + -1.4062912464141846 + ], + "logprob_reward": -0.2812912166118622, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812912166118622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5685736536979675, + 0.16724470257759094, + 0.08072645217180252, + 0.04812774807214737, + 0.03251093998551369, + 0.026786761358380318, + 0.01685931906104088, + 0.01252119243144989, + 0.03472214937210083, + 0.011927016079425812 + ], + "credit_map": [ + 0.04521707445383072, + 0.008667903020977974, + 0.2213779240846634, + 0.016079410910606384, + 0.011852147057652473, + 0.23490974307060242, + 0.012349914759397507, + 0.018119560554623604, + 0.02085011452436447, + 0.01565316878259182, + 0.028796393424272537, + 0.009654291905462742, + 0.01849757693707943, + 0.039803970605134964, + 0.016807399690151215, + 0.01718178577721119, + 0.013264073058962822, + 0.004990486428141594, + 0.015380141325294971, + 0.0031237523071467876, + 0.008848963305354118, + 0.018684711307287216, + 0.0050957356579601765, + 0.01133858785033226, + 0.006196732632815838, + 0.0020079936366528273, + 0.012520250864326954, + 0.002453173277899623, + 0.002675149589776993, + 0.01478555053472519, + 0.0030209901742637157, + 0.004467908758670092, + 0.005354494322091341, + 0.002342852298170328, + 0.0062617529183626175, + 0.0021334937773644924, + 0.002789014019072056, + 0.0060593304224312305, + 0.0033116217236965895, + 0.004258379805833101, + 0.005380810238420963, + 0.002706667175516486, + 0.004097470082342625, + 0.0017441939562559128, + 0.002898375503718853, + 0.004904044326394796, + 0.0025024733040481806, + 0.00255272607319057, + 0.002649401081725955, + 0.001269830041565001, + 0.0034397810231894255, + 0.0009346852893941104, + 0.0014388870913535357, + 0.0038456579204648733, + 0.0013270467752590775, + 0.00195403047837317, + 0.0019920675549656153, + 0.0008145108004100621, + 0.002652909606695175, + 0.0008453177288174629, + 0.001628764672204852, + 0.0017042545368894935, + 0.0014158682897686958, + 0.0014674996491521597, + 0.006353022530674934, + 0.002381711034104228, + 0.007720502559095621, + 0.0020338709000498056, + 0.0038586659356951714, + 0.006510364823043346, + 0.0023595900274813175, + 0.0035044199321419, + 0.0027182523626834154, + 0.0007855248986743391, + 0.0030509172938764095, + 0.0006849963101558387, + 0.0009707458084449172, + 0.0013282952131703496, + 0.0014495826326310635, + 0.0009387005702592432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2812912166118622 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979658544063568 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.02979658544063568, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979658544063568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6419992446899414, + 0.04637632891535759, + 0.03937171772122383, + 0.043482497334480286, + 0.11868677288293839, + 0.01930227130651474, + 0.03560846298933029, + 0.007226596586406231, + 0.01312751229852438, + 0.03481852635741234 + ], + "credit_map": [ + 0.048724621534347534, + 0.01502185221761465, + 0.23812684416770935, + 0.01853516511619091, + 0.015472056344151497, + 0.26266685128211975, + 0.018782787024974823, + 0.024669118225574493, + 0.0056666904129087925, + 0.0049532619304955006, + 0.00854196585714817, + 0.0023810709826648235, + 0.006355085410177708, + 0.010561984032392502, + 0.0038339842576533556, + 0.004082284402102232, + 0.0038902387022972107, + 0.0027447158936411142, + 0.009395203553140163, + 0.0019747228361666203, + 0.0040258364751935005, + 0.009110506623983383, + 0.003700300818309188, + 0.004530192818492651, + 0.0049078273586928844, + 0.005019139498472214, + 0.006226181518286467, + 0.005000790115445852, + 0.004893537610769272, + 0.005651035346090794, + 0.0068440306931734085, + 0.004939957056194544, + 0.01139773242175579, + 0.005100478883832693, + 0.04064776748418808, + 0.0049692499451339245, + 0.005569306667894125, + 0.04047640413045883, + 0.005099068861454725, + 0.005426760762929916, + 0.002980787307024002, + 0.0016219108365476131, + 0.003234582720324397, + 0.001664881594479084, + 0.003947240300476551, + 0.002393681090325117, + 0.0021168231032788754, + 0.0013423655182123184, + 0.0035748369991779327, + 0.0025821628514677286, + 0.007127413991838694, + 0.002161073498427868, + 0.005593560170382261, + 0.007251135539263487, + 0.004260312765836716, + 0.0030579676385968924, + 0.000978140509687364, + 0.0005779979983344674, + 0.0010829573730006814, + 0.000512843020260334, + 0.0009667380945757031, + 0.0012812901986762881, + 0.0009688697173260152, + 0.0008577599655836821, + 0.0016575607005506754, + 0.0011709454702213407, + 0.002037864178419113, + 0.0009671850129961967, + 0.002958054421469569, + 0.0015614554286003113, + 0.0017197688575834036, + 0.001054677995853126, + 0.008796075358986855, + 0.0023010699078440666, + 0.010626216419041157, + 0.0014682236360386014, + 0.0034443254116922617, + 0.00382575043477118, + 0.002565589267760515, + 0.001791273127309978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02979658544063568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076981201767921 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076981201767921, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076981201767921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4773458242416382, + 0.14931267499923706, + 0.06319870054721832, + 0.038841407746076584, + 0.044321659952402115, + 0.051951758563518524, + 0.04834427684545517, + 0.031613048166036606, + 0.04130460321903229, + 0.053766075521707535 + ], + "credit_map": [ + 0.03782615810632706, + 0.008411438204348087, + 0.18286754190921783, + 0.013991612941026688, + 0.012455619871616364, + 0.1950240433216095, + 0.012155503034591675, + 0.014613914303481579, + 0.018606171011924744, + 0.014269780367612839, + 0.02616686187684536, + 0.008703108876943588, + 0.017297161743044853, + 0.03510354459285736, + 0.013880511745810509, + 0.015285540372133255, + 0.010015073232352734, + 0.006558136083185673, + 0.010002397000789642, + 0.0033092384692281485, + 0.006794754881411791, + 0.014525393024086952, + 0.005830883048474789, + 0.006162822712212801, + 0.005753193516284227, + 0.002197351772338152, + 0.006792542990297079, + 0.0028850939124822617, + 0.003235945710912347, + 0.009743787348270416, + 0.0039255665615201, + 0.004307927563786507, + 0.006051645614206791, + 0.003570308443158865, + 0.007589269895106554, + 0.003598898882046342, + 0.0057843392714858055, + 0.005622311029583216, + 0.006216244772076607, + 0.005888642277568579, + 0.00833823811262846, + 0.006743387319147587, + 0.007822650484740734, + 0.003243213053792715, + 0.007774876896291971, + 0.009288119152188301, + 0.004326772876083851, + 0.0044144983403384686, + 0.005859584081918001, + 0.0021634858567267656, + 0.014784855768084526, + 0.0016781940357759595, + 0.0034633008763194084, + 0.01414555124938488, + 0.003524763509631157, + 0.002724542049691081, + 0.0046483841724693775, + 0.002209751633927226, + 0.005295505281537771, + 0.002926967339590192, + 0.0038677945267409086, + 0.0043234932236373425, + 0.003973670769482851, + 0.0043674795888364315, + 0.006681567989289761, + 0.003354705637320876, + 0.007654356770217419, + 0.002257648855447769, + 0.006079366430640221, + 0.005303302314132452, + 0.005019846837967634, + 0.004953811410814524, + 0.008429191075265408, + 0.005493371281772852, + 0.010072058066725731, + 0.0034290978219360113, + 0.008182537741959095, + 0.006866409908980131, + 0.0068334778770804405, + 0.004459935240447521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011076981201767921 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633168175816536 + ], + "opposite_logprob_tokens": [ + -3.651331663131714 + ], + "logprob_reward": -0.02633168175816536, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633168175816536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6091802716255188, + 0.13514253497123718, + 0.07508041709661484, + 0.022459082305431366, + 0.048110321164131165, + 0.030815398320555687, + 0.023763366043567657, + 0.010821373201906681, + 0.019754046574234962, + 0.02487318217754364 + ], + "credit_map": [ + 0.04763728007674217, + 0.011024685576558113, + 0.23705443739891052, + 0.01737631857395172, + 0.013116884976625443, + 0.2525448501110077, + 0.01355302706360817, + 0.016872791573405266, + 0.01712620072066784, + 0.01324949786067009, + 0.022045854479074478, + 0.007639124523848295, + 0.01816779188811779, + 0.03005801886320114, + 0.013155732303857803, + 0.013700319454073906, + 0.011019762605428696, + 0.009603261947631836, + 0.012597637251019478, + 0.004154609516263008, + 0.007496547419577837, + 0.016391748562455177, + 0.0061309002339839935, + 0.007685943506658077, + 0.002777189714834094, + 0.0020204659085720778, + 0.003687741933390498, + 0.0015278746141120791, + 0.00206844718195498, + 0.005482940003275871, + 0.002496806439012289, + 0.002397616859525442, + 0.005324491765350103, + 0.002611882286146283, + 0.014385560527443886, + 0.0027611118275672197, + 0.003625296289101243, + 0.013625926338136196, + 0.003021623007953167, + 0.002754427259787917, + 0.005360509268939495, + 0.0035277067217975855, + 0.005326415412127972, + 0.001821064273826778, + 0.0034616091288626194, + 0.005845379550009966, + 0.002769323531538248, + 0.0027033891528844833, + 0.004210266284644604, + 0.0013437385205179453, + 0.0041014463640749454, + 0.0013046935200691223, + 0.002768959617242217, + 0.003871235763654113, + 0.0029637145344167948, + 0.0031993112061172724, + 0.001448740833438933, + 0.0008363658562302589, + 0.0018873577937483788, + 0.0006884183385409415, + 0.0016072761500254273, + 0.0014014198677614331, + 0.0016417053993791342, + 0.0013100883224979043, + 0.003166880691424012, + 0.0013144196709617972, + 0.0031788900960236788, + 0.001078024972230196, + 0.002599850296974182, + 0.004385167267173529, + 0.0017603944288566709, + 0.0022704200819134712, + 0.0050318120047450066, + 0.0026396135799586773, + 0.006015230901539326, + 0.0020877141505479813, + 0.0022836332209408283, + 0.0023976529482752085, + 0.002291793469339609, + 0.0021257337648421526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02633168175816536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, static shot, shot from a level angle, captures the center of an empty circus tent. A single spotlight shines vertically from the soaring canvas dome, casting a bright circle of light on the sawdust-strewn floor, illuminating the surrounding rows of empty wooden benches and deep shadows. In the center of this luminous light, a performer in classic clown makeup—a large, exaggerated red smile framed by a pale base, with blue teardrop patterns adorning the corners of his eyes—wears a slightly askew, curly orange wig and pedals a unicycle that is comically small and contrasts sharply with his body. He wears a saturated, incongruous outfit: a loose, bright yellow silk shirt tied with a large blue and white striped bow at the collar, and baggy trousers held up by red suspenders, one leg adorned with white dots on a red background, the other with yellow stripes on a green background. He leans forward, his brows furrowed beneath the paint, beads of sweat oozing from his forehead. His legs pedal ferociously and clumsily, his entire body swaying violently with every jolt of the unicycle, his arms stretched out to the sides to maintain his precarious balance. Although the smile on his face remains constant, his focused and strained posture contrasts sharply with the vast emptiness around him, creating a surreal atmosphere that blends the comical, the poignant, and the obsessive.", + "seed": 25875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_2/index_6516_seed_25875_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10010823868215084, + "vlm_logprob": -0.10010823868215084, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.10010823868215084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35172906517982483 + ], + "opposite_logprob_tokens": [ + -1.2267290353775024 + ], + "logprob_reward": -0.35172906517982483, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35172906517982483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2003001570701599, + 0.15466508269309998, + 0.11672826111316681, + 0.13874174654483795, + 0.0639907717704773, + 0.08912904560565948, + 0.04540034011006355, + 0.04360327124595642, + 0.09590701013803482, + 0.051534298807382584 + ], + "credit_map": [ + 0.024337848648428917, + 0.030600234866142273, + 0.021588848903775215, + 0.022951077669858932, + 0.022871671244502068, + 0.02601437084376812, + 0.02239081636071205, + 0.02954529970884323, + 0.020698150619864464, + 0.018087325617671013, + 0.01939287595450878, + 0.018230250105261803, + 0.011284098960459232, + 0.027968760579824448, + 0.010368488729000092, + 0.028635140508413315, + 0.016893399879336357, + 0.01614205539226532, + 0.009792596101760864, + 0.013256245292723179, + 0.009828907437622547, + 0.015755580738186836, + 0.007324928417801857, + 0.027734549716114998, + 0.018738878890872, + 0.018094290047883987, + 0.009207429364323616, + 0.01281560305505991, + 0.028529804199934006, + 0.023971691727638245, + 0.012602697126567364, + 0.01478137169033289, + 0.009470955468714237, + 0.010548978112637997, + 0.005200737621635199, + 0.007564939558506012, + 0.0042170691303908825, + 0.009528602473437786, + 0.004282113164663315, + 0.013177386485040188, + 0.00843238178640604, + 0.011383318342268467, + 0.007875134237110615, + 0.010310784913599491, + 0.010784979909658432, + 0.014477647840976715, + 0.00699867308139801, + 0.01886613294482231, + 0.004754712805151939, + 0.006101763341575861, + 0.0027102171443402767, + 0.0043563698418438435, + 0.007852927781641483, + 0.006820191163569689, + 0.004976556170731783, + 0.00782760325819254, + 0.005071764346212149, + 0.0051268613897264, + 0.005583705380558968, + 0.005734338890761137, + 0.004466371610760689, + 0.006886905059218407, + 0.004477908834815025, + 0.006255422718822956, + 0.009752290323376656, + 0.013601803220808506, + 0.010773703455924988, + 0.010163191705942154, + 0.007775657810270786, + 0.014788717031478882, + 0.008715635165572166, + 0.02033601514995098, + 0.006154411938041449, + 0.005208533722907305, + 0.003534656250849366, + 0.004544961266219616, + 0.00742660416290164, + 0.009679930284619331, + 0.00489379744976759, + 0.010091411881148815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.35172906517982483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899098843336105 + ], + "opposite_logprob_tokens": [ + -2.578990936279297 + ], + "logprob_reward": -0.07899098843336105, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899098843336105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24290405213832855, + 0.18913710117340088, + 0.12160014361143112, + 0.15898732841014862, + 0.04603260010480881, + 0.04975618049502373, + 0.03094443492591381, + 0.08188837766647339, + 0.04702987149357796, + 0.031719937920570374 + ], + "credit_map": [ + 0.030663764104247093, + 0.04092758521437645, + 0.020279649645090103, + 0.022005099803209305, + 0.030823025852441788, + 0.03753650560975075, + 0.022695224732160568, + 0.03797318786382675, + 0.024679729714989662, + 0.01983022131025791, + 0.020383743569254875, + 0.02340129390358925, + 0.014888790436089039, + 0.031955454498529434, + 0.01359668467193842, + 0.040401168167591095, + 0.017097264528274536, + 0.014090893790125847, + 0.008713411167263985, + 0.013663602992892265, + 0.010458088479936123, + 0.016615338623523712, + 0.007972297258675098, + 0.032989244908094406, + 0.019566480070352554, + 0.02165709249675274, + 0.010960741899907589, + 0.015426917932927608, + 0.03398474305868149, + 0.02841932140290737, + 0.015309515409171581, + 0.013662509620189667, + 0.006536509841680527, + 0.006494720000773668, + 0.003417303552851081, + 0.004276905208826065, + 0.0038234007079154253, + 0.008644342422485352, + 0.003773836186155677, + 0.009065581485629082, + 0.004175280686467886, + 0.0051962994039058685, + 0.0034036110155284405, + 0.00589465256780386, + 0.007138595916330814, + 0.007284252438694239, + 0.004743216093629599, + 0.011920264922082424, + 0.003015002002939582, + 0.00408497080206871, + 0.0019021817715838552, + 0.0029592085629701614, + 0.003949219826608896, + 0.005157054401934147, + 0.00356094166636467, + 0.006315852981060743, + 0.014615220949053764, + 0.007123522460460663, + 0.005369181279093027, + 0.006736312061548233, + 0.007932999171316624, + 0.018654873594641685, + 0.00612630182877183, + 0.015329962596297264, + 0.004752512089908123, + 0.007513750344514847, + 0.004413091577589512, + 0.0050548226572573185, + 0.003603172255679965, + 0.006531758699566126, + 0.006406500935554504, + 0.008754259906709194, + 0.0024131513200700283, + 0.0032353203278034925, + 0.0020218363497406244, + 0.0031664068810641766, + 0.006612322758883238, + 0.00470348447561264, + 0.0036344751715660095, + 0.00593293784186244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07899098843336105 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540061958134174 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012540061958134174, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540061958134174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17264626920223236, + 0.11784916371107101, + 0.08358730375766754, + 0.0616670586168766, + 0.06368234008550644, + 0.20074865221977234, + 0.02778567187488079, + 0.08400426805019379, + 0.0477178692817688, + 0.1403113454580307 + ], + "credit_map": [ + 0.02149975672364235, + 0.028785143047571182, + 0.01770322024822235, + 0.021474670618772507, + 0.015164552256464958, + 0.020956577733159065, + 0.021938348188996315, + 0.025123998522758484, + 0.01716037094593048, + 0.012166484259068966, + 0.031334809958934784, + 0.01173852477222681, + 0.005209855269640684, + 0.015890486538410187, + 0.009663782082498074, + 0.014684844762086868, + 0.011233380995690823, + 0.01457051932811737, + 0.007021630182862282, + 0.012086323462426662, + 0.005445037968456745, + 0.016498154029250145, + 0.00592793058604002, + 0.010804329067468643, + 0.01008559763431549, + 0.0068258014507591724, + 0.00494903139770031, + 0.006304251030087471, + 0.006301946472376585, + 0.012377115897834301, + 0.006493091583251953, + 0.008330224081873894, + 0.007976510562002659, + 0.013138186186552048, + 0.0073885004967451096, + 0.009818951599299908, + 0.00445283530279994, + 0.00777775514870882, + 0.0067043546587228775, + 0.006425245199352503, + 0.018352117389440536, + 0.03467244654893875, + 0.02251141518354416, + 0.027730997651815414, + 0.011618733406066895, + 0.030267536640167236, + 0.018564565107226372, + 0.03703085333108902, + 0.0028470205143094063, + 0.0055675567127764225, + 0.0018337909132242203, + 0.003770921379327774, + 0.003871543100103736, + 0.003828441258519888, + 0.0022761214058846235, + 0.003790278686210513, + 0.012884377501904964, + 0.008152011781930923, + 0.00880905520170927, + 0.00959264300763607, + 0.005901928525418043, + 0.016214508563280106, + 0.009336111135780811, + 0.013113626278936863, + 0.0051101576536893845, + 0.005569023545831442, + 0.0053099035285413265, + 0.0048476108349859715, + 0.003336991649121046, + 0.006426724139600992, + 0.0051238639280200005, + 0.011993595398962498, + 0.033116329461336136, + 0.02260621078312397, + 0.01152210496366024, + 0.013128998689353466, + 0.007176612038165331, + 0.021032152697443962, + 0.010951820760965347, + 0.020777113735675812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.012540061958134174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639813400804996 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639813400804996, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639813400804996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19144564867019653, + 0.10873940587043762, + 0.11840202659368515, + 0.09004885703325272, + 0.09417637437582016, + 0.08347070962190628, + 0.04107856750488281, + 0.12613755464553833, + 0.07407306879758835, + 0.07242777943611145 + ], + "credit_map": [ + 0.025332093238830566, + 0.0299835242331028, + 0.02230336144566536, + 0.022093670442700386, + 0.016029275953769684, + 0.03350989893078804, + 0.01713871769607067, + 0.025055112317204475, + 0.012737519107758999, + 0.01246531493961811, + 0.015363190323114395, + 0.01678687147796154, + 0.008296831510961056, + 0.01705949567258358, + 0.013322080485522747, + 0.012708097696304321, + 0.01828715391457081, + 0.01940836012363434, + 0.01101730577647686, + 0.02009846456348896, + 0.006559351924806833, + 0.022440390661358833, + 0.007444718386977911, + 0.01314628031104803, + 0.016036836430430412, + 0.006934503559023142, + 0.004438940901309252, + 0.009414741769433022, + 0.006992978975176811, + 0.023804550990462303, + 0.008985770866274834, + 0.013440534472465515, + 0.013972927816212177, + 0.014827478677034378, + 0.009472722187638283, + 0.012864894233644009, + 0.006092200055718422, + 0.019036734476685524, + 0.007739262189716101, + 0.01017015241086483, + 0.007592724170535803, + 0.009179229848086834, + 0.013253750279545784, + 0.00940815731883049, + 0.0058456407859921455, + 0.01326863095164299, + 0.009368167258799076, + 0.015554404817521572, + 0.004134987015277147, + 0.0063185179606080055, + 0.002798061352223158, + 0.007356059737503529, + 0.005300101358443499, + 0.005695059895515442, + 0.0029888448771089315, + 0.006486938800662756, + 0.016105560585856438, + 0.01567455753684044, + 0.016054311767220497, + 0.01772962138056755, + 0.010452190414071083, + 0.018783172592520714, + 0.013038596138358116, + 0.018299536779522896, + 0.007929540239274502, + 0.00795629620552063, + 0.005410163663327694, + 0.008518370799720287, + 0.005832976661622524, + 0.011798432096838951, + 0.00598382530733943, + 0.020643463358283043, + 0.013524075038731098, + 0.011402984149754047, + 0.006611006800085306, + 0.007666618563234806, + 0.0040617515332996845, + 0.01153148990124464, + 0.005365089979022741, + 0.012264766730368137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.008639813400804996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static view from a level angle capturing the center of an empty circus tent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single spotlight shining vertically from the dome, casting a bright circle on a sawdust-strewn floor with rows of empty wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the performer wearing classic clown makeup with a large red smile, pale base, and blue teardrop patterns, along with a curly orange wig?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the performer wearing a loose yellow silk shirt with a blue and white striped bow, baggy trousers with red suspenders, and mismatched patterned legs (one red with white dots, one green with yellow stripes)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the performer pedaling a comically small unicycle with ferocious, clumsy motion while swaying violently to maintain balance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864126443862915 + ], + "opposite_logprob_tokens": [ + -3.0486412048339844 + ], + "logprob_reward": -0.04864126443862915, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864126443862915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1873919665813446, + 0.1486806720495224, + 0.10561566799879074, + 0.08382686972618103, + 0.15089064836502075, + 0.05746478959918022, + 0.035422127693891525, + 0.14623123407363892, + 0.04855429381132126, + 0.035921741276979446 + ], + "credit_map": [ + 0.023996220901608467, + 0.028880847617983818, + 0.02504899352788925, + 0.023245742544531822, + 0.014269757084548473, + 0.024311231449246407, + 0.022060859948396683, + 0.02557831071317196, + 0.017643019556999207, + 0.024216922000050545, + 0.01767924055457115, + 0.02434537746012211, + 0.007019341923296452, + 0.02137603424489498, + 0.015264734625816345, + 0.021135985851287842, + 0.01216172892600298, + 0.01570161059498787, + 0.011403298936784267, + 0.02004779502749443, + 0.006892429664731026, + 0.011806260794401169, + 0.01148736011236906, + 0.01611516997218132, + 0.014973321929574013, + 0.00814895797520876, + 0.005588430445641279, + 0.006946784909814596, + 0.008121737278997898, + 0.022744884714484215, + 0.005630149506032467, + 0.011672597378492355, + 0.015192447230219841, + 0.04153064265847206, + 0.01020100899040699, + 0.024061085656285286, + 0.009635663591325283, + 0.02063608728349209, + 0.014790726825594902, + 0.014842981472611427, + 0.006806230638176203, + 0.006976093631237745, + 0.006393959280103445, + 0.007320764008909464, + 0.003782299580052495, + 0.01108552236109972, + 0.005652527790516615, + 0.009447384625673294, + 0.0028041654732078314, + 0.008814379572868347, + 0.003102069953456521, + 0.0046137613244354725, + 0.004368926398456097, + 0.004198586102575064, + 0.0037493130657821894, + 0.0037709269672632217, + 0.02267252467572689, + 0.012364182621240616, + 0.011359935626387596, + 0.01129370741546154, + 0.016095060855150223, + 0.03508098050951958, + 0.01125025749206543, + 0.026114571839571, + 0.005584092810750008, + 0.0077905431389808655, + 0.00589416641741991, + 0.007506940979510546, + 0.0030210595577955246, + 0.0068953693844377995, + 0.004056516569107771, + 0.007805603090673685, + 0.00517552625387907, + 0.005932826083153486, + 0.003162628971040249, + 0.007217673119157553, + 0.0016286902828142047, + 0.0052534169517457485, + 0.0028700411785393953, + 0.0046809371560812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04864126443862915 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_tensor([3026, 3026])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_tensor([3026, 3026])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..69ce13c174336c343c0b396125b18c05a84f3560 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_tensor([3026, 3026])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from an overhead angle focuses on the corner of a dark wood desk. A pair of young, knuckled hands, trembling slightly with excitement and nervousness, hover over a thick, textured, off-white envelope containing an acceptance letter. A soft, yellow desk lamp casts a soft, warm hue from the side, highlighting the intricately designed, gilded school emblem at the center, which gleams metallically in the light. As if performing a sacred ritual, the hands carefully touch the seal of the envelope. With steady, measured movements, they firmly tear open the seal and slowly extract the neatly folded white letter paper. The entire scene is bathed in warm tones, with light and shadow delicately framing every detail. A complex mix of breathless anticipation and the joy of impending revelation permeates the air, a private moment filled with hope and dreams captured with pinpoint precision.", + "seed": 22237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_3026_seed_22237_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009321974881459028, + "vlm_logprob": -0.009321974881459028, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.009321974881459028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0060392278246581554 + ], + "opposite_logprob_tokens": [ + -6.256039142608643 + ], + "logprob_reward": -0.0060392278246581554, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0060392278246581554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3439835011959076, + 0.09472234547138214, + 0.03949504345655441, + 0.05742458254098892, + 0.03490929678082466, + 0.03593028336763382, + 0.0398128442466259, + 0.08074556291103363, + 0.19511497020721436, + 0.07786159962415695 + ], + "credit_map": [ + 0.020719362422823906, + 0.013656160794198513, + 0.06270143389701843, + 0.01697654090821743, + 0.013781125657260418, + 0.0693107545375824, + 0.02643108367919922, + 0.12040704488754272, + 0.009844688698649406, + 0.010109287686645985, + 0.01726289466023445, + 0.008180592209100723, + 0.0081495251506567, + 0.017913006246089935, + 0.012451949529349804, + 0.010810396634042263, + 0.003286126535385847, + 0.004109270870685577, + 0.003519708989188075, + 0.002969315042719245, + 0.0050320103764534, + 0.006774695124477148, + 0.005294007249176502, + 0.008509908802807331, + 0.005818060599267483, + 0.005127805285155773, + 0.008411237969994545, + 0.005018644500523806, + 0.008633906953036785, + 0.007613193243741989, + 0.006739306263625622, + 0.010062424466013908, + 0.003355480497702956, + 0.003164128866046667, + 0.006227720063179731, + 0.003759203013032675, + 0.005292544607073069, + 0.0033900695852935314, + 0.004796456545591354, + 0.004923692438751459, + 0.003811478614807129, + 0.004150803666561842, + 0.00802615936845541, + 0.0029880863148719072, + 0.0036616865545511246, + 0.0034750390332192183, + 0.004483063239604235, + 0.005333964247256517, + 0.0031973300501704216, + 0.003702949034050107, + 0.005162201821804047, + 0.003369482234120369, + 0.004038203973323107, + 0.007114994339644909, + 0.0059043592773377895, + 0.0073233237490057945, + 0.008822496049106121, + 0.007259663660079241, + 0.015518110245466232, + 0.007096418179571629, + 0.00539375701919198, + 0.015128989703953266, + 0.008466819301247597, + 0.013059306889772415, + 0.01033758744597435, + 0.011168064549565315, + 0.011698459275066853, + 0.01115171704441309, + 0.015005570836365223, + 0.06328520178794861, + 0.021584194153547287, + 0.0508841834962368, + 0.005344206467270851, + 0.007798978127539158, + 0.010295822285115719, + 0.007573310751467943, + 0.01471235416829586, + 0.010132970288395882, + 0.011562876403331757, + 0.010441086255013943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0060392278246581554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555954977869987 + ], + "opposite_logprob_tokens": [ + -4.387556076049805 + ], + "logprob_reward": -0.012555954977869987, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555954977869987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17234820127487183, + 0.04690754786133766, + 0.03969738632440567, + 0.01985032670199871, + 0.02272595837712288, + 0.050513241440057755, + 0.08975058794021606, + 0.1291085183620453, + 0.3738585114479065, + 0.05523969605565071 + ], + "credit_map": [ + 0.007451330777257681, + 0.008313634432852268, + 0.01732008345425129, + 0.007309773936867714, + 0.01203258614987135, + 0.05544380843639374, + 0.022582869976758957, + 0.041894104331731796, + 0.004720677621662617, + 0.004659381229430437, + 0.0066170827485620975, + 0.005346089601516724, + 0.005522239021956921, + 0.00702156824991107, + 0.0070995306596159935, + 0.005920976400375366, + 0.0031872375402599573, + 0.00426418986171484, + 0.005563077051192522, + 0.0031651086173951626, + 0.004479667171835899, + 0.0046370853669941425, + 0.007156668230891228, + 0.007244355510920286, + 0.002417352981865406, + 0.0030297248158603907, + 0.002606665948405862, + 0.0022086927201598883, + 0.0025800387375056744, + 0.002483461517840624, + 0.002085569081827998, + 0.0024388213641941547, + 0.003252533031627536, + 0.002784284995868802, + 0.0032311510294675827, + 0.0023688937071710825, + 0.003465682966634631, + 0.0017996405949816108, + 0.0028334977105259895, + 0.0029902749229222536, + 0.005912430118769407, + 0.005161685403436422, + 0.006909489165991545, + 0.005113227292895317, + 0.007001341786235571, + 0.004740056116133928, + 0.008996080607175827, + 0.006678933277726173, + 0.007881286554038525, + 0.006674193311482668, + 0.014347967691719532, + 0.015366930514574051, + 0.007243211381137371, + 0.012438415549695492, + 0.012533213943243027, + 0.013265370391309261, + 0.016965171322226524, + 0.012623324058949947, + 0.02607051469385624, + 0.012789350934326649, + 0.008768072351813316, + 0.0187260489910841, + 0.012164768762886524, + 0.02100127376616001, + 0.014510536566376686, + 0.014079869724810123, + 0.01938972808420658, + 0.012220367789268494, + 0.01972101628780365, + 0.1448521465063095, + 0.030795354396104813, + 0.118289515376091, + 0.004298980813473463, + 0.0051553938537836075, + 0.0064572799019515514, + 0.004469852428883314, + 0.010288123041391373, + 0.010885791853070259, + 0.0063443235121667385, + 0.007339953910559416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.012555954977869987 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096450749784708 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096450749784708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096450749784708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22718165814876556, + 0.14072197675704956, + 0.1331331878900528, + 0.0793551430106163, + 0.08123131096363068, + 0.035352569073438644, + 0.04464427009224892, + 0.11767315864562988, + 0.08113643527030945, + 0.05957023799419403 + ], + "credit_map": [ + 0.02143525891005993, + 0.015852555632591248, + 0.041804540902376175, + 0.019607078284025192, + 0.023483606055378914, + 0.02463747374713421, + 0.03124820999801159, + 0.049112945795059204, + 0.013009347021579742, + 0.013417626731097698, + 0.02260982058942318, + 0.017983775585889816, + 0.02064107358455658, + 0.014490343630313873, + 0.02203758805990219, + 0.016532408073544502, + 0.014498152770102024, + 0.02214197628200054, + 0.01936744898557663, + 0.009258989244699478, + 0.013241288252174854, + 0.014198776334524155, + 0.016954615712165833, + 0.023471955209970474, + 0.006796876434236765, + 0.00743718771263957, + 0.010705817490816116, + 0.0070068673230707645, + 0.009085973724722862, + 0.014418933540582657, + 0.011054838076233864, + 0.012848650105297565, + 0.011996710672974586, + 0.011247300542891026, + 0.015799416229128838, + 0.0066535621881484985, + 0.007826970890164375, + 0.0077537354081869125, + 0.008794909343123436, + 0.011158700101077557, + 0.0035536724608391523, + 0.0035340343602001667, + 0.007772160694003105, + 0.0028125266544520855, + 0.003795788623392582, + 0.0035249064676463604, + 0.004360976628959179, + 0.005998505745083094, + 0.005347566679120064, + 0.004734621848911047, + 0.006002834066748619, + 0.005977538414299488, + 0.00412996718659997, + 0.004637391772121191, + 0.007873873226344585, + 0.005940475966781378, + 0.01790320873260498, + 0.014186954125761986, + 0.016568824648857117, + 0.012690491043031216, + 0.009710610844194889, + 0.01771334558725357, + 0.011380652897059917, + 0.017519071698188782, + 0.006488583516329527, + 0.004840082488954067, + 0.0086439223960042, + 0.0059266663156449795, + 0.008267508819699287, + 0.014927835203707218, + 0.018230147659778595, + 0.013811691664159298, + 0.005884652491658926, + 0.005872081499546766, + 0.007432390935719013, + 0.006746226456016302, + 0.011154060252010822, + 0.0062047927640378475, + 0.009125563316047192, + 0.007150466553866863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004096450749784708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328341454267502 + ], + "opposite_logprob_tokens": [ + -3.7732834815979004 + ], + "logprob_reward": -0.02328341454267502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328341454267502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08502791821956635, + 0.039199694991111755, + 0.039555128663778305, + 0.02559507265686989, + 0.043097320944070816, + 0.032465044409036636, + 0.040600359439849854, + 0.07154837250709534, + 0.5606208443641663, + 0.06229029968380928 + ], + "credit_map": [ + 0.005340713541954756, + 0.005270822439342737, + 0.015291877090930939, + 0.007914637215435505, + 0.010665345005691051, + 0.016242211684584618, + 0.014851375482976437, + 0.009450932964682579, + 0.003873473731800914, + 0.003850548528134823, + 0.004784715361893177, + 0.0034874179400503635, + 0.003819406731054187, + 0.00843444000929594, + 0.005591144785284996, + 0.005358546040952206, + 0.003473821561783552, + 0.004248102195560932, + 0.0049447594210505486, + 0.004012123681604862, + 0.005818005185574293, + 0.004485602956265211, + 0.005969824269413948, + 0.006602890323847532, + 0.001960158348083496, + 0.0021782421972602606, + 0.002712119836360216, + 0.002436659764498472, + 0.0046445284970104694, + 0.004470341838896275, + 0.0036465299781411886, + 0.0035464875400066376, + 0.003944051451981068, + 0.00522579625248909, + 0.0067069013603031635, + 0.004601289983838797, + 0.006816311739385128, + 0.004515520296990871, + 0.005738710984587669, + 0.005548735149204731, + 0.003308334853500128, + 0.0039312769658863544, + 0.007476489990949631, + 0.0029986670706421137, + 0.0030792783945798874, + 0.0029673431999981403, + 0.003976917825639248, + 0.004726733546704054, + 0.0032721913885325193, + 0.002853961894288659, + 0.007617299444973469, + 0.004555963445454836, + 0.0034862966276705265, + 0.0071707298047840595, + 0.005748290568590164, + 0.0058956267312169075, + 0.011473425664007664, + 0.008654714561998844, + 0.012037237174808979, + 0.008415580727159977, + 0.00572608457878232, + 0.006107592023909092, + 0.007377430330961943, + 0.011756296269595623, + 0.024057602509856224, + 0.021955832839012146, + 0.030888576060533524, + 0.022579018026590347, + 0.03380812704563141, + 0.20753918588161469, + 0.052042074501514435, + 0.16775038838386536, + 0.003993322141468525, + 0.00523690041154623, + 0.008917955681681633, + 0.005070982035249472, + 0.011111597530543804, + 0.01020625326782465, + 0.009081264026463032, + 0.008672020398080349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.02328341454267502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006348263123072684 + ], + "opposite_logprob_tokens": [ + -7.375634670257568 + ], + "logprob_reward": -0.0006348263123072684, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006348263123072684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5676652789115906, + 0.05194986239075661, + 0.0780937522649765, + 0.034154247492551804, + 0.018433300778269768, + 0.030744286254048347, + 0.056353770196437836, + 0.028913328424096107, + 0.04851667210459709, + 0.08517549932003021 + ], + "credit_map": [ + 0.04254750534892082, + 0.025291739031672478, + 0.03993024677038193, + 0.023584455251693726, + 0.031220819801092148, + 0.1866747885942459, + 0.061212290078401566, + 0.15720340609550476, + 0.004855860956013203, + 0.005938568618148565, + 0.007056501694023609, + 0.006392240524291992, + 0.006178774405270815, + 0.006727307569235563, + 0.008589258417487144, + 0.006211351603269577, + 0.007177590858191252, + 0.010669992305338383, + 0.010774776339530945, + 0.00508528808131814, + 0.007632836699485779, + 0.010517762042582035, + 0.009578383527696133, + 0.016657119616866112, + 0.0029635631944984198, + 0.0029690819792449474, + 0.003623528406023979, + 0.0035251271910965443, + 0.006610761396586895, + 0.005221740808337927, + 0.004592822398990393, + 0.0046476200222969055, + 0.002023135544732213, + 0.0020180584397166967, + 0.00225503696128726, + 0.0019634368363767862, + 0.002368796616792679, + 0.002188068814575672, + 0.0032554608769714832, + 0.002361306454986334, + 0.00344132911413908, + 0.0034060871694236994, + 0.005200140178203583, + 0.0025863086339086294, + 0.002943228231742978, + 0.003639153903350234, + 0.00475606694817543, + 0.004771970212459564, + 0.005080582108348608, + 0.004836781416088343, + 0.008327297866344452, + 0.00686367554590106, + 0.005065066274255514, + 0.0076591032557189465, + 0.010346456430852413, + 0.008174804970622063, + 0.0032825618982315063, + 0.00285136210732162, + 0.0028729778714478016, + 0.002457626862451434, + 0.0031148919370025396, + 0.006308700889348984, + 0.0034779077395796776, + 0.004547297954559326, + 0.003501984290778637, + 0.002273538848385215, + 0.004184907302260399, + 0.004976560827344656, + 0.00417658407241106, + 0.012548528611660004, + 0.008004394359886646, + 0.008850174024701118, + 0.006251623388379812, + 0.006786600686609745, + 0.01132060308009386, + 0.01024361327290535, + 0.014719857834279537, + 0.01550200767815113, + 0.010638190433382988, + 0.0097130062058568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0006348263123072684 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an overhead angle focuses on the corner of a dark wood desk. A pair of young, knuckled hands, trembling slightly with excitement and nervousness, hover over a thick, textured, off-white envelope containing an acceptance letter. A soft, yellow desk lamp casts a soft, warm hue from the side, highlighting the intricately designed, gilded school emblem at the center, which gleams metallically in the light. As if performing a sacred ritual, the hands carefully touch the seal of the envelope. With steady, measured movements, they firmly tear open the seal and slowly extract the neatly folded white letter paper. The entire scene is bathed in warm tones, with light and shadow delicately framing every detail. A complex mix of breathless anticipation and the joy of impending revelation permeates the air, a private moment filled with hope and dreams captured with pinpoint precision.", + "seed": 22238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_3026_seed_22238_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011517667293082923, + "vlm_logprob": -0.011517667293082923, + "vlm_yes_no_margin": 5.95, + "vlm_token_credit": -0.011517667293082923, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003308776533231139 + ], + "opposite_logprob_tokens": [ + -7.6283087730407715 + ], + "logprob_reward": -0.003308776533231139, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.003308776533231139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2751840054988861, + 0.10667179524898529, + 0.024938497692346573, + 0.08109820634126663, + 0.05259067565202713, + 0.03399557247757912, + 0.07922378927469254, + 0.0998641699552536, + 0.11279608309268951, + 0.1336371898651123 + ], + "credit_map": [ + 0.03476740047335625, + 0.03634260594844818, + 0.0373169369995594, + 0.02762850746512413, + 0.023309191688895226, + 0.05404989793896675, + 0.033551931381225586, + 0.028217537328600883, + 0.010160560719668865, + 0.014096181839704514, + 0.018733210861682892, + 0.009273072704672813, + 0.006392895709723234, + 0.02692943811416626, + 0.010735144838690758, + 0.010351299308240414, + 0.003156200982630253, + 0.003519014921039343, + 0.004066647030413151, + 0.002623454201966524, + 0.00338590401224792, + 0.002870177151635289, + 0.0028194133192300797, + 0.002497687004506588, + 0.01339026354253292, + 0.010534042492508888, + 0.005887940526008606, + 0.004907447844743729, + 0.0073118493892252445, + 0.02726617455482483, + 0.006231998093426228, + 0.0055684917606413364, + 0.007350062020123005, + 0.008064175955951214, + 0.004465498495846987, + 0.004027002491056919, + 0.003557686461135745, + 0.0174001082777977, + 0.0033432377967983484, + 0.004382903687655926, + 0.0051686703227460384, + 0.0042994967661798, + 0.003486109897494316, + 0.002637564903125167, + 0.00436455849558115, + 0.007754799444228411, + 0.002776607172563672, + 0.0035077696666121483, + 0.014692998491227627, + 0.0073803565464913845, + 0.022996464744210243, + 0.007267671171575785, + 0.003912810236215591, + 0.011976117268204689, + 0.006179748103022575, + 0.004817627836018801, + 0.016280803829431534, + 0.010400990024209023, + 0.017843233421444893, + 0.01078717689961195, + 0.010000661946833134, + 0.010652191005647182, + 0.013147499412298203, + 0.010751614347100258, + 0.01246281061321497, + 0.013936448842287064, + 0.014942783862352371, + 0.017129406332969666, + 0.008317186497151852, + 0.012234224006533623, + 0.01917186938226223, + 0.014601359143853188, + 0.018050160259008408, + 0.02281639166176319, + 0.014215465635061264, + 0.01997165009379387, + 0.00843918789178133, + 0.013315928168594837, + 0.020321447402238846, + 0.016506953164935112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.003308776533231139 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420446764677763 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.01420446764677763, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420446764677763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26576441526412964, + 0.13778969645500183, + 0.041586004197597504, + 0.029318759217858315, + 0.04955660551786423, + 0.06385498493909836, + 0.030428890138864517, + 0.14466330409049988, + 0.1598207801580429, + 0.07721657305955887 + ], + "credit_map": [ + 0.026928644627332687, + 0.03753603249788284, + 0.037317853420972824, + 0.028530701994895935, + 0.01761048287153244, + 0.06289754807949066, + 0.028725283220410347, + 0.026217874139547348, + 0.013130820356309414, + 0.02024007961153984, + 0.017743393778800964, + 0.013572991825640202, + 0.0078502856194973, + 0.0392548032104969, + 0.013393742963671684, + 0.012603583745658398, + 0.006935310084372759, + 0.004538490436971188, + 0.006144543178379536, + 0.004688323009759188, + 0.003477875143289566, + 0.005350095219910145, + 0.0055041126906871796, + 0.004947258625179529, + 0.003884624456986785, + 0.003926526755094528, + 0.002542543690651655, + 0.0028337668627500534, + 0.0022079097107052803, + 0.008134336210787296, + 0.0030495563987642527, + 0.002739494200795889, + 0.010201689787209034, + 0.003555538132786751, + 0.010365718975663185, + 0.005713196471333504, + 0.0028838752768933773, + 0.006951467599719763, + 0.004693149123340845, + 0.005191968288272619, + 0.005891466047614813, + 0.008500838652253151, + 0.009299339726567268, + 0.0051896763034164906, + 0.004278942011296749, + 0.01756398007273674, + 0.006676713470369577, + 0.006454029120504856, + 0.0034978429321199656, + 0.0027758029755204916, + 0.0067893690429627895, + 0.001975201303139329, + 0.001541285659186542, + 0.008509204722940922, + 0.0026089432649314404, + 0.0027312415186315775, + 0.020505202934145927, + 0.016630953177809715, + 0.0204787515103817, + 0.020028943195939064, + 0.014244388788938522, + 0.017832035198807716, + 0.02068175934255123, + 0.014261271804571152, + 0.0169749204069376, + 0.02398921549320221, + 0.021069427952170372, + 0.029086068272590637, + 0.012879743240773678, + 0.01717306114733219, + 0.02233423851430416, + 0.016314100474119186, + 0.00754809333011508, + 0.014598456211388111, + 0.009469686076045036, + 0.014543121680617332, + 0.004680239129811525, + 0.009277424775063992, + 0.008392512798309326, + 0.008707037195563316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01420446764677763 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017122859135270119 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017122859135270119, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017122859135270119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17987515032291412, + 0.0853189080953598, + 0.05526399984955788, + 0.03468318656086922, + 0.0489492304623127, + 0.0894741415977478, + 0.029017459601163864, + 0.06625669449567795, + 0.21113401651382446, + 0.200027197599411 + ], + "credit_map": [ + 0.02446691505610943, + 0.01969948038458824, + 0.035197705030441284, + 0.017822610214352608, + 0.016088036820292473, + 0.017913227900862694, + 0.027364708483219147, + 0.021322481334209442, + 0.010754691436886787, + 0.009231793694198132, + 0.013732890598475933, + 0.008133809082210064, + 0.008190115913748741, + 0.011761274188756943, + 0.01167313102632761, + 0.011841203086078167, + 0.005908606573939323, + 0.00699741393327713, + 0.007447089068591595, + 0.006398390047252178, + 0.005770835094153881, + 0.00786258652806282, + 0.007281492929905653, + 0.0075975870713591576, + 0.00427172239869833, + 0.0038623877335339785, + 0.003523443592712283, + 0.003776148660108447, + 0.003982139751315117, + 0.006582637317478657, + 0.003713770769536495, + 0.00497093703597784, + 0.006466177757829428, + 0.005625027231872082, + 0.00884997472167015, + 0.0062321568839251995, + 0.0047104377299547195, + 0.005485559813678265, + 0.005403319839388132, + 0.006176574155688286, + 0.008330204524099827, + 0.013058437965810299, + 0.010948119685053825, + 0.006743692327290773, + 0.0061653475277125835, + 0.03026859275996685, + 0.0065390788950026035, + 0.007420672103762627, + 0.004075729288160801, + 0.0032114579807966948, + 0.0027079659048467875, + 0.0023839601781219244, + 0.0032242219895124435, + 0.0075399987399578094, + 0.0027032862417399883, + 0.003170839510858059, + 0.008931168355047703, + 0.007719825487583876, + 0.01140163280069828, + 0.008847150951623917, + 0.005713875871151686, + 0.006364959757775068, + 0.009526540525257587, + 0.0077515412122011185, + 0.02434639446437359, + 0.029266757890582085, + 0.027249546721577644, + 0.03646273538470268, + 0.015209915116429329, + 0.026454944163560867, + 0.02837248332798481, + 0.023771243169903755, + 0.02160640060901642, + 0.03649485111236572, + 0.023535335436463356, + 0.036354050040245056, + 0.011886845342814922, + 0.019880026578903198, + 0.024677198380231857, + 0.025592491030693054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0017122859135270119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809569403529167 + ], + "opposite_logprob_tokens": [ + -3.288095712661743 + ], + "logprob_reward": -0.03809569403529167, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809569403529167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10379486531019211, + 0.06382562965154648, + 0.030809704214334488, + 0.044694941490888596, + 0.03346037492156029, + 0.06862647831439972, + 0.09412342309951782, + 0.17266671359539032, + 0.2944470942020416, + 0.09355083853006363 + ], + "credit_map": [ + 0.013105817139148712, + 0.012257863767445087, + 0.020065445452928543, + 0.011434203945100307, + 0.009769967757165432, + 0.013476751744747162, + 0.013074866496026516, + 0.010609935969114304, + 0.007831594906747341, + 0.007757507264614105, + 0.008622179739177227, + 0.004631920251995325, + 0.005926871206611395, + 0.017737237736582756, + 0.005944894626736641, + 0.005373417865484953, + 0.0036169092636555433, + 0.0038334177806973457, + 0.0034909904934465885, + 0.0032539116218686104, + 0.0021034840028733015, + 0.00711814034730196, + 0.004208801314234734, + 0.003184050088748336, + 0.0076085906475782394, + 0.005208238027989864, + 0.0031261518597602844, + 0.00265552313067019, + 0.0031303009018301964, + 0.016695426777005196, + 0.0025185414124280214, + 0.0037521671038120985, + 0.006478360388427973, + 0.0036295857280492783, + 0.004271601792424917, + 0.0025669881142675877, + 0.0020122479181736708, + 0.009453926235437393, + 0.0021270031575113535, + 0.0029206573963165283, + 0.008042165078222752, + 0.009702851995825768, + 0.006316352169960737, + 0.0036389436572790146, + 0.004614338278770447, + 0.027828209102153778, + 0.003590590786188841, + 0.004893023520708084, + 0.008455303497612476, + 0.012235833331942558, + 0.018819468095898628, + 0.00471539469435811, + 0.004496632609516382, + 0.034478627145290375, + 0.004858419299125671, + 0.006063734646886587, + 0.02516353130340576, + 0.021198421716690063, + 0.02380627579987049, + 0.023454729467630386, + 0.016805313527584076, + 0.02173015847802162, + 0.02435613051056862, + 0.01615213230252266, + 0.03139389678835869, + 0.041676998138427734, + 0.040458645671606064, + 0.05401188135147095, + 0.02186550572514534, + 0.030394546687602997, + 0.0459664911031723, + 0.028679095208644867, + 0.009358893148601055, + 0.017651326954364777, + 0.0111234774813056, + 0.01824226602911949, + 0.005747648421674967, + 0.0098232077434659, + 0.010479815304279327, + 0.011124190874397755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.03809569403529167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002671123365871608 + ], + "opposite_logprob_tokens": [ + -8.250267028808594 + ], + "logprob_reward": -0.0002671123365871608, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.0002671123365871608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2298216074705124, + 0.14759252965450287, + 0.05356476828455925, + 0.03260793164372444, + 0.033535297960042953, + 0.035267338156700134, + 0.03257584199309349, + 0.07325078547000885, + 0.24365276098251343, + 0.11813118308782578 + ], + "credit_map": [ + 0.0375412292778492, + 0.027833085507154465, + 0.025072310119867325, + 0.018111050128936768, + 0.015287072397768497, + 0.06659407913684845, + 0.019074924290180206, + 0.020307833328843117, + 0.03107643686234951, + 0.015147787518799305, + 0.017104294151067734, + 0.008618555963039398, + 0.01028622966259718, + 0.04663020744919777, + 0.008254587650299072, + 0.010474436916410923, + 0.005815715063363314, + 0.007437993306666613, + 0.00972416065633297, + 0.004564250819385052, + 0.002928518457338214, + 0.01334038469940424, + 0.005049276631325483, + 0.0047044651582837105, + 0.004048510920256376, + 0.004042502027004957, + 0.006672993768006563, + 0.003222027327865362, + 0.003596685826778412, + 0.006505086086690426, + 0.0021105704363435507, + 0.0024095517583191395, + 0.0034318771213293076, + 0.004948889371007681, + 0.00594094255939126, + 0.0029279827140271664, + 0.002882620319724083, + 0.008458203636109829, + 0.002199812326580286, + 0.0027449687477201223, + 0.003407401032745838, + 0.005098555237054825, + 0.006254358217120171, + 0.003587424522265792, + 0.001997896935790777, + 0.009591859765350819, + 0.0024596103467047215, + 0.0028702300041913986, + 0.005865645129233599, + 0.003966858610510826, + 0.004922375548630953, + 0.00247591664083302, + 0.002701354445889592, + 0.007011008448898792, + 0.0030999728478491306, + 0.002532707992941141, + 0.009089484810829163, + 0.008988143876194954, + 0.008481464348733425, + 0.01222602091729641, + 0.006107387598603964, + 0.006466019432991743, + 0.009832341223955154, + 0.012059925124049187, + 0.029483996331691742, + 0.03615464270114899, + 0.030476249754428864, + 0.03833002224564552, + 0.01682726852595806, + 0.026858147233724594, + 0.03450377658009529, + 0.03101864457130432, + 0.015190079808235168, + 0.01844409480690956, + 0.013314774259924889, + 0.02159709483385086, + 0.007258575409650803, + 0.01045807171612978, + 0.01576605997979641, + 0.01610242761671543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0002671123365871608 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an overhead angle focuses on the corner of a dark wood desk. A pair of young, knuckled hands, trembling slightly with excitement and nervousness, hover over a thick, textured, off-white envelope containing an acceptance letter. A soft, yellow desk lamp casts a soft, warm hue from the side, highlighting the intricately designed, gilded school emblem at the center, which gleams metallically in the light. As if performing a sacred ritual, the hands carefully touch the seal of the envelope. With steady, measured movements, they firmly tear open the seal and slowly extract the neatly folded white letter paper. The entire scene is bathed in warm tones, with light and shadow delicately framing every detail. A complex mix of breathless anticipation and the joy of impending revelation permeates the air, a private moment filled with hope and dreams captured with pinpoint precision.", + "seed": 22239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_3026_seed_22239_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015173406945541501, + "vlm_logprob": -0.015173406945541501, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.015173406945541501, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006295137107372284 + ], + "opposite_logprob_tokens": [ + -6.131295204162598 + ], + "logprob_reward": -0.006295137107372284, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.006295137107372284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15169745683670044, + 0.0726442039012909, + 0.04124198481440544, + 0.2006208747625351, + 0.04370957240462303, + 0.16448116302490234, + 0.05699533224105835, + 0.16411134600639343, + 0.06317422538995743, + 0.04132390767335892 + ], + "credit_map": [ + 0.013881386257708073, + 0.01592291332781315, + 0.014583701267838478, + 0.033458366990089417, + 0.01462254486978054, + 0.020654158666729927, + 0.021426569670438766, + 0.017147796228528023, + 0.00994852650910616, + 0.010787584818899632, + 0.005880707874894142, + 0.014605306088924408, + 0.007577063050121069, + 0.008572005666792393, + 0.009191378951072693, + 0.006081631872802973, + 0.004580404609441757, + 0.0038880493957549334, + 0.004839893896132708, + 0.0043305824510753155, + 0.006951146759092808, + 0.008344156667590141, + 0.005225410684943199, + 0.0030823415145277977, + 0.014703362248837948, + 0.047196149826049805, + 0.011500312015414238, + 0.06653593480587006, + 0.016735777258872986, + 0.013615786097943783, + 0.023158006370067596, + 0.007175542414188385, + 0.00435385899618268, + 0.005671421531587839, + 0.0057941037230193615, + 0.005169330630451441, + 0.006008645985275507, + 0.006868744269013405, + 0.006032772362232208, + 0.0038106946740299463, + 0.013703061267733574, + 0.0354924276471138, + 0.023546462878584862, + 0.015216438099741936, + 0.017099132761359215, + 0.025468381121754646, + 0.014814198017120361, + 0.0191410593688488, + 0.004807670135051012, + 0.009109334088861942, + 0.006204699166119099, + 0.004726009909063578, + 0.007406257092952728, + 0.008592242375016212, + 0.009103477001190186, + 0.007045643869787455, + 0.011308601126074791, + 0.05888236686587334, + 0.013632923364639282, + 0.02862984500825405, + 0.01238663587719202, + 0.008814750239253044, + 0.01127760298550129, + 0.01917860470712185, + 0.004475062247365713, + 0.0058674318715929985, + 0.006082073785364628, + 0.005355654284358025, + 0.008615084923803806, + 0.009441141970455647, + 0.014628416858613491, + 0.008709360845386982, + 0.0036956188268959522, + 0.00514865992590785, + 0.004118403419852257, + 0.003432034980505705, + 0.005440474022179842, + 0.006962988059967756, + 0.00569620868191123, + 0.006829521618783474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.006295137107372284 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649149909615517 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649149909615517, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649149909615517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11808080226182938, + 0.05935516580939293, + 0.08058848977088928, + 0.1277138590812683, + 0.05550207570195198, + 0.11973436921834946, + 0.10221001505851746, + 0.16954801976680756, + 0.12901915609836578, + 0.03824808448553085 + ], + "credit_map": [ + 0.011919297277927399, + 0.01758621074259281, + 0.01676553301513195, + 0.014694816432893276, + 0.019986547529697418, + 0.014688864350318909, + 0.013477022759616375, + 0.008962505497038364, + 0.0042346459813416, + 0.0060062105767428875, + 0.006244931370019913, + 0.011159059591591358, + 0.009971815161406994, + 0.00793444737792015, + 0.008613563142716885, + 0.005190490745007992, + 0.005712933838367462, + 0.006197711918503046, + 0.010643761605024338, + 0.0108733419328928, + 0.013351407833397388, + 0.015516161918640137, + 0.009768243879079819, + 0.008524920791387558, + 0.010058517567813396, + 0.018921853974461555, + 0.009012575261294842, + 0.041596993803977966, + 0.010454697534441948, + 0.01306445524096489, + 0.019203342497348785, + 0.005401407368481159, + 0.004433152265846729, + 0.005760721862316132, + 0.0090628731995821, + 0.005157500971108675, + 0.011331445537507534, + 0.007039959076792002, + 0.006591571494936943, + 0.0061248429119586945, + 0.009036945179104805, + 0.033915866166353226, + 0.013741259463131428, + 0.014309004880487919, + 0.011433202773332596, + 0.01448450144380331, + 0.012023445218801498, + 0.010790133848786354, + 0.008733483962714672, + 0.013710270635783672, + 0.01211389061063528, + 0.0079568512737751, + 0.012044399045407772, + 0.017213385552167892, + 0.017189761623740196, + 0.013247957453131676, + 0.010591106489300728, + 0.06470437347888947, + 0.011327517218887806, + 0.032265208661556244, + 0.008960106410086155, + 0.008796659298241138, + 0.012124015018343925, + 0.020779021084308624, + 0.007781742140650749, + 0.014077093452215195, + 0.01535390131175518, + 0.01645648665726185, + 0.015359384939074516, + 0.020591219887137413, + 0.024513712152838707, + 0.014885617420077324, + 0.004522519186139107, + 0.0030837964732199907, + 0.0037710454780608416, + 0.003403858980163932, + 0.006174896843731403, + 0.005214303266257048, + 0.007116863038390875, + 0.004960800986737013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.008649149909615517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052552176639437675 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.0052552176639437675, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052552176639437675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19872742891311646, + 0.10054267197847366, + 0.08877474069595337, + 0.14051967859268188, + 0.0414423868060112, + 0.09833315014839172, + 0.04692225903272629, + 0.15930122137069702, + 0.030138889327645302, + 0.09529749304056168 + ], + "credit_map": [ + 0.014259674586355686, + 0.018317967653274536, + 0.02678254246711731, + 0.04530717059969902, + 0.0172115508466959, + 0.02635110728442669, + 0.028059756383299828, + 0.02243766002357006, + 0.007931383326649666, + 0.007988209836184978, + 0.012516667135059834, + 0.01563435234129429, + 0.013002724386751652, + 0.01806909218430519, + 0.016988839954137802, + 0.008411407470703125, + 0.006685481872409582, + 0.007450188510119915, + 0.011512496508657932, + 0.011467027477920055, + 0.013309406116604805, + 0.017119185999035835, + 0.014836708083748817, + 0.006394241936504841, + 0.00885231513530016, + 0.03370090201497078, + 0.007400688715279102, + 0.04517732933163643, + 0.012534107081592083, + 0.011951020918786526, + 0.015754075720906258, + 0.005149247124791145, + 0.0031759741250425577, + 0.004906407091766596, + 0.0049543967470526695, + 0.004577330779284239, + 0.006286590360105038, + 0.006370562128722668, + 0.007277160417288542, + 0.003893965855240822, + 0.008265556767582893, + 0.024888576939702034, + 0.013119256123900414, + 0.010174471884965897, + 0.009424563497304916, + 0.012762651778757572, + 0.009122961200773716, + 0.01057511568069458, + 0.003602677956223488, + 0.005964295938611031, + 0.005343773402273655, + 0.004732612054795027, + 0.005662610288709402, + 0.009461370296776295, + 0.006870636250823736, + 0.005284283310174942, + 0.009349687956273556, + 0.06128628924489021, + 0.010781676508486271, + 0.028618566691875458, + 0.009990611113607883, + 0.007854390889406204, + 0.01027605775743723, + 0.02114393375813961, + 0.0026894989423453808, + 0.002767822938039899, + 0.003142350586131215, + 0.002594316378235817, + 0.004346888046711683, + 0.004136689007282257, + 0.006449311040341854, + 0.004012011922895908, + 0.006776295602321625, + 0.010208960622549057, + 0.007214873097836971, + 0.005920593161135912, + 0.01116762962192297, + 0.02233309857547283, + 0.014724805019795895, + 0.016951240599155426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0052552176639437675 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054949671030044556 + ], + "opposite_logprob_tokens": [ + -2.9299497604370117 + ], + "logprob_reward": -0.054949671030044556, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054949671030044556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06604326516389847, + 0.040663573890924454, + 0.03330914303660393, + 0.32496941089630127, + 0.029833685606718063, + 0.0617351308465004, + 0.11690005660057068, + 0.2694360911846161, + 0.029319847002625465, + 0.027789821848273277 + ], + "credit_map": [ + 0.0071163843385875225, + 0.008062614127993584, + 0.008594281040132046, + 0.006699326913803816, + 0.009534232318401337, + 0.010783027857542038, + 0.007528617978096008, + 0.007724779192358255, + 0.0037105975206941366, + 0.004738116171211004, + 0.0041938964277505875, + 0.004244205076247454, + 0.00561833381652832, + 0.007775917183607817, + 0.006090420298278332, + 0.004292088560760021, + 0.0031307244207710028, + 0.003154398174956441, + 0.004933843854814768, + 0.0033125244081020355, + 0.004441404715180397, + 0.006005523260682821, + 0.003882669610902667, + 0.004448053892701864, + 0.01873735897243023, + 0.05700627714395523, + 0.015154315158724785, + 0.11543993651866913, + 0.025900306180119514, + 0.030942440032958984, + 0.05100298672914505, + 0.01078580692410469, + 0.0032814613077789545, + 0.0035729086957871914, + 0.003717887680977583, + 0.0032242373563349247, + 0.005037805996835232, + 0.003642079886049032, + 0.003769253846257925, + 0.0035880492068827152, + 0.005864439997822046, + 0.014573308639228344, + 0.00755787780508399, + 0.006309118587523699, + 0.0064415642991662025, + 0.008008144795894623, + 0.006260386668145657, + 0.006720291916280985, + 0.00858729425817728, + 0.016902131959795952, + 0.01231606025248766, + 0.01186873484402895, + 0.014047782868146896, + 0.022988328710198402, + 0.015727277845144272, + 0.014462451450526714, + 0.0132837425917387, + 0.1050911694765091, + 0.015642983838915825, + 0.048498574644327164, + 0.01587500236928463, + 0.013297300785779953, + 0.017555098980665207, + 0.040192194283008575, + 0.0017561748391017318, + 0.003133716993033886, + 0.0023913667537271976, + 0.002211608225479722, + 0.004330131225287914, + 0.00442374125123024, + 0.007067070342600346, + 0.004006035625934601, + 0.0023139468394219875, + 0.0023836903274059296, + 0.002011363860219717, + 0.0020380637142807245, + 0.00445200689136982, + 0.0047552743926644325, + 0.004617002792656422, + 0.005218472797423601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.054949671030044556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007178590167313814 + ], + "opposite_logprob_tokens": [ + -7.250717639923096 + ], + "logprob_reward": -0.0007178590167313814, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007178590167313814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11383505910634995, + 0.053601790219545364, + 0.03946417197585106, + 0.33292365074157715, + 0.043480370193719864, + 0.07413415610790253, + 0.07362643629312515, + 0.18180038034915924, + 0.0486641600728035, + 0.038469817489385605 + ], + "credit_map": [ + 0.013921071775257587, + 0.015308666042983532, + 0.010783761739730835, + 0.01863793283700943, + 0.015531294979155064, + 0.01431173738092184, + 0.014004159718751907, + 0.0113364327698946, + 0.007250166963785887, + 0.009684063494205475, + 0.005030352156609297, + 0.005602390971034765, + 0.006886778399348259, + 0.006944550201296806, + 0.006975253578275442, + 0.005228235851973295, + 0.0035062646493315697, + 0.008451052941381931, + 0.0034267278388142586, + 0.0047554317861795425, + 0.005171927157789469, + 0.004379026126116514, + 0.005683201365172863, + 0.0040905368514359, + 0.01834200881421566, + 0.07934325933456421, + 0.012703117914497852, + 0.11971618235111237, + 0.027416491881012917, + 0.020973702892661095, + 0.0443744920194149, + 0.010054385289549828, + 0.004413267597556114, + 0.005341160576790571, + 0.005356772802770138, + 0.004423287697136402, + 0.006115800701081753, + 0.008100483566522598, + 0.005667414516210556, + 0.004062183666974306, + 0.004623222630470991, + 0.011743723414838314, + 0.010035974904894829, + 0.008803436532616615, + 0.01064931321889162, + 0.009269055910408497, + 0.012463072314858437, + 0.006546349730342627, + 0.006826534867286682, + 0.014840671792626381, + 0.007773703895509243, + 0.007514822296798229, + 0.007506961468607187, + 0.012040422298014164, + 0.007593140471726656, + 0.009530171751976013, + 0.008649605326354504, + 0.06836199015378952, + 0.013702142052352428, + 0.03182340785861015, + 0.01349357794970274, + 0.009164297953248024, + 0.012181364931166172, + 0.0244239941239357, + 0.0035007528495043516, + 0.005580782890319824, + 0.005165773909538984, + 0.0032299053855240345, + 0.006840269546955824, + 0.007346420083194971, + 0.010645223781466484, + 0.006355030927807093, + 0.002809179713949561, + 0.004146105609834194, + 0.00342287914827466, + 0.003513558767735958, + 0.005190819967538118, + 0.008715634234249592, + 0.005499895196408033, + 0.0051717450842261314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.0007178590167313814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an overhead angle focuses on the corner of a dark wood desk. A pair of young, knuckled hands, trembling slightly with excitement and nervousness, hover over a thick, textured, off-white envelope containing an acceptance letter. A soft, yellow desk lamp casts a soft, warm hue from the side, highlighting the intricately designed, gilded school emblem at the center, which gleams metallically in the light. As if performing a sacred ritual, the hands carefully touch the seal of the envelope. With steady, measured movements, they firmly tear open the seal and slowly extract the neatly folded white letter paper. The entire scene is bathed in warm tones, with light and shadow delicately framing every detail. A complex mix of breathless anticipation and the joy of impending revelation permeates the air, a private moment filled with hope and dreams captured with pinpoint precision.", + "seed": 22240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_3/index_3026_seed_22240_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009989894193131477, + "vlm_logprob": -0.009989894193131477, + "vlm_yes_no_margin": 5.625, + "vlm_token_credit": -0.009989894193131477, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004793224390596151 + ], + "opposite_logprob_tokens": [ + -6.754793167114258 + ], + "logprob_reward": -0.004793224390596151, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.004793224390596151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.575969398021698, + 0.030846139416098595, + 0.03299972042441368, + 0.03598446026444435, + 0.03915490582585335, + 0.12600524723529816, + 0.10050531476736069, + 0.016974760219454765, + 0.022579310461878777, + 0.0189807191491127 + ], + "credit_map": [ + 0.040161844342947006, + 0.038315434008836746, + 0.18370658159255981, + 0.03149149939417839, + 0.04157941788434982, + 0.1540132761001587, + 0.056281596422195435, + 0.030419819056987762, + 0.0027141403406858444, + 0.004904724191874266, + 0.0065801795572042465, + 0.002425702288746834, + 0.00301528163254261, + 0.0057863835245370865, + 0.002842989284545183, + 0.0025767418555915356, + 0.0034522560890764, + 0.002787309931591153, + 0.005645832046866417, + 0.0023943346459418535, + 0.003150216769427061, + 0.004001064691692591, + 0.008912286721169949, + 0.0026564225554466248, + 0.0027100148145109415, + 0.0027622287161648273, + 0.006210650783032179, + 0.003579723183065653, + 0.00440091360360384, + 0.005011275410652161, + 0.008944063447415829, + 0.0023655961267650127, + 0.002692785579711199, + 0.0022688552271574736, + 0.007664965931326151, + 0.0030771768651902676, + 0.003245091065764427, + 0.004579903092235327, + 0.012984778732061386, + 0.002641355386003852, + 0.013316584751009941, + 0.01690167374908924, + 0.017139649018645287, + 0.010260390117764473, + 0.02486351691186428, + 0.017334245145320892, + 0.016044745221734047, + 0.010144449770450592, + 0.012611077167093754, + 0.006646647118031979, + 0.02052208036184311, + 0.010613000020384789, + 0.00782625563442707, + 0.021404339000582695, + 0.012202668935060501, + 0.008679253980517387, + 0.0015809895703569055, + 0.001071441569365561, + 0.0026449693832546473, + 0.001792603055946529, + 0.0013310987269505858, + 0.0014518225798383355, + 0.0052968463860452175, + 0.0018049919744953513, + 0.002353978343307972, + 0.003167701419442892, + 0.0026909320149570704, + 0.0027107291389256716, + 0.0026350023690611124, + 0.0024084802716970444, + 0.003622417338192463, + 0.002990070963278413, + 0.0021078670397400856, + 0.003132406622171402, + 0.0026403400115668774, + 0.002023436361923814, + 0.0018323780968785286, + 0.0021353322081267834, + 0.0021857761312276125, + 0.002923184772953391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.004793224390596151 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638058625161648 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.007638058625161648, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638058625161648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.435981810092926, + 0.039716366678476334, + 0.029893655329942703, + 0.04860572889447212, + 0.022126393392682076, + 0.05593501403927803, + 0.011943948455154896, + 0.08696004003286362, + 0.14354370534420013, + 0.12529334425926208 + ], + "credit_map": [ + 0.029455654323101044, + 0.008912495337426662, + 0.17018581926822662, + 0.018871009349822998, + 0.025349348783493042, + 0.148176908493042, + 0.019469356164336205, + 0.015561222098767757, + 0.0029888341668993235, + 0.007171357981860638, + 0.008285818621516228, + 0.0035603453870862722, + 0.0036705173552036285, + 0.00652370136231184, + 0.0040946113876998425, + 0.0034211815800517797, + 0.004073813557624817, + 0.002229826059192419, + 0.004278548527508974, + 0.00400521419942379, + 0.003587067825719714, + 0.003060935763642192, + 0.0065313116647303104, + 0.0021269372664391994, + 0.004249513149261475, + 0.004535418935120106, + 0.007081992924213409, + 0.0070680733770132065, + 0.007205679547041655, + 0.0050109196454286575, + 0.009482749737799168, + 0.003971381112933159, + 0.002079395344480872, + 0.0017967852763831615, + 0.0038133689668029547, + 0.0029807572718709707, + 0.002312184777110815, + 0.002881549298763275, + 0.004490063525736332, + 0.0017722898628562689, + 0.0059580933302640915, + 0.007573951501399279, + 0.008320054039359093, + 0.005426314659416676, + 0.010553045198321342, + 0.0061958893202245235, + 0.0064623067155480385, + 0.005445356480777264, + 0.0014275179710239172, + 0.0008074909565038979, + 0.001946494448930025, + 0.0020261274185031652, + 0.0010657522361725569, + 0.0011078683892264962, + 0.0022570351138710976, + 0.0013056615134701133, + 0.007013910915702581, + 0.004997886251658201, + 0.014743532985448837, + 0.010130412876605988, + 0.006033606827259064, + 0.005629315506666899, + 0.030640793964266777, + 0.007770577911287546, + 0.01392691396176815, + 0.015114382840692997, + 0.015750983729958534, + 0.020360784605145454, + 0.014303510077297688, + 0.02850387431681156, + 0.023388929665088654, + 0.01219432894140482, + 0.010349079966545105, + 0.027995267882943153, + 0.018351446837186813, + 0.012704669497907162, + 0.009871957823634148, + 0.008799389936029911, + 0.012917833402752876, + 0.024303698912262917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.007638058625161648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41463637351989746, + 0.10403495281934738, + 0.08607719838619232, + 0.04146099463105202, + 0.04940706491470337, + 0.08260450512170792, + 0.06777754426002502, + 0.044765654951334, + 0.060345686972141266, + 0.04889003187417984 + ], + "credit_map": [ + 0.036905765533447266, + 0.043458934873342514, + 0.10419473797082901, + 0.02855396829545498, + 0.04131526127457619, + 0.0835753083229065, + 0.04773229360580444, + 0.028900111094117165, + 0.007741803303360939, + 0.016031349077820778, + 0.018706200644373894, + 0.008566346950829029, + 0.014537869952619076, + 0.015613597817718983, + 0.014560628682374954, + 0.00827715639024973, + 0.00965043529868126, + 0.010639515705406666, + 0.012821437790989876, + 0.009051710367202759, + 0.010805277153849602, + 0.008288981392979622, + 0.013998879119753838, + 0.010820969007909298, + 0.00410389993339777, + 0.0037192548625171185, + 0.008205731399357319, + 0.00407783780246973, + 0.005576296243816614, + 0.005396685563027859, + 0.00641256757080555, + 0.0039687203243374825, + 0.004841981455683708, + 0.003047327511012554, + 0.008551356382668018, + 0.0044341981410980225, + 0.0054482268169522285, + 0.0037576111499220133, + 0.016045941039919853, + 0.003280423814430833, + 0.00918173510581255, + 0.008026533760130405, + 0.013578938320279121, + 0.0070931576192379, + 0.014371058903634548, + 0.016423659399151802, + 0.006771680898964405, + 0.007157743442803621, + 0.007798163685947657, + 0.004683707375079393, + 0.012996573001146317, + 0.00832771323621273, + 0.00601792149245739, + 0.012674780562520027, + 0.00972352921962738, + 0.005555157549679279, + 0.00432214979082346, + 0.0022668701130896807, + 0.007156149949878454, + 0.006457472685724497, + 0.0037758348044008017, + 0.004007148556411266, + 0.012682831846177578, + 0.0040971944108605385, + 0.00564613239839673, + 0.006685077212750912, + 0.006680784747004509, + 0.007037043571472168, + 0.00794191937893629, + 0.009485842660069466, + 0.008887258358299732, + 0.007981624454259872, + 0.00522557133808732, + 0.008996707387268543, + 0.007485501002520323, + 0.004990395158529282, + 0.003853256581351161, + 0.004524471703916788, + 0.004376620054244995, + 0.009437504224479198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0031922117341309786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369222208857536 + ], + "opposite_logprob_tokens": [ + -3.4086921215057373 + ], + "logprob_reward": -0.03369222208857536, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369222208857536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18819133937358856, + 0.0435606986284256, + 0.05678258091211319, + 0.05688682571053505, + 0.06502743810415268, + 0.09191183745861053, + 0.04720831289887428, + 0.05372966453433037, + 0.32219401001930237, + 0.07450729608535767 + ], + "credit_map": [ + 0.01970822364091873, + 0.025296706706285477, + 0.03229629248380661, + 0.017287857830524445, + 0.024424634873867035, + 0.02810157649219036, + 0.026658732444047928, + 0.014417306520044804, + 0.0034535808954387903, + 0.008526734076440334, + 0.00606820872053504, + 0.0039765844121575356, + 0.006537644658237696, + 0.004263328388333321, + 0.005872511770576239, + 0.004862105008214712, + 0.0065874201245605946, + 0.0030832362826913595, + 0.010826150886714458, + 0.0052472688257694244, + 0.00499152485281229, + 0.010986719280481339, + 0.010996434837579727, + 0.004063818603754044, + 0.00616159662604332, + 0.0033451097551733255, + 0.011161436326801777, + 0.003890876192599535, + 0.0040848408825695515, + 0.01141885481774807, + 0.013033165596425533, + 0.00379093992523849, + 0.006210988387465477, + 0.004393380135297775, + 0.011028409004211426, + 0.004790033213794231, + 0.006644992623478174, + 0.009137154556810856, + 0.018482903018593788, + 0.004339572507888079, + 0.009788012132048607, + 0.008255098015069962, + 0.012720568105578423, + 0.008226457051932812, + 0.011503908783197403, + 0.009173363447189331, + 0.015697259455919266, + 0.01654716022312641, + 0.004163934383541346, + 0.0023352927528321743, + 0.008433998562395573, + 0.004400813486427069, + 0.00410839868709445, + 0.007295093033462763, + 0.013195263221859932, + 0.0032755190040916204, + 0.006265297532081604, + 0.00347004272043705, + 0.007985583506524563, + 0.0055210222490131855, + 0.005675443448126316, + 0.004727459512650967, + 0.015451124869287014, + 0.004633682779967785, + 0.0309097021818161, + 0.03483780845999718, + 0.03890588507056236, + 0.04710446298122406, + 0.03470084071159363, + 0.06431373208761215, + 0.047561164945364, + 0.0238603875041008, + 0.00694256043061614, + 0.012978258542716503, + 0.009789151139557362, + 0.011668935418128967, + 0.007100848481059074, + 0.005956278182566166, + 0.010746910236775875, + 0.009324351325631142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.03369222208857536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from an overhead angle focusing on the corner of a dark wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of young, knuckled hands visible hovering over a thick, textured, off-white envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, yellow desk lamp casting a warm hue that highlights a gilded school emblem on the envelope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands perform the action of tearing open the envelope seal and extracting a neatly folded white letter paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene bathed in warm tones with light and shadow delicately framing the details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006337541271932423 + ], + "opposite_logprob_tokens": [ + -7.375633716583252 + ], + "logprob_reward": -0.0006337541271932423, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006337541271932423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3170994520187378, + 0.04704560711979866, + 0.04068468511104584, + 0.04009896516799927, + 0.0441448912024498, + 0.19669978320598602, + 0.02524750679731369, + 0.10769816488027573, + 0.10572516918182373, + 0.07555568963289261 + ], + "credit_map": [ + 0.03576986491680145, + 0.04294966161251068, + 0.052668649703264236, + 0.03040708601474762, + 0.04059769958257675, + 0.04079628363251686, + 0.05014749616384506, + 0.023762723430991173, + 0.003846298437565565, + 0.007740027736872435, + 0.007135193794965744, + 0.004080678801983595, + 0.005679350811988115, + 0.0038858288899064064, + 0.01013367623090744, + 0.004544553812593222, + 0.003966630436480045, + 0.004718911834061146, + 0.006518103647977114, + 0.003876818111166358, + 0.004374396055936813, + 0.005016990937292576, + 0.00831225048750639, + 0.0039005849976092577, + 0.003636283567175269, + 0.0035458204802125692, + 0.007495349273085594, + 0.0031373281963169575, + 0.004450189881026745, + 0.007223961874842644, + 0.007488989736884832, + 0.0031210393644869328, + 0.0052311900071799755, + 0.004450206644833088, + 0.006005294620990753, + 0.00479557178914547, + 0.006222731899470091, + 0.004173470661044121, + 0.008833475410938263, + 0.004432951100170612, + 0.022961441427469254, + 0.021785421296954155, + 0.029010793194174767, + 0.01456840243190527, + 0.03606598079204559, + 0.03560800105333328, + 0.018767600879073143, + 0.01793215423822403, + 0.003560339566320181, + 0.0021766319405287504, + 0.003927961457520723, + 0.0030631113331764936, + 0.002566450508311391, + 0.003301330143585801, + 0.004378181882202625, + 0.0022735013626515865, + 0.009035897441208363, + 0.006971822120249271, + 0.019279764965176582, + 0.011396265588700771, + 0.00722296629101038, + 0.007429406978189945, + 0.03611402586102486, + 0.010248019360005856, + 0.012775490991771221, + 0.014618536457419395, + 0.01795784756541252, + 0.011521118693053722, + 0.010428928770124912, + 0.013504876755177975, + 0.012948685325682163, + 0.011969693005084991, + 0.007838871330022812, + 0.01617870107293129, + 0.011217405088245869, + 0.008813715539872646, + 0.006216763518750668, + 0.006849798373878002, + 0.007044024299830198, + 0.011396413668990135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0006337541271932423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_4/index_tensor([4090, 4090])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_4/index_tensor([4090, 4090])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..62f4dda7711ada1243ef6d67b1ddba0e4f465292 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_4/index_tensor([4090, 4090])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly high-angle, static, extreme close-up focuses on a single, thick, ivory pillar candle standing on a dark, indeterminate surface. Its solitary flame flickers with a warm orange glow, forming a small, translucent pool of melted wax around its blackened wick and casting deep, dancing shadows, creating an atmosphere of tranquility and peace. Suddenly, this tranquility is shattered by an unseen weight plummeting from above. The entire frame trembles chaotically under the violent impact, simulating the brutal force of the impact. In an instant, the candle's form collapses, its flame extinguished in a final, bright puff of smoke, and the molten wax erupts explosively, splashing outward in a radial burst of golden liquid lines and droplets against the dark background, capturing the moment of shocking, destructive transformation.", + "seed": 30473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_4/index_4090_seed_30473_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03930147564969957, + "vlm_logprob": -0.03930147564969957, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03930147564969957, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026549018919467926 + ], + "opposite_logprob_tokens": [ + -5.151548862457275 + ], + "logprob_reward": -0.026549018919467926, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.026549018919467926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10335704684257507, + 0.0548771470785141, + 0.056806374341249466, + 0.052742958068847656, + 0.1199006661772728, + 0.046754974871873856, + 0.13683092594146729, + 0.2119435966014862, + 0.12509235739707947, + 0.09169399738311768 + ], + "credit_map": [ + 0.02262251265347004, + 0.006143788807094097, + 0.02127242088317871, + 0.006263359449803829, + 0.004395810421556234, + 0.019406110048294067, + 0.006261110305786133, + 0.01699194125831127, + 0.008610514923930168, + 0.003906863741576672, + 0.014419646933674812, + 0.0031196221243590117, + 0.003464434063062072, + 0.009755060076713562, + 0.003591909073293209, + 0.008009093813598156, + 0.006278864108026028, + 0.004238522611558437, + 0.013027379289269447, + 0.0030554700642824173, + 0.004069018177688122, + 0.0157221257686615, + 0.0034811669029295444, + 0.006933827418833971, + 0.011134572327136993, + 0.002164846286177635, + 0.014476929791271687, + 0.0029079432133585215, + 0.0017935583600774407, + 0.009522931650280952, + 0.002360031707212329, + 0.008382144384086132, + 0.01502310298383236, + 0.004354305565357208, + 0.03591481223702431, + 0.006461980752646923, + 0.0038518495857715607, + 0.023327140137553215, + 0.005638888571411371, + 0.025328587740659714, + 0.006774479057639837, + 0.003265906823799014, + 0.010325846262276173, + 0.0033876472152769566, + 0.0036028206814080477, + 0.010138476267457008, + 0.003289595479145646, + 0.005970199592411518, + 0.026300324127078056, + 0.00927670206874609, + 0.0347464457154274, + 0.007824207656085491, + 0.005867873318493366, + 0.03067886456847191, + 0.007241394836455584, + 0.014895117841660976, + 0.03465006500482559, + 0.02715473622083664, + 0.03311225399374962, + 0.024729566648602486, + 0.017887769266963005, + 0.021283617243170738, + 0.01703665778040886, + 0.03608892485499382, + 0.016793159767985344, + 0.015098639763891697, + 0.017913952469825745, + 0.01659260131418705, + 0.014430359937250614, + 0.013251273892819881, + 0.016782520338892937, + 0.0142298499122262, + 0.01209428533911705, + 0.015121926553547382, + 0.006614271085709333, + 0.01588376984000206, + 0.010159659199416637, + 0.01512347161769867, + 0.009029264561831951, + 0.007667344529181719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.026549018919467926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076476410031318665 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.0076476410031318665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076476410031318665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1021859273314476, + 0.056017521768808365, + 0.05307969078421593, + 0.04180850461125374, + 0.07799006253480911, + 0.04335985332727432, + 0.263523668050766, + 0.20305126905441284, + 0.12083055078983307, + 0.038152966648340225 + ], + "credit_map": [ + 0.011471842415630817, + 0.00971401296555996, + 0.02502612955868244, + 0.007510142866522074, + 0.008769377134740353, + 0.01773560605943203, + 0.0075864167883992195, + 0.014372394420206547, + 0.007236699108034372, + 0.004372599069029093, + 0.012316715903580189, + 0.004470363259315491, + 0.00423406483605504, + 0.010888508521020412, + 0.0046582212671637535, + 0.007840348407626152, + 0.008114506490528584, + 0.006562014576047659, + 0.010328453965485096, + 0.0033544949255883694, + 0.005168522708117962, + 0.008476518094539642, + 0.004533674567937851, + 0.006541504058986902, + 0.007121723145246506, + 0.0033755276817828417, + 0.008897858671844006, + 0.0026764250360429287, + 0.003331462386995554, + 0.008457646705210209, + 0.0025375448167324066, + 0.0054103173315525055, + 0.014562592841684818, + 0.002885658061131835, + 0.022755611687898636, + 0.0037534660659730434, + 0.0026691523380577564, + 0.019029218703508377, + 0.0029131092596799135, + 0.009421253576874733, + 0.004922797437757254, + 0.002421129262074828, + 0.011187342926859856, + 0.001992918783798814, + 0.0028851893730461597, + 0.013319745659828186, + 0.002617662539705634, + 0.004013067577034235, + 0.039107877761125565, + 0.027806637808680534, + 0.05504702031612396, + 0.016868747770786285, + 0.020041050389409065, + 0.05928511917591095, + 0.017537886276841164, + 0.02782933972775936, + 0.026884904131293297, + 0.0307833980768919, + 0.03438257798552513, + 0.019625307992100716, + 0.019163703545928, + 0.03042757324874401, + 0.016404999420046806, + 0.02537880279123783, + 0.018279563635587692, + 0.014043396338820457, + 0.014007568359375, + 0.015036522410809994, + 0.015443481504917145, + 0.011683792807161808, + 0.02146153524518013, + 0.010874693281948566, + 0.005224965512752533, + 0.005528741516172886, + 0.003982327412813902, + 0.004997740965336561, + 0.003771057352423668, + 0.005995908752083778, + 0.003826769534498453, + 0.0048254551365971565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.0076476410031318665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003893703222275 + ], + "opposite_logprob_tokens": [ + -2.6950390338897705 + ], + "logprob_reward": -0.07003893703222275, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003893703222275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15083427727222443, + 0.06365269422531128, + 0.0597446970641613, + 0.039656203240156174, + 0.06123363971710205, + 0.043151598423719406, + 0.06703424453735352, + 0.3125203549861908, + 0.09781360626220703, + 0.1043587327003479 + ], + "credit_map": [ + 0.03523087501525879, + 0.009232492186129093, + 0.029028238728642464, + 0.01052238792181015, + 0.010377057828009129, + 0.02435421757400036, + 0.009861955419182777, + 0.02222704514861107, + 0.008406417444348335, + 0.008312422782182693, + 0.00967464130371809, + 0.004845930729061365, + 0.006105812266469002, + 0.012516817077994347, + 0.0056090448051691055, + 0.008181600831449032, + 0.009823834523558617, + 0.005786139518022537, + 0.011372103355824947, + 0.003988041542470455, + 0.004449248779565096, + 0.01203227136284113, + 0.0047446247190237045, + 0.007548429071903229, + 0.0059261079877614975, + 0.003996833227574825, + 0.006282752845436335, + 0.0035887807607650757, + 0.004292231984436512, + 0.006689663976430893, + 0.003784520085901022, + 0.0050953105092048645, + 0.005781720858067274, + 0.0037902342155575752, + 0.015994848683476448, + 0.003038428956642747, + 0.004432956222444773, + 0.011691227555274963, + 0.00430232472717762, + 0.012201893143355846, + 0.004913310054689646, + 0.003583754412829876, + 0.011444580741226673, + 0.002392882714048028, + 0.0035276024136692286, + 0.010588515549898148, + 0.00245715887285769, + 0.004243793431669474, + 0.014763067476451397, + 0.004076551645994186, + 0.013359533622860909, + 0.005262549500912428, + 0.003025352954864502, + 0.011639249511063099, + 0.005583093035966158, + 0.00932483933866024, + 0.042933326214551926, + 0.05469173192977905, + 0.05365123599767685, + 0.030924610793590546, + 0.02465575933456421, + 0.04294665530323982, + 0.02388572134077549, + 0.038831308484077454, + 0.013341168873012066, + 0.012109808623790741, + 0.012437708675861359, + 0.012488365173339844, + 0.013449646532535553, + 0.008753852918744087, + 0.016964523121714592, + 0.008268516510725021, + 0.016098352149128914, + 0.013926692306995392, + 0.007364797871559858, + 0.015200668945908546, + 0.013238575309515, + 0.02016500197350979, + 0.01037313137203455, + 0.007991505786776543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.07003893703222275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907588571310043 + ], + "opposite_logprob_tokens": [ + -2.464075803756714 + ], + "logprob_reward": -0.08907588571310043, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907588571310043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13340219855308533, + 0.04458611086010933, + 0.07034393399953842, + 0.03196972981095314, + 0.04100274294614792, + 0.029775451868772507, + 0.06724964082241058, + 0.460985392332077, + 0.05982622131705284, + 0.060858625918626785 + ], + "credit_map": [ + 0.017503226175904274, + 0.008352527394890785, + 0.038658156991004944, + 0.007710854522883892, + 0.006705594249069691, + 0.02279500477015972, + 0.009067647159099579, + 0.022609179839491844, + 0.007493890356272459, + 0.003140898421406746, + 0.008909582160413265, + 0.0026495440397411585, + 0.0024946979247033596, + 0.011583879590034485, + 0.0031776651740074158, + 0.005135952495038509, + 0.010990478098392487, + 0.004308551549911499, + 0.012304887175559998, + 0.002895639045163989, + 0.004173868801444769, + 0.015761515125632286, + 0.004055289551615715, + 0.015853703022003174, + 0.004852629266679287, + 0.0017152026994153857, + 0.007174672093242407, + 0.0022656070068478584, + 0.002257481450214982, + 0.006390694994479418, + 0.0018011173233389854, + 0.005512325093150139, + 0.004635451827198267, + 0.0020687629003077745, + 0.009837960824370384, + 0.002205817960202694, + 0.002014154102653265, + 0.010559428483247757, + 0.0021702724043279886, + 0.007510898169130087, + 0.004835241939872503, + 0.00148515694309026, + 0.007431271485984325, + 0.0015375326620414853, + 0.001926409313455224, + 0.007809800561517477, + 0.0014213784597814083, + 0.0033286609686911106, + 0.010659456253051758, + 0.003706828225404024, + 0.017789453268051147, + 0.004535906482487917, + 0.002872372977435589, + 0.018470054492354393, + 0.0035613435320556164, + 0.005654221400618553, + 0.06311696022748947, + 0.07956542819738388, + 0.07193973660469055, + 0.05601043999195099, + 0.041759125888347626, + 0.06116402521729469, + 0.033567875623703, + 0.053861815482378006, + 0.006837350782006979, + 0.00839152280241251, + 0.007716725114732981, + 0.006100789178162813, + 0.008645731955766678, + 0.006459690630435944, + 0.008840451017022133, + 0.006833959836512804, + 0.008547736331820488, + 0.008123881183564663, + 0.0042816707864403725, + 0.010612482205033302, + 0.007304401136934757, + 0.010220471769571304, + 0.006503600161522627, + 0.005264385603368282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.08907588571310043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319589558057487 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.00319589558057487, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319589558057487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08495840430259705, + 0.02950330264866352, + 0.036906883120536804, + 0.042768508195877075, + 0.0883050262928009, + 0.03602006286382675, + 0.08939573913812637, + 0.3467791676521301, + 0.07255040109157562, + 0.1728125363588333 + ], + "credit_map": [ + 0.013363574631512165, + 0.004473008215427399, + 0.022768650203943253, + 0.006096822675317526, + 0.004870204720646143, + 0.016178904101252556, + 0.006770894397050142, + 0.010436337441205978, + 0.0050255367532372475, + 0.0025244569405913353, + 0.00677104527130723, + 0.0029807023238390684, + 0.0025747339241206646, + 0.00429385993629694, + 0.002480336930602789, + 0.0028526263777166605, + 0.004701642319560051, + 0.002888368209823966, + 0.009577498771250248, + 0.0024714325554668903, + 0.002599571365863085, + 0.008020997047424316, + 0.002504147356376052, + 0.004143224097788334, + 0.005095269531011581, + 0.0024762246757745743, + 0.010098246857523918, + 0.0030823287088423967, + 0.002691591391339898, + 0.007858822122216225, + 0.003166376845911145, + 0.008299644105136395, + 0.011882450431585312, + 0.004186098929494619, + 0.025334877893328667, + 0.005723272915929556, + 0.0035780509933829308, + 0.01677296683192253, + 0.004619221668690443, + 0.01620808243751526, + 0.004483480472117662, + 0.002039601095020771, + 0.009270875714719296, + 0.0028904711361974478, + 0.0019180930685251951, + 0.008015983738005161, + 0.0022776201367378235, + 0.0051239305175840855, + 0.015562299638986588, + 0.007057534530758858, + 0.019880041480064392, + 0.006754045840352774, + 0.0054341270588338375, + 0.018894493579864502, + 0.00705226743593812, + 0.008760925382375717, + 0.058217644691467285, + 0.038120564073324203, + 0.04751996323466301, + 0.045193180441856384, + 0.03845639526844025, + 0.04069234058260918, + 0.04230682924389839, + 0.036272212862968445, + 0.007904594764113426, + 0.01012706570327282, + 0.008061829023063183, + 0.010955914855003357, + 0.00875302217900753, + 0.007153369951993227, + 0.012135603465139866, + 0.007458989042788744, + 0.030026216059923172, + 0.023158298805356026, + 0.0133423563092947, + 0.018646135926246643, + 0.024372192099690437, + 0.03386830538511276, + 0.015324139967560768, + 0.014074874110519886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.00319589558057487 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle, static, extreme close-up focuses on a single, thick, ivory pillar candle standing on a dark, indeterminate surface. Its solitary flame flickers with a warm orange glow, forming a small, translucent pool of melted wax around its blackened wick and casting deep, dancing shadows, creating an atmosphere of tranquility and peace. Suddenly, this tranquility is shattered by an unseen weight plummeting from above. The entire frame trembles chaotically under the violent impact, simulating the brutal force of the impact. In an instant, the candle's form collapses, its flame extinguished in a final, bright puff of smoke, and the molten wax erupts explosively, splashing outward in a radial burst of golden liquid lines and droplets against the dark background, capturing the moment of shocking, destructive transformation.", + "seed": 30474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_4/index_4090_seed_30474_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06483644815161824, + "vlm_logprob": -0.06483644815161824, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06483644815161824, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028616229072213173 + ], + "opposite_logprob_tokens": [ + -4.403616428375244 + ], + "logprob_reward": -0.028616229072213173, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.028616229072213173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18239067494869232, + 0.05896588787436485, + 0.037558384239673615, + 0.06301900744438171, + 0.06968669593334198, + 0.14968647062778473, + 0.12141203135251999, + 0.1272048056125641, + 0.10223481059074402, + 0.0878412127494812 + ], + "credit_map": [ + 0.022222893312573433, + 0.0142216170206666, + 0.033873289823532104, + 0.012917228043079376, + 0.011895962990820408, + 0.0335712730884552, + 0.014888632111251354, + 0.038799770176410675, + 0.007004918996244669, + 0.0028759529814124107, + 0.009268016554415226, + 0.0035958397202193737, + 0.0052796658128499985, + 0.01116498839110136, + 0.0039223842322826385, + 0.0158541202545166, + 0.005047228652983904, + 0.003357960842549801, + 0.004289948381483555, + 0.0017907106084749103, + 0.0053807576186954975, + 0.006819422822445631, + 0.002932802075520158, + 0.007939552888274193, + 0.008398931473493576, + 0.004301330074667931, + 0.01179107092320919, + 0.00409504771232605, + 0.005044084507972002, + 0.011055920273065567, + 0.0041387686505913734, + 0.014193857088685036, + 0.010175704956054688, + 0.0052317907102406025, + 0.013400811702013016, + 0.00484350323677063, + 0.0052358307875692844, + 0.013107353821396828, + 0.005242893472313881, + 0.01244881097227335, + 0.016040263697504997, + 0.016851793974637985, + 0.020495254546403885, + 0.01618359610438347, + 0.02013743855059147, + 0.01551793608814478, + 0.02450806088745594, + 0.019952131435275078, + 0.010446586646139622, + 0.008852518163621426, + 0.014124357141554356, + 0.013283520936965942, + 0.014691207557916641, + 0.019492661580443382, + 0.020118186250329018, + 0.020402994006872177, + 0.014671320095658302, + 0.008659735321998596, + 0.012391100637614727, + 0.013378007337450981, + 0.033211201429367065, + 0.022875797003507614, + 0.008776145987212658, + 0.013241497799754143, + 0.011245859786868095, + 0.008734183385968208, + 0.010959566570818424, + 0.008118685334920883, + 0.021868672221899033, + 0.01380762830376625, + 0.011227688752114773, + 0.01627252995967865, + 0.0056995851919054985, + 0.00518727395683527, + 0.006784643977880478, + 0.004391225054860115, + 0.028128698468208313, + 0.02500814199447632, + 0.00451893825083971, + 0.008122698403894901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.028616229072213173 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963864270597696 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.005963864270597696, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963864270597696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06912973523139954, + 0.05622469633817673, + 0.06043713167309761, + 0.035731878131628036, + 0.144524484872818, + 0.17380572855472565, + 0.05109450966119766, + 0.23578794300556183, + 0.1416822075843811, + 0.03158166632056236 + ], + "credit_map": [ + 0.00796167366206646, + 0.004407831002026796, + 0.013234753161668777, + 0.004067338537424803, + 0.006735507398843765, + 0.012344151735305786, + 0.005378492176532745, + 0.014999986626207829, + 0.006737578194588423, + 0.003940669819712639, + 0.008696544915437698, + 0.0039390092715620995, + 0.004191633779555559, + 0.00955867301672697, + 0.004849158227443695, + 0.014311433769762516, + 0.00782979279756546, + 0.004786783363670111, + 0.0094895139336586, + 0.0033545882906764746, + 0.004533263389021158, + 0.010442739352583885, + 0.0031916294246912003, + 0.016808820888400078, + 0.0037866104394197464, + 0.002641151426360011, + 0.008066984824836254, + 0.0016735147219151258, + 0.0040390160866081715, + 0.007245850749313831, + 0.0018082335591316223, + 0.006470515858381987, + 0.017709994688630104, + 0.01669328287243843, + 0.013602088205516338, + 0.0062268879264593124, + 0.03201314061880112, + 0.024920614436268806, + 0.01214218232780695, + 0.02121627889573574, + 0.01598762720823288, + 0.012436210177838802, + 0.029922645539045334, + 0.022759072482585907, + 0.019881799817085266, + 0.017097430303692818, + 0.034283556044101715, + 0.0214373916387558, + 0.005319710820913315, + 0.003915628883987665, + 0.006044592242687941, + 0.004575609695166349, + 0.006641110870987177, + 0.008545784279704094, + 0.008346756920218468, + 0.00770531827583909, + 0.0352017767727375, + 0.019265230745077133, + 0.01907953992486, + 0.021687783300876617, + 0.07504095882177353, + 0.03756287321448326, + 0.010153796523809433, + 0.017795974388718605, + 0.010991227813065052, + 0.015016211196780205, + 0.016332319006323814, + 0.018538227304816246, + 0.027956517413258553, + 0.019411811605095863, + 0.012780813500285149, + 0.020655084401369095, + 0.0022778648417443037, + 0.0020434148609638214, + 0.002337635261937976, + 0.0017371104331687093, + 0.009543376974761486, + 0.0075486511923372746, + 0.002001004759222269, + 0.004092603921890259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.005963864270597696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521458566188812 + ], + "opposite_logprob_tokens": [ + -1.5021458864212036 + ], + "logprob_reward": -0.2521458566188812, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521458566188812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08175674825906754, + 0.05601722002029419, + 0.08536923676729202, + 0.05894188955426216, + 0.053599532693624496, + 0.074817955493927, + 0.038385920226573944, + 0.23987585306167603, + 0.05690617859363556, + 0.25432947278022766 + ], + "credit_map": [ + 0.011270900256931782, + 0.007074169348925352, + 0.011563504114747047, + 0.005327646620571613, + 0.010562493465840816, + 0.012842283584177494, + 0.007502276450395584, + 0.015613469295203686, + 0.0074587296694517136, + 0.0031436937861144543, + 0.012328552082180977, + 0.002299277577549219, + 0.003390922909602523, + 0.010550332255661488, + 0.003237367607653141, + 0.013608344830572605, + 0.00883581954985857, + 0.00355413812212646, + 0.01294639054685831, + 0.003358506830409169, + 0.005087367724627256, + 0.016523269936442375, + 0.005250889342278242, + 0.029812859371304512, + 0.006008557043969631, + 0.003969975747168064, + 0.008242820389568806, + 0.0031980094499886036, + 0.005817345809191465, + 0.01006097812205553, + 0.004240901675075293, + 0.01740330085158348, + 0.005570514127612114, + 0.0031874277628958225, + 0.008995364420115948, + 0.002704917686060071, + 0.0035620229318737984, + 0.010373739525675774, + 0.004815998021513224, + 0.014389545656740665, + 0.007574457209557295, + 0.00647983280941844, + 0.009195234626531601, + 0.00888566579669714, + 0.012533989734947681, + 0.009895860217511654, + 0.010171076282858849, + 0.010081835091114044, + 0.00407241377979517, + 0.0030031180940568447, + 0.004602601286023855, + 0.004214543849229813, + 0.0048305378295481205, + 0.004468993749469519, + 0.007645866367965937, + 0.005547844804823399, + 0.02349017933011055, + 0.018281128257513046, + 0.02034740336239338, + 0.03147901967167854, + 0.07498902827501297, + 0.03746308758854866, + 0.015429181046783924, + 0.018396828323602676, + 0.006401619873940945, + 0.009003883227705956, + 0.004990717861801386, + 0.005378309171646833, + 0.007742386776953936, + 0.009584342129528522, + 0.005810613743960857, + 0.007994304411113262, + 0.016317108646035194, + 0.013336222618818283, + 0.013403361663222313, + 0.011053385213017464, + 0.0804397463798523, + 0.0781485065817833, + 0.010610025376081467, + 0.031021108850836754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.2521458566188812 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026377776637673378 + ], + "opposite_logprob_tokens": [ + -3.6513776779174805 + ], + "logprob_reward": -0.026377776637673378, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026377776637673378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07280045002698898, + 0.0355343297123909, + 0.057364463806152344, + 0.029223524034023285, + 0.04680455103516579, + 0.14111854135990143, + 0.18263840675354004, + 0.2196532040834427, + 0.08811921626329422, + 0.12674327194690704 + ], + "credit_map": [ + 0.010247929953038692, + 0.0061731766909360886, + 0.01318939309567213, + 0.005734752397984266, + 0.004242038819938898, + 0.012070001102983952, + 0.005872380454093218, + 0.015270779840648174, + 0.005394146777689457, + 0.0026920372620224953, + 0.005133653525263071, + 0.0029307480435818434, + 0.0021375815849751234, + 0.005211784038692713, + 0.003362682880833745, + 0.008671694435179234, + 0.006493679713457823, + 0.00547611340880394, + 0.00848443154245615, + 0.0034688697196543217, + 0.009105252102017403, + 0.008809352293610573, + 0.004441711585968733, + 0.01108505204319954, + 0.0046204193495213985, + 0.0024674490559846163, + 0.005773910321295261, + 0.001848369836807251, + 0.002446685219183564, + 0.0050116474740207195, + 0.002020537620410323, + 0.005034503526985645, + 0.006327392999082804, + 0.003590195206925273, + 0.006934160366654396, + 0.0028992181178182364, + 0.004177011549472809, + 0.008639968000352383, + 0.003549919929355383, + 0.010686682537198067, + 0.012733014300465584, + 0.01079509500414133, + 0.022001802921295166, + 0.022529354318976402, + 0.015059485100209713, + 0.015919042751193047, + 0.02748337760567665, + 0.014597371220588684, + 0.01615094766020775, + 0.012098371982574463, + 0.02435765601694584, + 0.01668323576450348, + 0.0240483321249485, + 0.02962145023047924, + 0.03171784058213234, + 0.02796057052910328, + 0.020923011004924774, + 0.0159755926579237, + 0.014228061772882938, + 0.022252490743994713, + 0.07518628239631653, + 0.044457126408815384, + 0.011236381717026234, + 0.015394261106848717, + 0.007803445216268301, + 0.008924183435738087, + 0.008217968046665192, + 0.0075149559415876865, + 0.021680081263184547, + 0.012316902168095112, + 0.009934349916875362, + 0.01172732189297676, + 0.011876747012138367, + 0.011950663290917873, + 0.0071598985232412815, + 0.006403764244168997, + 0.03844080865383148, + 0.026157531887292862, + 0.008894364349544048, + 0.01585950143635273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.026377776637673378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078514158725739 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078514158725739, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078514158725739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035200342535972595, + 0.016877399757504463, + 0.022800466045737267, + 0.03217562660574913, + 0.026415757834911346, + 0.3447990119457245, + 0.15058347582817078, + 0.2521616816520691, + 0.07169299572706223, + 0.0472932867705822 + ], + "credit_map": [ + 0.0041390759870409966, + 0.002487236401066184, + 0.005635113921016455, + 0.0023051658645272255, + 0.0029999054968357086, + 0.006205345969647169, + 0.0028563952073454857, + 0.008572104386985302, + 0.0018479221034795046, + 0.0008639444713480771, + 0.0033079057466238737, + 0.0009181999485008419, + 0.0007827590452507138, + 0.003333563217893243, + 0.0011352564906701446, + 0.004687848500907421, + 0.002894860692322254, + 0.002588851610198617, + 0.003328464925289154, + 0.0011664925841614604, + 0.00299615366384387, + 0.004079581703990698, + 0.001818354707211256, + 0.003927707672119141, + 0.004628245253115892, + 0.0035192600917071104, + 0.0037375062238425016, + 0.0021044453606009483, + 0.0044223107397556305, + 0.005593257024884224, + 0.0030384985730051994, + 0.005132102873176336, + 0.0030292195733636618, + 0.0028504361398518085, + 0.003604804864153266, + 0.0015837537357583642, + 0.003943231422454119, + 0.004704533144831657, + 0.0023224561009556055, + 0.0043773227371275425, + 0.028790371492505074, + 0.02552533522248268, + 0.06134909763932228, + 0.046051379293203354, + 0.03689538687467575, + 0.036487285047769547, + 0.0636591911315918, + 0.046040941029787064, + 0.015006797388195992, + 0.011097746901214123, + 0.02418663166463375, + 0.016863416880369186, + 0.020774243399500847, + 0.016364140436053276, + 0.027578633278608322, + 0.018711868673563004, + 0.023305758833885193, + 0.019951041787862778, + 0.016093872487545013, + 0.03119507059454918, + 0.09276778250932693, + 0.04779130965471268, + 0.011440570466220379, + 0.009616258554160595, + 0.007714039646089077, + 0.0089419512078166, + 0.00819000881165266, + 0.006690944079309702, + 0.013295798562467098, + 0.009424609132111073, + 0.006811793427914381, + 0.010623854584991932, + 0.0031736036762595177, + 0.0032481993548572063, + 0.0029352481942623854, + 0.002613065531477332, + 0.01472814567387104, + 0.013490025885403156, + 0.0028017228469252586, + 0.00430327607318759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.011078514158725739 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle, static, extreme close-up focuses on a single, thick, ivory pillar candle standing on a dark, indeterminate surface. Its solitary flame flickers with a warm orange glow, forming a small, translucent pool of melted wax around its blackened wick and casting deep, dancing shadows, creating an atmosphere of tranquility and peace. Suddenly, this tranquility is shattered by an unseen weight plummeting from above. The entire frame trembles chaotically under the violent impact, simulating the brutal force of the impact. In an instant, the candle's form collapses, its flame extinguished in a final, bright puff of smoke, and the molten wax erupts explosively, splashing outward in a radial burst of golden liquid lines and droplets against the dark background, capturing the moment of shocking, destructive transformation.", + "seed": 30475, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_4/index_4090_seed_30475_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031588733196258545, + "vlm_logprob": -0.031588733196258545, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.031588733196258545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02287057600915432 + ], + "opposite_logprob_tokens": [ + -4.2728705406188965 + ], + "logprob_reward": -0.02287057600915432, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02287057600915432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1059083491563797, + 0.030215095728635788, + 0.017230316996574402, + 0.04302796721458435, + 0.03170238807797432, + 0.012395873665809631, + 0.022648228332400322, + 0.6155531406402588, + 0.07918142527341843, + 0.042137261480093 + ], + "credit_map": [ + 0.016299011185765266, + 0.004442088771611452, + 0.020618906244635582, + 0.005885231774300337, + 0.00656471960246563, + 0.01941005326807499, + 0.0077148545533418655, + 0.02497348189353943, + 0.0036575321573764086, + 0.0012527420185506344, + 0.005100609268993139, + 0.0014322026399895549, + 0.0015370119363069534, + 0.00571296364068985, + 0.002224839059635997, + 0.009297192096710205, + 0.0021871915087103844, + 0.0012901441659778357, + 0.0027496102266013622, + 0.0012071621604263783, + 0.0011334081646054983, + 0.003013036446645856, + 0.0017359728226438165, + 0.003913788124918938, + 0.0063888500444591045, + 0.001767017412930727, + 0.005333167966455221, + 0.0021118640433996916, + 0.0023824337404221296, + 0.008311864919960499, + 0.003610652405768633, + 0.013122109696269035, + 0.004385626874864101, + 0.0010213673813268542, + 0.00408179173246026, + 0.0013384654885157943, + 0.0013169486774131656, + 0.006978179328143597, + 0.0019262837013229728, + 0.010653721168637276, + 0.0016201449325308204, + 0.000803162925876677, + 0.002154106041416526, + 0.000815783510915935, + 0.0008128117769956589, + 0.0022646035067737103, + 0.0009337498340755701, + 0.00299150962382555, + 0.0017100496916100383, + 0.0008862542454153299, + 0.00254728551954031, + 0.0008044294081628323, + 0.001129303826019168, + 0.00631266925483942, + 0.0013244478031992912, + 0.0079337852075696, + 0.03892907872796059, + 0.023468490689992905, + 0.040902938693761826, + 0.014948000200092793, + 0.22865861654281616, + 0.11095904558897018, + 0.12456973642110825, + 0.03311721235513687, + 0.007093995343893766, + 0.004521969240158796, + 0.0056375195272266865, + 0.006902645342051983, + 0.009531628340482712, + 0.014402198605239391, + 0.023314647376537323, + 0.007776813115924597, + 0.0030982615426182747, + 0.005689399316906929, + 0.0036817502696067095, + 0.007896805182099342, + 0.006711374502629042, + 0.00435770396143198, + 0.007754844147711992, + 0.0029471174348145723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02287057600915432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962560884654522 + ], + "opposite_logprob_tokens": [ + -5.130962371826172 + ], + "logprob_reward": -0.005962560884654522, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962560884654522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05017753317952156, + 0.08633460104465485, + 0.07739651948213577, + 0.08385437726974487, + 0.03193175792694092, + 0.021487927064299583, + 0.01846979185938835, + 0.5298948884010315, + 0.055579088628292084, + 0.044873517006635666 + ], + "credit_map": [ + 0.004607067909091711, + 0.0033940558787435293, + 0.008203918114304543, + 0.002593931509181857, + 0.0033093788661062717, + 0.009459526278078556, + 0.004194579552859068, + 0.01441507413983345, + 0.01592138223350048, + 0.003019818803295493, + 0.011274575255811214, + 0.00455674109980464, + 0.005413230042904615, + 0.015862109139561653, + 0.006121659651398659, + 0.024165093898773193, + 0.01297382079064846, + 0.003155349288135767, + 0.00997608806937933, + 0.0032546904403716326, + 0.0037983220536261797, + 0.015590235590934753, + 0.004664372652769089, + 0.023983638733625412, + 0.012761945836246014, + 0.0024109382648020983, + 0.011157295666635036, + 0.0036748135462403297, + 0.004312420729547739, + 0.01718829572200775, + 0.005282075610011816, + 0.027066592127084732, + 0.004033194854855537, + 0.0013426716905087233, + 0.004771103151142597, + 0.0015164460055530071, + 0.0013182749971747398, + 0.006615393795073032, + 0.0023753484711050987, + 0.0099593261256814, + 0.0027007812168449163, + 0.001357675064355135, + 0.0036874592769891024, + 0.0011276103323325515, + 0.0012612859718501568, + 0.0036614565178751945, + 0.0018030975479632616, + 0.005888562649488449, + 0.0013444711221382022, + 0.0008878344669938087, + 0.0023196344263851643, + 0.0007849275134503841, + 0.000940874801017344, + 0.004859205801039934, + 0.0012804849538952112, + 0.006052358541637659, + 0.035834211856126785, + 0.021419165655970573, + 0.028595566749572754, + 0.01204665657132864, + 0.19508835673332214, + 0.036572523415088654, + 0.12441210448741913, + 0.07592631876468658, + 0.00480551365762949, + 0.0035265334881842136, + 0.004658652003854513, + 0.005870271008461714, + 0.006640289910137653, + 0.010495830327272415, + 0.013764713890850544, + 0.0058172824792563915, + 0.0030875822994858027, + 0.004969785921275616, + 0.00336117809638381, + 0.00957360677421093, + 0.0067839836701750755, + 0.004840731620788574, + 0.00893489457666874, + 0.0033217521850019693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.005962560884654522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003293186426163 + ], + "opposite_logprob_tokens": [ + -2.695032835006714 + ], + "logprob_reward": -0.07003293186426163, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003293186426163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045990876853466034, + 0.049578797072172165, + 0.03468134626746178, + 0.08867502957582474, + 0.08571219444274902, + 0.03067302703857422, + 0.013093321584165096, + 0.5307002663612366, + 0.09577445685863495, + 0.025120627135038376 + ], + "credit_map": [ + 0.007498729974031448, + 0.004432204179465771, + 0.009227818809449673, + 0.00288142217323184, + 0.004167815670371056, + 0.007291857618838549, + 0.003795055905357003, + 0.006695971824228764, + 0.00802011787891388, + 0.0031671347096562386, + 0.00870510283857584, + 0.002801997121423483, + 0.0030567399226129055, + 0.008220125921070576, + 0.004085784312337637, + 0.01152179203927517, + 0.0060201180167496204, + 0.002717572031542659, + 0.005449995398521423, + 0.001963574904948473, + 0.0026956817600876093, + 0.005473944824188948, + 0.002838537562638521, + 0.00752192223444581, + 0.013487888500094414, + 0.003509409725666046, + 0.01136558223515749, + 0.004013547208160162, + 0.0048033734783530235, + 0.017222367227077484, + 0.006339651066809893, + 0.027933206409215927, + 0.012236748822033405, + 0.002465509809553623, + 0.009889446198940277, + 0.0037315485533326864, + 0.0038742804899811745, + 0.01891072280704975, + 0.005777406040579081, + 0.028826531022787094, + 0.0045770322903990746, + 0.0013429014943540096, + 0.003381391754373908, + 0.001599977258592844, + 0.001992749283090234, + 0.00619090860709548, + 0.002460798481479287, + 0.009127267636358738, + 0.0012997727608308196, + 0.0006244386313483119, + 0.0017507323063910007, + 0.0005432522157207131, + 0.0006299465894699097, + 0.0035486374981701374, + 0.0007738853455521166, + 0.003922655712813139, + 0.03194940835237503, + 0.02085910178720951, + 0.037299007177352905, + 0.01567813940346241, + 0.16742844879627228, + 0.12920354306697845, + 0.08598863333463669, + 0.042293980717659, + 0.008911166340112686, + 0.005463557783514261, + 0.007291813846677542, + 0.008964961394667625, + 0.010451380163431168, + 0.019031157717108727, + 0.028195304796099663, + 0.0074651166796684265, + 0.002542164409533143, + 0.0030009718611836433, + 0.002122695092111826, + 0.004827972501516342, + 0.0034277408849447966, + 0.002828225726261735, + 0.004492313601076603, + 0.0018785435240715742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.07003293186426163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498126521706581 + ], + "opposite_logprob_tokens": [ + -2.929981231689453 + ], + "logprob_reward": -0.05498126521706581, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498126521706581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049970563501119614, + 0.09138239920139313, + 0.058958377689123154, + 0.012201367877423763, + 0.026802431792020798, + 0.009082570672035217, + 0.025742623955011368, + 0.25324541330337524, + 0.44776052236557007, + 0.02485373429954052 + ], + "credit_map": [ + 0.0063550276681780815, + 0.002801866503432393, + 0.010229585692286491, + 0.004574350081384182, + 0.002856234787032008, + 0.008592941798269749, + 0.004711431916803122, + 0.00984912645071745, + 0.015275878831744194, + 0.002696848940104246, + 0.010627315379679203, + 0.004149675369262695, + 0.004838393535465002, + 0.01781086064875126, + 0.0066396561451256275, + 0.029343776404857635, + 0.010280645452439785, + 0.0018448445480316877, + 0.00704475911334157, + 0.002629360882565379, + 0.003342200070619583, + 0.011547659523785114, + 0.004287061281502247, + 0.01798184961080551, + 0.00144589738920331, + 0.0005680694594047964, + 0.0021585107315331697, + 0.0006107579101808369, + 0.0007781835156492889, + 0.0022484585642814636, + 0.0009674450848251581, + 0.0034240453969687223, + 0.004101000260561705, + 0.0009689932921901345, + 0.0030934817623347044, + 0.001312243053689599, + 0.0013824478955939412, + 0.005573516245931387, + 0.0021606548689305782, + 0.008210092782974243, + 0.00098783103749156, + 0.0008838038193061948, + 0.0011649082880467176, + 0.0009556334116496146, + 0.0008634997648186982, + 0.0013840441824868321, + 0.0008703527273610234, + 0.001972497906535864, + 0.0022338549606502056, + 0.0014715447323396802, + 0.0031647151336073875, + 0.0017003469401970506, + 0.0017682720208540559, + 0.005665389820933342, + 0.002815599786117673, + 0.00692289974540472, + 0.014567790552973747, + 0.011285735294222832, + 0.018005428835749626, + 0.010145827196538448, + 0.08080395311117172, + 0.058201927691698074, + 0.04147034138441086, + 0.018764397129416466, + 0.03626162186264992, + 0.023748023435473442, + 0.03510696068406105, + 0.04073677584528923, + 0.046434976160526276, + 0.09893139451742172, + 0.13045525550842285, + 0.03608550876379013, + 0.0022237002849578857, + 0.002919855760410428, + 0.0020524142310023308, + 0.0034101863857358694, + 0.0052450597286224365, + 0.003382103517651558, + 0.0031601199880242348, + 0.0024602951016277075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.05498126521706581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040963320061564445 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.0040963320061564445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040963320061564445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05621273070573807, + 0.01947765238583088, + 0.04597051441669464, + 0.03486999496817589, + 0.042131341993808746, + 0.022663438692688942, + 0.007833646610379219, + 0.4304707944393158, + 0.20911075174808502, + 0.13125915825366974 + ], + "credit_map": [ + 0.005257450509816408, + 0.00232041347771883, + 0.009553579613566399, + 0.0027089500799775124, + 0.0028131038416177034, + 0.01199799682945013, + 0.003795814933255315, + 0.01776541955769062, + 0.004120653495192528, + 0.0012678178027272224, + 0.0027350864838808775, + 0.001544022699818015, + 0.001459235791116953, + 0.003309945808723569, + 0.0017589246854186058, + 0.003281964221969247, + 0.007795384153723717, + 0.0014257237780839205, + 0.005453160963952541, + 0.0019711717031896114, + 0.0025743627920746803, + 0.00907799880951643, + 0.0031994502060115337, + 0.014473264105618, + 0.005093554500490427, + 0.001287628896534443, + 0.004479874391108751, + 0.0015228877309709787, + 0.0017056668875738978, + 0.007499272469431162, + 0.002058009384199977, + 0.011223099194467068, + 0.0065204547718167305, + 0.0012568250531330705, + 0.005014275200664997, + 0.0017161312280222774, + 0.0019387839129194617, + 0.009096185676753521, + 0.0027848619502037764, + 0.013803824782371521, + 0.0024487667251378298, + 0.001006593112833798, + 0.004238869063556194, + 0.0012547321384772658, + 0.0013308754423633218, + 0.004592583980411291, + 0.0018174608703702688, + 0.00597355654463172, + 0.0010011520935222507, + 0.0006751629989594221, + 0.0014901545364409685, + 0.0006240852526389062, + 0.0006203626398928463, + 0.0015001348219811916, + 0.000676651659887284, + 0.0012459420831874013, + 0.030679449439048767, + 0.018970387056469917, + 0.027379997074604034, + 0.012294630520045757, + 0.1633947640657425, + 0.047316163778305054, + 0.08322757482528687, + 0.04720783606171608, + 0.01978934183716774, + 0.015320587903261185, + 0.018198784440755844, + 0.024036191403865814, + 0.02343754656612873, + 0.035666219890117645, + 0.05493176355957985, + 0.01773030497133732, + 0.009856123477220535, + 0.016136154532432556, + 0.008970256894826889, + 0.023558318614959717, + 0.02301858738064766, + 0.01607714407145977, + 0.02428954280912876, + 0.009353028610348701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0040963320061564445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle, static, extreme close-up focuses on a single, thick, ivory pillar candle standing on a dark, indeterminate surface. Its solitary flame flickers with a warm orange glow, forming a small, translucent pool of melted wax around its blackened wick and casting deep, dancing shadows, creating an atmosphere of tranquility and peace. Suddenly, this tranquility is shattered by an unseen weight plummeting from above. The entire frame trembles chaotically under the violent impact, simulating the brutal force of the impact. In an instant, the candle's form collapses, its flame extinguished in a final, bright puff of smoke, and the molten wax erupts explosively, splashing outward in a radial burst of golden liquid lines and droplets against the dark background, capturing the moment of shocking, destructive transformation.", + "seed": 30476, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_4/index_4090_seed_30476_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029546061204746364, + "vlm_logprob": -0.029546061204746364, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.029546061204746364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03869085758924484 + ], + "opposite_logprob_tokens": [ + -3.9136908054351807 + ], + "logprob_reward": -0.03869085758924484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03869085758924484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17122909426689148, + 0.04254531115293503, + 0.05576179549098015, + 0.035403382033109665, + 0.04642165079712868, + 0.06943690776824951, + 0.17979179322719574, + 0.19850492477416992, + 0.10922013968229294, + 0.0916849747300148 + ], + "credit_map": [ + 0.03755839914083481, + 0.013337573036551476, + 0.03120538592338562, + 0.01667892187833786, + 0.014641598798334599, + 0.019449353218078613, + 0.013207903131842613, + 0.025149958208203316, + 0.009805689565837383, + 0.003850563894957304, + 0.00824978481978178, + 0.0036580979358404875, + 0.004250336904078722, + 0.004371884744614363, + 0.002788890153169632, + 0.0055700670927762985, + 0.016384193673729897, + 0.0031515166629105806, + 0.010767599567770958, + 0.004160932265222073, + 0.0046622902154922485, + 0.0044815875589847565, + 0.0036079008132219315, + 0.008545774966478348, + 0.009893865324556828, + 0.002486718585714698, + 0.006405739579349756, + 0.002970183501020074, + 0.0023508293088525534, + 0.0033486662432551384, + 0.0026700871530920267, + 0.005277291405946016, + 0.013537849299609661, + 0.002905570901930332, + 0.009136999025940895, + 0.0030889634508639574, + 0.004039057530462742, + 0.003589617321267724, + 0.0030975383706390858, + 0.007026054430752993, + 0.013937357813119888, + 0.007153173442929983, + 0.00993986427783966, + 0.006235349457710981, + 0.007998409681022167, + 0.007003887556493282, + 0.005394532810896635, + 0.01177433505654335, + 0.02609516866505146, + 0.018180441111326218, + 0.029110154137015343, + 0.019528230652213097, + 0.02471081167459488, + 0.012123605236411095, + 0.03607930243015289, + 0.013964073732495308, + 0.01953968033194542, + 0.016162432730197906, + 0.026790600270032883, + 0.021808678284287453, + 0.029667284339666367, + 0.022841759026050568, + 0.02453288994729519, + 0.03716159611940384, + 0.012638663873076439, + 0.018676381558179855, + 0.0269590076059103, + 0.007340596988797188, + 0.011031857691705227, + 0.009207302704453468, + 0.007031621877104044, + 0.016334710642695427, + 0.01646309159696102, + 0.015817375853657722, + 0.010073664598166943, + 0.0077337599359452724, + 0.01046132855117321, + 0.013372478075325489, + 0.00849081575870514, + 0.009272463619709015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03869085758924484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00320825376547873 + ], + "opposite_logprob_tokens": [ + -5.753208160400391 + ], + "logprob_reward": -0.00320825376547873, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00320825376547873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1103917583823204, + 0.046121321618556976, + 0.05819901078939438, + 0.03278912976384163, + 0.06795736402273178, + 0.34044814109802246, + 0.18826185166835785, + 0.06222265958786011, + 0.04191521182656288, + 0.05169359967112541 + ], + "credit_map": [ + 0.019526658579707146, + 0.009596653282642365, + 0.016457464545965195, + 0.010656033642590046, + 0.01637202501296997, + 0.012798814103007317, + 0.009476203471422195, + 0.015507909469306469, + 0.00800677016377449, + 0.004003958310931921, + 0.008135494776070118, + 0.004616286139935255, + 0.006182029377669096, + 0.003963074646890163, + 0.004688503686338663, + 0.006525204051285982, + 0.013156184926629066, + 0.004328210838139057, + 0.01103671919554472, + 0.005903082899749279, + 0.005513755604624748, + 0.005136433523148298, + 0.005627733189612627, + 0.007496889214962721, + 0.005628314800560474, + 0.002924085594713688, + 0.005789653863757849, + 0.003242054022848606, + 0.002996092429384589, + 0.0030421344563364983, + 0.003246980020776391, + 0.005919813644140959, + 0.00971793569624424, + 0.005367266945540905, + 0.014365318231284618, + 0.006410352420061827, + 0.0068865432403981686, + 0.006826394237577915, + 0.0067074778489768505, + 0.011676075868308544, + 0.04030633345246315, + 0.01996106468141079, + 0.05392826348543167, + 0.025295943021774292, + 0.03352413326501846, + 0.04153475537896156, + 0.023344283923506737, + 0.1025533676147461, + 0.017633795738220215, + 0.018580131232738495, + 0.03378376364707947, + 0.02422863431274891, + 0.02640658989548683, + 0.017944347113370895, + 0.03632707521319389, + 0.013357523828744888, + 0.00856947060674429, + 0.006168777588754892, + 0.009056846611201763, + 0.007061626762151718, + 0.010233188048005104, + 0.007353351917117834, + 0.00583950150758028, + 0.007939898408949375, + 0.0046952879056334496, + 0.008249283768236637, + 0.004163367673754692, + 0.004706393461674452, + 0.006322034168988466, + 0.005493760108947754, + 0.005111538805067539, + 0.0031735494267195463, + 0.010463431477546692, + 0.008639877662062645, + 0.004391014110296965, + 0.006399241741746664, + 0.006111491937190294, + 0.006118595600128174, + 0.0058318148367106915, + 0.0037381313741207123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.00320825376547873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822157695889473 + ], + "opposite_logprob_tokens": [ + -3.5298221111297607 + ], + "logprob_reward": -0.029822157695889473, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822157695889473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06845957040786743, + 0.0654059424996376, + 0.03671782463788986, + 0.056861672550439835, + 0.10625778883695602, + 0.06010766327381134, + 0.26989713311195374, + 0.13497263193130493, + 0.15231887996196747, + 0.049000874161720276 + ], + "credit_map": [ + 0.012517140246927738, + 0.006759322248399258, + 0.011599654331803322, + 0.0070570083335042, + 0.008457805030047894, + 0.008581125177443027, + 0.004894724581390619, + 0.008592796511948109, + 0.011591415852308273, + 0.004995982628315687, + 0.011478626169264317, + 0.005616419017314911, + 0.008608036674559116, + 0.006116466596722603, + 0.006598041392862797, + 0.010400956496596336, + 0.007607069797813892, + 0.003246310167014599, + 0.007288122083991766, + 0.0027319060172885656, + 0.005000430624932051, + 0.003596472553908825, + 0.002863830653950572, + 0.004383684601634741, + 0.01256927102804184, + 0.0035634415689855814, + 0.009841668419539928, + 0.0038431533612310886, + 0.008176167495548725, + 0.005584822501987219, + 0.00440899096429348, + 0.00887416023761034, + 0.01491805911064148, + 0.008236956782639027, + 0.019908566027879715, + 0.012591798789799213, + 0.015233064070343971, + 0.008421671576797962, + 0.013878987170755863, + 0.013068695552647114, + 0.009346680715680122, + 0.004295560996979475, + 0.009631829336285591, + 0.004913834854960442, + 0.003911515697836876, + 0.007173605728894472, + 0.00415022810921073, + 0.016684409230947495, + 0.027503110468387604, + 0.019594673067331314, + 0.056334417313337326, + 0.030429299920797348, + 0.03704660385847092, + 0.021800849586725235, + 0.05816497653722763, + 0.019023209810256958, + 0.015103604644536972, + 0.012129678390920162, + 0.018890736624598503, + 0.0140169532969594, + 0.01273219846189022, + 0.019826045259833336, + 0.015167532488703728, + 0.027105901390314102, + 0.015985697507858276, + 0.0359155647456646, + 0.022349661216139793, + 0.013034515082836151, + 0.019163699820637703, + 0.022889181971549988, + 0.011502391658723354, + 0.011478174477815628, + 0.0111083984375, + 0.007911430671811104, + 0.008650150150060654, + 0.003989406395703554, + 0.005553442519158125, + 0.005409927573055029, + 0.003684577066451311, + 0.00269354903139174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.029822157695889473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005749642848969 + ], + "opposite_logprob_tokens": [ + -2.6950573921203613 + ], + "logprob_reward": -0.07005749642848969, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005749642848969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08977130055427551, + 0.03839864581823349, + 0.05403692275285721, + 0.035053472965955734, + 0.029629463329911232, + 0.04246694594621658, + 0.18916703760623932, + 0.29971805214881897, + 0.18165923655033112, + 0.04009894281625748 + ], + "credit_map": [ + 0.019184544682502747, + 0.005067487247288227, + 0.017418978735804558, + 0.0068354057148098946, + 0.008741357363760471, + 0.008527649566531181, + 0.006242832168936729, + 0.01775304228067398, + 0.010079728439450264, + 0.002522645751014352, + 0.006782036740332842, + 0.002726270817220211, + 0.004253995604813099, + 0.0033697853796184063, + 0.002486689016222954, + 0.006177492905408144, + 0.014103326946496964, + 0.003183131106197834, + 0.011617421172559261, + 0.005360172595828772, + 0.004188401624560356, + 0.003315551672130823, + 0.004532579332590103, + 0.007736337371170521, + 0.007796738762408495, + 0.002153550274670124, + 0.007700342684984207, + 0.002665913663804531, + 0.0030483747832477093, + 0.0027840184047818184, + 0.0021320271771401167, + 0.006772508379071951, + 0.007205480709671974, + 0.001765820081345737, + 0.006090892944484949, + 0.0020305777434259653, + 0.0033744950778782368, + 0.002356216311454773, + 0.00189245177898556, + 0.004913529381155968, + 0.006823998875916004, + 0.0032177302055060863, + 0.006815998815000057, + 0.0036088600754737854, + 0.004073513206094503, + 0.006366667803376913, + 0.0025315626990050077, + 0.009028609842061996, + 0.030723437666893005, + 0.01502616610378027, + 0.03620191663503647, + 0.020687488839030266, + 0.018350936472415924, + 0.019377868622541428, + 0.03495015576481819, + 0.013849068433046341, + 0.040382422506809235, + 0.035473719239234924, + 0.042506422847509384, + 0.03702262043952942, + 0.03925004228949547, + 0.031048007309436798, + 0.03206133842468262, + 0.041973475366830826, + 0.02450612746179104, + 0.03545377776026726, + 0.03025827370584011, + 0.016883734613656998, + 0.027039019390940666, + 0.018677232787013054, + 0.012510327622294426, + 0.016330745071172714, + 0.006150488276034594, + 0.005876339506357908, + 0.005538005381822586, + 0.0038635486271232367, + 0.005434644874185324, + 0.005442251916974783, + 0.0038877399638295174, + 0.0039059240370988846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.07005749642848969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly high-angle static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the framing an extreme close-up of a single thick ivory pillar candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the candle have a blackened wick and a small pool of melted wax around it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an unseen weight plummeting from above causing the frame to tremble chaotically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the candle collapse with its flame extinguished in a puff of smoke and molten wax erupting in a radial burst?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951540544629097 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951540544629097, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951540544629097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11134976148605347, + 0.05239792913198471, + 0.06505902111530304, + 0.05862188711762428, + 0.06094842776656151, + 0.04780447110533714, + 0.13234558701515198, + 0.2695557773113251, + 0.12515662610530853, + 0.076760433614254 + ], + "credit_map": [ + 0.025420349091291428, + 0.005216111429035664, + 0.023666493594646454, + 0.006318903993815184, + 0.010371820069849491, + 0.010128798894584179, + 0.006441574543714523, + 0.023785719648003578, + 0.009313292801380157, + 0.004047498106956482, + 0.008128348737955093, + 0.005438029766082764, + 0.006377924233675003, + 0.005122821778059006, + 0.004644379485398531, + 0.009325634688138962, + 0.011879130266606808, + 0.005182258319109678, + 0.010072990320622921, + 0.00824489165097475, + 0.008002487011253834, + 0.005718594416975975, + 0.0076062544248998165, + 0.008352421224117279, + 0.015896610915660858, + 0.0038829187396913767, + 0.010691622272133827, + 0.004803396295756102, + 0.006096352823078632, + 0.006089375354349613, + 0.004273495636880398, + 0.006888116244226694, + 0.014722941443324089, + 0.004366106353700161, + 0.01083561684936285, + 0.003647256875410676, + 0.006969416048377752, + 0.005961794871836901, + 0.0037138420157134533, + 0.010731454938650131, + 0.009640298783779144, + 0.003432184923440218, + 0.008254719898104668, + 0.00409549567848444, + 0.003548524109646678, + 0.0038511778693646193, + 0.003100314177572727, + 0.0118817538022995, + 0.0239491555839777, + 0.011926635168492794, + 0.01776159182190895, + 0.014904681593179703, + 0.016066301614046097, + 0.011937140487134457, + 0.01695367880165577, + 0.01884639821946621, + 0.031411830335855484, + 0.02318403869867325, + 0.03866804763674736, + 0.03234599903225899, + 0.03160921111702919, + 0.033458177000284195, + 0.03101493790745735, + 0.04786355048418045, + 0.01707821525633335, + 0.025496819987893105, + 0.02173977717757225, + 0.009460842236876488, + 0.01772640272974968, + 0.01466396450996399, + 0.0079424399882555, + 0.011048170737922192, + 0.016702918335795403, + 0.011236132122576237, + 0.009384473785758018, + 0.006735557224601507, + 0.006665929686278105, + 0.010650117881596088, + 0.0063934349454939365, + 0.008991873823106289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.005951540544629097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_5/index_tensor([170, 170])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_5/index_tensor([170, 170])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..890759dc127b9c08a79484664bacdba900d77790 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_5/index_tensor([170, 170])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, rendered in the ethereal style of classical ink painting. A magnificent black blade hangs horizontally, mere inches above the surface of perfectly still water, floating serenely. Its body, crafted from a material resembling polished black obsidian, absorbs the soft ambient light, imparting a deep matte finish. Intricate, enigmatic carvings, like ancient, fluid figures, are meticulously sculpted along the length of the blade, seemingly harboring a dormant energy. Beneath the enchanted weapon, the water is an undisturbed mirror, reflecting a hazy, gray-white void. From a point directly below the blade's tip, soft ripples expand outward, less like disturbed water than like delicate ink spread across rice paper; these concentric circles move with a slow, hypnotic grace, their feathered edges dissolving into the tranquil surface. The entire composition is defined by a heavily desaturated color grading, creating a monochromatic world of black, white, and infinite shades of gray, enhancing the minimalist atmosphere and evoking a powerful sense of calm and ancient power.", + "seed": 23160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_5/index_170_seed_23160_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03570897274184972, + "vlm_logprob": -0.03570897274184972, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.03570897274184972, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013277565129101276 + ], + "opposite_logprob_tokens": [ + -4.513277530670166 + ], + "logprob_reward": -0.013277565129101276, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013277565129101276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3619573414325714, + 0.18046934902668, + 0.053599435836076736, + 0.014820868149399757, + 0.035557277500629425, + 0.026006249710917473, + 0.01818431355059147, + 0.11127676069736481, + 0.08082281798124313, + 0.11730562150478363 + ], + "credit_map": [ + 0.03251749649643898, + 0.04073547199368477, + 0.060572005808353424, + 0.06936535984277725, + 0.04329310357570648, + 0.03157547861337662, + 0.050578776746988297, + 0.03331964462995529, + 0.01642344705760479, + 0.025021029636263847, + 0.019947396591305733, + 0.04613887891173363, + 0.017422879114747047, + 0.012413321062922478, + 0.025887051597237587, + 0.01721533201634884, + 0.005033680237829685, + 0.00744459917768836, + 0.006715567782521248, + 0.006980160251259804, + 0.006157833617180586, + 0.005002603400498629, + 0.009758100844919682, + 0.006506888195872307, + 0.0015102762263268232, + 0.0018240319332107902, + 0.0023780560586601496, + 0.0013425203505903482, + 0.0027088590431958437, + 0.001468057045713067, + 0.0015062171733006835, + 0.002082849619910121, + 0.003613931592553854, + 0.004761102609336376, + 0.0029061397071927786, + 0.002889418974518776, + 0.007718170993030071, + 0.0029907003045082092, + 0.00534858787432313, + 0.0053292205557227135, + 0.002419620519503951, + 0.0027652308344841003, + 0.0031575148459523916, + 0.001953009283170104, + 0.006001119036227465, + 0.002621726132929325, + 0.0040767258033156395, + 0.003011300228536129, + 0.0021373536437749863, + 0.0016394408885389566, + 0.0026186280883848667, + 0.0019511543214321136, + 0.002552845748141408, + 0.002269857795909047, + 0.0026286544743925333, + 0.0023863762617111206, + 0.01057704258710146, + 0.012371052987873554, + 0.0155410785228014, + 0.008298424072563648, + 0.025997573509812355, + 0.013352137058973312, + 0.013015229254961014, + 0.012124218046665192, + 0.00912386178970337, + 0.00785329844802618, + 0.020891206339001656, + 0.007166444323956966, + 0.009780904278159142, + 0.00687064602971077, + 0.008106671273708344, + 0.01102977804839611, + 0.016508981585502625, + 0.01045665517449379, + 0.01587641052901745, + 0.007497546263039112, + 0.023462045937776566, + 0.015529822558164597, + 0.013451483100652695, + 0.014522665180265903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.013277565129101276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005967300850898027 + ], + "opposite_logprob_tokens": [ + -5.130967140197754 + ], + "logprob_reward": -0.005967300850898027, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005967300850898027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18857450783252716, + 0.4342763423919678, + 0.022640906274318695, + 0.015817761421203613, + 0.04235018789768219, + 0.04925034940242767, + 0.03263609856367111, + 0.03505488485097885, + 0.0428253710269928, + 0.1365736573934555 + ], + "credit_map": [ + 0.01588202640414238, + 0.02763480693101883, + 0.02489161305129528, + 0.03853924572467804, + 0.02207918092608452, + 0.01534013170748949, + 0.026066236197948456, + 0.0181412510573864, + 0.04054728522896767, + 0.06047189235687256, + 0.042632926255464554, + 0.10642515867948532, + 0.04600437730550766, + 0.03172529861330986, + 0.06119101122021675, + 0.0452783964574337, + 0.002412102185189724, + 0.002611723029986024, + 0.002784585813060403, + 0.002883556531742215, + 0.0025348940398544073, + 0.0022051522973924875, + 0.0039016439113765955, + 0.003307249629870057, + 0.0017031580209732056, + 0.002624764572829008, + 0.0014853161992505193, + 0.0011827533598989248, + 0.003618542803451419, + 0.001342841307632625, + 0.0015353938797488809, + 0.0023249906953424215, + 0.003875028807669878, + 0.0050604878924787045, + 0.0042589944787323475, + 0.00305029540322721, + 0.0104951411485672, + 0.005832337774336338, + 0.004315415397286415, + 0.005462488625198603, + 0.004707156680524349, + 0.005239102989435196, + 0.0064872815273702145, + 0.005480923689901829, + 0.006101232022047043, + 0.005457561928778887, + 0.008539321832358837, + 0.007237768732011318, + 0.0041146790608763695, + 0.002970863599330187, + 0.005179433152079582, + 0.0033725411631166935, + 0.00414715800434351, + 0.004025069065392017, + 0.004965914413332939, + 0.0038604415021836758, + 0.0037070615217089653, + 0.004206969402730465, + 0.006709363777190447, + 0.003994350787252188, + 0.002905855420976877, + 0.003524772124364972, + 0.004636180587112904, + 0.005370328668504953, + 0.00555350910872221, + 0.0036898814141750336, + 0.010739261284470558, + 0.005101305432617664, + 0.004633340518921614, + 0.0038204623851925135, + 0.003968144301325083, + 0.005319467280060053, + 0.018716050311923027, + 0.013099941425025463, + 0.017937298864126205, + 0.009190472774207592, + 0.02504769153892994, + 0.017321359366178513, + 0.018030403181910515, + 0.017230436205863953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.005967300850898027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899902760982513 + ], + "opposite_logprob_tokens": [ + -2.5789990425109863 + ], + "logprob_reward": -0.07899902760982513, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899902760982513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30448660254478455, + 0.24928848445415497, + 0.010449984110891819, + 0.03125176578760147, + 0.04682183265686035, + 0.026450561359524727, + 0.0220338124781847, + 0.17440566420555115, + 0.10120145231485367, + 0.033609870821237564 + ], + "credit_map": [ + 0.02906075306236744, + 0.034599438309669495, + 0.055670566856861115, + 0.05624304339289665, + 0.03272050991654396, + 0.02639589086174965, + 0.04044184461236, + 0.029354551807045937, + 0.02203129231929779, + 0.03356381505727768, + 0.024134041741490364, + 0.06209664046764374, + 0.02682538516819477, + 0.017969850450754166, + 0.036457642912864685, + 0.02620982751250267, + 0.0010726111941039562, + 0.0020808863919228315, + 0.0015214025042951107, + 0.0009948508813977242, + 0.0011260490864515305, + 0.0011244025081396103, + 0.0010073529556393623, + 0.0015224290546029806, + 0.0027552819810807705, + 0.004132927395403385, + 0.003538890043273568, + 0.00268571893684566, + 0.007195177022367716, + 0.002859534462913871, + 0.002937272423878312, + 0.005146963521838188, + 0.0038697707932442427, + 0.006031588651239872, + 0.004415104165673256, + 0.0031672760378569365, + 0.011776059865951538, + 0.0046899341978132725, + 0.0066950973123312, + 0.006177000235766172, + 0.002166860969737172, + 0.002758729737251997, + 0.004237843677401543, + 0.0024521206505596638, + 0.0035493914037942886, + 0.0029623473528772593, + 0.004820306785404682, + 0.0035029612481594086, + 0.0018760350067168474, + 0.0022074037697166204, + 0.003755879821255803, + 0.0025885843206197023, + 0.0022914940491318703, + 0.0024262911174446344, + 0.003646333934739232, + 0.0032417913898825645, + 0.017156274989247322, + 0.020560001954436302, + 0.021732715889811516, + 0.01303166151046753, + 0.04227149859070778, + 0.022673578932881355, + 0.018077557906508446, + 0.0189023707062006, + 0.008345224894583225, + 0.016163544729351997, + 0.022083669900894165, + 0.00897449441254139, + 0.01174948737025261, + 0.008878711611032486, + 0.012206653133034706, + 0.012799665331840515, + 0.0033488457556813955, + 0.004463626071810722, + 0.0035980548709630966, + 0.0026414450258016586, + 0.007867689244449139, + 0.0050455485470592976, + 0.003110474906861782, + 0.0035341866314411163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.07899902760982513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789627879858017 + ], + "opposite_logprob_tokens": [ + -2.578962802886963 + ], + "logprob_reward": -0.0789627879858017, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789627879858017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10219034552574158, + 0.5039437413215637, + 0.05136983469128609, + 0.012960975989699364, + 0.06142445281147957, + 0.03236708045005798, + 0.04570521041750908, + 0.05683058127760887, + 0.026847751811146736, + 0.10636003315448761 + ], + "credit_map": [ + 0.009255900979042053, + 0.01670386828482151, + 0.010831868276000023, + 0.017662491649389267, + 0.015224510803818703, + 0.008915219455957413, + 0.014193016104400158, + 0.00940346997231245, + 0.047603070735931396, + 0.07528150826692581, + 0.049002837389707565, + 0.12573687732219696, + 0.04992321878671646, + 0.03712601959705353, + 0.06867565959692001, + 0.050594523549079895, + 0.004963334184139967, + 0.0056336079724133015, + 0.006119033321738243, + 0.007170645985752344, + 0.005896085873246193, + 0.0049599576741456985, + 0.009552925825119019, + 0.007074244320392609, + 0.001307376311160624, + 0.0019020078470930457, + 0.0018195835873484612, + 0.0010211174376308918, + 0.0026698748115450144, + 0.001392503036186099, + 0.0012830691412091255, + 0.001565444516018033, + 0.0052508944645524025, + 0.006644799839705229, + 0.005257067736238241, + 0.004974461160600185, + 0.0152735011652112, + 0.008201239630579948, + 0.00842437893152237, + 0.007398109883069992, + 0.003159720916301012, + 0.0030194376595318317, + 0.005227462388575077, + 0.00334031879901886, + 0.003110631136223674, + 0.0034659148659557104, + 0.006116801872849464, + 0.004926793742924929, + 0.005870386492460966, + 0.004087059758603573, + 0.007126562763005495, + 0.004323145840317011, + 0.0059426818042993546, + 0.005965317599475384, + 0.006955086253583431, + 0.005434965714812279, + 0.00490910466760397, + 0.006374816410243511, + 0.008292936719954014, + 0.00502964248880744, + 0.0113171162083745, + 0.005720702465623617, + 0.00781635195016861, + 0.007369911298155785, + 0.0024926448240876198, + 0.002913743956014514, + 0.003635722678154707, + 0.003041571006178856, + 0.004165695048868656, + 0.003347074380144477, + 0.0028439948800951242, + 0.004407305270433426, + 0.012102887965738773, + 0.013118850067257881, + 0.011384999379515648, + 0.006758386269211769, + 0.02613009884953499, + 0.01602969504892826, + 0.009071526117622852, + 0.011763595044612885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0789627879858017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013381821336224675 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013381821336224675, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013381821336224675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20787875354290009, + 0.3099167048931122, + 0.007436740677803755, + 0.03499350696802139, + 0.037882108241319656, + 0.01324434019625187, + 0.027409149333834648, + 0.13641983270645142, + 0.1347646564245224, + 0.0900542214512825 + ], + "credit_map": [ + 0.01881580613553524, + 0.02730773389339447, + 0.030077651143074036, + 0.04134804382920265, + 0.025102082639932632, + 0.016757162287831306, + 0.02974190190434456, + 0.01872837543487549, + 0.028756193816661835, + 0.04421811178326607, + 0.030566245317459106, + 0.07995007187128067, + 0.03081321343779564, + 0.02126958966255188, + 0.04439219459891319, + 0.029951075091958046, + 0.0007416271837428212, + 0.0011665697675198317, + 0.000860680709592998, + 0.000774687621742487, + 0.000770142360124737, + 0.0009065442718565464, + 0.0011083978461101651, + 0.0011080908589065075, + 0.0031412553507834673, + 0.005063027609139681, + 0.0032756286673247814, + 0.0024650203995406628, + 0.009163211099803448, + 0.004633104428648949, + 0.003040623851120472, + 0.00421163672581315, + 0.0034075574949383736, + 0.004470129031687975, + 0.003067744430154562, + 0.0025069445837289095, + 0.009806392714381218, + 0.004793544299900532, + 0.005206134170293808, + 0.004623661283403635, + 0.0014044902054592967, + 0.001382483053021133, + 0.0018664897652342916, + 0.0010371434036642313, + 0.0025131977163255215, + 0.0017616814002394676, + 0.0015333316987380385, + 0.0017455219058319926, + 0.0030939998105168343, + 0.002939781406894326, + 0.004222049377858639, + 0.003156147664412856, + 0.0033931208308786154, + 0.003349750069901347, + 0.00418793223798275, + 0.003066368866711855, + 0.013619255274534225, + 0.014356338419020176, + 0.02051389589905739, + 0.00977881159633398, + 0.030684830620884895, + 0.015909386798739433, + 0.01628958433866501, + 0.015267733484506607, + 0.012961925938725471, + 0.013706800527870655, + 0.027940182015299797, + 0.0111668286845088, + 0.02448217198252678, + 0.012306739576160908, + 0.014344348572194576, + 0.017855651676654816, + 0.012215939350426197, + 0.008693436160683632, + 0.01193985529243946, + 0.00531443115323782, + 0.019749360159039497, + 0.011787060648202896, + 0.009513425640761852, + 0.010840710252523422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0013381821336224675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, rendered in the ethereal style of classical ink painting. A magnificent black blade hangs horizontally, mere inches above the surface of perfectly still water, floating serenely. Its body, crafted from a material resembling polished black obsidian, absorbs the soft ambient light, imparting a deep matte finish. Intricate, enigmatic carvings, like ancient, fluid figures, are meticulously sculpted along the length of the blade, seemingly harboring a dormant energy. Beneath the enchanted weapon, the water is an undisturbed mirror, reflecting a hazy, gray-white void. From a point directly below the blade's tip, soft ripples expand outward, less like disturbed water than like delicate ink spread across rice paper; these concentric circles move with a slow, hypnotic grace, their feathered edges dissolving into the tranquil surface. The entire composition is defined by a heavily desaturated color grading, creating a monochromatic world of black, white, and infinite shades of gray, enhancing the minimalist atmosphere and evoking a powerful sense of calm and ancient power.", + "seed": 23161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_5/index_170_seed_23161_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.253799169883132, + "vlm_logprob": -0.253799169883132, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.253799169883132, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9015424251556396 + ], + "opposite_logprob_tokens": [ + -0.5265424251556396 + ], + "logprob_reward": -0.9015424251556396, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9015424251556396, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.40834298729896545, + 0.15272308886051178, + 0.07083074003458023, + 0.03902903199195862, + 0.07162997871637344, + 0.08161169290542603, + 0.0783204734325409, + 0.021151643246412277, + 0.026668839156627655, + 0.04969155415892601 + ], + "credit_map": [ + 0.04154084995388985, + 0.05677919462323189, + 0.05245397239923477, + 0.07699038088321686, + 0.04400528594851494, + 0.054503247141838074, + 0.0349762849509716, + 0.047093771398067474, + 0.015825998038053513, + 0.01785041205585003, + 0.019009673967957497, + 0.02746211178600788, + 0.020962946116924286, + 0.01803135685622692, + 0.012207280844449997, + 0.021373318508267403, + 0.006262132897973061, + 0.0073303640820086, + 0.01087678037583828, + 0.007176635786890984, + 0.008059603162109852, + 0.0115285012871027, + 0.006888340227305889, + 0.012708382681012154, + 0.003637285903096199, + 0.004449453204870224, + 0.005191820673644543, + 0.004406746011227369, + 0.004719505086541176, + 0.008048414252698421, + 0.003809776157140732, + 0.0047660283744335175, + 0.008688578382134438, + 0.006008247844874859, + 0.012647567316889763, + 0.007326513063162565, + 0.006985025014728308, + 0.01030341349542141, + 0.007070711348205805, + 0.012599923647940159, + 0.007050142157822847, + 0.007568417116999626, + 0.009557093493640423, + 0.0069847493432462215, + 0.016505591571331024, + 0.015297690406441689, + 0.00958734005689621, + 0.009060664102435112, + 0.008343295194208622, + 0.008481206372380257, + 0.013579480350017548, + 0.010210311971604824, + 0.007914610207080841, + 0.012895234860479832, + 0.008355109952390194, + 0.008541222661733627, + 0.002384158316999674, + 0.0020083978306502104, + 0.0040577491745352745, + 0.002651236252859235, + 0.0017583213048055768, + 0.002608594484627247, + 0.002848773030564189, + 0.0028344132006168365, + 0.0028296131640672684, + 0.0041656200774014, + 0.004293894860893488, + 0.0024164693895727396, + 0.0034537729807198048, + 0.004454897250980139, + 0.0021298134233802557, + 0.0029247577767819166, + 0.0060108257457613945, + 0.004881179891526699, + 0.006576906889677048, + 0.004085110500454903, + 0.00710343848913908, + 0.008231637068092823, + 0.0041163465939462185, + 0.008686109445989132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.9015424251556396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805961512029171 + ], + "opposite_logprob_tokens": [ + -4.634806156158447 + ], + "logprob_reward": -0.009805961512029171, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805961512029171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27447569370269775, + 0.23906107246875763, + 0.05578698590397835, + 0.06978835165500641, + 0.14750242233276367, + 0.058130308985710144, + 0.032105330377817154, + 0.028326943516731262, + 0.04188868775963783, + 0.0529341995716095 + ], + "credit_map": [ + 0.02656388096511364, + 0.038755569607019424, + 0.028474317863583565, + 0.03757627308368683, + 0.029355740174651146, + 0.03949108347296715, + 0.03231323882937431, + 0.04194561392068863, + 0.033682141453027725, + 0.02345253899693489, + 0.036482296884059906, + 0.023869335651397705, + 0.028820695355534554, + 0.03209863603115082, + 0.023333419114351273, + 0.03732200711965561, + 0.006605052854865789, + 0.006860820110887289, + 0.0059311105869710445, + 0.0051073115319013596, + 0.008735638111829758, + 0.008199644275009632, + 0.006170252338051796, + 0.008177152834832668, + 0.009790540672838688, + 0.008605108596384525, + 0.011413168162107468, + 0.006486339960247278, + 0.005221921484917402, + 0.009077171795070171, + 0.005752945318818092, + 0.013441153801977634, + 0.01885572262108326, + 0.01635207049548626, + 0.022227175533771515, + 0.01635526865720749, + 0.015772193670272827, + 0.020008891820907593, + 0.012828275561332703, + 0.02510281279683113, + 0.007198963314294815, + 0.007044450379908085, + 0.008177709765732288, + 0.005762362387031317, + 0.00768988486379385, + 0.007634500972926617, + 0.005545325577259064, + 0.009077110327780247, + 0.0033725916873663664, + 0.0038024659734219313, + 0.005433469079434872, + 0.003409932367503643, + 0.003100026398897171, + 0.005850581917911768, + 0.0031431782990694046, + 0.00399308605119586, + 0.00349926738999784, + 0.0027510561048984528, + 0.005358840338885784, + 0.003939507529139519, + 0.003439784748479724, + 0.0032796021550893784, + 0.003339288057759404, + 0.0027195969596505165, + 0.004556532017886639, + 0.004468480125069618, + 0.00493648461997509, + 0.003300685668364167, + 0.008339562453329563, + 0.005562615115195513, + 0.005913198925554752, + 0.004811130929738283, + 0.004672313574701548, + 0.004449380561709404, + 0.009411237202584743, + 0.00496094673871994, + 0.009833251126110554, + 0.010454198345541954, + 0.0039508226327598095, + 0.005202051717787981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009805961512029171 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814027667045593 + ], + "opposite_logprob_tokens": [ + -1.406402826309204 + ], + "logprob_reward": -0.2814027667045593, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814027667045593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22498393058776855, + 0.14299385249614716, + 0.10700245201587677, + 0.06118376925587654, + 0.11087427288293839, + 0.05892102047801018, + 0.03751805052161217, + 0.1400851458311081, + 0.07431527972221375, + 0.042122162878513336 + ], + "credit_map": [ + 0.028983818367123604, + 0.024595975875854492, + 0.027531588450074196, + 0.025085510686039925, + 0.027032237499952316, + 0.04172358289361, + 0.019194161519408226, + 0.03083706647157669, + 0.014866450801491737, + 0.012743334285914898, + 0.01763792894780636, + 0.025460338220000267, + 0.014543889090418816, + 0.018217245116829872, + 0.0215748380869627, + 0.01794983074069023, + 0.016588371247053146, + 0.015177291817963123, + 0.011877982877194881, + 0.008218861185014248, + 0.006528337020426989, + 0.015093166381120682, + 0.014970149844884872, + 0.01854829490184784, + 0.008259975351393223, + 0.006780613679438829, + 0.0071181426756083965, + 0.005828073713928461, + 0.006649370770901442, + 0.009325693361461163, + 0.005431084427982569, + 0.01179081853479147, + 0.015769386664032936, + 0.01414749026298523, + 0.0141140753403306, + 0.010864523239433765, + 0.0106252646073699, + 0.014674022793769836, + 0.012873574160039425, + 0.017805946990847588, + 0.005241486709564924, + 0.006451796740293503, + 0.005693541374057531, + 0.005116088781505823, + 0.010911288671195507, + 0.010612738318741322, + 0.008253772743046284, + 0.006640308536589146, + 0.005962768103927374, + 0.004964702762663364, + 0.006421057973057032, + 0.004270295146852732, + 0.002466700505465269, + 0.0048368340358138084, + 0.0030573424883186817, + 0.005538350902497768, + 0.022085033357143402, + 0.02387617528438568, + 0.025203488767147064, + 0.013294775038957596, + 0.011983991600573063, + 0.01835949905216694, + 0.009939351119101048, + 0.015342828817665577, + 0.011194633319973946, + 0.011427409015595913, + 0.012435058131814003, + 0.005765281617641449, + 0.009203632362186909, + 0.010740774683654308, + 0.0049506742507219315, + 0.00859781913459301, + 0.004643314052373171, + 0.005761208478361368, + 0.006706899497658014, + 0.003914287779480219, + 0.005915496032685041, + 0.006235658191144466, + 0.0035741752944886684, + 0.005371122155338526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.2814027667045593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037643045186996 + ], + "opposite_logprob_tokens": [ + -2.812037706375122 + ], + "logprob_reward": -0.062037643045186996, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037643045186996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4363057315349579, + 0.16810429096221924, + 0.09363511949777603, + 0.067112497985363, + 0.05006783828139305, + 0.028994273394346237, + 0.03394043445587158, + 0.0336720310151577, + 0.047723449766635895, + 0.0404442735016346 + ], + "credit_map": [ + 0.038832008838653564, + 0.04765963554382324, + 0.06536445021629333, + 0.08812850713729858, + 0.035572148859500885, + 0.04088817909359932, + 0.05825214087963104, + 0.06160866096615791, + 0.020117443054914474, + 0.01647905260324478, + 0.03381720557808876, + 0.02158159203827381, + 0.021629564464092255, + 0.01912846975028515, + 0.01488076988607645, + 0.02047019824385643, + 0.00915275327861309, + 0.010664951987564564, + 0.01109082531183958, + 0.01048677135258913, + 0.011523159220814705, + 0.014211772009730339, + 0.013191797770559788, + 0.013313088566064835, + 0.005582586396485567, + 0.006320280488580465, + 0.009198451414704323, + 0.007469176780432463, + 0.005177475977689028, + 0.012442531995475292, + 0.011277609504759312, + 0.009644392877817154, + 0.008232994936406612, + 0.0054121799767017365, + 0.006590197328478098, + 0.005326933693140745, + 0.0047228774055838585, + 0.005046976264566183, + 0.005447631701827049, + 0.009288047440350056, + 0.004864443093538284, + 0.0041200160048902035, + 0.003706625197082758, + 0.0022068044636398554, + 0.0025942560750991106, + 0.0046126400120556355, + 0.0033093297388404608, + 0.0035801585763692856, + 0.004467733204364777, + 0.003924362827092409, + 0.004277227912098169, + 0.003964265808463097, + 0.003920910879969597, + 0.005326438695192337, + 0.0028785718604922295, + 0.005180923268198967, + 0.0024657498579472303, + 0.0038125114515423775, + 0.006297350395470858, + 0.002949718153104186, + 0.003888368373736739, + 0.006304963491857052, + 0.00401232810690999, + 0.00394104141741991, + 0.004884012974798679, + 0.006451356690376997, + 0.007721206173300743, + 0.00419228058308363, + 0.004079937003552914, + 0.006687290966510773, + 0.007111798971891403, + 0.0065955668687820435, + 0.0036997392307966948, + 0.004131006076931953, + 0.006776467896997929, + 0.0036580231972038746, + 0.004886026959866285, + 0.008887797594070435, + 0.0038119948003441095, + 0.004593214951455593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.062037643045186996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207052998244762 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014207052998244762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207052998244762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37518203258514404, + 0.09887593984603882, + 0.06827802956104279, + 0.063265360891819, + 0.13822293281555176, + 0.07028192281723022, + 0.03514312207698822, + 0.03222249448299408, + 0.04349645972251892, + 0.07503169029951096 + ], + "credit_map": [ + 0.04085186868906021, + 0.05774741247296333, + 0.054268330335617065, + 0.051195938140153885, + 0.037629593163728714, + 0.05014736205339432, + 0.046161141246557236, + 0.03718039393424988, + 0.0098026217892766, + 0.010874010622501373, + 0.02091689594089985, + 0.013243340887129307, + 0.01151796244084835, + 0.011190002784132957, + 0.011498848907649517, + 0.009832256473600864, + 0.005828726105391979, + 0.007350222673267126, + 0.007913704961538315, + 0.008075566962361336, + 0.008294672705233097, + 0.007262625731527805, + 0.014870138838887215, + 0.008682379499077797, + 0.009829021990299225, + 0.010553935542702675, + 0.007780818268656731, + 0.004851466976106167, + 0.004786078818142414, + 0.009117812849581242, + 0.007680372800678015, + 0.008665851317346096, + 0.018001176416873932, + 0.012869940139353275, + 0.02045987918972969, + 0.01734079420566559, + 0.012696282006800175, + 0.012015837244689465, + 0.0210020262748003, + 0.023836996406316757, + 0.007239159662276506, + 0.00749708479270339, + 0.00847455020993948, + 0.006117916200309992, + 0.012808091938495636, + 0.011594796553254128, + 0.008442163467407227, + 0.008108154870569706, + 0.004824306350201368, + 0.004193054977804422, + 0.007172940298914909, + 0.004608471877872944, + 0.002822805428877473, + 0.004091702401638031, + 0.003467611502856016, + 0.003962229005992413, + 0.0034818542189896107, + 0.0033736033365130424, + 0.006117609795182943, + 0.0034013716503977776, + 0.0029100594110786915, + 0.004338275641202927, + 0.004588329698890448, + 0.004011392127722502, + 0.004383152350783348, + 0.004608448129147291, + 0.007227362599223852, + 0.002778711263090372, + 0.0087778614833951, + 0.006815359927713871, + 0.004188254941254854, + 0.00471730949357152, + 0.006778533570468426, + 0.009074934758245945, + 0.013236410915851593, + 0.005877795163542032, + 0.01128312572836876, + 0.01558582205325365, + 0.005316521506756544, + 0.007878548465669155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.014207052998244762 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, rendered in the ethereal style of classical ink painting. A magnificent black blade hangs horizontally, mere inches above the surface of perfectly still water, floating serenely. Its body, crafted from a material resembling polished black obsidian, absorbs the soft ambient light, imparting a deep matte finish. Intricate, enigmatic carvings, like ancient, fluid figures, are meticulously sculpted along the length of the blade, seemingly harboring a dormant energy. Beneath the enchanted weapon, the water is an undisturbed mirror, reflecting a hazy, gray-white void. From a point directly below the blade's tip, soft ripples expand outward, less like disturbed water than like delicate ink spread across rice paper; these concentric circles move with a slow, hypnotic grace, their feathered edges dissolving into the tranquil surface. The entire composition is defined by a heavily desaturated color grading, creating a monochromatic world of black, white, and infinite shades of gray, enhancing the minimalist atmosphere and evoking a powerful sense of calm and ancient power.", + "seed": 23162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_5/index_170_seed_23162_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09088902440853416, + "vlm_logprob": -0.09088902440853416, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.09088902440853416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18324851989746094 + ], + "opposite_logprob_tokens": [ + -1.808248519897461 + ], + "logprob_reward": -0.18324851989746094, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18324851989746094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38295355439186096, + 0.13978414237499237, + 0.04974597319960594, + 0.05132266879081726, + 0.10229504108428955, + 0.06564868986606598, + 0.07823797315359116, + 0.04665929079055786, + 0.051610250025987625, + 0.03174237161874771 + ], + "credit_map": [ + 0.05701416730880737, + 0.029514197260141373, + 0.03898794576525688, + 0.06138265132904053, + 0.03199101611971855, + 0.03702573850750923, + 0.06599605828523636, + 0.06104178726673126, + 0.021526122465729713, + 0.012899402529001236, + 0.01860959641635418, + 0.01350477896630764, + 0.011359907686710358, + 0.018117349594831467, + 0.015350640751421452, + 0.0284163486212492, + 0.0067777871154248714, + 0.004536327440291643, + 0.004952537827193737, + 0.006994451396167278, + 0.005355274770408869, + 0.0065944562666118145, + 0.009375069290399551, + 0.005160072818398476, + 0.007895185612142086, + 0.004204317461699247, + 0.005811687093228102, + 0.0066979676485061646, + 0.0036497851833701134, + 0.009582802653312683, + 0.007257807068526745, + 0.00622312119230628, + 0.01160074770450592, + 0.0142403868958354, + 0.009021345525979996, + 0.021731805056333542, + 0.011164148338139057, + 0.009185098111629486, + 0.01681550033390522, + 0.0085360212251544, + 0.006940755527466536, + 0.008608272299170494, + 0.006913325749337673, + 0.014527411200106144, + 0.0062759616412222385, + 0.006671634968370199, + 0.010154151357710361, + 0.005557185038924217, + 0.00959352683275938, + 0.010941627435386181, + 0.005577669478952885, + 0.012783593498170376, + 0.01002596691250801, + 0.010681498795747757, + 0.012181873433291912, + 0.006452226545661688, + 0.005644707009196281, + 0.005693904124200344, + 0.003317024791613221, + 0.007984454743564129, + 0.005737169645726681, + 0.005328609608113766, + 0.009066218510270119, + 0.0038872077129781246, + 0.006131603382527828, + 0.004090593196451664, + 0.005140420515090227, + 0.009447000920772552, + 0.00462848087772727, + 0.0065106903202831745, + 0.010709719732403755, + 0.004951746203005314, + 0.0034214910119771957, + 0.003776397556066513, + 0.002472124993801117, + 0.005292903166264296, + 0.003702065907418728, + 0.004001124761998653, + 0.0063284290954470634, + 0.0027478367555886507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.18324851989746094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009817766025662422 + ], + "opposite_logprob_tokens": [ + -4.634817600250244 + ], + "logprob_reward": -0.009817766025662422, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009817766025662422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08067911118268967, + 0.12049499154090881, + 0.01399935781955719, + 0.011438154615461826, + 0.05194372683763504, + 0.5079091787338257, + 0.07323392480611801, + 0.060192640870809555, + 0.02201746217906475, + 0.058091409504413605 + ], + "credit_map": [ + 0.013423854485154152, + 0.005541867110878229, + 0.010964118875563145, + 0.009707445278763771, + 0.005146583076566458, + 0.01086672767996788, + 0.009104346856474876, + 0.01592416875064373, + 0.020536959171295166, + 0.008608244359493256, + 0.014112651348114014, + 0.012182420119643211, + 0.009275603108108044, + 0.018649056553840637, + 0.014938068576157093, + 0.022191984578967094, + 0.001775631564669311, + 0.0016492659924551845, + 0.001053188811056316, + 0.0024090404622256756, + 0.0015561110340058804, + 0.0022049807012081146, + 0.002194506349042058, + 0.0011566318571567535, + 0.0014839026844128966, + 0.0012742943363264203, + 0.001234014518558979, + 0.0016162064857780933, + 0.0013012152630835772, + 0.0011884340783581138, + 0.002142228651791811, + 0.0011978588299825788, + 0.005944509524852037, + 0.007705386728048325, + 0.004908132366836071, + 0.0093689551576972, + 0.006962983403354883, + 0.0052299234084784985, + 0.007863936945796013, + 0.0039599002338945866, + 0.06298114359378815, + 0.07908814400434494, + 0.041734516620635986, + 0.08046351373195648, + 0.05923875793814659, + 0.0561196394264698, + 0.08560006320476532, + 0.042683426290750504, + 0.010100021958351135, + 0.0069190883077681065, + 0.007062042597681284, + 0.009469440206885338, + 0.004375855904072523, + 0.011536024510860443, + 0.015362920239567757, + 0.008408532477915287, + 0.0065520526841282845, + 0.00736256456002593, + 0.0059141600504517555, + 0.008776066824793816, + 0.007300495635718107, + 0.007246918044984341, + 0.01239254791289568, + 0.004647836089134216, + 0.002861438551917672, + 0.0025941571220755577, + 0.0021844443399459124, + 0.0033180834725499153, + 0.002252805046737194, + 0.002606293186545372, + 0.003649233840405941, + 0.0025510070845484734, + 0.008569536730647087, + 0.006339273881167173, + 0.007070292253047228, + 0.007441313937306404, + 0.0058109937235713005, + 0.006246886681765318, + 0.008872944861650467, + 0.0077401697635650635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.009817766025662422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991873621940613 + ], + "opposite_logprob_tokens": [ + -1.8049187660217285 + ], + "logprob_reward": -0.17991873621940613, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991873621940613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11538449674844742, + 0.10061932355165482, + 0.21512873470783234, + 0.03672178089618683, + 0.09032288938760757, + 0.36355942487716675, + 0.03219820559024811, + 0.023047959432005882, + 0.00974913127720356, + 0.013268033973872662 + ], + "credit_map": [ + 0.013013151474297047, + 0.009798085317015648, + 0.014438644982874393, + 0.019931267946958542, + 0.007716358173638582, + 0.014312042854726315, + 0.017573675140738487, + 0.018601270392537117, + 0.01908818446099758, + 0.0069435834884643555, + 0.012926685623824596, + 0.010240046307444572, + 0.00619141710922122, + 0.014552242122590542, + 0.00988497119396925, + 0.020792189985513687, + 0.040204163640737534, + 0.014860686846077442, + 0.0324699766933918, + 0.015089176595211029, + 0.01304295938462019, + 0.0303693525493145, + 0.023431535810232162, + 0.04566088318824768, + 0.004796348977833986, + 0.004925543908029795, + 0.004984826315194368, + 0.005111439619213343, + 0.00438982667401433, + 0.003642687574028969, + 0.005752921104431152, + 0.003118187189102173, + 0.010957496240735054, + 0.012494231574237347, + 0.00874185748398304, + 0.01521625928580761, + 0.010506004095077515, + 0.008748482912778854, + 0.015826022252440453, + 0.007832538336515427, + 0.045553211122751236, + 0.056116148829460144, + 0.030140602961182594, + 0.05630671605467796, + 0.042290978133678436, + 0.03803720325231552, + 0.062355637550354004, + 0.03275894746184349, + 0.003598520066589117, + 0.002022158121690154, + 0.002891122130677104, + 0.006188045255839825, + 0.0016232415800914168, + 0.0032732102554291487, + 0.010193298570811749, + 0.0024086101911962032, + 0.00276388693600893, + 0.0025213274639099836, + 0.0022988265845924616, + 0.0039720721542835236, + 0.002872639801353216, + 0.002327422611415386, + 0.004519841633737087, + 0.0017719421302899718, + 0.0010670720366761088, + 0.0007039690390229225, + 0.0012739410158246756, + 0.001551503548398614, + 0.0008427169523201883, + 0.0010829948587343097, + 0.002467212500050664, + 0.0007597210351377726, + 0.00163315050303936, + 0.0015597608871757984, + 0.0012974938144907355, + 0.001965832896530628, + 0.0016161395469680429, + 0.001581814605742693, + 0.0023588507901877165, + 0.0012549909297376871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.17991873621940613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896696776151657 + ], + "opposite_logprob_tokens": [ + -2.5789668560028076 + ], + "logprob_reward": -0.07896696776151657, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896696776151657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05847741663455963, + 0.35250887274742126, + 0.02464459091424942, + 0.011127643287181854, + 0.09220641851425171, + 0.35602614283561707, + 0.02798612415790558, + 0.031208952888846397, + 0.012652888894081116, + 0.03316091001033783 + ], + "credit_map": [ + 0.007947278209030628, + 0.00515862088650465, + 0.008051455020904541, + 0.004918838385492563, + 0.004725488368421793, + 0.009406279772520065, + 0.00680563086643815, + 0.011463821865618229, + 0.058978039771318436, + 0.027745965868234634, + 0.05011790618300438, + 0.030110063031315804, + 0.02751893363893032, + 0.04647703468799591, + 0.03778178617358208, + 0.0737791582942009, + 0.004516961984336376, + 0.002286267466843128, + 0.0027346392162144184, + 0.002217429457232356, + 0.002145912032574415, + 0.0034763282164931297, + 0.002846979536116123, + 0.004420071840286255, + 0.0012330318568274379, + 0.0009345654398202896, + 0.0011475029168650508, + 0.0022767295595258474, + 0.0011458988301455975, + 0.001411198521964252, + 0.002043906133621931, + 0.0009348098537884653, + 0.01104485522955656, + 0.013203151524066925, + 0.00938362069427967, + 0.015170110389590263, + 0.010167714208364487, + 0.009530787356197834, + 0.016720402985811234, + 0.006985775660723448, + 0.04353594034910202, + 0.052353568375110626, + 0.03140970692038536, + 0.0602676197886467, + 0.04140708968043327, + 0.04011962190270424, + 0.060210373252630234, + 0.02672223187983036, + 0.002741036470979452, + 0.0025303370784968138, + 0.002842481015250087, + 0.005741673056036234, + 0.002111358568072319, + 0.0025961410719901323, + 0.0060502407141029835, + 0.0033728585112839937, + 0.0037861608434468508, + 0.004622930195182562, + 0.0027053633239120245, + 0.004179141018539667, + 0.0045623485930264, + 0.00355412601493299, + 0.005225739441812038, + 0.0025731446221470833, + 0.0014543002471327782, + 0.001482554362155497, + 0.0010976685443893075, + 0.0017303851200267673, + 0.0017058037919923663, + 0.002059469697996974, + 0.002169632352888584, + 0.0009530750103294849, + 0.004533741623163223, + 0.004337391350418329, + 0.0033088927157223225, + 0.004384277388453484, + 0.004451422486454248, + 0.003691962221637368, + 0.004590271040797234, + 0.003862951882183552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.07896696776151657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024931321386247873 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.0024931321386247873, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024931321386247873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16072118282318115, + 0.06509685516357422, + 0.19556787610054016, + 0.03578340262174606, + 0.05193840339779854, + 0.302120566368103, + 0.07533439248800278, + 0.05557025969028473, + 0.02560555934906006, + 0.032261498272418976 + ], + "credit_map": [ + 0.015924004837870598, + 0.013159610331058502, + 0.019061490893363953, + 0.0237490963190794, + 0.014404603280127048, + 0.016978392377495766, + 0.03073645569384098, + 0.026707526296377182, + 0.007558778859674931, + 0.00608720863237977, + 0.00487189507111907, + 0.01117631234228611, + 0.007802796084433794, + 0.007383857388049364, + 0.014623796567320824, + 0.0055922106839716434, + 0.033103205263614655, + 0.015702703967690468, + 0.029222767800092697, + 0.014805513434112072, + 0.014041892252862453, + 0.02483833022415638, + 0.023475131019949913, + 0.04037833586335182, + 0.0043007382191717625, + 0.00343321799300611, + 0.004646002314984798, + 0.005182225722819567, + 0.0033440752886235714, + 0.004129673354327679, + 0.00663157319650054, + 0.0041158972308039665, + 0.005379390902817249, + 0.006253486033529043, + 0.006107182241976261, + 0.011051916517317295, + 0.005186660680919886, + 0.004759419709444046, + 0.007716063875705004, + 0.005484282970428467, + 0.036886733025312424, + 0.04502985253930092, + 0.02587302401661873, + 0.04936734586954117, + 0.03442849963903427, + 0.03446623682975769, + 0.051992859691381454, + 0.024076024070382118, + 0.008125943131744862, + 0.008661461062729359, + 0.006207556929439306, + 0.015577496029436588, + 0.008288154378533363, + 0.008640620857477188, + 0.013963748700916767, + 0.005869410932064056, + 0.006086207926273346, + 0.006764599587768316, + 0.004383950959891081, + 0.00985381193459034, + 0.007072594482451677, + 0.005349819548428059, + 0.011471236124634743, + 0.0045880405232310295, + 0.0031332005746662617, + 0.0020297435112297535, + 0.003277361858636141, + 0.004118766635656357, + 0.0021701764781028032, + 0.003059192094951868, + 0.005448764190077782, + 0.0023683521430939436, + 0.004827653989195824, + 0.0033810969907790422, + 0.0030434932559728622, + 0.00423970865085721, + 0.0033703125081956387, + 0.004371721297502518, + 0.004670324735343456, + 0.004357185680419207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.0024931321386247873 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, rendered in the ethereal style of classical ink painting. A magnificent black blade hangs horizontally, mere inches above the surface of perfectly still water, floating serenely. Its body, crafted from a material resembling polished black obsidian, absorbs the soft ambient light, imparting a deep matte finish. Intricate, enigmatic carvings, like ancient, fluid figures, are meticulously sculpted along the length of the blade, seemingly harboring a dormant energy. Beneath the enchanted weapon, the water is an undisturbed mirror, reflecting a hazy, gray-white void. From a point directly below the blade's tip, soft ripples expand outward, less like disturbed water than like delicate ink spread across rice paper; these concentric circles move with a slow, hypnotic grace, their feathered edges dissolving into the tranquil surface. The entire composition is defined by a heavily desaturated color grading, creating a monochromatic world of black, white, and infinite shades of gray, enhancing the minimalist atmosphere and evoking a powerful sense of calm and ancient power.", + "seed": 23163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_5/index_170_seed_23163_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10913276588544249, + "vlm_logprob": -0.10913276588544249, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.10913276588544249, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1159655898809433 + ], + "opposite_logprob_tokens": [ + -2.2409656047821045 + ], + "logprob_reward": -0.1159655898809433, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1159655898809433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16819821298122406, + 0.29349082708358765, + 0.3104851245880127, + 0.07206495106220245, + 0.008423307910561562, + 0.08533024042844772, + 0.012450918555259705, + 0.026736758649349213, + 0.01189021859318018, + 0.010929428040981293 + ], + "credit_map": [ + 0.04367811977863312, + 0.016267623752355576, + 0.052670009434223175, + 0.012315358966588974, + 0.007763104513287544, + 0.015366368927061558, + 0.006508690305054188, + 0.013628954067826271, + 0.10320445150136948, + 0.013261952437460423, + 0.12493466585874557, + 0.007828141562640667, + 0.006714733317494392, + 0.018121497705578804, + 0.0035664676688611507, + 0.01585891842842102, + 0.10969717055559158, + 0.014638436026871204, + 0.13136768341064453, + 0.008645741268992424, + 0.007144675124436617, + 0.019640501588582993, + 0.0036046106833964586, + 0.01574632339179516, + 0.017195751890540123, + 0.008856936357915401, + 0.01890629529953003, + 0.007284428924322128, + 0.004489189479500055, + 0.005331467837095261, + 0.0046416763216257095, + 0.005359208676964045, + 0.001700390363112092, + 0.0007166286231949925, + 0.002537073800340295, + 0.0006624356610700488, + 0.0004138318181503564, + 0.0008733814465813339, + 0.0006925942725501955, + 0.0008269728859886527, + 0.023582415655255318, + 0.0038773545529693365, + 0.029184363782405853, + 0.004596831742674112, + 0.003605297300964594, + 0.008829801343381405, + 0.005446968134492636, + 0.006207215134054422, + 0.0030684703961014748, + 0.0008089418988674879, + 0.00425631133839488, + 0.0006870118086226285, + 0.0004962493549101055, + 0.0011031939648091793, + 0.0010416028089821339, + 0.0009891384979709983, + 0.0055414107628166676, + 0.0022633718326687813, + 0.006541891023516655, + 0.002696462208405137, + 0.0014946539886295795, + 0.0028668707236647606, + 0.002289269119501114, + 0.0030428320169448853, + 0.00316648930311203, + 0.0005548974731937051, + 0.004278568085283041, + 0.0007009427645243704, + 0.0005199944716878235, + 0.0009448366472497582, + 0.0008527812315151095, + 0.00087171042105183, + 0.002117943949997425, + 0.0008010949823074043, + 0.0033634731080383062, + 0.000841652334202081, + 0.0004797181172762066, + 0.0010677282698452473, + 0.001212838920764625, + 0.0010449797846376896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.1159655898809433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023334426805377007 + ], + "opposite_logprob_tokens": [ + -3.773334503173828 + ], + "logprob_reward": -0.023334426805377007, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023334426805377007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05670144036412239, + 0.42700493335723877, + 0.2820947766304016, + 0.05705370008945465, + 0.0066489726305007935, + 0.02730396017432213, + 0.005967579782009125, + 0.0991123616695404, + 0.02891162782907486, + 0.009200588800013065 + ], + "credit_map": [ + 0.015074980445206165, + 0.003205486573278904, + 0.017848122864961624, + 0.002932690316811204, + 0.002861010143533349, + 0.006907348055392504, + 0.0026418245397508144, + 0.005229982081800699, + 0.15075309574604034, + 0.019542554393410683, + 0.1810528039932251, + 0.011553855612874031, + 0.00993253756314516, + 0.026850316673517227, + 0.005277187097817659, + 0.022042641416192055, + 0.09909746050834656, + 0.01323965098708868, + 0.11961890012025833, + 0.007752415258437395, + 0.00646245526149869, + 0.01748817227780819, + 0.003125010523945093, + 0.015310736373066902, + 0.01711312308907509, + 0.0032718742731958628, + 0.01980915293097496, + 0.002945518121123314, + 0.002256853273138404, + 0.005472575780004263, + 0.00222479784861207, + 0.00395980617031455, + 0.0014787479303777218, + 0.0003856563416775316, + 0.0018467929912731051, + 0.0005203181644901633, + 0.0003102217160630971, + 0.0006119541940279305, + 0.0008470021421089768, + 0.000648279266897589, + 0.006011814344674349, + 0.0024540191516280174, + 0.006879354827105999, + 0.002566496841609478, + 0.0015679456992074847, + 0.0028118209447711706, + 0.002659372752532363, + 0.0023531378246843815, + 0.0012103591579943895, + 0.00034925900399684906, + 0.0016024552751332521, + 0.000505574163980782, + 0.0002899109967984259, + 0.0007207178277894855, + 0.000847411691211164, + 0.0004418924800120294, + 0.030926711857318878, + 0.005222286097705364, + 0.03831837698817253, + 0.005733475089073181, + 0.0028032129630446434, + 0.005700408015400171, + 0.003440468106418848, + 0.006967431865632534, + 0.008987413719296455, + 0.0012873928062617779, + 0.010604695416986942, + 0.0019200164824724197, + 0.0011158800916746259, + 0.0017944665160030127, + 0.0015915888361632824, + 0.0016101765213534236, + 0.0026735919527709484, + 0.0005018584197387099, + 0.0032019310165196657, + 0.0005499829421751201, + 0.00034642504760995507, + 0.0006864821189083159, + 0.0007148373406380415, + 0.0005254801944829524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.023334426805377007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161135494709015 + ], + "opposite_logprob_tokens": [ + -1.7016113996505737 + ], + "logprob_reward": -0.20161135494709015, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161135494709015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08280223608016968, + 0.08317850530147552, + 0.2320026308298111, + 0.12542706727981567, + 0.011839731596410275, + 0.36203399300575256, + 0.003802067134529352, + 0.034713298082351685, + 0.05027780309319496, + 0.013922562822699547 + ], + "credit_map": [ + 0.022407516837120056, + 0.004401709884405136, + 0.02751067467033863, + 0.0034119572956115007, + 0.003944354597479105, + 0.010036848485469818, + 0.003523464547470212, + 0.007565714418888092, + 0.028355013579130173, + 0.003997634164988995, + 0.03434567525982857, + 0.0024908946361392736, + 0.0021111282985657454, + 0.005121216177940369, + 0.0013771983794867992, + 0.005379748065024614, + 0.0811644047498703, + 0.010868160054087639, + 0.09850761294364929, + 0.006334969308227301, + 0.005306418985128403, + 0.014242320321500301, + 0.0026819133199751377, + 0.01289683673530817, + 0.0362280011177063, + 0.010043147020041943, + 0.04189269244670868, + 0.007478197105228901, + 0.005696152802556753, + 0.010635149665176868, + 0.00465996190905571, + 0.008793771266937256, + 0.0022120673675090075, + 0.0008591058431193233, + 0.0038558596279472113, + 0.0007341104792430997, + 0.0006992334383539855, + 0.0012617867905646563, + 0.0008903417619876564, + 0.001327226054854691, + 0.10085124522447586, + 0.01618659310042858, + 0.12735164165496826, + 0.018065601587295532, + 0.011593579314649105, + 0.028955992311239243, + 0.023534046486020088, + 0.03549530357122421, + 0.0005931529449298978, + 0.00025703830760903656, + 0.0009714183979667723, + 0.00033133916440419853, + 0.00017092304187826812, + 0.0005086626042611897, + 0.000543937727343291, + 0.0004255951498635113, + 0.007155608385801315, + 0.0026946784928441048, + 0.00869840383529663, + 0.0031510163098573685, + 0.0018246896797791123, + 0.0038066150154918432, + 0.003006601007655263, + 0.004375683143734932, + 0.01535819936543703, + 0.0022797510027885437, + 0.018908483907580376, + 0.0025045829825103283, + 0.0015911529771983624, + 0.003558351192623377, + 0.0024948245845735073, + 0.003582457546144724, + 0.0020985822193324566, + 0.0008742454810999334, + 0.0030024643056094646, + 0.001129522337578237, + 0.00096161849796772, + 0.0018757448997348547, + 0.0024386467412114143, + 0.0015417388640344143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.20161135494709015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155620574951172 + ], + "opposite_logprob_tokens": [ + -1.7015562057495117 + ], + "logprob_reward": -0.20155620574951172, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155620574951172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10893940180540085, + 0.2513108253479004, + 0.04227009043097496, + 0.22691062092781067, + 0.0159998070448637, + 0.060085974633693695, + 0.011793427169322968, + 0.07935383170843124, + 0.1845259666442871, + 0.01881003938615322 + ], + "credit_map": [ + 0.018419597297906876, + 0.0068079326301813126, + 0.02459820732474327, + 0.0067327977158129215, + 0.006924942135810852, + 0.017849169671535492, + 0.008744661696255207, + 0.01886208914220333, + 0.08911990374326706, + 0.011955625377595425, + 0.1056034192442894, + 0.0068002245388925076, + 0.005706151016056538, + 0.015387802384793758, + 0.0036743171513080597, + 0.013063399121165276, + 0.013437706045806408, + 0.0021741853561252356, + 0.01666582003235817, + 0.0014255056157708168, + 0.0011998594272881746, + 0.0030033313669264317, + 0.001179271494038403, + 0.0031844123732298613, + 0.05848364904522896, + 0.024466292932629585, + 0.06544580310583115, + 0.019258996471762657, + 0.012447431683540344, + 0.017866820096969604, + 0.013773203827440739, + 0.015168421901762486, + 0.0038253781385719776, + 0.0010382053442299366, + 0.005420607049018145, + 0.000955958035774529, + 0.0006362603162415326, + 0.0015610981499776244, + 0.0012066385243088007, + 0.0013556588673964143, + 0.010481628589332104, + 0.0030438487883657217, + 0.012663316912949085, + 0.005270335357636213, + 0.00343835586681962, + 0.009080412797629833, + 0.008039663545787334, + 0.008068408817052841, + 0.0012078935978934169, + 0.0008985979366116226, + 0.0018716698978096247, + 0.0012401491403579712, + 0.0007922530057840049, + 0.0021156445145606995, + 0.001956978114321828, + 0.0017102400306612253, + 0.01695631816983223, + 0.008133034221827984, + 0.020289672538638115, + 0.007677921559661627, + 0.004258125554770231, + 0.007216134574264288, + 0.006305566988885403, + 0.00851705763489008, + 0.05832432955503464, + 0.008149810135364532, + 0.07431405037641525, + 0.007680778857320547, + 0.00449624378234148, + 0.012830546125769615, + 0.006944307591766119, + 0.01178589928895235, + 0.0016883991193026304, + 0.0008927615126594901, + 0.002484927885234356, + 0.003115639090538025, + 0.001529511995613575, + 0.0028255057986825705, + 0.004243058152496815, + 0.0020302371121942997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.20155620574951172 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of classical ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black blade hanging horizontally above a surface of still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the blade have a matte finish with intricate carvings along its length?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft ripples expanding outward from directly below the blade's tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color grading heavily desaturated, creating a monochromatic world of black, white, and gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196252044290304 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196252044290304, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196252044290304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08652499318122864, + 0.15069788694381714, + 0.38862329721450806, + 0.1510476917028427, + 0.05292696878314018, + 0.06748399138450623, + 0.009704128839075565, + 0.033729881048202515, + 0.02925461158156395, + 0.0300065279006958 + ], + "credit_map": [ + 0.02101661078631878, + 0.010052917525172234, + 0.022232888266444206, + 0.007489015813916922, + 0.005108160898089409, + 0.008202200755476952, + 0.004577607847750187, + 0.007845585234463215, + 0.05419601500034332, + 0.007590460125356913, + 0.06231749802827835, + 0.004294217098504305, + 0.003450438380241394, + 0.00978204794228077, + 0.0022179891820997, + 0.006849224213510752, + 0.13713814318180084, + 0.018284978345036507, + 0.16438814997673035, + 0.010740633122622967, + 0.008963590487837791, + 0.024636324495077133, + 0.004359619691967964, + 0.02011185698211193, + 0.028031015768647194, + 0.024799790233373642, + 0.0245800968259573, + 0.019396942108869553, + 0.012123066000640392, + 0.01438052486628294, + 0.012469389475882053, + 0.015266867354512215, + 0.012391818687319756, + 0.004436672665178776, + 0.017566876485943794, + 0.0032004108652472496, + 0.0023000589571893215, + 0.004721013363450766, + 0.003961827605962753, + 0.004348290152847767, + 0.007854721508920193, + 0.005411017686128616, + 0.008938156068325043, + 0.007208898197859526, + 0.005885621532797813, + 0.01250176876783371, + 0.008438581600785255, + 0.01124523114413023, + 0.0024450907949358225, + 0.0005355565808713436, + 0.00268068490549922, + 0.0007353331311605871, + 0.0005025349673815072, + 0.0010263113072142005, + 0.0009538737940602005, + 0.0008247435325756669, + 0.004422561731189489, + 0.003215859644114971, + 0.005369605030864477, + 0.003985396586358547, + 0.0027699798811227083, + 0.005615751724690199, + 0.0040052118711173534, + 0.004345510620623827, + 0.008816608227789402, + 0.0013462654314935207, + 0.011090947315096855, + 0.0015627102693542838, + 0.0008429936715401709, + 0.002014098223298788, + 0.0017407032428309321, + 0.001840283628553152, + 0.007128724362701178, + 0.0017376249888911843, + 0.009372719563543797, + 0.0018963550683110952, + 0.0012031920487061143, + 0.002799295587465167, + 0.0033840842079371214, + 0.002484532305970788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.003196252044290304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_tensor([7556, 7556])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_tensor([7556, 7556])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dda53028786a050a826814759d23e95bcf13c4f0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_tensor([7556, 7556])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with the shaky, visceral movement of a handheld camera, low-angle shots slowly tilt upwards from a shattered, barren landscape, following the rise of a vast, surreal vortex that inexorably distorts the sky. This spiraling vortex is a chaotic amalgamation of melting clocks and stretched canvases; classic pocket watches with tarnished brass casings and luxurious clocks with dark wood frames twist and drip like hot wax, their numbered faces blurred as their hands droop lifelessly. Interwoven with this temporal decay are vast expanses of raw fiber canvas, stretched so tight they appear to tear, their pale, textured surfaces contrasting starkly against the viscous flow. The entire impossible structure is illuminated from within by solid, glassy tendrils of lightning, which emit a cold, inner glow as they snake through the vortex, casting sharp, crystalline highlights across its molten surface. Above, the sky itself roils in an unnatural, dancing aurora, its ever-shifting veil of sickly yellows, deep purples, and blood-reds creating a chaotic and malevolent celestial backdrop, completing a profoundly nightmarish vision, the erratic camerawork immersing the viewer directly into a waking dream of a decaying reality.", + "seed": 27041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_7556_seed_27041_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3511525807902217, + "vlm_logprob": -0.3511525807902217, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.3511525807902217, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5044665336608887 + ], + "opposite_logprob_tokens": [ + -0.25446659326553345 + ], + "logprob_reward": -1.5044665336608887, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5044665336608887, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25324270129203796, + 0.03578415513038635, + 0.06311025470495224, + 0.020346280187368393, + 0.023942112922668457, + 0.09950269758701324, + 0.3318396210670471, + 0.12850432097911835, + 0.024870209395885468, + 0.018857670947909355 + ], + "credit_map": [ + 0.07311273366212845, + 0.028596626594662666, + 0.013267140835523605, + 0.008760448545217514, + 0.010327178984880447, + 0.020066482946276665, + 0.014458658173680305, + 0.08465343713760376, + 0.004133651033043861, + 0.007429687771946192, + 0.0037355441600084305, + 0.0018608042737469077, + 0.010766038671135902, + 0.0032422428485006094, + 0.002153717912733555, + 0.0024624669458717108, + 0.012625997886061668, + 0.005851014517247677, + 0.004489253740757704, + 0.004798667039722204, + 0.006473803427070379, + 0.02011931873857975, + 0.004056117497384548, + 0.0046960823237895966, + 0.0021518899593502283, + 0.0027827259618788958, + 0.0034434411209076643, + 0.0023623239248991013, + 0.0036861433181911707, + 0.0020070502068847418, + 0.002158994786441326, + 0.0017537117237225175, + 0.0018914032261818647, + 0.004009857773780823, + 0.004749039653688669, + 0.0030369823798537254, + 0.003409408964216709, + 0.0025737795513123274, + 0.0021050290670245886, + 0.0021666137035936117, + 0.012031981721520424, + 0.006970202550292015, + 0.019788073375821114, + 0.010827693156898022, + 0.030866743996739388, + 0.008664793334901333, + 0.006181378848850727, + 0.004171825014054775, + 0.04000164568424225, + 0.02121269889175892, + 0.0538673959672451, + 0.013415761291980743, + 0.11810030788183212, + 0.044174924492836, + 0.03123030811548233, + 0.00983657781034708, + 0.010256467387080193, + 0.033820439130067825, + 0.012595397420227528, + 0.011745823547244072, + 0.009716744534671307, + 0.017312021926045418, + 0.012145806103944778, + 0.020911632105708122, + 0.0028739955741912127, + 0.00499025359749794, + 0.003189806593582034, + 0.0033196096774190664, + 0.0022151677403599024, + 0.0033269254490733147, + 0.002700235927477479, + 0.002254214370623231, + 0.0027573774568736553, + 0.0017585409805178642, + 0.00209735962562263, + 0.0013222494162619114, + 0.002244197065010667, + 0.003698304295539856, + 0.002322894986718893, + 0.0026567461900413036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -1.5044665336608887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12711083889007568 + ], + "opposite_logprob_tokens": [ + -2.1271109580993652 + ], + "logprob_reward": -0.12711083889007568, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12711083889007568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6557385921478271, + 0.021663136780261993, + 0.06252457201480865, + 0.022222284227609634, + 0.010592622682452202, + 0.016376972198486328, + 0.12426432222127914, + 0.05855071544647217, + 0.022928791120648384, + 0.005137922707945108 + ], + "credit_map": [ + 0.20956552028656006, + 0.07325843721628189, + 0.02450402081012726, + 0.019549822434782982, + 0.025316834449768066, + 0.04024083539843559, + 0.03674493357539177, + 0.22655825316905975, + 0.0023008747957646847, + 0.003577597439289093, + 0.004435128532350063, + 0.0023561278358101845, + 0.0020374569576233625, + 0.0022821540478616953, + 0.0023419237695634365, + 0.0023318747989833355, + 0.007255876436829567, + 0.005207093432545662, + 0.006884596776217222, + 0.006918045226484537, + 0.005651075392961502, + 0.015492988750338554, + 0.0061441343277692795, + 0.008970770984888077, + 0.0019028217066079378, + 0.002801668830215931, + 0.002315910765901208, + 0.003185058943927288, + 0.004165996331721544, + 0.002725183730944991, + 0.0029382489155977964, + 0.0021873987279832363, + 0.0009577000164426863, + 0.0013023094506934285, + 0.0009213313460350037, + 0.0013659704709425569, + 0.002166260499507189, + 0.001835064496845007, + 0.0011792716104537249, + 0.000864714733324945, + 0.0015083312755450606, + 0.0017668585060164332, + 0.0025946993846446276, + 0.003665027441456914, + 0.0021622604690492153, + 0.002266742754727602, + 0.0013988211285322905, + 0.0010142321698367596, + 0.007474442943930626, + 0.008464165031909943, + 0.03440125286579132, + 0.006700568832457066, + 0.016994841396808624, + 0.010786005295813084, + 0.0342734158039093, + 0.005169637501239777, + 0.005160549655556679, + 0.010789891704916954, + 0.007067719008773565, + 0.004189522936940193, + 0.005044554825872183, + 0.01067686453461647, + 0.0067406208254396915, + 0.008880993351340294, + 0.0024035770911723375, + 0.004392349161207676, + 0.0027967083733528852, + 0.0026545797009021044, + 0.002357304794713855, + 0.0021934802643954754, + 0.0026981281116604805, + 0.0034326647873967886, + 0.0005953002255409956, + 0.0007328612846322358, + 0.0007884605438448489, + 0.000584383204113692, + 0.000615936703979969, + 0.0005078116664662957, + 0.0005498254904523492, + 0.0007633438217453659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.12711083889007568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084645055234432 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084645055234432, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084645055234432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23684300482273102, + 0.0556798130273819, + 0.04144920036196709, + 0.021961985155940056, + 0.04277995601296425, + 0.030326372012495995, + 0.40072157979011536, + 0.11788121610879898, + 0.02936355210840702, + 0.022993337363004684 + ], + "credit_map": [ + 0.07612375169992447, + 0.02698223665356636, + 0.012925935909152031, + 0.009724432602524757, + 0.01068254467099905, + 0.01244180928915739, + 0.013238951563835144, + 0.07472334802150726, + 0.0038851897697895765, + 0.011402600444853306, + 0.006404521409422159, + 0.00445248419418931, + 0.01754378341138363, + 0.00466594472527504, + 0.003409312106668949, + 0.003915977198630571, + 0.004355944227427244, + 0.00427972711622715, + 0.00658530043438077, + 0.004819642286747694, + 0.005321127828210592, + 0.005647400859743357, + 0.0040108091197907925, + 0.006429243832826614, + 0.002073027892038226, + 0.003145090537145734, + 0.0037309026811271906, + 0.0025905100628733635, + 0.003302912460640073, + 0.0031779815908521414, + 0.0023466015700250864, + 0.0015949569642543793, + 0.0029361594934016466, + 0.004829661920666695, + 0.008333748206496239, + 0.003854446578770876, + 0.012809175997972488, + 0.004935799166560173, + 0.0029077704530209303, + 0.002173191402107477, + 0.0036562562454491854, + 0.003619837574660778, + 0.002409028122201562, + 0.0037096438463777304, + 0.008271225728094578, + 0.003294572466984391, + 0.0025058258324861526, + 0.0028599833603948355, + 0.03842554986476898, + 0.02669662795960903, + 0.08143666386604309, + 0.01762317679822445, + 0.11688622832298279, + 0.04744144156575203, + 0.05945875495672226, + 0.012753120623528957, + 0.009292829781770706, + 0.02974468283355236, + 0.011522896587848663, + 0.010514329187572002, + 0.009508052840828896, + 0.020183004438877106, + 0.011950675398111343, + 0.01516474038362503, + 0.0032023382373154163, + 0.0035632874350994825, + 0.0038012226577848196, + 0.0033514525275677443, + 0.004081962164491415, + 0.0034046026412397623, + 0.004043047782033682, + 0.003915639594197273, + 0.0020882897078990936, + 0.0022322009317576885, + 0.003878391347825527, + 0.002873255405575037, + 0.002646539593115449, + 0.0025281417183578014, + 0.0026155305095016956, + 0.004130988381803036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.011084645055234432 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043079957365989685 + ], + "opposite_logprob_tokens": [ + -3.1680798530578613 + ], + "logprob_reward": -0.043079957365989685, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043079957365989685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2947935461997986, + 0.14789697527885437, + 0.04646007716655731, + 0.036257680505514145, + 0.06491777300834656, + 0.02785932458937168, + 0.3101542592048645, + 0.034921981394290924, + 0.01985214650630951, + 0.016886265948414803 + ], + "credit_map": [ + 0.0711149275302887, + 0.030916281044483185, + 0.02368149906396866, + 0.014410545118153095, + 0.01472331490367651, + 0.030973367393016815, + 0.015930842608213425, + 0.09304274618625641, + 0.03174006566405296, + 0.013946808874607086, + 0.010731499642133713, + 0.010209888219833374, + 0.04311003163456917, + 0.01712889038026333, + 0.01057781744748354, + 0.01045196782797575, + 0.0060248784720897675, + 0.006176974158734083, + 0.005148341879248619, + 0.006956170313060284, + 0.005934059154242277, + 0.006971732713282108, + 0.004481272306293249, + 0.004766646772623062, + 0.0036942092701792717, + 0.0050348700024187565, + 0.004171973560005426, + 0.00424620695412159, + 0.007568201515823603, + 0.0037474804557859898, + 0.00475032115355134, + 0.0030444152653217316, + 0.00465420912951231, + 0.006862183567136526, + 0.009396139532327652, + 0.005464856047183275, + 0.01937675103545189, + 0.008802651427686214, + 0.006308169104158878, + 0.004052810370922089, + 0.0028007838409394026, + 0.003852008143439889, + 0.0023008391726762056, + 0.003631700063124299, + 0.006526471581310034, + 0.0036836357321590185, + 0.002472552238032222, + 0.002591330325230956, + 0.03961828351020813, + 0.023384656757116318, + 0.04463757947087288, + 0.01452743448317051, + 0.11729858815670013, + 0.040426552295684814, + 0.020173491910099983, + 0.010087672621011734, + 0.0029973876662552357, + 0.002093349117785692, + 0.003166026435792446, + 0.002584449714049697, + 0.004955883603543043, + 0.008998268283903599, + 0.0036270725540816784, + 0.006499539595097303, + 0.0020287095103412867, + 0.002174605382606387, + 0.0032577980309724808, + 0.001978176413103938, + 0.0019206298748031259, + 0.0022587040439248085, + 0.002414259361103177, + 0.0038192628417164087, + 0.00300467386841774, + 0.0018649252597242594, + 0.0015126676298677921, + 0.0015058553544804454, + 0.0017231416422873735, + 0.0030222334899008274, + 0.002251213416457176, + 0.002001553773880005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.043079957365989685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002092897891998 + ], + "opposite_logprob_tokens": [ + -2.695020914077759 + ], + "logprob_reward": -0.07002092897891998, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002092897891998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19364137947559357, + 0.04158787429332733, + 0.0860700011253357, + 0.052061986178159714, + 0.04187643900513649, + 0.03603320196270943, + 0.16419383883476257, + 0.1366717666387558, + 0.0882575511932373, + 0.1596059501171112 + ], + "credit_map": [ + 0.058978594839572906, + 0.021504586562514305, + 0.012220410630106926, + 0.007654951419681311, + 0.011305376887321472, + 0.011180060915648937, + 0.010347696952521801, + 0.06044969707727432, + 0.007465589325875044, + 0.004925866611301899, + 0.004109311383217573, + 0.0032913838513195515, + 0.011289001442492008, + 0.004466217011213303, + 0.0029365450609475374, + 0.0031039598397910595, + 0.010918055661022663, + 0.006316856015473604, + 0.012460233643651009, + 0.0057648709043860435, + 0.006774339824914932, + 0.02803172916173935, + 0.005525475367903709, + 0.010278443805873394, + 0.004663033876568079, + 0.007831545546650887, + 0.007145175244659185, + 0.005101589951664209, + 0.012451378628611565, + 0.005623200908303261, + 0.005983136128634214, + 0.0032629244960844517, + 0.00478369602933526, + 0.004936708137392998, + 0.004436406772583723, + 0.004618074744939804, + 0.008787733502686024, + 0.005672297440469265, + 0.004465938545763493, + 0.004175583831965923, + 0.0032258343417197466, + 0.004266459960490465, + 0.004918450955301523, + 0.007631929591298103, + 0.00590496975928545, + 0.0036739977076649666, + 0.003381303744390607, + 0.003030254505574703, + 0.01951301470398903, + 0.009963169693946838, + 0.007923291996121407, + 0.013976175338029861, + 0.059461355209350586, + 0.014272858388721943, + 0.031115422025322914, + 0.007968544960021973, + 0.008519723080098629, + 0.023035326972603798, + 0.0344269834458828, + 0.009848889894783497, + 0.00932341255247593, + 0.010568717494606972, + 0.017610562965273857, + 0.023338155820965767, + 0.010803726501762867, + 0.007977167144417763, + 0.016650130972266197, + 0.008767440915107727, + 0.008134747855365276, + 0.011544258333742619, + 0.010980663821101189, + 0.013399412855505943, + 0.024913502857089043, + 0.009205029346048832, + 0.020980125293135643, + 0.011589989066123962, + 0.0196194089949131, + 0.0354791022837162, + 0.014364257454872131, + 0.02345452830195427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.07002092897891998 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with the shaky, visceral movement of a handheld camera, low-angle shots slowly tilt upwards from a shattered, barren landscape, following the rise of a vast, surreal vortex that inexorably distorts the sky. This spiraling vortex is a chaotic amalgamation of melting clocks and stretched canvases; classic pocket watches with tarnished brass casings and luxurious clocks with dark wood frames twist and drip like hot wax, their numbered faces blurred as their hands droop lifelessly. Interwoven with this temporal decay are vast expanses of raw fiber canvas, stretched so tight they appear to tear, their pale, textured surfaces contrasting starkly against the viscous flow. The entire impossible structure is illuminated from within by solid, glassy tendrils of lightning, which emit a cold, inner glow as they snake through the vortex, casting sharp, crystalline highlights across its molten surface. Above, the sky itself roils in an unnatural, dancing aurora, its ever-shifting veil of sickly yellows, deep purples, and blood-reds creating a chaotic and malevolent celestial backdrop, completing a profoundly nightmarish vision, the erratic camerawork immersing the viewer directly into a waking dream of a decaying reality.", + "seed": 27042, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_7556_seed_27042_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4263915229588747, + "vlm_logprob": -0.4263915229588747, + "vlm_yes_no_margin": 1.824999988079071, + "vlm_token_credit": -0.4263915229588747, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5039969682693481 + ], + "opposite_logprob_tokens": [ + -0.25399696826934814 + ], + "logprob_reward": -1.5039969682693481, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5039969682693481, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24513472616672516, + 0.14652197062969208, + 0.059803035110235214, + 0.11398337036371231, + 0.05771156772971153, + 0.0901993066072464, + 0.038734205067157745, + 0.10374773293733597, + 0.06530941277742386, + 0.07885469496250153 + ], + "credit_map": [ + 0.03262542188167572, + 0.027615100145339966, + 0.0346999429166317, + 0.041156746447086334, + 0.028155617415905, + 0.02179066836833954, + 0.026751982048153877, + 0.03233923763036728, + 0.021182671189308167, + 0.014311891049146652, + 0.018346793949604034, + 0.01926695927977562, + 0.016087451949715614, + 0.02806493081152439, + 0.015727974474430084, + 0.013533295132219791, + 0.005344014149159193, + 0.007939545437693596, + 0.007523617707192898, + 0.005784118548035622, + 0.009312346577644348, + 0.00808336678892374, + 0.009724581614136696, + 0.0060914428904652596, + 0.008184299804270267, + 0.009291027672588825, + 0.010340115986764431, + 0.01393844187259674, + 0.014965733513236046, + 0.013031275011599064, + 0.014740847051143646, + 0.029491635039448738, + 0.005216910038143396, + 0.007069124840199947, + 0.007590597029775381, + 0.009277726523578167, + 0.005862474907189608, + 0.00859204214066267, + 0.007142470218241215, + 0.00696022342890501, + 0.009662243537604809, + 0.007501034997403622, + 0.01863546296954155, + 0.013316546566784382, + 0.011233692988753319, + 0.005902155768126249, + 0.012681137770414352, + 0.011267032474279404, + 0.004012685734778643, + 0.0028179341461509466, + 0.005133777856826782, + 0.003675242653116584, + 0.006309961900115013, + 0.005216971971094608, + 0.005200464278459549, + 0.006367162801325321, + 0.016802804544568062, + 0.011291510425508022, + 0.013835630379617214, + 0.012867220677435398, + 0.010229586623609066, + 0.010081030428409576, + 0.01529745478183031, + 0.013342498801648617, + 0.013155522756278515, + 0.007705470081418753, + 0.013631106354296207, + 0.00568761583417654, + 0.00559861958026886, + 0.00827370397746563, + 0.003932547755539417, + 0.007324825040996075, + 0.013996027410030365, + 0.009355513378977776, + 0.020916523411870003, + 0.009533731266856194, + 0.009046357125043869, + 0.004010324366390705, + 0.005891681183129549, + 0.006104535888880491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -1.5039969682693481 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871355652809143 + ], + "opposite_logprob_tokens": [ + -1.1371355056762695 + ], + "logprob_reward": -0.3871355652809143, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3871355652809143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34429600834846497, + 0.13638442754745483, + 0.09028685837984085, + 0.06793134659528732, + 0.0633225291967392, + 0.059013161808252335, + 0.08087344467639923, + 0.05408339574933052, + 0.03805244341492653, + 0.06575638055801392 + ], + "credit_map": [ + 0.03806757926940918, + 0.035405658185482025, + 0.05446008965373039, + 0.0545729324221611, + 0.0436612144112587, + 0.034404169768095016, + 0.04794802516698837, + 0.03577634319663048, + 0.0204477459192276, + 0.010683394968509674, + 0.017154766246676445, + 0.015827080234885216, + 0.015923036262392998, + 0.023098789155483246, + 0.0151207922026515, + 0.018128827214241028, + 0.004832204896956682, + 0.007330321241170168, + 0.008014777675271034, + 0.008021065033972263, + 0.012586054392158985, + 0.0146756861358881, + 0.011946764774620533, + 0.022879986092448235, + 0.005076538305729628, + 0.007471876684576273, + 0.0064257290214300156, + 0.006255932152271271, + 0.008251970633864403, + 0.008580156601965427, + 0.00894713681191206, + 0.016922008246183395, + 0.006879542954266071, + 0.005740731488913298, + 0.006819983012974262, + 0.0069479565136134624, + 0.0069572897627949715, + 0.00950777716934681, + 0.009351605549454689, + 0.01111763995140791, + 0.00641972990706563, + 0.0041435034945607185, + 0.00872120726853609, + 0.012468572705984116, + 0.004981056787073612, + 0.0035962022375315428, + 0.004259640350937843, + 0.014423251152038574, + 0.010237820446491241, + 0.008821851573884487, + 0.011892067268490791, + 0.009791990742087364, + 0.009830676950514317, + 0.009014353156089783, + 0.010996435768902302, + 0.010288247838616371, + 0.00782003253698349, + 0.007162097375839949, + 0.006992862094193697, + 0.006762413308024406, + 0.006453148555010557, + 0.005636302288621664, + 0.006290617398917675, + 0.006965924054384232, + 0.005225583910942078, + 0.00463244691491127, + 0.004289858043193817, + 0.005446827970445156, + 0.0039949906058609486, + 0.0052316212095320225, + 0.004500983282923698, + 0.004730131942778826, + 0.009887201711535454, + 0.010183773003518581, + 0.012376740574836731, + 0.009729750454425812, + 0.006565166637301445, + 0.005427558906376362, + 0.005160942208021879, + 0.006425251252949238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.3871355652809143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20658250153064728, + 0.19522623717784882, + 0.02688014507293701, + 0.07600957900285721, + 0.08606266230344772, + 0.04145124554634094, + 0.1533595770597458, + 0.05925210192799568, + 0.08772721886634827, + 0.06744876503944397 + ], + "credit_map": [ + 0.027603141963481903, + 0.025105180218815804, + 0.027281729504466057, + 0.03272968530654907, + 0.024698074907064438, + 0.01925474964082241, + 0.029369793832302094, + 0.02054014429450035, + 0.038300495594739914, + 0.018917767331004143, + 0.030087843537330627, + 0.019696764647960663, + 0.01748860627412796, + 0.04220929369330406, + 0.019805172458291054, + 0.008720300160348415, + 0.002802477916702628, + 0.0036262189969420433, + 0.0030327625572681427, + 0.003182350192219019, + 0.004404949489980936, + 0.0034504453651607037, + 0.0034009027294814587, + 0.002980040153488517, + 0.005351515021175146, + 0.007589964661747217, + 0.007187970913946629, + 0.011668211780488491, + 0.010313056409358978, + 0.008207841776311398, + 0.010472597554326057, + 0.015218419022858143, + 0.006129686255007982, + 0.009884894825518131, + 0.01045678649097681, + 0.010299964807927608, + 0.009163361974060535, + 0.014506835490465164, + 0.012424318119883537, + 0.013196817599236965, + 0.0032918036449700594, + 0.0034375644754618406, + 0.0072157238610088825, + 0.007447709329426289, + 0.00476984353736043, + 0.0027936778496950865, + 0.005016065202653408, + 0.007478855084627867, + 0.014737442135810852, + 0.01364039070904255, + 0.021716143935918808, + 0.030515961349010468, + 0.019710559397935867, + 0.013689168728888035, + 0.027491673827171326, + 0.011858243495225906, + 0.008639706298708916, + 0.005611838772892952, + 0.006337447091937065, + 0.008936175145208836, + 0.006081596948206425, + 0.005776053760200739, + 0.011246859095990658, + 0.006622421555221081, + 0.015209699980914593, + 0.010737055912613869, + 0.011353690177202225, + 0.009027916938066483, + 0.008798500522971153, + 0.013775958679616451, + 0.007846115157008171, + 0.010978278703987598, + 0.01166526973247528, + 0.008326152339577675, + 0.017627732828259468, + 0.00865953415632248, + 0.007701040245592594, + 0.003663536161184311, + 0.00457370188087225, + 0.005231800489127636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0059540290385484695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054976411163806915 + ], + "opposite_logprob_tokens": [ + -2.929976463317871 + ], + "logprob_reward": -0.054976411163806915, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054976411163806915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33166971802711487, + 0.22736288607120514, + 0.05098455399274826, + 0.0985783189535141, + 0.038654670119285583, + 0.03920165076851845, + 0.05796761438250542, + 0.04713086038827896, + 0.05900327488780022, + 0.0494464673101902 + ], + "credit_map": [ + 0.03434940055012703, + 0.02845308557152748, + 0.05752911418676376, + 0.06390774995088577, + 0.03154565766453743, + 0.04336415231227875, + 0.04056200385093689, + 0.03195854648947716, + 0.04578638821840286, + 0.02060655504465103, + 0.03904052823781967, + 0.033558424562215805, + 0.024126984179019928, + 0.018146198242902756, + 0.028203150257468224, + 0.017894655466079712, + 0.005755416117608547, + 0.007277786731719971, + 0.007395862601697445, + 0.0065238638781011105, + 0.007349290885031223, + 0.006350611802190542, + 0.006714288145303726, + 0.0036174326669424772, + 0.007342811208218336, + 0.011308170855045319, + 0.00803947914391756, + 0.01609865576028824, + 0.012117588892579079, + 0.010694167576730251, + 0.015061330050230026, + 0.01791611686348915, + 0.002855842001736164, + 0.004073682706803083, + 0.004441957455128431, + 0.00532611133530736, + 0.004153120797127485, + 0.00685121538117528, + 0.005457128398120403, + 0.005495612509548664, + 0.0031653030309826136, + 0.00328190834261477, + 0.007192816585302353, + 0.00606233486905694, + 0.004650709684938192, + 0.003038042923435569, + 0.004196343012154102, + 0.0076141939498484135, + 0.0058083548210561275, + 0.004328463226556778, + 0.006807589437812567, + 0.007921069860458374, + 0.009954787790775299, + 0.006196700502187014, + 0.01062771026045084, + 0.006322940811514854, + 0.0060721635818481445, + 0.005339223891496658, + 0.006160115357488394, + 0.005354074761271477, + 0.005057814065366983, + 0.004908734932541847, + 0.00822903960943222, + 0.006009689997881651, + 0.010223663412034512, + 0.006282583344727755, + 0.0073517197743058205, + 0.00480378745123744, + 0.005741815082728863, + 0.01068736333400011, + 0.00481324503198266, + 0.00909909512847662, + 0.008236304856836796, + 0.006174658425152302, + 0.011751146987080574, + 0.008568626828491688, + 0.005121620837599039, + 0.0025324642192572355, + 0.0038038953207433224, + 0.003257747506722808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.054976411163806915 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989464104175568 + ], + "opposite_logprob_tokens": [ + -1.8048946857452393 + ], + "logprob_reward": -0.17989464104175568, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989464104175568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12887971103191376, + 0.0871792733669281, + 0.027539853006601334, + 0.03040079027414322, + 0.10523761063814163, + 0.048634324222803116, + 0.2524164319038391, + 0.06304722279310226, + 0.17737117409706116, + 0.07929357886314392 + ], + "credit_map": [ + 0.020068950951099396, + 0.013997562229633331, + 0.018394380807876587, + 0.01441364549100399, + 0.018782982602715492, + 0.012054570019245148, + 0.013690690509974957, + 0.01747693307697773, + 0.017204230651259422, + 0.008068975992500782, + 0.01475438755005598, + 0.012588405050337315, + 0.0074249194003641605, + 0.010413927026093006, + 0.00871841050684452, + 0.008006022311747074, + 0.002710531232878566, + 0.003684509778395295, + 0.004688022658228874, + 0.003563262987881899, + 0.003293655812740326, + 0.0027871490456163883, + 0.003425309434533119, + 0.0033874106593430042, + 0.002261810004711151, + 0.0033535566180944443, + 0.002846318995580077, + 0.0032012888696044683, + 0.004163688514381647, + 0.0037019490264356136, + 0.005093422252684832, + 0.005778755526989698, + 0.005907316692173481, + 0.010641236789524555, + 0.008294123224914074, + 0.01044515147805214, + 0.010872036218643188, + 0.021835483610630035, + 0.017794156447052956, + 0.019448107108473778, + 0.003118027001619339, + 0.0035107198636978865, + 0.005555361974984407, + 0.008787984028458595, + 0.005971265025436878, + 0.004540588241070509, + 0.003548412350937724, + 0.013601966202259064, + 0.025249185040593147, + 0.024761060252785683, + 0.038759469985961914, + 0.039690762758255005, + 0.03177238255739212, + 0.020756985992193222, + 0.0458836704492569, + 0.02554290182888508, + 0.007552513852715492, + 0.004869677592068911, + 0.006311428267508745, + 0.007030746433883905, + 0.008419472724199295, + 0.00942410621792078, + 0.009313156828284264, + 0.010126120410859585, + 0.02517264150083065, + 0.01660875231027603, + 0.016307638958096504, + 0.011243577115237713, + 0.016169583424925804, + 0.04340328648686409, + 0.012245307676494122, + 0.036220401525497437, + 0.012972589582204819, + 0.007453912869095802, + 0.013114667497575283, + 0.007160456385463476, + 0.006206533871591091, + 0.008722822181880474, + 0.008281008340418339, + 0.015381590463221073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.17989464104175568 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with the shaky, visceral movement of a handheld camera, low-angle shots slowly tilt upwards from a shattered, barren landscape, following the rise of a vast, surreal vortex that inexorably distorts the sky. This spiraling vortex is a chaotic amalgamation of melting clocks and stretched canvases; classic pocket watches with tarnished brass casings and luxurious clocks with dark wood frames twist and drip like hot wax, their numbered faces blurred as their hands droop lifelessly. Interwoven with this temporal decay are vast expanses of raw fiber canvas, stretched so tight they appear to tear, their pale, textured surfaces contrasting starkly against the viscous flow. The entire impossible structure is illuminated from within by solid, glassy tendrils of lightning, which emit a cold, inner glow as they snake through the vortex, casting sharp, crystalline highlights across its molten surface. Above, the sky itself roils in an unnatural, dancing aurora, its ever-shifting veil of sickly yellows, deep purples, and blood-reds creating a chaotic and malevolent celestial backdrop, completing a profoundly nightmarish vision, the erratic camerawork immersing the viewer directly into a waking dream of a decaying reality.", + "seed": 27043, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_7556_seed_27043_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.33158663399517535, + "vlm_logprob": -0.33158663399517535, + "vlm_yes_no_margin": 1.8000000059604644, + "vlm_token_credit": -0.33158663399517535, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9766684770584106 + ], + "opposite_logprob_tokens": [ + -0.47666850686073303 + ], + "logprob_reward": -0.9766684770584106, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9766684770584106, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18757539987564087, + 0.07770661264657974, + 0.031980011612176895, + 0.01736108586192131, + 0.04316694661974907, + 0.1745690107345581, + 0.06339031457901001, + 0.2715080976486206, + 0.029289811849594116, + 0.10345270484685898 + ], + "credit_map": [ + 0.012399217113852501, + 0.01530755590647459, + 0.010245315730571747, + 0.015199289657175541, + 0.021094778552651405, + 0.04448534920811653, + 0.035314321517944336, + 0.03352959081530571, + 0.005640689749270678, + 0.007087163161486387, + 0.006432499270886183, + 0.008273162879049778, + 0.009671125560998917, + 0.015424779616296291, + 0.013974803499877453, + 0.011202394962310791, + 0.0038386089727282524, + 0.002435815753415227, + 0.003364050993695855, + 0.0040731229819357395, + 0.004807241261005402, + 0.004977453965693712, + 0.004159622360020876, + 0.004324098117649555, + 0.0020926406141370535, + 0.0016743431333452463, + 0.0016927649267017841, + 0.0021377084776759148, + 0.0026267943903803825, + 0.0026800346095114946, + 0.002266270574182272, + 0.0021905316971242428, + 0.00469928327947855, + 0.004829178098589182, + 0.004392887931317091, + 0.004161985591053963, + 0.005266514606773853, + 0.009046437218785286, + 0.00658143125474453, + 0.004189232364296913, + 0.03065180778503418, + 0.024492021650075912, + 0.01151052862405777, + 0.005885884631425142, + 0.04837145283818245, + 0.027192726731300354, + 0.011260101571679115, + 0.01520449761301279, + 0.005597588140517473, + 0.007282906677573919, + 0.005478963255882263, + 0.005185744259506464, + 0.010207065381109715, + 0.012429582886397839, + 0.00955444946885109, + 0.007654016837477684, + 0.012765473686158657, + 0.10991311073303223, + 0.0046678874641656876, + 0.008404175750911236, + 0.01990305818617344, + 0.08823853731155396, + 0.014895038679242134, + 0.012720818631350994, + 0.002579539082944393, + 0.0047590602189302444, + 0.0025874606799334288, + 0.0036411546170711517, + 0.004706361796706915, + 0.004825579933822155, + 0.0033959532156586647, + 0.0027947037015110254, + 0.00840341579169035, + 0.013995000161230564, + 0.005683526862412691, + 0.012194919399917126, + 0.012596824206411839, + 0.008483500219881535, + 0.017649540677666664, + 0.02444598264992237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.9766684770584106 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270710527896881 + ], + "opposite_logprob_tokens": [ + -2.1270711421966553 + ], + "logprob_reward": -0.1270710527896881, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270710527896881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1970852166414261, + 0.0492297038435936, + 0.05482487753033638, + 0.038138289004564285, + 0.045285314321517944, + 0.2723293602466583, + 0.12840309739112854, + 0.1080714538693428, + 0.03689425438642502, + 0.0697384849190712 + ], + "credit_map": [ + 0.012871747836470604, + 0.015389909967780113, + 0.009932968765497208, + 0.016544371843338013, + 0.019883058965206146, + 0.04976782947778702, + 0.0373726487159729, + 0.0353226512670517, + 0.004331586882472038, + 0.004491102881729603, + 0.003008162369951606, + 0.006304489448666573, + 0.005613991525024176, + 0.00976870208978653, + 0.008693517185747623, + 0.0070181479677557945, + 0.0072886464186012745, + 0.005459248088300228, + 0.003984340000897646, + 0.007262306287884712, + 0.006133132614195347, + 0.009484228678047657, + 0.008419936522841454, + 0.006793034262955189, + 0.00469996128231287, + 0.004397556185722351, + 0.0025774918030947447, + 0.003433989593759179, + 0.00585133908316493, + 0.007518560625612736, + 0.004438639618456364, + 0.005220749415457249, + 0.004506917670369148, + 0.005116034299135208, + 0.0054895225912332535, + 0.0062583209946751595, + 0.0046733710914850235, + 0.007020163349807262, + 0.008112930692732334, + 0.004108045715838671, + 0.06298917531967163, + 0.043074917048215866, + 0.01430538110435009, + 0.00973755493760109, + 0.04637381434440613, + 0.04923474043607712, + 0.016933059319853783, + 0.02968067117035389, + 0.008611149154603481, + 0.01872364617884159, + 0.013995089568197727, + 0.015083856880664825, + 0.02156897634267807, + 0.019127091392874718, + 0.021534455940127373, + 0.00975882075726986, + 0.007003289181739092, + 0.03991229087114334, + 0.003822560887783766, + 0.0054008085280656815, + 0.006974797695875168, + 0.01596040651202202, + 0.007277405355125666, + 0.021719882264733315, + 0.004170351196080446, + 0.005232708994299173, + 0.0027415670920163393, + 0.004046642687171698, + 0.006004537921398878, + 0.00500442273914814, + 0.003709065727889538, + 0.005984954070299864, + 0.008335022255778313, + 0.008100961335003376, + 0.004087141714990139, + 0.0069044968113303185, + 0.0058603910729289055, + 0.00500067975372076, + 0.01993272639811039, + 0.011517065577208996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.1270710527896881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12111447006464005, + 0.096063032746315, + 0.023813731968402863, + 0.04438411444425583, + 0.03708050027489662, + 0.12542736530303955, + 0.06224089860916138, + 0.20334795117378235, + 0.04064946994185448, + 0.2458784282207489 + ], + "credit_map": [ + 0.011045307852327824, + 0.009795461781322956, + 0.012784416787326336, + 0.011828672140836716, + 0.01316158752888441, + 0.025204522535204887, + 0.018249768763780594, + 0.019044730812311172, + 0.007683477830141783, + 0.010815970599651337, + 0.009965646080672741, + 0.01516498439013958, + 0.010371418669819832, + 0.015623223967850208, + 0.013611560687422752, + 0.012826751917600632, + 0.0021383173298090696, + 0.002905048429965973, + 0.0018717226339504123, + 0.0036873153876513243, + 0.002638967242091894, + 0.0036292620934545994, + 0.003898004302754998, + 0.0030450960621237755, + 0.00647606048732996, + 0.006097347475588322, + 0.002997086849063635, + 0.004112331662327051, + 0.005941732786595821, + 0.007156323175877333, + 0.0059212688356637955, + 0.005681965034455061, + 0.004023459739983082, + 0.004118206910789013, + 0.0037731872871518135, + 0.0036829339805990458, + 0.005877614486962557, + 0.004053738433867693, + 0.0072707198560237885, + 0.004280638415366411, + 0.014837018214166164, + 0.02710665576159954, + 0.007088008802384138, + 0.005020492244511843, + 0.03861754387617111, + 0.01333128847181797, + 0.010536359623074532, + 0.008889994584023952, + 0.006261907517910004, + 0.007393382024019957, + 0.007316432893276215, + 0.006717201322317123, + 0.009489650838077068, + 0.008660583756864071, + 0.010360556654632092, + 0.006041180342435837, + 0.009200344793498516, + 0.07847471535205841, + 0.003296209964901209, + 0.005637382622808218, + 0.015434070490300655, + 0.07156791538000107, + 0.011201802641153336, + 0.008535504341125488, + 0.003303543198853731, + 0.00744332792237401, + 0.0036204522475600243, + 0.003687362652271986, + 0.007301831152290106, + 0.005313202738761902, + 0.004503021482378244, + 0.005476726219058037, + 0.02014574036002159, + 0.011547204107046127, + 0.00485599460080266, + 0.01342141442000866, + 0.011508280411362648, + 0.008975888602435589, + 0.09268742054700851, + 0.08273649215698242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038353741168976 + ], + "opposite_logprob_tokens": [ + -1.9103835821151733 + ], + "logprob_reward": -0.16038353741168976, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038353741168976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19009517133235931, + 0.10798671841621399, + 0.03264964744448662, + 0.03479653224349022, + 0.040735550224781036, + 0.1445924937725067, + 0.06404199451208115, + 0.09745002537965775, + 0.02808164246380329, + 0.25957024097442627 + ], + "credit_map": [ + 0.01582944020628929, + 0.019359327852725983, + 0.014492101967334747, + 0.015302632004022598, + 0.01896190643310547, + 0.04183247685432434, + 0.033860426396131516, + 0.030456874519586563, + 0.0051963371224701405, + 0.012048936448991299, + 0.007112368009984493, + 0.013751786202192307, + 0.02016443945467472, + 0.018656469881534576, + 0.01550174131989479, + 0.0155546385794878, + 0.0024162752088159323, + 0.0026873499155044556, + 0.0024088395293802023, + 0.0035313633270561695, + 0.0043542636558413506, + 0.004892575554549694, + 0.0074980901554226875, + 0.004860890097916126, + 0.0030994010157883167, + 0.0037199377547949553, + 0.0031711598858237267, + 0.004176231566816568, + 0.004928095731884241, + 0.006897836923599243, + 0.004349689930677414, + 0.004454178269952536, + 0.0031150789000093937, + 0.0036312045995146036, + 0.004311777651309967, + 0.0037962987553328276, + 0.010938560590147972, + 0.003976085223257542, + 0.0056650275364518166, + 0.005301518365740776, + 0.03471900150179863, + 0.0438787005841732, + 0.008159268647432327, + 0.006313360296189785, + 0.008900725282728672, + 0.017730217427015305, + 0.011471359059214592, + 0.013419868424534798, + 0.006196973845362663, + 0.00942464079707861, + 0.006742146331816912, + 0.00702475244179368, + 0.009363044053316116, + 0.008577877655625343, + 0.010604589246213436, + 0.0061079696752130985, + 0.004255683161318302, + 0.03214278072118759, + 0.0032822631765156984, + 0.0028746421448886395, + 0.008185061626136303, + 0.034144215285778046, + 0.006497751455754042, + 0.006067618262022734, + 0.002204903867095709, + 0.004326134920120239, + 0.002459001261740923, + 0.0036148785147815943, + 0.004506799858063459, + 0.004881981294602156, + 0.002860552631318569, + 0.0032273903489112854, + 0.02860971726477146, + 0.012599228881299496, + 0.006734913680702448, + 0.01219336036592722, + 0.01252242736518383, + 0.010400286875665188, + 0.10498291999101639, + 0.07152736932039261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.16038353741168976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706645369529724 + ], + "opposite_logprob_tokens": [ + -1.1370664834976196 + ], + "logprob_reward": -0.38706645369529724, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38706645369529724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12942418456077576, + 0.035187844187021255, + 0.034685246646404266, + 0.030605178326368332, + 0.03432578966021538, + 0.18673808872699738, + 0.18471793830394745, + 0.07601825892925262, + 0.0804823711514473, + 0.20781509578227997 + ], + "credit_map": [ + 0.01177005935460329, + 0.012557769194245338, + 0.013033698312938213, + 0.011905476450920105, + 0.016453657299280167, + 0.02171957492828369, + 0.022221362218260765, + 0.019762586802244186, + 0.00333677907474339, + 0.003379354253411293, + 0.0027917774859815836, + 0.003756467020139098, + 0.006373121403157711, + 0.006622702348977327, + 0.0044149491004645824, + 0.0045126923359930515, + 0.00393061526119709, + 0.003864535130560398, + 0.0029658719431608915, + 0.004266292788088322, + 0.004250054247677326, + 0.004982513841241598, + 0.005838464014232159, + 0.004586901515722275, + 0.0027157145086675882, + 0.0033018800895661116, + 0.0030247652903199196, + 0.003755163634195924, + 0.004537561908364296, + 0.005359900649636984, + 0.0040105483494699, + 0.003899645060300827, + 0.0030074957758188248, + 0.004083565436303616, + 0.0032233763486146927, + 0.003763532964512706, + 0.006749312859028578, + 0.005687193479388952, + 0.004274353384971619, + 0.0035369624383747578, + 0.011176413856446743, + 0.05552704259753227, + 0.005301936529576778, + 0.006216438952833414, + 0.04997514188289642, + 0.03903740644454956, + 0.010401013307273388, + 0.009102686308324337, + 0.011861106380820274, + 0.023534879088401794, + 0.015371683053672314, + 0.017545966431498528, + 0.04074111580848694, + 0.040521327406167984, + 0.0221428032964468, + 0.012999072670936584, + 0.005782639607787132, + 0.02067897655069828, + 0.0060428171418607235, + 0.0055631063878536224, + 0.006918306462466717, + 0.008737805299460888, + 0.007032695692032576, + 0.015261915512382984, + 0.006170094478875399, + 0.015295972116291523, + 0.007158036343753338, + 0.006565289571881294, + 0.018200598657131195, + 0.0104982890188694, + 0.0076585677452385426, + 0.008935521356761456, + 0.015441573224961758, + 0.017045767977833748, + 0.006537441164255142, + 0.01147172786295414, + 0.015006093308329582, + 0.011074219830334187, + 0.0563521645963192, + 0.07488609850406647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.38706645369529724 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with the shaky, visceral movement of a handheld camera, low-angle shots slowly tilt upwards from a shattered, barren landscape, following the rise of a vast, surreal vortex that inexorably distorts the sky. This spiraling vortex is a chaotic amalgamation of melting clocks and stretched canvases; classic pocket watches with tarnished brass casings and luxurious clocks with dark wood frames twist and drip like hot wax, their numbered faces blurred as their hands droop lifelessly. Interwoven with this temporal decay are vast expanses of raw fiber canvas, stretched so tight they appear to tear, their pale, textured surfaces contrasting starkly against the viscous flow. The entire impossible structure is illuminated from within by solid, glassy tendrils of lightning, which emit a cold, inner glow as they snake through the vortex, casting sharp, crystalline highlights across its molten surface. Above, the sky itself roils in an unnatural, dancing aurora, its ever-shifting veil of sickly yellows, deep purples, and blood-reds creating a chaotic and malevolent celestial backdrop, completing a profoundly nightmarish vision, the erratic camerawork immersing the viewer directly into a waking dream of a decaying reality.", + "seed": 27044, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_6/index_7556_seed_27044_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.5227728767320514, + "vlm_logprob": -0.5227728767320514, + "vlm_yes_no_margin": 1.849999988079071, + "vlm_token_credit": -0.5227728767320514, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4091484546661377 + ], + "opposite_logprob_tokens": [ + -0.2841483950614929 + ], + "logprob_reward": -1.4091484546661377, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4091484546661377, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09974678605794907, + 0.05900733172893524, + 0.05682260915637016, + 0.015512322075664997, + 0.07719641923904419, + 0.5048211812973022, + 0.042791008949279785, + 0.08121563494205475, + 0.02594994567334652, + 0.03693673759698868 + ], + "credit_map": [ + 0.020888103172183037, + 0.01398977730423212, + 0.010348220355808735, + 0.013026361353695393, + 0.008182416670024395, + 0.010389295406639576, + 0.010512305423617363, + 0.012410318478941917, + 0.0053284368477761745, + 0.006933982018381357, + 0.0063459426164627075, + 0.011755162850022316, + 0.005958252120763063, + 0.0098024383187294, + 0.006637670565396547, + 0.006245449185371399, + 0.005740576423704624, + 0.008995690383017063, + 0.0054275477305054665, + 0.007535897195339203, + 0.00725548667833209, + 0.006482536904513836, + 0.005205016583204269, + 0.010179858654737473, + 0.001321911346167326, + 0.0025624395348131657, + 0.002352282637730241, + 0.002196825109422207, + 0.0017859638901427388, + 0.001982598565518856, + 0.001409524236805737, + 0.0019007782684639096, + 0.013404753990471363, + 0.012030692771077156, + 0.007057712879031897, + 0.008289407938718796, + 0.00991333182901144, + 0.007326552178710699, + 0.010557367466390133, + 0.008616603910923004, + 0.023032572120428085, + 0.013395962305366993, + 0.06632880121469498, + 0.012084312736988068, + 0.053890667855739594, + 0.041662976145744324, + 0.1276734620332718, + 0.16675247251987457, + 0.004540524445474148, + 0.0033816173672676086, + 0.0070287068374454975, + 0.00544547475874424, + 0.003632698906585574, + 0.0066928560845553875, + 0.00661718612536788, + 0.00545194698497653, + 0.012156170792877674, + 0.005656380672007799, + 0.006784605327993631, + 0.008166217245161533, + 0.00714237941429019, + 0.020198224112391472, + 0.0133929792791605, + 0.007718678098171949, + 0.002876089885830879, + 0.00246194819919765, + 0.00204213778488338, + 0.0068327938206493855, + 0.0019706138409674168, + 0.0021797490771859884, + 0.005377579014748335, + 0.002209037309512496, + 0.002600409323349595, + 0.004339960403740406, + 0.006138673052191734, + 0.010443047620356083, + 0.003294224152341485, + 0.0025790557265281677, + 0.0041904449462890625, + 0.003350927261635661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -1.4091484546661377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539307594299316 + ], + "opposite_logprob_tokens": [ + -0.42893069982528687 + ], + "logprob_reward": -1.0539307594299316, + "yes_no_margin_reward": -0.6250000596046448, + "token_credit_reward": -1.0539307594299316, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1577323079109192, + 0.0906524509191513, + 0.06312867999076843, + 0.028079558163881302, + 0.1272611916065216, + 0.2664915919303894, + 0.06619955599308014, + 0.10128440707921982, + 0.05491199716925621, + 0.04425826668739319 + ], + "credit_map": [ + 0.027301320806145668, + 0.02228729985654354, + 0.01563357189297676, + 0.01989017240703106, + 0.0165080726146698, + 0.02134121023118496, + 0.0184322576969862, + 0.0163384098559618, + 0.008128093555569649, + 0.016587205231189728, + 0.01193560753017664, + 0.014899437315762043, + 0.009553609415888786, + 0.011168397963047028, + 0.0077161965891718864, + 0.010663905180990696, + 0.009578464552760124, + 0.007264584302902222, + 0.008596109226346016, + 0.007973880507051945, + 0.008214334025979042, + 0.0074905757792294025, + 0.006686013657599688, + 0.007324722595512867, + 0.002976472256705165, + 0.004025700036436319, + 0.006158345844596624, + 0.002891208976507187, + 0.003019551746547222, + 0.0028323540464043617, + 0.002732861554250121, + 0.003443061374127865, + 0.01207055151462555, + 0.010632185265421867, + 0.014714531600475311, + 0.01795533113181591, + 0.021484265103936195, + 0.01333000510931015, + 0.014675918966531754, + 0.02239840291440487, + 0.011628658510744572, + 0.014477407559752464, + 0.035172417759895325, + 0.011193271726369858, + 0.027576005086302757, + 0.025049740448594093, + 0.06158808246254921, + 0.07980599254369736, + 0.005259892903268337, + 0.006870490964502096, + 0.009814033284783363, + 0.0055036116391420364, + 0.006973562762141228, + 0.015581234358251095, + 0.006441599689424038, + 0.009755130857229233, + 0.012875063344836235, + 0.00691648805513978, + 0.005743641406297684, + 0.03140595182776451, + 0.012731376104056835, + 0.012440300546586514, + 0.014916214160621166, + 0.004255377221852541, + 0.008730705827474594, + 0.00443281652405858, + 0.009077761322259903, + 0.008415240794420242, + 0.0038234381936490536, + 0.00912582315504551, + 0.00572033179923892, + 0.005585877690464258, + 0.004299836233258247, + 0.00655197910964489, + 0.006344825029373169, + 0.009820896200835705, + 0.003985637798905373, + 0.003064890159294009, + 0.005570251494646072, + 0.004619951825588942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -1.0539307594299316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004105354659259319 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.004105354659259319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105354659259319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13096271455287933, + 0.058996137231588364, + 0.05060568079352379, + 0.02057068608701229, + 0.11026474088430405, + 0.3971973955631256, + 0.04240073263645172, + 0.06875038892030716, + 0.08406948298215866, + 0.036182060837745667 + ], + "credit_map": [ + 0.02234446071088314, + 0.01572866179049015, + 0.01257184986025095, + 0.015248946845531464, + 0.014947261661291122, + 0.02057853899896145, + 0.013761448673903942, + 0.015781544148921967, + 0.0048630088567733765, + 0.008241620846092701, + 0.006275818217545748, + 0.010199099779129028, + 0.006239024456590414, + 0.010818212293088436, + 0.005520838312804699, + 0.006838512606918812, + 0.005693334620445967, + 0.007037844974547625, + 0.005668079014867544, + 0.00670911418274045, + 0.006491244304925203, + 0.005573599133640528, + 0.0053960541263222694, + 0.008036409504711628, + 0.0029791935812681913, + 0.004105980508029461, + 0.0021969564259052277, + 0.0021753443870693445, + 0.001993520651012659, + 0.0024451413191854954, + 0.0016520003555342555, + 0.0030225501395761967, + 0.015080850571393967, + 0.01715255342423916, + 0.013644776307046413, + 0.016063449904322624, + 0.012550613842904568, + 0.009015364572405815, + 0.014851647429168224, + 0.011905485764145851, + 0.020510191097855568, + 0.016988620162010193, + 0.05096929520368576, + 0.0112874461337924, + 0.04062003642320633, + 0.03200891986489296, + 0.0963946208357811, + 0.12841825187206268, + 0.0048856912180781364, + 0.004826833959668875, + 0.006354969926178455, + 0.004938330035656691, + 0.004434152506291866, + 0.006300379056483507, + 0.005631816573441029, + 0.005028559826314449, + 0.008562205359339714, + 0.00481424992904067, + 0.004095647484064102, + 0.023270050063729286, + 0.006333186756819487, + 0.005357386544346809, + 0.012907497584819794, + 0.00341016729362309, + 0.011694756336510181, + 0.006340221036225557, + 0.011216356419026852, + 0.024307679384946823, + 0.004197991918772459, + 0.008223272860050201, + 0.011909454129636288, + 0.006179749965667725, + 0.0030339646618813276, + 0.004592734854668379, + 0.005766596179455519, + 0.010194259695708752, + 0.002594292163848877, + 0.002768490696325898, + 0.003908715210855007, + 0.00332300434820354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004105354659259319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372414782643318 + ], + "opposite_logprob_tokens": [ + -3.408724069595337 + ], + "logprob_reward": -0.03372414782643318, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372414782643318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11893847584724426, + 0.032480865716934204, + 0.0442931242287159, + 0.018527325242757797, + 0.11603560298681259, + 0.5303584337234497, + 0.025452561676502228, + 0.04432712867856026, + 0.05233790725469589, + 0.017248528078198433 + ], + "credit_map": [ + 0.019611075520515442, + 0.01858079992234707, + 0.010312974452972412, + 0.010820344090461731, + 0.016337234526872635, + 0.015400013886392117, + 0.015323958359658718, + 0.012552066706120968, + 0.0024005097802728415, + 0.00442825211212039, + 0.0033226811792701483, + 0.006555186118930578, + 0.003480166895315051, + 0.006558266468346119, + 0.0026334705762565136, + 0.00310233305208385, + 0.006182521115988493, + 0.006382642313838005, + 0.005839938763529062, + 0.00789997074753046, + 0.00582156190648675, + 0.0037355481181293726, + 0.004647466354072094, + 0.0037834797985851765, + 0.002167344093322754, + 0.0034165915567427874, + 0.002874127123504877, + 0.00256796064786613, + 0.002295007463544607, + 0.0013740769354626536, + 0.0019296255195513368, + 0.0019025916699320078, + 0.011178664863109589, + 0.01795051246881485, + 0.015490001998841763, + 0.02511664666235447, + 0.013571598567068577, + 0.010594958439469337, + 0.013186344876885414, + 0.008946877904236317, + 0.02634122408926487, + 0.017869941890239716, + 0.06776493787765503, + 0.01376293320208788, + 0.05615914613008499, + 0.04283503443002701, + 0.1326323300600052, + 0.17299291491508484, + 0.0029540695250034332, + 0.003865040373057127, + 0.0038311185780912638, + 0.0038634678348898888, + 0.0025155958719551563, + 0.002606268972158432, + 0.002894547302275896, + 0.0029224506579339504, + 0.005647570826113224, + 0.002992581343278289, + 0.00437383446842432, + 0.005444671027362347, + 0.003987752832472324, + 0.012603064998984337, + 0.005312985274940729, + 0.003964665345847607, + 0.010491185821592808, + 0.004785426426678896, + 0.006926314439624548, + 0.012634610757231712, + 0.002243990544229746, + 0.0037759814877063036, + 0.00834820605814457, + 0.003132190555334091, + 0.0016089397249743342, + 0.0020899316295981407, + 0.00245169410482049, + 0.0035189371556043625, + 0.0016925205709412694, + 0.001926514902152121, + 0.0018102189060300589, + 0.0021497709676623344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03372414782643318 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld with shaky movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low angle and slowly tilt upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central structure a vortex composed of melting clocks and stretched canvases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the clocks depicted with tarnished brass casings and dark wood frames that appear to twist and drip like hot wax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by solid, glassy tendrils of lightning emitting a cold inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295566707849503 + ], + "opposite_logprob_tokens": [ + -2.2379555702209473 + ], + "logprob_reward": -0.11295566707849503, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295566707849503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17953535914421082, + 0.11597220599651337, + 0.09605737775564194, + 0.03656585514545441, + 0.10808181762695312, + 0.13223594427108765, + 0.08372725546360016, + 0.09951165318489075, + 0.1068384051322937, + 0.04147415980696678 + ], + "credit_map": [ + 0.027112072333693504, + 0.030151937156915665, + 0.0143626369535923, + 0.012211904861032963, + 0.02861027419567108, + 0.025887755677103996, + 0.024209124967455864, + 0.01698964461684227, + 0.00994752999395132, + 0.018301144242286682, + 0.009160427376627922, + 0.013910643756389618, + 0.013896780088543892, + 0.027856599539518356, + 0.008437689393758774, + 0.014461391605436802, + 0.010134130716323853, + 0.015147116035223007, + 0.010681668296456337, + 0.01254945807158947, + 0.01701708883047104, + 0.011136683635413647, + 0.010011767037212849, + 0.00937946792691946, + 0.0044958703219890594, + 0.0077574606984853745, + 0.005461613181978464, + 0.0036763784009963274, + 0.004550635349005461, + 0.004312330856919289, + 0.002824746072292328, + 0.0034868167713284492, + 0.008795864880084991, + 0.009191468358039856, + 0.0061847493052482605, + 0.011874384246766567, + 0.021953824907541275, + 0.014243127778172493, + 0.015185326337814331, + 0.020653072744607925, + 0.010295470245182514, + 0.013632622547447681, + 0.016065193340182304, + 0.006388961803168058, + 0.013213532976806164, + 0.012121275998651981, + 0.02346806973218918, + 0.037050820887088776, + 0.009191669523715973, + 0.01135454885661602, + 0.010681034065783024, + 0.0077398899011313915, + 0.010935015976428986, + 0.011073248460888863, + 0.010043946094810963, + 0.012707898393273354, + 0.0070719243958592415, + 0.005453051067888737, + 0.008161347359418869, + 0.034604623913764954, + 0.016946915537118912, + 0.005932355299592018, + 0.017067305743694305, + 0.0042741307988762856, + 0.021020786836743355, + 0.009451552294194698, + 0.020640095695853233, + 0.01782059855759144, + 0.0059633259661495686, + 0.010869383811950684, + 0.012300917878746986, + 0.008771747350692749, + 0.0034369726199656725, + 0.0058135767467319965, + 0.006361343432217836, + 0.007903753779828548, + 0.0032069876324385405, + 0.004477133043110371, + 0.005161435343325138, + 0.0051129600033164024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.11295566707849503 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_tensor([8865, 8865])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_tensor([8865, 8865])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ab58a8ce9be2f197b8b7c61b66569448cc63b0f4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_tensor([8865, 8865])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot frames a lone candlelight against a deep, blurred black background, creating a scene of profound intimacy. Initially, the flame burns with a steady, teardrop-shaped glow, the main body of the flame a vibrant yellow-orange, gradually fading to a soft blue, licking the blackened wick. The process of extinguishing begins gracefully, the flame beginning to wither like a dying flower; its strong, bright edges soften and curl inward, while its tip teeters and droops downward. The pulses of light it emits are erratic, weakening with each flicker. With a final spasm, the flame collapses in on itself and disappears entirely, plunging the scene into near-total darkness. A tiny, glowing red spur on the wick is the final trace of light, before a thin, ethereal puff of white smoke rises deliberately, twisting and unfurling, melting into the void, leaving behind a palpable sense of quiet and finality.", + "seed": 16004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16004_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.008417788380756974, + "vlm_logprob": -0.008417788380756974, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.008417788380756974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02860267646610737 + ], + "opposite_logprob_tokens": [ + -4.403602600097656 + ], + "logprob_reward": -0.02860267646610737, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02860267646610737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33492475748062134, + 0.0554034449160099, + 0.05632760748267174, + 0.12193514406681061, + 0.0812298133969307, + 0.16313178837299347, + 0.02978585474193096, + 0.036678534001111984, + 0.07869844138622284, + 0.04188457876443863 + ], + "credit_map": [ + 0.019734254106879234, + 0.021751627326011658, + 0.14317071437835693, + 0.00998332817107439, + 0.004415805917233229, + 0.11595498770475388, + 0.006746916100382805, + 0.013167120516300201, + 0.004564798902720213, + 0.004299536347389221, + 0.021151741966605186, + 0.0024017083924263716, + 0.0012583986390382051, + 0.016453512012958527, + 0.0021651028655469418, + 0.0031086455564945936, + 0.0072044082917273045, + 0.008750122971832752, + 0.014334150590002537, + 0.005402446258813143, + 0.0023046527057886124, + 0.009838190861046314, + 0.003865610109642148, + 0.00462802778929472, + 0.0076813483610749245, + 0.010172028094530106, + 0.04166948422789574, + 0.009081016294658184, + 0.004910668823868036, + 0.038118135184049606, + 0.004512388724833727, + 0.005790071561932564, + 0.007949171587824821, + 0.021180331707000732, + 0.01743941940367222, + 0.00686770211905241, + 0.0055041005834937096, + 0.011108572594821453, + 0.0033144073095172644, + 0.007866105064749718, + 0.008336151018738747, + 0.018084513023495674, + 0.062060218304395676, + 0.0075488765724003315, + 0.008647174574434757, + 0.048530854284763336, + 0.002990286098793149, + 0.0069337207823991776, + 0.0027010568883270025, + 0.003597927512601018, + 0.00424936693161726, + 0.005937568377703428, + 0.005027787759900093, + 0.0025288977194577456, + 0.00238088914193213, + 0.003362360643222928, + 0.002849490847438574, + 0.004581835586577654, + 0.004221235401928425, + 0.004165735561400652, + 0.008127272129058838, + 0.0040492648258805275, + 0.005017440300434828, + 0.0036662574857473373, + 0.0053070103749632835, + 0.00862655509263277, + 0.019897935912013054, + 0.008252917788922787, + 0.01017654687166214, + 0.007589351385831833, + 0.010697252117097378, + 0.008150869980454445, + 0.0031729149632155895, + 0.006438442040234804, + 0.0058840918354690075, + 0.006513160187751055, + 0.006767624989151955, + 0.004380771424621344, + 0.005605392623692751, + 0.0031221790704876184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.02860267646610737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002837918698787689 + ], + "opposite_logprob_tokens": [ + -5.877838134765625 + ], + "logprob_reward": -0.002837918698787689, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002837918698787689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20931488275527954, + 0.05973679944872856, + 0.06626709550619125, + 0.10989213734865189, + 0.1434517204761505, + 0.29402467608451843, + 0.028416350483894348, + 0.02196699008345604, + 0.038525063544511795, + 0.028404317796230316 + ], + "credit_map": [ + 0.011369839310646057, + 0.010962649248540401, + 0.0811457559466362, + 0.006558246910572052, + 0.002654863288626075, + 0.08062051981687546, + 0.003872262081131339, + 0.012130755931138992, + 0.004305215552449226, + 0.004158396273851395, + 0.023516956716775894, + 0.002781142946332693, + 0.001011502230539918, + 0.019038768485188484, + 0.0018112417310476303, + 0.0031135736498981714, + 0.00765322195366025, + 0.009046025574207306, + 0.019757423549890518, + 0.006072493270039558, + 0.0021271102596074343, + 0.012255102396011353, + 0.004307813011109829, + 0.005047910381108522, + 0.007762568537145853, + 0.01649858057498932, + 0.034320149570703506, + 0.00942298024892807, + 0.0026371064595878124, + 0.030213557183742523, + 0.0032907500863075256, + 0.00574643025174737, + 0.007983325980603695, + 0.01710624247789383, + 0.052571818232536316, + 0.005081043113023043, + 0.004903118126094341, + 0.046083707362413406, + 0.003108944045379758, + 0.006613524165004492, + 0.014160404913127422, + 0.030245603993535042, + 0.11942517012357712, + 0.011569375172257423, + 0.014447025954723358, + 0.08839843422174454, + 0.004340630955994129, + 0.011438043788075447, + 0.0022355779074132442, + 0.0033494618255645037, + 0.004070294555276632, + 0.004999273922294378, + 0.004424256272614002, + 0.0030058950651437044, + 0.003647300647571683, + 0.002684288891032338, + 0.0017601697472855449, + 0.0030518488492816687, + 0.0021781830582767725, + 0.0024640019983053207, + 0.004519208800047636, + 0.002108703600242734, + 0.0037135325837880373, + 0.00217134109698236, + 0.0029107097070664167, + 0.0056871711276471615, + 0.008773823268711567, + 0.0035681433510035276, + 0.005848783999681473, + 0.004416882060468197, + 0.003317754017189145, + 0.004001797176897526, + 0.0022632679902017117, + 0.005107327364385128, + 0.004076697863638401, + 0.00416935607790947, + 0.0038181315176188946, + 0.0035398344043642282, + 0.002637211699038744, + 0.002792492276057601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.002837918698787689 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025032395496964455 + ], + "opposite_logprob_tokens": [ + -6.002503395080566 + ], + "logprob_reward": -0.0025032395496964455, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025032395496964455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06881649047136307, + 0.08302690833806992, + 0.05857470631599426, + 0.05788840353488922, + 0.07562246173620224, + 0.41287216544151306, + 0.04091363027691841, + 0.051495250314474106, + 0.0952908992767334, + 0.05549909546971321 + ], + "credit_map": [ + 0.00830259919166565, + 0.008538158610463142, + 0.015864936634898186, + 0.005710131023079157, + 0.004629014525562525, + 0.007649516686797142, + 0.006469021085649729, + 0.011653108522295952, + 0.00450326781719923, + 0.004197822418063879, + 0.02755148895084858, + 0.003653107210993767, + 0.002233096631243825, + 0.029903514310717583, + 0.003527027554810047, + 0.007457571104168892, + 0.005541146267205477, + 0.007933512330055237, + 0.015592708252370358, + 0.003117271699011326, + 0.0021077876444905996, + 0.01621508225798607, + 0.0029956938233226538, + 0.005071497522294521, + 0.005682359449565411, + 0.011041493155062199, + 0.007708892226219177, + 0.008817676454782486, + 0.00302032008767128, + 0.010044624097645283, + 0.0037011110689491034, + 0.007871923968195915, + 0.008465539664030075, + 0.018742714077234268, + 0.009435628540813923, + 0.005902489181607962, + 0.006370126269757748, + 0.016032401472330093, + 0.0031019137240946293, + 0.007571643218398094, + 0.022613300010561943, + 0.05228954553604126, + 0.1791333109140396, + 0.015255056321620941, + 0.025880897417664528, + 0.08246321231126785, + 0.00582785252481699, + 0.02940899133682251, + 0.0034611632581800222, + 0.005616617389023304, + 0.005603714380413294, + 0.0062165819108486176, + 0.006619304418563843, + 0.004901089705526829, + 0.005195444915443659, + 0.0032997094094753265, + 0.004547051154077053, + 0.0039884732104837894, + 0.0051809740252792835, + 0.004322791937738657, + 0.010172362439334393, + 0.004280880559235811, + 0.014031754806637764, + 0.00497095612809062, + 0.006618079263716936, + 0.018791161477565765, + 0.018776556476950645, + 0.011148309335112572, + 0.015363924205303192, + 0.009718051180243492, + 0.007031729444861412, + 0.007843072526156902, + 0.003918179776519537, + 0.010791629552841187, + 0.006806620862334967, + 0.007979098707437515, + 0.00845357496291399, + 0.005901585798710585, + 0.007369038183242083, + 0.004279363434761763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.0025032395496964455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198899397626519 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198899397626519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198899397626519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09192421287298203, + 0.022310150787234306, + 0.03863327205181122, + 0.026391569525003433, + 0.040851227939128876, + 0.22763025760650635, + 0.04849786311388016, + 0.3564692735671997, + 0.07760678976774216, + 0.06968533992767334 + ], + "credit_map": [ + 0.005234824027866125, + 0.004758304916322231, + 0.03503327816724777, + 0.0036948868073523045, + 0.0013563609682023525, + 0.03419555723667145, + 0.00216773827560246, + 0.0054832627065479755, + 0.0024041628930717707, + 0.0016494736773893237, + 0.004113384056836367, + 0.0014427717542275786, + 0.0007254724041558802, + 0.007543903309851885, + 0.001062522060237825, + 0.0033684612717479467, + 0.0029804683290421963, + 0.0034771235659718513, + 0.01474557165056467, + 0.0027160050813108683, + 0.0011550700291991234, + 0.009749134071171284, + 0.0015910034999251366, + 0.002218892564997077, + 0.0019516197498887777, + 0.003535731928423047, + 0.006927465554326773, + 0.0031776665709912777, + 0.0011320291087031364, + 0.0065537490881979465, + 0.0014549209736287594, + 0.0016583862015977502, + 0.0025859305169433355, + 0.005693033803254366, + 0.01048604492098093, + 0.003162097418680787, + 0.0021165101788938046, + 0.012506473809480667, + 0.0014102788409218192, + 0.0028908601962029934, + 0.01123928464949131, + 0.02311481535434723, + 0.09471631050109863, + 0.007939798757433891, + 0.011159438639879227, + 0.06795164197683334, + 0.0030096692498773336, + 0.008499289862811565, + 0.0028951806016266346, + 0.007383906282484531, + 0.008944650180637836, + 0.00612601637840271, + 0.0072260149754583836, + 0.004849259275943041, + 0.006386772729456425, + 0.004686061292886734, + 0.03299622982740402, + 0.02587432600557804, + 0.035760167986154556, + 0.03704829514026642, + 0.05237805098295212, + 0.028683265671133995, + 0.10415519773960114, + 0.03957373648881912, + 0.005150448065251112, + 0.012540999799966812, + 0.017034003511071205, + 0.007581318728625774, + 0.011512313969433308, + 0.008956721052527428, + 0.007452729158103466, + 0.007378255482763052, + 0.00496275257319212, + 0.01173634547740221, + 0.011571284383535385, + 0.007430470082908869, + 0.00940172839909792, + 0.008657274767756462, + 0.007019317243248224, + 0.008906169794499874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.002198899397626519 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946207791566849 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946207791566849, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946207791566849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11563587188720703, + 0.027735630050301552, + 0.011511523276567459, + 0.0459551177918911, + 0.04660170525312424, + 0.18329407274723053, + 0.07359007745981216, + 0.07060558348894119, + 0.34465283155441284, + 0.08041758835315704 + ], + "credit_map": [ + 0.0066605438478291035, + 0.006839916575700045, + 0.04919016361236572, + 0.0036357573699206114, + 0.0013858762104064226, + 0.04145918786525726, + 0.0019355302210897207, + 0.004528893157839775, + 0.0017371175345033407, + 0.0019599231891334057, + 0.011400328017771244, + 0.0010644850553944707, + 0.00040516964509151876, + 0.009464711882174015, + 0.0005547064356505871, + 0.0011491887271404266, + 0.0010766179766505957, + 0.0013660213444381952, + 0.0023652121890336275, + 0.0012536152498796582, + 0.0005395723856054246, + 0.003025980433449149, + 0.0007101154769770801, + 0.001174388569779694, + 0.0032204571180045605, + 0.006519603077322245, + 0.01560017466545105, + 0.0036301815416663885, + 0.000984362093731761, + 0.012335044331848621, + 0.0013838585000485182, + 0.0022814366966485977, + 0.0028730612248182297, + 0.006796462461352348, + 0.016735371202230453, + 0.002328573726117611, + 0.0017806576797738671, + 0.012835774570703506, + 0.0008834726177155972, + 0.0023683311883360147, + 0.009726612828671932, + 0.020996857434511185, + 0.07851286232471466, + 0.006844594608992338, + 0.01040769461542368, + 0.04393942654132843, + 0.002457576571032405, + 0.010408436879515648, + 0.004825843498110771, + 0.00834444910287857, + 0.010203829035162926, + 0.01302726473659277, + 0.01353819016367197, + 0.007819830439984798, + 0.010232537053525448, + 0.005598137155175209, + 0.006719223689287901, + 0.006523490883409977, + 0.00771030317991972, + 0.007413177751004696, + 0.010122252628207207, + 0.006187897641211748, + 0.018672259524464607, + 0.007256983779370785, + 0.024009039625525475, + 0.0626247301697731, + 0.06787711381912231, + 0.03597360849380493, + 0.054589785635471344, + 0.03875989094376564, + 0.03207366541028023, + 0.028744949027895927, + 0.005737232975661755, + 0.007311781868338585, + 0.025311442092061043, + 0.007508904207497835, + 0.007246468216180801, + 0.012766757979989052, + 0.004074872005730867, + 0.010460130870342255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.005946207791566849 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot frames a lone candlelight against a deep, blurred black background, creating a scene of profound intimacy. Initially, the flame burns with a steady, teardrop-shaped glow, the main body of the flame a vibrant yellow-orange, gradually fading to a soft blue, licking the blackened wick. The process of extinguishing begins gracefully, the flame beginning to wither like a dying flower; its strong, bright edges soften and curl inward, while its tip teeters and droops downward. The pulses of light it emits are erratic, weakening with each flicker. With a final spasm, the flame collapses in on itself and disappears entirely, plunging the scene into near-total darkness. A tiny, glowing red spur on the wick is the final trace of light, before a thin, ethereal puff of white smoke rises deliberately, twisting and unfurling, melting into the void, leaving behind a palpable sense of quiet and finality.", + "seed": 16005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16005_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011502451822161674, + "vlm_logprob": -0.011502451822161674, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.011502451822161674, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041284192353487015 + ], + "opposite_logprob_tokens": [ + -4.916284084320068 + ], + "logprob_reward": -0.041284192353487015, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.041284192353487015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.002829332137480378, + 0.0018744941335171461, + 0.0038719570729881525, + 0.013083732686936855, + 0.002985064871609211, + 0.005834217183291912, + 0.9515551328659058, + 0.006142976228147745, + 0.006407048553228378, + 0.005416075699031353 + ], + "credit_map": [ + 0.0003149135736748576, + 0.00015685279504396021, + 0.0006643965025432408, + 0.00022804480977356434, + 0.000151449057739228, + 0.0006240262882784009, + 0.00031918135937303305, + 0.00037046769284643233, + 0.00027975422563031316, + 6.519292946904898e-05, + 0.0005107623874209821, + 8.197046554414555e-05, + 7.507840200560167e-05, + 0.0003842823498416692, + 0.00015267515846062452, + 0.0003247781714890152, + 0.0005218144506216049, + 0.0003233866882510483, + 0.0009405401069670916, + 0.00027510724612511694, + 0.00027956216945312917, + 0.0006306957802735269, + 0.0003715846687555313, + 0.0005292660207487643, + 0.0014935035724192858, + 0.0005788629059679806, + 0.0031222361139953136, + 0.0008416809723712504, + 0.0008916105143725872, + 0.0024772922042757273, + 0.0023872065357863903, + 0.0012913396349176764, + 0.0002622041502036154, + 8.447532309219241e-05, + 0.0007863108185119927, + 0.00010473449947312474, + 0.00020501935796346515, + 0.000913263123948127, + 0.00033546643680892885, + 0.00029359100153669715, + 0.00040041212923824787, + 0.00014769523113500327, + 0.0019578393548727036, + 0.00014614325482398272, + 0.00037445404450409114, + 0.002171644475311041, + 0.00023461067758034915, + 0.0004014185396954417, + 0.05457505211234093, + 0.009012317284941673, + 0.3938393294811249, + 0.015272422693669796, + 0.028502263128757477, + 0.40230488777160645, + 0.010664237663149834, + 0.037384551018476486, + 0.0006355237565003335, + 0.0013076786417514086, + 0.0010527821723371744, + 0.0003593589353840798, + 0.0008734972216188908, + 0.0008471637265756726, + 0.00041066351695917547, + 0.0006563083734363317, + 0.0006628558621741831, + 0.000865063804667443, + 0.001293194480240345, + 0.0004714230017270893, + 0.001218662247993052, + 0.000992353307083249, + 0.0004195034853182733, + 0.00048399210209026933, + 0.0005033942288719118, + 0.0005554993404075503, + 0.0016840262105688453, + 0.0005617217393592, + 0.0005551860085688531, + 0.000634299183730036, + 0.0005616718553937972, + 0.0003602771903388202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.041284192353487015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005276918411254883 + ], + "opposite_logprob_tokens": [ + -5.255276679992676 + ], + "logprob_reward": -0.005276918411254883, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005276918411254883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.007467504125088453, + 0.003169678384438157, + 0.013750010170042515, + 0.0025776242837309837, + 0.006543624214828014, + 0.0034658797085285187, + 0.9531211256980896, + 0.002287814859300852, + 0.0041061947122216225, + 0.00351059646345675 + ], + "credit_map": [ + 0.0011025338899344206, + 0.0002844182890839875, + 0.0019053943688049912, + 0.0003719511441886425, + 0.00029163784347474575, + 0.002011386677622795, + 0.0005188842187635601, + 0.0009812979260459542, + 0.0004999970551580191, + 0.00021133168775122613, + 0.0005482127889990807, + 0.0003685987612698227, + 0.00023496786889154464, + 0.000396868446841836, + 0.0004925851826556027, + 0.00041711670928634703, + 0.0022841482423245907, + 0.001200278289616108, + 0.0026824630331248045, + 0.0011495609069243073, + 0.0009492282988503575, + 0.0021209069527685642, + 0.001565246027894318, + 0.0017981786513701081, + 0.0002267126546939835, + 9.434683306608349e-05, + 0.0007496774196624756, + 0.00012430235801730305, + 0.00014974887017160654, + 0.0006401400896720588, + 0.0003964719653595239, + 0.00019622381660155952, + 0.0007957520429044962, + 0.0001784689084161073, + 0.0018544899066910148, + 0.00023140640405472368, + 0.000380700541427359, + 0.0018623776268213987, + 0.0005083363503217697, + 0.000732092303223908, + 0.00031569538987241685, + 0.0001721909939078614, + 0.0009751154575496912, + 0.00012242594675626606, + 0.00024385802680626512, + 0.0009858803823590279, + 0.00025223311968147755, + 0.00039848050801083446, + 0.05470544099807739, + 0.008925476111471653, + 0.3946395218372345, + 0.015316816046833992, + 0.02848649211227894, + 0.40305042266845703, + 0.01069502905011177, + 0.0373019315302372, + 0.00026780925691127777, + 0.00047217344399541616, + 0.00029448079294525087, + 0.00012743797560688108, + 0.0003267109568696469, + 0.0002671661786735058, + 0.00016664693248458207, + 0.0003653891908470541, + 0.0004034576704725623, + 0.0006178568000905216, + 0.0006216800538823009, + 0.00038985401624813676, + 0.0008753625443205237, + 0.0005032263579778373, + 0.00033492446527816355, + 0.0003598328330554068, + 0.0003066422650590539, + 0.0004153256013523787, + 0.0007026875973679125, + 0.000386350613553077, + 0.000572013552300632, + 0.0003218292840756476, + 0.000542572233825922, + 0.0002631753741297871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.005276918411254883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025031205732375383 + ], + "opposite_logprob_tokens": [ + -6.002502918243408 + ], + "logprob_reward": -0.0025031205732375383, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025031205732375383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0028153453022241592, + 0.0016826912760734558, + 0.0019242243142798543, + 0.003982036840170622, + 0.0023652613162994385, + 0.03822140768170357, + 0.9360120892524719, + 0.004491792991757393, + 0.004323363769799471, + 0.004181862808763981 + ], + "credit_map": [ + 0.000308590562781319, + 0.00015308649744838476, + 0.0007074207533150911, + 0.000244572467636317, + 0.00014628045028075576, + 0.0005038644885644317, + 0.000403289363021031, + 0.00034824019530788064, + 0.00015834493387956172, + 6.891509110573679e-05, + 0.00033181224716827273, + 0.00013435905566439033, + 0.00012231190339662135, + 0.0003351072664372623, + 0.0002508974284864962, + 0.00028094317531213164, + 0.00023647781927138567, + 0.00013324088649824262, + 0.0003087351506110281, + 0.0001326852652709931, + 0.000139355513965711, + 0.0004413448041304946, + 0.00025161210214719176, + 0.0002807726268656552, + 0.0003214050375390798, + 0.00014902754628565162, + 0.0013407135847955942, + 0.00019426185463089496, + 0.00018978537991642952, + 0.0010394679848104715, + 0.0004754082765430212, + 0.00027196621522307396, + 0.00017664185725152493, + 8.317624451592565e-05, + 0.0007611353066749871, + 8.617330604465678e-05, + 0.0001558009535074234, + 0.0007469397387467325, + 0.00019290338968858123, + 0.0001624901924515143, + 0.0024689105339348316, + 0.0005944497534073889, + 0.013983212411403656, + 0.0006447468767873943, + 0.0020687072537839413, + 0.014840478077530861, + 0.000983552192337811, + 0.002637348836287856, + 0.05371754616498947, + 0.008759050630033016, + 0.38751259446144104, + 0.015070718713104725, + 0.027955086901783943, + 0.39587581157684326, + 0.01047943439334631, + 0.036641720682382584, + 0.0004698920820374042, + 0.001039354596287012, + 0.000704553269315511, + 0.00022563259699381888, + 0.000688271364197135, + 0.0006209746352396905, + 0.00024289978318847716, + 0.0005002141697332263, + 0.00041475435136817396, + 0.0005240216851234436, + 0.0011877057841047645, + 0.0003235595941077918, + 0.0005333191365934908, + 0.000867231865413487, + 0.00021718716016039252, + 0.00025558355264365673, + 0.00024121686874423176, + 0.0004648042668122798, + 0.0008381889783777297, + 0.0006650367868132889, + 0.0006376105011440814, + 0.00047259742859750986, + 0.0006249117432162166, + 0.00023749576939735562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0025031205732375383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031947072129696608 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031947072129696608, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031947072129696608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.002134365029633045, + 0.0010997281642630696, + 0.001456733327358961, + 0.0021062579471617937, + 0.0019819496665149927, + 0.0157554242759943, + 0.9382679462432861, + 0.007878792472183704, + 0.009972292929887772, + 0.01934647001326084 + ], + "credit_map": [ + 0.00027578361914493144, + 0.00010659195686457679, + 0.00046251763706095517, + 0.00017001094238366932, + 0.00011181164882145822, + 0.0004787435755133629, + 0.0002608077193144709, + 0.0002680982870515436, + 0.00013351145025808364, + 7.847337110433728e-05, + 0.0002419250668026507, + 9.31718823267147e-05, + 7.055824244162068e-05, + 0.00020559360564220697, + 0.00010660334373824298, + 0.00016989123832900077, + 0.00021224221563898027, + 0.00010899980406975374, + 0.0003065987257286906, + 0.00014766551612410694, + 0.00010868940444197506, + 0.0001875380694400519, + 0.0002055846416624263, + 0.00017941498663276434, + 0.0002931090712081641, + 0.00012892195081803948, + 0.00043600480421446264, + 0.0001570119202369824, + 0.00016070035053417087, + 0.00034752333885990083, + 0.00032140707480721176, + 0.0002615793782752007, + 0.00026858432102017105, + 6.824750016676262e-05, + 0.0004020839696750045, + 8.309493568958715e-05, + 0.0001450422132620588, + 0.0004774163826368749, + 0.00026704539777711034, + 0.00027043523732572794, + 0.0009497359860688448, + 0.00018590173567645252, + 0.0057644532062113285, + 0.00025326028116978705, + 0.0008244863129220903, + 0.006437873002141714, + 0.00031813859823159873, + 0.0010215756483376026, + 0.053757891058921814, + 0.008803420700132847, + 0.3880874216556549, + 0.015060666017234325, + 0.028141630813479424, + 0.3970161974430084, + 0.010528518818318844, + 0.0368722565472126, + 0.0008267700322903693, + 0.002016572980210185, + 0.0011049871100112796, + 0.0005613123648799956, + 0.0008895649807527661, + 0.0009586166706867516, + 0.0005652100080624223, + 0.0009557591984048486, + 0.0011246155481785536, + 0.001318879658356309, + 0.0016551818698644638, + 0.0009694157633930445, + 0.002099457196891308, + 0.001210461719892919, + 0.0007311575463972986, + 0.0008631243836134672, + 0.0011825929395854473, + 0.002018751809373498, + 0.003064521588385105, + 0.003279811469838023, + 0.0032503020484000444, + 0.0021786801517009735, + 0.00306047685444355, + 0.0013113345485180616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.0031947072129696608 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253320559859276 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253320559859276, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253320559859276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.008024396374821663, + 0.0029785765800625086, + 0.005042506381869316, + 0.014718013815581799, + 0.006583903916180134, + 0.0038664820604026318, + 0.8778680562973022, + 0.04872332140803337, + 0.008324640803039074, + 0.02387014590203762 + ], + "credit_map": [ + 0.0009513566037639976, + 0.000679220596794039, + 0.001523640938103199, + 0.0006545503274537623, + 0.000601380888838321, + 0.0012049522483721375, + 0.0015466418117284775, + 0.000862652319483459, + 0.0004089409194421023, + 0.00023297574080061167, + 0.0006391767528839409, + 0.000211631617275998, + 0.0002028500457527116, + 0.00043614028254523873, + 0.0003832288202829659, + 0.00046363211004063487, + 0.0006379231344908476, + 0.0004838111053686589, + 0.0008808648562990129, + 0.0004725016187876463, + 0.0004438014002516866, + 0.0005990915233269334, + 0.0008207870996557176, + 0.000703725207131356, + 0.0019214069470763206, + 0.0007504543173126876, + 0.0030793261248618364, + 0.0010044481605291367, + 0.0010844194330275059, + 0.0026146769523620605, + 0.0026727975346148014, + 0.0015904835890978575, + 0.0007847653469070792, + 0.00016812294779811054, + 0.0019136000191792846, + 0.00020084874995518476, + 0.0003430081414990127, + 0.001906678662635386, + 0.000490323465783149, + 0.0007765557966195047, + 0.0004679973644670099, + 0.00019878671446349472, + 0.0006864463211968541, + 0.00018567373626865447, + 0.00041750300442799926, + 0.0006880849250592291, + 0.0003293726476840675, + 0.000892616983037442, + 0.050347231328487396, + 0.00799029041081667, + 0.3641088902950287, + 0.014142727479338646, + 0.026168936863541603, + 0.3711787462234497, + 0.009702691808342934, + 0.034228529781103134, + 0.005820689257234335, + 0.011266649700701237, + 0.0070480285212397575, + 0.002871451433748007, + 0.00679813465103507, + 0.005471118725836277, + 0.002965133171528578, + 0.006482113152742386, + 0.0006711968453601003, + 0.001093115541152656, + 0.0020925470162183046, + 0.0005823432002216578, + 0.0012723824474960566, + 0.0015816628001630306, + 0.0004607073497027159, + 0.0005706859519705176, + 0.001940106856636703, + 0.002481916919350624, + 0.004011171869933605, + 0.0033334458712488413, + 0.003916155081242323, + 0.002790824742987752, + 0.003675836371257901, + 0.0017206863267347217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.005253320559859276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot frames a lone candlelight against a deep, blurred black background, creating a scene of profound intimacy. Initially, the flame burns with a steady, teardrop-shaped glow, the main body of the flame a vibrant yellow-orange, gradually fading to a soft blue, licking the blackened wick. The process of extinguishing begins gracefully, the flame beginning to wither like a dying flower; its strong, bright edges soften and curl inward, while its tip teeters and droops downward. The pulses of light it emits are erratic, weakening with each flicker. With a final spasm, the flame collapses in on itself and disappears entirely, plunging the scene into near-total darkness. A tiny, glowing red spur on the wick is the final trace of light, before a thin, ethereal puff of white smoke rises deliberately, twisting and unfurling, melting into the void, leaving behind a palpable sense of quiet and finality.", + "seed": 16006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16006_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007033959613181651, + "vlm_logprob": -0.007033959613181651, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.007033959613181651, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020180033519864082 + ], + "opposite_logprob_tokens": [ + -4.8951802253723145 + ], + "logprob_reward": -0.020180033519864082, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.020180033519864082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1346878558397293, + 0.041795242577791214, + 0.06198886036872864, + 0.08852329105138779, + 0.08291266113519669, + 0.2121211588382721, + 0.11520826071500778, + 0.044049493968486786, + 0.13292273879051208, + 0.0857904851436615 + ], + "credit_map": [ + 0.021481789648532867, + 0.006242992356419563, + 0.013720272108912468, + 0.011380613781511784, + 0.0211940985172987, + 0.029756169766187668, + 0.00941414199769497, + 0.021497787907719612, + 0.0043976809829473495, + 0.002412552246823907, + 0.005762279499322176, + 0.005261869169771671, + 0.006128861103206873, + 0.0077627599239349365, + 0.00588319543749094, + 0.004186046309769154, + 0.004336592275649309, + 0.0014120697742328048, + 0.00388537161052227, + 0.0026223689783364534, + 0.018485991284251213, + 0.02190956100821495, + 0.0036207609809935093, + 0.005716144572943449, + 0.010494858026504517, + 0.00327121140435338, + 0.007555054035037756, + 0.004129759501665831, + 0.015615892596542835, + 0.023061227053403854, + 0.009640561416745186, + 0.014754729345440865, + 0.01484849862754345, + 0.004844567738473415, + 0.01045869942754507, + 0.008830113336443901, + 0.008193312212824821, + 0.013027303852140903, + 0.009290453046560287, + 0.01341971568763256, + 0.03089289925992489, + 0.00821575615555048, + 0.02154964953660965, + 0.013720858842134476, + 0.03479558601975441, + 0.046346019953489304, + 0.014765668660402298, + 0.041834719479084015, + 0.019677599892020226, + 0.0034013616386801004, + 0.00997011736035347, + 0.004371010232716799, + 0.013329785317182541, + 0.024732958525419235, + 0.006132918410003185, + 0.03359251469373703, + 0.004000763408839703, + 0.004193295259028673, + 0.006289895623922348, + 0.005122564733028412, + 0.009013990871608257, + 0.007566946558654308, + 0.0042188456282019615, + 0.003643190022557974, + 0.01090127695351839, + 0.015791641548275948, + 0.016149457544088364, + 0.01590091921389103, + 0.01724505051970482, + 0.020449867472052574, + 0.013193934224545956, + 0.023290595039725304, + 0.005128016229718924, + 0.008574976585805416, + 0.005300756543874741, + 0.005228667985647917, + 0.020628975704312325, + 0.028657108545303345, + 0.0039751529693603516, + 0.008296833373606205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.020180033519864082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655119031667709 + ], + "opposite_logprob_tokens": [ + -5.379654884338379 + ], + "logprob_reward": -0.004655119031667709, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655119031667709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07773258537054062, + 0.03090781904757023, + 0.21133078634738922, + 0.07471746951341629, + 0.027920059859752655, + 0.20540162920951843, + 0.15944452583789825, + 0.06535746902227402, + 0.0904674082994461, + 0.05672029033303261 + ], + "credit_map": [ + 0.010067220777273178, + 0.002680196426808834, + 0.006116427015513182, + 0.0033098035492002964, + 0.014324811287224293, + 0.023194843903183937, + 0.004272047430276871, + 0.013767235912382603, + 0.004195918329060078, + 0.0010963493259623647, + 0.0030164665076881647, + 0.0026413146406412125, + 0.004656117409467697, + 0.007699631620198488, + 0.002952245296910405, + 0.004649774171411991, + 0.013210156932473183, + 0.003560776123777032, + 0.009526369161903858, + 0.005929905455559492, + 0.07091101258993149, + 0.08134961128234863, + 0.009626254439353943, + 0.01721671223640442, + 0.005965794436633587, + 0.004097179975360632, + 0.006840350106358528, + 0.005969456862658262, + 0.013929938897490501, + 0.019069993868470192, + 0.010374760255217552, + 0.008469990454614162, + 0.003975456580519676, + 0.0016773870447650552, + 0.0030024994630366564, + 0.0016520657809451222, + 0.0048741428181529045, + 0.0047128391452133656, + 0.002592560136690736, + 0.005433110054582357, + 0.028250200673937798, + 0.007755212485790253, + 0.019888754934072495, + 0.011772114783525467, + 0.03442302718758583, + 0.04770900309085846, + 0.016359394416213036, + 0.039243925362825394, + 0.026505058631300926, + 0.005688473116606474, + 0.014932805672287941, + 0.007160583510994911, + 0.018231285735964775, + 0.033271320164203644, + 0.007603551726788282, + 0.046051450073719025, + 0.005163721274584532, + 0.0054626972414553165, + 0.006984923966228962, + 0.005740865133702755, + 0.018424473702907562, + 0.011493014171719551, + 0.00629610288888216, + 0.005791673436760902, + 0.010129887610673904, + 0.01449377741664648, + 0.011293807998299599, + 0.012556958012282848, + 0.010532241314649582, + 0.012737400829792023, + 0.010254574008286, + 0.008468752726912498, + 0.002950319554656744, + 0.00567117752507329, + 0.003269579727202654, + 0.0033458652906119823, + 0.013410553336143494, + 0.019375862553715706, + 0.0026339006144553423, + 0.0060630301013588905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.004655119031667709 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032055205665528774 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.0032055205665528774, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032055205665528774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07288968563079834, + 0.05754920467734337, + 0.40099036693573, + 0.02241678722202778, + 0.053580090403556824, + 0.07454852759838104, + 0.09869296103715897, + 0.024952471256256104, + 0.09628871828317642, + 0.0980912446975708 + ], + "credit_map": [ + 0.008589822798967361, + 0.0034531631972640753, + 0.008129877969622612, + 0.006098480895161629, + 0.012615897692739964, + 0.012299691326916218, + 0.012689822353422642, + 0.009012926369905472, + 0.005328078754246235, + 0.003351398278027773, + 0.00622682785615325, + 0.0048712994903326035, + 0.011883411556482315, + 0.01311921700835228, + 0.007562111131846905, + 0.005206862930208445, + 0.02365042269229889, + 0.005095154512673616, + 0.014128975570201874, + 0.008379888720810413, + 0.14737381041049957, + 0.16547434031963348, + 0.008606200106441975, + 0.028281595557928085, + 0.0024431042838841677, + 0.0019957616459578276, + 0.0019377598073333502, + 0.0018283964600414038, + 0.003642852883785963, + 0.004232371691614389, + 0.003933868370950222, + 0.0024026716127991676, + 0.008829747326672077, + 0.002914999844506383, + 0.005492578726261854, + 0.003759243991225958, + 0.0068023670464754105, + 0.011359700001776218, + 0.0035229981876909733, + 0.010898451320827007, + 0.009443793445825577, + 0.002826835261657834, + 0.007517238613218069, + 0.004951185081154108, + 0.012676164507865906, + 0.01731281541287899, + 0.005977259017527103, + 0.013843237422406673, + 0.01742648147046566, + 0.003654500935226679, + 0.010109972208738327, + 0.004851971287280321, + 0.010602886788547039, + 0.019480248913168907, + 0.00540584372356534, + 0.027161043137311935, + 0.0028153490275144577, + 0.002026512986049056, + 0.0028392693493515253, + 0.0024477806873619556, + 0.00564965233206749, + 0.00487094558775425, + 0.0017936488147825003, + 0.0025093110743910074, + 0.011362874880433083, + 0.010342580266296864, + 0.012394447810947895, + 0.011423190124332905, + 0.01611201837658882, + 0.014662030152976513, + 0.009100489318370819, + 0.01089108269661665, + 0.004454043693840504, + 0.009793834760785103, + 0.00698145991191268, + 0.005688067991286516, + 0.0224031712859869, + 0.03195333853363991, + 0.005659120157361031, + 0.011158211156725883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.0032055205665528774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948696285486221 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948696285486221, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948696285486221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17323558032512665, + 0.044559139758348465, + 0.03143593668937683, + 0.021719293668866158, + 0.0552026703953743, + 0.10275465250015259, + 0.039947379380464554, + 0.1218109279870987, + 0.17310775816440582, + 0.23622667789459229 + ], + "credit_map": [ + 0.023362167179584503, + 0.00614686356857419, + 0.015910770744085312, + 0.01224646344780922, + 0.029998285695910454, + 0.043610818684101105, + 0.011825990863144398, + 0.03013422153890133, + 0.004511135630309582, + 0.0026213950477540493, + 0.005037663970142603, + 0.004275818821042776, + 0.00821872428059578, + 0.007368250284343958, + 0.008361692540347576, + 0.004164460115134716, + 0.0034072601702064276, + 0.0012967773946002126, + 0.0025131655856966972, + 0.0016189318848773837, + 0.007685227319598198, + 0.009414558298885822, + 0.002077938988804817, + 0.003422078210860491, + 0.002568845171481371, + 0.0019488303223624825, + 0.0030211308039724827, + 0.0021829945035278797, + 0.002867023227736354, + 0.003769214265048504, + 0.0025768287014216185, + 0.002784424927085638, + 0.009846890345215797, + 0.002618124708533287, + 0.006944654975086451, + 0.005613807123154402, + 0.006554461549967527, + 0.00888980645686388, + 0.004383183550089598, + 0.010351737961173058, + 0.014952365309000015, + 0.0038266070187091827, + 0.009740669280290604, + 0.005480160471051931, + 0.016837749630212784, + 0.022729845717549324, + 0.008863172493875027, + 0.020324088633060455, + 0.004444763530045748, + 0.0017438260838389397, + 0.0038273329846560955, + 0.0026092317420989275, + 0.005952144041657448, + 0.009508022107183933, + 0.003985393792390823, + 0.007876663468778133, + 0.011094991117715836, + 0.007598801515996456, + 0.015775762498378754, + 0.008538481779396534, + 0.033390533179044724, + 0.02013317495584488, + 0.011852066963911057, + 0.013427112251520157, + 0.02134755440056324, + 0.0317980982363224, + 0.016388164833188057, + 0.01766801066696644, + 0.03339323773980141, + 0.022076914086937904, + 0.015697108581662178, + 0.014738667756319046, + 0.010893520899116993, + 0.021146351471543312, + 0.011562343686819077, + 0.012994992546737194, + 0.06262022256851196, + 0.08782203495502472, + 0.008842977695167065, + 0.020344233140349388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005948696285486221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011804286623373628 + ], + "opposite_logprob_tokens": [ + -6.751180648803711 + ], + "logprob_reward": -0.0011804286623373628, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011804286623373628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03662277013063431, + 0.02333124354481697, + 0.17415763437747955, + 0.029948949813842773, + 0.018679186701774597, + 0.1950620412826538, + 0.058617789298295975, + 0.04781375452876091, + 0.17189382016658783, + 0.2438727766275406 + ], + "credit_map": [ + 0.005490655079483986, + 0.002416212810203433, + 0.004793067462742329, + 0.0032477877102792263, + 0.006148532498627901, + 0.006008541211485863, + 0.003521759295836091, + 0.004996211268007755, + 0.0026669332291930914, + 0.0009308939334005117, + 0.0023160255514085293, + 0.0014890491729602218, + 0.003970146644860506, + 0.005589901935309172, + 0.0026268826331943274, + 0.003741410095244646, + 0.010302099399268627, + 0.002745014848187566, + 0.007222283165901899, + 0.003799830097705126, + 0.06067105010151863, + 0.07013428956270218, + 0.005868362728506327, + 0.013414707966148853, + 0.003511585760861635, + 0.002397568430751562, + 0.0025968668051064014, + 0.0020802097860723734, + 0.005331391468644142, + 0.00575351994484663, + 0.004917745944112539, + 0.003360063536092639, + 0.0021887982729822397, + 0.0011959642870351672, + 0.0037980289198458195, + 0.0019169445149600506, + 0.0023549345787614584, + 0.0029488089494407177, + 0.0020675058476626873, + 0.0022082007490098476, + 0.02636890299618244, + 0.007279532961547375, + 0.019379623234272003, + 0.011023145169019699, + 0.031874291598796844, + 0.04351815581321716, + 0.018167907372117043, + 0.037450484931468964, + 0.009860790334641933, + 0.0017127544851973653, + 0.0051317098550498486, + 0.0022273792419582605, + 0.006645336747169495, + 0.012795702554285526, + 0.0030362363904714584, + 0.017207879573106766, + 0.004758047871291637, + 0.004318139515817165, + 0.005303000565618277, + 0.0036516410764306784, + 0.012783020734786987, + 0.008133403025567532, + 0.004440509248524904, + 0.004425994120538235, + 0.016614370048046112, + 0.02187882736325264, + 0.014500802382826805, + 0.020200112834572792, + 0.035991981625556946, + 0.029005635529756546, + 0.015244992449879646, + 0.01845710538327694, + 0.013868549838662148, + 0.0300095584243536, + 0.01666967384517193, + 0.015003252774477005, + 0.054632000625133514, + 0.07727742195129395, + 0.013813900761306286, + 0.022598447278141975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.0011804286623373628 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot frames a lone candlelight against a deep, blurred black background, creating a scene of profound intimacy. Initially, the flame burns with a steady, teardrop-shaped glow, the main body of the flame a vibrant yellow-orange, gradually fading to a soft blue, licking the blackened wick. The process of extinguishing begins gracefully, the flame beginning to wither like a dying flower; its strong, bright edges soften and curl inward, while its tip teeters and droops downward. The pulses of light it emits are erratic, weakening with each flicker. With a final spasm, the flame collapses in on itself and disappears entirely, plunging the scene into near-total darkness. A tiny, glowing red spur on the wick is the final trace of light, before a thin, ethereal puff of white smoke rises deliberately, twisting and unfurling, melting into the void, leaving behind a palpable sense of quiet and finality.", + "seed": 16007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_7/index_8865_seed_16007_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01619045315310359, + "vlm_logprob": -0.01619045315310359, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.01619045315310359, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06110459566116333 + ], + "opposite_logprob_tokens": [ + -3.3111045360565186 + ], + "logprob_reward": -0.06110459566116333, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06110459566116333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11835619062185287, + 0.0793754905462265, + 0.059129729866981506, + 0.20207048952579498, + 0.08276426047086716, + 0.2054881602525711, + 0.034525107592344284, + 0.08956732600927353, + 0.08473914116621017, + 0.04398408532142639 + ], + "credit_map": [ + 0.007519954815506935, + 0.021582631394267082, + 0.024419181048870087, + 0.015802960842847824, + 0.012773377820849419, + 0.01772412657737732, + 0.007802322041243315, + 0.010731635615229607, + 0.006509167607873678, + 0.011779548600316048, + 0.022412966936826706, + 0.010579565539956093, + 0.0038897430058568716, + 0.008891279809176922, + 0.00413883663713932, + 0.011174390092492104, + 0.0039032690692692995, + 0.0027912319637835026, + 0.014113327488303185, + 0.004170708823949099, + 0.0024330944288522005, + 0.020818715915083885, + 0.0035758402664214373, + 0.00732354586943984, + 0.021227248013019562, + 0.012597762979567051, + 0.038708727806806564, + 0.010124946013092995, + 0.005731130018830299, + 0.06738737225532532, + 0.006998961791396141, + 0.03929433599114418, + 0.007128710858523846, + 0.009266570210456848, + 0.016373464837670326, + 0.012960294261574745, + 0.009700325317680836, + 0.014099032618105412, + 0.004555782303214073, + 0.008680080994963646, + 0.00872747227549553, + 0.02923309989273548, + 0.021594885736703873, + 0.009566298685967922, + 0.018375152722001076, + 0.07628283649682999, + 0.006461605429649353, + 0.03524680808186531, + 0.0030649530235677958, + 0.00430724024772644, + 0.005308975465595722, + 0.006370335351675749, + 0.004597397521138191, + 0.0032410023268312216, + 0.003928281366825104, + 0.003706925315782428, + 0.004359882324934006, + 0.011275522410869598, + 0.02839835174381733, + 0.0047347042709589005, + 0.0056387740187346935, + 0.01344504114240408, + 0.0038603132124990225, + 0.017854731529951096, + 0.005893129389733076, + 0.010907934978604317, + 0.013601970858871937, + 0.006638139020651579, + 0.020853042602539062, + 0.013530975207686424, + 0.007045703008770943, + 0.0062682428397238255, + 0.002708151238039136, + 0.0048159826546907425, + 0.007579901721328497, + 0.004892219323664904, + 0.010357901453971863, + 0.005564842373132706, + 0.004482195246964693, + 0.003582888515666127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.06110459566116333 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807731956243515 + ], + "opposite_logprob_tokens": [ + -4.634807586669922 + ], + "logprob_reward": -0.009807731956243515, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807731956243515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08410376310348511, + 0.30416667461395264, + 0.04114308953285217, + 0.022955210879445076, + 0.08408407121896744, + 0.09616661816835403, + 0.04195025935769081, + 0.13892287015914917, + 0.14538179337978363, + 0.04112563654780388 + ], + "credit_map": [ + 0.006361529231071472, + 0.01779058203101158, + 0.017446206882596016, + 0.00814980547875166, + 0.005703468341380358, + 0.01859649457037449, + 0.004316089674830437, + 0.005739591084420681, + 0.030496032908558846, + 0.043540000915527344, + 0.11940222978591919, + 0.020159929990768433, + 0.007819757796823978, + 0.038692656904459, + 0.009340072050690651, + 0.03471601381897926, + 0.0033402261324226856, + 0.0017449199222028255, + 0.011662041768431664, + 0.0019163097022101283, + 0.0009404392330907285, + 0.01593870483338833, + 0.0015948923537507653, + 0.00400555832311511, + 0.002003347733989358, + 0.0018210429698228836, + 0.004767180420458317, + 0.0026233764365315437, + 0.0012308511650189757, + 0.005351103376597166, + 0.0014499323442578316, + 0.003708377480506897, + 0.005103596951812506, + 0.008469744585454464, + 0.017819782719016075, + 0.010591015219688416, + 0.004879408050328493, + 0.027254145592451096, + 0.003033063840121031, + 0.006933320313692093, + 0.0058497656136751175, + 0.015165874734520912, + 0.01173896249383688, + 0.016946088522672653, + 0.011578712612390518, + 0.014121203683316708, + 0.011701405979692936, + 0.009064609184861183, + 0.0037120431661605835, + 0.004929300397634506, + 0.005313914734870195, + 0.004622075706720352, + 0.005496661644428968, + 0.0060063498094677925, + 0.00528195733204484, + 0.00658795889467001, + 0.006891623139381409, + 0.020430350676178932, + 0.05462470278143883, + 0.005889496300369501, + 0.005517599638551474, + 0.01117048878222704, + 0.003256858792155981, + 0.03114176355302334, + 0.00911707617342472, + 0.018605679273605347, + 0.01988072134554386, + 0.010966337285935879, + 0.0360671766102314, + 0.0277237631380558, + 0.011484362185001373, + 0.011536695994436741, + 0.002535581588745117, + 0.0048175835981965065, + 0.0066653648391366005, + 0.003944054711610079, + 0.009406638331711292, + 0.006532704923301935, + 0.004322622437030077, + 0.0029010875150561333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.009807731956243515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649898037314415 + ], + "opposite_logprob_tokens": [ + -5.379650115966797 + ], + "logprob_reward": -0.004649898037314415, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649898037314415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5105072259902954, + 0.13280338048934937, + 0.028843097388744354, + 0.035860396921634674, + 0.04649043455719948, + 0.043416135013103485, + 0.0501420795917511, + 0.07299324125051498, + 0.05925510451197624, + 0.01968889683485031 + ], + "credit_map": [ + 0.01892714947462082, + 0.021766915917396545, + 0.20845162868499756, + 0.016286039724946022, + 0.01923610456287861, + 0.1986115574836731, + 0.008538219146430492, + 0.018689647316932678, + 0.009401229210197926, + 0.018529634922742844, + 0.04607108235359192, + 0.009538140147924423, + 0.007792232558131218, + 0.014639118686318398, + 0.0056660473346710205, + 0.02116590179502964, + 0.002546068746596575, + 0.002303287386894226, + 0.007887770421802998, + 0.002832392929121852, + 0.0017494423082098365, + 0.005007466301321983, + 0.0031086751259863377, + 0.003407994983717799, + 0.006198799703270197, + 0.0032111071050167084, + 0.0069733173586428165, + 0.003817436983808875, + 0.0022410047240555286, + 0.004868678282946348, + 0.0022601368837058544, + 0.006289920303970575, + 0.0029961937107145786, + 0.005487928166985512, + 0.0062036896124482155, + 0.006442403886467218, + 0.006263308692723513, + 0.010131077840924263, + 0.0031191811431199312, + 0.0058466545306146145, + 0.0021577656734734774, + 0.004647610243409872, + 0.006028453819453716, + 0.003988027106970549, + 0.006067356094717979, + 0.01315339095890522, + 0.0018740606028586626, + 0.005499470978975296, + 0.004153149668127298, + 0.006519855000078678, + 0.007000821642577648, + 0.006910861004143953, + 0.007447536569088697, + 0.005640202667564154, + 0.006644084583967924, + 0.0058255731128156185, + 0.0064152018167078495, + 0.011945109814405441, + 0.006994197145104408, + 0.006337851285934448, + 0.009583881124854088, + 0.015277407132089138, + 0.006453133188188076, + 0.009986459277570248, + 0.004862708505243063, + 0.007568264380097389, + 0.007844672538340092, + 0.005010239779949188, + 0.014226879924535751, + 0.00845624040812254, + 0.005620754789561033, + 0.005665348377078772, + 0.0013823187910020351, + 0.002416706643998623, + 0.0029579750262200832, + 0.0020312608685344458, + 0.0031144714448601007, + 0.004304931033402681, + 0.0015591697301715612, + 0.0019220643443986773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004649898037314415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031929248943924904 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031929248943924904, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031929248943924904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.400449275970459, + 0.05094972997903824, + 0.009028790518641472, + 0.008524206466972828, + 0.024514850229024887, + 0.02646813541650772, + 0.1065378189086914, + 0.11424154788255692, + 0.2001114934682846, + 0.05917411297559738 + ], + "credit_map": [ + 0.014022153802216053, + 0.009573887102305889, + 0.16701555252075195, + 0.009541787207126617, + 0.010763486847281456, + 0.16785405576229095, + 0.005434331018477678, + 0.016244037076830864, + 0.0038981889374554157, + 0.007433711551129818, + 0.015198287554085255, + 0.005679656285792589, + 0.0030757635831832886, + 0.0066919587552547455, + 0.0026207827031612396, + 0.006351384799927473, + 0.0008515733061358333, + 0.0006940083112567663, + 0.0018416530219838023, + 0.001072241342626512, + 0.0004423937643878162, + 0.001405046903528273, + 0.0008112768409773707, + 0.0019105973187834024, + 0.0006399899139069021, + 0.0006250916630961001, + 0.000978120369836688, + 0.0009097190923057497, + 0.0005584366153925657, + 0.0024223204236477613, + 0.00039895385270938277, + 0.0019915748853236437, + 0.0019319156417623162, + 0.0034906058572232723, + 0.004083708394318819, + 0.004050097893923521, + 0.002828834345564246, + 0.004411511588841677, + 0.0013811825774610043, + 0.0023369938135147095, + 0.001846727216616273, + 0.0029950966127216816, + 0.003151085926219821, + 0.004288792610168457, + 0.0038208565674722195, + 0.0036137031856924295, + 0.0033694461453706026, + 0.0033824308775365353, + 0.01006657537072897, + 0.01381669007241726, + 0.013415089808404446, + 0.018390625715255737, + 0.014412968419492245, + 0.012804115191102028, + 0.010742515325546265, + 0.012889246456325054, + 0.006403150502592325, + 0.01642432063817978, + 0.011993305757641792, + 0.0073499521240592, + 0.012876019813120365, + 0.034986965358257294, + 0.007951967418193817, + 0.016255872324109077, + 0.012622362934052944, + 0.025658583268523216, + 0.02832987532019615, + 0.016019603237509727, + 0.04506297782063484, + 0.04223410785198212, + 0.01487581804394722, + 0.015308191999793053, + 0.003550122259184718, + 0.007614451926201582, + 0.0080329068005085, + 0.003955858759582043, + 0.01277599111199379, + 0.01445950660854578, + 0.004283537156879902, + 0.0045017460361123085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0031929248943924904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep and blurred black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flame start as a steady, teardrop-shaped glow with a yellow-orange body fading to blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame eventually collapse and disappear, leaving only a tiny glowing red spur on the wick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a thin, ethereal puff of white smoke rise, twist, and melt into the void after the flame disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021971152164041996 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021971152164041996, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021971152164041996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3728369474411011, + 0.016659867018461227, + 0.0316617451608181, + 0.01938113011419773, + 0.018044019117951393, + 0.04000068083405495, + 0.09096332639455795, + 0.15640099346637726, + 0.19641529023647308, + 0.057636041194200516 + ], + "credit_map": [ + 0.013162166811525822, + 0.009284857660531998, + 0.1594620943069458, + 0.008037365972995758, + 0.008334727957844734, + 0.1567211151123047, + 0.004376579076051712, + 0.013458041474223137, + 0.001864847494289279, + 0.001761720865033567, + 0.0029877263586968184, + 0.003036782843992114, + 0.0015144452918320894, + 0.0014888972509652376, + 0.0013121827505528927, + 0.002693264279514551, + 0.0022881284821778536, + 0.0013710686471313238, + 0.006906153168529272, + 0.001524419873021543, + 0.0007330395746976137, + 0.01184416189789772, + 0.0009363586432300508, + 0.006058415398001671, + 0.002840905450284481, + 0.0013941468205302954, + 0.0034341781865805387, + 0.0021329170558601618, + 0.001071062171831727, + 0.004277435131371021, + 0.0011975147062912583, + 0.003032970940694213, + 0.001241697114892304, + 0.0017753125866875052, + 0.003208371577784419, + 0.003779553808271885, + 0.0022748408373445272, + 0.002624566899612546, + 0.0011076421942561865, + 0.0020320340991020203, + 0.0026502443943172693, + 0.004525875207036734, + 0.005314907059073448, + 0.003534953808411956, + 0.005721833556890488, + 0.010206588543951511, + 0.0026141214184463024, + 0.005432158708572388, + 0.007017016876488924, + 0.012738347984850407, + 0.011580183170735836, + 0.012612583115696907, + 0.012353762984275818, + 0.01152077317237854, + 0.012333915568888187, + 0.010806750506162643, + 0.00835154578089714, + 0.022810988128185272, + 0.06179540976881981, + 0.00531373918056488, + 0.005058718379586935, + 0.01546757947653532, + 0.0029558944515883923, + 0.03464711084961891, + 0.013573178090155125, + 0.024928351864218712, + 0.02829776145517826, + 0.014646206982433796, + 0.043735016137361526, + 0.03965996950864792, + 0.016321469098329544, + 0.015253349207341671, + 0.0038591441698372364, + 0.0066728852689266205, + 0.006564501207321882, + 0.006534787826240063, + 0.010908822529017925, + 0.01321981381624937, + 0.005129230674356222, + 0.004746858961880207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0021971152164041996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_8/index_tensor([7361, 7361])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_8/index_tensor([7361, 7361])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8d2ba88a3690d4437f727b6c705b47581a0c8f2b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_8/index_tensor([7361, 7361])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Viewed from a static, slightly high-angle close-up, the scene focuses intensely on the patients' faces as they regain consciousness. Their eyelids, thin and translucent, flutter with a series of rapid, involuntary tremors before slowly, heavily peeling back. The exposed irises are a hazy, unfocused blue that darts aimlessly under the cool, sterile glare of the overhead fluorescent lights, casting soft shadows across their pale features. They lie on starched white hospital pillows, a few wisps of damp light brown hair clinging to their foreheads, their expressions one of pure disorientation. The fabric of their pale blue, wrinkled hospital gowns tightens as their chests begin to slowly and unsteadily rise, culminating in a low, trembling, audible breath that visibly swirls through their bodies, creating a powerful and intimate atmosphere of vulnerability and bewildered re-entry into the world.", + "seed": 5677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_8/index_7361_seed_5677_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1685162384994328, + "vlm_logprob": -0.1685162384994328, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.1685162384994328, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19348439574241638 + ], + "opposite_logprob_tokens": [ + -1.8184844255447388 + ], + "logprob_reward": -0.19348439574241638, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19348439574241638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18325789272785187, + 0.17607621848583221, + 0.0860576331615448, + 0.09747140854597092, + 0.08162576705217361, + 0.07213401049375534, + 0.04857154190540314, + 0.05141063034534454, + 0.07163771241903305, + 0.13175716996192932 + ], + "credit_map": [ + 0.026741940528154373, + 0.012849213555455208, + 0.045751217752695084, + 0.02309085801243782, + 0.02240455336868763, + 0.014894149266183376, + 0.01907052844762802, + 0.018455417826771736, + 0.022196438163518906, + 0.015374777838587761, + 0.040358614176511765, + 0.023569880053400993, + 0.02192857302725315, + 0.011294150725007057, + 0.018582047894597054, + 0.022771727293729782, + 0.010405778884887695, + 0.0068054902367293835, + 0.01894879713654518, + 0.009393845684826374, + 0.013922031037509441, + 0.0056540644727647305, + 0.010435996577143669, + 0.010491630993783474, + 0.009764784947037697, + 0.006583427544683218, + 0.01816628687083721, + 0.018945250660181046, + 0.008828612044453621, + 0.009694715030491352, + 0.010228238068521023, + 0.01526008453220129, + 0.00984554085880518, + 0.005591029301285744, + 0.010628203861415386, + 0.02349216118454933, + 0.008781107142567635, + 0.009475688450038433, + 0.00711922999471426, + 0.006692800205200911, + 0.011739665642380714, + 0.005558199714869261, + 0.01202089712023735, + 0.007788844406604767, + 0.009597528725862503, + 0.010053492151200771, + 0.007134620565921068, + 0.008240760304033756, + 0.006671084091067314, + 0.0033857200760394335, + 0.008177157491445541, + 0.007274669129401445, + 0.00625901622697711, + 0.004378844518214464, + 0.005567662883549929, + 0.006857383996248245, + 0.006817677989602089, + 0.004776249639689922, + 0.008643160574138165, + 0.008147334679961205, + 0.008026094175875187, + 0.0037849938962608576, + 0.005029698833823204, + 0.006185415666550398, + 0.009561441838741302, + 0.0059403833001852036, + 0.00994555838406086, + 0.016613809391856194, + 0.006237885914742947, + 0.004584508016705513, + 0.0095349857583642, + 0.009219137020409107, + 0.021859217435121536, + 0.00989479385316372, + 0.01760312169790268, + 0.025444788858294487, + 0.012574398890137672, + 0.014576750807464123, + 0.014173462986946106, + 0.01563062146306038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.19348439574241638 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42898914217948914 + ], + "opposite_logprob_tokens": [ + -1.0539891719818115 + ], + "logprob_reward": -0.42898914217948914, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42898914217948914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22924144566059113, + 0.13162852823734283, + 0.11550363153219223, + 0.10956285893917084, + 0.052374567836523056, + 0.06551294028759003, + 0.053225453943014145, + 0.15010762214660645, + 0.06078309938311577, + 0.03205980733036995 + ], + "credit_map": [ + 0.016696643084287643, + 0.023234130814671516, + 0.027204008772969246, + 0.02615538239479065, + 0.05945001542568207, + 0.02158580720424652, + 0.03285359591245651, + 0.02206186018884182, + 0.015660075470805168, + 0.008251125924289227, + 0.01726473495364189, + 0.012507203966379166, + 0.033298201858997345, + 0.0156459528952837, + 0.015714989975094795, + 0.013286239467561245, + 0.01131629291921854, + 0.007725127507001162, + 0.02095215581357479, + 0.015012881718575954, + 0.018115174025297165, + 0.007110648322850466, + 0.013431617058813572, + 0.02183973230421543, + 0.011803038418293, + 0.006132299546152353, + 0.014904774725437164, + 0.014128024689853191, + 0.014377603307366371, + 0.007663257420063019, + 0.016448425129055977, + 0.024105437099933624, + 0.006840029265731573, + 0.004821722395718098, + 0.006919339299201965, + 0.0095518808811903, + 0.0055987355299293995, + 0.006700414232909679, + 0.005822682753205299, + 0.006119763944298029, + 0.007091113366186619, + 0.004668672569096088, + 0.007341975811868906, + 0.013428150676190853, + 0.009048892185091972, + 0.0059307594783604145, + 0.008444134145975113, + 0.009559247642755508, + 0.005553317256271839, + 0.005264113657176495, + 0.009271487593650818, + 0.007226886227726936, + 0.007831896655261517, + 0.004848208744078875, + 0.0061773755587637424, + 0.007052167784422636, + 0.02228861302137375, + 0.013724950142204762, + 0.016082264482975006, + 0.035160988569259644, + 0.020086172968149185, + 0.010849082842469215, + 0.016371779143810272, + 0.015543780289590359, + 0.006479828618466854, + 0.0055048163048923016, + 0.008236304856836796, + 0.008490313775837421, + 0.012160972692072392, + 0.005891092587262392, + 0.007285809610038996, + 0.006733962334692478, + 0.005118307191878557, + 0.002785870572552085, + 0.00433527072891593, + 0.005616957787424326, + 0.00400922866538167, + 0.0033008363097906113, + 0.003812291193753481, + 0.003081043716520071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.42898914217948914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036897659301758 + ], + "opposite_logprob_tokens": [ + -2.3503689765930176 + ], + "logprob_reward": -0.10036897659301758, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036897659301758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31642207503318787, + 0.12854337692260742, + 0.17087297141551971, + 0.11063671112060547, + 0.0333159938454628, + 0.05148499459028244, + 0.04448454827070236, + 0.045520007610321045, + 0.03594941273331642, + 0.06276994198560715 + ], + "credit_map": [ + 0.03269942104816437, + 0.02385122701525688, + 0.0497998408973217, + 0.05313936248421669, + 0.04460190609097481, + 0.022294729948043823, + 0.0383920855820179, + 0.0516434870660305, + 0.015604604035615921, + 0.012132300063967705, + 0.012499485164880753, + 0.01830451190471649, + 0.035547226667404175, + 0.005983950570225716, + 0.017606234177947044, + 0.010865068063139915, + 0.017464356496930122, + 0.012219320051372051, + 0.0379674956202507, + 0.020403200760483742, + 0.017028402537107468, + 0.007967684417963028, + 0.021932844072580338, + 0.03588966280221939, + 0.014329267665743828, + 0.008248860016465187, + 0.016747429966926575, + 0.01437720749527216, + 0.013584868982434273, + 0.012469674460589886, + 0.011692957952618599, + 0.019186438992619514, + 0.004216527566313744, + 0.0030156809370964766, + 0.004690040368586779, + 0.0035843905061483383, + 0.005673554725944996, + 0.004021391738206148, + 0.003586218226701021, + 0.004528190474957228, + 0.005930060055106878, + 0.0037071758415549994, + 0.00803607702255249, + 0.006972744595259428, + 0.007283938117325306, + 0.005015840753912926, + 0.006733653601258993, + 0.007805505767464638, + 0.004190842155367136, + 0.0028052199631929398, + 0.0067875999957323074, + 0.005567831452935934, + 0.0066931359469890594, + 0.004731378983706236, + 0.005527215544134378, + 0.008181323297321796, + 0.005756932310760021, + 0.0034424022305756807, + 0.006540585774928331, + 0.005712342448532581, + 0.008795756846666336, + 0.0020919484086334705, + 0.007828990928828716, + 0.005351049359887838, + 0.004396562930196524, + 0.003057647030800581, + 0.005394638981670141, + 0.004779642913490534, + 0.006151694804430008, + 0.0037338894326239824, + 0.0040848166681826115, + 0.004350520204752684, + 0.010015632957220078, + 0.004987466149032116, + 0.007559335790574551, + 0.016090506687760353, + 0.005123431794345379, + 0.00397838419303298, + 0.007570337969809771, + 0.007444845046848059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.10036897659301758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300059407949448 + ], + "opposite_logprob_tokens": [ + -2.2380006313323975 + ], + "logprob_reward": -0.11300059407949448, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300059407949448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24115711450576782, + 0.11356818675994873, + 0.11522993445396423, + 0.1272718459367752, + 0.0936918780207634, + 0.02771947532892227, + 0.07424555718898773, + 0.046694111078977585, + 0.05048852041363716, + 0.10993339866399765 + ], + "credit_map": [ + 0.028081543743610382, + 0.01939890719950199, + 0.03905868157744408, + 0.04913268983364105, + 0.028591029345989227, + 0.027405789121985435, + 0.02457754872739315, + 0.024910928681492805, + 0.014963292516767979, + 0.008812421932816505, + 0.01615188829600811, + 0.023933539167046547, + 0.017099468037486076, + 0.008523370139300823, + 0.013899181969463825, + 0.010185022838413715, + 0.014218268916010857, + 0.009874065406620502, + 0.016933072358369827, + 0.012334300205111504, + 0.016708407551050186, + 0.01425380352884531, + 0.01753167062997818, + 0.013376343064010143, + 0.016921797767281532, + 0.00928191002458334, + 0.02022295631468296, + 0.0159671101719141, + 0.015508398413658142, + 0.0159207284450531, + 0.01221246924251318, + 0.021236475557088852, + 0.014958291314542294, + 0.008394911885261536, + 0.009672381915152073, + 0.026152152568101883, + 0.006776249036192894, + 0.009386155754327774, + 0.00885129626840353, + 0.009500445798039436, + 0.004835978150367737, + 0.0022472478449344635, + 0.0038421067874878645, + 0.004352938383817673, + 0.0026864048559218645, + 0.003473312593996525, + 0.0027640960179269314, + 0.003517390927299857, + 0.00826037023216486, + 0.005846821237355471, + 0.009455904364585876, + 0.010268711484968662, + 0.008725586347281933, + 0.008750101551413536, + 0.009609735570847988, + 0.013328327797353268, + 0.0057943821884691715, + 0.0038339663296937943, + 0.004991591442376375, + 0.007640220690518618, + 0.0069071706384420395, + 0.005604899488389492, + 0.006040583364665508, + 0.005881296470761299, + 0.007472699508070946, + 0.0038766171783208847, + 0.007014427334070206, + 0.009902972728013992, + 0.006935986690223217, + 0.004242502152919769, + 0.005917373578995466, + 0.0051259445026516914, + 0.015597239136695862, + 0.009286867454648018, + 0.01373918354511261, + 0.022567708045244217, + 0.013526581227779388, + 0.012521202675998211, + 0.010718881152570248, + 0.011975731700658798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.11300059407949448 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738083902746439 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738083902746439, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738083902746439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28568586707115173, + 0.08826673030853271, + 0.11594665795564651, + 0.06151207908987999, + 0.07839755713939667, + 0.04713144153356552, + 0.07497488707304001, + 0.13793522119522095, + 0.07275960594415665, + 0.03738997504115105 + ], + "credit_map": [ + 0.03463783487677574, + 0.021013036370277405, + 0.04036720097064972, + 0.06425877660512924, + 0.03615676239132881, + 0.026194175705313683, + 0.027957657352089882, + 0.03510041534900665, + 0.007968224585056305, + 0.0067869434133172035, + 0.01683237962424755, + 0.010490858927369118, + 0.013041701167821884, + 0.0055446685291826725, + 0.01076576579362154, + 0.016836179420351982, + 0.00998013187199831, + 0.006323688663542271, + 0.021459335461258888, + 0.017590397968888283, + 0.015542414970695972, + 0.005854043178260326, + 0.014474592171609402, + 0.02472204715013504, + 0.0070426915772259235, + 0.004604344721883535, + 0.006982280407100916, + 0.011759283021092415, + 0.009489484131336212, + 0.004575321916490793, + 0.009259156882762909, + 0.007799510378390551, + 0.008497309871017933, + 0.005582406651228666, + 0.014448142610490322, + 0.015829551964998245, + 0.007007932756096125, + 0.007651899941265583, + 0.00885581225156784, + 0.010524498298764229, + 0.006586211733520031, + 0.004257326014339924, + 0.005712609272450209, + 0.008693895302712917, + 0.005283606704324484, + 0.005171871744096279, + 0.005309682339429855, + 0.006116233300417662, + 0.009119045920670033, + 0.009062650613486767, + 0.008383520878851414, + 0.01219120528548956, + 0.01042888779193163, + 0.006787460297346115, + 0.009945513680577278, + 0.009056593291461468, + 0.01920025609433651, + 0.013056631200015545, + 0.01419790182262659, + 0.03344794362783432, + 0.017693795263767242, + 0.012203012593090534, + 0.012353260070085526, + 0.015782417729496956, + 0.009508620947599411, + 0.006990436464548111, + 0.009265580214560032, + 0.014493953436613083, + 0.01132143009454012, + 0.005786821246147156, + 0.007565837353467941, + 0.007826922461390495, + 0.005331438034772873, + 0.0032957291696220636, + 0.004203710705041885, + 0.009820879437029362, + 0.004074976779520512, + 0.0029273892287164927, + 0.0034731884952634573, + 0.004262663424015045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.006738083902746439 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a static, slightly high-angle close-up, the scene focuses intensely on the patients' faces as they regain consciousness. Their eyelids, thin and translucent, flutter with a series of rapid, involuntary tremors before slowly, heavily peeling back. The exposed irises are a hazy, unfocused blue that darts aimlessly under the cool, sterile glare of the overhead fluorescent lights, casting soft shadows across their pale features. They lie on starched white hospital pillows, a few wisps of damp light brown hair clinging to their foreheads, their expressions one of pure disorientation. The fabric of their pale blue, wrinkled hospital gowns tightens as their chests begin to slowly and unsteadily rise, culminating in a low, trembling, audible breath that visibly swirls through their bodies, creating a powerful and intimate atmosphere of vulnerability and bewildered re-entry into the world.", + "seed": 5678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_8/index_7361_seed_5678_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18634614273905753, + "vlm_logprob": -0.18634614273905753, + "vlm_yes_no_margin": 2.250000011920929, + "vlm_token_credit": -0.18634614273905753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22344134747982025 + ], + "opposite_logprob_tokens": [ + -1.7234413623809814 + ], + "logprob_reward": -0.22344134747982025, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22344134747982025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09786621481180191, + 0.06327929347753525, + 0.07634010165929794, + 0.17564894258975983, + 0.31972837448120117, + 0.09467221796512604, + 0.07275764644145966, + 0.03540917858481407, + 0.0222021397203207, + 0.04209589585661888 + ], + "credit_map": [ + 0.00724913040176034, + 0.010652638971805573, + 0.011868334375321865, + 0.01159236766397953, + 0.012691876851022243, + 0.014584793709218502, + 0.010900544933974743, + 0.018326535820961, + 0.0052408017218112946, + 0.007769727148115635, + 0.007474011275917292, + 0.0060905893333256245, + 0.012569455429911613, + 0.009454848244786263, + 0.00778314471244812, + 0.0068967146798968315, + 0.004739629104733467, + 0.00938339065760374, + 0.01635972410440445, + 0.010329223237931728, + 0.008159879595041275, + 0.010533511638641357, + 0.010361170396208763, + 0.0064735715277493, + 0.01039077527821064, + 0.01125731598585844, + 0.058330029249191284, + 0.017276687547564507, + 0.013288216665387154, + 0.040832776576280594, + 0.011628609150648117, + 0.01264453399926424, + 0.044017013162374496, + 0.027043266221880913, + 0.08937422931194305, + 0.03364444896578789, + 0.027585307136178017, + 0.044721875339746475, + 0.023109449073672295, + 0.030232800170779228, + 0.008885666728019714, + 0.010871185921132565, + 0.021145006641745567, + 0.009974749758839607, + 0.010065412148833275, + 0.014370826072990894, + 0.010010846890509129, + 0.009348521940410137, + 0.010327176190912724, + 0.009547225199639797, + 0.010047133080661297, + 0.007446258328855038, + 0.008981899358332157, + 0.010081939399242401, + 0.006971040740609169, + 0.009354970417916775, + 0.0052489023655653, + 0.0034550009295344353, + 0.004207781050354242, + 0.0037524893414229155, + 0.003280664561316371, + 0.005597698036581278, + 0.004212314262986183, + 0.005654328968375921, + 0.003128441981971264, + 0.003000137396156788, + 0.0031070695258677006, + 0.0029023292008787394, + 0.002631210954859853, + 0.002743588760495186, + 0.002454219153150916, + 0.002235144143924117, + 0.001958355773240328, + 0.0022729465272277594, + 0.002469566883519292, + 0.0024483439046889544, + 0.015404917299747467, + 0.013041757978498936, + 0.0021386046428233385, + 0.002361403778195381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.22344134747982025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38718098402023315 + ], + "opposite_logprob_tokens": [ + -1.137181043624878 + ], + "logprob_reward": -0.38718098402023315, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38718098402023315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16321484744548798, + 0.07243616133928299, + 0.15782251954078674, + 0.272886723279953, + 0.07622003555297852, + 0.052944615483284, + 0.025934526696801186, + 0.035272013396024704, + 0.03797735273838043, + 0.10529119521379471 + ], + "credit_map": [ + 0.016636190935969353, + 0.016239553689956665, + 0.02657359652221203, + 0.015941543504595757, + 0.018202753737568855, + 0.027998894453048706, + 0.026869049295783043, + 0.014753276482224464, + 0.005500763189047575, + 0.006134221330285072, + 0.006675641983747482, + 0.007116544060409069, + 0.008712607435882092, + 0.011461304500699043, + 0.009067906998097897, + 0.017767177894711494, + 0.01071521732956171, + 0.006220080889761448, + 0.052632883191108704, + 0.009029421955347061, + 0.012131169438362122, + 0.05042317882180214, + 0.006780588999390602, + 0.009889987297356129, + 0.01678801141679287, + 0.011173399165272713, + 0.10472316294908524, + 0.018249033018946648, + 0.01479253638535738, + 0.08245373517274857, + 0.01104256883263588, + 0.013664303347468376, + 0.00684365164488554, + 0.006341471336781979, + 0.01492074504494667, + 0.0069202217273414135, + 0.009394816122949123, + 0.015973517671227455, + 0.006709026638418436, + 0.009116598404943943, + 0.006200732663273811, + 0.003808918409049511, + 0.00799624901264906, + 0.004916318692266941, + 0.007414547260850668, + 0.012794634327292442, + 0.0037064903881400824, + 0.006106734275817871, + 0.0030850900802761316, + 0.0023198137059807777, + 0.0033487463369965553, + 0.0018785824067890644, + 0.0029330432880669832, + 0.00371786137111485, + 0.003166354261338711, + 0.005485040135681629, + 0.007038223557174206, + 0.00437395554035902, + 0.00420818105340004, + 0.0030930666252970695, + 0.0036493500228971243, + 0.0045532649382948875, + 0.003445710288360715, + 0.00491026695817709, + 0.002913112984970212, + 0.004734291695058346, + 0.004553280305117369, + 0.005207907408475876, + 0.00483095133677125, + 0.004809329751878977, + 0.004898189101368189, + 0.006030295975506306, + 0.004692604299634695, + 0.004784189630299807, + 0.0030937364790588617, + 0.005350640043616295, + 0.04241160675883293, + 0.03467034548521042, + 0.005043566692620516, + 0.005244516767561436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.38718098402023315 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062102846801280975 + ], + "opposite_logprob_tokens": [ + -2.812102794647217 + ], + "logprob_reward": -0.062102846801280975, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062102846801280975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22846534848213196, + 0.14317673444747925, + 0.1143246591091156, + 0.0909559354186058, + 0.08427339047193527, + 0.04988763853907585, + 0.05024462193250656, + 0.04712647199630737, + 0.0168149434030056, + 0.17473024129867554 + ], + "credit_map": [ + 0.016343088820576668, + 0.02353154495358467, + 0.03130060434341431, + 0.026266921311616898, + 0.03667518123984337, + 0.042361173778772354, + 0.024339403957128525, + 0.02764742635190487, + 0.01070145983248949, + 0.015053384006023407, + 0.022295953705906868, + 0.015886355191469193, + 0.012581219896674156, + 0.02376454696059227, + 0.013572664931416512, + 0.029321158304810524, + 0.00897741224616766, + 0.007378504145890474, + 0.025448376312851906, + 0.00717489467933774, + 0.012093672528862953, + 0.02914794348180294, + 0.00754136499017477, + 0.01656249724328518, + 0.006696321535855532, + 0.009414334781467915, + 0.017558306455612183, + 0.009376640431582928, + 0.010080086998641491, + 0.021363141015172005, + 0.008509431034326553, + 0.007957665249705315, + 0.008424051105976105, + 0.0071578314527869225, + 0.012342814356088638, + 0.008005788549780846, + 0.008382037281990051, + 0.0224622655659914, + 0.00903528742492199, + 0.008463311940431595, + 0.009089328348636627, + 0.005553274415433407, + 0.007210415322333574, + 0.004534874111413956, + 0.00653563579544425, + 0.008437598124146461, + 0.004506781231611967, + 0.004019731655716896, + 0.010872461833059788, + 0.006981424055993557, + 0.005707588978111744, + 0.004622240085154772, + 0.005144844297319651, + 0.00711926631629467, + 0.004111680202186108, + 0.00568511663004756, + 0.007349975872784853, + 0.004468221217393875, + 0.00632260600104928, + 0.0037319795228540897, + 0.004809986799955368, + 0.006672152318060398, + 0.005131850950419903, + 0.00863969698548317, + 0.0014759630430489779, + 0.0019202334806323051, + 0.0023328838869929314, + 0.0018949914956465364, + 0.002056837547570467, + 0.002398327924311161, + 0.0020333859138190746, + 0.0027023199945688248, + 0.008186415769159794, + 0.006891866214573383, + 0.003797421930357814, + 0.007894209586083889, + 0.0741499587893486, + 0.05909867212176323, + 0.007315812632441521, + 0.0073958770371973515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.062102846801280975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25226330757141113 + ], + "opposite_logprob_tokens": [ + -1.5022633075714111 + ], + "logprob_reward": -0.25226330757141113, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25226330757141113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07237889617681503, + 0.0345633365213871, + 0.15165723860263824, + 0.5983412265777588, + 0.030782600864768028, + 0.0136234937235713, + 0.02132071740925312, + 0.026954198256134987, + 0.00792239885777235, + 0.04245588555932045 + ], + "credit_map": [ + 0.005918467417359352, + 0.004590225871652365, + 0.008307668380439281, + 0.006458400283008814, + 0.00866242591291666, + 0.014012079685926437, + 0.009863321669399738, + 0.01456630602478981, + 0.002465285360813141, + 0.00445302901789546, + 0.0048835002817213535, + 0.004985425155609846, + 0.004291267599910498, + 0.007083610165864229, + 0.003894107649102807, + 0.00250710966065526, + 0.010019225999712944, + 0.004233207553625107, + 0.05285915732383728, + 0.008478476665914059, + 0.012887641787528992, + 0.05217897146940231, + 0.005442777182906866, + 0.005557784345000982, + 0.03878394886851311, + 0.019621703773736954, + 0.2512983977794647, + 0.048003166913986206, + 0.024249697104096413, + 0.17231808602809906, + 0.018766120076179504, + 0.025300147011876106, + 0.002658422337844968, + 0.0022488029208034277, + 0.006139191798865795, + 0.0026503014378249645, + 0.003379221074283123, + 0.007788032293319702, + 0.0022274572402238846, + 0.0036911724600940943, + 0.0019057807512581348, + 0.0018491848604753613, + 0.0017799768829718232, + 0.001357278786599636, + 0.0018653753213584423, + 0.001888684812001884, + 0.00160735915414989, + 0.001369853038340807, + 0.0015716439811512828, + 0.002391248010098934, + 0.0023414769675582647, + 0.002124819904565811, + 0.002645009895786643, + 0.0028490922413766384, + 0.002842922927811742, + 0.004554502200335264, + 0.00500723859295249, + 0.0026710527017712593, + 0.003362303599715233, + 0.002029995433986187, + 0.0028349205385893583, + 0.0036429360043257475, + 0.0026829310227185488, + 0.004722819663584232, + 0.0006788722821511328, + 0.0010456598829478025, + 0.001162998378276825, + 0.0011207865318283439, + 0.0009240643703378737, + 0.0013824119232594967, + 0.0007038451731204987, + 0.0009037608979269862, + 0.0022686896845698357, + 0.0019889178220182657, + 0.0013525184476748109, + 0.002052080584689975, + 0.017194950953125954, + 0.013754058629274368, + 0.0017746618250384927, + 0.0020700066816061735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.25226330757141113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067422278225421906 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.0067422278225421906, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067422278225421906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2838737964630127, + 0.11110369861125946, + 0.05287258327007294, + 0.08909167349338531, + 0.037918027490377426, + 0.12611037492752075, + 0.06726669520139694, + 0.0499950647354126, + 0.023557182401418686, + 0.1582108736038208 + ], + "credit_map": [ + 0.018779445439577103, + 0.024207042530179024, + 0.05445689335465431, + 0.027704980224370956, + 0.03582822531461716, + 0.056078311055898666, + 0.027554690837860107, + 0.03926420956850052, + 0.009239939972758293, + 0.008470644243061543, + 0.02330600656569004, + 0.00915033370256424, + 0.013339533470571041, + 0.024219388142228127, + 0.009807277470827103, + 0.013570579700171947, + 0.005686830263584852, + 0.005748970899730921, + 0.009253923781216145, + 0.006993474904447794, + 0.00637405039742589, + 0.006800897419452667, + 0.005697615910321474, + 0.00631681876257062, + 0.0065871719270944595, + 0.010058666579425335, + 0.023443395271897316, + 0.012984267435967922, + 0.010724933817982674, + 0.008705534040927887, + 0.010038331151008606, + 0.006549373269081116, + 0.0038485899567604065, + 0.004587968345731497, + 0.006422131787985563, + 0.003668447956442833, + 0.005238266661763191, + 0.004788505844771862, + 0.004688495770096779, + 0.004675619769841433, + 0.012801319360733032, + 0.008196956478059292, + 0.0171683169901371, + 0.015104725956916809, + 0.015216290950775146, + 0.033911190927028656, + 0.010833748616278172, + 0.012877837754786015, + 0.01095207966864109, + 0.010731926187872887, + 0.007184503134340048, + 0.006960325874388218, + 0.008385741151869297, + 0.008431997150182724, + 0.006354274693876505, + 0.008265850134193897, + 0.006303963717073202, + 0.0056305211037397385, + 0.005055430810898542, + 0.004603879060596228, + 0.006495337001979351, + 0.010223980993032455, + 0.005373216699808836, + 0.006308736279606819, + 0.001757156103849411, + 0.002202835166826844, + 0.0029919303487986326, + 0.003075247397646308, + 0.003200679086148739, + 0.005556175485253334, + 0.0022577617783099413, + 0.0025153961032629013, + 0.014669371768832207, + 0.007860488258302212, + 0.007723903749138117, + 0.007661016192287207, + 0.06039995700120926, + 0.04375375807285309, + 0.007819618098437786, + 0.00832277163863182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0067422278225421906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a static, slightly high-angle close-up, the scene focuses intensely on the patients' faces as they regain consciousness. Their eyelids, thin and translucent, flutter with a series of rapid, involuntary tremors before slowly, heavily peeling back. The exposed irises are a hazy, unfocused blue that darts aimlessly under the cool, sterile glare of the overhead fluorescent lights, casting soft shadows across their pale features. They lie on starched white hospital pillows, a few wisps of damp light brown hair clinging to their foreheads, their expressions one of pure disorientation. The fabric of their pale blue, wrinkled hospital gowns tightens as their chests begin to slowly and unsteadily rise, culminating in a low, trembling, audible breath that visibly swirls through their bodies, creating a powerful and intimate atmosphere of vulnerability and bewildered re-entry into the world.", + "seed": 5679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_8/index_7361_seed_5679_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17833937518298626, + "vlm_logprob": -0.17833937518298626, + "vlm_yes_no_margin": 2.2749999880790712, + "vlm_token_credit": -0.17833937518298626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2759416401386261 + ], + "opposite_logprob_tokens": [ + -1.5259416103363037 + ], + "logprob_reward": -0.2759416401386261, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2759416401386261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18678879737854004, + 0.058634839951992035, + 0.0806560069322586, + 0.05859585478901863, + 0.13932634890079498, + 0.07580094784498215, + 0.0381411612033844, + 0.07860642671585083, + 0.037704940885305405, + 0.24574461579322815 + ], + "credit_map": [ + 0.0243093091994524, + 0.02105451375246048, + 0.017111260443925858, + 0.022590873762965202, + 0.02160237357020378, + 0.02954971045255661, + 0.023541131988167763, + 0.027029648423194885, + 0.008318133652210236, + 0.005060513038188219, + 0.005786202847957611, + 0.007712156046181917, + 0.006923720706254244, + 0.007965951226651669, + 0.009094228968024254, + 0.007773933932185173, + 0.006000912748277187, + 0.007501356769353151, + 0.00866986159235239, + 0.01045424398034811, + 0.009817153215408325, + 0.015453732572495937, + 0.01063157245516777, + 0.012127183377742767, + 0.004646181128919125, + 0.00466160336509347, + 0.00447392463684082, + 0.0062851374968886375, + 0.011237814091145992, + 0.010619853623211384, + 0.005598380230367184, + 0.011072964407503605, + 0.010416250675916672, + 0.01512986421585083, + 0.010228442959487438, + 0.01239470299333334, + 0.01859685592353344, + 0.032541561871767044, + 0.01599293015897274, + 0.024025753140449524, + 0.007153539452701807, + 0.007918711751699448, + 0.005609264597296715, + 0.008330085314810276, + 0.011470616795122623, + 0.013356704264879227, + 0.009131759405136108, + 0.012830282561480999, + 0.00288793514482677, + 0.00434150593355298, + 0.0021842799615114927, + 0.005062168929725885, + 0.00673547200858593, + 0.005488706286996603, + 0.006443014368414879, + 0.004998082760721445, + 0.006501622032374144, + 0.004638953134417534, + 0.006007219664752483, + 0.0056574540212750435, + 0.013344874605536461, + 0.01942557469010353, + 0.008013197220861912, + 0.015017539262771606, + 0.0022793239913880825, + 0.0020824300590902567, + 0.0016793388640508056, + 0.002577373292297125, + 0.013302193023264408, + 0.0090987179428339, + 0.002962814411148429, + 0.003722755005583167, + 0.022109700366854668, + 0.02123364992439747, + 0.02470472827553749, + 0.019782543182373047, + 0.03804890811443329, + 0.058404382318258286, + 0.026571888476610184, + 0.03488883748650551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.2759416401386261 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871278166770935 + ], + "opposite_logprob_tokens": [ + -1.1371278762817383 + ], + "logprob_reward": -0.3871278166770935, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3871278166770935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3307027518749237, + 0.1040114089846611, + 0.049677811563014984, + 0.06943085044622421, + 0.08796841651201248, + 0.04484182223677635, + 0.046453189104795456, + 0.07939775288105011, + 0.10004319995641708, + 0.08747278153896332 + ], + "credit_map": [ + 0.03083883970975876, + 0.03701586648821831, + 0.025011131539940834, + 0.035196367651224136, + 0.04686115309596062, + 0.06526754796504974, + 0.0441313162446022, + 0.04638052359223366, + 0.00942998193204403, + 0.012153169140219688, + 0.010237505659461021, + 0.017309391871094704, + 0.013279917649924755, + 0.015503489412367344, + 0.01108763087540865, + 0.015010319650173187, + 0.0044334460981190205, + 0.004222871270030737, + 0.0058007375337183475, + 0.0067713395692408085, + 0.005474673584103584, + 0.009756390005350113, + 0.00597852049395442, + 0.007239837199449539, + 0.005770891904830933, + 0.005153296049684286, + 0.006130974739789963, + 0.010625160299241543, + 0.0159467700868845, + 0.009226836264133453, + 0.0076277186162769794, + 0.008949199691414833, + 0.007612116169184446, + 0.009820453822612762, + 0.007605809718370438, + 0.0088064419105649, + 0.013058383949100971, + 0.018293770030140877, + 0.00937085971236229, + 0.013400576077401638, + 0.0037685048300772905, + 0.004646534100174904, + 0.004052549600601196, + 0.004970251582562923, + 0.008577157743275166, + 0.007032773923128843, + 0.004827494267374277, + 0.0069665564224123955, + 0.0035151236224919558, + 0.0049612196162343025, + 0.003514789743348956, + 0.006688629277050495, + 0.009258476085960865, + 0.0062047019600868225, + 0.007499666418880224, + 0.0048105791211128235, + 0.008074515499174595, + 0.00917387567460537, + 0.007042964454740286, + 0.010647748596966267, + 0.009350262582302094, + 0.015052341856062412, + 0.009273087605834007, + 0.010782964527606964, + 0.005214793607592583, + 0.0035613880027085543, + 0.0036139602307230234, + 0.004482266027480364, + 0.039732493460178375, + 0.030864624306559563, + 0.004918406717479229, + 0.007655269466340542, + 0.008258880116045475, + 0.00791183952242136, + 0.008695759810507298, + 0.010293561033904552, + 0.011790166608989239, + 0.019407322630286217, + 0.009546457789838314, + 0.011568796820938587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.3871278166770935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903372496366501 + ], + "opposite_logprob_tokens": [ + -2.579033613204956 + ], + "logprob_reward": -0.07903372496366501, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903372496366501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4216757118701935, + 0.1374741643667221, + 0.03986692428588867, + 0.06378156691789627, + 0.04383265599608421, + 0.06835394352674484, + 0.02930939570069313, + 0.03829996660351753, + 0.05459316074848175, + 0.10281245410442352 + ], + "credit_map": [ + 0.04831555485725403, + 0.0367748886346817, + 0.03988305851817131, + 0.04982084408402443, + 0.05282014608383179, + 0.08492163568735123, + 0.05274920165538788, + 0.05639040097594261, + 0.01900886744260788, + 0.014470028690993786, + 0.01201503537595272, + 0.013107331469655037, + 0.015019578859210014, + 0.025901786983013153, + 0.015005363151431084, + 0.02294616959989071, + 0.005496273282915354, + 0.0038250975776463747, + 0.004370271228253841, + 0.004332899581640959, + 0.005598914809525013, + 0.005568123888224363, + 0.005344677716493607, + 0.005330667365342379, + 0.005761560518294573, + 0.004571670200675726, + 0.0045617069117724895, + 0.004390845075249672, + 0.009699259884655476, + 0.01499908696860075, + 0.005433127749711275, + 0.014364311471581459, + 0.005671212915331125, + 0.0049938238225877285, + 0.003775298362597823, + 0.0047744205221533775, + 0.007470675744116306, + 0.006652656476944685, + 0.004755773115903139, + 0.005738795269280672, + 0.0077767642214894295, + 0.007184700574725866, + 0.006161481142044067, + 0.006642296444624662, + 0.01405409537255764, + 0.010916675440967083, + 0.005820704158395529, + 0.009797229431569576, + 0.002649449510499835, + 0.003986320458352566, + 0.0024786503054201603, + 0.004560449160635471, + 0.004214845132082701, + 0.0033180343452841043, + 0.004906661342829466, + 0.0031949831172823906, + 0.0041322228498756886, + 0.004176787566393614, + 0.0031473038252443075, + 0.0032529402524232864, + 0.004878861363977194, + 0.008803498931229115, + 0.004097362048923969, + 0.005810986738651991, + 0.003447747090831399, + 0.0029699315782636404, + 0.0021074768155813217, + 0.0030397134833037853, + 0.019696582108736038, + 0.015828154981136322, + 0.003299720585346222, + 0.0042038350366055965, + 0.01089177094399929, + 0.012178772129118443, + 0.008475687354803085, + 0.008351361379027367, + 0.015785688534379005, + 0.022416850551962852, + 0.012566357851028442, + 0.012145975604653358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07903372496366501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285501837730408 + ], + "opposite_logprob_tokens": [ + -2.017854928970337 + ], + "logprob_reward": -0.14285501837730408, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14285501837730408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21779823303222656, + 0.11440978199243546, + 0.05813177675008774, + 0.07250867038965225, + 0.10683921724557877, + 0.1128997653722763, + 0.03904193267226219, + 0.0391489639878273, + 0.1457994431257248, + 0.09342218935489655 + ], + "credit_map": [ + 0.029398901388049126, + 0.019152240827679634, + 0.022039541974663734, + 0.02491738088428974, + 0.026811450719833374, + 0.0397488959133625, + 0.0234245415776968, + 0.032305292785167694, + 0.0099545419216156, + 0.011135110631585121, + 0.009172312915325165, + 0.015263274312019348, + 0.013180150650441647, + 0.021161731332540512, + 0.015701724216341972, + 0.01884094439446926, + 0.0047537172213196754, + 0.005765800829976797, + 0.006048867478966713, + 0.009017015807330608, + 0.006724915932863951, + 0.010284588672220707, + 0.0067533571273088455, + 0.008783513680100441, + 0.006071292795240879, + 0.007818290963768959, + 0.004895175341516733, + 0.00693142507225275, + 0.010791217908263206, + 0.012142324820160866, + 0.005926723591983318, + 0.017932221293449402, + 0.01748531125485897, + 0.008404682390391827, + 0.01049557700753212, + 0.008165009319782257, + 0.013659259304404259, + 0.018460312858223915, + 0.009796010330319405, + 0.020373059436678886, + 0.008215983398258686, + 0.009084923192858696, + 0.008063912391662598, + 0.012195462360978127, + 0.02381611242890358, + 0.02117985486984253, + 0.017040401697158813, + 0.013303120620548725, + 0.00364646571688354, + 0.003040505573153496, + 0.003352491883561015, + 0.004108245484530926, + 0.005194997880607843, + 0.0070902579464018345, + 0.003831193782389164, + 0.008777773939073086, + 0.003246133215725422, + 0.0028181676752865314, + 0.00303070992231369, + 0.003068523248657584, + 0.005815980955958366, + 0.009693346917629242, + 0.004120789002627134, + 0.007355314213782549, + 0.005403650458902121, + 0.004356343299150467, + 0.0031984024681150913, + 0.0037138836923986673, + 0.06242511421442032, + 0.051616132259368896, + 0.0064871362410485744, + 0.008598778396844864, + 0.009104627184569836, + 0.010714084841310978, + 0.007296007592231035, + 0.006955650169402361, + 0.024309994652867317, + 0.017192138358950615, + 0.00718200346454978, + 0.010667680762708187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.14285501837730408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738675758242607 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738675758242607, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738675758242607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21439041197299957, + 0.11200876533985138, + 0.09107658267021179, + 0.11427085101604462, + 0.10124380141496658, + 0.05407406762242317, + 0.05444015562534332, + 0.09482935070991516, + 0.11467082798480988, + 0.04899515584111214 + ], + "credit_map": [ + 0.02240615151822567, + 0.022465582937002182, + 0.021388301625847816, + 0.027898622676730156, + 0.024372771382331848, + 0.03720013052225113, + 0.029133139178156853, + 0.029525699093937874, + 0.017133761197328568, + 0.009494603611528873, + 0.010069540701806545, + 0.013287996873259544, + 0.01433226652443409, + 0.014936785213649273, + 0.014276214875280857, + 0.018477600067853928, + 0.007804437540471554, + 0.010692572221159935, + 0.00930847693234682, + 0.017825409770011902, + 0.016418129205703735, + 0.009582356549799442, + 0.010634904727339745, + 0.008810300379991531, + 0.00870535708963871, + 0.008113251999020576, + 0.009648694656789303, + 0.009409266524016857, + 0.021107684820890427, + 0.024821246042847633, + 0.012173241935670376, + 0.020292110741138458, + 0.00999612919986248, + 0.012355633080005646, + 0.008731655776500702, + 0.011606750078499317, + 0.024532053619623184, + 0.013984153047204018, + 0.009253100492060184, + 0.010784323327243328, + 0.005639275535941124, + 0.003841537982225418, + 0.004022596869617701, + 0.005306679755449295, + 0.007797786500304937, + 0.011482364498078823, + 0.007072815205901861, + 0.008911012671887875, + 0.0057027461007237434, + 0.004161272197961807, + 0.005207395646721125, + 0.006848256569355726, + 0.007821952924132347, + 0.009114698506891727, + 0.00573631701990962, + 0.009847516193985939, + 0.014994812197983265, + 0.01165351364761591, + 0.00965769961476326, + 0.01020115613937378, + 0.010852326638996601, + 0.01678844541311264, + 0.009622648358345032, + 0.011058743111789227, + 0.007585620507597923, + 0.008515109308063984, + 0.00914602167904377, + 0.014284311793744564, + 0.030464738607406616, + 0.016020499169826508, + 0.015488551929593086, + 0.013165980577468872, + 0.004840235225856304, + 0.005953513085842133, + 0.004697798285633326, + 0.007223882246762514, + 0.007410881575196981, + 0.007840684615075588, + 0.005203571170568466, + 0.005824587307870388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.006738675758242607 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a static, slightly high-angle close-up, the scene focuses intensely on the patients' faces as they regain consciousness. Their eyelids, thin and translucent, flutter with a series of rapid, involuntary tremors before slowly, heavily peeling back. The exposed irises are a hazy, unfocused blue that darts aimlessly under the cool, sterile glare of the overhead fluorescent lights, casting soft shadows across their pale features. They lie on starched white hospital pillows, a few wisps of damp light brown hair clinging to their foreheads, their expressions one of pure disorientation. The fabric of their pale blue, wrinkled hospital gowns tightens as their chests begin to slowly and unsteadily rise, culminating in a low, trembling, audible breath that visibly swirls through their bodies, creating a powerful and intimate atmosphere of vulnerability and bewildered re-entry into the world.", + "seed": 5680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_8/index_7361_seed_5680_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15690170843154191, + "vlm_logprob": -0.15690170843154191, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.15690170843154191, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2654624283313751 + ], + "opposite_logprob_tokens": [ + -1.5154623985290527 + ], + "logprob_reward": -0.2654624283313751, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2654624283313751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21092499792575836, + 0.1752152293920517, + 0.03394508361816406, + 0.05920811742544174, + 0.0822315663099289, + 0.1631886065006256, + 0.012428386136889458, + 0.053410544991493225, + 0.14756716787815094, + 0.061880290508270264 + ], + "credit_map": [ + 0.017204420641064644, + 0.018779633566737175, + 0.0314028263092041, + 0.024788428097963333, + 0.03052877075970173, + 0.03562108054757118, + 0.034573156386613846, + 0.01802668906748295, + 0.01589740253984928, + 0.02367229573428631, + 0.023071106523275375, + 0.013375001028180122, + 0.033627357333898544, + 0.027458228170871735, + 0.01879873126745224, + 0.019315117970108986, + 0.004087075125426054, + 0.004327741451561451, + 0.0041095600463449955, + 0.003705481765791774, + 0.005343097262084484, + 0.005060830619186163, + 0.003927147015929222, + 0.003384150331839919, + 0.004331992939114571, + 0.006522521376609802, + 0.007071444299072027, + 0.006530740763992071, + 0.008380364626646042, + 0.0057934666983783245, + 0.008908551186323166, + 0.011669041588902473, + 0.008719682693481445, + 0.009447144344449043, + 0.008023951202630997, + 0.010224001482129097, + 0.011492449790239334, + 0.011745440773665905, + 0.007830959744751453, + 0.014747942797839642, + 0.016542082652449608, + 0.018588140606880188, + 0.019861765205860138, + 0.018904997035861015, + 0.022219711914658546, + 0.03208690509200096, + 0.016328873112797737, + 0.018656140193343163, + 0.0008759929914958775, + 0.0021406456362456083, + 0.0011804376263171434, + 0.0019793189130723476, + 0.002387700602412224, + 0.0009368964238092303, + 0.00189142522867769, + 0.0010359697043895721, + 0.0028979391790926456, + 0.007388686295598745, + 0.005181364715099335, + 0.005369902588427067, + 0.011415615677833557, + 0.007522632367908955, + 0.008614419028162956, + 0.005019985139369965, + 0.017843225970864296, + 0.017008092254400253, + 0.014964083209633827, + 0.015665603801608086, + 0.02527620643377304, + 0.012542564421892166, + 0.02135135419666767, + 0.022916046902537346, + 0.00405596848577261, + 0.007903430610895157, + 0.005545232445001602, + 0.008764457888901234, + 0.01354435458779335, + 0.005113091319799423, + 0.01117105595767498, + 0.00578270060941577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.2654624283313751 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871530592441559 + ], + "opposite_logprob_tokens": [ + -1.1371530294418335 + ], + "logprob_reward": -0.3871530592441559, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871530592441559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22452257573604584, + 0.08454394340515137, + 0.05767252296209335, + 0.06406523287296295, + 0.09204825013875961, + 0.19997023046016693, + 0.04727170988917351, + 0.017764993011951447, + 0.17938555777072906, + 0.032754987478256226 + ], + "credit_map": [ + 0.023269163444638252, + 0.023566532880067825, + 0.03869697079062462, + 0.020385274663567543, + 0.03237234801054001, + 0.02701706439256668, + 0.03656292334198952, + 0.022652296349406242, + 0.009190388023853302, + 0.011163941584527493, + 0.009938989765942097, + 0.011080280877649784, + 0.010599874891340733, + 0.01139345858246088, + 0.01196851022541523, + 0.009208494797348976, + 0.006270433776080608, + 0.007999690249562263, + 0.005461516324430704, + 0.006877930834889412, + 0.009545709937810898, + 0.005298098549246788, + 0.00861499272286892, + 0.0076041510328650475, + 0.00804953183978796, + 0.010159403085708618, + 0.006995923817157745, + 0.007063163910061121, + 0.010616135783493519, + 0.00842479057610035, + 0.007103768177330494, + 0.005652512423694134, + 0.013399077579379082, + 0.010441150516271591, + 0.01034446433186531, + 0.009617537260055542, + 0.01024954579770565, + 0.015482124872505665, + 0.008747895248234272, + 0.013766453601419926, + 0.022586721926927567, + 0.02429758757352829, + 0.024279307574033737, + 0.019501227885484695, + 0.028160372748970985, + 0.03766972944140434, + 0.019375614821910858, + 0.024099672213196754, + 0.003929686266928911, + 0.006974144838750362, + 0.005402435082942247, + 0.005289245396852493, + 0.00968698039650917, + 0.004933302756398916, + 0.006124521140009165, + 0.00493139261379838, + 0.0020657985005527735, + 0.0035135778598487377, + 0.0017336612800136209, + 0.0027139661833643913, + 0.003164514433592558, + 0.001510088099166751, + 0.0018598990282043815, + 0.0012034879764541984, + 0.02585017867386341, + 0.020281407982110977, + 0.016966918483376503, + 0.0218739565461874, + 0.026309896260499954, + 0.01535292249172926, + 0.03248051181435585, + 0.020269762724637985, + 0.0020343847572803497, + 0.002784139709547162, + 0.002853952581062913, + 0.006659842096269131, + 0.004526207223534584, + 0.0034715961664915085, + 0.0081761684268713, + 0.0022486967500299215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.3871530592441559 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790238082408905 + ], + "opposite_logprob_tokens": [ + -2.579023838043213 + ], + "logprob_reward": -0.0790238082408905, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790238082408905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33418264985084534, + 0.0758887231349945, + 0.05215134844183922, + 0.1022164449095726, + 0.06019614264369011, + 0.12366236001253128, + 0.02916870266199112, + 0.04576103761792183, + 0.13931193947792053, + 0.03746066987514496 + ], + "credit_map": [ + 0.03550150990486145, + 0.03866519033908844, + 0.04037369415163994, + 0.035348135977983475, + 0.05246296897530556, + 0.054947249591350555, + 0.051964715123176575, + 0.02491920068860054, + 0.005392579361796379, + 0.011724237352609634, + 0.009708624333143234, + 0.006903527770191431, + 0.014734298922121525, + 0.010884949937462807, + 0.008347288705408573, + 0.008193221874535084, + 0.005328948609530926, + 0.005877814255654812, + 0.006063646636903286, + 0.005641385447233915, + 0.006637779530137777, + 0.007454999256879091, + 0.008870904333889484, + 0.006275871768593788, + 0.007925266399979591, + 0.0137880714610219, + 0.00997491367161274, + 0.009965249337255955, + 0.018837349489331245, + 0.010099999606609344, + 0.013661288656294346, + 0.017964299768209457, + 0.005145671311765909, + 0.007295008283108473, + 0.006440536584705114, + 0.006189301144331694, + 0.006982789840549231, + 0.012724488973617554, + 0.006283743306994438, + 0.009134604595601559, + 0.01706208661198616, + 0.016098443418741226, + 0.014867282472550869, + 0.016168883070349693, + 0.018043121322989464, + 0.01511023286730051, + 0.013053063303232193, + 0.013259246945381165, + 0.00312254810705781, + 0.004131148103624582, + 0.0030546230264008045, + 0.003414725186303258, + 0.005859687924385071, + 0.0029579319525510073, + 0.003119701985269785, + 0.0035083345137536526, + 0.003034425899386406, + 0.005526989232748747, + 0.005530600901693106, + 0.0037287434097379446, + 0.007455506362020969, + 0.007778570055961609, + 0.007270473521202803, + 0.005435727536678314, + 0.01987035758793354, + 0.016660435125231743, + 0.013349429704248905, + 0.01756438799202442, + 0.021806420758366585, + 0.01118736993521452, + 0.02615409716963768, + 0.012719433754682541, + 0.0025744412560015917, + 0.0034341972786933184, + 0.0036162028554826975, + 0.0066022672690451145, + 0.00596584053710103, + 0.003721928922459483, + 0.007949777878820896, + 0.003596011083573103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0790238082408905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308429732918739 + ], + "opposite_logprob_tokens": [ + -3.1680843830108643 + ], + "logprob_reward": -0.04308429732918739, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308429732918739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3006970286369324, + 0.13065092265605927, + 0.11435167491436005, + 0.06630692631006241, + 0.031192636117339134, + 0.053423285484313965, + 0.05207617208361626, + 0.06302008032798767, + 0.1298307329416275, + 0.05845049396157265 + ], + "credit_map": [ + 0.03173176944255829, + 0.025774387642741203, + 0.049348801374435425, + 0.022118831053376198, + 0.04619782045483589, + 0.06734181195497513, + 0.02832346223294735, + 0.029860155656933784, + 0.010774670168757439, + 0.010668114759027958, + 0.018963439390063286, + 0.01260871160775423, + 0.016918500885367393, + 0.027360999956727028, + 0.015637559816241264, + 0.01771894469857216, + 0.010280241258442402, + 0.014683198183774948, + 0.014265947975218296, + 0.012268834747374058, + 0.020075598731637, + 0.01654813624918461, + 0.01563621684908867, + 0.010593513026833534, + 0.004279079847037792, + 0.010008232668042183, + 0.00749379163607955, + 0.006808142643421888, + 0.013185355812311172, + 0.005890055559575558, + 0.008562512695789337, + 0.010079757310450077, + 0.0034181142691522837, + 0.003847662592306733, + 0.0030397672671824694, + 0.004203684162348509, + 0.00378127908334136, + 0.005662471055984497, + 0.0038374534342437983, + 0.0034022091422230005, + 0.009692976251244545, + 0.005449624266475439, + 0.006683743093162775, + 0.006935250945389271, + 0.0064177303574979305, + 0.006870325654745102, + 0.006045293062925339, + 0.005328347906470299, + 0.003673794213682413, + 0.00768223125487566, + 0.007216074038296938, + 0.004664527717977762, + 0.010429996997117996, + 0.006216518580913544, + 0.005722649861127138, + 0.006470383144915104, + 0.0034998897463083267, + 0.009223094210028648, + 0.007405109703540802, + 0.006323650013655424, + 0.012343051843345165, + 0.00747581385076046, + 0.010024377144873142, + 0.00672509428113699, + 0.016957111656665802, + 0.016327964141964912, + 0.012785539031028748, + 0.015184340067207813, + 0.024363672360777855, + 0.010880737565457821, + 0.021297166123986244, + 0.012034210376441479, + 0.00553622841835022, + 0.007676542270928621, + 0.005942655727267265, + 0.006983700208365917, + 0.008647890761494637, + 0.008286992087960243, + 0.00753966486081481, + 0.007836826145648956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.04308429732918739 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients lying on starched white hospital pillows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients' eyelids flutter with rapid tremors before slowly peeling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed irises a hazy, unfocused blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the patients wearing pale blue, wrinkled hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784949012100697 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009784949012100697, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784949012100697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16334772109985352, + 0.15192219614982605, + 0.06204225867986679, + 0.0797615647315979, + 0.07480458170175552, + 0.08931218087673187, + 0.02791544236242771, + 0.08348127454519272, + 0.096317358314991, + 0.17109541594982147 + ], + "credit_map": [ + 0.009313461370766163, + 0.02087365835905075, + 0.020065560936927795, + 0.01630636863410473, + 0.029531385749578476, + 0.026189297437667847, + 0.024962086230516434, + 0.016105910763144493, + 0.013970848172903061, + 0.021179649978876114, + 0.020215893164277077, + 0.015765486285090446, + 0.023672081530094147, + 0.02245515026152134, + 0.02011367678642273, + 0.014549415558576584, + 0.0065366122871637344, + 0.009495952166616917, + 0.007242362014949322, + 0.007354585453867912, + 0.010402258485555649, + 0.007160118781030178, + 0.009118599817156792, + 0.004731771070510149, + 0.006125178653746843, + 0.009984476491808891, + 0.007520216051489115, + 0.008818287402391434, + 0.014612863771617413, + 0.0071843769401311874, + 0.011275563389062881, + 0.014240610413253307, + 0.005633773747831583, + 0.007720413152128458, + 0.008265219628810883, + 0.007297050207853317, + 0.009069888852536678, + 0.015638895332813263, + 0.008389015682041645, + 0.012790336273610592, + 0.009645337238907814, + 0.011104459874331951, + 0.011096912436187267, + 0.008529579266905785, + 0.01092034112662077, + 0.017627829685807228, + 0.008166185580193996, + 0.012221546843647957, + 0.0027768313884735107, + 0.003077481873333454, + 0.003110864432528615, + 0.0030091553926467896, + 0.004160434007644653, + 0.00468165148049593, + 0.00347539852373302, + 0.003623628057539463, + 0.005245828535407782, + 0.012368305586278439, + 0.007363040000200272, + 0.010125851258635521, + 0.0163818821310997, + 0.008868342265486717, + 0.01363949105143547, + 0.009488538838922977, + 0.01425464078783989, + 0.010598459281027317, + 0.008872303180396557, + 0.010363484732806683, + 0.013872398063540459, + 0.009375527501106262, + 0.015319598838686943, + 0.01366095058619976, + 0.009979493916034698, + 0.018751313909888268, + 0.014815188013017178, + 0.03267236799001694, + 0.02568036876618862, + 0.011131273582577705, + 0.046194903552532196, + 0.011870511807501316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.009784949012100697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_9/index_tensor([4892, 4892])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_9/index_tensor([4892, 4892])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..618903b7e2083ac2e4b2b6d9ff06755ea651f284 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_9/index_tensor([4892, 4892])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot establishes a remarkably tranquil seascape from the perspective of the coast, looking directly out to sea at sunset. In the immediate foreground, the coastline offers a rich texture, composed of damp, dark sand, interspersed with a variety of smooth, wet rocks and colorful pebbles that shimmer under the night sky. The central region is dominated by the deep blue ocean, its surface brimming with gentle, relentless energy. Small waves rhythmically roll towards the land, their crests catching the last rays of light in shimmering golden ripples. On the distant horizon, the sun, a fiery, molten orb, is partially obscured as it sinks majestically, casting a spectacular, warm glow across the sky in a gradation of fiery orange, deep burgundy, and soft purple. This golden, hour-long light creates a shimmering path across the water, guiding the viewer's gaze from shore to horizon, bathing the entire scene in a warm, tranquil glow, evoking a powerful atmosphere of quiet contemplation and the vast beauty of nature.", + "seed": 25407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_9/index_4892_seed_25407_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04072172669693828, + "vlm_logprob": -0.04072172669693828, + "vlm_yes_no_margin": 4.249999976158142, + "vlm_token_credit": -0.04072172669693828, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007778591942042112 + ], + "opposite_logprob_tokens": [ + -6.882778644561768 + ], + "logprob_reward": -0.007778591942042112, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.007778591942042112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3551410138607025, + 0.08683479577302933, + 0.07590323686599731, + 0.06378714740276337, + 0.07323276251554489, + 0.0772087424993515, + 0.06759852170944214, + 0.047961972653865814, + 0.05587583780288696, + 0.09645596891641617 + ], + "credit_map": [ + 0.027665093541145325, + 0.022673500701785088, + 0.06623877584934235, + 0.06688874959945679, + 0.01962527260184288, + 0.033636175096035004, + 0.04945635423064232, + 0.0689571276307106, + 0.008723953738808632, + 0.005979375448077917, + 0.012864543125033379, + 0.0206258837133646, + 0.006102210376411676, + 0.007889541797339916, + 0.01080594677478075, + 0.013843339867889881, + 0.0051232269033789635, + 0.007039079908281565, + 0.013768870383501053, + 0.013459925539791584, + 0.005892653949558735, + 0.005490132141858339, + 0.010519991628825665, + 0.014609354548156261, + 0.008578390814363956, + 0.006585605442523956, + 0.010308709926903248, + 0.009058447554707527, + 0.006621743552386761, + 0.0064623719081282616, + 0.007025653030723333, + 0.009146228432655334, + 0.009319601580500603, + 0.0075238426215946674, + 0.012681739404797554, + 0.006621366832405329, + 0.005737758707255125, + 0.005449291318655014, + 0.012095065787434578, + 0.01380409300327301, + 0.006496011279523373, + 0.005113646388053894, + 0.011521225795149803, + 0.01167722325772047, + 0.0053841108456254005, + 0.00727404560893774, + 0.016019538044929504, + 0.013722941279411316, + 0.006487439386546612, + 0.005603522527962923, + 0.009526207111775875, + 0.0073273442685604095, + 0.0041203973814845085, + 0.007693321444094181, + 0.016110537573695183, + 0.010729756206274033, + 0.005152949597686529, + 0.006132297217845917, + 0.006522121839225292, + 0.006643402855843306, + 0.005522429011762142, + 0.004835686180740595, + 0.006896542385220528, + 0.0062565430998802185, + 0.00491682393476367, + 0.005935957655310631, + 0.008632946759462357, + 0.009338357485830784, + 0.005329038482159376, + 0.005679404363036156, + 0.007389314007014036, + 0.00865399744361639, + 0.008984241634607315, + 0.009091678075492382, + 0.01502317376434803, + 0.02059844322502613, + 0.007776445709168911, + 0.008326956070959568, + 0.010514250956475735, + 0.016140781342983246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.007778591942042112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194594696164131 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194594696164131, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194594696164131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16181476414203644, + 0.12538941204547882, + 0.06661242246627808, + 0.04815739765763283, + 0.05707264319062233, + 0.07712312787771225, + 0.07354667037725449, + 0.20919668674468994, + 0.05496508255600929, + 0.12612171471118927 + ], + "credit_map": [ + 0.01799783483147621, + 0.013172385282814503, + 0.028591644018888474, + 0.023549936711788177, + 0.013256877660751343, + 0.01879892684519291, + 0.02028769999742508, + 0.026159487664699554, + 0.011514812707901001, + 0.01502072811126709, + 0.024415338411927223, + 0.02080547623336315, + 0.014377736486494541, + 0.009432815946638584, + 0.00983280036598444, + 0.01998971961438656, + 0.009273058734834194, + 0.005836629308760166, + 0.011340971104800701, + 0.011588209308683872, + 0.006271404214203358, + 0.006634168792515993, + 0.005791543051600456, + 0.00987644400447607, + 0.005836141761392355, + 0.003780884202569723, + 0.007848995737731457, + 0.007953709922730923, + 0.002947040367871523, + 0.004406594205647707, + 0.008368645794689655, + 0.007015393115580082, + 0.006191697437316179, + 0.005794333294034004, + 0.010927688330411911, + 0.0072357370518147945, + 0.0056314608082175255, + 0.006169234402477741, + 0.006156109273433685, + 0.008966390043497086, + 0.00919361598789692, + 0.009737672284245491, + 0.013401545584201813, + 0.007664174772799015, + 0.0074967374093830585, + 0.011156007647514343, + 0.007223043590784073, + 0.01125034224241972, + 0.008932316675782204, + 0.007498865015804768, + 0.011379905976355076, + 0.009043712168931961, + 0.005368917249143124, + 0.008977830410003662, + 0.01157863438129425, + 0.010766496881842613, + 0.018365085124969482, + 0.017964735627174377, + 0.031022528186440468, + 0.03366486728191376, + 0.015055560506880283, + 0.017226023599505424, + 0.03878897428512573, + 0.03710894286632538, + 0.004326095804572105, + 0.00518638314679265, + 0.010290858335793018, + 0.007833356969058514, + 0.0047656213864684105, + 0.007099858485162258, + 0.006520495284348726, + 0.008942417800426483, + 0.01358078233897686, + 0.013314093463122845, + 0.02117254212498665, + 0.018395092338323593, + 0.011301870457828045, + 0.015571891330182552, + 0.013204672373831272, + 0.01958078145980835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.014194594696164131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270521581172943 + ], + "opposite_logprob_tokens": [ + -2.127052068710327 + ], + "logprob_reward": -0.1270521581172943, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270521581172943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2869810461997986, + 0.10703998804092407, + 0.08144321292638779, + 0.09435388445854187, + 0.13229520618915558, + 0.10561064630746841, + 0.048222724348306656, + 0.07301129400730133, + 0.04082738980650902, + 0.030214661732316017 + ], + "credit_map": [ + 0.03417426720261574, + 0.024488365277647972, + 0.05119917541742325, + 0.04582379758358002, + 0.028318345546722412, + 0.02904578484594822, + 0.03153754398226738, + 0.042393751442432404, + 0.010343130677938461, + 0.0075116874650120735, + 0.018241124227643013, + 0.022993529215455055, + 0.00848449394106865, + 0.00956400204449892, + 0.015034612268209457, + 0.014867410995066166, + 0.010311356745660305, + 0.00772404158487916, + 0.014378435909748077, + 0.01394637394696474, + 0.009101083502173424, + 0.007500395178794861, + 0.007450522854924202, + 0.011031001806259155, + 0.012882922776043415, + 0.008585527539253235, + 0.01431251410394907, + 0.015970071777701378, + 0.012071824632585049, + 0.007937544025480747, + 0.010433442890644073, + 0.012160041369497776, + 0.015247679315507412, + 0.01876695826649666, + 0.01984531804919243, + 0.010603206232190132, + 0.02207128517329693, + 0.01399768516421318, + 0.012447597458958626, + 0.01931547001004219, + 0.009326033294200897, + 0.006382734980434179, + 0.0193832665681839, + 0.02008800394833088, + 0.0077420128509402275, + 0.00921222846955061, + 0.015482790768146515, + 0.017993571236729622, + 0.0038395009469240904, + 0.0047571659088134766, + 0.005120965652167797, + 0.004966127686202526, + 0.004757733549922705, + 0.005146474111825228, + 0.012075121514499187, + 0.007559636142104864, + 0.005731889978051186, + 0.006179288495332003, + 0.010541453957557678, + 0.010745334438979626, + 0.005609463434666395, + 0.006238745525479317, + 0.014308295212686062, + 0.013656822964549065, + 0.004179548472166061, + 0.004089346621185541, + 0.0065371314994990826, + 0.005256113596260548, + 0.005887812003493309, + 0.004376565106213093, + 0.004633292090147734, + 0.005867582280188799, + 0.0031786388717591763, + 0.0035674262326210737, + 0.004145225044339895, + 0.004406143445521593, + 0.003289635758846998, + 0.003415644634515047, + 0.0035696786362677813, + 0.0046422663144767284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.1270521581172943 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949288606643677 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949288606643677, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949288606643677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2727782726287842, + 0.10778600722551346, + 0.09609828889369965, + 0.07285600155591965, + 0.07321928441524506, + 0.09465109556913376, + 0.061553820967674255, + 0.08332906663417816, + 0.05265754088759422, + 0.0850706622004509 + ], + "credit_map": [ + 0.03462577983736992, + 0.040800172835588455, + 0.04213253781199455, + 0.03154735639691353, + 0.036077775061130524, + 0.027247129008173943, + 0.025448184460401535, + 0.03489932417869568, + 0.00971528422087431, + 0.012617374770343304, + 0.019208520650863647, + 0.016452018171548843, + 0.012005492113530636, + 0.007519307546317577, + 0.014230255037546158, + 0.016037749126553535, + 0.011492523364722729, + 0.01198054663836956, + 0.012948861345648766, + 0.012854427099227905, + 0.016442526131868362, + 0.008744418621063232, + 0.00937893521040678, + 0.012256041169166565, + 0.011279772035777569, + 0.006226980127394199, + 0.011903498321771622, + 0.009873009286820889, + 0.0068354420363903046, + 0.009609436616301537, + 0.0060790469869971275, + 0.011048806831240654, + 0.010641894303262234, + 0.008654492907226086, + 0.009960511699318886, + 0.006771579384803772, + 0.00878220610320568, + 0.010081598535180092, + 0.008324622176587582, + 0.010002369992434978, + 0.011108100414276123, + 0.008433504030108452, + 0.015165206044912338, + 0.013640081509947777, + 0.007913832552731037, + 0.012114354409277439, + 0.012562385760247707, + 0.013713628984987736, + 0.006832942366600037, + 0.004576291888952255, + 0.010675309225916862, + 0.007594097405672073, + 0.0057104043662548065, + 0.00805213674902916, + 0.0097982631996274, + 0.008314372971653938, + 0.008217627182602882, + 0.007986350916326046, + 0.01598413847386837, + 0.011055578477680683, + 0.010964073240756989, + 0.007955274544656277, + 0.00936341192573309, + 0.011802604421973228, + 0.00596317183226347, + 0.006191331893205643, + 0.007008123677223921, + 0.004933224990963936, + 0.00802605226635933, + 0.007937828078866005, + 0.005384387448430061, + 0.007213415578007698, + 0.009379871189594269, + 0.009673971682786942, + 0.013800233602523804, + 0.012064831331372261, + 0.011323451064527035, + 0.00828284677118063, + 0.009626221843063831, + 0.010919231921434402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005949288606643677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863400012254715 + ], + "opposite_logprob_tokens": [ + -3.0486340522766113 + ], + "logprob_reward": -0.04863400012254715, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863400012254715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2200072854757309, + 0.09752807021141052, + 0.05878790467977524, + 0.2403724491596222, + 0.07663049548864365, + 0.12267544120550156, + 0.047656651586294174, + 0.04484068974852562, + 0.04298284649848938, + 0.04851812869310379 + ], + "credit_map": [ + 0.02209511213004589, + 0.01768607273697853, + 0.03997036814689636, + 0.03147868439555168, + 0.018440216779708862, + 0.024656441062688828, + 0.027285190299153328, + 0.038395218551158905, + 0.00896320678293705, + 0.00766020780429244, + 0.017778364941477776, + 0.013934263028204441, + 0.009973562322556973, + 0.011094121262431145, + 0.0110652856528759, + 0.017059072852134705, + 0.00691324844956398, + 0.005251875147223473, + 0.010172172449529171, + 0.009999103844165802, + 0.0038177468813955784, + 0.005221514962613583, + 0.0073904963210225105, + 0.010021748021245003, + 0.03093705140054226, + 0.02562672644853592, + 0.06585703045129776, + 0.0266432985663414, + 0.02059002034366131, + 0.01879456639289856, + 0.015281925909221172, + 0.03664182499051094, + 0.008053644560277462, + 0.007246250286698341, + 0.012819168157875538, + 0.008405675180256367, + 0.005587851628661156, + 0.00740783428773284, + 0.012836625799536705, + 0.014273451641201973, + 0.010132561437785625, + 0.00885386485606432, + 0.02092391811311245, + 0.019742663949728012, + 0.007463234476745129, + 0.01131420861929655, + 0.021585920825600624, + 0.022659070789813995, + 0.0042913975194096565, + 0.0045501478016376495, + 0.00664296792820096, + 0.004891796037554741, + 0.0039516594260931015, + 0.003441320266574621, + 0.012156433425843716, + 0.007730934303253889, + 0.0036556681152433157, + 0.003329351544380188, + 0.008105123415589333, + 0.007841285318136215, + 0.003840814810246229, + 0.005033764056861401, + 0.005801116582006216, + 0.00723356893286109, + 0.003391384147107601, + 0.004594492260366678, + 0.006514366250485182, + 0.0082449521869421, + 0.0033248523250222206, + 0.0039029486943036318, + 0.0065951356664299965, + 0.006414719391614199, + 0.004423636011779308, + 0.004745709244161844, + 0.00912455189973116, + 0.008241003379225731, + 0.004389876965433359, + 0.005029487423598766, + 0.004818357527256012, + 0.007745508104562759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.04863400012254715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot establishes a remarkably tranquil seascape from the perspective of the coast, looking directly out to sea at sunset. In the immediate foreground, the coastline offers a rich texture, composed of damp, dark sand, interspersed with a variety of smooth, wet rocks and colorful pebbles that shimmer under the night sky. The central region is dominated by the deep blue ocean, its surface brimming with gentle, relentless energy. Small waves rhythmically roll towards the land, their crests catching the last rays of light in shimmering golden ripples. On the distant horizon, the sun, a fiery, molten orb, is partially obscured as it sinks majestically, casting a spectacular, warm glow across the sky in a gradation of fiery orange, deep burgundy, and soft purple. This golden, hour-long light creates a shimmering path across the water, guiding the viewer's gaze from shore to horizon, bathing the entire scene in a warm, tranquil glow, evoking a powerful atmosphere of quiet contemplation and the vast beauty of nature.", + "seed": 25408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_9/index_4892_seed_25408_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.052446423657238485, + "vlm_logprob": -0.052446423657238485, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.052446423657238485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270989146083593 + ], + "opposite_logprob_tokens": [ + -6.755270957946777 + ], + "logprob_reward": -0.005270989146083593, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005270989146083593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2563357353210449, + 0.046553172171115875, + 0.042300187051296234, + 0.03400547802448273, + 0.12022602558135986, + 0.03451763093471527, + 0.1518913060426712, + 0.09166264533996582, + 0.12195079773664474, + 0.10055701434612274 + ], + "credit_map": [ + 0.046001553535461426, + 0.05152542144060135, + 0.048641953617334366, + 0.03203834965825081, + 0.01526348851621151, + 0.022633634507656097, + 0.013783132657408714, + 0.026448190212249756, + 0.008263477124273777, + 0.00555056007578969, + 0.009232883341610432, + 0.006873204838484526, + 0.002371835755184293, + 0.004294070415198803, + 0.004369912203401327, + 0.005597227718681097, + 0.006640262436121702, + 0.005722088739275932, + 0.006520513910800219, + 0.007558782584965229, + 0.004127582535147667, + 0.004886024631559849, + 0.0029744317289441824, + 0.003870498389005661, + 0.006217451300472021, + 0.005221272818744183, + 0.004673514515161514, + 0.005083031952381134, + 0.0032459255307912827, + 0.0031661938410252333, + 0.002684491453692317, + 0.00371359521523118, + 0.02213391847908497, + 0.021475395187735558, + 0.02251194790005684, + 0.013021277263760567, + 0.009770452976226807, + 0.010679195635020733, + 0.007804465480148792, + 0.012829365208745003, + 0.007979696616530418, + 0.004857269581407309, + 0.004902521148324013, + 0.004082778934389353, + 0.0025027948431670666, + 0.0034568668343126774, + 0.0023904284462332726, + 0.004345274996012449, + 0.023471852764487267, + 0.02221681922674179, + 0.02237679623067379, + 0.011486547999083996, + 0.011550704948604107, + 0.016582844778895378, + 0.012668396346271038, + 0.03153734654188156, + 0.013244661502540112, + 0.008759971708059311, + 0.012636886909604073, + 0.009856454096734524, + 0.00933038629591465, + 0.008468267507851124, + 0.01622650772333145, + 0.01313950214534998, + 0.022470109164714813, + 0.01791311800479889, + 0.017359597608447075, + 0.013837647624313831, + 0.012449974194169044, + 0.014467274770140648, + 0.009929037652909756, + 0.013524038717150688, + 0.01636783964931965, + 0.009566279128193855, + 0.019976098090410233, + 0.014067727141082287, + 0.006848446559160948, + 0.007409907411783934, + 0.007514281664043665, + 0.018806425854563713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.005270989146083593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978341419249773 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.00978341419249773, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978341419249773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2047714740037918, + 0.08999909460544586, + 0.15237922966480255, + 0.0957389697432518, + 0.08325418829917908, + 0.05648192763328552, + 0.07362578809261322, + 0.16680465638637543, + 0.04258014261722565, + 0.03436455503106117 + ], + "credit_map": [ + 0.03981450945138931, + 0.031880155205726624, + 0.04099242761731148, + 0.028055403381586075, + 0.015164513140916824, + 0.01828267052769661, + 0.010609744116663933, + 0.019972050562500954, + 0.017648974433541298, + 0.011597882024943829, + 0.018166981637477875, + 0.014075036160647869, + 0.00578045891597867, + 0.007700853981077671, + 0.005851265974342823, + 0.00917764287441969, + 0.023088695481419563, + 0.022250397130846977, + 0.030643464997410774, + 0.0240329597145319, + 0.006547871977090836, + 0.010214408859610558, + 0.011956209316849709, + 0.023645225912332535, + 0.015499293804168701, + 0.01827118918299675, + 0.014055658131837845, + 0.014079705812036991, + 0.0060166302137076855, + 0.008118762634694576, + 0.006137964315712452, + 0.013559764251112938, + 0.015869155526161194, + 0.013480321504175663, + 0.016533980146050453, + 0.009946572594344616, + 0.0065152449533343315, + 0.0065930006094276905, + 0.007021764758974314, + 0.007294153328984976, + 0.01245055627077818, + 0.008751877583563328, + 0.008315518498420715, + 0.005809416528791189, + 0.005224015098065138, + 0.005170854274183512, + 0.0029052565805613995, + 0.007854430936276913, + 0.013652006164193153, + 0.007874231785535812, + 0.01869056187570095, + 0.007795536424964666, + 0.003888422856107354, + 0.0057020485401153564, + 0.003434482729062438, + 0.012588487938046455, + 0.026478895917534828, + 0.018939904868602753, + 0.03670654818415642, + 0.02319454774260521, + 0.012547794729471207, + 0.015436893329024315, + 0.012176292948424816, + 0.021323781460523605, + 0.007178815081715584, + 0.005558398552238941, + 0.007160266395658255, + 0.005142399575561285, + 0.003563126316294074, + 0.004345014691352844, + 0.0044464292004704475, + 0.005185691174119711, + 0.006159581709653139, + 0.003545232582837343, + 0.007548841182142496, + 0.00563817797228694, + 0.0022883042693138123, + 0.0023248419165611267, + 0.002461720257997513, + 0.004397853277623653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.00978341419249773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798975169658661 + ], + "opposite_logprob_tokens": [ + -1.8048975467681885 + ], + "logprob_reward": -0.1798975169658661, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798975169658661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39225703477859497, + 0.10111061483621597, + 0.06885381788015366, + 0.10662867873907089, + 0.07955225557088852, + 0.07923942059278488, + 0.04521545395255089, + 0.03198251500725746, + 0.04956761747598648, + 0.045592594891786575 + ], + "credit_map": [ + 0.08781475573778152, + 0.0590149462223053, + 0.06820239871740341, + 0.049591198563575745, + 0.02210007794201374, + 0.028910182416439056, + 0.03684120997786522, + 0.03978222236037254, + 0.016680333763360977, + 0.011614210903644562, + 0.0213377233594656, + 0.013403539545834064, + 0.008271482773125172, + 0.01105558406561613, + 0.005782265681773424, + 0.012965478003025055, + 0.01140100508928299, + 0.011850632727146149, + 0.01108231209218502, + 0.010059554129838943, + 0.005013637710362673, + 0.006471431348472834, + 0.005134387407451868, + 0.007840857841074467, + 0.01698075421154499, + 0.014746209606528282, + 0.02108297124505043, + 0.009786874055862427, + 0.006382883992046118, + 0.009637544862926006, + 0.0062055084854364395, + 0.021805930882692337, + 0.01498319674283266, + 0.010056929662823677, + 0.012381165288388729, + 0.00866775680333376, + 0.007437767460942268, + 0.007867719978094101, + 0.007145029958337545, + 0.011012692004442215, + 0.013524975627660751, + 0.007529325317591429, + 0.016485022380948067, + 0.012226161547005177, + 0.004670791327953339, + 0.006399665493518114, + 0.008862825110554695, + 0.009540659375488758, + 0.008516882546246052, + 0.0043308911845088005, + 0.00961469765752554, + 0.0046954224817454815, + 0.0027511173393577337, + 0.0035237097181379795, + 0.0032607296016067266, + 0.008522005751729012, + 0.005331083666533232, + 0.004376791883260012, + 0.00464622350409627, + 0.002723065437749028, + 0.0027096308767795563, + 0.0027776702772825956, + 0.004621817264705896, + 0.004796232562512159, + 0.008451813831925392, + 0.011866227723658085, + 0.005984405986964703, + 0.004802818410098553, + 0.007006380707025528, + 0.005355868488550186, + 0.0027513732202351093, + 0.0033487307373434305, + 0.007465495727956295, + 0.004908776376396418, + 0.007506831083446741, + 0.006170371081680059, + 0.004713315516710281, + 0.004666140303015709, + 0.0029606593307107687, + 0.007201005704700947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.1798975169658661 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253201816231012 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253201816231012, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253201816231012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11812490224838257, + 0.1559789478778839, + 0.06561284512281418, + 0.09588466584682465, + 0.1425260603427887, + 0.10048798471689224, + 0.12136702984571457, + 0.06324534118175507, + 0.058679141104221344, + 0.07809307426214218 + ], + "credit_map": [ + 0.020367734134197235, + 0.019270319491624832, + 0.014175591059029102, + 0.014607689343392849, + 0.015108388848602772, + 0.014391433447599411, + 0.00865449570119381, + 0.011549255810678005, + 0.027744797989726067, + 0.014508355408906937, + 0.033676281571388245, + 0.018606062978506088, + 0.008039132691919804, + 0.013009846210479736, + 0.009792591445147991, + 0.030601877719163895, + 0.010692748241126537, + 0.00906780268996954, + 0.008968045003712177, + 0.010935774073004723, + 0.006309618707746267, + 0.006069393828511238, + 0.0039017738308757544, + 0.009667693637311459, + 0.016803056001663208, + 0.01097038947045803, + 0.01215673703700304, + 0.01218241173774004, + 0.0072485171258449554, + 0.0075172120705246925, + 0.010690347291529179, + 0.01831599324941635, + 0.021759076043963432, + 0.020745031535625458, + 0.01782592572271824, + 0.021194329485297203, + 0.020814238116145134, + 0.014916115440428257, + 0.008387106470763683, + 0.01688423752784729, + 0.019511733204126358, + 0.01614098995923996, + 0.0182651337236166, + 0.009120790287852287, + 0.013848233968019485, + 0.010519614443182945, + 0.005263283848762512, + 0.007818200625479221, + 0.021821562200784683, + 0.008754569105803967, + 0.02322811260819435, + 0.008320290595293045, + 0.006117179989814758, + 0.010754575952887535, + 0.011416954919695854, + 0.03095378540456295, + 0.010173099115490913, + 0.0068520004861056805, + 0.01134710293263197, + 0.0064225224778056145, + 0.007320594508200884, + 0.006579521112143993, + 0.007704917341470718, + 0.006845587398856878, + 0.008853464387357235, + 0.009106358513236046, + 0.01010691188275814, + 0.006069174036383629, + 0.006691319402307272, + 0.005764339584857225, + 0.0053518689237535, + 0.006735706236213446, + 0.011149412021040916, + 0.007228486239910126, + 0.014623372815549374, + 0.011532106436789036, + 0.01101623848080635, + 0.007485958281904459, + 0.006496669724583626, + 0.008560829795897007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.005253201816231012 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202699616551399 + ], + "opposite_logprob_tokens": [ + -2.8120269775390625 + ], + "logprob_reward": -0.06202699616551399, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202699616551399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19944100081920624, + 0.048290930688381195, + 0.06115703284740448, + 0.09860095381736755, + 0.20564310252666473, + 0.06300631910562515, + 0.045756008476018906, + 0.06399744749069214, + 0.05437715724110603, + 0.15973012149333954 + ], + "credit_map": [ + 0.03716571629047394, + 0.01883368380367756, + 0.0502716600894928, + 0.02819802239537239, + 0.011233678087592125, + 0.015159985981881618, + 0.016784394159913063, + 0.021793844178318977, + 0.008244813419878483, + 0.006630716845393181, + 0.011130983009934425, + 0.006194357294589281, + 0.002966831438243389, + 0.003305072197690606, + 0.0036164093762636185, + 0.006201742682605982, + 0.010617128573358059, + 0.008370688185095787, + 0.010247431695461273, + 0.01194420550018549, + 0.005321796517819166, + 0.005446534138172865, + 0.0044426750391721725, + 0.004766570869833231, + 0.017419971525669098, + 0.012274631299078465, + 0.011763444170355797, + 0.010985346511006355, + 0.007009418681263924, + 0.009218544699251652, + 0.00910711195319891, + 0.020822476595640182, + 0.039662133902311325, + 0.0314803272485733, + 0.041679345071315765, + 0.025247404351830482, + 0.0119503578171134, + 0.016079483553767204, + 0.015271655283868313, + 0.024272367358207703, + 0.010650131851434708, + 0.009829916059970856, + 0.01055630762130022, + 0.007598200347274542, + 0.005686927121132612, + 0.006371733266860247, + 0.003801463870331645, + 0.008511635474860668, + 0.008658522740006447, + 0.005810365080833435, + 0.007873810827732086, + 0.00456456933170557, + 0.0031243327539414167, + 0.0036842254921793938, + 0.004905514419078827, + 0.007134661544114351, + 0.009663003496825695, + 0.007713269907981157, + 0.013411053456366062, + 0.007348080165684223, + 0.005885949358344078, + 0.006028347183018923, + 0.006498998962342739, + 0.007448737975209951, + 0.010022839531302452, + 0.008897029794752598, + 0.007896550931036472, + 0.005527537316083908, + 0.005494324956089258, + 0.006176971364766359, + 0.004700925666838884, + 0.0056609706953167915, + 0.02709738351404667, + 0.009000691585242748, + 0.03428458422422409, + 0.037475112825632095, + 0.009976173751056194, + 0.00874971691519022, + 0.018004639074206352, + 0.015141806565225124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.06202699616551399 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot establishes a remarkably tranquil seascape from the perspective of the coast, looking directly out to sea at sunset. In the immediate foreground, the coastline offers a rich texture, composed of damp, dark sand, interspersed with a variety of smooth, wet rocks and colorful pebbles that shimmer under the night sky. The central region is dominated by the deep blue ocean, its surface brimming with gentle, relentless energy. Small waves rhythmically roll towards the land, their crests catching the last rays of light in shimmering golden ripples. On the distant horizon, the sun, a fiery, molten orb, is partially obscured as it sinks majestically, casting a spectacular, warm glow across the sky in a gradation of fiery orange, deep burgundy, and soft purple. This golden, hour-long light creates a shimmering path across the water, guiding the viewer's gaze from shore to horizon, bathing the entire scene in a warm, tranquil glow, evoking a powerful atmosphere of quiet contemplation and the vast beauty of nature.", + "seed": 25409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_9/index_4892_seed_25409_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05455939183011651, + "vlm_logprob": -0.05455939183011651, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.05455939183011651, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0068658385425806046 + ], + "opposite_logprob_tokens": [ + -7.006865978240967 + ], + "logprob_reward": -0.0068658385425806046, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0068658385425806046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1525277942419052, + 0.12899760901927948, + 0.052283067256212234, + 0.06362900882959366, + 0.09177573025226593, + 0.10676467418670654, + 0.03972142934799194, + 0.21071040630340576, + 0.07655490934848785, + 0.07703535258769989 + ], + "credit_map": [ + 0.012419783510267735, + 0.013587079010903835, + 0.025384504348039627, + 0.027821136638522148, + 0.012963589280843735, + 0.015974318608641624, + 0.013153629377484322, + 0.031223759055137634, + 0.011210779659450054, + 0.012640907429158688, + 0.019477147608995438, + 0.025196382775902748, + 0.012735952623188496, + 0.0132532874122262, + 0.014082755893468857, + 0.02040039375424385, + 0.0038678552955389023, + 0.0038620121777057648, + 0.008704029023647308, + 0.010426958091557026, + 0.00495640654116869, + 0.0038161405827850103, + 0.0063437786884605885, + 0.010305888950824738, + 0.005316015332937241, + 0.006626174319535494, + 0.008407792076468468, + 0.009558437392115593, + 0.00824459083378315, + 0.006567221134901047, + 0.007671205792576075, + 0.011237574741244316, + 0.006048906594514847, + 0.007619055453687906, + 0.012714860960841179, + 0.015819156542420387, + 0.007374566048383713, + 0.01063112635165453, + 0.011893365532159805, + 0.019674694165587425, + 0.006570487283170223, + 0.009007732383906841, + 0.015167644247412682, + 0.02123282104730606, + 0.01160581037402153, + 0.009341735392808914, + 0.01865927129983902, + 0.015179164707660675, + 0.0028610220178961754, + 0.004984673578292131, + 0.005246278829872608, + 0.005793354008346796, + 0.005685311742126942, + 0.003700593486428261, + 0.005499476101249456, + 0.005950720049440861, + 0.017639731988310814, + 0.02804470993578434, + 0.02397342585027218, + 0.02510889619588852, + 0.03690464049577713, + 0.0278004240244627, + 0.01937636360526085, + 0.03186221048235893, + 0.008124078623950481, + 0.006821905728429556, + 0.014977681450545788, + 0.009457724168896675, + 0.01057338435202837, + 0.007868128828704357, + 0.005556029733270407, + 0.01317597832530737, + 0.005804682150483131, + 0.006031889468431473, + 0.010724078863859177, + 0.01526309922337532, + 0.010761694051325321, + 0.008034649305045605, + 0.010222578421235085, + 0.010192680172622204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0068658385425806046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20567744970321655, + 0.06455902755260468, + 0.08653724193572998, + 0.11645463109016418, + 0.12477325648069382, + 0.08910389244556427, + 0.09392084926366806, + 0.07563266158103943, + 0.09737109392881393, + 0.04596990346908569 + ], + "credit_map": [ + 0.02019331604242325, + 0.01522454060614109, + 0.032531872391700745, + 0.031635385006666183, + 0.01860727183520794, + 0.022165516391396523, + 0.029656535014510155, + 0.035663023591041565, + 0.0061070919036865234, + 0.005171388387680054, + 0.009931584820151329, + 0.009637869894504547, + 0.0037536299787461758, + 0.007811393588781357, + 0.006789919920265675, + 0.015356144867837429, + 0.00683566415682435, + 0.008827367797493935, + 0.011696175672113895, + 0.015219712629914284, + 0.007027022074908018, + 0.009395386092364788, + 0.009864924475550652, + 0.01767098903656006, + 0.009578889235854149, + 0.010432330891489983, + 0.014503033831715584, + 0.017143825069069862, + 0.015075251460075378, + 0.012076355516910553, + 0.014716517180204391, + 0.022928433492779732, + 0.008420675992965698, + 0.008499319665133953, + 0.018782535567879677, + 0.025423085317015648, + 0.009720328263938427, + 0.009578979574143887, + 0.02008741907775402, + 0.024260912090539932, + 0.006525985896587372, + 0.007047269027680159, + 0.013741627335548401, + 0.015140680596232414, + 0.0078109512105584145, + 0.006181566510349512, + 0.019911641255021095, + 0.012744167819619179, + 0.007799874991178513, + 0.009688124060630798, + 0.01213386282324791, + 0.016938073560595512, + 0.010352229699492455, + 0.012349504977464676, + 0.01102929376065731, + 0.013629885390400887, + 0.006955921649932861, + 0.007118123583495617, + 0.011623969301581383, + 0.011668147519230843, + 0.008086476475000381, + 0.006769271567463875, + 0.008994529023766518, + 0.0144162243232131, + 0.007342122495174408, + 0.007075142581015825, + 0.01721177063882351, + 0.016782239079475403, + 0.006591519806534052, + 0.009154131636023521, + 0.013684096746146679, + 0.01953006722033024, + 0.004805646371096373, + 0.003833184251561761, + 0.0068277036771178246, + 0.008723572827875614, + 0.003997118212282658, + 0.005143057554960251, + 0.0059497300535440445, + 0.006689891684800386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.009783769026398659 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156097412109375 + ], + "opposite_logprob_tokens": [ + -1.7015609741210938 + ], + "logprob_reward": -0.20156097412109375, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156097412109375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23006831109523773, + 0.14026102423667908, + 0.07215236127376556, + 0.046081796288490295, + 0.03579098358750343, + 0.05341590940952301, + 0.08353273570537567, + 0.13936078548431396, + 0.18189044296741486, + 0.01744571141898632 + ], + "credit_map": [ + 0.01841394416987896, + 0.03757833316922188, + 0.03552308678627014, + 0.026913553476333618, + 0.02809453383088112, + 0.028305642306804657, + 0.02535933256149292, + 0.029879866167902946, + 0.010785728693008423, + 0.01836865209043026, + 0.018848270177841187, + 0.026608463376760483, + 0.01325204223394394, + 0.015772664919495583, + 0.01475461758673191, + 0.021870579570531845, + 0.0047017051838338375, + 0.007670816499739885, + 0.009221689775586128, + 0.010352148674428463, + 0.010699555277824402, + 0.008676626719534397, + 0.00887551810592413, + 0.01195429265499115, + 0.002803829498589039, + 0.005408653989434242, + 0.007408494129776955, + 0.008226603269577026, + 0.005738973617553711, + 0.0037906861398369074, + 0.005224460735917091, + 0.007480093743652105, + 0.002911745570600033, + 0.0033311450388282537, + 0.005037267226725817, + 0.0061084008775651455, + 0.0025767963379621506, + 0.0029485090635716915, + 0.003779034595936537, + 0.009098082780838013, + 0.004007338080555201, + 0.005412120372056961, + 0.007056267932057381, + 0.008674874901771545, + 0.008473440073430538, + 0.006453358102589846, + 0.007488512434065342, + 0.005849994253367186, + 0.006212770473212004, + 0.00856977328658104, + 0.011426529847085476, + 0.015763483941555023, + 0.009460609406232834, + 0.01005635131150484, + 0.011796566657721996, + 0.010246643796563148, + 0.011103108525276184, + 0.014033264480531216, + 0.020113060250878334, + 0.016797414049506187, + 0.01766858622431755, + 0.014168361201882362, + 0.014549728482961655, + 0.030927257612347603, + 0.01462774258106947, + 0.01290486752986908, + 0.03241894021630287, + 0.029657278209924698, + 0.013244476169347763, + 0.016630584374070168, + 0.023417122662067413, + 0.038989417254924774, + 0.0013760854490101337, + 0.0017103287391364574, + 0.002452276647090912, + 0.0027680268976837397, + 0.002035376848652959, + 0.0015553456032648683, + 0.0029563133139163256, + 0.002591955941170454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.20156097412109375 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059498813934624195 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.0059498813934624195, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059498813934624195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2681056261062622, + 0.10578960180282593, + 0.07526854425668716, + 0.057705868035554886, + 0.11009862273931503, + 0.10779659450054169, + 0.06860660761594772, + 0.047314949333667755, + 0.12904523313045502, + 0.03026830218732357 + ], + "credit_map": [ + 0.02233564294874668, + 0.038121506571769714, + 0.055811863392591476, + 0.03196199610829353, + 0.023898091167211533, + 0.018295366317033768, + 0.02902255766093731, + 0.048658598214387894, + 0.007508488371968269, + 0.009917477145791054, + 0.017963092774152756, + 0.014344184659421444, + 0.009678179398179054, + 0.009440000168979168, + 0.011319718323647976, + 0.02561846189200878, + 0.006490199360996485, + 0.006924820598214865, + 0.012147990986704826, + 0.014079130254685879, + 0.006651555188000202, + 0.006992062088102102, + 0.009017331525683403, + 0.012965460307896137, + 0.004896362777799368, + 0.005201125983148813, + 0.010030089877545834, + 0.007687035948038101, + 0.00484011135995388, + 0.004673621617257595, + 0.008032904006540775, + 0.012344615533947945, + 0.006605790462344885, + 0.007244717329740524, + 0.016409046947956085, + 0.023653831332921982, + 0.008825746364891529, + 0.008909599855542183, + 0.015928687527775764, + 0.022521203383803368, + 0.005712717771530151, + 0.013172389939427376, + 0.011912285350263119, + 0.019157109782099724, + 0.02081788145005703, + 0.009365837089717388, + 0.01259282324463129, + 0.01506554801017046, + 0.005015948787331581, + 0.010282532311975956, + 0.010144169442355633, + 0.010518301278352737, + 0.011570588685572147, + 0.0055175586603581905, + 0.008767928928136826, + 0.006789578590542078, + 0.004968418274074793, + 0.005906012374907732, + 0.005484618246555328, + 0.005511704366654158, + 0.007577024400234222, + 0.006303237751126289, + 0.0057164509780704975, + 0.005847482476383448, + 0.012423688545823097, + 0.013318276964128017, + 0.022880150005221367, + 0.014447895810008049, + 0.017719309777021408, + 0.015616792254149914, + 0.012362180277705193, + 0.02027694508433342, + 0.0028197169303894043, + 0.0029921059031039476, + 0.0038836521562188864, + 0.004737499635666609, + 0.005047909449785948, + 0.0037154778838157654, + 0.003776347730308771, + 0.003295593662187457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0059498813934624195 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863649606704712 + ], + "opposite_logprob_tokens": [ + -3.0486364364624023 + ], + "logprob_reward": -0.04863649606704712, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863649606704712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19682830572128296, + 0.08146780729293823, + 0.10887560993432999, + 0.10404244810342789, + 0.16298158466815948, + 0.08847782015800476, + 0.05784100294113159, + 0.0827159732580185, + 0.04795412719249725, + 0.06881531327962875 + ], + "credit_map": [ + 0.019434496760368347, + 0.022619308903813362, + 0.029421208426356316, + 0.024827586486935616, + 0.01682189851999283, + 0.024804387241601944, + 0.029070544987916946, + 0.029828881844878197, + 0.008416754193603992, + 0.00951636303216219, + 0.009785838425159454, + 0.012469870038330555, + 0.008079084567725658, + 0.01219459343701601, + 0.009213712066411972, + 0.011791594326496124, + 0.009178614243865013, + 0.01066116988658905, + 0.016519883647561073, + 0.02065293677151203, + 0.010463286191225052, + 0.011404193937778473, + 0.011548172682523727, + 0.018447354435920715, + 0.007980572059750557, + 0.010060769505798817, + 0.013759260065853596, + 0.0182372834533453, + 0.014356909319758415, + 0.011430718936026096, + 0.009702041745185852, + 0.01851489581167698, + 0.012118838727474213, + 0.011034488677978516, + 0.023900499567389488, + 0.026187710464000702, + 0.013930775225162506, + 0.01448444277048111, + 0.018693137913942337, + 0.04263169690966606, + 0.004633090924471617, + 0.004420590121299028, + 0.01422642357647419, + 0.022912483662366867, + 0.00540481973439455, + 0.00480562774464488, + 0.016938231885433197, + 0.01513654924929142, + 0.004399423953145742, + 0.0045088366605341434, + 0.00971065554767847, + 0.011561824940145016, + 0.005016814451664686, + 0.005926159210503101, + 0.007594794034957886, + 0.009122498333454132, + 0.0072135054506361485, + 0.010533610358834267, + 0.009000904858112335, + 0.011891791597008705, + 0.014115317724645138, + 0.01132959220558405, + 0.008708440698683262, + 0.009922805242240429, + 0.004735312890261412, + 0.0032643526792526245, + 0.009686156176030636, + 0.006630150601267815, + 0.002959475852549076, + 0.004746959079056978, + 0.004666471853852272, + 0.011265247128903866, + 0.006902329158037901, + 0.005317981820553541, + 0.010381891392171383, + 0.010722770355641842, + 0.007073173299431801, + 0.00862472876906395, + 0.008079742081463337, + 0.011712699197232723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.04863649606704712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot establishes a remarkably tranquil seascape from the perspective of the coast, looking directly out to sea at sunset. In the immediate foreground, the coastline offers a rich texture, composed of damp, dark sand, interspersed with a variety of smooth, wet rocks and colorful pebbles that shimmer under the night sky. The central region is dominated by the deep blue ocean, its surface brimming with gentle, relentless energy. Small waves rhythmically roll towards the land, their crests catching the last rays of light in shimmering golden ripples. On the distant horizon, the sun, a fiery, molten orb, is partially obscured as it sinks majestically, casting a spectacular, warm glow across the sky in a gradation of fiery orange, deep burgundy, and soft purple. This golden, hour-long light creates a shimmering path across the water, guiding the viewer's gaze from shore to horizon, bathing the entire scene in a warm, tranquil glow, evoking a powerful atmosphere of quiet contemplation and the vast beauty of nature.", + "seed": 25410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000081/group_9/index_4892_seed_25410_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11266297791153193, + "vlm_logprob": -0.11266297791153193, + "vlm_yes_no_margin": 4.049999994039536, + "vlm_token_credit": -0.11266297791153193, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005444697570055723 + ], + "opposite_logprob_tokens": [ + -7.130444526672363 + ], + "logprob_reward": -0.005444697570055723, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.005444697570055723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26630839705467224, + 0.37362468242645264, + 0.2977394163608551, + 0.005541335325688124, + 0.0075495862402021885, + 0.013750506564974785, + 0.007494291756302118, + 0.007145389448851347, + 0.005952645093202591, + 0.014893774874508381 + ], + "credit_map": [ + 0.02552114985883236, + 0.10826543718576431, + 0.043895211070775986, + 0.01160410325974226, + 0.031683020293712616, + 0.033553726971149445, + 0.004553097765892744, + 0.007232641335576773, + 0.033217087388038635, + 0.1523606926202774, + 0.0658402293920517, + 0.015719451010227203, + 0.045820917934179306, + 0.04498288035392761, + 0.005486905109137297, + 0.01019652932882309, + 0.025147587060928345, + 0.1228935495018959, + 0.05342358723282814, + 0.012624901719391346, + 0.036841247230768204, + 0.03525647893548012, + 0.004187438637018204, + 0.007364614866673946, + 0.000674661248922348, + 0.0006915866979397833, + 0.001149065443314612, + 0.0006305421120487154, + 0.00047933391761034727, + 0.0004178688977845013, + 0.0008750256965868175, + 0.0006232511368580163, + 0.0006011268706060946, + 0.002555882791057229, + 0.0015742433024570346, + 0.00041736135608516634, + 0.0008152168011292815, + 0.000835952116176486, + 0.00036062573781237006, + 0.00038917717756703496, + 0.0010919772321358323, + 0.005051544401794672, + 0.0009111406398005784, + 0.0009628369007259607, + 0.0010262526338919997, + 0.0035802132915705442, + 0.0005589924403466284, + 0.0005675486172549427, + 0.0005118819535709918, + 0.0007037709001451731, + 0.001957409083843231, + 0.0007437086314894259, + 0.0009028416825458407, + 0.0012700827792286873, + 0.0004672073118854314, + 0.0009373901993967593, + 0.0006073248805478215, + 0.0009223836241289973, + 0.0008263237541541457, + 0.001270614448003471, + 0.0007208934985101223, + 0.0010955041507259011, + 0.0007342125754803419, + 0.0009681329247541726, + 0.0005378052592277527, + 0.000642719678580761, + 0.0008740796474739909, + 0.0013996647903695703, + 0.0006893111858516932, + 0.00046832620864734054, + 0.0006678907666355371, + 0.000672847731038928, + 0.0012201792560517788, + 0.003177032107487321, + 0.003912344109266996, + 0.0017184100579470396, + 0.0017793569713830948, + 0.0007918067858554423, + 0.0009526086505502462, + 0.001342037576250732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.005444697570055723 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628357969224453 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628357969224453, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628357969224453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13170431554317474, + 0.7081643342971802, + 0.09112352132797241, + 0.008187509141862392, + 0.008001328445971012, + 0.0038363162893801928, + 0.02892255038022995, + 0.006416224408894777, + 0.0056105186231434345, + 0.008033353835344315 + ], + "credit_map": [ + 0.01221717894077301, + 0.05254708603024483, + 0.021304132416844368, + 0.006548446137458086, + 0.014787710271775723, + 0.018222978338599205, + 0.0025056845042854548, + 0.003571105655282736, + 0.06317195296287537, + 0.2889919579029083, + 0.12452712655067444, + 0.029463032260537148, + 0.08708498626947403, + 0.08562254160642624, + 0.010114524513483047, + 0.01918824389576912, + 0.00781282503157854, + 0.03762264922261238, + 0.016250701621174812, + 0.0037968142423778772, + 0.011180034838616848, + 0.010960184969007969, + 0.0012255235342308879, + 0.0022747956681996584, + 0.0007570988382212818, + 0.0022226134315133095, + 0.0015399276744574308, + 0.0008578153210692108, + 0.0008763788500800729, + 0.000732100976165384, + 0.0005864784470759332, + 0.0006150968256406486, + 0.0008054382633417845, + 0.0029487754218280315, + 0.0006113023264333606, + 0.0005946176243014634, + 0.0007190883625298738, + 0.0015718576032668352, + 0.0004160561948083341, + 0.000334192969603464, + 0.00032395662856288254, + 0.0009820626582950354, + 0.00038005414535291493, + 0.00040023107430897653, + 0.000361116137355566, + 0.0006535360589623451, + 0.00040396430995315313, + 0.0003313954803161323, + 0.002507084282115102, + 0.010754610411822796, + 0.0051459455862641335, + 0.0017569675110280514, + 0.003097579348832369, + 0.0033433297649025917, + 0.0007775294361636043, + 0.0015395063674077392, + 0.0005465735448524356, + 0.0011809978168457747, + 0.0009188798139803112, + 0.0009970910614356399, + 0.000517427222803235, + 0.00076005031587556, + 0.0005043691489845514, + 0.0009908359497785568, + 0.0005138632841408253, + 0.0005310405977070332, + 0.0008619693689979613, + 0.0009282837272621691, + 0.0010000050533562899, + 0.0005239369929768145, + 0.0005477743688970804, + 0.0007036456954665482, + 0.0005381011287681758, + 0.00206318823620677, + 0.0018632252467796206, + 0.0008032378391362727, + 0.0008895770297385752, + 0.0007385974167846143, + 0.0004099101643078029, + 0.000727517472114414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007628357969224453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742983877658844 + ], + "opposite_logprob_tokens": [ + -0.974298357963562 + ], + "logprob_reward": -0.4742983877658844, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.4742983877658844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12088917940855026, + 0.30361753702163696, + 0.44070327281951904, + 0.005633760709315538, + 0.009235681034624577, + 0.015357482247054577, + 0.06128603592514992, + 0.03119567036628723, + 0.003132244572043419, + 0.008949159644544125 + ], + "credit_map": [ + 0.010577564127743244, + 0.04610458016395569, + 0.023352395743131638, + 0.0062419939786195755, + 0.014106003567576408, + 0.01259034126996994, + 0.003347658319398761, + 0.004568636883050203, + 0.02847517840564251, + 0.1255815476179123, + 0.048757657408714294, + 0.0120686711743474, + 0.037438541650772095, + 0.039020370692014694, + 0.004431873094290495, + 0.007843692786991596, + 0.03749031946063042, + 0.18214981257915497, + 0.07754016667604446, + 0.01857510767877102, + 0.054234493523836136, + 0.05329441279172897, + 0.00647769495844841, + 0.01094123162329197, + 0.0007323267636820674, + 0.0011242999462410808, + 0.0006101441103965044, + 0.0006177960312925279, + 0.0006869505159556866, + 0.000784504518378526, + 0.0005980960559099913, + 0.0004796424473170191, + 0.0007452101563103497, + 0.002999946940690279, + 0.0019363875035196543, + 0.000650134461466223, + 0.0012762994738295674, + 0.0006030318909324706, + 0.0005566360196098685, + 0.0004680337733589113, + 0.0014737340388819575, + 0.004962962586432695, + 0.00128475995734334, + 0.0014340459601953626, + 0.0010117217898368835, + 0.0033798301592469215, + 0.0009996341541409492, + 0.0008107930771075189, + 0.005122053436934948, + 0.02347891964018345, + 0.011659104377031326, + 0.003013694193214178, + 0.00614372082054615, + 0.007899653166532516, + 0.0010271633509546518, + 0.00294171879068017, + 0.0029421462677419186, + 0.0034165491815656424, + 0.004938800353556871, + 0.007626960054039955, + 0.0013662438141182065, + 0.0027304384857416153, + 0.003841996658593416, + 0.004332532174885273, + 0.00026283867191523314, + 0.00031605930416844785, + 0.00048344788956455886, + 0.0007418438908644021, + 0.00034045433858409524, + 0.000259849097346887, + 0.0003670593723654747, + 0.0003606919781304896, + 0.0006135930307209492, + 0.0025607021525502205, + 0.001959343208000064, + 0.0005183679168112576, + 0.0011866510612890124, + 0.0009908730862662196, + 0.0003624791861511767, + 0.0007571497117169201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.4742983877658844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949525628238916 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949525628238916, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949525628238916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15143044292926788, + 0.4102685749530792, + 0.3398948907852173, + 0.01854041963815689, + 0.018952040001749992, + 0.010356427170336246, + 0.015083994716405869, + 0.016871169209480286, + 0.011942670680582523, + 0.006659423466771841 + ], + "credit_map": [ + 0.012811598367989063, + 0.0591759979724884, + 0.027026016265153885, + 0.007461809553205967, + 0.016947191208600998, + 0.019777826964855194, + 0.003588154213503003, + 0.0046418411657214165, + 0.03650606423616409, + 0.16692006587982178, + 0.07264290750026703, + 0.017435461282730103, + 0.05039225146174431, + 0.04907296970486641, + 0.005987850017845631, + 0.01131096575409174, + 0.028293998911976814, + 0.13884621858596802, + 0.06258097290992737, + 0.014957226812839508, + 0.04219832643866539, + 0.03877418115735054, + 0.005448466632515192, + 0.008795500732958317, + 0.002243640134111047, + 0.005333768203854561, + 0.0015104170888662338, + 0.0012077607680112123, + 0.002593665150925517, + 0.002971164183691144, + 0.0014586506877094507, + 0.001221353537403047, + 0.001859170151874423, + 0.006992521695792675, + 0.0008124089799821377, + 0.0010188791202381253, + 0.0019903541542589664, + 0.0047373222187161446, + 0.0007462830399163067, + 0.0007951011648401618, + 0.0011649209773167968, + 0.004014354199171066, + 0.0006743228295817971, + 0.0005364935495890677, + 0.0013319734716787934, + 0.0018539840821176767, + 0.0003730252501554787, + 0.00040735240327194333, + 0.001452550059184432, + 0.00252262013964355, + 0.001926749711856246, + 0.000845397065859288, + 0.0033581077586859465, + 0.003129855263978243, + 0.0007888413383625448, + 0.0010598736116662621, + 0.001268132240511477, + 0.0030736129265278578, + 0.003033999353647232, + 0.0018610835541039705, + 0.0012393799843266606, + 0.002073694486171007, + 0.001274753361940384, + 0.0030465126037597656, + 0.0010892554419115186, + 0.0012623472139239311, + 0.0016971302684396505, + 0.0012552967527881265, + 0.0025306339375674725, + 0.0016163596883416176, + 0.001186475856229663, + 0.0013051717542111874, + 0.0008693173294886947, + 0.0010570461163297296, + 0.0007582420948892832, + 0.0012388289906084538, + 0.0006715989438816905, + 0.000618069723714143, + 0.0007405723445117474, + 0.0007057475158944726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.005949525628238916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a seascape viewed from the coast looking out to sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground composed of damp dark sand with smooth wet rocks and colorful pebbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean deep blue with small waves rolling towards the land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sun a fiery molten orb partially obscured on the horizon casting orange, burgundy, and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static low-angle wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999392062425613 + ], + "opposite_logprob_tokens": [ + -2.6949939727783203 + ], + "logprob_reward": -0.06999392062425613, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999392062425613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12064384669065475, + 0.14542046189308167, + 0.4801696240901947, + 0.02306152880191803, + 0.04357555881142616, + 0.013775003142654896, + 0.1137724220752716, + 0.01507066935300827, + 0.01004859060049057, + 0.034462280571460724 + ], + "credit_map": [ + 0.010137625969946384, + 0.043065205216407776, + 0.020857715979218483, + 0.00835734885185957, + 0.012375625781714916, + 0.014137420803308487, + 0.005308195017278194, + 0.0064047216437757015, + 0.012020568363368511, + 0.05805473402142525, + 0.027533547952771187, + 0.0065502286888659, + 0.017708735540509224, + 0.01659129559993744, + 0.0025812829844653606, + 0.0043800766579806805, + 0.041463352739810944, + 0.19818823039531708, + 0.08282755315303802, + 0.020517311990261078, + 0.0590549111366272, + 0.05820488557219505, + 0.0077745490707457066, + 0.01213890966027975, + 0.002920908387750387, + 0.005339927971363068, + 0.0029269284568727016, + 0.002251456957310438, + 0.0025109953712671995, + 0.0021879966370761395, + 0.0026419349014759064, + 0.0022813817486166954, + 0.003901904681697488, + 0.016699206084012985, + 0.005629369989037514, + 0.0018466226756572723, + 0.004314746242016554, + 0.007738030049949884, + 0.0016269379993900657, + 0.0018187419045716524, + 0.0013726179022341967, + 0.0026131835766136646, + 0.001690031262114644, + 0.0016651332844048738, + 0.0016323167365044355, + 0.002562304027378559, + 0.0012034437386319041, + 0.001035974477417767, + 0.010073372162878513, + 0.03947422280907631, + 0.02378152683377266, + 0.0075124879367649555, + 0.015912333503365517, + 0.0068003092892467976, + 0.003231084905564785, + 0.006987091153860092, + 0.0012166450032964349, + 0.002021754626184702, + 0.002154604997485876, + 0.002809632569551468, + 0.0013217766536399722, + 0.002006388967856765, + 0.0016564320540055633, + 0.0018834356451407075, + 0.0009776828810572624, + 0.0012595581356436014, + 0.0011787889525294304, + 0.001321172807365656, + 0.0017809945857152343, + 0.0011640770826488733, + 0.0009183117654174566, + 0.0014480049721896648, + 0.0028323440346866846, + 0.009832269512116909, + 0.008122614584863186, + 0.0036758268252015114, + 0.00377348973415792, + 0.0019982713274657726, + 0.0016705357702448964, + 0.0025569309946149588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.06999392062425613 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_0/index_tensor([9465, 9465])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_0/index_tensor([9465, 9465])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3e2a3b03987696441b76218511d87980b586bd11 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_0/index_tensor([9465, 9465])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from an extremely low angle, using high-speed photography to create a slow-motion effect, focuses on an old, dark brown wooden dining chair, its paint chipped, careening down at breakneck speed onto a rough, dusty concrete floor. The moment the chair legs impact the ground, the immense force of the impact causes it to shatter, sending countless shards of wood, sawdust, and dust scattering in all directions from the point of impact like a volcanic eruption. The slow motion capture clearly reveals the trajectory of each fragment, tumbling and spinning through the air, creating a still image imbued with a violent aesthetic. A cold, overhead spotlight pierces the dense dust, sharply illuminating the flying fragments and casting rapidly shifting shadows across their irregular surfaces. The scene's grim tone evokes a sense of primal destructive force and a momentary tension. The camera vibrates slightly yet violently at the moment of impact, further heightening the impact's realism.", + "seed": 28033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_0/index_9465_seed_28033_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04156634993851185, + "vlm_logprob": -0.04156634993851185, + "vlm_yes_no_margin": 3.824999976158142, + "vlm_token_credit": -0.04156634993851185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14489349722862244 + ], + "opposite_logprob_tokens": [ + -2.0198934078216553 + ], + "logprob_reward": -0.14489349722862244, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14489349722862244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1689981073141098, + 0.16022823750972748, + 0.05796414986252785, + 0.07966940104961395, + 0.06376610696315765, + 0.12324705719947815, + 0.13562873005867004, + 0.07661090791225433, + 0.07476159930229187, + 0.05912565439939499 + ], + "credit_map": [ + 0.02344018593430519, + 0.028272224590182304, + 0.02029602788388729, + 0.015246826224029064, + 0.02697703056037426, + 0.02653786912560463, + 0.016234038397669792, + 0.011993897147476673, + 0.018428977578878403, + 0.021457089111208916, + 0.016439486294984818, + 0.017613694071769714, + 0.01928676851093769, + 0.04465564712882042, + 0.012424229644238949, + 0.009922358207404613, + 0.00891341082751751, + 0.00798049196600914, + 0.004610504489392042, + 0.013115499168634415, + 0.005071832798421383, + 0.008666264824569225, + 0.0053032622672617435, + 0.004302884452044964, + 0.009426429867744446, + 0.013287506066262722, + 0.0075152479112148285, + 0.010452337563037872, + 0.01563887670636177, + 0.01048283651471138, + 0.006486875005066395, + 0.006379290018230677, + 0.006022085901349783, + 0.00633398350328207, + 0.004952057730406523, + 0.006613980047404766, + 0.009625486098229885, + 0.012070070020854473, + 0.0067292917519807816, + 0.011419148184359074, + 0.011265037581324577, + 0.017840100452303886, + 0.022011011838912964, + 0.015159975737333298, + 0.01911209709942341, + 0.012213052250444889, + 0.016957338899374008, + 0.008688452653586864, + 0.017207153141498566, + 0.01735648699104786, + 0.02237563021481037, + 0.013822576031088829, + 0.013577575795352459, + 0.026161879301071167, + 0.01059031579643488, + 0.014537113718688488, + 0.008337636478245258, + 0.008039114996790886, + 0.00876726396381855, + 0.0100333783775568, + 0.009591156616806984, + 0.009803560562431812, + 0.012665041722357273, + 0.009373757056891918, + 0.008462058380246162, + 0.012496084906160831, + 0.007052498869597912, + 0.007720906287431717, + 0.010285802185535431, + 0.012981683015823364, + 0.007240696810185909, + 0.008521869778633118, + 0.005525190848857164, + 0.006824519485235214, + 0.013234521262347698, + 0.005483380984514952, + 0.009999380446970463, + 0.006981773767620325, + 0.0052024307660758495, + 0.005874456837773323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.14489349722862244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011115296743810177 + ], + "opposite_logprob_tokens": [ + -4.511115074157715 + ], + "logprob_reward": -0.011115296743810177, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011115296743810177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2517334222793579, + 0.09131548553705215, + 0.0464821495115757, + 0.09644932299852371, + 0.08070546388626099, + 0.06648227572441101, + 0.18758231401443481, + 0.09251939505338669, + 0.047905005514621735, + 0.03882521763443947 + ], + "credit_map": [ + 0.03704351559281349, + 0.03513110429048538, + 0.05430973693728447, + 0.034866563975811005, + 0.018881898373365402, + 0.020665660500526428, + 0.03224583715200424, + 0.018589092418551445, + 0.011544247157871723, + 0.011030875146389008, + 0.011629779823124409, + 0.010918554849922657, + 0.013145359233021736, + 0.018701286986470222, + 0.006811760365962982, + 0.007533614989370108, + 0.0077458592131733894, + 0.006091988179832697, + 0.0038969172164797783, + 0.0072563206776976585, + 0.00441858172416687, + 0.007272315211594105, + 0.005107769276946783, + 0.0046923961490392685, + 0.0100295664742589, + 0.014161502942442894, + 0.007567699998617172, + 0.009541632607579231, + 0.021971801295876503, + 0.016783930361270905, + 0.008137396536767483, + 0.008255796507000923, + 0.008136619813740253, + 0.009032412432134151, + 0.006610070820897818, + 0.013407331891357899, + 0.01050614845007658, + 0.010487405583262444, + 0.011112886480987072, + 0.011412590742111206, + 0.00678580766543746, + 0.008765746839344501, + 0.007274425122886896, + 0.006083714310079813, + 0.009740319103002548, + 0.015290102921426296, + 0.005528128705918789, + 0.007014031987637281, + 0.01930549181997776, + 0.033393315970897675, + 0.02634473145008087, + 0.024833420291543007, + 0.026956707239151, + 0.022790908813476562, + 0.017307287082076073, + 0.01665046066045761, + 0.010234396904706955, + 0.010065100155770779, + 0.017235001549124718, + 0.014323992654681206, + 0.010169513523578644, + 0.009796488098800182, + 0.01288528274744749, + 0.007809618022292852, + 0.005178085528314114, + 0.0065782563760876656, + 0.00386615376919508, + 0.0054325382225215435, + 0.005610163789242506, + 0.007584266364574432, + 0.006812739185988903, + 0.006842803210020065, + 0.0036816855426877737, + 0.004501021467149258, + 0.006970666814595461, + 0.0043638646602630615, + 0.007887547835707664, + 0.005003451835364103, + 0.0029958609957247972, + 0.0034211166203022003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.011115296743810177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253794603049755 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253794603049755, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253794603049755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17145462334156036, + 0.04014257714152336, + 0.0415971465408802, + 0.045319125056266785, + 0.03780956193804741, + 0.15513110160827637, + 0.41369399428367615, + 0.03063824772834778, + 0.03968184441328049, + 0.024531785398721695 + ], + "credit_map": [ + 0.03325087949633598, + 0.023682430386543274, + 0.033756885677576065, + 0.019220158457756042, + 0.019443050026893616, + 0.014083394780755043, + 0.017355695366859436, + 0.01066211611032486, + 0.005091310013085604, + 0.004412179812788963, + 0.004835260100662708, + 0.006050714757293463, + 0.004640741739422083, + 0.00812644325196743, + 0.003346715122461319, + 0.003639208385720849, + 0.005475769285112619, + 0.005676679778844118, + 0.0029948630835860968, + 0.009004989638924599, + 0.004109290894120932, + 0.00402843439951539, + 0.005556093528866768, + 0.004751025699079037, + 0.005044715944677591, + 0.006201548036187887, + 0.0033524991013109684, + 0.0056977542117238045, + 0.009470202028751373, + 0.005823563318699598, + 0.0045782653614878654, + 0.005150573793798685, + 0.0030508169438689947, + 0.003962273243814707, + 0.003007556777447462, + 0.004582878667861223, + 0.005997917614877224, + 0.00553106889128685, + 0.003490624949336052, + 0.008186421357095242, + 0.01786649227142334, + 0.020332934334874153, + 0.034212857484817505, + 0.017337651923298836, + 0.021572111174464226, + 0.017536776140332222, + 0.017543340101838112, + 0.0087289372459054, + 0.03479718416929245, + 0.09534929692745209, + 0.03326529636979103, + 0.03997431695461273, + 0.09123571217060089, + 0.052484508603811264, + 0.03733216971158981, + 0.029255490750074387, + 0.0028774128295481205, + 0.004691106732934713, + 0.002513596322387457, + 0.003397194668650627, + 0.0043902043253183365, + 0.004564798902720213, + 0.006142682861536741, + 0.0020612499210983515, + 0.0037019439041614532, + 0.006972821429371834, + 0.006223250646144152, + 0.003942872863262892, + 0.0045972890220582485, + 0.004278336651623249, + 0.00523301912471652, + 0.0047323089092969894, + 0.002273018704727292, + 0.00402083108201623, + 0.003389537800103426, + 0.00175410695374012, + 0.0069878846406936646, + 0.002336686011403799, + 0.001988679403439164, + 0.001781037775799632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.005253794603049755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328725904226303 + ], + "opposite_logprob_tokens": [ + -3.773287296295166 + ], + "logprob_reward": -0.02328725904226303, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328725904226303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2769812345504761, + 0.03366074338555336, + 0.029908396303653717, + 0.027902238070964813, + 0.05868847295641899, + 0.3606572449207306, + 0.036905743181705475, + 0.01276133581995964, + 0.021469568833708763, + 0.14106504619121552 + ], + "credit_map": [ + 0.07037637382745743, + 0.03156641498208046, + 0.09354716539382935, + 0.02522297389805317, + 0.01783890463411808, + 0.015940172597765923, + 0.010799935087561607, + 0.011689301580190659, + 0.003884785808622837, + 0.004036357626318932, + 0.00578529667109251, + 0.004079796373844147, + 0.0027202016208320856, + 0.0066151912324130535, + 0.0038391826674342155, + 0.002699932549148798, + 0.004471858497709036, + 0.0030589911621063948, + 0.0036301000509411097, + 0.007391586899757385, + 0.0025540499482303858, + 0.0037210695445537567, + 0.0023263960611075163, + 0.0027543443720787764, + 0.003638666123151779, + 0.002755230525508523, + 0.0025863342452794313, + 0.0038641931023448706, + 0.005130154546350241, + 0.0035113156773149967, + 0.0029632514342665672, + 0.003453092183917761, + 0.009355402551591396, + 0.008541078306734562, + 0.007160651497542858, + 0.0059480625204741955, + 0.007933170534670353, + 0.0076845563016831875, + 0.004964911844581366, + 0.007100637070834637, + 0.03136492893099785, + 0.05518233776092529, + 0.0393947996199131, + 0.027602627873420715, + 0.062180209904909134, + 0.0785745233297348, + 0.033185530453920364, + 0.033172279596328735, + 0.00428026495501399, + 0.004973277449607849, + 0.00468028150498867, + 0.003648853860795498, + 0.005018331576138735, + 0.0064606270752847195, + 0.0038944853004068136, + 0.003949620760977268, + 0.001266229897737503, + 0.001691906712949276, + 0.0020662746392190456, + 0.0015467533376067877, + 0.0016695924568921328, + 0.0017971595516428351, + 0.0016823820769786835, + 0.0010410381946712732, + 0.0023837059270590544, + 0.0041695283725857735, + 0.0029022516682744026, + 0.0024754954501986504, + 0.0030074843671172857, + 0.0023287541698664427, + 0.002101609017699957, + 0.002100741257891059, + 0.012698634527623653, + 0.02010701410472393, + 0.027974573895335197, + 0.010490325279533863, + 0.0253627710044384, + 0.015267676673829556, + 0.0132321547716856, + 0.01593189872801304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.02328725904226303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281902074813843 + ], + "opposite_logprob_tokens": [ + -3.7732818126678467 + ], + "logprob_reward": -0.023281902074813843, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281902074813843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25282955169677734, + 0.10196539759635925, + 0.06451581418514252, + 0.04771645739674568, + 0.04430394247174263, + 0.15398560464382172, + 0.1410295069217682, + 0.05813690274953842, + 0.05483347550034523, + 0.0806833952665329 + ], + "credit_map": [ + 0.04305214062333107, + 0.04032411426305771, + 0.042787883430719376, + 0.023950379341840744, + 0.023282799869775772, + 0.025832073763012886, + 0.03630805015563965, + 0.017292063683271408, + 0.011788222938776016, + 0.012091412208974361, + 0.01468473020941019, + 0.01343311183154583, + 0.013710331171751022, + 0.019453654065728188, + 0.00921754539012909, + 0.007586383260786533, + 0.010070284828543663, + 0.007258748635649681, + 0.006481233984231949, + 0.015485179610550404, + 0.006447940599173307, + 0.006993390619754791, + 0.006552540697157383, + 0.00522648636251688, + 0.005381154827773571, + 0.006236822810024023, + 0.004687903914600611, + 0.007178077008575201, + 0.00579595984891057, + 0.005767974071204662, + 0.006540124304592609, + 0.006128434557467699, + 0.004476030357182026, + 0.0039038867689669132, + 0.00367949903011322, + 0.0051390561275184155, + 0.00644217524677515, + 0.007610219996422529, + 0.006399319972842932, + 0.006653747987002134, + 0.01866472326219082, + 0.01861797831952572, + 0.03212394192814827, + 0.01688661426305771, + 0.017836421728134155, + 0.021184522658586502, + 0.017133820801973343, + 0.011537553742527962, + 0.01431437861174345, + 0.026541532948613167, + 0.009675640612840652, + 0.013490984216332436, + 0.025418274104595184, + 0.025078797712922096, + 0.01753629557788372, + 0.008973600342869759, + 0.0060434043407440186, + 0.006782072596251965, + 0.0045577227137982845, + 0.007668902166187763, + 0.007683750241994858, + 0.007617647293955088, + 0.012019066140055656, + 0.005764336790889502, + 0.006562430411577225, + 0.00884647760540247, + 0.005066452547907829, + 0.006188775412738323, + 0.006192534696310759, + 0.008407202549278736, + 0.006265555042773485, + 0.007304041180759668, + 0.007217391859740019, + 0.009197157807648182, + 0.01690610684454441, + 0.006586809176951647, + 0.01911161281168461, + 0.008056063205003738, + 0.005064839962869883, + 0.008543407544493675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.023281902074813843 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, using high-speed photography to create a slow-motion effect, focuses on an old, dark brown wooden dining chair, its paint chipped, careening down at breakneck speed onto a rough, dusty concrete floor. The moment the chair legs impact the ground, the immense force of the impact causes it to shatter, sending countless shards of wood, sawdust, and dust scattering in all directions from the point of impact like a volcanic eruption. The slow motion capture clearly reveals the trajectory of each fragment, tumbling and spinning through the air, creating a still image imbued with a violent aesthetic. A cold, overhead spotlight pierces the dense dust, sharply illuminating the flying fragments and casting rapidly shifting shadows across their irregular surfaces. The scene's grim tone evokes a sense of primal destructive force and a momentary tension. The camera vibrates slightly yet violently at the moment of impact, further heightening the impact's realism.", + "seed": 28034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_0/index_9465_seed_28034_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06017154436558485, + "vlm_logprob": -0.06017154436558485, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.06017154436558485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06787180155515671 + ], + "opposite_logprob_tokens": [ + -2.8178718090057373 + ], + "logprob_reward": -0.06787180155515671, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06787180155515671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27752748131752014, + 0.044702641665935516, + 0.08864494413137436, + 0.19952011108398438, + 0.07335221022367477, + 0.13105632364749908, + 0.047997377812862396, + 0.05372197553515434, + 0.04104238748550415, + 0.042434535920619965 + ], + "credit_map": [ + 0.029125604778528214, + 0.04936124011874199, + 0.0364719033241272, + 0.03828180208802223, + 0.04162202775478363, + 0.03843527287244797, + 0.023176534101366997, + 0.021053101867437363, + 0.004206203389912844, + 0.00517272250726819, + 0.005166442599147558, + 0.00597843062132597, + 0.004197630099952221, + 0.007632540538907051, + 0.0065220738761126995, + 0.00582659849897027, + 0.011097660288214684, + 0.00901633221656084, + 0.009712428785860538, + 0.012297414243221283, + 0.017825307324528694, + 0.008206672966480255, + 0.012273252010345459, + 0.008215879090130329, + 0.034213390201330185, + 0.03215893357992172, + 0.05020253732800484, + 0.03881938010454178, + 0.013149282895028591, + 0.012696649879217148, + 0.009818601422011852, + 0.008461330085992813, + 0.009755546227097511, + 0.010134105570614338, + 0.00577602768316865, + 0.007465175818651915, + 0.0068540023639798164, + 0.013420689851045609, + 0.015513540245592594, + 0.004433119669556618, + 0.015735048800706863, + 0.017978716641664505, + 0.02351750247180462, + 0.02499418705701828, + 0.011064370162785053, + 0.012933927588164806, + 0.013241183944046497, + 0.011591393500566483, + 0.007373904809355736, + 0.0061365398578345776, + 0.0070383669808506966, + 0.005063406191766262, + 0.004397437442094088, + 0.00882199127227068, + 0.004301689565181732, + 0.004864042159169912, + 0.010417950339615345, + 0.0070410557091236115, + 0.006890753284096718, + 0.007096235174685717, + 0.005135754123330116, + 0.005058222450315952, + 0.004404047504067421, + 0.0076779574155807495, + 0.00735692773014307, + 0.005601337645202875, + 0.005432464182376862, + 0.006324823945760727, + 0.005074767861515284, + 0.003212890587747097, + 0.004786086268723011, + 0.0032530860044062138, + 0.007599055301398039, + 0.005251305643469095, + 0.004627612419426441, + 0.005906295031309128, + 0.0035631367936730385, + 0.007032206282019615, + 0.004245288670063019, + 0.004209632519632578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.06787180155515671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036908835172653 + ], + "opposite_logprob_tokens": [ + -2.3503689765930176 + ], + "logprob_reward": -0.10036908835172653, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036908835172653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.113280288875103, + 0.038724299520254135, + 0.10618002712726593, + 0.3046349883079529, + 0.16121253371238708, + 0.07085296511650085, + 0.07100388407707214, + 0.08136594295501709, + 0.025305738672614098, + 0.027439318597316742 + ], + "credit_map": [ + 0.012157462537288666, + 0.012022317387163639, + 0.01608702354133129, + 0.014694936573505402, + 0.016398955136537552, + 0.02160346508026123, + 0.010357547551393509, + 0.00995858758687973, + 0.0053210691548883915, + 0.0036203954368829727, + 0.004697971045970917, + 0.004613358061760664, + 0.00435468927025795, + 0.005981097463518381, + 0.004167912062257528, + 0.0059678079560399055, + 0.014572007581591606, + 0.014650107361376286, + 0.009784248657524586, + 0.01396248023957014, + 0.02172059565782547, + 0.00960682425647974, + 0.012335163541138172, + 0.009548601694405079, + 0.06005660444498062, + 0.05625087767839432, + 0.062262315303087234, + 0.0494835264980793, + 0.017495285719633102, + 0.02590268664062023, + 0.017202066257596016, + 0.01598164811730385, + 0.024501647800207138, + 0.030349869281053543, + 0.01080742385238409, + 0.016691990196704865, + 0.014027615077793598, + 0.021145975217223167, + 0.03175764158368111, + 0.011930366978049278, + 0.006597637198865414, + 0.009161626920104027, + 0.015266316942870617, + 0.013282573781907558, + 0.00449155829846859, + 0.006455480121076107, + 0.00867223646491766, + 0.0069255344569683075, + 0.009870411828160286, + 0.008806402795016766, + 0.014570721425116062, + 0.010649558156728745, + 0.007055635564029217, + 0.005905277095735073, + 0.006761117838323116, + 0.007384755183011293, + 0.013958071358501911, + 0.013013243675231934, + 0.011854175478219986, + 0.014219244942069054, + 0.00554331811144948, + 0.004603235516697168, + 0.00836244784295559, + 0.009812208823859692, + 0.004197929985821247, + 0.0031833869870752096, + 0.003156626597046852, + 0.004496149718761444, + 0.003302759723737836, + 0.0025915338192135096, + 0.0019935043528676033, + 0.0023838477209210396, + 0.004290981683880091, + 0.0032631470821797848, + 0.004729323089122772, + 0.005315771792083979, + 0.0023682410828769207, + 0.0021564571652561426, + 0.003013689536601305, + 0.0023017069324851036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.10036908835172653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632735185325146 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632735185325146, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632735185325146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0781666487455368, + 0.0497712567448616, + 0.08585225045681, + 0.339036226272583, + 0.09137115627527237, + 0.11009600013494492, + 0.06107869744300842, + 0.08219216018915176, + 0.0740472674369812, + 0.028388289734721184 + ], + "credit_map": [ + 0.012979639694094658, + 0.00929417833685875, + 0.008394994772970676, + 0.008725008927285671, + 0.009248395450413227, + 0.013826311565935612, + 0.008629155345261097, + 0.007068968378007412, + 0.00803457759320736, + 0.004354624543339014, + 0.006248099729418755, + 0.007834423333406448, + 0.006158872507512569, + 0.008351911790668964, + 0.005169776268303394, + 0.0036189749371260405, + 0.011726194061338902, + 0.007918743416666985, + 0.009338867850601673, + 0.013483724556863308, + 0.021504981443285942, + 0.006450140383094549, + 0.009381339885294437, + 0.0060482630506157875, + 0.0365193672478199, + 0.06819608062505722, + 0.0426953062415123, + 0.07825712114572525, + 0.027920767664909363, + 0.030293401330709457, + 0.02834501303732395, + 0.02680918201804161, + 0.012400914914906025, + 0.021123826503753662, + 0.0077132415026426315, + 0.007558546029031277, + 0.009859970770776272, + 0.010409427806735039, + 0.01430458389222622, + 0.008000645786523819, + 0.012327833101153374, + 0.015112047083675861, + 0.022395333275198936, + 0.023278014734387398, + 0.009263035841286182, + 0.009735828265547752, + 0.010200186632573605, + 0.0077837370336055756, + 0.007176377810537815, + 0.007353263441473246, + 0.01339658536016941, + 0.011236927472054958, + 0.004765921272337437, + 0.005893966183066368, + 0.0038744944613426924, + 0.007381167262792587, + 0.012724115513265133, + 0.010623142123222351, + 0.014238408766686916, + 0.014010882936418056, + 0.0062720016576349735, + 0.005272442474961281, + 0.0073502701707184315, + 0.011700903065502644, + 0.008298004046082497, + 0.008287516422569752, + 0.011532338336110115, + 0.012675192207098007, + 0.007835088297724724, + 0.005961067974567413, + 0.01001038122922182, + 0.009447692893445492, + 0.005687456112354994, + 0.003427680116146803, + 0.004586461465805769, + 0.004701056983321905, + 0.002419820986688137, + 0.002525119809433818, + 0.0025226145517081022, + 0.0025180818047374487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.007632735185325146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496795102953911 + ], + "opposite_logprob_tokens": [ + -2.9299678802490234 + ], + "logprob_reward": -0.05496795102953911, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496795102953911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14717546105384827, + 0.0920921340584755, + 0.1415206491947174, + 0.16506808996200562, + 0.06764043122529984, + 0.1430818736553192, + 0.0710163563489914, + 0.05379459634423256, + 0.10303936153650284, + 0.015571114607155323 + ], + "credit_map": [ + 0.01763341575860977, + 0.017681220546364784, + 0.01730462536215782, + 0.0186673142015934, + 0.015150495804846287, + 0.03035564348101616, + 0.010987034067511559, + 0.019395694136619568, + 0.010242074728012085, + 0.012895185500383377, + 0.012868016958236694, + 0.010953938588500023, + 0.010541193187236786, + 0.019280988723039627, + 0.006201724521815777, + 0.00910900067538023, + 0.021256349980831146, + 0.02476237528026104, + 0.013637833297252655, + 0.017475053668022156, + 0.02402665466070175, + 0.016168277710676193, + 0.0119136618450284, + 0.01228042971342802, + 0.02447216585278511, + 0.02221732772886753, + 0.04071464017033577, + 0.037792228162288666, + 0.010212826542556286, + 0.011269734241068363, + 0.008703029714524746, + 0.009686109609901905, + 0.010693447664380074, + 0.009789930656552315, + 0.004094903822988272, + 0.0048624202609062195, + 0.005189080256968737, + 0.00838664174079895, + 0.020040400326251984, + 0.004583599511533976, + 0.0154180359095335, + 0.02003345638513565, + 0.030426692217588425, + 0.028338436037302017, + 0.010436169803142548, + 0.011634315364062786, + 0.015345176681876183, + 0.011449575424194336, + 0.009276002645492554, + 0.008363080210983753, + 0.011053048074245453, + 0.008507286198437214, + 0.007034831214696169, + 0.012057868763804436, + 0.008955284021794796, + 0.005768945906311274, + 0.011226645670831203, + 0.00785691849887371, + 0.0075227790512144566, + 0.010119421407580376, + 0.0032585079316049814, + 0.003946408629417419, + 0.004243978764861822, + 0.005619932431727648, + 0.014562243595719337, + 0.012012328952550888, + 0.011508040130138397, + 0.01027553342282772, + 0.016032619401812553, + 0.014805799350142479, + 0.01133597269654274, + 0.012506806291639805, + 0.002577067818492651, + 0.001856088638305664, + 0.001860623829998076, + 0.0015902769519016147, + 0.0014498846139758825, + 0.0025286271702498198, + 0.0018237860640510917, + 0.0018847581231966615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.05496795102953911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001614570617676 + ], + "opposite_logprob_tokens": [ + -2.6950161457061768 + ], + "logprob_reward": -0.07001614570617676, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001614570617676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0669475644826889, + 0.10360404849052429, + 0.19388385117053986, + 0.2576466500759125, + 0.10345567762851715, + 0.11046749353408813, + 0.065077044069767, + 0.05422221124172211, + 0.02510327473282814, + 0.019592177122831345 + ], + "credit_map": [ + 0.009909632615745068, + 0.0062567745335400105, + 0.007577765733003616, + 0.0076769874431192875, + 0.006652181036770344, + 0.014308948069810867, + 0.0063147712498903275, + 0.008250503800809383, + 0.010819272138178349, + 0.006886475253850222, + 0.01584312878549099, + 0.013232412748038769, + 0.010523252189159393, + 0.023468634113669395, + 0.011869269423186779, + 0.010961608029901981, + 0.02295691706240177, + 0.021826518699526787, + 0.0184496957808733, + 0.03021678887307644, + 0.03845364227890968, + 0.017388945445418358, + 0.028701556846499443, + 0.015889788046479225, + 0.03524577617645264, + 0.04683443531394005, + 0.04525526985526085, + 0.0521385557949543, + 0.025735629722476006, + 0.019090503454208374, + 0.02012905478477478, + 0.013217415660619736, + 0.015301070176064968, + 0.02063373103737831, + 0.008155307732522488, + 0.008045073598623276, + 0.011051532812416553, + 0.012519708834588528, + 0.0172730702906847, + 0.010476188734173775, + 0.012059376575052738, + 0.013123827055096626, + 0.024476490914821625, + 0.02447444014251232, + 0.00883440300822258, + 0.009325196035206318, + 0.008822217583656311, + 0.00935154128819704, + 0.008057168684899807, + 0.007617570925503969, + 0.012565618380904198, + 0.011147323995828629, + 0.0046570212580263615, + 0.006856139283627272, + 0.005652911029756069, + 0.008523293770849705, + 0.006514337845146656, + 0.0063305264338850975, + 0.010005421005189419, + 0.012641207315027714, + 0.003744445974007249, + 0.0031226102728396654, + 0.005671058315783739, + 0.006192605476826429, + 0.0043580844067037106, + 0.00281509174965322, + 0.0029069569427520037, + 0.002543763257563114, + 0.0025202513206750154, + 0.003494629403576255, + 0.0036642823833972216, + 0.0028002155013382435, + 0.0026614763773977757, + 0.0019131662556901574, + 0.0022151649463921785, + 0.0032294599805027246, + 0.0014935873914510012, + 0.002573481062427163, + 0.0031659004744142294, + 0.0023399407509714365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.07001614570617676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, using high-speed photography to create a slow-motion effect, focuses on an old, dark brown wooden dining chair, its paint chipped, careening down at breakneck speed onto a rough, dusty concrete floor. The moment the chair legs impact the ground, the immense force of the impact causes it to shatter, sending countless shards of wood, sawdust, and dust scattering in all directions from the point of impact like a volcanic eruption. The slow motion capture clearly reveals the trajectory of each fragment, tumbling and spinning through the air, creating a still image imbued with a violent aesthetic. A cold, overhead spotlight pierces the dense dust, sharply illuminating the flying fragments and casting rapidly shifting shadows across their irregular surfaces. The scene's grim tone evokes a sense of primal destructive force and a momentary tension. The camera vibrates slightly yet violently at the moment of impact, further heightening the impact's realism.", + "seed": 28035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_0/index_9465_seed_28035_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.038184906262904404, + "vlm_logprob": -0.038184906262904404, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.038184906262904404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08087946474552155 + ], + "opposite_logprob_tokens": [ + -2.5808794498443604 + ], + "logprob_reward": -0.08087946474552155, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08087946474552155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10594518482685089, + 0.09042169898748398, + 0.14827492833137512, + 0.03675379604101181, + 0.18495112657546997, + 0.17051972448825836, + 0.05962521210312843, + 0.07565750181674957, + 0.043367236852645874, + 0.0844835713505745 + ], + "credit_map": [ + 0.012468477711081505, + 0.007967312820255756, + 0.02199864760041237, + 0.014061696827411652, + 0.013728526420891285, + 0.012718148529529572, + 0.015474718064069748, + 0.007527651730924845, + 0.016210904344916344, + 0.01400331873446703, + 0.01595756970345974, + 0.0086402278393507, + 0.010971811600029469, + 0.009473821148276329, + 0.00739814480766654, + 0.007765898946672678, + 0.018220072612166405, + 0.024095818400382996, + 0.015060098841786385, + 0.020856577903032303, + 0.014662541449069977, + 0.034032274037599564, + 0.009609395638108253, + 0.011738154105842113, + 0.005605727434158325, + 0.005863251630216837, + 0.004682364407926798, + 0.003757437923923135, + 0.00394921051338315, + 0.004151350352913141, + 0.003925292752683163, + 0.004819161724299192, + 0.02584594301879406, + 0.02407139167189598, + 0.017354246228933334, + 0.01794496923685074, + 0.03942285478115082, + 0.021237559616565704, + 0.02168234996497631, + 0.017391812056303024, + 0.028846940025687218, + 0.018807288259267807, + 0.02723989449441433, + 0.017342790961265564, + 0.01995665207505226, + 0.01691475138068199, + 0.027423463761806488, + 0.013987943530082703, + 0.009764159098267555, + 0.009588638320565224, + 0.010071302764117718, + 0.006165132392197847, + 0.007054764777421951, + 0.005511488299816847, + 0.007321002893149853, + 0.004148722626268864, + 0.009326531551778316, + 0.012008855119347572, + 0.011029661633074284, + 0.008495897985994816, + 0.014316247776150703, + 0.008135133422911167, + 0.0064378962852060795, + 0.005907273851335049, + 0.006639907136559486, + 0.0051710051484405994, + 0.00608719140291214, + 0.004871219862252474, + 0.005152757279574871, + 0.008175944909453392, + 0.003801864106208086, + 0.00346734537743032, + 0.011071078479290009, + 0.020202090963721275, + 0.01060375664383173, + 0.0074096317403018475, + 0.014335145242512226, + 0.007835375145077705, + 0.007143831811845303, + 0.005882652476429939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.08087946474552155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020613953471183777 + ], + "opposite_logprob_tokens": [ + -3.8956139087677 + ], + "logprob_reward": -0.020613953471183777, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020613953471183777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13728585839271545, + 0.0497979000210762, + 0.043132882565259933, + 0.06720422208309174, + 0.07284506410360336, + 0.04667821153998375, + 0.18929800391197205, + 0.14906898140907288, + 0.06415428221225739, + 0.18053467571735382 + ], + "credit_map": [ + 0.021283531561493874, + 0.011156738735735416, + 0.0178105179220438, + 0.010959676466882229, + 0.02584099769592285, + 0.023705750703811646, + 0.016565456986427307, + 0.009963181801140308, + 0.009338014759123325, + 0.005639792885631323, + 0.011658958159387112, + 0.006171184591948986, + 0.006884436588734388, + 0.0029500529635697603, + 0.0037388629280030727, + 0.003416591789573431, + 0.006929546594619751, + 0.007404870819300413, + 0.006324549671262503, + 0.004957098513841629, + 0.003450071671977639, + 0.007510651368647814, + 0.0028416148852556944, + 0.0037144795060157776, + 0.008446111343801022, + 0.00896017998456955, + 0.00568512175232172, + 0.004846875090152025, + 0.017611680552363396, + 0.007983587682247162, + 0.007634369656443596, + 0.006036291364580393, + 0.011003706604242325, + 0.012693275697529316, + 0.008361892774701118, + 0.005753485951572657, + 0.012346512638032436, + 0.008010897785425186, + 0.007316627539694309, + 0.007358662784099579, + 0.006802225019782782, + 0.004635973833501339, + 0.010245286859571934, + 0.0047455644235014915, + 0.005009669344872236, + 0.005192900542169809, + 0.005380595102906227, + 0.004665991757065058, + 0.023132171481847763, + 0.03605090454220772, + 0.025857005268335342, + 0.01651742309331894, + 0.03948795422911644, + 0.01689867489039898, + 0.02097604051232338, + 0.010377809405326843, + 0.015475506894290447, + 0.026269951835274696, + 0.022473203018307686, + 0.017209652811288834, + 0.02855803817510605, + 0.017152858898043633, + 0.01118332426995039, + 0.01074643898755312, + 0.008785516954958439, + 0.011646182276308537, + 0.009890278801321983, + 0.005914662033319473, + 0.0077512324787676334, + 0.01009361445903778, + 0.004868125077337027, + 0.00520466361194849, + 0.020911619067192078, + 0.043262094259262085, + 0.022631892934441566, + 0.014974674209952354, + 0.03881184756755829, + 0.0194011852145195, + 0.010362603701651096, + 0.010178753174841404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.020613953471183777 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096450749784708 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096450749784708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096450749784708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08799160271883011, + 0.07830747216939926, + 0.06144438311457634, + 0.07954950630664825, + 0.08346839249134064, + 0.08328694105148315, + 0.282148540019989, + 0.10536593198776245, + 0.03469536080956459, + 0.10374186187982559 + ], + "credit_map": [ + 0.011844517663121223, + 0.008680950850248337, + 0.02533232793211937, + 0.01231541857123375, + 0.008306160569190979, + 0.00657077319920063, + 0.008963249623775482, + 0.005978206172585487, + 0.01597600430250168, + 0.007785110734403133, + 0.009894889779388905, + 0.005807400681078434, + 0.014535663649439812, + 0.008881967514753342, + 0.00731788482517004, + 0.008108554407954216, + 0.007863873615860939, + 0.009068764746189117, + 0.00610645255073905, + 0.013055448420345783, + 0.006862543988972902, + 0.00927938986569643, + 0.004945486318320036, + 0.00426242733374238, + 0.012486085295677185, + 0.015058540739119053, + 0.009444820694625378, + 0.006938616745173931, + 0.009168709628283978, + 0.00988706760108471, + 0.010061074048280716, + 0.0065045892260968685, + 0.00860100518912077, + 0.012578388676047325, + 0.009208722040057182, + 0.00935306679457426, + 0.010347799398005009, + 0.010363518260419369, + 0.011872774921357632, + 0.011143118143081665, + 0.012906990014016628, + 0.011039859615266323, + 0.012080096639692783, + 0.009314446710050106, + 0.008683954365551472, + 0.008288580924272537, + 0.012952292338013649, + 0.008020716719329357, + 0.03542064130306244, + 0.053809091448783875, + 0.028970325365662575, + 0.023064838722348213, + 0.07462669163942337, + 0.02963358908891678, + 0.023253396153450012, + 0.013369977474212646, + 0.010336412116885185, + 0.018212441354990005, + 0.014411680400371552, + 0.010308608412742615, + 0.022449228912591934, + 0.01213928870856762, + 0.010751942172646523, + 0.0067563424818217754, + 0.0041913180612027645, + 0.0065163359977304935, + 0.005992388818413019, + 0.002786399330943823, + 0.004649365786463022, + 0.005113417282700539, + 0.002885841764509678, + 0.002560296328738332, + 0.011624136008322239, + 0.02373380959033966, + 0.013047759421169758, + 0.009486678056418896, + 0.02176227793097496, + 0.009782836772501469, + 0.008064640685915947, + 0.006239729002118111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.004096450749784708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204996630549431 + ], + "opposite_logprob_tokens": [ + -2.8120498657226562 + ], + "logprob_reward": -0.06204996630549431, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204996630549431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1088407039642334, + 0.11408169567584991, + 0.11411956697702408, + 0.057831645011901855, + 0.10866232216358185, + 0.09551943838596344, + 0.04681610316038132, + 0.06102924421429634, + 0.04753244295716286, + 0.24556683003902435 + ], + "credit_map": [ + 0.013960886746644974, + 0.009574479423463345, + 0.0192278902977705, + 0.012352816760540009, + 0.015539093874394894, + 0.018327657133340836, + 0.012004566378891468, + 0.007853311486542225, + 0.01865609548985958, + 0.019248152151703835, + 0.02124026045203209, + 0.010811399668455124, + 0.013684608042240143, + 0.011522223241627216, + 0.010210031643509865, + 0.008708920329809189, + 0.01825837418437004, + 0.016244860365986824, + 0.016368132084608078, + 0.024923056364059448, + 0.008768375031650066, + 0.013980469666421413, + 0.00783850159496069, + 0.007737800944596529, + 0.008105471730232239, + 0.010279795154929161, + 0.006575065664947033, + 0.005983592476695776, + 0.007714762818068266, + 0.0070456950925290585, + 0.0057929502800107, + 0.006334307603538036, + 0.013570102863013744, + 0.025607511401176453, + 0.013692758046090603, + 0.008373582735657692, + 0.01579534076154232, + 0.016170520335435867, + 0.008281650952994823, + 0.007170854602009058, + 0.009193390607833862, + 0.01754702441394329, + 0.00901179201900959, + 0.0070753502659499645, + 0.0214301198720932, + 0.015195054933428764, + 0.010858544148504734, + 0.005208163056522608, + 0.005884564481675625, + 0.008005204610526562, + 0.005328571889549494, + 0.004583057016134262, + 0.006014520768076181, + 0.004779950249940157, + 0.009197307750582695, + 0.003022929420694709, + 0.008024045266211033, + 0.008991905488073826, + 0.008093971759080887, + 0.00872000865638256, + 0.011984158307313919, + 0.0055785332806408405, + 0.005287919659167528, + 0.004348700400441885, + 0.005445602349936962, + 0.009736658073961735, + 0.005272460635751486, + 0.004822391085326672, + 0.009568879380822182, + 0.00573298055678606, + 0.003701183246448636, + 0.00325228669680655, + 0.03462333604693413, + 0.06077440083026886, + 0.030438128858804703, + 0.019009586423635483, + 0.04358983412384987, + 0.028585292398929596, + 0.013068431057035923, + 0.015477810055017471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.06204996630549431 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328469604253769 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.02328469604253769, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328469604253769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11189179867506027, + 0.06987544149160385, + 0.09846507757902145, + 0.06227649003267288, + 0.06759694218635559, + 0.060589008033275604, + 0.16003364324569702, + 0.2002115398645401, + 0.05478762835264206, + 0.11427249014377594 + ], + "credit_map": [ + 0.013175052590668201, + 0.01191648654639721, + 0.017270617187023163, + 0.01455145888030529, + 0.019261687994003296, + 0.01318768784403801, + 0.014150270260870457, + 0.008378533646464348, + 0.012749726884067059, + 0.008238988928496838, + 0.012392978183925152, + 0.006579175125807524, + 0.012498002499341965, + 0.005256866104900837, + 0.0054398225620388985, + 0.006719876546412706, + 0.015574387274682522, + 0.016398530453443527, + 0.010096135549247265, + 0.016918616369366646, + 0.008667432703077793, + 0.017639124765992165, + 0.005572844296693802, + 0.007598001975566149, + 0.009832393378019333, + 0.008146671578288078, + 0.005890346132218838, + 0.004921087995171547, + 0.009019376710057259, + 0.008948051370680332, + 0.0086140101775527, + 0.006904548034071922, + 0.00928462389856577, + 0.011162303388118744, + 0.008287634700536728, + 0.006730412598699331, + 0.011551104485988617, + 0.008343125693500042, + 0.007328726351261139, + 0.004909008741378784, + 0.009793645702302456, + 0.006430974695831537, + 0.011123157106339931, + 0.006969745736569166, + 0.006647398695349693, + 0.006875727791339159, + 0.007396256551146507, + 0.0053520952351391315, + 0.01692732237279415, + 0.030457500368356705, + 0.019452853128314018, + 0.015503526665270329, + 0.03656432777643204, + 0.014646683819591999, + 0.015026403591036797, + 0.011455007828772068, + 0.020582500845193863, + 0.034961897879838943, + 0.025615759193897247, + 0.01764187216758728, + 0.047785431146621704, + 0.021933075040578842, + 0.019464462995529175, + 0.012226498685777187, + 0.0063482956029474735, + 0.01028470415621996, + 0.008414137177169323, + 0.006811708677560091, + 0.006653949152678251, + 0.005876594223082066, + 0.006425309460610151, + 0.003972921520471573, + 0.01289642509073019, + 0.02446036785840988, + 0.01437933649867773, + 0.010430759750306606, + 0.022261783480644226, + 0.010829743929207325, + 0.012062234804034233, + 0.006951832212507725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.02328469604253769 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, using high-speed photography to create a slow-motion effect, focuses on an old, dark brown wooden dining chair, its paint chipped, careening down at breakneck speed onto a rough, dusty concrete floor. The moment the chair legs impact the ground, the immense force of the impact causes it to shatter, sending countless shards of wood, sawdust, and dust scattering in all directions from the point of impact like a volcanic eruption. The slow motion capture clearly reveals the trajectory of each fragment, tumbling and spinning through the air, creating a still image imbued with a violent aesthetic. A cold, overhead spotlight pierces the dense dust, sharply illuminating the flying fragments and casting rapidly shifting shadows across their irregular surfaces. The scene's grim tone evokes a sense of primal destructive force and a momentary tension. The camera vibrates slightly yet violently at the moment of impact, further heightening the impact's realism.", + "seed": 28036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_0/index_9465_seed_28036_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019476865977048875, + "vlm_logprob": -0.019476865977048875, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.019476865977048875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04028989374637604 + ], + "opposite_logprob_tokens": [ + -3.290289878845215 + ], + "logprob_reward": -0.04028989374637604, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04028989374637604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11854074895381927, + 0.0648355782032013, + 0.11022885143756866, + 0.22947217524051666, + 0.12127873301506042, + 0.17950862646102905, + 0.0453338697552681, + 0.04157925397157669, + 0.06811827421188354, + 0.02110392041504383 + ], + "credit_map": [ + 0.013553283177316189, + 0.01973867602646351, + 0.01790146715939045, + 0.011869149282574654, + 0.014916598796844482, + 0.014891711995005608, + 0.012300778180360794, + 0.013369081541895866, + 0.009564662352204323, + 0.010033694095909595, + 0.006225212477147579, + 0.010353447869420052, + 0.0075731766410171986, + 0.007206958252936602, + 0.006090464536100626, + 0.007787954993546009, + 0.017253121361136436, + 0.01326356828212738, + 0.00797740463167429, + 0.012346894480288029, + 0.013666047714650631, + 0.01902548223733902, + 0.012762767262756824, + 0.013933549635112286, + 0.033366743475198746, + 0.046749476343393326, + 0.01664072275161743, + 0.01905048079788685, + 0.03850753977894783, + 0.037123844027519226, + 0.024853738024830818, + 0.013179617002606392, + 0.010661034844815731, + 0.03648875653743744, + 0.005901418160647154, + 0.009465384297072887, + 0.010611582547426224, + 0.03212009742856026, + 0.006471795961260796, + 0.009558655321598053, + 0.02550387755036354, + 0.019208969548344612, + 0.04137522354722023, + 0.017961569130420685, + 0.018293006345629692, + 0.01555460412055254, + 0.02072622999548912, + 0.02088511921465397, + 0.004412244539707899, + 0.006283532828092575, + 0.004881625063717365, + 0.004798243287950754, + 0.006369896698743105, + 0.007321150042116642, + 0.007662890944629908, + 0.0036042830906808376, + 0.004018304869532585, + 0.005801053252071142, + 0.005287506151944399, + 0.00550388777628541, + 0.006589989643543959, + 0.006409260910004377, + 0.004817711655050516, + 0.0031515348237007856, + 0.007501327898353338, + 0.012107189744710922, + 0.005473337136209011, + 0.004710690584033728, + 0.013591942377388477, + 0.012471492402255535, + 0.006729509681463242, + 0.005532780662178993, + 0.0021523211617022753, + 0.0030525284819304943, + 0.002638659905642271, + 0.0028280499391257763, + 0.002744606463238597, + 0.003476352896541357, + 0.0024867444299161434, + 0.001724654110148549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.04028989374637604 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016097577288746834 + ], + "opposite_logprob_tokens": [ + -4.141097545623779 + ], + "logprob_reward": -0.016097577288746834, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016097577288746834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1835044026374817, + 0.19617928564548492, + 0.0672045648097992, + 0.1706993728876114, + 0.06631612032651901, + 0.04594730958342552, + 0.09971018880605698, + 0.10413602739572525, + 0.03493029996752739, + 0.03137242794036865 + ], + "credit_map": [ + 0.02245715819299221, + 0.015810396522283554, + 0.019051099196076393, + 0.021708151325583458, + 0.017707061022520065, + 0.023336218670010567, + 0.019006798043847084, + 0.04442752152681351, + 0.025509865954518318, + 0.031094811856746674, + 0.02034500613808632, + 0.021800538524985313, + 0.020881950855255127, + 0.026251718401908875, + 0.026127031072974205, + 0.02416837401688099, + 0.010103149339556694, + 0.009814360179007053, + 0.005990832112729549, + 0.007850578986108303, + 0.007652762345969677, + 0.007515815086662769, + 0.007898058742284775, + 0.010379004292190075, + 0.025410160422325134, + 0.03800921514630318, + 0.01237369142472744, + 0.01371079683303833, + 0.029070105403661728, + 0.02841491810977459, + 0.015046253800392151, + 0.008664224296808243, + 0.0066565158776938915, + 0.017881955951452255, + 0.005716424435377121, + 0.0058923764154314995, + 0.005991953890770674, + 0.015797428786754608, + 0.0030503147281706333, + 0.005329150706529617, + 0.0054161036387085915, + 0.005892509128898382, + 0.005156027153134346, + 0.006185370497405529, + 0.005503680091351271, + 0.007633165456354618, + 0.005126742646098137, + 0.005033711437135935, + 0.012027261778712273, + 0.012251393869519234, + 0.010825675912201405, + 0.009965669363737106, + 0.011694436892867088, + 0.024253055453300476, + 0.009421651251614094, + 0.009271047078073025, + 0.008100172504782677, + 0.013933265581727028, + 0.012080452404916286, + 0.0131065146997571, + 0.017712412402033806, + 0.01808156818151474, + 0.014576883055269718, + 0.006544765084981918, + 0.003304595360532403, + 0.0051353867165744305, + 0.004326574970036745, + 0.004130788147449493, + 0.006189887411892414, + 0.006000058725476265, + 0.002845579059794545, + 0.002997433068230748, + 0.003272371832281351, + 0.005879813805222511, + 0.002865260699763894, + 0.004837871063500643, + 0.003288299310952425, + 0.005402907729148865, + 0.0034404334146529436, + 0.0023854670580476522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.016097577288746834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253438837826252 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253438837826252, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253438837826252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06838669627904892, + 0.037578873336315155, + 0.09742515534162521, + 0.05907398834824562, + 0.12971006333827972, + 0.37301498651504517, + 0.10808926820755005, + 0.05120563134551048, + 0.044456467032432556, + 0.031058887019753456 + ], + "credit_map": [ + 0.006506815552711487, + 0.013279328122735023, + 0.009204582311213017, + 0.006362711079418659, + 0.011613499373197556, + 0.00907734502106905, + 0.006141482386738062, + 0.006200926378369331, + 0.0055484408512711525, + 0.007354292087256908, + 0.002663334598764777, + 0.004447478335350752, + 0.006078782491385937, + 0.005139254964888096, + 0.003277302486822009, + 0.0030699828639626503, + 0.018231108784675598, + 0.009951602667570114, + 0.00706541957333684, + 0.0066239419393241405, + 0.008873618207871914, + 0.02359260991215706, + 0.007225821726024151, + 0.015861013904213905, + 0.005434217397123575, + 0.0070715476758778095, + 0.004116758704185486, + 0.006330191157758236, + 0.011123592965304852, + 0.008898357860744, + 0.011345459148287773, + 0.004753860179334879, + 0.014005265198647976, + 0.0379522480070591, + 0.007129782345145941, + 0.010034991428256035, + 0.010824770666658878, + 0.033251047134399414, + 0.0065978458151221275, + 0.009914103895425797, + 0.0623958520591259, + 0.03991154208779335, + 0.10662094503641129, + 0.03124660812318325, + 0.029887882992625237, + 0.027417371049523354, + 0.03351598232984543, + 0.04201874881982803, + 0.0132830161601305, + 0.012527232058346272, + 0.009891819208860397, + 0.008306809701025486, + 0.02506455034017563, + 0.016938837245106697, + 0.013785726390779018, + 0.008291271515190601, + 0.0038646210450679064, + 0.005445513874292374, + 0.007596015930175781, + 0.007557162549346685, + 0.007468051742762327, + 0.008524375967681408, + 0.006490061990916729, + 0.004259821958839893, + 0.004168685059994459, + 0.009287288412451744, + 0.0034731237683445215, + 0.00478576123714447, + 0.006095086224377155, + 0.0075948662124574184, + 0.004360200837254524, + 0.004691451787948608, + 0.002747445832937956, + 0.0037348081823438406, + 0.004936644807457924, + 0.008405249565839767, + 0.0024073284585028887, + 0.003679914865642786, + 0.0033202823251485825, + 0.0018272098386660218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.005253438837826252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948222242295742 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948222242295742, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948222242295742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4155105650424957, + 0.06664913892745972, + 0.19472604990005493, + 0.040104348212480545, + 0.09875059127807617, + 0.046950437128543854, + 0.02951248176395893, + 0.035669583827257156, + 0.04511043429374695, + 0.02701643481850624 + ], + "credit_map": [ + 0.0389365553855896, + 0.05993707850575447, + 0.06198129802942276, + 0.03830215707421303, + 0.05911985784769058, + 0.04785642400383949, + 0.03242388740181923, + 0.07695330679416656, + 0.007009943015873432, + 0.012100120075047016, + 0.0054841917008161545, + 0.007746193092316389, + 0.011204185895621777, + 0.009199687279760838, + 0.0067244525998830795, + 0.007180358748883009, + 0.04320671781897545, + 0.009456117637455463, + 0.005685910582542419, + 0.0086772246286273, + 0.009218720719218254, + 0.068946972489357, + 0.011297781020402908, + 0.03823660686612129, + 0.004627660848200321, + 0.006156557239592075, + 0.0035200300626456738, + 0.003981208428740501, + 0.00722321355715394, + 0.005975866690278053, + 0.005776192992925644, + 0.0028436174616217613, + 0.009075253270566463, + 0.026131492108106613, + 0.0063296593725681305, + 0.008765396662056446, + 0.008323420770466328, + 0.025177044793963432, + 0.006117682438343763, + 0.008830642327666283, + 0.0053087822161614895, + 0.005975930951535702, + 0.008320498280227184, + 0.004934659693390131, + 0.006142133381217718, + 0.005293313879519701, + 0.005697520449757576, + 0.005277599208056927, + 0.0033410824835300446, + 0.0033674745354801416, + 0.003642513882368803, + 0.0029157456010580063, + 0.004593319725245237, + 0.005506019573658705, + 0.0039146370254457, + 0.002231690799817443, + 0.003620593110099435, + 0.004808290395885706, + 0.005028536077588797, + 0.004129454959183931, + 0.004956020973622799, + 0.005500289145857096, + 0.005239656660705805, + 0.00238674134016037, + 0.004324069246649742, + 0.008174831978976727, + 0.004532032646238804, + 0.004271463956683874, + 0.008512912318110466, + 0.008042437955737114, + 0.003987694159150124, + 0.0032649929635226727, + 0.003514740848913789, + 0.006003806367516518, + 0.0023507452569901943, + 0.0027289551217108965, + 0.003920688293874264, + 0.004903013352304697, + 0.0020504507701843977, + 0.0015440367860719562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.005948222242295742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old, dark brown wooden dining chair with chipped paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the chair careen down at breakneck speed onto a rough, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chair shatter upon impact, sending wood shards, sawdust, and dust scattering in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cold, overhead spotlight that pierces the dust and illuminates the flying fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit slight yet violent vibration at the moment of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795197769999504 + ], + "opposite_logprob_tokens": [ + -3.5297951698303223 + ], + "logprob_reward": -0.029795197769999504, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795197769999504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10746631026268005, + 0.12395328283309937, + 0.1057523638010025, + 0.09686847776174545, + 0.2644992470741272, + 0.07888051122426987, + 0.06533163040876389, + 0.06352119892835617, + 0.03981323912739754, + 0.05391369014978409 + ], + "credit_map": [ + 0.012889415957033634, + 0.01501659955829382, + 0.015334694646298885, + 0.011465540155768394, + 0.011571559123694897, + 0.014429043047130108, + 0.008835780434310436, + 0.01792367547750473, + 0.013184444047510624, + 0.02757832780480385, + 0.010107402689754963, + 0.011350798420608044, + 0.02078995481133461, + 0.015223303809762001, + 0.012865956872701645, + 0.012853102758526802, + 0.02249949797987938, + 0.006551922764629126, + 0.00590837886556983, + 0.004881684668362141, + 0.006701934151351452, + 0.03311501443386078, + 0.0075057451613247395, + 0.018588189035654068, + 0.015390443615615368, + 0.01873723976314068, + 0.007337616756558418, + 0.007813877426087856, + 0.01765824854373932, + 0.015795504674315453, + 0.00872281938791275, + 0.005412724334746599, + 0.03837047889828682, + 0.07077601552009583, + 0.01599527709186077, + 0.024101942777633667, + 0.021967602893710136, + 0.061850350350141525, + 0.012569096870720387, + 0.018868498504161835, + 0.009484384208917618, + 0.011809068731963634, + 0.013000481761991978, + 0.007292051799595356, + 0.012387423776090145, + 0.00784040242433548, + 0.00758108776062727, + 0.009485614486038685, + 0.008801652118563652, + 0.0076935128308832645, + 0.007496776524931192, + 0.005776152480393648, + 0.015038893558084965, + 0.01006960216909647, + 0.00605234457179904, + 0.00440269336104393, + 0.004560566507279873, + 0.007726202718913555, + 0.007655789609998465, + 0.007365213241428137, + 0.011998532339930534, + 0.011758419685065746, + 0.007083815522491932, + 0.005372658371925354, + 0.0037380261346697807, + 0.007703834213316441, + 0.004122202284634113, + 0.003311479464173317, + 0.006037722807377577, + 0.006917889229953289, + 0.004254519008100033, + 0.003727567847818136, + 0.004728676751255989, + 0.009253696538507938, + 0.006382253021001816, + 0.012315343134105206, + 0.006232027430087328, + 0.008182693272829056, + 0.004300596658140421, + 0.00251840241253376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.029795197769999504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_tensor([7123, 7123])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_tensor([7123, 7123])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c821b7e3ef389b81c3539969ea8bc4ed05b01296 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_tensor([7123, 7123])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from above, captured in extreme slow motion using an ultra-high-speed camera, captures the decisive moment. A pristine, flat, unfolded $100 bill rests against a matte black background. Benjamin Franklin's portrait, the blue security thread, and the delicate engravings are strikingly clear under the soft overhead light. Suddenly, a rapidly spinning bullet pierces the frame vertically from above. The intense friction between the brass tip and the air produces a fiery, almost incandescent orange glow, the sole dynamic source of light in the otherwise dark scene. The moment the bullet's tip contacts the bill, the paper fibers stretch and deform at a visible speed, then violently tear apart, creating a widening tear. Countless tiny paper and ink particles are blasted in all directions by the impact, the edges of the tear slightly curling and charring from the scorching heat of the bullet. The bullet penetrates effortlessly, continuing its unstoppable trajectory, leaving a rough-edged but perfectly shaped circular hole in the center of the banknote. The entire banknote trembles and undulates violently due to this instantaneous violent impact, forming ripples. The entire picture is full of power and the aesthetics of destruction, and the atmosphere is tense and shocking.", + "seed": 21057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_7123_seed_21057_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05887728463858366, + "vlm_logprob": -0.05887728463858366, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.05887728463858366, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.173951655626297 + ], + "opposite_logprob_tokens": [ + -1.9239516258239746 + ], + "logprob_reward": -0.173951655626297, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.173951655626297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14357401430606842, + 0.05721111223101616, + 0.1810946762561798, + 0.04209328815340996, + 0.13597524166107178, + 0.03436088562011719, + 0.143938347697258, + 0.06178499758243561, + 0.07192818075418472, + 0.12803928554058075 + ], + "credit_map": [ + 0.015738826245069504, + 0.01508689671754837, + 0.017647475004196167, + 0.03197305649518967, + 0.01859240047633648, + 0.011273176409304142, + 0.017284652218222618, + 0.01597752422094345, + 0.005783824250102043, + 0.0063071297481656075, + 0.004016757011413574, + 0.009362742304801941, + 0.0059092724695801735, + 0.00555899553000927, + 0.007411647122353315, + 0.012860739603638649, + 0.009062528610229492, + 0.05118270963430405, + 0.011139092966914177, + 0.015959788113832474, + 0.013472654856741428, + 0.06235538795590401, + 0.009353222325444221, + 0.008569302968680859, + 0.004432802554219961, + 0.004426939878612757, + 0.0041182925924658775, + 0.005422063637524843, + 0.00702112540602684, + 0.0065271626226603985, + 0.00584516953676939, + 0.004299729131162167, + 0.03165320307016373, + 0.041381172835826874, + 0.01117647159844637, + 0.015873609110713005, + 0.006872707977890968, + 0.012826087884604931, + 0.007290148641914129, + 0.008901827037334442, + 0.0032519183587282896, + 0.002902466570958495, + 0.002318721730262041, + 0.003917523194104433, + 0.004987992811948061, + 0.004935292061418295, + 0.004770975559949875, + 0.007275997661054134, + 0.010687446221709251, + 0.009781045839190483, + 0.02115892618894577, + 0.024716751649975777, + 0.014651519246399403, + 0.022290853783488274, + 0.015686649829149246, + 0.024965157732367516, + 0.00649780360981822, + 0.0048700361512601376, + 0.008768806234002113, + 0.011065180413424969, + 0.005681565031409264, + 0.006151206325739622, + 0.010000375099480152, + 0.008750028908252716, + 0.006215529982000589, + 0.007545117288827896, + 0.013030611909925938, + 0.012875746004283428, + 0.007985238917171955, + 0.007591931615024805, + 0.007358891423791647, + 0.009325115941464901, + 0.012479841709136963, + 0.015168076381087303, + 0.018492771312594414, + 0.017662106081843376, + 0.015717774629592896, + 0.012267135083675385, + 0.013109602965414524, + 0.023141972720623016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.173951655626297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373829692602158 + ], + "opposite_logprob_tokens": [ + -3.651373863220215 + ], + "logprob_reward": -0.026373829692602158, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373829692602158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11193794757127762, + 0.17785441875457764, + 0.0882604569196701, + 0.06373792141675949, + 0.10394024103879929, + 0.022515442222356796, + 0.09494644403457642, + 0.10586417466402054, + 0.08843820542097092, + 0.14250478148460388 + ], + "credit_map": [ + 0.02646665833890438, + 0.026220744475722313, + 0.009140396490693092, + 0.013202449306845665, + 0.012568109668791294, + 0.0073099867440760136, + 0.008216194808483124, + 0.00881340354681015, + 0.01226930320262909, + 0.025253770872950554, + 0.009727329947054386, + 0.018579483032226562, + 0.03177618980407715, + 0.029039055109024048, + 0.018025919795036316, + 0.03318335860967636, + 0.009708491154015064, + 0.009157071821391582, + 0.01644565537571907, + 0.014706484973430634, + 0.009532992728054523, + 0.011463525705039501, + 0.011265646666288376, + 0.005980588961392641, + 0.006365679670125246, + 0.009872464463114738, + 0.005517544690519571, + 0.00642995024099946, + 0.011138482950627804, + 0.00919010117650032, + 0.006795941386371851, + 0.008427759632468224, + 0.023160530254244804, + 0.03141244873404503, + 0.008968871086835861, + 0.008489479310810566, + 0.006486648693680763, + 0.010318818502128124, + 0.006902819965034723, + 0.008200633339583874, + 0.002842805115506053, + 0.002468824153766036, + 0.0024703810922801495, + 0.0029803437646478415, + 0.002762255258858204, + 0.002897365717217326, + 0.0028919330798089504, + 0.003201534738764167, + 0.006379071157425642, + 0.007858525961637497, + 0.010582497343420982, + 0.01655513420701027, + 0.008168286643922329, + 0.017501993104815483, + 0.012240906246006489, + 0.01566002517938614, + 0.007105535361915827, + 0.005687899421900511, + 0.013539968989789486, + 0.016684236004948616, + 0.00930794421583414, + 0.00919371098279953, + 0.029441749677062035, + 0.014903129078447819, + 0.005796143319457769, + 0.00674217427149415, + 0.009558582678437233, + 0.011774287559092045, + 0.008139463141560555, + 0.014520170167088509, + 0.022834930568933487, + 0.009072452783584595, + 0.012583755888044834, + 0.014349834062159061, + 0.022089418023824692, + 0.018265502527356148, + 0.017414899542927742, + 0.013725402764976025, + 0.013644627295434475, + 0.030431348830461502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.026373829692602158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002830429933965206 + ], + "opposite_logprob_tokens": [ + -5.877830505371094 + ], + "logprob_reward": -0.002830429933965206, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002830429933965206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05625331774353981, + 0.11689107865095139, + 0.21429812908172607, + 0.11050952225923538, + 0.04775567725300789, + 0.022630741819739342, + 0.06035766750574112, + 0.15916717052459717, + 0.050695955753326416, + 0.16144074499607086 + ], + "credit_map": [ + 0.008059852756559849, + 0.005138651002198458, + 0.006288126576691866, + 0.008923402987420559, + 0.006870621349662542, + 0.004463484510779381, + 0.00564091419801116, + 0.010868264362215996, + 0.007459468208253384, + 0.02076774649322033, + 0.007513131480664015, + 0.013207672163844109, + 0.01733868196606636, + 0.019865060225129128, + 0.009224075824022293, + 0.021515244618058205, + 0.018445491790771484, + 0.03951341658830643, + 0.02356434240937233, + 0.025870606303215027, + 0.0211323294788599, + 0.050279781222343445, + 0.018570363521575928, + 0.016921797767281532, + 0.01103757880628109, + 0.03161117807030678, + 0.006753702647984028, + 0.01616959273815155, + 0.008981854654848576, + 0.015917453914880753, + 0.009212830103933811, + 0.010825326666235924, + 0.008864021860063076, + 0.012506690807640553, + 0.005780385807156563, + 0.0038227997720241547, + 0.002906061941757798, + 0.0054335142485797405, + 0.0038547648582607508, + 0.0045874412171542645, + 0.003500057151541114, + 0.0016446566442027688, + 0.0026755225844681263, + 0.0023832081351429224, + 0.002612719079479575, + 0.00233702571131289, + 0.002839513123035431, + 0.004638040438294411, + 0.004243424627929926, + 0.005183995235711336, + 0.00803790707141161, + 0.011112576350569725, + 0.005309485364705324, + 0.009985082782804966, + 0.006563578732311726, + 0.009921618737280369, + 0.012449636124074459, + 0.009565934538841248, + 0.02100495994091034, + 0.020215967670083046, + 0.014564628712832928, + 0.016390791162848473, + 0.04393713176250458, + 0.021038124337792397, + 0.0038465505931526423, + 0.004124634899199009, + 0.009217667393386364, + 0.006464858539402485, + 0.004751929081976414, + 0.0068449960090219975, + 0.010074833407998085, + 0.005370483733713627, + 0.01395722571760416, + 0.014927626587450504, + 0.02693486586213112, + 0.020708303898572922, + 0.018788784742355347, + 0.01599923148751259, + 0.014935702085494995, + 0.03518899902701378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.002830429933965206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201516181230545 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201516181230545, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201516181230545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07163374871015549, + 0.048091862350702286, + 0.21962691843509674, + 0.05377493053674698, + 0.03233931213617325, + 0.03385342285037041, + 0.06754788011312485, + 0.07935769110918045, + 0.062440428882837296, + 0.3313337564468384 + ], + "credit_map": [ + 0.013305560685694218, + 0.010825404897332191, + 0.008046806789934635, + 0.01051696389913559, + 0.007288917899131775, + 0.0066032372415065765, + 0.0067530409432947636, + 0.008293818682432175, + 0.006046404596418142, + 0.004423584323376417, + 0.005003695376217365, + 0.009623057208955288, + 0.0056539252400398254, + 0.005574199371039867, + 0.004695792682468891, + 0.007071204017847776, + 0.014836338348686695, + 0.0628608763217926, + 0.010395203717052937, + 0.014718087390065193, + 0.015193275175988674, + 0.07934209704399109, + 0.01144435629248619, + 0.010836699977517128, + 0.006224546115845442, + 0.009973037987947464, + 0.004204414784908295, + 0.007086719386279583, + 0.0071599665097892284, + 0.008231764659285545, + 0.005042417906224728, + 0.005852061789482832, + 0.007749886251986027, + 0.009056363254785538, + 0.0021844918373972178, + 0.004332846496254206, + 0.0030804434791207314, + 0.0019026427762582898, + 0.0016990092117339373, + 0.0023336289450526237, + 0.003906165948137641, + 0.002999520394951105, + 0.002661189530044794, + 0.0037740010302513838, + 0.006281604524701834, + 0.004265219904482365, + 0.004929039627313614, + 0.005036684684455395, + 0.0077567812986671925, + 0.0042423983104527, + 0.006833578459918499, + 0.008585885167121887, + 0.005408932454884052, + 0.009774252772331238, + 0.0073336041532456875, + 0.01761244796216488, + 0.005763768684118986, + 0.005651697516441345, + 0.008378705009818077, + 0.01070871576666832, + 0.006672148127108812, + 0.00912050437182188, + 0.02346722222864628, + 0.009594927541911602, + 0.0054204510524868965, + 0.005069689825177193, + 0.009620933793485165, + 0.008983945474028587, + 0.0049660541117191315, + 0.008299958892166615, + 0.013293551281094551, + 0.006785841193050146, + 0.02865024283528328, + 0.030186688527464867, + 0.057752396911382675, + 0.041088610887527466, + 0.037055280059576035, + 0.033004384487867355, + 0.029324118047952652, + 0.0742720514535904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.002201516181230545 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890289917588234 + ], + "opposite_logprob_tokens": [ + -2.46402907371521 + ], + "logprob_reward": -0.0890289917588234, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890289917588234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07534752041101456, + 0.0604761466383934, + 0.22171859443187714, + 0.06538760662078857, + 0.053372472524642944, + 0.03685560077428818, + 0.12721118330955505, + 0.11196133494377136, + 0.12901197373867035, + 0.11865758895874023 + ], + "credit_map": [ + 0.009720058180391788, + 0.008088073693215847, + 0.008396604098379612, + 0.009421629831194878, + 0.009246486239135265, + 0.009201175533235073, + 0.0077474070712924, + 0.013526083901524544, + 0.006895687431097031, + 0.00604792358353734, + 0.0054940115660429, + 0.012696613557636738, + 0.008542981930077076, + 0.0059115095064044, + 0.004898429848253727, + 0.009988992474973202, + 0.018060972914099693, + 0.05601377785205841, + 0.0159164946526289, + 0.019784264266490936, + 0.016166361048817635, + 0.07315374910831451, + 0.010985718108713627, + 0.011637250892817974, + 0.008015288971364498, + 0.01084066741168499, + 0.00637791957706213, + 0.0069033242762088776, + 0.00948272179812193, + 0.010669811628758907, + 0.006507746409624815, + 0.006590123753994703, + 0.01404143963009119, + 0.01635507307946682, + 0.004310570657253265, + 0.0027405961882323027, + 0.0028035396244376898, + 0.005098931025713682, + 0.002970978617668152, + 0.005051343236118555, + 0.0030325795523822308, + 0.0030681113712489605, + 0.00484056631103158, + 0.00835910439491272, + 0.003779107239097357, + 0.004851861856877804, + 0.005201776511967182, + 0.0037224949337542057, + 0.009471331723034382, + 0.007044148165732622, + 0.01494854036718607, + 0.012624597176909447, + 0.011708380654454231, + 0.022621454671025276, + 0.021540839225053787, + 0.027251889929175377, + 0.01046628225594759, + 0.008184857666492462, + 0.015952739864587784, + 0.01830446533858776, + 0.011706085875630379, + 0.011525582522153854, + 0.02027301676571369, + 0.015548300929367542, + 0.009885630570352077, + 0.010889336466789246, + 0.010518289171159267, + 0.014842947013676167, + 0.013483013026416302, + 0.019753631204366684, + 0.03462466225028038, + 0.015014465898275375, + 0.010831551626324654, + 0.012312378734350204, + 0.017438098788261414, + 0.01668909378349781, + 0.013231681659817696, + 0.011676919646561146, + 0.011600781232118607, + 0.024877086281776428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.0890289917588234 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from above, captured in extreme slow motion using an ultra-high-speed camera, captures the decisive moment. A pristine, flat, unfolded $100 bill rests against a matte black background. Benjamin Franklin's portrait, the blue security thread, and the delicate engravings are strikingly clear under the soft overhead light. Suddenly, a rapidly spinning bullet pierces the frame vertically from above. The intense friction between the brass tip and the air produces a fiery, almost incandescent orange glow, the sole dynamic source of light in the otherwise dark scene. The moment the bullet's tip contacts the bill, the paper fibers stretch and deform at a visible speed, then violently tear apart, creating a widening tear. Countless tiny paper and ink particles are blasted in all directions by the impact, the edges of the tear slightly curling and charring from the scorching heat of the bullet. The bullet penetrates effortlessly, continuing its unstoppable trajectory, leaving a rough-edged but perfectly shaped circular hole in the center of the banknote. The entire banknote trembles and undulates violently due to this instantaneous violent impact, forming ripples. The entire picture is full of power and the aesthetics of destruction, and the atmosphere is tense and shocking.", + "seed": 21058, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_7123_seed_21058_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03633505657780915, + "vlm_logprob": -0.03633505657780915, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03633505657780915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10894069075584412 + ], + "opposite_logprob_tokens": [ + -2.358940601348877 + ], + "logprob_reward": -0.10894069075584412, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10894069075584412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1167183369398117, + 0.20099042356014252, + 0.0814022421836853, + 0.06649395823478699, + 0.03910280019044876, + 0.10102527588605881, + 0.06013127788901329, + 0.07997497916221619, + 0.07986967265605927, + 0.17429102957248688 + ], + "credit_map": [ + 0.018149860203266144, + 0.011763960123062134, + 0.021757936105132103, + 0.015582187101244926, + 0.009860863909125328, + 0.012114188633859158, + 0.011866271495819092, + 0.015623069368302822, + 0.02957175485789776, + 0.03550555557012558, + 0.04115495830774307, + 0.02817695029079914, + 0.008679606020450592, + 0.015434548258781433, + 0.01747947931289673, + 0.02498757652938366, + 0.010349633172154427, + 0.015215088613331318, + 0.016470640897750854, + 0.007523310836404562, + 0.004822577815502882, + 0.009464583359658718, + 0.008647977374494076, + 0.008908438496291637, + 0.006089650094509125, + 0.008379983715713024, + 0.007080006413161755, + 0.005542531609535217, + 0.007093304768204689, + 0.00967379193753004, + 0.012192676775157452, + 0.01044201198965311, + 0.0033216930460184813, + 0.004613598342984915, + 0.0028104253578931093, + 0.0039865728467702866, + 0.003691525664180517, + 0.007448046933859587, + 0.00723163690418005, + 0.005999303422868252, + 0.012189656496047974, + 0.018523627892136574, + 0.011376386508345604, + 0.010613715276122093, + 0.013102162629365921, + 0.012272028252482414, + 0.011513618752360344, + 0.011434082873165607, + 0.006664371583610773, + 0.0057962448336184025, + 0.008822347968816757, + 0.00812479853630066, + 0.004153115209192038, + 0.010385949164628983, + 0.00528846587985754, + 0.010895985178649426, + 0.00897172000259161, + 0.008476004004478455, + 0.01361530739814043, + 0.011910254135727882, + 0.00641636224463582, + 0.009746846742928028, + 0.008891986683011055, + 0.011946492828428745, + 0.008852933533489704, + 0.026079397648572922, + 0.008423535153269768, + 0.006990774534642696, + 0.00626398017629981, + 0.009446078911423683, + 0.005377347581088543, + 0.008435626514256, + 0.016706543043255806, + 0.052463363856077194, + 0.013143415562808514, + 0.017763299867510796, + 0.019063878804445267, + 0.02234802395105362, + 0.012072703801095486, + 0.020729796960949898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.10894069075584412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020613720640540123 + ], + "opposite_logprob_tokens": [ + -3.895613670349121 + ], + "logprob_reward": -0.020613720640540123, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020613720640540123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2071629911661148, + 0.08841019123792648, + 0.22422094643115997, + 0.04123653471469879, + 0.04827426001429558, + 0.09150528907775879, + 0.12526902556419373, + 0.03170883283019066, + 0.03636389225721359, + 0.10584801435470581 + ], + "credit_map": [ + 0.020244313403964043, + 0.031442586332559586, + 0.027787532657384872, + 0.020962350070476532, + 0.030317464843392372, + 0.023498112335801125, + 0.028471099212765694, + 0.024439549073576927, + 0.01076413318514824, + 0.010214874520897865, + 0.008001036942005157, + 0.00868893787264824, + 0.00863292533904314, + 0.010379962623119354, + 0.019819213077425957, + 0.011909102089703083, + 0.026639165356755257, + 0.04631393775343895, + 0.02865006774663925, + 0.01902112178504467, + 0.027149613946676254, + 0.022960033267736435, + 0.030256595462560654, + 0.023230422288179398, + 0.004011486191302538, + 0.0039808787405490875, + 0.005441005807369947, + 0.003932861611247063, + 0.006054346449673176, + 0.0049756672233343124, + 0.0077308835461735725, + 0.005109407007694244, + 0.002939214464277029, + 0.006212894339114428, + 0.0038362641353160143, + 0.005732086021453142, + 0.0067222947254776955, + 0.005795588716864586, + 0.01093971822410822, + 0.006096199620515108, + 0.009630415588617325, + 0.009623493999242783, + 0.008292151615023613, + 0.009951516054570675, + 0.009281902574002743, + 0.01923719234764576, + 0.012909140437841415, + 0.012579470872879028, + 0.009608646854758263, + 0.012886489741504192, + 0.010994520038366318, + 0.011550119146704674, + 0.013755434192717075, + 0.03384127467870712, + 0.013586544431746006, + 0.019046001136302948, + 0.00463945372030139, + 0.004334570374339819, + 0.005454654339700937, + 0.0033330535516142845, + 0.0025594583712518215, + 0.004133091773837805, + 0.0034268212039023638, + 0.003827728098258376, + 0.004005805589258671, + 0.006148370448499918, + 0.005048545077443123, + 0.004373948089778423, + 0.00330420327372849, + 0.006225645542144775, + 0.0028288967441767454, + 0.004428477492183447, + 0.008999012410640717, + 0.03271766006946564, + 0.008967290632426739, + 0.010283107869327068, + 0.011026907712221146, + 0.015148593112826347, + 0.006174619309604168, + 0.012530814856290817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.020613720640540123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015215021558105946 + ], + "opposite_logprob_tokens": [ + -6.501521587371826 + ], + "logprob_reward": -0.0015215021558105946, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015215021558105946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06127993389964104, + 0.11955051124095917, + 0.08794725686311722, + 0.0898970291018486, + 0.02941238507628441, + 0.07513464242219925, + 0.07093431800603867, + 0.049830105155706406, + 0.05873732268810272, + 0.35727646946907043 + ], + "credit_map": [ + 0.010150919668376446, + 0.005785120651125908, + 0.005910525098443031, + 0.006501374300569296, + 0.008469331078231335, + 0.008143797516822815, + 0.008155643939971924, + 0.008163219317793846, + 0.01791585236787796, + 0.015609752386808395, + 0.01977892406284809, + 0.01530485413968563, + 0.009548800066113472, + 0.011503343470394611, + 0.013033087365329266, + 0.016855895519256592, + 0.007232888136059046, + 0.011141948401927948, + 0.006809654179960489, + 0.006982557941228151, + 0.015616009943187237, + 0.009104857221245766, + 0.019275829195976257, + 0.011783505789935589, + 0.009079176932573318, + 0.007974325679242611, + 0.008704944513738155, + 0.007683878764510155, + 0.013331439346075058, + 0.01222359761595726, + 0.018648559227585793, + 0.012251111678779125, + 0.002840661210939288, + 0.004047447349876165, + 0.0024300855584442616, + 0.0024258028715848923, + 0.003977957647293806, + 0.004094614181667566, + 0.005529364105314016, + 0.004066452383995056, + 0.00491202250123024, + 0.004688092973083258, + 0.004843530710786581, + 0.008790293708443642, + 0.007834814488887787, + 0.019857686012983322, + 0.012438932433724403, + 0.011769267730414867, + 0.004762290045619011, + 0.004054751712828875, + 0.004920670296996832, + 0.00944400206208229, + 0.00648236321285367, + 0.01929027959704399, + 0.010897299274802208, + 0.011082658544182777, + 0.005645504221320152, + 0.005525526124984026, + 0.0076199062168598175, + 0.007383782882243395, + 0.004122814629226923, + 0.006148641929030418, + 0.006550437305122614, + 0.006833490449935198, + 0.00650995597243309, + 0.009637604467570782, + 0.007799563929438591, + 0.006029848009347916, + 0.006421692669391632, + 0.008172605186700821, + 0.005404672119766474, + 0.008761382661759853, + 0.0350087434053421, + 0.10635267198085785, + 0.027875134721398354, + 0.034364860504865646, + 0.04180968552827835, + 0.047779593616724014, + 0.021821828559041023, + 0.04226396232843399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.0015215021558105946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019452712731435895 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019452712731435895, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019452712731435895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06518242508172989, + 0.11434213072061539, + 0.09129815548658371, + 0.130986750125885, + 0.05803986266255379, + 0.08360056579113007, + 0.06265819072723389, + 0.045713864266872406, + 0.13176313042640686, + 0.21641485393047333 + ], + "credit_map": [ + 0.007751431316137314, + 0.008844427764415741, + 0.011464932933449745, + 0.00724665867164731, + 0.007182737812399864, + 0.007789809722453356, + 0.0076424009166657925, + 0.007260032929480076, + 0.017971739172935486, + 0.018947577103972435, + 0.02090354450047016, + 0.015788745135068893, + 0.006037305109202862, + 0.010404023341834545, + 0.009922826662659645, + 0.014366384595632553, + 0.0076254066079854965, + 0.011115865781903267, + 0.009385677985846996, + 0.006992099806666374, + 0.014312750659883022, + 0.014162133447825909, + 0.017648931592702866, + 0.010055293329060078, + 0.012098194099962711, + 0.015888722613453865, + 0.013906874693930149, + 0.0118491156026721, + 0.019365819171071053, + 0.019232945516705513, + 0.023208588361740112, + 0.015436504036188126, + 0.0043268585577607155, + 0.003990115597844124, + 0.003966882824897766, + 0.005804924760013819, + 0.005455611739307642, + 0.011708970181643963, + 0.013161152601242065, + 0.009625351056456566, + 0.00802402850240469, + 0.012971188873052597, + 0.00901112798601389, + 0.008645295165479183, + 0.014906314201653004, + 0.010517577640712261, + 0.009600439108908176, + 0.009924589656293392, + 0.005256436765193939, + 0.007196591701358557, + 0.005690751597285271, + 0.0070586251094937325, + 0.006210207939147949, + 0.012144912965595722, + 0.009746968746185303, + 0.009353703819215298, + 0.0060087284073233604, + 0.004746389575302601, + 0.006997969001531601, + 0.006183321587741375, + 0.0032238431740552187, + 0.006380931008607149, + 0.005328917410224676, + 0.006843768525868654, + 0.012409809045493603, + 0.04269982874393463, + 0.013690786436200142, + 0.014669869095087051, + 0.012625160627067089, + 0.017027495428919792, + 0.008197877556085587, + 0.010442313738167286, + 0.022662561386823654, + 0.058840587735176086, + 0.019491707906126976, + 0.022023377940058708, + 0.02554956078529358, + 0.02759304642677307, + 0.016284512355923653, + 0.02396952174603939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.0019452712731435895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865409806370735 + ], + "opposite_logprob_tokens": [ + -3.048654079437256 + ], + "logprob_reward": -0.04865409806370735, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865409806370735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08965031802654266, + 0.09955666214227676, + 0.1699702888727188, + 0.12329062819480896, + 0.11591318994760513, + 0.08296501636505127, + 0.11015672236680984, + 0.04561110585927963, + 0.09283450245857239, + 0.07005159556865692 + ], + "credit_map": [ + 0.008826470002532005, + 0.007757635321468115, + 0.017175208777189255, + 0.010428773239254951, + 0.013012147508561611, + 0.014663802459836006, + 0.007162805646657944, + 0.010623469017446041, + 0.011802159249782562, + 0.009060324169695377, + 0.02138635702431202, + 0.01080677006393671, + 0.008394923992455006, + 0.012068664655089378, + 0.012917127460241318, + 0.013120332732796669, + 0.01252009067684412, + 0.017389649525284767, + 0.021608246490359306, + 0.013378209434449673, + 0.02817218005657196, + 0.026148125529289246, + 0.03010699152946472, + 0.02064678631722927, + 0.014754303731024265, + 0.012636239640414715, + 0.015120785683393478, + 0.013669497333467007, + 0.013702236115932465, + 0.01477167010307312, + 0.021776895970106125, + 0.016858983784914017, + 0.011083790101110935, + 0.012633166275918484, + 0.0109994076192379, + 0.011604277417063713, + 0.011406837962567806, + 0.020378224551677704, + 0.02162480726838112, + 0.01618267595767975, + 0.00663764076307416, + 0.0067647420801222324, + 0.007164711598306894, + 0.008069450035691261, + 0.01115104928612709, + 0.012421936728060246, + 0.017458032816648483, + 0.013297450728714466, + 0.012222719378769398, + 0.016503777354955673, + 0.01526424940675497, + 0.013210822828114033, + 0.008952953852713108, + 0.016850873827934265, + 0.011859237216413021, + 0.015292080119252205, + 0.0043512796983122826, + 0.005359021481126547, + 0.0057713985443115234, + 0.007949569262564182, + 0.0038164211437106133, + 0.007044543977826834, + 0.004367702174931765, + 0.006951162591576576, + 0.010742377489805222, + 0.030641982331871986, + 0.008692074567079544, + 0.008709151297807693, + 0.007094105239957571, + 0.013192661106586456, + 0.004897516220808029, + 0.008864629082381725, + 0.006301105488091707, + 0.01107917819172144, + 0.00920406635850668, + 0.013307253830134869, + 0.006399804726243019, + 0.00889917928725481, + 0.00835218746215105, + 0.006508816033601761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.04865409806370735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from above, captured in extreme slow motion using an ultra-high-speed camera, captures the decisive moment. A pristine, flat, unfolded $100 bill rests against a matte black background. Benjamin Franklin's portrait, the blue security thread, and the delicate engravings are strikingly clear under the soft overhead light. Suddenly, a rapidly spinning bullet pierces the frame vertically from above. The intense friction between the brass tip and the air produces a fiery, almost incandescent orange glow, the sole dynamic source of light in the otherwise dark scene. The moment the bullet's tip contacts the bill, the paper fibers stretch and deform at a visible speed, then violently tear apart, creating a widening tear. Countless tiny paper and ink particles are blasted in all directions by the impact, the edges of the tear slightly curling and charring from the scorching heat of the bullet. The bullet penetrates effortlessly, continuing its unstoppable trajectory, leaving a rough-edged but perfectly shaped circular hole in the center of the banknote. The entire banknote trembles and undulates violently due to this instantaneous violent impact, forming ripples. The entire picture is full of power and the aesthetics of destruction, and the atmosphere is tense and shocking.", + "seed": 21059, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_7123_seed_21059_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04135949746705592, + "vlm_logprob": -0.04135949746705592, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04135949746705592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12045145779848099 + ], + "opposite_logprob_tokens": [ + -2.2454514503479004 + ], + "logprob_reward": -0.12045145779848099, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12045145779848099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1395687758922577, + 0.07146298140287399, + 0.05527396872639656, + 0.038154393434524536, + 0.06928206235170364, + 0.12252742052078247, + 0.050668492913246155, + 0.3251822888851166, + 0.07323414832353592, + 0.05464549735188484 + ], + "credit_map": [ + 0.02354707382619381, + 0.013661538250744343, + 0.011698367074131966, + 0.024086307734251022, + 0.01685093156993389, + 0.01656409539282322, + 0.013824918307363987, + 0.019335534423589706, + 0.007279064040631056, + 0.008526249788701534, + 0.006693591829389334, + 0.012242146767675877, + 0.01383521780371666, + 0.006290922407060862, + 0.009085721336305141, + 0.007510063238441944, + 0.005202391650527716, + 0.007285310886800289, + 0.004941900260746479, + 0.006469128653407097, + 0.012119406834244728, + 0.0091268140822649, + 0.005147501360625029, + 0.004981509409844875, + 0.005121776368469, + 0.004252416081726551, + 0.004027257207781076, + 0.00622867327183485, + 0.00483652064576745, + 0.007741210982203484, + 0.002259677741676569, + 0.0036868597380816936, + 0.01817278191447258, + 0.005337764509022236, + 0.006274474784731865, + 0.0073857917450368404, + 0.005651913117617369, + 0.017232589423656464, + 0.004195385612547398, + 0.00503135472536087, + 0.013371005654335022, + 0.009716435335576534, + 0.006114528514444828, + 0.004312068689614534, + 0.039298128336668015, + 0.013941893354058266, + 0.006159236654639244, + 0.02961411327123642, + 0.00852400716394186, + 0.006165761966258287, + 0.0069662826135754585, + 0.003512735478579998, + 0.007108686491847038, + 0.006988661829382181, + 0.005132253747433424, + 0.006270098965615034, + 0.015257145278155804, + 0.023223573341965675, + 0.034662969410419464, + 0.007843518629670143, + 0.09841796010732651, + 0.10251664370298386, + 0.014621926471590996, + 0.02863854356110096, + 0.006573291029781103, + 0.012102719396352768, + 0.006757303606718779, + 0.008720140904188156, + 0.010486740618944168, + 0.010736013762652874, + 0.004766299854964018, + 0.013091634027659893, + 0.004818716552108526, + 0.012987226247787476, + 0.00436085881665349, + 0.00718774925917387, + 0.006741682533174753, + 0.007237996906042099, + 0.004916779696941376, + 0.006394484080374241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.12045145779848099 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373829692602158 + ], + "opposite_logprob_tokens": [ + -3.651373863220215 + ], + "logprob_reward": -0.026373829692602158, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373829692602158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18674416840076447, + 0.1463974565267563, + 0.05838266387581825, + 0.04183468967676163, + 0.049996741116046906, + 0.2532443702220917, + 0.051211901009082794, + 0.13937820494174957, + 0.05729150027036667, + 0.015518255531787872 + ], + "credit_map": [ + 0.027369828894734383, + 0.013684296049177647, + 0.01369418017566204, + 0.049980003386735916, + 0.018291793763637543, + 0.016427725553512573, + 0.02436569333076477, + 0.022930648177862167, + 0.0128254946321249, + 0.019379690289497375, + 0.015792472288012505, + 0.02370617538690567, + 0.02610844559967518, + 0.014763126149773598, + 0.01786796748638153, + 0.01595407910645008, + 0.005965136457234621, + 0.008455743081867695, + 0.0031545832753181458, + 0.004582483321428299, + 0.01651797816157341, + 0.008480481803417206, + 0.0060325912199914455, + 0.00519366143271327, + 0.00517564220353961, + 0.004418789409101009, + 0.00438876124098897, + 0.006182726472616196, + 0.006810187362134457, + 0.009028327651321888, + 0.002262258902192116, + 0.003567997133359313, + 0.012249368242919445, + 0.005120945628732443, + 0.0038508239667862654, + 0.003469097428023815, + 0.006253672298043966, + 0.012126726098358631, + 0.0035048709250986576, + 0.003421236528083682, + 0.02881280891597271, + 0.022129341959953308, + 0.012577717192471027, + 0.006848458666354418, + 0.09353719651699066, + 0.026514429599046707, + 0.013276299461722374, + 0.04954812675714493, + 0.007843712344765663, + 0.006448453292250633, + 0.00776282325387001, + 0.004019712097942829, + 0.00829000398516655, + 0.0056469738483428955, + 0.006078275386244059, + 0.005121943540871143, + 0.010926398448646069, + 0.014537478797137737, + 0.008570178411900997, + 0.005182766355574131, + 0.04438405483961105, + 0.014913604594767094, + 0.007530051749199629, + 0.0333336666226387, + 0.007303536403924227, + 0.015290357172489166, + 0.00526095274835825, + 0.00484848627820611, + 0.006604162510484457, + 0.006908836774528027, + 0.0044904896058142185, + 0.006584676913917065, + 0.001882755197584629, + 0.0026696089189499617, + 0.0012380857951939106, + 0.0019033370772376657, + 0.0022750042844563723, + 0.001763547072187066, + 0.0017456549685448408, + 0.0020402621012181044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.026373829692602158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028255560901015997 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028255560901015997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028255560901015997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07742521911859512, + 0.05506698414683342, + 0.1745535135269165, + 0.03721519559621811, + 0.25876760482788086, + 0.08187762647867203, + 0.08426771312952042, + 0.16903091967105865, + 0.03411731496453285, + 0.027677854523062706 + ], + "credit_map": [ + 0.01087130419909954, + 0.00785908568650484, + 0.008950024843215942, + 0.01402287743985653, + 0.012439895421266556, + 0.007839813828468323, + 0.008953308686614037, + 0.006488921120762825, + 0.005527961999177933, + 0.006818774621933699, + 0.0035291078966110945, + 0.00660706777125597, + 0.0074131968431174755, + 0.006389255169779062, + 0.008491018787026405, + 0.010290606878697872, + 0.017744600772857666, + 0.021810946986079216, + 0.008668831549584866, + 0.010823413729667664, + 0.04768512025475502, + 0.03714965656399727, + 0.016537033021450043, + 0.014133945107460022, + 0.0055668894201517105, + 0.003880512434989214, + 0.003439940745010972, + 0.0033589673694223166, + 0.006681447383016348, + 0.010057113133370876, + 0.001845720806159079, + 0.002384610241279006, + 0.0407155342400074, + 0.04082506522536278, + 0.016478579491376877, + 0.018027151003479958, + 0.058257635682821274, + 0.028583461418747902, + 0.03872523456811905, + 0.017154961824417114, + 0.018151508644223213, + 0.005250493995845318, + 0.011491897515952587, + 0.003442871617153287, + 0.016183439642190933, + 0.0075851124711334705, + 0.003952444531023502, + 0.015819866210222244, + 0.010336820036172867, + 0.0076955766417086124, + 0.009782458655536175, + 0.006295241881161928, + 0.01828484609723091, + 0.008170313201844692, + 0.014352983795106411, + 0.009349481202661991, + 0.010200376622378826, + 0.01194819062948227, + 0.011472721584141254, + 0.0041289725340902805, + 0.059075020253658295, + 0.02373177371919155, + 0.008132049813866615, + 0.040341816842556, + 0.005426682066172361, + 0.004421189893037081, + 0.0036564397159963846, + 0.004392543341964483, + 0.0039977263659238815, + 0.003244555089622736, + 0.002713180612772703, + 0.006265002768486738, + 0.0036109855864197016, + 0.0037777742836624384, + 0.002293152967467904, + 0.0032324970234185457, + 0.003904647659510374, + 0.0026747388765215874, + 0.0036018153186887503, + 0.004582245368510485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.0028255560901015997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200921531766653 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002200921531766653, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200921531766653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08009234815835953, + 0.07154083251953125, + 0.04519224166870117, + 0.03297874704003334, + 0.13514019548892975, + 0.2031753808259964, + 0.03522494062781334, + 0.23568889498710632, + 0.13220460712909698, + 0.02876170352101326 + ], + "credit_map": [ + 0.012375965714454651, + 0.007772138342261314, + 0.004100677091628313, + 0.01239154301583767, + 0.010852237232029438, + 0.012037456035614014, + 0.008181726559996605, + 0.012380608357489109, + 0.007449022494256496, + 0.01010158471763134, + 0.0066003017127513885, + 0.011829166673123837, + 0.014585588127374649, + 0.005917645059525967, + 0.008839809335768223, + 0.006217725574970245, + 0.004145538900047541, + 0.007496200036257505, + 0.003007331397384405, + 0.0041539366357028484, + 0.01188532542437315, + 0.006207730621099472, + 0.004530841019004583, + 0.0037653420586138964, + 0.0034636787604540586, + 0.004075970966368914, + 0.0022689877077937126, + 0.00548104103654623, + 0.006956582423299551, + 0.005415051244199276, + 0.0024014597292989492, + 0.0029159823898226023, + 0.021289395168423653, + 0.020172489807009697, + 0.009864156134426594, + 0.0109982555732131, + 0.028553182259202003, + 0.015233206562697887, + 0.019538000226020813, + 0.009491541422903538, + 0.023971647024154663, + 0.018912307918071747, + 0.008697543293237686, + 0.00557416258379817, + 0.07792752981185913, + 0.021638084203004837, + 0.010781090706586838, + 0.03567304462194443, + 0.004231204278767109, + 0.004960482474416494, + 0.002892847638577223, + 0.0032253339886665344, + 0.007684281561523676, + 0.005222704727202654, + 0.002893032506108284, + 0.004115056712180376, + 0.0132500184699893, + 0.01736047863960266, + 0.019807754084467888, + 0.006436688359826803, + 0.07425793260335922, + 0.049119994044303894, + 0.01027880609035492, + 0.045177266001701355, + 0.017825281247496605, + 0.020671747624874115, + 0.014010732993483543, + 0.013817643746733665, + 0.016576509922742844, + 0.01962307095527649, + 0.01386299729347229, + 0.01581663079559803, + 0.0025319531559944153, + 0.0056165894493460655, + 0.0022360992152243853, + 0.004984972532838583, + 0.004192324820905924, + 0.0030918526463210583, + 0.002435646252706647, + 0.0036722682416439056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.002200921531766653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054945722222328186 + ], + "opposite_logprob_tokens": [ + -2.929945707321167 + ], + "logprob_reward": -0.054945722222328186, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054945722222328186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06086753308773041, + 0.052513353526592255, + 0.03384162113070488, + 0.01889335736632347, + 0.08654704689979553, + 0.2955354154109955, + 0.07228953391313553, + 0.2997199296951294, + 0.05294662341475487, + 0.026845593005418777 + ], + "credit_map": [ + 0.007060451898723841, + 0.006917102728039026, + 0.004655158147215843, + 0.011449167504906654, + 0.009219870902597904, + 0.008328765630722046, + 0.00699126860126853, + 0.006245751399546862, + 0.004707127809524536, + 0.0048621296882629395, + 0.005951914470642805, + 0.005833202973008156, + 0.009428300894796848, + 0.004943842068314552, + 0.0073555270209908485, + 0.009431312792003155, + 0.003195670200511813, + 0.00434112036600709, + 0.0021584010683000088, + 0.002705981954932213, + 0.008619222790002823, + 0.00690647354349494, + 0.003231180366128683, + 0.0026835743337869644, + 0.00211144145578146, + 0.0027208831161260605, + 0.002113315276801586, + 0.0023318463936448097, + 0.003327366430312395, + 0.0029454929754137993, + 0.001307299011386931, + 0.002035713754594326, + 0.022650012746453285, + 0.005792190786451101, + 0.007203532848507166, + 0.0074377343989908695, + 0.0071595823392271996, + 0.0267312191426754, + 0.004098966717720032, + 0.0054738144390285015, + 0.046477921307086945, + 0.027860002592206, + 0.017617907375097275, + 0.007399897091090679, + 0.11601206660270691, + 0.029739724472165108, + 0.017779331654310226, + 0.03264855593442917, + 0.01589042879641056, + 0.007676738779991865, + 0.015099123120307922, + 0.004932065960019827, + 0.009093032218515873, + 0.006811237893998623, + 0.0073472424410283566, + 0.005439662374556065, + 0.015362044796347618, + 0.014437203295528889, + 0.03391969949007034, + 0.005066543351858854, + 0.0891864076256752, + 0.1067112609744072, + 0.011485557071864605, + 0.02355121448636055, + 0.003910429775714874, + 0.007140772882848978, + 0.004441819619387388, + 0.00573268998414278, + 0.007523094769567251, + 0.008513971231877804, + 0.004140201490372419, + 0.01154364738613367, + 0.0027835064101964235, + 0.006015438120812178, + 0.0019486909732222557, + 0.003232925198972225, + 0.004064477048814297, + 0.002596532693132758, + 0.002523528877645731, + 0.003680494846776128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.054945722222328186 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from above, captured in extreme slow motion using an ultra-high-speed camera, captures the decisive moment. A pristine, flat, unfolded $100 bill rests against a matte black background. Benjamin Franklin's portrait, the blue security thread, and the delicate engravings are strikingly clear under the soft overhead light. Suddenly, a rapidly spinning bullet pierces the frame vertically from above. The intense friction between the brass tip and the air produces a fiery, almost incandescent orange glow, the sole dynamic source of light in the otherwise dark scene. The moment the bullet's tip contacts the bill, the paper fibers stretch and deform at a visible speed, then violently tear apart, creating a widening tear. Countless tiny paper and ink particles are blasted in all directions by the impact, the edges of the tear slightly curling and charring from the scorching heat of the bullet. The bullet penetrates effortlessly, continuing its unstoppable trajectory, leaving a rough-edged but perfectly shaped circular hole in the center of the banknote. The entire banknote trembles and undulates violently due to this instantaneous violent impact, forming ripples. The entire picture is full of power and the aesthetics of destruction, and the atmosphere is tense and shocking.", + "seed": 21060, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_1/index_7123_seed_21060_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01876833864953369, + "vlm_logprob": -0.01876833864953369, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.01876833864953369, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0374913364648819 + ], + "opposite_logprob_tokens": [ + -3.6624913215637207 + ], + "logprob_reward": -0.0374913364648819, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0374913364648819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14214354753494263, + 0.02731352671980858, + 0.031743988394737244, + 0.041304368525743484, + 0.07455527782440186, + 0.03343978151679039, + 0.5416198372840881, + 0.048128388822078705, + 0.03579680994153023, + 0.023954464122653008 + ], + "credit_map": [ + 0.010044996626675129, + 0.0157167948782444, + 0.010698090307414532, + 0.00925193540751934, + 0.033274758607149124, + 0.03296925500035286, + 0.01007361151278019, + 0.020114116370677948, + 0.0024598254822194576, + 0.004591220058500767, + 0.0028454079292714596, + 0.002158368704840541, + 0.007146342657506466, + 0.0035615614615380764, + 0.0024687950499355793, + 0.002082008169963956, + 0.003404093673452735, + 0.004911181051284075, + 0.0042100585997104645, + 0.0027398932725191116, + 0.005210903473198414, + 0.003909089136868715, + 0.004191147163510323, + 0.0031676243524998426, + 0.004211745224893093, + 0.005044904071837664, + 0.0056272707879543304, + 0.004646139685064554, + 0.0048287357203662395, + 0.006379988044500351, + 0.003548656590282917, + 0.007016931660473347, + 0.007482393644750118, + 0.016767501831054688, + 0.008487760089337826, + 0.006856945343315601, + 0.006786921061575413, + 0.011509882286190987, + 0.005153188016265631, + 0.011510689742863178, + 0.002668696688488126, + 0.007828330621123314, + 0.003263516118749976, + 0.0026886758860200644, + 0.0054674516431987286, + 0.004984306171536446, + 0.002512301318347454, + 0.0040265037678182125, + 0.09459086507558823, + 0.049477145075798035, + 0.12968045473098755, + 0.025024276226758957, + 0.028982464224100113, + 0.1638040542602539, + 0.014956280589103699, + 0.03510434925556183, + 0.0055981287732720375, + 0.016235027462244034, + 0.00767486821860075, + 0.0035566729493439198, + 0.005154167301952839, + 0.0038688830099999905, + 0.003447883762419224, + 0.002592760603874922, + 0.006427360698580742, + 0.006639320403337479, + 0.00612267293035984, + 0.00362230627797544, + 0.004240693058818579, + 0.0038387340027838945, + 0.002439452800899744, + 0.002466274658218026, + 0.0031001234892755747, + 0.003985053393989801, + 0.002213474363088608, + 0.0024251118302345276, + 0.00281655159778893, + 0.004588091745972633, + 0.0016731099458411336, + 0.003152951132506132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0374913364648819 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023310668766498566 + ], + "opposite_logprob_tokens": [ + -3.773310661315918 + ], + "logprob_reward": -0.023310668766498566, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023310668766498566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10682555288076401, + 0.0167863592505455, + 0.04649850353598595, + 0.04353877902030945, + 0.030131710693240166, + 0.019147275015711784, + 0.6254523992538452, + 0.07668981701135635, + 0.013903332874178886, + 0.02102627418935299 + ], + "credit_map": [ + 0.003941730596125126, + 0.03984162211418152, + 0.006062288302928209, + 0.003964108880609274, + 0.036245960742235184, + 0.006023594178259373, + 0.0056441291235387325, + 0.0051021212711930275, + 0.0012472766684368253, + 0.002325724810361862, + 0.0023809364065527916, + 0.001216833246871829, + 0.0038394001312553883, + 0.0023310319520533085, + 0.0018311301246285439, + 0.0016140260268002748, + 0.00544585008174181, + 0.00722883827984333, + 0.006686249747872353, + 0.003962490241974592, + 0.006547463592141867, + 0.006711853202432394, + 0.004512234590947628, + 0.005403525661677122, + 0.004209536127746105, + 0.006198414601385593, + 0.003790855873376131, + 0.004028132185339928, + 0.007347294595092535, + 0.008374670520424843, + 0.005438166670501232, + 0.004151707980781794, + 0.0013007220113649964, + 0.010393540374934673, + 0.0010462731588631868, + 0.0016161707462742925, + 0.004015084821730852, + 0.009046600200235844, + 0.0015036540571600199, + 0.0012096677673980594, + 0.0018093017861247063, + 0.0033573750406503677, + 0.0015771203907206655, + 0.0018466856563463807, + 0.00378561788238585, + 0.0023360750637948513, + 0.0024431333877146244, + 0.001991966972127557, + 0.11029113829135895, + 0.03375587239861488, + 0.14489707350730896, + 0.028444936498999596, + 0.03167411684989929, + 0.22061443328857422, + 0.016521165147423744, + 0.03925364464521408, + 0.011041088029742241, + 0.02990039251744747, + 0.013665477745234966, + 0.003914013970643282, + 0.00822629127651453, + 0.0042140125297009945, + 0.0026778625324368477, + 0.003050682134926319, + 0.0017814283492043614, + 0.0031596284825354815, + 0.0016146256821230054, + 0.0013429397949948907, + 0.0019287379691377282, + 0.0016329723875969648, + 0.0010571941966190934, + 0.0013858056627213955, + 0.003146962495520711, + 0.00457785977050662, + 0.0018623188370838761, + 0.0017339761834591627, + 0.002120842458680272, + 0.003704251954331994, + 0.001450314070098102, + 0.002429749583825469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.023310668766498566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001723829424008727 + ], + "opposite_logprob_tokens": [ + -6.376723766326904 + ], + "logprob_reward": -0.001723829424008727, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001723829424008727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07001303136348724, + 0.014547125436365604, + 0.023646211251616478, + 0.01972786709666252, + 0.024784985929727554, + 0.02754773572087288, + 0.7397032976150513, + 0.04032561555504799, + 0.010917133651673794, + 0.028787042945623398 + ], + "credit_map": [ + 0.003910204861313105, + 0.020077213644981384, + 0.0062573738396167755, + 0.004205172415822744, + 0.02031191624701023, + 0.004394763149321079, + 0.004741082899272442, + 0.006115305237472057, + 0.0012615957530215383, + 0.0022543969098478556, + 0.0007785410271026194, + 0.0007244418957270682, + 0.005078256130218506, + 0.0025024458300322294, + 0.0010994876502081752, + 0.0008479607058688998, + 0.0022347026970237494, + 0.003846649546176195, + 0.003181465668603778, + 0.0024570536334067583, + 0.0033150583039969206, + 0.00403937790542841, + 0.0021107594948261976, + 0.0024611453991383314, + 0.0021780994720757008, + 0.0014355420134961605, + 0.0033307294361293316, + 0.002020282670855522, + 0.002335249213501811, + 0.0026795153971761465, + 0.003193163312971592, + 0.002555286046117544, + 0.0017096686642616987, + 0.007132007274776697, + 0.0011445489944890141, + 0.001515259500592947, + 0.004529714118689299, + 0.00524940388277173, + 0.0021933966781944036, + 0.0013109857682138681, + 0.0019679239485412836, + 0.009124069474637508, + 0.0011872508330270648, + 0.0020496677607297897, + 0.005351609084755182, + 0.004177671391516924, + 0.0019347092602401972, + 0.0017548303585499525, + 0.13031883537769318, + 0.04015339910984039, + 0.17098554968833923, + 0.033878155052661896, + 0.03759697452187538, + 0.26081976294517517, + 0.01958654448390007, + 0.04636404663324356, + 0.00558453006669879, + 0.014998129568994045, + 0.006780754774808884, + 0.002103977371007204, + 0.00426689675077796, + 0.0024413687642663717, + 0.0022813603281974792, + 0.0018685976974666119, + 0.0020641612354665995, + 0.0011905424762517214, + 0.002241090638563037, + 0.0010488743428140879, + 0.0009649414569139481, + 0.0011944860452786088, + 0.001109896693378687, + 0.0011031405301764607, + 0.004192093387246132, + 0.006667150184512138, + 0.002436719136312604, + 0.002186323283240199, + 0.0029862315859645605, + 0.005386505741626024, + 0.0017029063310474157, + 0.0032291142269968987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.001723829424008727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015162649797275662 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015162649797275662, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015162649797275662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029580119997262955, + 0.006047764793038368, + 0.007719333749264479, + 0.009067114442586899, + 0.005153651814907789, + 0.0072255427949130535, + 0.891403317451477, + 0.026425320655107498, + 0.012628670781850815, + 0.004749140236526728 + ], + "credit_map": [ + 0.0014499896205961704, + 0.00806211307644844, + 0.002769890706986189, + 0.001571265165694058, + 0.00907161459326744, + 0.0021341084502637386, + 0.0018185431836172938, + 0.0027025952003896236, + 0.0004728538333438337, + 0.0007946829427964985, + 0.0006529121892526746, + 0.0005536454264074564, + 0.0013296196702867746, + 0.0008227977668866515, + 0.0008974451920948923, + 0.0005238077137619257, + 0.0007112016901373863, + 0.0014362172223627567, + 0.0009484538459219038, + 0.0006957383011467755, + 0.0012056251289322972, + 0.0011710188118740916, + 0.0006259576184675097, + 0.0009251216542907059, + 0.0009550746763125062, + 0.001023606862872839, + 0.0013726961333304644, + 0.0009339548414573073, + 0.0009720941307023168, + 0.001377722597680986, + 0.0008388401474803686, + 0.0015931247035041451, + 0.00040933984564617276, + 0.0014998919796198606, + 0.00040143608930520713, + 0.0003493329568300396, + 0.0005414655897766352, + 0.0010421830229461193, + 0.00033745638211257756, + 0.0005725457449443638, + 0.0007273572846315801, + 0.001470282324589789, + 0.0005159061984159052, + 0.0005539243575185537, + 0.001721601583994925, + 0.0007792329997755587, + 0.0006099595339037478, + 0.0008472781628370285, + 0.1567511111497879, + 0.05366918072104454, + 0.2089083045721054, + 0.04122035205364227, + 0.04597805067896843, + 0.3043835759162903, + 0.024043478071689606, + 0.0564492791891098, + 0.0032967450097203255, + 0.009481077082455158, + 0.004144346807152033, + 0.0014742650091648102, + 0.0028185227420181036, + 0.0022146746050566435, + 0.0016413750126957893, + 0.0013543162494897842, + 0.0024178093299269676, + 0.0028202186804264784, + 0.0018925715703517199, + 0.0009639344061724842, + 0.001561158336699009, + 0.001282534678466618, + 0.0007622939301654696, + 0.0009281500242650509, + 0.000718334165867418, + 0.0009786253795027733, + 0.00042516810935921967, + 0.0005991291254758835, + 0.0004471412394195795, + 0.0006578547763638198, + 0.00040515896398574114, + 0.0005177284474484622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0015162649797275662 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above of a pristine, flat, unfolded $100 bill on a matte black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a rapidly spinning bullet piercing the frame vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bullet's brass tip produce a fiery, incandescent orange glow upon contact with the bill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers stretch, deform, and tear apart with tiny particles blasting in all directions upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979959361255169 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979959361255169, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979959361255169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0401683934032917, + 0.008487025275826454, + 0.00665668211877346, + 0.0116126062348485, + 0.007004776503890753, + 0.01389584131538868, + 0.858790397644043, + 0.031505271792411804, + 0.006673041265457869, + 0.015206051990389824 + ], + "credit_map": [ + 0.001525679137557745, + 0.015077297575771809, + 0.0015831004129722714, + 0.0016984251560643315, + 0.014455491676926613, + 0.00242091272957623, + 0.0016765097388997674, + 0.0017309781396761537, + 0.0007423129864037037, + 0.0011732602724805474, + 0.0009876753902062774, + 0.0007366195786744356, + 0.0018054838292300701, + 0.0010515004396438599, + 0.0011575235985219479, + 0.0008326480747200549, + 0.000633496674709022, + 0.001029539736919105, + 0.000859336054418236, + 0.0006041439482942224, + 0.0011977010872215033, + 0.0008510244078934193, + 0.0005551095237024128, + 0.0009263301617465913, + 0.0010328606003895402, + 0.001127642230130732, + 0.001449869479984045, + 0.0015920937294140458, + 0.0014771653804928064, + 0.0012001446448266506, + 0.00233889976516366, + 0.001393930404447019, + 0.0006494333501905203, + 0.0012135310098528862, + 0.0007063220255076885, + 0.0006385786109603941, + 0.0012112631229683757, + 0.0011913226917386055, + 0.0006170610431581736, + 0.0007772642420604825, + 0.0011279778555035591, + 0.0031167182605713606, + 0.0010124476393684745, + 0.001341231749393046, + 0.0023260151501744986, + 0.002175142290070653, + 0.0013536292826756835, + 0.001442677341401577, + 0.15064547955989838, + 0.046196140348911285, + 0.19907288253307343, + 0.03956925496459007, + 0.043740082532167435, + 0.3029991090297699, + 0.02286580577492714, + 0.05370159447193146, + 0.0041115120984613895, + 0.01171208918094635, + 0.006386584136635065, + 0.001780419028364122, + 0.0034045183565467596, + 0.0016115759499371052, + 0.0012283660471439362, + 0.0012702043168246746, + 0.0009889150969684124, + 0.0011660282034426928, + 0.0010786900529637933, + 0.0004857531748712063, + 0.0007845614454708993, + 0.00092632498126477, + 0.0005258872406557202, + 0.0007168804877437651, + 0.0024038052652031183, + 0.003854437731206417, + 0.0018085918854922056, + 0.001401350717060268, + 0.0012703102547675371, + 0.0016333621460944414, + 0.0010720628779381514, + 0.0017621306469663978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.02979959361255169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_tensor([2646, 2646])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_tensor([2646, 2646])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d2d44f25fd8247a88aff1b867613cb77f546ca43 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_tensor([2646, 2646])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, statically captured in a 16:9 aspect ratio, captures the digital miracle of life, steeped in a refined, nostalgic gaming aesthetic. Against a solid background of dark blue and purple pixels, a compact, spherical bud, its surface composed of crimson and magenta pixels, is the sole focal point. Its surface, composed of blocks of crimson and magenta pixels, is subtly dithered to create a subtle three-dimensional effect. With a silent start, the outermost layer of pixelated petals at the bud's tip cautiously pokes open, revealing the more vibrant layers of color within. Then, with a fluid, yet distinctly digital rhythm, the entire flower visibly unfolds, layer upon layer of pixelated petals unfolding one by one, each with a distinct pixelated outline, as if meticulously painted by a skilled pixel artist. A soft, analog overhead light illuminates this slow blossoming process, casting simple, yet defined pixel shadows across the petals. Eventually, the flower fully opens into a perfect, symmetrical blossom with a vintage digital aesthetic, its limited yet saturated color palette contrasting beautifully with the deep background, creating a high-definition pixel art atmosphere that is both serene and captivating.", + "seed": 7989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_2646_seed_7989_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0844211552204797, + "vlm_logprob": -0.0844211552204797, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.0844211552204797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026909051463007927 + ], + "opposite_logprob_tokens": [ + -3.776909112930298 + ], + "logprob_reward": -0.026909051463007927, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026909051463007927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09262799471616745, + 0.32144078612327576, + 0.3619418740272522, + 0.028749793767929077, + 0.02537013217806816, + 0.016423750668764114, + 0.05757720395922661, + 0.06630143523216248, + 0.018076647073030472, + 0.011490358039736748 + ], + "credit_map": [ + 0.013932524248957634, + 0.0074625746347010136, + 0.015555676072835922, + 0.010084199719130993, + 0.006972594652324915, + 0.015597461722791195, + 0.011385185644030571, + 0.011637783609330654, + 0.04380731284618378, + 0.02386692725121975, + 0.046535030007362366, + 0.020258689299225807, + 0.028297273442149162, + 0.05216211825609207, + 0.02504882961511612, + 0.08146458864212036, + 0.07889597862958908, + 0.024700261652469635, + 0.04855811223387718, + 0.025182077661156654, + 0.022300275042653084, + 0.05959107354283333, + 0.022387640550732613, + 0.08032649010419846, + 0.003339659422636032, + 0.00219043530523777, + 0.004336249548941851, + 0.003137409221380949, + 0.002277497434988618, + 0.004572635516524315, + 0.0032101497054100037, + 0.005685757379978895, + 0.004425987601280212, + 0.0018844424048438668, + 0.004025634843856096, + 0.0019151297165080905, + 0.001451867981813848, + 0.0043232617899775505, + 0.0017695612041279674, + 0.005574246868491173, + 0.0017039075028151274, + 0.001703899004496634, + 0.002336602658033371, + 0.001474079443141818, + 0.0019924044609069824, + 0.002489949343726039, + 0.001622317242436111, + 0.0031005917116999626, + 0.004245045129209757, + 0.002092621987685561, + 0.007887393236160278, + 0.001576935756020248, + 0.0020653505343943834, + 0.0157436765730381, + 0.0014927454758435488, + 0.02247343212366104, + 0.009755831211805344, + 0.004777030553668737, + 0.009674405679106712, + 0.004951945971697569, + 0.004168608691543341, + 0.01260401587933302, + 0.004804558586329222, + 0.015565040521323681, + 0.002609464805573225, + 0.001246224856004119, + 0.002065821085125208, + 0.0013842334738001227, + 0.0013791577657684684, + 0.003525662003085017, + 0.0017954770009964705, + 0.004070607014000416, + 0.0014961780980229378, + 0.0012784722493961453, + 0.0008180459844879806, + 0.0009730812162160873, + 0.0012070653028786182, + 0.0020845176186412573, + 0.0014505047583952546, + 0.0021824927534908056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.026909051463007927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636283989995718 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636283989995718, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636283989995718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1940736174583435, + 0.06916298717260361, + 0.0665961280465126, + 0.1235644593834877, + 0.07342397421598434, + 0.13889893889427185, + 0.07085097581148148, + 0.07538417726755142, + 0.13282200694084167, + 0.05522270128130913 + ], + "credit_map": [ + 0.026857003569602966, + 0.01946582831442356, + 0.03576941415667534, + 0.02131141535937786, + 0.014461434446275234, + 0.03086548112332821, + 0.019520506262779236, + 0.025822538882493973, + 0.006513983476907015, + 0.005621643736958504, + 0.010071149095892906, + 0.0074608419090509415, + 0.006037456914782524, + 0.013366817496716976, + 0.0092135239392519, + 0.010877576656639576, + 0.01139308325946331, + 0.0058530764654278755, + 0.00926438719034195, + 0.010095611214637756, + 0.0048063634894788265, + 0.010764549486339092, + 0.007504324894398451, + 0.0069147446192801, + 0.030606713145971298, + 0.009314503520727158, + 0.021002279594540596, + 0.009466289542615414, + 0.005300451535731554, + 0.01708906702697277, + 0.007591891102492809, + 0.023193271830677986, + 0.011380156502127647, + 0.006472490262240171, + 0.00885970238596201, + 0.011012532748281956, + 0.006431880872696638, + 0.009616216644644737, + 0.008824567310512066, + 0.010826426558196545, + 0.020020728930830956, + 0.013969606719911098, + 0.01835552416741848, + 0.0124970069155097, + 0.01321118138730526, + 0.021082669496536255, + 0.015118221752345562, + 0.02464400790631771, + 0.007174678146839142, + 0.005056994501501322, + 0.014011743478477001, + 0.003253589617088437, + 0.005819406360387802, + 0.009913472458720207, + 0.003961416892707348, + 0.021659689024090767, + 0.006224352866411209, + 0.004823834635317326, + 0.007530501578003168, + 0.006347121205180883, + 0.005783569533377886, + 0.01668778993189335, + 0.005977261811494827, + 0.02200975827872753, + 0.011479640379548073, + 0.007708180695772171, + 0.013645161874592304, + 0.010782609693706036, + 0.008184644393622875, + 0.02899160422384739, + 0.012893042527139187, + 0.03913713991641998, + 0.006886737886816263, + 0.008274801075458527, + 0.003565338207408786, + 0.007151175290346146, + 0.007552444003522396, + 0.00802954938262701, + 0.004994669463485479, + 0.008767991326749325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.007636283989995718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00034290633630007505 + ], + "opposite_logprob_tokens": [ + -8.000343322753906 + ], + "logprob_reward": -0.00034290633630007505, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00034290633630007505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30945122241973877, + 0.18027858436107635, + 0.10849136114120483, + 0.08298076689243317, + 0.04351486265659332, + 0.042026225477457047, + 0.11315187811851501, + 0.05575951933860779, + 0.04661909490823746, + 0.017726434394717216 + ], + "credit_map": [ + 0.028208939358592033, + 0.03131891041994095, + 0.03519127517938614, + 0.047880131751298904, + 0.037124112248420715, + 0.05868016555905342, + 0.038042791187763214, + 0.033004891127347946, + 0.02031528949737549, + 0.01459901686757803, + 0.02185250259935856, + 0.021004866808652878, + 0.015182863920927048, + 0.02944815717637539, + 0.022630199790000916, + 0.03524567931890488, + 0.01824011653661728, + 0.012628651224076748, + 0.009847760200500488, + 0.011232422664761543, + 0.009731433354318142, + 0.01483881939202547, + 0.012946026399731636, + 0.019026128575205803, + 0.018299931660294533, + 0.009770200587809086, + 0.009708626195788383, + 0.010708814486861229, + 0.009098546579480171, + 0.0094283577054739, + 0.007891600020229816, + 0.008074695244431496, + 0.006179735995829105, + 0.0037067995872348547, + 0.006514957174658775, + 0.00454067625105381, + 0.003609949257224798, + 0.006097504869103432, + 0.0065727862529456615, + 0.006292452570050955, + 0.005780124105513096, + 0.004011315759271383, + 0.005700405221432447, + 0.004255134146660566, + 0.0040519582107663155, + 0.005893409252166748, + 0.004686808213591576, + 0.0076470691710710526, + 0.007117836736142635, + 0.005201211664825678, + 0.020499389618635178, + 0.004142399877309799, + 0.006246208678930998, + 0.022738296538591385, + 0.003963883966207504, + 0.043242648243904114, + 0.00627497024834156, + 0.005572094582021236, + 0.00750325620174408, + 0.005801254417747259, + 0.005014084745198488, + 0.009859674610197544, + 0.0056084152311086655, + 0.010125771164894104, + 0.0038636098615825176, + 0.005083590745925903, + 0.00496335607022047, + 0.005538987927138805, + 0.003661117982119322, + 0.008150466717779636, + 0.004764425568282604, + 0.010593537241220474, + 0.0027062075678259134, + 0.0019593292381614447, + 0.0020873958710581064, + 0.0018422439461573958, + 0.0016103263478726149, + 0.0035500838421285152, + 0.0012597040040418506, + 0.0027111442759633064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00034290633630007505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00018523407925385982 + ], + "opposite_logprob_tokens": [ + -8.625185012817383 + ], + "logprob_reward": -0.00018523407925385982, + "yes_no_margin_reward": 8.625, + "token_credit_reward": -0.00018523407925385982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13552910089492798, + 0.2514741122722626, + 0.11580362915992737, + 0.08737819641828537, + 0.06225934624671936, + 0.040093421936035156, + 0.07324454188346863, + 0.04734618589282036, + 0.15413951873779297, + 0.032731957733631134 + ], + "credit_map": [ + 0.01723521202802658, + 0.010937921702861786, + 0.021237479522824287, + 0.016905127093195915, + 0.010889137163758278, + 0.017780637368559837, + 0.020550906658172607, + 0.01999267190694809, + 0.03432116284966469, + 0.01636471413075924, + 0.03537201136350632, + 0.01755163073539734, + 0.016011331230401993, + 0.045966699719429016, + 0.02086370252072811, + 0.06502285599708557, + 0.015976237133145332, + 0.011597142554819584, + 0.014491546899080276, + 0.012102566659450531, + 0.0077011436223983765, + 0.01756611093878746, + 0.014366798102855682, + 0.02200208604335785, + 0.01581661030650139, + 0.008027654141187668, + 0.012476040050387383, + 0.00977235846221447, + 0.009653531014919281, + 0.014215821400284767, + 0.008279531262814999, + 0.009136653505265713, + 0.00857551023364067, + 0.0062150838784873486, + 0.009540333412587643, + 0.0060653528198599815, + 0.005493639037013054, + 0.00939896423369646, + 0.005861224140971899, + 0.01110923383384943, + 0.006657102145254612, + 0.00435575470328331, + 0.005069770850241184, + 0.003134601516649127, + 0.003504498628899455, + 0.005111917853355408, + 0.0038877970073372126, + 0.00837197806686163, + 0.010889878496527672, + 0.004952278453856707, + 0.014604717493057251, + 0.003959737252444029, + 0.006928791757673025, + 0.010378235951066017, + 0.003836104180663824, + 0.01769479364156723, + 0.008621764369308949, + 0.004325947258621454, + 0.007877502590417862, + 0.004641401581466198, + 0.003973627462983131, + 0.006791105028241873, + 0.003894811263307929, + 0.007220033090561628, + 0.0157537329941988, + 0.009891116991639137, + 0.01958487555384636, + 0.009903914295136929, + 0.008679977618157864, + 0.03481261804699898, + 0.009285807609558105, + 0.046227481216192245, + 0.004942244850099087, + 0.002766881138086319, + 0.0036397220101207495, + 0.0036662325728684664, + 0.0035855029709637165, + 0.005622620694339275, + 0.0027924329042434692, + 0.005716318730264902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.00018523407925385982 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38703230023384094 + ], + "opposite_logprob_tokens": [ + -1.1370322704315186 + ], + "logprob_reward": -0.38703230023384094, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38703230023384094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23354117572307587, + 0.2518080174922943, + 0.050890225917100906, + 0.048584237694740295, + 0.03721855953335762, + 0.10604625940322876, + 0.11358325183391571, + 0.04477786645293236, + 0.08761639147996902, + 0.025934068486094475 + ], + "credit_map": [ + 0.030190715566277504, + 0.01642017997801304, + 0.03746431693434715, + 0.02149791643023491, + 0.015077619813382626, + 0.05380812659859657, + 0.02329249680042267, + 0.03578977286815643, + 0.03421453386545181, + 0.016654958948493004, + 0.03507192060351372, + 0.024343255907297134, + 0.018037835136055946, + 0.04422400891780853, + 0.026666758581995964, + 0.05259472504258156, + 0.009870985522866249, + 0.003420502645894885, + 0.006976560223847628, + 0.004682470578700304, + 0.0030832896009087563, + 0.006471419241279364, + 0.006140949670225382, + 0.010244040749967098, + 0.006796073168516159, + 0.003099584486335516, + 0.006662874482572079, + 0.004403283353894949, + 0.0025707853492349386, + 0.008339647203683853, + 0.004759894218295813, + 0.011952094733715057, + 0.004001474007964134, + 0.0032996037043631077, + 0.004889084491878748, + 0.0039118193089962006, + 0.003547787433490157, + 0.006326382514089346, + 0.003836776129901409, + 0.007405626121908426, + 0.010485527105629444, + 0.008593416772782803, + 0.011810152791440487, + 0.009975596331059933, + 0.007515784353017807, + 0.019325844943523407, + 0.01013555284589529, + 0.02820436656475067, + 0.0247503574937582, + 0.008291874080896378, + 0.015472794882953167, + 0.004851466044783592, + 0.006594558712095022, + 0.02441064640879631, + 0.005668962839990854, + 0.02354259043931961, + 0.007325549144297838, + 0.003940618596971035, + 0.0075027854181826115, + 0.0037495545111596584, + 0.003456759499385953, + 0.006231175269931555, + 0.003991125617176294, + 0.008580293506383896, + 0.010028217919170856, + 0.006053472403436899, + 0.017388692125678062, + 0.006010349839925766, + 0.005097734276205301, + 0.01466717105358839, + 0.006459774915128946, + 0.021910974755883217, + 0.0035274175461381674, + 0.0030114450491964817, + 0.0021773981861770153, + 0.0030027905013412237, + 0.002818692708387971, + 0.0038316689897328615, + 0.002316495170816779, + 0.005248154979199171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.38703230023384094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, statically captured in a 16:9 aspect ratio, captures the digital miracle of life, steeped in a refined, nostalgic gaming aesthetic. Against a solid background of dark blue and purple pixels, a compact, spherical bud, its surface composed of crimson and magenta pixels, is the sole focal point. Its surface, composed of blocks of crimson and magenta pixels, is subtly dithered to create a subtle three-dimensional effect. With a silent start, the outermost layer of pixelated petals at the bud's tip cautiously pokes open, revealing the more vibrant layers of color within. Then, with a fluid, yet distinctly digital rhythm, the entire flower visibly unfolds, layer upon layer of pixelated petals unfolding one by one, each with a distinct pixelated outline, as if meticulously painted by a skilled pixel artist. A soft, analog overhead light illuminates this slow blossoming process, casting simple, yet defined pixel shadows across the petals. Eventually, the flower fully opens into a perfect, symmetrical blossom with a vintage digital aesthetic, its limited yet saturated color palette contrasting beautifully with the deep background, creating a high-definition pixel art atmosphere that is both serene and captivating.", + "seed": 7990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_2646_seed_7990_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.036784039958729406, + "vlm_logprob": -0.036784039958729406, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.036784039958729406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01617417484521866 + ], + "opposite_logprob_tokens": [ + -4.39117431640625 + ], + "logprob_reward": -0.01617417484521866, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01617417484521866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16187433898448944, + 0.07394091039896011, + 0.03951377421617508, + 0.10697393864393234, + 0.03238172084093094, + 0.10561054199934006, + 0.1935611218214035, + 0.11252911388874054, + 0.09016619622707367, + 0.08344832062721252 + ], + "credit_map": [ + 0.03064333088696003, + 0.010897643864154816, + 0.022865265607833862, + 0.02293592505156994, + 0.012374931015074253, + 0.02080298401415348, + 0.022571539506316185, + 0.0187827255576849, + 0.013462652452290058, + 0.0034080466721206903, + 0.014137671329081059, + 0.008391957730054855, + 0.004081178456544876, + 0.011273322626948357, + 0.00860133022069931, + 0.010584751144051552, + 0.005542373284697533, + 0.003551054745912552, + 0.006413683760911226, + 0.004905355628579855, + 0.003153269411996007, + 0.0049735791981220245, + 0.005763047840446234, + 0.005211412440985441, + 0.01152743399143219, + 0.006407173350453377, + 0.019792761653661728, + 0.014591582119464874, + 0.008102775551378727, + 0.02183285541832447, + 0.016099661588668823, + 0.008619684725999832, + 0.003949807491153479, + 0.002786634722724557, + 0.004065149929374456, + 0.005847939755767584, + 0.002994898473843932, + 0.00383500917814672, + 0.003995861392468214, + 0.004906420595943928, + 0.018834494054317474, + 0.007084325421601534, + 0.015683908015489578, + 0.010334978811442852, + 0.006114887539297342, + 0.00926085002720356, + 0.01092428807169199, + 0.02737281285226345, + 0.03249400854110718, + 0.016756098717451096, + 0.03101145103573799, + 0.018146751448512077, + 0.020495589822530746, + 0.019263919442892075, + 0.02235690876841545, + 0.033036403357982635, + 0.008889712393283844, + 0.008310492150485516, + 0.020189933478832245, + 0.011835337616503239, + 0.014697334729135036, + 0.015262536704540253, + 0.01332878042012453, + 0.020014984533190727, + 0.009421433322131634, + 0.010783972218632698, + 0.014081748202443123, + 0.008648362010717392, + 0.01207018457353115, + 0.008574591018259525, + 0.007650818210095167, + 0.01893508993089199, + 0.019646136090159416, + 0.01117750070989132, + 0.013357635587453842, + 0.005716401617974043, + 0.010371332056820393, + 0.007905874401330948, + 0.004888031631708145, + 0.010385408997535706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.01617417484521866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741754245012999 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741754245012999, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741754245012999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1954137235879898, + 0.08042887598276138, + 0.054958682507276535, + 0.11900550872087479, + 0.030164949595928192, + 0.0298911202698946, + 0.037331487983465195, + 0.3876396417617798, + 0.05239951238036156, + 0.012766546569764614 + ], + "credit_map": [ + 0.036898087710142136, + 0.015129044651985168, + 0.026790354400873184, + 0.023676874116063118, + 0.014244158752262592, + 0.028774335980415344, + 0.026351453736424446, + 0.023549409583210945, + 0.013383687473833561, + 0.004430864937603474, + 0.016231728717684746, + 0.005813946016132832, + 0.004301153123378754, + 0.015492148697376251, + 0.007307081017643213, + 0.013468264602124691, + 0.00787662435323, + 0.005024045240134001, + 0.009394361637532711, + 0.007260005921125412, + 0.006164806429296732, + 0.007390448357909918, + 0.005926417652517557, + 0.005921964533627033, + 0.01154597569257021, + 0.006918143946677446, + 0.018904762342572212, + 0.018747881054878235, + 0.00782198179513216, + 0.026279864832758904, + 0.019551897421479225, + 0.009234990924596786, + 0.005490574520081282, + 0.002095738658681512, + 0.004102945327758789, + 0.003764710621908307, + 0.0024511024821549654, + 0.0034591981675475836, + 0.002457541413605213, + 0.006343137007206678, + 0.002934354357421398, + 0.002333641517907381, + 0.003996557090431452, + 0.005788082256913185, + 0.003300098469480872, + 0.0026088689919561148, + 0.004614497069269419, + 0.004315017722547054, + 0.007536744233220816, + 0.0031470174435526133, + 0.006021927110850811, + 0.004128389526158571, + 0.002096105832606554, + 0.0036998249124735594, + 0.003841846715658903, + 0.0068596284836530685, + 0.0406971238553524, + 0.03691089153289795, + 0.075393445789814, + 0.032934825867414474, + 0.03409084677696228, + 0.05463990941643715, + 0.03806275129318237, + 0.07490982115268707, + 0.004139012657105923, + 0.005444242153316736, + 0.005859633907675743, + 0.005778622813522816, + 0.006674717180430889, + 0.006221114192157984, + 0.005735728889703751, + 0.012546435929834843, + 0.0015960084274411201, + 0.0016038290923461318, + 0.002049713861197233, + 0.001412268029525876, + 0.0013139646034687757, + 0.0014383797533810139, + 0.001711539807729423, + 0.0016408426454290748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.006741754245012999 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002374367177253589 + ], + "opposite_logprob_tokens": [ + -8.375237464904785 + ], + "logprob_reward": -0.0002374367177253589, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.0002374367177253589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28646448254585266, + 0.08529184013605118, + 0.06309770792722702, + 0.10200124233961105, + 0.04666724056005478, + 0.08448547124862671, + 0.08750718832015991, + 0.14136390388011932, + 0.064005546271801, + 0.03911532834172249 + ], + "credit_map": [ + 0.04025886952877045, + 0.019953230395913124, + 0.043784547597169876, + 0.028337400406599045, + 0.021253949031233788, + 0.04788715019822121, + 0.035948287695646286, + 0.0490410253405571, + 0.008853808045387268, + 0.005877404008060694, + 0.010983704589307308, + 0.013463624753057957, + 0.00709933415055275, + 0.011460977606475353, + 0.016752231866121292, + 0.010800753720104694, + 0.006632163189351559, + 0.006604034919291735, + 0.0063288332894444466, + 0.012956621125340462, + 0.007521298713982105, + 0.006762798875570297, + 0.010731220245361328, + 0.005560735240578651, + 0.010900058783590794, + 0.007590152323246002, + 0.014520280063152313, + 0.015128028579056263, + 0.008997617289423943, + 0.01864197477698326, + 0.01648615673184395, + 0.009736972860991955, + 0.005391519516706467, + 0.005424984265118837, + 0.00559644028544426, + 0.008251842111349106, + 0.006119662430137396, + 0.0041641113348305225, + 0.006389687303453684, + 0.005328995641320944, + 0.005866463761776686, + 0.00630006892606616, + 0.012257818132638931, + 0.01521975640207529, + 0.008829805999994278, + 0.011516890488564968, + 0.013342353515326977, + 0.011152311228215694, + 0.011177760548889637, + 0.007340323179960251, + 0.01668696664273739, + 0.008700649254024029, + 0.00862970668822527, + 0.008227762766182423, + 0.009574162773787975, + 0.017169853672385216, + 0.015736734494566917, + 0.015542642213404179, + 0.026525119319558144, + 0.014240480028092861, + 0.012171877548098564, + 0.018329694867134094, + 0.013922750018537045, + 0.024894600734114647, + 0.007816510275006294, + 0.009152080863714218, + 0.009673313237726688, + 0.007458499167114496, + 0.00866561010479927, + 0.006989333778619766, + 0.005930060986429453, + 0.008320141583681107, + 0.0048829191364347935, + 0.005005986895412207, + 0.0043480186723172665, + 0.004411384928971529, + 0.005060227587819099, + 0.004656055010855198, + 0.004675913602113724, + 0.006074824370443821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0002374367177253589 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004368066438473761 + ], + "opposite_logprob_tokens": [ + -7.750436782836914 + ], + "logprob_reward": -0.0004368066438473761, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004368066438473761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17779141664505005, + 0.09162525087594986, + 0.06234419345855713, + 0.07675950229167938, + 0.06416306644678116, + 0.03020005114376545, + 0.12205670773983002, + 0.2186640352010727, + 0.08273717015981674, + 0.07365866005420685 + ], + "credit_map": [ + 0.03141716122627258, + 0.011960147880017757, + 0.027064353227615356, + 0.020046304911375046, + 0.014130349270999432, + 0.027187669649720192, + 0.024861512705683708, + 0.021123899146914482, + 0.010496252216398716, + 0.005859046708792448, + 0.02002415619790554, + 0.010166486725211143, + 0.00598204554989934, + 0.015371482819318771, + 0.010433072224259377, + 0.013292699120938778, + 0.008941270411014557, + 0.004757335875183344, + 0.008098199963569641, + 0.009202091954648495, + 0.008072893135249615, + 0.007998812012374401, + 0.0077202944085001945, + 0.007553284987807274, + 0.008516527712345123, + 0.005960120353847742, + 0.01069544069468975, + 0.01294692698866129, + 0.006316617596894503, + 0.011908413842320442, + 0.011849205940961838, + 0.008566247299313545, + 0.010297013446688652, + 0.004626148380339146, + 0.010711874812841415, + 0.007542039733380079, + 0.004286864306777716, + 0.006385973654687405, + 0.006340974010527134, + 0.013972166925668716, + 0.0030378729570657015, + 0.002453917171806097, + 0.004393093753606081, + 0.004736432805657387, + 0.002482469193637371, + 0.0033025373704731464, + 0.004965742584317923, + 0.0048279850743710995, + 0.016331356018781662, + 0.009917672723531723, + 0.018515650182962418, + 0.021072953939437866, + 0.0095876045525074, + 0.012593189254403114, + 0.013559921644628048, + 0.02047833614051342, + 0.02012612298130989, + 0.01953084021806717, + 0.050120290368795395, + 0.020563045516610146, + 0.018995804712176323, + 0.02773107960820198, + 0.020180614665150642, + 0.0414162203669548, + 0.006391323171555996, + 0.011271673254668713, + 0.011943461373448372, + 0.012206095270812511, + 0.012152432464063168, + 0.008349050767719746, + 0.01165782567113638, + 0.008765301667153835, + 0.010792073793709278, + 0.010569019243121147, + 0.00720501272007823, + 0.008631959557533264, + 0.009778108447790146, + 0.008993305265903473, + 0.009256994351744652, + 0.008432180620729923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0004368066438473761 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033002734184265 + ], + "opposite_logprob_tokens": [ + -1.910330057144165 + ], + "logprob_reward": -0.16033002734184265, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033002734184265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09789804369211197, + 0.1441836655139923, + 0.031883321702480316, + 0.1195966973900795, + 0.01271742582321167, + 0.016753416508436203, + 0.04753808304667473, + 0.48080748319625854, + 0.020948169752955437, + 0.027673736214637756 + ], + "credit_map": [ + 0.022822173312306404, + 0.00479290122166276, + 0.01395150925964117, + 0.007737493608146906, + 0.0053675915114581585, + 0.01736660860478878, + 0.007149997632950544, + 0.018709776923060417, + 0.032243311405181885, + 0.005379972513765097, + 0.027164893224835396, + 0.00917870830744505, + 0.004364291206002235, + 0.028756191954016685, + 0.00911673903465271, + 0.02797955647110939, + 0.004785079043358564, + 0.002674967050552368, + 0.005596499890089035, + 0.0038661332800984383, + 0.0023825489915907383, + 0.004443598911166191, + 0.0033996980637311935, + 0.004734797868877649, + 0.014036067761480808, + 0.0067704347893595695, + 0.02133258245885372, + 0.016816822811961174, + 0.006753855384886265, + 0.026093043386936188, + 0.015086173079907894, + 0.012707719579339027, + 0.002368863672018051, + 0.0010022827191278338, + 0.001924646319821477, + 0.0013380831805989146, + 0.001188207184895873, + 0.0011650638189166784, + 0.0011715770233422518, + 0.002558701438829303, + 0.0032928704749792814, + 0.0012606345117092133, + 0.0025411343667656183, + 0.0017420785734429955, + 0.001384862931445241, + 0.0015431843930855393, + 0.002404679311439395, + 0.002583971945568919, + 0.0071839760057628155, + 0.00431100046262145, + 0.0077216788195073605, + 0.0046218764036893845, + 0.00445974338799715, + 0.005060227587819099, + 0.005449657794088125, + 0.00872992817312479, + 0.04889452084898949, + 0.04525964334607124, + 0.09546814858913422, + 0.03918817266821861, + 0.04320240020751953, + 0.06891081482172012, + 0.04705004021525383, + 0.09283369779586792, + 0.001843396807089448, + 0.003679088084027171, + 0.002718436997383833, + 0.0021040260326117277, + 0.0020130116026848555, + 0.0023349688854068518, + 0.0021678078919649124, + 0.004087433218955994, + 0.004848443903028965, + 0.002917952137067914, + 0.004167439416050911, + 0.0025021012406796217, + 0.0023331507109105587, + 0.003395377192646265, + 0.0024999682791531086, + 0.0050093038007617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.16033002734184265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, statically captured in a 16:9 aspect ratio, captures the digital miracle of life, steeped in a refined, nostalgic gaming aesthetic. Against a solid background of dark blue and purple pixels, a compact, spherical bud, its surface composed of crimson and magenta pixels, is the sole focal point. Its surface, composed of blocks of crimson and magenta pixels, is subtly dithered to create a subtle three-dimensional effect. With a silent start, the outermost layer of pixelated petals at the bud's tip cautiously pokes open, revealing the more vibrant layers of color within. Then, with a fluid, yet distinctly digital rhythm, the entire flower visibly unfolds, layer upon layer of pixelated petals unfolding one by one, each with a distinct pixelated outline, as if meticulously painted by a skilled pixel artist. A soft, analog overhead light illuminates this slow blossoming process, casting simple, yet defined pixel shadows across the petals. Eventually, the flower fully opens into a perfect, symmetrical blossom with a vintage digital aesthetic, its limited yet saturated color palette contrasting beautifully with the deep background, creating a high-definition pixel art atmosphere that is both serene and captivating.", + "seed": 7991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_2646_seed_7991_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04278882471844554, + "vlm_logprob": -0.04278882471844554, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.04278882471844554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03652314096689224 + ], + "opposite_logprob_tokens": [ + -3.5365231037139893 + ], + "logprob_reward": -0.03652314096689224, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03652314096689224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05345288664102554, + 0.01577259786427021, + 0.08433432877063751, + 0.041033949702978134, + 0.06884213536977768, + 0.03865111246705055, + 0.18034374713897705, + 0.07355716824531555, + 0.4295971095561981, + 0.014414911158382893 + ], + "credit_map": [ + 0.005755096673965454, + 0.0049002887681126595, + 0.011538324877619743, + 0.004849230404943228, + 0.004466014448553324, + 0.010595153085887432, + 0.0057434565387666225, + 0.005605330690741539, + 0.0015299390070140362, + 0.00130039406940341, + 0.0017886798596009612, + 0.0010749312350526452, + 0.0014952970668673515, + 0.0032580741681158543, + 0.0016197485383599997, + 0.003705536713823676, + 0.006675437092781067, + 0.0025687396991997957, + 0.03400305658578873, + 0.002481278032064438, + 0.004465376492589712, + 0.018448658287525177, + 0.002089639427140355, + 0.013602150604128838, + 0.002814146690070629, + 0.005391122307628393, + 0.006790380924940109, + 0.002797072986140847, + 0.004803248215466738, + 0.009222693741321564, + 0.005213180556893349, + 0.004002111963927746, + 0.005359996110200882, + 0.008820772171020508, + 0.016657471656799316, + 0.0038378301542252302, + 0.007600870449095964, + 0.005066871643066406, + 0.005446369759738445, + 0.016051962971687317, + 0.004871245939284563, + 0.0032961037941277027, + 0.007314824033528566, + 0.0024176998995244503, + 0.005119326990097761, + 0.004661288112401962, + 0.003519713645800948, + 0.0074509140104055405, + 0.026930006220936775, + 0.01793968491256237, + 0.03125324472784996, + 0.013251152820885181, + 0.017623940482735634, + 0.02275933139026165, + 0.01877327263355255, + 0.0318131223320961, + 0.005570323672145605, + 0.006360838655382395, + 0.011956198140978813, + 0.011530806310474873, + 0.006247405428439379, + 0.004791379440575838, + 0.01638292334973812, + 0.0107173016294837, + 0.03821304813027382, + 0.09132219105958939, + 0.039936427026987076, + 0.051627155393362045, + 0.06206170842051506, + 0.021541474387049675, + 0.06974254548549652, + 0.05515262112021446, + 0.002954212948679924, + 0.0012922004098072648, + 0.002373823896050453, + 0.0016491069691255689, + 0.0011967407772317529, + 0.0012372996425256133, + 0.0015149422688409686, + 0.0021965852938592434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.03652314096689224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078690066933632 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078690066933632, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078690066933632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09336365014314651, + 0.11625470221042633, + 0.12744471430778503, + 0.0502249151468277, + 0.16388651728630066, + 0.1795750856399536, + 0.04779362305998802, + 0.07808735221624374, + 0.06121259182691574, + 0.08215688169002533 + ], + "credit_map": [ + 0.011005120351910591, + 0.005063212011009455, + 0.010540691204369068, + 0.005309556610882282, + 0.00586860254406929, + 0.030306722968816757, + 0.006545226089656353, + 0.01872452162206173, + 0.005952364299446344, + 0.0036864280700683594, + 0.04193034768104553, + 0.0032891244627535343, + 0.004337636288255453, + 0.04538923129439354, + 0.003624259727075696, + 0.008045305497944355, + 0.009326174855232239, + 0.0024519129656255245, + 0.05329873785376549, + 0.0036434633657336235, + 0.0037213321775197983, + 0.033603500574827194, + 0.0028233625926077366, + 0.018576228991150856, + 0.003396490355953574, + 0.004869761876761913, + 0.01311781071126461, + 0.0024946474004536867, + 0.005050641018897295, + 0.014251716434955597, + 0.0034100953489542007, + 0.0036337526980787516, + 0.01364804431796074, + 0.012982973828911781, + 0.05154109373688698, + 0.00628836452960968, + 0.011905060149729252, + 0.012527476996183395, + 0.01068145502358675, + 0.04431205615401268, + 0.01111103966832161, + 0.006580617278814316, + 0.061332475394010544, + 0.006570992525666952, + 0.005921874660998583, + 0.012628719210624695, + 0.010557758621871471, + 0.06487160176038742, + 0.004710915964096785, + 0.0060358294285833836, + 0.007917387410998344, + 0.0048330966383218765, + 0.005236419383436441, + 0.006927825510501862, + 0.005728966556489468, + 0.006403180304914713, + 0.009412279352545738, + 0.007445197552442551, + 0.01273830235004425, + 0.009155618958175182, + 0.006291903555393219, + 0.006015024613589048, + 0.01409770455211401, + 0.01293132733553648, + 0.004590705502778292, + 0.008764245547354221, + 0.008951453492045403, + 0.007815783843398094, + 0.005759899970144033, + 0.0036135006230324507, + 0.013810262084007263, + 0.007906738668680191, + 0.01802051067352295, + 0.007127235643565655, + 0.01400356087833643, + 0.0077413534745574, + 0.007934101857244968, + 0.007834218442440033, + 0.006743260193616152, + 0.012752631679177284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.016078690066933632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004403813509270549 + ], + "opposite_logprob_tokens": [ + -7.75044059753418 + ], + "logprob_reward": -0.0004403813509270549, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004403813509270549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17837606370449066, + 0.04841068759560585, + 0.08581720292568207, + 0.09662129729986191, + 0.09177851676940918, + 0.09892310202121735, + 0.11426039040088654, + 0.224197119474411, + 0.02408752776682377, + 0.037528034299612045 + ], + "credit_map": [ + 0.01350001897662878, + 0.010616910643875599, + 0.061369262635707855, + 0.009681293740868568, + 0.010607797652482986, + 0.02906200848519802, + 0.011675122193992138, + 0.031863659620285034, + 0.00509260781109333, + 0.007049719803035259, + 0.007298972457647324, + 0.005133324768394232, + 0.007089926861226559, + 0.004169188439846039, + 0.0063915434293448925, + 0.006185404025018215, + 0.0066346293315291405, + 0.00982649251818657, + 0.019404327496886253, + 0.007851026952266693, + 0.010939182713627815, + 0.012035800144076347, + 0.010493513196706772, + 0.008632231503725052, + 0.005962064955383539, + 0.011686227284371853, + 0.020418690517544746, + 0.008717796765267849, + 0.011028769426047802, + 0.01864955760538578, + 0.013472934253513813, + 0.006685259751975536, + 0.007698338944464922, + 0.006050881929695606, + 0.026325305923819542, + 0.006739483214914799, + 0.006649910006672144, + 0.006361580453813076, + 0.008549245074391365, + 0.023403767496347427, + 0.008888685144484043, + 0.008515581488609314, + 0.020722759887576103, + 0.007438176777213812, + 0.00849006325006485, + 0.013507762923836708, + 0.0116471778601408, + 0.019712897017598152, + 0.019377846270799637, + 0.011101443320512772, + 0.02227497100830078, + 0.0074641890823841095, + 0.011309253983199596, + 0.011652691289782524, + 0.010297749191522598, + 0.020782243460416794, + 0.024118369445204735, + 0.019429445266723633, + 0.03959357365965843, + 0.030403491109609604, + 0.018471267074346542, + 0.01423339918255806, + 0.042511239647865295, + 0.03543633595108986, + 0.002537317341193557, + 0.003326686564832926, + 0.003726431168615818, + 0.0036344327963888645, + 0.0030727041885256767, + 0.002409733133390546, + 0.0025361336302012205, + 0.0028440887108445168, + 0.005609425250440836, + 0.004315042868256569, + 0.004868683870881796, + 0.004812181927263737, + 0.005214508157223463, + 0.003374736523255706, + 0.004391005728393793, + 0.004942451138049364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0004403813509270549 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005600076401606202 + ], + "opposite_logprob_tokens": [ + -7.5005598068237305 + ], + "logprob_reward": -0.0005600076401606202, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005600076401606202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1337907910346985, + 0.0464034341275692, + 0.05574076250195503, + 0.10614140331745148, + 0.0982164666056633, + 0.2019091248512268, + 0.1034470945596695, + 0.06687981635332108, + 0.1590450257062912, + 0.028426114469766617 + ], + "credit_map": [ + 0.01303910929709673, + 0.008299381472170353, + 0.03997200354933739, + 0.008901374414563179, + 0.006687848828732967, + 0.03475057706236839, + 0.010518294759094715, + 0.011622192338109016, + 0.003226148197427392, + 0.0032692153472453356, + 0.012182371690869331, + 0.0033314847387373447, + 0.006319844163954258, + 0.00772396894171834, + 0.0043866285122931, + 0.0059637753292918205, + 0.004911032505333424, + 0.004679449833929539, + 0.013312621042132378, + 0.004734786693006754, + 0.006684428546577692, + 0.008736798539757729, + 0.005724018439650536, + 0.006957626901566982, + 0.007296168711036444, + 0.013641996309161186, + 0.018112514168024063, + 0.010798703879117966, + 0.012302083894610405, + 0.018262868747115135, + 0.01701538823544979, + 0.0087116789072752, + 0.008655082434415817, + 0.009857451543211937, + 0.024135151877999306, + 0.007941857911646366, + 0.007229919079691172, + 0.00894575659185648, + 0.010271123610436916, + 0.021180124953389168, + 0.018029039725661278, + 0.013428840786218643, + 0.05084485188126564, + 0.012166558764874935, + 0.014463216997683048, + 0.023530954495072365, + 0.020321983844041824, + 0.04912368953227997, + 0.015570877119898796, + 0.010703885927796364, + 0.016421886160969734, + 0.009852565824985504, + 0.00783558189868927, + 0.015091463923454285, + 0.012772354297339916, + 0.015198477543890476, + 0.005425686482340097, + 0.007649398874491453, + 0.009551945142447948, + 0.009301909245550632, + 0.005997774191200733, + 0.005203970707952976, + 0.01326338853687048, + 0.010485740378499031, + 0.0126750273630023, + 0.02244018204510212, + 0.02373243123292923, + 0.021051861345767975, + 0.014998062513768673, + 0.008766078390181065, + 0.0343676321208477, + 0.021013755351305008, + 0.003157824743539095, + 0.0035908997524529696, + 0.0035710898227989674, + 0.003177389269694686, + 0.005371463485062122, + 0.0028988018166273832, + 0.003491866635158658, + 0.0031667789444327354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0005600076401606202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034190356731415 + ], + "opposite_logprob_tokens": [ + -1.9103418588638306 + ], + "logprob_reward": -0.16034190356731415, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034190356731415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04030296206474304, + 0.032411277294158936, + 0.02309073880314827, + 0.03002101741731167, + 0.09896798431873322, + 0.08261425793170929, + 0.18439596891403198, + 0.21327923238277435, + 0.24500583112239838, + 0.04991068318486214 + ], + "credit_map": [ + 0.006323341280221939, + 0.0030586186330765486, + 0.00614031171426177, + 0.003181266598403454, + 0.005470144562423229, + 0.005183496978133917, + 0.0053698038682341576, + 0.005575982853770256, + 0.002512754173949361, + 0.0021245297975838184, + 0.008687775582075119, + 0.0020969246979802847, + 0.002827243646606803, + 0.002919947262853384, + 0.002900198334828019, + 0.008341904729604721, + 0.002723642624914646, + 0.0027458558324724436, + 0.004780680872499943, + 0.0022077348548918962, + 0.0022688452154397964, + 0.0022472466807812452, + 0.0025759339332580566, + 0.0035407994873821735, + 0.0027978133875876665, + 0.004105645697563887, + 0.005522072780877352, + 0.0034017611760646105, + 0.004069065675139427, + 0.002789357677102089, + 0.0042641544714570045, + 0.0030711486469954252, + 0.006727817002683878, + 0.008062846027314663, + 0.031273894011974335, + 0.004240053240209818, + 0.008131640031933784, + 0.006389070767909288, + 0.004884819965809584, + 0.029257848858833313, + 0.009044202044606209, + 0.006588099990040064, + 0.01733330450952053, + 0.006293099839240313, + 0.00574695086106658, + 0.011537974700331688, + 0.0115024633705616, + 0.014568169601261616, + 0.03099805861711502, + 0.018030401319265366, + 0.03281096741557121, + 0.014817669987678528, + 0.016621867194771767, + 0.0221962071955204, + 0.016042623668909073, + 0.03287817910313606, + 0.016659576445817947, + 0.02011534571647644, + 0.031685970723629, + 0.03230006620287895, + 0.019945302978157997, + 0.012312591075897217, + 0.04837971553206444, + 0.03188066929578781, + 0.017675193026661873, + 0.04389173164963722, + 0.03098856285214424, + 0.029557514935731888, + 0.028365135192871094, + 0.014245842583477497, + 0.04979454725980759, + 0.030487319454550743, + 0.007036466617137194, + 0.005474899895489216, + 0.008505672216415405, + 0.007093826308846474, + 0.005919582210481167, + 0.0044263494201004505, + 0.005475328303873539, + 0.00597856380045414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.16034190356731415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, statically captured in a 16:9 aspect ratio, captures the digital miracle of life, steeped in a refined, nostalgic gaming aesthetic. Against a solid background of dark blue and purple pixels, a compact, spherical bud, its surface composed of crimson and magenta pixels, is the sole focal point. Its surface, composed of blocks of crimson and magenta pixels, is subtly dithered to create a subtle three-dimensional effect. With a silent start, the outermost layer of pixelated petals at the bud's tip cautiously pokes open, revealing the more vibrant layers of color within. Then, with a fluid, yet distinctly digital rhythm, the entire flower visibly unfolds, layer upon layer of pixelated petals unfolding one by one, each with a distinct pixelated outline, as if meticulously painted by a skilled pixel artist. A soft, analog overhead light illuminates this slow blossoming process, casting simple, yet defined pixel shadows across the petals. Eventually, the flower fully opens into a perfect, symmetrical blossom with a vintage digital aesthetic, its limited yet saturated color palette contrasting beautifully with the deep background, creating a high-definition pixel art atmosphere that is both serene and captivating.", + "seed": 7992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_10/index_2646_seed_7992_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13512182685662993, + "vlm_logprob": -0.13512182685662993, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.13512182685662993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09254354238510132 + ], + "opposite_logprob_tokens": [ + -2.467543601989746 + ], + "logprob_reward": -0.09254354238510132, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09254354238510132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2564694881439209, + 0.05239759758114815, + 0.02678954415023327, + 0.05882272496819496, + 0.08104917407035828, + 0.15722912549972534, + 0.23879985511302948, + 0.037607479840517044, + 0.07973460108041763, + 0.011100377887487411 + ], + "credit_map": [ + 0.01868632808327675, + 0.010762733407318592, + 0.0529123954474926, + 0.01889386959373951, + 0.008688251487910748, + 0.09202436357736588, + 0.014289787970483303, + 0.0402117557823658, + 0.007113961037248373, + 0.004782133270055056, + 0.007712071295827627, + 0.008694492280483246, + 0.004199571907520294, + 0.0066141714341938496, + 0.007128411438316107, + 0.006152786314487457, + 0.004966399632394314, + 0.0024934192188084126, + 0.004403235856443644, + 0.002422572113573551, + 0.0024964245967566967, + 0.004000754561275244, + 0.0024418532848358154, + 0.0035648862831294537, + 0.012821832671761513, + 0.002899513114243746, + 0.013798202387988567, + 0.0030014202930033207, + 0.0030256356112658978, + 0.00848185271024704, + 0.003325066063553095, + 0.011469201184809208, + 0.014551264233887196, + 0.006533142179250717, + 0.012531666085124016, + 0.006354546640068293, + 0.006131977774202824, + 0.01034974679350853, + 0.006556484382599592, + 0.018040349707007408, + 0.0515001080930233, + 0.011381793767213821, + 0.02680613286793232, + 0.009226638823747635, + 0.007629175204783678, + 0.020697545260190964, + 0.009739157743752003, + 0.020248569548130035, + 0.062340039759874344, + 0.01827658899128437, + 0.03893217444419861, + 0.011898239143192768, + 0.008235465735197067, + 0.05804235115647316, + 0.00989853497594595, + 0.031176451593637466, + 0.010490396991372108, + 0.0033968230709433556, + 0.0059295170940458775, + 0.0025488357059657574, + 0.0017926645232364535, + 0.0069402772933244705, + 0.002033054595813155, + 0.004475908353924751, + 0.005322489887475967, + 0.0020757950842380524, + 0.027322573587298393, + 0.0027867620810866356, + 0.00223686289973557, + 0.032215554267168045, + 0.0015099901938810945, + 0.006264576222747564, + 0.0013049730332568288, + 0.0012917257845401764, + 0.001874631503596902, + 0.0011909076711162925, + 0.0012547102523967624, + 0.0015549828531220555, + 0.0010460582561790943, + 0.0015823880676180124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.09254354238510132 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956872832030058 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956872832030058, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956872832030058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2973499000072479, + 0.0582856684923172, + 0.11713800579309464, + 0.19365406036376953, + 0.038951221853494644, + 0.06309093534946442, + 0.05245513841509819, + 0.06961953639984131, + 0.08645230531692505, + 0.023003267124295235 + ], + "credit_map": [ + 0.018400777131319046, + 0.016743170097470284, + 0.05889112874865532, + 0.02027216926217079, + 0.01662350632250309, + 0.09894858300685883, + 0.021721480414271355, + 0.04574907571077347, + 0.011713541112840176, + 0.0045746201649308205, + 0.008017932996153831, + 0.008824565447866917, + 0.003880487522110343, + 0.007252025417983532, + 0.007961074821650982, + 0.006061416584998369, + 0.018542107194662094, + 0.00966646522283554, + 0.02713337168097496, + 0.010956096462905407, + 0.005848139990121126, + 0.019118530675768852, + 0.008558093570172787, + 0.01731518842279911, + 0.014318582601845264, + 0.0034868603106588125, + 0.07060840725898743, + 0.0051810299046337605, + 0.0043455385603010654, + 0.07561295479536057, + 0.004414608236402273, + 0.015686053782701492, + 0.007774713449180126, + 0.0031517245806753635, + 0.0058553582057356834, + 0.003020233241841197, + 0.0027983200270682573, + 0.004568061325699091, + 0.0031974208541214466, + 0.008585387840867043, + 0.014526770450174809, + 0.005681758280843496, + 0.009157633408904076, + 0.005577775649726391, + 0.004606117494404316, + 0.007892024703323841, + 0.0061407326720654964, + 0.009508121758699417, + 0.012597039341926575, + 0.0036421765107661486, + 0.008769354782998562, + 0.0030523345340043306, + 0.0026315650902688503, + 0.011842888779938221, + 0.002599968807771802, + 0.007319805212318897, + 0.018358295783400536, + 0.004572156351059675, + 0.008976195007562637, + 0.0036121243610978127, + 0.003427499206736684, + 0.013112111017107964, + 0.004164664074778557, + 0.01339648850262165, + 0.011454061605036259, + 0.004154900088906288, + 0.03001948818564415, + 0.0028334613889455795, + 0.001672630081884563, + 0.021159719675779343, + 0.002127393614500761, + 0.013030649162828922, + 0.0031017472501844168, + 0.0025543896481394768, + 0.003391346661373973, + 0.0021059568971395493, + 0.001567330677062273, + 0.00268186186440289, + 0.0019349070498719811, + 0.005665725562721491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.005956872832030058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006350646144710481 + ], + "opposite_logprob_tokens": [ + -7.375635147094727 + ], + "logprob_reward": -0.0006350646144710481, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006350646144710481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2335757166147232, + 0.07596369087696075, + 0.027048960328102112, + 0.04470640420913696, + 0.020771050825715065, + 0.04648195207118988, + 0.047776367515325546, + 0.02800472266972065, + 0.46547427773475647, + 0.01019685622304678 + ], + "credit_map": [ + 0.014877993613481522, + 0.01848805509507656, + 0.034524306654930115, + 0.02256201207637787, + 0.016572246327996254, + 0.06798047572374344, + 0.018326804041862488, + 0.040243834257125854, + 0.008545318618416786, + 0.00684274872764945, + 0.013322305865585804, + 0.011034959927201271, + 0.006725414656102657, + 0.011992999352514744, + 0.009607280604541302, + 0.007892663590610027, + 0.004032993223518133, + 0.0027559271547943354, + 0.0046002515591681, + 0.002280833199620247, + 0.002602869179099798, + 0.004202083684504032, + 0.0027931397780776024, + 0.003780863480642438, + 0.0074381716549396515, + 0.0023180486168712378, + 0.01109965331852436, + 0.002657705219462514, + 0.0027921826113015413, + 0.008651587180793285, + 0.0026957555674016476, + 0.00705330166965723, + 0.002799014328047633, + 0.002177432645112276, + 0.0028451145626604557, + 0.0023502702824771404, + 0.0018255350878462195, + 0.0026849349960684776, + 0.002124110935255885, + 0.003964640200138092, + 0.01252046599984169, + 0.00373275950551033, + 0.006945929955691099, + 0.003308580256998539, + 0.0028683131095021963, + 0.006118063349276781, + 0.003492682706564665, + 0.0074951606802642345, + 0.010237238369882107, + 0.0033002514392137527, + 0.007728029973804951, + 0.003112231381237507, + 0.002368106273934245, + 0.011381944641470909, + 0.0025171220768243074, + 0.007131445221602917, + 0.008460772223770618, + 0.0026621087454259396, + 0.004033706150949001, + 0.0019231727346777916, + 0.0013998976210132241, + 0.005257863085716963, + 0.0015164857031777501, + 0.002750714309513569, + 0.034665901213884354, + 0.013651364482939243, + 0.1488438844680786, + 0.015739528462290764, + 0.011519319377839565, + 0.19727443158626556, + 0.007888486608862877, + 0.03589136525988579, + 0.0013781972229480743, + 0.001320457668043673, + 0.0014844776596873999, + 0.0012180160265415907, + 0.000807345611974597, + 0.0012261555530130863, + 0.0011289251269772649, + 0.001633281703107059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0006350646144710481 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00034195298212580383 + ], + "opposite_logprob_tokens": [ + -8.00034236907959 + ], + "logprob_reward": -0.00034195298212580383, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00034195298212580383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1082293838262558, + 0.07687105238437653, + 0.02873106487095356, + 0.07335107028484344, + 0.03219817206263542, + 0.06264828145503998, + 0.09650520980358124, + 0.03336988389492035, + 0.46551546454429626, + 0.022580459713935852 + ], + "credit_map": [ + 0.009302303194999695, + 0.0077354502864181995, + 0.012332123704254627, + 0.014321550726890564, + 0.00782463513314724, + 0.025846470147371292, + 0.011806093156337738, + 0.019060734659433365, + 0.010449036955833435, + 0.006805854849517345, + 0.012158454395830631, + 0.01239740289747715, + 0.006564239971339703, + 0.01014875341206789, + 0.009330454282462597, + 0.009016850963234901, + 0.004507073666900396, + 0.0026392920408397913, + 0.006079087965190411, + 0.002480069175362587, + 0.0016920696943998337, + 0.0040809353813529015, + 0.0024313924368470907, + 0.0048211426474153996, + 0.014857407659292221, + 0.0030533019453287125, + 0.016680745407938957, + 0.0031410441733896732, + 0.004338167607784271, + 0.013707409612834454, + 0.003965671639889479, + 0.013607313856482506, + 0.004792920313775539, + 0.002965414198115468, + 0.004926322493702173, + 0.0030728185083717108, + 0.0028048132080584764, + 0.004876359365880489, + 0.002830536337569356, + 0.005928986705839634, + 0.022493405267596245, + 0.004090338945388794, + 0.011345579288899899, + 0.0027647563256323338, + 0.002455510199069977, + 0.008216479793190956, + 0.002881525317206979, + 0.008400681428611279, + 0.020242439582943916, + 0.00659788167104125, + 0.016418127343058586, + 0.005925807170569897, + 0.004419885575771332, + 0.02352737821638584, + 0.005047758109867573, + 0.014325927942991257, + 0.008508999831974506, + 0.002754588844254613, + 0.004185694269835949, + 0.0022006258368492126, + 0.002356938784942031, + 0.005994858685880899, + 0.00217824918217957, + 0.005189923103898764, + 0.029563946649432182, + 0.010583294555544853, + 0.16923338174819946, + 0.01602858677506447, + 0.010658024810254574, + 0.1909727156162262, + 0.006962238810956478, + 0.031513262540102005, + 0.003131026169285178, + 0.002858779625967145, + 0.003367499215528369, + 0.002501363167539239, + 0.0025001023896038532, + 0.0026965709403157234, + 0.002002075547352433, + 0.0035230398643761873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.00034195298212580383 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video captured in a static close-up shot with a 16:9 aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a solid dark blue and purple pixelated surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the flower bud initially compact and spherical with a surface of crimson and magenta pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower unfold layer by layer from a closed bud into a fully open symmetrical blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft overhead light casting defined pixel shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761317014694214 + ], + "opposite_logprob_tokens": [ + -0.8261317014694214 + ], + "logprob_reward": -0.5761317014694214, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761317014694214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.257952481508255, + 0.05943889915943146, + 0.027597980573773384, + 0.026210863143205643, + 0.03289695084095001, + 0.05619316175580025, + 0.140522763133049, + 0.03326911851763725, + 0.32826560735702515, + 0.03765210509300232 + ], + "credit_map": [ + 0.01762801967561245, + 0.018529187887907028, + 0.05826311558485031, + 0.019965404644608498, + 0.015590967610478401, + 0.07932986319065094, + 0.019996901974081993, + 0.02864902652800083, + 0.008716518059372902, + 0.0072134118527174, + 0.009292063303291798, + 0.007358023896813393, + 0.005773143842816353, + 0.008742830716073513, + 0.005825319327414036, + 0.006517583504319191, + 0.0054090372286736965, + 0.0026173528749495745, + 0.00394902378320694, + 0.0023805340752005577, + 0.003493798431009054, + 0.0030004107393324375, + 0.0026290062814950943, + 0.00411881972104311, + 0.004767317324876785, + 0.0015872030053287745, + 0.005380496382713318, + 0.0021098237484693527, + 0.00200882856734097, + 0.004111989866942167, + 0.002837821841239929, + 0.003407382173463702, + 0.005787676200270653, + 0.00321741821244359, + 0.005882176570594311, + 0.002885023131966591, + 0.002683437429368496, + 0.0037702075205743313, + 0.002831367077305913, + 0.005839643534272909, + 0.011265111155807972, + 0.004772704094648361, + 0.00821471493691206, + 0.005187098868191242, + 0.00409608194604516, + 0.006354562938213348, + 0.006597664672881365, + 0.009705223143100739, + 0.04091605171561241, + 0.010153411887586117, + 0.024062642827630043, + 0.007558033335953951, + 0.007116259075701237, + 0.027729103341698647, + 0.007435296196490526, + 0.015551955439150333, + 0.004829744342714548, + 0.0026220271829515696, + 0.005049570929259062, + 0.0027461370918899775, + 0.002512034261599183, + 0.006838612258434296, + 0.003381952876225114, + 0.005289038177579641, + 0.020850716158747673, + 0.007172542158514261, + 0.11512314528226852, + 0.011338774114847183, + 0.007880264893174171, + 0.13684676587581635, + 0.005055363290011883, + 0.02399802766740322, + 0.004896033089607954, + 0.002848090371116996, + 0.0064062802121043205, + 0.0027928699273616076, + 0.0025419301819056273, + 0.0043122051283717155, + 0.002495335880666971, + 0.011359358206391335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.5761317014694214 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_11/index_tensor([2874, 2874])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_11/index_tensor([2874, 2874])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb68efff136022733bc56088e9156e49f9dbdcec --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_11/index_tensor([2874, 2874])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This extreme close-up, rendered in a hyperrealistic style, begins with the camera holding static over a glistening, translucent pool of melted wax on a thick, cream-colored candle. The liquid wax, emitting a warm golden glow from the invisible flame above, gently overflows its cratered edges. A small amount of wax is meticulously gathered into a perfect, viscous bead, its surface tension momentarily holding it together before the camera begins a slow, deliberate downward tilt, following the irresistible pull of gravity. The droplet begins to slide slowly along the textured, uneven side of the candle, leaving a shimmering, translucent trail behind the previously solidified droplet. As it nears the bottom edge, the droplet's own weight causes it to elongate, stretching into a teardrop shape, its neck tapering. The warm candlelight beautifully refracts through its increasingly elongated form. In the final moments, the connection breaks cleanly, and the droplet separates, plummeting a short distance to a dark, polished wooden tabletop below. Its rich texture is visible in the soft, flickering ambient light, creating an atmosphere of tranquility and concentration captured with intense macroscopic detail and physically accurate fluid dynamics.", + "seed": 6240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_11/index_2874_seed_6240_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05221449546515942, + "vlm_logprob": -0.05221449546515942, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.05221449546515942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07967866957187653 + ], + "opposite_logprob_tokens": [ + -2.704678773880005 + ], + "logprob_reward": -0.07967866957187653, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07967866957187653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04536210373044014, + 0.030017109587788582, + 0.020328009501099586, + 0.06053715571761131, + 0.02961680479347706, + 0.01326841115951538, + 0.01588173769414425, + 0.6439427733421326, + 0.01773591712117195, + 0.12331005185842514 + ], + "credit_map": [ + 0.003995460458099842, + 0.003416541963815689, + 0.005175392143428326, + 0.006691079121083021, + 0.006313540041446686, + 0.00393762206658721, + 0.008599868044257164, + 0.007232599426060915, + 0.003186205169185996, + 0.0030686974059790373, + 0.0034097477328032255, + 0.005959218833595514, + 0.0034425980411469936, + 0.0033008630853146315, + 0.004088978283107281, + 0.0035608012694865465, + 0.001745507470332086, + 0.0018401996931061149, + 0.0021640053018927574, + 0.0038086127024143934, + 0.004076446872204542, + 0.0015761656686663628, + 0.003023646306246519, + 0.0020934257190674543, + 0.003020000644028187, + 0.0056764776818454266, + 0.0037247745785862207, + 0.012673589400947094, + 0.015197237022221088, + 0.0037629413418471813, + 0.010650823824107647, + 0.00583131005987525, + 0.00230612070299685, + 0.003132084384560585, + 0.0031782190781086683, + 0.004383388441056013, + 0.004934763535857201, + 0.002514312043786049, + 0.005359209142625332, + 0.003808707697317004, + 0.0007910537533462048, + 0.001329658436588943, + 0.0011588913621380925, + 0.0024430803023278713, + 0.0024453052319586277, + 0.0009867260232567787, + 0.0027426877059042454, + 0.0013710085768252611, + 0.0011420748196542263, + 0.0014263776829466224, + 0.0015913568204268813, + 0.001996483886614442, + 0.0033821952529251575, + 0.0017656830605119467, + 0.0028576499316841364, + 0.0017199155408889055, + 0.020745810121297836, + 0.051728662103414536, + 0.03288691118359566, + 0.28309229016304016, + 0.024231886491179466, + 0.0190009493380785, + 0.08935604989528656, + 0.12290016561746597, + 0.0009054761612787843, + 0.0027634098660200834, + 0.0011241082102060318, + 0.0021167094819247723, + 0.0041657667607069016, + 0.003024949925020337, + 0.0022854311391711235, + 0.00135006511118263, + 0.007305804640054703, + 0.01203371211886406, + 0.015279150567948818, + 0.012406693771481514, + 0.021299147978425026, + 0.011196848005056381, + 0.014301949180662632, + 0.02948673442006111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.07967866957187653 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907894045114517 + ], + "opposite_logprob_tokens": [ + -2.464078903198242 + ], + "logprob_reward": -0.08907894045114517, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907894045114517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036084070801734924, + 0.03552056476473808, + 0.03233885392546654, + 0.1394258290529251, + 0.030531080439686775, + 0.017279349267482758, + 0.014601041562855244, + 0.4706077575683594, + 0.06399496644735336, + 0.15961648523807526 + ], + "credit_map": [ + 0.0027522919699549675, + 0.002693075919523835, + 0.0039563653990626335, + 0.006613112520426512, + 0.0066018239594995975, + 0.0035689566284418106, + 0.005414205137640238, + 0.004484234377741814, + 0.002767156809568405, + 0.0028585223481059074, + 0.0035593195352703333, + 0.0055411760695278645, + 0.0052998000755906105, + 0.0031884287018328905, + 0.007341352291405201, + 0.004964805208146572, + 0.001765527529641986, + 0.003461041720584035, + 0.0023117943201214075, + 0.004792974330484867, + 0.009354496374726295, + 0.0024271055590361357, + 0.004652965813875198, + 0.0035729464143514633, + 0.0071030561812222, + 0.013861455954611301, + 0.009628348052501678, + 0.026352841407060623, + 0.03696494176983833, + 0.009579063393175602, + 0.023068634793162346, + 0.012867464683949947, + 0.0023389861453324556, + 0.002963610691949725, + 0.00296955369412899, + 0.005455676931887865, + 0.004833029117435217, + 0.0021749294828623533, + 0.00542904669418931, + 0.004366244189441204, + 0.0014220501761883497, + 0.0014310465194284916, + 0.001480813603848219, + 0.003614466404542327, + 0.002310000127181411, + 0.0013939372729510069, + 0.003524200990796089, + 0.002102832542732358, + 0.0009155953302979469, + 0.0014962410787120461, + 0.001125568407587707, + 0.0016548956045880914, + 0.0038314973935484886, + 0.0013033533468842506, + 0.002579306485131383, + 0.001694582519121468, + 0.015921272337436676, + 0.03798677772283554, + 0.024209948256611824, + 0.20519199967384338, + 0.018974989652633667, + 0.013801150023937225, + 0.06436499953269958, + 0.09015657752752304, + 0.004246032331138849, + 0.011727931909263134, + 0.0033400841057300568, + 0.005861528217792511, + 0.017539191991090775, + 0.007214544806629419, + 0.009305429644882679, + 0.004760217387229204, + 0.008640527725219727, + 0.01283859834074974, + 0.02971663884818554, + 0.016957253217697144, + 0.01985621266067028, + 0.012015162035822868, + 0.018165482208132744, + 0.041426606476306915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.08907894045114517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810510411858559 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.03810510411858559, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810510411858559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05243495851755142, + 0.02697535790503025, + 0.018195349723100662, + 0.07977086305618286, + 0.06268768757581711, + 0.04935803264379501, + 0.03779756650328636, + 0.19890311360359192, + 0.10477681457996368, + 0.36910027265548706 + ], + "credit_map": [ + 0.0025674812495708466, + 0.005563076585531235, + 0.0037372044753283262, + 0.007653798442333937, + 0.014369550161063671, + 0.0025400903541594744, + 0.010291210375726223, + 0.005712540820240974, + 0.0025045110378414392, + 0.0027103929314762354, + 0.0035191650968044996, + 0.004985418636351824, + 0.004257436376065016, + 0.0019759046845138073, + 0.004391026217490435, + 0.002631501527503133, + 0.0015432442305609584, + 0.0023698031436651945, + 0.0016833905829116702, + 0.0032879759091883898, + 0.0030915201641619205, + 0.0013538122875615954, + 0.0029239272698760033, + 0.0019416760187596083, + 0.004460720345377922, + 0.007082626223564148, + 0.005831185728311539, + 0.014599604532122612, + 0.019284820184111595, + 0.006288554519414902, + 0.013910996727645397, + 0.008312342688441277, + 0.006179494317620993, + 0.007199067156761885, + 0.007253686897456646, + 0.012128157541155815, + 0.007964610122144222, + 0.00451783137395978, + 0.008095122873783112, + 0.009349710308015347, + 0.004040050320327282, + 0.0040178862400352955, + 0.004425823222845793, + 0.009146607480943203, + 0.008425089530646801, + 0.003353324718773365, + 0.010562104173004627, + 0.005387145094573498, + 0.0030623034108430147, + 0.0043152570724487305, + 0.0038312776014208794, + 0.004503135569393635, + 0.0060106501914560795, + 0.00291218189522624, + 0.007620032876729965, + 0.005542729515582323, + 0.007396025583148003, + 0.014992371201515198, + 0.012597719207406044, + 0.08573164790868759, + 0.009411776438355446, + 0.006504528224468231, + 0.029045242816209793, + 0.03322378918528557, + 0.011555898934602737, + 0.017578834667801857, + 0.011854988522827625, + 0.012433192692697048, + 0.011877521872520447, + 0.021803906187415123, + 0.010026556439697742, + 0.007645908743143082, + 0.02267495170235634, + 0.050683602690696716, + 0.04547376558184624, + 0.055256422609090805, + 0.07014234364032745, + 0.03601415455341339, + 0.05063282325863838, + 0.03822217881679535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.03810510411858559 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080332919955254 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080332919955254, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080332919955254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04117005690932274, + 0.018292928114533424, + 0.03426880016922951, + 0.07008903473615646, + 0.041618261486291885, + 0.024961179122328758, + 0.01632990874350071, + 0.24854299426078796, + 0.10872406512498856, + 0.3960028290748596 + ], + "credit_map": [ + 0.003292190609499812, + 0.0031336413230746984, + 0.00503925746306777, + 0.005853322334587574, + 0.007298959419131279, + 0.003566733095794916, + 0.007001162972301245, + 0.005984791088849306, + 0.0014896985376253724, + 0.002379482612013817, + 0.002036281395703554, + 0.003080303082242608, + 0.002240261062979698, + 0.001776355435140431, + 0.003406279720366001, + 0.0018842659192159772, + 0.00209396006539464, + 0.0027497494593262672, + 0.0026747137308120728, + 0.009994094260036945, + 0.003984218463301659, + 0.0023380883503705263, + 0.005584565456956625, + 0.004849412012845278, + 0.004017689265310764, + 0.006399023812264204, + 0.004505145829170942, + 0.01661485806107521, + 0.01363680511713028, + 0.004318395163863897, + 0.01216464675962925, + 0.008432472124695778, + 0.0037337560206651688, + 0.004219376016408205, + 0.005487261340022087, + 0.005818682257086039, + 0.006397916004061699, + 0.003630757564678788, + 0.0067506274208426476, + 0.005579883232712746, + 0.0023046545684337616, + 0.0019666904117912054, + 0.0022681672126054764, + 0.006630976218730211, + 0.002054814714938402, + 0.001828150823712349, + 0.0032242361921817064, + 0.0046834866516292095, + 0.0015967946965247393, + 0.0015074905240908265, + 0.0023006098344922066, + 0.0018259690841659904, + 0.0022518765181303024, + 0.0016874767607077956, + 0.002661139238625765, + 0.002498551271855831, + 0.008577886037528515, + 0.019964398816227913, + 0.011849397793412209, + 0.10726140439510345, + 0.010734875686466694, + 0.008479451760649681, + 0.03531317412853241, + 0.04636240750551224, + 0.00639056321233511, + 0.02471839264035225, + 0.00838947482407093, + 0.010069076903164387, + 0.019113559275865555, + 0.018043624237179756, + 0.014305485412478447, + 0.00769389234483242, + 0.02944495901465416, + 0.04822862893342972, + 0.054962921887636185, + 0.06178883835673332, + 0.07870898395776749, + 0.036752499639987946, + 0.04577677696943283, + 0.040339190512895584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.016080332919955254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812943026423454 + ], + "opposite_logprob_tokens": [ + -3.2881293296813965 + ], + "logprob_reward": -0.03812943026423454, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812943026423454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04567936435341835, + 0.022861311212182045, + 0.01729767397046089, + 0.05576081946492195, + 0.011431495659053326, + 0.014049778692424297, + 0.01809035800397396, + 0.541096031665802, + 0.03363660350441933, + 0.2400965839624405 + ], + "credit_map": [ + 0.004039218183606863, + 0.00321732135489583, + 0.005981561727821827, + 0.006093691568821669, + 0.006130320020020008, + 0.005648954771459103, + 0.0078273955732584, + 0.006740899756550789, + 0.0016884615179151297, + 0.0018864604644477367, + 0.002341326791793108, + 0.003685331204906106, + 0.003702405607327819, + 0.002246937481686473, + 0.0042784069664776325, + 0.0030319818761199713, + 0.0012363027781248093, + 0.0020998395048081875, + 0.0013676723465323448, + 0.0030200418550521135, + 0.003705183509737253, + 0.0011773896403610706, + 0.0029293233528733253, + 0.0017619208665564656, + 0.003163954708725214, + 0.004853539168834686, + 0.00403723306953907, + 0.011181065812706947, + 0.012397263199090958, + 0.003952715080231428, + 0.010092153213918209, + 0.006082892417907715, + 0.0013171127066016197, + 0.0008776429458521307, + 0.0015726794954389334, + 0.00207234057597816, + 0.0009871715446934104, + 0.0009743132977746427, + 0.0015254357131198049, + 0.0021047997288405895, + 0.0009021465084515512, + 0.001267935149371624, + 0.0011598656419664621, + 0.0028703438583761454, + 0.002426442224532366, + 0.0012061621528118849, + 0.0025827011559158564, + 0.0016341820592060685, + 0.001335974084213376, + 0.00212481408379972, + 0.0016799987060949206, + 0.0021415753290057182, + 0.003465069457888603, + 0.0016135607147589326, + 0.003381521673873067, + 0.0023478458169847727, + 0.017466779798269272, + 0.0431959293782711, + 0.028420405462384224, + 0.23820741474628448, + 0.020356252789497375, + 0.015293671749532223, + 0.07520366460084915, + 0.10295195132493973, + 0.002707102568820119, + 0.005390713457018137, + 0.0018136940198019147, + 0.0028859744779765606, + 0.008791628293693066, + 0.004933987278491259, + 0.004357321187853813, + 0.00275618233717978, + 0.012948093004524708, + 0.02475011721253395, + 0.03828535974025726, + 0.02443597838282585, + 0.03848384693264961, + 0.020583875477313995, + 0.027032572776079178, + 0.05357673764228821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.03812943026423454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extreme close-up, rendered in a hyperrealistic style, begins with the camera holding static over a glistening, translucent pool of melted wax on a thick, cream-colored candle. The liquid wax, emitting a warm golden glow from the invisible flame above, gently overflows its cratered edges. A small amount of wax is meticulously gathered into a perfect, viscous bead, its surface tension momentarily holding it together before the camera begins a slow, deliberate downward tilt, following the irresistible pull of gravity. The droplet begins to slide slowly along the textured, uneven side of the candle, leaving a shimmering, translucent trail behind the previously solidified droplet. As it nears the bottom edge, the droplet's own weight causes it to elongate, stretching into a teardrop shape, its neck tapering. The warm candlelight beautifully refracts through its increasingly elongated form. In the final moments, the connection breaks cleanly, and the droplet separates, plummeting a short distance to a dark, polished wooden tabletop below. Its rich texture is visible in the soft, flickering ambient light, creating an atmosphere of tranquility and concentration captured with intense macroscopic detail and physically accurate fluid dynamics.", + "seed": 6241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_11/index_2874_seed_6241_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031062374962493776, + "vlm_logprob": -0.031062374962493776, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.031062374962493776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02073037438094616 + ], + "opposite_logprob_tokens": [ + -4.145730495452881 + ], + "logprob_reward": -0.02073037438094616, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02073037438094616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14024481177330017, + 0.1666565239429474, + 0.04450632631778717, + 0.03705764561891556, + 0.06370184570550919, + 0.05618581920862198, + 0.2813541293144226, + 0.04162435233592987, + 0.14065353572368622, + 0.028014961630105972 + ], + "credit_map": [ + 0.00986588466912508, + 0.014936390332877636, + 0.0169544517993927, + 0.024641185998916626, + 0.024187346920371056, + 0.013876131735742092, + 0.01672566495835781, + 0.01905776746571064, + 0.008866299875080585, + 0.04813390597701073, + 0.012703530490398407, + 0.02874559350311756, + 0.02157389186322689, + 0.02189639024436474, + 0.012953922152519226, + 0.011782998219132423, + 0.003219207050278783, + 0.007344820071011782, + 0.003235505660995841, + 0.007933498360216618, + 0.007932662032544613, + 0.006508889142423868, + 0.005666954908519983, + 0.002664794446900487, + 0.003032921114936471, + 0.006283894646912813, + 0.0025788939092308283, + 0.004710844252258539, + 0.006949309725314379, + 0.007203252520412207, + 0.004124699626117945, + 0.002173836575821042, + 0.005215393844991922, + 0.007827681489288807, + 0.004483988042920828, + 0.009022921323776245, + 0.01143423467874527, + 0.010931474156677723, + 0.007093195337802172, + 0.007692965213209391, + 0.0044450354762375355, + 0.009265067987143993, + 0.003181473119184375, + 0.0027437382377684116, + 0.017526229843497276, + 0.007371008861809969, + 0.004450824111700058, + 0.007202447857707739, + 0.012938156723976135, + 0.015446188859641552, + 0.017166711390018463, + 0.02459583431482315, + 0.016654495149850845, + 0.09614662826061249, + 0.02244308777153492, + 0.07596307247877121, + 0.004647686146199703, + 0.006748522166162729, + 0.003721943823620677, + 0.005187691189348698, + 0.0069311619736254215, + 0.005672006402164698, + 0.0040289065800607204, + 0.004686437547206879, + 0.008300515823066235, + 0.007682971656322479, + 0.007809285074472427, + 0.009744013659656048, + 0.00518193282186985, + 0.05137922242283821, + 0.007100203540176153, + 0.04345541447401047, + 0.002434314461424947, + 0.0033142217434942722, + 0.0029419008642435074, + 0.00548664852976799, + 0.0040594302117824554, + 0.0033052179496735334, + 0.003693430218845606, + 0.002779800444841385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.02073037438094616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372564539313316 + ], + "opposite_logprob_tokens": [ + -3.4087257385253906 + ], + "logprob_reward": -0.03372564539313316, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372564539313316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09909144043922424, + 0.08693846315145493, + 0.01694108545780182, + 0.11901750415563583, + 0.34259018301963806, + 0.026380090042948723, + 0.10816002637147903, + 0.0606546513736248, + 0.06507162004709244, + 0.07515499740839005 + ], + "credit_map": [ + 0.009005431085824966, + 0.006023373920470476, + 0.009908890351653099, + 0.03057284653186798, + 0.009288708679378033, + 0.00690105278044939, + 0.019783878698945045, + 0.007607256527990103, + 0.0073724486865103245, + 0.010137669742107391, + 0.01108044758439064, + 0.01250215619802475, + 0.016262095421552658, + 0.01289580762386322, + 0.008875682950019836, + 0.007812152151018381, + 0.0016110409051179886, + 0.0019416494760662317, + 0.0014251100365072489, + 0.0034517233725637197, + 0.0026713383849710226, + 0.0026499915402382612, + 0.0021942739840596914, + 0.0009959567105397582, + 0.012087343260645866, + 0.01409683283418417, + 0.011672106571495533, + 0.02083468995988369, + 0.014720594510436058, + 0.023940537124872208, + 0.014097072184085846, + 0.0075683146715164185, + 0.01830645091831684, + 0.08978348225355148, + 0.009289216250181198, + 0.014664252288639545, + 0.1362287998199463, + 0.024652911350131035, + 0.025193627923727036, + 0.024471430107951164, + 0.0022926388774067163, + 0.00408180383965373, + 0.0018596287118270993, + 0.0023817922919988632, + 0.006023991387337446, + 0.0028382695745676756, + 0.003025439800694585, + 0.0038765212520956993, + 0.006918839644640684, + 0.0102014672011137, + 0.01302332617342472, + 0.02147834748029709, + 0.008365624584257603, + 0.01961520127952099, + 0.012064904905855656, + 0.01649230159819126, + 0.004135751165449619, + 0.005248756147921085, + 0.005457770545035601, + 0.0070861829444766045, + 0.009777243249118328, + 0.01320644561201334, + 0.006599067244678736, + 0.00914343073964119, + 0.004184406716376543, + 0.006766810547560453, + 0.006716455798596144, + 0.010357752442359924, + 0.004866232164204121, + 0.014104166999459267, + 0.006702410522848368, + 0.011373382993042469, + 0.0063049388118088245, + 0.007834182120859623, + 0.006198366172611713, + 0.010102505795657635, + 0.008944930508732796, + 0.015457632951438427, + 0.008378827013075352, + 0.011933603323996067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.03372564539313316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430564358830452 + ], + "opposite_logprob_tokens": [ + -3.1680564880371094 + ], + "logprob_reward": -0.0430564358830452, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430564358830452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10660210996866226, + 0.072696752846241, + 0.03323359414935112, + 0.09805990755558014, + 0.17135436832904816, + 0.0715443342924118, + 0.06831403821706772, + 0.07424625009298325, + 0.16556403040885925, + 0.13838455080986023 + ], + "credit_map": [ + 0.011715762317180634, + 0.012780915014445782, + 0.012366564013063908, + 0.01235265377908945, + 0.01752515509724617, + 0.015339277684688568, + 0.01338236965239048, + 0.011139425449073315, + 0.006682477425783873, + 0.01729716546833515, + 0.005905004683881998, + 0.011313050054013729, + 0.008796924725174904, + 0.009763771668076515, + 0.006507066544145346, + 0.006431301590055227, + 0.0028021379839628935, + 0.0045298319309949875, + 0.003219578182324767, + 0.005374807398766279, + 0.004847853910177946, + 0.006323768757283688, + 0.004011087119579315, + 0.0021245323587208986, + 0.011620215140283108, + 0.010614448226988316, + 0.006911607459187508, + 0.014615259133279324, + 0.012480268254876137, + 0.025832869112491608, + 0.009431888349354267, + 0.0065533737652003765, + 0.008666458539664745, + 0.04088138788938522, + 0.00613414216786623, + 0.010186116211116314, + 0.06350535899400711, + 0.01613815315067768, + 0.012325049377977848, + 0.013517710380256176, + 0.010370389558374882, + 0.009102264419198036, + 0.005850956775248051, + 0.0072511001490056515, + 0.013893790543079376, + 0.00899425707757473, + 0.008408626541495323, + 0.007672963198274374, + 0.0069656698033213615, + 0.006047799251973629, + 0.007284368388354778, + 0.012424884364008904, + 0.006410375237464905, + 0.01048622652888298, + 0.009078577160835266, + 0.00961615052074194, + 0.0081758638843894, + 0.008161747828125954, + 0.005100180860608816, + 0.00761695159599185, + 0.009641123004257679, + 0.0172208771109581, + 0.006826009601354599, + 0.011503509245812893, + 0.006752674467861652, + 0.010387619957327843, + 0.00845580268651247, + 0.017628204077482224, + 0.007923422381281853, + 0.05479361489415169, + 0.009226352907717228, + 0.05039636418223381, + 0.007827263325452805, + 0.00976461824029684, + 0.014678838662803173, + 0.01857767626643181, + 0.010026831179857254, + 0.03681056201457977, + 0.013290129601955414, + 0.027408648282289505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.0430564358830452 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028242485132068396 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.0028242485132068396, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028242485132068396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08902452141046524, + 0.07352185249328613, + 0.0374121256172657, + 0.05152314528822899, + 0.1353207528591156, + 0.08108382672071457, + 0.16800346970558167, + 0.10917478799819946, + 0.08012908697128296, + 0.1748064160346985 + ], + "credit_map": [ + 0.007440848741680384, + 0.01011534221470356, + 0.008099310100078583, + 0.01717352867126465, + 0.012941643595695496, + 0.008601481094956398, + 0.012442643754184246, + 0.01220971904695034, + 0.0042547849006950855, + 0.017134109511971474, + 0.005220956169068813, + 0.011434570886194706, + 0.008508712984621525, + 0.013091417960822582, + 0.007475678808987141, + 0.006401625461876392, + 0.0032702742610126734, + 0.005550325848162174, + 0.002799877431243658, + 0.006579587236046791, + 0.0044679767452180386, + 0.007345167919993401, + 0.005083054769784212, + 0.002315860940143466, + 0.0037477188743650913, + 0.007685823366045952, + 0.004507942125201225, + 0.007805241737514734, + 0.00754682719707489, + 0.011112751439213753, + 0.005869414657354355, + 0.0032474312465637922, + 0.006729363929480314, + 0.03375696763396263, + 0.004389885812997818, + 0.007341600023210049, + 0.0512881875038147, + 0.011146094650030136, + 0.007940408773720264, + 0.012728248722851276, + 0.00753441546112299, + 0.011876791715621948, + 0.005066295154392719, + 0.006224697455763817, + 0.019810860976576805, + 0.01297152042388916, + 0.006008787080645561, + 0.011590455658733845, + 0.008543177507817745, + 0.00983502995222807, + 0.011852688156068325, + 0.016818486154079437, + 0.010236642323434353, + 0.053374867886304855, + 0.013615109026432037, + 0.043727464973926544, + 0.00554041238501668, + 0.008241087198257446, + 0.008089082315564156, + 0.011095879599452019, + 0.010906706564128399, + 0.03044443577528, + 0.014143278822302818, + 0.020713914185762405, + 0.0061576636508107185, + 0.0061896201223134995, + 0.00705364253371954, + 0.0080705052241683, + 0.005099852103739977, + 0.020162254571914673, + 0.006402500905096531, + 0.020993053913116455, + 0.012781898491084576, + 0.009405248798429966, + 0.021487826481461525, + 0.021518338471651077, + 0.010670285671949387, + 0.04861735180020332, + 0.016158398240804672, + 0.03416706621646881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.0028242485132068396 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497517064213753 + ], + "opposite_logprob_tokens": [ + -2.9299752712249756 + ], + "logprob_reward": -0.05497517064213753, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497517064213753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1444258838891983, + 0.04641060158610344, + 0.012083609588444233, + 0.03619104623794556, + 0.10772787034511566, + 0.0965600237250328, + 0.44053688645362854, + 0.024367686361074448, + 0.07096034288406372, + 0.02073606848716736 + ], + "credit_map": [ + 0.009083726443350315, + 0.007795574609190226, + 0.013134585693478584, + 0.0465923547744751, + 0.009713360108435154, + 0.009748023934662342, + 0.02404952421784401, + 0.024308741092681885, + 0.002879153937101364, + 0.012259235605597496, + 0.0029433686286211014, + 0.007192226126790047, + 0.005235183052718639, + 0.008628102950751781, + 0.004042206797748804, + 0.003231124021112919, + 0.0010921740904450417, + 0.0012602610513567924, + 0.0013184043345972896, + 0.0019639008678495884, + 0.0017337955068796873, + 0.0019245672738179564, + 0.001823995728045702, + 0.0009665109100751579, + 0.0027889637276530266, + 0.004764333367347717, + 0.0029563254211097956, + 0.005384708754718304, + 0.006609441712498665, + 0.006286363583058119, + 0.00435616122558713, + 0.0030447475146502256, + 0.007520674727857113, + 0.02288784645497799, + 0.0053312936797738075, + 0.007804689928889275, + 0.03175682574510574, + 0.013415726833045483, + 0.008873059414327145, + 0.010137747041881084, + 0.005826877895742655, + 0.01153639517724514, + 0.006252137012779713, + 0.008621851913630962, + 0.02179073542356491, + 0.01644221879541874, + 0.01747015118598938, + 0.008619658648967743, + 0.021383946761488914, + 0.024218128994107246, + 0.027696019038558006, + 0.0328989252448082, + 0.027755487710237503, + 0.15440046787261963, + 0.03335864841938019, + 0.11882524192333221, + 0.002969109918922186, + 0.004082332365214825, + 0.002026803558692336, + 0.00320636760443449, + 0.0036688786931335926, + 0.0030702492222189903, + 0.0028716158121824265, + 0.002472328022122383, + 0.00479136174544692, + 0.0034267199225723743, + 0.005014528054744005, + 0.005272457841783762, + 0.0037680589593946934, + 0.02416164241731167, + 0.005184566602110863, + 0.019341005012392998, + 0.001639481750316918, + 0.0021569430828094482, + 0.002718325238674879, + 0.0037408005446195602, + 0.002142843557521701, + 0.003091598628088832, + 0.002860164502635598, + 0.002385911764577031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.05497517064213753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extreme close-up, rendered in a hyperrealistic style, begins with the camera holding static over a glistening, translucent pool of melted wax on a thick, cream-colored candle. The liquid wax, emitting a warm golden glow from the invisible flame above, gently overflows its cratered edges. A small amount of wax is meticulously gathered into a perfect, viscous bead, its surface tension momentarily holding it together before the camera begins a slow, deliberate downward tilt, following the irresistible pull of gravity. The droplet begins to slide slowly along the textured, uneven side of the candle, leaving a shimmering, translucent trail behind the previously solidified droplet. As it nears the bottom edge, the droplet's own weight causes it to elongate, stretching into a teardrop shape, its neck tapering. The warm candlelight beautifully refracts through its increasingly elongated form. In the final moments, the connection breaks cleanly, and the droplet separates, plummeting a short distance to a dark, polished wooden tabletop below. Its rich texture is visible in the soft, flickering ambient light, creating an atmosphere of tranquility and concentration captured with intense macroscopic detail and physically accurate fluid dynamics.", + "seed": 6242, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_11/index_2874_seed_6242_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.024072079360485076, + "vlm_logprob": -0.024072079360485076, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.024072079360485076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02182711660861969 + ], + "opposite_logprob_tokens": [ + -4.021827220916748 + ], + "logprob_reward": -0.02182711660861969, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02182711660861969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06176906079053879, + 0.03183446824550629, + 0.012797605246305466, + 0.040219057351350784, + 0.01736248843371868, + 0.12315654009580612, + 0.13986967504024506, + 0.23200401663780212, + 0.24002692103385925, + 0.1009601280093193 + ], + "credit_map": [ + 0.006361405365169048, + 0.006404532119631767, + 0.00822189450263977, + 0.01383501011878252, + 0.006324378773570061, + 0.006221267860382795, + 0.007465873844921589, + 0.006934698671102524, + 0.0031739375554025173, + 0.002872404409572482, + 0.0029965818393975496, + 0.0046235909685492516, + 0.0054574087262153625, + 0.0038941986858844757, + 0.0048959203995764256, + 0.003920426592230797, + 0.0014891054015606642, + 0.0015575597062706947, + 0.0012178111355751753, + 0.0018745942506939173, + 0.0017753777792677283, + 0.001837863470427692, + 0.0015361369587481022, + 0.001509157009422779, + 0.004668125882744789, + 0.003914202097803354, + 0.004372073337435722, + 0.007233561482280493, + 0.00571628101170063, + 0.004903592634946108, + 0.00586890522390604, + 0.003542316611856222, + 0.001604318036697805, + 0.0015380745753645897, + 0.0014125207671895623, + 0.0027554708067327738, + 0.0018404759466648102, + 0.0024212640710175037, + 0.0023809776175767183, + 0.0034093866124749184, + 0.01240890845656395, + 0.009801032021641731, + 0.01528269611299038, + 0.023777028545737267, + 0.00890051107853651, + 0.011406430043280125, + 0.013950616121292114, + 0.027629319578409195, + 0.011372733861207962, + 0.010995601303875446, + 0.01319157611578703, + 0.037921760231256485, + 0.01220019068568945, + 0.02212786115705967, + 0.01983804441988468, + 0.012221921235322952, + 0.018183916807174683, + 0.03666652366518974, + 0.02315998077392578, + 0.030013155192136765, + 0.028046734631061554, + 0.0521961972117424, + 0.024772968143224716, + 0.01896454021334648, + 0.0138908252120018, + 0.06032710522413254, + 0.017298396676778793, + 0.02262238599359989, + 0.05663689598441124, + 0.019524142146110535, + 0.024885937571525574, + 0.02484123781323433, + 0.006556481588631868, + 0.012835758738219738, + 0.013320853933691978, + 0.010564222931861877, + 0.019013509154319763, + 0.013254630379378796, + 0.015496354550123215, + 0.009918314404785633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.02182711660861969 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497268959879875 + ], + "opposite_logprob_tokens": [ + -2.9299726486206055 + ], + "logprob_reward": -0.05497268959879875, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497268959879875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058813005685806274, + 0.0359346829354763, + 0.03959411755204201, + 0.06839506328105927, + 0.02120882458984852, + 0.26140719652175903, + 0.09242740273475647, + 0.06426350772380829, + 0.2072807103395462, + 0.15067549049854279 + ], + "credit_map": [ + 0.006362412590533495, + 0.00472840154543519, + 0.007750295102596283, + 0.009879042394459248, + 0.006528002675622702, + 0.006196359172463417, + 0.00680661853402853, + 0.010561873205006123, + 0.003577398369088769, + 0.003298015333712101, + 0.003826526692137122, + 0.0050440565682947636, + 0.005019206553697586, + 0.004699390381574631, + 0.005428972188383341, + 0.005041113588958979, + 0.003367442637681961, + 0.0035493960604071617, + 0.006576609332114458, + 0.009667901322245598, + 0.0035835038870573044, + 0.003528547938913107, + 0.005536941811442375, + 0.003783774096518755, + 0.007122494280338287, + 0.005777871701866388, + 0.008000306785106659, + 0.01554778777062893, + 0.007628863211721182, + 0.006927323993295431, + 0.010608441196382046, + 0.006781972479075193, + 0.0016843808116391301, + 0.0016728314803913236, + 0.0017404129030182958, + 0.004238994792103767, + 0.0017813033191487193, + 0.003869386622682214, + 0.0036176005378365517, + 0.002603915287181735, + 0.02615233138203621, + 0.018597256392240524, + 0.032921869307756424, + 0.04101391136646271, + 0.01623542048037052, + 0.02790343016386032, + 0.029557960107922554, + 0.06902503967285156, + 0.007132347673177719, + 0.009478149004280567, + 0.007921119220554829, + 0.018681146204471588, + 0.010775147937238216, + 0.012661337852478027, + 0.018616894260048866, + 0.007161254994571209, + 0.006924229208379984, + 0.01579039730131626, + 0.005875581409782171, + 0.006752910558134317, + 0.009646383114159107, + 0.00855228677392006, + 0.007155014667659998, + 0.0035667018964886665, + 0.010502375662326813, + 0.051307860761880875, + 0.013888629153370857, + 0.0194365493953228, + 0.047801673412323, + 0.014260427094995975, + 0.025224661454558372, + 0.02485852874815464, + 0.013851262629032135, + 0.018379151821136475, + 0.015995435416698456, + 0.01621965505182743, + 0.031739525496959686, + 0.02386912703514099, + 0.018441256135702133, + 0.012180076912045479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.05497268959879875 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633799687027931 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633799687027931, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633799687027931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04802758991718292, + 0.0139870410785079, + 0.02831858955323696, + 0.026406407356262207, + 0.017526889219880104, + 0.26864975690841675, + 0.08101902157068253, + 0.21577738225460052, + 0.19982077181339264, + 0.10046662390232086 + ], + "credit_map": [ + 0.006423063576221466, + 0.00347824115306139, + 0.00687100924551487, + 0.011429094709455967, + 0.004361278377473354, + 0.00473802862688899, + 0.005752949044108391, + 0.004973922856152058, + 0.0015998204471543431, + 0.001266195671632886, + 0.0016411810647696257, + 0.0017370638670399785, + 0.0011524718720465899, + 0.0023079966194927692, + 0.0014363491209223866, + 0.002845962066203356, + 0.003279730910435319, + 0.0030161875765770674, + 0.002624654909595847, + 0.004178476519882679, + 0.0031206165440380573, + 0.004328755661845207, + 0.003031832631677389, + 0.004738333635032177, + 0.003305793972685933, + 0.002291004406288266, + 0.003403540002182126, + 0.003602940123528242, + 0.0017943454440683126, + 0.0037985474336892366, + 0.002618285594508052, + 0.0055919489823281765, + 0.0020801201462745667, + 0.001655035768635571, + 0.0018230914138257504, + 0.00224032299593091, + 0.0013983523240312934, + 0.0023534332867711782, + 0.0023768083192408085, + 0.0035997226368635893, + 0.021863065659999847, + 0.020027324557304382, + 0.03547604754567146, + 0.053309399634599686, + 0.016462447121739388, + 0.026202289387583733, + 0.03174909949302673, + 0.06356006860733032, + 0.0046608527190983295, + 0.007351523730903864, + 0.006797823589295149, + 0.024797111749649048, + 0.01163213700056076, + 0.00776261929422617, + 0.013093418441712856, + 0.004923528525978327, + 0.02237524464726448, + 0.027185408398509026, + 0.026280144229531288, + 0.03460778668522835, + 0.028823483735322952, + 0.028986088931560516, + 0.03061608411371708, + 0.016903143376111984, + 0.011230891570448875, + 0.04623027518391609, + 0.014707863330841064, + 0.01881171576678753, + 0.04160613566637039, + 0.014290763065218925, + 0.02606973797082901, + 0.02687337063252926, + 0.008889161050319672, + 0.0121197160333395, + 0.010177482850849628, + 0.009786141104996204, + 0.024503080174326897, + 0.012840773910284042, + 0.014246356673538685, + 0.007903899997472763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.007633799687027931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200683578848839 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200683578848839, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200683578848839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026692450046539307, + 0.011689230799674988, + 0.013740471564233303, + 0.016653506085276604, + 0.009776345454156399, + 0.20101210474967957, + 0.18188518285751343, + 0.03640199825167656, + 0.3611624836921692, + 0.14098618924617767 + ], + "credit_map": [ + 0.00203387881629169, + 0.002229876583442092, + 0.0037957215681672096, + 0.006192924454808235, + 0.0021686579566448927, + 0.004673364572227001, + 0.0029283498879522085, + 0.0026696757413446903, + 0.0008804071112535894, + 0.0011926102451980114, + 0.001273882808163762, + 0.002693731803447008, + 0.0016256010858342052, + 0.0015327046858146787, + 0.0012406266760081053, + 0.0012496671406552196, + 0.0012210826389491558, + 0.0013576492201536894, + 0.0011039660312235355, + 0.0025076712481677532, + 0.0015974477864801884, + 0.0026626386679708958, + 0.001685276860371232, + 0.001604740391485393, + 0.00196191668510437, + 0.001394251943565905, + 0.001523687969893217, + 0.0030476136598736048, + 0.0015740947565063834, + 0.0021823211573064327, + 0.0024525884073227644, + 0.0025170331355184317, + 0.0014443504624068737, + 0.0012852855725213885, + 0.0009353382629342377, + 0.001383354770950973, + 0.0011070112232118845, + 0.0012341662077233195, + 0.0011712531559169292, + 0.0012155857402831316, + 0.02012406289577484, + 0.015726236626505852, + 0.023369228467345238, + 0.03140605241060257, + 0.01422386709600687, + 0.021410437300801277, + 0.023222142830491066, + 0.05153008922934532, + 0.014940179884433746, + 0.016347616910934448, + 0.01771382987499237, + 0.04037225991487503, + 0.015718428418040276, + 0.0280468612909317, + 0.028800738975405693, + 0.019945278763771057, + 0.002616721438243985, + 0.006776424124836922, + 0.0033375469502061605, + 0.00472044525668025, + 0.004034166689962149, + 0.009160923771560192, + 0.0032995683141052723, + 0.002456201473250985, + 0.01680661551654339, + 0.08677586913108826, + 0.030816366896033287, + 0.027401752769947052, + 0.08380821347236633, + 0.029154600575566292, + 0.03541979938745499, + 0.050979286432266235, + 0.00989239290356636, + 0.03237232193350792, + 0.017215421423316002, + 0.013419918715953827, + 0.027524543926119804, + 0.012600995600223541, + 0.016212116926908493, + 0.011748485267162323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.002200683578848839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372610732913017 + ], + "opposite_logprob_tokens": [ + -3.408726215362549 + ], + "logprob_reward": -0.03372610732913017, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372610732913017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06080256775021553, + 0.027140136808156967, + 0.025496531277894974, + 0.036496829241514206, + 0.030380766838788986, + 0.07711921632289886, + 0.2942430377006531, + 0.15849891304969788, + 0.15381859242916107, + 0.13600338995456696 + ], + "credit_map": [ + 0.0057255118153989315, + 0.005937203764915466, + 0.0077921864576637745, + 0.00815440434962511, + 0.00684715760871768, + 0.007912595756351948, + 0.00886906124651432, + 0.009564448148012161, + 0.0035081675741821527, + 0.0026235703844577074, + 0.0029329326935112476, + 0.0036349082365632057, + 0.0030678831972181797, + 0.004599777981638908, + 0.003209462156519294, + 0.003563435049727559, + 0.002786977682262659, + 0.0021570916287600994, + 0.0025117695331573486, + 0.0031146788969635963, + 0.002444034907966852, + 0.005080842413008213, + 0.0033116734120994806, + 0.004089463036507368, + 0.003969169221818447, + 0.0034677814692258835, + 0.005055560264736414, + 0.007346602156758308, + 0.0032866133842617273, + 0.0032980674877762794, + 0.0054108016192913055, + 0.0046622357331216335, + 0.004169190302491188, + 0.003614885499700904, + 0.003817636286839843, + 0.004298848565667868, + 0.0034073241986334324, + 0.003978450782597065, + 0.0035227180924266577, + 0.0035717154387384653, + 0.00524148577824235, + 0.005618568975478411, + 0.011618796736001968, + 0.016925137490034103, + 0.0049499026499688625, + 0.007501519285142422, + 0.009859158657491207, + 0.015404644422233105, + 0.0254158116877079, + 0.025898698717355728, + 0.027867911383509636, + 0.06265491247177124, + 0.026353277266025543, + 0.0468139685690403, + 0.043895453214645386, + 0.03534300997853279, + 0.013610129244625568, + 0.031399037688970566, + 0.015605238266289234, + 0.022313842549920082, + 0.021552957594394684, + 0.020785385742783546, + 0.02352762594819069, + 0.009704701602458954, + 0.009112875908613205, + 0.037413641810417175, + 0.013325612992048264, + 0.0192499328404665, + 0.03515103459358215, + 0.01061183586716652, + 0.014883376657962799, + 0.014070287346839905, + 0.010656150989234447, + 0.02203962579369545, + 0.015229725278913975, + 0.013645811006426811, + 0.01883750967681408, + 0.017151745036244392, + 0.022995974868535995, + 0.015446854755282402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.03372610732913017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extreme close-up, rendered in a hyperrealistic style, begins with the camera holding static over a glistening, translucent pool of melted wax on a thick, cream-colored candle. The liquid wax, emitting a warm golden glow from the invisible flame above, gently overflows its cratered edges. A small amount of wax is meticulously gathered into a perfect, viscous bead, its surface tension momentarily holding it together before the camera begins a slow, deliberate downward tilt, following the irresistible pull of gravity. The droplet begins to slide slowly along the textured, uneven side of the candle, leaving a shimmering, translucent trail behind the previously solidified droplet. As it nears the bottom edge, the droplet's own weight causes it to elongate, stretching into a teardrop shape, its neck tapering. The warm candlelight beautifully refracts through its increasingly elongated form. In the final moments, the connection breaks cleanly, and the droplet separates, plummeting a short distance to a dark, polished wooden tabletop below. Its rich texture is visible in the soft, flickering ambient light, creating an atmosphere of tranquility and concentration captured with intense macroscopic detail and physically accurate fluid dynamics.", + "seed": 6243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_11/index_2874_seed_6243_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013340407563373447, + "vlm_logprob": -0.013340407563373447, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.013340407563373447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006446281913667917 + ], + "opposite_logprob_tokens": [ + -5.881446361541748 + ], + "logprob_reward": -0.006446281913667917, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006446281913667917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045688606798648834, + 0.17368800938129425, + 0.07757687568664551, + 0.02321302890777588, + 0.31860002875328064, + 0.08537834137678146, + 0.10326548665761948, + 0.12454772740602493, + 0.022293541580438614, + 0.025748390704393387 + ], + "credit_map": [ + 0.013752024620771408, + 0.00518363481387496, + 0.005189598072320223, + 0.0034649712033569813, + 0.007261788006871939, + 0.0038215748500078917, + 0.004317642655223608, + 0.0026973693165928125, + 0.03323958069086075, + 0.018909409642219543, + 0.010425098240375519, + 0.009832003153860569, + 0.012646769173443317, + 0.06330610066652298, + 0.010381168685853481, + 0.014947856776416302, + 0.007050287909805775, + 0.004354438744485378, + 0.007548322901129723, + 0.03261525556445122, + 0.0036337629426270723, + 0.005810148548334837, + 0.012905535288155079, + 0.0036591137759387493, + 0.002533103572204709, + 0.002225626492872834, + 0.0018786132568493485, + 0.0066498345695436, + 0.0021098509896546602, + 0.003214990021660924, + 0.003430719254538417, + 0.0011702881893143058, + 0.08733423799276352, + 0.023520249873399734, + 0.023355405777692795, + 0.11388767510652542, + 0.013441753573715687, + 0.013516214676201344, + 0.02999105677008629, + 0.01355341449379921, + 0.009934097528457642, + 0.015808777883648872, + 0.0055482094176113605, + 0.011625127866864204, + 0.022575266659259796, + 0.007708671968430281, + 0.009161507710814476, + 0.0030166776850819588, + 0.007393427658826113, + 0.026720266789197922, + 0.003911216743290424, + 0.005532297305762768, + 0.04040014371275902, + 0.005995550658553839, + 0.006926191970705986, + 0.0063863820396363735, + 0.00891197845339775, + 0.027780313044786453, + 0.03534528613090515, + 0.010475764982402325, + 0.01571042090654373, + 0.0061749001033604145, + 0.007851415313780308, + 0.012297626584768295, + 0.001584990881383419, + 0.002434094436466694, + 0.0040967026725411415, + 0.0037140040658414364, + 0.002343685133382678, + 0.002190342638641596, + 0.0035345908254384995, + 0.0023951278999447823, + 0.004945675376802683, + 0.0033945394679903984, + 0.003365576732903719, + 0.0024794277269393206, + 0.0020528335589915514, + 0.003076766850426793, + 0.003043923992663622, + 0.0033896451350301504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.006446281913667917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081271693110466 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081271693110466, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081271693110466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09123023599386215, + 0.21730154752731323, + 0.2203354835510254, + 0.0806644931435585, + 0.03343519568443298, + 0.09340814501047134, + 0.14185617864131927, + 0.07850035279989243, + 0.028527215123176575, + 0.014741087332367897 + ], + "credit_map": [ + 0.029306601732969284, + 0.009885217994451523, + 0.009927313774824142, + 0.005898359697312117, + 0.008244921453297138, + 0.008981207385659218, + 0.007836618460714817, + 0.011149993166327477, + 0.05019066482782364, + 0.02050556242465973, + 0.015608386136591434, + 0.012007900513708591, + 0.01557599101215601, + 0.07156674563884735, + 0.013797628693282604, + 0.018048666417598724, + 0.012194852344691753, + 0.011330056935548782, + 0.01669507846236229, + 0.08485543727874756, + 0.0076010217890143394, + 0.0492379255592823, + 0.028070131316781044, + 0.010350970551371574, + 0.006121685262769461, + 0.011407560668885708, + 0.007698203437030315, + 0.019824033603072166, + 0.009305603802204132, + 0.007429228629916906, + 0.015020116232335567, + 0.0038580612745136023, + 0.004183121491223574, + 0.002996626542881131, + 0.0023401607759296894, + 0.010470542125403881, + 0.0034261776600033045, + 0.0030895022209733725, + 0.0044910237193107605, + 0.0024380397517234087, + 0.008695702999830246, + 0.02347201108932495, + 0.0070287911221385, + 0.012272271327674389, + 0.021446848288178444, + 0.0059383343905210495, + 0.010282144881784916, + 0.004272046498954296, + 0.01040846761316061, + 0.037567880004644394, + 0.00460802624002099, + 0.007541104219853878, + 0.0564536452293396, + 0.008648637682199478, + 0.009088918566703796, + 0.007539514452219009, + 0.005765856243669987, + 0.014139085076749325, + 0.02458079159259796, + 0.007720967289060354, + 0.0075396220199763775, + 0.003733512246981263, + 0.006535801570862532, + 0.008484709076583385, + 0.0016415605787187815, + 0.003002624260261655, + 0.004138767719268799, + 0.0037319851107895374, + 0.003344736760482192, + 0.006635473109781742, + 0.003432234050706029, + 0.00259983423165977, + 0.0016631054459139705, + 0.0024492486845701933, + 0.001712504425086081, + 0.0021057329140603542, + 0.0016205867286771536, + 0.0013861662009730935, + 0.001985660521313548, + 0.0018180814804509282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.016081271693110466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194007031619549 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194007031619549, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194007031619549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15804779529571533, + 0.055084891617298126, + 0.1179187074303627, + 0.049206994473934174, + 0.04613865166902542, + 0.23139889538288116, + 0.058718353509902954, + 0.08714627474546432, + 0.14864587783813477, + 0.04769351705908775 + ], + "credit_map": [ + 0.05325452983379364, + 0.01627076230943203, + 0.016990363597869873, + 0.010202931240200996, + 0.017622632905840874, + 0.014946751296520233, + 0.012849166989326477, + 0.01591068133711815, + 0.012994343414902687, + 0.006086502689868212, + 0.005053066182881594, + 0.004057969897985458, + 0.00529594486579299, + 0.01389643270522356, + 0.0038829303812235594, + 0.0038177049718797207, + 0.007151711732149124, + 0.005148028954863548, + 0.010579370893537998, + 0.04951818659901619, + 0.004336333367973566, + 0.020066190510988235, + 0.015651680529117584, + 0.00546722486615181, + 0.004499883856624365, + 0.0041818926110863686, + 0.003811372211202979, + 0.015350338071584702, + 0.00385401607491076, + 0.007214778568595648, + 0.0075266361236572266, + 0.002768081845715642, + 0.010095605626702309, + 0.006268893834203482, + 0.003253980539739132, + 0.006605945527553558, + 0.005188239738345146, + 0.0032741441391408443, + 0.007941973395645618, + 0.0035098728258162737, + 0.022961314767599106, + 0.037309952080249786, + 0.031253088265657425, + 0.023007437586784363, + 0.057415299117565155, + 0.018050458282232285, + 0.027917969971895218, + 0.013483398593962193, + 0.004305120091885328, + 0.01255873218178749, + 0.00358711089938879, + 0.004552380181849003, + 0.019392814487218857, + 0.003771159565076232, + 0.007207529619336128, + 0.0033435134682804346, + 0.006965428590774536, + 0.004341499879956245, + 0.029067518189549446, + 0.010437747463583946, + 0.006342833396047354, + 0.005962080787867308, + 0.011712084524333477, + 0.012317092157900333, + 0.009486292488873005, + 0.021521901711821556, + 0.018133152276277542, + 0.022537721320986748, + 0.017359789460897446, + 0.0199735090136528, + 0.01871854066848755, + 0.020914992317557335, + 0.006471403408795595, + 0.007038808427751064, + 0.008601619862020016, + 0.005153985228389502, + 0.0052552116103470325, + 0.005187434144318104, + 0.00572893675416708, + 0.004256123676896095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.014194007031619549 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003618955845013261 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.003618955845013261, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003618955845013261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04259840026497841, + 0.08171754330396652, + 0.1803559958934784, + 0.02551363781094551, + 0.08021324872970581, + 0.07422707974910736, + 0.14865991473197937, + 0.28488096594810486, + 0.061198219656944275, + 0.02063501439988613 + ], + "credit_map": [ + 0.013480830937623978, + 0.004695067182183266, + 0.005238336045295, + 0.0030470704659819603, + 0.006433155387639999, + 0.0036142265889793634, + 0.0033157533034682274, + 0.002773960353806615, + 0.024127349257469177, + 0.007058513350784779, + 0.007197265513241291, + 0.004943040199577808, + 0.005806042812764645, + 0.020374005660414696, + 0.005414782091975212, + 0.006796546280384064, + 0.011182157322764397, + 0.00846611987799406, + 0.014999080449342728, + 0.075139619410038, + 0.006135978735983372, + 0.03251376003026962, + 0.024212589487433434, + 0.007706691976636648, + 0.002804673044010997, + 0.003075967077165842, + 0.0020409112330526114, + 0.007757685147225857, + 0.0020301046315580606, + 0.002423518802970648, + 0.004070705268532038, + 0.0013100722571834922, + 0.01857706718146801, + 0.005759084597229958, + 0.006163625977933407, + 0.02835087664425373, + 0.004244609735906124, + 0.004776041954755783, + 0.008768844418227673, + 0.0035731038078665733, + 0.008511703461408615, + 0.008493311703205109, + 0.00721649918705225, + 0.020854052156209946, + 0.006093007046729326, + 0.003663690760731697, + 0.014466136693954468, + 0.004928684327751398, + 0.011153504252433777, + 0.03818456456065178, + 0.005217642057687044, + 0.007440648041665554, + 0.058710794895887375, + 0.009342907927930355, + 0.009806905873119831, + 0.00880296528339386, + 0.019582493230700493, + 0.05122381076216698, + 0.09262704104185104, + 0.025814350694417953, + 0.029216839000582695, + 0.013382993638515472, + 0.021845726296305656, + 0.031187698245048523, + 0.0035858165938407183, + 0.005784942302852869, + 0.009841062128543854, + 0.01098155789077282, + 0.005866301245987415, + 0.010562430135905743, + 0.0087806461378932, + 0.005795466247946024, + 0.0018187249079346657, + 0.003807850880548358, + 0.0024531821254640818, + 0.002932809293270111, + 0.0020664921030402184, + 0.0020673645194619894, + 0.0032756158616393805, + 0.002212975174188614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.003618955845013261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a glistening, translucent pool of melted wax on a thick, cream-colored candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid wax overflow its cratered edges and form a viscous bead held by surface tension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a static hold and then perform a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wax droplet slide down the textured candle side, elongate into a teardrop shape, and separate cleanly to fall onto a dark, polished wooden tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a hyperrealistic style with intense macroscopic detail and physically accurate fluid dynamics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636152133345604 + ], + "opposite_logprob_tokens": [ + -3.6513614654541016 + ], + "logprob_reward": -0.02636152133345604, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636152133345604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25589650869369507, + 0.14451655745506287, + 0.017244378104805946, + 0.021814825013279915, + 0.0628105103969574, + 0.08301779627799988, + 0.21479247510433197, + 0.1462595909833908, + 0.0326639860868454, + 0.020983431488275528 + ], + "credit_map": [ + 0.08354616910219193, + 0.025969095528125763, + 0.02732132561504841, + 0.014918037690222263, + 0.021069975569844246, + 0.025490202009677887, + 0.022628311067819595, + 0.034953389316797256, + 0.030260931700468063, + 0.015086308121681213, + 0.009691464714705944, + 0.008268557488918304, + 0.010204288177192211, + 0.05017293989658356, + 0.009176300838589668, + 0.011655754409730434, + 0.002216281136497855, + 0.0015225718962028623, + 0.0036443863064050674, + 0.00279302173294127, + 0.001276438357308507, + 0.001565466052852571, + 0.002854133490473032, + 0.0013720780843868852, + 0.0021829092875123024, + 0.002646057168021798, + 0.001691373880021274, + 0.006615907419472933, + 0.0018551304237917066, + 0.003272866364568472, + 0.0025948965921998024, + 0.0009556817822158337, + 0.015583762899041176, + 0.0052978056482970715, + 0.004526258911937475, + 0.020690714940428734, + 0.004490535706281662, + 0.003088007913902402, + 0.005985756404697895, + 0.00314766401425004, + 0.008810433559119701, + 0.015070299617946148, + 0.0052875797264277935, + 0.012055051513016224, + 0.022920986637473106, + 0.007920601405203342, + 0.007962108589708805, + 0.002990725450217724, + 0.015569629147648811, + 0.056544043123722076, + 0.007375238463282585, + 0.01021833997219801, + 0.08594927191734314, + 0.013635479845106602, + 0.012504217214882374, + 0.012996233999729156, + 0.010691497474908829, + 0.02580557018518448, + 0.0482257679104805, + 0.01431773230433464, + 0.011978669092059135, + 0.007817800156772137, + 0.010793978348374367, + 0.016628559678792953, + 0.002217387082055211, + 0.004203148651868105, + 0.006015855818986893, + 0.005580737255513668, + 0.003463050816208124, + 0.00320248119533062, + 0.005125066731125116, + 0.0028562566731125116, + 0.0023663213942199945, + 0.002882935106754303, + 0.002686082385480404, + 0.002531701698899269, + 0.0026109570171684027, + 0.0026042377576231956, + 0.0028058812022209167, + 0.0024953139945864677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.02636152133345604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_12/index_tensor([9961, 9961])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_12/index_tensor([9961, 9961])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..297fb3122da38c9b0e255f70b53624379bde9596 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_12/index_tensor([9961, 9961])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, forward-moving close-up, framed by the dramatic contrast of classic black and white light and shadow, precisely captures a dancer twirling gracefully and seamlessly. Her hair is styled in a sleek, delicate ballet bun, a few strands of sweat-soaked black hair freed by the centrifugal force of the spin, gently brushing against her closed eyes and high nose. Her head tilts back slightly in time with the twirl, her jawline gracefully tensed, a faint, utterly captivated smile playing at the corners of her mouth, as if completely absorbed in the silent melody. A powerful spotlight from above and side, like Rembrandt's brush, sculpts the three-dimensional contours of her face, illuminating the curve of her cheekbones and her sweat-beaded forehead in the highlights while casting deep, soft shadows on the other side, intensifying her intense concentration and selflessness in her art. Her simple dark dance dress reveals a silky luster under the strong light, forming an extreme contrast with her fair skin. The whole picture presents a sculptural sense of tranquility in the dynamic, solidifying the momentary passion and emotion into a timeless portrait full of classical aesthetics and inner monologue.", + "seed": 11394, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_12/index_9961_seed_11394_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10415022606030107, + "vlm_logprob": -0.10415022606030107, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.10415022606030107, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2077997922897339 + ], + "opposite_logprob_tokens": [ + -1.8327997922897339 + ], + "logprob_reward": -0.2077997922897339, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.2077997922897339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09767621755599976, + 0.04159921407699585, + 0.04216752201318741, + 0.03726271912455559, + 0.17831631004810333, + 0.1988779604434967, + 0.08583854883909225, + 0.0939246416091919, + 0.06300824135541916, + 0.16132862865924835 + ], + "credit_map": [ + 0.009114915505051613, + 0.011720392853021622, + 0.012810732237994671, + 0.01762530393898487, + 0.015167034231126308, + 0.008677646517753601, + 0.013046947307884693, + 0.009513243101537228, + 0.005230858456343412, + 0.005027212668210268, + 0.005119933746755123, + 0.002837257692590356, + 0.010360345244407654, + 0.004481888376176357, + 0.0046423496678471565, + 0.0038993675261735916, + 0.004800969269126654, + 0.006583386100828648, + 0.005951269995421171, + 0.004402040969580412, + 0.00858340971171856, + 0.003432575147598982, + 0.004314763471484184, + 0.0040991040877997875, + 0.00343000004068017, + 0.004728642757982016, + 0.00492422841489315, + 0.007397386711090803, + 0.004453533794730902, + 0.004505821969360113, + 0.004230278544127941, + 0.003592829219996929, + 0.027126207947731018, + 0.011549015529453754, + 0.027500297874212265, + 0.014026593416929245, + 0.01786460541188717, + 0.05060518532991409, + 0.008967464789748192, + 0.02067694254219532, + 0.035187654197216034, + 0.017901787534356117, + 0.03226591274142265, + 0.01607804372906685, + 0.01821397989988327, + 0.024805841967463493, + 0.015357381664216518, + 0.03906737640500069, + 0.01397823728621006, + 0.0070471204817295074, + 0.01406546775251627, + 0.009026292711496353, + 0.011538970284163952, + 0.007634688168764114, + 0.005314059555530548, + 0.017233705148100853, + 0.015748566016554832, + 0.010266066528856754, + 0.012582036666572094, + 0.0038368115201592445, + 0.01357821561396122, + 0.009933609515428543, + 0.010002955794334412, + 0.01797638274729252, + 0.007859764620661736, + 0.007021902594715357, + 0.00814451090991497, + 0.006396452896296978, + 0.010897684842348099, + 0.00719812884926796, + 0.008139252662658691, + 0.007350540719926357, + 0.018238084390759468, + 0.0173873919993639, + 0.02481774613261223, + 0.025473402813076973, + 0.022616440430283546, + 0.015617032535374165, + 0.024673327803611755, + 0.012505185790359974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.2077997922897339 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20180515944957733 + ], + "opposite_logprob_tokens": [ + -1.7018051147460938 + ], + "logprob_reward": -0.20180515944957733, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20180515944957733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2367875576019287, + 0.10800248384475708, + 0.051676418632268906, + 0.026257742196321487, + 0.12744073569774628, + 0.2646216154098511, + 0.04550853371620178, + 0.06368311494588852, + 0.04662087932229042, + 0.029400909319519997 + ], + "credit_map": [ + 0.02759999781847, + 0.03776342049241066, + 0.041438471525907516, + 0.035407837480306625, + 0.03866363316774368, + 0.019456854090094566, + 0.023060625419020653, + 0.013396715745329857, + 0.017650017514824867, + 0.011243756860494614, + 0.01850232109427452, + 0.013017406687140465, + 0.01067277230322361, + 0.012977946549654007, + 0.009726130403578281, + 0.014212132431566715, + 0.005659482907503843, + 0.006188034545630217, + 0.007966414093971252, + 0.004624978639185429, + 0.008769872598350048, + 0.003925638273358345, + 0.005979838781058788, + 0.008562156930565834, + 0.002559768036007881, + 0.004387063439935446, + 0.003361272858455777, + 0.003637932240962982, + 0.0033541186712682247, + 0.002632076619192958, + 0.0036369406152516603, + 0.002688567852601409, + 0.015415281988680363, + 0.011795779690146446, + 0.01864198036491871, + 0.009997970424592495, + 0.020628737285733223, + 0.01695949025452137, + 0.008275280706584454, + 0.025726227089762688, + 0.047480400651693344, + 0.018956754356622696, + 0.04283599182963371, + 0.020531244575977325, + 0.023166101425886154, + 0.03435356542468071, + 0.021829210221767426, + 0.05546834319829941, + 0.010270416736602783, + 0.0026510613970458508, + 0.008139693178236485, + 0.0030582656618207693, + 0.0043905433267354965, + 0.0028502647764980793, + 0.0017038944642990828, + 0.012444394640624523, + 0.008235014975070953, + 0.0059177810326218605, + 0.008046284317970276, + 0.004333533346652985, + 0.009766439907252789, + 0.009746556170284748, + 0.005730974022299051, + 0.011906527914106846, + 0.007566576357930899, + 0.0032641044817864895, + 0.006752130575478077, + 0.005169863812625408, + 0.0055761937983334064, + 0.004809511825442314, + 0.006966340355575085, + 0.006516157183796167, + 0.0025283186696469784, + 0.0032966313883662224, + 0.003973804879933596, + 0.0037049981765449047, + 0.003917568828910589, + 0.0038637176621705294, + 0.004351131618022919, + 0.0037647385615855455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.20180515944957733 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634628098458052 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634628098458052, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634628098458052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22062280774116516, + 0.0830293521285057, + 0.07475509494543076, + 0.0543588325381279, + 0.14018864929676056, + 0.08644231408834457, + 0.0681597962975502, + 0.08481019735336304, + 0.07140373438596725, + 0.11622919142246246 + ], + "credit_map": [ + 0.020456422120332718, + 0.025612279772758484, + 0.025799430906772614, + 0.03741016238927841, + 0.03816137835383415, + 0.01989608258008957, + 0.029638100415468216, + 0.023648936301469803, + 0.010447307489812374, + 0.010505425743758678, + 0.012199606746435165, + 0.006184150464832783, + 0.02040839195251465, + 0.009324034675955772, + 0.00699589354917407, + 0.006964542903006077, + 0.007038208656013012, + 0.01075421180576086, + 0.010809307917952538, + 0.008383823558688164, + 0.013294973410665989, + 0.007240803446620703, + 0.006714826449751854, + 0.010518934577703476, + 0.005530751310288906, + 0.0087199155241251, + 0.008063438348472118, + 0.010559286922216415, + 0.006507105194032192, + 0.004128212109208107, + 0.005133617669343948, + 0.005716503132134676, + 0.017756447196006775, + 0.017022952437400818, + 0.01786503754556179, + 0.01581583358347416, + 0.02411104366183281, + 0.026970524340867996, + 0.007185768336057663, + 0.013461047783493996, + 0.013101274147629738, + 0.009062431752681732, + 0.012145389802753925, + 0.007237622048705816, + 0.013277323916554451, + 0.009660841897130013, + 0.008549133315682411, + 0.013408295810222626, + 0.00985371321439743, + 0.008366744965314865, + 0.00932241603732109, + 0.006984134204685688, + 0.009506313130259514, + 0.006506028585135937, + 0.007025927305221558, + 0.010594522580504417, + 0.018239913508296013, + 0.00712869968265295, + 0.012908106669783592, + 0.004139394499361515, + 0.00875105056911707, + 0.009224912151694298, + 0.004187293350696564, + 0.02023082785308361, + 0.011379209347069263, + 0.00928856898099184, + 0.01032558735460043, + 0.006850052159279585, + 0.008539446629583836, + 0.007372693158686161, + 0.00884043239057064, + 0.008807743899524212, + 0.011413308791816235, + 0.01371894683688879, + 0.017580615356564522, + 0.013351055793464184, + 0.016655677929520607, + 0.02168184332549572, + 0.014247697778046131, + 0.0075800446793437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.007634628098458052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197559155523777 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.003197559155523777, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197559155523777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15847675502300262, + 0.08060041069984436, + 0.21153217554092407, + 0.05836781859397888, + 0.09937437623739243, + 0.10246540606021881, + 0.07008372247219086, + 0.08728255331516266, + 0.054216042160987854, + 0.07760076969861984 + ], + "credit_map": [ + 0.014195005409419537, + 0.017958596348762512, + 0.021714866161346436, + 0.02147996798157692, + 0.033126350492239, + 0.01421349123120308, + 0.022217512130737305, + 0.013570956885814667, + 0.012175781652331352, + 0.007182650733739138, + 0.01093143317848444, + 0.008112117648124695, + 0.015320531092584133, + 0.0103294737637043, + 0.008991681970655918, + 0.007556728553026915, + 0.026417579501867294, + 0.023479899391531944, + 0.03183330222964287, + 0.02302779071033001, + 0.028488028794527054, + 0.03505943343043327, + 0.020572291687130928, + 0.02265382558107376, + 0.007033295929431915, + 0.0049056983552873135, + 0.0084568215534091, + 0.007485464680939913, + 0.005811214447021484, + 0.010953297838568687, + 0.006098553538322449, + 0.0076234666630625725, + 0.013240592554211617, + 0.006001516245305538, + 0.01332750916481018, + 0.00710960803553462, + 0.014235583133995533, + 0.022968146950006485, + 0.005773127544671297, + 0.01671828329563141, + 0.014114722609519958, + 0.01415302138775587, + 0.015722645446658134, + 0.011895280331373215, + 0.011246713809669018, + 0.013593517243862152, + 0.00798210222274065, + 0.013757405802607536, + 0.010800876654684544, + 0.008684555999934673, + 0.010002920404076576, + 0.008134978823363781, + 0.009128611534833908, + 0.006678737234324217, + 0.004572731908410788, + 0.012080298736691475, + 0.01767789013683796, + 0.008730234578251839, + 0.01363515853881836, + 0.0051918067038059235, + 0.011815235950052738, + 0.007351277396082878, + 0.006540547125041485, + 0.01634039916098118, + 0.009519434534013271, + 0.005059501156210899, + 0.007580540142953396, + 0.003909290302544832, + 0.006719384808093309, + 0.004987282212823629, + 0.009651340544223785, + 0.006789260543882847, + 0.009576202370226383, + 0.008374998345971107, + 0.012178133241832256, + 0.009291430935263634, + 0.01214041281491518, + 0.007040831726044416, + 0.01222926378250122, + 0.00676949555054307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.003197559155523777 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031399130821228 + ], + "opposite_logprob_tokens": [ + -2.350313901901245 + ], + "logprob_reward": -0.10031399130821228, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031399130821228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15086673200130463, + 0.05437924712896347, + 0.055728960782289505, + 0.06669743359088898, + 0.14191506803035736, + 0.08221518993377686, + 0.14276190102100372, + 0.0876198336482048, + 0.1454474776983261, + 0.07236812263727188 + ], + "credit_map": [ + 0.015969648957252502, + 0.023976875469088554, + 0.023611286655068398, + 0.01711522601544857, + 0.03115370310842991, + 0.015784669667482376, + 0.014316652901470661, + 0.008938678540289402, + 0.00657622842118144, + 0.0068908450193703175, + 0.007674812339246273, + 0.005158658139407635, + 0.013194331899285316, + 0.005860981997102499, + 0.004434007219970226, + 0.004589382093399763, + 0.005995095707476139, + 0.00650778878480196, + 0.008235503919422626, + 0.005049875937402248, + 0.010215666145086288, + 0.005526726599782705, + 0.006310130935162306, + 0.00788817834109068, + 0.006267703603953123, + 0.0064562284387648106, + 0.007291340734809637, + 0.017316406592726707, + 0.007851976901292801, + 0.0076065645553171635, + 0.00740443728864193, + 0.006502770818769932, + 0.020805880427360535, + 0.011869137175381184, + 0.023932240903377533, + 0.013924980536103249, + 0.017011160030961037, + 0.024688517674803734, + 0.006983546074479818, + 0.022699618712067604, + 0.009633391164243221, + 0.010682415217161179, + 0.012998169288039207, + 0.008508365601301193, + 0.00962564442306757, + 0.014312620274722576, + 0.0064036184921860695, + 0.01005097571760416, + 0.02460705116391182, + 0.01259431429207325, + 0.02167642116546631, + 0.017152173444628716, + 0.015603410080075264, + 0.014848427847027779, + 0.007939456030726433, + 0.02834065817296505, + 0.01791349984705448, + 0.007358500733971596, + 0.012859128415584564, + 0.004282666835933924, + 0.010447540320456028, + 0.006061083171516657, + 0.004445122089236975, + 0.024252304807305336, + 0.02360221929848194, + 0.012736411765217781, + 0.022930879145860672, + 0.011833056807518005, + 0.018361374735832214, + 0.01849822700023651, + 0.019700651988387108, + 0.01778465509414673, + 0.008587467484176159, + 0.008925600908696651, + 0.010376480408012867, + 0.011140018701553345, + 0.009869515895843506, + 0.00949291791766882, + 0.009032683447003365, + 0.0049434443935751915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.10031399130821228 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, forward-moving close-up, framed by the dramatic contrast of classic black and white light and shadow, precisely captures a dancer twirling gracefully and seamlessly. Her hair is styled in a sleek, delicate ballet bun, a few strands of sweat-soaked black hair freed by the centrifugal force of the spin, gently brushing against her closed eyes and high nose. Her head tilts back slightly in time with the twirl, her jawline gracefully tensed, a faint, utterly captivated smile playing at the corners of her mouth, as if completely absorbed in the silent melody. A powerful spotlight from above and side, like Rembrandt's brush, sculpts the three-dimensional contours of her face, illuminating the curve of her cheekbones and her sweat-beaded forehead in the highlights while casting deep, soft shadows on the other side, intensifying her intense concentration and selflessness in her art. Her simple dark dance dress reveals a silky luster under the strong light, forming an extreme contrast with her fair skin. The whole picture presents a sculptural sense of tranquility in the dynamic, solidifying the momentary passion and emotion into a timeless portrait full of classical aesthetics and inner monologue.", + "seed": 11395, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_12/index_9961_seed_11395_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03018042529001832, + "vlm_logprob": -0.03018042529001832, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03018042529001832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05863361433148384 + ], + "opposite_logprob_tokens": [ + -3.30863356590271 + ], + "logprob_reward": -0.05863361433148384, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05863361433148384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08205905556678772, + 0.1325398087501526, + 0.05526981130242348, + 0.09454701095819473, + 0.0548742301762104, + 0.11618014425039291, + 0.05564242973923683, + 0.14074870944023132, + 0.09189064800739288, + 0.17624808847904205 + ], + "credit_map": [ + 0.006868154741823673, + 0.009029948152601719, + 0.007004891987890005, + 0.019632477313280106, + 0.01084140408784151, + 0.013653362169861794, + 0.007040177937597036, + 0.00798864383250475, + 0.013417706824839115, + 0.026454145088791847, + 0.018799109384417534, + 0.026612380519509315, + 0.011877595447003841, + 0.01359068974852562, + 0.008080190978944302, + 0.013708007521927357, + 0.004686505533754826, + 0.0058404989540576935, + 0.007626038510352373, + 0.008598990738391876, + 0.010515094734728336, + 0.007303076330572367, + 0.004270302597433329, + 0.006429305765777826, + 0.010190490633249283, + 0.012766317464411259, + 0.013393886387348175, + 0.02062513865530491, + 0.012091659009456635, + 0.0073392875492572784, + 0.00811428390443325, + 0.010025957599282265, + 0.007988594472408295, + 0.00763722974807024, + 0.0066643040627241135, + 0.008042385801672935, + 0.008616112172603607, + 0.003945894539356232, + 0.005615567322820425, + 0.006364141590893269, + 0.01630755327641964, + 0.018917275592684746, + 0.015822630375623703, + 0.009609620086848736, + 0.020388873293995857, + 0.014615518972277641, + 0.011567419394850731, + 0.008951261639595032, + 0.00695871701464057, + 0.009348365478217602, + 0.007849992252886295, + 0.009955340065062046, + 0.006685628090053797, + 0.005698733497411013, + 0.004899017978459597, + 0.004246636759489775, + 0.016673561185598373, + 0.02091534063220024, + 0.018681177869439125, + 0.044933900237083435, + 0.009523984976112843, + 0.012518218718469143, + 0.005870734807103872, + 0.011631814762949944, + 0.01650300621986389, + 0.008749619126319885, + 0.013972176238894463, + 0.010380974970757961, + 0.01140176597982645, + 0.011836580000817776, + 0.007474689744412899, + 0.011571847833693027, + 0.035399261862039566, + 0.015699369832873344, + 0.025480112060904503, + 0.021575499325990677, + 0.011592898517847061, + 0.027829239144921303, + 0.012687276117503643, + 0.025984441861510277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.05863361433148384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061687409877777 + ], + "opposite_logprob_tokens": [ + -3.8956167697906494 + ], + "logprob_reward": -0.02061687409877777, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061687409877777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22428476810455322, + 0.20283041894435883, + 0.1319809854030609, + 0.1063687801361084, + 0.022605592384934425, + 0.025239720940589905, + 0.04150669649243355, + 0.07734230905771255, + 0.04007554426789284, + 0.12776517868041992 + ], + "credit_map": [ + 0.02386372722685337, + 0.02000396139919758, + 0.020244238898158073, + 0.03722105175256729, + 0.01768025942146778, + 0.03684680163860321, + 0.025906627997756004, + 0.04251810535788536, + 0.017730817198753357, + 0.04386149346828461, + 0.02674289420247078, + 0.04687397927045822, + 0.013813233934342861, + 0.020878298208117485, + 0.011347176507115364, + 0.021582525223493576, + 0.023615041747689247, + 0.02487228810787201, + 0.024449698626995087, + 0.017780795693397522, + 0.015780409798026085, + 0.00799684226512909, + 0.00842056144028902, + 0.009065349586308002, + 0.011884096078574657, + 0.01721871644258499, + 0.015274733304977417, + 0.02477925457060337, + 0.009823402389883995, + 0.005907201673835516, + 0.006802845746278763, + 0.01467853132635355, + 0.003592649707570672, + 0.0027597949374467134, + 0.00388721632771194, + 0.003826409811154008, + 0.0030720080249011517, + 0.001637724693864584, + 0.0021938011050224304, + 0.001635987893678248, + 0.0037348794285207987, + 0.004312308505177498, + 0.0032631494104862213, + 0.002808133838698268, + 0.00412690220400691, + 0.0018410492921248078, + 0.002122838981449604, + 0.0030304589308798313, + 0.004214688669890165, + 0.004571207799017429, + 0.005367347039282322, + 0.004390442278236151, + 0.007953844964504242, + 0.004252358805388212, + 0.007158080581575632, + 0.0035987284500151873, + 0.0077485740184783936, + 0.01211608573794365, + 0.008695350028574467, + 0.025163106620311737, + 0.00483194924890995, + 0.007525185588747263, + 0.00436062179505825, + 0.006901433225721121, + 0.005905980709940195, + 0.0029075867496430874, + 0.005854773335158825, + 0.0034311176277697086, + 0.0026897145435214043, + 0.009391942992806435, + 0.0021784109994769096, + 0.007716019172221422, + 0.01887674443423748, + 0.008346148766577244, + 0.028674159198999405, + 0.013651843182742596, + 0.0061334045603871346, + 0.024111082777380943, + 0.005475555546581745, + 0.022496242076158524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02061687409877777 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421081367880106 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.01421081367880106, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421081367880106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1901807188987732, + 0.14025640487670898, + 0.0727953091263771, + 0.08952530473470688, + 0.07125109434127808, + 0.0675092414021492, + 0.050748515874147415, + 0.08698886632919312, + 0.07825399935245514, + 0.1524905413389206 + ], + "credit_map": [ + 0.01579149253666401, + 0.018298804759979248, + 0.01942746341228485, + 0.03260939195752144, + 0.028921490535140038, + 0.02617303654551506, + 0.02952020801603794, + 0.019438831135630608, + 0.013188284821808338, + 0.031010597944259644, + 0.019075771793723106, + 0.03177424147725105, + 0.010056547820568085, + 0.014530162326991558, + 0.007529145106673241, + 0.013091660104691982, + 0.013557751663029194, + 0.01237523928284645, + 0.012008892372250557, + 0.009895337745547295, + 0.008768109604716301, + 0.004869233351200819, + 0.004377912729978561, + 0.00694283377379179, + 0.008717862889170647, + 0.013998141512274742, + 0.013701051473617554, + 0.020665308460593224, + 0.01081615500152111, + 0.005428739357739687, + 0.006145513150840998, + 0.010052536614239216, + 0.008216919377446175, + 0.008995473384857178, + 0.009109036065638065, + 0.009692381136119366, + 0.011799636296927929, + 0.005180289037525654, + 0.008828485384583473, + 0.009428867138922215, + 0.00926907267421484, + 0.017151057720184326, + 0.009374452754855156, + 0.007170084398239851, + 0.009982092306017876, + 0.005320429801940918, + 0.004499887116253376, + 0.004742164630442858, + 0.006134514231234789, + 0.011861502192914486, + 0.005986405536532402, + 0.0060646687634289265, + 0.004958535078912973, + 0.006130286492407322, + 0.004586456809192896, + 0.005026147235184908, + 0.0087293591350317, + 0.01370972953736782, + 0.010303227230906487, + 0.019386496394872665, + 0.0077805896289646626, + 0.011421706527471542, + 0.006304266396909952, + 0.009353497996926308, + 0.0070014167577028275, + 0.005327572114765644, + 0.009322280064225197, + 0.00857935007661581, + 0.006405647378414869, + 0.02008725516498089, + 0.005162084475159645, + 0.01636839285492897, + 0.020768040791153908, + 0.011781057342886925, + 0.023382816463708878, + 0.014365175738930702, + 0.0072911386378109455, + 0.03341985121369362, + 0.0080658458173275, + 0.03341660648584366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01421081367880106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491943072527647 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.002491943072527647, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491943072527647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2262224406003952, + 0.04094700515270233, + 0.05261583998799324, + 0.16014127433300018, + 0.05027679726481438, + 0.03995411470532417, + 0.057978883385658264, + 0.06467639654874802, + 0.05447450280189514, + 0.25271281599998474 + ], + "credit_map": [ + 0.02668842114508152, + 0.035405974835157394, + 0.023811649531126022, + 0.03566363826394081, + 0.021728627383708954, + 0.03487831726670265, + 0.0215495303273201, + 0.0264962799847126, + 0.002879936248064041, + 0.005814885720610619, + 0.005602304358035326, + 0.010778662748634815, + 0.003995444159954786, + 0.004363252781331539, + 0.0029056984931230545, + 0.004606818780303001, + 0.009147651493549347, + 0.010679478757083416, + 0.01179412566125393, + 0.008112650364637375, + 0.004002479836344719, + 0.002778700552880764, + 0.00255984696559608, + 0.0035409058909863234, + 0.021841244772076607, + 0.02371506206691265, + 0.02247287705540657, + 0.03722183033823967, + 0.014744959771633148, + 0.009991827420890331, + 0.010232635773718357, + 0.01992083340883255, + 0.008398338221013546, + 0.005093947984278202, + 0.00776250846683979, + 0.00904477946460247, + 0.006561452057212591, + 0.0032757436856627464, + 0.0051071662455797195, + 0.005032861605286598, + 0.006652454845607281, + 0.0071500311605632305, + 0.004750580061227083, + 0.00409359997138381, + 0.0058855656534433365, + 0.0030931863002479076, + 0.0029823470395058393, + 0.00534634618088603, + 0.007526015862822533, + 0.008997367694973946, + 0.007728096097707748, + 0.007059548050165176, + 0.007299334742128849, + 0.0069902073591947556, + 0.007159688975661993, + 0.005218619480729103, + 0.006862279959022999, + 0.010667538270354271, + 0.006664114072918892, + 0.020283794030547142, + 0.00392313115298748, + 0.007100010756403208, + 0.003231756156310439, + 0.00594377052038908, + 0.009426672011613846, + 0.004852199461311102, + 0.009967347607016563, + 0.0056854309514164925, + 0.002167344558984041, + 0.012001832947134972, + 0.0019109462155029178, + 0.008462732657790184, + 0.03766108676791191, + 0.018071459606289864, + 0.05719292536377907, + 0.02585715427994728, + 0.013541505672037601, + 0.0464978851377964, + 0.010228097438812256, + 0.04366269335150719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.002491943072527647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494888126850128 + ], + "opposite_logprob_tokens": [ + -2.9299488067626953 + ], + "logprob_reward": -0.05494888126850128, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494888126850128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0781775712966919, + 0.12493434548377991, + 0.057875022292137146, + 0.220791295170784, + 0.0601259246468544, + 0.05844448134303093, + 0.10678741335868835, + 0.11603854596614838, + 0.05838416889309883, + 0.11844121664762497 + ], + "credit_map": [ + 0.007754548452794552, + 0.008648055605590343, + 0.008193081244826317, + 0.011516044847667217, + 0.013345777057111263, + 0.011491108685731888, + 0.008825846016407013, + 0.0084031131118536, + 0.01276172325015068, + 0.024976080283522606, + 0.018112698569893837, + 0.02510342188179493, + 0.008310239762067795, + 0.013861384242773056, + 0.006874775979667902, + 0.014934021048247814, + 0.005344593897461891, + 0.0075721233151853085, + 0.008485565893352032, + 0.010606086812913418, + 0.00692377332597971, + 0.007015334442257881, + 0.003321938682347536, + 0.00860560778528452, + 0.024077938869595528, + 0.03379063308238983, + 0.03317036107182503, + 0.04844912886619568, + 0.027215568348765373, + 0.014115648344159126, + 0.014434397220611572, + 0.025537613779306412, + 0.006504574790596962, + 0.006839780602604151, + 0.007088785525411367, + 0.010030951350927353, + 0.008961118757724762, + 0.0034453903790563345, + 0.010027706623077393, + 0.007227618712931871, + 0.009887590073049068, + 0.00832606852054596, + 0.007355138659477234, + 0.005595176015049219, + 0.01184272300451994, + 0.003556292038410902, + 0.006293341983109713, + 0.00558815011754632, + 0.012329540215432644, + 0.018680773675441742, + 0.014917876571416855, + 0.009834918193519115, + 0.017574336379766464, + 0.01100985985249281, + 0.012655665166676044, + 0.009784446097910404, + 0.010992640629410744, + 0.01585850864648819, + 0.019313868135213852, + 0.032977983355522156, + 0.006978080607950687, + 0.014251542277634144, + 0.007633796893060207, + 0.008032131940126419, + 0.004641109611839056, + 0.006983791012316942, + 0.008645324967801571, + 0.009531502611935139, + 0.0049290284514427185, + 0.010171891190111637, + 0.006040810141712427, + 0.007440710905939341, + 0.01664203777909279, + 0.009027394466102123, + 0.018897367641329765, + 0.011926849372684956, + 0.005927570629864931, + 0.024223988875746727, + 0.007680768612772226, + 0.024115247651934624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.05494888126850128 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, forward-moving close-up, framed by the dramatic contrast of classic black and white light and shadow, precisely captures a dancer twirling gracefully and seamlessly. Her hair is styled in a sleek, delicate ballet bun, a few strands of sweat-soaked black hair freed by the centrifugal force of the spin, gently brushing against her closed eyes and high nose. Her head tilts back slightly in time with the twirl, her jawline gracefully tensed, a faint, utterly captivated smile playing at the corners of her mouth, as if completely absorbed in the silent melody. A powerful spotlight from above and side, like Rembrandt's brush, sculpts the three-dimensional contours of her face, illuminating the curve of her cheekbones and her sweat-beaded forehead in the highlights while casting deep, soft shadows on the other side, intensifying her intense concentration and selflessness in her art. Her simple dark dance dress reveals a silky luster under the strong light, forming an extreme contrast with her fair skin. The whole picture presents a sculptural sense of tranquility in the dynamic, solidifying the momentary passion and emotion into a timeless portrait full of classical aesthetics and inner monologue.", + "seed": 11396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_12/index_9961_seed_11396_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03543767655501142, + "vlm_logprob": -0.03543767655501142, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03543767655501142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11470354348421097 + ], + "opposite_logprob_tokens": [ + -2.48970365524292 + ], + "logprob_reward": -0.11470354348421097, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11470354348421097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057942844927310944, + 0.02953570894896984, + 0.022988416254520416, + 0.07129174470901489, + 0.11792340874671936, + 0.14765098690986633, + 0.1563352644443512, + 0.09986381977796555, + 0.26478341221809387, + 0.03168429434299469 + ], + "credit_map": [ + 0.00781114399433136, + 0.007368489168584347, + 0.0071968198753893375, + 0.007740927394479513, + 0.0059783379547297955, + 0.007291322574019432, + 0.006678418722003698, + 0.007877389900386333, + 0.003732299664989114, + 0.003741211025044322, + 0.004367123357951641, + 0.0029976600781083107, + 0.003139032516628504, + 0.0041703153401613235, + 0.003891815897077322, + 0.0034962520003318787, + 0.002990865148603916, + 0.0028495430015027523, + 0.003530737478286028, + 0.0027258596383035183, + 0.002357194200158119, + 0.002751880558207631, + 0.003306297119706869, + 0.0024760409723967314, + 0.009637708775699139, + 0.0072465864941477776, + 0.008852158673107624, + 0.010793143883347511, + 0.009003612212836742, + 0.00879259966313839, + 0.009595002047717571, + 0.007370939012616873, + 0.010939860716462135, + 0.010751605965197086, + 0.009829750284552574, + 0.008468292653560638, + 0.012798148207366467, + 0.025278670713305473, + 0.008908028714358807, + 0.030949067324399948, + 0.014646751806139946, + 0.006271030753850937, + 0.008905838243663311, + 0.00619741203263402, + 0.012999310158193111, + 0.03983084857463837, + 0.006460370030254126, + 0.05233944579958916, + 0.007949433289468288, + 0.009223041124641895, + 0.009513730183243752, + 0.007354154251515865, + 0.017208879813551903, + 0.05042554810643196, + 0.008068178780376911, + 0.046592310070991516, + 0.007759074214845896, + 0.009363044984638691, + 0.007819433696568012, + 0.004653103183954954, + 0.02355733886361122, + 0.019236749038100243, + 0.010630209930241108, + 0.01684488169848919, + 0.012699530459940434, + 0.0112483324483037, + 0.01952599734067917, + 0.00829357746988535, + 0.007555915508419275, + 0.10691937059164047, + 0.010798455215990543, + 0.08774228394031525, + 0.0036060851998627186, + 0.006570939440280199, + 0.005181693006306887, + 0.0030659271869808435, + 0.0027874798979610205, + 0.003802300663664937, + 0.002758412854745984, + 0.0039114574901759624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.11470354348421097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813884034752846 + ], + "opposite_logprob_tokens": [ + -3.2881388664245605 + ], + "logprob_reward": -0.03813884034752846, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813884034752846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03649108111858368, + 0.03362097591161728, + 0.03647318109869957, + 0.07859933376312256, + 0.10827197134494781, + 0.24673913419246674, + 0.044401615858078, + 0.033055808395147324, + 0.2951357662677765, + 0.08721114695072174 + ], + "credit_map": [ + 0.005635179579257965, + 0.0039129299111664295, + 0.00500975176692009, + 0.005045944359153509, + 0.00345911318436265, + 0.0046683973632752895, + 0.004028716124594212, + 0.004731049295514822, + 0.004659177735447884, + 0.0034332331269979477, + 0.006071276962757111, + 0.0028312811627984047, + 0.0035515702329576015, + 0.004676053300499916, + 0.0034932431299239397, + 0.004905139096081257, + 0.0035464800894260406, + 0.004663464613258839, + 0.004046646878123283, + 0.003720664419233799, + 0.005750463344156742, + 0.004623910412192345, + 0.004318667110055685, + 0.005802887491881847, + 0.0076850163750350475, + 0.009127609431743622, + 0.012090860866010189, + 0.011420856229960918, + 0.012075335718691349, + 0.010165954940021038, + 0.00684327632188797, + 0.009190422482788563, + 0.008927224203944206, + 0.013799422420561314, + 0.010577056556940079, + 0.010773600079119205, + 0.014421980828046799, + 0.019853439182043076, + 0.008246892131865025, + 0.021672356873750687, + 0.024336792528629303, + 0.011120286770164967, + 0.014317043125629425, + 0.010275560431182384, + 0.02259439416229725, + 0.06611236929893494, + 0.011757842265069485, + 0.08622485399246216, + 0.0012375429505482316, + 0.001501822960563004, + 0.0016203142004087567, + 0.0009913401445373893, + 0.0016936742467805743, + 0.0182865709066391, + 0.0012424279702827334, + 0.017827922478318214, + 0.0021612111013382673, + 0.00241659814491868, + 0.0022681488189846277, + 0.0011082204291597009, + 0.0016783942701295018, + 0.01067162211984396, + 0.0018486372428014874, + 0.01090297382324934, + 0.012993033975362778, + 0.007506232243031263, + 0.01830601692199707, + 0.006860223598778248, + 0.005977501627057791, + 0.12682588398456573, + 0.009629784151911736, + 0.10703711211681366, + 0.010060785338282585, + 0.016763383522629738, + 0.013172090984880924, + 0.010008272714912891, + 0.010504347272217274, + 0.010814648121595383, + 0.008322002366185188, + 0.00756561616435647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.03813884034752846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638413459062576 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.007638413459062576, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638413459062576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03239087387919426, + 0.019085409119725227, + 0.012671154923737049, + 0.018100643530488014, + 0.037554752081632614, + 0.05391407012939453, + 0.1437043696641922, + 0.02528253011405468, + 0.6294085383415222, + 0.027887631207704544 + ], + "credit_map": [ + 0.0034555441234260798, + 0.004660061094909906, + 0.0033467304892838, + 0.0038955514319241047, + 0.004280552733689547, + 0.004000801593065262, + 0.005122937727719545, + 0.003628693986684084, + 0.0019615215715020895, + 0.0022865855135023594, + 0.0030553743708878756, + 0.001951727899722755, + 0.0028621815145015717, + 0.0026033404283225536, + 0.0023325083311647177, + 0.002032169373705983, + 0.0011860976228490472, + 0.0015692802844569087, + 0.0016698422841727734, + 0.001285111764445901, + 0.001858323230408132, + 0.0019189822487533092, + 0.0017011500895023346, + 0.001482367399148643, + 0.0023688466753810644, + 0.0016954614548012614, + 0.0030899494886398315, + 0.0019117376068606973, + 0.002491673920303583, + 0.0025559188798069954, + 0.0014228769578039646, + 0.0025641783140599728, + 0.002919195219874382, + 0.002974680159240961, + 0.003656758926808834, + 0.004458736162632704, + 0.005051506217569113, + 0.0068350015208125114, + 0.003729037009179592, + 0.007929838262498379, + 0.005127848591655493, + 0.003131718374788761, + 0.0034245564602315426, + 0.0028842235915362835, + 0.0047260127030313015, + 0.013496740721166134, + 0.0026133169885724783, + 0.018509656190872192, + 0.004663609899580479, + 0.004300704225897789, + 0.0040641543455421925, + 0.00197526253759861, + 0.006630622316151857, + 0.06009645015001297, + 0.003644379787147045, + 0.05832918360829353, + 0.0017009356524795294, + 0.0032885188702493906, + 0.002548281801864505, + 0.002053014701232314, + 0.004609102848917246, + 0.004176621325314045, + 0.0022105819080024958, + 0.004695470444858074, + 0.02084175869822502, + 0.01126435399055481, + 0.033215489238500595, + 0.008283544331789017, + 0.010119544342160225, + 0.2831605076789856, + 0.011295139789581299, + 0.2512281537055969, + 0.0031479469034820795, + 0.0047434172593057156, + 0.00345298508182168, + 0.004492816515266895, + 0.004368910100311041, + 0.002565171802416444, + 0.0022114983294159174, + 0.0029048847500234842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.007638413459062576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006381620769388974 + ], + "opposite_logprob_tokens": [ + -7.375638008117676 + ], + "logprob_reward": -0.0006381620769388974, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006381620769388974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029941318556666374, + 0.015053687617182732, + 0.01828540302813053, + 0.032527320086956024, + 0.03266489878296852, + 0.37812548875808716, + 0.04338942840695381, + 0.05063018575310707, + 0.3770453929901123, + 0.022336890920996666 + ], + "credit_map": [ + 0.0037273543421179056, + 0.0023564291186630726, + 0.0047448365949094296, + 0.0036996074486523867, + 0.00223443447612226, + 0.0044156815856695175, + 0.003336714580655098, + 0.005426257383078337, + 0.0018546351930126548, + 0.0015916206175461411, + 0.002901126630604267, + 0.0012033400125801563, + 0.0013927785912528634, + 0.0024156244471669197, + 0.0016452607233077288, + 0.0020493001211434603, + 0.0021598804742097855, + 0.0027864957228302956, + 0.0017643322935327888, + 0.001780594582669437, + 0.0032209332566708326, + 0.002199273556470871, + 0.0024471490178257227, + 0.0019267419120296836, + 0.004068667534738779, + 0.002793138613924384, + 0.005180779378861189, + 0.004884785041213036, + 0.0035363531205803156, + 0.004109380301088095, + 0.004115470219403505, + 0.0038387442473322153, + 0.003442233894020319, + 0.003954228945076466, + 0.004308221396058798, + 0.0037907420191913843, + 0.005499782972037792, + 0.004833662416785955, + 0.0023649539798498154, + 0.004471068736165762, + 0.03708542510867119, + 0.016994645819067955, + 0.022108018398284912, + 0.016259469091892242, + 0.033967431634664536, + 0.10139498114585876, + 0.01781170442700386, + 0.13250376284122467, + 0.003160044550895691, + 0.004072561860084534, + 0.0025757018011063337, + 0.0024316217750310898, + 0.009736244566738605, + 0.008524705655872822, + 0.004175081849098206, + 0.00871346052736044, + 0.0028180615045130253, + 0.0034791429061442614, + 0.0036021836567670107, + 0.002954500261694193, + 0.005644435063004494, + 0.013886609114706516, + 0.0035414472222328186, + 0.014703800901770592, + 0.012207035906612873, + 0.00674565089866519, + 0.02065073326230049, + 0.0051570008508861065, + 0.005289012566208839, + 0.16987088322639465, + 0.0068235998041927814, + 0.1503014713525772, + 0.0027088436763733625, + 0.003382638329640031, + 0.003923311363905668, + 0.001397483516484499, + 0.0033306905534118414, + 0.003840807592496276, + 0.0015118289738893509, + 0.0022412848193198442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.0006381620769388974 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069423407316208 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069423407316208, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069423407316208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03891546651721001, + 0.024615895003080368, + 0.015935897827148438, + 0.02707013487815857, + 0.02468881942331791, + 0.03528458625078201, + 0.37096938490867615, + 0.05997259169816971, + 0.3841720521450043, + 0.018375247716903687 + ], + "credit_map": [ + 0.003871228778734803, + 0.004775735549628735, + 0.0040718126110732555, + 0.004082861356437206, + 0.004960423335433006, + 0.004658454563468695, + 0.00776259321719408, + 0.0047323559410870075, + 0.003015172202140093, + 0.0031562098301947117, + 0.004985315725207329, + 0.0019844258204102516, + 0.002970379078760743, + 0.003585439408197999, + 0.002095635049045086, + 0.002823315793648362, + 0.001484899315983057, + 0.0024277677293866873, + 0.002869744785130024, + 0.0016502737998962402, + 0.002141332719475031, + 0.002389857778325677, + 0.0014198299031704664, + 0.0015521912137046456, + 0.003202781779691577, + 0.0028659712988883257, + 0.004265244584530592, + 0.00432745274156332, + 0.003060033544898033, + 0.00336822378449142, + 0.003198199439793825, + 0.0027822263073176146, + 0.0018269260181114078, + 0.003272139234468341, + 0.0030569916125386953, + 0.0025364563334733248, + 0.0032582839485257864, + 0.004203240852802992, + 0.0018044828902930021, + 0.00473029725253582, + 0.0031400034204125404, + 0.003009582869708538, + 0.004489226266741753, + 0.002696469658985734, + 0.006572078913450241, + 0.005618670489639044, + 0.0033602179028093815, + 0.006398335564881563, + 0.01745101995766163, + 0.017259756103157997, + 0.012447423301637173, + 0.009460330940783024, + 0.03932498022913933, + 0.12875688076019287, + 0.017752094194293022, + 0.12851686775684357, + 0.0029949264135211706, + 0.004072884563356638, + 0.003872531931847334, + 0.00292004831135273, + 0.004480193834751844, + 0.018759651109576225, + 0.0037481181789189577, + 0.0191242303699255, + 0.011980956420302391, + 0.011543845757842064, + 0.019978441298007965, + 0.008152141235768795, + 0.008848704397678375, + 0.16444608569145203, + 0.007859719917178154, + 0.15136213600635529, + 0.0022100042551755905, + 0.0031934352591633797, + 0.0029353706631809473, + 0.0018509376095607877, + 0.00211484357714653, + 0.0023138420656323433, + 0.0018175146542489529, + 0.0019392974209040403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.016069423407316208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, forward-moving close-up, framed by the dramatic contrast of classic black and white light and shadow, precisely captures a dancer twirling gracefully and seamlessly. Her hair is styled in a sleek, delicate ballet bun, a few strands of sweat-soaked black hair freed by the centrifugal force of the spin, gently brushing against her closed eyes and high nose. Her head tilts back slightly in time with the twirl, her jawline gracefully tensed, a faint, utterly captivated smile playing at the corners of her mouth, as if completely absorbed in the silent melody. A powerful spotlight from above and side, like Rembrandt's brush, sculpts the three-dimensional contours of her face, illuminating the curve of her cheekbones and her sweat-beaded forehead in the highlights while casting deep, soft shadows on the other side, intensifying her intense concentration and selflessness in her art. Her simple dark dance dress reveals a silky luster under the strong light, forming an extreme contrast with her fair skin. The whole picture presents a sculptural sense of tranquility in the dynamic, solidifying the momentary passion and emotion into a timeless portrait full of classical aesthetics and inner monologue.", + "seed": 11397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_12/index_9961_seed_11397_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04350502979941666, + "vlm_logprob": -0.04350502979941666, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04350502979941666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12618102133274078 + ], + "opposite_logprob_tokens": [ + -2.501181125640869 + ], + "logprob_reward": -0.12618102133274078, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.12618102133274078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11605077236890793, + 0.08838514238595963, + 0.10573186725378036, + 0.050833143293857574, + 0.05317306146025658, + 0.03433077037334442, + 0.06388824433088303, + 0.0201200470328331, + 0.4221528470516205, + 0.04533403366804123 + ], + "credit_map": [ + 0.013511287048459053, + 0.01679445058107376, + 0.016794120892882347, + 0.009242544881999493, + 0.013836950995028019, + 0.019929615780711174, + 0.016379304230213165, + 0.009562499821186066, + 0.006009423173964024, + 0.007352846674621105, + 0.023144008591771126, + 0.007811776828020811, + 0.0059876637533307076, + 0.024154847487807274, + 0.006900403182953596, + 0.007024175953119993, + 0.02245614118874073, + 0.010039497166872025, + 0.020534943789243698, + 0.008732998743653297, + 0.006825381889939308, + 0.015101438388228416, + 0.010916301049292088, + 0.011125161312520504, + 0.005610190331935883, + 0.005515864584594965, + 0.005127699580043554, + 0.004352792166173458, + 0.005689849145710468, + 0.007839000783860683, + 0.006512316409498453, + 0.010185430757701397, + 0.004496052861213684, + 0.0029619161505252123, + 0.010955785401165485, + 0.0067565045319497585, + 0.00439161341637373, + 0.012389314360916615, + 0.005752928089350462, + 0.005468948278576136, + 0.00359452236443758, + 0.003300681244581938, + 0.005837404169142246, + 0.007409135345369577, + 0.003436815459281206, + 0.0034820563159883022, + 0.003473317250609398, + 0.0037968375254422426, + 0.005359193310141563, + 0.005451153498142958, + 0.008060267195105553, + 0.008615400642156601, + 0.010523769073188305, + 0.005434887483716011, + 0.009248237125575542, + 0.011195337399840355, + 0.002535902662202716, + 0.0025212347973138094, + 0.002736241789534688, + 0.0020909386221319437, + 0.003509090980514884, + 0.002080021658912301, + 0.0021551051177084446, + 0.0024915109388530254, + 0.03267863020300865, + 0.04406645894050598, + 0.06455014646053314, + 0.041698552668094635, + 0.0579938106238842, + 0.06790890544652939, + 0.05082792788743973, + 0.062428414821624756, + 0.004907193128019571, + 0.0030529324430972338, + 0.009685800410807133, + 0.004166628699749708, + 0.0036391287576407194, + 0.01029680110514164, + 0.003204986220225692, + 0.006380561273545027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.12618102133274078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048701900988817215 + ], + "opposite_logprob_tokens": [ + -3.0487020015716553 + ], + "logprob_reward": -0.048701900988817215, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048701900988817215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4743533432483673, + 0.07284997403621674, + 0.054057080298662186, + 0.03371646627783775, + 0.022998636588454247, + 0.09329267591238022, + 0.03830365464091301, + 0.10674846172332764, + 0.06713961064815521, + 0.036540113389492035 + ], + "credit_map": [ + 0.04781915619969368, + 0.07026904821395874, + 0.041962333023548126, + 0.0646444708108902, + 0.05840173363685608, + 0.05573677644133568, + 0.09290891140699387, + 0.04261090233922005, + 0.007470590528100729, + 0.008107616566121578, + 0.008313040249049664, + 0.010702008381485939, + 0.00836147554218769, + 0.009813130833208561, + 0.010765383951365948, + 0.009316726587712765, + 0.005898992996662855, + 0.007314861286431551, + 0.006932067219167948, + 0.004460242111235857, + 0.005425096023827791, + 0.007386641576886177, + 0.008618110790848732, + 0.008021067827939987, + 0.004727975931018591, + 0.004364171531051397, + 0.0034294137731194496, + 0.004665371496230364, + 0.002693043788895011, + 0.00483895605430007, + 0.004847649950534105, + 0.004149885382503271, + 0.0023832202423363924, + 0.0013314362149685621, + 0.003785220440477133, + 0.002279209904372692, + 0.0018171916017308831, + 0.003967212978750467, + 0.003535590833052993, + 0.0038995547220110893, + 0.009003550745546818, + 0.008384285494685173, + 0.01775786280632019, + 0.009039015509188175, + 0.01731187105178833, + 0.01152681652456522, + 0.010594004765152931, + 0.00967527274042368, + 0.0036371781025081873, + 0.002222054172307253, + 0.006215870846062899, + 0.006083318963646889, + 0.004804648458957672, + 0.004484774544835091, + 0.004337962716817856, + 0.006517848465591669, + 0.01328320149332285, + 0.01889892853796482, + 0.014350347220897675, + 0.00860487762838602, + 0.013410324230790138, + 0.010744627565145493, + 0.012824121862649918, + 0.014632024802267551, + 0.006565013434737921, + 0.007594110909849405, + 0.01042625866830349, + 0.007783889304846525, + 0.0040010190568864346, + 0.012260536663234234, + 0.009604516439139843, + 0.00890426803380251, + 0.005628860555589199, + 0.0030641560442745686, + 0.005159921478480101, + 0.00336665497161448, + 0.001830192981287837, + 0.007378194481134415, + 0.0037783083971589804, + 0.006333823781460524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.048701900988817215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674353027716279 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.00674353027716279, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674353027716279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2592187225818634, + 0.06608036905527115, + 0.07239441573619843, + 0.04556091129779816, + 0.043914031237363815, + 0.08075682073831558, + 0.0576636828482151, + 0.07697174698114395, + 0.17469695210456848, + 0.1227424144744873 + ], + "credit_map": [ + 0.030366696417331696, + 0.038825687021017075, + 0.034300655126571655, + 0.02228059247136116, + 0.030529817566275597, + 0.03156688064336777, + 0.04900806397199631, + 0.0223403237760067, + 0.005625569261610508, + 0.006711209658533335, + 0.01155058667063713, + 0.007931350730359554, + 0.006006225012242794, + 0.011738318018615246, + 0.010142686776816845, + 0.006374414544552565, + 0.008196375332772732, + 0.010179826058447361, + 0.007879727520048618, + 0.008403981104493141, + 0.006450533866882324, + 0.010638325475156307, + 0.009442681446671486, + 0.011202959343791008, + 0.005883015226572752, + 0.006409509107470512, + 0.004754145164042711, + 0.006825985852628946, + 0.005059224087744951, + 0.00626855343580246, + 0.0062365448102355, + 0.004123928025364876, + 0.004225492011755705, + 0.0028792244847863913, + 0.01018250361084938, + 0.0038811927661299706, + 0.004179276060312986, + 0.008931898511946201, + 0.0059840744361281395, + 0.003650364000350237, + 0.007822639308869839, + 0.006397522985935211, + 0.010158836841583252, + 0.00899256020784378, + 0.014498557895421982, + 0.015807850286364555, + 0.008458549156785011, + 0.008620298467576504, + 0.007473417557775974, + 0.0037604260724037886, + 0.008429198525846004, + 0.00870309118181467, + 0.006969055626541376, + 0.007899703457951546, + 0.006717170588672161, + 0.007711616810411215, + 0.005653486121445894, + 0.010765925981104374, + 0.009323370642960072, + 0.0069677201099693775, + 0.014042877592146397, + 0.008482808247208595, + 0.012509563006460667, + 0.009225988760590553, + 0.015469043515622616, + 0.01886621117591858, + 0.03883371502161026, + 0.021396096795797348, + 0.01247379556298256, + 0.024638213217258453, + 0.02707689441740513, + 0.015942949801683426, + 0.012300342321395874, + 0.01474814210087061, + 0.03265153616666794, + 0.014943839982151985, + 0.005774183198809624, + 0.021794795989990234, + 0.008869273588061333, + 0.011660289019346237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.00674353027716279 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201635157689452 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201635157689452, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201635157689452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16826659440994263, + 0.17284701764583588, + 0.07947906106710434, + 0.0365590862929821, + 0.04602424055337906, + 0.09190438687801361, + 0.0481477715075016, + 0.10964027792215347, + 0.1445709764957428, + 0.1025606244802475 + ], + "credit_map": [ + 0.016087539494037628, + 0.020764878019690514, + 0.014087549410760403, + 0.01751171238720417, + 0.01756558194756508, + 0.020199153572320938, + 0.043935131281614304, + 0.01811504177749157, + 0.01245992723852396, + 0.008547328412532806, + 0.053411826491355896, + 0.010820398107171059, + 0.0066648563370108604, + 0.06000356003642082, + 0.010206417180597782, + 0.010732699185609818, + 0.015384175814688206, + 0.006735595874488354, + 0.013625253923237324, + 0.006909062620252371, + 0.004707650281488895, + 0.011857401579618454, + 0.006985056679695845, + 0.013274865224957466, + 0.0051504625007510185, + 0.005375006701797247, + 0.004122960846871138, + 0.0038672625087201595, + 0.004747377708554268, + 0.004647530149668455, + 0.005134996958076954, + 0.003513486823067069, + 0.006532331928610802, + 0.003237537806853652, + 0.010387198999524117, + 0.00489111989736557, + 0.003858804004266858, + 0.006219726987183094, + 0.004977385979145765, + 0.005920133087784052, + 0.009559175930917263, + 0.008522373624145985, + 0.01753711886703968, + 0.01867612451314926, + 0.011039817705750465, + 0.009786655195057392, + 0.00796522106975317, + 0.008817901834845543, + 0.004942961502820253, + 0.004642747342586517, + 0.006307726260274649, + 0.007059487514197826, + 0.007745333015918732, + 0.005836750380694866, + 0.005360438488423824, + 0.0062523288652300835, + 0.01504630595445633, + 0.016403216868638992, + 0.012781576253473759, + 0.012148753739893436, + 0.011213154532015324, + 0.011633136309683323, + 0.013216732069849968, + 0.01719740219414234, + 0.012866334058344364, + 0.013772993348538876, + 0.026608334854245186, + 0.012663188390433788, + 0.014478461816906929, + 0.026064595207571983, + 0.017466381192207336, + 0.02065068855881691, + 0.01175493374466896, + 0.010268236510455608, + 0.016224512830376625, + 0.010068160481750965, + 0.007613314315676689, + 0.020372014492750168, + 0.00781098660081625, + 0.018448464572429657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.002201635157689452 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a slow, forward-moving close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dancer wearing a simple dark dance dress with a silky luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer's hair styled in a sleek ballet bun with a few sweat-soaked strands brushing against her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a powerful spotlight from above and side creating high contrast between highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dancer twirling gracefully with her head tilted back and a faint smile on her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033697061240673065 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.033697061240673065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697061240673065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28489020466804504, + 0.11324051767587662, + 0.14937055110931396, + 0.09319062530994415, + 0.04153149202466011, + 0.09453551471233368, + 0.06189665198326111, + 0.050564754754304886, + 0.05395088717341423, + 0.056828804314136505 + ], + "credit_map": [ + 0.030705001205205917, + 0.033190950751304626, + 0.01972171850502491, + 0.03430653735995293, + 0.03429596126079559, + 0.029663972556591034, + 0.0801100805401802, + 0.022895993664860725, + 0.010444515384733677, + 0.01412601862102747, + 0.022152282297611237, + 0.009783517569303513, + 0.01018962636590004, + 0.02443237043917179, + 0.013088323175907135, + 0.009023857302963734, + 0.024831173941493034, + 0.019373873248696327, + 0.023667721077799797, + 0.016166266053915024, + 0.01088129635900259, + 0.01941732130944729, + 0.016865132376551628, + 0.01816777139902115, + 0.009219656698405743, + 0.010494722053408623, + 0.010451091453433037, + 0.007583259139209986, + 0.012896131724119186, + 0.01374252699315548, + 0.015030642040073872, + 0.013772583566606045, + 0.005675392225384712, + 0.0021235973108559847, + 0.009118177928030491, + 0.0043468172661960125, + 0.0040209293365478516, + 0.007691246923059225, + 0.003935575485229492, + 0.004619755316525698, + 0.008899446576833725, + 0.007764553185552359, + 0.013185943476855755, + 0.026791924610733986, + 0.009122046642005444, + 0.009641309268772602, + 0.009972179308533669, + 0.009158113040030003, + 0.005740816704928875, + 0.006550013553351164, + 0.008573289960622787, + 0.010011676698923111, + 0.008950024843215942, + 0.005656509194523096, + 0.009074424393475056, + 0.007339892443269491, + 0.00490491883829236, + 0.005879949312657118, + 0.006256099324673414, + 0.005326434504240751, + 0.010741002857685089, + 0.005213148891925812, + 0.006528628524392843, + 0.005714575294405222, + 0.004083017352968454, + 0.005533729214221239, + 0.009233391843736172, + 0.004387767985463142, + 0.006177522242069244, + 0.009193276055157185, + 0.008178927935659885, + 0.007163255475461483, + 0.005374318454414606, + 0.004633714910596609, + 0.015263650566339493, + 0.007922136224806309, + 0.003342263400554657, + 0.01000442449003458, + 0.005039873532950878, + 0.005248422268778086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.033697061240673065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_tensor([5278, 5278])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_tensor([5278, 5278])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f793a5c357a4f8f8f960144352ece9a509288466 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_tensor([5278, 5278])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a high-definition vertical video blog shot, captured from a steady, mid-range still, the image focuses on the Baidu Bear Paw logo, composed precisely of countless luminous blue data blocks. It hovers quietly in the center of a deep, nearly black digital space. Each translucent cube emits a soft digital glow, creating a sense of high-tech sophistication. As the shot continues, the logo slowly disintegrates, its constituent blocks seemingly freed from their invisible constraints, gracefully and effortlessly drifting outwards, forming a dynamic nebula of countless blue points of light that flickers and flows in the darkness. Then, as if drawn by a central gravitational force, these scattered points of light begin to converge smoothly and orderly toward the center of the frame, swirling and rearranging, ultimately seamlessly reconstructing into a delicate model of the Earth. The contours of the land and ocean are clearly delineated by the data blocks of varying densities. The entire Earth model rotates in a gentle, steady rhythm at the center of the frame, its own light casting a faint, dynamic halo against the inky black background. The entire process is carried out under cool light and shadow, full of a sense of the future, technology and a quiet evolutionary beauty, creating a focused and poetic visual atmosphere.", + "seed": 32104, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_5278_seed_32104_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4429025293327868, + "vlm_logprob": -0.4429025293327868, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.4429025293327868, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5510226488113403 + ], + "opposite_logprob_tokens": [ + -0.30102261900901794 + ], + "logprob_reward": -1.5510226488113403, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5510226488113403, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3297293186187744, + 0.05250800400972366, + 0.18353460729122162, + 0.1820593923330307, + 0.0762602910399437, + 0.04241585358977318, + 0.034532178193330765, + 0.024510584771633148, + 0.025248948484659195, + 0.04920082911849022 + ], + "credit_map": [ + 0.027070743963122368, + 0.02374701201915741, + 0.10015519708395004, + 0.022083844989538193, + 0.021560529246926308, + 0.09798348695039749, + 0.022820428013801575, + 0.014308078214526176, + 0.008233068510890007, + 0.0033355592750012875, + 0.014762169681489468, + 0.0039551942609250546, + 0.0032755008433014154, + 0.005380018148571253, + 0.0048126233741641045, + 0.008753865957260132, + 0.02333679236471653, + 0.022086545825004578, + 0.02340279519557953, + 0.014599910005927086, + 0.02436954155564308, + 0.03533393517136574, + 0.021142730489373207, + 0.01926236040890217, + 0.016519198194146156, + 0.003071749582886696, + 0.06059132516384125, + 0.006987059023231268, + 0.00585629977285862, + 0.07297136634588242, + 0.0050635021179914474, + 0.010998889803886414, + 0.0070167719386518, + 0.004045929294079542, + 0.028087757527828217, + 0.003927048295736313, + 0.003820703597739339, + 0.010761312209069729, + 0.006630007643252611, + 0.011970756575465202, + 0.006096976343542337, + 0.0041990154422819614, + 0.00732684601098299, + 0.004101266153156757, + 0.0027713957242667675, + 0.007331311237066984, + 0.0040528299286961555, + 0.00653621181845665, + 0.0073937708511948586, + 0.0024978225119411945, + 0.005572777707129717, + 0.002147930208593607, + 0.0025874741841107607, + 0.003918771166354418, + 0.0030960661824792624, + 0.0073175630532205105, + 0.003508385270833969, + 0.0028422297909855843, + 0.0037682882975786924, + 0.0017543191788718104, + 0.0026811265852302313, + 0.0026301005855202675, + 0.00229025655426085, + 0.005035874899476767, + 0.0020282799378037453, + 0.0035638539120554924, + 0.0046232896856963634, + 0.002461550524458289, + 0.002656696131452918, + 0.00402837386354804, + 0.0027666573878377676, + 0.0031202463433146477, + 0.0039393180049955845, + 0.003871507477015257, + 0.009973213076591492, + 0.004591342527419329, + 0.004974916111677885, + 0.012104052118957043, + 0.005236930213868618, + 0.004509543534368277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -1.5510226488113403 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032834685407578945 + ], + "opposite_logprob_tokens": [ + -5.753283500671387 + ], + "logprob_reward": -0.0032834685407578945, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032834685407578945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18663330376148224, + 0.06078129634261131, + 0.21164952218532562, + 0.2761496305465698, + 0.05821877717971802, + 0.05642925947904587, + 0.042978685349226, + 0.024594251066446304, + 0.04723064973950386, + 0.035334646701812744 + ], + "credit_map": [ + 0.023879650980234146, + 0.0159732848405838, + 0.04804924502968788, + 0.01411927118897438, + 0.012689950875937939, + 0.027868561446666718, + 0.018149089068174362, + 0.025904247537255287, + 0.00783620961010456, + 0.006771673448383808, + 0.013327466323971748, + 0.006571066100150347, + 0.0062427399680018425, + 0.005862412974238396, + 0.007102837320417166, + 0.007066893391311169, + 0.026331396773457527, + 0.02583710104227066, + 0.02661261335015297, + 0.017266014590859413, + 0.02932604029774666, + 0.04043744504451752, + 0.02411271072924137, + 0.021726178005337715, + 0.024577226489782333, + 0.005407697521150112, + 0.09521211683750153, + 0.01105546671897173, + 0.009238343685865402, + 0.10816451162099838, + 0.007201021537184715, + 0.015293218195438385, + 0.006303048692643642, + 0.005669533275067806, + 0.010900578461587429, + 0.00752731878310442, + 0.004962129984050989, + 0.008494369685649872, + 0.009580944664776325, + 0.004780854098498821, + 0.009690922684967518, + 0.004569775424897671, + 0.01071211788803339, + 0.007502980995923281, + 0.002946041524410248, + 0.007226581685245037, + 0.007505094166845083, + 0.006275746040046215, + 0.008755112998187542, + 0.0035823462530970573, + 0.006343633867800236, + 0.002893177792429924, + 0.005314646288752556, + 0.006517122499644756, + 0.0036448019091039896, + 0.00592784071341157, + 0.0036833572667092085, + 0.002583335153758526, + 0.0035537939984351397, + 0.001992810284718871, + 0.0022952842991799116, + 0.0031245925929397345, + 0.0028254820499569178, + 0.004535594955086708, + 0.004207093268632889, + 0.0029628556221723557, + 0.00880683958530426, + 0.0059921471402049065, + 0.0028247220907360315, + 0.009114114567637444, + 0.008529833517968655, + 0.004793043248355389, + 0.003667521523311734, + 0.0034427554346621037, + 0.006232913583517075, + 0.004155464004725218, + 0.003278957214206457, + 0.006647791713476181, + 0.0036990868393331766, + 0.0042101554572582245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0032834685407578945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978896301239729 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.00978896301239729, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978896301239729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09881424158811569, + 0.07068925350904465, + 0.10613254457712173, + 0.2631419897079468, + 0.052275147289037704, + 0.1287192404270172, + 0.07913082838058472, + 0.051684848964214325, + 0.09645263850688934, + 0.052959341555833817 + ], + "credit_map": [ + 0.015211953781545162, + 0.013519730418920517, + 0.009559025056660175, + 0.00964148249477148, + 0.011286385357379913, + 0.016269870102405548, + 0.012738516554236412, + 0.01058727502822876, + 0.006421687081456184, + 0.009457400068640709, + 0.009422126226127148, + 0.0065399534069001675, + 0.010250367224216461, + 0.009381811134517193, + 0.01126005593687296, + 0.007955845445394516, + 0.011310853064060211, + 0.013655747286975384, + 0.011487041600048542, + 0.008652552030980587, + 0.015134790912270546, + 0.018754176795482635, + 0.011706413701176643, + 0.015430969186127186, + 0.02224062569439411, + 0.006855769548565149, + 0.08504558354616165, + 0.010813386179506779, + 0.011672653257846832, + 0.10101721435785294, + 0.008798482827842236, + 0.016698263585567474, + 0.004125576466321945, + 0.0038815366569906473, + 0.006338384468108416, + 0.006769735831767321, + 0.004315226338803768, + 0.007044803351163864, + 0.010412540286779404, + 0.009387345053255558, + 0.02254965715110302, + 0.009754640981554985, + 0.016239149495959282, + 0.014725052751600742, + 0.008415726944804192, + 0.024250347167253494, + 0.01588854007422924, + 0.016896111890673637, + 0.014800406992435455, + 0.007113611791282892, + 0.011351157911121845, + 0.006428057793527842, + 0.010786881670355797, + 0.012111295014619827, + 0.006952793337404728, + 0.00958662386983633, + 0.007601730525493622, + 0.004965567495673895, + 0.009126266464591026, + 0.004854175262153149, + 0.005290579050779343, + 0.0054894695058465, + 0.006162738893181086, + 0.00819432269781828, + 0.010926881805062294, + 0.006358595564961433, + 0.01762508973479271, + 0.01047355867922306, + 0.007209376432001591, + 0.016756143420934677, + 0.0148507971316576, + 0.012252197600901127, + 0.006520337890833616, + 0.004306442569941282, + 0.008505100384354591, + 0.005489114671945572, + 0.005812478251755238, + 0.008326615206897259, + 0.005917738191783428, + 0.008081513457000256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.00978896301239729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706726789474487 + ], + "opposite_logprob_tokens": [ + -2.1270673274993896 + ], + "logprob_reward": -0.12706726789474487, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706726789474487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11872097104787827, + 0.07815558463335037, + 0.16430282592773438, + 0.060677215456962585, + 0.08345071971416473, + 0.10842294991016388, + 0.12202858179807663, + 0.07889892905950546, + 0.11712011694908142, + 0.0682220607995987 + ], + "credit_map": [ + 0.016426730901002884, + 0.010654156096279621, + 0.026796791702508926, + 0.013808202929794788, + 0.012777936644852161, + 0.008802437223494053, + 0.016155414283275604, + 0.013299294747412205, + 0.010337382555007935, + 0.005026866681873798, + 0.02477843686938286, + 0.0046781618148088455, + 0.005720018409192562, + 0.013065211474895477, + 0.005323868244886398, + 0.009225639514625072, + 0.020306019112467766, + 0.017752844840288162, + 0.03119918517768383, + 0.012188941240310669, + 0.01969948038458824, + 0.03246978670358658, + 0.014894558116793633, + 0.015792004764080048, + 0.006318436004221439, + 0.002594259800389409, + 0.015562892891466618, + 0.0030435402877628803, + 0.0029103385750204325, + 0.0226027499884367, + 0.002595482161268592, + 0.005049520637840033, + 0.005545551888644695, + 0.00446335319429636, + 0.013638189993798733, + 0.010429687798023224, + 0.0049296156503260136, + 0.009635881520807743, + 0.013932442292571068, + 0.020875997841358185, + 0.016451779752969742, + 0.008624102920293808, + 0.020167840644717216, + 0.014415841549634933, + 0.008599594235420227, + 0.013376286253333092, + 0.011781218461692333, + 0.01500627864152193, + 0.014077480882406235, + 0.009763309732079506, + 0.01640397310256958, + 0.009889873676002026, + 0.01449713297188282, + 0.024233240634202957, + 0.014247446320950985, + 0.018916135653853416, + 0.01002335175871849, + 0.007048019673675299, + 0.013337206095457077, + 0.007172927260398865, + 0.008481409400701523, + 0.01014726608991623, + 0.010174880735576153, + 0.012513862922787666, + 0.010295278392732143, + 0.009029362350702286, + 0.01941661909222603, + 0.014188294298946857, + 0.009738481603562832, + 0.019775286316871643, + 0.02187502384185791, + 0.012801771983504295, + 0.006667308509349823, + 0.008921029046177864, + 0.009141089394688606, + 0.0072107198648154736, + 0.007199214771389961, + 0.01057280320674181, + 0.008082933723926544, + 0.010426966473460197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.12706726789474487 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233502984046936 + ], + "opposite_logprob_tokens": [ + -0.8983502984046936 + ], + "logprob_reward": -0.5233502984046936, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233502984046936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20932544767856598, + 0.04294047877192497, + 0.08384402096271515, + 0.10815240442752838, + 0.15333957970142365, + 0.14523664116859436, + 0.09541262686252594, + 0.09090464562177658, + 0.026249071583151817, + 0.04459504410624504 + ], + "credit_map": [ + 0.018732944503426552, + 0.010568047873675823, + 0.0701645165681839, + 0.013055391609668732, + 0.010796887800097466, + 0.06394845247268677, + 0.01280075591057539, + 0.00925846304744482, + 0.004108577035367489, + 0.0059736743569374084, + 0.0045608035288751125, + 0.004535180050879717, + 0.007529708556830883, + 0.006848617922514677, + 0.004838516470044851, + 0.0045453994534909725, + 0.012226586230099201, + 0.008618985302746296, + 0.01352092158049345, + 0.0062652346678078175, + 0.008253288455307484, + 0.018370943143963814, + 0.007972286082804203, + 0.008615778759121895, + 0.009196136146783829, + 0.005073553882539272, + 0.0323532298207283, + 0.005902566015720367, + 0.007707476615905762, + 0.03383351489901543, + 0.00654771039262414, + 0.007538209203630686, + 0.013873834162950516, + 0.007643647957593203, + 0.05139172822237015, + 0.014787211082875729, + 0.008359663188457489, + 0.018084462732076645, + 0.021041765809059143, + 0.01815726049244404, + 0.01953318901360035, + 0.013573914766311646, + 0.019856058061122894, + 0.02181532420217991, + 0.00885236170142889, + 0.02322394959628582, + 0.019144298508763313, + 0.019237546250224113, + 0.017780829221010208, + 0.009315237402915955, + 0.016177751123905182, + 0.005717829801142216, + 0.00792192667722702, + 0.009474389255046844, + 0.008703477680683136, + 0.020321184769272804, + 0.01326050329953432, + 0.00637244014069438, + 0.0164643581956625, + 0.006005608476698399, + 0.0058397832326591015, + 0.012426965869963169, + 0.008895031176507473, + 0.021639954298734665, + 0.0031987393740564585, + 0.0030244500376284122, + 0.003222157247364521, + 0.0030909599736332893, + 0.0019161690725013614, + 0.00389858684502542, + 0.0037029751110821962, + 0.00419503403827548, + 0.004444335587322712, + 0.007509768009185791, + 0.004719310440123081, + 0.006120555102825165, + 0.00461987080052495, + 0.0066047594882547855, + 0.005074836313724518, + 0.005501609295606613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.5233502984046936 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-definition vertical video blog shot, captured from a steady, mid-range still, the image focuses on the Baidu Bear Paw logo, composed precisely of countless luminous blue data blocks. It hovers quietly in the center of a deep, nearly black digital space. Each translucent cube emits a soft digital glow, creating a sense of high-tech sophistication. As the shot continues, the logo slowly disintegrates, its constituent blocks seemingly freed from their invisible constraints, gracefully and effortlessly drifting outwards, forming a dynamic nebula of countless blue points of light that flickers and flows in the darkness. Then, as if drawn by a central gravitational force, these scattered points of light begin to converge smoothly and orderly toward the center of the frame, swirling and rearranging, ultimately seamlessly reconstructing into a delicate model of the Earth. The contours of the land and ocean are clearly delineated by the data blocks of varying densities. The entire Earth model rotates in a gentle, steady rhythm at the center of the frame, its own light casting a faint, dynamic halo against the inky black background. The entire process is carried out under cool light and shadow, full of a sense of the future, technology and a quiet evolutionary beauty, creating a focused and poetic visual atmosphere.", + "seed": 32105, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_5278_seed_32105_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.35663587441667916, + "vlm_logprob": -0.35663587441667916, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.35663587441667916, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5430772304534912 + ], + "opposite_logprob_tokens": [ + -0.29307717084884644 + ], + "logprob_reward": -1.5430772304534912, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5430772304534912, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12343200296163559, + 0.04511696472764015, + 0.02597731165587902, + 0.15208286046981812, + 0.058833424001932144, + 0.1378922164440155, + 0.06692440807819366, + 0.16989874839782715, + 0.08930079638957977, + 0.13054127991199493 + ], + "credit_map": [ + 0.0154545484110713, + 0.010834407992661, + 0.022260526195168495, + 0.010814039967954159, + 0.01618565246462822, + 0.011298710480332375, + 0.014213891699910164, + 0.022370221093297005, + 0.004414438735693693, + 0.0039058898109942675, + 0.00565475644543767, + 0.005111569073051214, + 0.004549145232886076, + 0.006179089192301035, + 0.008980317041277885, + 0.006321759428828955, + 0.0040835668332874775, + 0.0033980717416852713, + 0.00445233890786767, + 0.0028185900300741196, + 0.0025446717627346516, + 0.0037258428055793047, + 0.002049955539405346, + 0.00290427403524518, + 0.02070322446525097, + 0.01319798268377781, + 0.03541288152337074, + 0.011733109131455421, + 0.018569335341453552, + 0.023001302033662796, + 0.010810739360749722, + 0.01865428313612938, + 0.008792655542492867, + 0.004765275400131941, + 0.009736772626638412, + 0.007082198280841112, + 0.005150164943188429, + 0.00945222470909357, + 0.0067894598469138145, + 0.007064671255648136, + 0.014806674793362617, + 0.00875453744083643, + 0.0256962813436985, + 0.01171291247010231, + 0.008185879327356815, + 0.036548711359500885, + 0.011447161436080933, + 0.020740056410431862, + 0.006185158621519804, + 0.0038477492053061724, + 0.01601858250796795, + 0.0064071668311953545, + 0.005102443974465132, + 0.011280125938355923, + 0.005492683034390211, + 0.012590503320097923, + 0.018117256462574005, + 0.008090763352811337, + 0.03300075978040695, + 0.009659837931394577, + 0.007115777116268873, + 0.048263948410749435, + 0.009858475998044014, + 0.03579191863536835, + 0.009561867453157902, + 0.00670635187998414, + 0.021557820960879326, + 0.0057593537494540215, + 0.00729003082960844, + 0.013906148262321949, + 0.007646208629012108, + 0.016873005777597427, + 0.005973210092633963, + 0.004037599544972181, + 0.01755429059267044, + 0.003331670770421624, + 0.003721822751685977, + 0.04697560518980026, + 0.005621044896543026, + 0.04332602769136429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -1.5430772304534912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047042411752045155 + ], + "opposite_logprob_tokens": [ + -5.379704475402832 + ], + "logprob_reward": -0.0047042411752045155, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0047042411752045155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23875835537910461, + 0.09567753225564957, + 0.08220711350440979, + 0.06552167236804962, + 0.05044620484113693, + 0.10999526083469391, + 0.05302667245268822, + 0.0749545618891716, + 0.05492997169494629, + 0.17448264360427856 + ], + "credit_map": [ + 0.02152908407151699, + 0.028285767883062363, + 0.03488931059837341, + 0.03633809834718704, + 0.03617217391729355, + 0.020972203463315964, + 0.03463223949074745, + 0.02593948319554329, + 0.009434254840016365, + 0.00721154036000371, + 0.015023736283183098, + 0.014603079296648502, + 0.009854702278971672, + 0.013525375165045261, + 0.013879170641303062, + 0.012145663611590862, + 0.017774458974599838, + 0.010646567679941654, + 0.007969116792082787, + 0.01110958494246006, + 0.010321661829948425, + 0.008209111168980598, + 0.010010281577706337, + 0.0061663296073675156, + 0.010762068443000317, + 0.0073334043845534325, + 0.011070811189711094, + 0.006236434448510408, + 0.009763534180819988, + 0.008447807282209396, + 0.005642306059598923, + 0.00626530684530735, + 0.0064013004302978516, + 0.0038122350815683603, + 0.01014712080359459, + 0.0056350985541939735, + 0.0041236234828829765, + 0.008778216317296028, + 0.004378873389214277, + 0.007169733755290508, + 0.007960498332977295, + 0.004316368605941534, + 0.02674323506653309, + 0.0058998968452215195, + 0.006337896920740604, + 0.03902783989906311, + 0.006584204267710447, + 0.013125321827828884, + 0.0053236824460327625, + 0.0043334621004760265, + 0.007810557261109352, + 0.004463955294340849, + 0.008357852697372437, + 0.008707401342689991, + 0.005416707135736942, + 0.00861305184662342, + 0.00795349758118391, + 0.005074625369161367, + 0.01330603938549757, + 0.004985921084880829, + 0.004482571966946125, + 0.017196275293827057, + 0.005502955988049507, + 0.016452673822641373, + 0.005378784146159887, + 0.003942600451409817, + 0.0148812560364604, + 0.0037044361233711243, + 0.005044810473918915, + 0.00673648901283741, + 0.004530575126409531, + 0.01071101613342762, + 0.004684747196733952, + 0.004100078251212835, + 0.018341470509767532, + 0.003459433326497674, + 0.0037311248015612364, + 0.07159485667943954, + 0.005694026127457619, + 0.06287692487239838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0047042411752045155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203657291829586 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.002203657291829586, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203657291829586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06373012065887451, + 0.08786766976118088, + 0.10293164104223251, + 0.150085911154747, + 0.04553505778312683, + 0.11764603108167648, + 0.05983247607946396, + 0.1033521220088005, + 0.05200446769595146, + 0.21701450645923615 + ], + "credit_map": [ + 0.010082375258207321, + 0.006644214503467083, + 0.010097187012434006, + 0.00694679468870163, + 0.007693143095821142, + 0.006703818216919899, + 0.006461429875344038, + 0.00910115335136652, + 0.008071988821029663, + 0.009409690275788307, + 0.009372087195515633, + 0.012630210258066654, + 0.008498153649270535, + 0.011566749773919582, + 0.015794074162840843, + 0.012524709105491638, + 0.014834274537861347, + 0.01650150492787361, + 0.014969639480113983, + 0.009547864086925983, + 0.013082391582429409, + 0.016700489446520805, + 0.008969361893832684, + 0.00832610297948122, + 0.020441317930817604, + 0.016309097409248352, + 0.03285473212599754, + 0.011471953243017197, + 0.018163498491048813, + 0.021257877349853516, + 0.009234814904630184, + 0.020352613180875778, + 0.007752548903226852, + 0.0033911089412868023, + 0.009079395793378353, + 0.00547410873696208, + 0.004260263405740261, + 0.004972670692950487, + 0.005327642895281315, + 0.005277314223349094, + 0.008459256961941719, + 0.0048815845511853695, + 0.028887439519166946, + 0.005839680787175894, + 0.006976592820137739, + 0.042269375175237656, + 0.007069649640470743, + 0.013262445107102394, + 0.006929097697138786, + 0.004030689597129822, + 0.006928481627255678, + 0.005906763952225447, + 0.006455283612012863, + 0.009575407020747662, + 0.008689663372933865, + 0.011317085474729538, + 0.011237439699470997, + 0.004923688247799873, + 0.016782307997345924, + 0.006381119601428509, + 0.0055192578583955765, + 0.029109368100762367, + 0.0065275137312710285, + 0.022871416062116623, + 0.0042511778883636, + 0.004062900319695473, + 0.013693038374185562, + 0.0036912886425852776, + 0.004196984693408012, + 0.007624067831784487, + 0.004685015883296728, + 0.009799991734325886, + 0.00873290654271841, + 0.008387154899537563, + 0.028821375221014023, + 0.0080172223970294, + 0.009251680225133896, + 0.07878576964139938, + 0.012796827591955662, + 0.06222157180309296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.002203657291829586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03445756062865257, + 0.052498649805784225, + 0.03424612805247307, + 0.06191612407565117, + 0.040646642446517944, + 0.02972280979156494, + 0.1424967348575592, + 0.1515108346939087, + 0.14889618754386902, + 0.30360838770866394 + ], + "credit_map": [ + 0.003728872165083885, + 0.0027866270393133163, + 0.004874350968748331, + 0.004173245746642351, + 0.0034097102470695972, + 0.004813587758690119, + 0.003997100051492453, + 0.006674063857644796, + 0.00564480060711503, + 0.0053338948637247086, + 0.006313398480415344, + 0.00582839036360383, + 0.008155561983585358, + 0.00796107854694128, + 0.0080341761931777, + 0.005227342713624239, + 0.003659923328086734, + 0.006736550945788622, + 0.004422320984303951, + 0.004634362645447254, + 0.003981899004429579, + 0.004894256126135588, + 0.0024208351969718933, + 0.003495976561680436, + 0.008163154125213623, + 0.008253417909145355, + 0.01036064699292183, + 0.005344299133867025, + 0.009839493781328201, + 0.008948965929448605, + 0.0043814475648105145, + 0.006624690722674131, + 0.004122829996049404, + 0.002795659238472581, + 0.0073092905804514885, + 0.004235684406012297, + 0.00464284885674715, + 0.007772279903292656, + 0.004844998940825462, + 0.004923047497868538, + 0.003021777607500553, + 0.0026451086159795523, + 0.0033766727428883314, + 0.0032400041818618774, + 0.0028956078458577394, + 0.005213076714426279, + 0.005311950575560331, + 0.004018608946353197, + 0.009268472902476788, + 0.005198071710765362, + 0.02590922825038433, + 0.007604083511978388, + 0.009627940133213997, + 0.05014967918395996, + 0.010437955148518085, + 0.024301284924149513, + 0.016835182905197144, + 0.01096694078296423, + 0.019701384007930756, + 0.015623685903847218, + 0.013505297712981701, + 0.02975628338754177, + 0.017285548150539398, + 0.027836503461003304, + 0.017043398693203926, + 0.011729693971574306, + 0.03003830462694168, + 0.011289562098681927, + 0.014970324002206326, + 0.022304514423012733, + 0.013924525119364262, + 0.027595840394496918, + 0.008375057019293308, + 0.00589889008551836, + 0.030221937224268913, + 0.006025705952197313, + 0.0062058186158537865, + 0.12921342253684998, + 0.007416583131998777, + 0.11025097966194153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.007627766579389572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556647658348083 + ], + "opposite_logprob_tokens": [ + -1.6005665063858032 + ], + "logprob_reward": -0.22556647658348083, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556647658348083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08996395021677017, + 0.04216267541050911, + 0.07756582647562027, + 0.1191239058971405, + 0.07159753143787384, + 0.04170829430222511, + 0.08066745102405548, + 0.08925667405128479, + 0.15976694226264954, + 0.2281867414712906 + ], + "credit_map": [ + 0.013743802905082703, + 0.006297019775956869, + 0.023329639807343483, + 0.008158311247825623, + 0.008551988750696182, + 0.010394296608865261, + 0.007492458447813988, + 0.011996441520750523, + 0.004636194556951523, + 0.0039938571862876415, + 0.005767001770436764, + 0.006169435102492571, + 0.005490248091518879, + 0.006305202376097441, + 0.004896898288279772, + 0.004903840366750956, + 0.009719477966427803, + 0.016383344307541847, + 0.00876118615269661, + 0.009386821649968624, + 0.010870853438973427, + 0.00956495851278305, + 0.005700304172933102, + 0.007178883533924818, + 0.016647346317768097, + 0.013103174977004528, + 0.018469860777258873, + 0.009888458997011185, + 0.017023451626300812, + 0.021296892315149307, + 0.006636351812630892, + 0.01605837605893612, + 0.009248321875929832, + 0.004006337374448776, + 0.014860891737043858, + 0.006151753477752209, + 0.005843895021826029, + 0.015441973693668842, + 0.006032068282365799, + 0.010012289509177208, + 0.003968259319663048, + 0.0026733039412647486, + 0.007388908416032791, + 0.003855066606774926, + 0.0045627569779753685, + 0.010380939580500126, + 0.004292025696486235, + 0.004587035626173019, + 0.005352100357413292, + 0.0040532806888222694, + 0.01627543196082115, + 0.005875850562006235, + 0.006652111187577248, + 0.025228513404726982, + 0.0072154696099460125, + 0.010014690458774567, + 0.0076775141060352325, + 0.00490840757265687, + 0.015232083387672901, + 0.007096914108842611, + 0.005098533350974321, + 0.02532912604510784, + 0.0067669544368982315, + 0.017147140577435493, + 0.01717452146112919, + 0.009166967123746872, + 0.04139109328389168, + 0.008677205070853233, + 0.010909725911915302, + 0.02529696188867092, + 0.011161306872963905, + 0.03598915785551071, + 0.005516646429896355, + 0.004166507162153721, + 0.022616667672991753, + 0.003561585210263729, + 0.00442577013745904, + 0.09919164329767227, + 0.005810433998703957, + 0.08289747685194016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.22556647658348083 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-definition vertical video blog shot, captured from a steady, mid-range still, the image focuses on the Baidu Bear Paw logo, composed precisely of countless luminous blue data blocks. It hovers quietly in the center of a deep, nearly black digital space. Each translucent cube emits a soft digital glow, creating a sense of high-tech sophistication. As the shot continues, the logo slowly disintegrates, its constituent blocks seemingly freed from their invisible constraints, gracefully and effortlessly drifting outwards, forming a dynamic nebula of countless blue points of light that flickers and flows in the darkness. Then, as if drawn by a central gravitational force, these scattered points of light begin to converge smoothly and orderly toward the center of the frame, swirling and rearranging, ultimately seamlessly reconstructing into a delicate model of the Earth. The contours of the land and ocean are clearly delineated by the data blocks of varying densities. The entire Earth model rotates in a gentle, steady rhythm at the center of the frame, its own light casting a faint, dynamic halo against the inky black background. The entire process is carried out under cool light and shadow, full of a sense of the future, technology and a quiet evolutionary beauty, creating a focused and poetic visual atmosphere.", + "seed": 32106, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_5278_seed_32106_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21408474510535597, + "vlm_logprob": -0.21408474510535597, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.21408474510535597, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8411806225776672 + ], + "opposite_logprob_tokens": [ + -0.7161806225776672 + ], + "logprob_reward": -0.8411806225776672, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.8411806225776672, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14084814488887787, + 0.030270922929048538, + 0.019637957215309143, + 0.016778910532593727, + 0.23265813291072845, + 0.03114822320640087, + 0.04230428487062454, + 0.02718081884086132, + 0.35632312297821045, + 0.10284943878650665 + ], + "credit_map": [ + 0.013465066440403461, + 0.003926875069737434, + 0.040854562073946, + 0.006021324545145035, + 0.006220920942723751, + 0.04833285138010979, + 0.004400671925395727, + 0.017625872045755386, + 0.003661108436062932, + 0.0008051727782003582, + 0.003836281131953001, + 0.0012038283748552203, + 0.0009129227255471051, + 0.009368755854666233, + 0.001337322173640132, + 0.009145531803369522, + 0.001865678932517767, + 0.0006118544843047857, + 0.005050388630479574, + 0.0007157745421864092, + 0.0006488210055977106, + 0.006720550823956728, + 0.0006772982305847108, + 0.003347590332850814, + 0.0017658884171396494, + 0.00043697000364772975, + 0.0038239355199038982, + 0.00043935448047704995, + 0.0005277483724057674, + 0.006521179806441069, + 0.00042819013469852507, + 0.002835644641891122, + 0.011189112439751625, + 0.0024387375451624393, + 0.08939807116985321, + 0.005487458780407906, + 0.0031144656240940094, + 0.10190711915493011, + 0.0022938137408345938, + 0.01682935282588005, + 0.004075915087014437, + 0.002664499916136265, + 0.004244137089699507, + 0.0036015415098518133, + 0.0026323855854570866, + 0.005542629398405552, + 0.002640517195686698, + 0.005746596492826939, + 0.002225285628810525, + 0.001648957491852343, + 0.0023962289560586214, + 0.0023709714878350496, + 0.0015906270127743483, + 0.015209761448204517, + 0.0021008034236729145, + 0.014761649072170258, + 0.00331552280113101, + 0.0026396899484097958, + 0.004610278178006411, + 0.004425141029059887, + 0.0022244462743401527, + 0.00378647493198514, + 0.0026249983347952366, + 0.0035542661789804697, + 0.019773876294493675, + 0.016278821974992752, + 0.11044302582740784, + 0.013095973990857601, + 0.03233463317155838, + 0.03343433514237404, + 0.013533050194382668, + 0.11742939800024033, + 0.013207593001425266, + 0.004135929513722658, + 0.039003968238830566, + 0.004130280576646328, + 0.0051414561457931995, + 0.00899308267980814, + 0.004644738510251045, + 0.02359238639473915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.8411806225776672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014283913187682629 + ], + "opposite_logprob_tokens": [ + -4.264284133911133 + ], + "logprob_reward": -0.014283913187682629, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014283913187682629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052748195827007294, + 0.05572420731186867, + 0.006488352082669735, + 0.03509780764579773, + 0.6363155841827393, + 0.03276506066322327, + 0.02138417586684227, + 0.03700686991214752, + 0.10806930810213089, + 0.014400428161025047 + ], + "credit_map": [ + 0.0069618141278624535, + 0.0025922891218215227, + 0.010625284165143967, + 0.0033713686279952526, + 0.0037234167102724314, + 0.008952243253588676, + 0.004799650050699711, + 0.01172212790697813, + 0.003057306632399559, + 0.0008738564793020487, + 0.016800694167613983, + 0.001340637682005763, + 0.0012300502276048064, + 0.02359609492123127, + 0.0011385392863303423, + 0.007687023840844631, + 0.001001570257358253, + 0.00035111888428218663, + 0.0011734713334590197, + 0.0006162882200442255, + 0.00043931687832809985, + 0.0011056079529225826, + 0.0006219549686647952, + 0.0011790238786488771, + 0.0037088431417942047, + 0.0005586289917118847, + 0.009293196722865105, + 0.0008826610282994807, + 0.0008305257651954889, + 0.013342988677322865, + 0.000790960795711726, + 0.005690000951290131, + 0.03398348391056061, + 0.006592309568077326, + 0.2577064633369446, + 0.015657899901270866, + 0.00859768409281969, + 0.2740509808063507, + 0.0061976187862455845, + 0.033529140055179596, + 0.0030004202853888273, + 0.0034605800174176693, + 0.004448122810572386, + 0.003612280124798417, + 0.002694692462682724, + 0.005648506805300713, + 0.003453488228842616, + 0.0064469678327441216, + 0.0027961344458162785, + 0.0014446746790781617, + 0.0019041151972487569, + 0.0029899445362389088, + 0.0016297346446663141, + 0.0034024256747215986, + 0.0029572017956525087, + 0.004259945824742317, + 0.0038297700230032206, + 0.003799241967499256, + 0.005043395329266787, + 0.006116398144513369, + 0.003945274278521538, + 0.004422246012836695, + 0.00578165240585804, + 0.004068892914801836, + 0.006528479512780905, + 0.006106069777160883, + 0.02952595055103302, + 0.004577955696731806, + 0.012671767733991146, + 0.011888002045452595, + 0.0050882804207503796, + 0.0316828116774559, + 0.0027072892989963293, + 0.0012033794773742557, + 0.003881925716996193, + 0.0010405360953882337, + 0.0017453139880672097, + 0.0012694657780230045, + 0.0011802752269431949, + 0.0013722419971600175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.014283913187682629 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595545070245862 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.00595545070245862, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595545070245862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12459460645914078, + 0.021767722442746162, + 0.009687263518571854, + 0.010657711885869503, + 0.29506081342697144, + 0.03426632285118103, + 0.021981047466397285, + 0.19004535675048828, + 0.25880923867225647, + 0.03313000500202179 + ], + "credit_map": [ + 0.01331010926514864, + 0.003456271020695567, + 0.04794921725988388, + 0.005092718172818422, + 0.00454043410718441, + 0.038775019347667694, + 0.0037834527902305126, + 0.00768738379701972, + 0.002548355609178543, + 0.00043354943045414984, + 0.004898119252175093, + 0.0006900859298184514, + 0.0005642873584292829, + 0.004201023373752832, + 0.0008462126716040075, + 0.0075860884971916676, + 0.0011014941846951842, + 0.0002930571208707988, + 0.001961302710697055, + 0.00038538165972568095, + 0.0004039216146338731, + 0.003415188519284129, + 0.00032546985312364995, + 0.0018014470115303993, + 0.0008502649143338203, + 0.0002958432014565915, + 0.0028925444930791855, + 0.0004496654437389225, + 0.00041200549458153546, + 0.0038104611448943615, + 0.0004100215155631304, + 0.00153690529987216, + 0.014497535303235054, + 0.0030898756813257933, + 0.11423178017139435, + 0.006989581044763327, + 0.003924721386283636, + 0.1289660930633545, + 0.002921644365414977, + 0.02043958008289337, + 0.0042381249368190765, + 0.0036573580000549555, + 0.00433517899364233, + 0.004292706027626991, + 0.004499650560319424, + 0.00436120480298996, + 0.003980118781328201, + 0.004901978652924299, + 0.0017585172317922115, + 0.002208021702244878, + 0.001794843003153801, + 0.0032195334788411856, + 0.002068657660856843, + 0.0029088573064655066, + 0.003702386748045683, + 0.004320231266319752, + 0.017223699018359184, + 0.02364485338330269, + 0.02270692214369774, + 0.031156418845057487, + 0.022729549556970596, + 0.02565179206430912, + 0.025228599086403847, + 0.021703513339161873, + 0.014811383560299873, + 0.0063850595615804195, + 0.09380657225847244, + 0.007664252072572708, + 0.011020038276910782, + 0.021111732348799706, + 0.0067893266677856445, + 0.09722085297107697, + 0.004876445047557354, + 0.0015812109922990203, + 0.012390810996294022, + 0.0017208221834152937, + 0.0016377379652112722, + 0.002756892703473568, + 0.0017290567047894, + 0.006437030620872974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.00595545070245862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865523427724838 + ], + "opposite_logprob_tokens": [ + -3.0486552715301514 + ], + "logprob_reward": -0.04865523427724838, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865523427724838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03945864737033844, + 0.01656201295554638, + 0.010955425910651684, + 0.023923980072140694, + 0.433783620595932, + 0.029758216813206673, + 0.0870041772723198, + 0.13545504212379456, + 0.17957265675067902, + 0.04352622106671333 + ], + "credit_map": [ + 0.00430596899241209, + 0.00207718787714839, + 0.012009277939796448, + 0.002742442302405834, + 0.002526750322431326, + 0.009140947833657265, + 0.0031387447379529476, + 0.0035173262003809214, + 0.0020796540193259716, + 0.0004903477965854108, + 0.005387342069298029, + 0.0006631299038417637, + 0.0006053714314475656, + 0.0020318033639341593, + 0.0005989668425172567, + 0.0047053988091647625, + 0.0013981923693791032, + 0.0002508095349185169, + 0.0020109997130930424, + 0.00038644822780042887, + 0.00037280897959135473, + 0.004037023521959782, + 0.0004213294305372983, + 0.0020778137259185314, + 0.0019296396058052778, + 0.00041475449688732624, + 0.006565427873283625, + 0.0006148848915472627, + 0.0005951903876848519, + 0.009699343703687191, + 0.00047295057447627187, + 0.003631787607446313, + 0.023149680346250534, + 0.004566774237900972, + 0.17684994637966156, + 0.010759773664176464, + 0.005881031509488821, + 0.186431884765625, + 0.004262913949787617, + 0.02188160829246044, + 0.0025970640126615763, + 0.003309558844193816, + 0.0036231311969459057, + 0.0029372896533459425, + 0.0037936719600111246, + 0.005606279242783785, + 0.003714245278388262, + 0.004176977090537548, + 0.006669922731816769, + 0.005581522360444069, + 0.008621502667665482, + 0.012226458638906479, + 0.004424923565238714, + 0.01846984215080738, + 0.011169156059622765, + 0.019840851426124573, + 0.01346123218536377, + 0.015632182359695435, + 0.014924340881407261, + 0.017545780166983604, + 0.018980782479047775, + 0.019446495920419693, + 0.014672585763037205, + 0.020791638642549515, + 0.01377202570438385, + 0.01490332093089819, + 0.030991779640316963, + 0.010097346268594265, + 0.033370330929756165, + 0.026441289111971855, + 0.011963562108576298, + 0.038033001124858856, + 0.012064153328537941, + 0.002231554826721549, + 0.009769287891685963, + 0.0020832084119319916, + 0.003190788207575679, + 0.0038823869545012712, + 0.00265584047883749, + 0.007649001199752092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.04865523427724838 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034850478172302 + ], + "opposite_logprob_tokens": [ + -1.9103485345840454 + ], + "logprob_reward": -0.16034850478172302, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034850478172302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06706322729587555, + 0.017011255025863647, + 0.00420853728428483, + 0.004120468162000179, + 0.435434490442276, + 0.010228529572486877, + 0.0164025891572237, + 0.023611584678292274, + 0.41105079650878906, + 0.010868514887988567 + ], + "credit_map": [ + 0.007820245809853077, + 0.0016191703034564853, + 0.026204636320471764, + 0.0026091060135513544, + 0.0021551537793129683, + 0.02136768400669098, + 0.001426099450327456, + 0.0038611325435340405, + 0.0015565520152449608, + 0.00035045770346187055, + 0.0058667100965976715, + 0.0006813061772845685, + 0.0004802743496838957, + 0.006098917685449123, + 0.0006479762960225344, + 0.0013290615752339363, + 0.0008232968975789845, + 0.00016602764662820846, + 0.000823390029836446, + 0.00029009574791416526, + 0.0001745360204949975, + 0.0007435030420310795, + 0.0003278124495409429, + 0.0008598754648119211, + 0.00038491334998980165, + 9.236906771548092e-05, + 0.0010491535067558289, + 0.000127435676404275, + 0.00012693044845946133, + 0.0015965934144333005, + 0.0001310424122493714, + 0.0006120302132330835, + 0.022503534331917763, + 0.004483146592974663, + 0.17490918934345245, + 0.010644431225955486, + 0.00586522975936532, + 0.18857215344905853, + 0.004259948153048754, + 0.02419685758650303, + 0.0014945033472031355, + 0.0008260412141680717, + 0.0015352941118180752, + 0.0012450529029592872, + 0.0009661110816523433, + 0.0015219644410535693, + 0.0008207567152567208, + 0.0018188052345067263, + 0.0016832971014082432, + 0.001405444461852312, + 0.0017367092659696937, + 0.001770518720149994, + 0.0012657183688133955, + 0.003671670565381646, + 0.0015331583563238382, + 0.0033360712695866823, + 0.001916516455821693, + 0.0035450924187898636, + 0.0033858041279017925, + 0.0026421525981277227, + 0.003382254857569933, + 0.0025515607558190823, + 0.0026482592802494764, + 0.0035399417392909527, + 0.022823810577392578, + 0.0183429978787899, + 0.12808524072170258, + 0.014914323575794697, + 0.036656755954027176, + 0.038780272006988525, + 0.01572350412607193, + 0.13572388887405396, + 0.0014965810114517808, + 0.0007379575981758535, + 0.001972497208043933, + 0.0008620929438620806, + 0.0008557559922337532, + 0.001311184256337583, + 0.0008489932515658438, + 0.0027834519278258085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.16034850478172302 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-definition vertical video blog shot, captured from a steady, mid-range still, the image focuses on the Baidu Bear Paw logo, composed precisely of countless luminous blue data blocks. It hovers quietly in the center of a deep, nearly black digital space. Each translucent cube emits a soft digital glow, creating a sense of high-tech sophistication. As the shot continues, the logo slowly disintegrates, its constituent blocks seemingly freed from their invisible constraints, gracefully and effortlessly drifting outwards, forming a dynamic nebula of countless blue points of light that flickers and flows in the darkness. Then, as if drawn by a central gravitational force, these scattered points of light begin to converge smoothly and orderly toward the center of the frame, swirling and rearranging, ultimately seamlessly reconstructing into a delicate model of the Earth. The contours of the land and ocean are clearly delineated by the data blocks of varying densities. The entire Earth model rotates in a gentle, steady rhythm at the center of the frame, its own light casting a faint, dynamic halo against the inky black background. The entire process is carried out under cool light and shadow, full of a sense of the future, technology and a quiet evolutionary beauty, creating a focused and poetic visual atmosphere.", + "seed": 32107, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_13/index_5278_seed_32107_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4881704572122544, + "vlm_logprob": -0.4881704572122544, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.4881704572122544, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.273922324180603 + ], + "opposite_logprob_tokens": [ + -0.39892229437828064 + ], + "logprob_reward": -1.273922324180603, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.273922324180603, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3358116149902344, + 0.0647919699549675, + 0.049071140587329865, + 0.09151104092597961, + 0.07654033601284027, + 0.16252774000167847, + 0.06040187552571297, + 0.07019588351249695, + 0.044366661459207535, + 0.044781818985939026 + ], + "credit_map": [ + 0.022763356566429138, + 0.01760186068713665, + 0.07151835411787033, + 0.019634289667010307, + 0.017363930121064186, + 0.12157388776540756, + 0.022591635584831238, + 0.04276430234313011, + 0.008575661107897758, + 0.004147662781178951, + 0.014230548404157162, + 0.006478897295892239, + 0.0036796126514673233, + 0.008638234809041023, + 0.008981002494692802, + 0.010060345754027367, + 0.004662640392780304, + 0.007745663169771433, + 0.007589562796056271, + 0.006128071341663599, + 0.007044147700071335, + 0.0034397528506815434, + 0.00610865605995059, + 0.006352648604661226, + 0.009267711080610752, + 0.005267282482236624, + 0.015687119215726852, + 0.00994432158768177, + 0.008368486538529396, + 0.02390814758837223, + 0.007668574806302786, + 0.011399395763874054, + 0.00897535216063261, + 0.004875236190855503, + 0.017042109742760658, + 0.00982007198035717, + 0.004867853596806526, + 0.01255127228796482, + 0.0057447729632258415, + 0.012663670815527439, + 0.02073024772107601, + 0.020854538306593895, + 0.018634958192706108, + 0.02770194225013256, + 0.016784338280558586, + 0.020824741572141647, + 0.015098517760634422, + 0.02189844846725464, + 0.005588453728705645, + 0.0056717172265052795, + 0.006845532916486263, + 0.008543564938008785, + 0.005523714702576399, + 0.013427210040390491, + 0.005778450518846512, + 0.009023231454193592, + 0.009444068185985088, + 0.008503680117428303, + 0.008619789034128189, + 0.008802459575235844, + 0.007073552347719669, + 0.011736761778593063, + 0.0070709786377847195, + 0.00894459430128336, + 0.00410633347928524, + 0.0055168322287499905, + 0.005701716057956219, + 0.007342275232076645, + 0.004844519775360823, + 0.006687390152364969, + 0.004662014078348875, + 0.005505579989403486, + 0.005173682700842619, + 0.007173578720539808, + 0.0047432598657906055, + 0.009532557800412178, + 0.004629512783139944, + 0.006028102245181799, + 0.0038109493907541037, + 0.0036901752464473248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -1.273922324180603 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015767297009006143 + ], + "opposite_logprob_tokens": [ + -6.501576900482178 + ], + "logprob_reward": -0.0015767297009006143, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015767297009006143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3470519185066223, + 0.062770314514637, + 0.0609121099114418, + 0.07248053699731827, + 0.08636460453271866, + 0.14242935180664062, + 0.04306606948375702, + 0.06550146639347076, + 0.059233736246824265, + 0.060189832001924515 + ], + "credit_map": [ + 0.023143041878938675, + 0.020880218595266342, + 0.07898688316345215, + 0.024922415614128113, + 0.019117895513772964, + 0.11908365786075592, + 0.020578712224960327, + 0.04033911973237991, + 0.008229613304138184, + 0.005138341803103685, + 0.012027053162455559, + 0.00687639182433486, + 0.005391126498579979, + 0.009842491708695889, + 0.005898183211684227, + 0.009367121383547783, + 0.00856888946145773, + 0.005552948452532291, + 0.011611591093242168, + 0.007458845619112253, + 0.003566988045349717, + 0.006315813399851322, + 0.01086820475757122, + 0.006968838162720203, + 0.007563536521047354, + 0.00531737832352519, + 0.011294248513877392, + 0.011523274704813957, + 0.006359984632581472, + 0.01458977535367012, + 0.008320222608745098, + 0.0075121247209608555, + 0.01840979978442192, + 0.006561041343957186, + 0.00975845754146576, + 0.013077874667942524, + 0.005544234532862902, + 0.01533565390855074, + 0.007545202039182186, + 0.010132347233593464, + 0.018744390457868576, + 0.01528338436037302, + 0.020410867407917976, + 0.025407984852790833, + 0.012230055406689644, + 0.018330425024032593, + 0.013286923058331013, + 0.01873531937599182, + 0.005365005694329739, + 0.004369506146758795, + 0.005285377148538828, + 0.006238238885998726, + 0.004387527704238892, + 0.0056924824602901936, + 0.005521158222109079, + 0.00620677787810564, + 0.0075747049413621426, + 0.009554537944495678, + 0.008344905450940132, + 0.011603808961808681, + 0.0065062870271503925, + 0.010211246088147163, + 0.00463858200237155, + 0.007067401893436909, + 0.006283140741288662, + 0.010262587107717991, + 0.006333428435027599, + 0.00797092542052269, + 0.0062093064188957214, + 0.008120030164718628, + 0.005863451864570379, + 0.008190873079001904, + 0.006704811472445726, + 0.009406968019902706, + 0.006177193019539118, + 0.009242091327905655, + 0.006931627634912729, + 0.007540987804532051, + 0.007715502288192511, + 0.006470654159784317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0015767297009006143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019481267081573606 + ], + "opposite_logprob_tokens": [ + -6.251948356628418 + ], + "logprob_reward": -0.0019481267081573606, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019481267081573606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3028930723667145, + 0.12318842113018036, + 0.07228757441043854, + 0.10063277184963226, + 0.11089963465929031, + 0.0824766606092453, + 0.04120805114507675, + 0.05251502990722656, + 0.07463336735963821, + 0.03926529735326767 + ], + "credit_map": [ + 0.022196996957063675, + 0.014944678172469139, + 0.06443070620298386, + 0.015070312656462193, + 0.015194470062851906, + 0.11142443865537643, + 0.01601746305823326, + 0.043614067137241364, + 0.016939368098974228, + 0.00706697441637516, + 0.02509947493672371, + 0.008036766201257706, + 0.008871505968272686, + 0.029656153172254562, + 0.008561914786696434, + 0.018956271931529045, + 0.009296063333749771, + 0.008482499979436398, + 0.012391503900289536, + 0.006472072098404169, + 0.00815475545823574, + 0.010417250916361809, + 0.008009335026144981, + 0.009064099751412868, + 0.009420877322554588, + 0.0060793617740273476, + 0.016192177310585976, + 0.01311876717954874, + 0.009352030232548714, + 0.024019166827201843, + 0.01100003533065319, + 0.011450362391769886, + 0.025151489302515984, + 0.012242143973708153, + 0.015714066103100777, + 0.01636902242898941, + 0.007948018610477448, + 0.011564468964934349, + 0.011448269709944725, + 0.010462168604135513, + 0.01179024949669838, + 0.011165843345224857, + 0.014814695343375206, + 0.011316554620862007, + 0.004289144184440374, + 0.009036378003656864, + 0.008223379030823708, + 0.01184042263776064, + 0.006129628978669643, + 0.003757854923605919, + 0.005483580287545919, + 0.005334366112947464, + 0.0037073250859975815, + 0.006108112633228302, + 0.004704265855252743, + 0.005982921924442053, + 0.008312474004924297, + 0.007570788264274597, + 0.006634368561208248, + 0.007432993035763502, + 0.004252287093549967, + 0.007120704744011164, + 0.005169411189854145, + 0.006022012792527676, + 0.008262323215603828, + 0.01223713718354702, + 0.008000987581908703, + 0.011456456035375595, + 0.008274482563138008, + 0.009872721508145332, + 0.007801857776939869, + 0.00872740987688303, + 0.0038454115856438875, + 0.005983391776680946, + 0.0049715591594576836, + 0.0068376376293599606, + 0.003913229797035456, + 0.005724797956645489, + 0.00376436416991055, + 0.004224912729114294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0019481267081573606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634410560131073 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634410560131073, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634410560131073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06263092160224915, + 0.09503751248121262, + 0.031617119908332825, + 0.04119572043418884, + 0.12617644667625427, + 0.21194101870059967, + 0.0650620236992836, + 0.141368106007576, + 0.10927098989486694, + 0.11570018529891968 + ], + "credit_map": [ + 0.006764489226043224, + 0.005616225767880678, + 0.0069270627573132515, + 0.00904414989054203, + 0.006461611948907375, + 0.012269997969269753, + 0.007269469555467367, + 0.008277908898890018, + 0.011365083046257496, + 0.007372972089797258, + 0.024171538650989532, + 0.00676796305924654, + 0.005525910295546055, + 0.016083331778645515, + 0.006874830462038517, + 0.016875887289643288, + 0.004439546726644039, + 0.0026083295233547688, + 0.004591320175677538, + 0.00401208596304059, + 0.0024529253132641315, + 0.005323007237166166, + 0.003938867244869471, + 0.004251035861670971, + 0.0038372355047613382, + 0.004299881402403116, + 0.005826891399919987, + 0.004649697802960873, + 0.006499285809695721, + 0.007348689716309309, + 0.004484051838517189, + 0.004249983001500368, + 0.028650445863604546, + 0.010689022950828075, + 0.013255644589662552, + 0.014838469214737415, + 0.019999219104647636, + 0.014097223989665508, + 0.012341227382421494, + 0.012305185198783875, + 0.025222979485988617, + 0.027035491541028023, + 0.028651272878050804, + 0.028969043865799904, + 0.019962569698691368, + 0.03160182386636734, + 0.02250666730105877, + 0.027991175651550293, + 0.005348915234208107, + 0.005760934203863144, + 0.010190663859248161, + 0.008084014058113098, + 0.008986609987914562, + 0.009269420057535172, + 0.007905982434749603, + 0.009515480138361454, + 0.01429053209722042, + 0.015250799246132374, + 0.017929548397660255, + 0.019489113241434097, + 0.021912861615419388, + 0.019999345764517784, + 0.014329306781291962, + 0.018166592344641685, + 0.007632570341229439, + 0.013123700395226479, + 0.015476422384381294, + 0.02647724375128746, + 0.010317563079297543, + 0.01463133841753006, + 0.011315750889480114, + 0.010296400636434555, + 0.01510744821280241, + 0.017386317253112793, + 0.018379246816039085, + 0.01687663607299328, + 0.012205197475850582, + 0.014411093667149544, + 0.008845752105116844, + 0.012488496489822865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.02634410560131073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical aspect ratio?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial scene focused on the Baidu Bear Paw logo composed of luminous blue data blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the logo disintegrate into scattered blue points of light that drift outwards before converging to form a model of the Earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final Earth model clearly delineated with visible contours of land and ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned as a steady, mid-range still shot throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370609998703003 + ], + "opposite_logprob_tokens": [ + -0.3870610296726227 + ], + "logprob_reward": -1.1370609998703003, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370609998703003, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16419394314289093, + 0.09862756729125977, + 0.07482727617025375, + 0.055748093873262405, + 0.2006448358297348, + 0.10430470108985901, + 0.06410333514213562, + 0.1145687997341156, + 0.03803921490907669, + 0.0849423035979271 + ], + "credit_map": [ + 0.013048646040260792, + 0.00897150207310915, + 0.04423639178276062, + 0.007528297137469053, + 0.009306726977229118, + 0.055757418274879456, + 0.010884607210755348, + 0.014460356906056404, + 0.01343002263456583, + 0.008504665456712246, + 0.022060459479689598, + 0.009658081457018852, + 0.007326225750148296, + 0.015157267451286316, + 0.00960621703416109, + 0.012884634546935558, + 0.007083018776029348, + 0.007707464508712292, + 0.010620883665978909, + 0.00882049836218357, + 0.009094116277992725, + 0.009259535931050777, + 0.013275008648633957, + 0.008966750465333462, + 0.0055262241512537, + 0.005478528328239918, + 0.008412930183112621, + 0.007714468985795975, + 0.0076781343668699265, + 0.009157523512840271, + 0.006338797975331545, + 0.005441485904157162, + 0.04702797159552574, + 0.016634928062558174, + 0.03806246817111969, + 0.030436141416430473, + 0.012921207584440708, + 0.014091436751186848, + 0.017080776393413544, + 0.024389903992414474, + 0.013276965357363224, + 0.012892510741949081, + 0.012883749790489674, + 0.013208657503128052, + 0.011306826956570148, + 0.01592920906841755, + 0.011005911976099014, + 0.013800867833197117, + 0.008682642132043839, + 0.006063886918127537, + 0.010662831366062164, + 0.007360544055700302, + 0.0070273433811962605, + 0.007222037296742201, + 0.008589559234678745, + 0.008494493551552296, + 0.019760245457291603, + 0.014760274440050125, + 0.013972265645861626, + 0.016692250967025757, + 0.008926033042371273, + 0.016935670748353004, + 0.009359316900372505, + 0.014162742532789707, + 0.0035860766656696796, + 0.0031691661570221186, + 0.005472489632666111, + 0.0064223469235002995, + 0.0036202918272465467, + 0.005598144140094519, + 0.004648077301681042, + 0.005522620398551226, + 0.010300416499376297, + 0.013582009822130203, + 0.009352486580610275, + 0.015920117497444153, + 0.008689917623996735, + 0.00990675762295723, + 0.008768877014517784, + 0.008421720936894417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -1.1370609998703003 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_14/index_tensor([601, 601])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_14/index_tensor([601, 601])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..526f3874bd66d968c492e0bdeab751ba4b8007de --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_14/index_tensor([601, 601])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, static, extreme close-up depicts an intricately patterned small lizard, its body a detailed mosaic of sandy beige and russet scales, perfectly mimicking the sun-drenched sandstone rock below. The midday sun casts sharp, microscopic shadows across the rock's weathered ochre surface. The lizard freezes mid-step, one delicate front leg suspended above the stone's grainy texture, its entire body lowered in a posture of pure, primal instinct. As a large, amorphous shadow glides swiftly and silently across the scene, the lizard, momentarily trapped in cool darkness, remains completely petrified. Its muscles tense, its beady black eyes wide and unblinking, its breath held as it desperately and silently attempts to merge with the stone, creating a palpable atmosphere of high-stakes tension and a frozen moment of survival.", + "seed": 28202, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_14/index_601_seed_28202_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10794543148949742, + "vlm_logprob": -0.10794543148949742, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.10794543148949742, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2858482003211975 + ], + "opposite_logprob_tokens": [ + -1.5358481407165527 + ], + "logprob_reward": -0.2858482003211975, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2858482003211975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09173569083213806, + 0.06043153628706932, + 0.03423376753926277, + 0.31083422899246216, + 0.11755526065826416, + 0.1177264153957367, + 0.14116699993610382, + 0.06705472618341446, + 0.022220809012651443, + 0.0370405875146389 + ], + "credit_map": [ + 0.010340113192796707, + 0.009340492077171803, + 0.02054370753467083, + 0.014854236505925655, + 0.0101041654124856, + 0.00949800293892622, + 0.009851651266217232, + 0.007203314919024706, + 0.009974333457648754, + 0.004732553381472826, + 0.007449989207088947, + 0.008927990682423115, + 0.00602934043854475, + 0.010905729606747627, + 0.007254172116518021, + 0.0051574185490608215, + 0.004873985890299082, + 0.00410323916003108, + 0.005913704168051481, + 0.002931127557530999, + 0.00432088878005743, + 0.00443409476429224, + 0.0033266814425587654, + 0.004330045077949762, + 0.01609145477414131, + 0.006716225296258926, + 0.014393552206456661, + 0.014575213193893433, + 0.02688770927488804, + 0.009674116037786007, + 0.11602240055799484, + 0.10647355020046234, + 0.014447786845266819, + 0.014960386790335178, + 0.009371208027005196, + 0.010237705893814564, + 0.015638088807463646, + 0.019358856603503227, + 0.017606189474463463, + 0.015935026109218597, + 0.011168821714818478, + 0.019172828644514084, + 0.006556012202054262, + 0.018799273297190666, + 0.008814094588160515, + 0.020296072587370872, + 0.01639500819146633, + 0.01652430184185505, + 0.018613865599036217, + 0.016896173357963562, + 0.01712932623922825, + 0.01780981943011284, + 0.020869189873337746, + 0.017460891976952553, + 0.01604168862104416, + 0.01634603925049305, + 0.007270009722560644, + 0.009641360491514206, + 0.008000349625945091, + 0.011670079082250595, + 0.009288633242249489, + 0.006804737728089094, + 0.008157319389283657, + 0.006222235504537821, + 0.00286286068148911, + 0.0033321089576929808, + 0.0034207201097160578, + 0.004460030701011419, + 0.0015490635996684432, + 0.002144434954971075, + 0.0016210428439080715, + 0.0028305461164563894, + 0.002785602817311883, + 0.0033644901122897863, + 0.0022948943078517914, + 0.005201458465307951, + 0.0061406842432916164, + 0.005121625028550625, + 0.006598412990570068, + 0.005533419083803892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.2858482003211975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422479934990406 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.01422479934990406, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422479934990406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07467558979988098, + 0.06587709486484528, + 0.06391913443803787, + 0.4394764304161072, + 0.041053399443626404, + 0.08292175084352493, + 0.11086306720972061, + 0.04865268990397453, + 0.05930722504854202, + 0.01325361616909504 + ], + "credit_map": [ + 0.008146797306835651, + 0.01015316229313612, + 0.008721292950212955, + 0.009802727960050106, + 0.012090546078979969, + 0.010518339462578297, + 0.00842369720339775, + 0.006819015834480524, + 0.008637919090688229, + 0.006758124101907015, + 0.008293376304209232, + 0.011651846580207348, + 0.008550704456865788, + 0.012474417686462402, + 0.005676428321748972, + 0.003834279952570796, + 0.008925794623792171, + 0.005096477922052145, + 0.012133406475186348, + 0.005185694899410009, + 0.00945163331925869, + 0.009479566477239132, + 0.006024994421750307, + 0.007621558383107185, + 0.0247185081243515, + 0.010183102451264858, + 0.023446284234523773, + 0.021333107724785805, + 0.03651849180459976, + 0.011485614813864231, + 0.16189803183078766, + 0.14989325404167175, + 0.004072058014571667, + 0.0050080157816410065, + 0.0028991533908993006, + 0.0032119813840836287, + 0.006513312924653292, + 0.006477874703705311, + 0.00790601596236229, + 0.004964986350387335, + 0.00478870514780283, + 0.01281223725527525, + 0.0038097284268587828, + 0.012412181124091148, + 0.016690215095877647, + 0.01186842005699873, + 0.012962106615304947, + 0.0075781503692269325, + 0.01536288857460022, + 0.014998448081314564, + 0.008894022554159164, + 0.014545360580086708, + 0.01709974929690361, + 0.012167854234576225, + 0.01839381828904152, + 0.009400924667716026, + 0.005187112372368574, + 0.005416988395154476, + 0.005846147891134024, + 0.007525722961872816, + 0.006058851256966591, + 0.005028097424656153, + 0.008850292302668095, + 0.004739473108202219, + 0.010016229934990406, + 0.009862176142632961, + 0.008159860968589783, + 0.011079289019107819, + 0.005436178762465715, + 0.004748726263642311, + 0.00500314449891448, + 0.005001617595553398, + 0.0011466164141893387, + 0.001220924430526793, + 0.000981340417638421, + 0.0018731505842879415, + 0.002504672622308135, + 0.0016703837318345904, + 0.001990435877814889, + 0.0018660910427570343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.01422479934990406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996463179588318 + ], + "opposite_logprob_tokens": [ + -1.8049646615982056 + ], + "logprob_reward": -0.17996463179588318, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996463179588318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2012404501438141, + 0.13309548795223236, + 0.02893468737602234, + 0.4807756245136261, + 0.01904210075736046, + 0.028278091922402382, + 0.03313129022717476, + 0.028847869485616684, + 0.0174181517213583, + 0.029236238449811935 + ], + "credit_map": [ + 0.022047849372029305, + 0.014081398956477642, + 0.05958372727036476, + 0.028892485424876213, + 0.0165270809084177, + 0.02992771752178669, + 0.020143883302807808, + 0.010036305524408817, + 0.012813017703592777, + 0.014709165319800377, + 0.01625722646713257, + 0.019156569615006447, + 0.020796390250325203, + 0.028869647532701492, + 0.012044133618474007, + 0.008449327200651169, + 0.0035610615741461515, + 0.0027406266890466213, + 0.004630399867892265, + 0.0021446857135742903, + 0.0031361898873001337, + 0.0038325090426951647, + 0.003608272410929203, + 0.005280941724777222, + 0.02392079308629036, + 0.010786257684230804, + 0.021134717389941216, + 0.021991737186908722, + 0.039389919489622116, + 0.012427524663507938, + 0.1819034218788147, + 0.16922129690647125, + 0.0018645462114363909, + 0.002665614942088723, + 0.001289258012548089, + 0.0013277530670166016, + 0.0029245673213154078, + 0.0032044686377048492, + 0.0033398529049009085, + 0.00242603849619627, + 0.0020651628728955984, + 0.004175545182079077, + 0.0014076696243137121, + 0.0038191922940313816, + 0.004174437373876572, + 0.0036159660667181015, + 0.005292096640914679, + 0.0037280230317264795, + 0.0036061739083379507, + 0.0038781138136982918, + 0.0036687818355858326, + 0.00448041595518589, + 0.004231680184602737, + 0.004844774026423693, + 0.004819734487682581, + 0.003601616946980357, + 0.0037226842250674963, + 0.003351649735122919, + 0.004387014079838991, + 0.0034014887642115355, + 0.003215976757928729, + 0.0036087511107325554, + 0.004569622687995434, + 0.0025906814262270927, + 0.002270875032991171, + 0.0030875077936798334, + 0.002653904724866152, + 0.0026724664494395256, + 0.001348336460068822, + 0.0016515747411176562, + 0.0012489738874137402, + 0.0024845132138580084, + 0.0021039280109107494, + 0.0031616510823369026, + 0.0017508761957287788, + 0.002291525946930051, + 0.006981818005442619, + 0.00428478978574276, + 0.004788515157997608, + 0.0038731349632143974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.17996463179588318 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464135454967618 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464135454967618, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464135454967618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17181184887886047, + 0.18293574452400208, + 0.11972030997276306, + 0.14368239045143127, + 0.04513012617826462, + 0.08899468928575516, + 0.05860341712832451, + 0.0903695598244667, + 0.03807898238301277, + 0.060672905296087265 + ], + "credit_map": [ + 0.015704788267612457, + 0.01141057163476944, + 0.04982277378439903, + 0.018453232944011688, + 0.01609255187213421, + 0.03497915342450142, + 0.014566527679562569, + 0.010782255791127682, + 0.01720869354903698, + 0.02051507867872715, + 0.02150874398648739, + 0.026874301955103874, + 0.02838219702243805, + 0.03577296808362007, + 0.019118813797831535, + 0.0135549521073699, + 0.016323300078511238, + 0.007657041773200035, + 0.024448931217193604, + 0.010807013139128685, + 0.013535480946302414, + 0.016977667808532715, + 0.013061348348855972, + 0.01690952107310295, + 0.0177586879581213, + 0.008916595950722694, + 0.021405670791864395, + 0.00829251203685999, + 0.0135494414716959, + 0.009447140619158745, + 0.032742928713560104, + 0.03156941384077072, + 0.0042122649028897285, + 0.00522878672927618, + 0.0029193253722041845, + 0.003398844040930271, + 0.00695997616276145, + 0.0071216425858438015, + 0.009068974293768406, + 0.006220312789082527, + 0.0050514694303274155, + 0.016876012086868286, + 0.003601131960749626, + 0.012423609383404255, + 0.013008225709199905, + 0.012876654975116253, + 0.015127928927540779, + 0.010029647499322891, + 0.00714776199311018, + 0.007733843754976988, + 0.006432708352804184, + 0.00825705286115408, + 0.00967319868505001, + 0.006885943002998829, + 0.007738879881799221, + 0.004734028596431017, + 0.011713250540196896, + 0.010309790261089802, + 0.013212691061198711, + 0.011933562345802784, + 0.009827451780438423, + 0.009984455071389675, + 0.014853985980153084, + 0.008534383028745651, + 0.0038235001266002655, + 0.007225092966109514, + 0.0065636057406663895, + 0.0061437697149813175, + 0.003665939439088106, + 0.0032909659203141928, + 0.002515745349228382, + 0.004850362427532673, + 0.005098611116409302, + 0.0053521934896707535, + 0.004009803757071495, + 0.006599614396691322, + 0.012015372514724731, + 0.008780446834862232, + 0.010452554561197758, + 0.008364308625459671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.00464135454967618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05504817143082619 + ], + "opposite_logprob_tokens": [ + -2.9300482273101807 + ], + "logprob_reward": -0.05504817143082619, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05504817143082619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0717807486653328, + 0.0911213830113411, + 0.0319637805223465, + 0.5982953906059265, + 0.040259137749671936, + 0.0881674587726593, + 0.01873675547540188, + 0.02510996162891388, + 0.017267975956201553, + 0.01729743182659149 + ], + "credit_map": [ + 0.008276790380477905, + 0.005949088372290134, + 0.01988944411277771, + 0.008808543905615807, + 0.00575611786916852, + 0.011812079697847366, + 0.007717195898294449, + 0.0035714854020625353, + 0.013014581054449081, + 0.009418841451406479, + 0.011775853112339973, + 0.015247093513607979, + 0.01392210740596056, + 0.016467276960611343, + 0.007036745548248291, + 0.004238884896039963, + 0.00419854698702693, + 0.00250706123188138, + 0.006337946746498346, + 0.0025849700905382633, + 0.003912811633199453, + 0.004078594967722893, + 0.003673373721539974, + 0.004670475609600544, + 0.024527553468942642, + 0.01184957753866911, + 0.014905999414622784, + 0.027887430042028427, + 0.049832385033369064, + 0.016785981133580208, + 0.2346101701259613, + 0.21789632737636566, + 0.004724064376205206, + 0.005062492098659277, + 0.003094257088378072, + 0.002696191892027855, + 0.005700633395463228, + 0.007007997948676348, + 0.006927402224391699, + 0.005046099424362183, + 0.005096676293760538, + 0.01427414733916521, + 0.004146439954638481, + 0.012264047749340534, + 0.013238906860351562, + 0.016301652416586876, + 0.013404575176537037, + 0.009441010653972626, + 0.0017808224074542522, + 0.002028407296165824, + 0.002574654296040535, + 0.001960719469934702, + 0.0024651475250720978, + 0.0030912458896636963, + 0.0028046099469065666, + 0.0020311486441642046, + 0.0036253887228667736, + 0.0030555452685803175, + 0.003280429169535637, + 0.0036914129741489887, + 0.0035448926500976086, + 0.0028420304879546165, + 0.0030314165633171797, + 0.002038846258074045, + 0.0023051390890032053, + 0.0030075800605118275, + 0.0026093136984854937, + 0.0034139235503971577, + 0.001186244422569871, + 0.0015290111768990755, + 0.001247064908966422, + 0.0019696997478604317, + 0.0013765390031039715, + 0.0014786573592573404, + 0.001178512698970735, + 0.0024076274130493402, + 0.0024187052622437477, + 0.0023760509211570024, + 0.003062026808038354, + 0.0029993110802024603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.05504817143082619 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static, extreme close-up depicts an intricately patterned small lizard, its body a detailed mosaic of sandy beige and russet scales, perfectly mimicking the sun-drenched sandstone rock below. The midday sun casts sharp, microscopic shadows across the rock's weathered ochre surface. The lizard freezes mid-step, one delicate front leg suspended above the stone's grainy texture, its entire body lowered in a posture of pure, primal instinct. As a large, amorphous shadow glides swiftly and silently across the scene, the lizard, momentarily trapped in cool darkness, remains completely petrified. Its muscles tense, its beady black eyes wide and unblinking, its breath held as it desperately and silently attempts to merge with the stone, creating a palpable atmosphere of high-stakes tension and a frozen moment of survival.", + "seed": 28203, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_14/index_601_seed_28203_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07172482833266258, + "vlm_logprob": -0.07172482833266258, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.07172482833266258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053027477115392685 + ], + "opposite_logprob_tokens": [ + -3.553027391433716 + ], + "logprob_reward": -0.053027477115392685, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.053027477115392685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38156652450561523, + 0.040521565824747086, + 0.04634690657258034, + 0.017738377675414085, + 0.02832324244081974, + 0.017310163006186485, + 0.13729950785636902, + 0.10793161392211914, + 0.05162561312317848, + 0.17133644223213196 + ], + "credit_map": [ + 0.1407797932624817, + 0.018739841878414154, + 0.00899241678416729, + 0.015682023018598557, + 0.009112873114645481, + 0.018947042524814606, + 0.013544597662985325, + 0.15576797723770142, + 0.005348858889192343, + 0.00486748805269599, + 0.006701878271996975, + 0.004957920406013727, + 0.004058747086673975, + 0.005633770488202572, + 0.0029966949950903654, + 0.005956211127340794, + 0.005964852403849363, + 0.005709656048566103, + 0.0059390016831457615, + 0.00486666290089488, + 0.006128845736384392, + 0.008480716496706009, + 0.0037060105241835117, + 0.005551164038479328, + 0.0033903077710419893, + 0.002255796454846859, + 0.0017631170339882374, + 0.0018400730332359672, + 0.0020924555137753487, + 0.0027617209125310183, + 0.0014596318360418081, + 0.002175276866182685, + 0.004930656868964434, + 0.0028643475379794836, + 0.004235507920384407, + 0.0025733569636940956, + 0.002724412363022566, + 0.005297684110701084, + 0.0021705308463424444, + 0.003526749089360237, + 0.0026895373594015837, + 0.0017391822766512632, + 0.0019199446542188525, + 0.0013664126163348556, + 0.002089750487357378, + 0.0030346079729497433, + 0.0019045270746573806, + 0.002566202310845256, + 0.021605568006634712, + 0.015440336428582668, + 0.01784530095756054, + 0.017863115295767784, + 0.008853938430547714, + 0.0284821018576622, + 0.011090567335486412, + 0.01611858606338501, + 0.015251045115292072, + 0.0072392551228404045, + 0.006211154628545046, + 0.003796725533902645, + 0.004409563262015581, + 0.0311567485332489, + 0.014085215516388416, + 0.025781916454434395, + 0.0066264597699046135, + 0.005085913930088282, + 0.00918416865170002, + 0.004183789249509573, + 0.005706591997295618, + 0.009754317812621593, + 0.004962562583386898, + 0.006121811922639608, + 0.007054806686937809, + 0.006884914357215166, + 0.010338595137000084, + 0.005842868704348803, + 0.011372004635632038, + 0.062423594295978546, + 0.0076303924433887005, + 0.05978928878903389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.053027477115392685 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900365442037582 + ], + "opposite_logprob_tokens": [ + -2.5790035724639893 + ], + "logprob_reward": -0.07900365442037582, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900365442037582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23913894593715668, + 0.0858406275510788, + 0.10343553870916367, + 0.05348709225654602, + 0.05996664986014366, + 0.0231022872030735, + 0.11312553286552429, + 0.0643153265118599, + 0.07032690197229385, + 0.18726110458374023 + ], + "credit_map": [ + 0.07655660063028336, + 0.012392055243253708, + 0.012382297776639462, + 0.012871273793280125, + 0.01512154284864664, + 0.014062085188925266, + 0.010030003264546394, + 0.08572306483983994, + 0.013494969345629215, + 0.013470413163304329, + 0.008915340527892113, + 0.009861285798251629, + 0.012018725275993347, + 0.012829631567001343, + 0.005342923570424318, + 0.009907333180308342, + 0.016662098467350006, + 0.013523483648896217, + 0.011621533893048763, + 0.010520607233047485, + 0.010523291304707527, + 0.017755601555109024, + 0.010485564358532429, + 0.01234335359185934, + 0.0077568781562149525, + 0.00626758998259902, + 0.005049108061939478, + 0.004649005830287933, + 0.0058687711134552956, + 0.010991502553224564, + 0.005453045479953289, + 0.007451181765645742, + 0.009146474301815033, + 0.006977437995374203, + 0.007723620161414146, + 0.00709056481719017, + 0.005652646068483591, + 0.012264405377209187, + 0.003698562504723668, + 0.007412936072796583, + 0.0033687474206089973, + 0.002667215885594487, + 0.0028501925989985466, + 0.0023825555108487606, + 0.003073896514251828, + 0.0029140361584722996, + 0.0029267678037285805, + 0.002918874379247427, + 0.016849232837557793, + 0.013227901421487331, + 0.018850579857826233, + 0.009849700145423412, + 0.009865078143775463, + 0.01863749884068966, + 0.008013279177248478, + 0.017832251265645027, + 0.008265411481261253, + 0.0045681605115532875, + 0.004576331935822964, + 0.003379909321665764, + 0.003445110283792019, + 0.01699240133166313, + 0.00786461029201746, + 0.015223382972180843, + 0.0067232525907456875, + 0.008107542060315609, + 0.008617457002401352, + 0.009823638945817947, + 0.007979683578014374, + 0.00688443798571825, + 0.0124882897362113, + 0.009702595882117748, + 0.006120756734162569, + 0.006970444694161415, + 0.007841946557164192, + 0.0056627485901117325, + 0.013705044984817505, + 0.06852513551712036, + 0.011006547138094902, + 0.06742846965789795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.07900365442037582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799979954957962 + ], + "opposite_logprob_tokens": [ + -1.8049980401992798 + ], + "logprob_reward": -0.1799979954957962, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799979954957962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3302743136882782, + 0.04953332990407944, + 0.048335835337638855, + 0.04033207148313522, + 0.05138782784342766, + 0.03740187734365463, + 0.21544718742370605, + 0.10136675089597702, + 0.047883883118629456, + 0.07803695648908615 + ], + "credit_map": [ + 0.11252205818891525, + 0.015763595700263977, + 0.013080263510346413, + 0.01641766168177128, + 0.01084961462765932, + 0.019984614104032516, + 0.013252700679004192, + 0.12840378284454346, + 0.008279194124042988, + 0.006213589105755091, + 0.005044042132794857, + 0.00446567265316844, + 0.00613939855247736, + 0.008818281814455986, + 0.003491506213322282, + 0.007081645540893078, + 0.005378199275583029, + 0.006026797462254763, + 0.005792213603854179, + 0.0046838391572237015, + 0.005466989241540432, + 0.007677859161049128, + 0.006564920302480459, + 0.006745014805346727, + 0.007570837624371052, + 0.003497085999697447, + 0.005346056539565325, + 0.0047533148899674416, + 0.0044441926293075085, + 0.006792635191231966, + 0.0026128448080271482, + 0.00531510217115283, + 0.00741627998650074, + 0.007381931412965059, + 0.008108151145279408, + 0.005274576134979725, + 0.006740981247276068, + 0.005737388040870428, + 0.004050014540553093, + 0.006678506266325712, + 0.005989508703351021, + 0.0038466062396764755, + 0.007055751513689756, + 0.003581720869988203, + 0.004650609567761421, + 0.004368601366877556, + 0.004239199683070183, + 0.0036698803305625916, + 0.03353819623589516, + 0.027023861184716225, + 0.03063439391553402, + 0.02253112941980362, + 0.014941692352294922, + 0.03630978614091873, + 0.015550201758742332, + 0.0349179282784462, + 0.012919800356030464, + 0.010130769573152065, + 0.009138060733675957, + 0.006982012651860714, + 0.008137721568346024, + 0.023152120411396027, + 0.012637116946280003, + 0.01826915331184864, + 0.005139526445418596, + 0.006463556084781885, + 0.005644599907100201, + 0.0054685152135789394, + 0.005474166013300419, + 0.007423833478242159, + 0.006154737435281277, + 0.00611494854092598, + 0.006012957077473402, + 0.005376123357564211, + 0.0071568177081644535, + 0.004736693110316992, + 0.009751618839800358, + 0.017073186114430428, + 0.011714933440089226, + 0.016214625909924507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.1799979954957962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289937525987625 + ], + "opposite_logprob_tokens": [ + -3.773289918899536 + ], + "logprob_reward": -0.023289937525987625, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289937525987625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10036126524209976, + 0.02736521139740944, + 0.01883205585181713, + 0.024199718609452248, + 0.0298673864454031, + 0.025491997599601746, + 0.3232499659061432, + 0.22488035261631012, + 0.08302713185548782, + 0.1427249014377594 + ], + "credit_map": [ + 0.032844316214323044, + 0.00568859139457345, + 0.005636536981910467, + 0.004691363777965307, + 0.005287663545459509, + 0.006691883783787489, + 0.004608334973454475, + 0.03491256758570671, + 0.004542083479464054, + 0.0031247977167367935, + 0.0030468315817415714, + 0.0027740232180804014, + 0.0031469613313674927, + 0.004768472630530596, + 0.0019842134788632393, + 0.003977827727794647, + 0.0024047743063420057, + 0.002107723383232951, + 0.002041274681687355, + 0.0015362279955297709, + 0.0024972755927592516, + 0.004205760546028614, + 0.0020013349130749702, + 0.0020376851316541433, + 0.00354829546995461, + 0.0033174471464008093, + 0.003235993906855583, + 0.002442916389554739, + 0.002205470111221075, + 0.0043148016557097435, + 0.001954833511263132, + 0.0031799597200006247, + 0.004029618110507727, + 0.005029609426856041, + 0.0038439957424998283, + 0.002969115972518921, + 0.004769756458699703, + 0.0034544533118605614, + 0.0023548253811895847, + 0.0034160101786255836, + 0.004141927231103182, + 0.0021692144218832254, + 0.005185646936297417, + 0.002199021400883794, + 0.002848298056051135, + 0.0032824298832565546, + 0.0028359792195260525, + 0.00282947882078588, + 0.048981476575136185, + 0.040077533572912216, + 0.043159205466508865, + 0.027681482955813408, + 0.025658342987298965, + 0.06077329069375992, + 0.021937323734164238, + 0.054981332272291183, + 0.03068278171122074, + 0.015192544087767601, + 0.009100444614887238, + 0.010238563641905785, + 0.012651070952415466, + 0.06542392075061798, + 0.029015110805630684, + 0.052575916051864624, + 0.010000765323638916, + 0.00748682115226984, + 0.010602292604744434, + 0.007216013502329588, + 0.011663597077131271, + 0.014761489816009998, + 0.009773573838174343, + 0.011522585526108742, + 0.0031946918461471796, + 0.0035487532150000334, + 0.0033403027337044477, + 0.0026386864483356476, + 0.009126669727265835, + 0.05870208889245987, + 0.004573565907776356, + 0.05760014057159424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.023289937525987625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305077105760574 + ], + "opposite_logprob_tokens": [ + -3.7733051776885986 + ], + "logprob_reward": -0.023305077105760574, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305077105760574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3617538511753082, + 0.07248899340629578, + 0.04133491590619087, + 0.0422324575483799, + 0.050004299730062485, + 0.03693539649248123, + 0.09092016518115997, + 0.17690081894397736, + 0.06330225616693497, + 0.06412690132856369 + ], + "credit_map": [ + 0.1237141415476799, + 0.017714375630021095, + 0.016443759202957153, + 0.017440514639019966, + 0.01479929219931364, + 0.020885594189167023, + 0.015668991953134537, + 0.13508716225624084, + 0.010227300226688385, + 0.011001870036125183, + 0.008102010935544968, + 0.008349305018782616, + 0.00940919853746891, + 0.012170476838946342, + 0.0048074424266815186, + 0.00842138472944498, + 0.006421764846891165, + 0.005244491621851921, + 0.003901988733559847, + 0.003622898366302252, + 0.004333042073994875, + 0.009400478564202785, + 0.003900352632626891, + 0.004509898833930492, + 0.005814393050968647, + 0.004830049816519022, + 0.00469500245526433, + 0.00298072281293571, + 0.003815822070464492, + 0.010470517911016941, + 0.0032750258687883615, + 0.006350921932607889, + 0.006888001225888729, + 0.006772555410861969, + 0.007889732718467712, + 0.004921594634652138, + 0.007459465879946947, + 0.005830793175846338, + 0.004822753835469484, + 0.005419401917606592, + 0.005691208876669407, + 0.005255242343991995, + 0.006708797533065081, + 0.002872767858207226, + 0.004397857002913952, + 0.004335339646786451, + 0.0037077278830111027, + 0.00396645488217473, + 0.012799497693777084, + 0.0083438316360116, + 0.015641162171959877, + 0.007136576808989048, + 0.009349574334919453, + 0.015678366646170616, + 0.006664604879915714, + 0.01530655100941658, + 0.024484779685735703, + 0.012787227518856525, + 0.005400712601840496, + 0.004757924471050501, + 0.008845345117151737, + 0.055941708385944366, + 0.02282789722084999, + 0.04185521975159645, + 0.006738340016454458, + 0.009381785057485104, + 0.008221883326768875, + 0.0074412161484360695, + 0.005233721807599068, + 0.00922670029103756, + 0.009226882830262184, + 0.007831722497940063, + 0.005208710208535194, + 0.005132511258125305, + 0.007209960836917162, + 0.004485090728849173, + 0.006835573818534613, + 0.01374673843383789, + 0.00836279708892107, + 0.01314552128314972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.023305077105760574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static, extreme close-up depicts an intricately patterned small lizard, its body a detailed mosaic of sandy beige and russet scales, perfectly mimicking the sun-drenched sandstone rock below. The midday sun casts sharp, microscopic shadows across the rock's weathered ochre surface. The lizard freezes mid-step, one delicate front leg suspended above the stone's grainy texture, its entire body lowered in a posture of pure, primal instinct. As a large, amorphous shadow glides swiftly and silently across the scene, the lizard, momentarily trapped in cool darkness, remains completely petrified. Its muscles tense, its beady black eyes wide and unblinking, its breath held as it desperately and silently attempts to merge with the stone, creating a palpable atmosphere of high-stakes tension and a frozen moment of survival.", + "seed": 28204, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_14/index_601_seed_28204_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0381416748277843, + "vlm_logprob": -0.0381416748277843, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0381416748277843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06899885088205338 + ], + "opposite_logprob_tokens": [ + -3.0689988136291504 + ], + "logprob_reward": -0.06899885088205338, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06899885088205338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18795187771320343, + 0.12657111883163452, + 0.054718539118766785, + 0.13296470046043396, + 0.08834574371576309, + 0.04184366762638092, + 0.08888708055019379, + 0.07962989807128906, + 0.06125177443027496, + 0.13783568143844604 + ], + "credit_map": [ + 0.028171567246317863, + 0.03708256036043167, + 0.022949250414967537, + 0.016501637175679207, + 0.022837059572339058, + 0.026334695518016815, + 0.01604706235229969, + 0.01802801340818405, + 0.019546816125512123, + 0.01675291731953621, + 0.013676730915904045, + 0.018492357805371284, + 0.018269138410687447, + 0.01392934937030077, + 0.013349293731153011, + 0.012554504908621311, + 0.0077385748736560345, + 0.007684566546231508, + 0.004483060911297798, + 0.007316765375435352, + 0.008502188138663769, + 0.008407720364630222, + 0.005272010341286659, + 0.005313642788678408, + 0.016797920688986778, + 0.016621731221675873, + 0.014633465558290482, + 0.01515849120914936, + 0.02243741974234581, + 0.018413549289107323, + 0.014886980876326561, + 0.014015116728842258, + 0.011334783397614956, + 0.011621765792369843, + 0.011231710202991962, + 0.010305657051503658, + 0.010848699137568474, + 0.012746920809149742, + 0.00809311494231224, + 0.012163076549768448, + 0.006545977666974068, + 0.003461560932919383, + 0.005148458294570446, + 0.004619854502379894, + 0.0038672543596476316, + 0.006929867900907993, + 0.003689930774271488, + 0.007580755744129419, + 0.022402528673410416, + 0.009788956493139267, + 0.00694331806153059, + 0.0042291623540222645, + 0.007198515348136425, + 0.011499466374516487, + 0.005277593154460192, + 0.021547531709074974, + 0.008938807062804699, + 0.0044409711845219135, + 0.013500002212822437, + 0.004974416457116604, + 0.002700534649193287, + 0.011619609780609608, + 0.004261875990778208, + 0.02919366955757141, + 0.007699471432715654, + 0.005733531434088945, + 0.010267938487231731, + 0.009853481315076351, + 0.00582962715998292, + 0.007800883613526821, + 0.005665441043674946, + 0.008401390165090561, + 0.018173452466726303, + 0.013905983418226242, + 0.029738249257206917, + 0.027717934921383858, + 0.011862372048199177, + 0.009654812514781952, + 0.01253849733620882, + 0.014244361780583858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.06899885088205338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268143489956856 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005268143489956856, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268143489956856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16735300421714783, + 0.05649883300065994, + 0.34086450934410095, + 0.05153336375951767, + 0.10660248249769211, + 0.027432257309556007, + 0.056935057044029236, + 0.06056305766105652, + 0.06550898402929306, + 0.06670840084552765 + ], + "credit_map": [ + 0.025151167064905167, + 0.03376496210694313, + 0.016592102125287056, + 0.0160886999219656, + 0.019463546574115753, + 0.02377762459218502, + 0.017201708629727364, + 0.015313209965825081, + 0.007909280247986317, + 0.0075717344880104065, + 0.00580564746633172, + 0.008356519043445587, + 0.006927493028342724, + 0.007495226804167032, + 0.005418156273663044, + 0.007014780770987272, + 0.05657263100147247, + 0.046689651906490326, + 0.03833784908056259, + 0.04732796549797058, + 0.05060257017612457, + 0.04961460083723068, + 0.027720430865883827, + 0.023998858407139778, + 0.008265509270131588, + 0.005661962553858757, + 0.005460287444293499, + 0.0038498963695019484, + 0.007956275716423988, + 0.007764352485537529, + 0.0055831968784332275, + 0.006991890724748373, + 0.015851078554987907, + 0.013945999555289745, + 0.017771389335393906, + 0.008146533742547035, + 0.009057826362550259, + 0.013851343654096127, + 0.007426643744111061, + 0.020551685243844986, + 0.0037522173952311277, + 0.00262834457680583, + 0.0029855251777917147, + 0.0028041223995387554, + 0.003860976779833436, + 0.004289171192795038, + 0.002074997639283538, + 0.005036904476583004, + 0.010567408986389637, + 0.007300885394215584, + 0.006347000133246183, + 0.005235821940004826, + 0.006983848288655281, + 0.007278177887201309, + 0.004972275346517563, + 0.008249643258750439, + 0.005949666723608971, + 0.0037448767106980085, + 0.007981323637068272, + 0.00472518103197217, + 0.005544518120586872, + 0.009777074679732323, + 0.004823985509574413, + 0.01801643706858158, + 0.006363467313349247, + 0.004855303093791008, + 0.00925416685640812, + 0.013096186332404613, + 0.009510740637779236, + 0.008452684618532658, + 0.004588297568261623, + 0.009388145059347153, + 0.004858476109802723, + 0.007232733070850372, + 0.015359879471361637, + 0.01683576963841915, + 0.005330195184797049, + 0.00509299710392952, + 0.00571030005812645, + 0.006288061384111643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.005268143489956856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908580988645554 + ], + "opposite_logprob_tokens": [ + -2.464085817337036 + ], + "logprob_reward": -0.08908580988645554, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908580988645554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12576012313365936, + 0.0834464356303215, + 0.26923754811286926, + 0.058741189539432526, + 0.1280236840248108, + 0.032480280846357346, + 0.06501895934343338, + 0.040887441486120224, + 0.11722397059202194, + 0.07918038964271545 + ], + "credit_map": [ + 0.020265039056539536, + 0.017704371362924576, + 0.013702897354960442, + 0.015362434089183807, + 0.02156086638569832, + 0.01235882192850113, + 0.013606221415102482, + 0.011199474334716797, + 0.011062482371926308, + 0.01158706285059452, + 0.007178191095590591, + 0.01234368048608303, + 0.013350728899240494, + 0.00919247791171074, + 0.01106336247175932, + 0.007668450009077787, + 0.04482675716280937, + 0.03755639120936394, + 0.02642066776752472, + 0.03687767684459686, + 0.04329850524663925, + 0.038998935371637344, + 0.023072678595781326, + 0.018185900524258614, + 0.007092442363500595, + 0.0076215388253331184, + 0.00570316519588232, + 0.005170427728444338, + 0.011655078269541264, + 0.009429670870304108, + 0.005915928166359663, + 0.0061529348604381084, + 0.016960816457867622, + 0.015985343605279922, + 0.016474537551403046, + 0.013323270715773106, + 0.01585506834089756, + 0.01821393333375454, + 0.0106221167370677, + 0.020588606595993042, + 0.005445756949484348, + 0.00454581156373024, + 0.003701204666867852, + 0.002879718318581581, + 0.0036756230983883142, + 0.005470839329063892, + 0.0024402362760156393, + 0.004321089014410973, + 0.011139040812849998, + 0.006038923282176256, + 0.009210190735757351, + 0.004908971954137087, + 0.00946792308241129, + 0.007553218398243189, + 0.006155131850391626, + 0.010545562021434307, + 0.005037834867835045, + 0.004395379684865475, + 0.0031808349303901196, + 0.0058941286988556385, + 0.005961544346064329, + 0.005889503285288811, + 0.0038490628357976675, + 0.006679156329482794, + 0.012875703163444996, + 0.011178144253790379, + 0.016154872253537178, + 0.023610558360815048, + 0.01756252720952034, + 0.012040733359754086, + 0.010329041630029678, + 0.013472386635839939, + 0.007629523519426584, + 0.008926418609917164, + 0.017704268917441368, + 0.019121848046779633, + 0.005021856166422367, + 0.004982218611985445, + 0.006096087396144867, + 0.009698163717985153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.08908580988645554 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740096490830183 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006740096490830183, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740096490830183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04452341049909592, + 0.029891353100538254, + 0.04748687520623207, + 0.07453542202711105, + 0.0855369120836258, + 0.028415361419320107, + 0.06894359737634659, + 0.030432961881160736, + 0.32345256209373474, + 0.2667815089225769 + ], + "credit_map": [ + 0.007346911355853081, + 0.005549798719584942, + 0.004881659988313913, + 0.004362689796835184, + 0.007584433536976576, + 0.005035816226154566, + 0.004302801098674536, + 0.005459299311041832, + 0.004279123153537512, + 0.004917302634567022, + 0.003258941462263465, + 0.004171431064605713, + 0.0034648971632122993, + 0.003419562941417098, + 0.0032746982760727406, + 0.00310539617203176, + 0.008732481859624386, + 0.005880002398043871, + 0.003991086967289448, + 0.0064294664189219475, + 0.006368552800267935, + 0.007303188554942608, + 0.004286035429686308, + 0.004496058914810419, + 0.013345912098884583, + 0.011306330561637878, + 0.0073107159696519375, + 0.006063451524823904, + 0.007713185157626867, + 0.01266295276582241, + 0.0057706148363649845, + 0.010362261906266212, + 0.01178447064012289, + 0.011051831766963005, + 0.01398448832333088, + 0.008361434563994408, + 0.009375805035233498, + 0.010451268404722214, + 0.0063125500455498695, + 0.01421506330370903, + 0.005025103222578764, + 0.0033167630899697542, + 0.004065952729433775, + 0.002553618745878339, + 0.003042219439521432, + 0.004691624082624912, + 0.0020820614881813526, + 0.003638016525655985, + 0.015963967889547348, + 0.004959445912390947, + 0.007216361816972494, + 0.005288831423968077, + 0.006407675798982382, + 0.009774674661457539, + 0.006000213790684938, + 0.0133324284106493, + 0.004926434718072414, + 0.0033706158865243196, + 0.0024050751235336065, + 0.0032698367722332478, + 0.004043402150273323, + 0.005689735058695078, + 0.0030632931739091873, + 0.003664568066596985, + 0.03822692856192589, + 0.03287183493375778, + 0.04382399842143059, + 0.06119949370622635, + 0.04257277026772499, + 0.03414440155029297, + 0.031967829912900925, + 0.038645315915346146, + 0.026728332042694092, + 0.028743252158164978, + 0.05761817842721939, + 0.06285031884908676, + 0.02034778892993927, + 0.018995491787791252, + 0.019386740401387215, + 0.03211141377687454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.006740096490830183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061547338962555 + ], + "opposite_logprob_tokens": [ + -3.895615577697754 + ], + "logprob_reward": -0.02061547338962555, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061547338962555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13125987350940704, + 0.22967170178890228, + 0.16651862859725952, + 0.0720653384923935, + 0.09350047260522842, + 0.02914557233452797, + 0.12533633410930634, + 0.045847088098526, + 0.03751678764820099, + 0.06913822889328003 + ], + "credit_map": [ + 0.016526883468031883, + 0.025725891813635826, + 0.016076337546110153, + 0.012584056705236435, + 0.014669155701994896, + 0.01537729799747467, + 0.015846936032176018, + 0.014453309588134289, + 0.04315665364265442, + 0.0329010970890522, + 0.021662183105945587, + 0.028398143127560616, + 0.03644225001335144, + 0.02658744342625141, + 0.01985258422791958, + 0.020671358332037926, + 0.026410195976495743, + 0.0231259074062109, + 0.015154454857110977, + 0.025268729776144028, + 0.026649191975593567, + 0.02301955781877041, + 0.014950883574783802, + 0.011939703486859798, + 0.009568529203534126, + 0.011302260681986809, + 0.007499537896364927, + 0.006708214059472084, + 0.011946437880396843, + 0.011890729889273643, + 0.006511062383651733, + 0.0066385651007294655, + 0.012958991341292858, + 0.011679026298224926, + 0.012435550801455975, + 0.010387391783297062, + 0.011930899694561958, + 0.011903718113899231, + 0.007677839137613773, + 0.01452705916017294, + 0.004481730982661247, + 0.0026616763789206743, + 0.0031407708302140236, + 0.003386064898222685, + 0.003067489480599761, + 0.006129744928330183, + 0.0026523347478359938, + 0.0036257596220821142, + 0.02465529926121235, + 0.011827650479972363, + 0.012023739516735077, + 0.008904518559575081, + 0.014672666788101196, + 0.01814982108771801, + 0.009432687424123287, + 0.02566996030509472, + 0.005019656848162413, + 0.004070523660629988, + 0.003789081471040845, + 0.00555056519806385, + 0.008040154352784157, + 0.008706402964890003, + 0.005663382820785046, + 0.00500732334330678, + 0.00430556433275342, + 0.004282371606677771, + 0.0057140979915857315, + 0.005461659748107195, + 0.004674246534705162, + 0.003903029253706336, + 0.00424694549292326, + 0.004928876645863056, + 0.007161042653024197, + 0.007047014310956001, + 0.015238623134791851, + 0.01587226428091526, + 0.004909324925392866, + 0.004765772260725498, + 0.005534121301025152, + 0.008610060438513756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.02061547338962555 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static, extreme close-up depicts an intricately patterned small lizard, its body a detailed mosaic of sandy beige and russet scales, perfectly mimicking the sun-drenched sandstone rock below. The midday sun casts sharp, microscopic shadows across the rock's weathered ochre surface. The lizard freezes mid-step, one delicate front leg suspended above the stone's grainy texture, its entire body lowered in a posture of pure, primal instinct. As a large, amorphous shadow glides swiftly and silently across the scene, the lizard, momentarily trapped in cool darkness, remains completely petrified. Its muscles tense, its beady black eyes wide and unblinking, its breath held as it desperately and silently attempts to merge with the stone, creating a palpable atmosphere of high-stakes tension and a frozen moment of survival.", + "seed": 28205, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_14/index_601_seed_28205_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09774872614070773, + "vlm_logprob": -0.09774872614070773, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.09774872614070773, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08231642842292786 + ], + "opposite_logprob_tokens": [ + -2.8323163986206055 + ], + "logprob_reward": -0.08231642842292786, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08231642842292786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13374081254005432, + 0.1028124988079071, + 0.10082457214593887, + 0.16831302642822266, + 0.0819322019815445, + 0.051905278116464615, + 0.12833184003829956, + 0.04398012161254883, + 0.11778610944747925, + 0.0703735202550888 + ], + "credit_map": [ + 0.02084304764866829, + 0.018943391740322113, + 0.01846717670559883, + 0.015153379179537296, + 0.01134873554110527, + 0.015170643106102943, + 0.01299557089805603, + 0.020818868651986122, + 0.014614425599575043, + 0.018310151994228363, + 0.014012839645147324, + 0.008183063939213753, + 0.01222675759345293, + 0.015337763354182243, + 0.0059275454841554165, + 0.014199960976839066, + 0.015303870663046837, + 0.014205999672412872, + 0.010860121808946133, + 0.00831367913633585, + 0.019977064803242683, + 0.013052255846560001, + 0.009051211178302765, + 0.010060366243124008, + 0.025517739355564117, + 0.019108712673187256, + 0.020820504054427147, + 0.024334857240319252, + 0.018962310627102852, + 0.030070239678025246, + 0.012911729514598846, + 0.01658692955970764, + 0.00954831950366497, + 0.011929669417440891, + 0.010061969049274921, + 0.016625555232167244, + 0.010733179748058319, + 0.009883626364171505, + 0.007043687626719475, + 0.0061061931774020195, + 0.00790457334369421, + 0.0079561248421669, + 0.004906448535621166, + 0.007706544362008572, + 0.007540000136941671, + 0.006216897163540125, + 0.005345550831407309, + 0.004329139832407236, + 0.01750173047184944, + 0.010664969682693481, + 0.0185657050460577, + 0.018261685967445374, + 0.017226556316018105, + 0.011673313565552235, + 0.021000854671001434, + 0.01343702245503664, + 0.007049525622278452, + 0.005537478718906641, + 0.006269203964620829, + 0.005036139395087957, + 0.005507935769855976, + 0.005232496187090874, + 0.006037214770913124, + 0.0033101276494562626, + 0.013173752464354038, + 0.01426623947918415, + 0.012731681577861309, + 0.01988823711872101, + 0.017446016892790794, + 0.011811558157205582, + 0.01682150363922119, + 0.0116471191868186, + 0.009531334042549133, + 0.012559343129396439, + 0.01564611680805683, + 0.008888023905456066, + 0.006806341931223869, + 0.004611017648130655, + 0.005900257267057896, + 0.006431089714169502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.08231642842292786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004657610785216093 + ], + "opposite_logprob_tokens": [ + -5.379657745361328 + ], + "logprob_reward": -0.004657610785216093, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004657610785216093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10352930426597595, + 0.17894983291625977, + 0.045974649488925934, + 0.05631294846534729, + 0.0801137387752533, + 0.11105143278837204, + 0.19378870725631714, + 0.031234165653586388, + 0.1113511174917221, + 0.08769410848617554 + ], + "credit_map": [ + 0.011992602609097958, + 0.014056496322154999, + 0.016004471108317375, + 0.013474787585437298, + 0.011536717414855957, + 0.01179492473602295, + 0.01121408399194479, + 0.013455209322273731, + 0.03169453889131546, + 0.01726437732577324, + 0.026338350027799606, + 0.0174812413752079, + 0.01991989091038704, + 0.03754550963640213, + 0.00995788536965847, + 0.018748046830296516, + 0.008702119812369347, + 0.003217791672796011, + 0.00580258434638381, + 0.0072203949093818665, + 0.005869896151125431, + 0.005628371611237526, + 0.00345780560746789, + 0.006075683049857616, + 0.010014784522354603, + 0.006173152942210436, + 0.008868146687746048, + 0.0062629044987261295, + 0.005690841935575008, + 0.011147229000926018, + 0.0032234021928161383, + 0.004932481795549393, + 0.009106820449233055, + 0.006671860348433256, + 0.009333302266895771, + 0.012126898393034935, + 0.009795140475034714, + 0.012566427700221539, + 0.00825275294482708, + 0.012260537594556808, + 0.01482173427939415, + 0.020671429112553596, + 0.013889714144170284, + 0.017874445766210556, + 0.012335301376879215, + 0.011196275241672993, + 0.010728659108281136, + 0.009533869102597237, + 0.025397950783371925, + 0.016870960593223572, + 0.02016281895339489, + 0.03202242776751518, + 0.029765602201223373, + 0.026462530717253685, + 0.023655369877815247, + 0.019451018422842026, + 0.006815758068114519, + 0.0028775958344340324, + 0.0047907703556120396, + 0.003184085013344884, + 0.00385667709633708, + 0.0037299676332622766, + 0.003045635996386409, + 0.002933674491941929, + 0.01098391879349947, + 0.014509514905512333, + 0.0114529924467206, + 0.015148078091442585, + 0.02087879553437233, + 0.012755661271512508, + 0.014891518279910088, + 0.010730636306107044, + 0.014128598384559155, + 0.010255374945700169, + 0.010667036287486553, + 0.010731104761362076, + 0.012794316746294498, + 0.008906724862754345, + 0.012609152123332024, + 0.0076017980463802814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.004657610785216093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135486841201782 + ], + "opposite_logprob_tokens": [ + -1.3135486841201782 + ], + "logprob_reward": -0.3135486841201782, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135486841201782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23892629146575928, + 0.15546540915966034, + 0.07178051024675369, + 0.041110049933195114, + 0.1633129119873047, + 0.14634448289871216, + 0.06348293274641037, + 0.031176002696156502, + 0.037502847611904144, + 0.05089856684207916 + ], + "credit_map": [ + 0.02788480930030346, + 0.027862772345542908, + 0.02902660332620144, + 0.027233285829424858, + 0.019432272762060165, + 0.029807452112436295, + 0.021922137588262558, + 0.05575694516301155, + 0.018407167866826057, + 0.018197642639279366, + 0.022495543584227562, + 0.015779802575707436, + 0.013948425650596619, + 0.034964434802532196, + 0.011626112274825573, + 0.02004627138376236, + 0.013551096431910992, + 0.007967453449964523, + 0.007349254563450813, + 0.009933626279234886, + 0.010309278033673763, + 0.0071645090356469154, + 0.005495540797710419, + 0.010009750723838806, + 0.007569642271846533, + 0.005602709949016571, + 0.005701867397874594, + 0.004443597048521042, + 0.005509183742105961, + 0.005497219506651163, + 0.0028799681458622217, + 0.003905860474333167, + 0.021085109561681747, + 0.017926551401615143, + 0.016364911571145058, + 0.023208647966384888, + 0.021312382072210312, + 0.02760978974401951, + 0.01416391134262085, + 0.021641599014401436, + 0.01739756017923355, + 0.025966130197048187, + 0.014220583252608776, + 0.024333123117685318, + 0.020709028467535973, + 0.01567763462662697, + 0.016267551109194756, + 0.01177287194877863, + 0.008167890831828117, + 0.004981354810297489, + 0.008186633698642254, + 0.010513756424188614, + 0.008870287798345089, + 0.009751014411449432, + 0.007524130400270224, + 0.005487858783453703, + 0.004867550451308489, + 0.003599377116188407, + 0.003635931294411421, + 0.004111087415367365, + 0.004371404182165861, + 0.0048048836179077625, + 0.0031261183321475983, + 0.002659651916474104, + 0.005609200336039066, + 0.005168265197426081, + 0.004162727389484644, + 0.004451252985745668, + 0.00481206551194191, + 0.0049661085940897465, + 0.005097529385238886, + 0.003235702170059085, + 0.007619230542331934, + 0.009948351420462132, + 0.008418973535299301, + 0.006362715736031532, + 0.006314149126410484, + 0.003981061279773712, + 0.004327062051743269, + 0.003927020821720362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.3135486841201782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002470642328262 + ], + "opposite_logprob_tokens": [ + -2.6950247287750244 + ], + "logprob_reward": -0.07002470642328262, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002470642328262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0887475535273552, + 0.18061210215091705, + 0.30467531085014343, + 0.04582253843545914, + 0.08195580542087555, + 0.11180252581834793, + 0.03271685540676117, + 0.04567127674818039, + 0.057608939707279205, + 0.050387121737003326 + ], + "credit_map": [ + 0.010961206629872322, + 0.010750777088105679, + 0.011420853435993195, + 0.010313354432582855, + 0.009984111413359642, + 0.009711342863738537, + 0.007938334718346596, + 0.017667565494775772, + 0.028247613459825516, + 0.025091951712965965, + 0.024702975526452065, + 0.014602774754166603, + 0.019096188247203827, + 0.04232200235128403, + 0.008707724511623383, + 0.017840873450040817, + 0.05322547256946564, + 0.03024451434612274, + 0.039310574531555176, + 0.028194038197398186, + 0.05570568889379501, + 0.0385248176753521, + 0.02880619652569294, + 0.03066399320960045, + 0.005396647844463587, + 0.005651095416396856, + 0.00621077511459589, + 0.006158720701932907, + 0.005895531736314297, + 0.009039167314767838, + 0.003238352946937084, + 0.0042322478257119656, + 0.00842041801661253, + 0.007892170920968056, + 0.011396024376153946, + 0.012541182339191437, + 0.010476588271558285, + 0.012515882030129433, + 0.007566037122160196, + 0.011147497221827507, + 0.016637539491057396, + 0.015928342938423157, + 0.011474010534584522, + 0.018198316916823387, + 0.012290509417653084, + 0.012539387680590153, + 0.011517479084432125, + 0.013216940686106682, + 0.005223724525421858, + 0.0032031016889959574, + 0.003540661185979843, + 0.004239215981215239, + 0.0052016922272741795, + 0.0046641649678349495, + 0.003889599349349737, + 0.0027546933852136135, + 0.008476915769279003, + 0.004406772553920746, + 0.0046193599700927734, + 0.00574851268902421, + 0.006123073399066925, + 0.007176347076892853, + 0.0048287794925272465, + 0.004291516728699207, + 0.006631479598581791, + 0.007239011116325855, + 0.006186546292155981, + 0.008102428168058395, + 0.009449365548789501, + 0.006325568072497845, + 0.008293972350656986, + 0.005380570422858, + 0.0067785452120006084, + 0.006834142841398716, + 0.009277543984353542, + 0.005917780566960573, + 0.006993240676820278, + 0.004355626180768013, + 0.0057426150888204575, + 0.004487625323235989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.07002470642328262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle high-angle and static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme close-up of a small lizard with sandy beige and russet scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard frozen mid-step with one front leg suspended above the stone surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a large, amorphous shadow glide swiftly and silently across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the lizard's beady black eyes wide and unblinking while it remains petrified?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819620095193386 + ], + "opposite_logprob_tokens": [ + -4.018196105957031 + ], + "logprob_reward": -0.01819620095193386, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819620095193386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1509871780872345, + 0.10052874684333801, + 0.13458696007728577, + 0.06513244658708572, + 0.14802879095077515, + 0.10905759036540985, + 0.05892834812402725, + 0.03677649796009064, + 0.13604232668876648, + 0.05993116647005081 + ], + "credit_map": [ + 0.019697731360793114, + 0.016323894262313843, + 0.018106749281287193, + 0.020892491564154625, + 0.014086438342928886, + 0.01620323397219181, + 0.021577881649136543, + 0.024098772555589676, + 0.012953159399330616, + 0.01638481207191944, + 0.011360150761902332, + 0.00929156132042408, + 0.01552776899188757, + 0.016912110149860382, + 0.006772711873054504, + 0.01132646668702364, + 0.02605767361819744, + 0.013593317940831184, + 0.017091870307922363, + 0.01614638790488243, + 0.019587071612477303, + 0.018687034025788307, + 0.00890626572072506, + 0.014517342671751976, + 0.006210532505065203, + 0.01034263800829649, + 0.007841718383133411, + 0.010395296849310398, + 0.00899194274097681, + 0.009413711726665497, + 0.0048502166755497456, + 0.007086385507136583, + 0.012003651820123196, + 0.016106421127915382, + 0.018484795466065407, + 0.02384566329419613, + 0.026347866281867027, + 0.021819932386279106, + 0.013371208682656288, + 0.016049260273575783, + 0.01352754794061184, + 0.01822979934513569, + 0.014051231555640697, + 0.016563313081860542, + 0.015521804802119732, + 0.01176531333476305, + 0.012118778191506863, + 0.007279803976416588, + 0.007218427490442991, + 0.00607402715831995, + 0.006145997438579798, + 0.009957121685147285, + 0.010781648568809032, + 0.0066344253718853, + 0.007609582506120205, + 0.004507120698690414, + 0.005512911360710859, + 0.0031603786628693342, + 0.003628564765676856, + 0.004246729891747236, + 0.005834394600242376, + 0.004861988127231598, + 0.0050408244132995605, + 0.004490705672651529, + 0.016225367784500122, + 0.016626931726932526, + 0.014194426126778126, + 0.02183069847524166, + 0.020270803943276405, + 0.015173614956438541, + 0.01957445591688156, + 0.012146023102104664, + 0.008769681677222252, + 0.009263522922992706, + 0.012605450116097927, + 0.008285768330097198, + 0.006683375220745802, + 0.00425186101347208, + 0.005279683507978916, + 0.0047918218187987804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.01819620095193386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_15/index_tensor([7970, 7970])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_15/index_tensor([7970, 7970])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..888329e2783873d6cc51a2395d4565a874aa419f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_15/index_tensor([7970, 7970])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, wide-angle lens slowly tilts right, revealing a surreal scene captured in stark, artistic black-and-white, 3D, Pixar-style. A vast, rolling sea, composed entirely of meticulously cultivated, high-standard farmland, stretches to the horizon. Rolling black cultivated hills and silvery crops mimic the undulations of waves beneath a high-contrast sky. A whimsical boat glides silently across this land-sea, its sleek, rounded hull of dark charcoal-gray wood and a friendly, stylized animal head perched on its prow. As the boat steadily approaches a distant village, its single, large, patchwork sail—a mosaic of various gray tones and textured fabrics—is gracefully and deliberately lowered, the canvas gently billowing before collapsing in neat folds onto the slightly bent mast. The destination, a village built in a distinctive ethnic minority style, becomes more clearly visible, its structures characterized by curved, dark-tiled roofs and intricate white gridwork, all organically nestled into the terraced slopes of the farmland hills. The entire dreamlike scene is defined by a dramatic play of light and shadow; vibrant white highlights the edges and rooftops of the boats, while deep, inky black pools in the furrows of the land, creating an atmosphere that is visually stunning and profoundly tranquil.", + "seed": 15875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_15/index_7970_seed_15875_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2642767712473869, + "vlm_logprob": -0.2642767712473869, + "vlm_yes_no_margin": 1.5, + "vlm_token_credit": -0.2642767712473869, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2848994731903076 + ], + "opposite_logprob_tokens": [ + -1.4098994731903076 + ], + "logprob_reward": -0.2848994731903076, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2848994731903076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04552815854549408, + 0.5559732913970947, + 0.07710561901330948, + 0.20351383090019226, + 0.048897884786129, + 0.011458991095423698, + 0.009407968260347843, + 0.018660107627511024, + 0.021186217665672302, + 0.008267891593277454 + ], + "credit_map": [ + 0.005247513297945261, + 0.005609021056443453, + 0.006932208314538002, + 0.005136936902999878, + 0.006126756779849529, + 0.0062612066976726055, + 0.004736847709864378, + 0.005477670580148697, + 0.02244376391172409, + 0.02479131892323494, + 0.01133925374597311, + 0.01497973594814539, + 0.21329234540462494, + 0.22334471344947815, + 0.021120313555002213, + 0.02466185949742794, + 0.010640179738402367, + 0.031712837517261505, + 0.0042928671464324, + 0.0026766580995172262, + 0.01800919696688652, + 0.005049033556133509, + 0.002204017946496606, + 0.0025208338629454374, + 0.02950025536119938, + 0.0318470224738121, + 0.0249023400247097, + 0.023798394948244095, + 0.025081805884838104, + 0.027182400226593018, + 0.02574409916996956, + 0.015457517467439175, + 0.007541223894804716, + 0.007882615551352501, + 0.007057977374643087, + 0.00387128465808928, + 0.005241413600742817, + 0.0047691259533166885, + 0.004378385376185179, + 0.00815585907548666, + 0.0017406492261216044, + 0.0019382965983822942, + 0.0010550243314355612, + 0.0012694214237853885, + 0.0015904224710538983, + 0.0012043360620737076, + 0.0010174302151426673, + 0.0016434097196906805, + 0.00108505564276129, + 0.0016570851439610124, + 0.001713076257146895, + 0.0011535531375557184, + 0.0008071660413406789, + 0.000781392038334161, + 0.0012119426392018795, + 0.0009986970108002424, + 0.0023960424587130547, + 0.0029761274345219135, + 0.003079964080825448, + 0.0025656609795987606, + 0.0022208071313798428, + 0.0015452471561729908, + 0.0017773456638678908, + 0.0020989126060158014, + 0.002587983850389719, + 0.003482707543298602, + 0.0031717270612716675, + 0.002881784224882722, + 0.002705738879740238, + 0.0015098853036761284, + 0.0020605288445949554, + 0.002785863820463419, + 0.0012351882178336382, + 0.0011229400988668203, + 0.0014196241972967982, + 0.0009026770712807775, + 0.00059851398691535, + 0.0011251987889409065, + 0.0008553502266295254, + 0.0010083985980600119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.2848994731903076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906050771474838 + ], + "opposite_logprob_tokens": [ + -2.4640605449676514 + ], + "logprob_reward": -0.08906050771474838, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906050771474838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05218283459544182, + 0.2832726538181305, + 0.2500680387020111, + 0.06934839487075806, + 0.07090060412883759, + 0.043363768607378006, + 0.03774449974298477, + 0.0205379705876112, + 0.07633034139871597, + 0.09625087678432465 + ], + "credit_map": [ + 0.00583277503028512, + 0.006839217152446508, + 0.009013095870614052, + 0.00549789285287261, + 0.006664041429758072, + 0.00693484814837575, + 0.005129991564899683, + 0.006270973477512598, + 0.014070474542677402, + 0.01879502460360527, + 0.0158762838691473, + 0.010421903803944588, + 0.0984998270869255, + 0.09678236395120621, + 0.017034702003002167, + 0.011792073957622051, + 0.028966359794139862, + 0.0731857568025589, + 0.03588217869400978, + 0.017449134960770607, + 0.052031341940164566, + 0.01771780662238598, + 0.011289742775261402, + 0.013545699417591095, + 0.009220488369464874, + 0.007585716899484396, + 0.006671158596873283, + 0.007003897335380316, + 0.008719926699995995, + 0.01370114367455244, + 0.008327966555953026, + 0.008118102326989174, + 0.010040921159088612, + 0.011368153616786003, + 0.009953654371201992, + 0.0061037070117890835, + 0.008464253507554531, + 0.007284655701369047, + 0.007225037552416325, + 0.010460223071277142, + 0.006400201935321093, + 0.00857592560350895, + 0.004920329432934523, + 0.006372090429067612, + 0.005091168452054262, + 0.0038079877849668264, + 0.004266918171197176, + 0.0039291419088840485, + 0.005174277815967798, + 0.008150268346071243, + 0.005277005489915609, + 0.004128019325435162, + 0.0040626381523907185, + 0.003350046230480075, + 0.0035596489906311035, + 0.004042595624923706, + 0.0031706125009804964, + 0.0034460793249309063, + 0.0029192459769546986, + 0.0025143823586404324, + 0.0019517490873113275, + 0.0016599461669102311, + 0.002090314170345664, + 0.0027856393717229366, + 0.01033269427716732, + 0.010815401561558247, + 0.006873755715787411, + 0.0048135449178516865, + 0.00895286351442337, + 0.016298003494739532, + 0.007605194114148617, + 0.010638881474733353, + 0.013299391604959965, + 0.009006507694721222, + 0.013291386887431145, + 0.008737470023334026, + 0.010400953702628613, + 0.017293786630034447, + 0.010525509715080261, + 0.013695870526134968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.08906050771474838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701991200447083 + ], + "opposite_logprob_tokens": [ + -2.1270198822021484 + ], + "logprob_reward": -0.12701991200447083, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701991200447083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029684877023100853, + 0.2632717490196228, + 0.46947336196899414, + 0.11843804270029068, + 0.035755932331085205, + 0.01530881691724062, + 0.013290982693433762, + 0.0124302227050066, + 0.016582725569605827, + 0.025763317942619324 + ], + "credit_map": [ + 0.004831343423575163, + 0.003508303314447403, + 0.0028491555713117123, + 0.0027082632295787334, + 0.0035262976307421923, + 0.004178726579993963, + 0.0033070051576942205, + 0.004775782581418753, + 0.01073875930160284, + 0.01472601667046547, + 0.008211127482354641, + 0.0066076829098165035, + 0.10100310295820236, + 0.10170304775238037, + 0.01023904513567686, + 0.01004297100007534, + 0.047867849469184875, + 0.18532466888427734, + 0.026062695309519768, + 0.021821632981300354, + 0.11553621292114258, + 0.034815508872270584, + 0.017506131902337074, + 0.020538631826639175, + 0.01630127988755703, + 0.019074849784374237, + 0.022193992510437965, + 0.011610424146056175, + 0.013195698149502277, + 0.012541692703962326, + 0.013555233366787434, + 0.009964869357645512, + 0.004807766526937485, + 0.004209192935377359, + 0.005822035484015942, + 0.0033136855345219374, + 0.003339411923661828, + 0.004388231784105301, + 0.002864598762243986, + 0.007011009380221367, + 0.0023611595388501883, + 0.0029845156241208315, + 0.0031218440271914005, + 0.001536075840704143, + 0.0015739728696644306, + 0.0009138575987890363, + 0.001174846664071083, + 0.0016425453359261155, + 0.0019097059266641736, + 0.002731332089751959, + 0.0015046170447021723, + 0.0011623096652328968, + 0.0016635960200801492, + 0.001359551795758307, + 0.0009750848985277116, + 0.001984784845262766, + 0.0014465874992311, + 0.002039529848843813, + 0.0015333883929997683, + 0.0018565842183306813, + 0.0015384767903015018, + 0.0013245445443317294, + 0.0010506465332582593, + 0.0016404640628024936, + 0.0016132217133417726, + 0.0016321628354489803, + 0.002037389436736703, + 0.0018393972422927618, + 0.0025854583363980055, + 0.0018277831841260195, + 0.0028558510821312666, + 0.0021914620883762836, + 0.003236655378714204, + 0.0027702797669917345, + 0.004323423374444246, + 0.0036454303190112114, + 0.002312681870535016, + 0.003913633059710264, + 0.0025424493942409754, + 0.0030187645461410284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.12701991200447083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270771324634552 + ], + "opposite_logprob_tokens": [ + -2.127077102661133 + ], + "logprob_reward": -0.1270771324634552, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270771324634552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07100335508584976, + 0.13168394565582275, + 0.27316272258758545, + 0.21884261071681976, + 0.12845298647880554, + 0.03284689038991928, + 0.024709858000278473, + 0.02369687519967556, + 0.05069126561284065, + 0.04490948095917702 + ], + "credit_map": [ + 0.006774124689400196, + 0.010218525305390358, + 0.007014267146587372, + 0.006801765412092209, + 0.009175396524369717, + 0.008845202624797821, + 0.009874841198325157, + 0.012299232184886932, + 0.00629601813852787, + 0.00949483085423708, + 0.0061164903454482555, + 0.0048841615207493305, + 0.04432279244065285, + 0.0452539436519146, + 0.008475016802549362, + 0.0068406881764531136, + 0.022928718477487564, + 0.09326346963644028, + 0.019197890534996986, + 0.01743166334927082, + 0.0704907700419426, + 0.02162882313132286, + 0.012918916530907154, + 0.01530247088521719, + 0.027792664244771004, + 0.028929855674505234, + 0.03335457295179367, + 0.02347331866621971, + 0.026456201449036598, + 0.03136524558067322, + 0.02341013215482235, + 0.02406061813235283, + 0.021537862718105316, + 0.022615235298871994, + 0.016602233052253723, + 0.010581469163298607, + 0.019300714135169983, + 0.012656954117119312, + 0.010553369298577309, + 0.014605148695409298, + 0.0055775875225663185, + 0.006926450412720442, + 0.00357494386844337, + 0.003920774441212416, + 0.003510806243866682, + 0.0030449614860117435, + 0.002392128575593233, + 0.0038992392364889383, + 0.003935935441404581, + 0.00414077565073967, + 0.0024381112307310104, + 0.0024624608922749758, + 0.0031112320721149445, + 0.002940327161923051, + 0.002409145003184676, + 0.0032718712463974953, + 0.003430632408708334, + 0.004193721804767847, + 0.0039109643548727036, + 0.0034300489351153374, + 0.0025674100033938885, + 0.001838919473811984, + 0.002133789239451289, + 0.0021913903765380383, + 0.0066732969135046005, + 0.006547627039253712, + 0.0045035844668745995, + 0.0037832967936992645, + 0.005612934939563274, + 0.011111307889223099, + 0.0048689162358641624, + 0.00759030319750309, + 0.006026557181030512, + 0.006452902685850859, + 0.007748553529381752, + 0.0065654306672513485, + 0.004934880416840315, + 0.005366533063352108, + 0.0036106836050748825, + 0.004203938879072666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.1270771324634552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933268308639526 + ], + "opposite_logprob_tokens": [ + -0.6933268308639526 + ], + "logprob_reward": -0.6933268308639526, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933268308639526, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05080084875226021, + 0.45145535469055176, + 0.14864543080329895, + 0.10754485428333282, + 0.07753200829029083, + 0.033068977296352386, + 0.04704896733164787, + 0.029104461893439293, + 0.022167015820741653, + 0.03263215348124504 + ], + "credit_map": [ + 0.006470109801739454, + 0.0055488646030426025, + 0.00535268010571599, + 0.006479427684098482, + 0.00516395503655076, + 0.005924602039158344, + 0.007969970814883709, + 0.007891236804425716, + 0.017704613506793976, + 0.01972663775086403, + 0.009158620610833168, + 0.01111062802374363, + 0.17854298651218414, + 0.18129411339759827, + 0.016001876443624496, + 0.017915841192007065, + 0.01660889759659767, + 0.05910223722457886, + 0.00905909575521946, + 0.006329293828457594, + 0.035404022783041, + 0.01008355151861906, + 0.005112385377287865, + 0.006945921573787928, + 0.017486393451690674, + 0.013788069598376751, + 0.012890796177089214, + 0.013183738104999065, + 0.014936360530555248, + 0.014179123565554619, + 0.011884377337992191, + 0.009195986203849316, + 0.011179990135133266, + 0.011289890855550766, + 0.010326859541237354, + 0.007658968213945627, + 0.0070193735882639885, + 0.007903644815087318, + 0.006795138586312532, + 0.015358131378889084, + 0.005686563439667225, + 0.007326869294047356, + 0.0041045560501515865, + 0.0035790561232715845, + 0.0033284469973295927, + 0.002290558535605669, + 0.002483590506017208, + 0.004269331693649292, + 0.006125636864453554, + 0.009396315552294254, + 0.009352942928671837, + 0.0049273353070020676, + 0.003918302711099386, + 0.00271532847546041, + 0.005291214678436518, + 0.005321886390447617, + 0.0037065681535750628, + 0.004907822702080011, + 0.005104172043502331, + 0.004077058285474777, + 0.0036032928619533777, + 0.0022140040528029203, + 0.003096247324720025, + 0.0023952943738549948, + 0.003494605189189315, + 0.0024531243834644556, + 0.0028520426712930202, + 0.001958623994141817, + 0.0026034764014184475, + 0.003474733792245388, + 0.002360299928113818, + 0.0029701082967221737, + 0.0045501417480409145, + 0.004153925459831953, + 0.0035613556392490864, + 0.0029737716540694237, + 0.004199433606117964, + 0.00515414634719491, + 0.0032218284904956818, + 0.004817543085664511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.6933268308639526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly tilts right, revealing a surreal scene captured in stark, artistic black-and-white, 3D, Pixar-style. A vast, rolling sea, composed entirely of meticulously cultivated, high-standard farmland, stretches to the horizon. Rolling black cultivated hills and silvery crops mimic the undulations of waves beneath a high-contrast sky. A whimsical boat glides silently across this land-sea, its sleek, rounded hull of dark charcoal-gray wood and a friendly, stylized animal head perched on its prow. As the boat steadily approaches a distant village, its single, large, patchwork sail—a mosaic of various gray tones and textured fabrics—is gracefully and deliberately lowered, the canvas gently billowing before collapsing in neat folds onto the slightly bent mast. The destination, a village built in a distinctive ethnic minority style, becomes more clearly visible, its structures characterized by curved, dark-tiled roofs and intricate white gridwork, all organically nestled into the terraced slopes of the farmland hills. The entire dreamlike scene is defined by a dramatic play of light and shadow; vibrant white highlights the edges and rooftops of the boats, while deep, inky black pools in the furrows of the land, creating an atmosphere that is visually stunning and profoundly tranquil.", + "seed": 15876, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_15/index_7970_seed_15876_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.122152316942811, + "vlm_logprob": -0.122152316942811, + "vlm_yes_no_margin": 3.325000011920929, + "vlm_token_credit": -0.122152316942811, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007551930844783783 + ], + "opposite_logprob_tokens": [ + -7.132552146911621 + ], + "logprob_reward": -0.007551930844783783, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.007551930844783783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45251327753067017, + 0.05543926730751991, + 0.028289228677749634, + 0.0310840904712677, + 0.031095949932932854, + 0.017987022176384926, + 0.04253889247775078, + 0.028352687135338783, + 0.16608762741088867, + 0.14661192893981934 + ], + "credit_map": [ + 0.02818123623728752, + 0.13956500589847565, + 0.027549458667635918, + 0.10792063176631927, + 0.07142527401447296, + 0.016601156443357468, + 0.04032042622566223, + 0.020950110629200935, + 0.009736529551446438, + 0.006596046034246683, + 0.004104803316295147, + 0.011244148947298527, + 0.010500003583729267, + 0.005476390942931175, + 0.004266131669282913, + 0.003515214193612337, + 0.0025239756796509027, + 0.0033715306781232357, + 0.004436394199728966, + 0.003449183888733387, + 0.007251360919326544, + 0.0020172172226011753, + 0.0014888044679537416, + 0.003750760108232498, + 0.002328107599169016, + 0.004895992111414671, + 0.004033172503113747, + 0.00525008887052536, + 0.005419102497398853, + 0.0026707476936280727, + 0.003807291155681014, + 0.002679589670151472, + 0.004849489312618971, + 0.007557202130556107, + 0.0021391473710536957, + 0.0026437698397785425, + 0.006531689781695604, + 0.0017240066081285477, + 0.0026134622748941183, + 0.003037183778360486, + 0.0014409308787435293, + 0.0031686145812273026, + 0.00349867413751781, + 0.001784040592610836, + 0.0036362733226269484, + 0.001418293104507029, + 0.0018921734299510717, + 0.0011480230605229735, + 0.007411304395645857, + 0.005571004934608936, + 0.00890070479363203, + 0.0029992966447025537, + 0.004482670687139034, + 0.003732476383447647, + 0.00343291531316936, + 0.006008516065776348, + 0.003660383401438594, + 0.009375715628266335, + 0.0027347097638994455, + 0.002610732801258564, + 0.0029211954679340124, + 0.0016856695292517543, + 0.0023140350822359324, + 0.003050245111808181, + 0.03672003746032715, + 0.017075516283512115, + 0.06102709099650383, + 0.010783610865473747, + 0.009981992654502392, + 0.010019090957939625, + 0.009160076268017292, + 0.011320206336677074, + 0.008927632123231888, + 0.04515527933835983, + 0.04079088196158409, + 0.009598616510629654, + 0.010502532124519348, + 0.009154044091701508, + 0.00582091836258769, + 0.016662027686834335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.007551930844783783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204963102936745 + ], + "opposite_logprob_tokens": [ + -2.812049627304077 + ], + "logprob_reward": -0.06204963102936745, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204963102936745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22040489315986633, + 0.035678617656230927, + 0.1049448698759079, + 0.023938272148370743, + 0.054905667901039124, + 0.03916430100798607, + 0.11718174815177917, + 0.08501537889242172, + 0.21551015973091125, + 0.10325605422258377 + ], + "credit_map": [ + 0.013563334941864014, + 0.06277131289243698, + 0.015601246617734432, + 0.05147615075111389, + 0.034621719270944595, + 0.009817957878112793, + 0.022298762574791908, + 0.010254432447254658, + 0.005046611651778221, + 0.004334956407546997, + 0.004856918938457966, + 0.004625222645699978, + 0.005531080067157745, + 0.004500037524849176, + 0.003250712063163519, + 0.0035330806858837605, + 0.005873981397598982, + 0.012575427070260048, + 0.025777820497751236, + 0.012094293721020222, + 0.032489944249391556, + 0.0060813287273049355, + 0.005130477249622345, + 0.004921606741845608, + 0.0030822022818028927, + 0.005418200511485338, + 0.003370363963767886, + 0.0023713703267276287, + 0.003047708887606859, + 0.0026613774243742228, + 0.0018856084207072854, + 0.0021014423109591007, + 0.014654678292572498, + 0.007486063055694103, + 0.004325595684349537, + 0.005179870408028364, + 0.006381345447152853, + 0.00960143469274044, + 0.003325245575979352, + 0.003951436374336481, + 0.0035229274071753025, + 0.006356125231832266, + 0.007120099849998951, + 0.004757049027830362, + 0.007824295200407505, + 0.0037519221659749746, + 0.0028050513938069344, + 0.003026836086064577, + 0.029153816401958466, + 0.023707708343863487, + 0.016580723226070404, + 0.00699183065444231, + 0.010271608829498291, + 0.010134677402675152, + 0.007603133097290993, + 0.012738262303173542, + 0.011486293748021126, + 0.027899110689759254, + 0.009132296778261662, + 0.00882298406213522, + 0.008021431975066662, + 0.005033469758927822, + 0.005457805935293436, + 0.009161986410617828, + 0.04305187612771988, + 0.022890031337738037, + 0.06191718578338623, + 0.016250574961304665, + 0.018923981115221977, + 0.018109170719981194, + 0.02028409019112587, + 0.014083260670304298, + 0.008342298679053783, + 0.02139950729906559, + 0.03705356642603874, + 0.009466823190450668, + 0.00760821346193552, + 0.008218706585466862, + 0.006106148473918438, + 0.0050607905723154545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.06204963102936745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327875606715679 + ], + "opposite_logprob_tokens": [ + -3.7732787132263184 + ], + "logprob_reward": -0.02327875606715679, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327875606715679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43439048528671265, + 0.03366217762231827, + 0.08904126286506653, + 0.024875933304429054, + 0.04981979355216026, + 0.02792113833129406, + 0.040252622216939926, + 0.03254735469818115, + 0.19583165645599365, + 0.07165750861167908 + ], + "credit_map": [ + 0.026669666171073914, + 0.1306767761707306, + 0.02999967522919178, + 0.11016479879617691, + 0.06266593933105469, + 0.015638750046491623, + 0.040426596999168396, + 0.018148289993405342, + 0.002911859890446067, + 0.0044104065746068954, + 0.004927846137434244, + 0.005773094482719898, + 0.003787453519180417, + 0.002844058908522129, + 0.004200424067676067, + 0.004807036370038986, + 0.004673223476856947, + 0.020644918084144592, + 0.004777919501066208, + 0.008742689155042171, + 0.03431880846619606, + 0.004983165301382542, + 0.0031972709111869335, + 0.007703273557126522, + 0.0021248243283480406, + 0.005346771329641342, + 0.00418297853320837, + 0.0028225670102983713, + 0.0036121916491538286, + 0.0021204976364970207, + 0.00273769604973495, + 0.0019284095615148544, + 0.007529950235038996, + 0.012646240182220936, + 0.0038841930218040943, + 0.005253646057099104, + 0.010961645282804966, + 0.002742600627243519, + 0.0029928458388894796, + 0.0038086778949946165, + 0.005603817291557789, + 0.0066656856797635555, + 0.0036240231711417437, + 0.002284100977703929, + 0.004303477704524994, + 0.0019071617862209678, + 0.0021349911112338305, + 0.0013978827046230435, + 0.004826671443879604, + 0.013585630804300308, + 0.004483195953071117, + 0.0032931165769696236, + 0.0065369270741939545, + 0.0025842429604381323, + 0.0027411303017288446, + 0.0022017143201082945, + 0.0037983166985213757, + 0.006959817372262478, + 0.003393610240891576, + 0.0028229085728526115, + 0.004372796043753624, + 0.002550876932218671, + 0.002841431414708495, + 0.005807600915431976, + 0.038889143615961075, + 0.019796794280409813, + 0.0692620649933815, + 0.01427625585347414, + 0.015433992259204388, + 0.013681191019713879, + 0.013624097220599651, + 0.010868122801184654, + 0.00506708724424243, + 0.008431482128798962, + 0.016716867685317993, + 0.008261739276349545, + 0.004092019516974688, + 0.005162008572369814, + 0.004856778774410486, + 0.01906953752040863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.02327875606715679 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890313908457756 + ], + "opposite_logprob_tokens": [ + -2.464031457901001 + ], + "logprob_reward": -0.0890313908457756, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890313908457756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18651318550109863, + 0.04435200244188309, + 0.0936681479215622, + 0.03690243139863014, + 0.10251758992671967, + 0.07593119144439697, + 0.14278873801231384, + 0.03906671702861786, + 0.17452144622802734, + 0.10373854637145996 + ], + "credit_map": [ + 0.013011625036597252, + 0.04925188794732094, + 0.014002060517668724, + 0.05305929481983185, + 0.02029714174568653, + 0.008006327785551548, + 0.019148876890540123, + 0.009735950268805027, + 0.005631300620734692, + 0.0056986259296536446, + 0.006074030417948961, + 0.005390426609665155, + 0.008132732473313808, + 0.004468283150345087, + 0.004365534987300634, + 0.004591064061969519, + 0.005231809802353382, + 0.01783839613199234, + 0.013241234235465527, + 0.009376181289553642, + 0.03515685722231865, + 0.005326110403984785, + 0.0033395281061530113, + 0.004158025607466698, + 0.003970679827034473, + 0.007179200183600187, + 0.0061775511130690575, + 0.004909150768071413, + 0.00615981500595808, + 0.0030772085301578045, + 0.003129228251054883, + 0.0022995939943939447, + 0.017635105177760124, + 0.015981337055563927, + 0.009318104013800621, + 0.007893061265349388, + 0.01390335988253355, + 0.02094254083931446, + 0.0068696946837008, + 0.009974373504519463, + 0.013212453573942184, + 0.026614082977175713, + 0.005686217453330755, + 0.0065388306975364685, + 0.011909330263733864, + 0.004216496832668781, + 0.004568269941955805, + 0.003185506444424391, + 0.030795596539974213, + 0.04780534654855728, + 0.011918514966964722, + 0.008894325233995914, + 0.019566558301448822, + 0.007871217094361782, + 0.007520598825067282, + 0.008416576310992241, + 0.0051115103997290134, + 0.010996138677001, + 0.0035826065577566624, + 0.004130431916564703, + 0.004117776174098253, + 0.0026537736412137747, + 0.002944913227111101, + 0.005529563408344984, + 0.038151275366544724, + 0.023306846618652344, + 0.050852254033088684, + 0.015814298763871193, + 0.013136246241629124, + 0.009863351471722126, + 0.014223428443074226, + 0.009173734113574028, + 0.0066672093234956264, + 0.03046836145222187, + 0.01418488472700119, + 0.008709343150258064, + 0.008955313824117184, + 0.009078346192836761, + 0.005275459960103035, + 0.02039962448179722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0890313908457756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42884987592697144 + ], + "opposite_logprob_tokens": [ + -1.0538499355316162 + ], + "logprob_reward": -0.42884987592697144, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42884987592697144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4328267276287079, + 0.05201471969485283, + 0.20911897718906403, + 0.03273645415902138, + 0.07670395821332932, + 0.08234584331512451, + 0.04015962779521942, + 0.011410501785576344, + 0.034577157348394394, + 0.028106005862355232 + ], + "credit_map": [ + 0.02711491845548153, + 0.11287546902894974, + 0.03489641845226288, + 0.124260313808918, + 0.0499454028904438, + 0.0163076464086771, + 0.047856561839580536, + 0.01957000605762005, + 0.005517492536455393, + 0.006545030511915684, + 0.007794306613504887, + 0.00605255551636219, + 0.010379881598055363, + 0.00510236993432045, + 0.004879884421825409, + 0.005743201822042465, + 0.016609594225883484, + 0.036391619592905045, + 0.04853418469429016, + 0.02614133059978485, + 0.03773622587323189, + 0.012451809830963612, + 0.009413805790245533, + 0.021840404719114304, + 0.0025404472835361958, + 0.004837010521441698, + 0.004580030217766762, + 0.006100513506680727, + 0.004684068728238344, + 0.0035645070020109415, + 0.0038276880513876677, + 0.0026021895464509726, + 0.01787048950791359, + 0.0110288355499506, + 0.005419465713202953, + 0.004065040964633226, + 0.009667281061410904, + 0.016253603622317314, + 0.0057294005528092384, + 0.006669843103736639, + 0.017839357256889343, + 0.03009386733174324, + 0.008695383556187153, + 0.0058593726716935635, + 0.008921206928789616, + 0.003789119189605117, + 0.0036445718724280596, + 0.003502968233078718, + 0.011560169979929924, + 0.011942370794713497, + 0.0035200691781938076, + 0.0020896890200674534, + 0.0034158213529735804, + 0.00247050984762609, + 0.001947235083207488, + 0.0032137641683220863, + 0.0023272631224244833, + 0.002880111802369356, + 0.0009772857883945107, + 0.0013198318192735314, + 0.0012398420367389917, + 0.0005952440551482141, + 0.0011881485115736723, + 0.0008827746496535838, + 0.004759770352393389, + 0.00580650893971324, + 0.0035053433384746313, + 0.003889302723109722, + 0.0040657538920640945, + 0.004275170620530844, + 0.0027988858055323362, + 0.005476420279592276, + 0.0023078692611306906, + 0.004907297436147928, + 0.010691302828490734, + 0.0020654115360230207, + 0.0018809143221005797, + 0.0018752994947135448, + 0.0013997482601553202, + 0.0029781630728393793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.42884987592697144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly tilts right, revealing a surreal scene captured in stark, artistic black-and-white, 3D, Pixar-style. A vast, rolling sea, composed entirely of meticulously cultivated, high-standard farmland, stretches to the horizon. Rolling black cultivated hills and silvery crops mimic the undulations of waves beneath a high-contrast sky. A whimsical boat glides silently across this land-sea, its sleek, rounded hull of dark charcoal-gray wood and a friendly, stylized animal head perched on its prow. As the boat steadily approaches a distant village, its single, large, patchwork sail—a mosaic of various gray tones and textured fabrics—is gracefully and deliberately lowered, the canvas gently billowing before collapsing in neat folds onto the slightly bent mast. The destination, a village built in a distinctive ethnic minority style, becomes more clearly visible, its structures characterized by curved, dark-tiled roofs and intricate white gridwork, all organically nestled into the terraced slopes of the farmland hills. The entire dreamlike scene is defined by a dramatic play of light and shadow; vibrant white highlights the edges and rooftops of the boats, while deep, inky black pools in the furrows of the land, creating an atmosphere that is visually stunning and profoundly tranquil.", + "seed": 15877, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_15/index_7970_seed_15877_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1084949279204011, + "vlm_logprob": -0.1084949279204011, + "vlm_yes_no_margin": 3.650000011920929, + "vlm_token_credit": -0.1084949279204011, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005814661271870136 + ], + "opposite_logprob_tokens": [ + -7.505814552307129 + ], + "logprob_reward": -0.005814661271870136, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.005814661271870136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09272231161594391, + 0.2007399946451187, + 0.0830148234963417, + 0.13344323635101318, + 0.09506353735923767, + 0.06111111864447594, + 0.10538636147975922, + 0.15922491252422333, + 0.029691573232412338, + 0.03960211202502251 + ], + "credit_map": [ + 0.012619772925972939, + 0.013374440371990204, + 0.010087041184306145, + 0.01840912364423275, + 0.011368158273398876, + 0.010343554429709911, + 0.009047413244843483, + 0.007472797762602568, + 0.05035034939646721, + 0.019929509609937668, + 0.01528763398528099, + 0.009852207265794277, + 0.015274892561137676, + 0.014952798373997211, + 0.020086167380213737, + 0.05500642582774162, + 0.016757773235440254, + 0.008125795051455498, + 0.006398490630090237, + 0.024470046162605286, + 0.004378593061119318, + 0.008708790875971317, + 0.008755351416766644, + 0.005419977009296417, + 0.01949026621878147, + 0.0150435920804739, + 0.011975795030593872, + 0.028982792049646378, + 0.013392088003456593, + 0.013079035095870495, + 0.01609080284833908, + 0.01538886409252882, + 0.015671346336603165, + 0.014749671332538128, + 0.00828844215720892, + 0.023057982325553894, + 0.015248185954988003, + 0.006086244247853756, + 0.006493116728961468, + 0.00546854455024004, + 0.015019402839243412, + 0.00904776994138956, + 0.004690574482083321, + 0.01570594683289528, + 0.003611714579164982, + 0.0045028505846858025, + 0.003764429595321417, + 0.0047684321179986, + 0.011434726417064667, + 0.013539186678826809, + 0.010447756387293339, + 0.016210319474339485, + 0.01947232149541378, + 0.011267676018178463, + 0.014124481938779354, + 0.008889892138540745, + 0.04690026491880417, + 0.015552490949630737, + 0.034050822257995605, + 0.007911366410553455, + 0.012187848798930645, + 0.020401407033205032, + 0.009548330679535866, + 0.012672385200858116, + 0.00364335672929883, + 0.002572821220383048, + 0.005915489979088306, + 0.0028084374498575926, + 0.005613903980702162, + 0.004324954468756914, + 0.001629008911550045, + 0.0031836028210818768, + 0.004916141275316477, + 0.00365107087418437, + 0.004734826739877462, + 0.003212187672033906, + 0.006120216567069292, + 0.0057798889465630054, + 0.004406444262713194, + 0.006781337317079306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.005814661271870136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062049854546785355 + ], + "opposite_logprob_tokens": [ + -2.8120498657226562 + ], + "logprob_reward": -0.062049854546785355, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062049854546785355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08978570997714996, + 0.10980434715747833, + 0.05264183133840561, + 0.1982518583536148, + 0.14925120770931244, + 0.09869883209466934, + 0.054896000772714615, + 0.054545026272535324, + 0.11922907829284668, + 0.07289611548185349 + ], + "credit_map": [ + 0.012746130116283894, + 0.0103695597499609, + 0.012447199784219265, + 0.009581515565514565, + 0.011143231764435768, + 0.013592946343123913, + 0.009924978017807007, + 0.009980151429772377, + 0.02680654264986515, + 0.010152855888009071, + 0.008894745260477066, + 0.00590110570192337, + 0.006938817445188761, + 0.009480069391429424, + 0.011141342110931873, + 0.030488857999444008, + 0.005926096346229315, + 0.005846931599080563, + 0.004882706329226494, + 0.014902273193001747, + 0.004837732296437025, + 0.005670024082064629, + 0.005350090097635984, + 0.005225979257375002, + 0.02812616527080536, + 0.017095908522605896, + 0.01827310211956501, + 0.059776999056339264, + 0.016798093914985657, + 0.01251825038343668, + 0.030589191243052483, + 0.015074150636792183, + 0.02796456404030323, + 0.020816070958971977, + 0.015037093311548233, + 0.03247714042663574, + 0.013799348846077919, + 0.013413178734481335, + 0.013362260535359383, + 0.012381541542708874, + 0.01588285155594349, + 0.01576656475663185, + 0.007315553724765778, + 0.02008182927966118, + 0.010051984339952469, + 0.014988569542765617, + 0.00817266758531332, + 0.00643881456926465, + 0.005762916058301926, + 0.009690351784229279, + 0.00481760548427701, + 0.006938091479241848, + 0.005176149308681488, + 0.00950124952942133, + 0.008421383798122406, + 0.004588251933455467, + 0.01114192046225071, + 0.003008440835401416, + 0.018255887553095818, + 0.0030105821788311005, + 0.003993723075836897, + 0.0037752375937998295, + 0.002340538427233696, + 0.009018698707222939, + 0.016085129231214523, + 0.012528127990663052, + 0.011690524406731129, + 0.013950265012681484, + 0.027922214940190315, + 0.016451239585876465, + 0.008818847127258778, + 0.011782737448811531, + 0.005784618202596903, + 0.005757912527769804, + 0.006609020754694939, + 0.0043639750219881535, + 0.007991756312549114, + 0.013840680941939354, + 0.007515238597989082, + 0.021032918244600296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.062049854546785355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634377270936966 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634377270936966, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634377270936966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09663534164428711, + 0.14618812501430511, + 0.09238012880086899, + 0.2732733190059662, + 0.05081778019666672, + 0.07488452643156052, + 0.06595326960086823, + 0.045009125024080276, + 0.11296473443508148, + 0.0418936125934124 + ], + "credit_map": [ + 0.01403502281755209, + 0.016190236434340477, + 0.014152546413242817, + 0.01759045198559761, + 0.00761818652972579, + 0.011829804629087448, + 0.00683062756434083, + 0.008388476446270943, + 0.019703920930624008, + 0.022876983508467674, + 0.01739051192998886, + 0.013621637597680092, + 0.01625414751470089, + 0.019275633618235588, + 0.022066175937652588, + 0.014999119564890862, + 0.012594077736139297, + 0.01232040673494339, + 0.008102644234895706, + 0.02430632896721363, + 0.006884945556521416, + 0.00947526190429926, + 0.009527609683573246, + 0.009168864227831364, + 0.03513742610812187, + 0.04019659385085106, + 0.026608102023601532, + 0.031115928664803505, + 0.038393646478652954, + 0.02564927004277706, + 0.031850408762693405, + 0.0443219430744648, + 0.007496593985706568, + 0.00866602174937725, + 0.0053389729000627995, + 0.006886725779622793, + 0.006499610375612974, + 0.005634532775729895, + 0.004691169131547213, + 0.005604160483926535, + 0.012559807859361172, + 0.012151304632425308, + 0.009224803186953068, + 0.013650396838784218, + 0.007947170175611973, + 0.007460279855877161, + 0.006629645358771086, + 0.005261123646050692, + 0.008736246265470982, + 0.01627141609787941, + 0.005202292464673519, + 0.008973523043096066, + 0.008492793887853622, + 0.007614779286086559, + 0.005321328062564135, + 0.005340893752872944, + 0.008266260847449303, + 0.003958989400416613, + 0.0033228546380996704, + 0.0031811725348234177, + 0.006707549095153809, + 0.013135535642504692, + 0.0026257000863552094, + 0.0038110665045678616, + 0.013686961494386196, + 0.011908529326319695, + 0.01500506792217493, + 0.010436048731207848, + 0.021891148760914803, + 0.01860746368765831, + 0.008063641376793385, + 0.01336587779223919, + 0.00485039409250021, + 0.003654983825981617, + 0.0034913583658635616, + 0.002550953533500433, + 0.004835064057260752, + 0.005777120590209961, + 0.004617475904524326, + 0.012116262689232826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.008634377270936966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896928489208221 + ], + "opposite_logprob_tokens": [ + -2.5789692401885986 + ], + "logprob_reward": -0.07896928489208221, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896928489208221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07309744507074356, + 0.10107069462537766, + 0.14278210699558258, + 0.22029733657836914, + 0.06913064420223236, + 0.11099458485841751, + 0.1676059067249298, + 0.023712458088994026, + 0.04510832577943802, + 0.04620051756501198 + ], + "credit_map": [ + 0.009698471985757351, + 0.013239062391221523, + 0.008398856967687607, + 0.01119072176516056, + 0.009794327430427074, + 0.00912077073007822, + 0.005075131542980671, + 0.00658009760081768, + 0.02418529987335205, + 0.014107869938015938, + 0.007443465292453766, + 0.0049324920400977135, + 0.00663555134087801, + 0.009401713497936726, + 0.010038926266133785, + 0.024325378239154816, + 0.018979646265506744, + 0.012835915200412273, + 0.011506838724017143, + 0.04421835020184517, + 0.009687396697700024, + 0.01770162396132946, + 0.014102938584983349, + 0.013749394565820694, + 0.02929837256669998, + 0.021602388471364975, + 0.019207630306482315, + 0.06664226204156876, + 0.021320197731256485, + 0.011837818659842014, + 0.03433192893862724, + 0.0160567294806242, + 0.009302124381065369, + 0.00981422420591116, + 0.008073311299085617, + 0.0064549436792731285, + 0.014328251592814922, + 0.008025472052395344, + 0.00541719188913703, + 0.007715126033872366, + 0.022303663194179535, + 0.017271125689148903, + 0.01112857274711132, + 0.021849436685442924, + 0.0092804916203022, + 0.012404737062752247, + 0.008668337017297745, + 0.008088221773505211, + 0.020302947610616684, + 0.02774169109761715, + 0.014191117137670517, + 0.023348333314061165, + 0.019934650510549545, + 0.027647405862808228, + 0.019408447667956352, + 0.015031308867037296, + 0.0028253435157239437, + 0.002220141002908349, + 0.003959409426897764, + 0.0022745849564671516, + 0.003262866288423538, + 0.003702583024278283, + 0.002122153528034687, + 0.0033453768119215965, + 0.004914470482617617, + 0.005750446580350399, + 0.0064628878608345985, + 0.004434763453900814, + 0.007673732936382294, + 0.00718270568177104, + 0.0034451864194124937, + 0.005244134459644556, + 0.005006616935133934, + 0.005686984863132238, + 0.004885759204626083, + 0.003370488528162241, + 0.007193894125521183, + 0.006303633097559214, + 0.003858080832287669, + 0.00989506021142006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.07896928489208221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870064616203308 + ], + "opposite_logprob_tokens": [ + -1.1370065212249756 + ], + "logprob_reward": -0.3870064616203308, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870064616203308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05088247358798981, + 0.17830421030521393, + 0.04831160977482796, + 0.08864232152700424, + 0.16879308223724365, + 0.10931580513715744, + 0.23234480619430542, + 0.05079269781708717, + 0.05454758554697037, + 0.018065443262457848 + ], + "credit_map": [ + 0.0073492564260959625, + 0.010574032552540302, + 0.0062344856560230255, + 0.008064123801887035, + 0.005063580349087715, + 0.005496790632605553, + 0.003952705767005682, + 0.0041474997997283936, + 0.040159691125154495, + 0.021388325840234756, + 0.013759035617113113, + 0.009328429587185383, + 0.015579132363200188, + 0.01542283222079277, + 0.01809091307222843, + 0.044575873762369156, + 0.010666701011359692, + 0.006063045002520084, + 0.0039355396293103695, + 0.006409840192645788, + 0.004347098525613546, + 0.00854902621358633, + 0.004054875113070011, + 0.004285484552383423, + 0.011911059729754925, + 0.011460654437541962, + 0.009142374619841576, + 0.012654904276132584, + 0.00957876443862915, + 0.008943449705839157, + 0.00999121367931366, + 0.014959893189370632, + 0.03317801281809807, + 0.01680115982890129, + 0.019362492486834526, + 0.056729622185230255, + 0.011759349144995213, + 0.005964452400803566, + 0.013860865496098995, + 0.011137131601572037, + 0.01198834553360939, + 0.017815664410591125, + 0.013195926323533058, + 0.01636243797838688, + 0.01142806001007557, + 0.016917632892727852, + 0.01281061116605997, + 0.008797123096883297, + 0.025641733780503273, + 0.03076283074915409, + 0.02065703272819519, + 0.04096032679080963, + 0.02727583609521389, + 0.033011406660079956, + 0.031863197684288025, + 0.022172437980771065, + 0.004981614649295807, + 0.003989148885011673, + 0.010203110985457897, + 0.004201202187687159, + 0.0072675966657698154, + 0.011653780937194824, + 0.003382046241313219, + 0.005114195868372917, + 0.005836683791130781, + 0.005755217745900154, + 0.008084744215011597, + 0.0052808658219873905, + 0.011879543773829937, + 0.008745796047151089, + 0.003739504376426339, + 0.005225229077041149, + 0.002057946752756834, + 0.0019937818869948387, + 0.00228657154366374, + 0.001470181392505765, + 0.0032630956266075373, + 0.002086285501718521, + 0.0016655916115269065, + 0.0032419892959296703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.3870064616203308 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly tilts right, revealing a surreal scene captured in stark, artistic black-and-white, 3D, Pixar-style. A vast, rolling sea, composed entirely of meticulously cultivated, high-standard farmland, stretches to the horizon. Rolling black cultivated hills and silvery crops mimic the undulations of waves beneath a high-contrast sky. A whimsical boat glides silently across this land-sea, its sleek, rounded hull of dark charcoal-gray wood and a friendly, stylized animal head perched on its prow. As the boat steadily approaches a distant village, its single, large, patchwork sail—a mosaic of various gray tones and textured fabrics—is gracefully and deliberately lowered, the canvas gently billowing before collapsing in neat folds onto the slightly bent mast. The destination, a village built in a distinctive ethnic minority style, becomes more clearly visible, its structures characterized by curved, dark-tiled roofs and intricate white gridwork, all organically nestled into the terraced slopes of the farmland hills. The entire dreamlike scene is defined by a dramatic play of light and shadow; vibrant white highlights the edges and rooftops of the boats, while deep, inky black pools in the furrows of the land, creating an atmosphere that is visually stunning and profoundly tranquil.", + "seed": 15878, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_15/index_7970_seed_15878_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10673234853893518, + "vlm_logprob": -0.10673234853893518, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.10673234853893518, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03889463096857071 + ], + "opposite_logprob_tokens": [ + -3.4138946533203125 + ], + "logprob_reward": -0.03889463096857071, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03889463096857071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18174704909324646, + 0.04395296424627304, + 0.15049199759960175, + 0.13534867763519287, + 0.1791910082101822, + 0.040516965091228485, + 0.04943288862705231, + 0.027360137552022934, + 0.030668437480926514, + 0.16128985583782196 + ], + "credit_map": [ + 0.01795835793018341, + 0.018396181985735893, + 0.043549831956624985, + 0.026627354323863983, + 0.014790180139243603, + 0.03291463479399681, + 0.012003535404801369, + 0.015506965108215809, + 0.007964645512402058, + 0.005545003805309534, + 0.00511241564527154, + 0.003922837320715189, + 0.006763699930161238, + 0.006034212186932564, + 0.005082293879240751, + 0.0035278529394418, + 0.022917285561561584, + 0.018860941752791405, + 0.015611083246767521, + 0.012458251789212227, + 0.02690006233751774, + 0.026909533888101578, + 0.013347258791327477, + 0.01348756905645132, + 0.03895227611064911, + 0.008728032000362873, + 0.008389129303395748, + 0.009210363030433655, + 0.00914190523326397, + 0.014903122559189796, + 0.01246389839798212, + 0.03355993703007698, + 0.021406715735793114, + 0.02114744856953621, + 0.009984682314097881, + 0.02205139771103859, + 0.024084534496068954, + 0.02287660539150238, + 0.028918104246258736, + 0.02872149460017681, + 0.005295964889228344, + 0.006675741635262966, + 0.004595078527927399, + 0.004510816652327776, + 0.006924851797521114, + 0.003864269470795989, + 0.0044160219840705395, + 0.004234216641634703, + 0.00319381570443511, + 0.006912352982908487, + 0.009997464716434479, + 0.009535177610814571, + 0.0047894734889268875, + 0.003965386189520359, + 0.00605620676651597, + 0.0049830032512545586, + 0.0033511498477309942, + 0.00549713708460331, + 0.0027437428943812847, + 0.002800872316583991, + 0.0034802155569195747, + 0.002230791375041008, + 0.0035003323573619127, + 0.0037558949552476406, + 0.006559939589351416, + 0.002759354654699564, + 0.004577016923576593, + 0.004667149391025305, + 0.0022179100196808577, + 0.0015221062349155545, + 0.004345928318798542, + 0.004019029904156923, + 0.017885511741042137, + 0.013621658086776733, + 0.026209531351923943, + 0.019974324852228165, + 0.018460314720869064, + 0.014706462621688843, + 0.029865272343158722, + 0.020566774532198906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.03889463096857071 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280582964420319 + ], + "opposite_logprob_tokens": [ + -2.017805814743042 + ], + "logprob_reward": -0.14280582964420319, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280582964420319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11230291426181793, + 0.14391735196113586, + 0.3765620291233063, + 0.03950538486242294, + 0.026496125385165215, + 0.04745131731033325, + 0.03545054420828819, + 0.035187479108572006, + 0.06642518937587738, + 0.116701640188694 + ], + "credit_map": [ + 0.00946076400578022, + 0.012129373848438263, + 0.02332954853773117, + 0.02014026790857315, + 0.008625014685094357, + 0.015317302197217941, + 0.010177401825785637, + 0.013123232871294022, + 0.018073685467243195, + 0.015843361616134644, + 0.012133349664509296, + 0.011846001259982586, + 0.018598122522234917, + 0.027930445969104767, + 0.018042245879769325, + 0.021450133994221687, + 0.0565786212682724, + 0.051987968385219574, + 0.05230248346924782, + 0.02861238829791546, + 0.06110811233520508, + 0.06781098991632462, + 0.0313979908823967, + 0.026763472706079483, + 0.006138158030807972, + 0.0028624439146369696, + 0.004019287880510092, + 0.005592133849859238, + 0.0043092514388263226, + 0.0053415787406265736, + 0.0054542915895581245, + 0.005788242910057306, + 0.0035859616473317146, + 0.003129686927422881, + 0.001660362584516406, + 0.002895205281674862, + 0.004238741006702185, + 0.0036964542232453823, + 0.00304250605404377, + 0.004247208125889301, + 0.00763612100854516, + 0.008953969925642014, + 0.0062122889794409275, + 0.005139047279953957, + 0.006705237552523613, + 0.0028306865133345127, + 0.004589703865349293, + 0.0053842621855437756, + 0.0030314880423247814, + 0.0055306581780314445, + 0.0034796977415680885, + 0.0037852851673960686, + 0.004995674826204777, + 0.003620510222390294, + 0.0062093716114759445, + 0.004797858651727438, + 0.0034916014410555363, + 0.0033099285792559385, + 0.004756751470267773, + 0.004519353155046701, + 0.004251259844750166, + 0.003278236137703061, + 0.006603515241295099, + 0.004976832773536444, + 0.011960574425756931, + 0.006243119947612286, + 0.00938485935330391, + 0.00928745698183775, + 0.006284966599196196, + 0.0041783214546740055, + 0.010725744999945164, + 0.008360149338841438, + 0.01333845779299736, + 0.009153307415544987, + 0.020459545776247978, + 0.014748143963515759, + 0.017443453893065453, + 0.01293033454567194, + 0.015099861659109592, + 0.01352854073047638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.14280582964420319 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636386133730412 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636386133730412, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636386133730412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13778051733970642, + 0.14357724785804749, + 0.1764800101518631, + 0.058633655309677124, + 0.13851991295814514, + 0.029336614534258842, + 0.14942730963230133, + 0.059570591896772385, + 0.04563888534903526, + 0.06103525310754776 + ], + "credit_map": [ + 0.014780910685658455, + 0.019759874790906906, + 0.020960843190550804, + 0.014031282626092434, + 0.01899542845785618, + 0.02077423594892025, + 0.016990086063742638, + 0.011487853713333607, + 0.01500172819942236, + 0.016660617664456367, + 0.015408921055495739, + 0.012199378572404385, + 0.020812848582863808, + 0.02685653418302536, + 0.02087012492120266, + 0.015767084434628487, + 0.021981284022331238, + 0.01928413100540638, + 0.04207289591431618, + 0.01350452657788992, + 0.019735489040613174, + 0.02194453775882721, + 0.012751158326864243, + 0.025205999612808228, + 0.01126693096011877, + 0.005616889335215092, + 0.005249883513897657, + 0.005838530138134956, + 0.006066642701625824, + 0.008749213069677353, + 0.005885436199605465, + 0.009960134513676167, + 0.01691845990717411, + 0.013852227479219437, + 0.009164512157440186, + 0.014795688912272453, + 0.02533627860248089, + 0.016570109874010086, + 0.02168995514512062, + 0.02019267901778221, + 0.004955680575221777, + 0.0030539643485099077, + 0.003620541887357831, + 0.0036514669191092253, + 0.0038206351455301046, + 0.0023402858059853315, + 0.0036900590639561415, + 0.004203980788588524, + 0.00794617272913456, + 0.012925426475703716, + 0.028187228366732597, + 0.01534580159932375, + 0.020548371598124504, + 0.01834881119430065, + 0.021038763225078583, + 0.02508673071861267, + 0.00438427971675992, + 0.005608005449175835, + 0.007188072893768549, + 0.005458869971334934, + 0.015131128951907158, + 0.007114253006875515, + 0.00914792064577341, + 0.0055380635894834995, + 0.0036951194051653147, + 0.0017232054378837347, + 0.010683879256248474, + 0.006126431282609701, + 0.004561533220112324, + 0.0033094093669205904, + 0.003874867456033826, + 0.011664440855383873, + 0.006328146439045668, + 0.0057298121973872185, + 0.008563674986362457, + 0.008622941561043262, + 0.009025508537888527, + 0.005611812695860863, + 0.009604023769497871, + 0.0075493305921554565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.008636386133730412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204458698630333 + ], + "opposite_logprob_tokens": [ + -2.812044620513916 + ], + "logprob_reward": -0.06204458698630333, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204458698630333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06149880960583687, + 0.05347215011715889, + 0.2610771358013153, + 0.142622709274292, + 0.09747162461280823, + 0.05039015784859657, + 0.15053394436836243, + 0.061654675751924515, + 0.05237870290875435, + 0.0688999816775322 + ], + "credit_map": [ + 0.009960196912288666, + 0.008602209389209747, + 0.00849818717688322, + 0.008040910586714745, + 0.006807935889810324, + 0.0065751271322369576, + 0.00525135500356555, + 0.007762888912111521, + 0.007794469129294157, + 0.006275613326579332, + 0.005947696976363659, + 0.005711798556149006, + 0.005588978994637728, + 0.01061239093542099, + 0.004547072574496269, + 0.006994131952524185, + 0.03175734728574753, + 0.02207036316394806, + 0.025340469554066658, + 0.012119024060666561, + 0.06358399242162704, + 0.0382995530962944, + 0.023761147633194923, + 0.04414527118206024, + 0.019579190760850906, + 0.01220203097909689, + 0.009051165543496609, + 0.019871966913342476, + 0.0126156285405159, + 0.015646973624825478, + 0.02842223085463047, + 0.02523353137075901, + 0.010085511952638626, + 0.007440689019858837, + 0.007428415585309267, + 0.011305011808872223, + 0.021074755117297173, + 0.014929918572306633, + 0.015312518924474716, + 0.00989480596035719, + 0.010395723395049572, + 0.006225604563951492, + 0.007285014260560274, + 0.005998961161822081, + 0.006509251426905394, + 0.003960005939006805, + 0.004721346776932478, + 0.005294255446642637, + 0.009167972020804882, + 0.008700123988091946, + 0.031521424651145935, + 0.0160562414675951, + 0.016602857038378716, + 0.014046799391508102, + 0.02119622565805912, + 0.033242326229810715, + 0.0038676110561937094, + 0.006481988821178675, + 0.006614925339818001, + 0.005383237265050411, + 0.014769787900149822, + 0.0056656659580767155, + 0.0122300386428833, + 0.006641422864049673, + 0.005864609498530626, + 0.00233271811157465, + 0.013340241275727749, + 0.007622265722602606, + 0.002724561607465148, + 0.0023486369755119085, + 0.0044906181283295155, + 0.013655057176947594, + 0.008097155950963497, + 0.005692465230822563, + 0.010919451713562012, + 0.009604604914784431, + 0.007893174886703491, + 0.006115452386438847, + 0.01222982443869114, + 0.008347860537469387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.06204458698630333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in stark black-and-white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast rolling sea composed entirely of cultivated farmland hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boat with a dark charcoal-gray hull and a stylized animal head on the prow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat's patchwork sail shown being lowered and collapsing onto the mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812803089618683 + ], + "opposite_logprob_tokens": [ + -1.406280279159546 + ], + "logprob_reward": -0.2812803089618683, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812803089618683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10684898495674133, + 0.05156997963786125, + 0.11938533931970596, + 0.2790296673774719, + 0.09618411213159561, + 0.12747003138065338, + 0.05156479403376579, + 0.06764280796051025, + 0.06485921144485474, + 0.03544509410858154 + ], + "credit_map": [ + 0.009138754568994045, + 0.00887314137071371, + 0.024572327733039856, + 0.018372774124145508, + 0.010255004279315472, + 0.02014916203916073, + 0.0073333666659891605, + 0.00815445464104414, + 0.007827484048902988, + 0.007990279234945774, + 0.005398080218583345, + 0.004390412010252476, + 0.005686750169843435, + 0.01013660803437233, + 0.004563585855066776, + 0.0055767809972167015, + 0.017275499179959297, + 0.017942508682608604, + 0.01871226169168949, + 0.012489530257880688, + 0.014152415096759796, + 0.015207779593765736, + 0.015300234779715538, + 0.00830510351806879, + 0.054926544427871704, + 0.019268207252025604, + 0.026301594451069832, + 0.03567365184426308, + 0.021979544311761856, + 0.02412826381623745, + 0.0372789166867733, + 0.0594729483127594, + 0.011545049957931042, + 0.012576869688928127, + 0.007821742445230484, + 0.015165591612458229, + 0.012478181160986423, + 0.012763725593686104, + 0.011826536618173122, + 0.012006417848169804, + 0.02255268767476082, + 0.017219798639416695, + 0.01566133089363575, + 0.013401361182332039, + 0.018509171903133392, + 0.009954302571713924, + 0.015213235281407833, + 0.014958149753510952, + 0.003100430127233267, + 0.00393359549343586, + 0.003420684952288866, + 0.0038339525926858187, + 0.007394380401819944, + 0.006372260395437479, + 0.01231697853654623, + 0.011192514561116695, + 0.004857400897890329, + 0.007907161489129066, + 0.005298811476677656, + 0.00607882346957922, + 0.01683332957327366, + 0.006761739030480385, + 0.013199209235608578, + 0.0067063323222100735, + 0.008104627020657063, + 0.00633225729689002, + 0.008180228993296623, + 0.010391185991466045, + 0.00912020355463028, + 0.005836110096424818, + 0.008539254777133465, + 0.008355340920388699, + 0.004100141115486622, + 0.0024536983110010624, + 0.005191904027014971, + 0.006027731113135815, + 0.003845800179988146, + 0.0035787008237093687, + 0.0057893493212759495, + 0.004457768052816391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.2812803089618683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_tensor([5087, 5087])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_tensor([5087, 5087])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fdca59d6a6b25117bf552017bf652316e21b1ad5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_tensor([5087, 5087])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking underwater close-up, panning downward and right in a smooth, steady rhythm, descends into the murky, green-brown waters of a river laden with suspended particles, focusing on a red-bellied piranha. Its silver-gray scales gleam with a cold metallic sheen in the faint light filtering through the surface, its belly a startling blood-red, a stark contrast to the gloomy surroundings. With a sudden, explosive snap, its powerful jaws flung open, revealing two rows of razor-sharp, tightly interlocked triangular teeth. This menacing movement churns the surrounding water, sending tiny particles of sand and mud forming a chaotic vortex in its path. As the camera pans, every subtle yet powerful twitch of its muscles beneath its gill covers and on its flanks is captured with pinpoint photorealism, vividly conveying the predator's primal, hair-trigger aggression and creating a breathtaking atmosphere of tension.", + "seed": 27712, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_5087_seed_27712_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17499346458353102, + "vlm_logprob": -0.17499346458353102, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.17499346458353102, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0034151107538491488 + ], + "opposite_logprob_tokens": [ + -7.003415107727051 + ], + "logprob_reward": -0.0034151107538491488, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0034151107538491488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1832929253578186, + 0.05587435141205788, + 0.10915815830230713, + 0.2536280155181885, + 0.04555241018533707, + 0.05379045009613037, + 0.05816049501299858, + 0.020486362278461456, + 0.05537796765565872, + 0.16467884182929993 + ], + "credit_map": [ + 0.01885882392525673, + 0.023918624967336655, + 0.04125964641571045, + 0.020414572209119797, + 0.02522663027048111, + 0.026753295212984085, + 0.010467784479260445, + 0.016393551602959633, + 0.00638712290674448, + 0.006708895321935415, + 0.007415486499667168, + 0.0053457776084542274, + 0.010767663829028606, + 0.006728582549840212, + 0.0057885111309587955, + 0.006732311099767685, + 0.016975652426481247, + 0.009742849506437778, + 0.01377162616699934, + 0.009483571164309978, + 0.01173757016658783, + 0.019078271463513374, + 0.012053175829350948, + 0.016315443441271782, + 0.07867908477783203, + 0.008975679986178875, + 0.09173541516065598, + 0.021862363442778587, + 0.017736241221427917, + 0.010454733856022358, + 0.016405513510107994, + 0.0077789886854588985, + 0.00424847612157464, + 0.00467716483399272, + 0.006086141802370548, + 0.004903234075754881, + 0.0036022812128067017, + 0.005104055628180504, + 0.007477038074284792, + 0.009454017505049706, + 0.003600513329729438, + 0.004655524622648954, + 0.006498883478343487, + 0.008532041683793068, + 0.006265007890760899, + 0.006212327629327774, + 0.008627144619822502, + 0.009399005211889744, + 0.004214882384985685, + 0.004900186788290739, + 0.0057918233796954155, + 0.008772307075560093, + 0.005195101257413626, + 0.004637296311557293, + 0.01256228145211935, + 0.012086617760360241, + 0.0013974527828395367, + 0.0017642304301261902, + 0.0022105451207607985, + 0.0031300601549446583, + 0.0018477232661098242, + 0.001605518045835197, + 0.004495054017752409, + 0.004035778343677521, + 0.004137199837714434, + 0.004160428419709206, + 0.009319203905761242, + 0.008600103668868542, + 0.006419765297323465, + 0.008022054098546505, + 0.007201273925602436, + 0.00751793896779418, + 0.013027236796915531, + 0.015695219859480858, + 0.025018366053700447, + 0.030308706685900688, + 0.015627646818757057, + 0.01742709055542946, + 0.02587788738310337, + 0.021696701645851135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.0034151107538491488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040998938493430614 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040998938493430614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040998938493430614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12190108746290207, + 0.05874079093337059, + 0.06417030841112137, + 0.35422664880752563, + 0.061722129583358765, + 0.07338092476129532, + 0.07351980358362198, + 0.014714030548930168, + 0.03052617609500885, + 0.14709803462028503 + ], + "credit_map": [ + 0.01482128445059061, + 0.01045604981482029, + 0.016918886452913284, + 0.009991698898375034, + 0.026893990114331245, + 0.02529049664735794, + 0.007321920711547136, + 0.01020675990730524, + 0.00558504881337285, + 0.006612374447286129, + 0.008137763477861881, + 0.006486451253294945, + 0.010077078826725483, + 0.009361913427710533, + 0.006727846339344978, + 0.005752319470047951, + 0.010799494571983814, + 0.006299227476119995, + 0.008533072657883167, + 0.006795939058065414, + 0.0073721809312701225, + 0.011941153556108475, + 0.005163934081792831, + 0.007265310734510422, + 0.10595926642417908, + 0.012632757425308228, + 0.12854096293449402, + 0.032116230577230453, + 0.02251121588051319, + 0.016690952703356743, + 0.021236268803477287, + 0.014538994058966637, + 0.004468332976102829, + 0.005389205180108547, + 0.00928410328924656, + 0.005771222058683634, + 0.00482726376503706, + 0.005956017412245274, + 0.011541735380887985, + 0.014484253711998463, + 0.004989320412278175, + 0.00738392723724246, + 0.010213086381554604, + 0.012230562046170235, + 0.008191854692995548, + 0.00691985571756959, + 0.010976909659802914, + 0.012475406751036644, + 0.0050858124159276485, + 0.005630548112094402, + 0.008713025599718094, + 0.013242830522358418, + 0.004999953787773848, + 0.004903637804090977, + 0.01738208346068859, + 0.013561917468905449, + 0.0008819366921670735, + 0.0010934374295175076, + 0.0015626013046130538, + 0.002640276215970516, + 0.0010410555405542254, + 0.001322506577707827, + 0.0031276338268071413, + 0.003044583136215806, + 0.0022135868202894926, + 0.003146107541397214, + 0.003851726185530424, + 0.004256462212651968, + 0.005507205612957478, + 0.004224022850394249, + 0.003502835752442479, + 0.003824227722361684, + 0.010536235757172108, + 0.0175419170409441, + 0.01746929995715618, + 0.027157621458172798, + 0.013249840587377548, + 0.015510734170675278, + 0.024130435660481453, + 0.021501947194337845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0040998938493430614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076298960484564304 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.0076298960484564304, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076298960484564304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13288339972496033, + 0.0822320282459259, + 0.2605544924736023, + 0.2952280342578888, + 0.05266706272959709, + 0.05350141227245331, + 0.03293078765273094, + 0.014457779936492443, + 0.029493482783436775, + 0.04605155810713768 + ], + "credit_map": [ + 0.013618738390505314, + 0.013259119354188442, + 0.03071240894496441, + 0.009481380693614483, + 0.0260903462767601, + 0.020923973992466927, + 0.0069146715104579926, + 0.011882752180099487, + 0.01010141707956791, + 0.009180216118693352, + 0.014570397324860096, + 0.007815886288881302, + 0.011371789500117302, + 0.012986390851438046, + 0.0077186692506074905, + 0.008487259969115257, + 0.029083890840411186, + 0.025280972942709923, + 0.02310408465564251, + 0.017232606187462807, + 0.03488515317440033, + 0.04496103897690773, + 0.033187899738550186, + 0.052818842232227325, + 0.08443194627761841, + 0.011143915355205536, + 0.10145222395658493, + 0.027578385546803474, + 0.015942871570587158, + 0.013180702924728394, + 0.01869257725775242, + 0.022805364802479744, + 0.0042161159217357635, + 0.006363610737025738, + 0.010162951424717903, + 0.006263220217078924, + 0.0043484424240887165, + 0.006100012920796871, + 0.007898589596152306, + 0.007314110640436411, + 0.0031467662192881107, + 0.0033606253564357758, + 0.004969948902726173, + 0.006914264988154173, + 0.005328480154275894, + 0.0047742282040417194, + 0.01351103000342846, + 0.011496064253151417, + 0.0016189274610951543, + 0.0018403506837785244, + 0.003305817721411586, + 0.005489532835781574, + 0.002496320754289627, + 0.0023542302660644054, + 0.008228042162954807, + 0.007597563322633505, + 0.0008944332366809249, + 0.001102991634979844, + 0.0019055656157433987, + 0.002305209869518876, + 0.0011130367638543248, + 0.0011969023616984487, + 0.0032123129349201918, + 0.0027273257728666067, + 0.0029829987324774265, + 0.002624709624797106, + 0.00525597995147109, + 0.005525730550289154, + 0.0020912380423396826, + 0.002096355427056551, + 0.004641829989850521, + 0.004274637438356876, + 0.0034646866843104362, + 0.0039253151044249535, + 0.008740813471376896, + 0.008319336920976639, + 0.003679430577903986, + 0.0035795108415186405, + 0.006848507560789585, + 0.00749395415186882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0076298960484564304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699482679367065 + ], + "opposite_logprob_tokens": [ + -3.4086995124816895 + ], + "logprob_reward": -0.033699482679367065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699482679367065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16259132325649261, + 0.06249609217047691, + 0.09661172330379486, + 0.1029907763004303, + 0.18908603489398956, + 0.09384379535913467, + 0.08986678719520569, + 0.025636645033955574, + 0.09630002826452255, + 0.08057678490877151 + ], + "credit_map": [ + 0.014341684989631176, + 0.013991410844027996, + 0.025599010288715363, + 0.01821288652718067, + 0.038279853761196136, + 0.024886976927518845, + 0.012770572677254677, + 0.014508932828903198, + 0.008280051872134209, + 0.007416584994643927, + 0.00767196761444211, + 0.007619334850460291, + 0.009042949415743351, + 0.00681375153362751, + 0.00829894095659256, + 0.0073525127954781055, + 0.012123234570026398, + 0.01050681620836258, + 0.00906155351549387, + 0.007267030421644449, + 0.012743504717946053, + 0.01651659607887268, + 0.009767978452146053, + 0.018625011667609215, + 0.01887214370071888, + 0.011056472547352314, + 0.01562188658863306, + 0.012179752811789513, + 0.008181183598935604, + 0.00994049571454525, + 0.01156616397202015, + 0.015572665259242058, + 0.011506050825119019, + 0.014005032368004322, + 0.03443130478262901, + 0.02502485178411007, + 0.017646661028265953, + 0.020521456375718117, + 0.030603483319282532, + 0.035347193479537964, + 0.006241851020604372, + 0.007297741714864969, + 0.010323193855583668, + 0.011081931181252003, + 0.009750555269420147, + 0.008654983714222908, + 0.021333305165171623, + 0.019160227850079536, + 0.006846373900771141, + 0.0071104844100773335, + 0.011469461023807526, + 0.012704714201390743, + 0.00773404398933053, + 0.008846551179885864, + 0.016644615679979324, + 0.018510542809963226, + 0.0017918347148224711, + 0.0024173040874302387, + 0.003807262284681201, + 0.00369820068590343, + 0.0035983985289931297, + 0.0031709305476397276, + 0.0037773095536977053, + 0.003375407075509429, + 0.010908388532698154, + 0.010199977084994316, + 0.01706138253211975, + 0.019145939499139786, + 0.0072786821983754635, + 0.005302103236317635, + 0.01341936644166708, + 0.012984178960323334, + 0.005872056819498539, + 0.00789846945554018, + 0.010383924469351768, + 0.015034734271466732, + 0.007091133389621973, + 0.006809063255786896, + 0.014772067777812481, + 0.012715335935354233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.033699482679367065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261229395866394 + ], + "opposite_logprob_tokens": [ + -0.5761229395866394 + ], + "logprob_reward": -0.8261229395866394, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261229395866394, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3112942576408386, + 0.059190984815359116, + 0.046864889562129974, + 0.10035036504268646, + 0.04116705805063248, + 0.10394984483718872, + 0.0826241746544838, + 0.0330873467028141, + 0.11362746357917786, + 0.1078435480594635 + ], + "credit_map": [ + 0.025052716955542564, + 0.025240503251552582, + 0.05573112145066261, + 0.024438511580228806, + 0.08400892466306686, + 0.055969882756471634, + 0.017676714807748795, + 0.02317589893937111, + 0.007075690198689699, + 0.005259085446596146, + 0.007743336260318756, + 0.0068405019119381905, + 0.007305910345166922, + 0.007852032780647278, + 0.007853349670767784, + 0.009261083789169788, + 0.007893277332186699, + 0.006158430129289627, + 0.008132806047797203, + 0.004802130628377199, + 0.005737140774726868, + 0.006006797309964895, + 0.0034176893532276154, + 0.0047166249714791775, + 0.023042399436235428, + 0.010140642523765564, + 0.023378895595669746, + 0.009519640356302261, + 0.00749750854447484, + 0.008830811828374863, + 0.006715900264680386, + 0.011224571615457535, + 0.002938761143013835, + 0.003114774590358138, + 0.010072881355881691, + 0.005075853783637285, + 0.003270162735134363, + 0.003905459074303508, + 0.006142439786344767, + 0.006646730005741119, + 0.005864668171852827, + 0.007488318718969822, + 0.010970360599458218, + 0.01357298344373703, + 0.009712892584502697, + 0.009251724928617477, + 0.025517970323562622, + 0.021570928394794464, + 0.007645017467439175, + 0.00770003255456686, + 0.01186373457312584, + 0.01199785154312849, + 0.007676501292735338, + 0.008684897795319557, + 0.0134721165522933, + 0.013584022410213947, + 0.002051099669188261, + 0.003092522732913494, + 0.004327309317886829, + 0.004471668042242527, + 0.003385781543329358, + 0.0038867800030857325, + 0.00585354445502162, + 0.006018645595759153, + 0.013142243959009647, + 0.01086929626762867, + 0.020607050508260727, + 0.023488793522119522, + 0.009277179837226868, + 0.00706548523157835, + 0.014180963858962059, + 0.01499644760042429, + 0.008188302628695965, + 0.009884817525744438, + 0.01351334061473608, + 0.019439609721302986, + 0.009627381339669228, + 0.009248205460608006, + 0.02006162516772747, + 0.0178802739828825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.8261229395866394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "striking underwater close-up, panning downward and right in a smooth, steady rhythm, descends into the murky, green-brown waters of a river laden with suspended particles, focusing on a red-bellied piranha. Its silver-gray scales gleam with a cold metallic sheen in the faint light filtering through the surface, its belly a startling blood-red, a stark contrast to the gloomy surroundings. With a sudden, explosive snap, its powerful jaws flung open, revealing two rows of razor-sharp, tightly interlocked triangular teeth. This menacing movement churns the surrounding water, sending tiny particles of sand and mud forming a chaotic vortex in its path. As the camera pans, every subtle yet powerful twitch of its muscles beneath its gill covers and on its flanks is captured with pinpoint photorealism, vividly conveying the predator's primal, hair-trigger aggression and creating a breathtaking atmosphere of tension.", + "seed": 27713, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_5087_seed_27713_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07376531902700663, + "vlm_logprob": -0.07376531902700663, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.07376531902700663, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004800698719918728 + ], + "opposite_logprob_tokens": [ + -6.754800796508789 + ], + "logprob_reward": -0.004800698719918728, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.004800698719918728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24917051196098328, + 0.08478743582963943, + 0.04629937931895256, + 0.057604871690273285, + 0.12214930355548859, + 0.06820262968540192, + 0.1055014580488205, + 0.10613330453634262, + 0.029952790588140488, + 0.13019828498363495 + ], + "credit_map": [ + 0.00998133048415184, + 0.008292448706924915, + 0.013073623180389404, + 0.09079013764858246, + 0.009349350817501545, + 0.014694968238472939, + 0.07983977347612381, + 0.023148881271481514, + 0.007905413396656513, + 0.005190678872168064, + 0.005974083673208952, + 0.007817904464900494, + 0.01667795144021511, + 0.008182992227375507, + 0.014895093627274036, + 0.018143322318792343, + 0.004173114895820618, + 0.004990588873624802, + 0.005061780568212271, + 0.007154583465307951, + 0.007590584456920624, + 0.005263111554086208, + 0.0067905280739068985, + 0.0052750869654119015, + 0.004055837634950876, + 0.0057127648033201694, + 0.0070293499156832695, + 0.006918344181030989, + 0.005598615389317274, + 0.013030353002250195, + 0.006277089007198811, + 0.008982516825199127, + 0.010197661817073822, + 0.006741927936673164, + 0.033313900232315063, + 0.025180991739034653, + 0.005604721140116453, + 0.009647182188928127, + 0.015532738529145718, + 0.01593017578125, + 0.00780548295006156, + 0.007928231731057167, + 0.008316857740283012, + 0.008618045598268509, + 0.007804206106811762, + 0.011038768105208874, + 0.0069210706278681755, + 0.009769965894520283, + 0.0128930127248168, + 0.014635520987212658, + 0.017151974141597748, + 0.01560294721275568, + 0.009420027956366539, + 0.012699134647846222, + 0.008612630888819695, + 0.014486206695437431, + 0.008087828755378723, + 0.021420950070023537, + 0.025220511481165886, + 0.019143221899867058, + 0.005049459636211395, + 0.009250858798623085, + 0.007207139860838652, + 0.010753335431218147, + 0.003499087644740939, + 0.004017144441604614, + 0.003479394596070051, + 0.004097774159163237, + 0.003635456319898367, + 0.0037779384292662144, + 0.003269337583333254, + 0.00417665857821703, + 0.017645636573433876, + 0.00986427441239357, + 0.041064172983169556, + 0.013035974465310574, + 0.011134344153106213, + 0.0101671377196908, + 0.007736514788120985, + 0.019550228491425514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.004800698719918728 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952725186944008 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952725186944008, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952725186944008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3029409945011139, + 0.044336382299661636, + 0.03673594444990158, + 0.13972973823547363, + 0.04646182060241699, + 0.03927239403128624, + 0.23496152460575104, + 0.02787383645772934, + 0.041312649846076965, + 0.08637470752000809 + ], + "credit_map": [ + 0.008092564530670643, + 0.005876481998711824, + 0.006359127350151539, + 0.06964214146137238, + 0.010289043188095093, + 0.014971268363296986, + 0.13244426250457764, + 0.055266086012125015, + 0.0036339296493679285, + 0.003775025252252817, + 0.0032391261775046587, + 0.004209144972264767, + 0.008284768089652061, + 0.0032301093451678753, + 0.011690361425280571, + 0.006273916456848383, + 0.0028577675111591816, + 0.0038499829825013876, + 0.0042773461900651455, + 0.006527423392981291, + 0.00432287622243166, + 0.004074361640959978, + 0.006448859348893166, + 0.00437732320278883, + 0.006059857085347176, + 0.009492726065218449, + 0.050155118107795715, + 0.018275056034326553, + 0.008876376785337925, + 0.02611829712986946, + 0.010046707466244698, + 0.01070559211075306, + 0.003957502078264952, + 0.00424365047365427, + 0.010203137993812561, + 0.008845400996506214, + 0.00359214935451746, + 0.005303349811583757, + 0.00585148623213172, + 0.004465143196284771, + 0.006218443159013987, + 0.0040396517142653465, + 0.0048553296364843845, + 0.004955004435032606, + 0.0043288604356348515, + 0.005492812022566795, + 0.004218250513076782, + 0.0051640416495501995, + 0.024898286908864975, + 0.03104858659207821, + 0.035491298884153366, + 0.03029921092092991, + 0.025432731956243515, + 0.027071867138147354, + 0.02272711880505085, + 0.03799242153763771, + 0.0026585350278764963, + 0.005700945388525724, + 0.006585389841347933, + 0.004343107808381319, + 0.0014013864565640688, + 0.0024304096587002277, + 0.0018761770334094763, + 0.002877884777262807, + 0.003191183554008603, + 0.005375316832214594, + 0.00572087150067091, + 0.005246876273304224, + 0.004985564388334751, + 0.006304823327809572, + 0.004389530047774315, + 0.006098479963839054, + 0.006665968801826239, + 0.012615537270903587, + 0.011272871866822243, + 0.010378609411418438, + 0.007332407403737307, + 0.011297358199954033, + 0.008788869716227055, + 0.018023082986474037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.005952725186944008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370259702205658 + ], + "opposite_logprob_tokens": [ + -3.4087026119232178 + ], + "logprob_reward": -0.03370259702205658, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370259702205658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22243142127990723, + 0.08170945197343826, + 0.08446294069290161, + 0.09300033748149872, + 0.08234865963459015, + 0.0719301849603653, + 0.10220526903867722, + 0.0383407361805439, + 0.18005605041980743, + 0.04351493716239929 + ], + "credit_map": [ + 0.009239556267857552, + 0.008122490718960762, + 0.008933122269809246, + 0.08440116047859192, + 0.008746161125600338, + 0.01661575771868229, + 0.03531670570373535, + 0.051056478172540665, + 0.005750442389398813, + 0.00666216341778636, + 0.008460392244160175, + 0.008155436255037785, + 0.014792130328714848, + 0.00590326264500618, + 0.020908119156956673, + 0.011077508330345154, + 0.007876693271100521, + 0.009238903410732746, + 0.00893707200884819, + 0.013590233400464058, + 0.011583981104195118, + 0.009219326078891754, + 0.010167527943849564, + 0.013849206268787384, + 0.006950118113309145, + 0.008833176456391811, + 0.011415233835577965, + 0.016674453392624855, + 0.009275206364691257, + 0.014250189065933228, + 0.014414661563932896, + 0.011187302879989147, + 0.010042796842753887, + 0.007518882397562265, + 0.011736160144209862, + 0.014688938856124878, + 0.0069624134339392185, + 0.012118165381252766, + 0.008523089811205864, + 0.010758213698863983, + 0.009999408386647701, + 0.008288996294140816, + 0.00837109424173832, + 0.010509288869798183, + 0.008391045965254307, + 0.01101607084274292, + 0.006937455851584673, + 0.008416824042797089, + 0.008929186500608921, + 0.01471638772636652, + 0.015474362298846245, + 0.013648628257215023, + 0.012261934578418732, + 0.011387449689209461, + 0.010834475047886372, + 0.014952843077480793, + 0.004010185599327087, + 0.006372127216309309, + 0.0064368294551968575, + 0.00811686459928751, + 0.002245101844891906, + 0.004710916895419359, + 0.0030975088011473417, + 0.0033512036316096783, + 0.015129194594919682, + 0.03084992803633213, + 0.030761422589421272, + 0.02199370600283146, + 0.023266462609171867, + 0.01750567927956581, + 0.0181986503303051, + 0.02235100418329239, + 0.004417139105498791, + 0.004284686874598265, + 0.004516669549047947, + 0.005652798339724541, + 0.005318505689501762, + 0.005656888708472252, + 0.004973024129867554, + 0.008695224300026894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.03370259702205658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306785389780998 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.04306785389780998, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306785389780998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22563311457633972, + 0.07075630128383636, + 0.042304906994104385, + 0.11193247884511948, + 0.21526749432086945, + 0.10193944722414017, + 0.01771017350256443, + 0.14092211425304413, + 0.036692045629024506, + 0.036841921508312225 + ], + "credit_map": [ + 0.008511231280863285, + 0.008151049725711346, + 0.011845572851598263, + 0.04731926694512367, + 0.01112870592623949, + 0.015386329032480717, + 0.08234678953886032, + 0.0409441776573658, + 0.004267606884241104, + 0.005824870429933071, + 0.005619496572762728, + 0.005046755541115999, + 0.019558202475309372, + 0.006455159746110439, + 0.010193385183811188, + 0.01379082165658474, + 0.002917862730100751, + 0.004202341660857201, + 0.004297920968383551, + 0.010244952514767647, + 0.004091597627848387, + 0.003950607031583786, + 0.007068736478686333, + 0.005530886352062225, + 0.008432495407760143, + 0.016125472262501717, + 0.010774760507047176, + 0.01494238805025816, + 0.011260895989835262, + 0.021390186622738838, + 0.016236329451203346, + 0.01276994775980711, + 0.016532979905605316, + 0.011979449540376663, + 0.045889969915151596, + 0.048512931913137436, + 0.014809702523052692, + 0.02573188580572605, + 0.027815692126750946, + 0.023994870483875275, + 0.014572052285075188, + 0.010612180456519127, + 0.012914491817355156, + 0.013740857131779194, + 0.009695213288068771, + 0.01723671332001686, + 0.01005287654697895, + 0.013115051202476025, + 0.0019307081820443273, + 0.002711385255679488, + 0.002773255342617631, + 0.0029681366868317127, + 0.0015964366029947996, + 0.0017648037755861878, + 0.0016390429809689522, + 0.002326403744518757, + 0.01157089602202177, + 0.02824004366993904, + 0.028867753222584724, + 0.02403605729341507, + 0.006885547656565905, + 0.014648351818323135, + 0.009884185157716274, + 0.01678927056491375, + 0.003429288510233164, + 0.0044871787540614605, + 0.004814520012587309, + 0.005447114817798138, + 0.003467586124315858, + 0.004948040936142206, + 0.003883552271872759, + 0.00621476024389267, + 0.003836951684206724, + 0.004092528019100428, + 0.0053078411146998405, + 0.00613044248893857, + 0.004458350129425526, + 0.0053349873051047325, + 0.0036624078638851643, + 0.004018411505967379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.04306785389780998 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28130272030830383 + ], + "opposite_logprob_tokens": [ + -1.4063026905059814 + ], + "logprob_reward": -0.28130272030830383, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28130272030830383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2692965269088745, + 0.10433964431285858, + 0.048972662538290024, + 0.10935063660144806, + 0.10505075007677078, + 0.11342410743236542, + 0.056512895971536636, + 0.08845794200897217, + 0.07853531837463379, + 0.026059575378894806 + ], + "credit_map": [ + 0.00960517767816782, + 0.008799267932772636, + 0.015868157148361206, + 0.10082045197486877, + 0.009816770441830158, + 0.01389274187386036, + 0.09829866141080856, + 0.012195288203656673, + 0.004875116515904665, + 0.004506058990955353, + 0.005650502163916826, + 0.007240782957524061, + 0.01714280992746353, + 0.009164099581539631, + 0.03607742115855217, + 0.019682852551341057, + 0.003930086735635996, + 0.005365324672311544, + 0.004848940297961235, + 0.007675766944885254, + 0.007450932636857033, + 0.005791058298200369, + 0.008115742355585098, + 0.005794815253466368, + 0.009213107638061047, + 0.011854122392833233, + 0.011081280186772346, + 0.009310564957559109, + 0.011776907369494438, + 0.022295961156487465, + 0.014452232979238033, + 0.01936645805835724, + 0.008350443094968796, + 0.005279968958348036, + 0.025087347254157066, + 0.019746599718928337, + 0.007601106073707342, + 0.01290205493569374, + 0.012811570428311825, + 0.013271656818687916, + 0.01964062824845314, + 0.01313453633338213, + 0.013785477727651596, + 0.0151238813996315, + 0.01325441338121891, + 0.015951121225953102, + 0.010572410188615322, + 0.011961637064814568, + 0.006963375490158796, + 0.00785553827881813, + 0.010392761789262295, + 0.007567345630377531, + 0.005512229632586241, + 0.005519920494407415, + 0.005134401377290487, + 0.007567324209958315, + 0.00992688350379467, + 0.011258863843977451, + 0.012211060151457787, + 0.02094445377588272, + 0.006532137747853994, + 0.010593794286251068, + 0.007895675487816334, + 0.009095068089663982, + 0.006957211531698704, + 0.009927746839821339, + 0.008806886151432991, + 0.011585772037506104, + 0.006992139853537083, + 0.013023141771554947, + 0.008291345089673996, + 0.012951078824698925, + 0.0036600667517632246, + 0.0019395789131522179, + 0.007814018987119198, + 0.0028853118419647217, + 0.0018802375998347998, + 0.0018903936725109816, + 0.0020620597060769796, + 0.003927908837795258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.28130272030830383 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking underwater close-up, panning downward and right in a smooth, steady rhythm, descends into the murky, green-brown waters of a river laden with suspended particles, focusing on a red-bellied piranha. Its silver-gray scales gleam with a cold metallic sheen in the faint light filtering through the surface, its belly a startling blood-red, a stark contrast to the gloomy surroundings. With a sudden, explosive snap, its powerful jaws flung open, revealing two rows of razor-sharp, tightly interlocked triangular teeth. This menacing movement churns the surrounding water, sending tiny particles of sand and mud forming a chaotic vortex in its path. As the camera pans, every subtle yet powerful twitch of its muscles beneath its gill covers and on its flanks is captured with pinpoint photorealism, vividly conveying the predator's primal, hair-trigger aggression and creating a breathtaking atmosphere of tension.", + "seed": 27714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_5087_seed_27714_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1621276183053851, + "vlm_logprob": -0.1621276183053851, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.1621276183053851, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004538592416793108 + ], + "opposite_logprob_tokens": [ + -6.379538536071777 + ], + "logprob_reward": -0.004538592416793108, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.004538592416793108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19039298593997955, + 0.37172284722328186, + 0.10124900937080383, + 0.04418279975652695, + 0.03970799967646599, + 0.0371360182762146, + 0.0750681608915329, + 0.03505893424153328, + 0.07022217661142349, + 0.03525908291339874 + ], + "credit_map": [ + 0.019306330010294914, + 0.024916432797908783, + 0.027311647310853004, + 0.02495601773262024, + 0.030172083526849747, + 0.024056842550635338, + 0.02232925407588482, + 0.01734437234699726, + 0.02043856494128704, + 0.02876030094921589, + 0.02320839650928974, + 0.09177643805742264, + 0.040809717029333115, + 0.06573153287172318, + 0.0437593013048172, + 0.057238586246967316, + 0.011116988956928253, + 0.00827194843441248, + 0.017684927210211754, + 0.010444740764796734, + 0.008942771703004837, + 0.01884731464087963, + 0.01092003844678402, + 0.015020277351140976, + 0.008593040518462658, + 0.004541680216789246, + 0.0063162087462842464, + 0.005934608168900013, + 0.004680592566728592, + 0.004599946551024914, + 0.004383789375424385, + 0.005132933147251606, + 0.0053138649091124535, + 0.004762950818985701, + 0.004619234707206488, + 0.004912346135824919, + 0.00580225046724081, + 0.003096776083111763, + 0.005468936637043953, + 0.005731637589633465, + 0.004501876421272755, + 0.003954235929995775, + 0.004667649045586586, + 0.008723203092813492, + 0.0032782009802758694, + 0.001905932673253119, + 0.005224126391112804, + 0.004880793858319521, + 0.011428226716816425, + 0.010378598235547543, + 0.013173103332519531, + 0.008853161707520485, + 0.009309662505984306, + 0.008143039420247078, + 0.005335595924407244, + 0.008446774445474148, + 0.0036911668721586466, + 0.004333938006311655, + 0.004936075769364834, + 0.004867513198405504, + 0.004281791392713785, + 0.005076746456325054, + 0.0038011139258742332, + 0.004070587456226349, + 0.007210489362478256, + 0.009355247020721436, + 0.008217831142246723, + 0.00884769856929779, + 0.010900159366428852, + 0.0055055334232747555, + 0.011824223212897778, + 0.00836099125444889, + 0.003914966247975826, + 0.002623156877234578, + 0.0036206268705427647, + 0.0044814166612923145, + 0.005533658433705568, + 0.004817392211407423, + 0.005445674993097782, + 0.004822191782295704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.004538592416793108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643371794372797 + ], + "opposite_logprob_tokens": [ + -5.379643440246582 + ], + "logprob_reward": -0.004643371794372797, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643371794372797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18570679426193237, + 0.31763574481010437, + 0.10255876928567886, + 0.06958592683076859, + 0.047978706657886505, + 0.04537228122353554, + 0.10098259150981903, + 0.04872271046042442, + 0.054319366812705994, + 0.027137065306305885 + ], + "credit_map": [ + 0.012041107751429081, + 0.01839369907975197, + 0.029403729364275932, + 0.051269590854644775, + 0.02074737474322319, + 0.01913347654044628, + 0.020932570099830627, + 0.013785247690975666, + 0.017152687534689903, + 0.0233715008944273, + 0.022713234648108482, + 0.07829934358596802, + 0.03471565619111061, + 0.0582919716835022, + 0.03677918016910553, + 0.04631216824054718, + 0.009351285174489021, + 0.010708802379667759, + 0.009828679263591766, + 0.009187080897390842, + 0.012354142963886261, + 0.02375432848930359, + 0.009941093623638153, + 0.017433354631066322, + 0.010912342928349972, + 0.006752688903361559, + 0.008702510967850685, + 0.013828271068632603, + 0.007281549274921417, + 0.0064512938261032104, + 0.007492600940167904, + 0.008164673112332821, + 0.003700400236994028, + 0.0051932078786194324, + 0.005925067700445652, + 0.009164961986243725, + 0.005934859160333872, + 0.005032732151448727, + 0.006308877374976873, + 0.006718604825437069, + 0.004380981903523207, + 0.005083772353827953, + 0.0041003143414855, + 0.005768631584942341, + 0.005622956436127424, + 0.0050977058708667755, + 0.008482574485242367, + 0.006835343781858683, + 0.011248234659433365, + 0.00896367710083723, + 0.02018011175096035, + 0.013889923691749573, + 0.013262509368360043, + 0.011659798212349415, + 0.008993548341095448, + 0.012784796766936779, + 0.0053642671555280685, + 0.005383693613111973, + 0.0061566573567688465, + 0.007010230328887701, + 0.008409889414906502, + 0.006416207645088434, + 0.004137201234698296, + 0.005844559986144304, + 0.005596249829977751, + 0.006045113783329725, + 0.006944105960428715, + 0.007476145401597023, + 0.008931904099881649, + 0.004681402817368507, + 0.007792224176228046, + 0.00685222027823329, + 0.003920732997357845, + 0.0026755311992019415, + 0.002225355478003621, + 0.002445775084197521, + 0.0038216914981603622, + 0.003647332079708576, + 0.0049818349070847034, + 0.003418811596930027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.004643371794372797 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810166195034981 + ], + "opposite_logprob_tokens": [ + -3.2881016731262207 + ], + "logprob_reward": -0.03810166195034981, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810166195034981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0739753246307373, + 0.08066587150096893, + 0.42310741543769836, + 0.048087891191244125, + 0.029614267870783806, + 0.06137794256210327, + 0.12161488085985184, + 0.0546015240252018, + 0.06847650557756424, + 0.03847840428352356 + ], + "credit_map": [ + 0.007364414632320404, + 0.01007289718836546, + 0.010943787172436714, + 0.015135440975427628, + 0.0073802475817501545, + 0.0076856487430632114, + 0.008037162944674492, + 0.00735572911798954, + 0.00548937264829874, + 0.00837546307593584, + 0.006783535238355398, + 0.022754965350031853, + 0.007996316999197006, + 0.007894373498857021, + 0.008569860830903053, + 0.01280198059976101, + 0.05908269062638283, + 0.03972272574901581, + 0.04537669196724892, + 0.03897921368479729, + 0.04097060114145279, + 0.08172224462032318, + 0.043734271079301834, + 0.07351895421743393, + 0.006090627983212471, + 0.00443026190623641, + 0.00838563870638609, + 0.006467999890446663, + 0.005988198798149824, + 0.005666918121278286, + 0.004017329309135675, + 0.007040916010737419, + 0.004274351056665182, + 0.002604737412184477, + 0.005552519578486681, + 0.00292330258525908, + 0.003558212425559759, + 0.00271442043595016, + 0.003982726950198412, + 0.0040039969608187675, + 0.006854965351521969, + 0.006699119694530964, + 0.007773569319397211, + 0.007155946921557188, + 0.007409791462123394, + 0.007658287882804871, + 0.009368646889925003, + 0.008457611314952374, + 0.01508145872503519, + 0.01287157367914915, + 0.023536715656518936, + 0.016893679276108742, + 0.017390668392181396, + 0.014346897602081299, + 0.00900179147720337, + 0.012492101639509201, + 0.005980501417070627, + 0.006201792508363724, + 0.0064271725714206696, + 0.007833057083189487, + 0.00697584543377161, + 0.008397637866437435, + 0.00612140167504549, + 0.006664114538580179, + 0.0065473648719489574, + 0.009316420182585716, + 0.0072637274861335754, + 0.009178293868899345, + 0.011040719226002693, + 0.008229796774685383, + 0.0074630738236010075, + 0.009437109343707561, + 0.004454693756997585, + 0.004547008778899908, + 0.0034815736580640078, + 0.004817445296794176, + 0.0053254603408277035, + 0.005107850302010775, + 0.005185228306800127, + 0.005559143610298634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.03810166195034981 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700254887342453 + ], + "opposite_logprob_tokens": [ + -2.6950254440307617 + ], + "logprob_reward": -0.0700254887342453, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700254887342453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25398796796798706, + 0.14328625798225403, + 0.19914411008358002, + 0.03943458944559097, + 0.04110484942793846, + 0.03900548443198204, + 0.0917065218091011, + 0.054101862013339996, + 0.09849004447460175, + 0.039738304913043976 + ], + "credit_map": [ + 0.019611520692706108, + 0.027962882071733475, + 0.03308199346065521, + 0.06043393909931183, + 0.03555889427661896, + 0.02307683229446411, + 0.030296260491013527, + 0.02396564371883869, + 0.010081424377858639, + 0.014806917868554592, + 0.009614456444978714, + 0.023548908531665802, + 0.01777656376361847, + 0.029553702101111412, + 0.018312692642211914, + 0.019591599702835083, + 0.01880216971039772, + 0.01563619263470173, + 0.034040097147226334, + 0.023996107280254364, + 0.016462191939353943, + 0.03273095563054085, + 0.027973055839538574, + 0.029503341764211655, + 0.006595345214009285, + 0.0037733495701104403, + 0.00631914334371686, + 0.0034259173553436995, + 0.004223771393299103, + 0.005815107375383377, + 0.003948017489165068, + 0.005333939101547003, + 0.006898277904838324, + 0.00445577222853899, + 0.005488344468176365, + 0.005090069025754929, + 0.003433556528761983, + 0.002631857991218567, + 0.00673662731423974, + 0.006370343267917633, + 0.0038918321952223778, + 0.006091299466788769, + 0.003417887957766652, + 0.004331980831921101, + 0.005918651353567839, + 0.004032009746879339, + 0.006024462636560202, + 0.00529736140742898, + 0.014179552905261517, + 0.014646600000560284, + 0.0110741276293993, + 0.011276321485638618, + 0.015257367864251137, + 0.008610709570348263, + 0.007985769771039486, + 0.008676068857312202, + 0.005692501086741686, + 0.006046833470463753, + 0.006439709570258856, + 0.006822350434958935, + 0.008595801889896393, + 0.008557932451367378, + 0.0054782298393547535, + 0.00646850373595953, + 0.010214284993708134, + 0.012261699885129929, + 0.011829947121441364, + 0.012617886066436768, + 0.017336281016469002, + 0.009037151001393795, + 0.01326887309551239, + 0.01192392222583294, + 0.004462044686079025, + 0.004300766158849001, + 0.002686598338186741, + 0.0031351889483630657, + 0.004449445754289627, + 0.00648553716018796, + 0.007433804217725992, + 0.006784920580685139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0700254887342453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933289766311646 + ], + "opposite_logprob_tokens": [ + -0.6933289766311646 + ], + "logprob_reward": -0.6933289766311646, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933289766311646, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2439848929643631, + 0.3151578903198242, + 0.09035541862249374, + 0.027678988873958588, + 0.02276737429201603, + 0.0437299944460392, + 0.09797544032335281, + 0.043910007923841476, + 0.07853399962186813, + 0.03590599447488785 + ], + "credit_map": [ + 0.016979245468974113, + 0.029427319765090942, + 0.034721795469522476, + 0.05895226448774338, + 0.03551970794796944, + 0.020342057570815086, + 0.028845954686403275, + 0.01919655129313469, + 0.01727432943880558, + 0.025149770081043243, + 0.023692741990089417, + 0.07486120611429214, + 0.03381604701280594, + 0.051548440009355545, + 0.03526223823428154, + 0.05355311557650566, + 0.011152631603181362, + 0.008507282473146915, + 0.01225739624351263, + 0.009353134781122208, + 0.010637804865837097, + 0.013865707442164421, + 0.009136274456977844, + 0.015445186756551266, + 0.0037923133932054043, + 0.0030586535576730967, + 0.002756812609732151, + 0.0036225863732397556, + 0.003807573579251766, + 0.0038798179011791945, + 0.00329670705832541, + 0.003464524867013097, + 0.0026382561773061752, + 0.002376714488491416, + 0.004675527103245258, + 0.0026294472627341747, + 0.0023698704317212105, + 0.0018200434278696775, + 0.0029125777073204517, + 0.0033449381589889526, + 0.0032094717025756836, + 0.0026268172077834606, + 0.003840457648038864, + 0.010443762876093388, + 0.0026534688659012318, + 0.002037458587437868, + 0.011872434988617897, + 0.007046121638268232, + 0.01159614883363247, + 0.01157927792519331, + 0.017698198556900024, + 0.013899675570428371, + 0.012480385601520538, + 0.01183635275810957, + 0.008746035397052765, + 0.01013937033712864, + 0.004050714895129204, + 0.004611640237271786, + 0.0072675240226089954, + 0.006159274373203516, + 0.004496246576309204, + 0.004678398370742798, + 0.005780194420367479, + 0.006866016425192356, + 0.008304949849843979, + 0.010642081499099731, + 0.010022695176303387, + 0.00928560271859169, + 0.012118078768253326, + 0.006444096099585295, + 0.012053106911480427, + 0.009663385339081287, + 0.005217325873672962, + 0.004516479559242725, + 0.0030795037746429443, + 0.0028849965892732143, + 0.005528719630092382, + 0.0049630142748355865, + 0.005199362523853779, + 0.004516592714935541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.6933289766311646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "striking underwater close-up, panning downward and right in a smooth, steady rhythm, descends into the murky, green-brown waters of a river laden with suspended particles, focusing on a red-bellied piranha. Its silver-gray scales gleam with a cold metallic sheen in the faint light filtering through the surface, its belly a startling blood-red, a stark contrast to the gloomy surroundings. With a sudden, explosive snap, its powerful jaws flung open, revealing two rows of razor-sharp, tightly interlocked triangular teeth. This menacing movement churns the surrounding water, sending tiny particles of sand and mud forming a chaotic vortex in its path. As the camera pans, every subtle yet powerful twitch of its muscles beneath its gill covers and on its flanks is captured with pinpoint photorealism, vividly conveying the predator's primal, hair-trigger aggression and creating a breathtaking atmosphere of tension.", + "seed": 27715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_2/index_5087_seed_27715_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1760220223106444, + "vlm_logprob": -0.1760220223106444, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.1760220223106444, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007744879927486181 + ], + "opposite_logprob_tokens": [ + -6.007744789123535 + ], + "logprob_reward": -0.007744879927486181, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007744879927486181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.431377112865448, + 0.061919741332530975, + 0.12116128206253052, + 0.0689154714345932, + 0.05263198912143707, + 0.05685395374894142, + 0.07051222771406174, + 0.03341636806726456, + 0.05909697711467743, + 0.04411488398909569 + ], + "credit_map": [ + 0.014621189795434475, + 0.009600374847650528, + 0.01758589595556259, + 0.016446352005004883, + 0.025074336677789688, + 0.17077438533306122, + 0.02450079470872879, + 0.15277378261089325, + 0.006947786081582308, + 0.005396614316850901, + 0.010774107649922371, + 0.0074358293786644936, + 0.00478941947221756, + 0.007532432675361633, + 0.009435748681426048, + 0.009607801213860512, + 0.02018277905881405, + 0.014185476116836071, + 0.02038738876581192, + 0.014903726987540722, + 0.0151376249268651, + 0.015218811109662056, + 0.01085519976913929, + 0.010290280915796757, + 0.0042920466512441635, + 0.005283166654407978, + 0.013674110174179077, + 0.007379868533462286, + 0.007526726461946964, + 0.01533321663737297, + 0.007857279852032661, + 0.007569060195237398, + 0.004747175145894289, + 0.004015544429421425, + 0.006593699567019939, + 0.00775133678689599, + 0.006819137837737799, + 0.010398547165095806, + 0.005109270103275776, + 0.007197274826467037, + 0.0046260543167591095, + 0.005081307142972946, + 0.009815569035708904, + 0.009135738015174866, + 0.007463852409273386, + 0.004550394602119923, + 0.007422413676977158, + 0.008758628740906715, + 0.005617754999548197, + 0.005868272855877876, + 0.016537390649318695, + 0.011832758784294128, + 0.00502215838059783, + 0.006557971704751253, + 0.007358686998486519, + 0.011717233806848526, + 0.003986351657658815, + 0.005147701129317284, + 0.003947212360799313, + 0.005350941326469183, + 0.0030578849837183952, + 0.004219463095068932, + 0.004105894360691309, + 0.003600918687880039, + 0.007284882012754679, + 0.005898341070860624, + 0.0055781034752726555, + 0.007166799623519182, + 0.009552031755447388, + 0.00730705913156271, + 0.008473249152302742, + 0.007836504839360714, + 0.0035903602838516235, + 0.0032418267801404, + 0.004005224909633398, + 0.00611980352550745, + 0.004469855222851038, + 0.00805988721549511, + 0.006223422009497881, + 0.008404508233070374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.007744879927486181 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046440837904810905 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.0046440837904810905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046440837904810905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3037072420120239, + 0.016561385244131088, + 0.04884026199579239, + 0.3202253580093384, + 0.033196140080690384, + 0.0509495735168457, + 0.11782623082399368, + 0.06436081230640411, + 0.032670821994543076, + 0.011662090197205544 + ], + "credit_map": [ + 0.007052665110677481, + 0.005296017974615097, + 0.007003289181739092, + 0.008619978092610836, + 0.017443373799324036, + 0.12584449350833893, + 0.015507005155086517, + 0.1169404685497284, + 0.002022468950599432, + 0.001900319242849946, + 0.00265638274140656, + 0.00267084501683712, + 0.0014007798163220286, + 0.0022216392681002617, + 0.001848929445259273, + 0.0018400232074782252, + 0.005264136474579573, + 0.005886696744710207, + 0.007468952797353268, + 0.006473227869719267, + 0.005752163473516703, + 0.005699747707694769, + 0.005810596514493227, + 0.006484747864305973, + 0.022497907280921936, + 0.016779137775301933, + 0.07170948386192322, + 0.03366219252347946, + 0.027287762612104416, + 0.07947271317243576, + 0.03259412199258804, + 0.03622206300497055, + 0.003246221225708723, + 0.0028970662970095873, + 0.003945492673665285, + 0.005076525267213583, + 0.004327438771724701, + 0.005861458368599415, + 0.003245728090405464, + 0.004596211481839418, + 0.003895670408383012, + 0.0033732035662978888, + 0.010745031759142876, + 0.00937514379620552, + 0.0048621348105371, + 0.004273691214621067, + 0.006886688061058521, + 0.007538013160228729, + 0.006691895425319672, + 0.007206019014120102, + 0.03805785998702049, + 0.02961082011461258, + 0.004961682483553886, + 0.004619188606739044, + 0.010174593888223171, + 0.016504183411598206, + 0.006865355651825666, + 0.012984517961740494, + 0.006038120947778225, + 0.008045035414397717, + 0.0071598244830966, + 0.007537121884524822, + 0.008162864483892918, + 0.007567976601421833, + 0.0024291707668453455, + 0.0024861497804522514, + 0.0019082584185525775, + 0.003615544643253088, + 0.004057110752910376, + 0.007671427447348833, + 0.005998922046273947, + 0.004504240117967129, + 0.0009797303937375546, + 0.0012802748242393136, + 0.0013620161917060614, + 0.001507901819422841, + 0.0015188887482509017, + 0.0017860857769846916, + 0.0016274322988465428, + 0.001599760726094246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0046440837904810905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494978278875351 + ], + "opposite_logprob_tokens": [ + -2.9299497604370117 + ], + "logprob_reward": -0.05494978278875351, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494978278875351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3089136481285095, + 0.07663099467754364, + 0.049481380730867386, + 0.20385469496250153, + 0.020353736355900764, + 0.048113733530044556, + 0.2027321606874466, + 0.03427061438560486, + 0.04015437141060829, + 0.015494623221457005 + ], + "credit_map": [ + 0.01072840578854084, + 0.012846033088862896, + 0.015420634299516678, + 0.014858934096992016, + 0.017915893346071243, + 0.11631391942501068, + 0.01705031841993332, + 0.10377954691648483, + 0.007245025131851435, + 0.013493409380316734, + 0.009890502318739891, + 0.014700688421726227, + 0.0074658989906311035, + 0.010069889016449451, + 0.008017519488930702, + 0.005748067982494831, + 0.007561757229268551, + 0.006070648320019245, + 0.0071253604255616665, + 0.0053123231045901775, + 0.007200969848781824, + 0.006038488820195198, + 0.005735279060900211, + 0.004436559975147247, + 0.014583857730031013, + 0.010666953399777412, + 0.046691346913576126, + 0.02124418132007122, + 0.017132554203271866, + 0.051370635628700256, + 0.019466567784547806, + 0.022698620334267616, + 0.0023582936264574528, + 0.002346014603972435, + 0.0027065526228398085, + 0.003312586108222604, + 0.0022544898092746735, + 0.001790326088666916, + 0.0022390899248421192, + 0.003346386132761836, + 0.004561241716146469, + 0.0042493510991334915, + 0.010768705978989601, + 0.008353612385690212, + 0.004626825451850891, + 0.003922431264072657, + 0.005023447796702385, + 0.006608124356716871, + 0.012533408589661121, + 0.013415772467851639, + 0.04512554779648781, + 0.04670777916908264, + 0.01836351677775383, + 0.016713494434952736, + 0.023346755653619766, + 0.02652590163052082, + 0.005308297462761402, + 0.004502858500927687, + 0.0049499208107590675, + 0.0039433762431144714, + 0.003204571781679988, + 0.004392806440591812, + 0.003809405956417322, + 0.004159378819167614, + 0.0024774917401373386, + 0.0034969609696418047, + 0.0031726763118058443, + 0.004530949983745813, + 0.004839444998651743, + 0.00971619039773941, + 0.005350201856344938, + 0.006570461206138134, + 0.0014981948770582676, + 0.0013847737573087215, + 0.0014382988447323442, + 0.0020586815662682056, + 0.0020121338311582804, + 0.002646793145686388, + 0.002070945454761386, + 0.002384803257882595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.05494978278875351 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496309697628021 + ], + "opposite_logprob_tokens": [ + -2.9299631118774414 + ], + "logprob_reward": -0.05496309697628021, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496309697628021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7200120687484741, + 0.04054403677582741, + 0.04787789285182953, + 0.08453519642353058, + 0.033131856471300125, + 0.015566802583634853, + 0.027629954740405083, + 0.012229038402438164, + 0.011021400801837444, + 0.007451733108609915 + ], + "credit_map": [ + 0.024836644530296326, + 0.017520194873213768, + 0.054020240902900696, + 0.027714526280760765, + 0.038442350924015045, + 0.2675052881240845, + 0.037627700716257095, + 0.2523452341556549, + 0.005139810033142567, + 0.006258491426706314, + 0.006607354152947664, + 0.004176116548478603, + 0.003445619950070977, + 0.00423827301710844, + 0.004558289889246225, + 0.006120087578892708, + 0.006179644260555506, + 0.006233547814190388, + 0.006936881225556135, + 0.007007046137005091, + 0.006026935763657093, + 0.005749186500906944, + 0.0050770374946296215, + 0.004667615983635187, + 0.00583005091175437, + 0.0038721899036318064, + 0.018903987482190132, + 0.009036468341946602, + 0.0072678858414292336, + 0.021721141412854195, + 0.008923404850065708, + 0.008980076760053635, + 0.002747051417827606, + 0.0025925037916749716, + 0.004260338842868805, + 0.004932255484163761, + 0.004662145394831896, + 0.006502465810626745, + 0.002856040373444557, + 0.0045790602453053, + 0.0013327402994036674, + 0.001702609471976757, + 0.002394885988906026, + 0.0025411148089915514, + 0.002441765507683158, + 0.001408067182637751, + 0.0017051949398592114, + 0.002040427178144455, + 0.0021287177223712206, + 0.002142592566087842, + 0.00580597436055541, + 0.004538948647677898, + 0.0025520934723317623, + 0.0035591789055615664, + 0.002667810069397092, + 0.004234643187373877, + 0.0014671309618279338, + 0.0015625759260728955, + 0.0016873389249667525, + 0.001612538006156683, + 0.0010828976519405842, + 0.0016563518438488245, + 0.0016343879979103804, + 0.0015258191851899028, + 0.0005785877583548427, + 0.0007113264291547239, + 0.0005920492694713175, + 0.0011780692730098963, + 0.001725862268358469, + 0.002225737553089857, + 0.001782227074727416, + 0.0022275426890701056, + 0.0006553073180839419, + 0.0005284123471938074, + 0.0006195245659910142, + 0.0007609312306158245, + 0.0006342322449199855, + 0.0017562918365001678, + 0.000980070442892611, + 0.0015169635880738497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.05496309697628021 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in murky, green-brown water with suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-bellied piranha with silver-gray scales and a blood-red belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piranha open its jaws explosively to reveal two rows of razor-sharp, triangular teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the piranha's movement churn the water to create a chaotic vortex of sand and mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan downward and right in a smooth, steady rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.757808268070221 + ], + "opposite_logprob_tokens": [ + -0.632808268070221 + ], + "logprob_reward": -0.757808268070221, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.757808268070221, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1231742650270462, + 0.025458037853240967, + 0.05230294167995453, + 0.4992923140525818, + 0.04093518853187561, + 0.07733447104692459, + 0.16277897357940674, + 0.006014679092913866, + 0.004918858874589205, + 0.007790236733853817 + ], + "credit_map": [ + 0.005222867708653212, + 0.004990224726498127, + 0.012132765725255013, + 0.00821771938353777, + 0.007141073700040579, + 0.040131885558366776, + 0.007950798608362675, + 0.03738692030310631, + 0.0031021684408187866, + 0.0038528686854988337, + 0.004084959626197815, + 0.0029026123229414225, + 0.002024282468482852, + 0.002745509147644043, + 0.0031243176199495792, + 0.0036213200073689222, + 0.007184308487921953, + 0.006265957374125719, + 0.00814739428460598, + 0.00655771791934967, + 0.005606936756521463, + 0.006715054158121347, + 0.005367218516767025, + 0.00645835418254137, + 0.03524639829993248, + 0.027615729719400406, + 0.10779852420091629, + 0.05347650498151779, + 0.04571095108985901, + 0.11696244031190872, + 0.054384104907512665, + 0.05809767171740532, + 0.004721353761851788, + 0.004570109769701958, + 0.004366228822618723, + 0.0065537625923752785, + 0.005325030535459518, + 0.0042169177904725075, + 0.00460127042606473, + 0.006580512039363384, + 0.005446350201964378, + 0.006011412013322115, + 0.014409089460968971, + 0.012461848556995392, + 0.008712748996913433, + 0.007192541845142841, + 0.010521910153329372, + 0.012578570283949375, + 0.009310324676334858, + 0.011113601736724377, + 0.053096968680620193, + 0.038067881017923355, + 0.005739492364227772, + 0.009692552499473095, + 0.012381098233163357, + 0.02337704598903656, + 0.0005809352151118219, + 0.0010899427579715848, + 0.0007965609547682106, + 0.000897513993550092, + 0.0004361086175777018, + 0.0007644279976375401, + 0.0006638056947849691, + 0.0007853838615119457, + 0.0004666267486754805, + 0.00037386416806839406, + 0.000386555417208001, + 0.0007228708127513528, + 0.0007722672889940441, + 0.0005818904028274119, + 0.0010053125442937016, + 0.0006094714626669884, + 0.000699110678397119, + 0.0006324914284050465, + 0.0007130444282665849, + 0.0010459612822160125, + 0.0007280223071575165, + 0.0011297852033749223, + 0.0012648781994357705, + 0.0015769436722621322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.757808268070221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_tensor([9597, 9597])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_tensor([9597, 9597])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..43e9fbab5c6b327dafc2a8f6f93ad68baca597a0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_tensor([9597, 9597])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, sweeping shot from left to right reveals a wide shot of an enchanting garden, rendered in the warm, detailed style of Pixar animation, with every leaf and petal perfectly formed. Ethereal tendrils of soft mist drift lazily across the frame, weaving between the stems of oversized ferns and soft-petaled flowers, catching the light as they move. The ground is completely enshrouded in a vibrant, velvety carpet of moss, its surface a rich tapestry of emerald and chartreuse that shimmers in the soft light. This lush moss extends to the edge of a tranquil pond, whose perfectly still, glassy surface perfectly reflects the soft golden light of the hidden sun and the contours of the surrounding foliage. The entire scene is bathed in a warm, diffused glow, creating an atmosphere of magical tranquility. A subtle layer of film grain has been applied to the original animation, adding a delicate, nostalgic texture and enhancing the feeling of a cherished, half-remembered dream.", + "seed": 6275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_9597_seed_6275_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3305676645599306, + "vlm_logprob": -0.3305676645599306, + "vlm_yes_no_margin": 2.5999999761581423, + "vlm_token_credit": -0.3305676645599306, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08657814562320709 + ], + "opposite_logprob_tokens": [ + -2.586578130722046 + ], + "logprob_reward": -0.08657814562320709, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08657814562320709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0571213960647583, + 0.07235698401927948, + 0.12836362421512604, + 0.0690055638551712, + 0.17303118109703064, + 0.09041138738393784, + 0.06917747855186462, + 0.05370606109499931, + 0.08870810270309448, + 0.1981182098388672 + ], + "credit_map": [ + 0.0044595166109502316, + 0.006345413625240326, + 0.00700162211433053, + 0.007686194498091936, + 0.009181768633425236, + 0.006074816454201937, + 0.009526354260742664, + 0.00684570986777544, + 0.003723691450431943, + 0.004534326493740082, + 0.009188845753669739, + 0.00849167164415121, + 0.006456353701651096, + 0.004671479109674692, + 0.02161707915365696, + 0.013673532754182816, + 0.011730239726603031, + 0.020618563517928123, + 0.012376010417938232, + 0.01857462152838707, + 0.011341704986989498, + 0.014750540256500244, + 0.022531069815158844, + 0.016440877690911293, + 0.005254434421658516, + 0.0069311088882386684, + 0.005288952961564064, + 0.009974073618650436, + 0.011383673176169395, + 0.006006197072565556, + 0.012140253558754921, + 0.012026868760585785, + 0.0125439977273345, + 0.025709154084324837, + 0.018955716863274574, + 0.013277691788971424, + 0.022524740546941757, + 0.027068423107266426, + 0.02436705492436886, + 0.028584405779838562, + 0.007063095457851887, + 0.007002354133874178, + 0.01905769109725952, + 0.016428394243121147, + 0.009660664014518261, + 0.00984448567032814, + 0.010826673358678818, + 0.010528027079999447, + 0.007196859456598759, + 0.0065330942161381245, + 0.01029956340789795, + 0.013705234974622726, + 0.007739141117781401, + 0.009191053919494152, + 0.007650140207260847, + 0.006862392649054527, + 0.006336874328553677, + 0.0049066259525716305, + 0.006781158037483692, + 0.010694028809666634, + 0.004128850530833006, + 0.005467256531119347, + 0.009277146309614182, + 0.0061141205951571465, + 0.007551481947302818, + 0.013646026141941547, + 0.016312751919031143, + 0.01373674813657999, + 0.008847293443977833, + 0.008007670752704144, + 0.013333464972674847, + 0.00727266538888216, + 0.017039360478520393, + 0.020581552758812904, + 0.043326131999492645, + 0.028775913640856743, + 0.022804105654358864, + 0.0173081923276186, + 0.028512131422758102, + 0.019770825281739235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.08657814562320709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062678813934326 + ], + "opposite_logprob_tokens": [ + -0.281267911195755 + ], + "logprob_reward": -1.4062678813934326, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062678813934326, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08411568403244019, + 0.04824821278452873, + 0.03484877943992615, + 0.1491653025150299, + 0.16371004283428192, + 0.04285121709108353, + 0.05589071288704872, + 0.12683290243148804, + 0.09391316026449203, + 0.20042406022548676 + ], + "credit_map": [ + 0.010602598078548908, + 0.008256078697741032, + 0.013332759961485863, + 0.008563018403947353, + 0.008790557272732258, + 0.008899357169866562, + 0.012409847229719162, + 0.013261455111205578, + 0.0030748711433261633, + 0.005019026342779398, + 0.0057295868173241615, + 0.007532982155680656, + 0.006753078196197748, + 0.003111728932708502, + 0.011738014407455921, + 0.005288924556225538, + 0.002932043978944421, + 0.0023294175043702126, + 0.0030724345706403255, + 0.00633062981069088, + 0.0025641575921326876, + 0.0038785163778811693, + 0.0058419606648385525, + 0.007899614050984383, + 0.012505218386650085, + 0.02194267138838768, + 0.01060603465884924, + 0.016240021213889122, + 0.031926706433296204, + 0.013634979724884033, + 0.024458544328808784, + 0.01785111054778099, + 0.011347398161888123, + 0.026740578934550285, + 0.01633533649146557, + 0.01258094422519207, + 0.01190976146608591, + 0.009617896750569344, + 0.03555859625339508, + 0.039619509130716324, + 0.004181239288300276, + 0.0053568631410598755, + 0.005453323945403099, + 0.004164544399827719, + 0.005734277423471212, + 0.005639267154037952, + 0.0071531860157847404, + 0.00516850920394063, + 0.0055958121083676815, + 0.006991358008235693, + 0.00861344113945961, + 0.010914070531725883, + 0.006367741618305445, + 0.006157652474939823, + 0.006229250226169825, + 0.005021380260586739, + 0.008779320865869522, + 0.014644854702055454, + 0.019854633137583733, + 0.01344645582139492, + 0.014417806640267372, + 0.019699493423104286, + 0.02030179090797901, + 0.015688538551330566, + 0.00693185580894351, + 0.009525328874588013, + 0.014853201806545258, + 0.012977913953363895, + 0.009974511340260506, + 0.012044147588312626, + 0.018471509218215942, + 0.009134681895375252, + 0.011376041918992996, + 0.03206655755639076, + 0.042566828429698944, + 0.035989321768283844, + 0.01696016825735569, + 0.014939247630536556, + 0.03476163372397423, + 0.011764252558350563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -1.4062678813934326 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552187778055668 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552187778055668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552187778055668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06001044437289238, + 0.1418319046497345, + 0.08166149258613586, + 0.21141286194324493, + 0.11177170276641846, + 0.0686325803399086, + 0.05778160318732262, + 0.06718800961971283, + 0.0510321743786335, + 0.1486772894859314 + ], + "credit_map": [ + 0.0061401682905852795, + 0.006928734946995974, + 0.006429092492908239, + 0.008759289979934692, + 0.009253802709281445, + 0.005945512093603611, + 0.009135673753917217, + 0.007418165914714336, + 0.009268456138670444, + 0.012382137589156628, + 0.022030070424079895, + 0.021429238840937614, + 0.012351025827229023, + 0.013731223531067371, + 0.02816448174417019, + 0.02247525565326214, + 0.006373743060976267, + 0.01124001294374466, + 0.009786441922187805, + 0.012941273860633373, + 0.008555777370929718, + 0.007868926972150803, + 0.010725586675107479, + 0.014169732108712196, + 0.017632436007261276, + 0.030091481283307076, + 0.01717686653137207, + 0.021012570708990097, + 0.054612044245004654, + 0.022714491933584213, + 0.02838798798620701, + 0.01978495717048645, + 0.00672942865639925, + 0.015126574784517288, + 0.00875304825603962, + 0.011109127663075924, + 0.00954581331461668, + 0.015703285112977028, + 0.021552659571170807, + 0.023251760751008987, + 0.005066090263426304, + 0.005514178890734911, + 0.006261407397687435, + 0.006774102803319693, + 0.008345244452357292, + 0.017513398081064224, + 0.009050504304468632, + 0.010107654146850109, + 0.006002080161124468, + 0.006251790560781956, + 0.008905109949409962, + 0.011646565049886703, + 0.006107620894908905, + 0.0073674689047038555, + 0.006504322402179241, + 0.0049966382794082165, + 0.005311564076691866, + 0.00701460987329483, + 0.00918116606771946, + 0.006657002028077841, + 0.008260332979261875, + 0.008876259438693523, + 0.013317707926034927, + 0.008569365367293358, + 0.003963267896324396, + 0.006584817543625832, + 0.00769864022731781, + 0.007330785505473614, + 0.0054220957681536674, + 0.007456051651388407, + 0.007221420295536518, + 0.005355089902877808, + 0.008588932454586029, + 0.019123660400509834, + 0.021957596763968468, + 0.02669835276901722, + 0.015950601547956467, + 0.01905939355492592, + 0.026969578117132187, + 0.01032915711402893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.012552187778055668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046377950347959995 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.0046377950347959995, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046377950347959995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14223915338516235, + 0.14939525723457336, + 0.052810899913311005, + 0.12105663120746613, + 0.1551319658756256, + 0.03877949342131615, + 0.0711032971739769, + 0.08049339056015015, + 0.052370909601449966, + 0.13661903142929077 + ], + "credit_map": [ + 0.01054572407156229, + 0.011341995559632778, + 0.01834033615887165, + 0.02335859276354313, + 0.01569048874080181, + 0.013399395160377026, + 0.02837848849594593, + 0.021184131503105164, + 0.006583670154213905, + 0.007957677356898785, + 0.012179072014987469, + 0.04284647852182388, + 0.008587788790464401, + 0.007188546005636454, + 0.052797283977270126, + 0.011254736222326756, + 0.003876553848385811, + 0.008599120192229748, + 0.0057825143449008465, + 0.007623655721545219, + 0.005040227435529232, + 0.005190375726670027, + 0.009384209290146828, + 0.007314248476177454, + 0.011984611861407757, + 0.018919888883829117, + 0.010476578027009964, + 0.012050539255142212, + 0.025520190596580505, + 0.012677519582211971, + 0.017094818875193596, + 0.0123324915766716, + 0.008258187212049961, + 0.014967799186706543, + 0.03126654401421547, + 0.014492673799395561, + 0.012458210811018944, + 0.009600451216101646, + 0.028747109696269035, + 0.03534099459648132, + 0.004167739301919937, + 0.0038035118486732244, + 0.0055145882070064545, + 0.004009678959846497, + 0.005062189418822527, + 0.004504194017499685, + 0.005885927006602287, + 0.005831661634147167, + 0.009431558661162853, + 0.008358391001820564, + 0.012903117574751377, + 0.011317376047372818, + 0.006211909465491772, + 0.008892801590263844, + 0.007447063457220793, + 0.00654107891023159, + 0.007939094677567482, + 0.009510636329650879, + 0.012751669622957706, + 0.018512146547436714, + 0.0072744498029351234, + 0.00894866045564413, + 0.008994238451123238, + 0.006562489550560713, + 0.0048319208435714245, + 0.010429182089865208, + 0.006252862047404051, + 0.00785153079777956, + 0.0063931322656571865, + 0.005672563798725605, + 0.005492596421390772, + 0.005447127390652895, + 0.0089089535176754, + 0.014909365214407444, + 0.018036609515547752, + 0.02430572360754013, + 0.017873775213956833, + 0.019889691844582558, + 0.022732479497790337, + 0.009962430223822594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0046377950347959995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280231297016144 + ], + "opposite_logprob_tokens": [ + -2.0178022384643555 + ], + "logprob_reward": -0.14280231297016144, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280231297016144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04069840908050537, + 0.02870762161910534, + 0.10155387222766876, + 0.03482982888817787, + 0.037751443684101105, + 0.04020417481660843, + 0.10765180736780167, + 0.240620419383049, + 0.15651291608810425, + 0.21146954596042633 + ], + "credit_map": [ + 0.0036754135508090258, + 0.0046562268398702145, + 0.005497555714100599, + 0.005271118599921465, + 0.0045723519288003445, + 0.004084613639861345, + 0.007690253667533398, + 0.005250873975455761, + 0.002333512529730797, + 0.002679002471268177, + 0.004722093231976032, + 0.0063486224971711636, + 0.004429392050951719, + 0.0022521691862493753, + 0.0027772211469709873, + 0.0031656038481742144, + 0.009241835214197636, + 0.014011279679834843, + 0.011599192395806313, + 0.01580946333706379, + 0.009638618677854538, + 0.010375176556408405, + 0.017875634133815765, + 0.013002662919461727, + 0.004428715445101261, + 0.0037456867285072803, + 0.0020206118933856487, + 0.0032333952840417624, + 0.005042286589741707, + 0.002879476873204112, + 0.004912888165563345, + 0.008566762320697308, + 0.003200738923624158, + 0.004507961217314005, + 0.00715629430487752, + 0.003958428278565407, + 0.004531418904662132, + 0.006128040142357349, + 0.004060519393533468, + 0.004208039492368698, + 0.003897266462445259, + 0.0053349151276052, + 0.004736937582492828, + 0.004452534019947052, + 0.00503268837928772, + 0.0055842893198132515, + 0.006731593050062656, + 0.0044339438900351524, + 0.014946797862648964, + 0.013588055036962032, + 0.019830161705613136, + 0.013849979266524315, + 0.01226201094686985, + 0.014142610132694244, + 0.00942913442850113, + 0.009603049606084824, + 0.02382093295454979, + 0.026811737567186356, + 0.03498566150665283, + 0.07604389637708664, + 0.01745706982910633, + 0.019182806834578514, + 0.023500358685851097, + 0.01881793700158596, + 0.012478397227823734, + 0.023543771356344223, + 0.0230184905230999, + 0.025447746738791466, + 0.015557815320789814, + 0.014085696078836918, + 0.02802090160548687, + 0.014360091648995876, + 0.01849135011434555, + 0.03298911452293396, + 0.0402168408036232, + 0.029306873679161072, + 0.024308588355779648, + 0.018924536183476448, + 0.031301550567150116, + 0.01593068800866604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.14280231297016144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping shot from left to right reveals a wide shot of an enchanting garden, rendered in the warm, detailed style of Pixar animation, with every leaf and petal perfectly formed. Ethereal tendrils of soft mist drift lazily across the frame, weaving between the stems of oversized ferns and soft-petaled flowers, catching the light as they move. The ground is completely enshrouded in a vibrant, velvety carpet of moss, its surface a rich tapestry of emerald and chartreuse that shimmers in the soft light. This lush moss extends to the edge of a tranquil pond, whose perfectly still, glassy surface perfectly reflects the soft golden light of the hidden sun and the contours of the surrounding foliage. The entire scene is bathed in a warm, diffused glow, creating an atmosphere of magical tranquility. A subtle layer of film grain has been applied to the original animation, adding a delicate, nostalgic texture and enhancing the feeling of a cherished, half-remembered dream.", + "seed": 6276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_9597_seed_6276_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06796756330877543, + "vlm_logprob": -0.06796756330877543, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.06796756330877543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21090194582939148 + ], + "opposite_logprob_tokens": [ + -1.7109019756317139 + ], + "logprob_reward": -0.21090194582939148, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21090194582939148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09267973899841309, + 0.0444447286427021, + 0.047209128737449646, + 0.039819370955228806, + 0.05202125385403633, + 0.05701679363846779, + 0.03633144125342369, + 0.4944041073322296, + 0.07974772900342941, + 0.05632569268345833 + ], + "credit_map": [ + 0.01088736392557621, + 0.005954261403530836, + 0.008256207220256329, + 0.006454077549278736, + 0.008706836961209774, + 0.008376096375286579, + 0.021756598725914955, + 0.022288301959633827, + 0.004951652139425278, + 0.006895496044307947, + 0.007469614036381245, + 0.004718569107353687, + 0.00552399642765522, + 0.004108612425625324, + 0.0046065314672887325, + 0.006170257460325956, + 0.0065295579843223095, + 0.006342354696244001, + 0.007088712882250547, + 0.005545626860111952, + 0.009079388342797756, + 0.0034622508101165295, + 0.0042562163434922695, + 0.00490502268075943, + 0.0031351102516055107, + 0.004716236609965563, + 0.004418038763105869, + 0.006865896750241518, + 0.003971153870224953, + 0.003795717377215624, + 0.004021457862108946, + 0.008895762264728546, + 0.006228169426321983, + 0.007319038268178701, + 0.009924877434968948, + 0.007240729406476021, + 0.0076947868801653385, + 0.004129765555262566, + 0.005248062312602997, + 0.0042358278296887875, + 0.00529664009809494, + 0.008825377561151981, + 0.0054284255020320415, + 0.0056117926724255085, + 0.015178841538727283, + 0.00655745342373848, + 0.00514935003593564, + 0.004968908615410328, + 0.0033801859244704247, + 0.009199243038892746, + 0.004024973139166832, + 0.005190781317651272, + 0.0032545183785259724, + 0.004511031787842512, + 0.0036447220481932163, + 0.0031259863171726465, + 0.02934310771524906, + 0.021916072815656662, + 0.19146515429019928, + 0.020296605303883553, + 0.021049760282039642, + 0.024956943467259407, + 0.03064042702317238, + 0.15473604202270508, + 0.012748063541948795, + 0.015019749291241169, + 0.011465935967862606, + 0.007915643975138664, + 0.010615439154207706, + 0.007696540094912052, + 0.007309356704354286, + 0.006976999808102846, + 0.006432363297790289, + 0.009999033063650131, + 0.006706991698592901, + 0.007782838307321072, + 0.006724794860929251, + 0.005570666398853064, + 0.006322501227259636, + 0.006786503829061985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.21090194582939148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903302997350693 + ], + "opposite_logprob_tokens": [ + -2.4640331268310547 + ], + "logprob_reward": -0.08903302997350693, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903302997350693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16822977364063263, + 0.15939579904079437, + 0.04609958827495575, + 0.0400545559823513, + 0.12218660116195679, + 0.04591215029358864, + 0.02412855625152588, + 0.07947523146867752, + 0.2654358446598053, + 0.04908191040158272 + ], + "credit_map": [ + 0.008735214360058308, + 0.009191900491714478, + 0.010041169822216034, + 0.007791795302182436, + 0.018327439203858376, + 0.00891376193612814, + 0.049642737954854965, + 0.05558574199676514, + 0.017463315278291702, + 0.05069241672754288, + 0.03365156427025795, + 0.008314951322972775, + 0.011282951571047306, + 0.016429755836725235, + 0.014894329011440277, + 0.0066664982587099075, + 0.004241259768605232, + 0.007566593121737242, + 0.005505985580384731, + 0.005678175017237663, + 0.007060777861624956, + 0.004202562849968672, + 0.004777251277118921, + 0.007066979072988033, + 0.0037705234717577696, + 0.006933126598596573, + 0.004084089305251837, + 0.008147994987666607, + 0.006618216168135405, + 0.00287422863766551, + 0.004244034644216299, + 0.0033823377452790737, + 0.009641759097576141, + 0.018769575282931328, + 0.012757412157952785, + 0.03153613954782486, + 0.017621662467718124, + 0.004628218710422516, + 0.017684398218989372, + 0.009547417983412743, + 0.003791203023865819, + 0.008442366495728493, + 0.006869908422231674, + 0.006638387218117714, + 0.009212004952132702, + 0.0033362500835210085, + 0.004360074643045664, + 0.003261950332671404, + 0.0021431008353829384, + 0.0038831301499158144, + 0.0035569851752370596, + 0.003336805384606123, + 0.0030807582661509514, + 0.0031440802849829197, + 0.002576174447312951, + 0.002407520543783903, + 0.0049781957641243935, + 0.006946533918380737, + 0.023704644292593002, + 0.0065702954307198524, + 0.007273636292666197, + 0.004871360957622528, + 0.007582919672131538, + 0.01754763536155224, + 0.03922497481107712, + 0.03307552635669708, + 0.036043696105480194, + 0.017797699198126793, + 0.0625263974070549, + 0.021737653762102127, + 0.03039732202887535, + 0.024632541462779045, + 0.00691836467012763, + 0.00802194606512785, + 0.00498711783438921, + 0.003150663338601589, + 0.010703147388994694, + 0.005133625585585833, + 0.005125280935317278, + 0.005041759926825762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.08903302997350693 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633917964994907 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633917964994907, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633917964994907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2031169980764389, + 0.07879702746868134, + 0.08676434308290482, + 0.04711879789829254, + 0.08342953771352768, + 0.042503535747528076, + 0.025939900428056717, + 0.2510131299495697, + 0.08141957968473434, + 0.09989716857671738 + ], + "credit_map": [ + 0.023339420557022095, + 0.01824478805065155, + 0.015857286751270294, + 0.02231978252530098, + 0.04074292629957199, + 0.012115048244595528, + 0.039196424186229706, + 0.03130132332444191, + 0.007692072540521622, + 0.010654083453118801, + 0.012900173664093018, + 0.006677920930087566, + 0.00795185100287199, + 0.0060464730486273766, + 0.011274084448814392, + 0.015600373037159443, + 0.010065957903862, + 0.01185608934611082, + 0.011859819293022156, + 0.012726468034088612, + 0.014025984331965446, + 0.008737336844205856, + 0.009802060201764107, + 0.0076906271278858185, + 0.005863520782440901, + 0.007082548923790455, + 0.004443925339728594, + 0.009096999652683735, + 0.005300307646393776, + 0.003422463545575738, + 0.0056249285116791725, + 0.006284104660153389, + 0.009124984964728355, + 0.01813441701233387, + 0.00818374939262867, + 0.01682690531015396, + 0.009162208996713161, + 0.005206321366131306, + 0.009033438749611378, + 0.0077575151808559895, + 0.003502242499962449, + 0.009709992446005344, + 0.005318115931004286, + 0.0038476113695651293, + 0.008266055956482887, + 0.004058735445141792, + 0.004081389866769314, + 0.0037193919997662306, + 0.0026025765109807253, + 0.005695538595318794, + 0.0036783814430236816, + 0.003737885970622301, + 0.0019094324670732021, + 0.0019532134756445885, + 0.003104583593085408, + 0.0032582904677838087, + 0.014865200966596603, + 0.011938336305320263, + 0.09351348876953125, + 0.009907929226756096, + 0.012599573470652103, + 0.013240501284599304, + 0.01712617836892605, + 0.07782189548015594, + 0.008786043152213097, + 0.012512615881860256, + 0.010903456248342991, + 0.009879122488200665, + 0.012866154313087463, + 0.00867000687867403, + 0.010341319255530834, + 0.00746086286380887, + 0.010024341754615307, + 0.023246459662914276, + 0.010188397020101547, + 0.006986524909734726, + 0.02108990214765072, + 0.009851755574345589, + 0.00932071078568697, + 0.009189074859023094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.007633917964994907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024857595562934875 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024857595562934875, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024857595562934875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032416682690382004, + 0.041910067200660706, + 0.009138374589383602, + 0.008654574863612652, + 0.03134983777999878, + 0.00754907913506031, + 0.007708530407398939, + 0.7884839177131653, + 0.046464014798402786, + 0.026324886828660965 + ], + "credit_map": [ + 0.003089385572820902, + 0.0029796690214425325, + 0.0034170786384493113, + 0.002071102149784565, + 0.00514113437384367, + 0.0025695483200252056, + 0.00747881131246686, + 0.005669953301548958, + 0.003671612124890089, + 0.010117915458977222, + 0.008802703581750393, + 0.0036230450496077538, + 0.003496576100587845, + 0.0035381221678107977, + 0.004096315708011389, + 0.004563778173178434, + 0.0007654564687982202, + 0.0012051654048264027, + 0.000824099755845964, + 0.0014546632301062346, + 0.0012486687628552318, + 0.0010309742065146565, + 0.0012391009368002415, + 0.0013702457072213292, + 0.0011661207536235452, + 0.0014108078321442008, + 0.0008194434340111911, + 0.0017866569105535746, + 0.0013863305794075131, + 0.000509783101733774, + 0.0008160138386301696, + 0.0007594184135086834, + 0.0019430844113230705, + 0.004958569537848234, + 0.003131116973236203, + 0.008580698631703854, + 0.004612673539668322, + 0.0012594234431162477, + 0.004730490036308765, + 0.0021337815560400486, + 0.0006498595466837287, + 0.0011199197033420205, + 0.0011795663740485907, + 0.0011107631726190448, + 0.00142732635140419, + 0.0005855599301867187, + 0.0007997153443284333, + 0.0006763690034858882, + 0.0007389262318611145, + 0.002243359573185444, + 0.0009699410293251276, + 0.0010761891026049852, + 0.0006746627041138709, + 0.0006151635898277164, + 0.0007263147854246199, + 0.0006639734492637217, + 0.04696238785982132, + 0.028401950374245644, + 0.30688902735710144, + 0.02784355729818344, + 0.036338306963443756, + 0.041228391230106354, + 0.04993968456983566, + 0.2508805990219116, + 0.0042151291854679585, + 0.006094183307141066, + 0.004026911687105894, + 0.004242132417857647, + 0.011393800377845764, + 0.004077627789229155, + 0.006173061206936836, + 0.006241166964173317, + 0.003230978734791279, + 0.003658160800114274, + 0.003530144225805998, + 0.0020734460558742285, + 0.005683243274688721, + 0.0024530147202312946, + 0.0023957807570695877, + 0.003300120122730732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.0024857595562934875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783163219690323 + ], + "opposite_logprob_tokens": [ + -3.529783248901367 + ], + "logprob_reward": -0.029783163219690323, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783163219690323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07179681956768036, + 0.0706191286444664, + 0.03772619739174843, + 0.04312397539615631, + 0.1489267796278, + 0.050611503422260284, + 0.028031393885612488, + 0.3500138223171234, + 0.09739644825458527, + 0.10175388306379318 + ], + "credit_map": [ + 0.006811014376580715, + 0.009182685054838657, + 0.006802531890571117, + 0.008230946026742458, + 0.015818368643522263, + 0.005078207701444626, + 0.010495319031178951, + 0.009377756156027317, + 0.008047694340348244, + 0.022745750844478607, + 0.01490020751953125, + 0.003970860503613949, + 0.004933205433189869, + 0.0071906703524291515, + 0.006177205592393875, + 0.0026535294018685818, + 0.00472182547673583, + 0.0075098928064107895, + 0.002810077741742134, + 0.003973409067839384, + 0.008541535586118698, + 0.0036159995943307877, + 0.0033602940384298563, + 0.0031931621488183737, + 0.004966742359101772, + 0.007033599074929953, + 0.0037197826895862818, + 0.009642867371439934, + 0.006372859235852957, + 0.0028763022273778915, + 0.004886413924396038, + 0.0036254103761166334, + 0.009429650381207466, + 0.02538173831999302, + 0.014784720726311207, + 0.04058755189180374, + 0.023693740367889404, + 0.006197931244969368, + 0.02010066993534565, + 0.008750765584409237, + 0.005709103308618069, + 0.009336783550679684, + 0.008931633085012436, + 0.007106638513505459, + 0.007783601991832256, + 0.0029455111362040043, + 0.004732655826956034, + 0.004065581131726503, + 0.003377610817551613, + 0.008831844665110111, + 0.003237004391849041, + 0.0041157277300953865, + 0.0023364939261227846, + 0.0018595445435494184, + 0.0023853471502661705, + 0.0018878187984228134, + 0.02069265954196453, + 0.017416566610336304, + 0.1334564983844757, + 0.013819235377013683, + 0.01615280471742153, + 0.017615247517824173, + 0.02166665345430374, + 0.10919416695833206, + 0.0188134778290987, + 0.014683626592159271, + 0.01402646116912365, + 0.011950531974434853, + 0.012885638512670994, + 0.007674498483538628, + 0.008122626692056656, + 0.00923958420753479, + 0.014562523923814297, + 0.01574469357728958, + 0.016707075759768486, + 0.011472712270915508, + 0.013651147484779358, + 0.00857127271592617, + 0.008001391775906086, + 0.01304306648671627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.029783163219690323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping shot from left to right reveals a wide shot of an enchanting garden, rendered in the warm, detailed style of Pixar animation, with every leaf and petal perfectly formed. Ethereal tendrils of soft mist drift lazily across the frame, weaving between the stems of oversized ferns and soft-petaled flowers, catching the light as they move. The ground is completely enshrouded in a vibrant, velvety carpet of moss, its surface a rich tapestry of emerald and chartreuse that shimmers in the soft light. This lush moss extends to the edge of a tranquil pond, whose perfectly still, glassy surface perfectly reflects the soft golden light of the hidden sun and the contours of the surrounding foliage. The entire scene is bathed in a warm, diffused glow, creating an atmosphere of magical tranquility. A subtle layer of film grain has been applied to the original animation, adding a delicate, nostalgic texture and enhancing the feeling of a cherished, half-remembered dream.", + "seed": 6277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_9597_seed_6277_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21427364759147166, + "vlm_logprob": -0.21427364759147166, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.21427364759147166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23750253021717072 + ], + "opposite_logprob_tokens": [ + -1.6125025749206543 + ], + "logprob_reward": -0.23750253021717072, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23750253021717072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09120308607816696, + 0.04900384321808815, + 0.18001152575016022, + 0.09359055757522583, + 0.09200523048639297, + 0.09146974235773087, + 0.09490848332643509, + 0.07817311584949493, + 0.09565553814172745, + 0.13397881388664246 + ], + "credit_map": [ + 0.00756271556019783, + 0.01775660552084446, + 0.009039412252604961, + 0.015877898782491684, + 0.010358466766774654, + 0.009175659157335758, + 0.012416246347129345, + 0.009016087278723717, + 0.005138704553246498, + 0.00393748888745904, + 0.005812776740640402, + 0.005993287544697523, + 0.006950858514755964, + 0.005532761570066214, + 0.008351859636604786, + 0.007286105304956436, + 0.03395724296569824, + 0.01419518981128931, + 0.020147455856204033, + 0.022179681807756424, + 0.024930888786911964, + 0.022369636222720146, + 0.0231320858001709, + 0.019099341705441475, + 0.006537206936627626, + 0.025278059765696526, + 0.0265923161059618, + 0.0060843657702207565, + 0.006723205558955669, + 0.0064231776632368565, + 0.00820529367774725, + 0.0077469227835536, + 0.008863434195518494, + 0.010186522267758846, + 0.01336812973022461, + 0.007110835053026676, + 0.016158420592546463, + 0.01589115895330906, + 0.010081377811729908, + 0.01034535001963377, + 0.009793839417397976, + 0.010327747091650963, + 0.011013164184987545, + 0.00934253353625536, + 0.014384530484676361, + 0.014409319497644901, + 0.010881897062063217, + 0.011316711083054543, + 0.008077497594058514, + 0.010914942249655724, + 0.013555489480495453, + 0.00804928969591856, + 0.01766500622034073, + 0.012683859094977379, + 0.01039460115134716, + 0.013567802496254444, + 0.009854333475232124, + 0.012476989068090916, + 0.010875929147005081, + 0.010937870480120182, + 0.011353796347975731, + 0.010474334470927715, + 0.005440038163214922, + 0.00675982516258955, + 0.00990068819373846, + 0.012886648066341877, + 0.018226662650704384, + 0.00962879043072462, + 0.018805643543601036, + 0.010786916129291058, + 0.007402938324958086, + 0.008017245680093765, + 0.013269513845443726, + 0.017694829031825066, + 0.010640581138432026, + 0.011420045979321003, + 0.031004561111330986, + 0.02695409394800663, + 0.011704826727509499, + 0.011290359310805798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.23750253021717072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576151967048645 + ], + "opposite_logprob_tokens": [ + -0.826151967048645 + ], + "logprob_reward": -0.576151967048645, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576151967048645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14114896953105927, + 0.09344401955604553, + 0.0545797199010849, + 0.08397768437862396, + 0.0536758117377758, + 0.02674379199743271, + 0.06653058528900146, + 0.1341572105884552, + 0.06610570847988129, + 0.2796365022659302 + ], + "credit_map": [ + 0.011903074570000172, + 0.02254904992878437, + 0.02184978872537613, + 0.020149091258645058, + 0.014582359232008457, + 0.019949469715356827, + 0.015464185737073421, + 0.014701939187943935, + 0.010117852129042149, + 0.007671746890991926, + 0.012154380790889263, + 0.011749695986509323, + 0.010595367290079594, + 0.013463158160448074, + 0.014268448576331139, + 0.01342336367815733, + 0.009422270581126213, + 0.004392120521515608, + 0.007038549054414034, + 0.004467763938009739, + 0.01100180484354496, + 0.00871066004037857, + 0.005822700448334217, + 0.0037238462828099728, + 0.008702918887138367, + 0.023409923538565636, + 0.01536855660378933, + 0.005842174403369427, + 0.009821872226893902, + 0.00793901551514864, + 0.00781633984297514, + 0.0050768726505339146, + 0.005866461433470249, + 0.006437028758227825, + 0.00876096822321415, + 0.004300667904317379, + 0.01244458556175232, + 0.008191595785319805, + 0.003934063948690891, + 0.0037404343020170927, + 0.002684619976207614, + 0.0024399743415415287, + 0.0033975266851484776, + 0.0031991174910217524, + 0.00396169675514102, + 0.003868497209623456, + 0.0037959858309477568, + 0.003396372776478529, + 0.0061081224121153355, + 0.008112812414765358, + 0.0058350032195448875, + 0.007495346944779158, + 0.010697656311094761, + 0.00947036873549223, + 0.009306344203650951, + 0.009504921734333038, + 0.027784006670117378, + 0.013588803820312023, + 0.00979074276983738, + 0.006397100631147623, + 0.037999823689460754, + 0.01681239902973175, + 0.008526318706572056, + 0.013257998041808605, + 0.007393320556730032, + 0.004856073297560215, + 0.006613538134843111, + 0.007818496786057949, + 0.015121538192033768, + 0.010284927673637867, + 0.006933688651770353, + 0.007084115874022245, + 0.021983977407217026, + 0.026044681668281555, + 0.0171335618942976, + 0.023897133767604828, + 0.09838338196277618, + 0.051818426698446274, + 0.01678260788321495, + 0.02359270118176937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.576151967048645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358038187026978 + ], + "opposite_logprob_tokens": [ + -3.651358127593994 + ], + "logprob_reward": -0.026358038187026978, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358038187026978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16247831284999847, + 0.07083823531866074, + 0.04496396705508232, + 0.13997994363307953, + 0.046506576240062714, + 0.02637479268014431, + 0.038959525525569916, + 0.12225237488746643, + 0.13654941320419312, + 0.2110968679189682 + ], + "credit_map": [ + 0.01440555788576603, + 0.02181369811296463, + 0.015689270570874214, + 0.03907231241464615, + 0.01523477490991354, + 0.012580524198710918, + 0.02394053339958191, + 0.019741641357541084, + 0.007147427648305893, + 0.006115634460002184, + 0.009224935434758663, + 0.011694719083607197, + 0.006060665473341942, + 0.01037935446947813, + 0.011392188258469105, + 0.00882331095635891, + 0.004649774637073278, + 0.004936667624861002, + 0.006075864192098379, + 0.005276502575725317, + 0.006498144473880529, + 0.005317425820976496, + 0.007411660626530647, + 0.004797923844307661, + 0.012185873463749886, + 0.018509069457650185, + 0.025912173092365265, + 0.015889348462224007, + 0.019309286028146744, + 0.01721629500389099, + 0.01700669154524803, + 0.01395120657980442, + 0.00469610933214426, + 0.006508787628263235, + 0.006350444629788399, + 0.003002665936946869, + 0.009296148084104061, + 0.007456326391547918, + 0.004851889330893755, + 0.004344204906374216, + 0.0037154408637434244, + 0.0029839524067938328, + 0.003320855787023902, + 0.0029981245752424, + 0.003935887943953276, + 0.003898544469848275, + 0.0023398729972541332, + 0.0031821143347769976, + 0.0040424298495054245, + 0.005001354962587357, + 0.0052725947462022305, + 0.0032213255763053894, + 0.006467520724982023, + 0.006298081483691931, + 0.002963767619803548, + 0.005692448932677507, + 0.015795832499861717, + 0.015076516196131706, + 0.011528832837939262, + 0.012842627242207527, + 0.02089545503258705, + 0.017878828570246696, + 0.01347391027957201, + 0.014760371297597885, + 0.012001545168459415, + 0.013472161255776882, + 0.021283570677042007, + 0.010952528566122055, + 0.027158619835972786, + 0.017846781760454178, + 0.016514217481017113, + 0.017319995909929276, + 0.01592591591179371, + 0.021551596000790596, + 0.013185064308345318, + 0.02074023336172104, + 0.0656852051615715, + 0.03955400735139847, + 0.013729606755077839, + 0.02072526142001152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.026358038187026978 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1215142011642456, + 0.01945321075618267, + 0.3019475042819977, + 0.022194553166627884, + 0.025954032316803932, + 0.024824533611536026, + 0.03594490885734558, + 0.07870045304298401, + 0.06910877674818039, + 0.300357848405838 + ], + "credit_map": [ + 0.010629736818373203, + 0.017070870846509933, + 0.012352286837995052, + 0.03187040239572525, + 0.014023168943822384, + 0.009745107963681221, + 0.011486377567052841, + 0.014336250722408295, + 0.002432487439364195, + 0.0014902678085491061, + 0.0024395303335040808, + 0.002290882868692279, + 0.0017192537197843194, + 0.0029314246494323015, + 0.003495881799608469, + 0.0026534826029092073, + 0.054842688143253326, + 0.023846358060836792, + 0.034394629299640656, + 0.030872423201799393, + 0.054994869977235794, + 0.041485123336315155, + 0.03461623936891556, + 0.026895184069871902, + 0.0023572221398353577, + 0.0025024202186614275, + 0.0028562061488628387, + 0.002609741175547242, + 0.0025849384255707264, + 0.002459127688780427, + 0.0036982258316129446, + 0.0031266696751117706, + 0.0028261400293558836, + 0.0031102753710001707, + 0.004014838021248579, + 0.002516950713470578, + 0.004307224880903959, + 0.004014801699668169, + 0.0028774517122656107, + 0.0022863498888909817, + 0.002430290449410677, + 0.002431959379464388, + 0.002497185254469514, + 0.0030320819932967424, + 0.0036107145715504885, + 0.0034890668466687202, + 0.003997369669377804, + 0.0033358661457896233, + 0.0036587449721992016, + 0.0033868474420160055, + 0.004671451169997454, + 0.0025575142353773117, + 0.004803376272320747, + 0.005056943278759718, + 0.004460061900317669, + 0.007349971681833267, + 0.01547893788665533, + 0.008717143908143044, + 0.007205459289252758, + 0.005846383515745401, + 0.01652967557311058, + 0.009236915968358517, + 0.00671025924384594, + 0.008975685574114323, + 0.0071768732741475105, + 0.005531439557671547, + 0.008433923125267029, + 0.005803925916552544, + 0.0221721101552248, + 0.00971123669296503, + 0.005815670359879732, + 0.004463599994778633, + 0.02329205721616745, + 0.03142617270350456, + 0.019676946103572845, + 0.022638974711298943, + 0.09899559617042542, + 0.05791259929537773, + 0.018333565443754196, + 0.028081953525543213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.02980329655110836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015524059534073 + ], + "opposite_logprob_tokens": [ + -1.701552391052246 + ], + "logprob_reward": -0.2015524059534073, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015524059534073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06632108241319656, + 0.023292869329452515, + 0.1342604011297226, + 0.06327680498361588, + 0.024707457050681114, + 0.038102418184280396, + 0.0652969628572464, + 0.18937255442142487, + 0.12357447296380997, + 0.2717949151992798 + ], + "credit_map": [ + 0.006980997044593096, + 0.011560983024537563, + 0.00653016148135066, + 0.008489154279232025, + 0.006260869558900595, + 0.009454356506466866, + 0.008879903703927994, + 0.008164661005139351, + 0.0014870095765218139, + 0.001988466363400221, + 0.0034990068525075912, + 0.004887663293629885, + 0.002053968608379364, + 0.0020511006005108356, + 0.0037283594720065594, + 0.0035972967743873596, + 0.023012178018689156, + 0.011069325730204582, + 0.01474302913993597, + 0.01531431544572115, + 0.025114063173532486, + 0.01799662970006466, + 0.015290767885744572, + 0.011720110662281513, + 0.004733725916594267, + 0.017315872013568878, + 0.013499435037374496, + 0.004680708982050419, + 0.006357223726809025, + 0.005407385993748903, + 0.006941959727555513, + 0.004340501967817545, + 0.00235383422113955, + 0.0026619520504027605, + 0.0037884986959397793, + 0.003336425870656967, + 0.002391841495409608, + 0.003191958647221327, + 0.003732082201167941, + 0.003250865964218974, + 0.003912326879799366, + 0.0029480699449777603, + 0.004766830708831549, + 0.0040411390364170074, + 0.005328522529453039, + 0.005783730186522007, + 0.005695483181625605, + 0.005626322701573372, + 0.006077147088944912, + 0.0041791233234107494, + 0.005379417911171913, + 0.005397069733589888, + 0.00740288058295846, + 0.015474786050617695, + 0.009404194541275501, + 0.011982345953583717, + 0.044737931340932846, + 0.015534581616520882, + 0.015878954902291298, + 0.015593435615301132, + 0.04970031976699829, + 0.021053681150078773, + 0.01171006541699171, + 0.015163580887019634, + 0.01729559153318405, + 0.0112740034237504, + 0.009644596837460995, + 0.008259305730462074, + 0.043286632746458054, + 0.01879146508872509, + 0.008037892170250416, + 0.006984994746744633, + 0.021551908925175667, + 0.029999420046806335, + 0.020284932106733322, + 0.016754034906625748, + 0.09059379994869232, + 0.05268155038356781, + 0.01652318425476551, + 0.02340609021484852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.2015524059534073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping shot from left to right reveals a wide shot of an enchanting garden, rendered in the warm, detailed style of Pixar animation, with every leaf and petal perfectly formed. Ethereal tendrils of soft mist drift lazily across the frame, weaving between the stems of oversized ferns and soft-petaled flowers, catching the light as they move. The ground is completely enshrouded in a vibrant, velvety carpet of moss, its surface a rich tapestry of emerald and chartreuse that shimmers in the soft light. This lush moss extends to the edge of a tranquil pond, whose perfectly still, glassy surface perfectly reflects the soft golden light of the hidden sun and the contours of the surrounding foliage. The entire scene is bathed in a warm, diffused glow, creating an atmosphere of magical tranquility. A subtle layer of film grain has been applied to the original animation, adding a delicate, nostalgic texture and enhancing the feeling of a cherished, half-remembered dream.", + "seed": 6278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_3/index_9597_seed_6278_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08244628580287099, + "vlm_logprob": -0.08244628580287099, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.08244628580287099, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08489162474870682 + ], + "opposite_logprob_tokens": [ + -2.5848915576934814 + ], + "logprob_reward": -0.08489162474870682, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08489162474870682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.595679521560669, + 0.014513423666357994, + 0.05341288819909096, + 0.021846819669008255, + 0.04777853563427925, + 0.035256411880254745, + 0.09912862628698349, + 0.046968553215265274, + 0.05450534075498581, + 0.030909907072782516 + ], + "credit_map": [ + 0.028726987540721893, + 0.025376474484801292, + 0.01452383492141962, + 0.015361630357801914, + 0.020735209807753563, + 0.016067907214164734, + 0.24663710594177246, + 0.22825036942958832, + 0.0015920239966362715, + 0.002520486479625106, + 0.0015991788823157549, + 0.0023310366086661816, + 0.0013839284656569362, + 0.0012347898446023464, + 0.0014191237278282642, + 0.002432856010273099, + 0.004688422195613384, + 0.005214018747210503, + 0.003428205382078886, + 0.004805432632565498, + 0.004004259128123522, + 0.005317597184330225, + 0.014355748891830444, + 0.011599206365644932, + 0.0026396506000310183, + 0.0026188192423433065, + 0.002433724934235215, + 0.0033480555284768343, + 0.0030440480913966894, + 0.0024576482828706503, + 0.0027319537475705147, + 0.0025729199405759573, + 0.004050174728035927, + 0.006397637072950602, + 0.0030391905456781387, + 0.007711529266089201, + 0.009663796052336693, + 0.005443572998046875, + 0.005765869747847319, + 0.005706764757633209, + 0.0030101805459707975, + 0.005327755119651556, + 0.0036350805312395096, + 0.0051076412200927734, + 0.0050186337903141975, + 0.006320801097899675, + 0.0035102020483464003, + 0.0033261182252317667, + 0.008179813623428345, + 0.012384597212076187, + 0.00828489102423191, + 0.013537969440221786, + 0.010114711709320545, + 0.011377186514437199, + 0.028005672618746758, + 0.007243777625262737, + 0.004690872505307198, + 0.0063749016262590885, + 0.007703815586864948, + 0.006193101406097412, + 0.005438182037323713, + 0.0067815217189490795, + 0.0055268192663788795, + 0.004259340465068817, + 0.007026380393654108, + 0.007691349368542433, + 0.005470576696097851, + 0.005416210275143385, + 0.009607180953025818, + 0.0069506908766925335, + 0.005140434484928846, + 0.007202513981610537, + 0.005257942248135805, + 0.005722986068576574, + 0.004135812167078257, + 0.0032256427221000195, + 0.002634241944178939, + 0.002410379471257329, + 0.003961792681366205, + 0.0035611079074442387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.08489162474870682 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154762268066406 + ], + "opposite_logprob_tokens": [ + -1.701547622680664 + ], + "logprob_reward": -0.20154762268066406, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154762268066406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8070166110992432, + 0.010884183458983898, + 0.01673084869980812, + 0.022021140903234482, + 0.03301653265953064, + 0.03210858628153801, + 0.030579544603824615, + 0.01125367172062397, + 0.008566603064537048, + 0.027822325006127357 + ], + "credit_map": [ + 0.03675462305545807, + 0.030572380870580673, + 0.019776489585638046, + 0.02203499525785446, + 0.024266500025987625, + 0.021891731768846512, + 0.3379886746406555, + 0.3137311637401581, + 0.0017322278581559658, + 0.001378262764774263, + 0.001438126782886684, + 0.0016139920335263014, + 0.0008456089999526739, + 0.0009363022400066257, + 0.0018004972953349352, + 0.0011391656007617712, + 0.002436423674225807, + 0.0020296538714319468, + 0.0009473446407355368, + 0.001389387296512723, + 0.0024839986581355333, + 0.002239991445094347, + 0.0029264388140290976, + 0.0022776101250201464, + 0.0023944047279655933, + 0.0032405764795839787, + 0.0019281426211819053, + 0.0028365987818688154, + 0.003357473760843277, + 0.002253743587061763, + 0.0024659635964781046, + 0.003544238628819585, + 0.0028931882698088884, + 0.00395631417632103, + 0.0021777828224003315, + 0.006304504815489054, + 0.0044284663163125515, + 0.0046297712251544, + 0.004027637653052807, + 0.004598869010806084, + 0.0038179822731763124, + 0.004966137930750847, + 0.003015032270923257, + 0.006058921571820974, + 0.003773269010707736, + 0.0022327338811010122, + 0.004684120416641235, + 0.003560387995094061, + 0.0022999108768999577, + 0.003512582741677761, + 0.0013457261957228184, + 0.005947605241090059, + 0.0023655034601688385, + 0.002176532754674554, + 0.010640313848853111, + 0.0022913720458745956, + 0.001415603095665574, + 0.0012060458539053798, + 0.0011321069905534387, + 0.0013088186969980597, + 0.0020118453539907932, + 0.0017377479234710336, + 0.0011979335686191916, + 0.0012435702374204993, + 0.0008856389904394746, + 0.001432317541912198, + 0.0007423229981213808, + 0.0009342905832454562, + 0.0012021793518215418, + 0.0010431766277179122, + 0.0011695567518472672, + 0.0011571202194318175, + 0.003757441183552146, + 0.003789599286392331, + 0.0026872800663113594, + 0.0026843096129596233, + 0.00314657692797482, + 0.004047940485179424, + 0.0036571489181369543, + 0.004052028991281986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.20154762268066406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633917964994907 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633917964994907, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633917964994907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39512181282043457, + 0.04060488939285278, + 0.09509793668985367, + 0.02222953550517559, + 0.060652054846286774, + 0.09345399588346481, + 0.14217215776443481, + 0.0282637607306242, + 0.03170524537563324, + 0.09069865196943283 + ], + "credit_map": [ + 0.026279065757989883, + 0.025153474882245064, + 0.01671256124973297, + 0.01523340493440628, + 0.016329314559698105, + 0.01587122119963169, + 0.14477908611297607, + 0.13476364314556122, + 0.006114454008638859, + 0.006886620074510574, + 0.005436580162495375, + 0.004607534967362881, + 0.0037181859370321035, + 0.003577850293368101, + 0.004189441446214914, + 0.006074217613786459, + 0.006899380590766668, + 0.006421910133212805, + 0.004767133388668299, + 0.006857807748019695, + 0.005316900555044413, + 0.007092122919857502, + 0.03062029369175434, + 0.0271223783493042, + 0.0026941304095089436, + 0.0027645595837384462, + 0.002991087967529893, + 0.002848153468221426, + 0.003093958832323551, + 0.002441192278638482, + 0.0024300571531057358, + 0.002966394415125251, + 0.007690909318625927, + 0.007009095977991819, + 0.005409961566329002, + 0.011789631098508835, + 0.005804114043712616, + 0.006444857455790043, + 0.006199523340910673, + 0.010303962975740433, + 0.013661479577422142, + 0.011390653438866138, + 0.00960657000541687, + 0.018481392413377762, + 0.0098073435947299, + 0.006267703603953123, + 0.012872331775724888, + 0.011366516351699829, + 0.011828096583485603, + 0.016463514417409897, + 0.0066527980379760265, + 0.027406888082623482, + 0.010891689918935299, + 0.010801480151712894, + 0.04818972945213318, + 0.009937946684658527, + 0.0038613213691860437, + 0.003682187059894204, + 0.002682384802028537, + 0.0040252115577459335, + 0.004507506266236305, + 0.0036343028768897057, + 0.003191856201738119, + 0.0026789891999214888, + 0.0038989828899502754, + 0.0041086068376898766, + 0.002956725424155593, + 0.003323783865198493, + 0.0052002654410898685, + 0.0042719910852611065, + 0.0033972891978919506, + 0.004547597840428352, + 0.012722749263048172, + 0.014481955207884312, + 0.007968195714056492, + 0.009258837439119816, + 0.009946960024535656, + 0.012943465262651443, + 0.012137612327933311, + 0.011238868348300457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.007633917964994907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052477470599114895 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052477470599114895, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052477470599114895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49730539321899414, + 0.03904625028371811, + 0.09245477616786957, + 0.03167508542537689, + 0.05939830467104912, + 0.053236253559589386, + 0.08305832743644714, + 0.053603462874889374, + 0.025185750797390938, + 0.06503640860319138 + ], + "credit_map": [ + 0.023196451365947723, + 0.022282417863607407, + 0.012533357366919518, + 0.01387467049062252, + 0.02097373642027378, + 0.015422657132148743, + 0.19947893917560577, + 0.18954312801361084, + 0.005262015853077173, + 0.0044446527026593685, + 0.004606224596500397, + 0.005242110230028629, + 0.004554669372737408, + 0.0038286473136395216, + 0.006110277026891708, + 0.004997648298740387, + 0.01383283082395792, + 0.010233603417873383, + 0.008768107742071152, + 0.009127961471676826, + 0.01199436467140913, + 0.012562515214085579, + 0.010051516816020012, + 0.015883872285485268, + 0.0035105186980217695, + 0.004466948565095663, + 0.0026606100145727396, + 0.003974092658609152, + 0.004427176900207996, + 0.0036988877691328526, + 0.005316967610269785, + 0.003619879251345992, + 0.0043028779327869415, + 0.006584770046174526, + 0.00348506192676723, + 0.010670334100723267, + 0.010392390191555023, + 0.010265951976180077, + 0.006367553025484085, + 0.0073293596506118774, + 0.005452936980873346, + 0.006674991454929113, + 0.0050421529449522495, + 0.010361110791563988, + 0.006183105520904064, + 0.005725843366235495, + 0.007318316493183374, + 0.006477788090705872, + 0.006677687168121338, + 0.010447647422552109, + 0.005234872456640005, + 0.012656054459512234, + 0.008021880872547626, + 0.008438683114945889, + 0.02533140778541565, + 0.006250085774809122, + 0.0058828997425735, + 0.006384111940860748, + 0.008066263049840927, + 0.006111692171543837, + 0.008107796311378479, + 0.007626052014529705, + 0.006284966599196196, + 0.005139677785336971, + 0.002839296590536833, + 0.002276209183037281, + 0.002395658753812313, + 0.0024682837538421154, + 0.004100697580724955, + 0.00406977254897356, + 0.003555201692506671, + 0.003480627667158842, + 0.008748956955969334, + 0.009363478049635887, + 0.006205946207046509, + 0.00649869441986084, + 0.007572139613330364, + 0.009485987015068531, + 0.008943448774516582, + 0.00821774359792471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0052477470599114895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, sweeping shot from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in the warm, detailed style of Pixar animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ethereal tendrils of soft mist drifting across the frame among oversized ferns and soft-petaled flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered in a vibrant, velvety carpet of moss with a rich tapestry of emerald and chartreuse colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a tranquil pond with a perfectly still, glassy surface reflecting the soft golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291051656007767 + ], + "opposite_logprob_tokens": [ + -2.237910509109497 + ], + "logprob_reward": -0.11291051656007767, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291051656007767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3529927134513855, + 0.037510741502046585, + 0.1847715526819229, + 0.07999954372644424, + 0.05125442147254944, + 0.05906334146857262, + 0.03343135118484497, + 0.0197177454829216, + 0.014754008501768112, + 0.16650457680225372 + ], + "credit_map": [ + 0.01603664830327034, + 0.02425088733434677, + 0.012886331416666508, + 0.012480069883167744, + 0.02976718731224537, + 0.013790560886263847, + 0.12956106662750244, + 0.11421997100114822, + 0.004452291410416365, + 0.008034038357436657, + 0.0026184830348938704, + 0.0037659539375454187, + 0.008158121258020401, + 0.0026229512877762318, + 0.003810452064499259, + 0.004048449452966452, + 0.023205749690532684, + 0.023878930136561394, + 0.016458680853247643, + 0.021042972803115845, + 0.023976393043994904, + 0.021740371361374855, + 0.018433615565299988, + 0.03603481873869896, + 0.00988369807600975, + 0.015686919912695885, + 0.004645878914743662, + 0.007323634810745716, + 0.01876106671988964, + 0.010667023248970509, + 0.006994748022407293, + 0.006036573555320501, + 0.004107893910259008, + 0.006038561463356018, + 0.0025945233646780252, + 0.007844535633921623, + 0.009283975698053837, + 0.00668962812051177, + 0.00750005804002285, + 0.007195241749286652, + 0.005840735509991646, + 0.009325429797172546, + 0.00411591213196516, + 0.014124208129942417, + 0.006586292292922735, + 0.004702206235378981, + 0.008450878784060478, + 0.005917674396187067, + 0.0031805355101823807, + 0.004047816153615713, + 0.003322325414046645, + 0.005602586083114147, + 0.004805785603821278, + 0.0051696812734007835, + 0.004645349457859993, + 0.002657270058989525, + 0.0023171177599579096, + 0.0027281770016998053, + 0.0022814676631242037, + 0.0033106673508882523, + 0.0026037406641989946, + 0.0026105858851224184, + 0.0020259881857782602, + 0.001839998411014676, + 0.001475044758990407, + 0.0012264250544831157, + 0.001726700458675623, + 0.0015376535011455417, + 0.0020943062845617533, + 0.0028408567886799574, + 0.0018221616046503186, + 0.0020308583043515682, + 0.01916692778468132, + 0.02067769318819046, + 0.018203457817435265, + 0.01812492124736309, + 0.02028789557516575, + 0.022671688348054886, + 0.02720264345407486, + 0.02016933262348175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.11291051656007767 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_4/index_tensor([3013, 3013])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_4/index_tensor([3013, 3013])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c0345bd8a8a98aa26f1c0adbaab5dcb1ace847b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_4/index_tensor([3013, 3013])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, bird's-eye view captures hyper-realistic CGI time-lapse footage of a massive Victoria Amazon water lily growing above still, dark water. The sequence begins with a tightly curled, spiky bud, its deep green form shimmering in soft, diffuse sunlight. The bud deliberately unfurls in a graceful, accelerated motion, revealing its intricately ribbed reddish-purple underside before its waxy, vibrant green topside begins its relentless expansion. As the leaf rapidly unfurls into a perfect circle, its surface reveals lifelike detail, every vein and subtle ripple rendered in 4K resolution, while its waxy texture causes tiny water droplets to dance and coalesce across its expanding surface. Finally, the massive, fully formed pad rests majestically on the water's surface, its raised edges forming sharp, clean edges against the dark liquid. Complex patterns of light filter through the surrounding clear water, creating a serene and awe-inspiring atmosphere of natural wonder.", + "seed": 31067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_4/index_3013_seed_31067_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03698745016008616, + "vlm_logprob": -0.03698745016008616, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03698745016008616, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11872588098049164 + ], + "opposite_logprob_tokens": [ + -2.2437257766723633 + ], + "logprob_reward": -0.11872588098049164, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11872588098049164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15077121555805206, + 0.05194408446550369, + 0.46253013610839844, + 0.16554296016693115, + 0.03223401680588722, + 0.038092393428087234, + 0.04968889057636261, + 0.009925218299031258, + 0.019694002345204353, + 0.019577108323574066 + ], + "credit_map": [ + 0.021334689110517502, + 0.009316523559391499, + 0.019918622449040413, + 0.010985680855810642, + 0.016193827614188194, + 0.02247784473001957, + 0.01733352616429329, + 0.03321050852537155, + 0.005576407071202993, + 0.004077191464602947, + 0.009943712502717972, + 0.005066295620054007, + 0.004401768557727337, + 0.0075057633221149445, + 0.004356785677373409, + 0.011016163975000381, + 0.07439838349819183, + 0.03728485479950905, + 0.051490891724824905, + 0.03811326622962952, + 0.0383799783885479, + 0.06342010200023651, + 0.04991647228598595, + 0.10952615737915039, + 0.013594888150691986, + 0.008335906080901623, + 0.060255199670791626, + 0.007503994274884462, + 0.008318917825818062, + 0.015079372562468052, + 0.006653717253357172, + 0.04580097645521164, + 0.001809844863601029, + 0.0017225380288437009, + 0.0029138396494090557, + 0.0011158199049532413, + 0.001714392681606114, + 0.014134547673165798, + 0.0020070457831025124, + 0.006815986707806587, + 0.0033907697070389986, + 0.004687660373747349, + 0.00381515477783978, + 0.0033326419070363045, + 0.0038447207771241665, + 0.009145507588982582, + 0.002539520151913166, + 0.007336417678743601, + 0.002978369826450944, + 0.002871518488973379, + 0.003499027341604233, + 0.0036669594701379538, + 0.0030512819066643715, + 0.01842479780316353, + 0.003159934189170599, + 0.012036999687552452, + 0.0012444606982171535, + 0.0010081276996061206, + 0.0008271709666587412, + 0.0012535072164610028, + 0.001451433403417468, + 0.0009756634244695306, + 0.001965143484994769, + 0.001199710532091558, + 0.0016544407699257135, + 0.0017043413827195764, + 0.0028099913615733385, + 0.0021159949246793985, + 0.0024079266004264355, + 0.0027532337699085474, + 0.002647821791470051, + 0.003600252093747258, + 0.0018710698932409286, + 0.0023960599210113287, + 0.002425176789984107, + 0.0034404750913381577, + 0.0021162196062505245, + 0.0016209714813157916, + 0.002998140174895525, + 0.0027089950162917376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.11872588098049164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636165712028742 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636165712028742, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636165712028742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3002481460571289, + 0.039558157324790955, + 0.425992488861084, + 0.05686332657933235, + 0.014813560992479324, + 0.07724474370479584, + 0.012505877763032913, + 0.04392760619521141, + 0.015225586481392384, + 0.01362051535397768 + ], + "credit_map": [ + 0.037196673452854156, + 0.03082428127527237, + 0.030690103769302368, + 0.029016127809882164, + 0.039695389568805695, + 0.03603452071547508, + 0.029904337599873543, + 0.06688673049211502, + 0.00355223729275167, + 0.003913209307938814, + 0.004673675633966923, + 0.0035517781507223845, + 0.006666678469628096, + 0.0049976203590631485, + 0.004077110905200243, + 0.008125844411551952, + 0.06663978099822998, + 0.03411390632390976, + 0.04793556407094002, + 0.036436453461647034, + 0.03359128162264824, + 0.05947887524962425, + 0.0444762222468853, + 0.10332043468952179, + 0.0047580343671143055, + 0.00455316761508584, + 0.01372405793517828, + 0.005365219432860613, + 0.004153488669544458, + 0.009758150205016136, + 0.004573057405650616, + 0.009978150017559528, + 0.0012530365493148565, + 0.0012111279647797346, + 0.0018345842836424708, + 0.001019210903905332, + 0.0012496478157117963, + 0.004968234337866306, + 0.0014087948948144913, + 0.0018689241260290146, + 0.019692759960889816, + 0.009299029596149921, + 0.006866377778351307, + 0.005811929237097502, + 0.006556431297212839, + 0.014706550166010857, + 0.005668361205607653, + 0.008643300272524357, + 0.0009110257960855961, + 0.0018815564690157771, + 0.0008048415184020996, + 0.0012052837992087007, + 0.0019221624825149775, + 0.002818272914737463, + 0.0013780378503724933, + 0.001584697631187737, + 0.0050149355083703995, + 0.003991211764514446, + 0.004567882977426052, + 0.0035336462315171957, + 0.0038615132216364145, + 0.01233199518173933, + 0.004721487872302532, + 0.005904934834688902, + 0.001524544321000576, + 0.0016926645766943693, + 0.0015045589534565806, + 0.00116225506644696, + 0.0022499135229736567, + 0.002693260321393609, + 0.0022240348625928164, + 0.0021743543911725283, + 0.0015316706849262118, + 0.0013055949239060283, + 0.0028361869044601917, + 0.0016592987813055515, + 0.0015285257250070572, + 0.001758068217895925, + 0.0012073416728526354, + 0.0017938289092853665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.007636165712028742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738794036209583 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738794036209583, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738794036209583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11910231411457062, + 0.09437434375286102, + 0.22308725118637085, + 0.0633702501654625, + 0.011439552530646324, + 0.10889983922243118, + 0.02055049128830433, + 0.29032647609710693, + 0.01332923024892807, + 0.055520329624414444 + ], + "credit_map": [ + 0.011860586702823639, + 0.011741850525140762, + 0.010524659417569637, + 0.012336008250713348, + 0.01710645854473114, + 0.01464211754500866, + 0.01342612225562334, + 0.02746448665857315, + 0.009750689379870892, + 0.011167743243277073, + 0.010459689423441887, + 0.011330975219607353, + 0.013357216492295265, + 0.011142871342599392, + 0.010162063874304295, + 0.017003072425723076, + 0.03466935083270073, + 0.017674105241894722, + 0.02494427002966404, + 0.018644534051418304, + 0.017587166279554367, + 0.03152519837021828, + 0.02353530190885067, + 0.054507285356521606, + 0.006169636268168688, + 0.005354885943233967, + 0.017773861065506935, + 0.00462214183062315, + 0.004988629370927811, + 0.008169440552592278, + 0.004770237952470779, + 0.011521411128342152, + 0.0011775994207710028, + 0.0014329677214846015, + 0.001665915478952229, + 0.001221805578097701, + 0.0013174175983294845, + 0.0016077925683930516, + 0.0015712539898231626, + 0.0014447978464886546, + 0.02453983947634697, + 0.012684646993875504, + 0.011383053846657276, + 0.008276763372123241, + 0.010331234894692898, + 0.01889491267502308, + 0.008273993618786335, + 0.014515383169054985, + 0.002275473903864622, + 0.0018557956209406257, + 0.002219310961663723, + 0.003543037222698331, + 0.002784613985568285, + 0.0022606204729527235, + 0.0038115375209599733, + 0.0018000975251197815, + 0.012490873225033283, + 0.00655022356659174, + 0.008519568480551243, + 0.008635063655674458, + 0.009959248825907707, + 0.1319909542798996, + 0.012181795202195644, + 0.09999867528676987, + 0.0013924279483035207, + 0.0011696801520884037, + 0.0018624977674335241, + 0.00176041794475168, + 0.00166981085203588, + 0.001703873393125832, + 0.0019102315418422222, + 0.001860288204625249, + 0.005610921885818243, + 0.006386906374245882, + 0.007582042831927538, + 0.010297810658812523, + 0.0056570712476968765, + 0.005127063952386379, + 0.007215235847979784, + 0.007643264252692461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.006738794036209583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189241047948599 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.003189241047948599, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189241047948599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2793009281158447, + 0.060445964336395264, + 0.13490544259548187, + 0.1474844366312027, + 0.03530104458332062, + 0.11231131851673126, + 0.052648406475782394, + 0.12123283743858337, + 0.03212831914424896, + 0.024241328239440918 + ], + "credit_map": [ + 0.041363731026649475, + 0.016181549057364464, + 0.03572620451450348, + 0.020757975056767464, + 0.031093694269657135, + 0.04175026714801788, + 0.0308795478194952, + 0.061547957360744476, + 0.005808481015264988, + 0.003847884712740779, + 0.013271482661366463, + 0.005808331537991762, + 0.005314699839800596, + 0.009015616029500961, + 0.005701151676476002, + 0.011678321287035942, + 0.020297763869166374, + 0.010932596400380135, + 0.01643960177898407, + 0.01237886119633913, + 0.012391099706292152, + 0.018316026777029037, + 0.014691248536109924, + 0.029458247125148773, + 0.010156500153243542, + 0.008423042483627796, + 0.050714652985334396, + 0.00925287976861, + 0.008469825610518456, + 0.017205169424414635, + 0.008804515935480595, + 0.03445785120129585, + 0.0032947808504104614, + 0.004614533390849829, + 0.004423612728714943, + 0.0033923557493835688, + 0.003426098031923175, + 0.007251448929309845, + 0.005082776769995689, + 0.0038154390640556812, + 0.026064427569508553, + 0.014358322136104107, + 0.007674071937799454, + 0.0068700374104082584, + 0.008639135397970676, + 0.030575696378946304, + 0.008051409386098385, + 0.010078209452331066, + 0.005232557654380798, + 0.005376837216317654, + 0.0049988585524261, + 0.008605833165347576, + 0.006568943616002798, + 0.007608856074512005, + 0.010008671320974827, + 0.004247851669788361, + 0.005406067240983248, + 0.003177033504471183, + 0.0037711712066084146, + 0.004445734433829784, + 0.004414374008774757, + 0.054186079651117325, + 0.005707541014999151, + 0.040124837309122086, + 0.0039540925063192844, + 0.002797806402668357, + 0.002497936598956585, + 0.004097577650099993, + 0.004535073414444923, + 0.005135322455316782, + 0.004625903908163309, + 0.00448460690677166, + 0.0024602331686764956, + 0.002650193404406309, + 0.0029181342106312513, + 0.004265951458364725, + 0.003425883362069726, + 0.0021922846790403128, + 0.003838135628029704, + 0.0024905127938836813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.003189241047948599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864716902375221 + ], + "opposite_logprob_tokens": [ + -3.048647165298462 + ], + "logprob_reward": -0.04864716902375221, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864716902375221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2416956126689911, + 0.05440959334373474, + 0.2472316175699234, + 0.12452040612697601, + 0.052478987723588943, + 0.08695103973150253, + 0.04142341390252113, + 0.09875228255987167, + 0.02729650028049946, + 0.025240564718842506 + ], + "credit_map": [ + 0.020779166370630264, + 0.017064746469259262, + 0.044642142951488495, + 0.021520527079701424, + 0.023383431136608124, + 0.03707408532500267, + 0.027123186737298965, + 0.05010833218693733, + 0.004892581142485142, + 0.005398571025580168, + 0.009972193278372288, + 0.004601658321917057, + 0.007838407531380653, + 0.006573798600584269, + 0.005660451017320156, + 0.009471932426095009, + 0.03916455805301666, + 0.019736334681510925, + 0.027140570804476738, + 0.018516618758440018, + 0.019183726981282234, + 0.03444332629442215, + 0.025314398109912872, + 0.0637320801615715, + 0.01743956282734871, + 0.013948057778179646, + 0.028579657897353172, + 0.00802626833319664, + 0.008499642834067345, + 0.01934792846441269, + 0.009274815209209919, + 0.019404474645853043, + 0.005802893079817295, + 0.007657806854695082, + 0.006635919213294983, + 0.004936939105391502, + 0.006651544012129307, + 0.006484129000455141, + 0.008285680785775185, + 0.006024077069014311, + 0.009777894243597984, + 0.01550975814461708, + 0.010382558219134808, + 0.0074366009794175625, + 0.012166531756520271, + 0.009700223803520203, + 0.010100957937538624, + 0.01187652163207531, + 0.003846548032015562, + 0.004058347083628178, + 0.0047983285039663315, + 0.007654913701117039, + 0.0038382583297789097, + 0.004185510333627462, + 0.009727983735501766, + 0.0033135232515633106, + 0.006009074859321117, + 0.0028548436239361763, + 0.003810401074588299, + 0.0038201259449124336, + 0.003862014738842845, + 0.04430457949638367, + 0.003021042328327894, + 0.031070200726389885, + 0.0028767366893589497, + 0.002792459912598133, + 0.0027456439565867186, + 0.003658168250694871, + 0.004171573091298342, + 0.0035903656389564276, + 0.004066470079123974, + 0.0033950807992368937, + 0.0023253769613802433, + 0.0025109860580414534, + 0.0031711000483483076, + 0.0037729491014033556, + 0.0037468820810317993, + 0.002378948498517275, + 0.0046954043209552765, + 0.002638918114826083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04864716902375221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, bird's-eye view captures hyper-realistic CGI time-lapse footage of a massive Victoria Amazon water lily growing above still, dark water. The sequence begins with a tightly curled, spiky bud, its deep green form shimmering in soft, diffuse sunlight. The bud deliberately unfurls in a graceful, accelerated motion, revealing its intricately ribbed reddish-purple underside before its waxy, vibrant green topside begins its relentless expansion. As the leaf rapidly unfurls into a perfect circle, its surface reveals lifelike detail, every vein and subtle ripple rendered in 4K resolution, while its waxy texture causes tiny water droplets to dance and coalesce across its expanding surface. Finally, the massive, fully formed pad rests majestically on the water's surface, its raised edges forming sharp, clean edges against the dark liquid. Complex patterns of light filter through the surrounding clear water, creating a serene and awe-inspiring atmosphere of natural wonder.", + "seed": 31068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_4/index_3013_seed_31068_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05441850158385932, + "vlm_logprob": -0.05441850158385932, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.05441850158385932, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2071339637041092 + ], + "opposite_logprob_tokens": [ + -1.7071340084075928 + ], + "logprob_reward": -0.2071339637041092, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2071339637041092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1154552474617958, + 0.2210991233587265, + 0.2142874300479889, + 0.106842540204525, + 0.06148834526538849, + 0.05852540582418442, + 0.041108883917331696, + 0.09576486796140671, + 0.05403933674097061, + 0.031388863921165466 + ], + "credit_map": [ + 0.02036181092262268, + 0.00967469159513712, + 0.019900137558579445, + 0.011865878477692604, + 0.010645187459886074, + 0.015547861345112324, + 0.01329729612916708, + 0.014162370935082436, + 0.036945145577192307, + 0.02086237631738186, + 0.02737310528755188, + 0.017198596149683, + 0.019847378134727478, + 0.029869696125388145, + 0.027322975918650627, + 0.04167984426021576, + 0.049336399883031845, + 0.01297038421034813, + 0.013490501791238785, + 0.013677998445928097, + 0.018943024799227715, + 0.06247899681329727, + 0.02445662021636963, + 0.0189334936439991, + 0.012890773825347424, + 0.009479993022978306, + 0.015301342122256756, + 0.011506347917020321, + 0.007704867050051689, + 0.01766832172870636, + 0.011491627432405949, + 0.02079925313591957, + 0.007001448888331652, + 0.0076859076507389545, + 0.006714490242302418, + 0.008238330483436584, + 0.007473061792552471, + 0.006740650627762079, + 0.009389175102114677, + 0.008245280012488365, + 0.010042120702564716, + 0.006112317554652691, + 0.006237647496163845, + 0.007888616994023323, + 0.005299726966768503, + 0.004415393341332674, + 0.007920375093817711, + 0.010609196498990059, + 0.004095611162483692, + 0.0036495663225650787, + 0.0034442441537976265, + 0.006176987197250128, + 0.005299304146319628, + 0.006044031120836735, + 0.005192074924707413, + 0.007207061164081097, + 0.008059258572757244, + 0.021471751853823662, + 0.005550985690206289, + 0.010874445550143719, + 0.02039606310427189, + 0.006697278935462236, + 0.0077238259837031364, + 0.014991250820457935, + 0.0074964286759495735, + 0.006438296753913164, + 0.008060316555202007, + 0.00808208528906107, + 0.006049510091543198, + 0.005803035106509924, + 0.005623057950288057, + 0.006486598402261734, + 0.0028342439327389, + 0.005969952791929245, + 0.003492928110063076, + 0.004173563327640295, + 0.004884431604295969, + 0.0028357484843581915, + 0.004164216108620167, + 0.0030337770003825426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.2071339637041092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747556384652853 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747556384652853, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747556384652853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1641329824924469, + 0.20513954758644104, + 0.20431840419769287, + 0.0529804602265358, + 0.025104926899075508, + 0.032911770045757294, + 0.032359346747398376, + 0.2220860719680786, + 0.042214006185531616, + 0.018752554431557655 + ], + "credit_map": [ + 0.023590736091136932, + 0.016623370349407196, + 0.025379380211234093, + 0.018588261678814888, + 0.01986105553805828, + 0.017277361825108528, + 0.0227394737303257, + 0.020073339343070984, + 0.03517238423228264, + 0.0157309640198946, + 0.025203395634889603, + 0.023690272122621536, + 0.022179152816534042, + 0.022369178012013435, + 0.02643207460641861, + 0.03436212241649628, + 0.06015004962682724, + 0.009196464903652668, + 0.008581877686083317, + 0.007095139939337969, + 0.014608587138354778, + 0.06170850619673729, + 0.021489866077899933, + 0.021487904712557793, + 0.003728418843820691, + 0.002522811759263277, + 0.003551441477611661, + 0.003448857693001628, + 0.003565858816727996, + 0.016240553930401802, + 0.005831744521856308, + 0.014090778306126595, + 0.0031399638392031193, + 0.0044261920265853405, + 0.0022045813966542482, + 0.0032417625188827515, + 0.0023238016292452812, + 0.0031553914304822683, + 0.0027925202157348394, + 0.0038207119796425104, + 0.0033683518413454294, + 0.0043347361497581005, + 0.0036278809420764446, + 0.005809888243675232, + 0.0028582923114299774, + 0.0033730599097907543, + 0.004823165014386177, + 0.004716395866125822, + 0.0031949083786457777, + 0.0063827550038695335, + 0.0027657507453113794, + 0.004348340444266796, + 0.0042405505664646626, + 0.0033161358442157507, + 0.0036424638237804174, + 0.004468444734811783, + 0.01859399676322937, + 0.05174396559596062, + 0.012668594717979431, + 0.026259558275341988, + 0.0470651239156723, + 0.014602258801460266, + 0.016852296888828278, + 0.03430027887225151, + 0.005656396970152855, + 0.005045418161898851, + 0.005772905424237251, + 0.006381341256201267, + 0.004384883213788271, + 0.00531552592292428, + 0.00482726376503706, + 0.004830274265259504, + 0.0019490037811920047, + 0.0035087596625089645, + 0.0017318149330094457, + 0.0022904234938323498, + 0.0025738768745213747, + 0.0016424359055235982, + 0.0025787947233766317, + 0.002477445872500539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.006747556384652853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059529622085392475 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.0059529622085392475, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059529622085392475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12247084081172943, + 0.12379451096057892, + 0.1309456080198288, + 0.13347604870796204, + 0.14778541028499603, + 0.045011624693870544, + 0.07285560667514801, + 0.13887465000152588, + 0.05222427472472191, + 0.03256145864725113 + ], + "credit_map": [ + 0.020475754514336586, + 0.014589344151318073, + 0.015430268831551075, + 0.011728595942258835, + 0.013733288273215294, + 0.013811861164867878, + 0.018008796498179436, + 0.01469291653484106, + 0.017636960372328758, + 0.01510450430214405, + 0.01125277392566204, + 0.014011920429766178, + 0.014876948669552803, + 0.01584370993077755, + 0.021385274827480316, + 0.0136824119836092, + 0.028325820341706276, + 0.01693437248468399, + 0.009658515453338623, + 0.009935248643159866, + 0.01676003821194172, + 0.016276702284812927, + 0.013353799469769001, + 0.01970110833644867, + 0.00956558994948864, + 0.007468355819582939, + 0.009255998767912388, + 0.007953180000185966, + 0.00829561147838831, + 0.043540000915527344, + 0.016363536939024925, + 0.031033767387270927, + 0.019510118290781975, + 0.01687966287136078, + 0.01424679160118103, + 0.018762290477752686, + 0.013135549612343311, + 0.02014307491481304, + 0.024107668548822403, + 0.021000241860747337, + 0.006102443672716618, + 0.005820929538458586, + 0.0049639069475233555, + 0.005750387441366911, + 0.00440100347623229, + 0.004410915542393923, + 0.006801695562899113, + 0.006760342046618462, + 0.007712468504905701, + 0.006923356559127569, + 0.009349139407277107, + 0.010248380713164806, + 0.006858489941805601, + 0.008619757369160652, + 0.013860284350812435, + 0.009283722378313541, + 0.0115651311352849, + 0.030739696696400642, + 0.007123502437025309, + 0.016842108219861984, + 0.029741069301962852, + 0.011063524521887302, + 0.010262236930429935, + 0.02153736725449562, + 0.007236899342387915, + 0.0054534119553864, + 0.008506565354764462, + 0.005546195898205042, + 0.005648978054523468, + 0.0071558402851223946, + 0.006578977219760418, + 0.006097408477216959, + 0.003357179230079055, + 0.006110081914812326, + 0.0028453334234654903, + 0.003241288475692272, + 0.004895790480077267, + 0.004423641599714756, + 0.004721627105027437, + 0.0029665157198905945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.0059529622085392475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036128980573266745 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.0036128980573266745, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036128980573266745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049599193036556244, + 0.04688611626625061, + 0.22014449536800385, + 0.06892428547143936, + 0.05116230249404907, + 0.031108742579817772, + 0.11965020000934601, + 0.2350071668624878, + 0.12505017220973969, + 0.052467308938503265 + ], + "credit_map": [ + 0.009020053781569004, + 0.004820453003048897, + 0.006743161473423243, + 0.00412095058709383, + 0.006307985633611679, + 0.007005852647125721, + 0.0055756838992238045, + 0.006005049683153629, + 0.0062609449960291386, + 0.005274344235658646, + 0.00559480395168066, + 0.003996357321739197, + 0.0056251417845487595, + 0.005351589061319828, + 0.0057508572936058044, + 0.00903207529336214, + 0.07054390013217926, + 0.008881423622369766, + 0.008735326118767262, + 0.005501062609255314, + 0.014682304114103317, + 0.06189288571476936, + 0.021386686712503433, + 0.028520895168185234, + 0.004821313079446554, + 0.005334293004125357, + 0.005403856746852398, + 0.006658686324954033, + 0.004342552740126848, + 0.018638767302036285, + 0.008203874342143536, + 0.015520951710641384, + 0.006947725545614958, + 0.006770224776118994, + 0.004914279095828533, + 0.00572045287117362, + 0.005296020302921534, + 0.0070930528454482555, + 0.006892614532262087, + 0.007527932059019804, + 0.004736192990094423, + 0.0028474547434598207, + 0.00366984773427248, + 0.00412429915741086, + 0.0029175472445786, + 0.002685053739696741, + 0.004237998742610216, + 0.005890347063541412, + 0.014568304643034935, + 0.01028028316795826, + 0.015320315957069397, + 0.017148274928331375, + 0.012022733688354492, + 0.014805469661951065, + 0.020647037774324417, + 0.014857782982289791, + 0.019245265051722527, + 0.053662028163671494, + 0.013670363463461399, + 0.026043763384222984, + 0.05013198032975197, + 0.018437538295984268, + 0.017384493723511696, + 0.03643173351883888, + 0.013258770108222961, + 0.01903199404478073, + 0.015752727165818214, + 0.017427880316972733, + 0.016816891729831696, + 0.015655912458896637, + 0.012352202087640762, + 0.014753790572285652, + 0.004526111762970686, + 0.010145188309252262, + 0.005091249477118254, + 0.007283839397132397, + 0.007886374369263649, + 0.0036174096167087555, + 0.007836803793907166, + 0.00608033174648881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.0036128980573266745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048645127564668655 + ], + "opposite_logprob_tokens": [ + -3.04864501953125 + ], + "logprob_reward": -0.048645127564668655, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048645127564668655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1467788964509964, + 0.037550002336502075, + 0.13321492075920105, + 0.1540757119655609, + 0.051413774490356445, + 0.023504652082920074, + 0.05059193819761276, + 0.18188896775245667, + 0.17145907878875732, + 0.04952201247215271 + ], + "credit_map": [ + 0.026836765930056572, + 0.009149173274636269, + 0.02431335486471653, + 0.009602445177733898, + 0.011753604747354984, + 0.022213641554117203, + 0.0171740110963583, + 0.025735916569828987, + 0.005319464486092329, + 0.0036472026258707047, + 0.003916976507753134, + 0.004531143698841333, + 0.0050232564099133015, + 0.005045396741479635, + 0.00610326137393713, + 0.0039633046835660934, + 0.03923170268535614, + 0.008958148770034313, + 0.006127473432570696, + 0.005270523019134998, + 0.012151769362390041, + 0.03131747990846634, + 0.012952730990946293, + 0.017205102369189262, + 0.008983195759356022, + 0.005564352963119745, + 0.006142387632280588, + 0.0070856050588190556, + 0.006489896215498447, + 0.05478902906179428, + 0.018708854913711548, + 0.04631238803267479, + 0.007310851942747831, + 0.007566964253783226, + 0.0048741125501692295, + 0.0052568051032722, + 0.007884891703724861, + 0.006390400230884552, + 0.005779758561402559, + 0.006349994335323572, + 0.002287583891302347, + 0.003996541723608971, + 0.002152289729565382, + 0.004429131746292114, + 0.0022583077661693096, + 0.001996138831600547, + 0.003235690761357546, + 0.003148971125483513, + 0.006007755175232887, + 0.00693931058049202, + 0.005422364920377731, + 0.007579414173960686, + 0.005487390328198671, + 0.005750768352299929, + 0.00711371935904026, + 0.006291217636317015, + 0.01543372217565775, + 0.0387902557849884, + 0.008832654915750027, + 0.02173311822116375, + 0.03777246177196503, + 0.017350973561406136, + 0.012611618265509605, + 0.02936418540775776, + 0.016570139676332474, + 0.03652431070804596, + 0.018520455807447433, + 0.025656668469309807, + 0.024536756798624992, + 0.012989046052098274, + 0.0166885145008564, + 0.019973212853074074, + 0.005354992114007473, + 0.009413674473762512, + 0.005124762188643217, + 0.006373888347297907, + 0.005858296528458595, + 0.00575631856918335, + 0.0064414613880217075, + 0.0051986235193908215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.048645127564668655 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, bird's-eye view captures hyper-realistic CGI time-lapse footage of a massive Victoria Amazon water lily growing above still, dark water. The sequence begins with a tightly curled, spiky bud, its deep green form shimmering in soft, diffuse sunlight. The bud deliberately unfurls in a graceful, accelerated motion, revealing its intricately ribbed reddish-purple underside before its waxy, vibrant green topside begins its relentless expansion. As the leaf rapidly unfurls into a perfect circle, its surface reveals lifelike detail, every vein and subtle ripple rendered in 4K resolution, while its waxy texture causes tiny water droplets to dance and coalesce across its expanding surface. Finally, the massive, fully formed pad rests majestically on the water's surface, its raised edges forming sharp, clean edges against the dark liquid. Complex patterns of light filter through the surrounding clear water, creating a serene and awe-inspiring atmosphere of natural wonder.", + "seed": 31069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_4/index_3013_seed_31069_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08214948065578938, + "vlm_logprob": -0.08214948065578938, + "vlm_yes_no_margin": 4.249999988079071, + "vlm_token_credit": -0.08214948065578938, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3527901768684387 + ], + "opposite_logprob_tokens": [ + -1.227790117263794 + ], + "logprob_reward": -0.3527901768684387, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3527901768684387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03063233382999897, + 0.05714922398328781, + 0.10114909708499908, + 0.1191684752702713, + 0.05772540345788002, + 0.021164188161492348, + 0.015158351510763168, + 0.11589423567056656, + 0.46461576223373413, + 0.01734289899468422 + ], + "credit_map": [ + 0.004518196452409029, + 0.002323649823665619, + 0.003703501308336854, + 0.005416134838014841, + 0.0027485389728099108, + 0.003305835649371147, + 0.003641424933448434, + 0.004975051619112492, + 0.01375923864543438, + 0.002385820960626006, + 0.023232560604810715, + 0.003051802283152938, + 0.0028420027811080217, + 0.003806841792538762, + 0.005954820197075605, + 0.0021161362528800964, + 0.02810046449303627, + 0.005152075085788965, + 0.01177277322858572, + 0.0065286653116345406, + 0.006917653605341911, + 0.027257723733782768, + 0.007228503469377756, + 0.008191234432160854, + 0.010622527450323105, + 0.005841508973389864, + 0.039973754435777664, + 0.007212307304143906, + 0.00545602198690176, + 0.03713828697800636, + 0.0056657325476408005, + 0.007258351426571608, + 0.005886235740035772, + 0.009036479517817497, + 0.00851211417466402, + 0.006535765714943409, + 0.014544462785124779, + 0.0036834427155554295, + 0.004961241502314806, + 0.004565659444779158, + 0.0020381538197398186, + 0.0013897132594138384, + 0.005298446863889694, + 0.0019477286841720343, + 0.004057538229972124, + 0.0024174009449779987, + 0.0016018389724195004, + 0.00241336552426219, + 0.0013338319258764386, + 0.001518594566732645, + 0.0023790907580405474, + 0.0024084586184471846, + 0.0019073588773608208, + 0.0016716571990400553, + 0.002059758407995105, + 0.0018796022050082684, + 0.007937878370285034, + 0.009258637204766273, + 0.010458359494805336, + 0.008540607988834381, + 0.03568084537982941, + 0.005903147626668215, + 0.00850044097751379, + 0.029614314436912537, + 0.034842509776353836, + 0.0465729795396328, + 0.12519988417625427, + 0.020225122570991516, + 0.16531196236610413, + 0.017992569133639336, + 0.03377246856689453, + 0.02069827727973461, + 0.002575357910245657, + 0.0026332412380725145, + 0.0020718337036669254, + 0.0018267270643264055, + 0.0020709580276161432, + 0.0018663399387151003, + 0.0021283654496073723, + 0.0021700761280953884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.3527901768684387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036245384253561497 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.0036245384253561497, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036245384253561497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05458233878016472, + 0.41275617480278015, + 0.14504505693912506, + 0.15148447453975677, + 0.012599049136042595, + 0.016090501099824905, + 0.011267171241343021, + 0.02747110277414322, + 0.16117429733276367, + 0.00752982497215271 + ], + "credit_map": [ + 0.0067628915421664715, + 0.004204835742712021, + 0.006479594856500626, + 0.007194181438535452, + 0.006181604694575071, + 0.0058583226054906845, + 0.006702690850943327, + 0.01119821984320879, + 0.0997256264090538, + 0.016390196979045868, + 0.16862663626670837, + 0.02700391784310341, + 0.013935244642198086, + 0.031061047688126564, + 0.039726559072732925, + 0.016286971047520638, + 0.03789224103093147, + 0.008223431184887886, + 0.016251208260655403, + 0.011938348412513733, + 0.012252568267285824, + 0.039890021085739136, + 0.01195768266916275, + 0.006639560684561729, + 0.013196373358368874, + 0.008067815564572811, + 0.049796152859926224, + 0.008707907050848007, + 0.0065484922379255295, + 0.048973385244607925, + 0.006844379007816315, + 0.009349972009658813, + 0.0016825763741508126, + 0.001575001748278737, + 0.0020140144042670727, + 0.0017876846250146627, + 0.0018438183469697833, + 0.0012257066555321217, + 0.00134399498347193, + 0.0011262515326961875, + 0.0018111361423507333, + 0.001757766935043037, + 0.0028180601075291634, + 0.0018677618354558945, + 0.0032746342476457357, + 0.0016437574522569776, + 0.0014221281744539738, + 0.001495257718488574, + 0.0010062485234811902, + 0.00143782130908221, + 0.0019175586057826877, + 0.0013539326610043645, + 0.0011083855060860515, + 0.001078783767297864, + 0.0008831615559756756, + 0.0024812803603708744, + 0.002447701757773757, + 0.0025635831989347935, + 0.0032344984356313944, + 0.0022227298468351364, + 0.007025036960840225, + 0.001618959242478013, + 0.0023362948559224606, + 0.006022298242896795, + 0.01230732910335064, + 0.016632182523608208, + 0.04273542761802673, + 0.00713958777487278, + 0.05676687881350517, + 0.006589390803128481, + 0.01166345365345478, + 0.007340043317526579, + 0.0014607051853090525, + 0.0010837314184755087, + 0.0007414058782160282, + 0.0007974054315127432, + 0.0009300229139626026, + 0.0007284712046384811, + 0.000895946694072336, + 0.0008921363041736186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.0036245384253561497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031947072129696608 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031947072129696608, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031947072129696608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037544019520282745, + 0.31010836362838745, + 0.06465785205364227, + 0.05379858613014221, + 0.019333936274051666, + 0.028855206444859505, + 0.02461947686970234, + 0.16341499984264374, + 0.28716644644737244, + 0.010501149110496044 + ], + "credit_map": [ + 0.004521208815276623, + 0.002726237988099456, + 0.003931951709091663, + 0.007196529768407345, + 0.0033678687177598476, + 0.004150418099015951, + 0.004374632611870766, + 0.0072751715779304504, + 0.0731271281838417, + 0.011456260457634926, + 0.13152258098125458, + 0.018604716286063194, + 0.011965321376919746, + 0.020414620637893677, + 0.030668148770928383, + 0.012349617667496204, + 0.009317383170127869, + 0.003985017538070679, + 0.015710124745965004, + 0.00713313976302743, + 0.004952039569616318, + 0.00892607867717743, + 0.007202820852398872, + 0.007431244943290949, + 0.004673187620937824, + 0.005054210778325796, + 0.015297203324735165, + 0.0035661503206938505, + 0.0033623811323195696, + 0.01576399989426136, + 0.0028142693918198347, + 0.0032671811059117317, + 0.002353839809074998, + 0.002891974290832877, + 0.0032759823370724916, + 0.0019996752962470055, + 0.003366414224728942, + 0.001879544579423964, + 0.0012103142216801643, + 0.002356191398575902, + 0.0032735916320234537, + 0.0026634063106030226, + 0.003489407477900386, + 0.0035746258217841387, + 0.00370120070874691, + 0.006066611036658287, + 0.003001271514222026, + 0.0030850928742438555, + 0.0016697078244760633, + 0.0022972419392317533, + 0.0049010817892849445, + 0.002525682095438242, + 0.002527765231207013, + 0.0017108137253671885, + 0.0019273128127679229, + 0.007059871684759855, + 0.011227631010115147, + 0.013716126792132854, + 0.011165495961904526, + 0.011297406628727913, + 0.05515403300523758, + 0.007846160791814327, + 0.01183445006608963, + 0.04117368906736374, + 0.021372059360146523, + 0.02917824499309063, + 0.07583092898130417, + 0.012187041342258453, + 0.10274815559387207, + 0.01120477169752121, + 0.021907800808548927, + 0.012737426906824112, + 0.0018033967353403568, + 0.001554020680487156, + 0.0012261675437912345, + 0.001086508040316403, + 0.0012106827925890684, + 0.000986689468845725, + 0.0014210648369044065, + 0.0012126198271289468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.0031947072129696608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486948622390628 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.002486948622390628, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486948622390628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030379515141248703, + 0.026557348668575287, + 0.06176847219467163, + 0.09406431764364243, + 0.015521757304668427, + 0.032575637102127075, + 0.035670336335897446, + 0.327674925327301, + 0.31978780031204224, + 0.05599989742040634 + ], + "credit_map": [ + 0.0041754175908863544, + 0.0036530825309455395, + 0.0028847334906458855, + 0.004601226653903723, + 0.004319046624004841, + 0.0028119164053350687, + 0.0035017323680222034, + 0.004432360641658306, + 0.004010487347841263, + 0.0033159409649670124, + 0.0032431159634143114, + 0.0037865452468395233, + 0.0031076681334525347, + 0.002543423557654023, + 0.003949099685996771, + 0.0026010675355792046, + 0.008333084173500538, + 0.004778847564011812, + 0.011361228302121162, + 0.006640464998781681, + 0.006346132140606642, + 0.010615020990371704, + 0.005437307525426149, + 0.00825638510286808, + 0.008451557718217373, + 0.005353784188628197, + 0.031118158251047134, + 0.005436030216515064, + 0.004409278277307749, + 0.02908891811966896, + 0.004415096715092659, + 0.0057914890348911285, + 0.0015394581714645028, + 0.0020111072808504105, + 0.0022650195751339197, + 0.0021956737618893385, + 0.002169294748455286, + 0.0015980564057826996, + 0.0019151951419189572, + 0.0018279521027579904, + 0.003126035910099745, + 0.0026078096125274897, + 0.007629011292010546, + 0.0036042556166648865, + 0.0056472341530025005, + 0.0037206257693469524, + 0.0025000886525958776, + 0.0037405742332339287, + 0.002477033296599984, + 0.005518946330994368, + 0.005552816670387983, + 0.004763191565871239, + 0.0049185543321073055, + 0.0025722836144268513, + 0.00338550703600049, + 0.00648199999704957, + 0.021516049280762672, + 0.027232171967625618, + 0.02208595536649227, + 0.02261986769735813, + 0.11120244115591049, + 0.015602393075823784, + 0.024530254304409027, + 0.08288579434156418, + 0.024251187220215797, + 0.032322634011507034, + 0.08246146142482758, + 0.01489483006298542, + 0.11453316360712051, + 0.01260156650096178, + 0.025106515735387802, + 0.013616456650197506, + 0.008856565691530704, + 0.00648729270324111, + 0.006768961437046528, + 0.0059976777993142605, + 0.006080647464841604, + 0.0069248792715370655, + 0.009436696767807007, + 0.005447179079055786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.002486948622390628 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865103214979172 + ], + "opposite_logprob_tokens": [ + -3.0486509799957275 + ], + "logprob_reward": -0.04865103214979172, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865103214979172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041069358587265015, + 0.09384428709745407, + 0.07625168561935425, + 0.046464815735816956, + 0.043299756944179535, + 0.01550906803458929, + 0.02496350184082985, + 0.5496298670768738, + 0.08340407907962799, + 0.025563592091202736 + ], + "credit_map": [ + 0.004401244688779116, + 0.005721484310925007, + 0.004913016688078642, + 0.005262479651719332, + 0.006727559491991997, + 0.003066475735977292, + 0.004956540651619434, + 0.00602055573835969, + 0.022450242191553116, + 0.004432064015418291, + 0.036318324506282806, + 0.005682620219886303, + 0.005694624036550522, + 0.005685153417289257, + 0.01049182377755642, + 0.0030894342344254255, + 0.024729274213314056, + 0.0038934112526476383, + 0.003742792410776019, + 0.004410612862557173, + 0.004559170920401812, + 0.025264084339141846, + 0.00564881507307291, + 0.004003522917628288, + 0.00474931625649333, + 0.0022054777946323156, + 0.013996021822094917, + 0.0033461025450378656, + 0.002300102496519685, + 0.014036275446414948, + 0.0027977675199508667, + 0.0030337488278746605, + 0.004651636350899935, + 0.005643364042043686, + 0.005762979853898287, + 0.004173713270574808, + 0.010735654272139072, + 0.0034436662681400776, + 0.004803490824997425, + 0.004085248801857233, + 0.0016091252909973264, + 0.0016857136506587267, + 0.0016357321292161942, + 0.002469580387696624, + 0.0016696221427991986, + 0.002374488627538085, + 0.002319976920261979, + 0.0017448273720219731, + 0.0019126542611047626, + 0.0025609731674194336, + 0.003938321955502033, + 0.0034779259003698826, + 0.0029965403955429792, + 0.0026719304732978344, + 0.002916154218837619, + 0.004488998558372259, + 0.03536328300833702, + 0.04608360305428505, + 0.03510601446032524, + 0.03976232931017876, + 0.18980450928211212, + 0.026165327057242393, + 0.041893377900123596, + 0.13545142114162445, + 0.007453239057213068, + 0.009328932501375675, + 0.01777203381061554, + 0.005420881789177656, + 0.027082735672593117, + 0.0036882939748466015, + 0.008046726696193218, + 0.004611227195709944, + 0.004612219985574484, + 0.0037831319496035576, + 0.0021690421272069216, + 0.0034304335713386536, + 0.00440102769061923, + 0.002130460226908326, + 0.0027535930275917053, + 0.0022836823482066393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.04865103214979172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, bird's-eye view captures hyper-realistic CGI time-lapse footage of a massive Victoria Amazon water lily growing above still, dark water. The sequence begins with a tightly curled, spiky bud, its deep green form shimmering in soft, diffuse sunlight. The bud deliberately unfurls in a graceful, accelerated motion, revealing its intricately ribbed reddish-purple underside before its waxy, vibrant green topside begins its relentless expansion. As the leaf rapidly unfurls into a perfect circle, its surface reveals lifelike detail, every vein and subtle ripple rendered in 4K resolution, while its waxy texture causes tiny water droplets to dance and coalesce across its expanding surface. Finally, the massive, fully formed pad rests majestically on the water's surface, its raised edges forming sharp, clean edges against the dark liquid. Complex patterns of light filter through the surrounding clear water, creating a serene and awe-inspiring atmosphere of natural wonder.", + "seed": 31070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_4/index_3013_seed_31070_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024058731086552144, + "vlm_logprob": -0.024058731086552144, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.024058731086552144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06482206284999847 + ], + "opposite_logprob_tokens": [ + -2.81482195854187 + ], + "logprob_reward": -0.06482206284999847, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06482206284999847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1130005344748497, + 0.07412116229534149, + 0.15155303478240967, + 0.04584411159157753, + 0.08486732840538025, + 0.34359410405158997, + 0.06059764698147774, + 0.05228298157453537, + 0.037285979837179184, + 0.0368531197309494 + ], + "credit_map": [ + 0.027370546013116837, + 0.012746589258313179, + 0.02042314037680626, + 0.009946221485733986, + 0.008825583383440971, + 0.018647270277142525, + 0.0064375088550150394, + 0.008603674359619617, + 0.011172828264534473, + 0.005128012504428625, + 0.020683566108345985, + 0.005346102640032768, + 0.005486087407916784, + 0.01477017067372799, + 0.005248832516372204, + 0.006285566836595535, + 0.013136346824467182, + 0.011576725170016289, + 0.014107256196439266, + 0.031285155564546585, + 0.01265079714357853, + 0.04332052171230316, + 0.013455163687467575, + 0.012021076865494251, + 0.00584156671538949, + 0.0046873558312654495, + 0.0062663801945745945, + 0.0049303756095469, + 0.006542168091982603, + 0.0041071027517318726, + 0.0038909607101231813, + 0.009578204713761806, + 0.013014934957027435, + 0.009991742670536041, + 0.01157340407371521, + 0.009338146075606346, + 0.009347639977931976, + 0.008582188747823238, + 0.016155071556568146, + 0.0068642026744782925, + 0.12547840178012848, + 0.018252700567245483, + 0.03893200308084488, + 0.008328884840011597, + 0.014479883946478367, + 0.014958668500185013, + 0.013513981364667416, + 0.10964958369731903, + 0.006495214067399502, + 0.01052886713296175, + 0.006266269367188215, + 0.007839932106435299, + 0.006750491447746754, + 0.005418501794338226, + 0.010995998978614807, + 0.006302378606051207, + 0.005500608589500189, + 0.0062394277192652225, + 0.005499293562024832, + 0.010850110091269016, + 0.00630845595151186, + 0.004551652818918228, + 0.008141712285578251, + 0.00519172428175807, + 0.00419586943462491, + 0.00308672315441072, + 0.005211502779275179, + 0.005307547282427549, + 0.003547536674886942, + 0.004763967823237181, + 0.005994174629449844, + 0.005178660620003939, + 0.0037341206334531307, + 0.0034601022489368916, + 0.005939696449786425, + 0.005244901869446039, + 0.0036474871449172497, + 0.003428306896239519, + 0.007542926352471113, + 0.0038555802311748266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.06482206284999847 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644440021365881 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644440021365881, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644440021365881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20458702743053436, + 0.1718100607395172, + 0.20110760629177094, + 0.062097664922475815, + 0.03602632135152817, + 0.11158396303653717, + 0.07083163410425186, + 0.05912482365965843, + 0.04986265301704407, + 0.03296829015016556 + ], + "credit_map": [ + 0.03219294548034668, + 0.018556127324700356, + 0.037819404155015945, + 0.020018938928842545, + 0.01811549812555313, + 0.04004283621907234, + 0.020326385274529457, + 0.017514880746603012, + 0.02032133936882019, + 0.010529431514441967, + 0.03879424184560776, + 0.01123941782861948, + 0.014882120303809643, + 0.05316420644521713, + 0.009985125623643398, + 0.012894165702164173, + 0.025919001549482346, + 0.017769599333405495, + 0.023484472185373306, + 0.03179964795708656, + 0.014017507433891296, + 0.0572015605866909, + 0.016347384080290794, + 0.014568418264389038, + 0.009770753793418407, + 0.006157843861728907, + 0.00896831601858139, + 0.00885093491524458, + 0.007589215412735939, + 0.006303914822638035, + 0.004377222154289484, + 0.0100794592872262, + 0.005435664672404528, + 0.003522333689033985, + 0.005623439326882362, + 0.003994140774011612, + 0.004388525616377592, + 0.0032530336175113916, + 0.005663234740495682, + 0.004145941697061062, + 0.021823903545737267, + 0.00870941486209631, + 0.010779195465147495, + 0.013045402243733406, + 0.011780600063502789, + 0.006497777532786131, + 0.01323251985013485, + 0.025715135037899017, + 0.0053602722473442554, + 0.0058975666761398315, + 0.013901540078222752, + 0.0070638395845890045, + 0.006884991656988859, + 0.0047278874553740025, + 0.012435643933713436, + 0.01455988734960556, + 0.006108660250902176, + 0.006470746826380491, + 0.005600623320788145, + 0.012277990579605103, + 0.0073745641857385635, + 0.005790282506495714, + 0.010063335299491882, + 0.005438607186079025, + 0.003973647952079773, + 0.009447930380702019, + 0.004788741935044527, + 0.009024790488183498, + 0.006309421267360449, + 0.004832478705793619, + 0.007091003004461527, + 0.00439463509246707, + 0.004095603711903095, + 0.002761622192338109, + 0.005035890731960535, + 0.00364019931294024, + 0.0028699045069515705, + 0.003905062098056078, + 0.007198642008006573, + 0.0034613606985658407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004644440021365881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638394996523857 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638394996523857, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638394996523857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15205495059490204, + 0.1452629417181015, + 0.1463342010974884, + 0.0725964829325676, + 0.0698830783367157, + 0.13209931552410126, + 0.05339517071843147, + 0.12199698388576508, + 0.05181579664349556, + 0.05456104129552841 + ], + "credit_map": [ + 0.021773703396320343, + 0.012908443808555603, + 0.019155360758304596, + 0.01846328377723694, + 0.022411124780774117, + 0.021299604326486588, + 0.017627598717808723, + 0.018415844067931175, + 0.013411232270300388, + 0.013992753811180592, + 0.02706013061106205, + 0.014641928486526012, + 0.020271750167012215, + 0.027435805648565292, + 0.012630215846002102, + 0.015819130465388298, + 0.012690327130258083, + 0.011046463623642921, + 0.015918249264359474, + 0.027192682027816772, + 0.015641285106539726, + 0.038506947457790375, + 0.01441358681768179, + 0.010924656875431538, + 0.007713030092418194, + 0.012203743681311607, + 0.009345784783363342, + 0.011398726142942905, + 0.008263658732175827, + 0.006459362339228392, + 0.00605251407250762, + 0.011159664019942284, + 0.010133715346455574, + 0.0077538141049444675, + 0.008550378493964672, + 0.007908674888312817, + 0.008153926581144333, + 0.008048268966376781, + 0.013437771238386631, + 0.0058965315110981464, + 0.018562667071819305, + 0.011278386227786541, + 0.012689933180809021, + 0.01777445524930954, + 0.017313484102487564, + 0.01178356446325779, + 0.02938142418861389, + 0.013315393589437008, + 0.00539392651990056, + 0.0045239548198878765, + 0.0082613006234169, + 0.0074496762827038765, + 0.005714808590710163, + 0.004650262650102377, + 0.011592934839427471, + 0.005808310117572546, + 0.011700031347572803, + 0.014301443472504616, + 0.012459767051041126, + 0.019857661798596382, + 0.016207531094551086, + 0.016654176637530327, + 0.01839473843574524, + 0.01242163497954607, + 0.005916535388678312, + 0.004708434455096722, + 0.008602578192949295, + 0.008503199554979801, + 0.006157068070024252, + 0.0037503368221223354, + 0.009196526370942593, + 0.004981114994734526, + 0.007483943365514278, + 0.006439887452870607, + 0.006314766593277454, + 0.008058618754148483, + 0.0049080802127718925, + 0.006125743500888348, + 0.009109271690249443, + 0.006120729260146618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008638394996523857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409122696146369 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.00409122696146369, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409122696146369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11599143594503403, + 0.11236105114221573, + 0.07843068242073059, + 0.09246118366718292, + 0.08410664647817612, + 0.21840395033359528, + 0.042344462126493454, + 0.10216674208641052, + 0.03243756294250488, + 0.12129627168178558 + ], + "credit_map": [ + 0.021560121327638626, + 0.012462937273085117, + 0.021542327478528023, + 0.011582808569073677, + 0.008814145810902119, + 0.020948760211467743, + 0.008587643504142761, + 0.010492689907550812, + 0.013048633001744747, + 0.005604891572147608, + 0.029895750805735588, + 0.0065641580149531364, + 0.006969661917537451, + 0.03878336772322655, + 0.004716681782156229, + 0.0067779021337628365, + 0.011109179817140102, + 0.007728709373623133, + 0.014723026193678379, + 0.010455386713147163, + 0.010357637889683247, + 0.008341587148606777, + 0.006939837243407965, + 0.008775320835411549, + 0.011039421893656254, + 0.010319514200091362, + 0.009730488061904907, + 0.011366854421794415, + 0.015453190542757511, + 0.010901371948421001, + 0.010779853910207748, + 0.012870484963059425, + 0.008331055752933025, + 0.011022466234862804, + 0.011189714074134827, + 0.009023387916386127, + 0.01407594420015812, + 0.008808343671262264, + 0.014772887341678143, + 0.006882844027131796, + 0.060683682560920715, + 0.014915294013917446, + 0.03286636993288994, + 0.017484625801444054, + 0.0163431316614151, + 0.009752880781888962, + 0.02210818976163864, + 0.0442497693002224, + 0.0050762309692800045, + 0.005030983593314886, + 0.004743245430290699, + 0.004815475549548864, + 0.005960495676845312, + 0.0036514829844236374, + 0.008146382868289948, + 0.0049201673828065395, + 0.00935298204421997, + 0.010037287138402462, + 0.00904866587370634, + 0.021790750324726105, + 0.015553299337625504, + 0.008507296442985535, + 0.015707440674304962, + 0.012169023975729942, + 0.004370673559606075, + 0.0031447617802768946, + 0.003944303374737501, + 0.004945292603224516, + 0.004104292951524258, + 0.002636473625898361, + 0.00574662908911705, + 0.0035451350267976522, + 0.011761768721044064, + 0.012116340920329094, + 0.013119274750351906, + 0.018203917890787125, + 0.010734288953244686, + 0.017929211258888245, + 0.025103077292442322, + 0.012328379787504673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.00409122696146369 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static high-angle bird's-eye view throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sequence begin with a tightly curled, spiky deep green bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bud unfurl to reveal a reddish-purple underside before expanding its green topside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the final leaf a massive, fully formed circular pad with raised edges resting on dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny water droplets visible dancing and coalescing on the waxy surface of the leaf?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097530603408813 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097530603408813, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097530603408813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08148089051246643, + 0.03442969173192978, + 0.09610059857368469, + 0.04355185478925705, + 0.03140713647007942, + 0.5274460315704346, + 0.036813247948884964, + 0.07007408887147903, + 0.03914394974708557, + 0.0395524799823761 + ], + "credit_map": [ + 0.019653894007205963, + 0.0076190694235265255, + 0.015073073096573353, + 0.006310936529189348, + 0.006416701711714268, + 0.014284034259617329, + 0.004382710438221693, + 0.00774047477170825, + 0.0043524508364498615, + 0.0026917464565485716, + 0.008891381323337555, + 0.0025400102604180574, + 0.003498268313705921, + 0.007574174087494612, + 0.002669460605829954, + 0.0022122033406049013, + 0.009462282061576843, + 0.005062888842076063, + 0.009545281529426575, + 0.020705144852399826, + 0.006496241316199303, + 0.03072778694331646, + 0.007108742371201515, + 0.006992239970713854, + 0.007232177536934614, + 0.004530551843345165, + 0.007044222205877304, + 0.004728932399302721, + 0.005779423750936985, + 0.0049314540810883045, + 0.003568124258890748, + 0.00573697267100215, + 0.005106172524392605, + 0.003580079646781087, + 0.004207157529890537, + 0.0035014713648706675, + 0.005830152425915003, + 0.0026852001901715994, + 0.0038298796862363815, + 0.0026670247316360474, + 0.13703282177448273, + 0.03156759962439537, + 0.04835203289985657, + 0.05125518888235092, + 0.0479799248278141, + 0.021497465670108795, + 0.07221544533967972, + 0.11754553765058517, + 0.005036163609474897, + 0.003023371798917651, + 0.007104568649083376, + 0.0037081080954521894, + 0.003677471773698926, + 0.0037735707592219114, + 0.0054956646636128426, + 0.004994327202439308, + 0.009023143909871578, + 0.008300036191940308, + 0.007638996932655573, + 0.01303800754249096, + 0.010188030079007149, + 0.006091068498790264, + 0.009041721932590008, + 0.006753089372068644, + 0.004082845989614725, + 0.007801201660186052, + 0.004961919505149126, + 0.005704226903617382, + 0.003931713290512562, + 0.003404778428375721, + 0.004743786994367838, + 0.004513476509600878, + 0.00576868886128068, + 0.004284977912902832, + 0.00486705731600523, + 0.004645098466426134, + 0.004215957596898079, + 0.005140767898410559, + 0.006968297529965639, + 0.003661637892946601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.038097530603408813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_5/index_tensor([5985, 5985])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_5/index_tensor([5985, 5985])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9a5b4c19e98063ee464ecbc24ea57961ddca5ad9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_5/index_tensor([5985, 5985])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a motorcyclist clad in a black, high-tech armored suit, its carbon fiber joints gleaming with faint circuitry. A horizontal streak of neon blue light emanates from the dark visor of his full-face helmet. He's at the helm of an angular, futuristic black motorcycle, accelerating fiercely through a world saturated with digital noise. His wheels kick up not dust but pixelated particles gleaming with data errors. His surroundings are a vast ruined landscape, the sky a dark static field of pulsing green and purple data streams, the ground a blend of cracked asphalt and a digital ground that periodically dissolves into voxel blocks. As the camera follows closely and tilts upward, the rider decisively charges toward a crumbling ancient wall composed of massive, weathered stone blocks, their ancient surfaces digitally eroded, their textures shifting between stone and garbled text. Using a sloping cliff as a jumping platform, the motorcycle leaps into the air, carving a sharp arc against the chaotic digital sky, its body surfacing in an elegant and powerful silhouette at its peak. At this moment, the entire picture is occasionally torn apart by strong glitch art effects, with horizontal scan lines, color channel separation and brief mosaics appearing, as if the signal interference of this broken world, further intensifying the tense, psychedelic and surreal visual impact brought about by this high-speed leap between ancient ruins and future collapse.", + "seed": 13880, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_5/index_5985_seed_13880_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011362335528247058, + "vlm_logprob": -0.011362335528247058, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.011362335528247058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016160450875759125 + ], + "opposite_logprob_tokens": [ + -4.391160488128662 + ], + "logprob_reward": -0.016160450875759125, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016160450875759125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04166843742132187, + 0.01318461075425148, + 0.02713286690413952, + 0.04620315507054329, + 0.2904720902442932, + 0.024786151945590973, + 0.02100282348692417, + 0.03095114789903164, + 0.4035671651363373, + 0.1010316014289856 + ], + "credit_map": [ + 0.005653184838593006, + 0.006678829435259104, + 0.0039335316978394985, + 0.0038315497804433107, + 0.0064148553647100925, + 0.003808875335380435, + 0.005844666622579098, + 0.005502941552549601, + 0.001699602697044611, + 0.0016347228083759546, + 0.0018755303462967277, + 0.0012126414803788066, + 0.0017478065565228462, + 0.0021472552325576544, + 0.001170956646092236, + 0.0016960956854745746, + 0.0021597566083073616, + 0.0033610749524086714, + 0.0030381788965314627, + 0.0032000334467738867, + 0.0039389897137880325, + 0.0034676839131861925, + 0.004409619141370058, + 0.0035575279034674168, + 0.007957328110933304, + 0.0059733022935688496, + 0.003706873394548893, + 0.005633113905787468, + 0.006232772022485733, + 0.008306908421218395, + 0.005125750321894884, + 0.003267107531428337, + 0.04302091524004936, + 0.030239364132285118, + 0.07010255753993988, + 0.04056045040488243, + 0.017278244718909264, + 0.022446321323513985, + 0.03208518773317337, + 0.034739039838314056, + 0.0031511487904936075, + 0.00290509476326406, + 0.0034191536251455545, + 0.0035032122395932674, + 0.0031696094665676355, + 0.0032048297580331564, + 0.003029887331649661, + 0.0024032164365053177, + 0.0026511549949645996, + 0.002597094513475895, + 0.0028964069206267595, + 0.002164829056710005, + 0.0017889583250507712, + 0.0027912408113479614, + 0.0032581030391156673, + 0.002855033613741398, + 0.003159204963594675, + 0.005308101419359446, + 0.004140131175518036, + 0.002665355335921049, + 0.004591585136950016, + 0.0029291845858097076, + 0.004261253867298365, + 0.003896331414580345, + 0.016757700592279434, + 0.03011777065694332, + 0.15765279531478882, + 0.02515379711985588, + 0.017065972089767456, + 0.12309170514345169, + 0.02076268196105957, + 0.012964696623384953, + 0.006239382084459066, + 0.0052724420092999935, + 0.00316395191475749, + 0.003919514827430248, + 0.007914218120276928, + 0.011732610873878002, + 0.03445974737405777, + 0.02832973375916481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.016160450875759125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001049682730808854 + ], + "opposite_logprob_tokens": [ + -6.876049518585205 + ], + "logprob_reward": -0.001049682730808854, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001049682730808854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10432733595371246, + 0.06171790882945061, + 0.04530800133943558, + 0.08164574205875397, + 0.1202160194516182, + 0.028997572138905525, + 0.06526171416044235, + 0.06995362788438797, + 0.3449499309062958, + 0.07762213051319122 + ], + "credit_map": [ + 0.014593171887099743, + 0.016846448183059692, + 0.0100395567715168, + 0.010491831228137016, + 0.015054522082209587, + 0.009895972907543182, + 0.015969332307577133, + 0.011436501517891884, + 0.0053524477407336235, + 0.015328705310821533, + 0.00722226407378912, + 0.005397581495344639, + 0.010207434184849262, + 0.007069116923958063, + 0.004854848142713308, + 0.00628550723195076, + 0.003774560522288084, + 0.007561309263110161, + 0.004844330716878176, + 0.00574758043512702, + 0.007403576746582985, + 0.006490066181868315, + 0.005797366611659527, + 0.0036892108619213104, + 0.01361522264778614, + 0.011625140905380249, + 0.005735538899898529, + 0.0085885776206851, + 0.009502814151346684, + 0.019169291481375694, + 0.007985022850334644, + 0.005424125120043755, + 0.01806095615029335, + 0.01097879558801651, + 0.02606084942817688, + 0.014515146613121033, + 0.013187509961426258, + 0.011065896600484848, + 0.012682382017374039, + 0.013664476573467255, + 0.002505349460989237, + 0.00525396503508091, + 0.002060077851638198, + 0.004372015129774809, + 0.005997133441269398, + 0.0030093875247985125, + 0.0028160742949694395, + 0.002983570797368884, + 0.010455813258886337, + 0.006628798320889473, + 0.011657466180622578, + 0.010650613345205784, + 0.00467453571036458, + 0.005173027515411377, + 0.007535239215940237, + 0.008486224338412285, + 0.006255859509110451, + 0.010858302935957909, + 0.005181606858968735, + 0.0035377987660467625, + 0.005605341866612434, + 0.02026979625225067, + 0.004168586805462837, + 0.014076336286962032, + 0.013785974122583866, + 0.026183025911450386, + 0.13534966111183167, + 0.023183461278676987, + 0.016992701217532158, + 0.09969016164541245, + 0.018893884494900703, + 0.010871071368455887, + 0.004536072723567486, + 0.0040564690716564655, + 0.0023546938318759203, + 0.003271569265052676, + 0.004758070223033428, + 0.008901014924049377, + 0.02762109786272049, + 0.02212313935160637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.001049682730808854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787445440888405 + ], + "opposite_logprob_tokens": [ + -3.529787540435791 + ], + "logprob_reward": -0.029787445440888405, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787445440888405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07472603768110275, + 0.030174344778060913, + 0.040235623717308044, + 0.031462300568819046, + 0.1872517466545105, + 0.031874772161245346, + 0.05257305130362511, + 0.1688053458929062, + 0.31704553961753845, + 0.06585120409727097 + ], + "credit_map": [ + 0.01166333258152008, + 0.011298660188913345, + 0.007837588898837566, + 0.006425820756703615, + 0.009489119052886963, + 0.008755755610764027, + 0.009971978142857552, + 0.009283789433538914, + 0.002383411629125476, + 0.005655550863593817, + 0.002862482098862529, + 0.0025269226171076298, + 0.005105513148009777, + 0.006301043555140495, + 0.002157963113859296, + 0.0031814593821763992, + 0.002831430174410343, + 0.0046251011081039906, + 0.004172376357018948, + 0.004352551884949207, + 0.004972236230969429, + 0.008280782029032707, + 0.007139052264392376, + 0.00386209599673748, + 0.005800230428576469, + 0.004093572963029146, + 0.003199269063770771, + 0.003318590810522437, + 0.003753378288820386, + 0.004875751677900553, + 0.003515235148370266, + 0.002906275214627385, + 0.025943707674741745, + 0.016247110441327095, + 0.04083751142024994, + 0.02298390120267868, + 0.022690847516059875, + 0.01623275876045227, + 0.01996825449168682, + 0.022347668185830116, + 0.0028279474936425686, + 0.003562572179362178, + 0.0033931785728782415, + 0.0038458541966974735, + 0.00610547699034214, + 0.0043653324246406555, + 0.0037342901341617107, + 0.004040124360471964, + 0.006201231386512518, + 0.012916742824018002, + 0.015971602872014046, + 0.004989758133888245, + 0.0033020456321537495, + 0.003140325890854001, + 0.002982700476422906, + 0.0030686459504067898, + 0.01091605145484209, + 0.016713660210371017, + 0.011110547930002213, + 0.007409296929836273, + 0.01078540738672018, + 0.05998962000012398, + 0.007837851531803608, + 0.04404291883111, + 0.015010316856205463, + 0.02644844353199005, + 0.13069239258766174, + 0.023195097222924232, + 0.013147632591426373, + 0.08365408331155777, + 0.015492092818021774, + 0.009405498392879963, + 0.003966219257563353, + 0.003476664423942566, + 0.002787622157484293, + 0.0034365172032266855, + 0.006843768060207367, + 0.007913406938314438, + 0.020985087379813194, + 0.01644192449748516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.029787445440888405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863331276923418 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.00863331276923418, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863331276923418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03754724562168121, + 0.025630107149481773, + 0.026241010054945946, + 0.018491579219698906, + 0.05237759277224541, + 0.07814522087574005, + 0.10340556502342224, + 0.10732769966125488, + 0.4781339466571808, + 0.07270000874996185 + ], + "credit_map": [ + 0.00401509553194046, + 0.006209667772054672, + 0.0034768031910061836, + 0.004008824005723, + 0.006561072543263435, + 0.004064847249537706, + 0.005363985896110535, + 0.003846951061859727, + 0.002757173730060458, + 0.005197709426283836, + 0.0022859424352645874, + 0.00289708748459816, + 0.004593251273036003, + 0.0024710684083402157, + 0.002594950143247843, + 0.0028329237829893827, + 0.0023084261920303106, + 0.003368485253304243, + 0.004588106647133827, + 0.002668040571734309, + 0.0029821419157087803, + 0.0030915301758795977, + 0.004141667392104864, + 0.0030926114413887262, + 0.0019882475025951862, + 0.0024394893553107977, + 0.0023385530803352594, + 0.0022624859120696783, + 0.002512375358492136, + 0.0023068501614034176, + 0.0028376588597893715, + 0.0018059187568724155, + 0.006988089997321367, + 0.006007085554301739, + 0.010860789567232132, + 0.006026402115821838, + 0.0060853720642626286, + 0.004598664585500956, + 0.005104001145809889, + 0.006707186345010996, + 0.006704007275402546, + 0.008756249211728573, + 0.008853434585034847, + 0.011540718376636505, + 0.013012454845011234, + 0.011998766101896763, + 0.00821951124817133, + 0.009060078300535679, + 0.012043326161801815, + 0.028563762083649635, + 0.027279891073703766, + 0.00901503674685955, + 0.005986470729112625, + 0.00614065071567893, + 0.005838352721184492, + 0.008538083173334599, + 0.014545176178216934, + 0.02653714269399643, + 0.014174363575875759, + 0.009328425861895084, + 0.01563110575079918, + 0.008172541856765747, + 0.01121253240853548, + 0.007726415991783142, + 0.019455568864941597, + 0.03400576114654541, + 0.18321849405765533, + 0.030444011092185974, + 0.02491709217429161, + 0.14482839405536652, + 0.025544511154294014, + 0.015720125287771225, + 0.0042375680059194565, + 0.003707998199388385, + 0.003234696574509144, + 0.004335626028478146, + 0.009403123520314693, + 0.008329612202942371, + 0.022644566372036934, + 0.016806814819574356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.00863331276923418 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011807858245447278 + ], + "opposite_logprob_tokens": [ + -6.751180648803711 + ], + "logprob_reward": -0.0011807858245447278, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011807858245447278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08019212633371353, + 0.03760043904185295, + 0.02315928228199482, + 0.06319151073694229, + 0.13079887628555298, + 0.06546054035425186, + 0.15604621171951294, + 0.12162754684686661, + 0.1992127150297165, + 0.12271075695753098 + ], + "credit_map": [ + 0.012438164092600346, + 0.011643655598163605, + 0.007913723587989807, + 0.007570561021566391, + 0.010516712442040443, + 0.008702279068529606, + 0.011953734792768955, + 0.009453294798731804, + 0.002573711099103093, + 0.007687161210924387, + 0.0033030607737600803, + 0.004123811610043049, + 0.006006462499499321, + 0.0063952174969017506, + 0.0035586142912507057, + 0.0039523993618786335, + 0.0018221139907836914, + 0.002579549327492714, + 0.003294105641543865, + 0.0025156871415674686, + 0.0028593854513019323, + 0.0031443736515939236, + 0.004331638105213642, + 0.002612427109852433, + 0.010877334512770176, + 0.009065966121852398, + 0.0046247621066868305, + 0.007816974073648453, + 0.008160512894392014, + 0.013333282433450222, + 0.004924600478261709, + 0.004388079512864351, + 0.018790658563375473, + 0.012928200885653496, + 0.02923133410513401, + 0.019023900851607323, + 0.011898159980773926, + 0.009704369120299816, + 0.014633054845035076, + 0.014589197002351284, + 0.006124383769929409, + 0.010848745703697205, + 0.01013407576829195, + 0.00955355167388916, + 0.010565944947302341, + 0.005577287636697292, + 0.006604406516999006, + 0.0060521396808326244, + 0.016621999442577362, + 0.02755785919725895, + 0.023546913638710976, + 0.023043284192681313, + 0.014076044782996178, + 0.013082761317491531, + 0.020614510402083397, + 0.017502833157777786, + 0.007109637837857008, + 0.0080445921048522, + 0.007148031610995531, + 0.004585120361298323, + 0.008628856390714645, + 0.042392149567604065, + 0.00846331100910902, + 0.035255853086709976, + 0.026957880705595016, + 0.02211243100464344, + 0.03634864464402199, + 0.029240615665912628, + 0.025828953832387924, + 0.037298936396837234, + 0.012232771143317223, + 0.00919246580451727, + 0.006832021288573742, + 0.00626706937327981, + 0.004403967410326004, + 0.007054298650473356, + 0.016272660344839096, + 0.013972875662147999, + 0.037925805896520615, + 0.029982062056660652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.0011807858245447278 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a motorcyclist clad in a black, high-tech armored suit, its carbon fiber joints gleaming with faint circuitry. A horizontal streak of neon blue light emanates from the dark visor of his full-face helmet. He's at the helm of an angular, futuristic black motorcycle, accelerating fiercely through a world saturated with digital noise. His wheels kick up not dust but pixelated particles gleaming with data errors. His surroundings are a vast ruined landscape, the sky a dark static field of pulsing green and purple data streams, the ground a blend of cracked asphalt and a digital ground that periodically dissolves into voxel blocks. As the camera follows closely and tilts upward, the rider decisively charges toward a crumbling ancient wall composed of massive, weathered stone blocks, their ancient surfaces digitally eroded, their textures shifting between stone and garbled text. Using a sloping cliff as a jumping platform, the motorcycle leaps into the air, carving a sharp arc against the chaotic digital sky, its body surfacing in an elegant and powerful silhouette at its peak. At this moment, the entire picture is occasionally torn apart by strong glitch art effects, with horizontal scan lines, color channel separation and brief mosaics appearing, as if the signal interference of this broken world, further intensifying the tense, psychedelic and surreal visual impact brought about by this high-speed leap between ancient ruins and future collapse.", + "seed": 13881, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_5/index_5985_seed_13881_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012573106144554914, + "vlm_logprob": -0.012573106144554914, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.012573106144554914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014494366943836212 + ], + "opposite_logprob_tokens": [ + -4.3894944190979 + ], + "logprob_reward": -0.014494366943836212, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014494366943836212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08666303753852844, + 0.06871303170919418, + 0.05081052705645561, + 0.047177016735076904, + 0.13137003779411316, + 0.40146610140800476, + 0.05966789275407791, + 0.04855291172862053, + 0.07221962511539459, + 0.033359769731760025 + ], + "credit_map": [ + 0.011655779555439949, + 0.0071713742800056934, + 0.013175171799957752, + 0.010516054928302765, + 0.011482774280011654, + 0.009563334286212921, + 0.010974869132041931, + 0.012123685330152512, + 0.010829850099980831, + 0.009693424217402935, + 0.005973489489406347, + 0.006573147606104612, + 0.013465877622365952, + 0.00887307059019804, + 0.005867789499461651, + 0.007436392828822136, + 0.0056176395155489445, + 0.008858525194227695, + 0.005526883527636528, + 0.005704845301806927, + 0.007711182814091444, + 0.00509499991312623, + 0.006330086383968592, + 0.005966373719274998, + 0.004244891926646233, + 0.007395343855023384, + 0.004575666505843401, + 0.004558297339826822, + 0.00520233903080225, + 0.011379341594874859, + 0.0046179331839084625, + 0.005203207489103079, + 0.012909946031868458, + 0.01685907132923603, + 0.013755363412201405, + 0.010943938046693802, + 0.021478183567523956, + 0.021885722875595093, + 0.01762799173593521, + 0.015909822657704353, + 0.027888676151633263, + 0.06698815524578094, + 0.0395335927605629, + 0.041514575481414795, + 0.039830442517995834, + 0.033912818878889084, + 0.11985204368829727, + 0.0319458469748497, + 0.004466987680643797, + 0.009367820806801319, + 0.005253994837403297, + 0.00622722739353776, + 0.007658052258193493, + 0.006049503106623888, + 0.007860613986849785, + 0.012783695012331009, + 0.00922553613781929, + 0.00337410275824368, + 0.006963891908526421, + 0.005960735492408276, + 0.0029252434615045786, + 0.0060838377103209496, + 0.006833079271018505, + 0.007186491508036852, + 0.009628674015402794, + 0.012917393818497658, + 0.013608875684440136, + 0.008949322625994682, + 0.008281901478767395, + 0.007175512611865997, + 0.005670635029673576, + 0.005987308919429779, + 0.0028343305457383394, + 0.004493760876357555, + 0.0043530091643333435, + 0.004832274280488491, + 0.005038002505898476, + 0.0039705983363091946, + 0.004017215222120285, + 0.003820581827312708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.014494366943836212 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015200738562271 + ], + "opposite_logprob_tokens": [ + -6.501520156860352 + ], + "logprob_reward": -0.0015200738562271, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015200738562271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1032906323671341, + 0.04906390979886055, + 0.0679992064833641, + 0.11362079530954361, + 0.06848468631505966, + 0.2841760516166687, + 0.04586053639650345, + 0.05387603119015694, + 0.1845661848783493, + 0.02906191349029541 + ], + "credit_map": [ + 0.01789030060172081, + 0.008028092794120312, + 0.01711283251643181, + 0.010354222729802132, + 0.010072456672787666, + 0.012717094272375107, + 0.007555243093520403, + 0.01956038922071457, + 0.005152391269803047, + 0.00881043542176485, + 0.0049020228907465935, + 0.005859343335032463, + 0.00844649039208889, + 0.005968464072793722, + 0.00395641103386879, + 0.00596834858879447, + 0.009920859709382057, + 0.011201952584087849, + 0.00636458583176136, + 0.005858522839844227, + 0.007534236181527376, + 0.007477807346731424, + 0.007176158484071493, + 0.012465083040297031, + 0.0182260163128376, + 0.016286801546812057, + 0.01621522568166256, + 0.01083441823720932, + 0.009458071552217007, + 0.016637466847896576, + 0.014855962246656418, + 0.011106833815574646, + 0.006710655055940151, + 0.00899714045226574, + 0.00728776678442955, + 0.006197184789925814, + 0.011206344701349735, + 0.00876940879970789, + 0.011635486967861652, + 0.007680703420192003, + 0.01692112348973751, + 0.07096253335475922, + 0.013637697324156761, + 0.01905474252998829, + 0.022448135539889336, + 0.020744189620018005, + 0.10028356313705444, + 0.020124085247516632, + 0.003667463781312108, + 0.004898684099316597, + 0.003218932542949915, + 0.003568654879927635, + 0.00440299091860652, + 0.003902942407876253, + 0.0048186322674155235, + 0.017382238060235977, + 0.00765258027240634, + 0.006499935872852802, + 0.006927694659680128, + 0.010273803025484085, + 0.006910381838679314, + 0.004844256676733494, + 0.006176508963108063, + 0.004590871278196573, + 0.023336365818977356, + 0.03603445366024971, + 0.029460033401846886, + 0.030511053279042244, + 0.024675529450178146, + 0.010293742641806602, + 0.015258803963661194, + 0.014996207319200039, + 0.0035086949355900288, + 0.0034083451610058546, + 0.005158870946615934, + 0.003979752771556377, + 0.004522462375462055, + 0.0021964525803923607, + 0.003668476128950715, + 0.002618858590722084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.0015200738562271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808823600411415 + ], + "opposite_logprob_tokens": [ + -3.288088321685791 + ], + "logprob_reward": -0.03808823600411415, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808823600411415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10286889225244522, + 0.04645811393857002, + 0.16447940468788147, + 0.06291933357715607, + 0.05525138974189758, + 0.14586660265922546, + 0.10885188728570938, + 0.06041638180613518, + 0.19902637600898743, + 0.05386165529489517 + ], + "credit_map": [ + 0.021230893209576607, + 0.006790203507989645, + 0.01457702461630106, + 0.011561927385628223, + 0.009115850552916527, + 0.015527467243373394, + 0.00951705127954483, + 0.01454847026616335, + 0.006871397607028484, + 0.005878386553376913, + 0.004130984656512737, + 0.004236475098878145, + 0.009483695961534977, + 0.004916409961879253, + 0.004450573120266199, + 0.0064901914447546005, + 0.02260574884712696, + 0.01817876100540161, + 0.013550043106079102, + 0.010844510048627853, + 0.020308995619416237, + 0.02196822687983513, + 0.014905964955687523, + 0.04211714118719101, + 0.007294978015124798, + 0.00988486036658287, + 0.00649411091580987, + 0.006297621410340071, + 0.008481914177536964, + 0.010187511332333088, + 0.00767472991719842, + 0.0066036079078912735, + 0.006120109930634499, + 0.007869387976825237, + 0.005304224323481321, + 0.0059644258581101894, + 0.008486103266477585, + 0.007396661210805178, + 0.008529460057616234, + 0.005581016652286053, + 0.009381795302033424, + 0.03556143119931221, + 0.0072306012734770775, + 0.010644122026860714, + 0.009937937371432781, + 0.01065825019031763, + 0.052287399768829346, + 0.010165054351091385, + 0.008786691352725029, + 0.011525586247444153, + 0.019151289016008377, + 0.014192931354045868, + 0.013620567508041859, + 0.012057160027325153, + 0.018970930948853493, + 0.01054673083126545, + 0.009416080079972744, + 0.007999726571142673, + 0.007425392046570778, + 0.00837931502610445, + 0.008639130741357803, + 0.006252930965274572, + 0.006370347924530506, + 0.005933460779488087, + 0.024704033508896828, + 0.03812868893146515, + 0.029897885397076607, + 0.02992318943142891, + 0.028436359018087387, + 0.012065943330526352, + 0.017116272822022438, + 0.018754014745354652, + 0.006683114916086197, + 0.00875102262943983, + 0.007221759296953678, + 0.007811050862073898, + 0.007110434118658304, + 0.004359462298452854, + 0.0066984654404222965, + 0.005226349923759699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.03808823600411415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944430362433195 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944430362433195, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944430362433195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039463140070438385, + 0.02767811343073845, + 0.04341500997543335, + 0.11776099354028702, + 0.023817671462893486, + 0.23443767428398132, + 0.04175709933042526, + 0.07032755017280579, + 0.2796391248703003, + 0.1217036247253418 + ], + "credit_map": [ + 0.004710403271019459, + 0.003613928100094199, + 0.004702088888734579, + 0.005111991427838802, + 0.006875021383166313, + 0.0056632887572050095, + 0.005003589671105146, + 0.003782827639952302, + 0.0027986003551632166, + 0.004143696743994951, + 0.0024954623077064753, + 0.0028211656026542187, + 0.00663012033328414, + 0.0029897813219577074, + 0.0025970791466534138, + 0.003202208550646901, + 0.003920027054846287, + 0.006605267990380526, + 0.004238474648445845, + 0.006622347515076399, + 0.005125894211232662, + 0.00440696207806468, + 0.008063000626862049, + 0.004433034919202328, + 0.01749338209629059, + 0.020134486258029938, + 0.018002185970544815, + 0.011669708415865898, + 0.011828395538032055, + 0.015150552615523338, + 0.014546369202435017, + 0.00893590971827507, + 0.003958486020565033, + 0.002943414496257901, + 0.0029345545917749405, + 0.0019423033809289336, + 0.0033876041416078806, + 0.0033248397521674633, + 0.0027285863179713488, + 0.0025978826452046633, + 0.013658150099217892, + 0.06408233940601349, + 0.006360014900565147, + 0.013633733615279198, + 0.01552502065896988, + 0.01797565445303917, + 0.08572578430175781, + 0.017476974055171013, + 0.00390973687171936, + 0.004061416257172823, + 0.006105710752308369, + 0.004903210327029228, + 0.0052896286360919476, + 0.005217761266976595, + 0.006311172619462013, + 0.005958467721939087, + 0.012760168872773647, + 0.008986424654722214, + 0.011339847929775715, + 0.007386649027466774, + 0.006245786789804697, + 0.009218608029186726, + 0.0071798148564994335, + 0.0072102476842701435, + 0.036709144711494446, + 0.04531428962945938, + 0.05319564789533615, + 0.04395555704832077, + 0.037677865475416183, + 0.017263362184166908, + 0.022772671654820442, + 0.022750599309802055, + 0.014721561223268509, + 0.01645663194358349, + 0.01954381726682186, + 0.0196345504373312, + 0.015961451455950737, + 0.009704955853521824, + 0.016278760507702827, + 0.009401889517903328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.005944430362433195 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818423556163907 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818423556163907, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818423556163907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05229007452726364, + 0.029967889189720154, + 0.03803197294473648, + 0.15933474898338318, + 0.03536958619952202, + 0.06578729301691055, + 0.104290671646595, + 0.061015430837869644, + 0.3903270661830902, + 0.06358525156974792 + ], + "credit_map": [ + 0.007746960036456585, + 0.003912006039172411, + 0.008209923282265663, + 0.0068354252725839615, + 0.005875117145478725, + 0.006625708658248186, + 0.005341700278222561, + 0.007743236143141985, + 0.0037121265195310116, + 0.004682906437665224, + 0.0035048776771873236, + 0.0033390920143574476, + 0.005042407661676407, + 0.0030249704141169786, + 0.0020057924557477236, + 0.004655716475099325, + 0.0048056538216769695, + 0.004040396306663752, + 0.0028913000132888556, + 0.003355704015120864, + 0.004965932574123144, + 0.005135274026542902, + 0.003780433675274253, + 0.009057280607521534, + 0.02160489931702614, + 0.024007046595215797, + 0.02134084329009056, + 0.016065383329987526, + 0.01561751589179039, + 0.023966768756508827, + 0.019562657922506332, + 0.017169641330838203, + 0.0038728867657482624, + 0.0045401970855891705, + 0.0035969121381640434, + 0.0032871889416128397, + 0.004889131989330053, + 0.005238295532763004, + 0.005618652794510126, + 0.004326322581619024, + 0.008282073773443699, + 0.008971085771918297, + 0.006070416886359453, + 0.011108576320111752, + 0.0058812168426811695, + 0.008364029228687286, + 0.011231787502765656, + 0.005878106690943241, + 0.009576437994837761, + 0.006602298002690077, + 0.01007323618978262, + 0.00914765615016222, + 0.010848438367247581, + 0.009515451267361641, + 0.01174682006239891, + 0.036780331283807755, + 0.009759072214365005, + 0.011603412218391895, + 0.00890764594078064, + 0.008182483725249767, + 0.006103317718952894, + 0.003949637990444899, + 0.007351136300712824, + 0.005158721935003996, + 0.04544738680124283, + 0.07686300575733185, + 0.04964182525873184, + 0.06987334787845612, + 0.05387571081519127, + 0.02734031341969967, + 0.03412579372525215, + 0.033159688115119934, + 0.005677434150129557, + 0.009785127826035023, + 0.011405957862734795, + 0.00970056839287281, + 0.009155531413853168, + 0.005009391810745001, + 0.007892666384577751, + 0.004958574194461107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.002818423556163907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a motorcyclist clad in a black, high-tech armored suit, its carbon fiber joints gleaming with faint circuitry. A horizontal streak of neon blue light emanates from the dark visor of his full-face helmet. He's at the helm of an angular, futuristic black motorcycle, accelerating fiercely through a world saturated with digital noise. His wheels kick up not dust but pixelated particles gleaming with data errors. His surroundings are a vast ruined landscape, the sky a dark static field of pulsing green and purple data streams, the ground a blend of cracked asphalt and a digital ground that periodically dissolves into voxel blocks. As the camera follows closely and tilts upward, the rider decisively charges toward a crumbling ancient wall composed of massive, weathered stone blocks, their ancient surfaces digitally eroded, their textures shifting between stone and garbled text. Using a sloping cliff as a jumping platform, the motorcycle leaps into the air, carving a sharp arc against the chaotic digital sky, its body surfacing in an elegant and powerful silhouette at its peak. At this moment, the entire picture is occasionally torn apart by strong glitch art effects, with horizontal scan lines, color channel separation and brief mosaics appearing, as if the signal interference of this broken world, further intensifying the tense, psychedelic and surreal visual impact brought about by this high-speed leap between ancient ruins and future collapse.", + "seed": 13882, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_5/index_5985_seed_13882_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01762098746839911, + "vlm_logprob": -0.01762098746839911, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01762098746839911, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030442385002970695 + ], + "opposite_logprob_tokens": [ + -3.655442476272583 + ], + "logprob_reward": -0.030442385002970695, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030442385002970695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10792333632707596, + 0.062217142432928085, + 0.06956726312637329, + 0.13862444460391998, + 0.065143883228302, + 0.07517105340957642, + 0.13368144631385803, + 0.11065230518579483, + 0.15993723273277283, + 0.07708191871643066 + ], + "credit_map": [ + 0.01794770173728466, + 0.011358947493135929, + 0.006519421935081482, + 0.007872894406318665, + 0.025153620168566704, + 0.019864074885845184, + 0.011885424144566059, + 0.00732125760987401, + 0.009892071597278118, + 0.005410749930888414, + 0.01031013485044241, + 0.0044675967656075954, + 0.009072563610970974, + 0.008830790407955647, + 0.007214422803372145, + 0.007018812466412783, + 0.009646869264543056, + 0.009727921336889267, + 0.00884957890957594, + 0.006498331669718027, + 0.009895969182252884, + 0.013318678364157677, + 0.006637899670749903, + 0.004992015194147825, + 0.020214753225445747, + 0.01736212708055973, + 0.02406180463731289, + 0.012146140448749065, + 0.02265329100191593, + 0.02007257379591465, + 0.009252037853002548, + 0.012861726805567741, + 0.011292495764791965, + 0.00505107594653964, + 0.008863447234034538, + 0.006197082810103893, + 0.009047240018844604, + 0.01055910810828209, + 0.007095694541931152, + 0.007037733215838671, + 0.010708292946219444, + 0.005625337362289429, + 0.008871922269463539, + 0.010897071100771427, + 0.01353665255010128, + 0.010207031853497028, + 0.0074911657720804214, + 0.007833580486476421, + 0.02692274935543537, + 0.012505301274359226, + 0.020694831386208534, + 0.014598905108869076, + 0.016366617754101753, + 0.022770851850509644, + 0.008510048501193523, + 0.011312149465084076, + 0.022695662453770638, + 0.010869450867176056, + 0.012929622083902359, + 0.012979577295482159, + 0.012801741249859333, + 0.012739053927361965, + 0.010532550513744354, + 0.01510464958846569, + 0.025200584903359413, + 0.01770426705479622, + 0.03513232618570328, + 0.018385272473096848, + 0.01870102807879448, + 0.01226743869483471, + 0.013576207682490349, + 0.018970103934407234, + 0.006146831437945366, + 0.009631718508899212, + 0.006703171879053116, + 0.010397172532975674, + 0.012838807888329029, + 0.009304170496761799, + 0.011048459447920322, + 0.011011584661900997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.030442385002970695 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013465156080201268 + ], + "opposite_logprob_tokens": [ + -6.626346588134766 + ], + "logprob_reward": -0.0013465156080201268, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013465156080201268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0921582505106926, + 0.07476806640625, + 0.0710562989115715, + 0.08233289420604706, + 0.14562132954597473, + 0.0261479914188385, + 0.08603058010339737, + 0.2902982831001282, + 0.060394659638404846, + 0.07119159400463104 + ], + "credit_map": [ + 0.01598515547811985, + 0.012993157841265202, + 0.004783549811691046, + 0.008328499272465706, + 0.021401183679699898, + 0.014212978072464466, + 0.007734759245067835, + 0.0067189643159508705, + 0.012068458832800388, + 0.006831525359302759, + 0.009081303142011166, + 0.005985150579363108, + 0.01564255729317665, + 0.00993076991289854, + 0.009640115313231945, + 0.005588187836110592, + 0.010578084737062454, + 0.007857635617256165, + 0.010501647368073463, + 0.006946289446204901, + 0.010505483485758305, + 0.011880713514983654, + 0.0070640891790390015, + 0.00572235370054841, + 0.011931238695979118, + 0.009891897439956665, + 0.01416971255093813, + 0.006018154323101044, + 0.013408002443611622, + 0.01066402904689312, + 0.00660221092402935, + 0.009647644124925137, + 0.02190626598894596, + 0.0076228962279856205, + 0.023985590785741806, + 0.013757188804447651, + 0.01680133305490017, + 0.027217481285333633, + 0.016559554263949394, + 0.017771020531654358, + 0.0043313829228281975, + 0.0019784460309892893, + 0.003116015577688813, + 0.003438354004174471, + 0.003574973903596401, + 0.00398944690823555, + 0.0028261144179850817, + 0.002893258584663272, + 0.014498136006295681, + 0.00847384799271822, + 0.01371489092707634, + 0.011045861057937145, + 0.01000258419662714, + 0.012684915214776993, + 0.007452483754605055, + 0.00815785862505436, + 0.0671389251947403, + 0.03458590433001518, + 0.034063927829265594, + 0.02933746762573719, + 0.027369096875190735, + 0.032247595489025116, + 0.024418380111455917, + 0.041136983782052994, + 0.008314925245940685, + 0.005717077292501926, + 0.009320608340203762, + 0.009926301427185535, + 0.0056830174289643764, + 0.0058832610957324505, + 0.007792972028255463, + 0.007756496779620647, + 0.006829957012087107, + 0.009677075780928135, + 0.005476821679621935, + 0.007853803224861622, + 0.011167905293405056, + 0.01451385859400034, + 0.007685045711696148, + 0.007987122051417828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0013465156080201268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304901137948036 + ], + "opposite_logprob_tokens": [ + -3.1680490970611572 + ], + "logprob_reward": -0.04304901137948036, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304901137948036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08698587864637375, + 0.14017809927463531, + 0.07354691624641418, + 0.11788084357976913, + 0.0857291892170906, + 0.05860547721385956, + 0.10035127401351929, + 0.1546112596988678, + 0.10298772156238556, + 0.0791233628988266 + ], + "credit_map": [ + 0.01235730480402708, + 0.010502660647034645, + 0.0074714901857078075, + 0.00813406053930521, + 0.02061905898153782, + 0.012403671629726887, + 0.009898662567138672, + 0.0055989595130085945, + 0.02409655787050724, + 0.01350030954927206, + 0.017149237915873528, + 0.009113547392189503, + 0.02053593471646309, + 0.025443730875849724, + 0.015329653397202492, + 0.01500911358743906, + 0.01251451950520277, + 0.0068220277316868305, + 0.011095001362264156, + 0.003680221736431122, + 0.011741472408175468, + 0.017087751999497414, + 0.006081533618271351, + 0.0045243860222399235, + 0.01814202219247818, + 0.0145764434710145, + 0.0197847131639719, + 0.008970996364951134, + 0.012847119010984898, + 0.019304154440760612, + 0.009994303807616234, + 0.01426109578460455, + 0.013430265709757805, + 0.005447562783956528, + 0.014690708369016647, + 0.010168846696615219, + 0.014711876399815083, + 0.011866413056850433, + 0.008256680332124233, + 0.007156841456890106, + 0.010089893825352192, + 0.005454958416521549, + 0.011261495761573315, + 0.007282469421625137, + 0.006586591247469187, + 0.007484743371605873, + 0.0058542038314044476, + 0.004591119010001421, + 0.013967500068247318, + 0.008453347720205784, + 0.014406118541955948, + 0.015100231394171715, + 0.02000221237540245, + 0.012128110975027084, + 0.00754641043022275, + 0.008747345767915249, + 0.039394866675138474, + 0.01703133061528206, + 0.018562929704785347, + 0.013870250433683395, + 0.013454659841954708, + 0.016008490696549416, + 0.013287021778523922, + 0.023001715540885925, + 0.013114667497575283, + 0.013402283191680908, + 0.017444370314478874, + 0.014665654860436916, + 0.013623884879052639, + 0.009428592398762703, + 0.00999735202640295, + 0.011310916393995285, + 0.006132323760539293, + 0.010464318096637726, + 0.006223016418516636, + 0.011186745949089527, + 0.011492906138300896, + 0.011043344624340534, + 0.009414228610694408, + 0.013166481629014015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.04304901137948036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106967218220234 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.01106967218220234, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106967218220234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05698335915803909, + 0.05870024487376213, + 0.02987929992377758, + 0.07484422624111176, + 0.04039212688803673, + 0.045692458748817444, + 0.14688964188098907, + 0.07501238584518433, + 0.23214945197105408, + 0.23945674300193787 + ], + "credit_map": [ + 0.008265906944870949, + 0.007945165038108826, + 0.005821166560053825, + 0.00691984174773097, + 0.009291213005781174, + 0.007452401332557201, + 0.006830171681940556, + 0.004457498900592327, + 0.009508405812084675, + 0.005498835816979408, + 0.007347120437771082, + 0.003926550038158894, + 0.009173635393381119, + 0.009449987672269344, + 0.0062085590325295925, + 0.00758716044947505, + 0.004739928059279919, + 0.0030865769367665052, + 0.003958277869969606, + 0.0025356123223900795, + 0.004026877693831921, + 0.00714454147964716, + 0.0026759421452879906, + 0.001711545861326158, + 0.01174909807741642, + 0.00852181389927864, + 0.013007847592234612, + 0.005645955912768841, + 0.007856757380068302, + 0.014353391714394093, + 0.0066691054962575436, + 0.007040258962661028, + 0.0053915358148515224, + 0.002851803321391344, + 0.007802671752870083, + 0.004253210965543985, + 0.006093503907322884, + 0.005242560990154743, + 0.004083750303834677, + 0.004673096816986799, + 0.006240621209144592, + 0.003463535336777568, + 0.006460242904722691, + 0.004940483719110489, + 0.006695502903312445, + 0.008342897519469261, + 0.0038576885126531124, + 0.005691490136086941, + 0.02360452339053154, + 0.014546002261340618, + 0.02805675007402897, + 0.020358992740511894, + 0.02074531465768814, + 0.015081614255905151, + 0.010171161033213139, + 0.01432530302554369, + 0.01203206554055214, + 0.009609059430658817, + 0.008186954073607922, + 0.009326995350420475, + 0.010617322288453579, + 0.008325982838869095, + 0.008289335295557976, + 0.008624686859548092, + 0.03146304190158844, + 0.033018339425325394, + 0.033478993922472, + 0.040556374937295914, + 0.031806305050849915, + 0.0131223788484931, + 0.025750285014510155, + 0.022953759878873825, + 0.02066401019692421, + 0.03238306939601898, + 0.021364351734519005, + 0.024188408628106117, + 0.039543669670820236, + 0.028553856536746025, + 0.03133585676550865, + 0.04142357036471367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.01106967218220234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09720584750175476, + 0.043632231652736664, + 0.04328353330492973, + 0.02931305766105652, + 0.045085255056619644, + 0.02289739064872265, + 0.12834402918815613, + 0.21283112466335297, + 0.28037458658218384, + 0.09703300893306732 + ], + "credit_map": [ + 0.014413628727197647, + 0.010306674055755138, + 0.006130574736744165, + 0.0075059207156300545, + 0.026613175868988037, + 0.015110519714653492, + 0.011575303971767426, + 0.005550042726099491, + 0.007313400506973267, + 0.003775431774556637, + 0.005787943489849567, + 0.003520699916407466, + 0.007521983236074448, + 0.007989788427948952, + 0.0038996259681880474, + 0.0038233541417866945, + 0.006987887900322676, + 0.00599804800003767, + 0.005912846885621548, + 0.004891615826636553, + 0.005688577890396118, + 0.006276839412748814, + 0.004221889656037092, + 0.003305827034637332, + 0.004429933615028858, + 0.003171475138515234, + 0.0053351870737969875, + 0.0025313610676676035, + 0.0037561294157058, + 0.005219135899096727, + 0.0023772725835442543, + 0.002492558676749468, + 0.0072745890356600285, + 0.0035564503632485867, + 0.005825459025800228, + 0.00528482673689723, + 0.007313018199056387, + 0.006525566801428795, + 0.004121179226785898, + 0.005184161942452192, + 0.00415871711447835, + 0.0017747587990015745, + 0.00256602605804801, + 0.0024562967009842396, + 0.002664579777047038, + 0.0035743936896324158, + 0.0025753402151167393, + 0.0031272757332772017, + 0.023099878802895546, + 0.015393441542983055, + 0.019911503419280052, + 0.01931728608906269, + 0.021366914734244347, + 0.00992906466126442, + 0.008153142407536507, + 0.011172771453857422, + 0.051050085574388504, + 0.023859219625592232, + 0.02570062130689621, + 0.02003472112119198, + 0.019091680645942688, + 0.021506456658244133, + 0.018393581733107567, + 0.03319472447037697, + 0.0377294197678566, + 0.05523385480046272, + 0.045148152858018875, + 0.03734226152300835, + 0.01720421575009823, + 0.030257247388362885, + 0.02037471905350685, + 0.03708470240235329, + 0.00903245061635971, + 0.011072111316025257, + 0.0064897737465798855, + 0.007695040199905634, + 0.017973802983760834, + 0.010925122536718845, + 0.01497562788426876, + 0.01886906288564205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.002197353169322014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a motorcyclist clad in a black, high-tech armored suit, its carbon fiber joints gleaming with faint circuitry. A horizontal streak of neon blue light emanates from the dark visor of his full-face helmet. He's at the helm of an angular, futuristic black motorcycle, accelerating fiercely through a world saturated with digital noise. His wheels kick up not dust but pixelated particles gleaming with data errors. His surroundings are a vast ruined landscape, the sky a dark static field of pulsing green and purple data streams, the ground a blend of cracked asphalt and a digital ground that periodically dissolves into voxel blocks. As the camera follows closely and tilts upward, the rider decisively charges toward a crumbling ancient wall composed of massive, weathered stone blocks, their ancient surfaces digitally eroded, their textures shifting between stone and garbled text. Using a sloping cliff as a jumping platform, the motorcycle leaps into the air, carving a sharp arc against the chaotic digital sky, its body surfacing in an elegant and powerful silhouette at its peak. At this moment, the entire picture is occasionally torn apart by strong glitch art effects, with horizontal scan lines, color channel separation and brief mosaics appearing, as if the signal interference of this broken world, further intensifying the tense, psychedelic and surreal visual impact brought about by this high-speed leap between ancient ruins and future collapse.", + "seed": 13883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_5/index_5985_seed_13883_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03954555112868548, + "vlm_logprob": -0.03954555112868548, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03954555112868548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13440120220184326 + ], + "opposite_logprob_tokens": [ + -2.134401321411133 + ], + "logprob_reward": -0.13440120220184326, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13440120220184326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1318328082561493, + 0.07889380306005478, + 0.23959824442863464, + 0.3176283538341522, + 0.02706475928425789, + 0.05039847642183304, + 0.016636844724416733, + 0.03055398166179657, + 0.020487423986196518, + 0.0869053304195404 + ], + "credit_map": [ + 0.021138722077012062, + 0.02942892722785473, + 0.011229370720684528, + 0.030102061107754707, + 0.01094784215092659, + 0.009398086927831173, + 0.008530037477612495, + 0.011057756841182709, + 0.011885781772434711, + 0.01037769578397274, + 0.011049647815525532, + 0.006830698810517788, + 0.007248037960380316, + 0.017017539590597153, + 0.006025334820151329, + 0.008459064178168774, + 0.04299144074320793, + 0.03286734223365784, + 0.013514244928956032, + 0.01794753037393093, + 0.04402685537934303, + 0.05087791383266449, + 0.020969806239008904, + 0.016403106972575188, + 0.023241255432367325, + 0.025511668995022774, + 0.07092507928609848, + 0.019967172294855118, + 0.03134901449084282, + 0.10235367715358734, + 0.013798721134662628, + 0.030481768772006035, + 0.004103630315512419, + 0.0026749316602945328, + 0.001953973202034831, + 0.0030726075638085604, + 0.004298929590731859, + 0.002016290556639433, + 0.004922239109873772, + 0.004022156819701195, + 0.0067488690838217735, + 0.007109809201210737, + 0.005298367701470852, + 0.005021062679588795, + 0.010980003513395786, + 0.004558860789984465, + 0.0058297025971114635, + 0.004851803183555603, + 0.0026002335362136364, + 0.002090511377900839, + 0.001996441977098584, + 0.0013087574625387788, + 0.0018214721931144595, + 0.0015109461965039372, + 0.001842685043811798, + 0.003465797519311309, + 0.0037741647101938725, + 0.005365883931517601, + 0.003161235013976693, + 0.002797904424369335, + 0.004907037131488323, + 0.00404071481898427, + 0.0037132606375962496, + 0.002793779131025076, + 0.0025539242196828127, + 0.002348950831219554, + 0.001978949410840869, + 0.0021413248032331467, + 0.004075706470757723, + 0.0023659085854887962, + 0.002208692254498601, + 0.0028139690402895212, + 0.012592654675245285, + 0.016157275065779686, + 0.017925959080457687, + 0.020952414721250534, + 0.004330792464315891, + 0.004482951946556568, + 0.007074360735714436, + 0.003388922894373536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.13440120220184326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004112834110856056 + ], + "opposite_logprob_tokens": [ + -5.504112720489502 + ], + "logprob_reward": -0.004112834110856056, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004112834110856056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08179526031017303, + 0.12808328866958618, + 0.3213798999786377, + 0.25142836570739746, + 0.0353841669857502, + 0.05527851730585098, + 0.01953337900340557, + 0.023037327453494072, + 0.031788382679224014, + 0.052291374653577805 + ], + "credit_map": [ + 0.012050394900143147, + 0.0177827849984169, + 0.007346761878579855, + 0.018450960516929626, + 0.007401190232485533, + 0.0060162669979035854, + 0.005738044623285532, + 0.007008863613009453, + 0.019336897879838943, + 0.016961239278316498, + 0.01877463236451149, + 0.011390731669962406, + 0.01197660993784666, + 0.026310764253139496, + 0.009397276677191257, + 0.0139351487159729, + 0.05808866024017334, + 0.045571811497211456, + 0.01911154016852379, + 0.02400979772210121, + 0.05408311262726784, + 0.07093508541584015, + 0.027056599035859108, + 0.02252330258488655, + 0.018590722233057022, + 0.019004687666893005, + 0.05478017404675484, + 0.014926424250006676, + 0.023065127432346344, + 0.08574939519166946, + 0.012671960517764091, + 0.022639870643615723, + 0.005324079655110836, + 0.003756071673706174, + 0.0036673916038125753, + 0.0043035345152020454, + 0.004035025369375944, + 0.003146826522424817, + 0.006334356032311916, + 0.004816885571926832, + 0.007231784984469414, + 0.008352032862603664, + 0.006542662624269724, + 0.005187662784010172, + 0.011874250136315823, + 0.006343106739223003, + 0.005369692575186491, + 0.004377329722046852, + 0.0025962747167795897, + 0.0020318401511758566, + 0.0020369922276586294, + 0.0015477582346647978, + 0.003011850407347083, + 0.0018180535407736897, + 0.0020229967776685953, + 0.0044676135294139385, + 0.0041696554981172085, + 0.00462712999433279, + 0.002073929412290454, + 0.0024367168080061674, + 0.0030920107383280993, + 0.0030250269919633865, + 0.001882784767076373, + 0.0017300740582868457, + 0.004142590798437595, + 0.004576792009174824, + 0.0030543920584023, + 0.004233645740896463, + 0.00470451544970274, + 0.003895313013345003, + 0.0032889507710933685, + 0.003892183303833008, + 0.007739388383924961, + 0.008573747240006924, + 0.011196763254702091, + 0.009456885978579521, + 0.0030576596036553383, + 0.003779858583584428, + 0.005593869835138321, + 0.0028932050336152315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.004112834110856056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430586040019989 + ], + "opposite_logprob_tokens": [ + -3.1680586338043213 + ], + "logprob_reward": -0.0430586040019989, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430586040019989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009458920918405056, + 0.12325246632099152, + 0.053972277790308, + 0.687365710735321, + 0.013958314433693886, + 0.034905869513750076, + 0.008770040236413479, + 0.010277699679136276, + 0.011848405934870243, + 0.046190280467271805 + ], + "credit_map": [ + 0.0013212114572525024, + 0.0014499312965199351, + 0.001003425451926887, + 0.0013887217501178384, + 0.0009657911723479629, + 0.0009523945045657456, + 0.0008735377923585474, + 0.0015039072604849935, + 0.0175749771296978, + 0.01605873741209507, + 0.018783489242196083, + 0.011144067160785198, + 0.012681189924478531, + 0.025504188612103462, + 0.008907110430300236, + 0.012598704546689987, + 0.009869745932519436, + 0.007695058360695839, + 0.003218604251742363, + 0.00402280455455184, + 0.008912676945328712, + 0.011838437989354134, + 0.004622234962880611, + 0.003792713861912489, + 0.05029929801821709, + 0.05261125788092613, + 0.15423592925071716, + 0.04084426909685135, + 0.0625181570649147, + 0.23303234577178955, + 0.02966007962822914, + 0.06416435539722443, + 0.001898614107631147, + 0.0011926075676456094, + 0.0014998370315879583, + 0.002146382350474596, + 0.0015546526992693543, + 0.001173084368929267, + 0.002568703144788742, + 0.0019244338618591428, + 0.004197530914098024, + 0.005560262128710747, + 0.003606407903134823, + 0.003592859720811248, + 0.00776535551995039, + 0.003898064838722348, + 0.003526227083057165, + 0.0027591618709266186, + 0.0010079651838168502, + 0.001014988636597991, + 0.0010356177808716893, + 0.0006288684089668095, + 0.0014208406209945679, + 0.0006599267944693565, + 0.0008220543386414647, + 0.002179778413847089, + 0.0017019793158397079, + 0.002058579120784998, + 0.0010514752939343452, + 0.0011523142457008362, + 0.001099065295420587, + 0.0011756642488762736, + 0.0012137015582993627, + 0.0008249205420725048, + 0.0013186342548578978, + 0.0017277825390920043, + 0.0009408670011907816, + 0.0018864177400246263, + 0.0007493471493944526, + 0.0015335568459704518, + 0.0006187925464473665, + 0.0030730084981769323, + 0.008267976343631744, + 0.0051058270037174225, + 0.012149317190051079, + 0.007704666815698147, + 0.0032702875323593616, + 0.003032899461686611, + 0.0034395940601825714, + 0.0032197125256061554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.0430586040019989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08759178966283798, + 0.09024689346551895, + 0.021758785471320152, + 0.36238500475883484, + 0.03769440948963165, + 0.06401891261339188, + 0.028589775785803795, + 0.0650615468621254, + 0.08387625217437744, + 0.1587766408920288 + ], + "credit_map": [ + 0.014798829331994057, + 0.018124639987945557, + 0.006765325553715229, + 0.01614246517419815, + 0.009361602365970612, + 0.007468721829354763, + 0.006810526363551617, + 0.008119680918753147, + 0.013168547302484512, + 0.012765731662511826, + 0.01287261862307787, + 0.008418435230851173, + 0.00910684373229742, + 0.01939954236149788, + 0.0055884686298668385, + 0.008926703594624996, + 0.0038259425200521946, + 0.0034287313465029, + 0.002311925869435072, + 0.0022423581685870886, + 0.003703288035467267, + 0.003781278384849429, + 0.001402537920512259, + 0.001062722527422011, + 0.026970308274030685, + 0.028137940913438797, + 0.07634810358285904, + 0.02254381589591503, + 0.033745814114809036, + 0.1223851665854454, + 0.019477227702736855, + 0.03277663514018059, + 0.005998819135129452, + 0.003510870039463043, + 0.003985649440437555, + 0.005376485176384449, + 0.003710022196173668, + 0.003214010503143072, + 0.00681654829531908, + 0.005082007497549057, + 0.007629204075783491, + 0.009896926581859589, + 0.006830750498920679, + 0.006711302790790796, + 0.013291511684656143, + 0.006582408212125301, + 0.007205420173704624, + 0.005871388129889965, + 0.00427560368552804, + 0.00456354720517993, + 0.002985564526170492, + 0.0032907489221543074, + 0.0028594518080353737, + 0.0031179755460470915, + 0.0036267603281885386, + 0.003870123764500022, + 0.007676332723349333, + 0.01109126303344965, + 0.008520145900547504, + 0.007832201197743416, + 0.0060555110685527325, + 0.007884932681918144, + 0.007734810933470726, + 0.008266345597803593, + 0.011620628647506237, + 0.010810954496264458, + 0.007857074961066246, + 0.010156570002436638, + 0.009961419738829136, + 0.010800854302942753, + 0.010286826640367508, + 0.012381917797029018, + 0.027393434196710587, + 0.02529938332736492, + 0.04017426073551178, + 0.04069390892982483, + 0.005425895564258099, + 0.004557058215141296, + 0.009697497822344303, + 0.00553519930690527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.012538178823888302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a motorcyclist wearing a black high-tech armored suit with carbon fiber joints and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle angular, futuristic, and black, ridden by the motorcyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the environment feature a vast ruined landscape with a dark static sky of pulsing green and purple data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the rider closely and tilt upward as the motorcycle leaps into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are strong glitch art effects visible, including horizontal scan lines, color channel separation, and mosaics?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616936504840851 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616936504840851, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616936504840851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0646456703543663, + 0.025583140552043915, + 0.0229097418487072, + 0.21321691572666168, + 0.05744009092450142, + 0.04641152173280716, + 0.03206096217036247, + 0.05976070091128349, + 0.07697100937366486, + 0.4010002017021179 + ], + "credit_map": [ + 0.011161988601088524, + 0.009737581014633179, + 0.007968957535922527, + 0.010914294049143791, + 0.008533027954399586, + 0.005047050304710865, + 0.006548329722136259, + 0.0047344365157186985, + 0.0035953836049884558, + 0.004600528161972761, + 0.0025342553853988647, + 0.0024463438894599676, + 0.002770233666524291, + 0.004677072167396545, + 0.0021092405077069998, + 0.002850083401426673, + 0.0035716614220291376, + 0.003811259288340807, + 0.002148168161511421, + 0.002045349683612585, + 0.004816813860088587, + 0.0033245128579437733, + 0.002081753918901086, + 0.0011102206772193313, + 0.015793392434716225, + 0.014934595674276352, + 0.050440575927495956, + 0.01241842471063137, + 0.0178309865295887, + 0.07250436395406723, + 0.008860813453793526, + 0.02043377235531807, + 0.006930130068212748, + 0.006359181832522154, + 0.005281305406242609, + 0.008851218968629837, + 0.00745006836950779, + 0.0053774467669427395, + 0.009835703298449516, + 0.007355034817010164, + 0.006264287047088146, + 0.005648378748446703, + 0.0056114220060408115, + 0.005196841899305582, + 0.007788931485265493, + 0.005433143116533756, + 0.0045589120127260685, + 0.005909606348723173, + 0.005963140167295933, + 0.003572839079424739, + 0.002419725526124239, + 0.0036119972355663776, + 0.005625264719128609, + 0.0032614071387797594, + 0.0035907751880586147, + 0.004015814047306776, + 0.008349896408617496, + 0.010451906360685825, + 0.004647466354072094, + 0.0059927720576524734, + 0.010275688022375107, + 0.007821138948202133, + 0.00789606012403965, + 0.004325771704316139, + 0.009878164157271385, + 0.00916067324578762, + 0.010174436494708061, + 0.0134449927136302, + 0.00900339987128973, + 0.0075190006755292416, + 0.008893053978681564, + 0.008897294290363789, + 0.06928611546754837, + 0.0498720221221447, + 0.09677672386169434, + 0.076491579413414, + 0.027528060600161552, + 0.021610256284475327, + 0.03474792465567589, + 0.024687469005584717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.003616936504840851 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_tensor([988, 988])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_tensor([988, 988])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5bfe04623d9a1340aa26f6234f09ebadb6973c13 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_tensor([988, 988])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed tracking shot, low to the ground, hurtles forward at dizzying speed along two parallel steel rails. In the golden late afternoon sun, the cold rails reflect a warm, dazzling glow, stretching endlessly into the distance before converging at a vanishing point on the horizon. Below the camera, the dark wooden sleepers and the gravel ballast between them transform into blurred, dynamic streaks, swiftly sweeping backwards, their individual forms barely discernible, leaving only a strong trail of movement. On either side of the tracks, withered grass sway gently in the breeze, bringing a touch of vitality to the vast, somewhat desolate landscape. The entire scene is imbued with a powerful sense of movement and visual impact. Through the extreme depiction of speed and the delicate capture of light and shadow, it creates a powerful sense of destiny and a dynamic aesthetic about journey, time, and endless progress.", + "seed": 15342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_988_seed_15342_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12352901911363005, + "vlm_logprob": -0.12352901911363005, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.12352901911363005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03473397344350815 + ], + "opposite_logprob_tokens": [ + -3.909734010696411 + ], + "logprob_reward": -0.03473397344350815, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03473397344350815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18706487119197845, + 0.06057461351156235, + 0.05611500144004822, + 0.19763922691345215, + 0.10534115135669708, + 0.024502253159880638, + 0.03917814791202545, + 0.01996459998190403, + 0.2328636795282364, + 0.07675641775131226 + ], + "credit_map": [ + 0.021017655730247498, + 0.02411796525120735, + 0.035862818360328674, + 0.02082103304564953, + 0.010976921766996384, + 0.040408845990896225, + 0.02219671756029129, + 0.01166293304413557, + 0.006163585465401411, + 0.0066723208874464035, + 0.013722383417189121, + 0.006650328170508146, + 0.0052537573501467705, + 0.00963712576776743, + 0.007936420850455761, + 0.004538697190582752, + 0.004717679228633642, + 0.009065837599337101, + 0.008099888451397419, + 0.008414804935455322, + 0.006096202414482832, + 0.006767675280570984, + 0.006566427182406187, + 0.006386488676071167, + 0.01633336767554283, + 0.02329983189702034, + 0.03708716109395027, + 0.04574945196509361, + 0.01315420400351286, + 0.02248774655163288, + 0.024874133989214897, + 0.014653348363935947, + 0.008479252457618713, + 0.00955609604716301, + 0.014073039405047894, + 0.011922655627131462, + 0.012903566472232342, + 0.012832120060920715, + 0.014105395413935184, + 0.021469034254550934, + 0.0029962637927383184, + 0.0028639051597565413, + 0.003652511164546013, + 0.002924675587564707, + 0.0019571168813854456, + 0.0037237645592540503, + 0.003289572661742568, + 0.0030944442842155695, + 0.004561308305710554, + 0.003884074045345187, + 0.0035174137447029352, + 0.005481604021042585, + 0.003220777027308941, + 0.0071137091144919395, + 0.006918446626514196, + 0.004480819217860699, + 0.001865440048277378, + 0.0025483565405011177, + 0.0026811412535607815, + 0.004031581338495016, + 0.0013325801119208336, + 0.0025160920340567827, + 0.003264911472797394, + 0.001724498812109232, + 0.018656080588698387, + 0.023969311267137527, + 0.041714973747730255, + 0.05922576040029526, + 0.011615610681474209, + 0.03280393034219742, + 0.030593372881412506, + 0.014284659177064896, + 0.006103099789470434, + 0.0069946362636983395, + 0.018907051533460617, + 0.012835143133997917, + 0.003886767430230975, + 0.008312921039760113, + 0.011436857283115387, + 0.008279948495328426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.03473397344350815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033719997853040695 + ], + "opposite_logprob_tokens": [ + -3.408720016479492 + ], + "logprob_reward": -0.033719997853040695, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033719997853040695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31090620160102844, + 0.059137485921382904, + 0.07461250573396683, + 0.06967318803071976, + 0.03668476641178131, + 0.03579108044505119, + 0.10769118368625641, + 0.03740958124399185, + 0.0360787957906723, + 0.2320152372121811 + ], + "credit_map": [ + 0.029108097776770592, + 0.027373474091291428, + 0.04426031932234764, + 0.09340870380401611, + 0.017223788425326347, + 0.034082863479852676, + 0.04976947605609894, + 0.015679480507969856, + 0.00764299975708127, + 0.005542988423258066, + 0.01117967814207077, + 0.007857554592192173, + 0.005230782553553581, + 0.008602321147918701, + 0.008004879578948021, + 0.00507627846673131, + 0.005788714624941349, + 0.008007084019482136, + 0.014839712530374527, + 0.012603409588336945, + 0.007517004851251841, + 0.007301827892661095, + 0.011251069605350494, + 0.007303683552891016, + 0.006690334063023329, + 0.010924549773335457, + 0.0098361074924469, + 0.014255845919251442, + 0.004877064842730761, + 0.008023436181247234, + 0.008737524971365929, + 0.006328323855996132, + 0.0042173294350504875, + 0.0038104273844510317, + 0.0031891618855297565, + 0.005068718455731869, + 0.003245291532948613, + 0.006153551395982504, + 0.005048141814768314, + 0.005952144041657448, + 0.00443511875346303, + 0.0037481372710317373, + 0.007285290863364935, + 0.0050553930923342705, + 0.0020739499013870955, + 0.004944090731441975, + 0.004167764447629452, + 0.004081336781382561, + 0.01403815858066082, + 0.011405473574995995, + 0.009980268776416779, + 0.019556816667318344, + 0.009192300029098988, + 0.020729532465338707, + 0.01328512467443943, + 0.009503514505922794, + 0.0027830866165459156, + 0.004345593508332968, + 0.006324639543890953, + 0.005920291878283024, + 0.002595159923657775, + 0.0057755643501877785, + 0.006367105059325695, + 0.003298137802630663, + 0.002453416585922241, + 0.0033428394235670567, + 0.005412997677922249, + 0.0058234562166035175, + 0.00216170446947217, + 0.006774614565074444, + 0.0057714516296982765, + 0.004338317085057497, + 0.017358966171741486, + 0.019846484065055847, + 0.044724829494953156, + 0.04509054496884346, + 0.014861938543617725, + 0.02738923579454422, + 0.03236742317676544, + 0.030375812202692032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.033719997853040695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631670217961073 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631670217961073, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631670217961073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5313717126846313, + 0.05988757312297821, + 0.018497565761208534, + 0.03858739510178566, + 0.10738102346658707, + 0.05738653615117073, + 0.06512782722711563, + 0.03219333291053772, + 0.04176589101552963, + 0.04780110716819763 + ], + "credit_map": [ + 0.04809505492448807, + 0.057370640337467194, + 0.09135062247514725, + 0.11549853533506393, + 0.05106118321418762, + 0.07211709767580032, + 0.05765427649021149, + 0.03822433575987816, + 0.005976788699626923, + 0.0095960833132267, + 0.010230019688606262, + 0.00938528310507536, + 0.005492187105119228, + 0.006455960217863321, + 0.008179779164493084, + 0.0045714727602899075, + 0.0015165936201810837, + 0.0023962887935340405, + 0.0036093229427933693, + 0.002989834640175104, + 0.0017292246920987964, + 0.0018458543345332146, + 0.0022929117549210787, + 0.002117534400895238, + 0.0037036905996501446, + 0.004174892324954271, + 0.005657526198774576, + 0.006975144147872925, + 0.003018151968717575, + 0.005746682174503803, + 0.004800408612936735, + 0.004510895814746618, + 0.01033414714038372, + 0.01517130620777607, + 0.012868039309978485, + 0.019713295623660088, + 0.00983522366732359, + 0.01477216649800539, + 0.014036456122994423, + 0.010650379583239555, + 0.005325336009263992, + 0.005576799623668194, + 0.008441086858510971, + 0.010859474539756775, + 0.004523912910372019, + 0.007682004943490028, + 0.00732237845659256, + 0.007655543740838766, + 0.006915292702615261, + 0.006169168744236231, + 0.005758343730121851, + 0.010613354854285717, + 0.005806248169392347, + 0.010838366113603115, + 0.009115460328757763, + 0.009911592118442059, + 0.002459077164530754, + 0.0033366214483976364, + 0.004922494757920504, + 0.005864189937710762, + 0.002728966297581792, + 0.004878494888544083, + 0.004607280250638723, + 0.003396208630874753, + 0.0036078253760933876, + 0.004697955679148436, + 0.006328871939331293, + 0.009465335868299007, + 0.0031221951358020306, + 0.0052973320707678795, + 0.0060469359159469604, + 0.003199439961463213, + 0.004302625078707933, + 0.0038626021705567837, + 0.008882099762558937, + 0.01149167213588953, + 0.002392837544903159, + 0.005969444755464792, + 0.006015518680214882, + 0.00488430866971612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.007631670217961073 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233675837516785 + ], + "opposite_logprob_tokens": [ + -0.8983675837516785 + ], + "logprob_reward": -0.5233675837516785, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233675837516785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16026580333709717, + 0.08359546214342117, + 0.08396010845899582, + 0.07042870670557022, + 0.1264505833387375, + 0.09837042540311813, + 0.16277962923049927, + 0.05937834829092026, + 0.02557937614619732, + 0.12919151782989502 + ], + "credit_map": [ + 0.014911028556525707, + 0.017262928187847137, + 0.03682040050625801, + 0.02358441986143589, + 0.009369468316435814, + 0.027460884302854538, + 0.02140158787369728, + 0.009455088526010513, + 0.009722656570374966, + 0.00810622051358223, + 0.018473150208592415, + 0.01276132371276617, + 0.006287402007728815, + 0.011003547348082066, + 0.011986196972429752, + 0.0052549573592841625, + 0.0075560021214187145, + 0.009258666075766087, + 0.01456591859459877, + 0.009657145477831364, + 0.0074866218492388725, + 0.012370184995234013, + 0.013269851915538311, + 0.0097957169637084, + 0.006730698049068451, + 0.009657044894993305, + 0.013056165538728237, + 0.011596616357564926, + 0.004925709217786789, + 0.009584740735590458, + 0.010139109566807747, + 0.004738622810691595, + 0.013944705948233604, + 0.017715591937303543, + 0.012520772404968739, + 0.01882629282772541, + 0.0132526196539402, + 0.01696254312992096, + 0.017475802451372147, + 0.015752263367176056, + 0.012563792988657951, + 0.009655178524553776, + 0.013353895395994186, + 0.014121460728347301, + 0.007543381303548813, + 0.017151618376374245, + 0.011120411567389965, + 0.012860688380897045, + 0.01782958209514618, + 0.016204435378313065, + 0.015800487250089645, + 0.031201481819152832, + 0.012600592337548733, + 0.028581062331795692, + 0.02011885680258274, + 0.020443128421902657, + 0.005751952063292265, + 0.006378222722560167, + 0.010011401027441025, + 0.012026899494230747, + 0.0034206241834908724, + 0.007621112745255232, + 0.007898170500993729, + 0.006269964389503002, + 0.0021391217596828938, + 0.003820171346887946, + 0.0035781068727374077, + 0.004140250384807587, + 0.0019131494918838143, + 0.0035273649264127016, + 0.004274147562682629, + 0.0021870643831789494, + 0.009747409261763096, + 0.011846650391817093, + 0.022794941440224648, + 0.02784978225827217, + 0.00864787120372057, + 0.015364081598818302, + 0.019519474357366562, + 0.013421321287751198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.5233675837516785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15300390124320984, + 0.12084361165761948, + 0.12933349609375, + 0.25184834003448486, + 0.054759759455919266, + 0.026405299082398415, + 0.04218820482492447, + 0.09429606050252914, + 0.06125756725668907, + 0.0660637766122818 + ], + "credit_map": [ + 0.01803642138838768, + 0.028419896960258484, + 0.023340363055467606, + 0.021674158051609993, + 0.012780718505382538, + 0.020444093272089958, + 0.01832422986626625, + 0.009984036907553673, + 0.012645742855966091, + 0.013270861469209194, + 0.021206924691796303, + 0.01675323396921158, + 0.009731561876833439, + 0.020473942160606384, + 0.016991177573800087, + 0.00977017916738987, + 0.010009163990616798, + 0.016592537984251976, + 0.024497801437973976, + 0.014726851135492325, + 0.013611911796033382, + 0.018478861078619957, + 0.019149914383888245, + 0.012266448698937893, + 0.022080881521105766, + 0.03798000514507294, + 0.035353317856788635, + 0.05051927641034126, + 0.014733048155903816, + 0.02693365514278412, + 0.031160470098257065, + 0.03308768942952156, + 0.0053097279742360115, + 0.006607790943235159, + 0.00553490687161684, + 0.008699147030711174, + 0.0066461204551160336, + 0.008230877108871937, + 0.007345307618379593, + 0.006385885179042816, + 0.0024188042152673006, + 0.0030238700564950705, + 0.0053478265181183815, + 0.004807982128113508, + 0.001611383748240769, + 0.0038511836901307106, + 0.002829834586009383, + 0.0025144144892692566, + 0.0035601884592324495, + 0.005266213323920965, + 0.004323745146393776, + 0.005938392132520676, + 0.004512422252446413, + 0.007582828868180513, + 0.006642617750912905, + 0.004361799918115139, + 0.010997812263667583, + 0.010581685230135918, + 0.015820158645510674, + 0.01977473311126232, + 0.006406787317246199, + 0.010654536075890064, + 0.013152355328202248, + 0.006908003240823746, + 0.005162233021110296, + 0.008634218014776707, + 0.00846927147358656, + 0.01279690582305193, + 0.005012227222323418, + 0.008321382105350494, + 0.008009674027562141, + 0.0048516555689275265, + 0.004817352630198002, + 0.0073198298923671246, + 0.010781126096844673, + 0.014632979407906532, + 0.004840539768338203, + 0.006528016179800034, + 0.009863439947366714, + 0.007280490826815367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0181918703019619 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed tracking shot, low to the ground, hurtles forward at dizzying speed along two parallel steel rails. In the golden late afternoon sun, the cold rails reflect a warm, dazzling glow, stretching endlessly into the distance before converging at a vanishing point on the horizon. Below the camera, the dark wooden sleepers and the gravel ballast between them transform into blurred, dynamic streaks, swiftly sweeping backwards, their individual forms barely discernible, leaving only a strong trail of movement. On either side of the tracks, withered grass sway gently in the breeze, bringing a touch of vitality to the vast, somewhat desolate landscape. The entire scene is imbued with a powerful sense of movement and visual impact. Through the extreme depiction of speed and the delicate capture of light and shadow, it creates a powerful sense of destiny and a dynamic aesthetic about journey, time, and endless progress.", + "seed": 15343, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_988_seed_15343_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12278265226632357, + "vlm_logprob": -0.12278265226632357, + "vlm_yes_no_margin": 3.699999952316284, + "vlm_token_credit": -0.12278265226632357, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03870886191725731 + ], + "opposite_logprob_tokens": [ + -4.038708686828613 + ], + "logprob_reward": -0.03870886191725731, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03870886191725731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18480284512043, + 0.08032887428998947, + 0.039222538471221924, + 0.04166873171925545, + 0.0624166876077652, + 0.11207391321659088, + 0.27944493293762207, + 0.026320233941078186, + 0.03528190776705742, + 0.13843931257724762 + ], + "credit_map": [ + 0.03754911571741104, + 0.014465782791376114, + 0.026154914870858192, + 0.01840803772211075, + 0.017018385231494904, + 0.035864949226379395, + 0.02070019580423832, + 0.014641458168625832, + 0.004665420390665531, + 0.005522646941244602, + 0.006392952054738998, + 0.01172163151204586, + 0.006744627840816975, + 0.023455863818526268, + 0.007287962362170219, + 0.014537771232426167, + 0.004409747663885355, + 0.005418651271611452, + 0.005007003899663687, + 0.005689387209713459, + 0.0056066373363137245, + 0.004230738617479801, + 0.003961777314543724, + 0.004898591432720423, + 0.004031802993267775, + 0.004220409784466028, + 0.0036163043696433306, + 0.007455015555024147, + 0.005731011275202036, + 0.005130875390022993, + 0.005833775270730257, + 0.005649535451084375, + 0.00603364035487175, + 0.0054841190576553345, + 0.01851518638432026, + 0.006695327814668417, + 0.004918231628835201, + 0.007313106674700975, + 0.0055684540420770645, + 0.007888623513281345, + 0.013147135265171528, + 0.009665329940617085, + 0.014858396723866463, + 0.02064555510878563, + 0.007922889664769173, + 0.006748805288225412, + 0.011762356385588646, + 0.027323443442583084, + 0.0142727792263031, + 0.01098804734647274, + 0.054193079471588135, + 0.02134784497320652, + 0.014979938045144081, + 0.09685694426298141, + 0.012648933567106724, + 0.054157380014657974, + 0.0026353048160672188, + 0.0025014348793774843, + 0.004767227452248335, + 0.002768312580883503, + 0.002270381897687912, + 0.0032922402024269104, + 0.002227946650236845, + 0.005857386160641909, + 0.003894009394571185, + 0.004021704196929932, + 0.0035965649876743555, + 0.004913875367492437, + 0.003901520511135459, + 0.007187565788626671, + 0.0031655917409807444, + 0.004601075313985348, + 0.014895846135914326, + 0.0110471798107028, + 0.019929518923163414, + 0.027362508699297905, + 0.014867719262838364, + 0.012193139642477036, + 0.015027528628706932, + 0.02311587706208229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.03870886191725731 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430784709751606 + ], + "opposite_logprob_tokens": [ + -3.1680784225463867 + ], + "logprob_reward": -0.0430784709751606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430784709751606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27016332745552063, + 0.1205144003033638, + 0.03244228661060333, + 0.056420814245939255, + 0.04004361853003502, + 0.12893551588058472, + 0.09954893589019775, + 0.06374796479940414, + 0.027095675468444824, + 0.16108745336532593 + ], + "credit_map": [ + 0.030834240838885307, + 0.015397577546536922, + 0.05495726689696312, + 0.025448622182011604, + 0.022145351395010948, + 0.06722307205200195, + 0.02824651449918747, + 0.02591068670153618, + 0.010108006186783314, + 0.00747773377224803, + 0.00814528577029705, + 0.0059028430841863155, + 0.008153143338859081, + 0.04049037769436836, + 0.010821614414453506, + 0.029415400698781013, + 0.0047559286467731, + 0.0031133824959397316, + 0.004206966608762741, + 0.005698018241673708, + 0.005083189345896244, + 0.002842227229848504, + 0.003621665295213461, + 0.0031209068838506937, + 0.005294304806739092, + 0.007788198068737984, + 0.005528590641915798, + 0.0075597865507006645, + 0.006860034540295601, + 0.008098552003502846, + 0.00706928875297308, + 0.008222057484090328, + 0.003251878311857581, + 0.0034525722730904818, + 0.003465373069047928, + 0.004688770044595003, + 0.003345609176903963, + 0.006244708318263292, + 0.006801513489335775, + 0.008793194778263569, + 0.01971174031496048, + 0.01353888213634491, + 0.01497170701622963, + 0.026132872328162193, + 0.012006787583231926, + 0.009326047264039516, + 0.01229770015925169, + 0.020949773490428925, + 0.006280742585659027, + 0.004672460723668337, + 0.02766953594982624, + 0.013645774684846401, + 0.005702582653611898, + 0.023576008155941963, + 0.006653512828052044, + 0.011348320171236992, + 0.008119337260723114, + 0.0061101727187633514, + 0.007190511096268892, + 0.010122522711753845, + 0.006781943142414093, + 0.006001137662678957, + 0.00727015919983387, + 0.012152177281677723, + 0.0016670101322233677, + 0.002206864533945918, + 0.0026457265485078096, + 0.004265068098902702, + 0.0025532927829772234, + 0.002516485285013914, + 0.0051177977584302425, + 0.006123431026935577, + 0.03012513928115368, + 0.020140066742897034, + 0.021608352661132812, + 0.021476659923791885, + 0.017054542899131775, + 0.013223892077803612, + 0.018215211108326912, + 0.019243594259023666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0430784709751606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059498813934624195 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.0059498813934624195, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059498813934624195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20440572500228882, + 0.2267315834760666, + 0.059175219386816025, + 0.05638991296291351, + 0.04984266310930252, + 0.07512921094894409, + 0.10420099645853043, + 0.09953446686267853, + 0.03432494029402733, + 0.09026523679494858 + ], + "credit_map": [ + 0.02688714489340782, + 0.015708863735198975, + 0.03282804787158966, + 0.020880138501524925, + 0.018524453043937683, + 0.05359064042568207, + 0.019921880215406418, + 0.01606455259025097, + 0.014231218956410885, + 0.015115803107619286, + 0.022810010239481926, + 0.013091888278722763, + 0.017757441848516464, + 0.08692675083875656, + 0.01752016693353653, + 0.03927832096815109, + 0.009044519625604153, + 0.007134126964956522, + 0.007922534830868244, + 0.006931179203093052, + 0.008101089857518673, + 0.007551611866801977, + 0.005797708872705698, + 0.006692447233945131, + 0.004203157965093851, + 0.008260446600615978, + 0.005396171472966671, + 0.0070896758697927, + 0.007305460516363382, + 0.008726119995117188, + 0.008058414794504642, + 0.007350462954491377, + 0.004052168224006891, + 0.004193791188299656, + 0.004328819457441568, + 0.006103758700191975, + 0.006328368093818426, + 0.010180581361055374, + 0.007189391180872917, + 0.007465784903615713, + 0.006761184893548489, + 0.005529589951038361, + 0.006115240044891834, + 0.015351863577961922, + 0.007074818946421146, + 0.005584633909165859, + 0.011785210110247135, + 0.01692667230963707, + 0.005315813701599836, + 0.004331404343247414, + 0.013525877147912979, + 0.009325440973043442, + 0.007572932634502649, + 0.03594909980893135, + 0.008592440746724606, + 0.019587988033890724, + 0.009274978190660477, + 0.008398622274398804, + 0.01926143281161785, + 0.012345945462584496, + 0.01141528133302927, + 0.010016734711825848, + 0.008715422824025154, + 0.020106052979826927, + 0.0034149421844631433, + 0.0038345635402947664, + 0.003548450069501996, + 0.003580555785447359, + 0.004347321577370167, + 0.006874565500766039, + 0.0036978526040911674, + 0.005026688799262047, + 0.008418070152401924, + 0.008651490323245525, + 0.015114382840692997, + 0.02169712446630001, + 0.008604875765740871, + 0.008532825857400894, + 0.00967354979366064, + 0.009572913870215416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0059498813934624195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233553647994995 + ], + "opposite_logprob_tokens": [ + -0.8983553647994995 + ], + "logprob_reward": -0.5233553647994995, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233553647994995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12163333594799042, + 0.09985146671533585, + 0.10014212131500244, + 0.0431569442152977, + 0.16611690819263458, + 0.08202798664569855, + 0.12559771537780762, + 0.07025554031133652, + 0.07681851834058762, + 0.11439952999353409 + ], + "credit_map": [ + 0.024635925889015198, + 0.008027583360671997, + 0.020056260749697685, + 0.0168551504611969, + 0.010303591378033161, + 0.02121455781161785, + 0.01007572840899229, + 0.010464529506862164, + 0.007853148505091667, + 0.005890342872589827, + 0.009336965158581734, + 0.009132592007517815, + 0.009019485674798489, + 0.03248054161667824, + 0.008491214364767075, + 0.017647165805101395, + 0.014527563937008381, + 0.0106288380920887, + 0.011749393306672573, + 0.018673371523618698, + 0.016912968829274178, + 0.007272414863109589, + 0.010792342945933342, + 0.009585214778780937, + 0.003308197483420372, + 0.005021664313971996, + 0.004342341795563698, + 0.006288838107138872, + 0.0076676467433571815, + 0.005487239919602871, + 0.004397341050207615, + 0.00664367014542222, + 0.011834989301860332, + 0.011701331473886967, + 0.05705196037888527, + 0.01752341166138649, + 0.009635783731937408, + 0.015467123128473759, + 0.014463554136455059, + 0.028438735753297806, + 0.011133451014757156, + 0.011356792412698269, + 0.009849583730101585, + 0.013253627344965935, + 0.009513769298791885, + 0.008236164227128029, + 0.008765911683440208, + 0.00991868507117033, + 0.007158439140766859, + 0.006544433068484068, + 0.032339632511138916, + 0.016463683918118477, + 0.00606777286157012, + 0.03220962733030319, + 0.008737881667912006, + 0.016076242551207542, + 0.008231423795223236, + 0.008702699095010757, + 0.007439425680786371, + 0.009561989456415176, + 0.012215855531394482, + 0.007023507729172707, + 0.007060396485030651, + 0.010020241141319275, + 0.007738061249256134, + 0.006205808371305466, + 0.005820580292493105, + 0.009275661781430244, + 0.008495922200381756, + 0.009231176227331161, + 0.012058083899319172, + 0.01799321360886097, + 0.020200351253151894, + 0.010046928189694881, + 0.01549682766199112, + 0.0178163293749094, + 0.00933851022273302, + 0.008116899989545345, + 0.012774608097970486, + 0.020609067752957344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.5233553647994995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028206822462379932 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028206822462379932, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028206822462379932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2801669239997864, + 0.18860100209712982, + 0.08279979974031448, + 0.022838054224848747, + 0.05940960347652435, + 0.17809538543224335, + 0.04373881220817566, + 0.042708076536655426, + 0.019776269793510437, + 0.081866055727005 + ], + "credit_map": [ + 0.019305264577269554, + 0.01947655901312828, + 0.04309152066707611, + 0.023830266669392586, + 0.025632843375205994, + 0.0953216552734375, + 0.03440893441438675, + 0.019099878147244453, + 0.01354576088488102, + 0.010802350006997585, + 0.012073138728737831, + 0.014281248673796654, + 0.01656186953186989, + 0.05157362297177315, + 0.024663148447871208, + 0.045099854469299316, + 0.011869771406054497, + 0.011570064350962639, + 0.009793012402951717, + 0.00944213755428791, + 0.013684812933206558, + 0.008000719361007214, + 0.009191237390041351, + 0.009248043410480022, + 0.002301062922924757, + 0.00191000837367028, + 0.0020609991624951363, + 0.0035206573083996773, + 0.0026993274223059416, + 0.0025279070250689983, + 0.003313760505989194, + 0.004504330921918154, + 0.0044379341416060925, + 0.004603965673595667, + 0.006090771872550249, + 0.005928493104875088, + 0.005658864974975586, + 0.01795426942408085, + 0.005969429388642311, + 0.008765880018472672, + 0.02225293219089508, + 0.01655394956469536, + 0.024239839985966682, + 0.030998600646853447, + 0.012959111481904984, + 0.011941560544073582, + 0.016294604167342186, + 0.04285479336977005, + 0.002830308396369219, + 0.0034362771548330784, + 0.003867103485390544, + 0.007235747762024403, + 0.0051854862831532955, + 0.011822900734841824, + 0.00508961221203208, + 0.004271377343684435, + 0.005058330483734608, + 0.004193465691059828, + 0.00956286583095789, + 0.004649726673960686, + 0.003653276478871703, + 0.0036781972739845514, + 0.0031759561970829964, + 0.0087362602353096, + 0.003142543137073517, + 0.0020152160432189703, + 0.0017186967888846993, + 0.0019444124773144722, + 0.0024211315903812647, + 0.0030392531771212816, + 0.0024054741952568293, + 0.003089543664827943, + 0.00821156706660986, + 0.009376199916005135, + 0.009298634715378284, + 0.010293370112776756, + 0.012110822834074497, + 0.007592960726469755, + 0.014757122844457626, + 0.01022538635879755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0028206822462379932 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed tracking shot, low to the ground, hurtles forward at dizzying speed along two parallel steel rails. In the golden late afternoon sun, the cold rails reflect a warm, dazzling glow, stretching endlessly into the distance before converging at a vanishing point on the horizon. Below the camera, the dark wooden sleepers and the gravel ballast between them transform into blurred, dynamic streaks, swiftly sweeping backwards, their individual forms barely discernible, leaving only a strong trail of movement. On either side of the tracks, withered grass sway gently in the breeze, bringing a touch of vitality to the vast, somewhat desolate landscape. The entire scene is imbued with a powerful sense of movement and visual impact. Through the extreme depiction of speed and the delicate capture of light and shadow, it creates a powerful sense of destiny and a dynamic aesthetic about journey, time, and endless progress.", + "seed": 15344, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_988_seed_15344_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14654333973303438, + "vlm_logprob": -0.14654333973303438, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.14654333973303438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03870370239019394 + ], + "opposite_logprob_tokens": [ + -4.038703918457031 + ], + "logprob_reward": -0.03870370239019394, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03870370239019394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4259263873100281, + 0.044220976531505585, + 0.07297786325216293, + 0.04088742658495903, + 0.05752109736204147, + 0.0311574786901474, + 0.055069804191589355, + 0.06276565790176392, + 0.16405339539051056, + 0.045419834554195404 + ], + "credit_map": [ + 0.03476560860872269, + 0.019131703302264214, + 0.09510137885808945, + 0.07265090942382812, + 0.02272137440741062, + 0.025110989809036255, + 0.03297281637787819, + 0.12347163259983063, + 0.003828601446002722, + 0.0037369346246123314, + 0.005745808593928814, + 0.008033540099859238, + 0.0047929296270012856, + 0.0041720191948115826, + 0.007530106697231531, + 0.006381040904670954, + 0.0035427073016762733, + 0.004213860258460045, + 0.010336547158658504, + 0.019590632990002632, + 0.005276641808450222, + 0.014222558587789536, + 0.009514175355434418, + 0.006280749104917049, + 0.002652085153385997, + 0.0021661720238626003, + 0.005320937838405371, + 0.006002948619425297, + 0.0043885172344744205, + 0.005024688318371773, + 0.010797852650284767, + 0.004534230101853609, + 0.004141442012041807, + 0.004184860736131668, + 0.008113859221339226, + 0.006895558442920446, + 0.005552601534873247, + 0.012346941977739334, + 0.005802766419947147, + 0.01048307679593563, + 0.0015040236758068204, + 0.0013062459183856845, + 0.0077437711879611015, + 0.0030929138883948326, + 0.0018463145243003964, + 0.003337611211463809, + 0.010002922266721725, + 0.0023236770648509264, + 0.004345673136413097, + 0.0074893697164952755, + 0.007925896905362606, + 0.006322088651359081, + 0.004549544304609299, + 0.007532481104135513, + 0.010658323764801025, + 0.006246431730687618, + 0.009272094815969467, + 0.005864777602255344, + 0.00702121714130044, + 0.008813017047941685, + 0.005200558807700872, + 0.006819260772317648, + 0.009311901405453682, + 0.010462840087711811, + 0.014516787603497505, + 0.01214076392352581, + 0.02436855435371399, + 0.03551120683550835, + 0.012655164115130901, + 0.016306808218359947, + 0.02198714390397072, + 0.02656697854399681, + 0.002705380553379655, + 0.0031598052009940147, + 0.003629410406574607, + 0.0104597182944417, + 0.0027577511500567198, + 0.004691197071224451, + 0.011217844672501087, + 0.00679872976616025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.03870370239019394 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521248161792755 + ], + "opposite_logprob_tokens": [ + -1.5021247863769531 + ], + "logprob_reward": -0.2521248161792755, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521248161792755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5013101696968079, + 0.04360440373420715, + 0.076255202293396, + 0.042350392788648605, + 0.03295335918664932, + 0.12402770668268204, + 0.03258446231484413, + 0.05381022021174431, + 0.0761895403265953, + 0.016914531588554382 + ], + "credit_map": [ + 0.048348162323236465, + 0.021918155252933502, + 0.1428556889295578, + 0.058262959122657776, + 0.02252844162285328, + 0.028036363422870636, + 0.046838510781526566, + 0.13252189755439758, + 0.002746620448306203, + 0.0033931599464267492, + 0.003483655396848917, + 0.008304656483232975, + 0.0046060350723564625, + 0.003787271911278367, + 0.009079222567379475, + 0.008203784003853798, + 0.0037834655959159136, + 0.004316286649554968, + 0.009264718741178513, + 0.018369393423199654, + 0.00657091336324811, + 0.013716056011617184, + 0.010879523120820522, + 0.009354852139949799, + 0.002215460641309619, + 0.0033486976753920317, + 0.004844008944928646, + 0.00877048447728157, + 0.003500637598335743, + 0.005209269002079964, + 0.00733954319730401, + 0.0071222963742911816, + 0.0022606661077588797, + 0.003528747009113431, + 0.003880507545545697, + 0.0046953074634075165, + 0.00418702932074666, + 0.006395032163709402, + 0.003313519060611725, + 0.0046925535425543785, + 0.005017653573304415, + 0.006270465441048145, + 0.0174226313829422, + 0.008069216273725033, + 0.005745981819927692, + 0.009082037024199963, + 0.031018322333693504, + 0.04140140861272812, + 0.0023793333675712347, + 0.003903910517692566, + 0.00483056902885437, + 0.0038458751514554024, + 0.0027237283065915108, + 0.0042840829119086266, + 0.006867820862680674, + 0.003749146359041333, + 0.006939566694200039, + 0.0035313805565238, + 0.005377982277423143, + 0.013366039842367172, + 0.0038339269813150167, + 0.004376950208097696, + 0.005555438343435526, + 0.01082894392311573, + 0.006995172239840031, + 0.006320525426417589, + 0.011742380447685719, + 0.012999732047319412, + 0.007183937821537256, + 0.010849318467080593, + 0.009034241549670696, + 0.011064237914979458, + 0.0014074911596253514, + 0.0014917367370799184, + 0.0015921044396236539, + 0.0035466512199491262, + 0.0014567894395440817, + 0.002145054517313838, + 0.0021262383088469505, + 0.003148466581478715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.2521248161792755 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782824665307999 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782824665307999, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782824665307999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4446444511413574, + 0.06417624652385712, + 0.07179088145494461, + 0.03467216715216637, + 0.08119868487119675, + 0.03512434661388397, + 0.0423286035656929, + 0.08286084234714508, + 0.07588355243206024, + 0.06732022762298584 + ], + "credit_map": [ + 0.037230830639600754, + 0.018665332347154617, + 0.10564859956502914, + 0.060908857733011246, + 0.02638694830238819, + 0.02639184519648552, + 0.05779029428958893, + 0.11162175238132477, + 0.0036212748382240534, + 0.004283457063138485, + 0.008368025533854961, + 0.012885689735412598, + 0.006120926234871149, + 0.007498463615775108, + 0.012484034523367882, + 0.008914382196962833, + 0.005093123763799667, + 0.009113418869674206, + 0.007546361070126295, + 0.011042178608477116, + 0.009370621293783188, + 0.012940614484250546, + 0.00844732765108347, + 0.008237246423959732, + 0.0023265595082193613, + 0.0022088508121669292, + 0.0030969951767474413, + 0.0051312996074557304, + 0.00357830454595387, + 0.0049130879342556, + 0.006746810395270586, + 0.006670261267572641, + 0.006369087379425764, + 0.007996723987162113, + 0.010085321962833405, + 0.013564997352659702, + 0.010121327824890614, + 0.010585987009108067, + 0.012876137159764767, + 0.009599106386303902, + 0.00366790103726089, + 0.0032253721728920937, + 0.004203712567687035, + 0.005595856346189976, + 0.0032706735655665398, + 0.005995848216116428, + 0.00437963055446744, + 0.004785353317856789, + 0.003161162603646517, + 0.0057947938330471516, + 0.005922021809965372, + 0.004987271968275309, + 0.0037609071005135775, + 0.005778120830655098, + 0.009011178277432919, + 0.003913150168955326, + 0.011851190589368343, + 0.008150826208293438, + 0.010003670118749142, + 0.010898425243794918, + 0.00794578343629837, + 0.009962046518921852, + 0.012406465597450733, + 0.011642439290881157, + 0.005359115079045296, + 0.006546981167048216, + 0.007942798547446728, + 0.01811482198536396, + 0.009452144615352154, + 0.009990920312702656, + 0.006390422582626343, + 0.012086346745491028, + 0.0037198103964328766, + 0.004437790252268314, + 0.007079559378325939, + 0.013014176860451698, + 0.007591564673930407, + 0.006810109131038189, + 0.009210220538079739, + 0.01545700617134571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.009782824665307999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42891231179237366 + ], + "opposite_logprob_tokens": [ + -1.0539122819900513 + ], + "logprob_reward": -0.42891231179237366, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42891231179237366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17608220875263214, + 0.0565275140106678, + 0.07609480619430542, + 0.1066533699631691, + 0.0656711682677269, + 0.05899476259946823, + 0.13156254589557648, + 0.11238677054643631, + 0.12802954018115997, + 0.08799725770950317 + ], + "credit_map": [ + 0.017072442919015884, + 0.008648691698908806, + 0.03544658049941063, + 0.029325349256396294, + 0.012013551779091358, + 0.013243897818028927, + 0.02169630117714405, + 0.03863539919257164, + 0.004449596628546715, + 0.003790148999541998, + 0.004741992801427841, + 0.01108537893742323, + 0.004524203483015299, + 0.004384176805615425, + 0.014179588295519352, + 0.009372428059577942, + 0.003502530511468649, + 0.005779415834695101, + 0.009858079254627228, + 0.017802298069000244, + 0.0077850003726780415, + 0.015634004026651382, + 0.00809257011860609, + 0.007640906143933535, + 0.006173043977469206, + 0.006051888223737478, + 0.01549954991787672, + 0.016440153121948242, + 0.00883586797863245, + 0.013519931584596634, + 0.0234499704092741, + 0.016682954505085945, + 0.004861024674028158, + 0.0076150670647621155, + 0.0075078546069562435, + 0.009620973840355873, + 0.008286606520414352, + 0.010212983936071396, + 0.009355940856039524, + 0.008210715837776661, + 0.0048119486309587955, + 0.004725792445242405, + 0.006016079802066088, + 0.008028155192732811, + 0.004016835708171129, + 0.007166469469666481, + 0.014069369062781334, + 0.010160114616155624, + 0.01022607646882534, + 0.015332578681409359, + 0.019314639270305634, + 0.015907825902104378, + 0.011547351256012917, + 0.02013697847723961, + 0.026298251003026962, + 0.01279884111136198, + 0.016077306121587753, + 0.010675911791622639, + 0.0131128691136837, + 0.021309029310941696, + 0.009049666114151478, + 0.014202467165887356, + 0.012337762862443924, + 0.015621756203472614, + 0.008714628405869007, + 0.01147137675434351, + 0.016108445823192596, + 0.018091045320034027, + 0.015239368192851543, + 0.02511618100106716, + 0.012789424508810043, + 0.02049906551837921, + 0.003039003349840641, + 0.003186821937561035, + 0.007791378069669008, + 0.02711494080722332, + 0.004789106547832489, + 0.0045335134491324425, + 0.022314565256237984, + 0.015227931551635265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.42891231179237366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193043638020754 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.003193043638020754, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193043638020754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37881481647491455, + 0.05715305358171463, + 0.053744565695524216, + 0.04921116679906845, + 0.02755882777273655, + 0.250351220369339, + 0.0355374701321125, + 0.042829860001802444, + 0.07826229929924011, + 0.02653668448328972 + ], + "credit_map": [ + 0.02422185242176056, + 0.020251501351594925, + 0.06165159121155739, + 0.09565065056085587, + 0.023087821900844574, + 0.021214326843619347, + 0.028186358511447906, + 0.10455069690942764, + 0.0036636979784816504, + 0.004909341223537922, + 0.005056804046034813, + 0.007363387383520603, + 0.005932217929512262, + 0.008517220616340637, + 0.012644309550523758, + 0.009066075086593628, + 0.003477302845567465, + 0.002963852370157838, + 0.009191642515361309, + 0.016580069437623024, + 0.00492518674582243, + 0.005925671197474003, + 0.005999254062771797, + 0.0046815876848995686, + 0.003542644437402487, + 0.003033022629097104, + 0.006738938391208649, + 0.007028118707239628, + 0.0042704688385128975, + 0.006510104052722454, + 0.009813808836042881, + 0.008274060674011707, + 0.0019684985745698214, + 0.0031778374686837196, + 0.003097363281995058, + 0.0041483002714812756, + 0.0037910209503024817, + 0.004686524160206318, + 0.0037076259031891823, + 0.002981657860800624, + 0.013744566589593887, + 0.010094049386680126, + 0.05675426498055458, + 0.024440821260213852, + 0.013354730792343616, + 0.018493765965104103, + 0.0928243026137352, + 0.020644713193178177, + 0.0027568398509174585, + 0.0038040922954678535, + 0.004707715939730406, + 0.004375615157186985, + 0.0032276485580950975, + 0.004552761558443308, + 0.008323405869305134, + 0.0037893913686275482, + 0.003950418438762426, + 0.0029483826365321875, + 0.004024202469736338, + 0.010398061014711857, + 0.003231606911867857, + 0.004176175221800804, + 0.0046578506007790565, + 0.009443159215152264, + 0.005372895859181881, + 0.00622732937335968, + 0.00912272185087204, + 0.014845592901110649, + 0.009040913544595242, + 0.012172380462288857, + 0.006276899483054876, + 0.015203570947051048, + 0.002015184611082077, + 0.002572637517005205, + 0.002380646765232086, + 0.005532565992325544, + 0.00309597235172987, + 0.004050748888403177, + 0.002782771596685052, + 0.004106157459318638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.003193043638020754 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed tracking shot, low to the ground, hurtles forward at dizzying speed along two parallel steel rails. In the golden late afternoon sun, the cold rails reflect a warm, dazzling glow, stretching endlessly into the distance before converging at a vanishing point on the horizon. Below the camera, the dark wooden sleepers and the gravel ballast between them transform into blurred, dynamic streaks, swiftly sweeping backwards, their individual forms barely discernible, leaving only a strong trail of movement. On either side of the tracks, withered grass sway gently in the breeze, bringing a touch of vitality to the vast, somewhat desolate landscape. The entire scene is imbued with a powerful sense of movement and visual impact. Through the extreme depiction of speed and the delicate capture of light and shadow, it creates a powerful sense of destiny and a dynamic aesthetic about journey, time, and endless progress.", + "seed": 15345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_6/index_988_seed_15345_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21867206431925296, + "vlm_logprob": -0.21867206431925296, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.21867206431925296, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04141755774617195 + ], + "opposite_logprob_tokens": [ + -3.7914175987243652 + ], + "logprob_reward": -0.04141755774617195, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04141755774617195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24601390957832336, + 0.045903392136096954, + 0.03611104190349579, + 0.09176277369260788, + 0.19311249256134033, + 0.0331578366458416, + 0.029226772487163544, + 0.11244785785675049, + 0.18724489212036133, + 0.025018997490406036 + ], + "credit_map": [ + 0.009320215322077274, + 0.010811416432261467, + 0.04390335828065872, + 0.01568673551082611, + 0.010424945503473282, + 0.012530326843261719, + 0.10356385260820389, + 0.03977307677268982, + 0.0026454883627593517, + 0.0055444627068936825, + 0.007320851553231478, + 0.007967323064804077, + 0.0041000451892614365, + 0.003616963978856802, + 0.0065998174250125885, + 0.008108440786600113, + 0.0022126827389001846, + 0.0049207936972379684, + 0.0041938843205571175, + 0.0063201445154845715, + 0.004698203410953283, + 0.003926337230950594, + 0.006065422669053078, + 0.003773573786020279, + 0.005945789162069559, + 0.009394501335918903, + 0.01539894100278616, + 0.01217268779873848, + 0.010474936105310917, + 0.009316802024841309, + 0.018863048404455185, + 0.010196066461503506, + 0.014311249367892742, + 0.01691737398505211, + 0.06653314083814621, + 0.013247889466583729, + 0.007861576974391937, + 0.0115094855427742, + 0.012484654784202576, + 0.05024711415171623, + 0.002639652928337455, + 0.004192284308373928, + 0.0050230794586241245, + 0.002730998210608959, + 0.007008398417383432, + 0.005225386004894972, + 0.0035544023849070072, + 0.002783633768558502, + 0.002330261282622814, + 0.0030109253711998463, + 0.005917469039559364, + 0.0055368137545883656, + 0.003039992181584239, + 0.003421480068936944, + 0.003558228025212884, + 0.002411601832136512, + 0.009440433233976364, + 0.012713707983493805, + 0.02137940749526024, + 0.021855419501662254, + 0.010698082856833935, + 0.012334268540143967, + 0.011330821551382542, + 0.012695719487965107, + 0.013751931488513947, + 0.027198562398552895, + 0.04481935873627663, + 0.02953219600021839, + 0.022422151640057564, + 0.011552564799785614, + 0.024862732738256454, + 0.013105395250022411, + 0.004070701543241739, + 0.0030196555890142918, + 0.004822025541216135, + 0.0031900042667984962, + 0.0029332544654607773, + 0.0027955209370702505, + 0.0015555188292637467, + 0.0026323157362639904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.04141755774617195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373567387461662 + ], + "opposite_logprob_tokens": [ + -3.408735752105713 + ], + "logprob_reward": -0.03373567387461662, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373567387461662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30251678824424744, + 0.34105461835861206, + 0.01276382990181446, + 0.0286377165466547, + 0.03265193849802017, + 0.090377077460289, + 0.052655983716249466, + 0.09851690381765366, + 0.022291716188192368, + 0.018533315509557724 + ], + "credit_map": [ + 0.011706999503076077, + 0.01279873214662075, + 0.052824780344963074, + 0.02496233768761158, + 0.018061671406030655, + 0.018648238852620125, + 0.11209863424301147, + 0.05141543969511986, + 0.01481548324227333, + 0.0181717686355114, + 0.08557102829217911, + 0.1550934910774231, + 0.018582578748464584, + 0.015359445475041866, + 0.018510285764932632, + 0.014950617216527462, + 0.0011431473540142179, + 0.0016336583066731691, + 0.0014759354526177049, + 0.0017351651331409812, + 0.0016582277603447437, + 0.0015769641613587737, + 0.00152100611012429, + 0.002019727835431695, + 0.0027978005819022655, + 0.004625626839697361, + 0.003930840641260147, + 0.003469899296760559, + 0.004043803550302982, + 0.004093729890882969, + 0.0026233242824673653, + 0.0030526963528245687, + 0.002512165578082204, + 0.003250978421419859, + 0.009303862228989601, + 0.004532936494797468, + 0.0024997051805257797, + 0.002999009797349572, + 0.0018273642053827643, + 0.005725922994315624, + 0.0080996323376894, + 0.011515522375702858, + 0.011225477792322636, + 0.00859089009463787, + 0.018248818814754486, + 0.013758782297372818, + 0.010019776411354542, + 0.00891819130629301, + 0.004243128467351198, + 0.004295798949897289, + 0.009332467801868916, + 0.00798215251415968, + 0.004526730161160231, + 0.005978365428745747, + 0.005663396790623665, + 0.010633951053023338, + 0.010385907255113125, + 0.012042161077260971, + 0.01938617043197155, + 0.01849665865302086, + 0.009374380111694336, + 0.009487144649028778, + 0.008754470385611057, + 0.010590022429823875, + 0.0017035540658980608, + 0.002511709462851286, + 0.0035973843187093735, + 0.004365440923720598, + 0.0021701636724174023, + 0.002049390459433198, + 0.0033808029256761074, + 0.0025132750160992146, + 0.002087796339765191, + 0.0025248227175325155, + 0.0024942955933511257, + 0.002060661558061838, + 0.002503572730347514, + 0.003703637281432748, + 0.0014999794075265527, + 0.0016585533739998937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.03373567387461662 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294363170862198 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.023294363170862198, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294363170862198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37047359347343445, + 0.11021793633699417, + 0.047810524702072144, + 0.05759963393211365, + 0.06772179901599884, + 0.08527012914419174, + 0.09538911283016205, + 0.077699676156044, + 0.06215449050068855, + 0.025663083419203758 + ], + "credit_map": [ + 0.019218267872929573, + 0.025490038096904755, + 0.06228841468691826, + 0.03731778636574745, + 0.03130248934030533, + 0.0312684141099453, + 0.1154111996293068, + 0.04817700758576393, + 0.005070722661912441, + 0.008655233308672905, + 0.02136107347905636, + 0.04024745523929596, + 0.010735171847045422, + 0.008535807952284813, + 0.007235056255012751, + 0.008377423509955406, + 0.0032236429397016764, + 0.0060471948236227036, + 0.006099444814026356, + 0.008462192490696907, + 0.005289990454912186, + 0.005987458396703005, + 0.009018057957291603, + 0.003682543057948351, + 0.004490576218813658, + 0.00872680265456438, + 0.008296924643218517, + 0.008300688117742538, + 0.007387996651232243, + 0.006633335258811712, + 0.006730858236551285, + 0.007032447960227728, + 0.005594894289970398, + 0.005742385517805815, + 0.019885098561644554, + 0.006407411769032478, + 0.005028394050896168, + 0.006406975910067558, + 0.004402292426675558, + 0.014254346489906311, + 0.007715227548032999, + 0.00897759385406971, + 0.009646239690482616, + 0.008294602856040001, + 0.01962929777801037, + 0.014359882101416588, + 0.01043820008635521, + 0.006209088955074549, + 0.006109181791543961, + 0.00888893287628889, + 0.018861347809433937, + 0.01772208884358406, + 0.011189604178071022, + 0.011012365110218525, + 0.012854673899710178, + 0.008750909939408302, + 0.006627011112868786, + 0.00660322979092598, + 0.012419008649885654, + 0.009659718722105026, + 0.012018784880638123, + 0.009721208363771439, + 0.009350722655653954, + 0.01129999291151762, + 0.003939840942621231, + 0.006440537050366402, + 0.011770558543503284, + 0.013951001688838005, + 0.007478820160031319, + 0.005599450785666704, + 0.007241076789796352, + 0.00573320547118783, + 0.0029818061739206314, + 0.0022130191791802645, + 0.006048310548067093, + 0.0030266044195741415, + 0.0039307307451963425, + 0.0030373851768672466, + 0.00243417383171618, + 0.001991051249206066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.023294363170862198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743272662162781 + ], + "opposite_logprob_tokens": [ + -0.4743272662162781 + ], + "logprob_reward": -0.9743272662162781, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743272662162781, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1473757028579712, + 0.31706294417381287, + 0.036090247333049774, + 0.039560627192258835, + 0.1531829982995987, + 0.03828924521803856, + 0.10572072118520737, + 0.10866310447454453, + 0.03201578930020332, + 0.022038672119379044 + ], + "credit_map": [ + 0.007145442999899387, + 0.011532173492014408, + 0.010201283730566502, + 0.01600801944732666, + 0.010114004835486412, + 0.009112559258937836, + 0.05005420744419098, + 0.03320800140500069, + 0.013879305683076382, + 0.017818698659539223, + 0.07939169555902481, + 0.14254049956798553, + 0.018085258081555367, + 0.014198627322912216, + 0.01711471565067768, + 0.014034129679203033, + 0.0022400536108762026, + 0.003630557795986533, + 0.00526385335251689, + 0.005719160661101341, + 0.003397096646949649, + 0.0035587784368544817, + 0.006734437309205532, + 0.005546308122575283, + 0.00254820310510695, + 0.0038334298878908157, + 0.005221142433583736, + 0.005036704242229462, + 0.004804173018783331, + 0.005391914397478104, + 0.00893029198050499, + 0.003794769523665309, + 0.011847048997879028, + 0.011292319744825363, + 0.05142216011881828, + 0.011933690868318081, + 0.006845832336694002, + 0.009890850633382797, + 0.009687487967312336, + 0.040263596922159195, + 0.002688534324988723, + 0.0034737405367195606, + 0.004978197626769543, + 0.004372199531644583, + 0.005912511143833399, + 0.007014538627117872, + 0.0065720705315470695, + 0.0032774510327726603, + 0.006193549372255802, + 0.01020585373044014, + 0.020609838888049126, + 0.018999824300408363, + 0.012223616242408752, + 0.013012423180043697, + 0.013785413466393948, + 0.010690201073884964, + 0.007793464232236147, + 0.009166423231363297, + 0.01805965043604374, + 0.01340305246412754, + 0.017450157552957535, + 0.014603580348193645, + 0.013227211311459541, + 0.014959569089114666, + 0.0020335649605840445, + 0.003206754568964243, + 0.005949366372078657, + 0.0068938229233026505, + 0.003820819081738591, + 0.002820475958287716, + 0.004628426861017942, + 0.002662558574229479, + 0.002298054052516818, + 0.0023954869247972965, + 0.0030737430788576603, + 0.002636853838339448, + 0.004107549786567688, + 0.004516424145549536, + 0.001356022898107767, + 0.0016545375110581517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.9743272662162781 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned low to the ground and moving forward at high speed along parallel steel rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the steel rails reflect a warm, dazzling glow from a golden late afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dark wooden sleepers and gravel ballast visible as blurred, dynamic streaks moving backwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is withered grass visible swaying gently on either side of the tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rails stretch into the distance and converge at a vanishing point on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7084572315216064, + 0.020201226696372032, + 0.0171817597001791, + 0.03784382715821266, + 0.04577101394534111, + 0.025258462876081467, + 0.029945015907287598, + 0.039042189717292786, + 0.05199628323316574, + 0.024302955716848373 + ], + "credit_map": [ + 0.0247288066893816, + 0.01953628472983837, + 0.13766250014305115, + 0.04727936536073685, + 0.03160249441862106, + 0.0351458340883255, + 0.30744117498397827, + 0.10506079345941544, + 0.0012050087098032236, + 0.0029077460058033466, + 0.0021157795563340187, + 0.0027793042827397585, + 0.0020587474573403597, + 0.0022672482300549746, + 0.0033017443493008614, + 0.0035656490363180637, + 0.0014748189132660627, + 0.002064394298940897, + 0.0023184919264167547, + 0.0027228323742747307, + 0.0022674163337796926, + 0.00229459791444242, + 0.0021493593230843544, + 0.0018898497801274061, + 0.0037238923832774162, + 0.00571157643571496, + 0.005780714098364115, + 0.004216798115521669, + 0.004321914166212082, + 0.006402033381164074, + 0.003897865768522024, + 0.003789037698879838, + 0.0029165223240852356, + 0.004295120947062969, + 0.010509287938475609, + 0.005200030282139778, + 0.004055339843034744, + 0.004439695738255978, + 0.004216288682073355, + 0.010138731449842453, + 0.0023157952819019556, + 0.0031362443696707487, + 0.0031577167101204395, + 0.002759132068604231, + 0.0036575766280293465, + 0.0037428992800414562, + 0.0031568356789648533, + 0.003332266118377447, + 0.002691410481929779, + 0.002522370545193553, + 0.004689881112426519, + 0.003999343607574701, + 0.003157584462314844, + 0.003284078324213624, + 0.003642522729933262, + 0.005957828834652901, + 0.004256289452314377, + 0.0042510731145739555, + 0.007893353700637817, + 0.0056793056428432465, + 0.004458525683730841, + 0.004154296591877937, + 0.0034169210121035576, + 0.004932427778840065, + 0.003134901402518153, + 0.00609892513602972, + 0.009789303876459599, + 0.011591358110308647, + 0.0065768142230808735, + 0.004098596516996622, + 0.006782936863601208, + 0.003923454787582159, + 0.0023102129343897104, + 0.0029710831586271524, + 0.004225562792271376, + 0.0023676559794694185, + 0.005204200744628906, + 0.0028980732895433903, + 0.0020281532779335976, + 0.0022980140056461096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.02058546058833599 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_tensor([4321, 4321])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_tensor([4321, 4321])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e4298595ead7167fa58ef2b04df856b5d818a7fa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_tensor([4321, 4321])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a world steeped in the hallmarks of Disney Pixar animation, a close-up of the polished chrome nozzle of a steam cleaner shifts dizzyingly rapidly between sharp ice crystal details and a blurred background. The steam cleaner itself boasts a sleek, friendly design, housed in a bright lemon-yellow plastic shell and fitted with a chunky, ergonomic black handle. When activated, a torrent of what should be high-temperature, high-pressure steam erupts from the nozzle. However, this swirling cloud of steam doesn't clean anything. Instead, upon contact with the surrounding air, it miraculously condenses into countless shimmering ice crystals. These crystals, intricately structured, take on the forms of geometrically perfect snowflakes and feathery, delicate frost patterns. They hang suspended in mid-air, shimmering with a translucent, prismatic brilliance like diamond dust. The camera rapidly locks onto the surreal details of a single ice crystal, rendering its edges sharp and clear, while the warm, slightly cluttered Pixar-esque garage background behind it (filled with wooden workbenches and colorful toolboxes) completely blurs into a soft mass of color. Then, the focus quickly pulls back to a blur, allowing the entire shimmering ice crystal cloud to take center stage. With each focus shift, a subtle, brief flicker appears across the screen, as if magic has disturbed the stability of reality, creating a strange atmosphere full of wonder, whimsy, and visual shock.", + "seed": 30348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30348_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08336853543296456, + "vlm_logprob": -0.08336853543296456, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.08336853543296456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10220937430858612 + ], + "opposite_logprob_tokens": [ + -2.3522093296051025 + ], + "logprob_reward": -0.10220937430858612, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10220937430858612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28090333938598633, + 0.05715826898813248, + 0.04651279374957085, + 0.060467369854450226, + 0.040311358869075775, + 0.1535239964723587, + 0.08395414054393768, + 0.10927018523216248, + 0.12145191431045532, + 0.046446602791547775 + ], + "credit_map": [ + 0.03670251742005348, + 0.0103776128962636, + 0.08994705229997635, + 0.011826650239527225, + 0.009064586833119392, + 0.022438453510403633, + 0.008946554735302925, + 0.09159991890192032, + 0.014331697486341, + 0.009874104522168636, + 0.008257508277893066, + 0.005476682912558317, + 0.0036200706381350756, + 0.0036892914213240147, + 0.006907632574439049, + 0.005001284182071686, + 0.004646207205951214, + 0.008050553500652313, + 0.011141560971736908, + 0.0025944060180336237, + 0.0020954362116754055, + 0.004158772993832827, + 0.0030929872300475836, + 0.010732866823673248, + 0.014661546796560287, + 0.004132329486310482, + 0.005254500079900026, + 0.0028061717748641968, + 0.002827384974807501, + 0.008117317222058773, + 0.004362634848803282, + 0.01830548793077469, + 0.005098366178572178, + 0.004090857692062855, + 0.0027133224066346884, + 0.0033641369082033634, + 0.001766261993907392, + 0.004581325687468052, + 0.010363943874835968, + 0.008333141915500164, + 0.02022876776754856, + 0.01902633160352707, + 0.018589671701192856, + 0.016679387539625168, + 0.017689282074570656, + 0.021752148866653442, + 0.02257627248764038, + 0.01698214001953602, + 0.008545429445803165, + 0.009502039290964603, + 0.008292069658637047, + 0.010339786298573017, + 0.016298547387123108, + 0.01037555281072855, + 0.011337360367178917, + 0.009263358078897, + 0.012414952740073204, + 0.011751433834433556, + 0.012701045721769333, + 0.010423235595226288, + 0.02260342612862587, + 0.006739342585206032, + 0.012316951528191566, + 0.020319798961281776, + 0.0192073006182909, + 0.01293972134590149, + 0.012470060028135777, + 0.011480102315545082, + 0.02220529317855835, + 0.014186238870024681, + 0.015872376039624214, + 0.013090821914374828, + 0.004950140602886677, + 0.00749176787212491, + 0.0036858036182820797, + 0.007086237892508507, + 0.00700558302924037, + 0.005744062829762697, + 0.0052858940325677395, + 0.005197115242481232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.10220937430858612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647731505334377 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647731505334377, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647731505334377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6914587020874023, + 0.11465191096067429, + 0.026946676895022392, + 0.011442974209785461, + 0.025736279785633087, + 0.026505276560783386, + 0.025913668796420097, + 0.035694342106580734, + 0.025950701907277107, + 0.01569952443242073 + ], + "credit_map": [ + 0.08967168629169464, + 0.0214585792273283, + 0.22768688201904297, + 0.029067223891615868, + 0.022308850660920143, + 0.05511362850666046, + 0.021327102556824684, + 0.2248247116804123, + 0.01352695096284151, + 0.015615910291671753, + 0.02795219235122204, + 0.007677172310650349, + 0.012055556289851665, + 0.007298927288502455, + 0.010903691872954369, + 0.01962149702012539, + 0.0028652916662395, + 0.0014030648162588477, + 0.007396313827484846, + 0.0012552703265100718, + 0.0009534177952446043, + 0.0025013729464262724, + 0.0018077557906508446, + 0.008764185011386871, + 0.0012463132152333856, + 0.0013538763159886003, + 0.0016433732816949487, + 0.000861982349306345, + 0.0009863530285656452, + 0.002267958130687475, + 0.0011758571490645409, + 0.0019072593422606587, + 0.0030983095057308674, + 0.002418381394818425, + 0.0019075090531259775, + 0.0019371391972526908, + 0.0010585678974166512, + 0.0017429060535505414, + 0.007761807646602392, + 0.005811659153550863, + 0.003476054174825549, + 0.0034360664431005716, + 0.003512799507007003, + 0.0027821562252938747, + 0.0030262728687375784, + 0.003494278760626912, + 0.0038467596750706434, + 0.0029308858793228865, + 0.003925944212824106, + 0.002442165045067668, + 0.003876115195453167, + 0.0023876025807112455, + 0.0034210816957056522, + 0.0035605267621576786, + 0.00289167370647192, + 0.003408556804060936, + 0.0042629181407392025, + 0.004907035268843174, + 0.004288191441446543, + 0.0044671897776424885, + 0.005867979023605585, + 0.002774512395262718, + 0.005061641335487366, + 0.004064867738634348, + 0.0044705490581691265, + 0.002369658090174198, + 0.003233515191823244, + 0.002731022424995899, + 0.0022909902036190033, + 0.002893101656809449, + 0.002205764641985297, + 0.005756097380071878, + 0.001413106918334961, + 0.002579523017629981, + 0.0011440294329077005, + 0.0019822572357952595, + 0.0028266801964491606, + 0.002129036234691739, + 0.0016860472969710827, + 0.0019388427026569843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.008647731505334377 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028190179727971554 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028190179727971554, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028190179727971554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21472060680389404, + 0.08867186307907104, + 0.06067653372883797, + 0.03799833729863167, + 0.03846528008580208, + 0.14089342951774597, + 0.07612840086221695, + 0.1141851618885994, + 0.15719205141067505, + 0.07106836140155792 + ], + "credit_map": [ + 0.028404615819454193, + 0.007828578352928162, + 0.06747233122587204, + 0.010031708516180515, + 0.010052187368273735, + 0.01647796295583248, + 0.009314758703112602, + 0.06513845175504684, + 0.010362425819039345, + 0.00863201916217804, + 0.02262292243540287, + 0.005831143353134394, + 0.0105816088616848, + 0.007167433854192495, + 0.006045373156666756, + 0.017428936436772346, + 0.008945965208113194, + 0.0037891811225563288, + 0.017400821670889854, + 0.0035144665744155645, + 0.002632573712617159, + 0.004656067118048668, + 0.004198763519525528, + 0.015538694337010384, + 0.004257861990481615, + 0.003570268163457513, + 0.005674439482390881, + 0.0028092055581510067, + 0.0024718069471418858, + 0.009245147928595543, + 0.004291829653084278, + 0.00567777780815959, + 0.0064771585166454315, + 0.007766412105411291, + 0.0041571413166821, + 0.0033976277336478233, + 0.0030423488933593035, + 0.004889242816716433, + 0.004559612832963467, + 0.004175734706223011, + 0.01976247876882553, + 0.01699773594737053, + 0.017705975100398064, + 0.015634529292583466, + 0.014636210165917873, + 0.019228950142860413, + 0.02138415537774563, + 0.015543391928076744, + 0.009858536534011364, + 0.008196454495191574, + 0.009895413182675838, + 0.006741809193044901, + 0.011510477401316166, + 0.01032650750130415, + 0.009976837784051895, + 0.009622365236282349, + 0.014430289156734943, + 0.015359662473201752, + 0.014864044263958931, + 0.015368122607469559, + 0.01669369637966156, + 0.008904756046831608, + 0.014665624126791954, + 0.013898965902626514, + 0.01903449557721615, + 0.017031274735927582, + 0.011857029050588608, + 0.017385195940732956, + 0.028214629739522934, + 0.019247455522418022, + 0.023141220211982727, + 0.021280750632286072, + 0.008805692195892334, + 0.01212537381798029, + 0.005614590365439653, + 0.009249201975762844, + 0.010032445192337036, + 0.009712202474474907, + 0.007330062799155712, + 0.008198797702789307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0028190179727971554 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038262844085693 + ], + "opposite_logprob_tokens": [ + -1.910382628440857 + ], + "logprob_reward": -0.16038262844085693, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038262844085693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24496924877166748, + 0.03585747629404068, + 0.07556048035621643, + 0.10824994742870331, + 0.085657499730587, + 0.05156168341636658, + 0.16022521257400513, + 0.13525673747062683, + 0.07166311144828796, + 0.03099854849278927 + ], + "credit_map": [ + 0.0322636216878891, + 0.009440235793590546, + 0.07756836712360382, + 0.009897630661725998, + 0.008655845187604427, + 0.020826298743486404, + 0.008386136963963509, + 0.07793112844228745, + 0.004881254862993956, + 0.007308714557439089, + 0.0061354925855994225, + 0.0024320653174072504, + 0.004513781517744064, + 0.0030437957029789686, + 0.002970591653138399, + 0.004571786615997553, + 0.01679255999624729, + 0.006320618558675051, + 0.007257707882672548, + 0.003410539822652936, + 0.0038262244779616594, + 0.010581083595752716, + 0.004638373386114836, + 0.022733379155397415, + 0.025767963379621506, + 0.010027647018432617, + 0.009007375687360764, + 0.005752816330641508, + 0.005491866264492273, + 0.01258940901607275, + 0.007059967610985041, + 0.03255290910601616, + 0.012884643860161304, + 0.012553964741528034, + 0.008923281915485859, + 0.005703656002879143, + 0.005962458439171314, + 0.013856201432645321, + 0.010250288061797619, + 0.01552300900220871, + 0.010419915430247784, + 0.005041934549808502, + 0.0066821337677538395, + 0.0051149241626262665, + 0.004696995951235294, + 0.006633538752794266, + 0.007374567445367575, + 0.005597678478807211, + 0.0318254716694355, + 0.02245202288031578, + 0.017751839011907578, + 0.0202915258705616, + 0.014580853283405304, + 0.015769055113196373, + 0.014544841833412647, + 0.02300960011780262, + 0.0108281085267663, + 0.012025315314531326, + 0.016918564215302467, + 0.010781756602227688, + 0.022079359740018845, + 0.011077631264925003, + 0.014373860321938992, + 0.03717214986681938, + 0.011078561656177044, + 0.006551623810082674, + 0.008807491511106491, + 0.006545600946992636, + 0.007312208414077759, + 0.008314764127135277, + 0.00792972557246685, + 0.015123138204216957, + 0.0029016488697379827, + 0.004606145434081554, + 0.0029618716798722744, + 0.0044572073966264725, + 0.004739786498248577, + 0.003911762032657862, + 0.003026270307600498, + 0.004393858835101128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.16038262844085693 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278392493724823 + ], + "opposite_logprob_tokens": [ + -2.0177838802337646 + ], + "logprob_reward": -0.14278392493724823, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278392493724823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12837255001068115, + 0.08800806850194931, + 0.04508206620812416, + 0.1161268875002861, + 0.1228342056274414, + 0.06115715950727463, + 0.049893818795681, + 0.0988948792219162, + 0.20085158944129944, + 0.08877868950366974 + ], + "credit_map": [ + 0.016921384260058403, + 0.005269088316708803, + 0.03949534147977829, + 0.005056499503552914, + 0.006199517287313938, + 0.01112035010010004, + 0.004963948857039213, + 0.039346449077129364, + 0.010283251293003559, + 0.00917584728449583, + 0.02501414529979229, + 0.005765928421169519, + 0.005302038509398699, + 0.006629565730690956, + 0.007479603867977858, + 0.01835770159959793, + 0.00978576485067606, + 0.004192452412098646, + 0.005824450869113207, + 0.002632659627124667, + 0.00355923967435956, + 0.005903536453843117, + 0.0031384651083499193, + 0.010045505128800869, + 0.02352573722600937, + 0.008795920759439468, + 0.012972341850399971, + 0.007218428887426853, + 0.006693942006677389, + 0.0156141621991992, + 0.009106920100748539, + 0.03219945728778839, + 0.010571404360234737, + 0.010992138646543026, + 0.00622542155906558, + 0.007944977842271328, + 0.004952187649905682, + 0.008854632265865803, + 0.04063887894153595, + 0.03265459090471268, + 0.012704750522971153, + 0.007352162152528763, + 0.0074547333642840385, + 0.006353534758090973, + 0.005675410386174917, + 0.007509133778512478, + 0.007744819391518831, + 0.006362630985677242, + 0.006442508660256863, + 0.007822580635547638, + 0.007037417963147163, + 0.00509087136015296, + 0.00540899857878685, + 0.005058564245700836, + 0.005445938091725111, + 0.007586952298879623, + 0.012091120705008507, + 0.01002501230686903, + 0.012218263931572437, + 0.007863380014896393, + 0.018317757174372673, + 0.008681645616889, + 0.009135058149695396, + 0.020562665536999702, + 0.03565913066267967, + 0.014815245755016804, + 0.026767682284116745, + 0.019797326996922493, + 0.01742514967918396, + 0.01836320199072361, + 0.021607397124171257, + 0.04641648381948471, + 0.01280222274363041, + 0.011278736405074596, + 0.008024639450013638, + 0.014032729901373386, + 0.009165192022919655, + 0.01101761031895876, + 0.010006878525018692, + 0.012450695037841797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.14278392493724823 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a world steeped in the hallmarks of Disney Pixar animation, a close-up of the polished chrome nozzle of a steam cleaner shifts dizzyingly rapidly between sharp ice crystal details and a blurred background. The steam cleaner itself boasts a sleek, friendly design, housed in a bright lemon-yellow plastic shell and fitted with a chunky, ergonomic black handle. When activated, a torrent of what should be high-temperature, high-pressure steam erupts from the nozzle. However, this swirling cloud of steam doesn't clean anything. Instead, upon contact with the surrounding air, it miraculously condenses into countless shimmering ice crystals. These crystals, intricately structured, take on the forms of geometrically perfect snowflakes and feathery, delicate frost patterns. They hang suspended in mid-air, shimmering with a translucent, prismatic brilliance like diamond dust. The camera rapidly locks onto the surreal details of a single ice crystal, rendering its edges sharp and clear, while the warm, slightly cluttered Pixar-esque garage background behind it (filled with wooden workbenches and colorful toolboxes) completely blurs into a soft mass of color. Then, the focus quickly pulls back to a blur, allowing the entire shimmering ice crystal cloud to take center stage. With each focus shift, a subtle, brief flicker appears across the screen, as if magic has disturbed the stability of reality, creating a strange atmosphere full of wonder, whimsy, and visual shock.", + "seed": 30349, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30349_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0546440891455859, + "vlm_logprob": -0.0546440891455859, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.0546440891455859, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08174272626638412 + ], + "opposite_logprob_tokens": [ + -2.581742763519287 + ], + "logprob_reward": -0.08174272626638412, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08174272626638412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09904681891202927, + 0.06664401292800903, + 0.26187506318092346, + 0.047563619911670685, + 0.10588344186544418, + 0.033892638981342316, + 0.07575158029794693, + 0.14581218361854553, + 0.12331986427307129, + 0.04021081700921059 + ], + "credit_map": [ + 0.01485308539122343, + 0.014943122863769531, + 0.01753058284521103, + 0.008898554369807243, + 0.010633689351379871, + 0.009637393057346344, + 0.009941676631569862, + 0.012608702294528484, + 0.007816017605364323, + 0.007127588149160147, + 0.012979206629097462, + 0.010196956805884838, + 0.005917922593653202, + 0.0057604494504630566, + 0.010587937198579311, + 0.0062579261139035225, + 0.01594737358391285, + 0.005170354153960943, + 0.10001043230295181, + 0.01447608508169651, + 0.008385480381548405, + 0.005528043489903212, + 0.10025608539581299, + 0.012101175263524055, + 0.008005770854651928, + 0.005149172618985176, + 0.005154345650225878, + 0.008640303276479244, + 0.004341555759310722, + 0.005207976792007685, + 0.004697150085121393, + 0.006367339286953211, + 0.008582201786339283, + 0.010323584079742432, + 0.01777668297290802, + 0.027429848909378052, + 0.014979743398725986, + 0.005463197827339172, + 0.014718789607286453, + 0.006609372794628143, + 0.006039366591721773, + 0.00398879311978817, + 0.004940449725836515, + 0.0027145049534738064, + 0.0041525475680828094, + 0.004531684797257185, + 0.003613573033362627, + 0.003911713138222694, + 0.010475131683051586, + 0.00672578951343894, + 0.016526928171515465, + 0.008437995798885822, + 0.00836829375475645, + 0.008411368355154991, + 0.007143452297896147, + 0.009662616066634655, + 0.014543288387358189, + 0.009233511053025723, + 0.04667573794722557, + 0.004994253162294626, + 0.006934942211955786, + 0.045429300516843796, + 0.008871136233210564, + 0.009129993617534637, + 0.02395710162818432, + 0.016115594655275345, + 0.02538974955677986, + 0.008440162055194378, + 0.013733533211052418, + 0.010986457578837872, + 0.013276885263621807, + 0.011420364491641521, + 0.004247317556291819, + 0.003468718146905303, + 0.006423152517527342, + 0.0036337063647806644, + 0.004682118073105812, + 0.004370979964733124, + 0.00573448371142149, + 0.007650338113307953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.08174272626638412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270752124488354 + ], + "opposite_logprob_tokens": [ + -5.255270957946777 + ], + "logprob_reward": -0.005270752124488354, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270752124488354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16728544235229492, + 0.10612437129020691, + 0.17063972353935242, + 0.022936929017305374, + 0.03921745344996452, + 0.036886025220155716, + 0.040724948048591614, + 0.3355676531791687, + 0.0471489280462265, + 0.03346853703260422 + ], + "credit_map": [ + 0.02135522849857807, + 0.02508118376135826, + 0.016752006486058235, + 0.020651381462812424, + 0.02045164816081524, + 0.020095612853765488, + 0.016504928469657898, + 0.026393450796604156, + 0.014262962155044079, + 0.011237305589020252, + 0.01825835183262825, + 0.016545046120882034, + 0.011473439633846283, + 0.01003018394112587, + 0.01488414779305458, + 0.009432932361960411, + 0.009665158577263355, + 0.00530497869476676, + 0.06225736066699028, + 0.010047370567917824, + 0.00608125701546669, + 0.004457180853933096, + 0.0645524337887764, + 0.008273975923657417, + 0.0027864589355885983, + 0.0026570837944746017, + 0.002710629953071475, + 0.0032990318723022938, + 0.0021179048344492912, + 0.0020097659435123205, + 0.0034266251605004072, + 0.003929425962269306, + 0.004515405278652906, + 0.004078129306435585, + 0.004476030822843313, + 0.006491995416581631, + 0.005165511276572943, + 0.0038521247915923595, + 0.005665790289640427, + 0.004972466733306646, + 0.005516681354492903, + 0.003689014818519354, + 0.0043495637364685535, + 0.00428055739030242, + 0.003105451352894306, + 0.0042050606571137905, + 0.007275562733411789, + 0.00446413503959775, + 0.006077928002923727, + 0.0033921045251190662, + 0.007956803776323795, + 0.0035258368588984013, + 0.005891142878681421, + 0.006010250188410282, + 0.004245378077030182, + 0.003625499317422509, + 0.030681187286973, + 0.01586136966943741, + 0.12833525240421295, + 0.011173054575920105, + 0.01471793558448553, + 0.10599590092897415, + 0.011779233813285828, + 0.017023708671331406, + 0.01187232043594122, + 0.003264560829848051, + 0.01237731333822012, + 0.0030768204014748335, + 0.0035787848755717278, + 0.004269199911504984, + 0.003821165068075061, + 0.00488876411691308, + 0.0029983657877892256, + 0.0020569697953760624, + 0.008326786570250988, + 0.0022308845072984695, + 0.0034198956564068794, + 0.004091816022992134, + 0.003431187476962805, + 0.006912633311003447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.005270752124488354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031953013967722654 + ], + "opposite_logprob_tokens": [ + -5.753195285797119 + ], + "logprob_reward": -0.0031953013967722654, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031953013967722654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05538048595190048, + 0.0669357106089592, + 0.1619771271944046, + 0.06711413711309433, + 0.05823155865073204, + 0.08636008203029633, + 0.11144737154245377, + 0.2529385983943939, + 0.07610885053873062, + 0.06350606679916382 + ], + "credit_map": [ + 0.0071176099590957165, + 0.007136891596019268, + 0.008072754368185997, + 0.007431939244270325, + 0.006234709173440933, + 0.006005716975778341, + 0.005577569827437401, + 0.007803292013704777, + 0.005428168457001448, + 0.006660137325525284, + 0.015503652393817902, + 0.008157315663993359, + 0.006726779509335756, + 0.007866227068006992, + 0.010959689505398273, + 0.00563373789191246, + 0.01208385918289423, + 0.007216128520667553, + 0.04701782763004303, + 0.015629159286618233, + 0.008272006176412106, + 0.006538829300552607, + 0.05371876060962677, + 0.011500570923089981, + 0.008619183674454689, + 0.005890985950827599, + 0.007413551211357117, + 0.010669635608792305, + 0.004798572510480881, + 0.006608646363019943, + 0.009754719212651253, + 0.013358842581510544, + 0.007803929038345814, + 0.0041311560198664665, + 0.009323458187282085, + 0.014612031169235706, + 0.0037975285667926073, + 0.0044361804611980915, + 0.009304146282374859, + 0.004823131021112204, + 0.015083885751664639, + 0.010675577446818352, + 0.009499401785433292, + 0.00798401702195406, + 0.011515210382640362, + 0.011103099212050438, + 0.010436530224978924, + 0.010062364861369133, + 0.016850562766194344, + 0.012663733214139938, + 0.01419382356107235, + 0.017379630357027054, + 0.010823954828083515, + 0.008676329627633095, + 0.01572287827730179, + 0.015136458911001682, + 0.023220794275403023, + 0.011372808367013931, + 0.08658420294523239, + 0.0092709856107831, + 0.011457137763500214, + 0.08932936191558838, + 0.007994074374437332, + 0.013709218241274357, + 0.00946824997663498, + 0.008223249576985836, + 0.012316749431192875, + 0.007610965054482222, + 0.012426656670868397, + 0.007961275056004524, + 0.01203478779643774, + 0.0060669188387691975, + 0.006372304633259773, + 0.004156776238232851, + 0.009232631884515285, + 0.004265886265784502, + 0.00697745755314827, + 0.007366866338998079, + 0.007657619658857584, + 0.0174765195697546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.0031953013967722654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129947379231453 + ], + "opposite_logprob_tokens": [ + -2.23799467086792 + ], + "logprob_reward": -0.1129947379231453, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129947379231453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054794155061244965, + 0.022996118292212486, + 0.12735417485237122, + 0.0359533354640007, + 0.05129401385784149, + 0.01735575683414936, + 0.07419119030237198, + 0.4338100254535675, + 0.0790594294667244, + 0.1031917929649353 + ], + "credit_map": [ + 0.01149740070104599, + 0.007996405474841595, + 0.007803504820913076, + 0.004301465582102537, + 0.0060685561038553715, + 0.005282687488943338, + 0.005668750032782555, + 0.0061753797344863415, + 0.0026274407282471657, + 0.0025561945512890816, + 0.0039501977153122425, + 0.0034178101923316717, + 0.0029232732485979795, + 0.0022278071846812963, + 0.0032204980961978436, + 0.0020728977397084236, + 0.007482807617634535, + 0.003681748639792204, + 0.04689163342118263, + 0.008403974585235119, + 0.0050382730551064014, + 0.0035429352428764105, + 0.04606456682085991, + 0.006248227320611477, + 0.0050321342423558235, + 0.003987861797213554, + 0.004451550543308258, + 0.0062178042717278, + 0.0029897578060626984, + 0.0027498165145516396, + 0.004816770553588867, + 0.005707638803869486, + 0.005671067629009485, + 0.004177385941147804, + 0.007180691696703434, + 0.01061500608921051, + 0.006521961651742458, + 0.005005704239010811, + 0.006802764721214771, + 0.005319430027157068, + 0.002496057190001011, + 0.0020878163632005453, + 0.001976862782612443, + 0.001838133204728365, + 0.0022886847145855427, + 0.0020389705896377563, + 0.0021972788963466883, + 0.00243195379152894, + 0.013693227432668209, + 0.0055490476079285145, + 0.015865707769989967, + 0.0033608246594667435, + 0.006457120645791292, + 0.012488404288887978, + 0.004425753373652697, + 0.01235110592097044, + 0.05273178592324257, + 0.03636045381426811, + 0.13633961975574493, + 0.020001213997602463, + 0.021786272525787354, + 0.09228043258190155, + 0.0324469693005085, + 0.04186326265335083, + 0.017902767285704613, + 0.009484576992690563, + 0.01620054431259632, + 0.005019763018935919, + 0.007154207676649094, + 0.006645418703556061, + 0.006431609857827425, + 0.010220540687441826, + 0.010660608299076557, + 0.006623681169003248, + 0.01876552402973175, + 0.00597794447094202, + 0.0068506477400660515, + 0.010769483633339405, + 0.0070214662700891495, + 0.03652244433760643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.1129947379231453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001692801713943 + ], + "opposite_logprob_tokens": [ + -2.695016860961914 + ], + "logprob_reward": -0.07001692801713943, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001692801713943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04830010235309601, + 0.04382549226284027, + 0.4006563425064087, + 0.07968620955944061, + 0.0993662104010582, + 0.06081768497824669, + 0.04151003062725067, + 0.1305394321680069, + 0.07659280300140381, + 0.01870565116405487 + ], + "credit_map": [ + 0.007773167919367552, + 0.005804191343486309, + 0.00870304275304079, + 0.004098627250641584, + 0.004989305976778269, + 0.005120390560477972, + 0.005154436454176903, + 0.006656935904175043, + 0.004726638551801443, + 0.004427776671946049, + 0.006828255020081997, + 0.005675089545547962, + 0.006430597510188818, + 0.004845843184739351, + 0.0059616402722895145, + 0.004929651506245136, + 0.02321617491543293, + 0.00951700285077095, + 0.1471097767353058, + 0.025722311809659004, + 0.013690500520169735, + 0.009433474391698837, + 0.15239520370960236, + 0.019571904093027115, + 0.011265778914093971, + 0.007905598729848862, + 0.00963895209133625, + 0.012780473567545414, + 0.006059582345187664, + 0.007857517339289188, + 0.010531257838010788, + 0.013647050596773624, + 0.009737463667988777, + 0.010005293413996696, + 0.013153032399713993, + 0.017069343477487564, + 0.014499383047223091, + 0.008536695502698421, + 0.01485314778983593, + 0.011511855758726597, + 0.01050716545432806, + 0.0064820293337106705, + 0.007630784995853901, + 0.005746661219745874, + 0.007207754533737898, + 0.010456499643623829, + 0.005420527420938015, + 0.0073662688955664635, + 0.007154725957661867, + 0.004167656414210796, + 0.006814572494477034, + 0.003113010199740529, + 0.005792809184640646, + 0.004886251874268055, + 0.00395779637619853, + 0.00562320789322257, + 0.012313214130699635, + 0.00565013661980629, + 0.04543919116258621, + 0.004114829003810883, + 0.005502177402377129, + 0.04340435564517975, + 0.0036885461304336786, + 0.01042698323726654, + 0.019772179424762726, + 0.005230730399489403, + 0.020639382302761078, + 0.0046789320185780525, + 0.004631910007447004, + 0.006287384312599897, + 0.00632928591221571, + 0.00902299489825964, + 0.0025987117551267147, + 0.0012205473612993956, + 0.0034157379996031523, + 0.0012437026016414165, + 0.0021369329188019037, + 0.002215890446677804, + 0.002949117450043559, + 0.0029250106308609247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.07001692801713943 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a world steeped in the hallmarks of Disney Pixar animation, a close-up of the polished chrome nozzle of a steam cleaner shifts dizzyingly rapidly between sharp ice crystal details and a blurred background. The steam cleaner itself boasts a sleek, friendly design, housed in a bright lemon-yellow plastic shell and fitted with a chunky, ergonomic black handle. When activated, a torrent of what should be high-temperature, high-pressure steam erupts from the nozzle. However, this swirling cloud of steam doesn't clean anything. Instead, upon contact with the surrounding air, it miraculously condenses into countless shimmering ice crystals. These crystals, intricately structured, take on the forms of geometrically perfect snowflakes and feathery, delicate frost patterns. They hang suspended in mid-air, shimmering with a translucent, prismatic brilliance like diamond dust. The camera rapidly locks onto the surreal details of a single ice crystal, rendering its edges sharp and clear, while the warm, slightly cluttered Pixar-esque garage background behind it (filled with wooden workbenches and colorful toolboxes) completely blurs into a soft mass of color. Then, the focus quickly pulls back to a blur, allowing the entire shimmering ice crystal cloud to take center stage. With each focus shift, a subtle, brief flicker appears across the screen, as if magic has disturbed the stability of reality, creating a strange atmosphere full of wonder, whimsy, and visual shock.", + "seed": 30350, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30350_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0276554387062788, + "vlm_logprob": -0.0276554387062788, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.0276554387062788, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0721849650144577 + ], + "opposite_logprob_tokens": [ + -2.6971850395202637 + ], + "logprob_reward": -0.0721849650144577, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0721849650144577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10901530086994171, + 0.028234289959073067, + 0.1394958645105362, + 0.14547084271907806, + 0.050809938460588455, + 0.07007461041212082, + 0.11913358420133591, + 0.19286023080348969, + 0.05804555490612984, + 0.08685978502035141 + ], + "credit_map": [ + 0.02000967226922512, + 0.01306171901524067, + 0.022185485810041428, + 0.0074367825873196125, + 0.013748342171311378, + 0.01044423971325159, + 0.012797430157661438, + 0.009331629611551762, + 0.0042116111144423485, + 0.003921848721802235, + 0.004160730168223381, + 0.002267081057652831, + 0.003713103709742427, + 0.0024780614767223597, + 0.003669830271974206, + 0.0038120232056826353, + 0.029386820271611214, + 0.02106727845966816, + 0.048670753836631775, + 0.011764219030737877, + 0.005041309166699648, + 0.006991381291300058, + 0.007371907122433186, + 0.009202202782034874, + 0.05029234290122986, + 0.013288727961480618, + 0.048136889934539795, + 0.007175212260335684, + 0.005287508945912123, + 0.004812321625649929, + 0.006335470359772444, + 0.010142376646399498, + 0.009755146689713001, + 0.01028541661798954, + 0.007526951842010021, + 0.006913306657224894, + 0.0038364576175808907, + 0.0034855420235544443, + 0.00426843948662281, + 0.004738680087029934, + 0.010150019079446793, + 0.011648126877844334, + 0.00913272611796856, + 0.007482429035007954, + 0.00529518024995923, + 0.006936156190931797, + 0.009040091186761856, + 0.010389882139861584, + 0.015778979286551476, + 0.017979232594370842, + 0.01509725209325552, + 0.018049510195851326, + 0.009952163323760033, + 0.008449563756585121, + 0.016888611018657684, + 0.01693827658891678, + 0.023024966940283775, + 0.02754184603691101, + 0.021561117842793465, + 0.024629078805446625, + 0.02288166619837284, + 0.021628975868225098, + 0.02650684490799904, + 0.02508573606610298, + 0.006040755659341812, + 0.008639651350677013, + 0.007804486900568008, + 0.00850451085716486, + 0.009786659851670265, + 0.0048536150716245174, + 0.005109917372465134, + 0.007305959239602089, + 0.008341807872056961, + 0.010946973226964474, + 0.012659041211009026, + 0.011771168559789658, + 0.006947470363229513, + 0.009343160316348076, + 0.013678260147571564, + 0.013171903789043427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.0721849650144577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420846302062273 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.01420846302062273, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420846302062273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18071207404136658, + 0.058293357491493225, + 0.13630960881710052, + 0.10812276601791382, + 0.06388036906719208, + 0.10615331679582596, + 0.14524699747562408, + 0.06246161088347435, + 0.07536941021680832, + 0.06345050781965256 + ], + "credit_map": [ + 0.023293284699320793, + 0.018675915896892548, + 0.025032833218574524, + 0.015775874257087708, + 0.034712158143520355, + 0.020647242665290833, + 0.01755654066801071, + 0.025018222630023956, + 0.009686267003417015, + 0.007451608311384916, + 0.007026007864624262, + 0.00619771983474493, + 0.010755207389593124, + 0.004785298369824886, + 0.0054703098721802235, + 0.006920939311385155, + 0.013073135167360306, + 0.02405085600912571, + 0.02711600251495838, + 0.01603609509766102, + 0.008245048113167286, + 0.012631415389478207, + 0.014918227680027485, + 0.020238833501935005, + 0.03676269203424454, + 0.009657605551183224, + 0.03435293585062027, + 0.005605950020253658, + 0.004767202772200108, + 0.004945763852447271, + 0.006404829677194357, + 0.005625783000141382, + 0.015199541114270687, + 0.0140640614554286, + 0.01164583582431078, + 0.00608007051050663, + 0.003909967839717865, + 0.004781093914061785, + 0.0034991460852324963, + 0.004700654186308384, + 0.016546398401260376, + 0.017905108630657196, + 0.01491836179047823, + 0.009227195754647255, + 0.008105402812361717, + 0.013446937315165997, + 0.010968991555273533, + 0.015034917742013931, + 0.021684007719159126, + 0.02378346212208271, + 0.01942436583340168, + 0.01753517985343933, + 0.017766792327165604, + 0.009818066842854023, + 0.020535480231046677, + 0.014699629507958889, + 0.008843609131872654, + 0.006494974251836538, + 0.007444070186465979, + 0.0060481298714876175, + 0.005719235632568598, + 0.0049063777551054955, + 0.010096616111695766, + 0.012908600270748138, + 0.012536131776869297, + 0.0099706519395113, + 0.00902609247714281, + 0.009025275707244873, + 0.012705545872449875, + 0.004898953251540661, + 0.007176566869020462, + 0.010030186735093594, + 0.0066368794068694115, + 0.008034855127334595, + 0.017398791387677193, + 0.009371960535645485, + 0.0038731531240046024, + 0.005440487992018461, + 0.007080537267029285, + 0.005613839719444513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.01420846302062273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538296170532703 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538296170532703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538296170532703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035488665103912354, + 0.014649768359959126, + 0.02128850854933262, + 0.03537193313241005, + 0.03398079797625542, + 0.0514247752726078, + 0.1436256468296051, + 0.5222620368003845, + 0.10563874989748001, + 0.0362691693007946 + ], + "credit_map": [ + 0.006326968315988779, + 0.005721708294004202, + 0.006563879083842039, + 0.0023885294795036316, + 0.00408310629427433, + 0.0035148293245583773, + 0.0033268204424530268, + 0.003562822937965393, + 0.002522006630897522, + 0.0023153896909207106, + 0.0020134281367063522, + 0.0018986890790984035, + 0.001712891273200512, + 0.0013975176261737943, + 0.0012092564720660448, + 0.0015805884031578898, + 0.002889604540541768, + 0.004076090175658464, + 0.0026890526060014963, + 0.0030537513084709644, + 0.0016613808693364263, + 0.0024760160595178604, + 0.001710534212179482, + 0.0027320771478116512, + 0.004814627580344677, + 0.007550607435405254, + 0.00691881263628602, + 0.0029048658907413483, + 0.0027827953454107046, + 0.003302369499579072, + 0.003160814056172967, + 0.003937041386961937, + 0.00736780883744359, + 0.0066732382401824, + 0.0060984487645328045, + 0.0036175507120788097, + 0.0022747907787561417, + 0.001868453691713512, + 0.0024235770106315613, + 0.003656926332041621, + 0.008268970064818859, + 0.009609070606529713, + 0.007835868746042252, + 0.006303231697529554, + 0.0035981761757284403, + 0.003734569763764739, + 0.00592507841065526, + 0.006149807944893837, + 0.021562503650784492, + 0.0259719118475914, + 0.019397977739572525, + 0.02272074855864048, + 0.016837947070598602, + 0.00804787129163742, + 0.016194652765989304, + 0.012892024591565132, + 0.07350798696279526, + 0.07846308499574661, + 0.05724672973155975, + 0.06324081867933273, + 0.06729736179113388, + 0.04798385873436928, + 0.08029144257307053, + 0.0542307011783123, + 0.010225938633084297, + 0.01264998223632574, + 0.014943784102797508, + 0.017007648944854736, + 0.017192570492625237, + 0.010430279187858105, + 0.009295960888266563, + 0.013892577961087227, + 0.0035876447800546885, + 0.005066034849733114, + 0.005786150228232145, + 0.004683046601712704, + 0.0027865786105394363, + 0.0049197617918252945, + 0.004736607428640127, + 0.00470334105193615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.012538296170532703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064966097474098 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016064966097474098, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064966097474098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08839171379804611, + 0.05202416703104973, + 0.1861530840396881, + 0.0941527932882309, + 0.12091045826673508, + 0.05912788212299347, + 0.21505935490131378, + 0.050642795860767365, + 0.06650256365537643, + 0.06703516095876694 + ], + "credit_map": [ + 0.01559880468994379, + 0.0206404197961092, + 0.013345192186534405, + 0.008031592704355717, + 0.007896275259554386, + 0.008945021778345108, + 0.007192202843725681, + 0.006742200814187527, + 0.01008695363998413, + 0.00873273890465498, + 0.011188194155693054, + 0.005663329269737005, + 0.004436849150806665, + 0.004091271199285984, + 0.0031246596481651068, + 0.004700167570263147, + 0.04979830980300903, + 0.026519570499658585, + 0.03564334288239479, + 0.021570418030023575, + 0.007965859025716782, + 0.011548119597136974, + 0.013965657912194729, + 0.019141795113682747, + 0.029702162370085716, + 0.017106378450989723, + 0.020066218450665474, + 0.006316167768090963, + 0.005344938952475786, + 0.004837793298065662, + 0.00474591925740242, + 0.006033211946487427, + 0.034915752708911896, + 0.025710565969347954, + 0.019518155604600906, + 0.011939899995923042, + 0.005716805811971426, + 0.006855883169919252, + 0.006844807416200638, + 0.00940859317779541, + 0.008963276632130146, + 0.011026333086192608, + 0.007635381538420916, + 0.005715012550354004, + 0.003947695717215538, + 0.0057342564687132835, + 0.007488237228244543, + 0.008617689833045006, + 0.031110357493162155, + 0.03821811079978943, + 0.027428975328803062, + 0.030509812757372856, + 0.02337235026061535, + 0.013848290778696537, + 0.02837747149169445, + 0.022193986922502518, + 0.006798254791647196, + 0.004496337845921516, + 0.007463833317160606, + 0.00483344029635191, + 0.003644094569608569, + 0.007309114094823599, + 0.008235763758420944, + 0.007861954160034657, + 0.005355996545404196, + 0.008623434230685234, + 0.008552242070436478, + 0.011179954744875431, + 0.010823541320860386, + 0.0071451845578849316, + 0.005811789073050022, + 0.009010417386889458, + 0.00798498373478651, + 0.00891350582242012, + 0.016423502936959267, + 0.007429773919284344, + 0.004252863582223654, + 0.005862012039870024, + 0.009444729425013065, + 0.006723794154822826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.016064966097474098 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328050322830677 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.02328050322830677, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328050322830677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035956211388111115, + 0.019114116206765175, + 0.058260418474674225, + 0.12418732792139053, + 0.058013916015625, + 0.17073599994182587, + 0.1586207002401352, + 0.14418268203735352, + 0.10626621544361115, + 0.12466243654489517 + ], + "credit_map": [ + 0.006981608457863331, + 0.0033358640503138304, + 0.008404245600104332, + 0.002467396203428507, + 0.004647503141313791, + 0.003406162140890956, + 0.003438499290496111, + 0.0032749308738857508, + 0.003817545948550105, + 0.003063374664634466, + 0.0031158733181655407, + 0.0014874930493533611, + 0.0031008278019726276, + 0.00151402922347188, + 0.001297171227633953, + 0.0017178007401525974, + 0.010712671093642712, + 0.010480712167918682, + 0.006835269741714001, + 0.0070613669231534, + 0.0033629562240093946, + 0.004668382927775383, + 0.005537843331694603, + 0.009601209312677383, + 0.0465502068400383, + 0.015123960562050343, + 0.03428683429956436, + 0.007788386661559343, + 0.004693292547017336, + 0.004681475926190615, + 0.006390704773366451, + 0.004672459792345762, + 0.014503763988614082, + 0.01412374060600996, + 0.008342442102730274, + 0.005127080250531435, + 0.0027387102600187063, + 0.003091769991442561, + 0.003249716479331255, + 0.006836689542979002, + 0.02255188301205635, + 0.01791517250239849, + 0.028349192813038826, + 0.02209814079105854, + 0.012679696083068848, + 0.015818137675523758, + 0.024872254580259323, + 0.02645152248442173, + 0.024949131533503532, + 0.027558835223317146, + 0.020051132887601852, + 0.024212488904595375, + 0.015869366005063057, + 0.00986864510923624, + 0.019327623769640923, + 0.01678348146378994, + 0.020675845444202423, + 0.017862437292933464, + 0.01697245053946972, + 0.01636572554707527, + 0.017048107460141182, + 0.01973072439432144, + 0.020448269322514534, + 0.015079116448760033, + 0.010465634055435658, + 0.01680324785411358, + 0.014657003805041313, + 0.020952990278601646, + 0.0140186483040452, + 0.00738926324993372, + 0.01037670485675335, + 0.011602712795138359, + 0.009314225986599922, + 0.01704874262213707, + 0.027362190186977386, + 0.020155716687440872, + 0.008251353166997433, + 0.013075080700218678, + 0.01500917412340641, + 0.014445945620536804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.02328050322830677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a world steeped in the hallmarks of Disney Pixar animation, a close-up of the polished chrome nozzle of a steam cleaner shifts dizzyingly rapidly between sharp ice crystal details and a blurred background. The steam cleaner itself boasts a sleek, friendly design, housed in a bright lemon-yellow plastic shell and fitted with a chunky, ergonomic black handle. When activated, a torrent of what should be high-temperature, high-pressure steam erupts from the nozzle. However, this swirling cloud of steam doesn't clean anything. Instead, upon contact with the surrounding air, it miraculously condenses into countless shimmering ice crystals. These crystals, intricately structured, take on the forms of geometrically perfect snowflakes and feathery, delicate frost patterns. They hang suspended in mid-air, shimmering with a translucent, prismatic brilliance like diamond dust. The camera rapidly locks onto the surreal details of a single ice crystal, rendering its edges sharp and clear, while the warm, slightly cluttered Pixar-esque garage background behind it (filled with wooden workbenches and colorful toolboxes) completely blurs into a soft mass of color. Then, the focus quickly pulls back to a blur, allowing the entire shimmering ice crystal cloud to take center stage. With each focus shift, a subtle, brief flicker appears across the screen, as if magic has disturbed the stability of reality, creating a strange atmosphere full of wonder, whimsy, and visual shock.", + "seed": 30351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_7/index_4321_seed_30351_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1639159596990794, + "vlm_logprob": -0.1639159596990794, + "vlm_yes_no_margin": 3.125000023841858, + "vlm_token_credit": -0.1639159596990794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14669169485569 + ], + "opposite_logprob_tokens": [ + -2.0216917991638184 + ], + "logprob_reward": -0.14669169485569, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14669169485569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04423155635595322, + 0.05501260980963707, + 0.017872508615255356, + 0.07049314677715302, + 0.05749310925602913, + 0.02493082918226719, + 0.08308163285255432, + 0.037868961691856384, + 0.5638096332550049, + 0.045206036418676376 + ], + "credit_map": [ + 0.00510284723713994, + 0.006880465894937515, + 0.004154631868004799, + 0.0035964325070381165, + 0.0036623028572648764, + 0.007687248755246401, + 0.003547695931047201, + 0.009599926881492138, + 0.006249117199331522, + 0.006279489025473595, + 0.003939654678106308, + 0.004713522735983133, + 0.009247684851288795, + 0.011336726136505604, + 0.006456030532717705, + 0.006790378596633673, + 0.003179794643074274, + 0.0024421396665275097, + 0.002195375505834818, + 0.002264360897243023, + 0.002425247337669134, + 0.0017206902848556638, + 0.001705641858279705, + 0.00193925725761801, + 0.009128677658736706, + 0.009809264913201332, + 0.010213285684585571, + 0.008199688978493214, + 0.008047875948250294, + 0.009918669238686562, + 0.008943451568484306, + 0.006232226733118296, + 0.005973608698695898, + 0.004666289780288935, + 0.006090175826102495, + 0.012277566827833652, + 0.007272625807672739, + 0.004888053983449936, + 0.0071657272055745125, + 0.009159058332443237, + 0.0027574924752116203, + 0.0028709324542433023, + 0.0035775606520473957, + 0.0032043340615928173, + 0.0033992326352745295, + 0.0036377240903675556, + 0.002980581484735012, + 0.002502970164641738, + 0.005562781821936369, + 0.009360725060105324, + 0.019810710102319717, + 0.0155025664716959, + 0.013293696567416191, + 0.00526735233142972, + 0.009456763043999672, + 0.004827034659683704, + 0.0034875108394771814, + 0.005076071713119745, + 0.004454759880900383, + 0.0052961488254368305, + 0.005916826892644167, + 0.004067752510309219, + 0.006870494224131107, + 0.0026993954088538885, + 0.040458083152770996, + 0.04311816766858101, + 0.0703599601984024, + 0.07805121690034866, + 0.11666104197502136, + 0.08573011308908463, + 0.08216071128845215, + 0.047270286828279495, + 0.007549989502876997, + 0.0027936124242842197, + 0.006238914094865322, + 0.005491861142218113, + 0.0033583759795874357, + 0.006947718560695648, + 0.006660835817456245, + 0.00616472028195858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.14669169485569 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018214227631688118 + ], + "opposite_logprob_tokens": [ + -4.018214225769043 + ], + "logprob_reward": -0.018214227631688118, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018214227631688118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19691628217697144, + 0.11110419034957886, + 0.06255754828453064, + 0.14047230780124664, + 0.06150023266673088, + 0.049156952649354935, + 0.03116554021835327, + 0.046330127865076065, + 0.27619075775146484, + 0.024606097489595413 + ], + "credit_map": [ + 0.020312819629907608, + 0.022911012172698975, + 0.016016438603401184, + 0.01441569160670042, + 0.01788422092795372, + 0.04773006960749626, + 0.017682719975709915, + 0.039963312447071075, + 0.016371386125683784, + 0.014538192190229893, + 0.009186661802232265, + 0.010206840932369232, + 0.020257117226719856, + 0.014398601837456226, + 0.010606247931718826, + 0.0155391376465559, + 0.005968624260276556, + 0.007299419958144426, + 0.00473097525537014, + 0.007159326691180468, + 0.014468529261648655, + 0.008782490156590939, + 0.007363044656813145, + 0.006785138975828886, + 0.015856552869081497, + 0.018634650856256485, + 0.018442951142787933, + 0.017461024224758148, + 0.02040153555572033, + 0.020979546010494232, + 0.016002902761101723, + 0.012693136930465698, + 0.005461415741592646, + 0.006347368471324444, + 0.00424151960760355, + 0.013137930072844028, + 0.007313922978937626, + 0.007792857475578785, + 0.007264483720064163, + 0.009940738789737225, + 0.005073778796941042, + 0.005203546490520239, + 0.005332796834409237, + 0.006636058911681175, + 0.008591177873313427, + 0.006675360724329948, + 0.007036284077912569, + 0.004607948008924723, + 0.0024048665072768927, + 0.0032631896901875734, + 0.007495251018553972, + 0.005462282337248325, + 0.004596411250531673, + 0.001982539426535368, + 0.003580169752240181, + 0.0023808323312550783, + 0.0037109307013452053, + 0.007240119855850935, + 0.005583921913057566, + 0.007707774173468351, + 0.007309632375836372, + 0.0035064236726611853, + 0.00838776957243681, + 0.0028835556004196405, + 0.018756186589598656, + 0.021578466519713402, + 0.03381674736738205, + 0.03758135437965393, + 0.05979199334979057, + 0.040409334003925323, + 0.04067787528038025, + 0.02357877977192402, + 0.005426586605608463, + 0.0016013352433219552, + 0.004225029144436121, + 0.0024381238035857677, + 0.0013915825402364135, + 0.0034631197340786457, + 0.00242252997122705, + 0.003637790447100997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.018214227631688118 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036210936959832907 + ], + "opposite_logprob_tokens": [ + -5.6286211013793945 + ], + "logprob_reward": -0.0036210936959832907, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036210936959832907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03492143005132675, + 0.031170794740319252, + 0.011868420988321304, + 0.09781435132026672, + 0.056543100625276566, + 0.015320944599807262, + 0.02689313143491745, + 0.07287991791963577, + 0.5911266803741455, + 0.06146128103137016 + ], + "credit_map": [ + 0.004751876927912235, + 0.004044878296554089, + 0.0030535035766661167, + 0.002784453332424164, + 0.002911574440076947, + 0.007106224540621042, + 0.0029309196397662163, + 0.007337997667491436, + 0.003894251771271229, + 0.003970413934439421, + 0.002100309357047081, + 0.00281696580350399, + 0.006146360654383898, + 0.004742834251374006, + 0.003501567989587784, + 0.003998089581727982, + 0.0012711186427623034, + 0.0011137268738821149, + 0.0009022208978421986, + 0.0016053665895015001, + 0.0019234421197324991, + 0.0013600070960819721, + 0.0021465092431753874, + 0.0015460283029824495, + 0.010741980746388435, + 0.013614078052341938, + 0.012577087618410587, + 0.012947624549269676, + 0.013155391439795494, + 0.015177158638834953, + 0.0111501794308424, + 0.008450844325125217, + 0.005209063179790974, + 0.005279549863189459, + 0.004232166800647974, + 0.013821913860738277, + 0.006908175069838762, + 0.004592270124703646, + 0.004224055912345648, + 0.012275902554392815, + 0.0013260026462376118, + 0.002151239663362503, + 0.001768665388226509, + 0.0021655254531651735, + 0.0023435158655047417, + 0.002036828314885497, + 0.0018538598669692874, + 0.0016753067029640079, + 0.0036124165635555983, + 0.0040845004841685295, + 0.004949064925312996, + 0.003594738431274891, + 0.003091910621151328, + 0.0021643158979713917, + 0.0032796128652989864, + 0.0021165707148611546, + 0.005830449517816305, + 0.007632130291312933, + 0.011117243207991123, + 0.012134170159697533, + 0.011879357509315014, + 0.005937893874943256, + 0.01345072966068983, + 0.004897942766547203, + 0.04527367651462555, + 0.0457637719810009, + 0.07335180789232254, + 0.08422385901212692, + 0.11351614445447922, + 0.09519000351428986, + 0.08408738672733307, + 0.049720000475645065, + 0.004877838771790266, + 0.003475533565506339, + 0.014615694992244244, + 0.007447820156812668, + 0.003676292020827532, + 0.00391398137435317, + 0.007044985890388489, + 0.01640913262963295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.0036210936959832907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328616142272949 + ], + "opposite_logprob_tokens": [ + -0.7578616142272949 + ], + "logprob_reward": -0.6328616142272949, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328616142272949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03972053900361061, + 0.021112458780407906, + 0.0374545082449913, + 0.18784624338150024, + 0.36908474564552307, + 0.03997986018657684, + 0.037066735327243805, + 0.13240933418273926, + 0.08171382546424866, + 0.05361174792051315 + ], + "credit_map": [ + 0.00589884864166379, + 0.00414271792396903, + 0.0035724176559597254, + 0.001898444490507245, + 0.003151010489091277, + 0.009298454970121384, + 0.0029297072906047106, + 0.008828938938677311, + 0.0024186191149055958, + 0.003092904342338443, + 0.0015747901052236557, + 0.002128015272319317, + 0.004498514346778393, + 0.0023729666136205196, + 0.002385335275903344, + 0.002641313476487994, + 0.00884388666599989, + 0.0038466807454824448, + 0.005146374925971031, + 0.0047909896820783615, + 0.0032384744845330715, + 0.0033551102969795465, + 0.003660371759906411, + 0.004572620149701834, + 0.023070769384503365, + 0.024868732318282127, + 0.027069484815001488, + 0.022805126383900642, + 0.02804524637758732, + 0.02359229326248169, + 0.0222316887229681, + 0.016162903979420662, + 0.02972226031124592, + 0.02600693888962269, + 0.028600789606571198, + 0.09878619015216827, + 0.0440358966588974, + 0.029000496491789818, + 0.028404122218489647, + 0.08452802151441574, + 0.005001730751246214, + 0.004535641521215439, + 0.0042627411894500256, + 0.004880464170128107, + 0.0063380831852555275, + 0.004900672473013401, + 0.006160505581647158, + 0.003900021081790328, + 0.003440575208514929, + 0.0035902324598282576, + 0.007456734776496887, + 0.005965229123830795, + 0.006576456595212221, + 0.0028781206347048283, + 0.004284156020730734, + 0.00287523097358644, + 0.009637841954827309, + 0.016036048531532288, + 0.01684938184916973, + 0.020081758499145508, + 0.022830471396446228, + 0.012849856168031693, + 0.02276858314871788, + 0.011355385184288025, + 0.0057423231191933155, + 0.004293377976864576, + 0.010371646843850613, + 0.01083308458328247, + 0.0086408955976367, + 0.009416012093424797, + 0.011261221952736378, + 0.021155254915356636, + 0.010934954509139061, + 0.0035880934447050095, + 0.008304155431687832, + 0.006145545281469822, + 0.0029949883464723825, + 0.00642639072611928, + 0.0071358042769134045, + 0.008081814274191856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.6328616142272949 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Disney Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the steam cleaner bright lemon-yellow with a chunky black handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the steam emitted from the nozzle condense into geometrically perfect snowflakes and frost patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a warm, cluttered garage with wooden workbenches and colorful toolboxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly shift focus between a sharp ice crystal and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819116808474064 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819116808474064, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819116808474064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.007807363290339708, + 0.005892553366720676, + 0.005802243482321501, + 0.03392040729522705, + 0.013727126643061638, + 0.016382021829485893, + 0.028957549482584, + 0.037220925092697144, + 0.8364963531494141, + 0.013793437741696835 + ], + "credit_map": [ + 0.0008616980048827827, + 0.0009003796149045229, + 0.0007294370443560183, + 0.0005621478776447475, + 0.0007499995408579707, + 0.0017562448047101498, + 0.000878016056958586, + 0.001369441393762827, + 0.0014262181939557195, + 0.00086675223428756, + 0.0005726995295844972, + 0.0005701002082787454, + 0.0005472051561810076, + 0.0007251031347550452, + 0.0005715360748581588, + 0.0006129395915195346, + 0.0005554677336476743, + 0.0009242461528629065, + 0.0005130586796440184, + 0.0007261141436174512, + 0.0008917912491597235, + 0.0007532772142440081, + 0.0008071943884715438, + 0.0006310948519967496, + 0.00536119332537055, + 0.004098091274499893, + 0.004725244827568531, + 0.00432443805038929, + 0.003694098209962249, + 0.004432415124028921, + 0.0038144467398524284, + 0.0034704809077084064, + 0.002217894187197089, + 0.002332014963030815, + 0.001825033570639789, + 0.0014189074281603098, + 0.001667479518800974, + 0.001444944296963513, + 0.001754723722115159, + 0.0010661290725693107, + 0.0023603651206940413, + 0.0023734914138913155, + 0.0019509918056428432, + 0.002175296423956752, + 0.00258451490662992, + 0.0011270863469690084, + 0.0021579875610768795, + 0.001652288599871099, + 0.0019840344320982695, + 0.004231103230267763, + 0.007119028829038143, + 0.004989467095583677, + 0.0034340114798396826, + 0.0023898810613900423, + 0.0032035044860094786, + 0.001606520963832736, + 0.0020616543479263783, + 0.004872849676758051, + 0.004990953952074051, + 0.005981112830340862, + 0.006545247510075569, + 0.0032690316438674927, + 0.006939939688891172, + 0.002560137305408716, + 0.05985330417752266, + 0.06430656462907791, + 0.10212753713130951, + 0.1159578263759613, + 0.17092737555503845, + 0.12629817426204681, + 0.12007260322570801, + 0.07695304602384567, + 0.002735947724431753, + 0.0008106250897981226, + 0.002088566543534398, + 0.0017345697851851583, + 0.0010466831736266613, + 0.001768575981259346, + 0.0012512346729636192, + 0.0023572363425046206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.01819116808474064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_8/index_tensor([3688, 3688])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_8/index_tensor([3688, 3688])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ed895f3efa40932ba7f8de0c5ce74ac389d17904 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_8/index_tensor([3688, 3688])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly low-angle close-up slowly zooms in on clusters of vibrant yellow wolfsbane flowers, rendered in Studio Ghibli's signature hand-drawn style. Set in a sun-drenched alpine meadow, lush green blades of grass, each with a soft watercolor texture, sway gently in the breeze. In the background, rolling hills fade into a hazy, distant mountain range, while the azure sky is dotted with fluffy, white Ghibli clouds. The camera focuses on the central flower head, a dense sphere of tightly closed buds. In incredibly smooth, high-frame-rate animation, each delicate petal begins to unfold in graceful, deliberate slowness, their movements as smooth as silk. The buttery yellow petals, outlined with delicate, organic brushstrokes, peel back one by one, revealing a soft, magical inner glow emanating from the core of each flower, as if painted with light itself. As the flowers fully open, this gentle, warm glow becomes more intense, casting a subtle radiance that contrasts beautifully with the bright, golden sunlight bathing the entire meadow, creating a profoundly serene, natural spectacle that captures life’s fleeting, magical moments in bloom in stunning detail.", + "seed": 3396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_8/index_3688_seed_3396_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038425090606324375, + "vlm_logprob": -0.038425090606324375, + "vlm_yes_no_margin": 4.124999976158142, + "vlm_token_credit": -0.038425090606324375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014270045794546604 + ], + "opposite_logprob_tokens": [ + -4.514269828796387 + ], + "logprob_reward": -0.014270045794546604, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014270045794546604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09292557090520859, + 0.048365384340286255, + 0.01756143569946289, + 0.04193093255162239, + 0.041516389697790146, + 0.06979672610759735, + 0.38188403844833374, + 0.13773809373378754, + 0.11186908930540085, + 0.05641232058405876 + ], + "credit_map": [ + 0.01259489357471466, + 0.009055458940565586, + 0.012561810202896595, + 0.009384741075336933, + 0.012010493315756321, + 0.012806505896151066, + 0.013513452373445034, + 0.010998216457664967, + 0.00621812092140317, + 0.00486466521397233, + 0.005987071897834539, + 0.00553502980619669, + 0.004963382612913847, + 0.005593820475041866, + 0.007436608895659447, + 0.0077666835859417915, + 0.0018516841810196638, + 0.0015299274818971753, + 0.0016486954409629107, + 0.002683386905118823, + 0.002024100162088871, + 0.001898209098726511, + 0.0038845310918986797, + 0.002040900755673647, + 0.003978809807449579, + 0.005074834451079369, + 0.004012217745184898, + 0.006123340222984552, + 0.006715942174196243, + 0.004614572040736675, + 0.006080375052988529, + 0.005330842919647694, + 0.004210060462355614, + 0.002335526514798403, + 0.0038925595581531525, + 0.005755782127380371, + 0.003593654837459326, + 0.003722633235156536, + 0.012313369661569595, + 0.0056928060948848724, + 0.004880459513515234, + 0.009206949733197689, + 0.009758816100656986, + 0.014136169105768204, + 0.007250250782817602, + 0.008740300312638283, + 0.007449943106621504, + 0.008373834192752838, + 0.0645408183336258, + 0.02161969617009163, + 0.028298836201429367, + 0.06347636878490448, + 0.024078799411654472, + 0.024141045287251472, + 0.12160183489322662, + 0.03412662073969841, + 0.009159645065665245, + 0.011440624482929707, + 0.010091286152601242, + 0.030076170340180397, + 0.019329974427819252, + 0.009776202961802483, + 0.03441596403717995, + 0.013448218815028667, + 0.008036796003580093, + 0.007459643762558699, + 0.007360225543379784, + 0.03038041666150093, + 0.010491856373846531, + 0.00644840020686388, + 0.031122125685214996, + 0.010569625534117222, + 0.003496797988191247, + 0.004048779606819153, + 0.0043646409176290035, + 0.00884654838591814, + 0.008418537676334381, + 0.004675221163779497, + 0.01272508967667818, + 0.009836706332862377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.014270045794546604 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023295527324080467 + ], + "opposite_logprob_tokens": [ + -3.7732956409454346 + ], + "logprob_reward": -0.023295527324080467, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023295527324080467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22270455956459045, + 0.07980820536613464, + 0.027184724807739258, + 0.07073239237070084, + 0.0607161819934845, + 0.03134211152791977, + 0.29717615246772766, + 0.1436024159193039, + 0.03530622273683548, + 0.0314270444214344 + ], + "credit_map": [ + 0.02842705510556698, + 0.020117061212658882, + 0.030030259862542152, + 0.024613410234451294, + 0.022574525326490402, + 0.03110405243933201, + 0.02884555049240589, + 0.036992624402046204, + 0.010141037404537201, + 0.006791985128074884, + 0.010853655636310577, + 0.007819480262696743, + 0.009832991287112236, + 0.00830092653632164, + 0.012306115590035915, + 0.013762010261416435, + 0.003968766424804926, + 0.0025700251571834087, + 0.0026481677778065205, + 0.002529539866372943, + 0.0032750896643847227, + 0.003803837113082409, + 0.00444598775357008, + 0.003943312913179398, + 0.009187637828290462, + 0.008305191993713379, + 0.008310028351843357, + 0.010945591144263744, + 0.008746490813791752, + 0.006456152535974979, + 0.011476260609924793, + 0.00730503536760807, + 0.011542531661689281, + 0.005868928041309118, + 0.006478869356215, + 0.008240006864070892, + 0.004755858797580004, + 0.007153588812798262, + 0.009974457323551178, + 0.00670194486156106, + 0.004175489768385887, + 0.0024413163773715496, + 0.004665010143071413, + 0.0042772614397108555, + 0.002648755442351103, + 0.003990253433585167, + 0.005574163980782032, + 0.0035698600113391876, + 0.049213603138923645, + 0.017327696084976196, + 0.0196136012673378, + 0.050921909511089325, + 0.020417233929038048, + 0.017854204401373863, + 0.09670116752386093, + 0.02512674778699875, + 0.009933185763657093, + 0.015524921007454395, + 0.009716183878481388, + 0.03614811971783638, + 0.011303959414362907, + 0.007170930039137602, + 0.04375031962990761, + 0.010054804384708405, + 0.004277420230209827, + 0.002972537185996771, + 0.0030861732084304094, + 0.00471075251698494, + 0.006080390885472298, + 0.0031154383905231953, + 0.006477986928075552, + 0.0045855226926505566, + 0.003078822512179613, + 0.003014857415109873, + 0.0028445736970752478, + 0.004726058337837458, + 0.005050852429121733, + 0.0036297377664595842, + 0.005637494847178459, + 0.003444648114964366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.023295527324080467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178999780677259 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.001178999780677259, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178999780677259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06788953393697739, + 0.029282651841640472, + 0.020779971033334732, + 0.021238986402750015, + 0.048725418746471405, + 0.0939323902130127, + 0.5466813445091248, + 0.07561424374580383, + 0.044064417481422424, + 0.051791008561849594 + ], + "credit_map": [ + 0.00872427225112915, + 0.01006652694195509, + 0.008269958198070526, + 0.009397992864251137, + 0.007635828107595444, + 0.008103419095277786, + 0.008627050556242466, + 0.007064480800181627, + 0.0026895576156675816, + 0.0030627369415014982, + 0.002910108072683215, + 0.0055241151712834835, + 0.0035719031002372503, + 0.0033815784845501184, + 0.004998079035431147, + 0.0031445729546248913, + 0.0036682283971458673, + 0.00157349300570786, + 0.002411846537142992, + 0.0026323567144572735, + 0.003022930584847927, + 0.002222300972789526, + 0.0028005349449813366, + 0.002448278944939375, + 0.0021416039671748877, + 0.0024164868518710136, + 0.0023648133501410484, + 0.004106001928448677, + 0.002828017110005021, + 0.0023772846907377243, + 0.0024164372589439154, + 0.002588341711089015, + 0.005032540298998356, + 0.00414216797798872, + 0.004420234356075525, + 0.00544515997171402, + 0.004892522934824228, + 0.00669274153187871, + 0.010545548982918262, + 0.007554501760751009, + 0.01305409986525774, + 0.007036915048956871, + 0.009748026728630066, + 0.0119017343968153, + 0.008825593627989292, + 0.00785602442920208, + 0.02488066628575325, + 0.01062933262437582, + 0.09145285189151764, + 0.029513010755181313, + 0.03614218160510063, + 0.09524138271808624, + 0.03476117178797722, + 0.032113589346408844, + 0.18101777136325836, + 0.04643944278359413, + 0.005307338200509548, + 0.008483083918690681, + 0.006508630234748125, + 0.013381517492234707, + 0.01425231620669365, + 0.00668911961838603, + 0.012916994281113148, + 0.008075246587395668, + 0.0038954499177634716, + 0.004259027075022459, + 0.003627261845394969, + 0.0064235287718474865, + 0.006808795500546694, + 0.003805504646152258, + 0.009342016652226448, + 0.005902833305299282, + 0.0038732292596250772, + 0.006055784411728382, + 0.004467118065804243, + 0.007262813858687878, + 0.007951813749969006, + 0.0038863476365804672, + 0.011177748441696167, + 0.007116150576621294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.001178999780677259 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340389624238014 + ], + "opposite_logprob_tokens": [ + -3.6513404846191406 + ], + "logprob_reward": -0.026340389624238014, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340389624238014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040617767721414566, + 0.0207800455391407, + 0.05002385005354881, + 0.04099687933921814, + 0.061431098729372025, + 0.07629279047250748, + 0.5968513488769531, + 0.07690007239580154, + 0.026200685650110245, + 0.009905524551868439 + ], + "credit_map": [ + 0.004207978490740061, + 0.005075556226074696, + 0.004054460674524307, + 0.0056809550151228905, + 0.004927538335323334, + 0.003943122923374176, + 0.007565306965261698, + 0.005162850953638554, + 0.0018341023242101073, + 0.002109826309606433, + 0.002710997825488448, + 0.003591823158785701, + 0.002247262978926301, + 0.00222012004815042, + 0.0037545086815953255, + 0.0023114033974707127, + 0.004573991056531668, + 0.004119107034057379, + 0.006111299153417349, + 0.008034917525947094, + 0.0055459048599004745, + 0.005021152086555958, + 0.009969085454940796, + 0.006648393347859383, + 0.004889999516308308, + 0.004047273192554712, + 0.004477208945900202, + 0.008141737431287766, + 0.005631349980831146, + 0.0034968708641827106, + 0.004949417430907488, + 0.005363017320632935, + 0.0076308180578053, + 0.005659283604472876, + 0.004569924436509609, + 0.0069181593134999275, + 0.0057493154890835285, + 0.0064342743717134, + 0.015023868530988693, + 0.009445452131330967, + 0.008885153569281101, + 0.0072644297033548355, + 0.009599552489817142, + 0.01083461008965969, + 0.006501303054392338, + 0.01180779468268156, + 0.011693321168422699, + 0.009706621058285236, + 0.09761278331279755, + 0.03652045503258705, + 0.04262867569923401, + 0.10077866166830063, + 0.04264084994792938, + 0.036611348390579224, + 0.18701381981372833, + 0.053044743835926056, + 0.005990053527057171, + 0.014793661423027515, + 0.005356478039175272, + 0.008191438391804695, + 0.01859958842396736, + 0.007355863694101572, + 0.011115211993455887, + 0.005497782025486231, + 0.0030705120880156755, + 0.004034866578876972, + 0.0017736381851136684, + 0.00390974897891283, + 0.004608607850968838, + 0.002067799912765622, + 0.0046243444085121155, + 0.00211116811260581, + 0.0010818925220519304, + 0.0012860385468229651, + 0.0007875273586250842, + 0.0012689913855865598, + 0.001439058338291943, + 0.001017601229250431, + 0.0017171110957860947, + 0.0013073045993223786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.026340389624238014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704049050807953 + ], + "opposite_logprob_tokens": [ + -2.127040386199951 + ], + "logprob_reward": -0.12704049050807953, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704049050807953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061453528702259064, + 0.04164809361100197, + 0.021183665841817856, + 0.036881282925605774, + 0.08402572572231293, + 0.0821438729763031, + 0.15866543352603912, + 0.3503284454345703, + 0.1050284132361412, + 0.05864154174923897 + ], + "credit_map": [ + 0.006959869991987944, + 0.006786982528865337, + 0.006358198821544647, + 0.006787254940718412, + 0.008345145732164383, + 0.0070731425657868385, + 0.012674931436777115, + 0.006468003150075674, + 0.005203488282859325, + 0.0045948512852191925, + 0.004546222742646933, + 0.005714226048439741, + 0.004754969850182533, + 0.004545931238681078, + 0.006549864076077938, + 0.005738536361604929, + 0.002056888537481427, + 0.002592829754576087, + 0.0022924148943275213, + 0.002909514121711254, + 0.002827920252457261, + 0.0019567757844924927, + 0.0038220041897147894, + 0.0027253178413957357, + 0.006384357810020447, + 0.0031428865622729063, + 0.004212273750454187, + 0.004349092021584511, + 0.004487406928092241, + 0.0028147564735263586, + 0.006982675287872553, + 0.004507832229137421, + 0.014303825795650482, + 0.008921289816498756, + 0.005146587733179331, + 0.007856151089072227, + 0.005797318648546934, + 0.00614843238145113, + 0.024906931445002556, + 0.010945191606879234, + 0.009703876450657845, + 0.007233886513859034, + 0.009911958128213882, + 0.012416189536452293, + 0.006170124281197786, + 0.015373299829661846, + 0.01082801353186369, + 0.010506522841751575, + 0.02194576896727085, + 0.013024427928030491, + 0.018317965790629387, + 0.023340176790952682, + 0.013374625705182552, + 0.013188737444579601, + 0.03755773603916168, + 0.01791599951684475, + 0.027688996866345406, + 0.056748129427433014, + 0.01764795370399952, + 0.05853450670838356, + 0.07188555598258972, + 0.025135548785328865, + 0.07277265191078186, + 0.01991509087383747, + 0.008781972341239452, + 0.009066198952496052, + 0.01032753475010395, + 0.02060270495712757, + 0.01249178871512413, + 0.00878163892775774, + 0.022343836724758148, + 0.012632740661501884, + 0.003913481719791889, + 0.004669953137636185, + 0.0057564787566661835, + 0.008626180700957775, + 0.006124760489910841, + 0.0048172553069889545, + 0.013157711364328861, + 0.011575720272958279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.12704049050807953 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle close-up slowly zooms in on clusters of vibrant yellow wolfsbane flowers, rendered in Studio Ghibli's signature hand-drawn style. Set in a sun-drenched alpine meadow, lush green blades of grass, each with a soft watercolor texture, sway gently in the breeze. In the background, rolling hills fade into a hazy, distant mountain range, while the azure sky is dotted with fluffy, white Ghibli clouds. The camera focuses on the central flower head, a dense sphere of tightly closed buds. In incredibly smooth, high-frame-rate animation, each delicate petal begins to unfold in graceful, deliberate slowness, their movements as smooth as silk. The buttery yellow petals, outlined with delicate, organic brushstrokes, peel back one by one, revealing a soft, magical inner glow emanating from the core of each flower, as if painted with light itself. As the flowers fully open, this gentle, warm glow becomes more intense, casting a subtle radiance that contrasts beautifully with the bright, golden sunlight bathing the entire meadow, creating a profoundly serene, natural spectacle that captures life’s fleeting, magical moments in bloom in stunning detail.", + "seed": 3397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_8/index_3688_seed_3397_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03756935533601791, + "vlm_logprob": -0.03756935533601791, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03756935533601791, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018436027690768242 + ], + "opposite_logprob_tokens": [ + -4.643435955047607 + ], + "logprob_reward": -0.018436027690768242, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.018436027690768242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10495395958423615, + 0.04448596015572548, + 0.017895633354783058, + 0.02730308100581169, + 0.018107756972312927, + 0.028955785557627678, + 0.6749258637428284, + 0.02500057779252529, + 0.024666287004947662, + 0.03370518982410431 + ], + "credit_map": [ + 0.007960098795592785, + 0.018896717578172684, + 0.011350661516189575, + 0.019609935581684113, + 0.013382181525230408, + 0.009915950708091259, + 0.014991270378232002, + 0.008847137913107872, + 0.006600420922040939, + 0.004553084261715412, + 0.00590879563242197, + 0.005849391222000122, + 0.004756729118525982, + 0.0051446035504341125, + 0.006088173482567072, + 0.00558476010337472, + 0.0011074229842051864, + 0.002323092892765999, + 0.0019965381361544132, + 0.002220534486696124, + 0.0029133071657270193, + 0.0012348209274932742, + 0.0029838066548109055, + 0.003116109874099493, + 0.0022139076609164476, + 0.003529525129124522, + 0.00429904367774725, + 0.004376160446554422, + 0.002451986772939563, + 0.003164836438372731, + 0.0043370649218559265, + 0.0029305561911314726, + 0.0012204552767798305, + 0.0038336515426635742, + 0.002459184033796191, + 0.0025056451559066772, + 0.002298937179148197, + 0.0012212832225486636, + 0.002493610605597496, + 0.0020749906543642282, + 0.002603095257654786, + 0.004206235054880381, + 0.003476756624877453, + 0.005157878156751394, + 0.002723214915022254, + 0.0024911384098231792, + 0.005452443845570087, + 0.0028450230602174997, + 0.03296302631497383, + 0.023609532043337822, + 0.23499861359596252, + 0.23715944588184357, + 0.035498932003974915, + 0.04914696514606476, + 0.035079192370176315, + 0.026470104232430458, + 0.0016990213189274073, + 0.005574097391217947, + 0.0031728949397802353, + 0.003475240897387266, + 0.004020361229777336, + 0.0018986896611750126, + 0.003144135931506753, + 0.002016135724261403, + 0.0024551500100642443, + 0.0038897141348570585, + 0.002832801314070821, + 0.0030915027018636465, + 0.003240299178287387, + 0.003663195762783289, + 0.002988846506923437, + 0.0025047766976058483, + 0.00304609932936728, + 0.007009363733232021, + 0.003503089537844062, + 0.004196918569505215, + 0.004985559266060591, + 0.0029305561911314726, + 0.004318270832300186, + 0.0037153351586312056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.018436027690768242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370640054345131 + ], + "opposite_logprob_tokens": [ + -3.4087064266204834 + ], + "logprob_reward": -0.03370640054345131, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370640054345131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23155222833156586, + 0.07911421358585358, + 0.09168203175067902, + 0.2443603277206421, + 0.07279784977436066, + 0.06162199378013611, + 0.10724502056837082, + 0.03844946250319481, + 0.030073316767811775, + 0.04310360923409462 + ], + "credit_map": [ + 0.03093363530933857, + 0.035592928528785706, + 0.024116234853863716, + 0.03541760891675949, + 0.023247364908456802, + 0.025615505874156952, + 0.027891233563423157, + 0.028737720102071762, + 0.005098438356071711, + 0.00868333037942648, + 0.009690920822322369, + 0.020694876089692116, + 0.007330659776926041, + 0.006678206846117973, + 0.01172730140388012, + 0.009210477583110332, + 0.010257641784846783, + 0.012839464470744133, + 0.012980171479284763, + 0.01424239668995142, + 0.009441491216421127, + 0.00850563496351242, + 0.011637119576334953, + 0.011778105050325394, + 0.023293327540159225, + 0.029516980051994324, + 0.030181970447301865, + 0.03130017966032028, + 0.021917689591646194, + 0.0318739078938961, + 0.04345937445759773, + 0.03281687572598457, + 0.005122804082930088, + 0.010142706334590912, + 0.012326891534030437, + 0.011851753108203411, + 0.007942755706608295, + 0.006973846349865198, + 0.011900290846824646, + 0.006536802276968956, + 0.006679666228592396, + 0.005175186786800623, + 0.006027704104781151, + 0.013695545494556427, + 0.006416646298021078, + 0.007936613634228706, + 0.009573590010404587, + 0.0061170440167188644, + 0.0048661078326404095, + 0.006099367048591375, + 0.03302662819623947, + 0.038542091846466064, + 0.0065960753709077835, + 0.004787549376487732, + 0.00810759887099266, + 0.005219609010964632, + 0.002364558633416891, + 0.008721188642084599, + 0.004955780226737261, + 0.004845963791012764, + 0.008829338476061821, + 0.002405020408332348, + 0.0035459769424051046, + 0.002781637478619814, + 0.002118841279298067, + 0.0036345412954688072, + 0.00396511098369956, + 0.004446001257747412, + 0.004105017054826021, + 0.003628264646977186, + 0.00494897598400712, + 0.0032265624031424522, + 0.0050614457577466965, + 0.006163991987705231, + 0.00475891912356019, + 0.007343628443777561, + 0.004674241412431002, + 0.003363017225638032, + 0.0075804730877280235, + 0.004157892893999815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.03370640054345131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017135950038209558 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017135950038209558, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017135950038209558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18748584389686584, + 0.1833723932504654, + 0.02161242999136448, + 0.05633823573589325, + 0.030582115054130554, + 0.0777365192770958, + 0.31773051619529724, + 0.01816198043525219, + 0.03865714743733406, + 0.06832291185855865 + ], + "credit_map": [ + 0.01737932674586773, + 0.03180922195315361, + 0.02262183465063572, + 0.023959346115589142, + 0.018842756748199463, + 0.024496838450431824, + 0.028068626299500465, + 0.02030785381793976, + 0.021839043125510216, + 0.021191485226154327, + 0.023707382380962372, + 0.039758242666721344, + 0.017247477546334267, + 0.014296382665634155, + 0.028774024918675423, + 0.016558324918150902, + 0.0017120520351454616, + 0.003282674588263035, + 0.002675955183804035, + 0.0031607048586010933, + 0.002367465989664197, + 0.0019457631278783083, + 0.003402917180210352, + 0.0030648927204310894, + 0.005877266172319651, + 0.009188441559672356, + 0.006350908428430557, + 0.007228125352412462, + 0.005845705047249794, + 0.008106274530291557, + 0.007030477747321129, + 0.006711025256663561, + 0.0023715756833553314, + 0.004805270116776228, + 0.003830775385722518, + 0.005067997146397829, + 0.00379629316739738, + 0.002116753254085779, + 0.005417929496616125, + 0.0031755149830132723, + 0.008085590787231922, + 0.007789469324052334, + 0.0097582396119833, + 0.014815538190305233, + 0.006974672898650169, + 0.01289463322609663, + 0.009443399496376514, + 0.007974959909915924, + 0.017974894493818283, + 0.010814109817147255, + 0.10615896433591843, + 0.08922146260738373, + 0.02146558277308941, + 0.04267709329724312, + 0.014619335532188416, + 0.014799022115767002, + 0.0013543276581913233, + 0.0035017626360058784, + 0.0019975118339061737, + 0.002989887958392501, + 0.002583963330835104, + 0.0014583278680220246, + 0.002439014846459031, + 0.001837180694565177, + 0.0025689024478197098, + 0.005372729152441025, + 0.006163546349853277, + 0.005204855464398861, + 0.005465865135192871, + 0.004175023641437292, + 0.006127296946942806, + 0.003578920615836978, + 0.006772538181394339, + 0.008197065442800522, + 0.008096104487776756, + 0.011990835890173912, + 0.006973211653530598, + 0.00520915724337101, + 0.013541002757847309, + 0.007542981766164303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0017135950038209558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368830308318138 + ], + "opposite_logprob_tokens": [ + -3.4086883068084717 + ], + "logprob_reward": -0.03368830308318138, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368830308318138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09166844189167023, + 0.03863280639052391, + 0.04505862668156624, + 0.09603144228458405, + 0.05317595973610878, + 0.06779822707176208, + 0.29684048891067505, + 0.12225654721260071, + 0.09795824438333511, + 0.09057924896478653 + ], + "credit_map": [ + 0.010339701548218727, + 0.011238050647079945, + 0.010338902473449707, + 0.01601744070649147, + 0.00878955703228712, + 0.011618532240390778, + 0.012124858796596527, + 0.011201397515833378, + 0.002919991733506322, + 0.002830869983881712, + 0.006101199425756931, + 0.00972065795212984, + 0.003065812634304166, + 0.0032431965228170156, + 0.007116040214896202, + 0.003635038621723652, + 0.004679189529269934, + 0.005360111594200134, + 0.006132462061941624, + 0.008960152044892311, + 0.005328851751983166, + 0.004581121262162924, + 0.006334992591291666, + 0.0036817432846874, + 0.005437043495476246, + 0.017943358048796654, + 0.009736866690218449, + 0.01920371875166893, + 0.012984835542738438, + 0.005861269310116768, + 0.016787109896540642, + 0.008077230304479599, + 0.004159694537520409, + 0.007021768018603325, + 0.007534265983849764, + 0.007245922461152077, + 0.00540332542732358, + 0.004437873139977455, + 0.011792331002652645, + 0.005580774042755365, + 0.005552695132791996, + 0.012539023533463478, + 0.007657054346054792, + 0.008796991780400276, + 0.010180627927184105, + 0.007354132365435362, + 0.009052559733390808, + 0.006665138062089682, + 0.012842169031500816, + 0.02798723243176937, + 0.019630931317806244, + 0.08509672433137894, + 0.09547535330057144, + 0.01453934796154499, + 0.023414012044668198, + 0.017854683101177216, + 0.008092389442026615, + 0.024314992129802704, + 0.01636750064790249, + 0.01579228974878788, + 0.020058168098330498, + 0.01133627351373434, + 0.015566154383122921, + 0.01072877086699009, + 0.00891706719994545, + 0.015402224846184254, + 0.010003509931266308, + 0.01096024364233017, + 0.014224371872842312, + 0.01270629558712244, + 0.015182648785412312, + 0.010561870411038399, + 0.007145849987864494, + 0.013939579017460346, + 0.007945893332362175, + 0.012928803451359272, + 0.011183444410562515, + 0.005836601369082928, + 0.022495277225971222, + 0.009103790856897831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.03368830308318138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030245035886765 + ], + "opposite_logprob_tokens": [ + -2.3503024578094482 + ], + "logprob_reward": -0.10030245035886765, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030245035886765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07555902749300003, + 0.04016020894050598, + 0.04448407515883446, + 0.04942285642027855, + 0.03325698524713516, + 0.06591198593378067, + 0.43607720732688904, + 0.13699817657470703, + 0.05962999537587166, + 0.05849948897957802 + ], + "credit_map": [ + 0.007569876033812761, + 0.009234648197889328, + 0.00982382707297802, + 0.014161447063088417, + 0.007491155993193388, + 0.006834027357399464, + 0.0106406444683671, + 0.009803400374948978, + 0.003941396251320839, + 0.003719138912856579, + 0.008241523057222366, + 0.00633954256772995, + 0.0034294442739337683, + 0.0032731329556554556, + 0.007025666534900665, + 0.004190364852547646, + 0.0031091594137251377, + 0.007688740734010935, + 0.006491906475275755, + 0.008903119713068008, + 0.00477914372459054, + 0.0032239879947155714, + 0.006175481714308262, + 0.004112537018954754, + 0.005298506934195757, + 0.007683738600462675, + 0.005101227201521397, + 0.00793333351612091, + 0.005615250673145056, + 0.006140368524938822, + 0.007573282811790705, + 0.004077151417732239, + 0.0023315479047596455, + 0.005558762699365616, + 0.004124866798520088, + 0.0045260475017130375, + 0.003483510809019208, + 0.002876276383176446, + 0.0073530604131519794, + 0.0030029097106307745, + 0.004793606232851744, + 0.013399682007730007, + 0.006996371317654848, + 0.008515623398125172, + 0.009100908413529396, + 0.005640587769448757, + 0.010555228218436241, + 0.006909977179020643, + 0.018109772354364395, + 0.01702445186674595, + 0.12477092444896698, + 0.1606859266757965, + 0.05552516132593155, + 0.02365596406161785, + 0.021339673548936844, + 0.01496534701436758, + 0.008108201436698437, + 0.02673918753862381, + 0.016728952527046204, + 0.019699906930327415, + 0.02629670687019825, + 0.011194982565939426, + 0.016533924266695976, + 0.011696308851242065, + 0.005529877729713917, + 0.007771427743136883, + 0.0062935929745435715, + 0.006871496792882681, + 0.00861307792365551, + 0.00810009241104126, + 0.010084345936775208, + 0.00636608200147748, + 0.003926085773855448, + 0.008525214157998562, + 0.0060714720748364925, + 0.009060959331691265, + 0.006192394997924566, + 0.0037878539878875017, + 0.0142600042745471, + 0.00667550228536129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.10030245035886765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle close-up slowly zooms in on clusters of vibrant yellow wolfsbane flowers, rendered in Studio Ghibli's signature hand-drawn style. Set in a sun-drenched alpine meadow, lush green blades of grass, each with a soft watercolor texture, sway gently in the breeze. In the background, rolling hills fade into a hazy, distant mountain range, while the azure sky is dotted with fluffy, white Ghibli clouds. The camera focuses on the central flower head, a dense sphere of tightly closed buds. In incredibly smooth, high-frame-rate animation, each delicate petal begins to unfold in graceful, deliberate slowness, their movements as smooth as silk. The buttery yellow petals, outlined with delicate, organic brushstrokes, peel back one by one, revealing a soft, magical inner glow emanating from the core of each flower, as if painted with light itself. As the flowers fully open, this gentle, warm glow becomes more intense, casting a subtle radiance that contrasts beautifully with the bright, golden sunlight bathing the entire meadow, creating a profoundly serene, natural spectacle that captures life’s fleeting, magical moments in bloom in stunning detail.", + "seed": 3398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_8/index_3688_seed_3398_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020913706975989044, + "vlm_logprob": -0.020913706975989044, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.020913706975989044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02139768749475479 + ], + "opposite_logprob_tokens": [ + -4.021397590637207 + ], + "logprob_reward": -0.02139768749475479, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02139768749475479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1889570653438568, + 0.1020335704088211, + 0.07303883135318756, + 0.14188653230667114, + 0.08413577824831009, + 0.07151997089385986, + 0.07864213734865189, + 0.08141981065273285, + 0.10937213897705078, + 0.0689942017197609 + ], + "credit_map": [ + 0.03188435733318329, + 0.02191111445426941, + 0.017787979915738106, + 0.02209901623427868, + 0.020664798095822334, + 0.016674920916557312, + 0.02914552018046379, + 0.02878936380147934, + 0.020377742126584053, + 0.013265686109662056, + 0.011012828908860683, + 0.019871247932314873, + 0.012084590271115303, + 0.007014280650764704, + 0.009986775927245617, + 0.008420423604547977, + 0.0076669855043292046, + 0.007223016582429409, + 0.010322732850909233, + 0.012120605446398258, + 0.006141441408544779, + 0.009376861155033112, + 0.007939893752336502, + 0.012247296050190926, + 0.011516904458403587, + 0.009364534169435501, + 0.013199611566960812, + 0.0233242716640234, + 0.009681710042059422, + 0.019062485545873642, + 0.014462383463978767, + 0.041274625808000565, + 0.0062935566529631615, + 0.00803930964320898, + 0.009837247431278229, + 0.00941366795450449, + 0.02237490378320217, + 0.005528723355382681, + 0.007455198094248772, + 0.015193172730505466, + 0.005073751322925091, + 0.00933292880654335, + 0.011473036371171474, + 0.01062099914997816, + 0.009581669233739376, + 0.005358765367418528, + 0.00931682251393795, + 0.01076200045645237, + 0.014202761463820934, + 0.011770022101700306, + 0.011354819871485233, + 0.0069418917410075665, + 0.011024735867977142, + 0.011179082095623016, + 0.0052033388055861, + 0.006965480744838715, + 0.011942042969167233, + 0.012514021247625351, + 0.00821587536484003, + 0.012505995109677315, + 0.009913498535752296, + 0.009063301607966423, + 0.009444297291338444, + 0.007820778526365757, + 0.010779289528727531, + 0.017716804519295692, + 0.009128909558057785, + 0.016341155394911766, + 0.017314312979578972, + 0.00842508114874363, + 0.020316660404205322, + 0.009349917992949486, + 0.008150913752615452, + 0.009369508363306522, + 0.005623429547995329, + 0.010516422800719738, + 0.013292679563164711, + 0.007580882869660854, + 0.009066574275493622, + 0.0053937919437885284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.02139768749475479 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108889013528824 + ], + "opposite_logprob_tokens": [ + -3.2881088256835938 + ], + "logprob_reward": -0.038108889013528824, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108889013528824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20555253326892853, + 0.1230495274066925, + 0.24665619432926178, + 0.043286170810461044, + 0.07011426240205765, + 0.06395017355680466, + 0.02910207025706768, + 0.04268825799226761, + 0.051715172827243805, + 0.1238856092095375 + ], + "credit_map": [ + 0.02244378626346588, + 0.019439157098531723, + 0.017132030799984932, + 0.03525301069021225, + 0.040137797594070435, + 0.01769920252263546, + 0.02829654887318611, + 0.02515099197626114, + 0.01140776090323925, + 0.014288902282714844, + 0.014883049763739109, + 0.02875412255525589, + 0.0207328200340271, + 0.0083074439316988, + 0.014041196554899216, + 0.010634230449795723, + 0.03454200550913811, + 0.030066953971982002, + 0.03917107358574867, + 0.03585663437843323, + 0.01944003626704216, + 0.03079313039779663, + 0.026143139228224754, + 0.030643224716186523, + 0.002600695239380002, + 0.003158294828608632, + 0.005009650718420744, + 0.006912039127200842, + 0.004265894647687674, + 0.00379898096434772, + 0.004460922442376614, + 0.013079696334898472, + 0.006001411937177181, + 0.009761537425220013, + 0.009109987877309322, + 0.011459670029580593, + 0.012105342000722885, + 0.00660316739231348, + 0.008344918489456177, + 0.006728230509907007, + 0.005967163946479559, + 0.008180292323231697, + 0.008883611299097538, + 0.007042191922664642, + 0.006957814563065767, + 0.00733821839094162, + 0.007352019194513559, + 0.012228861451148987, + 0.0030098932329565287, + 0.00545527832582593, + 0.004252017010003328, + 0.003528051311150193, + 0.004359888378530741, + 0.0025030486285686493, + 0.0032412572763860226, + 0.002752636792138219, + 0.0062828813679516315, + 0.006036372855305672, + 0.005627433303743601, + 0.004886613227427006, + 0.00538628688082099, + 0.004118743352591991, + 0.005351806990802288, + 0.004998121410608292, + 0.004544616676867008, + 0.006100427359342575, + 0.004821650218218565, + 0.006783151533454657, + 0.0054164729081094265, + 0.004545974545180798, + 0.012372951954603195, + 0.007129929028451443, + 0.01325072068721056, + 0.01276981458067894, + 0.013784735463559628, + 0.016721030697226524, + 0.034931574016809464, + 0.007805399131029844, + 0.011867250315845013, + 0.012755084782838821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.038108889013528824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017134760273620486 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017134760273620486, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017134760273620486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2491959184408188, + 0.13339246809482574, + 0.16405227780342102, + 0.0801742896437645, + 0.047856107354164124, + 0.036806438118219376, + 0.05205671489238739, + 0.05379805713891983, + 0.0830845832824707, + 0.09958314895629883 + ], + "credit_map": [ + 0.049176331609487534, + 0.023220699280500412, + 0.02108275145292282, + 0.024805432185530663, + 0.028846491128206253, + 0.020995348691940308, + 0.03528415039181709, + 0.045784708112478256, + 0.035130374133586884, + 0.019736286252737045, + 0.013488535769283772, + 0.020725449547171593, + 0.01077333278954029, + 0.010891648940742016, + 0.011459650471806526, + 0.011187192983925343, + 0.022782307118177414, + 0.01740000955760479, + 0.020606376230716705, + 0.02143365517258644, + 0.01733006164431572, + 0.025943703949451447, + 0.014711654745042324, + 0.023844512179493904, + 0.0057284897193312645, + 0.006225451361387968, + 0.004567782860249281, + 0.007450353354215622, + 0.022632764652371407, + 0.013820440508425236, + 0.006584332790225744, + 0.01316467858850956, + 0.004214046988636255, + 0.007170529570430517, + 0.007381849456578493, + 0.006124366074800491, + 0.007782000582665205, + 0.005981834139674902, + 0.0041263522580266, + 0.005075126886367798, + 0.0030174762941896915, + 0.004805875942111015, + 0.0048116291873157024, + 0.0050811488181352615, + 0.005116108804941177, + 0.002482031472027302, + 0.004678164608776569, + 0.006814002990722656, + 0.007045079488307238, + 0.01007495354861021, + 0.00773403886705637, + 0.006242270581424236, + 0.007294023875147104, + 0.005107845179736614, + 0.004310048650950193, + 0.0042484537698328495, + 0.007487042341381311, + 0.00698802899569273, + 0.005108206998556852, + 0.01019623689353466, + 0.006585553288459778, + 0.004607124719768763, + 0.0076435343362390995, + 0.005182329565286636, + 0.014618043787777424, + 0.00886202696710825, + 0.009441470727324486, + 0.010936688631772995, + 0.008438754826784134, + 0.009962134063243866, + 0.011768072843551636, + 0.009057391434907913, + 0.012520825490355492, + 0.012101223692297935, + 0.008661074563860893, + 0.015500891022384167, + 0.025027761235833168, + 0.006957684643566608, + 0.011015541851520538, + 0.007798145059496164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0017134760273620486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10787490010261536, + 0.03549618646502495, + 0.027032669633626938, + 0.09458599239587784, + 0.12391629815101624, + 0.04104743152856827, + 0.10091879963874817, + 0.10459145903587341, + 0.13247500360012054, + 0.2320612668991089 + ], + "credit_map": [ + 0.009493378922343254, + 0.007299124263226986, + 0.009608098305761814, + 0.016960186883807182, + 0.008872653357684612, + 0.006744654383510351, + 0.027949374169111252, + 0.020947422832250595, + 0.006227955687791109, + 0.005447237752377987, + 0.0038110443856567144, + 0.0065870569087564945, + 0.0038239352870732546, + 0.0024212419521063566, + 0.004178424831479788, + 0.002999290358275175, + 0.0036874564830213785, + 0.003067566780373454, + 0.004161471500992775, + 0.004085074178874493, + 0.002453532302752137, + 0.003181502455845475, + 0.003005715785548091, + 0.003390348982065916, + 0.00688251294195652, + 0.007942544296383858, + 0.005204944405704737, + 0.00924000609666109, + 0.03439430519938469, + 0.01580997183918953, + 0.0076167164370417595, + 0.0074949851259589195, + 0.01021299883723259, + 0.020166048780083656, + 0.010806573554873466, + 0.018409594893455505, + 0.026046879589557648, + 0.013480949215590954, + 0.014480583369731903, + 0.010312664322555065, + 0.004011680372059345, + 0.007998401299118996, + 0.005168917588889599, + 0.0058906059712171555, + 0.0057528698816895485, + 0.0045932685025036335, + 0.003655653912574053, + 0.003976030740886927, + 0.01556684821844101, + 0.0187520869076252, + 0.012491680681705475, + 0.011044710874557495, + 0.01446469221264124, + 0.012913218699395657, + 0.0068261041305959225, + 0.008859459310770035, + 0.018890002742409706, + 0.013146990910172462, + 0.008120311424136162, + 0.013644776307046413, + 0.01608159951865673, + 0.01281013898551464, + 0.012971404008567333, + 0.008926237002015114, + 0.010945899412035942, + 0.020473171025514603, + 0.010079062543809414, + 0.020198199898004532, + 0.018372654914855957, + 0.011421016417443752, + 0.025231055915355682, + 0.015753943473100662, + 0.03526976332068443, + 0.02489456720650196, + 0.024228982627391815, + 0.03609679266810417, + 0.03437105566263199, + 0.01884458027780056, + 0.03129340335726738, + 0.027062125504016876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.005251067224889994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809741511940956 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809741511940956, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809741511940956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10383262485265732, + 0.020153433084487915, + 0.02242872305214405, + 0.0838986411690712, + 0.13447213172912598, + 0.07763414829969406, + 0.10757161676883698, + 0.15608923137187958, + 0.2168087661266327, + 0.07711073756217957 + ], + "credit_map": [ + 0.00658063730224967, + 0.007402115501463413, + 0.00953345000743866, + 0.016403716057538986, + 0.008389982394874096, + 0.007378837559372187, + 0.0265492033213377, + 0.021594684571027756, + 0.003395714098587632, + 0.002030519535765052, + 0.0035127862356603146, + 0.002618170343339443, + 0.0016046927776187658, + 0.0021470715291798115, + 0.0022082924842834473, + 0.0026361853815615177, + 0.002957273507490754, + 0.0020972886122763157, + 0.0032862697262316942, + 0.00346946413628757, + 0.002758603310212493, + 0.002380147110670805, + 0.002346782013773918, + 0.003132894868031144, + 0.004949470050632954, + 0.004298186860978603, + 0.006250354927033186, + 0.010315226390957832, + 0.012419023551046848, + 0.015953175723552704, + 0.00835313368588686, + 0.021360067650675774, + 0.011879155412316322, + 0.023179329931735992, + 0.010949281044304371, + 0.018064578995108604, + 0.027540748938918114, + 0.015815826132893562, + 0.015293186530470848, + 0.011750021018087864, + 0.00759796891361475, + 0.01710428111255169, + 0.009059581905603409, + 0.01079104095697403, + 0.011737235821783543, + 0.009713848121464252, + 0.005011748522520065, + 0.0066184489987790585, + 0.011207343079149723, + 0.021372154355049133, + 0.011814546771347523, + 0.017095249146223068, + 0.013841531239449978, + 0.009961755014955997, + 0.013497940264642239, + 0.008781105279922485, + 0.017114387825131416, + 0.02403554692864418, + 0.014200124889612198, + 0.029984286054968834, + 0.024024732410907745, + 0.01367008313536644, + 0.020417196676135063, + 0.012642878107726574, + 0.01675226539373398, + 0.03227412328124046, + 0.016041457653045654, + 0.035294245928525925, + 0.02943512797355652, + 0.018275292590260506, + 0.041393522173166275, + 0.027342723682522774, + 0.009416542015969753, + 0.009260844439268112, + 0.00815613754093647, + 0.012991064228117466, + 0.014419113285839558, + 0.004827444441616535, + 0.010663160122931004, + 0.007376433350145817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.03809741511940956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle close-up slowly zooms in on clusters of vibrant yellow wolfsbane flowers, rendered in Studio Ghibli's signature hand-drawn style. Set in a sun-drenched alpine meadow, lush green blades of grass, each with a soft watercolor texture, sway gently in the breeze. In the background, rolling hills fade into a hazy, distant mountain range, while the azure sky is dotted with fluffy, white Ghibli clouds. The camera focuses on the central flower head, a dense sphere of tightly closed buds. In incredibly smooth, high-frame-rate animation, each delicate petal begins to unfold in graceful, deliberate slowness, their movements as smooth as silk. The buttery yellow petals, outlined with delicate, organic brushstrokes, peel back one by one, revealing a soft, magical inner glow emanating from the core of each flower, as if painted with light itself. As the flowers fully open, this gentle, warm glow becomes more intense, casting a subtle radiance that contrasts beautifully with the bright, golden sunlight bathing the entire meadow, creating a profoundly serene, natural spectacle that captures life’s fleeting, magical moments in bloom in stunning detail.", + "seed": 3399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_8/index_3688_seed_3399_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014985118270851671, + "vlm_logprob": -0.014985118270851671, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.014985118270851671, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012878225184977055 + ], + "opposite_logprob_tokens": [ + -4.88787841796875 + ], + "logprob_reward": -0.012878225184977055, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012878225184977055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039250824600458145, + 0.02220807783305645, + 0.039445385336875916, + 0.0185532309114933, + 0.04008175805211067, + 0.3214043378829956, + 0.05174202471971512, + 0.34351786971092224, + 0.0646926537156105, + 0.059103865176439285 + ], + "credit_map": [ + 0.004440268501639366, + 0.0034775007516145706, + 0.0066515011712908745, + 0.0044646249152719975, + 0.004236782435327768, + 0.005435777362436056, + 0.004535806830972433, + 0.006008562166243792, + 0.0021449837367981672, + 0.0024299800861626863, + 0.0036770140286535025, + 0.0025629072915762663, + 0.003412309568375349, + 0.002681303769350052, + 0.002928833244368434, + 0.002370745874941349, + 0.005694653373211622, + 0.004001215100288391, + 0.0032597335521131754, + 0.004230067133903503, + 0.004135912284255028, + 0.0034949304535984993, + 0.004681307822465897, + 0.009947564452886581, + 0.0014068668242543936, + 0.002178036840632558, + 0.0012471850495785475, + 0.0019598144572228193, + 0.004932204261422157, + 0.00371378636918962, + 0.0016051086131483316, + 0.0015102274483069777, + 0.003571499837562442, + 0.0035329367965459824, + 0.00534919835627079, + 0.006260556634515524, + 0.004701980389654636, + 0.0037965215742588043, + 0.005870027933269739, + 0.006999035365879536, + 0.015908140689134598, + 0.07968919724225998, + 0.02330968715250492, + 0.07542238384485245, + 0.054000504314899445, + 0.019946157932281494, + 0.03378463536500931, + 0.01934363879263401, + 0.0028827530331909657, + 0.008988766930997372, + 0.006030606105923653, + 0.014143170788884163, + 0.004515701439231634, + 0.0029967068694531918, + 0.007488219533115625, + 0.004696102347224951, + 0.022526364773511887, + 0.029590027406811714, + 0.01382108312100172, + 0.1363476812839508, + 0.03641330078244209, + 0.029291963204741478, + 0.05398847535252571, + 0.021538998931646347, + 0.0039046360179781914, + 0.004191403742879629, + 0.0020510642789304256, + 0.011868128553032875, + 0.026096351444721222, + 0.0027299923822283745, + 0.003337015863507986, + 0.010514069348573685, + 0.0032850292045623064, + 0.003829000983387232, + 0.007614471949636936, + 0.014102049171924591, + 0.005202800035476685, + 0.006785998586565256, + 0.005117244552820921, + 0.013167277909815311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.012878225184977055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329378016293049 + ], + "opposite_logprob_tokens": [ + -3.7732937335968018 + ], + "logprob_reward": -0.02329378016293049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329378016293049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07897292077541351, + 0.04061645269393921, + 0.03675997629761696, + 0.014404930174350739, + 0.027676939964294434, + 0.31769606471061707, + 0.058104272931814194, + 0.2420668751001358, + 0.05498502403497696, + 0.12871652841567993 + ], + "credit_map": [ + 0.008730651810765266, + 0.008672337047755718, + 0.012003104202449322, + 0.005989397410303354, + 0.007072357926517725, + 0.014403259381651878, + 0.007321421056985855, + 0.014780393801629543, + 0.00448885653167963, + 0.004902363289147615, + 0.007568820379674435, + 0.004743888974189758, + 0.0047408584505319595, + 0.004152435809373856, + 0.005280730780214071, + 0.004738500341773033, + 0.00822865404188633, + 0.0026525624562054873, + 0.002968646353110671, + 0.0025194003246724606, + 0.0029219314455986023, + 0.008080613799393177, + 0.0035756751894950867, + 0.005812494084239006, + 0.002331929048523307, + 0.001012952532619238, + 0.0010500455973669887, + 0.0008954947115853429, + 0.002755486173555255, + 0.004319226369261742, + 0.0008995930547825992, + 0.0011402037926018238, + 0.0037415705155581236, + 0.0029852171428501606, + 0.0026199854910373688, + 0.004384600557386875, + 0.0031007803045213223, + 0.0032635421957820654, + 0.0030139239970594645, + 0.0045673237182199955, + 0.014945217408239841, + 0.07791948318481445, + 0.022114159539341927, + 0.07293149083852768, + 0.05582151189446449, + 0.02225552685558796, + 0.0332774855196476, + 0.01843121275305748, + 0.003430041018873453, + 0.008401389233767986, + 0.005535703152418137, + 0.01004337053745985, + 0.010038203559815884, + 0.006284306291490793, + 0.008626440539956093, + 0.005744820926338434, + 0.014248049817979336, + 0.027839960530400276, + 0.009103789925575256, + 0.0958549976348877, + 0.02648325264453888, + 0.01683058962225914, + 0.03741500899195671, + 0.014291252009570599, + 0.0031979773193597794, + 0.006262570153921843, + 0.001989342039451003, + 0.009335619397461414, + 0.020593442022800446, + 0.002581604989245534, + 0.003079504007473588, + 0.007944971323013306, + 0.00754616828635335, + 0.00794602744281292, + 0.007823980413377285, + 0.0380907841026783, + 0.008578546345233917, + 0.006772197782993317, + 0.00963092315942049, + 0.04232792928814888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.02329378016293049 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334729720838368 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.001334729720838368, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334729720838368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06963009387254715, + 0.08429338037967682, + 0.038472771644592285, + 0.027004925534129143, + 0.02967323735356331, + 0.08835924416780472, + 0.04277217388153076, + 0.18135522305965424, + 0.25319036841392517, + 0.18524858355522156 + ], + "credit_map": [ + 0.009773106314241886, + 0.007237381301820278, + 0.012498659081757069, + 0.006385247688740492, + 0.0057883248664438725, + 0.010115637443959713, + 0.007901079952716827, + 0.009930655360221863, + 0.0075376383028924465, + 0.010379194281995296, + 0.017245642840862274, + 0.010737797245383263, + 0.010608242824673653, + 0.007393101695924997, + 0.010623381473124027, + 0.009768379852175713, + 0.009235956706106663, + 0.0024115846026688814, + 0.0028758496046066284, + 0.002806343138217926, + 0.003251028247177601, + 0.011372869834303856, + 0.0031189140863716602, + 0.003400219837203622, + 0.0036357627250254154, + 0.002737620146945119, + 0.0024460407439619303, + 0.002143662190064788, + 0.0040115672163665295, + 0.004745521117001772, + 0.0024449904449284077, + 0.004839758388698101, + 0.004259102512151003, + 0.0025928309187293053, + 0.0032376779709011316, + 0.004187962505966425, + 0.003447594353929162, + 0.003516593249514699, + 0.0034447424113750458, + 0.004986729007214308, + 0.004518480971455574, + 0.020593522116541862, + 0.007859439589083195, + 0.019900156185030937, + 0.01431137602776289, + 0.0062822019681334496, + 0.009435337036848068, + 0.005458722356706858, + 0.002178944880142808, + 0.007380068302154541, + 0.0036098307464271784, + 0.010722335427999496, + 0.006011212710291147, + 0.0023023621179163456, + 0.007003533188253641, + 0.003563886508345604, + 0.010876676067709923, + 0.018956877291202545, + 0.006964873522520065, + 0.07214561849832535, + 0.020253708586096764, + 0.01317069586366415, + 0.028077932074666023, + 0.010908831842243671, + 0.01569346711039543, + 0.015118645504117012, + 0.006736052222549915, + 0.04861082509160042, + 0.10797955095767975, + 0.01055612601339817, + 0.009256831370294094, + 0.03923885151743889, + 0.00839427299797535, + 0.010663793422281742, + 0.01798057183623314, + 0.05810096487402916, + 0.011007559485733509, + 0.009172015823423862, + 0.010119284503161907, + 0.059810105711221695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.001334729720838368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625163532793522 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625163532793522, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625163532793522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07273831963539124, + 0.09253992885351181, + 0.1219891905784607, + 0.039317335933446884, + 0.062056008726358414, + 0.14599119126796722, + 0.06449659913778305, + 0.06132587417960167, + 0.23347793519496918, + 0.10606759041547775 + ], + "credit_map": [ + 0.0062909494154155254, + 0.006006550509482622, + 0.012712497264146805, + 0.009487476199865341, + 0.009883862920105457, + 0.008873493410646915, + 0.011986314319074154, + 0.007497174199670553, + 0.009355605579912663, + 0.01119941659271717, + 0.019355902448296547, + 0.010147097520530224, + 0.011953689157962799, + 0.008656815625727177, + 0.011495443992316723, + 0.010375955142080784, + 0.02640058845281601, + 0.009168111719191074, + 0.008730906061828136, + 0.008582902140915394, + 0.010441767051815987, + 0.03552037477493286, + 0.008243240416049957, + 0.014901301823556423, + 0.0033506792970001698, + 0.006808877922594547, + 0.0032123138662427664, + 0.005222272127866745, + 0.00832326989620924, + 0.0038821781054139137, + 0.004752393811941147, + 0.003765349742025137, + 0.006230113562196493, + 0.007523340173065662, + 0.005257345736026764, + 0.0097733773291111, + 0.00984057318419218, + 0.005815141834318638, + 0.010175162926316261, + 0.007440952118486166, + 0.007665760815143585, + 0.040895976126194, + 0.008534394204616547, + 0.03178509324789047, + 0.023673729971051216, + 0.009450617246329784, + 0.01501584704965353, + 0.00896978098899126, + 0.004679650533944368, + 0.01127163041383028, + 0.0033994042314589024, + 0.019238635897636414, + 0.005427082069218159, + 0.007014714181423187, + 0.008643506094813347, + 0.0048219747841358185, + 0.004390563815832138, + 0.0070136296562850475, + 0.003482608590275049, + 0.01945585384964943, + 0.006135760806500912, + 0.007472604047507048, + 0.008794254623353481, + 0.004580592270940542, + 0.010681379586458206, + 0.058183491230010986, + 0.008949590846896172, + 0.042937394231557846, + 0.07266231626272202, + 0.008708402514457703, + 0.009735355153679848, + 0.02162003517150879, + 0.004569838754832745, + 0.005606046412140131, + 0.0083721112459898, + 0.03567834198474884, + 0.006063827313482761, + 0.003626407589763403, + 0.005260051693767309, + 0.036890968680381775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.007625163532793522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects clusters of vibrant yellow wolfsbane flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with Studio Ghibli's signature hand-drawn aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched alpine meadow with lush green grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow zoom-in on the central flower head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flower petals unfold slowly to reveal a soft inner glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793692752718925 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793692752718925, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793692752718925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03088737092912197, + 0.037289105355739594, + 0.010051912628114223, + 0.031534843146800995, + 0.017634393647313118, + 0.126276895403862, + 0.02318987064063549, + 0.12501010298728943, + 0.5591581463813782, + 0.03896736726164818 + ], + "credit_map": [ + 0.0031811839435249567, + 0.0034576398320496082, + 0.004378142766654491, + 0.0034981691278517246, + 0.004625680390745401, + 0.003949819598346949, + 0.004532193299382925, + 0.003264539409428835, + 0.004113788716495037, + 0.0034930077381432056, + 0.00782580766826868, + 0.004026737995445728, + 0.006054390221834183, + 0.0033016277011483908, + 0.004448740277439356, + 0.004025002475827932, + 0.0014212758978828788, + 0.0013203538255766034, + 0.001321168034337461, + 0.0011843941174447536, + 0.00121526257134974, + 0.0012453473173081875, + 0.0010436680167913437, + 0.0013004419161006808, + 0.0032811034470796585, + 0.0025675171054899693, + 0.0017982355784624815, + 0.0014259585877880454, + 0.009069742634892464, + 0.008326146751642227, + 0.002009031595662236, + 0.0030571059323847294, + 0.0012613940052688122, + 0.0019028224050998688, + 0.0012976500438526273, + 0.002730247564613819, + 0.0025259561371058226, + 0.002333502285182476, + 0.004062617663294077, + 0.001520200283266604, + 0.008257457055151463, + 0.03146635740995407, + 0.007533158175647259, + 0.029634032398462296, + 0.018491404131054878, + 0.010340170934796333, + 0.01271146908402443, + 0.007842847146093845, + 0.001430951408110559, + 0.0032230536453425884, + 0.0018781175604090095, + 0.007600625976920128, + 0.0019503120565786958, + 0.0014679592568427324, + 0.0036539596039801836, + 0.001984888454899192, + 0.007365916855633259, + 0.012063413858413696, + 0.005700602661818266, + 0.04973597824573517, + 0.014045449905097485, + 0.009015440009534359, + 0.019555864855647087, + 0.0075274198316037655, + 0.027285778895020485, + 0.0825369656085968, + 0.014626703225076199, + 0.10905209183692932, + 0.2136131078004837, + 0.02035852149128914, + 0.0233905091881752, + 0.06829443573951721, + 0.0025066265370696783, + 0.004839020781219006, + 0.006838732399046421, + 0.005996511783450842, + 0.005153784062713385, + 0.004006559029221535, + 0.004279434680938721, + 0.0053466977551579475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.029793692752718925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_9/index_tensor([9507, 9507])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_9/index_tensor([9507, 9507])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7eb071d6dccbc1fb03f3ca078fa4e0e4e21c654b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_9/index_tensor([9507, 9507])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, overhead long shot, captured in Tim Burton's signature gothic fantasy style, shows a massive, distorted rocket slowly ascending from a crude launch pad constructed of rusted scaffolding and bizarre pipes. Rather than a smooth, streamlined form, the rocket's body takes on a grotesque, spiraling silhouette. Its deep black metal shell is studded with rivets of varying sizes and haphazard welds, like the work of a crazed blacksmith. As the camera slowly tilts upward, we see thick, inky black smoke billowing from the base of the rocket, rather than a bright flame. This plume stands out against the gloomy, deep purple sky, dotted with swirling clouds. The rocket climbs upwards with a heavy and laborious posture, trembling and shaking, casting a stretched, twisting, grotesque shadow on the barren and cracked ground. As the rocket rises, the shadow dances and deforms. The whole scene is filled with an absurd, grand and slightly uneasy industrial fairy tale atmosphere in desaturated cold colors.", + "seed": 32450, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_9/index_9507_seed_32450_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13566308952867984, + "vlm_logprob": -0.13566308952867984, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.13566308952867984, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.48710358142852783 + ], + "opposite_logprob_tokens": [ + -0.9871035814285278 + ], + "logprob_reward": -0.48710358142852783, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.48710358142852783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08514665067195892, + 0.16774189472198486, + 0.06472866237163544, + 0.050561849027872086, + 0.04457911103963852, + 0.3170347213745117, + 0.0887768566608429, + 0.05550990626215935, + 0.07395315170288086, + 0.05196717008948326 + ], + "credit_map": [ + 0.008639843203127384, + 0.01368632260710001, + 0.007670044898986816, + 0.010404380038380623, + 0.012220274657011032, + 0.014253038913011551, + 0.008789336308836937, + 0.009483418427407742, + 0.030925577506422997, + 0.03305136412382126, + 0.00887721311300993, + 0.00792478397488594, + 0.009385027922689915, + 0.06175416335463524, + 0.0056988950818777084, + 0.01012488640844822, + 0.010626324452459812, + 0.008241478353738785, + 0.009977908805012703, + 0.009092199616134167, + 0.004327306989580393, + 0.0107796099036932, + 0.004643172491341829, + 0.00704067200422287, + 0.0057694269344210625, + 0.007510602008551359, + 0.0053578587248921394, + 0.011627459898591042, + 0.0028807902708649635, + 0.0061631640419363976, + 0.005225021857768297, + 0.00602752948179841, + 0.006093386095017195, + 0.0035902007948607206, + 0.004608366638422012, + 0.004897573497146368, + 0.007753229700028896, + 0.005909913219511509, + 0.0069243647158145905, + 0.00480208033695817, + 0.10715090483427048, + 0.009103422053158283, + 0.12418622523546219, + 0.025285759940743446, + 0.009013024158775806, + 0.0196228064596653, + 0.00838425848633051, + 0.014288345351815224, + 0.013745791278779507, + 0.008434334769845009, + 0.02167101390659809, + 0.008359785191714764, + 0.009393871761858463, + 0.008290213532745838, + 0.010672166012227535, + 0.008209685795009136, + 0.005983633454889059, + 0.006091449875384569, + 0.009162267670035362, + 0.006418414413928986, + 0.004526711069047451, + 0.010856697335839272, + 0.005480945110321045, + 0.006989793386310339, + 0.007696333806961775, + 0.007019348442554474, + 0.015743885189294815, + 0.007324243430048227, + 0.0083332359790802, + 0.011992847546935081, + 0.009133496321737766, + 0.006709773559123278, + 0.004585242364555597, + 0.005069095175713301, + 0.009314668364822865, + 0.005141252186149359, + 0.007474503014236689, + 0.0068840025924146175, + 0.0064179301261901855, + 0.007080481853336096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.48710358142852783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900112122297287 + ], + "opposite_logprob_tokens": [ + -2.5790011882781982 + ], + "logprob_reward": -0.07900112122297287, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900112122297287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03228366747498512, + 0.10903041064739227, + 0.03786927089095116, + 0.02794293314218521, + 0.06184709072113037, + 0.5223089456558228, + 0.07836861908435822, + 0.04212411493062973, + 0.04320019111037254, + 0.04502473771572113 + ], + "credit_map": [ + 0.0023130038753151894, + 0.007592952344566584, + 0.002636911114677787, + 0.004761905409395695, + 0.0060507128946483135, + 0.003359528025612235, + 0.0028204810805618763, + 0.0027481766883283854, + 0.016970301046967506, + 0.018146319314837456, + 0.005394984968006611, + 0.00749551085755229, + 0.012318889610469341, + 0.03391621634364128, + 0.008404537104070187, + 0.00638364814221859, + 0.00596641143783927, + 0.004723540507256985, + 0.005111756268888712, + 0.0060811592265963554, + 0.0029483961407095194, + 0.007048410829156637, + 0.0027279932983219624, + 0.003261605743318796, + 0.003719134023413062, + 0.003865892067551613, + 0.0023216756526380777, + 0.00606294721364975, + 0.0017187796765938401, + 0.0031413461547344923, + 0.003566552186384797, + 0.0035466093104332685, + 0.008475604467093945, + 0.004329313989728689, + 0.0045737894251942635, + 0.008530057966709137, + 0.00958190020173788, + 0.009707039222121239, + 0.010996335186064243, + 0.005653052125126123, + 0.14387045800685883, + 0.01468689925968647, + 0.21422883868217468, + 0.04760117083787918, + 0.01701599545776844, + 0.05517030879855156, + 0.0108241718262434, + 0.018911117687821388, + 0.022303646430373192, + 0.00667308596894145, + 0.013073423877358437, + 0.007595630828291178, + 0.008042111061513424, + 0.007618923205882311, + 0.0076293800957500935, + 0.005432418081909418, + 0.0050955177284777164, + 0.004338930826634169, + 0.008939003571867943, + 0.004868648946285248, + 0.005774331744760275, + 0.005689399782568216, + 0.0034499838948249817, + 0.003968304954469204, + 0.004950779024511576, + 0.003181000705808401, + 0.009586874395608902, + 0.005014199297875166, + 0.003967205993831158, + 0.007099107839167118, + 0.005556927528232336, + 0.0038441026117652655, + 0.004485271405428648, + 0.004644560627639294, + 0.004683200735598803, + 0.004810745362192392, + 0.010403062216937542, + 0.0071812220849096775, + 0.0037733230274170637, + 0.005043353419750929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.07900112122297287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003071159124374 + ], + "opposite_logprob_tokens": [ + -2.695030689239502 + ], + "logprob_reward": -0.07003071159124374, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003071159124374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05792063847184181, + 0.157791867852211, + 0.028278086334466934, + 0.0505952462553978, + 0.07029180973768234, + 0.2562592625617981, + 0.17524145543575287, + 0.048914581537246704, + 0.08612196147441864, + 0.06858508288860321 + ], + "credit_map": [ + 0.00504324259236455, + 0.016479073092341423, + 0.003514687530696392, + 0.0049163526855409145, + 0.007200303953140974, + 0.011771177873015404, + 0.004518017638474703, + 0.004477792419493198, + 0.020915884524583817, + 0.03617003187537193, + 0.009931933134794235, + 0.008637289516627789, + 0.011144384741783142, + 0.050586339086294174, + 0.01171308383345604, + 0.008692934177815914, + 0.0035567472223192453, + 0.003511585295200348, + 0.0045145913027226925, + 0.004681124351918697, + 0.002048359252512455, + 0.004048902541399002, + 0.0025747031904757023, + 0.0033420752733945847, + 0.009053192101418972, + 0.008168516680598259, + 0.0047712065279483795, + 0.00988178700208664, + 0.0034017765428870916, + 0.005371791776269674, + 0.0054461159743368626, + 0.00450086547061801, + 0.01045571081340313, + 0.0043175402097404, + 0.010316174477338791, + 0.011067400686442852, + 0.00846260879188776, + 0.008642210625112057, + 0.010601072572171688, + 0.006429099012166262, + 0.08122645318508148, + 0.007465618196874857, + 0.10472666472196579, + 0.021388346329331398, + 0.007370798848569393, + 0.018908079713582993, + 0.005782728083431721, + 0.009390612132847309, + 0.02482837252318859, + 0.011981911957263947, + 0.05303167551755905, + 0.017840536311268806, + 0.015330604277551174, + 0.018986547365784645, + 0.017992978915572166, + 0.015248855575919151, + 0.005426271352916956, + 0.003255808725953102, + 0.010391213931143284, + 0.004110402427613735, + 0.003213509451597929, + 0.013193205930292606, + 0.0037941255141049623, + 0.005530048627406359, + 0.007810404058545828, + 0.008063065819442272, + 0.01697176694869995, + 0.01079882588237524, + 0.01301782950758934, + 0.010587096214294434, + 0.01011607050895691, + 0.00875691045075655, + 0.009966439567506313, + 0.006995342671871185, + 0.015600461512804031, + 0.005185032729059458, + 0.007454092614352703, + 0.009990489110350609, + 0.004567507188767195, + 0.008825723081827164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.07003071159124374 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038088925182819366 + ], + "opposite_logprob_tokens": [ + -3.2880890369415283 + ], + "logprob_reward": -0.038088925182819366, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088925182819366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04462196305394173, + 0.09561946988105774, + 0.07212351262569427, + 0.03849896043539047, + 0.033538300544023514, + 0.1270693838596344, + 0.15097667276859283, + 0.21377962827682495, + 0.15420082211494446, + 0.06957131624221802 + ], + "credit_map": [ + 0.004270348697900772, + 0.008384013548493385, + 0.003750622970983386, + 0.004082279279828072, + 0.007055800408124924, + 0.007629461586475372, + 0.003921230789273977, + 0.005528208799660206, + 0.012178878299891949, + 0.01997969299554825, + 0.006880698725581169, + 0.006203177385032177, + 0.009329026564955711, + 0.02882508933544159, + 0.004583847243338823, + 0.007639062590897083, + 0.009873786941170692, + 0.008109752088785172, + 0.01126918476074934, + 0.012190687470138073, + 0.005729028955101967, + 0.01145752239972353, + 0.006250813137739897, + 0.007242740597575903, + 0.006441532634198666, + 0.005045956466346979, + 0.00525427982211113, + 0.007605846505612135, + 0.0022343944292515516, + 0.004122571088373661, + 0.00423111068084836, + 0.003563268342986703, + 0.0039863889105618, + 0.0024878920521587133, + 0.0033091676887124777, + 0.0047097401693463326, + 0.005678780842572451, + 0.003921826370060444, + 0.005821076221764088, + 0.0036234271246939898, + 0.024416347965598106, + 0.00952711421996355, + 0.02484031766653061, + 0.016459962353110313, + 0.011490230448544025, + 0.012478329241275787, + 0.010579812340438366, + 0.01727726310491562, + 0.015999017283320427, + 0.019658833742141724, + 0.028543399646878242, + 0.016744006425142288, + 0.019465310499072075, + 0.016389423981308937, + 0.02121763490140438, + 0.01295903418213129, + 0.02009052410721779, + 0.025981415063142776, + 0.02668527513742447, + 0.022453512996435165, + 0.01971898227930069, + 0.04370801895856857, + 0.02768898941576481, + 0.02745291404426098, + 0.01631855219602585, + 0.011407192796468735, + 0.02948380634188652, + 0.01310124434530735, + 0.019452176988124847, + 0.028048500418663025, + 0.017969615757465363, + 0.018419720232486725, + 0.006474366877228022, + 0.00714518316090107, + 0.013554532080888748, + 0.006740102544426918, + 0.01044890470802784, + 0.01150008849799633, + 0.005400660913437605, + 0.008307482115924358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.038088925182819366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091108217835426 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091108217835426, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091108217835426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043680548667907715, + 0.23291370272636414, + 0.1042170450091362, + 0.029151950031518936, + 0.07320448011159897, + 0.3087243437767029, + 0.04743190482258797, + 0.07209522277116776, + 0.04249504953622818, + 0.046085719019174576 + ], + "credit_map": [ + 0.005616495385766029, + 0.007048250176012516, + 0.003336126683279872, + 0.0037458057049661875, + 0.007317871320992708, + 0.007394479122012854, + 0.004073271527886391, + 0.005148251075297594, + 0.0423721969127655, + 0.04892943799495697, + 0.010664604604244232, + 0.010223058052361012, + 0.013065495528280735, + 0.08861790597438812, + 0.006181655917316675, + 0.012859348207712173, + 0.019266316667199135, + 0.014832362532615662, + 0.01156656350940466, + 0.018175456672906876, + 0.006011696066707373, + 0.017007386311888695, + 0.00601225858554244, + 0.011345005594193935, + 0.00461038202047348, + 0.004479659255594015, + 0.0029157642275094986, + 0.0052173323929309845, + 0.001935903332196176, + 0.0037285673897713423, + 0.0028281959239393473, + 0.003436148166656494, + 0.012103688903152943, + 0.00520758330821991, + 0.010080551728606224, + 0.008873891085386276, + 0.006858049891889095, + 0.009181761182844639, + 0.012080205604434013, + 0.008818752132356167, + 0.0894373431801796, + 0.009009766392409801, + 0.12690170109272003, + 0.02734161913394928, + 0.009357579052448273, + 0.029805738478899002, + 0.006224561017006636, + 0.010646047070622444, + 0.008571582846343517, + 0.004266610834747553, + 0.010320414789021015, + 0.004531229380518198, + 0.004980366677045822, + 0.005857020150870085, + 0.004855428356677294, + 0.004049248527735472, + 0.006545376498252153, + 0.008950716815888882, + 0.008650935254991055, + 0.008040874265134335, + 0.007446222007274628, + 0.013990945182740688, + 0.010006947442889214, + 0.008463202975690365, + 0.006267608609050512, + 0.002908233320340514, + 0.010716798715293407, + 0.004198241513222456, + 0.004102088510990143, + 0.006095937453210354, + 0.003451661905273795, + 0.004754479043185711, + 0.004967817105352879, + 0.0045101214200258255, + 0.012933005578815937, + 0.003295235801488161, + 0.00414043664932251, + 0.007701735943555832, + 0.0031766542233526707, + 0.005360710434615612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.004091108217835426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead long shot, captured in Tim Burton's signature gothic fantasy style, shows a massive, distorted rocket slowly ascending from a crude launch pad constructed of rusted scaffolding and bizarre pipes. Rather than a smooth, streamlined form, the rocket's body takes on a grotesque, spiraling silhouette. Its deep black metal shell is studded with rivets of varying sizes and haphazard welds, like the work of a crazed blacksmith. As the camera slowly tilts upward, we see thick, inky black smoke billowing from the base of the rocket, rather than a bright flame. This plume stands out against the gloomy, deep purple sky, dotted with swirling clouds. The rocket climbs upwards with a heavy and laborious posture, trembling and shaking, casting a stretched, twisting, grotesque shadow on the barren and cracked ground. As the rocket rises, the shadow dances and deforms. The whole scene is filled with an absurd, grand and slightly uneasy industrial fairy tale atmosphere in desaturated cold colors.", + "seed": 32451, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_9/index_9507_seed_32451_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16793957646004856, + "vlm_logprob": -0.16793957646004856, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.16793957646004856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5355564951896667 + ], + "opposite_logprob_tokens": [ + -0.9105564951896667 + ], + "logprob_reward": -0.5355564951896667, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5355564951896667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2252913862466812, + 0.06278398633003235, + 0.053971294313669205, + 0.06112675368785858, + 0.1402398943901062, + 0.14436134696006775, + 0.08733655512332916, + 0.058224648237228394, + 0.1348017007112503, + 0.03186245262622833 + ], + "credit_map": [ + 0.016283005475997925, + 0.016903212293982506, + 0.021898161619901657, + 0.030051421374082565, + 0.04947558790445328, + 0.01862773299217224, + 0.04442738741636276, + 0.027624880895018578, + 0.007316696457564831, + 0.007315962575376034, + 0.007668635807931423, + 0.0063637555576860905, + 0.009214515797793865, + 0.00742638623341918, + 0.007172532845288515, + 0.010305502451956272, + 0.0075940643437206745, + 0.007253504823893309, + 0.00539606437087059, + 0.0036404782440513372, + 0.004761659074574709, + 0.010854274965822697, + 0.0040535153821110725, + 0.01041773147881031, + 0.004908973351120949, + 0.007075431291013956, + 0.005323837511241436, + 0.010476640425622463, + 0.011071795597672462, + 0.008624701760709286, + 0.006341989617794752, + 0.007303387857973576, + 0.013729780912399292, + 0.035483211278915405, + 0.00991284754127264, + 0.013969724997878075, + 0.01672377809882164, + 0.016919082030653954, + 0.013211014680564404, + 0.020290467888116837, + 0.015722131356596947, + 0.009301954880356789, + 0.015623086132109165, + 0.018730297684669495, + 0.02630775049328804, + 0.016327237710356712, + 0.02009924314916134, + 0.02224963717162609, + 0.0053181517869234085, + 0.00829598493874073, + 0.006224760320037603, + 0.016869816929101944, + 0.012413498014211655, + 0.010965545661747456, + 0.011336042545735836, + 0.015912752598524094, + 0.004693123046308756, + 0.007618578616529703, + 0.00460423156619072, + 0.008749778382480145, + 0.010048671625554562, + 0.007715636398643255, + 0.008415824733674526, + 0.006378805264830589, + 0.009998908266425133, + 0.019663112238049507, + 0.020161544904112816, + 0.014314819127321243, + 0.01412453968077898, + 0.0255585890263319, + 0.016254836693406105, + 0.014725346118211746, + 0.0031686185393482447, + 0.0030930947978049517, + 0.003039751900359988, + 0.008003687486052513, + 0.005395143758505583, + 0.00363905425183475, + 0.0032255970872938633, + 0.002297503175213933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.5355564951896667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032962262630463 + ], + "opposite_logprob_tokens": [ + -2.350329637527466 + ], + "logprob_reward": -0.10032962262630463, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032962262630463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1274980753660202, + 0.09072020649909973, + 0.03706062212586403, + 0.2007090449333191, + 0.253184050321579, + 0.02563738077878952, + 0.08892849087715149, + 0.0463097020983696, + 0.05859635770320892, + 0.07135611027479172 + ], + "credit_map": [ + 0.015230012126266956, + 0.01282615028321743, + 0.015068786218762398, + 0.015954207628965378, + 0.01859978958964348, + 0.010347484610974789, + 0.021591676399111748, + 0.01787995733320713, + 0.007296763360500336, + 0.011795422993600368, + 0.00923407357186079, + 0.01060398668050766, + 0.01883171685039997, + 0.009620897471904755, + 0.012332525104284286, + 0.011004828847944736, + 0.004379927180707455, + 0.005187428090721369, + 0.002636417280882597, + 0.0034931697882711887, + 0.00636295136064291, + 0.0055625298991799355, + 0.003538883291184902, + 0.005899315234273672, + 0.021206341683864594, + 0.02840806357562542, + 0.013386735692620277, + 0.028726495802402496, + 0.03275854513049126, + 0.02786635421216488, + 0.018391365185379982, + 0.02996513806283474, + 0.025450067594647408, + 0.05942637100815773, + 0.017373478040099144, + 0.02702685073018074, + 0.031009012833237648, + 0.030379632487893105, + 0.02390500158071518, + 0.03861361742019653, + 0.003881055861711502, + 0.002652011113241315, + 0.002842950401827693, + 0.0027946303598582745, + 0.0038510651793330908, + 0.002808715682476759, + 0.0032893677707761526, + 0.0035175830125808716, + 0.007504464592784643, + 0.016438262537121773, + 0.0076555972918868065, + 0.01450609602034092, + 0.01590237393975258, + 0.008511095307767391, + 0.010676548816263676, + 0.0077340565621852875, + 0.0031418928410857916, + 0.0057904901914298534, + 0.004162401892244816, + 0.006235490553081036, + 0.006545690819621086, + 0.007769285701215267, + 0.008875735104084015, + 0.0037887180224061012, + 0.006478545721620321, + 0.005376711022108793, + 0.0064431591890752316, + 0.008006769232451916, + 0.006265046074986458, + 0.008709611371159554, + 0.009140610694885254, + 0.008175903931260109, + 0.009068354964256287, + 0.009794626384973526, + 0.010045976378023624, + 0.01009121723473072, + 0.009762540459632874, + 0.007850785739719868, + 0.009569963440299034, + 0.005172648932784796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.10032962262630463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032908618450165 + ], + "opposite_logprob_tokens": [ + -2.3503291606903076 + ], + "logprob_reward": -0.10032908618450165, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032908618450165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09832921624183655, + 0.04206770285964012, + 0.03260092809796333, + 0.2329111099243164, + 0.2670789062976837, + 0.019410453736782074, + 0.15000469982624054, + 0.04780595377087593, + 0.03833317011594772, + 0.07145790755748749 + ], + "credit_map": [ + 0.013076988980174065, + 0.009848359040915966, + 0.010177964344620705, + 0.012254631146788597, + 0.01778174377977848, + 0.006888945586979389, + 0.013382947072386742, + 0.014917641878128052, + 0.004375741351395845, + 0.007380042690783739, + 0.004503382369875908, + 0.003937380388379097, + 0.00690354872494936, + 0.0048750219866633415, + 0.0037582104559987783, + 0.00633437093347311, + 0.0022632693871855736, + 0.00419607013463974, + 0.0025453746784478426, + 0.0032377587631344795, + 0.005071908701211214, + 0.0052752248011529446, + 0.0035018976777791977, + 0.006509424187242985, + 0.02894471399486065, + 0.03436511382460594, + 0.017351943999528885, + 0.027634475380182266, + 0.039559535682201385, + 0.033413439989089966, + 0.021284565329551697, + 0.030357331037521362, + 0.02718105912208557, + 0.06160307675600052, + 0.01738535799086094, + 0.02709120512008667, + 0.03129333630204201, + 0.034224435687065125, + 0.025195345282554626, + 0.04310505837202072, + 0.0025652276817709208, + 0.001905931392684579, + 0.0018266538390889764, + 0.0020834796596318483, + 0.003499446902424097, + 0.001752272597514093, + 0.002788413781672716, + 0.0029890278819948435, + 0.009383108466863632, + 0.01614302210509777, + 0.015063110738992691, + 0.02222256548702717, + 0.020385580137372017, + 0.017889508977532387, + 0.01969752460718155, + 0.029220281168818474, + 0.003328819992020726, + 0.005616988521069288, + 0.004503491334617138, + 0.006227520294487476, + 0.007510414347052574, + 0.008536460809409618, + 0.007776640355587006, + 0.004305622074753046, + 0.003500438528135419, + 0.0037644593976438046, + 0.004124299623072147, + 0.0069481679238379, + 0.00516135897487402, + 0.005095565225929022, + 0.004876099061220884, + 0.004862781148403883, + 0.007460103835910559, + 0.009570316411554813, + 0.009694072417914867, + 0.01606704480946064, + 0.009245923720300198, + 0.006113939918577671, + 0.00815479177981615, + 0.005151709541678429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.10032908618450165 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029177367687225 + ], + "opposite_logprob_tokens": [ + -2.3502917289733887 + ], + "logprob_reward": -0.10029177367687225, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029177367687225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0774901807308197, + 0.075364850461483, + 0.07902474701404572, + 0.09471222013235092, + 0.16170726716518402, + 0.04206915572285652, + 0.16589920222759247, + 0.06774977594614029, + 0.19941550493240356, + 0.03656712919473648 + ], + "credit_map": [ + 0.014206476509571075, + 0.007519071456044912, + 0.011399069800972939, + 0.008024804294109344, + 0.00693715363740921, + 0.006300434470176697, + 0.010486503131687641, + 0.012616664171218872, + 0.007446670904755592, + 0.010376980528235435, + 0.008179089054465294, + 0.008296872489154339, + 0.010442912578582764, + 0.009061908349394798, + 0.008496616967022419, + 0.013063795864582062, + 0.012106329202651978, + 0.014314561150968075, + 0.009199953638017178, + 0.005304337479174137, + 0.006941395346075296, + 0.010799734853208065, + 0.00569218210875988, + 0.014666243456304073, + 0.012379881925880909, + 0.012834885157644749, + 0.00981232337653637, + 0.009557527489960194, + 0.015342005528509617, + 0.014105129055678844, + 0.006969735026359558, + 0.01371072418987751, + 0.01660502515733242, + 0.03525346890091896, + 0.011397640220820904, + 0.017990853637456894, + 0.017462924122810364, + 0.019490767270326614, + 0.01612146571278572, + 0.027385104447603226, + 0.005276001524180174, + 0.0032842636574059725, + 0.005172303412109613, + 0.00401467178016901, + 0.007780775427818298, + 0.004353930242359638, + 0.006237220484763384, + 0.005949986632913351, + 0.014344785362482071, + 0.030366074293851852, + 0.010013856925070286, + 0.02536354959011078, + 0.023492295295000076, + 0.01933249458670616, + 0.020173586905002594, + 0.02281254343688488, + 0.004093721974641085, + 0.0072549269534647465, + 0.005227198824286461, + 0.009633266367018223, + 0.009115180000662804, + 0.01390388235449791, + 0.008400012739002705, + 0.010121583938598633, + 0.022358546033501625, + 0.020610062405467033, + 0.03834174945950508, + 0.022477606311440468, + 0.018933143466711044, + 0.027002358809113503, + 0.024204108864068985, + 0.02548791654407978, + 0.0023685190826654434, + 0.004639580845832825, + 0.0028103082440793514, + 0.010963096283376217, + 0.005903662648051977, + 0.0037388347554951906, + 0.003863964695483446, + 0.0022791598457843065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.10029177367687225 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031909046228975058 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031909046228975058, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031909046228975058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09789398312568665, + 0.03791538253426552, + 0.019507307559251785, + 0.09421229362487793, + 0.1674320548772812, + 0.03328609839081764, + 0.1592472493648529, + 0.08582711964845657, + 0.2772952914237976, + 0.02738322876393795 + ], + "credit_map": [ + 0.009105119854211807, + 0.009414571337401867, + 0.01014204602688551, + 0.011415227316319942, + 0.019961025565862656, + 0.008212495595216751, + 0.01719317026436329, + 0.01245032623410225, + 0.003774739569053054, + 0.008612188510596752, + 0.005310756154358387, + 0.003630619030445814, + 0.0030534029938280582, + 0.005572892725467682, + 0.0026621047873049974, + 0.005298678297549486, + 0.0014578921254724264, + 0.003051822306588292, + 0.0019490249687805772, + 0.001842903788201511, + 0.0032354057766497135, + 0.001982719637453556, + 0.002060630125924945, + 0.003926908131688833, + 0.006879716645926237, + 0.012801161035895348, + 0.005388622637838125, + 0.01692943274974823, + 0.01722588576376438, + 0.013029461726546288, + 0.009445196948945522, + 0.012512817047536373, + 0.018803047016263008, + 0.036863263696432114, + 0.011578282341361046, + 0.01754840835928917, + 0.01995295099914074, + 0.020559601485729218, + 0.01561838760972023, + 0.026508113369345665, + 0.0033140359446406364, + 0.0019265073351562023, + 0.0030328885186463594, + 0.005151128396391869, + 0.0049391030333936214, + 0.0042893728241324425, + 0.005936424247920513, + 0.004696636460721493, + 0.00853921938687563, + 0.009589388966560364, + 0.01539789978414774, + 0.02911960892379284, + 0.019070038571953773, + 0.021572502329945564, + 0.02013966254889965, + 0.0358189195394516, + 0.00587215693667531, + 0.007423080503940582, + 0.005811826325953007, + 0.00923986081033945, + 0.014484051614999771, + 0.023180613294243813, + 0.010876322165131569, + 0.008939210325479507, + 0.03092177025973797, + 0.03610268235206604, + 0.02918328158557415, + 0.0268352460116148, + 0.035662222653627396, + 0.04360925406217575, + 0.03491591289639473, + 0.04006493464112282, + 0.0030644601210951805, + 0.0024712090380489826, + 0.004926537629216909, + 0.0033584425691515207, + 0.0040770648047327995, + 0.003400517860427499, + 0.00300137628801167, + 0.003083620686084032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.0031909046228975058 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead long shot, captured in Tim Burton's signature gothic fantasy style, shows a massive, distorted rocket slowly ascending from a crude launch pad constructed of rusted scaffolding and bizarre pipes. Rather than a smooth, streamlined form, the rocket's body takes on a grotesque, spiraling silhouette. Its deep black metal shell is studded with rivets of varying sizes and haphazard welds, like the work of a crazed blacksmith. As the camera slowly tilts upward, we see thick, inky black smoke billowing from the base of the rocket, rather than a bright flame. This plume stands out against the gloomy, deep purple sky, dotted with swirling clouds. The rocket climbs upwards with a heavy and laborious posture, trembling and shaking, casting a stretched, twisting, grotesque shadow on the barren and cracked ground. As the rocket rises, the shadow dances and deforms. The whole scene is filled with an absurd, grand and slightly uneasy industrial fairy tale atmosphere in desaturated cold colors.", + "seed": 32452, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_9/index_9507_seed_32452_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10223132707178592, + "vlm_logprob": -0.10223132707178592, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.10223132707178592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26649099588394165 + ], + "opposite_logprob_tokens": [ + -1.5164909362792969 + ], + "logprob_reward": -0.26649099588394165, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26649099588394165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05411093309521675, + 0.08061947673559189, + 0.09027531743049622, + 0.18083781003952026, + 0.13618621230125427, + 0.0421660915017128, + 0.13927003741264343, + 0.17225968837738037, + 0.0465204119682312, + 0.05775397643446922 + ], + "credit_map": [ + 0.005361041519790888, + 0.006186283193528652, + 0.008416173048317432, + 0.006522832438349724, + 0.007641601841896772, + 0.005384952295571566, + 0.007424702402204275, + 0.007173345889896154, + 0.007396393921226263, + 0.007717925123870373, + 0.008807873353362083, + 0.010949907824397087, + 0.013005988672375679, + 0.0073334090411663055, + 0.017291400581598282, + 0.008116578683257103, + 0.010144614614546299, + 0.012831541709601879, + 0.008967068046331406, + 0.011910717003047466, + 0.012882024981081486, + 0.008527789264917374, + 0.01708899438381195, + 0.007922574877738953, + 0.021167483180761337, + 0.01510099321603775, + 0.020927628502249718, + 0.026925817131996155, + 0.018331483006477356, + 0.025981320068240166, + 0.02282017469406128, + 0.029582904651761055, + 0.011300641112029552, + 0.009770544245839119, + 0.015046235173940659, + 0.0219822246581316, + 0.008662613108754158, + 0.028063084930181503, + 0.022991210222244263, + 0.018369661644101143, + 0.004003595560789108, + 0.0051832799799740314, + 0.0045659481547772884, + 0.005751316901296377, + 0.003922735340893269, + 0.005646411795169115, + 0.005294171627610922, + 0.007798631209880114, + 0.008616074919700623, + 0.012494780123233795, + 0.033407725393772125, + 0.011572827585041523, + 0.013811977580189705, + 0.013234312646090984, + 0.011358961462974548, + 0.03477337956428528, + 0.018690163269639015, + 0.02034168690443039, + 0.018038712441921234, + 0.023736516013741493, + 0.024549352005124092, + 0.018415380269289017, + 0.021412678062915802, + 0.02707521617412567, + 0.007318391930311918, + 0.005055712535977364, + 0.003804674604907632, + 0.005403181537985802, + 0.005241242703050375, + 0.006631395313888788, + 0.005120450165122747, + 0.007945365272462368, + 0.004752218257635832, + 0.00530683808028698, + 0.008528421632945538, + 0.00534097058698535, + 0.0041564712300896645, + 0.014994105324149132, + 0.006138717755675316, + 0.008536229841411114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.26649099588394165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295194178819656 + ], + "opposite_logprob_tokens": [ + -2.2379519939422607 + ], + "logprob_reward": -0.11295194178819656, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295194178819656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09517686069011688, + 0.03601005673408508, + 0.08627864718437195, + 0.10371950268745422, + 0.07383615523576736, + 0.09774889796972275, + 0.09147036820650101, + 0.1520014852285385, + 0.18084639310836792, + 0.08291158825159073 + ], + "credit_map": [ + 0.008218424394726753, + 0.010484541766345501, + 0.012463399209082127, + 0.013625155203044415, + 0.012807467021048069, + 0.009550890885293484, + 0.012143239378929138, + 0.01588374748826027, + 0.004412073642015457, + 0.0034074571449309587, + 0.003619685536250472, + 0.0035391359124332666, + 0.00683273421600461, + 0.0037571333814412355, + 0.005818940699100494, + 0.004622895270586014, + 0.01057517435401678, + 0.01485681813210249, + 0.00843023881316185, + 0.009048854000866413, + 0.014345573261380196, + 0.008914933539927006, + 0.011122052557766438, + 0.0089850053191185, + 0.012077424675226212, + 0.01000513881444931, + 0.018652446568012238, + 0.009502114728093147, + 0.010822348296642303, + 0.014770865440368652, + 0.012040236964821815, + 0.015848923474550247, + 0.005368764046579599, + 0.008399514481425285, + 0.00965044740587473, + 0.010812577791512012, + 0.008655563928186893, + 0.015389549545943737, + 0.008363372646272182, + 0.007196363527327776, + 0.008980226702988148, + 0.011149384081363678, + 0.012388919480144978, + 0.014636914245784283, + 0.007296522613614798, + 0.022138865664601326, + 0.012079501524567604, + 0.0090785576030612, + 0.0058946083299815655, + 0.010212032124400139, + 0.01589374616742134, + 0.011038084514439106, + 0.013797777704894543, + 0.01584634557366371, + 0.010302732698619366, + 0.008485037833452225, + 0.013897404074668884, + 0.015935905277729034, + 0.014834877103567123, + 0.022806823253631592, + 0.018880357965826988, + 0.02058526873588562, + 0.02162415347993374, + 0.02343670278787613, + 0.022215189412236214, + 0.03185282275080681, + 0.01197450514882803, + 0.009811708703637123, + 0.013060578145086765, + 0.05358947440981865, + 0.008995306678116322, + 0.02934679575264454, + 0.008054232224822044, + 0.007775537203997374, + 0.007588214706629515, + 0.015460766851902008, + 0.011284762062132359, + 0.01024155505001545, + 0.012881261296570301, + 0.009625253267586231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.11295194178819656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062058817595243454 + ], + "opposite_logprob_tokens": [ + -2.812058925628662 + ], + "logprob_reward": -0.062058817595243454, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062058817595243454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06982079893350601, + 0.022853637114167213, + 0.03767402470111847, + 0.060586705803871155, + 0.04396497458219528, + 0.03905574604868889, + 0.21216866374015808, + 0.2809816300868988, + 0.15901054441928864, + 0.07388319075107574 + ], + "credit_map": [ + 0.005835636984556913, + 0.006551662459969521, + 0.012020444497466087, + 0.011713854968547821, + 0.007688640616834164, + 0.006303605157881975, + 0.011030222289264202, + 0.008676734752953053, + 0.003059644717723131, + 0.0020396304316818714, + 0.0029478082433342934, + 0.0024837309028953314, + 0.0033671727869659662, + 0.002255439292639494, + 0.0032017999328672886, + 0.0034984119702130556, + 0.0036298835184425116, + 0.007843520492315292, + 0.003514892654493451, + 0.0035144472494721413, + 0.006676055490970612, + 0.003875895868986845, + 0.005957794841378927, + 0.002661533886566758, + 0.007115083746612072, + 0.0062329876236617565, + 0.005153171718120575, + 0.00812006089836359, + 0.010234242305159569, + 0.00776074780151248, + 0.007319030351936817, + 0.008651381358504295, + 0.005088456906378269, + 0.004121922887861729, + 0.007156361360102892, + 0.004576021805405617, + 0.0030031483620405197, + 0.008581960573792458, + 0.005141716916114092, + 0.006295389030128717, + 0.004264444578438997, + 0.003819715930148959, + 0.004069472663104534, + 0.003770416136831045, + 0.0037361704744398594, + 0.007719940971583128, + 0.004176028538495302, + 0.007499555591493845, + 0.0127077242359519, + 0.015023470856249332, + 0.06425350904464722, + 0.017431320622563362, + 0.01519125234335661, + 0.017469413578510284, + 0.014400326646864414, + 0.055691663175821304, + 0.03020559251308441, + 0.032744258642196655, + 0.030291829258203506, + 0.039542291313409805, + 0.03632880374789238, + 0.032497331500053406, + 0.03489283472299576, + 0.04447871819138527, + 0.02408308908343315, + 0.0234881192445755, + 0.013448758982121944, + 0.010297170840203762, + 0.011712154373526573, + 0.04283084720373154, + 0.009726383723318577, + 0.023424021899700165, + 0.00465928902849555, + 0.00557264918461442, + 0.009169629774987698, + 0.009726333431899548, + 0.008080492727458477, + 0.017942577600479126, + 0.009224151261150837, + 0.009508068673312664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.062058817595243454 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062031254172325134 + ], + "opposite_logprob_tokens": [ + -2.8120312690734863 + ], + "logprob_reward": -0.062031254172325134, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062031254172325134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09992823749780655, + 0.09399213641881943, + 0.11572566628456116, + 0.1095237284898758, + 0.0544792041182518, + 0.08771709352731705, + 0.0869968831539154, + 0.1161797046661377, + 0.03084683232009411, + 0.20461057126522064 + ], + "credit_map": [ + 0.008323820307850838, + 0.006826758850365877, + 0.02206696942448616, + 0.013354633003473282, + 0.010075305588543415, + 0.009799767285585403, + 0.011388293467462063, + 0.018092676997184753, + 0.012335115112364292, + 0.010672091506421566, + 0.0144965136423707, + 0.01392584852874279, + 0.00890586618334055, + 0.008742626756429672, + 0.011178623884916306, + 0.01373544242233038, + 0.011567110195755959, + 0.024265358224511147, + 0.007963800802826881, + 0.010042872279882431, + 0.021035289391875267, + 0.011576109565794468, + 0.018028022721409798, + 0.011247093789279461, + 0.017318513244390488, + 0.008973953314125538, + 0.009668543934822083, + 0.01655867137014866, + 0.010009764693677425, + 0.01443934254348278, + 0.013979168608784676, + 0.01857576332986355, + 0.004462170414626598, + 0.006258917041122913, + 0.007229706738144159, + 0.008396662771701813, + 0.005423370748758316, + 0.011328762397170067, + 0.005640959832817316, + 0.005738652776926756, + 0.008083297871053219, + 0.010694683529436588, + 0.007425899617373943, + 0.011433841660618782, + 0.010668794624507427, + 0.010465661995112896, + 0.015145545825362206, + 0.013799373060464859, + 0.005542043596506119, + 0.008068806491792202, + 0.013663118705153465, + 0.011350813321769238, + 0.007530326955020428, + 0.01013964507728815, + 0.008925371803343296, + 0.021776756271719933, + 0.012683432549238205, + 0.013832636177539825, + 0.010919895954430103, + 0.016234315931797028, + 0.017773261293768883, + 0.01260182075202465, + 0.015550839714705944, + 0.016583487391471863, + 0.00266129057854414, + 0.0052604516968131065, + 0.0037520236801356077, + 0.002141189994290471, + 0.0024890596978366375, + 0.008025312796235085, + 0.0017509699100628495, + 0.0047665308229625225, + 0.015772728249430656, + 0.01658402942121029, + 0.02859618328511715, + 0.023111939430236816, + 0.022079700604081154, + 0.033758409321308136, + 0.02500794641673565, + 0.03969961777329445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.062031254172325134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623625919222832 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623625919222832, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623625919222832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14577047526836395, + 0.09044003486633301, + 0.1687796264886856, + 0.13922205567359924, + 0.07660157233476639, + 0.07614784687757492, + 0.10247201472520828, + 0.05561237037181854, + 0.05520519241690636, + 0.0897488221526146 + ], + "credit_map": [ + 0.010333333164453506, + 0.009711244143545628, + 0.04198814556002617, + 0.017982278019189835, + 0.010928417555987835, + 0.019314510747790337, + 0.01587461680173874, + 0.019637934863567352, + 0.011209486052393913, + 0.00873861275613308, + 0.009674763306975365, + 0.008963719010353088, + 0.011692846193909645, + 0.008761397562921047, + 0.018331168219447136, + 0.013068041764199734, + 0.020907914265990257, + 0.034502871334552765, + 0.014091025106608868, + 0.015200565569102764, + 0.02588796801865101, + 0.019083727151155472, + 0.024224426597356796, + 0.014881135895848274, + 0.02175474911928177, + 0.010595890693366528, + 0.02176850102841854, + 0.016196539625525475, + 0.00991418119519949, + 0.018780043348670006, + 0.017233796417713165, + 0.02297835424542427, + 0.005840581841766834, + 0.005643058568239212, + 0.01016616728156805, + 0.010996627621352673, + 0.005183547735214233, + 0.02072446048259735, + 0.00845454353839159, + 0.009592593647539616, + 0.005711372010409832, + 0.009586866945028305, + 0.007765654940158129, + 0.013926930725574493, + 0.008756926283240318, + 0.009322919882833958, + 0.01222727820277214, + 0.008849901147186756, + 0.0067331064492464066, + 0.0071717542596161366, + 0.027146104723215103, + 0.008845302276313305, + 0.007926714606583118, + 0.008561125956475735, + 0.007138676941394806, + 0.028949221596121788, + 0.006582625210285187, + 0.008236370049417019, + 0.009479169733822346, + 0.007207990158349276, + 0.005694698542356491, + 0.005378887057304382, + 0.006610919255763292, + 0.006421709433197975, + 0.005638716276735067, + 0.008961094543337822, + 0.00522468239068985, + 0.004171143751591444, + 0.005057991947978735, + 0.015287036076188087, + 0.004460591357201338, + 0.006403936538845301, + 0.008636721409857273, + 0.00684754503890872, + 0.011352971196174622, + 0.012389592826366425, + 0.007571403868496418, + 0.016859255731105804, + 0.012290130369365215, + 0.013801198452711105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.007623625919222832 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead long shot, captured in Tim Burton's signature gothic fantasy style, shows a massive, distorted rocket slowly ascending from a crude launch pad constructed of rusted scaffolding and bizarre pipes. Rather than a smooth, streamlined form, the rocket's body takes on a grotesque, spiraling silhouette. Its deep black metal shell is studded with rivets of varying sizes and haphazard welds, like the work of a crazed blacksmith. As the camera slowly tilts upward, we see thick, inky black smoke billowing from the base of the rocket, rather than a bright flame. This plume stands out against the gloomy, deep purple sky, dotted with swirling clouds. The rocket climbs upwards with a heavy and laborious posture, trembling and shaking, casting a stretched, twisting, grotesque shadow on the barren and cracked ground. As the rocket rises, the shadow dances and deforms. The whole scene is filled with an absurd, grand and slightly uneasy industrial fairy tale atmosphere in desaturated cold colors.", + "seed": 32453, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000115/group_9/index_9507_seed_32453_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12660162281244994, + "vlm_logprob": -0.12660162281244994, + "vlm_yes_no_margin": 2.999999988079071, + "vlm_token_credit": -0.12660162281244994, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4025290608406067 + ], + "opposite_logprob_tokens": [ + -1.152529001235962 + ], + "logprob_reward": -0.4025290608406067, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.4025290608406067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042788684368133545, + 0.031110821291804314, + 0.016369948163628578, + 0.03212619945406914, + 0.08240670710802078, + 0.033572837710380554, + 0.08047719299793243, + 0.07845017313957214, + 0.09917369484901428, + 0.5035237073898315 + ], + "credit_map": [ + 0.002865814371034503, + 0.004336854442954063, + 0.008486730046570301, + 0.0035187567118555307, + 0.0058549270033836365, + 0.0055281734094023705, + 0.0036083669401705265, + 0.008589062839746475, + 0.003473196877166629, + 0.005567832384258509, + 0.0037266435101628304, + 0.0034712585620582104, + 0.0034427859354764223, + 0.004621211905032396, + 0.002986437175422907, + 0.0038214593660086393, + 0.002193233696743846, + 0.0020957321394234896, + 0.002230314537882805, + 0.0018759706290438771, + 0.0019759596325457096, + 0.001665269024670124, + 0.0013663372956216335, + 0.00296713225543499, + 0.004090421833097935, + 0.004988310392946005, + 0.0031490291003137827, + 0.005986662581562996, + 0.003605108242481947, + 0.0034737009555101395, + 0.00331517425365746, + 0.0035177955869585276, + 0.00762401707470417, + 0.012852538377046585, + 0.01551845483481884, + 0.01355875376611948, + 0.006446287967264652, + 0.006634971126914024, + 0.01039503887295723, + 0.009376651607453823, + 0.0058843521401286125, + 0.004004787188023329, + 0.005663220304995775, + 0.0025171416345983744, + 0.0038505601696670055, + 0.0039003239944577217, + 0.004677030723541975, + 0.0030754241161048412, + 0.010697740130126476, + 0.009420732967555523, + 0.014763090759515762, + 0.005944516975432634, + 0.008399664424359798, + 0.011619486846029758, + 0.006646733731031418, + 0.012985230423510075, + 0.007260615937411785, + 0.020097441971302032, + 0.006297105923295021, + 0.004153114277869463, + 0.009426996111869812, + 0.015348309651017189, + 0.0061507197096943855, + 0.009715874679386616, + 0.013144700787961483, + 0.01607682555913925, + 0.01573425903916359, + 0.010920929722487926, + 0.010221895761787891, + 0.012453211471438408, + 0.009321854449808598, + 0.011300025507807732, + 0.025054777041077614, + 0.17972984910011292, + 0.20934252440929413, + 0.0165955051779747, + 0.013016181997954845, + 0.018481630831956863, + 0.010978936217725277, + 0.030324308201670647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.4025290608406067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127068430185318 + ], + "opposite_logprob_tokens": [ + -2.127068519592285 + ], + "logprob_reward": -0.127068430185318, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127068430185318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0497884675860405, + 0.02867310307919979, + 0.03657141327857971, + 0.031073585152626038, + 0.05264875292778015, + 0.03288609907031059, + 0.14810103178024292, + 0.41657838225364685, + 0.05084576457738876, + 0.15283332765102386 + ], + "credit_map": [ + 0.004858893342316151, + 0.005391827318817377, + 0.004597941413521767, + 0.0075813899748027325, + 0.007451173849403858, + 0.0068184747360646725, + 0.0054084197618067265, + 0.007680350914597511, + 0.00507002929225564, + 0.0032787644304335117, + 0.004012644290924072, + 0.0022430233657360077, + 0.003356828121468425, + 0.004011703189462423, + 0.0028829581569880247, + 0.003817155258730054, + 0.003476976417005062, + 0.004788131918758154, + 0.003208027919754386, + 0.00538205960765481, + 0.005475725047290325, + 0.003652118844911456, + 0.0035224517341703176, + 0.007065922953188419, + 0.006499812938272953, + 0.003792783245444298, + 0.003386595519259572, + 0.0032889072317630053, + 0.0040365345776081085, + 0.003171459771692753, + 0.002279620384797454, + 0.004617875907570124, + 0.005016363691538572, + 0.01376334484666586, + 0.0066716019064188, + 0.010098946280777454, + 0.005054578650742769, + 0.003735779784619808, + 0.0032094542402774096, + 0.005098691210150719, + 0.004260981921106577, + 0.004552077502012253, + 0.007387956138700247, + 0.0027947702910751104, + 0.0027463010046631098, + 0.003323201322928071, + 0.004318013787269592, + 0.003502798266708851, + 0.01943363808095455, + 0.01971382647752762, + 0.01791805773973465, + 0.009260295890271664, + 0.015572745352983475, + 0.026772817596793175, + 0.01999855600297451, + 0.019431112334132195, + 0.05169980973005295, + 0.06965584307909012, + 0.04362623021006584, + 0.03667127341032028, + 0.06156342104077339, + 0.06893211603164673, + 0.04848223552107811, + 0.035947490483522415, + 0.008708510547876358, + 0.009248526766896248, + 0.009277399629354477, + 0.0050408681854605675, + 0.004614261444658041, + 0.005118584726005793, + 0.003364994190633297, + 0.005472627002745867, + 0.006959233433008194, + 0.05833422765135765, + 0.05931137874722481, + 0.006316938437521458, + 0.006766915787011385, + 0.004545892588794231, + 0.00540844677016139, + 0.0051903072744607925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.127068430185318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904796838760376 + ], + "opposite_logprob_tokens": [ + -2.464047908782959 + ], + "logprob_reward": -0.08904796838760376, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904796838760376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719747632741928, + 0.031332019716501236, + 0.06355858594179153, + 0.05215942859649658, + 0.05765290930867195, + 0.03202493116259575, + 0.10431589186191559, + 0.1030769869685173, + 0.11462202668190002, + 0.26928243041038513 + ], + "credit_map": [ + 0.016599705442786217, + 0.01802530512213707, + 0.026820262894034386, + 0.02622750960290432, + 0.017568770796060562, + 0.020561303943395615, + 0.021583644673228264, + 0.02458827570080757, + 0.0038061318919062614, + 0.003678929526358843, + 0.0032360642217099667, + 0.0034717516973614693, + 0.0032190256752073765, + 0.004678148776292801, + 0.004432095214724541, + 0.0048098755069077015, + 0.006354035343974829, + 0.008372952230274677, + 0.006950992159545422, + 0.009189664386212826, + 0.009233181364834309, + 0.0065579540096223354, + 0.005494464188814163, + 0.011405342258512974, + 0.0107444291934371, + 0.006991754285991192, + 0.0061594899743795395, + 0.0071612996980547905, + 0.005135411396622658, + 0.004998605698347092, + 0.004317023325711489, + 0.006651422940194607, + 0.006006581708788872, + 0.012167071923613548, + 0.00822844635695219, + 0.007954048924148083, + 0.006623667664825916, + 0.0046083214692771435, + 0.005451750010251999, + 0.006613022182136774, + 0.005108179058879614, + 0.004351374693214893, + 0.006579533684998751, + 0.0022909832186996937, + 0.003963204100728035, + 0.003567362204194069, + 0.002898912411183119, + 0.0032653824891895056, + 0.011975342407822609, + 0.012736192904412746, + 0.013771841302514076, + 0.007955922745168209, + 0.011529160663485527, + 0.015612154267728329, + 0.019659874960780144, + 0.01107540912926197, + 0.0189010389149189, + 0.01544856559485197, + 0.01589851826429367, + 0.00887583289295435, + 0.010603931732475758, + 0.009202546440064907, + 0.010707911103963852, + 0.013438650406897068, + 0.0193658247590065, + 0.0179620198905468, + 0.01776779815554619, + 0.010180453769862652, + 0.010142235085368156, + 0.014280658215284348, + 0.007671787869185209, + 0.017251260578632355, + 0.011600755155086517, + 0.11080129444599152, + 0.11099085211753845, + 0.008452859707176685, + 0.009429411962628365, + 0.0055467514321208, + 0.00632096640765667, + 0.006139557808637619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.08904796838760376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067365444265306 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.0067365444265306, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067365444265306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03572837635874748, + 0.021745888516306877, + 0.045312583446502686, + 0.04337134212255478, + 0.0992787778377533, + 0.057595111429691315, + 0.1560782492160797, + 0.11888615787029266, + 0.15086844563484192, + 0.2711350917816162 + ], + "credit_map": [ + 0.00398454163223505, + 0.003235162701457739, + 0.004617546685039997, + 0.004364237189292908, + 0.005273693241178989, + 0.004691043868660927, + 0.005083645228296518, + 0.0044785006903111935, + 0.003207965986803174, + 0.0023574181832373142, + 0.0026640198193490505, + 0.0020691000390797853, + 0.0021695445757359266, + 0.0032002278603613377, + 0.001941180438734591, + 0.004136429168283939, + 0.005116978194564581, + 0.005899081472307444, + 0.005072014406323433, + 0.006600118707865477, + 0.005692740902304649, + 0.004223514813929796, + 0.003950113896280527, + 0.008758014999330044, + 0.0078008780255913734, + 0.006146593485027552, + 0.00526244705542922, + 0.0061403438448905945, + 0.004429336171597242, + 0.004944725427776575, + 0.0039229500107467175, + 0.004724064376205206, + 0.007999135181307793, + 0.024504967033863068, + 0.012797529809176922, + 0.022823479026556015, + 0.00780761381611228, + 0.006145484745502472, + 0.00801592692732811, + 0.0091846389696002, + 0.009296958334743977, + 0.009044286794960499, + 0.011722897179424763, + 0.004733756184577942, + 0.005841795355081558, + 0.006129297893494368, + 0.005389340687543154, + 0.005436768755316734, + 0.02092072181403637, + 0.019400225952267647, + 0.022915957495570183, + 0.010802599601447582, + 0.014473271556198597, + 0.027614213526248932, + 0.019761160016059875, + 0.020190071314573288, + 0.01511236559599638, + 0.029032841324806213, + 0.015308406203985214, + 0.007483609952032566, + 0.012098405510187149, + 0.021867841482162476, + 0.008799087256193161, + 0.009183590300381184, + 0.019834095612168312, + 0.036275483667850494, + 0.023596739396452904, + 0.015134655870497227, + 0.012646470218896866, + 0.018006542697548866, + 0.009894117712974548, + 0.015480306930840015, + 0.023415522649884224, + 0.062074512243270874, + 0.09955549240112305, + 0.013342350721359253, + 0.01097120065242052, + 0.017106544226408005, + 0.010671778582036495, + 0.03399767726659775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.0067365444265306 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an overhead long shot perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rocket's body shaped as a grotesque, spiraling silhouette rather than a smooth streamlined form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rocket's shell deep black metal studded with rivets of varying sizes and haphazard welds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly tilt upward while the rocket ascends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky gloomy and deep purple with swirling clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076261102221906185 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.0076261102221906185, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076261102221906185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06296029686927795, + 0.0443740040063858, + 0.02982618287205696, + 0.042281340807676315, + 0.11995631456375122, + 0.05257781594991684, + 0.06272696703672409, + 0.10771944373846054, + 0.17789193987846375, + 0.29968568682670593 + ], + "credit_map": [ + 0.006816213019192219, + 0.01012538280338049, + 0.005517618730664253, + 0.007504184264689684, + 0.012610502541065216, + 0.0076447585597634315, + 0.006669828202575445, + 0.006071806885302067, + 0.007036574184894562, + 0.007011451758444309, + 0.0052861496806144714, + 0.005038733594119549, + 0.004311180207878351, + 0.005893928464502096, + 0.003559046657755971, + 0.0062369415536522865, + 0.003091050311923027, + 0.004543852061033249, + 0.0030990461818873882, + 0.004552049562335014, + 0.004659874830394983, + 0.0027908796910196543, + 0.002631649374961853, + 0.004457779694348574, + 0.006298427004367113, + 0.005852206144481897, + 0.004288135562092066, + 0.005428943317383528, + 0.006400649435818195, + 0.004599062725901604, + 0.0038018221966922283, + 0.005612095817923546, + 0.010637015104293823, + 0.03139745071530342, + 0.01430661790072918, + 0.025538839399814606, + 0.008051643148064613, + 0.008022928610444069, + 0.01026863418519497, + 0.01173318736255169, + 0.009084844030439854, + 0.007576351519674063, + 0.011051279492676258, + 0.003776267170906067, + 0.0050694639794528484, + 0.005815945565700531, + 0.005802503786981106, + 0.004401162266731262, + 0.008905226364731789, + 0.00870903767645359, + 0.00890109222382307, + 0.0037257240619510412, + 0.005732607562094927, + 0.011147872544825077, + 0.008056833408772945, + 0.007548574823886156, + 0.014738023281097412, + 0.015313925221562386, + 0.02047409862279892, + 0.008866028860211372, + 0.010855844244360924, + 0.01313987746834755, + 0.008415949530899525, + 0.015915703028440475, + 0.021463733166456223, + 0.03510419651865959, + 0.032619621604681015, + 0.014133174903690815, + 0.012439938262104988, + 0.026326710358262062, + 0.01274155918508768, + 0.02306300587952137, + 0.014610799960792065, + 0.10264129191637039, + 0.12077614665031433, + 0.011080191470682621, + 0.008357994258403778, + 0.017332840710878372, + 0.008004394359886646, + 0.01688201352953911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.0076261102221906185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_0/index_tensor([9926, 9926])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_0/index_tensor([9926, 9926])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fbde8738b7301ba94959e23abe9f05bbaad6f7ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_0/index_tensor([9926, 9926])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, mid-close-up shot from a horizontal angle, captured in a textured stop-motion style, captures a white-bearded grandfather sculpted entirely from clay. Fingerprints and traces of hand-molding are evident on his surface, and his lush, exaggerated white clay beard nearly covers half of his rounded cheeks. His face radiates pure, contented joy, his eyes narrowed into slits by a wide smile. He wears a handmade red clay sweater, its texture roughened with tooling, and sits on a simple, dark brown clay wooden chair. A soft, warm light slants down from the upper right of the frame, casting gentle, sharp shadows on his clay body and the chair, highlighting the texture of the miniature model. He holds a bright yellow clay corncob in his hands, its kernels distinct yet slightly irregular. With a pause and pause, typical of frame-by-frame animation, he clumsily and haltingly brings the corn to his mouth, then happily \"gnaws\" on it. His head sways slightly and discontinuously from side to side as he chews, his whole body trembling with happiness. The background is a solid, warm beige wall, and the overall scene is warm and bright, imbued with childlike innocence, warmth, and the simple joy of handicraft.", + "seed": 3387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_0/index_9926_seed_3387_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13041272200644016, + "vlm_logprob": -0.13041272200644016, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.13041272200644016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5265986323356628 + ], + "opposite_logprob_tokens": [ + -0.9015986323356628 + ], + "logprob_reward": -0.5265986323356628, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5265986323356628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23844264447689056, + 0.1209789365530014, + 0.07627464830875397, + 0.11834553629159927, + 0.07369284331798553, + 0.02420438639819622, + 0.11061225831508636, + 0.11662942171096802, + 0.03331118822097778, + 0.0875081717967987 + ], + "credit_map": [ + 0.038437969982624054, + 0.029388712719082832, + 0.02756158635020256, + 0.02834644913673401, + 0.038933273404836655, + 0.028065141290426254, + 0.023402871564030647, + 0.02430664747953415, + 0.01942763477563858, + 0.013836565427482128, + 0.01513576228171587, + 0.01517850998789072, + 0.015489770099520683, + 0.012405836954712868, + 0.015845200046896935, + 0.013659653253853321, + 0.013027955777943134, + 0.007699835579842329, + 0.011010130867362022, + 0.006839935667812824, + 0.015060633420944214, + 0.008037242107093334, + 0.008993702009320259, + 0.005605211015790701, + 0.027402067556977272, + 0.013870884664356709, + 0.03396838530898094, + 0.012296541593968868, + 0.008476534858345985, + 0.01053594145923853, + 0.006492787040770054, + 0.005302383564412594, + 0.011101835407316685, + 0.007237027864903212, + 0.008480795659124851, + 0.009937578812241554, + 0.010662995278835297, + 0.01084450725466013, + 0.009504389017820358, + 0.0059237172827124596, + 0.004104861058294773, + 0.002132331021130085, + 0.0022851950488984585, + 0.0033414149656891823, + 0.0028670765459537506, + 0.0033796268980950117, + 0.0038533739279955626, + 0.0022405076306313276, + 0.014907345175743103, + 0.013650836423039436, + 0.014917721040546894, + 0.008295854553580284, + 0.021404830738902092, + 0.01642821542918682, + 0.01251890603452921, + 0.008488547056913376, + 0.01795496791601181, + 0.012365742586553097, + 0.023486705496907234, + 0.014171578921377659, + 0.018723223358392715, + 0.009613469243049622, + 0.012421186082065105, + 0.007892557419836521, + 0.005729171447455883, + 0.00467434898018837, + 0.003528204280883074, + 0.0028287244495004416, + 0.005858121905475855, + 0.004528851248323917, + 0.003551824949681759, + 0.002611941657960415, + 0.014544741250574589, + 0.015599786303937435, + 0.010876563377678394, + 0.01104638073593378, + 0.009945687837898731, + 0.00861323717981577, + 0.008322724141180515, + 0.008559053763747215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.5265986323356628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812403604388237 + ], + "opposite_logprob_tokens": [ + -3.2881240844726562 + ], + "logprob_reward": -0.03812403604388237, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812403604388237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12431935966014862, + 0.06248351186513901, + 0.2172601968050003, + 0.08077074587345123, + 0.07114841789007187, + 0.0681343674659729, + 0.04220329225063324, + 0.19195674359798431, + 0.04825133830308914, + 0.09347201883792877 + ], + "credit_map": [ + 0.020974118262529373, + 0.015848873183131218, + 0.013803229667246342, + 0.011095542460680008, + 0.017864394932985306, + 0.0189377274364233, + 0.01148294284939766, + 0.014312544837594032, + 0.011657878756523132, + 0.005789810325950384, + 0.009922564029693604, + 0.006553732790052891, + 0.007958474569022655, + 0.006460004951804876, + 0.007349011488258839, + 0.00679204473271966, + 0.03792142868041992, + 0.033266324549913406, + 0.025127369910478592, + 0.022252816706895828, + 0.03159376233816147, + 0.02290729060769081, + 0.022910764440894127, + 0.02128046564757824, + 0.013781954534351826, + 0.010930354706943035, + 0.011852187104523182, + 0.009796654805541039, + 0.008904606103897095, + 0.008761979639530182, + 0.006547203753143549, + 0.010195819661021233, + 0.011948798783123493, + 0.007049700245261192, + 0.008298588916659355, + 0.008447385393083096, + 0.009730185382068157, + 0.010883917100727558, + 0.008395792916417122, + 0.00639405008405447, + 0.010279427282512188, + 0.007527465000748634, + 0.009628950618207455, + 0.007542837876826525, + 0.008622686378657818, + 0.006298545282334089, + 0.011594045907258987, + 0.006640422623604536, + 0.0066702282056212425, + 0.00522947870194912, + 0.0045498148538172245, + 0.004463864490389824, + 0.0046511124819517136, + 0.004310396034270525, + 0.00679096020758152, + 0.005537443328648806, + 0.0307985357940197, + 0.02211942709982395, + 0.015460285358130932, + 0.027875851839780807, + 0.03736370801925659, + 0.01998184435069561, + 0.02079838328063488, + 0.01755872368812561, + 0.008025073446333408, + 0.0055355834774672985, + 0.00551254115998745, + 0.005363646894693375, + 0.006667222362011671, + 0.008092316798865795, + 0.005262922961264849, + 0.003792033763602376, + 0.016351180151104927, + 0.01340043731033802, + 0.008354471996426582, + 0.011789137497544289, + 0.01263642217963934, + 0.01080478448420763, + 0.00979679636657238, + 0.010338801890611649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.03812403604388237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1097056120634079, + 0.11873709410429001, + 0.15977083146572113, + 0.08148007839918137, + 0.060257263481616974, + 0.09177234768867493, + 0.0792180523276329, + 0.15001128613948822, + 0.046743620187044144, + 0.1023038774728775 + ], + "credit_map": [ + 0.018950678408145905, + 0.015099805779755116, + 0.017299065366387367, + 0.015881387516856194, + 0.011667538434267044, + 0.011882736347615719, + 0.009184137918055058, + 0.009740252047777176, + 0.020414788275957108, + 0.013901122845709324, + 0.0190852340310812, + 0.014431623741984367, + 0.014056884683668613, + 0.011618884280323982, + 0.013948989100754261, + 0.011279548518359661, + 0.03187311068177223, + 0.023494377732276917, + 0.028307542204856873, + 0.017972923815250397, + 0.015654858201742172, + 0.017807308584451675, + 0.01283484697341919, + 0.011825836263597012, + 0.013663573190569878, + 0.010047330521047115, + 0.011326995678246021, + 0.010235413908958435, + 0.00870135985314846, + 0.012040695175528526, + 0.007961460389196873, + 0.007503237575292587, + 0.010673697106540203, + 0.009230304509401321, + 0.008686179295182228, + 0.008447650820016861, + 0.005643445532768965, + 0.005758094601333141, + 0.005346208345144987, + 0.006471677683293819, + 0.014447400346398354, + 0.012116771191358566, + 0.012305903248488903, + 0.011202570050954819, + 0.01292468886822462, + 0.00868216436356306, + 0.013193456456065178, + 0.006899379193782806, + 0.011408373713493347, + 0.008895850740373135, + 0.010034740902483463, + 0.006634548306465149, + 0.014803330413997173, + 0.012067632749676704, + 0.007980944588780403, + 0.0073926206678152084, + 0.02305389754474163, + 0.017164824530482292, + 0.03209131583571434, + 0.024884289130568504, + 0.014415464363992214, + 0.012534973211586475, + 0.016392935067415237, + 0.00947357527911663, + 0.008167844265699387, + 0.006051007658243179, + 0.005628960207104683, + 0.004685213789343834, + 0.008556384593248367, + 0.005368730518966913, + 0.004658761899918318, + 0.0036267144605517387, + 0.01744791306555271, + 0.01864279806613922, + 0.010698379948735237, + 0.013529905118048191, + 0.012868457473814487, + 0.009457483887672424, + 0.01015164889395237, + 0.009507277049124241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.014206582680344582 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495079979300499 + ], + "opposite_logprob_tokens": [ + -2.929950714111328 + ], + "logprob_reward": -0.05495079979300499, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495079979300499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07772916555404663, + 0.0625510960817337, + 0.06825318187475204, + 0.08640795201063156, + 0.083245649933815, + 0.05142577737569809, + 0.20385053753852844, + 0.10479191690683365, + 0.09243301302194595, + 0.1693117320537567 + ], + "credit_map": [ + 0.013729436323046684, + 0.007954850792884827, + 0.015197334811091423, + 0.013898846693336964, + 0.006591229233890772, + 0.0073690530844032764, + 0.0064105684868991375, + 0.006577842868864536, + 0.01058056391775608, + 0.006589317694306374, + 0.009403612464666367, + 0.006626708898693323, + 0.006042483262717724, + 0.0056410888209939, + 0.006730707827955484, + 0.010936606675386429, + 0.013175821863114834, + 0.012281781062483788, + 0.013827534392476082, + 0.008023274131119251, + 0.005765838548541069, + 0.006753673776984215, + 0.004988483153283596, + 0.003436767030507326, + 0.015459108166396618, + 0.009282806888222694, + 0.019936736673116684, + 0.009459658525884151, + 0.009319855831563473, + 0.007369592785835266, + 0.008054857142269611, + 0.007525323890149593, + 0.012952741235494614, + 0.010919137857854366, + 0.013234460726380348, + 0.01323851477354765, + 0.008414193987846375, + 0.008209110237658024, + 0.007385834585875273, + 0.00889164675027132, + 0.007982270792126656, + 0.0048973835073411465, + 0.007804210763424635, + 0.0070696501061320305, + 0.005985594820231199, + 0.006752727087587118, + 0.005386065691709518, + 0.005547873675823212, + 0.02796732448041439, + 0.026385081931948662, + 0.023733284324407578, + 0.018359452486038208, + 0.029689449816942215, + 0.023184286430478096, + 0.026877952739596367, + 0.027653681114315987, + 0.016882777214050293, + 0.014831919223070145, + 0.015853097662329674, + 0.013646796345710754, + 0.014957125298678875, + 0.010615634731948376, + 0.00975984800606966, + 0.008244710974395275, + 0.014752012677490711, + 0.014860089868307114, + 0.011938146315515041, + 0.011611164547502995, + 0.01210686843842268, + 0.010268738493323326, + 0.009113667532801628, + 0.0077823130413889885, + 0.025789687409996986, + 0.031180469319224358, + 0.016675157472491264, + 0.023195669054985046, + 0.017730670049786568, + 0.02389327809214592, + 0.01658703200519085, + 0.01425976399332285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.05495079979300499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818355917930603 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.01818355917930603, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818355917930603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1315702348947525, + 0.10715838521718979, + 0.060548342764377594, + 0.3139388859272003, + 0.13572447001934052, + 0.04988792911171913, + 0.09236535429954529, + 0.030545219779014587, + 0.02948620356619358, + 0.04877500608563423 + ], + "credit_map": [ + 0.02101157233119011, + 0.01611226238310337, + 0.019886212423443794, + 0.01685209758579731, + 0.01355673372745514, + 0.015504978597164154, + 0.011875871568918228, + 0.0167704988270998, + 0.018519693985581398, + 0.014362362213432789, + 0.016157248988747597, + 0.015770576894283295, + 0.01154295913875103, + 0.007893679663538933, + 0.01089287456125021, + 0.012018989771604538, + 0.011551175266504288, + 0.005397730506956577, + 0.014419974759221077, + 0.0078367805108428, + 0.006157076917588711, + 0.007646611891686916, + 0.0040664272382855415, + 0.0034725642763078213, + 0.07295007258653641, + 0.0402708426117897, + 0.09712032973766327, + 0.03561512380838394, + 0.019225135445594788, + 0.01866755448281765, + 0.015016447752714157, + 0.015073367394506931, + 0.02186519466340542, + 0.015319354832172394, + 0.03149018809199333, + 0.02498828060925007, + 0.010767142288386822, + 0.00989558920264244, + 0.0112155107781291, + 0.010183202102780342, + 0.007673370651900768, + 0.00478547066450119, + 0.004970958456397057, + 0.005856633186340332, + 0.004917089827358723, + 0.008376024663448334, + 0.00587622681632638, + 0.00743215624243021, + 0.012623876333236694, + 0.010609094984829426, + 0.014645654708147049, + 0.00877923984080553, + 0.016776124015450478, + 0.011507026851177216, + 0.009988883510231972, + 0.0074354493990540504, + 0.004531975369900465, + 0.003242155537009239, + 0.004704499151557684, + 0.004759825300425291, + 0.004967995919287205, + 0.0033456250093877316, + 0.0027295751497149467, + 0.002263569040223956, + 0.004755575209856033, + 0.002583988243713975, + 0.0026497477665543556, + 0.002183051547035575, + 0.0043898820877075195, + 0.0076911295764148235, + 0.002609682735055685, + 0.002623145468533039, + 0.008204462006688118, + 0.010697255842387676, + 0.0060403537936508656, + 0.006407400127500296, + 0.004993823356926441, + 0.0053048934787511826, + 0.003857232863083482, + 0.003269582986831665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.01818355917930603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-close-up shot from a horizontal angle, captured in a textured stop-motion style, captures a white-bearded grandfather sculpted entirely from clay. Fingerprints and traces of hand-molding are evident on his surface, and his lush, exaggerated white clay beard nearly covers half of his rounded cheeks. His face radiates pure, contented joy, his eyes narrowed into slits by a wide smile. He wears a handmade red clay sweater, its texture roughened with tooling, and sits on a simple, dark brown clay wooden chair. A soft, warm light slants down from the upper right of the frame, casting gentle, sharp shadows on his clay body and the chair, highlighting the texture of the miniature model. He holds a bright yellow clay corncob in his hands, its kernels distinct yet slightly irregular. With a pause and pause, typical of frame-by-frame animation, he clumsily and haltingly brings the corn to his mouth, then happily \"gnaws\" on it. His head sways slightly and discontinuously from side to side as he chews, his whole body trembling with happiness. The background is a solid, warm beige wall, and the overall scene is warm and bright, imbued with childlike innocence, warmth, and the simple joy of handicraft.", + "seed": 3388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_0/index_9926_seed_3388_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05437268186360598, + "vlm_logprob": -0.05437268186360598, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.05437268186360598, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20583981275558472 + ], + "opposite_logprob_tokens": [ + -1.7058398723602295 + ], + "logprob_reward": -0.20583981275558472, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20583981275558472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.079271599650383, + 0.09588604420423508, + 0.06851042062044144, + 0.05327196419239044, + 0.0364295169711113, + 0.022274300456047058, + 0.02867240644991398, + 0.030683865770697594, + 0.5235540866851807, + 0.061445750296115875 + ], + "credit_map": [ + 0.010641953907907009, + 0.008662821725010872, + 0.010571013204753399, + 0.007558657322078943, + 0.00800621509552002, + 0.01547328196465969, + 0.009984617121517658, + 0.008373039774596691, + 0.011107571423053741, + 0.011764589697122574, + 0.01360586192458868, + 0.008327428251504898, + 0.01727883331477642, + 0.010801587253808975, + 0.009842226281762123, + 0.013157946988940239, + 0.011090255342423916, + 0.005870453082025051, + 0.007833534851670265, + 0.009759695269167423, + 0.009038968943059444, + 0.008340410888195038, + 0.0070351180620491505, + 0.009541981853544712, + 0.007403682917356491, + 0.005060749594122171, + 0.008462228812277317, + 0.006210939958691597, + 0.008623515255749226, + 0.005591707769781351, + 0.006599918939173222, + 0.005319219082593918, + 0.008105089887976646, + 0.0030570465605705976, + 0.00538078136742115, + 0.004597531631588936, + 0.004108270164579153, + 0.0035343568306416273, + 0.004035488236695528, + 0.0036109525244683027, + 0.0033038125839084387, + 0.001948166755028069, + 0.003226023865863681, + 0.003191830823197961, + 0.0021568704396486282, + 0.002372798277065158, + 0.003140403190627694, + 0.0029343953356146812, + 0.002961777849122882, + 0.0022024984937161207, + 0.004253815859556198, + 0.0038176465313881636, + 0.0038539718370884657, + 0.004795133601874113, + 0.0041473787277936935, + 0.0026401837822049856, + 0.006251730490475893, + 0.002290942706167698, + 0.004443126264959574, + 0.004252995830029249, + 0.004567566327750683, + 0.0026294372510164976, + 0.002698129042983055, + 0.0035499376244843006, + 0.06923668086528778, + 0.05148809775710106, + 0.07025489211082458, + 0.06789784133434296, + 0.06575091928243637, + 0.08840630203485489, + 0.05640784278512001, + 0.05411152541637421, + 0.013920504599809647, + 0.0056676845997571945, + 0.009745941497385502, + 0.007111486047506332, + 0.006516166031360626, + 0.007211427204310894, + 0.005540890619158745, + 0.005731651559472084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.20583981275558472 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865340419113636 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.00865340419113636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865340419113636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12544842064380646, + 0.15444710850715637, + 0.0926792249083519, + 0.06419625878334045, + 0.06579294800758362, + 0.020048659294843674, + 0.017662160098552704, + 0.01663532853126526, + 0.22068732976913452, + 0.22240254282951355 + ], + "credit_map": [ + 0.0113257747143507, + 0.02017061971127987, + 0.01750790886580944, + 0.01087141316384077, + 0.016570528969168663, + 0.016641849651932716, + 0.01986885257065296, + 0.01249146368354559, + 0.01556006632745266, + 0.017803918570280075, + 0.021979613229632378, + 0.01653001829981804, + 0.02563166432082653, + 0.0155845507979393, + 0.018494518473744392, + 0.022862745448946953, + 0.010530461557209492, + 0.006876146886497736, + 0.010473269037902355, + 0.015463748015463352, + 0.009770222939550877, + 0.011699534021317959, + 0.013246092945337296, + 0.014619742520153522, + 0.004829909186810255, + 0.005820828955620527, + 0.009107121266424656, + 0.006334678735584021, + 0.014333710074424744, + 0.0068987091071903706, + 0.010143599472939968, + 0.006727691739797592, + 0.009659997187554836, + 0.007031351327896118, + 0.008982312865555286, + 0.006128834094852209, + 0.008737755939364433, + 0.008555327542126179, + 0.009182863868772984, + 0.0075145005248487, + 0.002370661124587059, + 0.0019603592809289694, + 0.0027995959389954805, + 0.002322128275409341, + 0.0023520675022155046, + 0.0026657339185476303, + 0.002637112746015191, + 0.0029409974813461304, + 0.002183094620704651, + 0.001314140623435378, + 0.002799788024276495, + 0.0018649388803169131, + 0.002845659153535962, + 0.002776126144453883, + 0.001674265251494944, + 0.0022041446063667536, + 0.002633546246215701, + 0.0014064451679587364, + 0.002402402926236391, + 0.001967300660908222, + 0.002759070135653019, + 0.001959659159183502, + 0.0018965969793498516, + 0.0016103066736832261, + 0.029305025935173035, + 0.02058688923716545, + 0.029586268588900566, + 0.0287080816924572, + 0.029601754620671272, + 0.03512502834200859, + 0.026092350482940674, + 0.02168191224336624, + 0.04556051269173622, + 0.021654367446899414, + 0.03479987755417824, + 0.020957188680768013, + 0.03379758074879646, + 0.023845894262194633, + 0.017199615016579628, + 0.024587482213974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.00865340419113636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548303231596947 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.012548303231596947, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548303231596947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14393210411071777, + 0.08152715861797333, + 0.1587512493133545, + 0.1044640764594078, + 0.07603508979082108, + 0.08540080487728119, + 0.0651288777589798, + 0.16164523363113403, + 0.05449923872947693, + 0.06861621141433716 + ], + "credit_map": [ + 0.017426902428269386, + 0.017152393236756325, + 0.022020680829882622, + 0.013802342116832733, + 0.01472262293100357, + 0.024359285831451416, + 0.018490610644221306, + 0.015957258641719818, + 0.008879200555384159, + 0.008720049634575844, + 0.011163935996592045, + 0.010876928456127644, + 0.010213203728199005, + 0.007719495799392462, + 0.009476319886744022, + 0.014478029683232307, + 0.017532138153910637, + 0.010247837752103806, + 0.01793937385082245, + 0.026434293016791344, + 0.016640830785036087, + 0.014054441824555397, + 0.03568476811051369, + 0.02021755836904049, + 0.020661823451519012, + 0.009145217947661877, + 0.013483998365700245, + 0.010864236392080784, + 0.014403780922293663, + 0.01043365616351366, + 0.012556890025734901, + 0.012914475984871387, + 0.017980379983782768, + 0.0068520354107022285, + 0.01000881101936102, + 0.007171812001615763, + 0.009464581497013569, + 0.009841623716056347, + 0.006255329120904207, + 0.0084605161100626, + 0.020270822569727898, + 0.006658358033746481, + 0.013814814388751984, + 0.00922753568738699, + 0.00891025923192501, + 0.008592776954174042, + 0.007526971399784088, + 0.010399271734058857, + 0.012343773618340492, + 0.0045496318489313126, + 0.010932422243058681, + 0.008138876408338547, + 0.007863775826990604, + 0.00818354357033968, + 0.007266918662935495, + 0.0058499411679804325, + 0.040762435644865036, + 0.015575969591736794, + 0.02999291941523552, + 0.02160712517797947, + 0.016078896820545197, + 0.015700049698352814, + 0.010399772785604, + 0.011528050526976585, + 0.007224991451948881, + 0.004563506692647934, + 0.00800938718020916, + 0.006355006247758865, + 0.010898319073021412, + 0.005062396638095379, + 0.008039799518883228, + 0.004345832858234644, + 0.016000308096408844, + 0.004588149022310972, + 0.010617062449455261, + 0.00723089836537838, + 0.007541273254901171, + 0.009237081743776798, + 0.006140580866485834, + 0.007260861340910196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.012548303231596947 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038089726120233536 + ], + "opposite_logprob_tokens": [ + -3.2880897521972656 + ], + "logprob_reward": -0.038089726120233536, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089726120233536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2671186029911041, + 0.0385754369199276, + 0.4021640717983246, + 0.04641585424542427, + 0.04012871906161308, + 0.029431208968162537, + 0.028783444315195084, + 0.035444699227809906, + 0.07294274866580963, + 0.03899521380662918 + ], + "credit_map": [ + 0.025034088641405106, + 0.038714464753866196, + 0.03732007369399071, + 0.0217155609279871, + 0.0418812595307827, + 0.03401155769824982, + 0.04119829088449478, + 0.027243319898843765, + 0.005709780380129814, + 0.004230339080095291, + 0.005322012584656477, + 0.0048658414743840694, + 0.005394194275140762, + 0.0041238111443817616, + 0.004969375673681498, + 0.003960082307457924, + 0.054889943450689316, + 0.031239161267876625, + 0.04460291936993599, + 0.06433231383562088, + 0.04634815827012062, + 0.04819899797439575, + 0.051145195960998535, + 0.061407383531332016, + 0.008646372705698013, + 0.005450887605547905, + 0.00612270925194025, + 0.005865957587957382, + 0.006311115808784962, + 0.005376068875193596, + 0.0039011426270008087, + 0.004741600714623928, + 0.005285551305860281, + 0.004817288834601641, + 0.005726090166717768, + 0.004745566286146641, + 0.0055024875327944756, + 0.005051919724792242, + 0.005244371481239796, + 0.003755448153242469, + 0.004851455334573984, + 0.002382307080551982, + 0.004368915688246489, + 0.0035715627018362284, + 0.0036686782259494066, + 0.0032873577438294888, + 0.003822443075478077, + 0.0034784905146807432, + 0.003694098675623536, + 0.002228548750281334, + 0.004620264749974012, + 0.003014611778780818, + 0.0036251309793442488, + 0.004493245389312506, + 0.0028587679844349623, + 0.004248779267072678, + 0.004352484829723835, + 0.0032438181806355715, + 0.004490189719945192, + 0.0051920292899012566, + 0.005707115866243839, + 0.004651508294045925, + 0.0036672421265393496, + 0.004140312317758799, + 0.013123455457389355, + 0.007155818864703178, + 0.01084111351519823, + 0.007863100618124008, + 0.009914439171552658, + 0.009464729577302933, + 0.007985508069396019, + 0.006594585254788399, + 0.006452775560319424, + 0.003691396676003933, + 0.006213248707354069, + 0.004947943612933159, + 0.005424097646027803, + 0.0041876789182424545, + 0.0028605435509234667, + 0.005217531230300665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.038089726120233536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732163019478321 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732163019478321, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732163019478321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22815090417861938, + 0.18652839958667755, + 0.10460777580738068, + 0.04626510292291641, + 0.11587989330291748, + 0.04194372147321701, + 0.05714552849531174, + 0.04460519924759865, + 0.07330536842346191, + 0.10156808793544769 + ], + "credit_map": [ + 0.025329995900392532, + 0.03414333611726761, + 0.032907743006944656, + 0.018701665103435516, + 0.029401585459709167, + 0.03687019646167755, + 0.027134617790579796, + 0.023661769926548004, + 0.03449569642543793, + 0.021110909059643745, + 0.02778772823512554, + 0.024454807862639427, + 0.020668916404247284, + 0.02603256143629551, + 0.016041966155171394, + 0.015935828909277916, + 0.017468014732003212, + 0.00692081730812788, + 0.01162672694772482, + 0.014475399628281593, + 0.010477086529135704, + 0.009846248663961887, + 0.019097160547971725, + 0.014696327969431877, + 0.0057116923853755, + 0.005292093846946955, + 0.006907409988343716, + 0.005029789637774229, + 0.009797563776373863, + 0.00577633548527956, + 0.004673490766435862, + 0.0030767328571528196, + 0.02281060628592968, + 0.010587669909000397, + 0.0177183598279953, + 0.012636062689125538, + 0.01374081987887621, + 0.010708184912800789, + 0.012472199276089668, + 0.01520600263029337, + 0.00890251062810421, + 0.004011715296655893, + 0.007333028130233288, + 0.004418498836457729, + 0.0045557147823274136, + 0.004219798371195793, + 0.004462823271751404, + 0.0040396335534751415, + 0.009320015087723732, + 0.004817818757146597, + 0.008972005918622017, + 0.005713781341910362, + 0.00785765890032053, + 0.006862873677164316, + 0.006199018098413944, + 0.00740236509591341, + 0.006198016460984945, + 0.004340402316302061, + 0.006489124149084091, + 0.006320313550531864, + 0.006603391841053963, + 0.005250690039247274, + 0.004475274123251438, + 0.004927992820739746, + 0.010503267869353294, + 0.005663817282766104, + 0.011560797691345215, + 0.008446470834314823, + 0.01259735506027937, + 0.006200636737048626, + 0.013162322342395782, + 0.005170708987861872, + 0.0258808471262455, + 0.009170331992208958, + 0.01667643152177334, + 0.010586628690361977, + 0.01030099205672741, + 0.011063873767852783, + 0.009307552129030228, + 0.008581439033150673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.006732163019478321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-close-up shot from a horizontal angle, captured in a textured stop-motion style, captures a white-bearded grandfather sculpted entirely from clay. Fingerprints and traces of hand-molding are evident on his surface, and his lush, exaggerated white clay beard nearly covers half of his rounded cheeks. His face radiates pure, contented joy, his eyes narrowed into slits by a wide smile. He wears a handmade red clay sweater, its texture roughened with tooling, and sits on a simple, dark brown clay wooden chair. A soft, warm light slants down from the upper right of the frame, casting gentle, sharp shadows on his clay body and the chair, highlighting the texture of the miniature model. He holds a bright yellow clay corncob in his hands, its kernels distinct yet slightly irregular. With a pause and pause, typical of frame-by-frame animation, he clumsily and haltingly brings the corn to his mouth, then happily \"gnaws\" on it. His head sways slightly and discontinuously from side to side as he chews, his whole body trembling with happiness. The background is a solid, warm beige wall, and the overall scene is warm and bright, imbued with childlike innocence, warmth, and the simple joy of handicraft.", + "seed": 3389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_0/index_9926_seed_3389_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10048029590398073, + "vlm_logprob": -0.10048029590398073, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.10048029590398073, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3904053270816803 + ], + "opposite_logprob_tokens": [ + -1.140405297279358 + ], + "logprob_reward": -0.3904053270816803, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3904053270816803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36917829513549805, + 0.03596697747707367, + 0.027199337258934975, + 0.05449119210243225, + 0.08627389371395111, + 0.019498048350214958, + 0.06900644302368164, + 0.0769389271736145, + 0.034034255892038345, + 0.22741256654262543 + ], + "credit_map": [ + 0.051872048527002335, + 0.038825131952762604, + 0.053088556975126266, + 0.04060190171003342, + 0.04881707951426506, + 0.04884589836001396, + 0.04502633213996887, + 0.042101357132196426, + 0.008198259398341179, + 0.0025528811383992434, + 0.006947685964405537, + 0.0030300249345600605, + 0.0034276768565177917, + 0.0028306564781814814, + 0.005421761889010668, + 0.0035580326803028584, + 0.004019237123429775, + 0.0030941965524107218, + 0.003885893849655986, + 0.003076578490436077, + 0.0036302225198596716, + 0.0032341547776013613, + 0.002815768588334322, + 0.0034432862885296345, + 0.007683318108320236, + 0.005406259559094906, + 0.0071622454561293125, + 0.005289389751851559, + 0.011009554378688335, + 0.005950660910457373, + 0.005990485195070505, + 0.005999275948852301, + 0.015156818553805351, + 0.013512875884771347, + 0.011911309324204922, + 0.00873189140111208, + 0.012293325737118721, + 0.009326022118330002, + 0.007824394851922989, + 0.0075172604992985725, + 0.0028108914848417044, + 0.0020634985994547606, + 0.0028598320204764605, + 0.0027360618114471436, + 0.002352530835196376, + 0.0019326541805639863, + 0.002183878794312477, + 0.0025586995761841536, + 0.014072523452341557, + 0.005193973425775766, + 0.010434228926897049, + 0.008380329236388206, + 0.009359595365822315, + 0.005767268128693104, + 0.008301778696477413, + 0.007496749050915241, + 0.014010302722454071, + 0.007152941543608904, + 0.014662419445812702, + 0.00831636879593134, + 0.012216736562550068, + 0.007021027151495218, + 0.007713376544415951, + 0.0058457497507333755, + 0.005305910017341375, + 0.004690240137279034, + 0.005537147633731365, + 0.0041773333214223385, + 0.0034221564419567585, + 0.00345384213142097, + 0.0052351090125739574, + 0.0022125146351754665, + 0.03789541870355606, + 0.02214084565639496, + 0.03468593955039978, + 0.0281258262693882, + 0.022226683795452118, + 0.024855300784111023, + 0.03849780559539795, + 0.018984749913215637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.3904053270816803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635374292731285 + ], + "opposite_logprob_tokens": [ + -3.6513538360595703 + ], + "logprob_reward": -0.02635374292731285, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635374292731285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32206159830093384, + 0.09344851970672607, + 0.040679216384887695, + 0.040284883230924606, + 0.14678965508937836, + 0.04845764487981796, + 0.07739516347646713, + 0.048802126199007034, + 0.04592662677168846, + 0.13615454733371735 + ], + "credit_map": [ + 0.06475944072008133, + 0.026811741292476654, + 0.054679110646247864, + 0.038973938673734665, + 0.03499346971511841, + 0.03498454764485359, + 0.029686681926250458, + 0.03717269003391266, + 0.0164189413189888, + 0.008475941605865955, + 0.01499001681804657, + 0.014603999443352222, + 0.010061496868729591, + 0.008479994721710682, + 0.007418469991534948, + 0.01299966499209404, + 0.005232238210737705, + 0.003814434865489602, + 0.005608853884041309, + 0.004605860449373722, + 0.0064431605860590935, + 0.004531696438789368, + 0.005825057625770569, + 0.004617918748408556, + 0.005777311976999044, + 0.004355785436928272, + 0.005332127679139376, + 0.0038158113602548838, + 0.006511202082037926, + 0.0057892766781151295, + 0.004776493646204472, + 0.003926876466721296, + 0.01945364661514759, + 0.024898892268538475, + 0.01716952584683895, + 0.01559523493051529, + 0.02283070608973503, + 0.01595854014158249, + 0.014545571990311146, + 0.016337547451257706, + 0.008046160452067852, + 0.005614043679088354, + 0.007190268021076918, + 0.00393324438482523, + 0.009714048355817795, + 0.004656414035707712, + 0.004574734251946211, + 0.004728736821562052, + 0.011403865180909634, + 0.006527738180011511, + 0.01029900461435318, + 0.00862534437328577, + 0.01164587028324604, + 0.00772037822753191, + 0.009651442058384418, + 0.011521527543663979, + 0.008239169605076313, + 0.006142305675894022, + 0.008692304603755474, + 0.005744475405663252, + 0.006547191645950079, + 0.003883717581629753, + 0.003906655590981245, + 0.005646311677992344, + 0.007196422200649977, + 0.005499706137925386, + 0.007481532637029886, + 0.005386418662965298, + 0.004783192649483681, + 0.004867417272180319, + 0.006401284597814083, + 0.004310660995543003, + 0.026843836531043053, + 0.011129074729979038, + 0.023073939606547356, + 0.01684214174747467, + 0.013012032955884933, + 0.012297515757381916, + 0.02061677910387516, + 0.01233924925327301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.02635374292731285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541946023702621 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541946023702621, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541946023702621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2312200963497162, + 0.15280580520629883, + 0.09150875359773636, + 0.05152038857340813, + 0.06655772030353546, + 0.06578329205513, + 0.06601845473051071, + 0.1150999665260315, + 0.04129437357187271, + 0.11819116026163101 + ], + "credit_map": [ + 0.047828201204538345, + 0.020073920488357544, + 0.03259250894188881, + 0.026353130117058754, + 0.027229422703385353, + 0.019203677773475647, + 0.02775689959526062, + 0.030182339251041412, + 0.025713425129652023, + 0.015592849813401699, + 0.022320114076137543, + 0.020656118169426918, + 0.025564653798937798, + 0.012394381687045097, + 0.015155111439526081, + 0.015409143641591072, + 0.01773506961762905, + 0.009945297613739967, + 0.012937475927174091, + 0.011480840854346752, + 0.01373661495745182, + 0.007202625274658203, + 0.008740589022636414, + 0.009730246849358082, + 0.008424287661910057, + 0.005719576962292194, + 0.007450946141034365, + 0.005381396971642971, + 0.009413613937795162, + 0.005449654534459114, + 0.004411350004374981, + 0.005269566085189581, + 0.01666870340704918, + 0.0073015447705984116, + 0.011488298885524273, + 0.0065470910631120205, + 0.006136578507721424, + 0.0048720454797148705, + 0.006421864032745361, + 0.007121589966118336, + 0.008128339424729347, + 0.00826910138130188, + 0.010270440950989723, + 0.005641113501042128, + 0.012984056957066059, + 0.006858190521597862, + 0.008220822550356388, + 0.005411224905401468, + 0.008506371639668941, + 0.005263867788016796, + 0.008968573063611984, + 0.00832996517419815, + 0.010528101585805416, + 0.005446247756481171, + 0.007593593094497919, + 0.011381737887859344, + 0.02698039449751377, + 0.009904484264552593, + 0.022926893085241318, + 0.018128391355276108, + 0.010514257475733757, + 0.009357616305351257, + 0.006703097373247147, + 0.010584830306470394, + 0.005833643022924662, + 0.005555875599384308, + 0.006003622896969318, + 0.0044238027185201645, + 0.006421535275876522, + 0.004282202105969191, + 0.005098944064229727, + 0.0036747471895068884, + 0.020125586539506912, + 0.00968702882528305, + 0.01906507834792137, + 0.012138023041188717, + 0.01189730130136013, + 0.010518690571188927, + 0.02489662542939186, + 0.009862832725048065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012541946023702621 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062033385038375854 + ], + "opposite_logprob_tokens": [ + -2.8120334148406982 + ], + "logprob_reward": -0.062033385038375854, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062033385038375854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18895739316940308, + 0.07192462682723999, + 0.06276099383831024, + 0.14308522641658783, + 0.08646777272224426, + 0.05153583362698555, + 0.14157164096832275, + 0.10134049504995346, + 0.08046618103981018, + 0.07188988476991653 + ], + "credit_map": [ + 0.048267364501953125, + 0.013505227863788605, + 0.034242648631334305, + 0.019161446020007133, + 0.012859020382165909, + 0.021115703508257866, + 0.022886255756020546, + 0.016919724643230438, + 0.011762394569814205, + 0.008898071013391018, + 0.011345201171934605, + 0.005418354645371437, + 0.013191032223403454, + 0.005976902320981026, + 0.008707917295396328, + 0.006624743342399597, + 0.012595897540450096, + 0.006802170537412167, + 0.008005380630493164, + 0.005725703202188015, + 0.008044436573982239, + 0.005918941460549831, + 0.0082035968080163, + 0.007464863825589418, + 0.017403561621904373, + 0.017623409628868103, + 0.01927204243838787, + 0.01230894960463047, + 0.029590731486678123, + 0.01867988519370556, + 0.015624991618096828, + 0.012581642717123032, + 0.022046400234103203, + 0.009585415944457054, + 0.016335761174559593, + 0.0072542703710496426, + 0.009404333308339119, + 0.008423333056271076, + 0.005631862673908472, + 0.007786390371620655, + 0.006861214060336351, + 0.005586610641330481, + 0.007235908415168524, + 0.005135006736963987, + 0.009380420669913292, + 0.005031564272940159, + 0.005988556891679764, + 0.006316549610346556, + 0.021258484572172165, + 0.011052978225052357, + 0.017868559807538986, + 0.02819676138460636, + 0.022692963480949402, + 0.010324077680706978, + 0.015363646671175957, + 0.014814167283475399, + 0.02070034295320511, + 0.010573036968708038, + 0.020587457343935966, + 0.011624793522059917, + 0.014894882217049599, + 0.006557848304510117, + 0.007162412162870169, + 0.00923970714211464, + 0.017452485859394073, + 0.008077342994511127, + 0.012728957459330559, + 0.010692656971514225, + 0.008594647981226444, + 0.007933548651635647, + 0.006850399076938629, + 0.008136136457324028, + 0.012181823141872883, + 0.007267470005899668, + 0.011914363130927086, + 0.008598131127655506, + 0.007827281951904297, + 0.007180139888077974, + 0.009311840869486332, + 0.007608833257108927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.062033385038375854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067078448832035 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011067078448832035, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067078448832035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43296509981155396, + 0.07822339981794357, + 0.03294858708977699, + 0.04409131035208702, + 0.14617162942886353, + 0.06098739802837372, + 0.08043044060468674, + 0.055039070546627045, + 0.03314446657896042, + 0.03599858656525612 + ], + "credit_map": [ + 0.10472063720226288, + 0.03647920861840248, + 0.07404256612062454, + 0.053223978728055954, + 0.0370466522872448, + 0.039052192121744156, + 0.03345292806625366, + 0.05494697391986847, + 0.014914074912667274, + 0.008005342446267605, + 0.012775581330060959, + 0.010498426854610443, + 0.007394429761916399, + 0.00695218937471509, + 0.008870474062860012, + 0.008812886662781239, + 0.006120374891906977, + 0.0030666159000247717, + 0.005815356969833374, + 0.002830595476552844, + 0.003690931014716625, + 0.0027802789118140936, + 0.00548511603847146, + 0.0031593230087310076, + 0.007879995740950108, + 0.0047208103351294994, + 0.006223625503480434, + 0.0051938914693892, + 0.005223585292696953, + 0.005025689490139484, + 0.003803979605436325, + 0.006019738037139177, + 0.039510101079940796, + 0.012824642471969128, + 0.0274793840944767, + 0.011218422092497349, + 0.012577908113598824, + 0.013296433724462986, + 0.012353677302598953, + 0.01691107079386711, + 0.008033724501729012, + 0.004769847262650728, + 0.007538951002061367, + 0.005701231770217419, + 0.008902247063815594, + 0.00540523836389184, + 0.014362482354044914, + 0.006273683160543442, + 0.02039766125380993, + 0.006015582475811243, + 0.014046811498701572, + 0.00736261997371912, + 0.010082113556563854, + 0.006513641215860844, + 0.008322316221892834, + 0.007689701393246651, + 0.010912170633673668, + 0.004539743531495333, + 0.008937542326748371, + 0.007697932422161102, + 0.007234790828078985, + 0.0037551952991634607, + 0.005044495686888695, + 0.006917208433151245, + 0.008059490472078323, + 0.003356514498591423, + 0.006109218113124371, + 0.0035187157336622477, + 0.003626400139182806, + 0.0027350964955985546, + 0.0032739981543272734, + 0.0024650348350405693, + 0.006753177382051945, + 0.0034914857242256403, + 0.005326112266629934, + 0.003712838049978018, + 0.005305828992277384, + 0.003328284714370966, + 0.004466087557375431, + 0.0036147735081613064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.011067078448832035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-close-up shot from a horizontal angle, captured in a textured stop-motion style, captures a white-bearded grandfather sculpted entirely from clay. Fingerprints and traces of hand-molding are evident on his surface, and his lush, exaggerated white clay beard nearly covers half of his rounded cheeks. His face radiates pure, contented joy, his eyes narrowed into slits by a wide smile. He wears a handmade red clay sweater, its texture roughened with tooling, and sits on a simple, dark brown clay wooden chair. A soft, warm light slants down from the upper right of the frame, casting gentle, sharp shadows on his clay body and the chair, highlighting the texture of the miniature model. He holds a bright yellow clay corncob in his hands, its kernels distinct yet slightly irregular. With a pause and pause, typical of frame-by-frame animation, he clumsily and haltingly brings the corn to his mouth, then happily \"gnaws\" on it. His head sways slightly and discontinuously from side to side as he chews, his whole body trembling with happiness. The background is a solid, warm beige wall, and the overall scene is warm and bright, imbued with childlike innocence, warmth, and the simple joy of handicraft.", + "seed": 3390, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_0/index_9926_seed_3390_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10706277526915073, + "vlm_logprob": -0.10706277526915073, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.10706277526915073, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39010000228881836 + ], + "opposite_logprob_tokens": [ + -1.1401000022888184 + ], + "logprob_reward": -0.39010000228881836, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.39010000228881836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16567938029766083, + 0.048076871782541275, + 0.13914501667022705, + 0.05365309119224548, + 0.07463844865560532, + 0.07463119179010391, + 0.05708323046565056, + 0.16106122732162476, + 0.1976921558380127, + 0.02833942510187626 + ], + "credit_map": [ + 0.027949342504143715, + 0.023051993921399117, + 0.020859548822045326, + 0.018924418836832047, + 0.027173103764653206, + 0.012263105250895023, + 0.020788375288248062, + 0.014669477008283138, + 0.007322933059185743, + 0.007916325703263283, + 0.00577964773401618, + 0.006046089809387922, + 0.0075636617839336395, + 0.004937773570418358, + 0.003831505309790373, + 0.004678929224610329, + 0.02081257477402687, + 0.013307154178619385, + 0.015067428350448608, + 0.01665225252509117, + 0.017333468422293663, + 0.017093483358621597, + 0.016182756051421165, + 0.022695893421769142, + 0.008451391942799091, + 0.009755668230354786, + 0.006694283336400986, + 0.007033403497189283, + 0.00624949811026454, + 0.005492634139955044, + 0.004058897960931063, + 0.005917313974350691, + 0.014054792933166027, + 0.006524210330098867, + 0.00963087659329176, + 0.007045374717563391, + 0.006417648401111364, + 0.013715732842683792, + 0.006623587105423212, + 0.010626222938299179, + 0.013783866539597511, + 0.007612981367856264, + 0.010991609655320644, + 0.007537228520959616, + 0.008271611295640469, + 0.009410081431269646, + 0.007671566680073738, + 0.009352244436740875, + 0.009459033608436584, + 0.00954513531178236, + 0.006205999292433262, + 0.005900857038795948, + 0.006492953281849623, + 0.006394514814019203, + 0.004391639027744532, + 0.008693099953234196, + 0.02488134242594242, + 0.021091915667057037, + 0.018832165747880936, + 0.01811131089925766, + 0.03065081499516964, + 0.017826126888394356, + 0.014000842347741127, + 0.01566670462489128, + 0.030537769198417664, + 0.03454473242163658, + 0.019955480471253395, + 0.01985538750886917, + 0.03584235534071922, + 0.019946947693824768, + 0.018707040697336197, + 0.018302440643310547, + 0.004526409320533276, + 0.003545833518728614, + 0.0041152420453727245, + 0.002923449734225869, + 0.0050075603649020195, + 0.002854227088391781, + 0.0023568738251924515, + 0.0030098280403763056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.39010000228881836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003571093082428 + ], + "opposite_logprob_tokens": [ + -2.695035696029663 + ], + "logprob_reward": -0.07003571093082428, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003571093082428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11253030598163605, + 0.039649590849876404, + 0.11946603655815125, + 0.06369810551404953, + 0.10645779222249985, + 0.07822868227958679, + 0.14143212139606476, + 0.09351396560668945, + 0.2026989459991455, + 0.04232441261410713 + ], + "credit_map": [ + 0.020629137754440308, + 0.011245373636484146, + 0.021586624905467033, + 0.011749116703867912, + 0.015921812504529953, + 0.00949561782181263, + 0.011645503342151642, + 0.01025712862610817, + 0.005838404875248671, + 0.006240206770598888, + 0.005155530758202076, + 0.005028350278735161, + 0.00539912236854434, + 0.003962766844779253, + 0.003367109689861536, + 0.004658104386180639, + 0.019130606204271317, + 0.012798432260751724, + 0.011061523109674454, + 0.01255208533257246, + 0.019798755645751953, + 0.00918787531554699, + 0.018662720918655396, + 0.01627405360341072, + 0.009717831388115883, + 0.009032453410327435, + 0.007345682475715876, + 0.006797079928219318, + 0.010070666670799255, + 0.0055755991488695145, + 0.006704916246235371, + 0.008453883230686188, + 0.01804168149828911, + 0.014955599792301655, + 0.011917129158973694, + 0.009859328158199787, + 0.010841487906873226, + 0.01739671267569065, + 0.009252509102225304, + 0.014193352311849594, + 0.014364504255354404, + 0.007487721741199493, + 0.010865808464586735, + 0.00831797905266285, + 0.008808028884232044, + 0.01196281984448433, + 0.007792211137712002, + 0.008629622869193554, + 0.022643089294433594, + 0.014896183274686337, + 0.015166182070970535, + 0.012008072808384895, + 0.017704572528600693, + 0.01830310933291912, + 0.014155944809317589, + 0.026554979383945465, + 0.014742700383067131, + 0.009119076654314995, + 0.012351884506642818, + 0.011216223239898682, + 0.013653602451086044, + 0.011098873801529408, + 0.009746489115059376, + 0.011585129424929619, + 0.03343845531344414, + 0.02496972121298313, + 0.022654615342617035, + 0.017989514395594597, + 0.04038262739777565, + 0.023661624640226364, + 0.01684340089559555, + 0.02275899611413479, + 0.006513727363198996, + 0.0055878362618386745, + 0.005369739141315222, + 0.003761525033041835, + 0.007271237671375275, + 0.005193956196308136, + 0.004634207114577293, + 0.003992187790572643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.07003571093082428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300768807530403 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023300768807530403, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300768807530403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2138017863035202, + 0.1389448493719101, + 0.055606432259082794, + 0.17312198877334595, + 0.04615716636180878, + 0.07018701732158661, + 0.07690860331058502, + 0.08019060641527176, + 0.09019312262535095, + 0.05488839000463486 + ], + "credit_map": [ + 0.04182931408286095, + 0.04822414368391037, + 0.037399597465991974, + 0.023032594472169876, + 0.017107754945755005, + 0.017883017659187317, + 0.013637673109769821, + 0.01468771044164896, + 0.023881783708930016, + 0.03685199096798897, + 0.025451622903347015, + 0.015174385160207748, + 0.00999389961361885, + 0.013841362670063972, + 0.005772547330707312, + 0.007977254688739777, + 0.008381699211895466, + 0.004964019171893597, + 0.0063872444443404675, + 0.007897785864770412, + 0.005490290466696024, + 0.0091404989361763, + 0.005038837902247906, + 0.008306053467094898, + 0.03322386369109154, + 0.026152990758419037, + 0.023840755224227905, + 0.016008179634809494, + 0.017207587137818336, + 0.022075479850172997, + 0.013733058236539364, + 0.020880073308944702, + 0.007596888113766909, + 0.005999173503369093, + 0.0070713479071855545, + 0.00667007640004158, + 0.0042693233117461205, + 0.006346984766423702, + 0.0031370134092867374, + 0.005066359881311655, + 0.01313280500471592, + 0.007583136670291424, + 0.011310137808322906, + 0.009808029979467392, + 0.007140651345252991, + 0.0051533435471355915, + 0.007690807804465294, + 0.008368110284209251, + 0.01156242098659277, + 0.010348374955356121, + 0.009686597622931004, + 0.008655368350446224, + 0.007335734087973833, + 0.010000490583479404, + 0.00753879314288497, + 0.01178082637488842, + 0.01186181791126728, + 0.01021778304129839, + 0.010209424421191216, + 0.008442864753305912, + 0.015478415414690971, + 0.008719190023839474, + 0.006878008600324392, + 0.008383100852370262, + 0.013308202847838402, + 0.014247721061110497, + 0.009245043620467186, + 0.009444053284823895, + 0.015059137716889381, + 0.009021741338074207, + 0.009320092387497425, + 0.010547135025262833, + 0.009510833770036697, + 0.008732031099498272, + 0.007747822906821966, + 0.005265349522233009, + 0.006258878391236067, + 0.006181071978062391, + 0.005725218914449215, + 0.005467177834361792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.023300768807530403 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693719655275345 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.033693719655275345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693719655275345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09810183942317963, + 0.11905620992183685, + 0.24234060943126678, + 0.05898992717266083, + 0.05773304030299187, + 0.11926639080047607, + 0.08643832057714462, + 0.11916236579418182, + 0.05904844030737877, + 0.03986288234591484 + ], + "credit_map": [ + 0.01845661737024784, + 0.014902553521096706, + 0.02276833914220333, + 0.01261352188885212, + 0.006743582896888256, + 0.007636619731783867, + 0.006606205366551876, + 0.008374393917620182, + 0.021589994430541992, + 0.020250720903277397, + 0.01947895810008049, + 0.011413132771849632, + 0.010731548070907593, + 0.01964176446199417, + 0.006832428742200136, + 0.009117655456066132, + 0.045491501688957214, + 0.017219532281160355, + 0.03651609644293785, + 0.03297964483499527, + 0.018386390060186386, + 0.046239398419857025, + 0.0183276180177927, + 0.027180392295122147, + 0.009299644269049168, + 0.011285687796771526, + 0.005740056280046701, + 0.005453614518046379, + 0.007368145976215601, + 0.007935365661978722, + 0.003764489432796836, + 0.008142923004925251, + 0.01127710286527872, + 0.005176302045583725, + 0.011609680950641632, + 0.006569263059645891, + 0.00506618432700634, + 0.006832852028310299, + 0.004037759732455015, + 0.007163890637457371, + 0.019192872568964958, + 0.011062199249863625, + 0.01710664853453636, + 0.019618522375822067, + 0.015005904249846935, + 0.0139227993786335, + 0.010388300754129887, + 0.012969134375452995, + 0.01245679147541523, + 0.008837458677589893, + 0.011426442302763462, + 0.008148434571921825, + 0.009672966785728931, + 0.012682071886956692, + 0.010788733139634132, + 0.01242541242390871, + 0.01805116981267929, + 0.015343518927693367, + 0.010852137580513954, + 0.014312224462628365, + 0.018283259123563766, + 0.020624810829758644, + 0.010055635124444962, + 0.011639595963060856, + 0.008425011299550533, + 0.00749759329482913, + 0.007194364443421364, + 0.0061130221001803875, + 0.009727060794830322, + 0.008499860763549805, + 0.006473593879491091, + 0.005117929074913263, + 0.007037668954581022, + 0.005050707142800093, + 0.005115064792335033, + 0.004315763246268034, + 0.007409319281578064, + 0.003196170087903738, + 0.004154527094215155, + 0.0035836549941450357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.033693719655275345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white-bearded grandfather sculpted entirely from clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the grandfather wearing a handmade red clay sweater with a roughened texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the grandfather sitting on a simple dark brown clay wooden chair while holding a bright yellow clay corncob?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform the action of clumsily bringing the corn to his mouth and gnawing on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a solid warm beige wall with soft warm light slanting from the upper right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183674663305283 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183674663305283, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183674663305283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1808820217847824, + 0.08915001153945923, + 0.12386126816272736, + 0.0947280153632164, + 0.04740021005272865, + 0.06419748812913895, + 0.05127834156155586, + 0.10942180454730988, + 0.0865267664194107, + 0.15255406498908997 + ], + "credit_map": [ + 0.03293512389063835, + 0.03325200453400612, + 0.023532705381512642, + 0.03171981871128082, + 0.014569370076060295, + 0.015509522520005703, + 0.009145303629338741, + 0.020218167454004288, + 0.015850959345698357, + 0.01041437964886427, + 0.013601499609649181, + 0.0072357854805886745, + 0.009465341456234455, + 0.02016746625304222, + 0.005803041160106659, + 0.006611532997339964, + 0.020270677283406258, + 0.014620010741055012, + 0.013964035548269749, + 0.016787089407444, + 0.016254132613539696, + 0.008491646498441696, + 0.019633499905467033, + 0.013840170577168465, + 0.016627399250864983, + 0.015098958276212215, + 0.013880058191716671, + 0.009825143031775951, + 0.011112012900412083, + 0.010123424232006073, + 0.010237781330943108, + 0.007823234423995018, + 0.008803926408290863, + 0.006385782267898321, + 0.007043664809316397, + 0.005203121807426214, + 0.004921075887978077, + 0.005113653838634491, + 0.003950745332986116, + 0.00597823690623045, + 0.011023188009858131, + 0.0052038757130503654, + 0.010321840643882751, + 0.009644163772463799, + 0.009043165482580662, + 0.005629874300211668, + 0.005560123827308416, + 0.007771250791847706, + 0.007717607542872429, + 0.006640969309955835, + 0.004767241422086954, + 0.0057294173166155815, + 0.007523388601839542, + 0.0071014706045389175, + 0.00569975795224309, + 0.006098486483097076, + 0.017705386504530907, + 0.012509624473750591, + 0.011824219487607479, + 0.011531326919794083, + 0.019060730934143066, + 0.01528290007263422, + 0.009186073206365108, + 0.012321542017161846, + 0.01245349831879139, + 0.014869887381792068, + 0.009079725481569767, + 0.008934641256928444, + 0.01113435160368681, + 0.010571116581559181, + 0.010294592939317226, + 0.009188944473862648, + 0.028669457882642746, + 0.014583433978259563, + 0.03121759183704853, + 0.016514113172888756, + 0.014577156864106655, + 0.016761019825935364, + 0.011702830903232098, + 0.018528450280427933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.018183674663305283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_tensor([4986, 4986])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_tensor([4986, 4986])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b69b1b8238ce9cc11ee8e633436e2e050d95ed2f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_tensor([4986, 4986])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, low-angle tracking shot follows a poet dressed in Tang Dynasty attire as he journeys through a vast, mist-shrouded bamboo forest. The poet's face is thin and pensive, his long, dark hair held in a bun with a simple wooden hairpin. He wears an indigo silk round-necked robe, its wide sleeves swaying gently with his leisurely steps. His left hand gently grasps a scroll of yellowed bamboo slips, while his right hand occasionally strokes the emerald green bamboo stalks beside him, dappled with morning dew. The key light source comes from above, with beams of golden sunlight penetrating the dense bamboo canopy and milky mist, creating a distinct Tyndall effect. The poet and the ground, carpeted with wet fallen leaves, are dappled with shifting, interplay of light and dark. In the background, the overlapping, varying shades of bamboo shadows, looming through the mist, create a sense of infinite depth. The entire picture is dominated by the cool green of the bamboo forest and the soft white of the mist, illuminated by the warm sunlight, presenting a tranquil, ethereal and cinematic atmosphere full of oriental classical aesthetics.", + "seed": 21727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_4986_seed_21727_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06333756847307086, + "vlm_logprob": -0.06333756847307086, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.06333756847307086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09168204665184021 + ], + "opposite_logprob_tokens": [ + -2.841681957244873 + ], + "logprob_reward": -0.09168204665184021, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09168204665184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1666063815355301, + 0.09426040947437286, + 0.188412144780159, + 0.08443625271320343, + 0.07133615016937256, + 0.03830316662788391, + 0.08515291661024094, + 0.022763697430491447, + 0.13686910271644592, + 0.11185980588197708 + ], + "credit_map": [ + 0.023281730711460114, + 0.021815931424498558, + 0.024935968220233917, + 0.02153169736266136, + 0.01785639300942421, + 0.014422302134335041, + 0.025667713955044746, + 0.017094654962420464, + 0.0111421262845397, + 0.01655626855790615, + 0.01584271714091301, + 0.007981697097420692, + 0.010645551607012749, + 0.010274622589349747, + 0.010058037005364895, + 0.011759387329220772, + 0.015773793682456017, + 0.036172498017549515, + 0.019836846739053726, + 0.0305266622453928, + 0.02198575995862484, + 0.021333759650588036, + 0.026467429473996162, + 0.016315393149852753, + 0.014132804237306118, + 0.009030824527144432, + 0.009343242272734642, + 0.006853118073195219, + 0.00869725365191698, + 0.008760723285377026, + 0.00912354327738285, + 0.0184947457164526, + 0.011576288379728794, + 0.004998843185603619, + 0.009831386618316174, + 0.0062810867093503475, + 0.0042286948300898075, + 0.009600207209587097, + 0.006356905680149794, + 0.018462728708982468, + 0.006135527044534683, + 0.0030847846064716578, + 0.0059284633025527, + 0.004382580518722534, + 0.004467042628675699, + 0.004616623278707266, + 0.005449384450912476, + 0.004238762427121401, + 0.008616359904408455, + 0.008049988187849522, + 0.011022104881703854, + 0.006615718826651573, + 0.008340329863131046, + 0.011957368813455105, + 0.0075100986286997795, + 0.02304094471037388, + 0.002138014417141676, + 0.00348889478482306, + 0.002827288582921028, + 0.0027210272382944822, + 0.0033408813178539276, + 0.002844689879566431, + 0.003103147726505995, + 0.0022997527848929167, + 0.016311340034008026, + 0.024276355281472206, + 0.01739034615457058, + 0.015066890977323055, + 0.013922435231506824, + 0.010608297772705555, + 0.014859575778245926, + 0.024433860555291176, + 0.012663384899497032, + 0.03189299255609512, + 0.010928754694759846, + 0.00785814505070448, + 0.007480702828615904, + 0.01156007219105959, + 0.005273660644888878, + 0.024202095344662666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.09168204665184021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209466606378555 + ], + "opposite_logprob_tokens": [ + -2.8120946884155273 + ], + "logprob_reward": -0.06209466606378555, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209466606378555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10471697151660919, + 0.12255964428186417, + 0.12287013232707977, + 0.09011378884315491, + 0.1908237189054489, + 0.04634285718202591, + 0.09526506066322327, + 0.06263984739780426, + 0.0717836394906044, + 0.09288439154624939 + ], + "credit_map": [ + 0.01659989170730114, + 0.011767425574362278, + 0.013062390498816967, + 0.013578144833445549, + 0.010520407930016518, + 0.013097483664751053, + 0.01092198584228754, + 0.015169234946370125, + 0.016123071312904358, + 0.0156116196885705, + 0.016996417194604874, + 0.00933422427624464, + 0.01852240599691868, + 0.020306382328271866, + 0.016878308728337288, + 0.008787213824689388, + 0.010546698234975338, + 0.01841595210134983, + 0.0138700595125556, + 0.02341412752866745, + 0.01902972348034382, + 0.012545357458293438, + 0.015395322814583778, + 0.009652886539697647, + 0.024149909615516663, + 0.00966357346624136, + 0.009601139463484287, + 0.005971400532871485, + 0.009627248160541058, + 0.009368124417960644, + 0.009764025919139385, + 0.01196836307644844, + 0.025986114516854286, + 0.014454352669417858, + 0.02763393707573414, + 0.01542087085545063, + 0.015309317037463188, + 0.024089375510811806, + 0.019892608746886253, + 0.04803714528679848, + 0.007633519358932972, + 0.003870446002110839, + 0.006229551509022713, + 0.0053731161169707775, + 0.004356664139777422, + 0.005984073039144278, + 0.005281167570501566, + 0.007614318747073412, + 0.008392330259084702, + 0.010080019012093544, + 0.010757162235677242, + 0.009730786085128784, + 0.0080374451354146, + 0.013823608867824078, + 0.008767038583755493, + 0.0256766639649868, + 0.008334645070135593, + 0.0078107379376888275, + 0.012179026380181313, + 0.005495134741067886, + 0.006962914019823074, + 0.00835633184760809, + 0.007068360224366188, + 0.006432693917304277, + 0.007577098906040192, + 0.015232428908348083, + 0.007340533193200827, + 0.008927296847105026, + 0.007703704759478569, + 0.006078239530324936, + 0.008859194815158844, + 0.010065142996609211, + 0.006829245947301388, + 0.015445208176970482, + 0.006587670184671879, + 0.010371769778430462, + 0.014348114840686321, + 0.009125296957790852, + 0.012698880396783352, + 0.017478207126259804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.06209466606378555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496840178966522 + ], + "opposite_logprob_tokens": [ + -2.9299683570861816 + ], + "logprob_reward": -0.05496840178966522, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496840178966522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09245621412992477, + 0.096498042345047, + 0.0973808690905571, + 0.09333976358175278, + 0.16618692874908447, + 0.06165881082415581, + 0.133184015750885, + 0.07247480750083923, + 0.032662998884916306, + 0.15415753424167633 + ], + "credit_map": [ + 0.016111938282847404, + 0.009974176995456219, + 0.013232040219008923, + 0.009101581759750843, + 0.008908042684197426, + 0.008202697150409222, + 0.011756557039916515, + 0.015169176273047924, + 0.011913622729480267, + 0.012902816757559776, + 0.016520299017429352, + 0.008218296803534031, + 0.013054790906608105, + 0.016115279868245125, + 0.009681614115834236, + 0.00809132494032383, + 0.01141135673969984, + 0.012095192447304726, + 0.010439539328217506, + 0.015590837225317955, + 0.015634845942258835, + 0.010211251676082611, + 0.01214019674807787, + 0.009857650846242905, + 0.026981420814990997, + 0.009174935519695282, + 0.010551883839070797, + 0.006738225929439068, + 0.01016377191990614, + 0.00910574197769165, + 0.008598999120295048, + 0.012024780735373497, + 0.021767381578683853, + 0.012818212620913982, + 0.022978849709033966, + 0.014559089206159115, + 0.013881942257285118, + 0.020744564011693, + 0.016307421028614044, + 0.04312945902347565, + 0.008673550561070442, + 0.005527863744646311, + 0.008504698984324932, + 0.0075969090685248375, + 0.00656112190335989, + 0.008513587526977062, + 0.007566146552562714, + 0.00871493760496378, + 0.022221269086003304, + 0.01502107735723257, + 0.015786731615662575, + 0.006573863793164492, + 0.008393554948270321, + 0.01823810487985611, + 0.010415920987725258, + 0.03653348982334137, + 0.0103440061211586, + 0.0175425186753273, + 0.00789256114512682, + 0.007160151842981577, + 0.007320033386349678, + 0.010036081075668335, + 0.0052186367101967335, + 0.006960817147046328, + 0.0035659612622112036, + 0.007303289603441954, + 0.003916420973837376, + 0.0033043401781469584, + 0.00344885908998549, + 0.002960309386253357, + 0.0027765841223299503, + 0.0053872340358793736, + 0.019125472754240036, + 0.05176576226949692, + 0.006486900616437197, + 0.0066607557237148285, + 0.015236727893352509, + 0.011076412163674831, + 0.009543200023472309, + 0.03426230326294899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.05496840178966522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14419855177402496, + 0.055974796414375305, + 0.1485595703125, + 0.08473387360572815, + 0.10473725944757462, + 0.056789908558130264, + 0.07922433316707611, + 0.0559711828827858, + 0.09051039814949036, + 0.17930008471012115 + ], + "credit_map": [ + 0.023922178894281387, + 0.014156393706798553, + 0.027553711086511612, + 0.0126192020252347, + 0.01389271393418312, + 0.01821039617061615, + 0.015815677121281624, + 0.01802828349173069, + 0.006948943715542555, + 0.006822496186941862, + 0.007201254367828369, + 0.005676457192748785, + 0.0050615048967301846, + 0.008704771287739277, + 0.0074740322306752205, + 0.008085331879556179, + 0.015489687211811543, + 0.025649920105934143, + 0.015950296074151993, + 0.022510500624775887, + 0.01590971276164055, + 0.016908826306462288, + 0.020499160513281822, + 0.015641478821635246, + 0.024880021810531616, + 0.0095674442127347, + 0.010175736621022224, + 0.004552964586764574, + 0.00633628573268652, + 0.010518886148929596, + 0.004519464913755655, + 0.014183072373270988, + 0.009790830314159393, + 0.011270926333963871, + 0.014237172901630402, + 0.010347716510295868, + 0.010886440984904766, + 0.021802254021167755, + 0.010077555663883686, + 0.016324358060956, + 0.008198532275855541, + 0.0051726531237363815, + 0.0078033641912043095, + 0.004358363337814808, + 0.004044756758958101, + 0.009543278254568577, + 0.005812456365674734, + 0.01185650285333395, + 0.008109314367175102, + 0.00787444319576025, + 0.010825623758137226, + 0.008495265617966652, + 0.007217098493129015, + 0.011374951340258121, + 0.007640118710696697, + 0.017687518149614334, + 0.005870835389941931, + 0.006028949748724699, + 0.011364849284291267, + 0.00480298139154911, + 0.004352213349193335, + 0.010755150578916073, + 0.00560608459636569, + 0.007190123200416565, + 0.009152835234999657, + 0.02096227928996086, + 0.011322296224534512, + 0.009515756741166115, + 0.00747071485966444, + 0.01353777851909399, + 0.0078037334606051445, + 0.010745004750788212, + 0.022004391998052597, + 0.05513947084546089, + 0.013795153237879276, + 0.011258863843977451, + 0.011323921382427216, + 0.020591164007782936, + 0.00878656655550003, + 0.036400552839040756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.007631433662027121 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003112941980362 + ], + "opposite_logprob_tokens": [ + -2.350311279296875 + ], + "logprob_reward": -0.1003112941980362, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003112941980362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11738470941781998, + 0.04413332790136337, + 0.1196916252374649, + 0.04752441495656967, + 0.02963639609515667, + 0.06187557056546211, + 0.11293800920248032, + 0.06776200979948044, + 0.06527732312679291, + 0.333776593208313 + ], + "credit_map": [ + 0.015006466768682003, + 0.015018672682344913, + 0.01807796023786068, + 0.012358958832919598, + 0.014482114464044571, + 0.012044742703437805, + 0.012359363958239555, + 0.01803642511367798, + 0.006977352313697338, + 0.007111205253750086, + 0.005450056400150061, + 0.0052245366387069225, + 0.005234152544289827, + 0.0039728484116494656, + 0.0037616544868797064, + 0.006401521619409323, + 0.01656324230134487, + 0.01956149935722351, + 0.013996616005897522, + 0.011793014593422413, + 0.015190313570201397, + 0.012923719361424446, + 0.014799526892602444, + 0.014863699674606323, + 0.0115495091304183, + 0.005660795606672764, + 0.004886866547167301, + 0.0032118731178343296, + 0.0071086762472987175, + 0.00490907346829772, + 0.0037375744432210922, + 0.0064600445330142975, + 0.0030953192617744207, + 0.002699183998629451, + 0.007387000601738691, + 0.002699175849556923, + 0.0025789944920688868, + 0.006343714892864227, + 0.0024489678908139467, + 0.0023840400390326977, + 0.008429708890616894, + 0.0065887076780200005, + 0.008320095017552376, + 0.00736960768699646, + 0.004545609932392836, + 0.006759510841220617, + 0.011417421512305737, + 0.008444909006357193, + 0.016715675592422485, + 0.01024975161999464, + 0.012666133232414722, + 0.008327674120664597, + 0.009737889282405376, + 0.016683991998434067, + 0.00901417899876833, + 0.029542721807956696, + 0.010308273136615753, + 0.012431715615093708, + 0.00978888850659132, + 0.0057455128990113735, + 0.005859339144080877, + 0.010388481430709362, + 0.00402451679110527, + 0.009215285070240498, + 0.007752107921987772, + 0.010691902600228786, + 0.009337430819869041, + 0.005879038944840431, + 0.00591711001470685, + 0.0054641179740428925, + 0.005818528588861227, + 0.014417083002626896, + 0.06241687014698982, + 0.11953260004520416, + 0.019490739330649376, + 0.01315913163125515, + 0.028847025707364082, + 0.03195692226290703, + 0.011672377586364746, + 0.04670093581080437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.1003112941980362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle tracking shot follows a poet dressed in Tang Dynasty attire as he journeys through a vast, mist-shrouded bamboo forest. The poet's face is thin and pensive, his long, dark hair held in a bun with a simple wooden hairpin. He wears an indigo silk round-necked robe, its wide sleeves swaying gently with his leisurely steps. His left hand gently grasps a scroll of yellowed bamboo slips, while his right hand occasionally strokes the emerald green bamboo stalks beside him, dappled with morning dew. The key light source comes from above, with beams of golden sunlight penetrating the dense bamboo canopy and milky mist, creating a distinct Tyndall effect. The poet and the ground, carpeted with wet fallen leaves, are dappled with shifting, interplay of light and dark. In the background, the overlapping, varying shades of bamboo shadows, looming through the mist, create a sense of infinite depth. The entire picture is dominated by the cool green of the bamboo forest and the soft white of the mist, illuminated by the warm sunlight, presenting a tranquil, ethereal and cinematic atmosphere full of oriental classical aesthetics.", + "seed": 21728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_4986_seed_21728_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07168984301388263, + "vlm_logprob": -0.07168984301388263, + "vlm_yes_no_margin": 3.024999976158142, + "vlm_token_credit": -0.07168984301388263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16375885903835297 + ], + "opposite_logprob_tokens": [ + -2.1637587547302246 + ], + "logprob_reward": -0.16375885903835297, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.16375885903835297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27858787775039673, + 0.0896451324224472, + 0.06180352345108986, + 0.04186292737722397, + 0.07007578015327454, + 0.03958626091480255, + 0.18372558057308197, + 0.12830767035484314, + 0.08134076744318008, + 0.025064406916499138 + ], + "credit_map": [ + 0.024590108543634415, + 0.030307898297905922, + 0.021422628313302994, + 0.02131490223109722, + 0.051659923046827316, + 0.044636692851781845, + 0.02420618385076523, + 0.060449570417404175, + 0.013924302533268929, + 0.009091051295399666, + 0.010774577967822552, + 0.009822061285376549, + 0.013731420040130615, + 0.010975277051329613, + 0.011292830109596252, + 0.010033623315393925, + 0.008031458593904972, + 0.0072830477729439735, + 0.010769550688564777, + 0.006222188472747803, + 0.006792858708649874, + 0.007543932180851698, + 0.005701767280697823, + 0.009458727203309536, + 0.006564561743289232, + 0.007593090180307627, + 0.005663167219609022, + 0.0046541569754481316, + 0.0039052581414580345, + 0.006319036707282066, + 0.0031001302413642406, + 0.004063533619046211, + 0.011795049533247948, + 0.01056423969566822, + 0.009803607128560543, + 0.00739227095618844, + 0.007567336782813072, + 0.005568792577832937, + 0.009579768404364586, + 0.007804724387824535, + 0.006694694980978966, + 0.003843340091407299, + 0.005390065256506205, + 0.0044327471405267715, + 0.005472210235893726, + 0.0030756036285310984, + 0.005550846923142672, + 0.005126758012920618, + 0.049957070499658585, + 0.013080389238893986, + 0.04763351008296013, + 0.022294528782367706, + 0.008856558240950108, + 0.016382083296775818, + 0.00998660083860159, + 0.015534848906099796, + 0.008929213508963585, + 0.012259234674274921, + 0.031782325357198715, + 0.006531435530632734, + 0.004893044475466013, + 0.02223164588212967, + 0.003854940878227353, + 0.03782584145665169, + 0.013802397064864635, + 0.009647886268794537, + 0.011136393994092941, + 0.005453975405544043, + 0.0068848636001348495, + 0.01634170487523079, + 0.008599217049777508, + 0.009474337100982666, + 0.002674736548215151, + 0.0022456436417996883, + 0.0029961091931909323, + 0.005948767066001892, + 0.0025938707403838634, + 0.0025960979983210564, + 0.0019022325286641717, + 0.004106953274458647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.16375885903835297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499108135700226 + ], + "opposite_logprob_tokens": [ + -2.9299910068511963 + ], + "logprob_reward": -0.05499108135700226, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499108135700226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12990021705627441, + 0.06744115799665451, + 0.14101272821426392, + 0.04882507026195526, + 0.23496097326278687, + 0.05450696498155594, + 0.039213452488183975, + 0.07892683148384094, + 0.16227447986602783, + 0.042938172817230225 + ], + "credit_map": [ + 0.0160044077783823, + 0.01531648077070713, + 0.015274399891495705, + 0.015507849864661694, + 0.01919851452112198, + 0.01717500388622284, + 0.009953543543815613, + 0.02147001028060913, + 0.008024556562304497, + 0.00618730578571558, + 0.00798010267317295, + 0.009771701879799366, + 0.010271145962178707, + 0.009404454380273819, + 0.009559056721627712, + 0.006242829840630293, + 0.013514098711311817, + 0.017654791474342346, + 0.013861844316124916, + 0.01116290595382452, + 0.027411412447690964, + 0.019481847062706947, + 0.026873042806982994, + 0.011052771471440792, + 0.005339766386896372, + 0.00818354170769453, + 0.00393311120569706, + 0.0044822609052062035, + 0.00849897414445877, + 0.0076952725648880005, + 0.004712033085525036, + 0.005980103276669979, + 0.025764847174286842, + 0.026759466156363487, + 0.03068484552204609, + 0.019944556057453156, + 0.029111716896295547, + 0.023568006232380867, + 0.042194779962301254, + 0.03693274408578873, + 0.009442671202123165, + 0.012763423845171928, + 0.007608160842210054, + 0.006119923200458288, + 0.004434350878000259, + 0.004032118711620569, + 0.0044218203984200954, + 0.005684493575245142, + 0.008180985227227211, + 0.0037042880430817604, + 0.006601777393370867, + 0.004445008933544159, + 0.00436383718624711, + 0.003390747820958495, + 0.005574661772698164, + 0.0029521428514271975, + 0.007874293252825737, + 0.007457050960510969, + 0.008179789409041405, + 0.006964354310184717, + 0.01080302707850933, + 0.016984086483716965, + 0.006389071699231863, + 0.01427515409886837, + 0.022887209430336952, + 0.014727186411619186, + 0.01939423382282257, + 0.010153215378522873, + 0.007716975640505552, + 0.04755769670009613, + 0.008542914874851704, + 0.031295035034418106, + 0.005913378670811653, + 0.004445851780474186, + 0.006199026480317116, + 0.010853172279894352, + 0.004261094145476818, + 0.003973537124693394, + 0.0034555813763290644, + 0.0038365251384675503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.05499108135700226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002570480108261 + ], + "opposite_logprob_tokens": [ + -2.695025682449341 + ], + "logprob_reward": -0.07002570480108261, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002570480108261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11259540170431137, + 0.07644737511873245, + 0.17341630160808563, + 0.03777045011520386, + 0.08226902037858963, + 0.07132090628147125, + 0.057336632162332535, + 0.13075438141822815, + 0.08462175726890564, + 0.17346780002117157 + ], + "credit_map": [ + 0.013540479354560375, + 0.013703336007893085, + 0.015507793985307217, + 0.012470441870391369, + 0.018454503268003464, + 0.014976856298744678, + 0.009651359170675278, + 0.014290637336671352, + 0.007219416089355946, + 0.007097991183400154, + 0.00759806577116251, + 0.01219992060214281, + 0.012038771994411945, + 0.011018043383955956, + 0.0122817512601614, + 0.006993415299803019, + 0.014644145034253597, + 0.020026395097374916, + 0.013797915540635586, + 0.014545487239956856, + 0.035719454288482666, + 0.023421315476298332, + 0.03503884747624397, + 0.0162227563560009, + 0.003950139973312616, + 0.004641388542950153, + 0.0030402718111872673, + 0.005747883580625057, + 0.006080599967390299, + 0.005346724763512611, + 0.0053689437918365, + 0.0035944960545748472, + 0.01068880595266819, + 0.009949053637683392, + 0.00956928264349699, + 0.008691644296050072, + 0.013376410119235516, + 0.008770430460572243, + 0.009919144213199615, + 0.011304250918328762, + 0.013314615935087204, + 0.017029544338583946, + 0.010797916911542416, + 0.0080154649913311, + 0.004819209687411785, + 0.005185585934668779, + 0.00499697495251894, + 0.007161599118262529, + 0.014343279413878918, + 0.005407506600022316, + 0.013679281808435917, + 0.0066980295814573765, + 0.0037506488151848316, + 0.0038514656480401754, + 0.004397235345095396, + 0.005209181923419237, + 0.012864325195550919, + 0.01212648581713438, + 0.03714009374380112, + 0.010931028984487057, + 0.00871432013809681, + 0.012342344038188457, + 0.006895983126014471, + 0.029739810153841972, + 0.013046128675341606, + 0.007937251590192318, + 0.009902829304337502, + 0.004869858734309673, + 0.005036101210862398, + 0.022660059854388237, + 0.005691245198249817, + 0.015478277578949928, + 0.016014112159609795, + 0.014946612529456615, + 0.022040363401174545, + 0.05592332035303116, + 0.017674431204795837, + 0.011038880795240402, + 0.020970504730939865, + 0.01485956646502018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.07002570480108261 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633799687027931 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633799687027931, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633799687027931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3350042402744293, + 0.21064724028110504, + 0.05378330126404762, + 0.0325249545276165, + 0.053931090980768204, + 0.038423482328653336, + 0.05294324457645416, + 0.03550982102751732, + 0.14426778256893158, + 0.04296481981873512 + ], + "credit_map": [ + 0.03160968795418739, + 0.02761579304933548, + 0.0320417657494545, + 0.02267446182668209, + 0.06616553664207458, + 0.05163286253809929, + 0.024340316653251648, + 0.07892382144927979, + 0.03832292556762695, + 0.025459300726652145, + 0.03280341625213623, + 0.019809149205684662, + 0.019721508026123047, + 0.028792114928364754, + 0.017040668055415154, + 0.0286981463432312, + 0.00937443133443594, + 0.006058491300791502, + 0.0065718889236450195, + 0.004048328381031752, + 0.007939118891954422, + 0.0062413932755589485, + 0.006902453023940325, + 0.006647198926657438, + 0.00547489058226347, + 0.006298320833593607, + 0.0037721171975135803, + 0.0035869546700268984, + 0.0026165719609707594, + 0.004487360827624798, + 0.002501019276678562, + 0.0037877194117754698, + 0.006220912095159292, + 0.006899354979395866, + 0.00745175126940012, + 0.007454343605786562, + 0.006022307556122541, + 0.00456786947324872, + 0.007799563929438591, + 0.007514988072216511, + 0.00527455797418952, + 0.006890946999192238, + 0.0053738076239824295, + 0.006262767128646374, + 0.00405667582526803, + 0.0027952217496931553, + 0.003421985311433673, + 0.004347518552094698, + 0.015715790912508965, + 0.0043730060569942, + 0.013947155326604843, + 0.00608136085793376, + 0.0022541240323334932, + 0.003910468425601721, + 0.002446784172207117, + 0.004214550368487835, + 0.003605764592066407, + 0.004503641277551651, + 0.0031592920422554016, + 0.002839795546606183, + 0.002711889101192355, + 0.008604830130934715, + 0.0017881764797493815, + 0.008296431042253971, + 0.01722487062215805, + 0.012191579677164555, + 0.016042470932006836, + 0.009328704327344894, + 0.00790271908044815, + 0.04385679215192795, + 0.008244793862104416, + 0.029475849121809006, + 0.0054828147403895855, + 0.0037830350920557976, + 0.005701621528714895, + 0.008398259058594704, + 0.003705588635057211, + 0.0043131085112690926, + 0.004876419901847839, + 0.006703971419483423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.007633799687027931 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203977018594742 + ], + "opposite_logprob_tokens": [ + -2.812039852142334 + ], + "logprob_reward": -0.06203977018594742, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203977018594742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08529683202505112, + 0.022825712338089943, + 0.03659570589661598, + 0.030132222920656204, + 0.022347576916217804, + 0.021112021058797836, + 0.14654035866260529, + 0.25100910663604736, + 0.17669205367565155, + 0.20744845271110535 + ], + "credit_map": [ + 0.010715377517044544, + 0.0072852750308811665, + 0.00894245132803917, + 0.006221801973879337, + 0.01528457272797823, + 0.013167878612875938, + 0.006233408581465483, + 0.01744607463479042, + 0.002555255312472582, + 0.0027795094065368176, + 0.00260718259960413, + 0.0027745983097702265, + 0.003206368535757065, + 0.003236413700506091, + 0.0028466936200857162, + 0.002819688292220235, + 0.005385453812777996, + 0.005248095374554396, + 0.006097477860748768, + 0.003904137061908841, + 0.004914411343634129, + 0.004306550603359938, + 0.003599938005208969, + 0.0031396402046084404, + 0.0044143605045974255, + 0.0050789956003427505, + 0.003883840749040246, + 0.0033259023912250996, + 0.0030466716270893812, + 0.004937966354191303, + 0.0021127916406840086, + 0.003331696381792426, + 0.0040473854169249535, + 0.0028977105394005775, + 0.002651074668392539, + 0.00220612995326519, + 0.003421276807785034, + 0.0017519546672701836, + 0.0024811318144202232, + 0.002890912815928459, + 0.003815112169831991, + 0.0032019473146647215, + 0.0026131211780011654, + 0.0028210633900016546, + 0.0023554747458547354, + 0.0015702227829024196, + 0.0023272340185940266, + 0.0024078458081930876, + 0.040661249309778214, + 0.009836205281317234, + 0.043577954173088074, + 0.019229266792535782, + 0.005737551022320986, + 0.0110681913793087, + 0.005460373591631651, + 0.010969553142786026, + 0.022541331127285957, + 0.025831950828433037, + 0.06685119867324829, + 0.016574077308177948, + 0.009601256810128689, + 0.03684726729989052, + 0.007232425268739462, + 0.06552959233522415, + 0.015187135897576809, + 0.010864563286304474, + 0.017697572708129883, + 0.008267759345471859, + 0.008033392019569874, + 0.05021514743566513, + 0.009175281040370464, + 0.057251203805208206, + 0.018888454884290695, + 0.018088847398757935, + 0.024510467424988747, + 0.06363686174154282, + 0.022364506497979164, + 0.01409810408949852, + 0.02692524716258049, + 0.01893596164882183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.06203977018594742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle tracking shot follows a poet dressed in Tang Dynasty attire as he journeys through a vast, mist-shrouded bamboo forest. The poet's face is thin and pensive, his long, dark hair held in a bun with a simple wooden hairpin. He wears an indigo silk round-necked robe, its wide sleeves swaying gently with his leisurely steps. His left hand gently grasps a scroll of yellowed bamboo slips, while his right hand occasionally strokes the emerald green bamboo stalks beside him, dappled with morning dew. The key light source comes from above, with beams of golden sunlight penetrating the dense bamboo canopy and milky mist, creating a distinct Tyndall effect. The poet and the ground, carpeted with wet fallen leaves, are dappled with shifting, interplay of light and dark. In the background, the overlapping, varying shades of bamboo shadows, looming through the mist, create a sense of infinite depth. The entire picture is dominated by the cool green of the bamboo forest and the soft white of the mist, illuminated by the warm sunlight, presenting a tranquil, ethereal and cinematic atmosphere full of oriental classical aesthetics.", + "seed": 21729, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_4986_seed_21729_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0743156545329839, + "vlm_logprob": -0.0743156545329839, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.0743156545329839, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09032034873962402 + ], + "opposite_logprob_tokens": [ + -2.715320348739624 + ], + "logprob_reward": -0.09032034873962402, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09032034873962402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09872036427259445, + 0.04179874807596207, + 0.0773518905043602, + 0.04032107815146446, + 0.10766784101724625, + 0.08981097489595413, + 0.06496808677911758, + 0.03140952065587044, + 0.3147578835487366, + 0.13319361209869385 + ], + "credit_map": [ + 0.008368442766368389, + 0.010880246758460999, + 0.013490656390786171, + 0.0167339276522398, + 0.017528513446450233, + 0.010452614165842533, + 0.013960556127130985, + 0.007305407430976629, + 0.004412826616317034, + 0.0037121109198778868, + 0.005296827759593725, + 0.006264000199735165, + 0.004828140139579773, + 0.008136431686580181, + 0.0048811412416398525, + 0.004267267417162657, + 0.01411247905343771, + 0.0092406515032053, + 0.00897816102951765, + 0.010031894780695438, + 0.007451718207448721, + 0.01638621836900711, + 0.005475062411278486, + 0.005675707943737507, + 0.007258233148604631, + 0.0033936600666493177, + 0.006736279930919409, + 0.004574112128466368, + 0.005574165843427181, + 0.004952164366841316, + 0.003996554296463728, + 0.003835907205939293, + 0.009764603339135647, + 0.007103392388671637, + 0.009804563596844673, + 0.018477967008948326, + 0.014492261223495007, + 0.014229604043066502, + 0.020942943170666695, + 0.01285250298678875, + 0.011697321198880672, + 0.008709506131708622, + 0.009734019637107849, + 0.01058871764689684, + 0.012499661184847355, + 0.012473261915147305, + 0.010251644998788834, + 0.013856843113899231, + 0.006240108981728554, + 0.006438982207328081, + 0.005981799680739641, + 0.008036519400775433, + 0.010348742827773094, + 0.009089176543056965, + 0.008436537347733974, + 0.010396216064691544, + 0.0032984616700559855, + 0.002177337184548378, + 0.006303243339061737, + 0.004629521630704403, + 0.0032653158996254206, + 0.004544673953205347, + 0.002900737803429365, + 0.0042902277782559395, + 0.050767917186021805, + 0.034138936549425125, + 0.04937632009387016, + 0.03983255475759506, + 0.043473511934280396, + 0.03660832345485687, + 0.03048679418861866, + 0.030073538422584534, + 0.00956054124981165, + 0.006607415620237589, + 0.012667964212596416, + 0.012295881286263466, + 0.015256552025675774, + 0.03749791532754898, + 0.011541217565536499, + 0.027766117826104164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.09032034873962402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003575474023819 + ], + "opposite_logprob_tokens": [ + -2.3503575325012207 + ], + "logprob_reward": -0.1003575474023819, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003575474023819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13138213753700256, + 0.0619659461081028, + 0.09395790845155716, + 0.10278038680553436, + 0.13844145834445953, + 0.07111484557390213, + 0.04625916853547096, + 0.02048538066446781, + 0.17773547768592834, + 0.15587729215621948 + ], + "credit_map": [ + 0.014047040604054928, + 0.012541436590254307, + 0.02064066007733345, + 0.015049267560243607, + 0.02909538522362709, + 0.013790345750749111, + 0.013331269845366478, + 0.012886726297438145, + 0.006854357197880745, + 0.005295475013554096, + 0.010060206055641174, + 0.010387582704424858, + 0.009453780017793179, + 0.005593923851847649, + 0.008873110637068748, + 0.005447508301585913, + 0.012479946948587894, + 0.013840476050972939, + 0.009356779046356678, + 0.012156082317233086, + 0.018447743728756905, + 0.01190026756376028, + 0.00819619931280613, + 0.007580408826470375, + 0.009075714275240898, + 0.007238524034619331, + 0.014600506983697414, + 0.013817554339766502, + 0.01980847492814064, + 0.011442797258496284, + 0.015094389207661152, + 0.011702421121299267, + 0.015974201261997223, + 0.012585615739226341, + 0.011105798184871674, + 0.020380791276693344, + 0.02040385641157627, + 0.017666999250650406, + 0.026876673102378845, + 0.013447516597807407, + 0.0076015847735106945, + 0.005858688149601221, + 0.005998227279633284, + 0.009425055235624313, + 0.011928382329642773, + 0.00950238574296236, + 0.010923760011792183, + 0.009876759722828865, + 0.004609478637576103, + 0.0033586788922548294, + 0.004058923106640577, + 0.00921949464827776, + 0.005785502027720213, + 0.007127334363758564, + 0.005798539146780968, + 0.006301217246800661, + 0.0018211425049230456, + 0.0016799494624137878, + 0.0033494748640805483, + 0.0035619831178337336, + 0.0024676427710801363, + 0.00262992805801332, + 0.0018065180629491806, + 0.0031687412410974503, + 0.028123848140239716, + 0.01869780942797661, + 0.027761410921812057, + 0.023924486711621284, + 0.023424027487635612, + 0.021274404600262642, + 0.018181489780545235, + 0.016348008066415787, + 0.013088892214000225, + 0.010680422186851501, + 0.015779761597514153, + 0.022293100133538246, + 0.02685331366956234, + 0.02671854756772518, + 0.016850745305418968, + 0.02361249551177025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.1003575474023819 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904305845499039 + ], + "opposite_logprob_tokens": [ + -2.464043140411377 + ], + "logprob_reward": -0.08904305845499039, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904305845499039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1183062344789505, + 0.15670116245746613, + 0.07616658508777618, + 0.1305178701877594, + 0.17982690036296844, + 0.06074849143624306, + 0.06057235971093178, + 0.035203248262405396, + 0.10081202536821365, + 0.08114514499902725 + ], + "credit_map": [ + 0.010602308437228203, + 0.010032731108367443, + 0.018281439319252968, + 0.018991392105817795, + 0.016881171613931656, + 0.012922622263431549, + 0.013404500670731068, + 0.01719006709754467, + 0.017321132123470306, + 0.012428157031536102, + 0.019208867102861404, + 0.021541299298405647, + 0.018049627542495728, + 0.03892473131418228, + 0.01336949784308672, + 0.015857866033911705, + 0.010569680482149124, + 0.011826876550912857, + 0.008840607479214668, + 0.011531532742083073, + 0.01118080411106348, + 0.009143548086285591, + 0.0067682405933737755, + 0.006305294577032328, + 0.014996669255197048, + 0.008418653160333633, + 0.019934920594096184, + 0.01451766025274992, + 0.023674272000789642, + 0.015306539833545685, + 0.018208734691143036, + 0.015460420399904251, + 0.019508250057697296, + 0.014390953816473484, + 0.017163731157779694, + 0.02943035215139389, + 0.026775671169161797, + 0.02205676957964897, + 0.02619866281747818, + 0.02430250681936741, + 0.00881499145179987, + 0.005464029964059591, + 0.005664310418069363, + 0.00660515995696187, + 0.010331536643207073, + 0.007514889817684889, + 0.008286384865641594, + 0.008067191578447819, + 0.004569632466882467, + 0.006590902805328369, + 0.005883814767003059, + 0.007387036923319101, + 0.009392805397510529, + 0.007891220040619373, + 0.010577021166682243, + 0.008279928006231785, + 0.0034450029488652945, + 0.0022538448683917522, + 0.00862905289977789, + 0.004073140677064657, + 0.0032943394035100937, + 0.004752689972519875, + 0.0030553797259926796, + 0.005699795205146074, + 0.015247604809701443, + 0.010168868117034435, + 0.015491042286157608, + 0.012495364062488079, + 0.01343983132392168, + 0.013423395343124866, + 0.010273815132677555, + 0.010272108018398285, + 0.0068033295683562756, + 0.005596962757408619, + 0.011515859514474869, + 0.011268861591815948, + 0.012651241384446621, + 0.011934970505535603, + 0.009356577880680561, + 0.012017342261970043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.08904305845499039 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028208012226969004 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028208012226969004, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028208012226969004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19799178838729858, + 0.0334608368575573, + 0.08978721499443054, + 0.07432689517736435, + 0.09770100563764572, + 0.07315107434988022, + 0.03824572637677193, + 0.184698224067688, + 0.12460863590240479, + 0.0860285833477974 + ], + "credit_map": [ + 0.024111801758408546, + 0.01334304828196764, + 0.03806447610259056, + 0.024261288344860077, + 0.02802894078195095, + 0.03173494711518288, + 0.018692051991820335, + 0.019755259156227112, + 0.0032140673138201237, + 0.0026020463556051254, + 0.004582036752253771, + 0.00426775636151433, + 0.003970850259065628, + 0.006074314936995506, + 0.004119619727134705, + 0.004630147945135832, + 0.01626185141503811, + 0.008231072686612606, + 0.013606123626232147, + 0.0063203428871929646, + 0.007811759598553181, + 0.025411885231733322, + 0.005405500065535307, + 0.006738688796758652, + 0.010796659626066685, + 0.01002668496221304, + 0.010117949917912483, + 0.00862465426325798, + 0.01114795170724392, + 0.008959735743701458, + 0.008073793724179268, + 0.006579472683370113, + 0.01561109721660614, + 0.013193893246352673, + 0.012316583655774593, + 0.013067108578979969, + 0.010215952061116695, + 0.012233704328536987, + 0.010879085399210453, + 0.010183583945035934, + 0.01181809976696968, + 0.009429345838725567, + 0.008997892960906029, + 0.008100082166492939, + 0.00911498162895441, + 0.009543024934828281, + 0.006102893967181444, + 0.010044761002063751, + 0.004993542563170195, + 0.005674258805811405, + 0.003550375811755657, + 0.006951343268156052, + 0.0033004386350512505, + 0.005466490983963013, + 0.003945688717067242, + 0.004363590385764837, + 0.02060774713754654, + 0.011238886043429375, + 0.046779222786426544, + 0.020660027861595154, + 0.015531769953668118, + 0.02654326520860195, + 0.01578032784163952, + 0.027556980028748512, + 0.021114785224199295, + 0.0170772522687912, + 0.0236294437199831, + 0.010496426373720169, + 0.01480951253324747, + 0.015819484367966652, + 0.008616359904408455, + 0.01304538082331419, + 0.005835758056491613, + 0.005109881982207298, + 0.011191711761057377, + 0.007530120201408863, + 0.011511812917888165, + 0.022232001647353172, + 0.007051353342831135, + 0.015565944835543633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0028208012226969004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903651684522629 + ], + "opposite_logprob_tokens": [ + -2.464036464691162 + ], + "logprob_reward": -0.08903651684522629, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903651684522629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.277336061000824, + 0.2798535227775574, + 0.07197560369968414, + 0.03988783434033394, + 0.042884062975645065, + 0.038631781935691833, + 0.04251626506447792, + 0.028029322624206543, + 0.06487265974283218, + 0.11401289701461792 + ], + "credit_map": [ + 0.030077867209911346, + 0.013761835172772408, + 0.04974793270230293, + 0.03325231745839119, + 0.0324455127120018, + 0.038939639925956726, + 0.029861288145184517, + 0.049249693751335144, + 0.05402626842260361, + 0.017044516280293465, + 0.07166986167430878, + 0.03041338361799717, + 0.027455223724246025, + 0.027227897197008133, + 0.02700851857662201, + 0.02500787563621998, + 0.01188577339053154, + 0.006810027174651623, + 0.015357030555605888, + 0.008265840820968151, + 0.007779667153954506, + 0.01316466648131609, + 0.004981272853910923, + 0.0037313313223421574, + 0.005241907201707363, + 0.0025632164906710386, + 0.0045720916241407394, + 0.005120780784636736, + 0.0058946311473846436, + 0.005522811785340309, + 0.0046966224908828735, + 0.006275778170675039, + 0.004083866253495216, + 0.004988487344235182, + 0.006360301282256842, + 0.006404704414308071, + 0.004408129025250673, + 0.005031653679907322, + 0.006510984152555466, + 0.005095938686281443, + 0.0034654568880796432, + 0.004098181612789631, + 0.004464356694370508, + 0.005246771965175867, + 0.006767970975488424, + 0.004856156650930643, + 0.00483237998560071, + 0.00490051181986928, + 0.0037252926267683506, + 0.0035998118110001087, + 0.0045308563858270645, + 0.00655874190852046, + 0.005679279565811157, + 0.005868606269359589, + 0.007083837408572435, + 0.005469843279570341, + 0.003726379247382283, + 0.0023199382703751326, + 0.003951468970626593, + 0.004149844869971275, + 0.003177231177687645, + 0.00457434868440032, + 0.0029486205894500017, + 0.0031814922112971544, + 0.010973500087857246, + 0.0075686234049499035, + 0.00932589266449213, + 0.007371193263679743, + 0.009025937877595425, + 0.007666487246751785, + 0.0062149870209395885, + 0.006726050283759832, + 0.008986810222268105, + 0.008646376430988312, + 0.012813687324523926, + 0.015017630532383919, + 0.0168314166367054, + 0.02144656330347061, + 0.012544887140393257, + 0.017725536599755287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.08903651684522629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle tracking shot follows a poet dressed in Tang Dynasty attire as he journeys through a vast, mist-shrouded bamboo forest. The poet's face is thin and pensive, his long, dark hair held in a bun with a simple wooden hairpin. He wears an indigo silk round-necked robe, its wide sleeves swaying gently with his leisurely steps. His left hand gently grasps a scroll of yellowed bamboo slips, while his right hand occasionally strokes the emerald green bamboo stalks beside him, dappled with morning dew. The key light source comes from above, with beams of golden sunlight penetrating the dense bamboo canopy and milky mist, creating a distinct Tyndall effect. The poet and the ground, carpeted with wet fallen leaves, are dappled with shifting, interplay of light and dark. In the background, the overlapping, varying shades of bamboo shadows, looming through the mist, create a sense of infinite depth. The entire picture is dominated by the cool green of the bamboo forest and the soft white of the mist, illuminated by the warm sunlight, presenting a tranquil, ethereal and cinematic atmosphere full of oriental classical aesthetics.", + "seed": 21730, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_1/index_4986_seed_21730_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06338069047778845, + "vlm_logprob": -0.06338069047778845, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.06338069047778845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15644806623458862 + ], + "opposite_logprob_tokens": [ + -2.1564481258392334 + ], + "logprob_reward": -0.15644806623458862, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15644806623458862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10210397094488144, + 0.1512252688407898, + 0.113573357462883, + 0.11191023886203766, + 0.1298854500055313, + 0.07504317164421082, + 0.16517673432826996, + 0.03997233510017395, + 0.08597422391176224, + 0.025135237723588943 + ], + "credit_map": [ + 0.012110200710594654, + 0.012789566069841385, + 0.01352197490632534, + 0.009377304464578629, + 0.01784995011985302, + 0.012493029236793518, + 0.01345048937946558, + 0.010511455126106739, + 0.03282275050878525, + 0.007785596419125795, + 0.038265161216259, + 0.011034389957785606, + 0.020535361021757126, + 0.02390683814883232, + 0.010652903467416763, + 0.006222282536327839, + 0.010412344709038734, + 0.013875216245651245, + 0.021777205169200897, + 0.01139325276017189, + 0.021162105724215508, + 0.015603996813297272, + 0.010079400613904, + 0.009269841946661472, + 0.016105156391859055, + 0.011460443027317524, + 0.017750388011336327, + 0.015231234952807426, + 0.014190608635544777, + 0.017389003187417984, + 0.00959382951259613, + 0.010189575143158436, + 0.01567937061190605, + 0.021654129028320312, + 0.01090217288583517, + 0.01618695631623268, + 0.013259151950478554, + 0.01924094744026661, + 0.016474146395921707, + 0.01648857071995735, + 0.007847022265195847, + 0.008162095211446285, + 0.009034186601638794, + 0.00786608923226595, + 0.014812560752034187, + 0.008686215616762638, + 0.01031353510916233, + 0.00832146406173706, + 0.018569182604551315, + 0.02605820819735527, + 0.011918511241674423, + 0.018824752420186996, + 0.020612819120287895, + 0.01859459839761257, + 0.0227435901761055, + 0.02785508707165718, + 0.003303843317553401, + 0.004787508398294449, + 0.004328570328652859, + 0.004973106551915407, + 0.00852576456964016, + 0.004944507498294115, + 0.0044485838152468204, + 0.004660450853407383, + 0.012732211500406265, + 0.006843599956482649, + 0.007064564153552055, + 0.008443973027169704, + 0.008201604709029198, + 0.024012502282857895, + 0.00824031699448824, + 0.010435459204018116, + 0.0025673112832009792, + 0.0037806062027812004, + 0.002787527395412326, + 0.004315026104450226, + 0.002917373785749078, + 0.003126610768958926, + 0.0026431744918227196, + 0.0029976079240441322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.15644806623458862 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813711926341057 + ], + "opposite_logprob_tokens": [ + -3.288137197494507 + ], + "logprob_reward": -0.03813711926341057, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813711926341057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2129659801721573, + 0.24044737219810486, + 0.0909406840801239, + 0.05022365599870682, + 0.08281980454921722, + 0.05600031465291977, + 0.08409783244132996, + 0.07720047235488892, + 0.0716581866145134, + 0.03364577144384384 + ], + "credit_map": [ + 0.03847913816571236, + 0.024772103875875473, + 0.028176497668027878, + 0.015071730129420757, + 0.0389467291533947, + 0.02832029014825821, + 0.018848786130547523, + 0.02035069838166237, + 0.04279237985610962, + 0.015264423564076424, + 0.054211635142564774, + 0.022365828976035118, + 0.03408629447221756, + 0.04857271537184715, + 0.01444285549223423, + 0.008711213245987892, + 0.010790294967591763, + 0.00985329132527113, + 0.01913587935268879, + 0.007824333384633064, + 0.014969920739531517, + 0.012453543953597546, + 0.007825239561498165, + 0.008088173344731331, + 0.005024776794016361, + 0.008241270668804646, + 0.004114195704460144, + 0.005332241766154766, + 0.008741437457501888, + 0.007547394838184118, + 0.005914578679949045, + 0.005307752639055252, + 0.009238790720701218, + 0.01160410325974226, + 0.01052682101726532, + 0.00986377615481615, + 0.008523602038621902, + 0.013096781447529793, + 0.009912597946822643, + 0.010053325444459915, + 0.009019885212182999, + 0.00757063552737236, + 0.0046356297098100185, + 0.00480293994769454, + 0.007272184360772371, + 0.008758693002164364, + 0.006547767203301191, + 0.007392575033009052, + 0.010743703693151474, + 0.012221312150359154, + 0.007151945028454065, + 0.011341365985572338, + 0.007568641100078821, + 0.009588150307536125, + 0.010465354658663273, + 0.01501736044883728, + 0.009857979603111744, + 0.006903484463691711, + 0.014645267277956009, + 0.00704380264505744, + 0.010387604124844074, + 0.01363759022206068, + 0.006927247624844313, + 0.00779749546200037, + 0.008063564077019691, + 0.006254558917135, + 0.005859248340129852, + 0.00737377954646945, + 0.006871350109577179, + 0.019057603552937508, + 0.007333709392696619, + 0.010844369418919086, + 0.002476013731211424, + 0.005661513190716505, + 0.0033240457996726036, + 0.006150571163743734, + 0.004670496564358473, + 0.003667530370876193, + 0.004450404085218906, + 0.003245194675400853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03813711926341057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104645907878876 + ], + "opposite_logprob_tokens": [ + -3.28810453414917 + ], + "logprob_reward": -0.038104645907878876, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104645907878876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21790476143360138, + 0.18413542211055756, + 0.08947660773992538, + 0.17441120743751526, + 0.07946760207414627, + 0.07155594974756241, + 0.04990555718541145, + 0.06350723654031754, + 0.02853572741150856, + 0.041099969297647476 + ], + "credit_map": [ + 0.028806904330849648, + 0.038601431995630264, + 0.027346037328243256, + 0.022548463195562363, + 0.03195084631443024, + 0.02770267054438591, + 0.024668341502547264, + 0.01628006622195244, + 0.03687926381826401, + 0.018486080691218376, + 0.04110019654035568, + 0.020992493256926537, + 0.017819534987211227, + 0.02266697771847248, + 0.017970971763134003, + 0.008219915442168713, + 0.010821830481290817, + 0.008280803449451923, + 0.019223690032958984, + 0.00859758909791708, + 0.016609815880656242, + 0.01080278679728508, + 0.008702965453267097, + 0.006437122821807861, + 0.03240055590867996, + 0.018765799701213837, + 0.03512096032500267, + 0.0185394324362278, + 0.022647574543952942, + 0.020727502182126045, + 0.013217804953455925, + 0.012991566210985184, + 0.0089016268029809, + 0.01215401478111744, + 0.005935611668974161, + 0.008912505581974983, + 0.009683385491371155, + 0.012280418537557125, + 0.010391829535365105, + 0.011208205483853817, + 0.006118724122643471, + 0.008308622054755688, + 0.0072019221261143684, + 0.006609298754483461, + 0.011779009364545345, + 0.011991375125944614, + 0.008239476010203362, + 0.011307520791888237, + 0.006721844896674156, + 0.008540046401321888, + 0.0032066553831100464, + 0.005754739046096802, + 0.007706015836447477, + 0.006669912952929735, + 0.00602660933509469, + 0.0052797324024140835, + 0.007614508271217346, + 0.005523184314370155, + 0.009506179951131344, + 0.005512464325875044, + 0.010569868609309196, + 0.012005929835140705, + 0.0062835728749632835, + 0.006491528823971748, + 0.0040324777364730835, + 0.003241217928007245, + 0.002738850424066186, + 0.00295368698425591, + 0.0036830094177275896, + 0.004184618592262268, + 0.0030809591989964247, + 0.004620908293873072, + 0.0032111064065247774, + 0.004718093201518059, + 0.005872922483831644, + 0.004675593692809343, + 0.008940632455050945, + 0.0042325654067099094, + 0.0038876784965395927, + 0.005561378784477711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.038104645907878876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196357689797878 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196357689797878, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196357689797878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16084381937980652, + 0.17602872848510742, + 0.23222514986991882, + 0.1581435650587082, + 0.04344583675265312, + 0.04063641279935837, + 0.029796626418828964, + 0.058042626827955246, + 0.07364046573638916, + 0.027196813374757767 + ], + "credit_map": [ + 0.019273627549409866, + 0.016151048243045807, + 0.01820072904229164, + 0.013536881655454636, + 0.026020754128694534, + 0.02174534648656845, + 0.024816278368234634, + 0.021099139004945755, + 0.02347112074494362, + 0.011043132282793522, + 0.029950503259897232, + 0.014324747025966644, + 0.032459158450365067, + 0.04243132844567299, + 0.012875143438577652, + 0.00947357714176178, + 0.033649981021881104, + 0.01844205893576145, + 0.018968157470226288, + 0.01756538264453411, + 0.06607571244239807, + 0.03459492325782776, + 0.01978290267288685, + 0.02314600721001625, + 0.02977868542075157, + 0.01703970693051815, + 0.03205972909927368, + 0.010474760085344315, + 0.023010186851024628, + 0.02285485342144966, + 0.009737320244312286, + 0.013188301585614681, + 0.006412869784981012, + 0.005681363400071859, + 0.0038074946496635675, + 0.006054533179849386, + 0.0037219454534351826, + 0.006779166404157877, + 0.004962550941854715, + 0.00602590711787343, + 0.0073294807225465775, + 0.004962774459272623, + 0.004088683519512415, + 0.004616247955709696, + 0.006368853151798248, + 0.00539162615314126, + 0.0036745022516697645, + 0.004204243887215853, + 0.0036635349970310926, + 0.00442085973918438, + 0.0018879742128774524, + 0.0032857886981219053, + 0.0033211291301995516, + 0.004674240946769714, + 0.0038422138895839453, + 0.004700884688645601, + 0.009743846952915192, + 0.005199647042900324, + 0.008920573629438877, + 0.0051505072042346, + 0.004530345089733601, + 0.013904436491429806, + 0.004867925774306059, + 0.0057253409177064896, + 0.012019661255180836, + 0.004087374079972506, + 0.007830231450498104, + 0.007015119772404432, + 0.008498732931911945, + 0.01971837505698204, + 0.006215312983840704, + 0.00825564842671156, + 0.0027826877776533365, + 0.003732153680175543, + 0.003977573476731777, + 0.003206440946087241, + 0.00445534149184823, + 0.003197771729901433, + 0.003142615780234337, + 0.0027022273279726505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014196357689797878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a steady, low-angle tracking shot following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a poet wearing Tang Dynasty attire with a thin, pensive face and long dark hair in a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the poet wearing an indigo silk round-necked robe with wide sleeves, holding a scroll of yellowed bamboo slips in his left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast, mist-shrouded bamboo forest with wet fallen leaves on the ground and golden sunlight beams creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the poet's right hand occasionally stroking the emerald green bamboo stalks beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700172632932663 + ], + "opposite_logprob_tokens": [ + -2.6950173377990723 + ], + "logprob_reward": -0.0700172632932663, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700172632932663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2007126361131668, + 0.05488201975822449, + 0.10889355838298798, + 0.05876042693853378, + 0.06342838704586029, + 0.16887257993221283, + 0.18746314942836761, + 0.04689266160130501, + 0.07336398214101791, + 0.03673054650425911 + ], + "credit_map": [ + 0.015474825166165829, + 0.023860029876232147, + 0.017879273742437363, + 0.013847174122929573, + 0.04975336790084839, + 0.031544897705316544, + 0.024550920352339745, + 0.02380215935409069, + 0.0050543504767119884, + 0.006530158221721649, + 0.0038829208351671696, + 0.010304814204573631, + 0.00905348639935255, + 0.007824106141924858, + 0.00827038288116455, + 0.003961800131946802, + 0.01255003921687603, + 0.010391389019787312, + 0.007332792971283197, + 0.009096385911107063, + 0.03243110701441765, + 0.018676629289984703, + 0.008566477335989475, + 0.009848734363913536, + 0.009760621003806591, + 0.007034188136458397, + 0.009771750308573246, + 0.0065521192736923695, + 0.009006379172205925, + 0.007271425798535347, + 0.005836979486048222, + 0.003526961663737893, + 0.0060330224223434925, + 0.010811660438776016, + 0.00559655949473381, + 0.010036444291472435, + 0.0069173951633274555, + 0.0081710834056139, + 0.008564617484807968, + 0.007297610864043236, + 0.019012533128261566, + 0.016746321693062782, + 0.017827656120061874, + 0.022728340700268745, + 0.024439048022031784, + 0.02062463015317917, + 0.027573365718126297, + 0.019920682534575462, + 0.020530907437205315, + 0.028601882979273796, + 0.011875851079821587, + 0.022730804979801178, + 0.022896045818924904, + 0.02252497710287571, + 0.026872511953115463, + 0.03143016993999481, + 0.004154005087912083, + 0.005463765002787113, + 0.00590539863333106, + 0.004407891537994146, + 0.010449773631989956, + 0.0056359064765274525, + 0.005524227861315012, + 0.00535169430077076, + 0.008040008135139942, + 0.007653811015188694, + 0.007478435058146715, + 0.008084754459559917, + 0.008600723929703236, + 0.01552851777523756, + 0.008743268437683582, + 0.00923446286469698, + 0.0029558748938143253, + 0.005312304012477398, + 0.00412787077948451, + 0.0033773728646337986, + 0.00595957413315773, + 0.004708836320787668, + 0.0030660235788673162, + 0.007222686428576708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0700172632932663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_tensor([1097, 1097])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_tensor([1097, 1097])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..88066b8d97bc07bf49693ab7992b1a35b6e7ed46 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_tensor([1097, 1097])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, majestic wide-angle shot, framed in the style of a traditional Chinese landscape painting, captures a massive bronze dragon statue, its powerful, serpentine body coiled around the highest, mist-shrouded peak of a jagged mountain range. Initially, its metallic scales gleam with a deep, polished bronze sheen, its fierce head, flowing beard, bare teeth, facing the sky. Time-lapse begins, compressing centuries into a single moment. Sun and moon streak across the sky in fleeting arcs of light, while ink-like clouds swirl over the peaks like a celestial river. The seasons shift with breathtaking regularity: snow falls and melts in an instant, while sparse, gnarled pines cling to the cliffs, flickering between the vibrant green of spring and the chill of winter. In this endless cycle, the slow and relentless process of erosion becomes clearly visible on the dragon; light moisture begins to leave marks along its form, and over time these marks transform into a vibrant green patina that visibly runs along the bronze, pools in the crevices of its intricate scales, and flows from the corners of its eyes, depicting the relentless passage of time and transforming the majestic figure into an ancient, weathered monument, a timeless mountain it commands.", + "seed": 12857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_1097_seed_12857_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.42176311612129214, + "vlm_logprob": -0.42176311612129214, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.42176311612129214, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021355098113417625 + ], + "opposite_logprob_tokens": [ + -4.146355152130127 + ], + "logprob_reward": -0.021355098113417625, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021355098113417625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1877550184726715, + 0.21949100494384766, + 0.071574866771698, + 0.07115741074085236, + 0.08585522323846817, + 0.07837449014186859, + 0.09309045225381851, + 0.036309000104665756, + 0.04299424588680267, + 0.11339825391769409 + ], + "credit_map": [ + 0.026247451081871986, + 0.017422819510102272, + 0.036595530807971954, + 0.019810575991868973, + 0.019067591056227684, + 0.02438051998615265, + 0.02058694325387478, + 0.023643603548407555, + 0.02618340216577053, + 0.014266489073634148, + 0.044796839356422424, + 0.02254195511341095, + 0.02043073996901512, + 0.02895328775048256, + 0.031644802540540695, + 0.03067351132631302, + 0.009419540874660015, + 0.0064788442105054855, + 0.020437750965356827, + 0.00668455520644784, + 0.004679744131863117, + 0.008712205104529858, + 0.004288636613637209, + 0.010873599909245968, + 0.008265306241810322, + 0.006142667960375547, + 0.017737045884132385, + 0.006882393732666969, + 0.00682037090882659, + 0.007506304886192083, + 0.0067971269600093365, + 0.01100620161741972, + 0.008810466155409813, + 0.014497791416943073, + 0.010633818805217743, + 0.009161890484392643, + 0.013937205076217651, + 0.007836217060685158, + 0.013371018692851067, + 0.007606823928654194, + 0.009537004865705967, + 0.008506403304636478, + 0.008743089623749256, + 0.01077643409371376, + 0.010208296589553356, + 0.00960242934525013, + 0.01082510594278574, + 0.010175731964409351, + 0.013303297571837902, + 0.010753011330962181, + 0.012815812602639198, + 0.006798755377531052, + 0.014351249672472477, + 0.013445130549371243, + 0.011465560644865036, + 0.010157649405300617, + 0.008299930952489376, + 0.0036689687985926867, + 0.003505715634673834, + 0.0032935207709670067, + 0.004093383438885212, + 0.0034139291383326054, + 0.0051398989744484425, + 0.004893654957413673, + 0.006548806093633175, + 0.0054867565631866455, + 0.005774060729891062, + 0.005203032400459051, + 0.006126396358013153, + 0.0033431826159358025, + 0.00431927852332592, + 0.006192740518599749, + 0.023612024262547493, + 0.011115461587905884, + 0.014659975655376911, + 0.00872456468641758, + 0.01578601449728012, + 0.010406238958239555, + 0.009389455430209637, + 0.019704531878232956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.021355098113417625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371976688504219 + ], + "opposite_logprob_tokens": [ + -3.408719778060913 + ], + "logprob_reward": -0.03371976688504219, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371976688504219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10261960327625275, + 0.08815242350101471, + 0.10364503413438797, + 0.07533316314220428, + 0.10049919784069061, + 0.11693872511386871, + 0.183846578001976, + 0.059753354638814926, + 0.07771922647953033, + 0.0914926752448082 + ], + "credit_map": [ + 0.014377766288816929, + 0.011311749927699566, + 0.015518633648753166, + 0.009752117097377777, + 0.011965795420110226, + 0.014190426096320152, + 0.014150476083159447, + 0.011352633126080036, + 0.010818335227668285, + 0.008907567709684372, + 0.014406699687242508, + 0.009247868321835995, + 0.009921573102474213, + 0.009322619996964931, + 0.014311203733086586, + 0.011216552928090096, + 0.013729674741625786, + 0.01035735011100769, + 0.028475400060415268, + 0.00910427886992693, + 0.007149842567741871, + 0.011531274765729904, + 0.006638687569648027, + 0.016658520326018333, + 0.007383656222373247, + 0.007624106481671333, + 0.015079258941113949, + 0.008441455662250519, + 0.012399191036820412, + 0.0074456981383264065, + 0.008929089643061161, + 0.008030702359974384, + 0.011396218091249466, + 0.013376183807849884, + 0.017918795347213745, + 0.009452124126255512, + 0.011066392064094543, + 0.010297332890331745, + 0.011863653548061848, + 0.015128491446375847, + 0.012311969883739948, + 0.012661119922995567, + 0.014492222107946873, + 0.022619713097810745, + 0.015762507915496826, + 0.012130183167755604, + 0.012336039915680885, + 0.01462495420128107, + 0.023174621164798737, + 0.008348730392754078, + 0.013240590691566467, + 0.007841452024877071, + 0.013594601303339005, + 0.050847962498664856, + 0.012773549184203148, + 0.054025061428546906, + 0.016231099143624306, + 0.0061563244089484215, + 0.006958952639251947, + 0.005143248941749334, + 0.005096072796732187, + 0.0051256148144602776, + 0.007128935772925615, + 0.007913103327155113, + 0.01841040700674057, + 0.009133979678153992, + 0.008104817941784859, + 0.007469897624105215, + 0.009626905433833599, + 0.005028895102441311, + 0.0075617399998009205, + 0.01238257996737957, + 0.022814098745584488, + 0.008143736980855465, + 0.01088528148829937, + 0.0048454562202095985, + 0.009534459561109543, + 0.008531882427632809, + 0.006056091748178005, + 0.020681660622358322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.03371976688504219 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290520533919334 + ], + "opposite_logprob_tokens": [ + -3.7732906341552734 + ], + "logprob_reward": -0.023290520533919334, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290520533919334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05476637929677963, + 0.044876836240291595, + 0.059764888137578964, + 0.1706351637840271, + 0.07344989478588104, + 0.0541662834584713, + 0.3797000050544739, + 0.07155076414346695, + 0.06828341633081436, + 0.022806374356150627 + ], + "credit_map": [ + 0.006282369140535593, + 0.004963853396475315, + 0.010979284532368183, + 0.005405528470873833, + 0.007250364404171705, + 0.008198742754757404, + 0.006235796958208084, + 0.00545043358579278, + 0.0062420545145869255, + 0.0034290447365492582, + 0.005159176420420408, + 0.003937185276299715, + 0.007206729147583246, + 0.005585001781582832, + 0.007286306470632553, + 0.006031332071870565, + 0.008296510204672813, + 0.005274069029837847, + 0.01503579318523407, + 0.00546574592590332, + 0.0036754952743649483, + 0.00972528476268053, + 0.0035636238753795624, + 0.008728357031941414, + 0.019094865769147873, + 0.016395296901464462, + 0.038266621530056, + 0.016130920499563217, + 0.021952053532004356, + 0.018593821674585342, + 0.019268527626991272, + 0.020933035761117935, + 0.007005004677921534, + 0.01058330200612545, + 0.012006121687591076, + 0.006542328279465437, + 0.011923292651772499, + 0.00772811658680439, + 0.009168188087642193, + 0.00849353801459074, + 0.006744363345205784, + 0.006990528199821711, + 0.0071608456782996655, + 0.007257664110511541, + 0.006738133262842894, + 0.006451682653278112, + 0.006371007766574621, + 0.0064520505256950855, + 0.05840407311916351, + 0.018093599006533623, + 0.03344036266207695, + 0.011394969187676907, + 0.02279144898056984, + 0.1025347039103508, + 0.024339284747838974, + 0.10870154201984406, + 0.02346085011959076, + 0.006654346827417612, + 0.008552069775760174, + 0.004939164966344833, + 0.004880634602159262, + 0.006935343146324158, + 0.00821127463132143, + 0.007917081005871296, + 0.015513869002461433, + 0.007253376301378012, + 0.008117477409541607, + 0.005521424580365419, + 0.006861691828817129, + 0.004590085707604885, + 0.006873746402561665, + 0.013551734387874603, + 0.004308377858251333, + 0.002527043456211686, + 0.0037905999924987555, + 0.0021837095264345407, + 0.0027603269554674625, + 0.0018502947641536593, + 0.0021235765889286995, + 0.0032624423038214445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.023290520533919334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556695342063904 + ], + "opposite_logprob_tokens": [ + -1.6005669832229614 + ], + "logprob_reward": -0.22556695342063904, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556695342063904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06636011600494385, + 0.07703080028295517, + 0.050731923431158066, + 0.051886092871427536, + 0.08735091239213943, + 0.08477157354354858, + 0.36740928888320923, + 0.11533621698617935, + 0.05417467653751373, + 0.044948410242795944 + ], + "credit_map": [ + 0.009096239693462849, + 0.004777099471539259, + 0.009946711361408234, + 0.0070585994981229305, + 0.007722413167357445, + 0.01015370711684227, + 0.009473667480051517, + 0.008131679147481918, + 0.009870003908872604, + 0.006459137890487909, + 0.012491998262703419, + 0.008688691072165966, + 0.007419212721288204, + 0.008023003116250038, + 0.01240334752947092, + 0.01167540904134512, + 0.0060085332952439785, + 0.005025365389883518, + 0.01052284985780716, + 0.005663150921463966, + 0.005185027606785297, + 0.006468150764703751, + 0.0053502898663282394, + 0.006508554797619581, + 0.0056693060323596, + 0.004850567784160376, + 0.00753833306953311, + 0.005923764314502478, + 0.0069694216363132, + 0.006076438818126917, + 0.0069424607791006565, + 0.007915796712040901, + 0.007891629822552204, + 0.011912626214325428, + 0.011717277579009533, + 0.010728998109698296, + 0.012022105976939201, + 0.007497549057006836, + 0.016850287094712257, + 0.008730433881282806, + 0.00829114206135273, + 0.008759697899222374, + 0.010343202389776707, + 0.013385922648012638, + 0.01169732678681612, + 0.00796736590564251, + 0.01152098923921585, + 0.012805928476154804, + 0.05117025226354599, + 0.01456679217517376, + 0.028006862848997116, + 0.010579734109342098, + 0.022380933165550232, + 0.10390782356262207, + 0.024673083797097206, + 0.11212379485368729, + 0.019695639610290527, + 0.01422030571848154, + 0.009997345507144928, + 0.017538564279675484, + 0.01592301018536091, + 0.009277740493416786, + 0.017924118787050247, + 0.010759493336081505, + 0.010429282672703266, + 0.0055305249989032745, + 0.008567803539335728, + 0.0061079529114067554, + 0.005397533532232046, + 0.004556708969175816, + 0.007302809506654739, + 0.006282060407102108, + 0.01083091739565134, + 0.003534759394824505, + 0.004454452078789473, + 0.0027586957439780235, + 0.0035020180512219667, + 0.0042815436609089375, + 0.004468858242034912, + 0.01111716777086258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.22556695342063904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048832416534424 + ], + "opposite_logprob_tokens": [ + -0.1798831820487976 + ], + "logprob_reward": -1.8048832416534424, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048832416534424, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10068657249212265, + 0.055366672575473785, + 0.06125057488679886, + 0.12044286727905273, + 0.14267738163471222, + 0.14199303090572357, + 0.2527294456958771, + 0.04687805473804474, + 0.029667239636182785, + 0.04830816015601158 + ], + "credit_map": [ + 0.012404551729559898, + 0.009365925565361977, + 0.012932678684592247, + 0.012455828487873077, + 0.012509464286267757, + 0.012858971953392029, + 0.015071914531290531, + 0.013087237253785133, + 0.007083081174641848, + 0.0037249920424073935, + 0.010393808595836163, + 0.006110691465437412, + 0.006307093892246485, + 0.005463784094899893, + 0.00851435773074627, + 0.0077688624151051044, + 0.006898196879774332, + 0.0063932971097528934, + 0.013112090528011322, + 0.008431772701442242, + 0.006759344600141048, + 0.006918368861079216, + 0.005744975060224533, + 0.006992531009018421, + 0.014203913509845734, + 0.011234568431973457, + 0.023592425510287285, + 0.01160630863159895, + 0.0160042867064476, + 0.012039165012538433, + 0.017365694046020508, + 0.01439649984240532, + 0.014542482793331146, + 0.019542280584573746, + 0.027197230607271194, + 0.013224898837506771, + 0.013835377991199493, + 0.017881710082292557, + 0.018717292696237564, + 0.01773611083626747, + 0.014198074117302895, + 0.017809739336371422, + 0.02901001088321209, + 0.019784172996878624, + 0.01432056538760662, + 0.013432583771646023, + 0.017131468281149864, + 0.016306424513459206, + 0.031612928956747055, + 0.010044433176517487, + 0.013530950993299484, + 0.00793384574353695, + 0.017823370173573494, + 0.07339636981487274, + 0.016698386520147324, + 0.08168916404247284, + 0.009876476600766182, + 0.005294832866638899, + 0.004500590730458498, + 0.005926708225160837, + 0.006457641255110502, + 0.00499225128442049, + 0.005307223182171583, + 0.004522331524640322, + 0.00519348680973053, + 0.003274193499237299, + 0.003475500037893653, + 0.003453002078458667, + 0.003227368462830782, + 0.002398644108325243, + 0.0041533890180289745, + 0.004491657018661499, + 0.013019006699323654, + 0.004397030919790268, + 0.0042991978116333485, + 0.0022705947048962116, + 0.00617505656555295, + 0.004553617909550667, + 0.0033003799617290497, + 0.010293274186551571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -1.8048832416534424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, majestic wide-angle shot, framed in the style of a traditional Chinese landscape painting, captures a massive bronze dragon statue, its powerful, serpentine body coiled around the highest, mist-shrouded peak of a jagged mountain range. Initially, its metallic scales gleam with a deep, polished bronze sheen, its fierce head, flowing beard, bare teeth, facing the sky. Time-lapse begins, compressing centuries into a single moment. Sun and moon streak across the sky in fleeting arcs of light, while ink-like clouds swirl over the peaks like a celestial river. The seasons shift with breathtaking regularity: snow falls and melts in an instant, while sparse, gnarled pines cling to the cliffs, flickering between the vibrant green of spring and the chill of winter. In this endless cycle, the slow and relentless process of erosion becomes clearly visible on the dragon; light moisture begins to leave marks along its form, and over time these marks transform into a vibrant green patina that visibly runs along the bronze, pools in the crevices of its intricate scales, and flows from the corners of its eyes, depicting the relentless passage of time and transforming the majestic figure into an ancient, weathered monument, a timeless mountain it commands.", + "seed": 12858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_1097_seed_12858_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11137831769883633, + "vlm_logprob": -0.11137831769883633, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.11137831769883633, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05167060345411301 + ], + "opposite_logprob_tokens": [ + -3.176670551300049 + ], + "logprob_reward": -0.05167060345411301, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05167060345411301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037525810301303864, + 0.03357033059000969, + 0.019993510097265244, + 0.013888624496757984, + 0.012598013505339622, + 0.0338638536632061, + 0.7804465889930725, + 0.051730211824178696, + 0.010756169445812702, + 0.005626819096505642 + ], + "credit_map": [ + 0.005322195123881102, + 0.0059642004780471325, + 0.0034441682510077953, + 0.005492658354341984, + 0.004463652148842812, + 0.0034442534670233727, + 0.0036846494767814875, + 0.005710037425160408, + 0.005134573206305504, + 0.0037889026571065187, + 0.004522571340203285, + 0.0048187472857534885, + 0.002900113817304373, + 0.0029010374564677477, + 0.0033589638769626617, + 0.006145424209535122, + 0.0023662042804062366, + 0.0030869143083691597, + 0.0022830027155578136, + 0.001985798357054591, + 0.0019981798250228167, + 0.0031656366772949696, + 0.0018855854868888855, + 0.0032221912406384945, + 0.0016968926647678018, + 0.0015347206499427557, + 0.0015754877822473645, + 0.0018142390763387084, + 0.0010806486243382096, + 0.0024431522469967604, + 0.0012322855181992054, + 0.002511200262233615, + 0.0016426794463768601, + 0.001384683302603662, + 0.0013695652596652508, + 0.0009516797144897282, + 0.0011109481565654278, + 0.00235987501218915, + 0.001351251034066081, + 0.002427332801744342, + 0.0037460336461663246, + 0.004302154295146465, + 0.0050323414616286755, + 0.0044156513176858425, + 0.0025525433011353016, + 0.002445886144414544, + 0.005943557247519493, + 0.005425692070275545, + 0.166795551776886, + 0.02217804081737995, + 0.023440921679139137, + 0.02512199990451336, + 0.036506034433841705, + 0.32547470927238464, + 0.019465940073132515, + 0.1614634245634079, + 0.0055222949013113976, + 0.010410922579467297, + 0.006727777887135744, + 0.005825717467814684, + 0.005820561200380325, + 0.005426103714853525, + 0.005170680116862059, + 0.006826159544289112, + 0.0013839089078828692, + 0.0015338258817791939, + 0.0013280026614665985, + 0.0012273916509002447, + 0.0012736612698063254, + 0.0012866876786574721, + 0.001221817801706493, + 0.0015008751070126891, + 0.0006651309668086469, + 0.0007329414365813136, + 0.0008187588537111878, + 0.000463992590084672, + 0.0005773066077381372, + 0.0008548169862478971, + 0.0006002764566801488, + 0.0009135958389379084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.05167060345411301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029827596619725227 + ], + "opposite_logprob_tokens": [ + -3.52982759475708 + ], + "logprob_reward": -0.029827596619725227, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029827596619725227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07127565890550613, + 0.05700039491057396, + 0.024720454588532448, + 0.025507096201181412, + 0.014425673522055149, + 0.10012341290712357, + 0.49226969480514526, + 0.16341005265712738, + 0.032257772982120514, + 0.01900981180369854 + ], + "credit_map": [ + 0.008058175444602966, + 0.012156889773905277, + 0.010955405421555042, + 0.011347574181854725, + 0.00901294220238924, + 0.00470738997682929, + 0.0070348577573895454, + 0.008002419956028461, + 0.006819071248173714, + 0.009444590657949448, + 0.007760648615658283, + 0.010390320792794228, + 0.006512181833386421, + 0.0058763111010193825, + 0.005036226939409971, + 0.005161044653505087, + 0.002756536239758134, + 0.0033382726833224297, + 0.0021075266413390636, + 0.001636217231862247, + 0.0037164336536079645, + 0.003602550597861409, + 0.0019979847129434347, + 0.005564933177083731, + 0.0031386250630021095, + 0.0026670382358133793, + 0.0025387743953615427, + 0.0019720345735549927, + 0.001798753859475255, + 0.005676208529621363, + 0.0024915486574172974, + 0.005224111955612898, + 0.0017136353999376297, + 0.0011432324536144733, + 0.002026828471571207, + 0.0008754644077271223, + 0.001160675659775734, + 0.003243160666897893, + 0.001437450642697513, + 0.0028252247720956802, + 0.01378204021602869, + 0.00802368950098753, + 0.011281199753284454, + 0.0080475565046072, + 0.00896510947495699, + 0.023615090176463127, + 0.01256220880895853, + 0.013846522197127342, + 0.10408086329698563, + 0.013965156860649586, + 0.014992744661867619, + 0.01579148694872856, + 0.02319907210767269, + 0.2043512463569641, + 0.013101808726787567, + 0.10278728604316711, + 0.021705366671085358, + 0.030432378873229027, + 0.021332500502467155, + 0.0176868699491024, + 0.017161382362246513, + 0.014676595106720924, + 0.014504507184028625, + 0.025910459458827972, + 0.0037496236618608236, + 0.005500337574630976, + 0.003986404277384281, + 0.002801677444949746, + 0.003931522835046053, + 0.004562218673527241, + 0.0035617786925286055, + 0.004164205864071846, + 0.002373110968619585, + 0.0042045703157782555, + 0.0024042425211519003, + 0.001895615947432816, + 0.002352149924263358, + 0.0015026916516944766, + 0.0020689961966127157, + 0.002208433346822858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.029827596619725227 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337059386074543 + ], + "opposite_logprob_tokens": [ + -3.408705949783325 + ], + "logprob_reward": -0.0337059386074543, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337059386074543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.004213548265397549, + 0.001564435544423759, + 0.0036321887746453285, + 0.001092435559257865, + 0.0009347927989438176, + 0.004831399768590927, + 0.9749589562416077, + 0.006147031672298908, + 0.0016561128431931138, + 0.0009690396836958826 + ], + "credit_map": [ + 0.0004687891050707549, + 0.000643944600597024, + 0.0006551900296472013, + 0.0007174703641794622, + 0.00048356331535615027, + 0.0003284791891928762, + 0.00042786935227923095, + 0.00048824248369783163, + 0.00019703742873389274, + 0.00024220062186941504, + 0.00018717171042226255, + 0.00021822945564053953, + 0.00015038867422845215, + 0.00018685699615161866, + 0.0001782658655429259, + 0.00020428502466529608, + 0.00044963363325223327, + 0.0003675743064377457, + 0.00028383947210386395, + 0.00025294272927567363, + 0.0004740556760225445, + 0.0006585838855244219, + 0.000315859098918736, + 0.0008297000895254314, + 0.00014080801338423043, + 9.989742829930037e-05, + 0.00010560181544860825, + 5.685384167009033e-05, + 8.875747880665585e-05, + 0.0002389333094470203, + 0.000113813322968781, + 0.00024777042563073337, + 8.630754018668085e-05, + 0.00013369259249884635, + 0.0001227817847393453, + 0.0001435716258129105, + 8.75393016031012e-05, + 8.509292092639953e-05, + 0.00014236773131415248, + 0.0001334393600700423, + 0.0008569539058953524, + 0.0005075232475064695, + 0.00038815324660390615, + 0.00041332675027661026, + 0.0004618879174813628, + 0.0007781037129461765, + 0.0005483372951857746, + 0.0008771138964220881, + 0.21004286408424377, + 0.02846129611134529, + 0.029141876846551895, + 0.0310536976903677, + 0.045731909573078156, + 0.407398521900177, + 0.024308055639266968, + 0.1988208144903183, + 0.0008207427454181015, + 0.0011387785198166966, + 0.0007935046451166272, + 0.0006666443077847362, + 0.0006403079023584723, + 0.0005689901299774647, + 0.0005015652277506888, + 0.0010164988925680518, + 0.00024727333220653236, + 0.00017697674047667533, + 0.00019044539658352733, + 0.00016497375327162445, + 0.00017616736295167357, + 0.0002688695094548166, + 0.0001510808797320351, + 0.0002803260285872966, + 0.00011263840133324265, + 0.00020030175801366568, + 0.00013602222315967083, + 9.861802391242236e-05, + 0.00011184289178345352, + 9.165686788037419e-05, + 0.00011360934149706736, + 0.00010435032163513824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.0337059386074543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813131511211395 + ], + "opposite_logprob_tokens": [ + -1.406313180923462 + ], + "logprob_reward": -0.2813131511211395, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813131511211395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05422850325703621, + 0.024796700105071068, + 0.016375472769141197, + 0.011729536578059196, + 0.011138134635984898, + 0.021977094933390617, + 0.772176206111908, + 0.05344454199075699, + 0.02816622518002987, + 0.005967568140476942 + ], + "credit_map": [ + 0.004810872487723827, + 0.008768298663198948, + 0.008293394930660725, + 0.00996019784361124, + 0.006907083559781313, + 0.004658382385969162, + 0.004982346203178167, + 0.005847925785928965, + 0.0029916963540017605, + 0.0037585939280688763, + 0.0037399844732135534, + 0.004688215907663107, + 0.002431547036394477, + 0.0027541082818061113, + 0.0017794595332816243, + 0.0026530935429036617, + 0.0018593805143609643, + 0.001296594855375588, + 0.0017892925534397364, + 0.0012367499293759465, + 0.001354483189061284, + 0.003805941203609109, + 0.0014240855816751719, + 0.0036089452914893627, + 0.0014963692519813776, + 0.001236306386999786, + 0.0012194165028631687, + 0.0008813748718239367, + 0.0010007143719121814, + 0.0025258255191147327, + 0.0011078648967668414, + 0.0022616644855588675, + 0.0009592624264769256, + 0.0017367793479934335, + 0.0014626298798248172, + 0.0017772272694855928, + 0.0011346336686983705, + 0.0011797824408859015, + 0.001393045298755169, + 0.0014947742456570268, + 0.00287289428524673, + 0.003296124516054988, + 0.002499082824215293, + 0.0029675541445612907, + 0.0018205699743703008, + 0.002669187029823661, + 0.002737423637881875, + 0.003114258637651801, + 0.16723975539207458, + 0.022046776488423347, + 0.02334527112543583, + 0.024467481300234795, + 0.03609733656048775, + 0.32222071290016174, + 0.01919061690568924, + 0.1575683057308197, + 0.007995585910975933, + 0.008184983395040035, + 0.0074879382736980915, + 0.0067255860194563866, + 0.004551765508949757, + 0.004712022375315428, + 0.004046787973493338, + 0.00973987765610218, + 0.0031347889453172684, + 0.00518708536401391, + 0.0034325814340263605, + 0.002466178499162197, + 0.0033399721141904593, + 0.003786125686019659, + 0.003325552213937044, + 0.003493941854685545, + 0.0007732104859314859, + 0.001094264560379088, + 0.0007757237181067467, + 0.0004924460081383586, + 0.0008052957127802074, + 0.000817561405710876, + 0.0004965697880834341, + 0.0007124960538931191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.2813131511211395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037429869174957 + ], + "opposite_logprob_tokens": [ + -1.9103742837905884 + ], + "logprob_reward": -0.16037429869174957, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037429869174957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011299711652100086, + 0.007806272245943546, + 0.00405673636123538, + 0.0036491421051323414, + 0.00382405798882246, + 0.019066397100687027, + 0.9274924397468567, + 0.009997513145208359, + 0.00963606871664524, + 0.003171671647578478 + ], + "credit_map": [ + 0.0009654918685555458, + 0.001274386653676629, + 0.0017551402561366558, + 0.0022571543231606483, + 0.0010223397985100746, + 0.0011702771298587322, + 0.0012182279024273157, + 0.0016366931376978755, + 0.0010437027085572481, + 0.0013665042351931334, + 0.0011596613330766559, + 0.0012659543426707387, + 0.00084325997158885, + 0.0006564865470863879, + 0.0006530101527459919, + 0.0008176926057785749, + 0.0004551004385575652, + 0.0004976363270543516, + 0.0004897306789644063, + 0.0005947270547039807, + 0.0005401508533395827, + 0.0005795162287540734, + 0.00040009000804275274, + 0.0004997846554033458, + 0.0003467000788077712, + 0.0006807673489674926, + 0.00047261189320124686, + 0.0005699298344552517, + 0.0004612608754541725, + 0.00033359273220412433, + 0.00033549495856277645, + 0.00044878400512970984, + 0.00042969826608896255, + 0.0004557338252197951, + 0.00044951261952519417, + 0.0005234278505668044, + 0.0003203564556315541, + 0.0005001649260520935, + 0.00042738032061606646, + 0.0007177836378104985, + 0.003563126316294074, + 0.0015910584479570389, + 0.0016279404517263174, + 0.0016228490276262164, + 0.0015787719748914242, + 0.0026860060170292854, + 0.0029172138310968876, + 0.0034794288221746683, + 0.199186310172081, + 0.026805415749549866, + 0.02789384499192238, + 0.029549099504947662, + 0.043483879417181015, + 0.3872176706790924, + 0.02310638315975666, + 0.19024984538555145, + 0.0011998177506029606, + 0.0010410184040665627, + 0.0017767174867913127, + 0.0013526531402021646, + 0.0006594907026737928, + 0.0013135309563949704, + 0.0007400367176160216, + 0.001914247521199286, + 0.0010460857301950455, + 0.0013412920525297523, + 0.0018299114890396595, + 0.0011079636169597507, + 0.000977306510321796, + 0.0009267028071917593, + 0.001051644328981638, + 0.0013551615411415696, + 0.0003655011532828212, + 0.00031610974110662937, + 0.0005203671171329916, + 0.0002682232588995248, + 0.0002802348753903061, + 0.0004942616797052324, + 0.0003493941330816597, + 0.0005775794270448387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.16037429869174957 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, majestic wide-angle shot, framed in the style of a traditional Chinese landscape painting, captures a massive bronze dragon statue, its powerful, serpentine body coiled around the highest, mist-shrouded peak of a jagged mountain range. Initially, its metallic scales gleam with a deep, polished bronze sheen, its fierce head, flowing beard, bare teeth, facing the sky. Time-lapse begins, compressing centuries into a single moment. Sun and moon streak across the sky in fleeting arcs of light, while ink-like clouds swirl over the peaks like a celestial river. The seasons shift with breathtaking regularity: snow falls and melts in an instant, while sparse, gnarled pines cling to the cliffs, flickering between the vibrant green of spring and the chill of winter. In this endless cycle, the slow and relentless process of erosion becomes clearly visible on the dragon; light moisture begins to leave marks along its form, and over time these marks transform into a vibrant green patina that visibly runs along the bronze, pools in the crevices of its intricate scales, and flows from the corners of its eyes, depicting the relentless passage of time and transforming the majestic figure into an ancient, weathered monument, a timeless mountain it commands.", + "seed": 12859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_1097_seed_12859_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18701693341135978, + "vlm_logprob": -0.18701693341135978, + "vlm_yes_no_margin": 2.1499999880790712, + "vlm_token_credit": -0.18701693341135978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05545752868056297 + ], + "opposite_logprob_tokens": [ + -3.055457592010498 + ], + "logprob_reward": -0.05545752868056297, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05545752868056297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23543067276477814, + 0.11909478902816772, + 0.06610647588968277, + 0.12182999402284622, + 0.03892771527171135, + 0.028049372136592865, + 0.026996128261089325, + 0.0766947865486145, + 0.13249677419662476, + 0.15437327325344086 + ], + "credit_map": [ + 0.026862671598792076, + 0.024304356426000595, + 0.03613942861557007, + 0.023559831082820892, + 0.02035929635167122, + 0.024962635710835457, + 0.040145158767700195, + 0.03909730166196823, + 0.011543111875653267, + 0.022526763379573822, + 0.015397794544696808, + 0.017513955011963844, + 0.0130795668810606, + 0.009172831661999226, + 0.016019925475120544, + 0.013840838335454464, + 0.005479019600898027, + 0.009085891768336296, + 0.008466285653412342, + 0.009364032186567783, + 0.008197789080440998, + 0.006968965288251638, + 0.010885351337492466, + 0.0076591381803154945, + 0.009975103661417961, + 0.016108782961964607, + 0.01609342359006405, + 0.026087487116456032, + 0.009421407245099545, + 0.009877263568341732, + 0.023044219240546227, + 0.01122230663895607, + 0.0046306392177939415, + 0.004966145846992731, + 0.005023730918765068, + 0.004868031945079565, + 0.004816575441509485, + 0.004628265276551247, + 0.005059275310486555, + 0.004935051314532757, + 0.002737032948061824, + 0.0024418660905212164, + 0.005569028668105602, + 0.0038099794182926416, + 0.003393829334527254, + 0.003832409856840968, + 0.003173444652929902, + 0.0030917814001441, + 0.0026256730780005455, + 0.003347738878801465, + 0.0034791170619428158, + 0.004433934576809406, + 0.0028498682659119368, + 0.00279424199834466, + 0.0033061823341995478, + 0.004159372299909592, + 0.008941702544689178, + 0.010364565066993237, + 0.008257306180894375, + 0.008052735589444637, + 0.009863853454589844, + 0.007690588477998972, + 0.00838091317564249, + 0.015143118798732758, + 0.019442757591605186, + 0.014782479964196682, + 0.016189534217119217, + 0.013345425017178059, + 0.015234584920108318, + 0.020436497405171394, + 0.011298752389848232, + 0.02176673710346222, + 0.03598680719733238, + 0.008640163578093052, + 0.010852952487766743, + 0.008487315848469734, + 0.01306027453392744, + 0.05750272050499916, + 0.006018823944032192, + 0.013824215158820152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.05545752868056297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983870357275009 + ], + "opposite_logprob_tokens": [ + -3.529838800430298 + ], + "logprob_reward": -0.02983870357275009, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983870357275009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18520788848400116, + 0.09645416587591171, + 0.1153007224202156, + 0.055938661098480225, + 0.0648244321346283, + 0.07742509245872498, + 0.06806397438049316, + 0.07138772308826447, + 0.19458968937397003, + 0.07080770283937454 + ], + "credit_map": [ + 0.018732571974396706, + 0.017326412722468376, + 0.025842461735010147, + 0.01753995567560196, + 0.026641322299838066, + 0.019836828112602234, + 0.02690497227013111, + 0.03238334879279137, + 0.010230776853859425, + 0.010156642645597458, + 0.019044233486056328, + 0.012878398410975933, + 0.010828706435859203, + 0.00908400397747755, + 0.012096334248781204, + 0.012135064229369164, + 0.009199636988341808, + 0.010980707593262196, + 0.017543096095323563, + 0.014792905189096928, + 0.010970535688102245, + 0.011572610586881638, + 0.020256396383047104, + 0.01998482272028923, + 0.004777923226356506, + 0.006441899575293064, + 0.011581260710954666, + 0.008114935830235481, + 0.00541066424921155, + 0.004909934941679239, + 0.007243921514600515, + 0.007458114065229893, + 0.0084158256649971, + 0.0055162482894957066, + 0.009318801574409008, + 0.005517061334103346, + 0.009703584015369415, + 0.010924573987722397, + 0.006667113397270441, + 0.008761215955018997, + 0.008682575076818466, + 0.004460862837731838, + 0.010323265567421913, + 0.010400256142020226, + 0.012194392271339893, + 0.01042486447840929, + 0.011319348588585854, + 0.009619519114494324, + 0.007305987179279327, + 0.005127682816237211, + 0.01300547830760479, + 0.0071923742070794106, + 0.008954628370702267, + 0.008190239779651165, + 0.008771738037467003, + 0.009515844285488129, + 0.009505638852715492, + 0.006556806154549122, + 0.00865147914737463, + 0.0059395721182227135, + 0.007650952320545912, + 0.007592128124088049, + 0.006316705606877804, + 0.01917443983256817, + 0.03591512143611908, + 0.01805236004292965, + 0.016215693205595016, + 0.015980832278728485, + 0.023526040837168694, + 0.04287179931998253, + 0.01845536381006241, + 0.023572465404868126, + 0.014081176370382309, + 0.007264947984367609, + 0.008474186062812805, + 0.006048324517905712, + 0.007521283812820911, + 0.012751094996929169, + 0.005610055755823851, + 0.00905662216246128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.02983870357275009 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204582005739212 + ], + "opposite_logprob_tokens": [ + -2.8120458126068115 + ], + "logprob_reward": -0.06204582005739212, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204582005739212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1859900951385498, + 0.07834014296531677, + 0.037140678614377975, + 0.07089084386825562, + 0.07502970844507217, + 0.1113017201423645, + 0.08805970102548599, + 0.08820900321006775, + 0.0889679342508316, + 0.17607012391090393 + ], + "credit_map": [ + 0.018302427604794502, + 0.016138853505253792, + 0.02507174387574196, + 0.024057934060692787, + 0.017162013798952103, + 0.016105573624372482, + 0.033009469509124756, + 0.03614209219813347, + 0.006630796939134598, + 0.01118330005556345, + 0.011905898340046406, + 0.009464566595852375, + 0.007046809419989586, + 0.008940095081925392, + 0.013611000031232834, + 0.009557672776281834, + 0.003703850321471691, + 0.004285954870283604, + 0.006400666665285826, + 0.005276055075228214, + 0.003313567955046892, + 0.003030091291293502, + 0.005540349055081606, + 0.005590145476162434, + 0.006602239795029163, + 0.006200436968356371, + 0.012042530812323093, + 0.008474905043840408, + 0.009170548059046268, + 0.006464621517807245, + 0.012395261786878109, + 0.009540294297039509, + 0.009256853722035885, + 0.005667293909937143, + 0.011084244586527348, + 0.006735767237842083, + 0.008243456482887268, + 0.012285526841878891, + 0.009212610311806202, + 0.012543954886496067, + 0.012372041121125221, + 0.013740777038037777, + 0.017017610371112823, + 0.01790042594075203, + 0.01193955633789301, + 0.015201892703771591, + 0.012038345448672771, + 0.011091065593063831, + 0.010071066208183765, + 0.009768490679562092, + 0.014932211488485336, + 0.008930005133152008, + 0.010147913359105587, + 0.010700281709432602, + 0.01128125749528408, + 0.012228471226990223, + 0.012247334234416485, + 0.007299632765352726, + 0.010781189426779747, + 0.006192194297909737, + 0.009028722532093525, + 0.011080369353294373, + 0.00916525349020958, + 0.022414304316043854, + 0.014562247321009636, + 0.009929206222295761, + 0.0067492201924324036, + 0.0066635701805353165, + 0.011370625346899033, + 0.017802266404032707, + 0.006182290613651276, + 0.015708506107330322, + 0.040145162492990494, + 0.010708796791732311, + 0.01484397891908884, + 0.010691704228520393, + 0.016336647793650627, + 0.05761069431900978, + 0.011592474766075611, + 0.014140667393803596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.06204582005739212 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742857813835144 + ], + "opposite_logprob_tokens": [ + -0.9742857813835144 + ], + "logprob_reward": -0.4742857813835144, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742857813835144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16324378550052643, + 0.13619983196258545, + 0.04289945587515831, + 0.07523675262928009, + 0.09301634132862091, + 0.038848601281642914, + 0.04252680763602257, + 0.10458210110664368, + 0.14723071455955505, + 0.1562156230211258 + ], + "credit_map": [ + 0.01673887111246586, + 0.012377454899251461, + 0.03273002430796623, + 0.012662992812693119, + 0.021604029461741447, + 0.015300770290195942, + 0.02190910279750824, + 0.029920538887381554, + 0.015081735327839851, + 0.014281366020441055, + 0.02675393596291542, + 0.012634385377168655, + 0.016521502286195755, + 0.0161585733294487, + 0.015797706320881844, + 0.018970627337694168, + 0.004024960566312075, + 0.005624289624392986, + 0.006321027874946594, + 0.005031870678067207, + 0.00640752213075757, + 0.004002061672508717, + 0.004461915232241154, + 0.0070258029736578465, + 0.006888949777930975, + 0.007468173746019602, + 0.016147339716553688, + 0.009022827260196209, + 0.00757468119263649, + 0.006945830304175615, + 0.010972333140671253, + 0.010216617956757545, + 0.011296579614281654, + 0.008153089322149754, + 0.015975363552570343, + 0.008843972347676754, + 0.012361433357000351, + 0.013526692055165768, + 0.01060391217470169, + 0.01225530169904232, + 0.00528066186234355, + 0.003718993393704295, + 0.006275787949562073, + 0.003691653721034527, + 0.004574121441692114, + 0.00628889724612236, + 0.0039834072813391685, + 0.005035076756030321, + 0.004918181803077459, + 0.0038977747317403555, + 0.008298899978399277, + 0.00490071065723896, + 0.004953280091285706, + 0.004783091600984335, + 0.005304853897541761, + 0.005470009054988623, + 0.013326335698366165, + 0.010190379805862904, + 0.014405888505280018, + 0.013690381310880184, + 0.009066019207239151, + 0.012762417085468769, + 0.011705311015248299, + 0.019435366615653038, + 0.025112725794315338, + 0.010739678516983986, + 0.012196473777294159, + 0.01249031350016594, + 0.01758778840303421, + 0.03167509287595749, + 0.01894100196659565, + 0.018487635999917984, + 0.03501396253705025, + 0.009132714942097664, + 0.009106488898396492, + 0.009008646942675114, + 0.016719738021492958, + 0.056675177067518234, + 0.007462501525878906, + 0.013096380047500134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.4742857813835144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345683336257935 + ], + "opposite_logprob_tokens": [ + -1.3134567737579346 + ], + "logprob_reward": -0.31345683336257935, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31345683336257935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16813428699970245, + 0.09904799610376358, + 0.11256000399589539, + 0.04590645059943199, + 0.07365298271179199, + 0.03448863327503204, + 0.04579126089811325, + 0.09661794453859329, + 0.08383217453956604, + 0.2399682253599167 + ], + "credit_map": [ + 0.013404909521341324, + 0.027077477425336838, + 0.01779342070221901, + 0.027861030772328377, + 0.020981432870030403, + 0.01026997622102499, + 0.03110687993466854, + 0.019639169797301292, + 0.010885432362556458, + 0.012644189409911633, + 0.015191026963293552, + 0.014396293088793755, + 0.010393165983259678, + 0.0097268745303154, + 0.012805921956896782, + 0.013005100190639496, + 0.010163196362555027, + 0.010538849048316479, + 0.013879009522497654, + 0.017115745693445206, + 0.010093960911035538, + 0.011909324675798416, + 0.0231472235172987, + 0.01571270450949669, + 0.005484824068844318, + 0.003878704272210598, + 0.011305896565318108, + 0.0037155873142182827, + 0.005549273919314146, + 0.00552093330770731, + 0.0047156549990177155, + 0.005735579412430525, + 0.008105571381747723, + 0.007611182052642107, + 0.013080032542347908, + 0.00982452929019928, + 0.00829043984413147, + 0.008201687596738338, + 0.009410453960299492, + 0.009129096753895283, + 0.004495273809880018, + 0.0030164779163897038, + 0.004414689261466265, + 0.003806366352364421, + 0.003924533259123564, + 0.006306763272732496, + 0.004255952313542366, + 0.004268581047654152, + 0.005239366553723812, + 0.003984805662184954, + 0.006226351018995047, + 0.005233230534940958, + 0.004026470705866814, + 0.007398929912596941, + 0.006053475197404623, + 0.007628634572029114, + 0.012591330334544182, + 0.009567017666995525, + 0.013227819465100765, + 0.010088076815009117, + 0.010314494371414185, + 0.010810423642396927, + 0.00954447966068983, + 0.020474305376410484, + 0.014311248436570168, + 0.005298318341374397, + 0.00637470418587327, + 0.0050914352759718895, + 0.0073749287985265255, + 0.016847658902406693, + 0.00566020468249917, + 0.022873684763908386, + 0.05369102582335472, + 0.014361393637955189, + 0.02043122425675392, + 0.01209753006696701, + 0.019771844148635864, + 0.08378607034683228, + 0.012704512104392052, + 0.023124637082219124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.31345683336257935 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, majestic wide-angle shot, framed in the style of a traditional Chinese landscape painting, captures a massive bronze dragon statue, its powerful, serpentine body coiled around the highest, mist-shrouded peak of a jagged mountain range. Initially, its metallic scales gleam with a deep, polished bronze sheen, its fierce head, flowing beard, bare teeth, facing the sky. Time-lapse begins, compressing centuries into a single moment. Sun and moon streak across the sky in fleeting arcs of light, while ink-like clouds swirl over the peaks like a celestial river. The seasons shift with breathtaking regularity: snow falls and melts in an instant, while sparse, gnarled pines cling to the cliffs, flickering between the vibrant green of spring and the chill of winter. In this endless cycle, the slow and relentless process of erosion becomes clearly visible on the dragon; light moisture begins to leave marks along its form, and over time these marks transform into a vibrant green patina that visibly runs along the bronze, pools in the crevices of its intricate scales, and flows from the corners of its eyes, depicting the relentless passage of time and transforming the majestic figure into an ancient, weathered monument, a timeless mountain it commands.", + "seed": 12860, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_10/index_1097_seed_12860_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14414450526237488, + "vlm_logprob": -0.14414450526237488, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.14414450526237488, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04679756611585617 + ], + "opposite_logprob_tokens": [ + -3.296797513961792 + ], + "logprob_reward": -0.04679756611585617, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04679756611585617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12505334615707397, + 0.07584816217422485, + 0.09937752038240433, + 0.043286651372909546, + 0.04204557463526726, + 0.10134043544530869, + 0.20869046449661255, + 0.09165776520967484, + 0.11192655563354492, + 0.10077346861362457 + ], + "credit_map": [ + 0.015655938535928726, + 0.015310552902519703, + 0.01340495701879263, + 0.013687020167708397, + 0.014864996075630188, + 0.02423194795846939, + 0.010625545866787434, + 0.01727239042520523, + 0.00905209593474865, + 0.00919001642614603, + 0.008643130771815777, + 0.009502840228378773, + 0.012850532308220863, + 0.009778950363397598, + 0.009855522774159908, + 0.006975074764341116, + 0.011557908728718758, + 0.012916496954858303, + 0.01606065407395363, + 0.009900390170514584, + 0.017919395118951797, + 0.012630384415388107, + 0.006885421462357044, + 0.011506862938404083, + 0.005915143061429262, + 0.0054820217192173, + 0.005525959189981222, + 0.004506983328610659, + 0.004338157828897238, + 0.009146163240075111, + 0.004015030805021524, + 0.004357189871370792, + 0.004027353599667549, + 0.004270570352673531, + 0.004929654765874147, + 0.006627619732171297, + 0.006244518328458071, + 0.00413966691121459, + 0.006994826719164848, + 0.00481136329472065, + 0.01168012898415327, + 0.015575366094708443, + 0.011297122575342655, + 0.012189456261694431, + 0.010924316942691803, + 0.008894888684153557, + 0.01816372014582157, + 0.012615436688065529, + 0.019370289519429207, + 0.028084736317396164, + 0.02770519256591797, + 0.025370106101036072, + 0.03149642050266266, + 0.020154103636741638, + 0.03435109183192253, + 0.022158527746796608, + 0.007374383043497801, + 0.005956695880740881, + 0.01875321939587593, + 0.012782922014594078, + 0.011390832252800465, + 0.006549790035933256, + 0.0152474669739604, + 0.01360245794057846, + 0.007612317800521851, + 0.008174760267138481, + 0.0191664882004261, + 0.024657800793647766, + 0.00885248463600874, + 0.006573439575731754, + 0.022059274837374687, + 0.014829984866082668, + 0.007746384013444185, + 0.006507462356239557, + 0.015936587005853653, + 0.022174537181854248, + 0.009875030256807804, + 0.007374210748821497, + 0.01842529885470867, + 0.012733961455523968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.04679756611585617 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308178275823593 + ], + "opposite_logprob_tokens": [ + -3.168081760406494 + ], + "logprob_reward": -0.04308178275823593, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308178275823593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16679537296295166, + 0.05408661440014839, + 0.16593125462532043, + 0.02055884525179863, + 0.03365730121731758, + 0.07666785269975662, + 0.05870324373245239, + 0.14294752478599548, + 0.16360358893871307, + 0.11704839766025543 + ], + "credit_map": [ + 0.020541368052363396, + 0.02974008210003376, + 0.024169841781258583, + 0.01894606091082096, + 0.017032058909535408, + 0.024837380275130272, + 0.014942904002964497, + 0.016585679724812508, + 0.007291148882359266, + 0.004978598561137915, + 0.007068202365189791, + 0.006101213861256838, + 0.005232763011008501, + 0.010814906097948551, + 0.007583742495626211, + 0.005016037728637457, + 0.019432375207543373, + 0.02247709035873413, + 0.028641782701015472, + 0.014247984625399113, + 0.028269106522202492, + 0.02067793719470501, + 0.015098867006599903, + 0.017086109146475792, + 0.002397728618234396, + 0.0021080116275697947, + 0.002912948140874505, + 0.0021149946842342615, + 0.002734866226091981, + 0.0038627402391284704, + 0.0023041972890496254, + 0.002123358892276883, + 0.0032832894939929247, + 0.003969989717006683, + 0.004320244770497084, + 0.004185670055449009, + 0.005130719393491745, + 0.004042870365083218, + 0.0042475564405322075, + 0.004476957954466343, + 0.007862457074224949, + 0.009686866775155067, + 0.00947828684002161, + 0.008357653394341469, + 0.008993844501674175, + 0.012593060731887817, + 0.009489135816693306, + 0.010206547565758228, + 0.004820183850824833, + 0.005307183135300875, + 0.014437802135944366, + 0.006439364515244961, + 0.0074254050850868225, + 0.006269284524023533, + 0.006050996948033571, + 0.007953023537993431, + 0.011195258237421513, + 0.011540703475475311, + 0.026827383786439896, + 0.019230108708143234, + 0.018796201795339584, + 0.014281493611633778, + 0.023533664643764496, + 0.017542704939842224, + 0.012003637850284576, + 0.010853256098926067, + 0.02610044926404953, + 0.033864766359329224, + 0.01293504610657692, + 0.010613102465867996, + 0.03396371752023697, + 0.023269614204764366, + 0.009174017235636711, + 0.009141575545072556, + 0.01696489378809929, + 0.025562873110175133, + 0.011973054148256779, + 0.009758894331753254, + 0.019808616489171982, + 0.014664468355476856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.04308178275823593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292095482349396 + ], + "opposite_logprob_tokens": [ + -2.2379209995269775 + ], + "logprob_reward": -0.11292095482349396, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292095482349396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09149070829153061, + 0.036153290420770645, + 0.056991688907146454, + 0.032320667058229446, + 0.04251587763428688, + 0.08666572719812393, + 0.1265140175819397, + 0.25918030738830566, + 0.09756734222173691, + 0.17060044407844543 + ], + "credit_map": [ + 0.010383830405771732, + 0.017613034695386887, + 0.012934788130223751, + 0.009042221121490002, + 0.010080419480800629, + 0.013456366956233978, + 0.008841766975820065, + 0.009138273075222969, + 0.003483677050098777, + 0.003900787327438593, + 0.005921951029449701, + 0.005297166295349598, + 0.00567163061350584, + 0.0032706758938729763, + 0.004516027867794037, + 0.004091374110430479, + 0.007024420890957117, + 0.007667811121791601, + 0.008629210293293, + 0.004979648627340794, + 0.011518335901200771, + 0.00632466608658433, + 0.005020521581172943, + 0.005827066954225302, + 0.003504789900034666, + 0.004609644878655672, + 0.004124408587813377, + 0.004094243980944157, + 0.004615799058228731, + 0.003347387770190835, + 0.004114378709346056, + 0.003910011611878872, + 0.00473380321636796, + 0.004934732336550951, + 0.006861893925815821, + 0.004561993293464184, + 0.005962036084383726, + 0.00563351484015584, + 0.005287895444780588, + 0.004540006164461374, + 0.008049022406339645, + 0.00910742674022913, + 0.00990806333720684, + 0.01169003825634718, + 0.008097766898572445, + 0.0111954715102911, + 0.015261860564351082, + 0.013356068171560764, + 0.010795592330396175, + 0.010441035032272339, + 0.03395944833755493, + 0.014982400462031364, + 0.01398595329374075, + 0.010375866666436195, + 0.01329862605780363, + 0.018675098195672035, + 0.024664312601089478, + 0.02177862823009491, + 0.0409916490316391, + 0.03806247562170029, + 0.031363584101200104, + 0.021125636994838715, + 0.04252345487475395, + 0.038670532405376434, + 0.007245422340929508, + 0.006856053601950407, + 0.014335897751152515, + 0.021174658089876175, + 0.006863652262836695, + 0.007157475687563419, + 0.021023251116275787, + 0.01291092112660408, + 0.01310709398239851, + 0.010113262571394444, + 0.026764478534460068, + 0.03987293317914009, + 0.016358258202672005, + 0.01244171429425478, + 0.029718415811657906, + 0.022224294021725655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.11292095482349396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903422951698303 + ], + "opposite_logprob_tokens": [ + -2.46403431892395 + ], + "logprob_reward": -0.08903422951698303, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903422951698303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06631767749786377, + 0.031403038650751114, + 0.06746302545070648, + 0.05295024439692497, + 0.044346410781145096, + 0.04719040170311928, + 0.07503122091293335, + 0.25064587593078613, + 0.22148264944553375, + 0.14316949248313904 + ], + "credit_map": [ + 0.010287843644618988, + 0.01353226788341999, + 0.0060034096240997314, + 0.006348655093461275, + 0.005955814849585295, + 0.010893254540860653, + 0.00507524237036705, + 0.00822118017822504, + 0.0040961625054478645, + 0.00364043191075325, + 0.0033148988150060177, + 0.003438635030761361, + 0.0031941330526024103, + 0.006217897403985262, + 0.0032263039611279964, + 0.004274572711437941, + 0.007773068733513355, + 0.009255727753043175, + 0.009007447399199009, + 0.008041695691645145, + 0.009728697128593922, + 0.008984589949250221, + 0.007931803353130817, + 0.006739992648363113, + 0.005509844981133938, + 0.00553049286827445, + 0.00604616804048419, + 0.005109188612550497, + 0.008601880632340908, + 0.009036492556333542, + 0.005913767032325268, + 0.007202405482530594, + 0.0052274856716394424, + 0.006954933051019907, + 0.006753132678568363, + 0.005020462442189455, + 0.00457634124904871, + 0.005382581148296595, + 0.005250757094472647, + 0.005180712323635817, + 0.005126140080392361, + 0.005990949459373951, + 0.0060064229182899, + 0.005673404783010483, + 0.005147661082446575, + 0.006559690460562706, + 0.006406389642506838, + 0.006279739551246166, + 0.006812368985265493, + 0.005784007254987955, + 0.020304309204220772, + 0.008054611273109913, + 0.006867977324873209, + 0.005274500697851181, + 0.009935335256159306, + 0.011998113244771957, + 0.01867716573178768, + 0.01506996899843216, + 0.05528805032372475, + 0.03217611461877823, + 0.03138267993927002, + 0.020949730649590492, + 0.03958497196435928, + 0.03751718997955322, + 0.015940867364406586, + 0.014213740825653076, + 0.031805939972400665, + 0.051713377237319946, + 0.017984570935368538, + 0.01600305549800396, + 0.045340653508901596, + 0.028480427339673042, + 0.010694059543311596, + 0.012120231986045837, + 0.019995249807834625, + 0.033294517546892166, + 0.011400743387639523, + 0.012923362664878368, + 0.024999380111694336, + 0.017741940915584564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.08903422951698303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze dragon statue with a serpentine body coiled around a mountain peak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video framed in the style of a traditional Chinese landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a time-lapse effect where sun and moon streak across the sky and seasons shift rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon's surface visibly transforming from a polished bronze sheen to a vibrant green patina due to erosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static wide-angle view of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288879930973053 + ], + "opposite_logprob_tokens": [ + -1.053887963294983 + ], + "logprob_reward": -0.4288879930973053, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288879930973053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07042039930820465, + 0.06443079560995102, + 0.20318545401096344, + 0.019844116643071175, + 0.01341970730572939, + 0.026289641857147217, + 0.07578208297491074, + 0.10769452154636383, + 0.3858144283294678, + 0.03311874344944954 + ], + "credit_map": [ + 0.008890791796147823, + 0.009562565945088863, + 0.008667969144880772, + 0.007550753187388182, + 0.008286858908832073, + 0.01236305758357048, + 0.0073357997462153435, + 0.007762614171952009, + 0.0074348049238324165, + 0.007513477001339197, + 0.01293968316167593, + 0.006973159499466419, + 0.009139086119830608, + 0.006368509028106928, + 0.007395041175186634, + 0.006667036563158035, + 0.024171218276023865, + 0.031086420640349388, + 0.03739679977297783, + 0.018707405775785446, + 0.03201468661427498, + 0.02313978783786297, + 0.018699193373322487, + 0.017969967797398567, + 0.0020897292997688055, + 0.0021692882291972637, + 0.0021344772540032864, + 0.002178588882088661, + 0.0031189818400889635, + 0.003635184606537223, + 0.0021013508085161448, + 0.0024165171198546886, + 0.001466405694372952, + 0.0015350081957876682, + 0.002212997991591692, + 0.0018791137263178825, + 0.001426075934432447, + 0.0014710521791130304, + 0.0017249069642275572, + 0.0017041475512087345, + 0.002867539646103978, + 0.0026592090725898743, + 0.002679943572729826, + 0.0036003347486257553, + 0.0032005212269723415, + 0.004743887111544609, + 0.0031091824639588594, + 0.003429025411605835, + 0.006197186186909676, + 0.006043338682502508, + 0.0205705426633358, + 0.009124374948441982, + 0.0076486640609800816, + 0.005467146635055542, + 0.008286260068416595, + 0.012444572523236275, + 0.011179226450622082, + 0.013743509538471699, + 0.017049210146069527, + 0.013957495801150799, + 0.014035198837518692, + 0.013207403011620045, + 0.01164294220507145, + 0.012879549525678158, + 0.028378644958138466, + 0.024770516902208328, + 0.06504590809345245, + 0.08558711409568787, + 0.027091393247246742, + 0.023142436519265175, + 0.07850418984889984, + 0.05329426750540733, + 0.002672476228326559, + 0.0026529189199209213, + 0.004694829694926739, + 0.007381688803434372, + 0.0024367808364331722, + 0.003334975801408291, + 0.005639214534312487, + 0.004305860958993435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.4288879930973053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_11/index_tensor([472, 472])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_11/index_tensor([472, 472])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c7f9189c980dc25f4df9956e9f17294a2f725432 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_11/index_tensor([472, 472])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Rendered in hyperrealistic 3D animation, the film captures the intense culinary transformation in dramatic slow motion, with extreme close-up shots. The static camera focuses on a perfectly spherical drop of golden amber seasoning oil. As it descends to the scorching cast-iron grill below, its surface shimmers, revealing tiny flecks of suspended black vanilla. The grill's surface is a rough, matte-black metal landscape, its raised ribs a menacing cherry red from the intense heat. Visible, shimmering waves of heat radiate upward from the metal, distorting the air and creating a tangible sensation of intense heat. When the drop makes contact in exquisite slow motion, it vaporizes and evaporates instantly upon impact. Instead of exploding into a gentle puff of smoke, it forms a dramatic, billowing column of dense white smoke that expands violently upward and outward, swirling and churning in complex turbulent patterns. The entire scene is illuminated from below by incandescent metal, casting a powerful, warm glow that accentuates the translucent quality of the falling drop and the voluminous textures of the smoke, creating a visceral and powerful atmosphere of focused, elemental transformation.", + "seed": 29700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_11/index_472_seed_29700_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11655570529401302, + "vlm_logprob": -0.11655570529401302, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.11655570529401302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08661487698554993 + ], + "opposite_logprob_tokens": [ + -2.5866148471832275 + ], + "logprob_reward": -0.08661487698554993, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08661487698554993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29787567257881165, + 0.0940907821059227, + 0.05704016238451004, + 0.05629533901810646, + 0.03742814064025879, + 0.09542527049779892, + 0.0699758380651474, + 0.10243290662765503, + 0.15381401777267456, + 0.035621874034404755 + ], + "credit_map": [ + 0.037426989525556564, + 0.028336964547634125, + 0.03841852396726608, + 0.03503643348813057, + 0.046251554042100906, + 0.03653169423341751, + 0.03131844103336334, + 0.04455511271953583, + 0.007625334430485964, + 0.008122725412249565, + 0.012039998546242714, + 0.018794456496834755, + 0.012258639559149742, + 0.009569969028234482, + 0.012392486445605755, + 0.01328718475997448, + 0.007405494339764118, + 0.006255460437387228, + 0.004367968533188105, + 0.006527825724333525, + 0.009584717452526093, + 0.008271458558738232, + 0.006591032259166241, + 0.00803621020168066, + 0.008765067905187607, + 0.004254844971001148, + 0.0075946589931845665, + 0.004926952533423901, + 0.00702048372477293, + 0.006128154695034027, + 0.007642501499503851, + 0.009962677024304867, + 0.005576740019023418, + 0.0024498908314853907, + 0.006480494048446417, + 0.0029075895436108112, + 0.004554776940494776, + 0.006075191777199507, + 0.003544859355315566, + 0.005838599056005478, + 0.011557088233530521, + 0.013862013816833496, + 0.010330201126635075, + 0.012740414589643478, + 0.01249037403613329, + 0.010418755933642387, + 0.012450168840587139, + 0.011576257646083832, + 0.008966639637947083, + 0.007215417455881834, + 0.008898870088160038, + 0.00804983451962471, + 0.008941524662077427, + 0.009590741246938705, + 0.009440061636269093, + 0.008872750215232372, + 0.008616982027888298, + 0.018121780827641487, + 0.006054587662220001, + 0.007129158359020948, + 0.03203655034303665, + 0.008577429689466953, + 0.006522082723677158, + 0.015374342910945415, + 0.011794577352702618, + 0.04557133838534355, + 0.009521638974547386, + 0.015663975849747658, + 0.03214263543486595, + 0.011320480145514011, + 0.016273997724056244, + 0.011525389738380909, + 0.004280745051801205, + 0.00393869960680604, + 0.008644073270261288, + 0.0032644392922520638, + 0.0030347337014973164, + 0.004283625166863203, + 0.0032187493052333593, + 0.004956812597811222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.08661487698554993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820306226611137 + ], + "opposite_logprob_tokens": [ + -3.529820203781128 + ], + "logprob_reward": -0.029820306226611137, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820306226611137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17589738965034485, + 0.09962321072816849, + 0.2216193675994873, + 0.12489032745361328, + 0.06482579559087753, + 0.07113903760910034, + 0.04391384869813919, + 0.07528311014175415, + 0.07417663186788559, + 0.04863131046295166 + ], + "credit_map": [ + 0.020904934033751488, + 0.013299836777150631, + 0.027015121653676033, + 0.021787848323583603, + 0.021673357114195824, + 0.02170450985431671, + 0.016808250918984413, + 0.032703518867492676, + 0.014733239077031612, + 0.008157560601830482, + 0.012812497094273567, + 0.012492071837186813, + 0.008668753318488598, + 0.012306256219744682, + 0.012520433403551579, + 0.01793239265680313, + 0.028192341327667236, + 0.016990331932902336, + 0.033441778272390366, + 0.024261975660920143, + 0.016018450260162354, + 0.031263597309589386, + 0.021186262369155884, + 0.0502646341919899, + 0.02025853656232357, + 0.006185100879520178, + 0.024344557896256447, + 0.01058612298220396, + 0.00839583296328783, + 0.01593027077615261, + 0.012419324368238449, + 0.02677057310938835, + 0.010533427819609642, + 0.003974195569753647, + 0.01441260240972042, + 0.005011482629925013, + 0.005244402214884758, + 0.008004658855497837, + 0.005791928619146347, + 0.011853089556097984, + 0.011409713886678219, + 0.007559735327959061, + 0.006152193993330002, + 0.007976899854838848, + 0.008432512171566486, + 0.012692841701209545, + 0.008372332900762558, + 0.00854280311614275, + 0.005455038044601679, + 0.003806145628914237, + 0.0083343256264925, + 0.004553273785859346, + 0.005960673559457064, + 0.005617404356598854, + 0.0054007722064852715, + 0.0047862110659480095, + 0.005757872015237808, + 0.013922552578151226, + 0.004564329516142607, + 0.003263491438701749, + 0.025220364332199097, + 0.005998759530484676, + 0.0039765131659805775, + 0.012579222209751606, + 0.006171283312141895, + 0.020335577428340912, + 0.005650733597576618, + 0.008300856687128544, + 0.013978028669953346, + 0.005065402016043663, + 0.007027474232017994, + 0.007647271268069744, + 0.004658518824726343, + 0.006673308555036783, + 0.00801008939743042, + 0.004768750164657831, + 0.005378318019211292, + 0.006276008207350969, + 0.006472205743193626, + 0.006394107826054096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.029820306226611137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289098083972931 + ], + "opposite_logprob_tokens": [ + -1.0539097785949707 + ], + "logprob_reward": -0.4289098083972931, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289098083972931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1418333351612091, + 0.11664626747369766, + 0.08080976456403732, + 0.06340084969997406, + 0.18789689242839813, + 0.05849352106451988, + 0.05031618848443031, + 0.07938608527183533, + 0.1515451818704605, + 0.0696718841791153 + ], + "credit_map": [ + 0.018255850300192833, + 0.013087690807878971, + 0.01750989258289337, + 0.012824783101677895, + 0.02252868376672268, + 0.017189469188451767, + 0.010990902781486511, + 0.029446061700582504, + 0.01092181634157896, + 0.008065943606197834, + 0.01686050556600094, + 0.021484218537807465, + 0.011101918295025826, + 0.014121140353381634, + 0.014133135788142681, + 0.01995759643614292, + 0.011425202712416649, + 0.008942815475165844, + 0.008286540396511555, + 0.00988048780709505, + 0.011613738723099232, + 0.009944580495357513, + 0.010615343227982521, + 0.01010105013847351, + 0.008497006259858608, + 0.0033733348827809095, + 0.009562634862959385, + 0.006140050478279591, + 0.005680281203240156, + 0.008565669879317284, + 0.009280133061110973, + 0.012301742099225521, + 0.03127725049853325, + 0.007862645201385021, + 0.05072277784347534, + 0.014708240516483784, + 0.013913780450820923, + 0.019044457003474236, + 0.014305289834737778, + 0.03606245666742325, + 0.007337911985814571, + 0.009538063779473305, + 0.004824968986213207, + 0.006112787406891584, + 0.006852560676634312, + 0.007932884618639946, + 0.006363816559314728, + 0.009530529379844666, + 0.00544318649917841, + 0.00491140503436327, + 0.007991153746843338, + 0.005891671869903803, + 0.008060281164944172, + 0.005605423357337713, + 0.005457655061036348, + 0.006955410819500685, + 0.006433088798075914, + 0.013752387836575508, + 0.005599721800535917, + 0.005953576415777206, + 0.02310233749449253, + 0.007628297433257103, + 0.004951881244778633, + 0.011964797973632812, + 0.011853822506964207, + 0.04584271088242531, + 0.008984771557152271, + 0.009834174066781998, + 0.033702071756124496, + 0.012953276745975018, + 0.01800292730331421, + 0.010371437296271324, + 0.00640154629945755, + 0.011263341642916203, + 0.011386930011212826, + 0.008141933009028435, + 0.008093290030956268, + 0.007613418623805046, + 0.007358180359005928, + 0.009413240477442741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.4289098083972931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629659026861191 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629659026861191, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629659026861191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07872748374938965, + 0.04460838809609413, + 0.06248690187931061, + 0.04502946883440018, + 0.04928598552942276, + 0.1934482306241989, + 0.16226112842559814, + 0.1397571563720703, + 0.12919731438159943, + 0.09519796818494797 + ], + "credit_map": [ + 0.009370691142976284, + 0.00799176748842001, + 0.00987374596297741, + 0.00910459365695715, + 0.013874845579266548, + 0.008851501159369946, + 0.007684712298214436, + 0.011975624598562717, + 0.0041299620643258095, + 0.004355739802122116, + 0.005048674531280994, + 0.0072098784148693085, + 0.006182990036904812, + 0.004835055209696293, + 0.006159876007586718, + 0.006686214357614517, + 0.006984970066696405, + 0.005516256205737591, + 0.007022024132311344, + 0.009262945502996445, + 0.008025710470974445, + 0.008286669850349426, + 0.004666042979806662, + 0.01272228080779314, + 0.007096814922988415, + 0.003218947909772396, + 0.007830473594367504, + 0.005074750632047653, + 0.006152631714940071, + 0.004493224900215864, + 0.005801191087812185, + 0.005361434072256088, + 0.007054962683469057, + 0.0035499606747180223, + 0.006229964550584555, + 0.004178139381110668, + 0.007784159388393164, + 0.007764365524053574, + 0.005030305124819279, + 0.0076941270381212234, + 0.017649203538894653, + 0.030534742400050163, + 0.01679893583059311, + 0.022656982764601707, + 0.024608615785837173, + 0.024697760120034218, + 0.025753764435648918, + 0.03074822761118412, + 0.018002858385443687, + 0.015212561935186386, + 0.01997256651520729, + 0.018836835399270058, + 0.023698123171925545, + 0.021914774551987648, + 0.02448328398168087, + 0.020140130072832108, + 0.013401124626398087, + 0.023025499656796455, + 0.013560608960688114, + 0.010456420481204987, + 0.03567308932542801, + 0.01326335035264492, + 0.008854638785123825, + 0.021522430703043938, + 0.00941022764891386, + 0.03717930242419243, + 0.00881497748196125, + 0.010021143592894077, + 0.026439011096954346, + 0.01187907438725233, + 0.015255698934197426, + 0.010197880677878857, + 0.008071144111454487, + 0.016075853258371353, + 0.01096217893064022, + 0.01233243104070425, + 0.011407122947275639, + 0.014193164184689522, + 0.012887297198176384, + 0.009268778376281261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.007629659026861191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980387583374977 + ], + "opposite_logprob_tokens": [ + -3.529803991317749 + ], + "logprob_reward": -0.02980387583374977, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980387583374977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14977705478668213, + 0.21501757204532623, + 0.10768283903598785, + 0.06016004458069801, + 0.06967638432979584, + 0.0570467934012413, + 0.032989878207445145, + 0.16474393010139465, + 0.07298008352518082, + 0.06992538273334503 + ], + "credit_map": [ + 0.014633332379162312, + 0.011298579163849354, + 0.023744352161884308, + 0.018098454922437668, + 0.017491605132818222, + 0.018510721623897552, + 0.02108823135495186, + 0.024911783635616302, + 0.02510501630604267, + 0.010913289152085781, + 0.026267625391483307, + 0.035565052181482315, + 0.01719091460108757, + 0.028915857896208763, + 0.025755058974027634, + 0.04530477151274681, + 0.010879622772336006, + 0.013711606152355671, + 0.014592955820262432, + 0.011277158744633198, + 0.009946199133992195, + 0.012404228560626507, + 0.008090411312878132, + 0.026780670508742332, + 0.010290739126503468, + 0.0034722567070275545, + 0.008282240480184555, + 0.004275607410818338, + 0.0050333715043962, + 0.008860223926603794, + 0.005704699549823999, + 0.014240910299122334, + 0.009960099123418331, + 0.004337850492447615, + 0.011368739418685436, + 0.00593142956495285, + 0.006362416781485081, + 0.014101662673056126, + 0.005879691336303949, + 0.01173450704663992, + 0.005179268307983875, + 0.011364554986357689, + 0.005972717422991991, + 0.007125383708626032, + 0.0068251281045377254, + 0.006344419438391924, + 0.007671513594686985, + 0.006563811097294092, + 0.004601708613336086, + 0.0026870593428611755, + 0.007347817067056894, + 0.0039115892723202705, + 0.003449266077950597, + 0.003844890045002103, + 0.0033837261144071817, + 0.003763827495276928, + 0.01303043495863676, + 0.03049786575138569, + 0.009892841801047325, + 0.007567024789750576, + 0.05552247166633606, + 0.013111068867146969, + 0.008112017065286636, + 0.027010222896933556, + 0.005135739222168922, + 0.021300040185451508, + 0.004431165289133787, + 0.005649221129715443, + 0.01615641824901104, + 0.004872992634773254, + 0.008458837866783142, + 0.006975676864385605, + 0.0057747857645154, + 0.010235392488539219, + 0.010025864467024803, + 0.007678295485675335, + 0.008555864915251732, + 0.011666909791529179, + 0.007843637838959694, + 0.008144642226397991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.02980387583374977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in hyperrealistic 3D animation, the film captures the intense culinary transformation in dramatic slow motion, with extreme close-up shots. The static camera focuses on a perfectly spherical drop of golden amber seasoning oil. As it descends to the scorching cast-iron grill below, its surface shimmers, revealing tiny flecks of suspended black vanilla. The grill's surface is a rough, matte-black metal landscape, its raised ribs a menacing cherry red from the intense heat. Visible, shimmering waves of heat radiate upward from the metal, distorting the air and creating a tangible sensation of intense heat. When the drop makes contact in exquisite slow motion, it vaporizes and evaporates instantly upon impact. Instead of exploding into a gentle puff of smoke, it forms a dramatic, billowing column of dense white smoke that expands violently upward and outward, swirling and churning in complex turbulent patterns. The entire scene is illuminated from below by incandescent metal, casting a powerful, warm glow that accentuates the translucent quality of the falling drop and the voluminous textures of the smoke, creating a visceral and powerful atmosphere of focused, elemental transformation.", + "seed": 29701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_11/index_472_seed_29701_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11342077814042568, + "vlm_logprob": -0.11342077814042568, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.11342077814042568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22933268547058105 + ], + "opposite_logprob_tokens": [ + -1.604332685470581 + ], + "logprob_reward": -0.22933268547058105, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22933268547058105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05912505462765694, + 0.029259109869599342, + 0.05522986873984337, + 0.031028274446725845, + 0.07277994602918625, + 0.059165436774492264, + 0.4328416585922241, + 0.18146802484989166, + 0.030560649931430817, + 0.04854194074869156 + ], + "credit_map": [ + 0.005610990338027477, + 0.008959473110735416, + 0.007591179572045803, + 0.00607295474037528, + 0.007617018185555935, + 0.00965828262269497, + 0.005121375434100628, + 0.00849378202110529, + 0.002915271557867527, + 0.0035190435592085123, + 0.004086242523044348, + 0.003450038842856884, + 0.0037988920230418444, + 0.0038441859651356936, + 0.0032574280630797148, + 0.004388007335364819, + 0.006791131105273962, + 0.011109269224107265, + 0.0035294087138026953, + 0.002889322815462947, + 0.016391977667808533, + 0.008117248304188251, + 0.003165390808135271, + 0.0032361233606934547, + 0.0030164134223014116, + 0.005601299460977316, + 0.0026701337192207575, + 0.0030079924035817385, + 0.008253109641373158, + 0.0037416666746139526, + 0.002134376671165228, + 0.002603281056508422, + 0.009988663718104362, + 0.009463555179536343, + 0.005991247482597828, + 0.00716339610517025, + 0.014426334761083126, + 0.009010621346533298, + 0.007425020914524794, + 0.009311105124652386, + 0.004065185785293579, + 0.007069900166243315, + 0.006321864202618599, + 0.0059854635037481785, + 0.01112019456923008, + 0.0077253482304513454, + 0.010534824803471565, + 0.006342656444758177, + 0.027289286255836487, + 0.0698808804154396, + 0.035720691084861755, + 0.044321343302726746, + 0.054474469274282455, + 0.08990591764450073, + 0.07785117626190186, + 0.0333978533744812, + 0.01744043454527855, + 0.0067908745259046555, + 0.014176808297634125, + 0.02838853746652603, + 0.012261396273970604, + 0.03060532733798027, + 0.05978066101670265, + 0.01202398631721735, + 0.0021719359792768955, + 0.004053463693708181, + 0.004362368956208229, + 0.004370398353785276, + 0.0058538177981972694, + 0.0030859855469316244, + 0.0029358058236539364, + 0.003726875875145197, + 0.006200599018484354, + 0.005898030009120703, + 0.007809499744325876, + 0.005320772062987089, + 0.006080760154873133, + 0.0050398604944348335, + 0.006181363947689533, + 0.00601105485111475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.22933268547058105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811543062329292 + ], + "opposite_logprob_tokens": [ + -3.2881155014038086 + ], + "logprob_reward": -0.03811543062329292, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811543062329292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15093661844730377, + 0.06001172959804535, + 0.046435609459877014, + 0.09655696153640747, + 0.18052619695663452, + 0.07723496109247208, + 0.07464762032032013, + 0.16163550317287445, + 0.11293718963861465, + 0.03907761722803116 + ], + "credit_map": [ + 0.012770426459610462, + 0.018211383372545242, + 0.021905547007918358, + 0.017047669738531113, + 0.01756136864423752, + 0.025270409882068634, + 0.012373575009405613, + 0.025796236470341682, + 0.005685012321919203, + 0.0066621992737054825, + 0.010914039798080921, + 0.007003033068031073, + 0.006418728269636631, + 0.006389601156115532, + 0.010083179920911789, + 0.006855938117951155, + 0.004942309111356735, + 0.005851345602422953, + 0.006107908673584461, + 0.006343627814203501, + 0.00788880418986082, + 0.005408930592238903, + 0.005408835597336292, + 0.0044838497415184975, + 0.009132515639066696, + 0.009363513439893723, + 0.013522847555577755, + 0.015774426981806755, + 0.011032686568796635, + 0.013075706548988819, + 0.011658627539873123, + 0.012996639125049114, + 0.01781274564564228, + 0.032006751745939255, + 0.02158527821302414, + 0.02098226733505726, + 0.023816388100385666, + 0.021782957017421722, + 0.018160494044423103, + 0.02437930554151535, + 0.006768275052309036, + 0.010477339848876, + 0.013254323974251747, + 0.007273021154105663, + 0.008324624970555305, + 0.01028995681554079, + 0.00779750244691968, + 0.013049919158220291, + 0.006380832754075527, + 0.013715392909944057, + 0.006259863264858723, + 0.004521428141742945, + 0.012762442231178284, + 0.01824324019253254, + 0.006576136685907841, + 0.006188284605741501, + 0.013931517489254475, + 0.01406414620578289, + 0.01564030535519123, + 0.026481151580810547, + 0.017307283356785774, + 0.018363093957304955, + 0.0390913225710392, + 0.01675669476389885, + 0.006611993070691824, + 0.015912756323814392, + 0.009178252890706062, + 0.015293893404304981, + 0.032196421176195145, + 0.013253532350063324, + 0.008179115131497383, + 0.012311220169067383, + 0.004269877914339304, + 0.004963856190443039, + 0.006765265949070454, + 0.005295503884553909, + 0.004900749307125807, + 0.00482435105368495, + 0.004011921584606171, + 0.004046092741191387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.03811543062329292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136956691741943 + ], + "opposite_logprob_tokens": [ + -1.4063695669174194 + ], + "logprob_reward": -0.28136956691741943, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136956691741943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11185760051012039, + 0.057919684797525406, + 0.046916376799345016, + 0.04479965567588806, + 0.07141487300395966, + 0.05432922765612602, + 0.21730934083461761, + 0.24305656552314758, + 0.07208523899316788, + 0.08031142503023148 + ], + "credit_map": [ + 0.010175623930990696, + 0.012036485597491264, + 0.015948547050356865, + 0.011836711317300797, + 0.01699080690741539, + 0.019780315458774567, + 0.010356277227401733, + 0.0147328432649374, + 0.007107973098754883, + 0.006443118676543236, + 0.00768602080643177, + 0.006192290224134922, + 0.006156593561172485, + 0.008764141239225864, + 0.006774298381060362, + 0.008795250207185745, + 0.005330579821020365, + 0.0069899545051157475, + 0.00493147736415267, + 0.004856955260038376, + 0.007890024222433567, + 0.006239160429686308, + 0.004302482586354017, + 0.00637574540451169, + 0.0031754819210618734, + 0.006852573715150356, + 0.005341028794646263, + 0.005410061217844486, + 0.010655440390110016, + 0.0049066427163779736, + 0.0037034135311841965, + 0.004755015950649977, + 0.006846519652754068, + 0.012246651574969292, + 0.007884906604886055, + 0.007567488122731447, + 0.01168640423566103, + 0.008517402224242687, + 0.006304262671619654, + 0.010361239314079285, + 0.003460826352238655, + 0.005540300626307726, + 0.009435190819203854, + 0.005654544569551945, + 0.006623538210988045, + 0.0065685659646987915, + 0.008392187766730785, + 0.008654077537357807, + 0.017220526933670044, + 0.02743559144437313, + 0.017566382884979248, + 0.01783493533730507, + 0.04940253123641014, + 0.02986723557114601, + 0.03928543999791145, + 0.018696701154112816, + 0.020183473825454712, + 0.021316571161150932, + 0.021108195185661316, + 0.03805699571967125, + 0.022575275972485542, + 0.031176244840025902, + 0.0655522495508194, + 0.023087574169039726, + 0.004596422892063856, + 0.009249616414308548, + 0.008784662932157516, + 0.01376313716173172, + 0.014105749316513538, + 0.006326699163764715, + 0.006794444285333157, + 0.008464513346552849, + 0.007279739715158939, + 0.011720805428922176, + 0.01082660723477602, + 0.009853605180978775, + 0.011355436407029629, + 0.01019827090203762, + 0.010147655382752419, + 0.008929303847253323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.28136956691741943 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409455131739378 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.00409455131739378, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409455131739378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08577575534582138, + 0.04382941499352455, + 0.05286714807152748, + 0.03031756728887558, + 0.04843514785170555, + 0.04854775220155716, + 0.21860647201538086, + 0.22014813125133514, + 0.1402209848165512, + 0.11125160753726959 + ], + "credit_map": [ + 0.008761846460402012, + 0.008585281670093536, + 0.011675519868731499, + 0.009440883062779903, + 0.012102564796805382, + 0.01680860109627247, + 0.007889272645115852, + 0.010511794127523899, + 0.00424313684925437, + 0.005931411869823933, + 0.006283258553594351, + 0.004784686490893364, + 0.005863227881491184, + 0.006387387402355671, + 0.004157265182584524, + 0.006179042626172304, + 0.005825731437653303, + 0.007653707638382912, + 0.005231526680290699, + 0.004900029860436916, + 0.012282485142350197, + 0.008054441772401333, + 0.004712249152362347, + 0.004206981975585222, + 0.0024272697046399117, + 0.004104187246412039, + 0.003444479312747717, + 0.0035362299531698227, + 0.007010716013610363, + 0.0037329706829041243, + 0.0033668698742985725, + 0.0026948468293994665, + 0.005074831657111645, + 0.0057190400548279285, + 0.005494670942425728, + 0.004175323061645031, + 0.01028365921229124, + 0.006155385635793209, + 0.006686612032353878, + 0.004845628514885902, + 0.0034347260370850563, + 0.005313963629305363, + 0.005794692784547806, + 0.005119487643241882, + 0.009721131063997746, + 0.006082702428102493, + 0.007360581308603287, + 0.005720466375350952, + 0.014764041639864445, + 0.04247965291142464, + 0.016514193266630173, + 0.021977560594677925, + 0.02523653954267502, + 0.053455427289009094, + 0.027630798518657684, + 0.016548272222280502, + 0.02088957279920578, + 0.02725127339363098, + 0.029290001839399338, + 0.027405941858887672, + 0.025048287585377693, + 0.029444633051753044, + 0.029777169227600098, + 0.031041258946061134, + 0.007771874777972698, + 0.021348366513848305, + 0.013077386654913425, + 0.015804192051291466, + 0.035028308629989624, + 0.019448142498731613, + 0.012286689132452011, + 0.01545602548867464, + 0.008885664865374565, + 0.016042793169617653, + 0.011219979263842106, + 0.013997790403664112, + 0.0185843538492918, + 0.009988675825297832, + 0.019119758158922195, + 0.013412603177130222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.00409455131739378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22446209192276, + 0.06968674063682556, + 0.08687923848628998, + 0.07862328737974167, + 0.11154551804065704, + 0.06499763578176498, + 0.14324374496936798, + 0.07474661618471146, + 0.06395549327135086, + 0.08185959607362747 + ], + "credit_map": [ + 0.01889737695455551, + 0.024988388642668724, + 0.0289770420640707, + 0.025548076257109642, + 0.025966713204979897, + 0.037393614649772644, + 0.02618093229830265, + 0.03650997579097748, + 0.006747149862349033, + 0.009852474555373192, + 0.00866809394210577, + 0.008175864815711975, + 0.009500021114945412, + 0.009076099842786789, + 0.007986684329807758, + 0.00968035776168108, + 0.007378425914794207, + 0.01443930808454752, + 0.00761239742860198, + 0.007706762291491032, + 0.023037035018205643, + 0.012566221877932549, + 0.007192701101303101, + 0.006946391426026821, + 0.006915336009114981, + 0.00975579209625721, + 0.01139503438025713, + 0.012140247039496899, + 0.010244582779705524, + 0.007186233066022396, + 0.011422262527048588, + 0.009563800878822803, + 0.012317005544900894, + 0.02447253093123436, + 0.012126983143389225, + 0.01083463802933693, + 0.015129178762435913, + 0.012362373061478138, + 0.010853414423763752, + 0.013449404388666153, + 0.005493274889886379, + 0.008787034079432487, + 0.008342529647052288, + 0.0063664973713457584, + 0.012025446631014347, + 0.006952996831387281, + 0.008982224389910698, + 0.008047637529671192, + 0.008738035336136818, + 0.02309011109173298, + 0.015793144702911377, + 0.01564410887658596, + 0.018215321004390717, + 0.024175990372896194, + 0.026275508105754852, + 0.011311529204249382, + 0.006787281483411789, + 0.00795370526611805, + 0.007670758292078972, + 0.011447867378592491, + 0.010488261468708515, + 0.00721826171502471, + 0.013795355334877968, + 0.009385132230818272, + 0.0038987004663795233, + 0.006734748370945454, + 0.009130598977208138, + 0.011277489364147186, + 0.012533346191048622, + 0.006434309296309948, + 0.005759912077337503, + 0.008186395280063152, + 0.00953960046172142, + 0.010066287592053413, + 0.012032954953610897, + 0.009094063192605972, + 0.010497421026229858, + 0.008606713265180588, + 0.011712439358234406, + 0.010310125537216663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.01419165637344122 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in hyperrealistic 3D animation, the film captures the intense culinary transformation in dramatic slow motion, with extreme close-up shots. The static camera focuses on a perfectly spherical drop of golden amber seasoning oil. As it descends to the scorching cast-iron grill below, its surface shimmers, revealing tiny flecks of suspended black vanilla. The grill's surface is a rough, matte-black metal landscape, its raised ribs a menacing cherry red from the intense heat. Visible, shimmering waves of heat radiate upward from the metal, distorting the air and creating a tangible sensation of intense heat. When the drop makes contact in exquisite slow motion, it vaporizes and evaporates instantly upon impact. Instead of exploding into a gentle puff of smoke, it forms a dramatic, billowing column of dense white smoke that expands violently upward and outward, swirling and churning in complex turbulent patterns. The entire scene is illuminated from below by incandescent metal, casting a powerful, warm glow that accentuates the translucent quality of the falling drop and the voluminous textures of the smoke, creating a visceral and powerful atmosphere of focused, elemental transformation.", + "seed": 29702, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_11/index_472_seed_29702_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20445265974849464, + "vlm_logprob": -0.20445265974849464, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.20445265974849464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3221854269504547 + ], + "opposite_logprob_tokens": [ + -1.3221853971481323 + ], + "logprob_reward": -0.3221854269504547, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3221854269504547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2630222737789154, + 0.07248872518539429, + 0.0939466580748558, + 0.03154641017317772, + 0.13590586185455322, + 0.1738802045583725, + 0.06253072619438171, + 0.04026978835463524, + 0.11303126811981201, + 0.013378056697547436 + ], + "credit_map": [ + 0.03325902298092842, + 0.04313832148909569, + 0.03163732960820198, + 0.04400567710399628, + 0.04546138644218445, + 0.022575797513127327, + 0.02315276674926281, + 0.0197919849306345, + 0.008936158381402493, + 0.005416456609964371, + 0.005259072408080101, + 0.01481824740767479, + 0.010494278743863106, + 0.014338088221848011, + 0.007964021526277065, + 0.005262397229671478, + 0.014490882866084576, + 0.010419606231153011, + 0.008486536331474781, + 0.009818289428949356, + 0.01502606738358736, + 0.012972129508852959, + 0.013618619181215763, + 0.0091145234182477, + 0.004998786840587854, + 0.003139042528346181, + 0.0037221149541437626, + 0.00354090821929276, + 0.003232058137655258, + 0.006402015686035156, + 0.0026773768477141857, + 0.0038341055624186993, + 0.02170356549322605, + 0.01362860668450594, + 0.017592793330550194, + 0.01786590926349163, + 0.017116902396082878, + 0.024852676317095757, + 0.012210498563945293, + 0.010934915393590927, + 0.025651078671216965, + 0.023679863661527634, + 0.02176019735634327, + 0.027116699144244194, + 0.022855039685964584, + 0.01940564624965191, + 0.02020953595638275, + 0.013202141970396042, + 0.007489786483347416, + 0.008013230748474598, + 0.005329051986336708, + 0.006378029473125935, + 0.012185030616819859, + 0.008784814737737179, + 0.007849602960050106, + 0.006501177325844765, + 0.0043218787759542465, + 0.006953893695026636, + 0.0032994206994771957, + 0.004385457374155521, + 0.006224541459232569, + 0.007129161152988672, + 0.004713218659162521, + 0.003242218168452382, + 0.009341810829937458, + 0.017408009618520737, + 0.004729048814624548, + 0.012373216450214386, + 0.02702728472650051, + 0.010609763674438, + 0.018542300909757614, + 0.012999827042222023, + 0.0014061438851058483, + 0.0015698890201747417, + 0.0012088451767340302, + 0.002104351529851556, + 0.0020832186564803123, + 0.001158453058451414, + 0.0018847788451239467, + 0.001962376059964299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.3221854269504547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358967646956444 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026358967646956444, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358967646956444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7850043773651123, + 0.05077454075217247, + 0.02089226245880127, + 0.027573131024837494, + 0.005134861450642347, + 0.015190775506198406, + 0.04371240362524986, + 0.016347665339708328, + 0.026465659961104393, + 0.008904268965125084 + ], + "credit_map": [ + 0.09455226361751556, + 0.14226585626602173, + 0.09320373088121414, + 0.11592228710651398, + 0.15103617310523987, + 0.06635314971208572, + 0.06346571445465088, + 0.05820520594716072, + 0.007443342823535204, + 0.004232514183968306, + 0.009114885702729225, + 0.004879537038505077, + 0.005639794282615185, + 0.00678597204387188, + 0.005610463675111532, + 0.0070680356584489346, + 0.0038950855378061533, + 0.0015355271752923727, + 0.004759279079735279, + 0.001256483606994152, + 0.0020394299644976854, + 0.002958758734166622, + 0.0014904034323990345, + 0.0029572939965873957, + 0.005135177168995142, + 0.002079539932310581, + 0.005932772532105446, + 0.002080256352201104, + 0.0016776768025010824, + 0.004971265327185392, + 0.0017017662758007646, + 0.003994675353169441, + 0.0012274003820493817, + 0.00044098522630520165, + 0.0006406194879673421, + 0.00039928240585140884, + 0.0005235861171968281, + 0.0007666407618671656, + 0.0003927287762053311, + 0.00074361840961501, + 0.002878772094845772, + 0.0015429971972480416, + 0.00254274089820683, + 0.0017555449157953262, + 0.001392559614032507, + 0.002264905720949173, + 0.0011679315939545631, + 0.0016453225398436189, + 0.005002946592867374, + 0.00729120010510087, + 0.00327779701910913, + 0.004362388513982296, + 0.008498109877109528, + 0.005990723613649607, + 0.005384904332458973, + 0.003904331475496292, + 0.0020359426271170378, + 0.0021046984475106, + 0.0014532735804095864, + 0.0017342878272756934, + 0.001888809259980917, + 0.0016755824908614159, + 0.0025988391134887934, + 0.0028562324587255716, + 0.0020983978174626827, + 0.004020522348582745, + 0.0013624151470139623, + 0.0028634851332753897, + 0.006473455112427473, + 0.0026926773134618998, + 0.003795998403802514, + 0.003158707171678543, + 0.0007368579972535372, + 0.0014410928124561906, + 0.0006042644963599741, + 0.0013460341142490506, + 0.0016844458878040314, + 0.0011092914501205087, + 0.0009448291966691613, + 0.0010374529520049691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.026358967646956444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328393816947937 + ], + "opposite_logprob_tokens": [ + -0.7578393816947937 + ], + "logprob_reward": -0.6328393816947937, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328393816947937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6079806089401245, + 0.03213465213775635, + 0.03583384305238724, + 0.022757960483431816, + 0.09542237222194672, + 0.08027642220258713, + 0.03789367154240608, + 0.029533084481954575, + 0.0386626236140728, + 0.019504819065332413 + ], + "credit_map": [ + 0.07188095897436142, + 0.10844559222459793, + 0.07778804004192352, + 0.08818171173334122, + 0.1130317822098732, + 0.05372197926044464, + 0.04908015578985214, + 0.04585034400224686, + 0.004720048978924751, + 0.003656677668914199, + 0.005228738766163588, + 0.0037729116156697273, + 0.004445538390427828, + 0.003588021732866764, + 0.004181119613349438, + 0.0025415942072868347, + 0.004553230945020914, + 0.0034090306144207716, + 0.004314466379582882, + 0.004157993011176586, + 0.007509798277169466, + 0.004083039239048958, + 0.004347165580838919, + 0.0034591173753142357, + 0.00392333697527647, + 0.0019703079015016556, + 0.003601956646889448, + 0.0021340721286833286, + 0.0015522626927122474, + 0.00499238446354866, + 0.0017787283286452293, + 0.0028049105312675238, + 0.015143034979701042, + 0.009604368358850479, + 0.013507400639355183, + 0.012135682627558708, + 0.013138257898390293, + 0.014134678058326244, + 0.00866231881082058, + 0.00909662526100874, + 0.011599360033869743, + 0.011520739644765854, + 0.0077277678065001965, + 0.013843175955116749, + 0.012115586549043655, + 0.008251195773482323, + 0.009632518514990807, + 0.00558607280254364, + 0.004595662467181683, + 0.005469462834298611, + 0.00295926327817142, + 0.004245289601385593, + 0.00779131380841136, + 0.00423892168328166, + 0.004770600702613592, + 0.0038231562357395887, + 0.0046874783001840115, + 0.002997060539200902, + 0.0029262895695865154, + 0.006294780410826206, + 0.0025838720612227917, + 0.0023664701730012894, + 0.003881626995280385, + 0.0037955075968056917, + 0.004587632603943348, + 0.005923922639340162, + 0.0026581345591694117, + 0.00450570322573185, + 0.008705886080861092, + 0.0032843947410583496, + 0.005204528104513884, + 0.0037924188654869795, + 0.0012597523164004087, + 0.002666713669896126, + 0.0018176023149862885, + 0.0036124580074101686, + 0.0038340382743626833, + 0.001934571424499154, + 0.0024368076119571924, + 0.0019428736995905638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.6328393816947937 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107992883771658 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.01107992883771658, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107992883771658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20279552042484283, + 0.04456331208348274, + 0.06277046352624893, + 0.056468669325113297, + 0.0485856756567955, + 0.044855695217847824, + 0.18241798877716064, + 0.15195678174495697, + 0.16959840059280396, + 0.03598745912313461 + ], + "credit_map": [ + 0.028038011863827705, + 0.031037956476211548, + 0.027596574276685715, + 0.027669541537761688, + 0.03479733690619469, + 0.01890162006020546, + 0.019796527922153473, + 0.014957956969738007, + 0.006683068349957466, + 0.005866209510713816, + 0.005871289409697056, + 0.005122268572449684, + 0.005798494443297386, + 0.00567198358476162, + 0.005763160064816475, + 0.003786840010434389, + 0.008625322952866554, + 0.0051317373290658, + 0.007696287706494331, + 0.004539808724075556, + 0.013458269648253918, + 0.011843325570225716, + 0.006449282635003328, + 0.005026439670473337, + 0.009763904847204685, + 0.004485635552555323, + 0.00846760906279087, + 0.005194428842514753, + 0.0053591420873999596, + 0.011766244657337666, + 0.00407023262232542, + 0.007361475843936205, + 0.0081821633502841, + 0.004685982596129179, + 0.007637665141373873, + 0.006241563241928816, + 0.006478509400039911, + 0.00600172346457839, + 0.004655578639358282, + 0.004702489357441664, + 0.008766688406467438, + 0.0033589880913496017, + 0.007947355508804321, + 0.004711437504738569, + 0.005306681152433157, + 0.007277187891304493, + 0.0033886320888996124, + 0.004098729696124792, + 0.026245368644595146, + 0.044184159487485886, + 0.017325397580862045, + 0.01412924099713564, + 0.030062854290008545, + 0.025422971695661545, + 0.01433519646525383, + 0.010712827555835247, + 0.018485143780708313, + 0.028933821246027946, + 0.01748039573431015, + 0.01513500977307558, + 0.02221132442355156, + 0.01849774830043316, + 0.015868594869971275, + 0.0153447724878788, + 0.016646360978484154, + 0.02884131483733654, + 0.007594055030494928, + 0.019487988203763962, + 0.03996866196393967, + 0.01539277657866478, + 0.023085109889507294, + 0.018582141026854515, + 0.0031039509922266006, + 0.004880394320935011, + 0.003123028902336955, + 0.005974659696221352, + 0.005227568093687296, + 0.0034942373167723417, + 0.005065334029495716, + 0.0051182894967496395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.01107992883771658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979959361255169 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979959361255169, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979959361255169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8085187077522278, + 0.04897269234061241, + 0.017817694693803787, + 0.006502398755401373, + 0.017228420823812485, + 0.023973261937499046, + 0.013734105974435806, + 0.012793146073818207, + 0.04108051955699921, + 0.009379031136631966 + ], + "credit_map": [ + 0.09735606610774994, + 0.14582134783267975, + 0.09542213380336761, + 0.11787522584199905, + 0.15604877471923828, + 0.06946772336959839, + 0.06549850106239319, + 0.06102893874049187, + 0.008286988362669945, + 0.0034373258240520954, + 0.0066732317209243774, + 0.006106844171881676, + 0.005057636182755232, + 0.007589195854961872, + 0.005243796389549971, + 0.006577674765139818, + 0.003652190323919058, + 0.0016429214738309383, + 0.002322951564565301, + 0.0015999850584194064, + 0.002076780889183283, + 0.002727237995713949, + 0.001670756028033793, + 0.0021248720586299896, + 0.0009774016216397285, + 0.00046544100041501224, + 0.0010673266369849443, + 0.00047057803021743894, + 0.0007443822105415165, + 0.0008813844178803265, + 0.0005080985138192773, + 0.0013877860037609935, + 0.0031037849839776754, + 0.0018491812516003847, + 0.0018065216718241572, + 0.0018235493917018175, + 0.0027084534522145987, + 0.0031113745644688606, + 0.0014497331576421857, + 0.0013758232817053795, + 0.003531623864546418, + 0.0029314199928194284, + 0.002820208901539445, + 0.0034367137122899294, + 0.0032770035322755575, + 0.00355538004077971, + 0.0026722392067313194, + 0.0017486701253801584, + 0.0018470458453521132, + 0.0019935593008995056, + 0.001961679896339774, + 0.0012905659386888146, + 0.002309571485966444, + 0.0017045354470610619, + 0.0014752439456060529, + 0.0011519038816913962, + 0.0018128957599401474, + 0.0016453092684969306, + 0.0016151517629623413, + 0.0019144220277667046, + 0.0018076683627441525, + 0.0011400732910260558, + 0.0014727216912433505, + 0.0013849041424691677, + 0.0035085545387119055, + 0.00702308677136898, + 0.002507427940145135, + 0.004566228948533535, + 0.010631483979523182, + 0.003759213024750352, + 0.005495368503034115, + 0.0035891602747142315, + 0.0008992277435027063, + 0.0017659945879131556, + 0.0008908255840651691, + 0.0013767520431429148, + 0.0012848343467339873, + 0.0006927300128154457, + 0.0015095886774361134, + 0.0009590787813067436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02979959361255169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in hyperrealistic 3D animation, the film captures the intense culinary transformation in dramatic slow motion, with extreme close-up shots. The static camera focuses on a perfectly spherical drop of golden amber seasoning oil. As it descends to the scorching cast-iron grill below, its surface shimmers, revealing tiny flecks of suspended black vanilla. The grill's surface is a rough, matte-black metal landscape, its raised ribs a menacing cherry red from the intense heat. Visible, shimmering waves of heat radiate upward from the metal, distorting the air and creating a tangible sensation of intense heat. When the drop makes contact in exquisite slow motion, it vaporizes and evaporates instantly upon impact. Instead of exploding into a gentle puff of smoke, it forms a dramatic, billowing column of dense white smoke that expands violently upward and outward, swirling and churning in complex turbulent patterns. The entire scene is illuminated from below by incandescent metal, casting a powerful, warm glow that accentuates the translucent quality of the falling drop and the voluminous textures of the smoke, creating a visceral and powerful atmosphere of focused, elemental transformation.", + "seed": 29703, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_11/index_472_seed_29703_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13730702977627515, + "vlm_logprob": -0.13730702977627515, + "vlm_yes_no_margin": 2.749999988079071, + "vlm_token_credit": -0.13730702977627515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2866555452346802 + ], + "opposite_logprob_tokens": [ + -1.4116555452346802 + ], + "logprob_reward": -0.2866555452346802, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2866555452346802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28036728501319885, + 0.10203105211257935, + 0.05248606204986572, + 0.11514507979154587, + 0.11833526939153671, + 0.072019062936306, + 0.06998344510793686, + 0.07242253422737122, + 0.046090733259916306, + 0.07111949473619461 + ], + "credit_map": [ + 0.01915971376001835, + 0.014623227529227734, + 0.02334224246442318, + 0.014040513895452023, + 0.024038849398493767, + 0.10342812538146973, + 0.014684492722153664, + 0.06705010682344437, + 0.00507226400077343, + 0.00953417643904686, + 0.00541737861931324, + 0.006567050237208605, + 0.01683960296213627, + 0.027592061087489128, + 0.007868978194892406, + 0.023139536380767822, + 0.005974441766738892, + 0.005725047085434198, + 0.0062738386914134026, + 0.005204702727496624, + 0.009744237177073956, + 0.0068688662722706795, + 0.005212455987930298, + 0.007482469547539949, + 0.018733171746134758, + 0.015886997804045677, + 0.016488797962665558, + 0.0087681720033288, + 0.014602850191295147, + 0.013335938565433025, + 0.012237337417900562, + 0.015091805718839169, + 0.011774600483477116, + 0.011535354889929295, + 0.009955678135156631, + 0.011960582807660103, + 0.016081951558589935, + 0.0288649071007967, + 0.012161999940872192, + 0.016000188887119293, + 0.008820431306958199, + 0.007569595240056515, + 0.010507086291909218, + 0.0076769194565713406, + 0.007133452221751213, + 0.008962883614003658, + 0.007510133553296328, + 0.01383855752646923, + 0.0036262646317481995, + 0.009864171035587788, + 0.009727217257022858, + 0.012113107368350029, + 0.013027889654040337, + 0.00687227351590991, + 0.008341449312865734, + 0.006411067675799131, + 0.006680042017251253, + 0.006385297980159521, + 0.005720233544707298, + 0.009168007411062717, + 0.010273950174450874, + 0.0068321311846375465, + 0.019704801961779594, + 0.007658067159354687, + 0.00305470684543252, + 0.004976971074938774, + 0.005392968188971281, + 0.005132302641868591, + 0.007420971989631653, + 0.011540582403540611, + 0.004495519213378429, + 0.004076706245541573, + 0.007126353681087494, + 0.008131857961416245, + 0.006260112393647432, + 0.009786107577383518, + 0.01124005950987339, + 0.01259969174861908, + 0.010580110363662243, + 0.005395194049924612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.2866555452346802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043076302856206894 + ], + "opposite_logprob_tokens": [ + -3.168076276779175 + ], + "logprob_reward": -0.043076302856206894, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043076302856206894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2118036150932312, + 0.18169811367988586, + 0.18172232806682587, + 0.055075109004974365, + 0.0973704531788826, + 0.08641190826892853, + 0.033520083874464035, + 0.05686736851930618, + 0.02781745046377182, + 0.06771352142095566 + ], + "credit_map": [ + 0.01832457073032856, + 0.015108256600797176, + 0.04409820958971977, + 0.02037239260971546, + 0.01686684414744377, + 0.0377868227660656, + 0.021556656807661057, + 0.03768986836075783, + 0.016728898510336876, + 0.015981465578079224, + 0.026012543588876724, + 0.03143737092614174, + 0.013472937978804111, + 0.02591230347752571, + 0.026162175461649895, + 0.025990432128310204, + 0.02824472263455391, + 0.019476555287837982, + 0.023644696921110153, + 0.01592707820236683, + 0.02814324200153351, + 0.02883368358016014, + 0.013111206702888012, + 0.0243411622941494, + 0.007203073240816593, + 0.005863728001713753, + 0.01133324671536684, + 0.004729870706796646, + 0.005334520246833563, + 0.007005150429904461, + 0.006336580030620098, + 0.007268944755196571, + 0.009171863086521626, + 0.0086547015234828, + 0.008042588829994202, + 0.009892052970826626, + 0.015422692522406578, + 0.023609625175595284, + 0.008547003380954266, + 0.01402993779629469, + 0.009921335615217686, + 0.006301662418991327, + 0.018245946615934372, + 0.011125488206744194, + 0.007785371504724026, + 0.01178470067679882, + 0.006915327627211809, + 0.01433208491653204, + 0.003055892651900649, + 0.003535397117957473, + 0.006824827753007412, + 0.004086302127689123, + 0.005175123456865549, + 0.003247085027396679, + 0.003159719053655863, + 0.0044357385486364365, + 0.005003856960684061, + 0.005852566100656986, + 0.005369438324123621, + 0.00941577646881342, + 0.008946581743657589, + 0.005037609953433275, + 0.012064415961503983, + 0.005177132319658995, + 0.0025090849958360195, + 0.002312067896127701, + 0.0025351527146995068, + 0.0034025292843580246, + 0.004804262891411781, + 0.004407574888318777, + 0.0038124797865748405, + 0.004034301731735468, + 0.006651972886174917, + 0.005750314798206091, + 0.015607619658112526, + 0.008336217142641544, + 0.007652296219021082, + 0.008423620834946632, + 0.008656156249344349, + 0.00663532642647624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.043076302856206894 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134613633155823 + ], + "opposite_logprob_tokens": [ + -1.3134613037109375 + ], + "logprob_reward": -0.3134613633155823, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134613633155823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11649251729249954, + 0.1260078400373459, + 0.093777596950531, + 0.07106108218431473, + 0.14792895317077637, + 0.12336716055870056, + 0.05740925669670105, + 0.1524418592453003, + 0.03669095039367676, + 0.0748228058218956 + ], + "credit_map": [ + 0.014012851752340794, + 0.011427313089370728, + 0.013010618276894093, + 0.010843081399798393, + 0.017084253951907158, + 0.017316987738013268, + 0.017561517655849457, + 0.01523588877171278, + 0.010568127036094666, + 0.007438828703016043, + 0.014620568603277206, + 0.013107776641845703, + 0.012060683220624924, + 0.02352459542453289, + 0.02170451730489731, + 0.022982748225331306, + 0.01445679645985365, + 0.008844828233122826, + 0.011594954878091812, + 0.00774553744122386, + 0.011635363101959229, + 0.01733091101050377, + 0.007187776267528534, + 0.014981434680521488, + 0.007530116010457277, + 0.008969499729573727, + 0.011005689390003681, + 0.007412491366267204, + 0.009124954231083393, + 0.010225234553217888, + 0.008010785095393658, + 0.008782313205301762, + 0.014951424673199654, + 0.011768894270062447, + 0.019464394077658653, + 0.013537242077291012, + 0.01845713146030903, + 0.034716952592134476, + 0.014175543561577797, + 0.020857367664575577, + 0.013149067759513855, + 0.01083263661712408, + 0.014376338571310043, + 0.02906624600291252, + 0.013312741182744503, + 0.011767828837037086, + 0.016436846926808357, + 0.014425449073314667, + 0.003949317149817944, + 0.007418544963002205, + 0.005986601579934359, + 0.0072787366807460785, + 0.010176511481404305, + 0.008136136457324028, + 0.008217770606279373, + 0.0062456377781927586, + 0.008688398636877537, + 0.013880258426070213, + 0.008517701178789139, + 0.023196278139948845, + 0.022982386872172356, + 0.012030239216983318, + 0.04796627536416054, + 0.015180325135588646, + 0.0034367528278380632, + 0.0034884668421000242, + 0.0038619181141257286, + 0.004144951235502958, + 0.004891171585768461, + 0.005997891072183847, + 0.00461210822686553, + 0.006257691420614719, + 0.006271513178944588, + 0.0071901860646903515, + 0.013364624232053757, + 0.011163723655045033, + 0.01012058649212122, + 0.011253596283495426, + 0.009707455523312092, + 0.005751125980168581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.3134613633155823 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1261511743068695, + 0.09220395237207413, + 0.03828301280736923, + 0.0692613422870636, + 0.22571979463100433, + 0.22626671195030212, + 0.0464545302093029, + 0.10083611309528351, + 0.040016453713178635, + 0.03480692580342293 + ], + "credit_map": [ + 0.007861534133553505, + 0.01158717181533575, + 0.010913001373410225, + 0.009677774272859097, + 0.014356452040374279, + 0.03668683394789696, + 0.011156492866575718, + 0.023911911994218826, + 0.006989189889281988, + 0.00794204417616129, + 0.006089235655963421, + 0.005785792134702206, + 0.016234878450632095, + 0.023681210353970528, + 0.007203380577266216, + 0.0182782132178545, + 0.004319655708968639, + 0.003891525324434042, + 0.0037900826428085566, + 0.0035138351377099752, + 0.005908524617552757, + 0.008087304420769215, + 0.003709525801241398, + 0.005062558688223362, + 0.011182152666151524, + 0.007897150702774525, + 0.00737575301900506, + 0.0060094986110925674, + 0.009417809545993805, + 0.008110422641038895, + 0.01129803154617548, + 0.007970526814460754, + 0.018986383453011513, + 0.02527875266969204, + 0.029655810445547104, + 0.03080880641937256, + 0.04020768031477928, + 0.02870764210820198, + 0.030864179134368896, + 0.021210549399256706, + 0.015660742297768593, + 0.01956673339009285, + 0.024326806887984276, + 0.047139737755060196, + 0.026384735479950905, + 0.03137042000889778, + 0.03202807530760765, + 0.029789457097649574, + 0.0035983240231871605, + 0.004876351449638605, + 0.010269938968122005, + 0.006111709866672754, + 0.005410292185842991, + 0.005510139279067516, + 0.0053762709721922874, + 0.005301504395902157, + 0.005954824388027191, + 0.00763780577108264, + 0.009226977825164795, + 0.014170740731060505, + 0.012340443208813667, + 0.011447275057435036, + 0.029956357553601265, + 0.010101678781211376, + 0.004317266400903463, + 0.003664413932710886, + 0.005825902801007032, + 0.0038603730499744415, + 0.0049233767203986645, + 0.005590802524238825, + 0.00470031937584281, + 0.007133998908102512, + 0.0035681540612131357, + 0.003507592948153615, + 0.005950482562184334, + 0.004786801058799028, + 0.004366811830550432, + 0.005505173932760954, + 0.004069427493959665, + 0.0030524805188179016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.00525094848126173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of golden amber seasoning oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a rough, matte-black cast-iron grill with raised ribs glowing cherry red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil drop contain visible tiny flecks of suspended black vanilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil drop vaporize instantly upon impact to form a dense, billowing column of white smoke with turbulent patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated from below by incandescent metal, casting a warm glow on the drop and smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038090988993644714 + ], + "opposite_logprob_tokens": [ + -3.288090944290161 + ], + "logprob_reward": -0.038090988993644714, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038090988993644714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33428511023521423, + 0.11770471930503845, + 0.12108809500932693, + 0.11410250514745712, + 0.05253578722476959, + 0.06522837281227112, + 0.06771586090326309, + 0.029809240251779556, + 0.04378689080476761, + 0.05374341458082199 + ], + "credit_map": [ + 0.017132865265011787, + 0.012260007672011852, + 0.02499358169734478, + 0.01737336814403534, + 0.019959792494773865, + 0.13270927965641022, + 0.016153721138834953, + 0.09370247274637222, + 0.0111674340441823, + 0.008509689010679722, + 0.011241079308092594, + 0.006029197014868259, + 0.013079526834189892, + 0.03280262276530266, + 0.007540405262261629, + 0.02733476646244526, + 0.018443893641233444, + 0.011711716651916504, + 0.013456014916300774, + 0.011392008513212204, + 0.01561700738966465, + 0.024192223325371742, + 0.010941477492451668, + 0.015333752147853374, + 0.024285979568958282, + 0.013284978456795216, + 0.016449030488729477, + 0.00879860483109951, + 0.01078864000737667, + 0.01248399168252945, + 0.013326182961463928, + 0.014685100875794888, + 0.005448921583592892, + 0.004882861860096455, + 0.0052468315698206425, + 0.005350813269615173, + 0.008759801276028156, + 0.010248021222651005, + 0.0055808210745453835, + 0.007017714902758598, + 0.004822594113647938, + 0.005062610376626253, + 0.01327822357416153, + 0.010098523460328579, + 0.006042481400072575, + 0.01035603042691946, + 0.007316941395401955, + 0.008250972256064415, + 0.00490817055106163, + 0.006362751126289368, + 0.01939072459936142, + 0.008719407953321934, + 0.006803952157497406, + 0.0062470813281834126, + 0.007494856137782335, + 0.007788915187120438, + 0.002346957800909877, + 0.0030842002015560865, + 0.003744998946785927, + 0.004019814543426037, + 0.004339530598372221, + 0.0044725071638822556, + 0.004261150490492582, + 0.0035400802735239267, + 0.003470874158665538, + 0.0044850013218820095, + 0.003424080554395914, + 0.004428508225828409, + 0.005768482107669115, + 0.008203747682273388, + 0.006096708122640848, + 0.007909487001597881, + 0.004840519279241562, + 0.004904341418296099, + 0.008817276917397976, + 0.007391667924821377, + 0.007909484207630157, + 0.008208281360566616, + 0.007321546785533428, + 0.004350295290350914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.038090988993644714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_12/index_tensor([2627, 2627])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_12/index_tensor([2627, 2627])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b22f825ff535723e1290894ef345ab8b1376b2af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_12/index_tensor([2627, 2627])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle close-up captures a pristine, paper-thin sheet of ice nestled against a dark, damp landscape. The ice's surface, almost completely transparent, reveals the granular texture of the soil and a few trapped microscopic air bubbles beneath. In the cool, low-angle morning light, the ice shimmers and sparkles. The scene begins in absolute silence, a fragile moment of perfection. Then, driven by an unseen vibration, a thin crack appears silently and swiftly, bursting out from one side. This initial crack immediately spiders outward, spreading a complex network of interconnected fractures across the surface in a silent, explosive manner. The once-solid sheet shatters into dozens of jagged, geometric sections. These sections, as their momentum runs out, delicately shift and stabilize with a final, tiny crunch, settling into new positions on the exposed ground. Their freshly shattered edges catch the light, glittering like broken glass and creating a quietly beautiful atmosphere of destruction.", + "seed": 16082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_12/index_2627_seed_16082_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01862964075990021, + "vlm_logprob": -0.01862964075990021, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.01862964075990021, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02691473811864853 + ], + "opposite_logprob_tokens": [ + -3.7769148349761963 + ], + "logprob_reward": -0.02691473811864853, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02691473811864853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3620165288448334, + 0.04977850243449211, + 0.04472126439213753, + 0.09431842714548111, + 0.04140340909361839, + 0.06671956181526184, + 0.1224709302186966, + 0.07491526007652283, + 0.10562090575695038, + 0.03803522512316704 + ], + "credit_map": [ + 0.027601104229688644, + 0.1277453452348709, + 0.014124465174973011, + 0.01730051077902317, + 0.12434497475624084, + 0.019384516403079033, + 0.014713766053318977, + 0.016801845282316208, + 0.007767483592033386, + 0.00715692900121212, + 0.005242497660219669, + 0.005237360019236803, + 0.007219077553600073, + 0.004319846164435148, + 0.0072790952399373055, + 0.005556214135140181, + 0.007071360480040312, + 0.005533467512577772, + 0.004551936872303486, + 0.004913117736577988, + 0.007509950082749128, + 0.005662441719323397, + 0.003041387302801013, + 0.006437605246901512, + 0.011859859339892864, + 0.010976213961839676, + 0.009011813439428806, + 0.013485001400113106, + 0.020248880609869957, + 0.008581478148698807, + 0.011008290573954582, + 0.009146888740360737, + 0.004949767142534256, + 0.004911978729069233, + 0.005461426917463541, + 0.0041334922425448895, + 0.00964345596730709, + 0.00337035208940506, + 0.004233474377542734, + 0.004699459299445152, + 0.008184410631656647, + 0.0071151601150631905, + 0.009769724681973457, + 0.011503084562718868, + 0.010019948706030846, + 0.0070276944898068905, + 0.006968929432332516, + 0.00613061198964715, + 0.013927345164120197, + 0.013761123642325401, + 0.013984394259750843, + 0.01980600878596306, + 0.02807355299592018, + 0.011311247944831848, + 0.010841364040970802, + 0.010765887796878815, + 0.00988432765007019, + 0.008278502151370049, + 0.012374424375593662, + 0.015282218344509602, + 0.009847347624599934, + 0.005793449003249407, + 0.006972654722630978, + 0.006482334341853857, + 0.014167610555887222, + 0.012994298711419106, + 0.018532488495111465, + 0.01151486299932003, + 0.014672856777906418, + 0.013473552651703358, + 0.011394407600164413, + 0.008870833553373814, + 0.004888426978141069, + 0.003829237539321184, + 0.005333544686436653, + 0.005429443903267384, + 0.005196199286729097, + 0.004739495925605297, + 0.005339476745575666, + 0.0032793970312923193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.02691473811864853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018224529922008514 + ], + "opposite_logprob_tokens": [ + -4.018224716186523 + ], + "logprob_reward": -0.018224529922008514, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018224529922008514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42542076110839844, + 0.07054268568754196, + 0.09230177104473114, + 0.051010847091674805, + 0.07409847527742386, + 0.039157409220933914, + 0.02960233762860298, + 0.05617478862404823, + 0.10405641049146652, + 0.05763449892401695 + ], + "credit_map": [ + 0.0349636934697628, + 0.14658226072788239, + 0.019306553527712822, + 0.018323317170143127, + 0.14315763115882874, + 0.02412918210029602, + 0.017021629959344864, + 0.021936457604169846, + 0.010958151891827583, + 0.013845132663846016, + 0.005703355651348829, + 0.006091895047575235, + 0.014287590980529785, + 0.006581001915037632, + 0.006355339661240578, + 0.0067202127538621426, + 0.01871759630739689, + 0.012213866226375103, + 0.007764099631458521, + 0.010106015019118786, + 0.013208496384322643, + 0.011436608619987965, + 0.006006822921335697, + 0.012848266400396824, + 0.00720372935757041, + 0.005641767755150795, + 0.00582260312512517, + 0.006848805584013462, + 0.006306084804236889, + 0.004598530475050211, + 0.008943667635321617, + 0.005645655561238527, + 0.009212687611579895, + 0.008022966794669628, + 0.010174022056162357, + 0.006218540016561747, + 0.018095143139362335, + 0.005827389191836119, + 0.008110115304589272, + 0.008437608368694782, + 0.005093513987958431, + 0.0042645675130188465, + 0.0040474021807312965, + 0.0035312846302986145, + 0.008711384609341621, + 0.005279198754578829, + 0.00468571949750185, + 0.003544332692399621, + 0.0033531938679516315, + 0.0030754420440644026, + 0.004078899044543505, + 0.004318052902817726, + 0.0037054680287837982, + 0.0036719662602990866, + 0.003074340056627989, + 0.004324975423514843, + 0.008691729977726936, + 0.007267904467880726, + 0.009923618286848068, + 0.008111249655485153, + 0.006705829408019781, + 0.005043927580118179, + 0.005241962615400553, + 0.005188564769923687, + 0.013320455327630043, + 0.00803549773991108, + 0.01219264231622219, + 0.010035938583314419, + 0.00808941014111042, + 0.011812573298811913, + 0.014048370532691479, + 0.02652151882648468, + 0.007177525665611029, + 0.006570060737431049, + 0.01091901957988739, + 0.006018041167408228, + 0.006680060178041458, + 0.007891292683780193, + 0.006152375601232052, + 0.006226121447980404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.018224529922008514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033734407275915146 + ], + "opposite_logprob_tokens": [ + -3.4087343215942383 + ], + "logprob_reward": -0.033734407275915146, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033734407275915146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2817046046257019, + 0.06861089169979095, + 0.08135241270065308, + 0.11088750511407852, + 0.061245694756507874, + 0.049192558974027634, + 0.18878528475761414, + 0.04118366166949272, + 0.06908635795116425, + 0.047951020300388336 + ], + "credit_map": [ + 0.019411958754062653, + 0.09424757212400436, + 0.01525068935006857, + 0.01777586154639721, + 0.09011820703744888, + 0.013263758271932602, + 0.017630506306886673, + 0.014006054028868675, + 0.011471103876829147, + 0.016406062990427017, + 0.006047247443348169, + 0.005301313940435648, + 0.016807399690151215, + 0.005108400713652372, + 0.003104695351794362, + 0.00436466746032238, + 0.01462713722139597, + 0.010480218566954136, + 0.009183158166706562, + 0.010382646694779396, + 0.01332065649330616, + 0.008247666992247105, + 0.005479087121784687, + 0.009631835855543613, + 0.015489594079554081, + 0.01985796168446541, + 0.01179887168109417, + 0.015544387511909008, + 0.013236955739557743, + 0.009438489563763142, + 0.015252523124217987, + 0.010268720798194408, + 0.007973340339958668, + 0.005512353032827377, + 0.008146430365741253, + 0.008349842391908169, + 0.008606730960309505, + 0.006341900210827589, + 0.008132177405059338, + 0.008182916790246964, + 0.00572836771607399, + 0.006521651521325111, + 0.005612281151115894, + 0.004459024406969547, + 0.011699836701154709, + 0.005299119744449854, + 0.004736215341836214, + 0.005136061459779739, + 0.023682845756411552, + 0.02408113144338131, + 0.019166741520166397, + 0.02955450676381588, + 0.033006563782691956, + 0.02422177977859974, + 0.01389098446816206, + 0.021180730313062668, + 0.00476980721578002, + 0.005003785714507103, + 0.005157237406820059, + 0.008327518589794636, + 0.0043251351453363895, + 0.0035879898350685835, + 0.0061906008049845695, + 0.0038215890526771545, + 0.00723459105938673, + 0.007146837655454874, + 0.012888031080365181, + 0.008665747009217739, + 0.00590507872402668, + 0.007760617882013321, + 0.011304497718811035, + 0.008180953562259674, + 0.006245031952857971, + 0.0054053799249231815, + 0.00509692681953311, + 0.007946478202939034, + 0.007241883780807257, + 0.0050603668205440044, + 0.007486738730221987, + 0.0034682180266827345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.033734407275915146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078514158725739 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078514158725739, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078514158725739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2571190595626831, + 0.06446601450443268, + 0.14544391632080078, + 0.24273765087127686, + 0.061163291335105896, + 0.03168891370296478, + 0.08258388191461563, + 0.022765465080738068, + 0.03591635078191757, + 0.056115467101335526 + ], + "credit_map": [ + 0.026347285136580467, + 0.0769486352801323, + 0.011631305329501629, + 0.020053045824170113, + 0.07527150958776474, + 0.015834348276257515, + 0.015385251492261887, + 0.01564767397940159, + 0.006440408993512392, + 0.01567697338759899, + 0.004677019082009792, + 0.006480701267719269, + 0.016200117766857147, + 0.0048095607198774815, + 0.005558394826948643, + 0.004622834734618664, + 0.028794536367058754, + 0.022080756723880768, + 0.013671904802322388, + 0.01839909888803959, + 0.025234095752239227, + 0.01631508581340313, + 0.009055528789758682, + 0.011892922222614288, + 0.03165094181895256, + 0.026278123259544373, + 0.02253938652575016, + 0.035128384828567505, + 0.05319606140255928, + 0.022333048284053802, + 0.029801437631249428, + 0.02181023359298706, + 0.007827061228454113, + 0.007828501984477043, + 0.006420905236154795, + 0.007341544609516859, + 0.009822554886341095, + 0.006071419455111027, + 0.008335762657225132, + 0.007515544071793556, + 0.003904802957549691, + 0.0028593617025762796, + 0.003066783770918846, + 0.0030727474950253963, + 0.00920951273292303, + 0.003639424918219447, + 0.003291632980108261, + 0.002644649241119623, + 0.010009371675550938, + 0.009684901684522629, + 0.010889444500207901, + 0.013475366868078709, + 0.015313356183469296, + 0.009381858631968498, + 0.006467695813626051, + 0.007361882366240025, + 0.0025324609596282244, + 0.0024688539560884237, + 0.0028306504245847464, + 0.004150726832449436, + 0.003498455509543419, + 0.0019875881262123585, + 0.0036924739833921194, + 0.0016042551724240184, + 0.003780810395255685, + 0.003182175336405635, + 0.003940023016184568, + 0.0037298775278031826, + 0.003979780245572329, + 0.003521938808262348, + 0.00859580747783184, + 0.005185937043279409, + 0.007573464419692755, + 0.00599799957126379, + 0.014835176058113575, + 0.006085065193474293, + 0.005546570289880037, + 0.0038734981790184975, + 0.005647105630487204, + 0.006556587293744087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.011078514158725739 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031960143242031336 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031960143242031336, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031960143242031336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20703540742397308, + 0.08085068315267563, + 0.08534422516822815, + 0.10367658734321594, + 0.019016476348042488, + 0.03243356943130493, + 0.19009296596050262, + 0.01785704679787159, + 0.036116331815719604, + 0.22757676243782043 + ], + "credit_map": [ + 0.02272968366742134, + 0.0550026074051857, + 0.011651638895273209, + 0.01922314427793026, + 0.05861195921897888, + 0.013261635787785053, + 0.011820513755083084, + 0.014734200201928616, + 0.010160468518733978, + 0.020044270902872086, + 0.005309266038239002, + 0.007015707902610302, + 0.020853975787758827, + 0.00600391486659646, + 0.005637869238853455, + 0.005825194530189037, + 0.018332675099372864, + 0.012225984595716, + 0.00807238183915615, + 0.01065076980739832, + 0.012579511851072311, + 0.008596636354923248, + 0.00520500261336565, + 0.009681250900030136, + 0.014277134090662003, + 0.01625468209385872, + 0.0072269318625330925, + 0.011198047548532486, + 0.026163777336478233, + 0.009736452251672745, + 0.010347742587327957, + 0.008471811190247536, + 0.002199970418587327, + 0.0021806408185511827, + 0.0018156495643779635, + 0.002446003956720233, + 0.0035619947593659163, + 0.0021169993560761213, + 0.003011944703757763, + 0.0016832706751301885, + 0.0036913093645125628, + 0.004343417938798666, + 0.002684442326426506, + 0.003851803485304117, + 0.009428227320313454, + 0.002684111474081874, + 0.002833831124007702, + 0.0029164222069084644, + 0.023858124390244484, + 0.022145304828882217, + 0.01976986974477768, + 0.02880595624446869, + 0.0400468148291111, + 0.022884409874677658, + 0.01419318187981844, + 0.018389292061328888, + 0.0022195240017026663, + 0.0018608203390613198, + 0.0016321918228641152, + 0.003066417295485735, + 0.0036421515978872776, + 0.001528666471131146, + 0.002470711013302207, + 0.00143656344152987, + 0.004233218263834715, + 0.004718174692243338, + 0.0047974050976336, + 0.00445884233340621, + 0.004093576688319445, + 0.004038181621581316, + 0.005945935845375061, + 0.0038309949450194836, + 0.030988631770014763, + 0.0212639681994915, + 0.06095806509256363, + 0.024169839918613434, + 0.017355751246213913, + 0.014905004762113094, + 0.027374928817152977, + 0.030560556799173355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0031960143242031336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a pristine, paper-thin sheet of ice nestled against a dark, damp landscape. The ice's surface, almost completely transparent, reveals the granular texture of the soil and a few trapped microscopic air bubbles beneath. In the cool, low-angle morning light, the ice shimmers and sparkles. The scene begins in absolute silence, a fragile moment of perfection. Then, driven by an unseen vibration, a thin crack appears silently and swiftly, bursting out from one side. This initial crack immediately spiders outward, spreading a complex network of interconnected fractures across the surface in a silent, explosive manner. The once-solid sheet shatters into dozens of jagged, geometric sections. These sections, as their momentum runs out, delicately shift and stabilize with a final, tiny crunch, settling into new positions on the exposed ground. Their freshly shattered edges catch the light, glittering like broken glass and creating a quietly beautiful atmosphere of destruction.", + "seed": 16083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_12/index_2627_seed_16083_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02447016052901745, + "vlm_logprob": -0.02447016052901745, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02447016052901745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015273164957761765 + ], + "opposite_logprob_tokens": [ + -4.890273094177246 + ], + "logprob_reward": -0.015273164957761765, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015273164957761765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19612161815166473, + 0.0548405759036541, + 0.06398280709981918, + 0.05600470304489136, + 0.09648443013429642, + 0.13374383747577667, + 0.15751735866069794, + 0.11140085011720657, + 0.053704582154750824, + 0.07619918882846832 + ], + "credit_map": [ + 0.01583223044872284, + 0.03292910009622574, + 0.02031087875366211, + 0.020376279950141907, + 0.038128942251205444, + 0.031222989782691002, + 0.020519813522696495, + 0.016801388934254646, + 0.005338728427886963, + 0.008720737881958485, + 0.005626640748232603, + 0.005659777671098709, + 0.012334243394434452, + 0.006625887006521225, + 0.005769877228885889, + 0.004764683078974485, + 0.0063963658176362514, + 0.00878599938005209, + 0.0046778954565525055, + 0.009065667167305946, + 0.010447761043906212, + 0.010105437599122524, + 0.008370314724743366, + 0.006133366841822863, + 0.005450279917567968, + 0.008682495914399624, + 0.0064870561473071575, + 0.007960600778460503, + 0.009680680930614471, + 0.007015995681285858, + 0.004842828493565321, + 0.005884760525077581, + 0.011769820004701614, + 0.015860015526413918, + 0.00683445343747735, + 0.013632629066705704, + 0.01740831695497036, + 0.01162712275981903, + 0.008333470672369003, + 0.011018602177500725, + 0.017758669331669807, + 0.026066523045301437, + 0.01279326993972063, + 0.023173566907644272, + 0.017554091289639473, + 0.008752129040658474, + 0.015521973371505737, + 0.012123625725507736, + 0.03373561054468155, + 0.04205626994371414, + 0.016827067360281944, + 0.012486991472542286, + 0.013405654579401016, + 0.01394121814519167, + 0.012451575137674809, + 0.012612968683242798, + 0.016595318913459778, + 0.021139470860362053, + 0.010536866262555122, + 0.01044860202819109, + 0.015963222831487656, + 0.013984359800815582, + 0.012333717197179794, + 0.010399295017123222, + 0.01036459393799305, + 0.0046911099925637245, + 0.004366289358586073, + 0.0038460169453173876, + 0.013615972362458706, + 0.008700241334736347, + 0.0036608753725886345, + 0.004459481220692396, + 0.01098549459129572, + 0.007098446600139141, + 0.0058699301443994045, + 0.0076670520938932896, + 0.011399722658097744, + 0.014382390305399895, + 0.012423998676240444, + 0.006372153293341398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.015273164957761765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02639472857117653 + ], + "opposite_logprob_tokens": [ + -3.651394844055176 + ], + "logprob_reward": -0.02639472857117653, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02639472857117653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18926402926445007, + 0.07173808664083481, + 0.038409698754549026, + 0.09434272348880768, + 0.028037140145897865, + 0.16159752011299133, + 0.0936795026063919, + 0.047225549817085266, + 0.14208249747753143, + 0.13362321257591248 + ], + "credit_map": [ + 0.02026332914829254, + 0.02581707574427128, + 0.021690506488084793, + 0.0230676531791687, + 0.028976207599043846, + 0.0309706199914217, + 0.022027188912034035, + 0.01645144820213318, + 0.007724635303020477, + 0.01081935502588749, + 0.008276736363768578, + 0.007660121191293001, + 0.013162929564714432, + 0.008822569623589516, + 0.009188664145767689, + 0.006083076819777489, + 0.004535030573606491, + 0.00572969950735569, + 0.0028329091146588326, + 0.004966138396412134, + 0.006387878209352493, + 0.005486079957336187, + 0.0055033257231116295, + 0.002968636341392994, + 0.011184649541974068, + 0.013181574642658234, + 0.015564166940748692, + 0.012574275955557823, + 0.011462226510047913, + 0.012692293152213097, + 0.008087901398539543, + 0.009595634415745735, + 0.002963859587907791, + 0.005144651513546705, + 0.001847588922828436, + 0.0037319117691367865, + 0.006317897699773312, + 0.0027150444220751524, + 0.0025558816269040108, + 0.0027603048365563154, + 0.02281511202454567, + 0.030632521957159042, + 0.014789527282118797, + 0.030228883028030396, + 0.016463087871670723, + 0.014418063685297966, + 0.01716439239680767, + 0.015085933730006218, + 0.018680285662412643, + 0.024699507281184196, + 0.010930943302810192, + 0.007339332718402147, + 0.008205984719097614, + 0.009896598756313324, + 0.007153500337153673, + 0.006773347035050392, + 0.008539591915905476, + 0.007645251229405403, + 0.004716525785624981, + 0.004427528474479914, + 0.005803362000733614, + 0.005358782596886158, + 0.006558847613632679, + 0.004175661597400904, + 0.019390348345041275, + 0.010367605835199356, + 0.0068501816131174564, + 0.007129622623324394, + 0.033960431814193726, + 0.04744238778948784, + 0.007819442078471184, + 0.009122473187744617, + 0.016934679821133614, + 0.012237855233252048, + 0.011271473951637745, + 0.011662879027426243, + 0.022312624379992485, + 0.030192915350198746, + 0.01831417717039585, + 0.01069660484790802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.02639472857117653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06211427226662636 + ], + "opposite_logprob_tokens": [ + -2.8121142387390137 + ], + "logprob_reward": -0.06211427226662636, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06211427226662636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09708840399980545, + 0.047715794295072556, + 0.03315625712275505, + 0.037764910608530045, + 0.06379200518131256, + 0.14325585961341858, + 0.08302967250347137, + 0.05880802124738693, + 0.08533485978841782, + 0.35005417466163635 + ], + "credit_map": [ + 0.007634775247424841, + 0.015024085529148579, + 0.008480238728225231, + 0.011789357289671898, + 0.0221599992364645, + 0.011363862082362175, + 0.009264207445085049, + 0.01137189194560051, + 0.005025691818445921, + 0.005953798536211252, + 0.006772846449166536, + 0.0059541515074670315, + 0.006670370232313871, + 0.007213105913251638, + 0.005940537434071302, + 0.00418529799208045, + 0.0034964806400239468, + 0.005369136575609446, + 0.0031360185239464045, + 0.003716624341905117, + 0.00697935139760375, + 0.004253172781318426, + 0.0032482820097357035, + 0.002957194810733199, + 0.004706744570285082, + 0.004591300152242184, + 0.006286379881203175, + 0.004040899220854044, + 0.0044905031099915504, + 0.004886923357844353, + 0.004043471068143845, + 0.004718691576272249, + 0.008224280551075935, + 0.009494497440755367, + 0.008014203980565071, + 0.006829277146607637, + 0.009709603153169155, + 0.007481114007532597, + 0.005376653280109167, + 0.00866237748414278, + 0.021204784512519836, + 0.023175708949565887, + 0.015620706602931023, + 0.023060236126184464, + 0.016937753185629845, + 0.012237798422574997, + 0.017262376844882965, + 0.013756508007645607, + 0.01585572212934494, + 0.021601226180791855, + 0.010613000020384789, + 0.007977818138897419, + 0.007708830758929253, + 0.006409816909581423, + 0.006060880608856678, + 0.006802388001233339, + 0.00935607310384512, + 0.0057409778237342834, + 0.004183839075267315, + 0.004791441839188337, + 0.009298798628151417, + 0.010389507748186588, + 0.00956783164292574, + 0.005479560233652592, + 0.013748682104051113, + 0.011253253556787968, + 0.011453882791101933, + 0.006285707466304302, + 0.012568753212690353, + 0.015850896015763283, + 0.008171970024704933, + 0.006001726258546114, + 0.05865221843123436, + 0.036677755415439606, + 0.025759391486644745, + 0.0323152057826519, + 0.044314585626125336, + 0.08450578153133392, + 0.03958522900938988, + 0.028244052082300186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.06211427226662636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076931729912758 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.016076931729912758, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076931729912758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08103159815073013, + 0.020742405205965042, + 0.018456362187862396, + 0.01747039332985878, + 0.07182969152927399, + 0.23178397119045258, + 0.18422673642635345, + 0.09243690967559814, + 0.25294622778892517, + 0.029075689613819122 + ], + "credit_map": [ + 0.009013504721224308, + 0.011277087032794952, + 0.00803303997963667, + 0.012557951733469963, + 0.013153507374227047, + 0.011551269330084324, + 0.009284311905503273, + 0.0061609274707734585, + 0.0016170624876394868, + 0.0028708840254694223, + 0.0029589785262942314, + 0.002513713203370571, + 0.0035130034666508436, + 0.003270912915468216, + 0.002047523856163025, + 0.0019503269577398896, + 0.0012984892819076777, + 0.0029183391015976667, + 0.001722597749903798, + 0.0018281497759744525, + 0.004290694836527109, + 0.002829350996762514, + 0.0019232142949476838, + 0.0016455277800559998, + 0.00140465481672436, + 0.0025153178721666336, + 0.0025699827820062637, + 0.0017443251563236117, + 0.002658181358128786, + 0.0023797189351171255, + 0.0019480636110529304, + 0.0022501496132463217, + 0.007993744686245918, + 0.011269349604845047, + 0.006094146985560656, + 0.010264091193675995, + 0.01237325556576252, + 0.009541679173707962, + 0.005676587577909231, + 0.008616838604211807, + 0.031666386872529984, + 0.042124081403017044, + 0.025312045589089394, + 0.04540431872010231, + 0.023753108456730843, + 0.01729743555188179, + 0.025007983669638634, + 0.021218622103333473, + 0.034888170659542084, + 0.04459448903799057, + 0.025877270847558975, + 0.01546448189765215, + 0.01600046083331108, + 0.01777336187660694, + 0.01421885471791029, + 0.015409640967845917, + 0.013836462050676346, + 0.014863058924674988, + 0.007593836635351181, + 0.007422497030347586, + 0.015322227030992508, + 0.011668355204164982, + 0.01299306470900774, + 0.008737407624721527, + 0.034958697855472565, + 0.014604290947318077, + 0.013195873238146305, + 0.012164496816694736, + 0.05266028642654419, + 0.09350939840078354, + 0.01324201188981533, + 0.018611179664731026, + 0.004242705646902323, + 0.0034695786889642477, + 0.0036021401174366474, + 0.0024143881164491177, + 0.003495949087664485, + 0.006043733563274145, + 0.0032408800907433033, + 0.0025663133710622787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.016076931729912758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024917051196098328 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024917051196098328, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024917051196098328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09666300565004349, + 0.04066404327750206, + 0.026013610884547234, + 0.03233465924859047, + 0.04408830404281616, + 0.12096553295850754, + 0.2241993248462677, + 0.1659352332353592, + 0.05674203857779503, + 0.19239427149295807 + ], + "credit_map": [ + 0.009758034721016884, + 0.013436047360301018, + 0.011147677898406982, + 0.01437537744641304, + 0.013637025840580463, + 0.01432084385305643, + 0.011706746183335781, + 0.008281256072223186, + 0.003926631063222885, + 0.004896080121397972, + 0.005925120320171118, + 0.004182819277048111, + 0.006565351039171219, + 0.0059785787016153336, + 0.004752736538648605, + 0.004436729941517115, + 0.0023560882546007633, + 0.0034964163787662983, + 0.0024296885821968317, + 0.0030017525423318148, + 0.004330086056143045, + 0.004404030274599791, + 0.0032993017230182886, + 0.00269624637439847, + 0.003581797704100609, + 0.0046896180137991905, + 0.004084293730556965, + 0.004384231753647327, + 0.0055184923112392426, + 0.003767885733395815, + 0.0029012798331677914, + 0.003407059470191598, + 0.004875075537711382, + 0.007073850370943546, + 0.004262305796146393, + 0.005617611110210419, + 0.00949170719832182, + 0.004415848292410374, + 0.00420362176373601, + 0.004148284904658794, + 0.01776188053190708, + 0.020176969468593597, + 0.013838658109307289, + 0.021487709134817123, + 0.013973873108625412, + 0.011418825946748257, + 0.012341043911874294, + 0.009966581128537655, + 0.044712066650390625, + 0.05878234654664993, + 0.025778507813811302, + 0.01903858222067356, + 0.01850849948823452, + 0.023793824017047882, + 0.017722615972161293, + 0.015862872824072838, + 0.026307348161935806, + 0.0333271324634552, + 0.011851855553686619, + 0.013759583234786987, + 0.024982688948512077, + 0.024361543357372284, + 0.018073800951242447, + 0.013271287083625793, + 0.0064870435744524, + 0.0035061610396951437, + 0.0034841764718294144, + 0.0035763748455792665, + 0.01166246086359024, + 0.01970628835260868, + 0.003627097699791193, + 0.004692436661571264, + 0.03468950465321541, + 0.021361056715250015, + 0.01758939027786255, + 0.0129323098808527, + 0.01789402961730957, + 0.05616961419582367, + 0.018075184896588326, + 0.013683167286217213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0024917051196098328 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a pristine, paper-thin sheet of ice nestled against a dark, damp landscape. The ice's surface, almost completely transparent, reveals the granular texture of the soil and a few trapped microscopic air bubbles beneath. In the cool, low-angle morning light, the ice shimmers and sparkles. The scene begins in absolute silence, a fragile moment of perfection. Then, driven by an unseen vibration, a thin crack appears silently and swiftly, bursting out from one side. This initial crack immediately spiders outward, spreading a complex network of interconnected fractures across the surface in a silent, explosive manner. The once-solid sheet shatters into dozens of jagged, geometric sections. These sections, as their momentum runs out, delicately shift and stabilize with a final, tiny crunch, settling into new positions on the exposed ground. Their freshly shattered edges catch the light, glittering like broken glass and creating a quietly beautiful atmosphere of destruction.", + "seed": 16084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_12/index_2627_seed_16084_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017433135118335486, + "vlm_logprob": -0.017433135118335486, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.017433135118335486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02648528479039669 + ], + "opposite_logprob_tokens": [ + -3.7764852046966553 + ], + "logprob_reward": -0.02648528479039669, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02648528479039669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2616710364818573, + 0.060609668493270874, + 0.1598849892616272, + 0.08315733075141907, + 0.11047451943159103, + 0.05907844752073288, + 0.0506826713681221, + 0.03637860715389252, + 0.03574911504983902, + 0.14231353998184204 + ], + "credit_map": [ + 0.036655936390161514, + 0.04004422575235367, + 0.025707591325044632, + 0.03415722772479057, + 0.03387739881873131, + 0.030514489859342575, + 0.03980085998773575, + 0.02091333270072937, + 0.00960560329258442, + 0.007557079661637545, + 0.008276074193418026, + 0.007012892048805952, + 0.007567229680716991, + 0.00786123052239418, + 0.006234518252313137, + 0.006495045032352209, + 0.026395224034786224, + 0.023882104083895683, + 0.022530050948262215, + 0.027827490121126175, + 0.013982555828988552, + 0.013153919018805027, + 0.021259862929582596, + 0.010853808373212814, + 0.020148511976003647, + 0.013552114367485046, + 0.008556162007153034, + 0.00587027333676815, + 0.010990548878908157, + 0.007389493752270937, + 0.00870607141405344, + 0.007944166660308838, + 0.012291842140257359, + 0.00897744856774807, + 0.01994701474905014, + 0.012726305983960629, + 0.010313810780644417, + 0.015395380556583405, + 0.014322489500045776, + 0.016500240191817284, + 0.004450618755072355, + 0.004689816851168871, + 0.012087732553482056, + 0.007169939577579498, + 0.004635393153876066, + 0.013799979351460934, + 0.005508909467607737, + 0.0067360615357756615, + 0.0053830984979867935, + 0.007755137514322996, + 0.007909457199275494, + 0.00784751120954752, + 0.0047252788208425045, + 0.005375883542001247, + 0.006504274904727936, + 0.005182032939046621, + 0.005039629060775042, + 0.004865544382482767, + 0.006343048997223377, + 0.0036734072491526604, + 0.004068800713866949, + 0.00497845932841301, + 0.004121027886867523, + 0.003288690699264407, + 0.004363677464425564, + 0.005013410001993179, + 0.004956170916557312, + 0.003664115210995078, + 0.003464357927441597, + 0.006363837514072657, + 0.004154219292104244, + 0.0037693295162171125, + 0.01585538312792778, + 0.010398552753031254, + 0.016307519748806953, + 0.016909852623939514, + 0.008802345022559166, + 0.027269424870610237, + 0.015469826757907867, + 0.03130064159631729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02648528479039669 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256290078163147 + ], + "opposite_logprob_tokens": [ + -4.3875627517700195 + ], + "logprob_reward": -0.01256290078163147, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256290078163147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2224009484052658, + 0.12671753764152527, + 0.09207674115896225, + 0.08335734903812408, + 0.07635602355003357, + 0.06249247118830681, + 0.028168275952339172, + 0.05478319525718689, + 0.05052216351032257, + 0.2031252682209015 + ], + "credit_map": [ + 0.029083237051963806, + 0.025694703683257103, + 0.041303325444459915, + 0.02488682046532631, + 0.03078017756342888, + 0.021951822564005852, + 0.02842218428850174, + 0.020278669893741608, + 0.02509317919611931, + 0.02179647609591484, + 0.01721082255244255, + 0.017236759886145592, + 0.011586077511310577, + 0.012103507295250893, + 0.011981968767940998, + 0.00970874447375536, + 0.013618610799312592, + 0.01469248067587614, + 0.014757227152585983, + 0.012378256767988205, + 0.008577724918723106, + 0.011204667389392853, + 0.009486421011388302, + 0.0073613557033240795, + 0.02134316973388195, + 0.013773815706372261, + 0.008462379686534405, + 0.008899078704416752, + 0.008749128319323063, + 0.005721048917621374, + 0.011434619314968586, + 0.004974107723683119, + 0.010957124643027782, + 0.006609249394387007, + 0.00781184621155262, + 0.007661040406674147, + 0.006518082693219185, + 0.011366598308086395, + 0.012626837007701397, + 0.012805250473320484, + 0.006026069633662701, + 0.0059881312772631645, + 0.009728535078465939, + 0.006853375118225813, + 0.005691954400390387, + 0.013217815198004246, + 0.00853735115379095, + 0.006449240725487471, + 0.0018293822649866343, + 0.004300733562558889, + 0.005497641861438751, + 0.004410934634506702, + 0.0022763742599636316, + 0.003639336908236146, + 0.0031927565578371286, + 0.003021115902811289, + 0.006907348521053791, + 0.0071233720518648624, + 0.009709629230201244, + 0.005126503761857748, + 0.005910136736929417, + 0.00894231628626585, + 0.006634260527789593, + 0.0044296300038695335, + 0.00546808447688818, + 0.0060086604207754135, + 0.007395559921860695, + 0.00491408770903945, + 0.004890399985015392, + 0.011691355146467686, + 0.004665970336645842, + 0.005488040391355753, + 0.021115999668836594, + 0.01997482217848301, + 0.05212923884391785, + 0.013973726890981197, + 0.016877954825758934, + 0.031035175547003746, + 0.019226238131523132, + 0.028792118653655052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01256290078163147 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298262070864439 + ], + "opposite_logprob_tokens": [ + -3.5298261642456055 + ], + "logprob_reward": -0.0298262070864439, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298262070864439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23116515576839447, + 0.10321008414030075, + 0.11371740698814392, + 0.10311000049114227, + 0.15809205174446106, + 0.0677027776837349, + 0.04131032153964043, + 0.040186312049627304, + 0.05321637541055679, + 0.08828945457935333 + ], + "credit_map": [ + 0.04663068801164627, + 0.030226703733205795, + 0.0269060917198658, + 0.01773960143327713, + 0.03759294003248215, + 0.02829507179558277, + 0.0238084364682436, + 0.019965611398220062, + 0.027948793023824692, + 0.018678952008485794, + 0.013123895041644573, + 0.0101926913484931, + 0.010231026448309422, + 0.007143138442188501, + 0.00807822123169899, + 0.007813364267349243, + 0.030654607340693474, + 0.018340110778808594, + 0.010896758176386356, + 0.013513968326151371, + 0.015936966985464096, + 0.007210971787571907, + 0.010611099191009998, + 0.00655292347073555, + 0.022941872477531433, + 0.0175771526992321, + 0.009028667584061623, + 0.00891572143882513, + 0.01170979905873537, + 0.008812522515654564, + 0.01497676596045494, + 0.009147495031356812, + 0.027391979470849037, + 0.015799004584550858, + 0.02306973934173584, + 0.01723596639931202, + 0.013916926458477974, + 0.01800389587879181, + 0.025520876049995422, + 0.017153671011328697, + 0.004234872292727232, + 0.0069105480797588825, + 0.017833029851317406, + 0.009273421950638294, + 0.008758129552006721, + 0.008170348592102528, + 0.006614529527723789, + 0.0059078955091536045, + 0.0044979494996368885, + 0.008506765589118004, + 0.00621011620387435, + 0.004546933341771364, + 0.004267971031367779, + 0.005661664996296167, + 0.00387157266959548, + 0.0037473479751497507, + 0.005412027705460787, + 0.004816844593733549, + 0.007637908682227135, + 0.0035916855558753014, + 0.0044220201671123505, + 0.005916238762438297, + 0.004329739138484001, + 0.00405984977260232, + 0.008301718160510063, + 0.010695356875658035, + 0.006152789108455181, + 0.005350021179765463, + 0.006612471770495176, + 0.007221732288599014, + 0.0041621094569563866, + 0.004720177501440048, + 0.007720598019659519, + 0.010819999501109123, + 0.01328635960817337, + 0.01155990082770586, + 0.005326706450432539, + 0.011283602565526962, + 0.010678209364414215, + 0.017614077776670456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0298262070864439 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194594696164131 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194594696164131, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194594696164131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12755249440670013, + 0.07133876532316208, + 0.09596145153045654, + 0.20906059443950653, + 0.0836089551448822, + 0.0701899304986, + 0.08096329122781754, + 0.04910214990377426, + 0.04661015793681145, + 0.16561219096183777 + ], + "credit_map": [ + 0.01961975358426571, + 0.01691957749426365, + 0.02335541322827339, + 0.015306923538446426, + 0.015253393910825253, + 0.013460619375109673, + 0.010754704475402832, + 0.012882113456726074, + 0.008146091364324093, + 0.007891779765486717, + 0.009124787524342537, + 0.010259740985929966, + 0.01121435221284628, + 0.007291544694453478, + 0.011388115584850311, + 0.006022354122251272, + 0.015340487472712994, + 0.011568808928132057, + 0.009824524633586407, + 0.009993479587137699, + 0.01669601909816265, + 0.010161638259887695, + 0.0128162307664752, + 0.009560252539813519, + 0.046070851385593414, + 0.0347210094332695, + 0.017077157273888588, + 0.017874622717499733, + 0.02598821185529232, + 0.01625010184943676, + 0.03201783075928688, + 0.019060811027884483, + 0.012438931502401829, + 0.010881727561354637, + 0.010315533727407455, + 0.0077661555260419846, + 0.008550175465643406, + 0.01131213828921318, + 0.012215101160109043, + 0.010129192844033241, + 0.005033894907683134, + 0.005537738092243671, + 0.013352129608392715, + 0.00788069237023592, + 0.009748438373208046, + 0.01241899561136961, + 0.00945259165018797, + 0.006765446625649929, + 0.010197359137237072, + 0.007832994684576988, + 0.015931261703372, + 0.008434198796749115, + 0.007046801503747702, + 0.015659548342227936, + 0.007329165935516357, + 0.0085319634526968, + 0.006402392406016588, + 0.006958093028515577, + 0.008827608078718185, + 0.0037786918692290783, + 0.006223611533641815, + 0.007028528954833746, + 0.005138361360877752, + 0.0047448622062802315, + 0.007738934829831123, + 0.008780947886407375, + 0.005564631428569555, + 0.004508424084633589, + 0.005533156450837851, + 0.006473962217569351, + 0.003906410653144121, + 0.004103692248463631, + 0.018226370215415955, + 0.016115056350827217, + 0.02087286487221718, + 0.023197973147034645, + 0.012322971597313881, + 0.023936448618769646, + 0.020145418122410774, + 0.030795088037848473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.014194594696164131 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15679502487182617, + 0.054888833314180374, + 0.08265095204114914, + 0.27256178855895996, + 0.06463632732629776, + 0.09538368135690689, + 0.07204718887805939, + 0.04579654708504677, + 0.045168437063694, + 0.11007124930620193 + ], + "credit_map": [ + 0.022083137184381485, + 0.022784363478422165, + 0.028922919183969498, + 0.021846095100045204, + 0.013925086706876755, + 0.015575236640870571, + 0.014105730690062046, + 0.01755244843661785, + 0.010718859732151031, + 0.006492767482995987, + 0.008022889494895935, + 0.011218216270208359, + 0.004311491269618273, + 0.0036663871724158525, + 0.007258509751409292, + 0.0031997065525501966, + 0.00956379622220993, + 0.009581390768289566, + 0.010645223781466484, + 0.010858478955924511, + 0.009531828574836254, + 0.01163959689438343, + 0.013320120051503181, + 0.00751051539555192, + 0.057546667754650116, + 0.051717501133680344, + 0.02079106867313385, + 0.022891776636242867, + 0.034416209906339645, + 0.021865850314497948, + 0.03937653824687004, + 0.0239561814814806, + 0.008139941841363907, + 0.0083544896915555, + 0.007460351102054119, + 0.006568227428942919, + 0.006678577978163958, + 0.010804879479110241, + 0.00891786627471447, + 0.007711987942457199, + 0.008377634920179844, + 0.009142111986875534, + 0.01549780648201704, + 0.011111089959740639, + 0.011574775911867619, + 0.01898704282939434, + 0.010742193087935448, + 0.009951015934348106, + 0.009179805405437946, + 0.008519483730196953, + 0.012928077019751072, + 0.007943264208734035, + 0.00575689971446991, + 0.012868626043200493, + 0.0072392551228404045, + 0.007611775770783424, + 0.006290195509791374, + 0.006683773826807737, + 0.0075427452102303505, + 0.004184841178357601, + 0.006009338889271021, + 0.0058897570706903934, + 0.005602388642728329, + 0.003593503963202238, + 0.006132069975137711, + 0.006695867981761694, + 0.006626949645578861, + 0.004213280510157347, + 0.005772218573838472, + 0.007878164760768414, + 0.0041767023503780365, + 0.0036731816362589598, + 0.013272784650325775, + 0.009842284955084324, + 0.02093232050538063, + 0.009230056777596474, + 0.008059432730078697, + 0.014620776288211346, + 0.017508652061223984, + 0.01660492829978466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.004096688237041235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a pristine, paper-thin sheet of ice nestled against a dark, damp landscape. The ice's surface, almost completely transparent, reveals the granular texture of the soil and a few trapped microscopic air bubbles beneath. In the cool, low-angle morning light, the ice shimmers and sparkles. The scene begins in absolute silence, a fragile moment of perfection. Then, driven by an unseen vibration, a thin crack appears silently and swiftly, bursting out from one side. This initial crack immediately spiders outward, spreading a complex network of interconnected fractures across the surface in a silent, explosive manner. The once-solid sheet shatters into dozens of jagged, geometric sections. These sections, as their momentum runs out, delicately shift and stabilize with a final, tiny crunch, settling into new positions on the exposed ground. Their freshly shattered edges catch the light, glittering like broken glass and creating a quietly beautiful atmosphere of destruction.", + "seed": 16085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_12/index_2627_seed_16085_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016079941927455366, + "vlm_logprob": -0.016079941927455366, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.016079941927455366, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008262262679636478 + ], + "opposite_logprob_tokens": [ + -5.0082621574401855 + ], + "logprob_reward": -0.008262262679636478, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008262262679636478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40673959255218506, + 0.0958041101694107, + 0.07873018085956573, + 0.054503683000802994, + 0.13143612444400787, + 0.11316744983196259, + 0.06756456196308136, + 0.019650550559163094, + 0.01578119397163391, + 0.01662256382405758 + ], + "credit_map": [ + 0.03358272463083267, + 0.09065654128789902, + 0.029801493510603905, + 0.026815535500645638, + 0.040876761078834534, + 0.03483562543988228, + 0.10875783115625381, + 0.041413094848394394, + 0.010628868825733662, + 0.014359809458255768, + 0.011605469509959221, + 0.009798021987080574, + 0.010458054952323437, + 0.00849205069243908, + 0.019894326105713844, + 0.010567515157163143, + 0.00524461455643177, + 0.005840362049639225, + 0.006461726035922766, + 0.013859707862138748, + 0.012374801561236382, + 0.00791650079190731, + 0.012864002957940102, + 0.014168466441333294, + 0.011114546097815037, + 0.004677572287619114, + 0.005889871157705784, + 0.006654022727161646, + 0.009229443967342377, + 0.004953363444656134, + 0.00578457722440362, + 0.006200285162776709, + 0.024941211566329002, + 0.013998913578689098, + 0.019053911790251732, + 0.014683942310512066, + 0.018316572532057762, + 0.018565885722637177, + 0.009069758467376232, + 0.0128059396520257, + 0.011070698499679565, + 0.01632840558886528, + 0.014209138229489326, + 0.014122724533081055, + 0.01234394870698452, + 0.013930092565715313, + 0.015116981230676174, + 0.016045469790697098, + 0.010524229146540165, + 0.005244421772658825, + 0.017681477591395378, + 0.004630628507584333, + 0.006801661569625139, + 0.01186335552483797, + 0.0044165486469864845, + 0.0064022415317595005, + 0.002864016918465495, + 0.0015671913279220462, + 0.006095581687986851, + 0.0014277806039899588, + 0.0018336190842092037, + 0.0026686065830290318, + 0.001264367369003594, + 0.0019293855875730515, + 0.001971048302948475, + 0.0022025760263204575, + 0.0034303010907024145, + 0.001712079276330769, + 0.0015082024037837982, + 0.001765637774951756, + 0.001558152143843472, + 0.0016331971855834126, + 0.002286091912537813, + 0.0015436147805303335, + 0.0036651073023676872, + 0.0027265124954283237, + 0.0014887653524056077, + 0.0019277471583336592, + 0.00144861894659698, + 0.0015361064579337835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.008262262679636478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011101385578513145 + ], + "opposite_logprob_tokens": [ + -4.511101245880127 + ], + "logprob_reward": -0.011101385578513145, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011101385578513145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.450807124376297, + 0.10053221136331558, + 0.07057172805070877, + 0.06423678994178772, + 0.060335468500852585, + 0.10171329230070114, + 0.08335819095373154, + 0.018020806834101677, + 0.03580930083990097, + 0.014615043997764587 + ], + "credit_map": [ + 0.03470667451620102, + 0.09499474614858627, + 0.03460092842578888, + 0.030456596985459328, + 0.05604538321495056, + 0.04183493182063103, + 0.10738251358270645, + 0.05078534781932831, + 0.008860776200890541, + 0.01029505766928196, + 0.01642153598368168, + 0.010341205634176731, + 0.01594686135649681, + 0.01253538765013218, + 0.016876842826604843, + 0.009254546836018562, + 0.007977641187608242, + 0.005587558262050152, + 0.008275103755295277, + 0.011496230959892273, + 0.007707204669713974, + 0.010024158284068108, + 0.012342900969088078, + 0.007160924840718508, + 0.011583658866584301, + 0.006282373797148466, + 0.0058929393999278545, + 0.004761938937008381, + 0.016246207058429718, + 0.005139992106705904, + 0.0044495644979178905, + 0.009880115278065205, + 0.00709143141284585, + 0.007039867341518402, + 0.008563208393752575, + 0.008922290988266468, + 0.007470693439245224, + 0.009022076614201069, + 0.004786036908626556, + 0.0074398634023964405, + 0.011408413760364056, + 0.012181046418845654, + 0.014317062683403492, + 0.010799597948789597, + 0.010035008192062378, + 0.014623010531067848, + 0.013465090654790401, + 0.01488406676799059, + 0.015435373410582542, + 0.0052155619487166405, + 0.03152596205472946, + 0.005825441796332598, + 0.00480528874322772, + 0.009994889609515667, + 0.004400437697768211, + 0.006155235692858696, + 0.0027336475905030966, + 0.0016293247463181615, + 0.003643602365627885, + 0.001309079467318952, + 0.002200551563873887, + 0.0028556198813021183, + 0.001186143490485847, + 0.002462836680933833, + 0.004989830777049065, + 0.0029851021245121956, + 0.009066643193364143, + 0.00391389150172472, + 0.0029485030099749565, + 0.004271815996617079, + 0.003657125635072589, + 0.0039763860404491425, + 0.001851011416874826, + 0.001957283355295658, + 0.0027559378650039434, + 0.00146100006531924, + 0.0014050911413505673, + 0.0019608826842159033, + 0.001516410382464528, + 0.0017074281349778175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011101385578513145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054998978972435 + ], + "opposite_logprob_tokens": [ + -2.9299988746643066 + ], + "logprob_reward": -0.054998978972435, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054998978972435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2441871017217636, + 0.08604365587234497, + 0.0643651932477951, + 0.09855086356401443, + 0.16342763602733612, + 0.10808242112398148, + 0.10777481645345688, + 0.05407959595322609, + 0.04345978423953056, + 0.03002893179655075 + ], + "credit_map": [ + 0.02598338946700096, + 0.02976244129240513, + 0.02774863690137863, + 0.02318774349987507, + 0.033841412514448166, + 0.03297021612524986, + 0.043669428676366806, + 0.02702384442090988, + 0.0071249655447900295, + 0.013228842988610268, + 0.010434776544570923, + 0.007816941477358341, + 0.0130327008664608, + 0.010357983410358429, + 0.014757209457457066, + 0.009290230460464954, + 0.008077294565737247, + 0.00613327557221055, + 0.006276782136410475, + 0.009048882871866226, + 0.008231714367866516, + 0.006290297955274582, + 0.012318155728280544, + 0.00798879750072956, + 0.009853200055658817, + 0.01761247217655182, + 0.016152506694197655, + 0.012063849717378616, + 0.009846893139183521, + 0.009535287506878376, + 0.01146780140697956, + 0.012018851935863495, + 0.03242441266775131, + 0.012608797289431095, + 0.029764700680971146, + 0.01827072538435459, + 0.017341647297143936, + 0.023010311648249626, + 0.010287783108651638, + 0.019719261676073074, + 0.012531658634543419, + 0.01566876471042633, + 0.01691313646733761, + 0.014771388843655586, + 0.012228308245539665, + 0.012011499144136906, + 0.012610108591616154, + 0.011347560212016106, + 0.015006600879132748, + 0.009164963848888874, + 0.026969537138938904, + 0.008604263886809349, + 0.011074990034103394, + 0.015578865073621273, + 0.010017599910497665, + 0.011357996612787247, + 0.0072335475124418736, + 0.007065956946462393, + 0.01331475842744112, + 0.004745581652969122, + 0.005750981159508228, + 0.00499758031219244, + 0.005141424015164375, + 0.005829767324030399, + 0.005610947962850332, + 0.005093614105135202, + 0.009269517846405506, + 0.004727819934487343, + 0.004128690343350172, + 0.004904392175376415, + 0.005109105724841356, + 0.004615698475390673, + 0.0036828573793172836, + 0.003980128560215235, + 0.006088092923164368, + 0.002982790581882, + 0.004180644638836384, + 0.004142529331147671, + 0.0020742022898048162, + 0.00289768329821527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.054998978972435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019408691441640258 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019408691441640258, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019408691441640258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2927156388759613, + 0.0624472014605999, + 0.1677079200744629, + 0.05445071682333946, + 0.061886731535196304, + 0.19580872356891632, + 0.07168900966644287, + 0.021803610026836395, + 0.02978619933128357, + 0.04170427471399307 + ], + "credit_map": [ + 0.02002483420073986, + 0.05231830105185509, + 0.01863281987607479, + 0.02561098523437977, + 0.03367653116583824, + 0.018648259341716766, + 0.0911829024553299, + 0.03262099623680115, + 0.005234700161963701, + 0.009058774448931217, + 0.010604886338114738, + 0.007716979365795851, + 0.006707572378218174, + 0.0070013138465583324, + 0.010680310428142548, + 0.005442663561552763, + 0.016692886129021645, + 0.012274105101823807, + 0.014944059774279594, + 0.03451043739914894, + 0.023916024714708328, + 0.014743715524673462, + 0.024124493822455406, + 0.026502186432480812, + 0.005527305416762829, + 0.007281285710632801, + 0.0048594046384096146, + 0.004498034715652466, + 0.011611763387918472, + 0.004231508821249008, + 0.004289871081709862, + 0.012151536531746387, + 0.010214847512543201, + 0.007120006717741489, + 0.007720586843788624, + 0.008545209653675556, + 0.009227699600160122, + 0.0066133285872638226, + 0.005809645634144545, + 0.0066354041919112206, + 0.020936787128448486, + 0.02147984877228737, + 0.026873966678977013, + 0.021917933598160744, + 0.01975182071328163, + 0.02488297037780285, + 0.030370008200407028, + 0.02959538996219635, + 0.01106740441173315, + 0.005777247715741396, + 0.01835671253502369, + 0.004973289091140032, + 0.007457956671714783, + 0.01138759683817625, + 0.005241692997515202, + 0.007427101489156485, + 0.0028011323884129524, + 0.003252263879403472, + 0.003769191214814782, + 0.002225901698693633, + 0.0025382465682923794, + 0.002875371603295207, + 0.002070167101919651, + 0.002271336270496249, + 0.0037975721061229706, + 0.0038956983480602503, + 0.005766657646745443, + 0.002734910696744919, + 0.003283851547166705, + 0.003553545568138361, + 0.004191359039396048, + 0.0025626029819250107, + 0.0050847167149186134, + 0.004045199602842331, + 0.010490821674466133, + 0.005633933003991842, + 0.005007355939596891, + 0.005666668526828289, + 0.0027276345063000917, + 0.003047939157113433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0019408691441640258 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a paper-thin sheet of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice surface almost completely transparent, revealing granular soil texture and trapped air bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature cool, low-angle morning light causing the ice to shimmer and sparkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thin crack appear silently and swiftly, spreading outward to shatter the ice into dozens of jagged geometric sections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shattered ice sections shift and stabilize on the ground with glittering edges that catch the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096213262528181 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096213262528181, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096213262528181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35272863507270813, + 0.057970739901065826, + 0.08305072784423828, + 0.12734994292259216, + 0.1520991027355194, + 0.06581874191761017, + 0.07718843966722488, + 0.036659300327301025, + 0.024805301800370216, + 0.022329140454530716 + ], + "credit_map": [ + 0.02655339613556862, + 0.09290041029453278, + 0.02229195088148117, + 0.02394825778901577, + 0.029906420037150383, + 0.014368551783263683, + 0.11112940311431885, + 0.03163022920489311, + 0.00485029723495245, + 0.006453643087297678, + 0.006355824414640665, + 0.006602533161640167, + 0.010367861948907375, + 0.0051381210796535015, + 0.010407966561615467, + 0.007794487755745649, + 0.005583587102591991, + 0.005893060006201267, + 0.01015155017375946, + 0.018017591908574104, + 0.00844536628574133, + 0.00632887938991189, + 0.014983740635216236, + 0.01364694070070982, + 0.01583137735724449, + 0.017475318163633347, + 0.022269876673817635, + 0.014149912633001804, + 0.015065725892782211, + 0.012929915450513363, + 0.013662773184478283, + 0.01596502959728241, + 0.028408968821167946, + 0.01732243038713932, + 0.018582582473754883, + 0.018857797607779503, + 0.019894404336810112, + 0.02102196402847767, + 0.012563680298626423, + 0.015447255223989487, + 0.011888464912772179, + 0.006349933333694935, + 0.012850936502218246, + 0.006630196236073971, + 0.005645590368658304, + 0.00894667487591505, + 0.00607376778498292, + 0.0074331676587462425, + 0.02069716900587082, + 0.005036155693233013, + 0.021049687638878822, + 0.004511600825935602, + 0.006447502877563238, + 0.008799409493803978, + 0.004540861584246159, + 0.006106041371822357, + 0.004817772191017866, + 0.0034375814720988274, + 0.011826899833977222, + 0.0032912336755543947, + 0.0034735468216240406, + 0.004216385539621115, + 0.002547679003328085, + 0.0030481957364827394, + 0.0035303093027323484, + 0.0023100893013179302, + 0.006421862170100212, + 0.0022878549061715603, + 0.002366956090554595, + 0.0034315623342990875, + 0.002255810424685478, + 0.002200852846726775, + 0.0025286225136369467, + 0.001914361258968711, + 0.005190444644540548, + 0.001776152173988521, + 0.0027585229836404324, + 0.0032212981022894382, + 0.0016367239877581596, + 0.0033030128106474876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.004096213262528181 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_tensor([7377, 7377])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_tensor([7377, 7377])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..acc6ebcabea81f04440770909e8bad3fbd6b5c9c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_tensor([7377, 7377])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, long shot shot from near the surface, the camera tilts steadily upward to track a massive humpback whale as it breaks through the deep indigo water with explosive force. Its vast, water-soaked body, its dark gray skin dotted with barnacles and scars of time, twists gracefully through the air in a majestic, graceful arc, its enormous, wing-like white pectoral fins fully extended. The golden late-afternoon sun captures every stream of water pouring down from its body, painting the air of spray into a shimmering golden mist, where droplets hover momentarily like countless diamonds. As the whale reaches the apex of its arc and begins its unstoppable plunge back into the water, sending up a deafening wall of white foam, the scene is filled with awe at the power of nature and an atmosphere of breathtaking tranquility and grandeur.", + "seed": 27570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_7377_seed_27570_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04456168469041586, + "vlm_logprob": -0.04456168469041586, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.04456168469041586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0478479228913784 + ], + "opposite_logprob_tokens": [ + -3.5478479862213135 + ], + "logprob_reward": -0.0478479228913784, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0478479228913784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11986314505338669, + 0.06800719350576401, + 0.07980406284332275, + 0.14008405804634094, + 0.031223222613334656, + 0.05917992815375328, + 0.10838180780410767, + 0.2285681962966919, + 0.09016567468643188, + 0.07472272217273712 + ], + "credit_map": [ + 0.020528040826320648, + 0.018837707117199898, + 0.009104812517762184, + 0.014238467440009117, + 0.016235338523983955, + 0.021465029567480087, + 0.009746593423187733, + 0.009707154706120491, + 0.0071558961644768715, + 0.010135143995285034, + 0.00857679732143879, + 0.008093398064374924, + 0.009912066161632538, + 0.011221790686249733, + 0.007596989627927542, + 0.005315105896443129, + 0.010411793366074562, + 0.011575463227927685, + 0.008999249897897243, + 0.012068373151123524, + 0.007918302901089191, + 0.011226369068026543, + 0.008085441775619984, + 0.009519065730273724, + 0.01854492537677288, + 0.01600658893585205, + 0.01711631938815117, + 0.026589231565594673, + 0.009691997431218624, + 0.021653439849615097, + 0.017955375835299492, + 0.012526172213256359, + 0.004820462316274643, + 0.004948947578668594, + 0.003482838859781623, + 0.004227783065289259, + 0.0028730223421007395, + 0.00361843453720212, + 0.004393981769680977, + 0.0028577533084899187, + 0.010769342072308064, + 0.01171104982495308, + 0.005386361386626959, + 0.006380988750606775, + 0.005938094109296799, + 0.007681798655539751, + 0.007003351580351591, + 0.004308940842747688, + 0.014702409505844116, + 0.011033791117370129, + 0.013747784309089184, + 0.028079431504011154, + 0.01140142884105444, + 0.00510827824473381, + 0.01592281460762024, + 0.008385872468352318, + 0.03496921807527542, + 0.019936753436923027, + 0.058559853583574295, + 0.022434692829847336, + 0.02901403047144413, + 0.013779575936496258, + 0.01968231610953808, + 0.030191760510206223, + 0.012837794609367847, + 0.011359267868101597, + 0.012555026449263096, + 0.019708378240466118, + 0.009145786054432392, + 0.004625153262168169, + 0.010296599008142948, + 0.00963767059147358, + 0.008300724439322948, + 0.0088414391502738, + 0.008304113522171974, + 0.01225385069847107, + 0.012750139459967613, + 0.00854711327701807, + 0.008867840282619, + 0.006857502274215221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.0478479228913784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020615939050912857 + ], + "opposite_logprob_tokens": [ + -3.895616054534912 + ], + "logprob_reward": -0.020615939050912857, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020615939050912857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04299461096525192, + 0.05060122162103653, + 0.01866724342107773, + 0.1028188094496727, + 0.04396533593535423, + 0.03664614260196686, + 0.031207647174596786, + 0.5019832849502563, + 0.12073373794555664, + 0.05038191378116608 + ], + "credit_map": [ + 0.00626466516405344, + 0.005991405341774225, + 0.0045874519273638725, + 0.003906513564288616, + 0.006601483095437288, + 0.006913437973707914, + 0.0038337160367518663, + 0.004895937163382769, + 0.006161818280816078, + 0.005872833076864481, + 0.011526310816407204, + 0.00555632496252656, + 0.007978907786309719, + 0.004166457802057266, + 0.003960845991969109, + 0.005377722438424826, + 0.0021645203232765198, + 0.0026671630330383778, + 0.002386289881542325, + 0.0030256379395723343, + 0.0018817618256434798, + 0.0029993595089763403, + 0.001762544154189527, + 0.0017799672205001116, + 0.013314582407474518, + 0.01175612024962902, + 0.012582401745021343, + 0.017037102952599525, + 0.007766394410282373, + 0.017234617844223976, + 0.014830225147306919, + 0.00829736515879631, + 0.006904545705765486, + 0.006662692409008741, + 0.005731460638344288, + 0.005433980841189623, + 0.003868279280140996, + 0.0036547956988215446, + 0.005411459133028984, + 0.006298125721514225, + 0.00515275401994586, + 0.006176726892590523, + 0.0045597669668495655, + 0.005813727155327797, + 0.004106706473976374, + 0.0030490257777273655, + 0.0050705051980912685, + 0.002716928021982312, + 0.004855369683355093, + 0.003720645559951663, + 0.004852245096117258, + 0.0034651942551136017, + 0.003422392997890711, + 0.0030538158025592566, + 0.004341599531471729, + 0.0034963837824761868, + 0.0752328634262085, + 0.04247674718499184, + 0.12379222363233566, + 0.050181321799755096, + 0.06774375587701797, + 0.030642664059996605, + 0.04426493123173714, + 0.06764879822731018, + 0.01710282452404499, + 0.01747119054198265, + 0.01678440347313881, + 0.021917227655649185, + 0.012432173825800419, + 0.007970062084496021, + 0.014653952792286873, + 0.012401903979480267, + 0.008031720295548439, + 0.007028189487755299, + 0.0070622884668409824, + 0.006703100632876158, + 0.005827931221574545, + 0.005660627502948046, + 0.005329724866896868, + 0.0047383299097418785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.020615939050912857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07009972631931305 + ], + "opposite_logprob_tokens": [ + -2.6950998306274414 + ], + "logprob_reward": -0.07009972631931305, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07009972631931305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11114374548196793, + 0.07829008251428604, + 0.059519875794649124, + 0.060409314930438995, + 0.027945037931203842, + 0.0714726448059082, + 0.03770215064287186, + 0.11187849193811417, + 0.12214343994855881, + 0.31949523091316223 + ], + "credit_map": [ + 0.01378126721829176, + 0.016417739912867546, + 0.007463269867002964, + 0.01200421154499054, + 0.023201443254947662, + 0.01755346916615963, + 0.010196778923273087, + 0.010525557212531567, + 0.008437477983534336, + 0.008939897641539574, + 0.016955174505710602, + 0.00985037162899971, + 0.007969084195792675, + 0.009592845104634762, + 0.006709515117108822, + 0.009835709817707539, + 0.009426715783774853, + 0.005874135997146368, + 0.006103155668824911, + 0.007855100557208061, + 0.008094522170722485, + 0.007257428485900164, + 0.00783799309283495, + 0.007070817984640598, + 0.008859342895448208, + 0.006089308299124241, + 0.005322055891156197, + 0.009972183965146542, + 0.004472183994948864, + 0.011497088707983494, + 0.0072236331179738045, + 0.006973512936383486, + 0.0049193622544407845, + 0.00555326696485281, + 0.003244950668886304, + 0.002963961800560355, + 0.00268042110837996, + 0.0026608146727085114, + 0.0030236891470849514, + 0.0028985689859837294, + 0.012961627915501595, + 0.014209666289389133, + 0.005857914686203003, + 0.008608760312199593, + 0.008235609158873558, + 0.009558407589793205, + 0.006117563229054213, + 0.005923089105635881, + 0.0061928099021315575, + 0.0060956827364861965, + 0.003819334087893367, + 0.003994170576334, + 0.003198414808139205, + 0.00394654693081975, + 0.00543734896928072, + 0.005017836578190327, + 0.017065273597836494, + 0.010621914640069008, + 0.026308482512831688, + 0.010692457668483257, + 0.013930385932326317, + 0.009328708983957767, + 0.010836394503712654, + 0.013094865716993809, + 0.012991257943212986, + 0.01647205650806427, + 0.016502896323800087, + 0.02120264805853367, + 0.015623337589204311, + 0.008646564558148384, + 0.011386818252503872, + 0.019317850470542908, + 0.03506121784448624, + 0.04733353853225708, + 0.04243660718202591, + 0.05834333971142769, + 0.0584881417453289, + 0.02625308558344841, + 0.026398953050374985, + 0.025180311873555183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.07009972631931305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004459947347641 + ], + "opposite_logprob_tokens": [ + -2.69504451751709 + ], + "logprob_reward": -0.07004459947347641, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004459947347641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07469236850738525, + 0.06694172322750092, + 0.018078476190567017, + 0.04649000242352486, + 0.041670724749565125, + 0.030667919665575027, + 0.16464029252529144, + 0.34583979845046997, + 0.06718392670154572, + 0.14379478991031647 + ], + "credit_map": [ + 0.007477238308638334, + 0.011310193687677383, + 0.006347307935357094, + 0.007046490907669067, + 0.016247134655714035, + 0.010700161568820477, + 0.007881107740104198, + 0.007682740222662687, + 0.008697896264493465, + 0.0077555072493851185, + 0.00968991406261921, + 0.00933866947889328, + 0.008589247241616249, + 0.008274666033685207, + 0.007380034774541855, + 0.007215786259621382, + 0.002572879893705249, + 0.0017004781402647495, + 0.0017736685695126653, + 0.0027076269034296274, + 0.002139621414244175, + 0.0023238048888742924, + 0.002221684670075774, + 0.002638710429891944, + 0.005271101836115122, + 0.0057693468406796455, + 0.005856639705598354, + 0.008806666359305382, + 0.004106015432626009, + 0.00569336861371994, + 0.007170374039560556, + 0.003816491225734353, + 0.007190148811787367, + 0.008013155311346054, + 0.004566347226500511, + 0.004857712425291538, + 0.0033136485144495964, + 0.005601017270237207, + 0.004708067514002323, + 0.003420626511797309, + 0.0051583112217485905, + 0.004874367266893387, + 0.003648797981441021, + 0.004705804865807295, + 0.002179632429033518, + 0.002659620251506567, + 0.005039042327553034, + 0.0024023442529141903, + 0.026889661327004433, + 0.018391279503703117, + 0.023753363639116287, + 0.04534539580345154, + 0.01056238729506731, + 0.009017355740070343, + 0.019117137417197227, + 0.011563709937036037, + 0.051703207194805145, + 0.029433853924274445, + 0.08407125622034073, + 0.034644726663827896, + 0.047674667090177536, + 0.02149083837866783, + 0.03041265346109867, + 0.04640859737992287, + 0.008364384062588215, + 0.009948590770363808, + 0.010008309036493301, + 0.01217895932495594, + 0.006993410177528858, + 0.003924294374883175, + 0.009203484281897545, + 0.006562494207173586, + 0.01799105294048786, + 0.017151037231087685, + 0.018273688852787018, + 0.028228946030139923, + 0.01900268904864788, + 0.011366530321538448, + 0.015064738690853119, + 0.016716107726097107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.07004459947347641 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200235716998577 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200235716998577, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200235716998577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20804086327552795, + 0.20819438993930817, + 0.02484656125307083, + 0.12068752944469452, + 0.02997482381761074, + 0.09606456756591797, + 0.09116118401288986, + 0.11058028042316437, + 0.046364519745111465, + 0.06408528238534927 + ], + "credit_map": [ + 0.03298302739858627, + 0.029643898829817772, + 0.01763984002172947, + 0.023213598877191544, + 0.03585999831557274, + 0.02821926213800907, + 0.02195669710636139, + 0.018524542450904846, + 0.023080771788954735, + 0.022963346913456917, + 0.03550935909152031, + 0.02467937208712101, + 0.024440385401248932, + 0.02414790913462639, + 0.022524405270814896, + 0.030848832800984383, + 0.004059977363795042, + 0.0023188884370028973, + 0.002626921981573105, + 0.003280108096078038, + 0.0033421253319829702, + 0.003778598504140973, + 0.0025498284958302975, + 0.0028901132754981518, + 0.019022498279809952, + 0.01693904399871826, + 0.013883569277822971, + 0.01902061142027378, + 0.00887710228562355, + 0.014946987852454185, + 0.016497856006026268, + 0.011499856598675251, + 0.005054568871855736, + 0.00562323909252882, + 0.0033933687955141068, + 0.003175942227244377, + 0.0032262278255075216, + 0.003765219822525978, + 0.003018884686753154, + 0.002717372728511691, + 0.016257764771580696, + 0.01761946827173233, + 0.008446497842669487, + 0.011431155726313591, + 0.010163030587136745, + 0.01264322642236948, + 0.011416075751185417, + 0.0080873416736722, + 0.014598766341805458, + 0.010565850883722305, + 0.013521174900233746, + 0.023493869230151176, + 0.0063728513196110725, + 0.006108597852289677, + 0.008859273977577686, + 0.007640809752047062, + 0.01668478362262249, + 0.009134095162153244, + 0.026939988136291504, + 0.010813829489052296, + 0.015196625143289566, + 0.007175974082201719, + 0.010576472617685795, + 0.01405851449817419, + 0.005945487879216671, + 0.00805159192532301, + 0.006453393492847681, + 0.009551271796226501, + 0.00452180951833725, + 0.0021980698220431805, + 0.004425255581736565, + 0.005217639729380608, + 0.00721256248652935, + 0.00806910265237093, + 0.009229179471731186, + 0.012323861941695213, + 0.009209158830344677, + 0.0066586267203092575, + 0.006476385053247213, + 0.004906405229121447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.014200235716998577 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long shot shot from near the surface, the camera tilts steadily upward to track a massive humpback whale as it breaks through the deep indigo water with explosive force. Its vast, water-soaked body, its dark gray skin dotted with barnacles and scars of time, twists gracefully through the air in a majestic, graceful arc, its enormous, wing-like white pectoral fins fully extended. The golden late-afternoon sun captures every stream of water pouring down from its body, painting the air of spray into a shimmering golden mist, where droplets hover momentarily like countless diamonds. As the whale reaches the apex of its arc and begins its unstoppable plunge back into the water, sending up a deafening wall of white foam, the scene is filled with awe at the power of nature and an atmosphere of breathtaking tranquility and grandeur.", + "seed": 27571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_7377_seed_27571_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04886234952136874, + "vlm_logprob": -0.04886234952136874, + "vlm_yes_no_margin": 3.7249999761581423, + "vlm_token_credit": -0.04886234952136874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03206966072320938 + ], + "opposite_logprob_tokens": [ + -4.157069683074951 + ], + "logprob_reward": -0.03206966072320938, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03206966072320938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22585053741931915, + 0.18524320423603058, + 0.09356874972581863, + 0.07505057007074356, + 0.06589507311582565, + 0.08121376484632492, + 0.0811198279261589, + 0.04647812619805336, + 0.10934537649154663, + 0.036234818398952484 + ], + "credit_map": [ + 0.025010984390974045, + 0.039829738438129425, + 0.016390908509492874, + 0.021300628781318665, + 0.04555540159344673, + 0.03249621018767357, + 0.02189718186855316, + 0.02336948551237583, + 0.02509569749236107, + 0.04280194267630577, + 0.019176911562681198, + 0.014171131886541843, + 0.030944019556045532, + 0.021361736580729485, + 0.016997167840600014, + 0.014694591984152794, + 0.013555539771914482, + 0.014648803509771824, + 0.014221400022506714, + 0.015559110790491104, + 0.011935796588659286, + 0.00803039874881506, + 0.008142721839249134, + 0.0074749840423464775, + 0.012271862477064133, + 0.009891972877085209, + 0.009993384592235088, + 0.007469303905963898, + 0.011238301172852516, + 0.00909274723380804, + 0.009529349394142628, + 0.005563645623624325, + 0.014008156023919582, + 0.011799199506640434, + 0.00917890202254057, + 0.009870577603578568, + 0.0055535235442221165, + 0.004862593021243811, + 0.005373590625822544, + 0.005248530767858028, + 0.011526795104146004, + 0.008836916647851467, + 0.017779573798179626, + 0.017008554190397263, + 0.0071044364012777805, + 0.0046276128850877285, + 0.006274580955505371, + 0.008055299520492554, + 0.013163603842258453, + 0.00814682524651289, + 0.012378296814858913, + 0.011368622072041035, + 0.009451593272387981, + 0.006534409709274769, + 0.009373042732477188, + 0.0107034333050251, + 0.004410000052303076, + 0.005182008258998394, + 0.005873979069292545, + 0.006817916873842478, + 0.006800275761634111, + 0.00435760198161006, + 0.005324226804077625, + 0.007712115999311209, + 0.01139088999480009, + 0.016538042575120926, + 0.010000636801123619, + 0.01053933147341013, + 0.023298267275094986, + 0.007884383201599121, + 0.016673678532242775, + 0.013020148500800133, + 0.005918464157730341, + 0.005956897046416998, + 0.0063381618820130825, + 0.00484652491286397, + 0.004605663474649191, + 0.002602494554594159, + 0.0031842475291341543, + 0.002782368566840887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03206966072320938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807968512177467 + ], + "opposite_logprob_tokens": [ + -4.63480806350708 + ], + "logprob_reward": -0.009807968512177467, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807968512177467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20901785790920258, + 0.04580004885792732, + 0.09176620841026306, + 0.05022088438272476, + 0.06274770200252533, + 0.07712464034557343, + 0.06951243430376053, + 0.11889735609292984, + 0.17512667179107666, + 0.09978613257408142 + ], + "credit_map": [ + 0.023250911384820938, + 0.03664380684494972, + 0.01777487061917782, + 0.024516407400369644, + 0.036513108760118484, + 0.027576789259910583, + 0.02068382501602173, + 0.022058160975575447, + 0.007503335829824209, + 0.006413975264877081, + 0.004454514943063259, + 0.007778793573379517, + 0.004839576780796051, + 0.004378848243504763, + 0.0072938306257128716, + 0.003137177787721157, + 0.012917393818497658, + 0.01150923315435648, + 0.009257127530872822, + 0.01207155454903841, + 0.019246617332100868, + 0.007194892503321171, + 0.01154184341430664, + 0.008027547970414162, + 0.008876635693013668, + 0.007813885807991028, + 0.008276152424514294, + 0.0042756483890116215, + 0.006847502663731575, + 0.0055276453495025635, + 0.005315519869327545, + 0.003287895815446973, + 0.011432909406721592, + 0.007775379344820976, + 0.009976246394217014, + 0.00904775783419609, + 0.007926886901259422, + 0.004859486594796181, + 0.007100739050656557, + 0.004628303926438093, + 0.008952677249908447, + 0.007700592745095491, + 0.010634951293468475, + 0.013937784358859062, + 0.010344427078962326, + 0.005151552148163319, + 0.007600435521453619, + 0.012802221812307835, + 0.011346746236085892, + 0.007454833015799522, + 0.011739026755094528, + 0.011334262788295746, + 0.007961730472743511, + 0.0050566960126161575, + 0.00775267044082284, + 0.0068664695136249065, + 0.011374907568097115, + 0.01520793791860342, + 0.012001355178654194, + 0.018313512206077576, + 0.01920160837471485, + 0.011561660096049309, + 0.011484525166451931, + 0.019751856103539467, + 0.02148679830133915, + 0.022910846397280693, + 0.02027086913585663, + 0.02371014654636383, + 0.031439997255802155, + 0.009975487366318703, + 0.027148045599460602, + 0.0181844811886549, + 0.01807083748281002, + 0.011813730001449585, + 0.01753346063196659, + 0.0166012030094862, + 0.011199045926332474, + 0.008444925770163536, + 0.00904118362814188, + 0.007081751246005297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.009807968512177467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271408647298813 + ], + "opposite_logprob_tokens": [ + -2.127140760421753 + ], + "logprob_reward": -0.1271408647298813, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1271408647298813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14776085317134857, + 0.07926572859287262, + 0.052131082862615585, + 0.0749548003077507, + 0.0819009318947792, + 0.053089406341314316, + 0.27947524189949036, + 0.027472853660583496, + 0.10682711750268936, + 0.09712199121713638 + ], + "credit_map": [ + 0.018624112010002136, + 0.024934522807598114, + 0.010424371808767319, + 0.013504639267921448, + 0.019122378900647163, + 0.018748870119452477, + 0.024532491341233253, + 0.017869461327791214, + 0.010752561502158642, + 0.014388876967132092, + 0.008432255126535892, + 0.005731948185712099, + 0.01281935628503561, + 0.007917472161352634, + 0.008694473654031754, + 0.01052878424525261, + 0.007270786911249161, + 0.007854850962758064, + 0.004519723355770111, + 0.006694206967949867, + 0.01140066608786583, + 0.004819346126168966, + 0.005895568057894707, + 0.0036759336944669485, + 0.013025019317865372, + 0.010761730372905731, + 0.013300507329404354, + 0.007175152190029621, + 0.008794874884188175, + 0.010160041972994804, + 0.006432604510337114, + 0.005304872989654541, + 0.01972547546029091, + 0.017857547849416733, + 0.008931695483624935, + 0.011160334572196007, + 0.007495939265936613, + 0.006166422739624977, + 0.006610839162021875, + 0.00395267503336072, + 0.00669948011636734, + 0.004240700043737888, + 0.008981365710496902, + 0.010024107061326504, + 0.008033971302211285, + 0.003846066305413842, + 0.00589550519362092, + 0.005368211772292852, + 0.05073392018675804, + 0.025844689458608627, + 0.05454308167099953, + 0.03888455778360367, + 0.024030359461903572, + 0.02075195126235485, + 0.03126528859138489, + 0.03342139720916748, + 0.002811277750879526, + 0.004394069314002991, + 0.0024349933955818415, + 0.00357696320861578, + 0.00430448679253459, + 0.002048729220405221, + 0.0034664159175008535, + 0.0044359187595546246, + 0.010717959143221378, + 0.01572353020310402, + 0.007967422716319561, + 0.008063258603215218, + 0.02598566561937332, + 0.007919912226498127, + 0.010915582068264484, + 0.0195337925106287, + 0.013089874759316444, + 0.01303994283080101, + 0.018376577645540237, + 0.009873415343463421, + 0.010787762701511383, + 0.01261440385133028, + 0.010688506998121738, + 0.008651507087051868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.1271408647298813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003815472126007 + ], + "opposite_logprob_tokens": [ + -2.695038080215454 + ], + "logprob_reward": -0.07003815472126007, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003815472126007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1645049750804901, + 0.22773751616477966, + 0.050171758979558945, + 0.04181985929608345, + 0.0490831658244133, + 0.07421047240495682, + 0.161650612950325, + 0.09293206036090851, + 0.0834236592054367, + 0.0544658787548542 + ], + "credit_map": [ + 0.01791621930897236, + 0.033778876066207886, + 0.013777181506156921, + 0.01839098520576954, + 0.024116618558764458, + 0.023587817326188087, + 0.01704748533666134, + 0.015889788046479225, + 0.03215838223695755, + 0.04848599433898926, + 0.021452410146594048, + 0.02688891813158989, + 0.04102421551942825, + 0.02596944011747837, + 0.021041985601186752, + 0.010716171935200691, + 0.0073007442988455296, + 0.008813122287392616, + 0.004080542828887701, + 0.005444646812975407, + 0.010136827826499939, + 0.004902561195194721, + 0.006227436475455761, + 0.0032658788841217756, + 0.006594175472855568, + 0.005528001580387354, + 0.004996000789105892, + 0.0039922031573951244, + 0.006821522489190102, + 0.005565437953919172, + 0.004790542647242546, + 0.003531971015036106, + 0.011254314333200455, + 0.008079607039690018, + 0.007104960735887289, + 0.00899877492338419, + 0.002778390422463417, + 0.0034913632553070784, + 0.004132071975618601, + 0.00324367918074131, + 0.008946421556174755, + 0.008445437997579575, + 0.013793595135211945, + 0.012433466501533985, + 0.0069231088273227215, + 0.006137022282928228, + 0.009311458095908165, + 0.008219962939620018, + 0.030225910246372223, + 0.01569998450577259, + 0.03600810840725899, + 0.019037261605262756, + 0.01196080818772316, + 0.012581193819642067, + 0.017064260318875313, + 0.019073085859417915, + 0.010307018645107746, + 0.010403615422546864, + 0.01299735065549612, + 0.014568665996193886, + 0.01393705140799284, + 0.0059920125640928745, + 0.009776485152542591, + 0.014949866570532322, + 0.00758341234177351, + 0.012657828629016876, + 0.00812317430973053, + 0.008696255274116993, + 0.020868580788373947, + 0.0067407675087451935, + 0.00984665285795927, + 0.008906986564397812, + 0.007210212759673595, + 0.00744536891579628, + 0.008755579590797424, + 0.006726498249918222, + 0.007571076974272728, + 0.00581005122512579, + 0.005545112304389477, + 0.005401979200541973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.07003815472126007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1878775954246521, + 0.11731664091348648, + 0.05046571418642998, + 0.09195970743894577, + 0.2283044308423996, + 0.1094500720500946, + 0.10562866926193237, + 0.019976213574409485, + 0.06492416560649872, + 0.024096811190247536 + ], + "credit_map": [ + 0.023186447098851204, + 0.03412482887506485, + 0.017639964818954468, + 0.020229080691933632, + 0.025922106578946114, + 0.022197214886546135, + 0.023734215646982193, + 0.020843738690018654, + 0.015120333060622215, + 0.02526649832725525, + 0.011665395461022854, + 0.01382028590887785, + 0.02020934410393238, + 0.012891769409179688, + 0.011379104107618332, + 0.0069639128632843494, + 0.008683878928422928, + 0.007917397655546665, + 0.005860086064785719, + 0.0052946885116398335, + 0.006907283794134855, + 0.0052373092621564865, + 0.005136201158165932, + 0.005428867880254984, + 0.015149971470236778, + 0.014804919250309467, + 0.008880448527634144, + 0.010991004295647144, + 0.011950520798563957, + 0.012205008417367935, + 0.008733781054615974, + 0.009244059212505817, + 0.0440983846783638, + 0.03744172677397728, + 0.035223305225372314, + 0.04691937938332558, + 0.019429786130785942, + 0.017815247178077698, + 0.015696734189987183, + 0.01167986635118723, + 0.014403234235942364, + 0.01050986535847187, + 0.015735793858766556, + 0.02313712053000927, + 0.014652998186647892, + 0.008408603258430958, + 0.01250960398465395, + 0.010092858225107193, + 0.023050582036376, + 0.0077970391139388084, + 0.026685088872909546, + 0.012341572903096676, + 0.008224006742238998, + 0.006455490365624428, + 0.009896739386022091, + 0.011178153567016125, + 0.001918548601679504, + 0.002781219780445099, + 0.0018181113991886377, + 0.0025492426939308643, + 0.0034820749424397945, + 0.0015465623000636697, + 0.0025956169702112675, + 0.0032848373521119356, + 0.007376756053417921, + 0.00907802488654852, + 0.005105108022689819, + 0.005168677773326635, + 0.014342564158141613, + 0.00492228614166379, + 0.006731255911290646, + 0.01219948846846819, + 0.0033799181692302227, + 0.003142470261082053, + 0.003984942100942135, + 0.003480833489447832, + 0.003018979448825121, + 0.002827175660058856, + 0.002251860685646534, + 0.0020106325391680002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.005255098920315504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long shot shot from near the surface, the camera tilts steadily upward to track a massive humpback whale as it breaks through the deep indigo water with explosive force. Its vast, water-soaked body, its dark gray skin dotted with barnacles and scars of time, twists gracefully through the air in a majestic, graceful arc, its enormous, wing-like white pectoral fins fully extended. The golden late-afternoon sun captures every stream of water pouring down from its body, painting the air of spray into a shimmering golden mist, where droplets hover momentarily like countless diamonds. As the whale reaches the apex of its arc and begins its unstoppable plunge back into the water, sending up a deafening wall of white foam, the scene is filled with awe at the power of nature and an atmosphere of breathtaking tranquility and grandeur.", + "seed": 27572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_7377_seed_27572_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09645925927907228, + "vlm_logprob": -0.09645925927907228, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.09645925927907228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06340223550796509 + ], + "opposite_logprob_tokens": [ + -3.1884021759033203 + ], + "logprob_reward": -0.06340223550796509, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06340223550796509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10971327126026154, + 0.522262454032898, + 0.0727127343416214, + 0.04936577007174492, + 0.03084350749850273, + 0.10072077065706253, + 0.017193011939525604, + 0.028177861124277115, + 0.0185433067381382, + 0.05046726018190384 + ], + "credit_map": [ + 0.015046318992972374, + 0.009210930205881596, + 0.01124102994799614, + 0.013435271568596363, + 0.010872360318899155, + 0.017179151996970177, + 0.018423454836010933, + 0.014304766431450844, + 0.019729815423488617, + 0.0087975412607193, + 0.19182874262332916, + 0.01818075217306614, + 0.01209423877298832, + 0.0390467643737793, + 0.033508941531181335, + 0.19907574355602264, + 0.007165517657995224, + 0.010504495352506638, + 0.011198117397725582, + 0.010030407458543777, + 0.008782211691141129, + 0.00482431473210454, + 0.008281421847641468, + 0.011926251463592052, + 0.005146407056599855, + 0.0038488060235977173, + 0.014626538380980492, + 0.011541514657437801, + 0.0027598224114626646, + 0.0033437227830290794, + 0.004075308330357075, + 0.004023651592433453, + 0.0038572880439460278, + 0.003986374009400606, + 0.0059575834311544895, + 0.004038122016936541, + 0.0029087732546031475, + 0.0021138181909918785, + 0.004150169901549816, + 0.003831378649920225, + 0.01990658789873123, + 0.010730203241109848, + 0.021599257364869118, + 0.013785144314169884, + 0.006135471165180206, + 0.005860462319105864, + 0.009197802282869816, + 0.013505848124623299, + 0.002312401309609413, + 0.001398478983901441, + 0.004655370954424143, + 0.0024538771249353886, + 0.0013653978239744902, + 0.001510603935457766, + 0.0016126781702041626, + 0.0018842036370187998, + 0.0039005784783512354, + 0.004397318698465824, + 0.005085616372525692, + 0.0035976844374090433, + 0.002590862102806568, + 0.002066468819975853, + 0.0027470081113278866, + 0.003792325733229518, + 0.0023600030690431595, + 0.0030539829749614, + 0.0029890225268900394, + 0.0040892125107347965, + 0.0016740580322220922, + 0.0012876857072114944, + 0.001658931840211153, + 0.0014304100768640637, + 0.008161398582160473, + 0.009222042746841908, + 0.008816100656986237, + 0.006784067023545504, + 0.004930329043418169, + 0.0033506143372505903, + 0.005033820401877165, + 0.004168888553977013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.06340223550796509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901555299758911 + ], + "opposite_logprob_tokens": [ + -2.5790154933929443 + ], + "logprob_reward": -0.07901555299758911, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901555299758911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052831631153821945, + 0.49399176239967346, + 0.05602346360683441, + 0.10749740898609161, + 0.12201470136642456, + 0.05351601913571358, + 0.04949203506112099, + 0.029082275927066803, + 0.010175701230764389, + 0.025375017896294594 + ], + "credit_map": [ + 0.0035224156454205513, + 0.006171492859721184, + 0.007196735590696335, + 0.008490607142448425, + 0.006359852384775877, + 0.006002027075737715, + 0.007093191146850586, + 0.007995306514203548, + 0.01839715987443924, + 0.008681558072566986, + 0.18485510349273682, + 0.018066171556711197, + 0.009387478232383728, + 0.03636300191283226, + 0.02970845252275467, + 0.18853282928466797, + 0.005421947222203016, + 0.008418982848525047, + 0.007936570793390274, + 0.00875094160437584, + 0.005567833781242371, + 0.0038227050099521875, + 0.006687389221042395, + 0.009417093358933926, + 0.010607035830616951, + 0.006686028093099594, + 0.035558588802814484, + 0.02686336264014244, + 0.005186803638935089, + 0.006958332844078541, + 0.008181427605450153, + 0.007455827668309212, + 0.02224617637693882, + 0.014348814263939857, + 0.01511469017714262, + 0.016492923721671104, + 0.01141450461000204, + 0.020160207524895668, + 0.008360136300325394, + 0.013877243734896183, + 0.00941700953990221, + 0.006173449568450451, + 0.011396845802664757, + 0.007089925929903984, + 0.0036688661202788353, + 0.0033291755244135857, + 0.005617532879114151, + 0.006823207717388868, + 0.008586132898926735, + 0.006642378866672516, + 0.007304062135517597, + 0.006395306438207626, + 0.004707517102360725, + 0.004063017200678587, + 0.005549403373152018, + 0.006244211457669735, + 0.004496469162404537, + 0.005630555562674999, + 0.004483858123421669, + 0.004162595607340336, + 0.002263591857627034, + 0.0022373562678694725, + 0.0033465398009866476, + 0.0024613081477582455, + 0.0011445548152551055, + 0.0015002675354480743, + 0.00164525315631181, + 0.00212301523424685, + 0.000734379980713129, + 0.0008988137706182897, + 0.0012627061223611236, + 0.0008667101501487195, + 0.003325637197121978, + 0.0037462005857378244, + 0.0047780415043234825, + 0.0029840620700269938, + 0.002572357887402177, + 0.0026583876460790634, + 0.0024053258821368217, + 0.0029050037264823914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.07901555299758911 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271701455116272 + ], + "opposite_logprob_tokens": [ + -2.1271700859069824 + ], + "logprob_reward": -0.1271701455116272, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1271701455116272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08190546184778214, + 0.11417131870985031, + 0.1710664927959442, + 0.08721546083688736, + 0.15156190097332, + 0.14441964030265808, + 0.12535825371742249, + 0.03578205034136772, + 0.03633929044008255, + 0.05218009278178215 + ], + "credit_map": [ + 0.010183748789131641, + 0.010164282284677029, + 0.008085044100880623, + 0.01055760309100151, + 0.010338295251131058, + 0.01160690188407898, + 0.01168152503669262, + 0.009288066998124123, + 0.00751864816993475, + 0.008796206675469875, + 0.023478347808122635, + 0.015034817159175873, + 0.01228715106844902, + 0.009589709341526031, + 0.013720992021262646, + 0.0237454641610384, + 0.016281459480524063, + 0.01745760440826416, + 0.02185894176363945, + 0.020129196345806122, + 0.02412550523877144, + 0.016697147861123085, + 0.02792026288807392, + 0.026596391573548317, + 0.009071652777493, + 0.009984995238482952, + 0.015066524967551231, + 0.015188371762633324, + 0.006503789220005274, + 0.007867284119129181, + 0.010690547525882721, + 0.012842291034758091, + 0.017855104058980942, + 0.02173798158764839, + 0.019629577174782753, + 0.01669493317604065, + 0.016579484567046165, + 0.022372422739863396, + 0.01648886501789093, + 0.020203547552227974, + 0.019032340496778488, + 0.015989258885383606, + 0.02652592584490776, + 0.020845284685492516, + 0.016439780592918396, + 0.012666883878409863, + 0.016819342970848083, + 0.016100827604532242, + 0.018045486882328987, + 0.016216399148106575, + 0.022035421803593636, + 0.01597057841718197, + 0.011436988599598408, + 0.009093538857996464, + 0.013644141145050526, + 0.018915701657533646, + 0.005399420391768217, + 0.004568038508296013, + 0.007325024344027042, + 0.005627208389341831, + 0.0036771248560398817, + 0.002958165016025305, + 0.0026847049593925476, + 0.003542365040630102, + 0.0055244239047169685, + 0.006247047800570726, + 0.006063648033887148, + 0.007974495179951191, + 0.0032178291585296392, + 0.0024609097745269537, + 0.002482103882357478, + 0.002368834102526307, + 0.007654090877622366, + 0.007641488220542669, + 0.009663963690400124, + 0.006974558811634779, + 0.005511854309588671, + 0.004691831301897764, + 0.004125241190195084, + 0.005917069967836142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.1271701455116272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162324607372284 + ], + "opposite_logprob_tokens": [ + -1.7016232013702393 + ], + "logprob_reward": -0.20162324607372284, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162324607372284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07545152306556702, + 0.47183793783187866, + 0.07591280341148376, + 0.053966693580150604, + 0.06719702482223511, + 0.06100721284747124, + 0.03367293253540993, + 0.10949159413576126, + 0.024035148322582245, + 0.027427075430750847 + ], + "credit_map": [ + 0.011703544296324253, + 0.007592030800879002, + 0.00880347564816475, + 0.008828209713101387, + 0.008182554505765438, + 0.010365701280534267, + 0.010749464854598045, + 0.009226547554135323, + 0.01730278506875038, + 0.006167517974972725, + 0.1757606714963913, + 0.015438783913850784, + 0.008848367258906364, + 0.03515796735882759, + 0.028250984847545624, + 0.1849108785390854, + 0.007272924296557903, + 0.010242783464491367, + 0.011309144087135792, + 0.009951232932507992, + 0.009708954021334648, + 0.007753158453851938, + 0.008851542137563229, + 0.010823074728250504, + 0.006606522016227245, + 0.0048020933754742146, + 0.014207158237695694, + 0.009437158703804016, + 0.003461492247879505, + 0.005628436803817749, + 0.005101035349071026, + 0.0047228061594069, + 0.010137470439076424, + 0.00873861089348793, + 0.008229116909205914, + 0.0072829630225896835, + 0.0074194432236254215, + 0.009917094372212887, + 0.006528952158987522, + 0.008943386375904083, + 0.009276400320231915, + 0.005292029585689306, + 0.01024439837783575, + 0.009007016196846962, + 0.006486337166279554, + 0.006877753417938948, + 0.006118724588304758, + 0.007704563904553652, + 0.003976549953222275, + 0.004139064345508814, + 0.005950684193521738, + 0.004379904363304377, + 0.0040560499764978886, + 0.0027109908405691385, + 0.004003523848950863, + 0.004456169903278351, + 0.01185363158583641, + 0.013605829328298569, + 0.02113650180399418, + 0.015466901473701, + 0.011900145560503006, + 0.008347424678504467, + 0.015075982548296452, + 0.012105190195143223, + 0.0025855819694697857, + 0.0037762110587209463, + 0.003436693921685219, + 0.0055004628375172615, + 0.0019458383321762085, + 0.0014479972887784243, + 0.0032973343040794134, + 0.002045030239969492, + 0.0038464206736534834, + 0.00402342202141881, + 0.005728418473154306, + 0.004420139826834202, + 0.0023798628244549036, + 0.0022765472531318665, + 0.0025486620143055916, + 0.0022036039736121893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.20162324607372284 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085116304457188 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011085116304457188, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085116304457188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0766688883304596, + 0.2890499532222748, + 0.1331784725189209, + 0.019492097198963165, + 0.026879815384745598, + 0.06743782758712769, + 0.25690630078315735, + 0.017342032864689827, + 0.09404603391885757, + 0.018998604267835617 + ], + "credit_map": [ + 0.008178829215466976, + 0.01075418759137392, + 0.007548105902969837, + 0.012666316702961922, + 0.008065106347203255, + 0.008253959938883781, + 0.01016591489315033, + 0.011036469601094723, + 0.010433395393192768, + 0.004392387345433235, + 0.10649999231100082, + 0.010448470711708069, + 0.007539627142250538, + 0.020883755758404732, + 0.016558537259697914, + 0.1122937798500061, + 0.012395588681101799, + 0.017282631248235703, + 0.017852772027254105, + 0.019785836338996887, + 0.014972005970776081, + 0.011469871737062931, + 0.020154070109128952, + 0.019265692681074142, + 0.0017087535234168172, + 0.0016883405623957515, + 0.0043957410380244255, + 0.00437389500439167, + 0.0014076955849304795, + 0.0017117559909820557, + 0.001958983251824975, + 0.0022469302639365196, + 0.004197678063064814, + 0.0036503009032458067, + 0.004947675857692957, + 0.003051436971873045, + 0.0023607045877724886, + 0.0019356798147782683, + 0.00345211336389184, + 0.003284223610535264, + 0.008725007064640522, + 0.007788165006786585, + 0.011714999563992023, + 0.009466731920838356, + 0.005675274413079023, + 0.007076396141201258, + 0.005499714054167271, + 0.011491537094116211, + 0.03991471603512764, + 0.03647419437766075, + 0.03666995093226433, + 0.03294497728347778, + 0.023540424183011055, + 0.020311914384365082, + 0.03048129566013813, + 0.03656882047653198, + 0.002600503619760275, + 0.0026839340571314096, + 0.002695042174309492, + 0.00266172643750906, + 0.0016107730334624648, + 0.0012778050731867552, + 0.00179993303027004, + 0.002012314274907112, + 0.009312517009675503, + 0.020005274564027786, + 0.020646994933485985, + 0.014888392761349678, + 0.009161208756268024, + 0.006235809065401554, + 0.008366218768060207, + 0.0054296161979436874, + 0.00323965260758996, + 0.0031189844012260437, + 0.0040518296882510185, + 0.0026380429044365883, + 0.001857336494140327, + 0.001308953738771379, + 0.001313281012699008, + 0.001470523769967258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.011085116304457188 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long shot shot from near the surface, the camera tilts steadily upward to track a massive humpback whale as it breaks through the deep indigo water with explosive force. Its vast, water-soaked body, its dark gray skin dotted with barnacles and scars of time, twists gracefully through the air in a majestic, graceful arc, its enormous, wing-like white pectoral fins fully extended. The golden late-afternoon sun captures every stream of water pouring down from its body, painting the air of spray into a shimmering golden mist, where droplets hover momentarily like countless diamonds. As the whale reaches the apex of its arc and begins its unstoppable plunge back into the water, sending up a deafening wall of white foam, the scene is filled with awe at the power of nature and an atmosphere of breathtaking tranquility and grandeur.", + "seed": 27573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_13/index_7377_seed_27573_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.037037721974775194, + "vlm_logprob": -0.037037721974775194, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.037037721974775194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03928280249238014 + ], + "opposite_logprob_tokens": [ + -3.78928279876709 + ], + "logprob_reward": -0.03928280249238014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03928280249238014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12236032634973526, + 0.1382983773946762, + 0.10356741398572922, + 0.06143699958920479, + 0.08891359716653824, + 0.053201086819171906, + 0.08599281311035156, + 0.02128821425139904, + 0.2728385925292969, + 0.05210261791944504 + ], + "credit_map": [ + 0.014231855981051922, + 0.010897013358771801, + 0.017849067226052284, + 0.014963259920477867, + 0.014657868072390556, + 0.018779857084155083, + 0.0172255989164114, + 0.013755793683230877, + 0.011366009712219238, + 0.012365937232971191, + 0.03867568075656891, + 0.008489342406392097, + 0.01348112802952528, + 0.012648104690015316, + 0.027695175260305405, + 0.013576997444033623, + 0.012522423639893532, + 0.007591722998768091, + 0.020241130143404007, + 0.005033882334828377, + 0.032047707587480545, + 0.01099520642310381, + 0.007084288168698549, + 0.00805105734616518, + 0.0072363633662462234, + 0.007689499296247959, + 0.016813477501273155, + 0.007807867135852575, + 0.00969737209379673, + 0.00415613641962409, + 0.0028326765168458223, + 0.0052036102861166, + 0.008341762237250805, + 0.01927926577627659, + 0.013252047821879387, + 0.007766224909573793, + 0.01341505628079176, + 0.009403235279023647, + 0.005978496745228767, + 0.011477507650852203, + 0.007180585991591215, + 0.006271463353186846, + 0.004937490448355675, + 0.008551718667149544, + 0.00837794691324234, + 0.008442620746791363, + 0.005107334349304438, + 0.004331924486905336, + 0.011955655179917812, + 0.0131163215264678, + 0.01178736798465252, + 0.006306536495685577, + 0.013910546898841858, + 0.012193534523248672, + 0.007532780524343252, + 0.009190071374177933, + 0.0033830387983471155, + 0.0021088379435241222, + 0.0030753682367503643, + 0.001865749480202794, + 0.002209804020822048, + 0.0035715934354811907, + 0.002174025634303689, + 0.0028997960034757853, + 0.0958016887307167, + 0.008373583666980267, + 0.010574296116828918, + 0.009978693909943104, + 0.01020408421754837, + 0.12035176903009415, + 0.01087765209376812, + 0.0066768410615623, + 0.006018022540956736, + 0.004917310085147619, + 0.007074763532727957, + 0.004134542308747768, + 0.0067922924645245075, + 0.008231708779931068, + 0.00642487732693553, + 0.008509100414812565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.03928280249238014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607298240065575 + ], + "opposite_logprob_tokens": [ + -3.8956072330474854 + ], + "logprob_reward": -0.020607298240065575, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607298240065575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1253073513507843, + 0.12869878113269806, + 0.08286616206169128, + 0.04664449021220207, + 0.10077101737260818, + 0.07980253547430038, + 0.08533937484025955, + 0.020232174545526505, + 0.19484373927116394, + 0.1354943960905075 + ], + "credit_map": [ + 0.014863650314509869, + 0.010529250837862492, + 0.01910383254289627, + 0.014228252694010735, + 0.018336838111281395, + 0.014513304457068443, + 0.018985429778695107, + 0.014746779575943947, + 0.013944010250270367, + 0.006750347558408976, + 0.04614843428134918, + 0.009116428904235363, + 0.009459968656301498, + 0.009485112503170967, + 0.02130427397787571, + 0.01249020267277956, + 0.015629705041646957, + 0.0070047760382294655, + 0.011773636564612389, + 0.0068862042389810085, + 0.017988013103604317, + 0.009527681395411491, + 0.00639093155041337, + 0.007665218785405159, + 0.004758700728416443, + 0.004274426493793726, + 0.007663913071155548, + 0.006380645092576742, + 0.0057861777022480965, + 0.004655213560909033, + 0.006685218308120966, + 0.006440198980271816, + 0.009357220493257046, + 0.015516825020313263, + 0.012662647292017937, + 0.008478340692818165, + 0.014298953115940094, + 0.013353148475289345, + 0.008530517108738422, + 0.01857336238026619, + 0.010590041056275368, + 0.010785023681819439, + 0.006457550451159477, + 0.01142116542905569, + 0.012604858726263046, + 0.012633445672690868, + 0.006855090148746967, + 0.008455360308289528, + 0.016020290553569794, + 0.009405286982655525, + 0.010250704362988472, + 0.0055685630068182945, + 0.01033327542245388, + 0.01734502613544464, + 0.006487537641078234, + 0.009928693063557148, + 0.003075147047638893, + 0.002362450584769249, + 0.0024944893084466457, + 0.001814546762034297, + 0.0018799945246428251, + 0.0036617799196392298, + 0.0020631439983844757, + 0.00288062309846282, + 0.07006311416625977, + 0.00612198282033205, + 0.006100798957049847, + 0.005121604539453983, + 0.006591848097741604, + 0.08757134526968002, + 0.007904624566435814, + 0.005368400365114212, + 0.0174705907702446, + 0.015744514763355255, + 0.01505748089402914, + 0.009955354034900665, + 0.021774593740701675, + 0.030092064291238785, + 0.015577186830341816, + 0.009822613559663296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.020607298240065575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310382157564163 + ], + "opposite_logprob_tokens": [ + -3.1681039333343506 + ], + "logprob_reward": -0.04310382157564163, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310382157564163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06382013857364655, + 0.07378476113080978, + 0.14849761128425598, + 0.40647533535957336, + 0.05291847139596939, + 0.017792807891964912, + 0.031580355018377304, + 0.010555427521467209, + 0.1407385617494583, + 0.05383649095892906 + ], + "credit_map": [ + 0.008018768392503262, + 0.004081318620592356, + 0.009293384850025177, + 0.009695854038000107, + 0.008060025982558727, + 0.00799601897597313, + 0.008187960833311081, + 0.00848680641502142, + 0.007999490946531296, + 0.005352574400603771, + 0.020014077425003052, + 0.00618593767285347, + 0.007223848253488541, + 0.007820461876690388, + 0.007755909115076065, + 0.011432467959821224, + 0.020422114059329033, + 0.010765793733298779, + 0.027639130130410194, + 0.007241336163133383, + 0.048913683742284775, + 0.012919044122099876, + 0.009182557463645935, + 0.011413948610424995, + 0.03165436536073685, + 0.02091672644019127, + 0.1101667732000351, + 0.04003634303808212, + 0.12651674449443817, + 0.012696861289441586, + 0.013203435577452183, + 0.05128408968448639, + 0.005468789022415876, + 0.009853837080299854, + 0.006550924852490425, + 0.005598657298833132, + 0.009412179701030254, + 0.00532170245423913, + 0.0041823736391961575, + 0.006530006416141987, + 0.0023152800276875496, + 0.0022785235196352005, + 0.0015882061561569571, + 0.0024574475828558207, + 0.002984193153679371, + 0.0024896461982280016, + 0.0016004386125132442, + 0.0020790721755474806, + 0.00483853230252862, + 0.0038449338171631098, + 0.0046529704704880714, + 0.0027702897787094116, + 0.004276287741959095, + 0.004335280042141676, + 0.0030492835212498903, + 0.0038127766456454992, + 0.0015311628812924027, + 0.0012880268041044474, + 0.0015592658892273903, + 0.0009174176957458258, + 0.0011474842904135585, + 0.0013158536748960614, + 0.0011009677546098828, + 0.00169524853117764, + 0.05144856870174408, + 0.004056817386299372, + 0.004044975154101849, + 0.003409360535442829, + 0.004649866838008165, + 0.06463854014873505, + 0.004858490079641342, + 0.003631945699453354, + 0.006550281308591366, + 0.00547341164201498, + 0.007908442057669163, + 0.004273509606719017, + 0.008289240300655365, + 0.010282603092491627, + 0.0060988785699009895, + 0.004960125312209129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.04310382157564163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790020003914833 + ], + "opposite_logprob_tokens": [ + -2.5790019035339355 + ], + "logprob_reward": -0.0790020003914833, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790020003914833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06403055787086487, + 0.03575383499264717, + 0.14303135871887207, + 0.1602870225906372, + 0.15776799619197845, + 0.11580760776996613, + 0.07396100461483002, + 0.026513567194342613, + 0.16573454439640045, + 0.057112544775009155 + ], + "credit_map": [ + 0.005718796048313379, + 0.00596681097522378, + 0.006496758200228214, + 0.006205781362950802, + 0.005936207249760628, + 0.005145586561411619, + 0.014869996346533298, + 0.013690615072846413, + 0.003591610351577401, + 0.0032715382985770702, + 0.005141764413565397, + 0.0030743989627808332, + 0.0040507204830646515, + 0.005548855289816856, + 0.006309837102890015, + 0.0047651068307459354, + 0.01096945907920599, + 0.00856764055788517, + 0.04652136564254761, + 0.012235353700816631, + 0.036688484251499176, + 0.011018594726920128, + 0.00783704686909914, + 0.00919341016560793, + 0.012779848650097847, + 0.010334182530641556, + 0.037244174629449844, + 0.013161104172468185, + 0.050754670053720474, + 0.006436611991375685, + 0.007853216491639614, + 0.021723207086324692, + 0.014796358533203602, + 0.02778441831469536, + 0.015409709885716438, + 0.021086636930704117, + 0.0212920643389225, + 0.02064678817987442, + 0.01817907951772213, + 0.01857292465865612, + 0.015905141830444336, + 0.01674264669418335, + 0.008242425508797169, + 0.016161566600203514, + 0.018307285383343697, + 0.01655454933643341, + 0.01052898820489645, + 0.013365001417696476, + 0.009943819604814053, + 0.010627955198287964, + 0.010059593245387077, + 0.006030811462551355, + 0.010844188742339611, + 0.01027874555438757, + 0.006918243132531643, + 0.009257650002837181, + 0.003720268839970231, + 0.0039842440746724606, + 0.0034507657401263714, + 0.002640477381646633, + 0.0026836609467864037, + 0.0028479560278356075, + 0.0035572818014770746, + 0.0036289095878601074, + 0.05894540995359421, + 0.0047118328511714935, + 0.007454493548721075, + 0.005311037879437208, + 0.005757172126322985, + 0.07393934577703476, + 0.005434154067188501, + 0.004181094001978636, + 0.006913261488080025, + 0.006592524703592062, + 0.005293204449117184, + 0.004934945609420538, + 0.008238472044467926, + 0.012034129351377487, + 0.006723247934132814, + 0.006382754538208246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.0790020003914833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle long shot taken from near the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward to track a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin dark gray, dotted with barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the whale's enormous white pectoral fins fully extended like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden late-afternoon sun creating a shimmering golden mist of water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0743912011384964, + 0.05926206335425377, + 0.02186298370361328, + 0.31707754731178284, + 0.01410861685872078, + 0.011989539489150047, + 0.015550944022834301, + 0.009641007520258427, + 0.43607768416404724, + 0.04003845155239105 + ], + "credit_map": [ + 0.006914875470101833, + 0.004764759913086891, + 0.00890065636485815, + 0.01078137569129467, + 0.00964338518679142, + 0.008047650568187237, + 0.014357348904013634, + 0.010981149971485138, + 0.0057086944580078125, + 0.00599904078990221, + 0.012468373402953148, + 0.00460202619433403, + 0.007336861453950405, + 0.00667736167088151, + 0.00979443546384573, + 0.006675266195088625, + 0.004307415336370468, + 0.0015953443944454193, + 0.0012570236576721072, + 0.0010491542052477598, + 0.006086644716560841, + 0.0032717823050916195, + 0.002192507730796933, + 0.0021031114738434553, + 0.024231739342212677, + 0.01703750155866146, + 0.08059802651405334, + 0.027803519740700722, + 0.10305406153202057, + 0.010499578900635242, + 0.01145803090184927, + 0.04239508509635925, + 0.0015017882687970996, + 0.003109989920631051, + 0.001527224201709032, + 0.0010234484216198325, + 0.002023237757384777, + 0.0021459010895341635, + 0.0011110164923593402, + 0.0016660114051774144, + 0.0017349794507026672, + 0.0016550520667806268, + 0.0013489428674802184, + 0.001192666357383132, + 0.001917050452902913, + 0.0016747638583183289, + 0.0008492798660881817, + 0.0016168035799637437, + 0.0036606185603886843, + 0.001115029794164002, + 0.0012180577032268047, + 0.0008977151592262089, + 0.0014854386681690812, + 0.004411932080984116, + 0.0008008594159036875, + 0.0019612922333180904, + 0.001502892468124628, + 0.0014056647196412086, + 0.0014120362466201186, + 0.0008695055148564279, + 0.0009873481467366219, + 0.001463047694414854, + 0.0007971823797561228, + 0.0012033303501084447, + 0.1597381830215454, + 0.012831123545765877, + 0.011819712817668915, + 0.009782986715435982, + 0.014972815290093422, + 0.20105497539043427, + 0.01513481792062521, + 0.010743063874542713, + 0.004597849678248167, + 0.0034812618978321552, + 0.005809691734611988, + 0.0035485411062836647, + 0.0059961858205497265, + 0.006991337984800339, + 0.00480202492326498, + 0.004811559338122606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.00319268717430532 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_14/index_tensor([9417, 9417])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_14/index_tensor([9417, 9417])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1f22d9709591b6c2fd970e567ba9ddc42586aba1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_14/index_tensor([9417, 9417])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In Hayao Miyazaki's captivating, hand-drawn Ghibli style, a static, intimate close-up captures a pair of slender, young hands cupped together in a tranquil underwater realm, meticulously detailed. Steadily and deliberately, these hands guide two shy, bioluminescent sea creatures toward each other. One creature nestles in the palm of the left hand, radiating a soft, ethereal azure glow. Its gelatinous body is shaped like a small, translucent bell, with inquisitive black eyes. The other, guided by the gentle fingers of the right hand, radiates a warm golden light. Its form is a slightly intricate, star-shaped droplet that trembles timidly at approach. As the two creatures cautiously approach, their respective radiance intensifies, casting a magical, pulsating glow that illuminates the subtle details of the characters' skin and the shimmering water particles between them. The immediate background is a lush underwater world with soft, painterly textures; out-of-focus, vibrant corals in soft shades of pink and deep purple sway gently, while the tentacles of a large, soft-edged sea anemone ripple in the subtle currents. The scene is primarily illuminated by a magical, blended blue and gold light emanating from the creatures, creating soft highlights and gentle shadows on the hands, enveloping the moment in an atmosphere of pure, quietly enchanting beauty and capturing a profound sense of wonder through the sheer tenderness of the movement.", + "seed": 17261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_14/index_9417_seed_17261_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.006025435472838581, + "vlm_logprob": -0.006025435472838581, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.006025435472838581, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0136126559227705 + ], + "opposite_logprob_tokens": [ + -4.513612747192383 + ], + "logprob_reward": -0.0136126559227705, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0136126559227705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08090958744287491, + 0.06264439970254898, + 0.07353267073631287, + 0.05262696370482445, + 0.04371387138962746, + 0.21050426363945007, + 0.06625554710626602, + 0.08055908977985382, + 0.24525849521160126, + 0.08399510383605957 + ], + "credit_map": [ + 0.010762104764580727, + 0.0076720924116671085, + 0.0061546433717012405, + 0.012738768942654133, + 0.015312675386667252, + 0.008429699577391148, + 0.008877416141331196, + 0.010962185449898243, + 0.008224446326494217, + 0.005175263620913029, + 0.004982147831469774, + 0.008979576639831066, + 0.015603938139975071, + 0.007072095759212971, + 0.006515908986330032, + 0.0060910191386938095, + 0.009032178670167923, + 0.0053455461747944355, + 0.007823311723768711, + 0.01089003961533308, + 0.013010548427700996, + 0.0053722369484603405, + 0.009527531452476978, + 0.012531278654932976, + 0.008482780307531357, + 0.0049552880227565765, + 0.005788277834653854, + 0.0053151268512010574, + 0.0064064981415867805, + 0.0036579498555511236, + 0.011239472776651382, + 0.006781570613384247, + 0.004463717807084322, + 0.003524221247062087, + 0.0035584368743002415, + 0.00566253112629056, + 0.0036676593590527773, + 0.005052146967500448, + 0.008081748150289059, + 0.009703406132757664, + 0.01720234379172325, + 0.010759581811726093, + 0.007758143823593855, + 0.014443239197134972, + 0.010234818793833256, + 0.014344162307679653, + 0.08227274566888809, + 0.053489215672016144, + 0.006273497827351093, + 0.005224835593253374, + 0.0073202308267354965, + 0.0117656784132123, + 0.0064040618017315865, + 0.0065042925998568535, + 0.012520622462034225, + 0.010242330841720104, + 0.005462434608489275, + 0.005391969345510006, + 0.0065342276357114315, + 0.006619257386773825, + 0.005513621959835291, + 0.007499411702156067, + 0.017425715923309326, + 0.026112450286746025, + 0.028020339086651802, + 0.02434905432164669, + 0.037107255309820175, + 0.041561517864465714, + 0.02113983780145645, + 0.025281811133027077, + 0.020323164761066437, + 0.04747552424669266, + 0.007447370793670416, + 0.008097060024738312, + 0.009333244524896145, + 0.008221819996833801, + 0.013155377469956875, + 0.011237045750021935, + 0.010914597660303116, + 0.015588587149977684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.0136126559227705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746372207999229 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746372207999229, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746372207999229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19578658044338226, + 0.08913356065750122, + 0.0791359692811966, + 0.028361007571220398, + 0.05413699895143509, + 0.1125061959028244, + 0.07355400174856186, + 0.06867682188749313, + 0.1340508908033371, + 0.16465795040130615 + ], + "credit_map": [ + 0.029379339888691902, + 0.01707819290459156, + 0.017230713739991188, + 0.022338448092341423, + 0.020599976181983948, + 0.01540573313832283, + 0.027537524700164795, + 0.04621665179729462, + 0.01416624803096056, + 0.007265136577188969, + 0.012120982632040977, + 0.011346504092216492, + 0.012288445606827736, + 0.009076284244656563, + 0.013552786782383919, + 0.009317176416516304, + 0.01177313644438982, + 0.00712245050817728, + 0.010732642374932766, + 0.01379377394914627, + 0.006866335403174162, + 0.006259826477617025, + 0.009515796788036823, + 0.01307199988514185, + 0.004549590405076742, + 0.0026722478214651346, + 0.0026891580782830715, + 0.003112721722573042, + 0.0030139866285026073, + 0.002606745809316635, + 0.005446023773401976, + 0.004270534496754408, + 0.0051997085101902485, + 0.004243358504027128, + 0.004742738790810108, + 0.007727430667728186, + 0.00692824786528945, + 0.006217369809746742, + 0.010022173635661602, + 0.009055973030626774, + 0.008970950730144978, + 0.006982195656746626, + 0.008136470802128315, + 0.00858412403613329, + 0.00616942485794425, + 0.00789527129381895, + 0.04062310978770256, + 0.02514464780688286, + 0.00795359443873167, + 0.004921601619571447, + 0.006827387027442455, + 0.008866195566952229, + 0.005361469928175211, + 0.009223971515893936, + 0.01852417178452015, + 0.011875616386532784, + 0.006792854052037001, + 0.006831809412688017, + 0.009400323033332825, + 0.009090712293982506, + 0.00895506888628006, + 0.005509426351636648, + 0.008811657316982746, + 0.013284971006214619, + 0.013783990405499935, + 0.012828730046749115, + 0.02230946533381939, + 0.020291989669203758, + 0.014306753873825073, + 0.01494696456938982, + 0.014112715609371662, + 0.021470285952091217, + 0.020704220980405807, + 0.01961999572813511, + 0.017582448199391365, + 0.015582423657178879, + 0.029382547363638878, + 0.015641631558537483, + 0.027747642248868942, + 0.01839703880250454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.006746372207999229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193043638020754 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.003193043638020754, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193043638020754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0809960588812828, + 0.057599715888500214, + 0.08122406154870987, + 0.045257654041051865, + 0.025090495124459267, + 0.09125415235757828, + 0.043948616832494736, + 0.1960960179567337, + 0.28974729776382446, + 0.08878596127033234 + ], + "credit_map": [ + 0.009226388297975063, + 0.008280621841549873, + 0.00731087289750576, + 0.009051174856722355, + 0.0065401047468185425, + 0.006906175520271063, + 0.011271304450929165, + 0.02240942232310772, + 0.006224975921213627, + 0.004750892054289579, + 0.004475385881960392, + 0.009366641752421856, + 0.008253425359725952, + 0.0058357734233140945, + 0.00854079145938158, + 0.010151829570531845, + 0.012167706154286861, + 0.006372695788741112, + 0.01002869289368391, + 0.012918642722070217, + 0.010953699238598347, + 0.006115570664405823, + 0.010985365137457848, + 0.011681688949465752, + 0.008656270802021027, + 0.005811300594359636, + 0.0060424781404435635, + 0.005704470910131931, + 0.0033877999521791935, + 0.0033362898975610733, + 0.005707825534045696, + 0.006611215882003307, + 0.00388561119325459, + 0.002080376259982586, + 0.0019551031291484833, + 0.0030495484825223684, + 0.001993721816688776, + 0.002018927363678813, + 0.0038746274076402187, + 0.006232581567019224, + 0.011970245279371738, + 0.005956549197435379, + 0.005295171868056059, + 0.00600449088960886, + 0.004795615095645189, + 0.0050229281187057495, + 0.03433114290237427, + 0.017878012731671333, + 0.005057929549366236, + 0.002727925544604659, + 0.003204178297892213, + 0.004720017779618502, + 0.002771135652437806, + 0.004906181246042252, + 0.012846785597503185, + 0.007714465726166964, + 0.03793017193675041, + 0.016860757023096085, + 0.008094139397144318, + 0.013503070920705795, + 0.00947213638573885, + 0.005125242751091719, + 0.06538709253072739, + 0.03972340747714043, + 0.03255218639969826, + 0.028005501255393028, + 0.040934961289167404, + 0.04377570003271103, + 0.0263070035725832, + 0.0298842191696167, + 0.027512310072779655, + 0.060775406658649445, + 0.008511250838637352, + 0.007628426421433687, + 0.012733074836432934, + 0.008795258589088917, + 0.01137924287468195, + 0.009587370790541172, + 0.01874314248561859, + 0.011408200487494469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.003193043638020754 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633760545402765 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004633760545402765, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633760545402765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11065205186605453, + 0.05800824984908104, + 0.06412864476442337, + 0.0434500016272068, + 0.02647869661450386, + 0.04337571933865547, + 0.04735760763287544, + 0.11065518110990524, + 0.18353226780891418, + 0.31236153841018677 + ], + "credit_map": [ + 0.01301424577832222, + 0.00976630114018917, + 0.011648735031485558, + 0.01596018858253956, + 0.012584221549332142, + 0.008871489204466343, + 0.012721465900540352, + 0.02608540654182434, + 0.009317731484770775, + 0.005573805887252092, + 0.007622560020536184, + 0.007927322760224342, + 0.006809374783188105, + 0.0037758667021989822, + 0.010461789555847645, + 0.006519796792417765, + 0.011617643758654594, + 0.006489353254437447, + 0.009746681898832321, + 0.009218767285346985, + 0.004477715119719505, + 0.004127531312406063, + 0.009059080854058266, + 0.00939186941832304, + 0.008453983813524246, + 0.004353987984359264, + 0.0058789742179214954, + 0.005282927770167589, + 0.004111850168555975, + 0.0023861422669142485, + 0.008134868927299976, + 0.004847267176955938, + 0.003199410391971469, + 0.00261327950283885, + 0.0023576882667839527, + 0.003566541476175189, + 0.0030881818383932114, + 0.002502082148566842, + 0.003993079531937838, + 0.005158433225005865, + 0.004924645647406578, + 0.0028582094237208366, + 0.002283307723701, + 0.0036359287332743406, + 0.002333859447389841, + 0.0035054723266512156, + 0.015641214326024055, + 0.008193080313503742, + 0.0060576023533940315, + 0.0044181933626532555, + 0.007364208810031414, + 0.007241395767778158, + 0.002950479509308934, + 0.00307073793374002, + 0.010186905972659588, + 0.00606808764860034, + 0.013806726783514023, + 0.009798354469239712, + 0.010459071956574917, + 0.0121015515178442, + 0.00892574805766344, + 0.006621668115258217, + 0.016739344224333763, + 0.032202716916799545, + 0.019771046936511993, + 0.016666557639837265, + 0.0345032699406147, + 0.02909150905907154, + 0.014951621182262897, + 0.032509759068489075, + 0.013159513473510742, + 0.022878987714648247, + 0.06334732472896576, + 0.03523680940270424, + 0.044335462152957916, + 0.03186321631073952, + 0.032268308103084564, + 0.03173764422535896, + 0.03299129381775856, + 0.04058149456977844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.004633760545402765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019413450499996543 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.0019413450499996543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019413450499996543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10364113748073578, + 0.09432094544172287, + 0.11341343075037003, + 0.03859628364443779, + 0.03575974702835083, + 0.14283022284507751, + 0.042942531406879425, + 0.08515613526105881, + 0.22345805168151855, + 0.11988156288862228 + ], + "credit_map": [ + 0.019053485244512558, + 0.00982503779232502, + 0.012617130763828754, + 0.010377799160778522, + 0.00748105626553297, + 0.006678122095763683, + 0.01171161513775587, + 0.025896890088915825, + 0.013747156597673893, + 0.009643802419304848, + 0.012450629845261574, + 0.014288377948105335, + 0.008416950702667236, + 0.009625560604035854, + 0.014319288544356823, + 0.011829173192381859, + 0.018014393746852875, + 0.010702231898903847, + 0.014946597628295422, + 0.01720217429101467, + 0.012784147635102272, + 0.009412512183189392, + 0.014131492003798485, + 0.016219884157180786, + 0.008585302159190178, + 0.004227719269692898, + 0.004574025981128216, + 0.005058031529188156, + 0.0026248774956911802, + 0.0028154293540865183, + 0.005701575428247452, + 0.005009321495890617, + 0.005658975802361965, + 0.0025714628864079714, + 0.00311636202968657, + 0.004568669945001602, + 0.003102079965174198, + 0.0026577147655189037, + 0.006898806896060705, + 0.007185674738138914, + 0.017784960567951202, + 0.012762767262756824, + 0.008596567437052727, + 0.009525202214717865, + 0.00817814376205206, + 0.008525341749191284, + 0.04743297025561333, + 0.03002428635954857, + 0.00498523423448205, + 0.004116978030651808, + 0.005883036646991968, + 0.006341657135635614, + 0.0037159344647079706, + 0.004555652383714914, + 0.005690926220268011, + 0.007653112057596445, + 0.009451163001358509, + 0.008562101982533932, + 0.01329366210848093, + 0.012779202312231064, + 0.00909604225307703, + 0.016707265749573708, + 0.006702470127493143, + 0.008564233779907227, + 0.023476989939808846, + 0.020347535610198975, + 0.03745153173804283, + 0.03517156094312668, + 0.018429722636938095, + 0.04309212416410446, + 0.018290629610419273, + 0.02719796635210514, + 0.011924355290830135, + 0.01121657993644476, + 0.014976198785007, + 0.011373935267329216, + 0.015255928970873356, + 0.013696824200451374, + 0.023097729310393333, + 0.018340012058615685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.0019413450499996543 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Hayao Miyazaki's captivating, hand-drawn Ghibli style, a static, intimate close-up captures a pair of slender, young hands cupped together in a tranquil underwater realm, meticulously detailed. Steadily and deliberately, these hands guide two shy, bioluminescent sea creatures toward each other. One creature nestles in the palm of the left hand, radiating a soft, ethereal azure glow. Its gelatinous body is shaped like a small, translucent bell, with inquisitive black eyes. The other, guided by the gentle fingers of the right hand, radiates a warm golden light. Its form is a slightly intricate, star-shaped droplet that trembles timidly at approach. As the two creatures cautiously approach, their respective radiance intensifies, casting a magical, pulsating glow that illuminates the subtle details of the characters' skin and the shimmering water particles between them. The immediate background is a lush underwater world with soft, painterly textures; out-of-focus, vibrant corals in soft shades of pink and deep purple sway gently, while the tentacles of a large, soft-edged sea anemone ripple in the subtle currents. The scene is primarily illuminated by a magical, blended blue and gold light emanating from the creatures, creating soft highlights and gentle shadows on the hands, enveloping the moment in an atmosphere of pure, quietly enchanting beauty and capturing a profound sense of wonder through the sheer tenderness of the movement.", + "seed": 17262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_14/index_9417_seed_17262_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09627428264357149, + "vlm_logprob": -0.09627428264357149, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.09627428264357149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4323004186153412 + ], + "opposite_logprob_tokens": [ + -1.0573004484176636 + ], + "logprob_reward": -0.4323004186153412, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4323004186153412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09194774925708771, + 0.3870501220226288, + 0.0596727691590786, + 0.05237279087305069, + 0.03923449292778969, + 0.1677161455154419, + 0.017843687906861305, + 0.010462191887199879, + 0.06703346222639084, + 0.10666655004024506 + ], + "credit_map": [ + 0.012817339040338993, + 0.010747430846095085, + 0.0068498896434903145, + 0.007808179594576359, + 0.011693852953612804, + 0.015301210805773735, + 0.01211815606802702, + 0.014611689373850822, + 0.02501303143799305, + 0.019396696239709854, + 0.0838036760687828, + 0.012552364729344845, + 0.012546940706670284, + 0.017096154391765594, + 0.07515902817249298, + 0.14148226380348206, + 0.00883153174072504, + 0.010159917175769806, + 0.0066824606619775295, + 0.004839600995182991, + 0.009012277238070965, + 0.005956009961664677, + 0.004934081342071295, + 0.009256891906261444, + 0.004933159798383713, + 0.0028250636532902718, + 0.014046387746930122, + 0.004137852694839239, + 0.003684631083160639, + 0.0031541166827082634, + 0.004570314195007086, + 0.015021268278360367, + 0.004440562799572945, + 0.0018680187640711665, + 0.009043385274708271, + 0.0018271562876179814, + 0.002695654984563589, + 0.003415189916267991, + 0.006791543215513229, + 0.00915298331528902, + 0.00972242746502161, + 0.006581771653145552, + 0.02592436783015728, + 0.007815351709723473, + 0.01941876858472824, + 0.061065755784511566, + 0.029572434723377228, + 0.00761528592556715, + 0.002044201595708728, + 0.00222964258864522, + 0.00396490003913641, + 0.0011789759155362844, + 0.0017129601910710335, + 0.004732056055217981, + 0.0008429212030023336, + 0.0011380311334505677, + 0.0018515016417950392, + 0.0011067125014960766, + 0.0013786383206024766, + 0.0017385386163368821, + 0.0010113903554156423, + 0.0010581181850284338, + 0.0011057762894779444, + 0.0012115169083699584, + 0.01023663580417633, + 0.005449925549328327, + 0.005901936441659927, + 0.007246864959597588, + 0.007171798963099718, + 0.014551693573594093, + 0.007381194271147251, + 0.00909341685473919, + 0.006211862899363041, + 0.003806841792538762, + 0.02062487229704857, + 0.0037315862718969584, + 0.014841404743492603, + 0.04349271580576897, + 0.002429649466648698, + 0.011527623049914837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.4323004186153412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307988420128822 + ], + "opposite_logprob_tokens": [ + -3.773308038711548 + ], + "logprob_reward": -0.023307988420128822, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307988420128822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10089604556560516, + 0.4559539556503296, + 0.031467948108911514, + 0.06643297523260117, + 0.013352086767554283, + 0.09114473313093185, + 0.035433266311883926, + 0.04069500416517258, + 0.14168846607208252, + 0.022935539484024048 + ], + "credit_map": [ + 0.009373996406793594, + 0.008358667604625225, + 0.021328894421458244, + 0.006725259590893984, + 0.006231566425412893, + 0.008044402115046978, + 0.026196211576461792, + 0.014637043699622154, + 0.025980757549405098, + 0.020017551258206367, + 0.07153201103210449, + 0.013936290517449379, + 0.01371181383728981, + 0.019265083596110344, + 0.12360095977783203, + 0.16790948808193207, + 0.004878494888544083, + 0.0050009372644126415, + 0.003030130872502923, + 0.002803550101816654, + 0.003832598915323615, + 0.0033706175163388252, + 0.002724927384406328, + 0.005826692562550306, + 0.006060120649635792, + 0.002884349785745144, + 0.021214885637164116, + 0.0026573745999485254, + 0.004691165406256914, + 0.003582635661587119, + 0.0056703295558691025, + 0.01967211440205574, + 0.002090561669319868, + 0.0018677766202017665, + 0.0023843119852244854, + 0.0012954316334798932, + 0.001375266001559794, + 0.0008893050253391266, + 0.0021285556722432375, + 0.0013208786258473992, + 0.005543747916817665, + 0.00312406150624156, + 0.01693267561495304, + 0.004112684167921543, + 0.01000682357698679, + 0.03369101881980896, + 0.014063049107789993, + 0.003670673817396164, + 0.0032794419676065445, + 0.001804723055101931, + 0.009641402401030064, + 0.0018929153447970748, + 0.003243819810450077, + 0.012381134554743767, + 0.0018270357977598906, + 0.0013627945445477962, + 0.003283367957919836, + 0.0026798632461577654, + 0.005394604988396168, + 0.002145546954125166, + 0.006768239662051201, + 0.014689000323414803, + 0.002147592604160309, + 0.0035867919214069843, + 0.009486263617873192, + 0.005042629782110453, + 0.008059142157435417, + 0.005957615561783314, + 0.02735765092074871, + 0.052182093262672424, + 0.006305341608822346, + 0.02729773335158825, + 0.0016457193996757269, + 0.0015367581509053707, + 0.0020209637004882097, + 0.0014487243024632335, + 0.0036489798221737146, + 0.008077910169959068, + 0.0009043961181305349, + 0.0036520895082503557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.023307988420128822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075684800744057 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075684800744057, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075684800744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11054163426160812, + 0.2814998924732208, + 0.01935075782239437, + 0.05527682974934578, + 0.008782744407653809, + 0.054389867931604385, + 0.013410935178399086, + 0.020158646628260612, + 0.4233320951461792, + 0.013256502337753773 + ], + "credit_map": [ + 0.00656433729454875, + 0.004566521383821964, + 0.007526850327849388, + 0.0035611139610409737, + 0.004057047888636589, + 0.005292634945362806, + 0.03432304412126541, + 0.044650088995695114, + 0.016309062018990517, + 0.012361240573227406, + 0.047172367572784424, + 0.00840853899717331, + 0.008012412115931511, + 0.01174277625977993, + 0.07268363982439041, + 0.10480985045433044, + 0.0026367532555013895, + 0.0036864455323666334, + 0.0021709762513637543, + 0.0015015259850770235, + 0.0028288986068218946, + 0.0017720006871968508, + 0.001545824808999896, + 0.003208333160728216, + 0.005253716837614775, + 0.00237919669598341, + 0.016505815088748932, + 0.0031082406640052795, + 0.003736997488886118, + 0.0032817982137203217, + 0.00501421419903636, + 0.01599685288965702, + 0.00199816282838583, + 0.001162032363936305, + 0.0019163385732099414, + 0.0006822369177825749, + 0.0007481857319362462, + 0.0004953259485773742, + 0.0007271256763488054, + 0.0010533357271924615, + 0.0032362332567572594, + 0.0022199128288775682, + 0.0059010544791817665, + 0.0028274140786379576, + 0.006879280786961317, + 0.019344812259078026, + 0.011158676818013191, + 0.002822482492774725, + 0.0017110687913373113, + 0.0018688075942918658, + 0.0024052138905972242, + 0.001208942150697112, + 0.0014244684716686606, + 0.002879808656871319, + 0.0009898344287648797, + 0.000922790786717087, + 0.0035131981130689383, + 0.0016992533346638083, + 0.001883111777715385, + 0.0025216226931661367, + 0.0019036398734897375, + 0.003046089317649603, + 0.0034400662407279015, + 0.002151666209101677, + 0.02737363986670971, + 0.015558832325041294, + 0.02299388125538826, + 0.017484689131379128, + 0.08023037016391754, + 0.15155236423015594, + 0.02791605331003666, + 0.08022227138280869, + 0.001295745256356895, + 0.0009480287553742528, + 0.0028359265998005867, + 0.0011729738907888532, + 0.001065155491232872, + 0.0033431772608309984, + 0.0010809804080054164, + 0.0015145145589485765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.011075684800744057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106967218220234 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.01106967218220234, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106967218220234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06458647549152374, + 0.6241360902786255, + 0.027087246999144554, + 0.019852176308631897, + 0.027683692052960396, + 0.04433204606175423, + 0.1173076182603836, + 0.01325889490544796, + 0.042848311364650726, + 0.018907498568296432 + ], + "credit_map": [ + 0.008333364501595497, + 0.006558919791132212, + 0.015482103452086449, + 0.0036847307346761227, + 0.004684762563556433, + 0.0066691115498542786, + 0.00655266223475337, + 0.012620813213288784, + 0.036724310368299484, + 0.029121816158294678, + 0.10877100378274918, + 0.020064223557710648, + 0.019640590995550156, + 0.02668636478483677, + 0.15522409975528717, + 0.22790364921092987, + 0.006372383795678616, + 0.004467028193175793, + 0.003297994378954172, + 0.0017968490719795227, + 0.0030287206172943115, + 0.002162401797249913, + 0.0018334132619202137, + 0.0041284505277872086, + 0.00254735816270113, + 0.0020301672630012035, + 0.0044916896149516106, + 0.0016324674943462014, + 0.0013465526280924678, + 0.0011616083793342113, + 0.001866019912995398, + 0.004776311572641134, + 0.003099010558798909, + 0.0012480550212785602, + 0.007061068434268236, + 0.001056683948263526, + 0.0021538797300308943, + 0.003169203642755747, + 0.003877533134073019, + 0.006018253043293953, + 0.003517711767926812, + 0.0034475838765501976, + 0.0043986220844089985, + 0.0027378350496292114, + 0.0055909729562699795, + 0.014630420133471489, + 0.007306025829166174, + 0.002702865516766906, + 0.010224945843219757, + 0.00465615838766098, + 0.034409113228321075, + 0.0046835956163704395, + 0.011551523581147194, + 0.044717345386743546, + 0.003761770436540246, + 0.0033031466882675886, + 0.001616851193830371, + 0.001214181655086577, + 0.0011621256126090884, + 0.0018099348526448011, + 0.0012280488153919578, + 0.002069084672257304, + 0.0026153596118092537, + 0.0015433060470968485, + 0.0030747572891414165, + 0.0020509061869233847, + 0.002934869611635804, + 0.002316925907507539, + 0.00716580543667078, + 0.013581208884716034, + 0.004746550228446722, + 0.0069772833958268166, + 0.001791526679880917, + 0.0015842290595173836, + 0.0020588990300893784, + 0.0012740347301587462, + 0.0029905762057751417, + 0.004511735867708921, + 0.0013567887945100665, + 0.00333970645442605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.01106967218220234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036176491994410753 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036176491994410753, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036176491994410753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09863276034593582, + 0.15246234834194183, + 0.07489793747663498, + 0.07869894057512283, + 0.02308097667992115, + 0.053345032036304474, + 0.06657274812459946, + 0.09205980598926544, + 0.2637573480606079, + 0.09649213403463364 + ], + "credit_map": [ + 0.007233693264424801, + 0.007064609322696924, + 0.006343329790979624, + 0.004678383469581604, + 0.006091611459851265, + 0.009115319699048996, + 0.024320989847183228, + 0.033784832805395126, + 0.008953670039772987, + 0.006438425742089748, + 0.01782536692917347, + 0.0050416127778589725, + 0.005387912038713694, + 0.007021258119493723, + 0.04671455919742584, + 0.055079542100429535, + 0.012098888866603374, + 0.012223400175571442, + 0.007532232441008091, + 0.006873787846416235, + 0.0086261797696352, + 0.006913496647030115, + 0.007098739501088858, + 0.013531212694942951, + 0.007147831376641989, + 0.0030314172618091106, + 0.025134984403848648, + 0.0034897164441645145, + 0.005955092143267393, + 0.004528073128312826, + 0.006330219097435474, + 0.023081602528691292, + 0.00352055043913424, + 0.0019828681834042072, + 0.00429503433406353, + 0.001176856691017747, + 0.001767186913639307, + 0.0015398768009617925, + 0.0031037055887281895, + 0.0056948973797261715, + 0.003561558434739709, + 0.0024390488397330046, + 0.007934518158435822, + 0.0027272480074316263, + 0.005827351473271847, + 0.017618359997868538, + 0.010221769101917744, + 0.003015178022906184, + 0.005865192506462336, + 0.004000700078904629, + 0.017780791968107224, + 0.003011103952303529, + 0.006854888517409563, + 0.02437232993543148, + 0.002242322312667966, + 0.002445424674078822, + 0.006949693895876408, + 0.0060574873350560665, + 0.010677997954189777, + 0.005797591060400009, + 0.013944385573267937, + 0.030292676761746407, + 0.010284860618412495, + 0.00805511325597763, + 0.018059730529785156, + 0.010031890124082565, + 0.01431253645569086, + 0.013593688607215881, + 0.04990927875041962, + 0.09459114819765091, + 0.015089014545083046, + 0.0481700524687767, + 0.0075556389056146145, + 0.0040445090271532536, + 0.028638970106840134, + 0.004895078483968973, + 0.00880226120352745, + 0.035958100110292435, + 0.0031981191132217646, + 0.00339945824816823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0036176491994410753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Hayao Miyazaki's captivating, hand-drawn Ghibli style, a static, intimate close-up captures a pair of slender, young hands cupped together in a tranquil underwater realm, meticulously detailed. Steadily and deliberately, these hands guide two shy, bioluminescent sea creatures toward each other. One creature nestles in the palm of the left hand, radiating a soft, ethereal azure glow. Its gelatinous body is shaped like a small, translucent bell, with inquisitive black eyes. The other, guided by the gentle fingers of the right hand, radiates a warm golden light. Its form is a slightly intricate, star-shaped droplet that trembles timidly at approach. As the two creatures cautiously approach, their respective radiance intensifies, casting a magical, pulsating glow that illuminates the subtle details of the characters' skin and the shimmering water particles between them. The immediate background is a lush underwater world with soft, painterly textures; out-of-focus, vibrant corals in soft shades of pink and deep purple sway gently, while the tentacles of a large, soft-edged sea anemone ripple in the subtle currents. The scene is primarily illuminated by a magical, blended blue and gold light emanating from the creatures, creating soft highlights and gentle shadows on the hands, enveloping the moment in an atmosphere of pure, quietly enchanting beauty and capturing a profound sense of wonder through the sheer tenderness of the movement.", + "seed": 17263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_14/index_9417_seed_17263_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.008165819616988302, + "vlm_logprob": -0.008165819616988302, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.008165819616988302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021961353719234467 + ], + "opposite_logprob_tokens": [ + -3.8969614505767822 + ], + "logprob_reward": -0.021961353719234467, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021961353719234467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14218993484973907, + 0.1912686675786972, + 0.0662083774805069, + 0.05266902223229408, + 0.0908646285533905, + 0.06948313117027283, + 0.05413874238729477, + 0.12298646569252014, + 0.08112893253564835, + 0.12906210124492645 + ], + "credit_map": [ + 0.011683671735227108, + 0.006946928333491087, + 0.03272665664553642, + 0.008331263437867165, + 0.007911360822618008, + 0.02198193036019802, + 0.01206168718636036, + 0.04054642468690872, + 0.008950348012149334, + 0.01068909466266632, + 0.019202105700969696, + 0.011727752164006233, + 0.02222946472465992, + 0.05031018331646919, + 0.01752459444105625, + 0.050635114312171936, + 0.006146646104753017, + 0.006701245903968811, + 0.0058348760940134525, + 0.0055773393251001835, + 0.0056199608370661736, + 0.019963324069976807, + 0.006649631541222334, + 0.009715345688164234, + 0.004375138785690069, + 0.00282906461507082, + 0.007031493354588747, + 0.004601835738867521, + 0.004984148778021336, + 0.007181040942668915, + 0.007258692756295204, + 0.014407606795430183, + 0.010997545905411243, + 0.007473656442016363, + 0.013584033586084843, + 0.00818781740963459, + 0.005948496516793966, + 0.02525382675230503, + 0.009194123558700085, + 0.010225129313766956, + 0.007173215504735708, + 0.007562556769698858, + 0.01166035607457161, + 0.007455641869455576, + 0.008218389935791492, + 0.010928971692919731, + 0.00764358788728714, + 0.008840407244861126, + 0.0053587970323860645, + 0.004891531076282263, + 0.009343346580862999, + 0.0045641567558050156, + 0.009342689998447895, + 0.008770494721829891, + 0.005824670195579529, + 0.0060430532321333885, + 0.010676666162908077, + 0.011869612149894238, + 0.021639810875058174, + 0.010124358348548412, + 0.01999463327229023, + 0.018441220745444298, + 0.013894296251237392, + 0.01634586602449417, + 0.0074034035205841064, + 0.01074419915676117, + 0.010331875644624233, + 0.006626729853451252, + 0.015539949759840965, + 0.01040627434849739, + 0.008355925790965557, + 0.01172057818621397, + 0.013165459968149662, + 0.009754571132361889, + 0.019643783569335938, + 0.010165934450924397, + 0.007701626047492027, + 0.028309980407357216, + 0.012838299386203289, + 0.027482446283102036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.021961353719234467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959480069577694 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.005959480069577694, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959480069577694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6091834902763367, + 0.08360966295003891, + 0.05696871876716614, + 0.015385953709483147, + 0.05412719026207924, + 0.038056593388319016, + 0.04779946804046631, + 0.04444563388824463, + 0.017876874655485153, + 0.032546453177928925 + ], + "credit_map": [ + 0.04836234822869301, + 0.01963280886411667, + 0.2425539791584015, + 0.027560170739889145, + 0.018453817814588547, + 0.07279647886753082, + 0.045207429677248, + 0.1346164047718048, + 0.005242711398750544, + 0.0043966518715023994, + 0.006638019345700741, + 0.005851007532328367, + 0.0066892351023852825, + 0.02148907631635666, + 0.008833829313516617, + 0.024469124153256416, + 0.005945979617536068, + 0.0045829834416508675, + 0.007544469553977251, + 0.006826071999967098, + 0.00534505769610405, + 0.008595116436481476, + 0.0061815823428332806, + 0.0119474520906806, + 0.0018272262532263994, + 0.0011242673499509692, + 0.0024684087838977575, + 0.0018279977375641465, + 0.0018072141101583838, + 0.001774756470695138, + 0.002396827330812812, + 0.002159255323931575, + 0.006438650190830231, + 0.005084298085421324, + 0.00742502324283123, + 0.005943663418292999, + 0.0035556182265281677, + 0.015310369431972504, + 0.004240532871335745, + 0.006129033397883177, + 0.0040016318671405315, + 0.003871920984238386, + 0.0054559591226279736, + 0.00420398311689496, + 0.0039820922538638115, + 0.006230663508176804, + 0.004999476484954357, + 0.00531086465343833, + 0.004297612700611353, + 0.004061487503349781, + 0.006295077968388796, + 0.005092785228043795, + 0.004829449579119682, + 0.006033956538885832, + 0.011003758758306503, + 0.006185338366776705, + 0.004180406220257282, + 0.005995235871523619, + 0.005336694419384003, + 0.004614363424479961, + 0.0076040830463171005, + 0.006726080551743507, + 0.005095283035188913, + 0.004893479403108358, + 0.0017315122531726956, + 0.0012412337819114327, + 0.001684995018877089, + 0.0011840327642858028, + 0.0022417602594941854, + 0.004641827195882797, + 0.0019956419710069895, + 0.0031558694317936897, + 0.003533826442435384, + 0.0026956750079989433, + 0.00668664276599884, + 0.003789209295064211, + 0.002396957017481327, + 0.0048000807873904705, + 0.0037621723022311926, + 0.004881886765360832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005959480069577694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36805295944213867, + 0.10210223495960236, + 0.06923720985651016, + 0.08758030086755753, + 0.05729624256491661, + 0.0698070228099823, + 0.069977767765522, + 0.09923193603754044, + 0.036516278982162476, + 0.04019811749458313 + ], + "credit_map": [ + 0.027339043095707893, + 0.01178477518260479, + 0.14326132833957672, + 0.01614750362932682, + 0.01390643510967493, + 0.06694019585847855, + 0.028755927458405495, + 0.059917718172073364, + 0.0065442523919045925, + 0.008279518224298954, + 0.010344036854803562, + 0.00825617928057909, + 0.01050520222634077, + 0.026848457753658295, + 0.008228507824242115, + 0.023096082732081413, + 0.008798868395388126, + 0.004432858899235725, + 0.008024247363209724, + 0.005600732285529375, + 0.006221909541636705, + 0.01810326613485813, + 0.006483264733105898, + 0.011572063900530338, + 0.009187902323901653, + 0.007251571398228407, + 0.010223357006907463, + 0.010056945495307446, + 0.009252549149096012, + 0.012659109197556973, + 0.011755974031984806, + 0.017192894592881203, + 0.0060241431929171085, + 0.0061049493961036205, + 0.006630750373005867, + 0.005498205311596394, + 0.005265492480248213, + 0.015440035611391068, + 0.005310512613505125, + 0.007022150792181492, + 0.010536383837461472, + 0.005744850262999535, + 0.00833329651504755, + 0.007486406248062849, + 0.006633173208683729, + 0.015306537970900536, + 0.007239073049277067, + 0.00852730218321085, + 0.007083333563059568, + 0.0065457457676529884, + 0.011102485470473766, + 0.0068611400201916695, + 0.007325705606490374, + 0.00951436348259449, + 0.012119618244469166, + 0.009425374679267406, + 0.012707622721791267, + 0.010688398033380508, + 0.017668483778834343, + 0.010043446905910969, + 0.012053157202899456, + 0.011098957620561123, + 0.01184599194675684, + 0.013125879690051079, + 0.0034170737490057945, + 0.0050325277261435986, + 0.008233651518821716, + 0.003015462076291442, + 0.0032318909652531147, + 0.005101691000163555, + 0.002712520770728588, + 0.005771464202553034, + 0.003433593548834324, + 0.0025461141485720873, + 0.005821551661938429, + 0.0037349630147218704, + 0.002927096327766776, + 0.010088622570037842, + 0.0031433214899152517, + 0.008502854034304619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.005251660011708736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942771211266518 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942771211266518, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942771211266518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5145638585090637, + 0.06505785137414932, + 0.038979675620794296, + 0.11257466673851013, + 0.02151079662144184, + 0.04879292845726013, + 0.04149336740374565, + 0.046667661517858505, + 0.026417458429932594, + 0.083941750228405 + ], + "credit_map": [ + 0.03698728606104851, + 0.01761331781744957, + 0.192708358168602, + 0.022702695801854134, + 0.023739269003272057, + 0.0727313831448555, + 0.04519268125295639, + 0.1028888002038002, + 0.00520735327154398, + 0.006570229306817055, + 0.006656750105321407, + 0.0067878845147788525, + 0.005663727410137653, + 0.014142145402729511, + 0.007929371669888496, + 0.012100383639335632, + 0.004072512965649366, + 0.0037691076286137104, + 0.004669921472668648, + 0.00388656510040164, + 0.003147020936012268, + 0.006031169556081295, + 0.005229509435594082, + 0.008173867128789425, + 0.010212461464107037, + 0.010809351690113544, + 0.014178548939526081, + 0.009220433421432972, + 0.010832170024514198, + 0.01795368082821369, + 0.021458445116877556, + 0.0179095771163702, + 0.002523208037018776, + 0.0022056102752685547, + 0.00398855609819293, + 0.0023555809166282415, + 0.0023316950537264347, + 0.002735213842242956, + 0.002645050408318639, + 0.00272588012740016, + 0.007412604056298733, + 0.0038474511820822954, + 0.008088509552180767, + 0.004457008559256792, + 0.005482254549860954, + 0.00830357987433672, + 0.0052877143025398254, + 0.005913805216550827, + 0.004784139338880777, + 0.003247245214879513, + 0.008345345966517925, + 0.003962506540119648, + 0.004999655764549971, + 0.006224502809345722, + 0.004727690946310759, + 0.005202278029173613, + 0.00316801224835217, + 0.007121456321328878, + 0.004812892992049456, + 0.005702456459403038, + 0.007728742901235819, + 0.005436270963400602, + 0.006347897928208113, + 0.006349926348775625, + 0.002927627647295594, + 0.002514780266210437, + 0.002297664061188698, + 0.0020786134991794825, + 0.005765747744590044, + 0.003065273631364107, + 0.004142648074775934, + 0.003625099780037999, + 0.007683487143367529, + 0.007049056235700846, + 0.01565304771065712, + 0.010168616659939289, + 0.0058808098547160625, + 0.014515376649796963, + 0.010267728939652443, + 0.012723617255687714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.005942771211266518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017138330731540918 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017138330731540918, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017138330731540918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2313159555196762, + 0.07323778420686722, + 0.06034855544567108, + 0.04362938925623894, + 0.06344158947467804, + 0.04990225285291672, + 0.11239346861839294, + 0.10442233830690384, + 0.024715635925531387, + 0.23659302294254303 + ], + "credit_map": [ + 0.02148941345512867, + 0.009224428795278072, + 0.08398129045963287, + 0.01390577107667923, + 0.00837614107877016, + 0.015692701563239098, + 0.017402175813913345, + 0.06124402955174446, + 0.0070883301086723804, + 0.008916687220335007, + 0.006091018673032522, + 0.007042164448648691, + 0.012894763611257076, + 0.010201489552855492, + 0.011981035582721233, + 0.009022296406328678, + 0.005896353628486395, + 0.007845019921660423, + 0.0073161656036973, + 0.007068932522088289, + 0.005295102950185537, + 0.01166460383683443, + 0.007202728185802698, + 0.008059646934270859, + 0.003633923828601837, + 0.004118717275559902, + 0.006053506396710873, + 0.005805327091366053, + 0.005014622118324041, + 0.006247721146792173, + 0.007102993782609701, + 0.005652579478919506, + 0.007086094468832016, + 0.0055499328300356865, + 0.008047791197896004, + 0.007121686358004808, + 0.004549730569124222, + 0.01950567401945591, + 0.00443645752966404, + 0.007144217845052481, + 0.005904266610741615, + 0.005877669434994459, + 0.00488736666738987, + 0.005988899152725935, + 0.005100774113088846, + 0.010707007721066475, + 0.005159704480320215, + 0.006276565138250589, + 0.012025510892271996, + 0.008426858112215996, + 0.019592419266700745, + 0.011579008772969246, + 0.014328758232295513, + 0.016340889036655426, + 0.016886666417121887, + 0.013213357888162136, + 0.00922852661460638, + 0.015442398376762867, + 0.017408017069101334, + 0.011351482942700386, + 0.011543111875653267, + 0.015873432159423828, + 0.010606702417135239, + 0.012968668714165688, + 0.0017536876257508993, + 0.002126578474417329, + 0.0036165500059723854, + 0.0019243985880166292, + 0.006067637819796801, + 0.002944977255538106, + 0.00257105496712029, + 0.003710750490427017, + 0.022845664992928505, + 0.018320487812161446, + 0.04275612533092499, + 0.024917732924222946, + 0.01663685403764248, + 0.0448099784553051, + 0.02635004371404648, + 0.039956141263246536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0017138330731540918 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Hayao Miyazaki's captivating, hand-drawn Ghibli style, a static, intimate close-up captures a pair of slender, young hands cupped together in a tranquil underwater realm, meticulously detailed. Steadily and deliberately, these hands guide two shy, bioluminescent sea creatures toward each other. One creature nestles in the palm of the left hand, radiating a soft, ethereal azure glow. Its gelatinous body is shaped like a small, translucent bell, with inquisitive black eyes. The other, guided by the gentle fingers of the right hand, radiates a warm golden light. Its form is a slightly intricate, star-shaped droplet that trembles timidly at approach. As the two creatures cautiously approach, their respective radiance intensifies, casting a magical, pulsating glow that illuminates the subtle details of the characters' skin and the shimmering water particles between them. The immediate background is a lush underwater world with soft, painterly textures; out-of-focus, vibrant corals in soft shades of pink and deep purple sway gently, while the tentacles of a large, soft-edged sea anemone ripple in the subtle currents. The scene is primarily illuminated by a magical, blended blue and gold light emanating from the creatures, creating soft highlights and gentle shadows on the hands, enveloping the moment in an atmosphere of pure, quietly enchanting beauty and capturing a profound sense of wonder through the sheer tenderness of the movement.", + "seed": 17264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_14/index_9417_seed_17264_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.00957152913324535, + "vlm_logprob": -0.00957152913324535, + "vlm_yes_no_margin": 5.3, + "vlm_token_credit": -0.00957152913324535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032053615897893906 + ], + "opposite_logprob_tokens": [ + -3.5320537090301514 + ], + "logprob_reward": -0.032053615897893906, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.032053615897893906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06935271620750427, + 0.04515194147825241, + 0.18675565719604492, + 0.2169070839881897, + 0.045702408999204636, + 0.10793314129114151, + 0.07978209108114243, + 0.13859474658966064, + 0.06069058179855347, + 0.049129635095596313 + ], + "credit_map": [ + 0.007760223001241684, + 0.008980391547083855, + 0.009249192662537098, + 0.00970335304737091, + 0.007547477725893259, + 0.004913229960948229, + 0.011882748454809189, + 0.0093160942196846, + 0.004776489920914173, + 0.004574720747768879, + 0.009077779948711395, + 0.007059263996779919, + 0.004646362736821175, + 0.005275618750602007, + 0.005377786234021187, + 0.004363918211311102, + 0.03136320039629936, + 0.013896273449063301, + 0.07271680980920792, + 0.02208874374628067, + 0.009131417609751225, + 0.009544814005494118, + 0.01676035486161709, + 0.011254034005105495, + 0.03314123675227165, + 0.06142178177833557, + 0.012596815824508667, + 0.010079837404191494, + 0.032607994973659515, + 0.015291835181415081, + 0.014528398402035236, + 0.03723917901515961, + 0.00576800387352705, + 0.011736405082046986, + 0.00531624723225832, + 0.004024858120828867, + 0.004158936906605959, + 0.003981153015047312, + 0.0035092218313366175, + 0.007207583636045456, + 0.013491704128682613, + 0.01260756328701973, + 0.012644882313907146, + 0.009135312400758266, + 0.011187905445694923, + 0.0217649694532156, + 0.009503696113824844, + 0.017597107216715813, + 0.007993181236088276, + 0.007996266707777977, + 0.012676096521317959, + 0.008479634299874306, + 0.0101669542491436, + 0.0130863506346941, + 0.011579357087612152, + 0.007804248947650194, + 0.02410845085978508, + 0.04911397397518158, + 0.015129118226468563, + 0.007767801638692617, + 0.010937390848994255, + 0.012239107862114906, + 0.007359609007835388, + 0.011939306743443012, + 0.008457164280116558, + 0.010834968648850918, + 0.008073386736214161, + 0.005302421282976866, + 0.007834182120859623, + 0.008945622481405735, + 0.006373357959091663, + 0.004869476426392794, + 0.014317793771624565, + 0.008220498450100422, + 0.005487489514052868, + 0.002423569792881608, + 0.005135964602231979, + 0.004858049564063549, + 0.003977200947701931, + 0.004709065426141024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.032053615897893906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636994123458862 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007636994123458862, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636994123458862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14134244620800018, + 0.039594974368810654, + 0.07998307049274445, + 0.24266308546066284, + 0.027983371168375015, + 0.04091985896229744, + 0.12529544532299042, + 0.06924985349178314, + 0.15239474177360535, + 0.08057313412427902 + ], + "credit_map": [ + 0.018688751384615898, + 0.016848603263497353, + 0.02089434675872326, + 0.017468107864260674, + 0.016515051946043968, + 0.012338966131210327, + 0.01645820401608944, + 0.022130416706204414, + 0.007073028478771448, + 0.004205749370157719, + 0.005328845232725143, + 0.004167160950601101, + 0.005001355428248644, + 0.004923620261251926, + 0.0038884158711880445, + 0.005006798543035984, + 0.014507882297039032, + 0.023440899327397346, + 0.006698833778500557, + 0.006553552579134703, + 0.008341243490576744, + 0.006385373417288065, + 0.006245148368179798, + 0.00781013909727335, + 0.030333442613482475, + 0.061041731387376785, + 0.017945025116205215, + 0.011921572498977184, + 0.04344530031085014, + 0.019078895449638367, + 0.01984287053346634, + 0.03905424475669861, + 0.0033683208748698235, + 0.0037351329810917377, + 0.005050749983638525, + 0.0027056101243942976, + 0.002865487476810813, + 0.00437778327614069, + 0.0027211022097617388, + 0.0031591830775141716, + 0.00432526646181941, + 0.004211699590086937, + 0.006077006924897432, + 0.0034148849081248045, + 0.005853382870554924, + 0.00666514178737998, + 0.005867903586477041, + 0.00450457539409399, + 0.016891730949282646, + 0.029928093776106834, + 0.010710814967751503, + 0.01134405191987753, + 0.010841009207069874, + 0.01417599618434906, + 0.00982268899679184, + 0.021581055596470833, + 0.010464600287377834, + 0.022374602034687996, + 0.010914621874690056, + 0.00448848120868206, + 0.0069334558211266994, + 0.005297909490764141, + 0.0045379046350717545, + 0.00423828512430191, + 0.01710878126323223, + 0.02496243268251419, + 0.029128359630703926, + 0.012258036062121391, + 0.017918821424245834, + 0.021286455914378166, + 0.013172600418329239, + 0.016559256240725517, + 0.024874769151210785, + 0.0118899280205369, + 0.008984626270830631, + 0.004074765369296074, + 0.0053789191879332066, + 0.00632012402638793, + 0.006842655595391989, + 0.01220734603703022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.007636994123458862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021967582870274782 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021967582870274782, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021967582870274782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15229931473731995, + 0.0320395827293396, + 0.14771883189678192, + 0.06097155064344406, + 0.03618132695555687, + 0.10232571512460709, + 0.065477155148983, + 0.31626954674720764, + 0.03658618405461311, + 0.05013083294034004 + ], + "credit_map": [ + 0.018858961760997772, + 0.01781584694981575, + 0.02489805966615677, + 0.02061651647090912, + 0.018591051921248436, + 0.015951499342918396, + 0.015274171717464924, + 0.02029319852590561, + 0.0032920693047344685, + 0.004474563058465719, + 0.0038566584698855877, + 0.005279563367366791, + 0.0045905872248113155, + 0.002733084373176098, + 0.004350266419351101, + 0.00346278864890337, + 0.024963146075606346, + 0.05123598501086235, + 0.011349708773195744, + 0.013508870266377926, + 0.012642642483115196, + 0.009141821414232254, + 0.008615463972091675, + 0.016261184588074684, + 0.00886488426476717, + 0.01183266006410122, + 0.006610844749957323, + 0.004049731884151697, + 0.006762090139091015, + 0.00568751385435462, + 0.006135051604360342, + 0.011028775945305824, + 0.0037930775433778763, + 0.0031141070649027824, + 0.005723237991333008, + 0.004358998965471983, + 0.004793747793883085, + 0.0059089576825499535, + 0.005222604610025883, + 0.0032665934413671494, + 0.013285805471241474, + 0.01197302620857954, + 0.012131978757679462, + 0.007872558198869228, + 0.01378476433455944, + 0.01772785186767578, + 0.01111396960914135, + 0.01443576905876398, + 0.008441709913313389, + 0.01361018605530262, + 0.006506007164716721, + 0.005640106741338968, + 0.008960818871855736, + 0.007025931961834431, + 0.005306018982082605, + 0.00998637918382883, + 0.05324903875589371, + 0.11524805426597595, + 0.04188607260584831, + 0.01952357031404972, + 0.02528129704296589, + 0.020128007978200912, + 0.014483404345810413, + 0.02647010050714016, + 0.0042928713373839855, + 0.005271075293421745, + 0.006052033044397831, + 0.003389932680875063, + 0.00427444651722908, + 0.004948165267705917, + 0.0042859395034611225, + 0.004071720875799656, + 0.017011428251862526, + 0.008401097729802132, + 0.005960430949926376, + 0.0025120519567281008, + 0.0032521458342671394, + 0.003275464754551649, + 0.003240925492718816, + 0.006477289367467165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0021967582870274782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633404780179262 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633404780179262, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633404780179262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2890035808086395, + 0.052668675780296326, + 0.1371808499097824, + 0.08318203687667847, + 0.054306760430336, + 0.04596254602074623, + 0.08411888778209686, + 0.18664519488811493, + 0.04559236764907837, + 0.021339135244488716 + ], + "credit_map": [ + 0.03396246209740639, + 0.030008608475327492, + 0.0451955571770668, + 0.0660916119813919, + 0.029457394033670425, + 0.018809381872415543, + 0.03561126068234444, + 0.02986731193959713, + 0.009128017351031303, + 0.004591898061335087, + 0.008491401560604572, + 0.005372615996748209, + 0.008562334813177586, + 0.005506431683897972, + 0.005325790960341692, + 0.005690181162208319, + 0.030316131189465523, + 0.032147008925676346, + 0.024502897635102272, + 0.011595095507800579, + 0.009348836727440357, + 0.008294627070426941, + 0.008598247542977333, + 0.012377995066344738, + 0.013017460703849792, + 0.01984473504126072, + 0.00662389537319541, + 0.004616933409124613, + 0.014814982190728188, + 0.006005393341183662, + 0.006162392906844616, + 0.012096238322556019, + 0.008271558210253716, + 0.016827115789055824, + 0.004343890585005283, + 0.0031441524624824524, + 0.0042060185223817825, + 0.004458538256585598, + 0.0032434514723718166, + 0.00981203094124794, + 0.006019530352205038, + 0.005812058225274086, + 0.005305963102728128, + 0.003411494428291917, + 0.004550916142761707, + 0.00937572494149208, + 0.004427393898367882, + 0.007059465628117323, + 0.01131597813218832, + 0.019261391833424568, + 0.009194628335535526, + 0.004665161017328501, + 0.006476601120084524, + 0.015187772922217846, + 0.005010200664401054, + 0.013007158413529396, + 0.03196404129266739, + 0.06902041286230087, + 0.024374229833483696, + 0.011020471341907978, + 0.014015065506100655, + 0.013179630972445011, + 0.007583169266581535, + 0.015488154254853725, + 0.005716727115213871, + 0.0069173588417470455, + 0.00622249161824584, + 0.0044140443205833435, + 0.006370157469063997, + 0.005886662285774946, + 0.005694448947906494, + 0.004370474256575108, + 0.004418663214892149, + 0.002606772817671299, + 0.0025485216174274683, + 0.0014838642673566937, + 0.0023868647404015064, + 0.002241480629891157, + 0.0026359125040471554, + 0.0030170553363859653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.004633404780179262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two slender, young hands cupped together visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature a translucent, bell-shaped gelatinous form with black eyes radiating soft azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other creature a star-shaped droplet radiating warm golden light that trembles as it approaches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature out-of-focus vibrant corals in pink and purple alongside rippling sea anemone tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a blended blue and gold magical glow that intensifies as the creatures approach?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013368725776672363 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013368725776672363, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013368725776672363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10321105271577835, + 0.0516689196228981, + 0.10790418833494186, + 0.057664114981889725, + 0.04247384890913963, + 0.04682343453168869, + 0.14784550666809082, + 0.04001154378056526, + 0.06603553891181946, + 0.336361825466156 + ], + "credit_map": [ + 0.011626210063695908, + 0.009192164987325668, + 0.016548218205571175, + 0.020208587870001793, + 0.013133984059095383, + 0.009540889412164688, + 0.012017626315355301, + 0.01094337273389101, + 0.006749283988028765, + 0.005414905026555061, + 0.010839896276593208, + 0.007374226581305265, + 0.00591532327234745, + 0.00537735503166914, + 0.005207802169024944, + 0.004790129140019417, + 0.01903042383491993, + 0.030282825231552124, + 0.015438116155564785, + 0.010860485956072807, + 0.007417700253427029, + 0.00595920579507947, + 0.00808529183268547, + 0.010830137878656387, + 0.006159846670925617, + 0.008348984643816948, + 0.009336883202195168, + 0.005025129299610853, + 0.009415973909199238, + 0.004965832456946373, + 0.00844531785696745, + 0.005966148339211941, + 0.006356223952025175, + 0.011525794863700867, + 0.003983460832387209, + 0.002748195081949234, + 0.004016544204205275, + 0.003239833749830723, + 0.0033808990847319365, + 0.0072228978388011456, + 0.004673955496400595, + 0.00546263949945569, + 0.005123250186443329, + 0.003793165786191821, + 0.006190200336277485, + 0.007461258210241795, + 0.007076697889715433, + 0.007042267359793186, + 0.023148376494646072, + 0.040312204509973526, + 0.01589025929570198, + 0.008494756184518337, + 0.012581003829836845, + 0.02147557958960533, + 0.008169900625944138, + 0.017773419618606567, + 0.005545428954064846, + 0.01183103397488594, + 0.005562750622630119, + 0.0025584022514522076, + 0.004320703912526369, + 0.004209923092275858, + 0.002711298642680049, + 0.0032720041926950216, + 0.007648473605513573, + 0.010197813622653484, + 0.015672314912080765, + 0.006478726398199797, + 0.006465951446443796, + 0.006701229140162468, + 0.0062629240565001965, + 0.00660810898989439, + 0.11204425990581512, + 0.05420689284801483, + 0.03912735730409622, + 0.014467615634202957, + 0.01880692131817341, + 0.026109332218766212, + 0.022581035271286964, + 0.04901839420199394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.0013368725776672363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_15/index_tensor([4063, 4063])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_15/index_tensor([4063, 4063])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a8df48dcc7e273c689b54ae70027b98ff3705b11 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_15/index_tensor([4063, 4063])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic long-range shot, captured from a fixed-altitude drone, captures the first rays of golden sunlight piercing the mist, illuminating the majestic Great Wall nestled between towering mountains. The ancient, mottled gray brickwork of the wall winds its way like a giant dragon, its weathered surface richly textured and cast long, pale blue shadows in the slanting morning light. Against this majestic and timeless backdrop, two incredibly tiny figures advance along the broad path atop the wall. The leading figure, draped in a crimson cloak, stands out against the gray landscape, his gait steady. Following closely behind, a second figure, clad in dark clothing, maintains a few paces behind, following steadily and calmly. Their movements appear insignificant against the vast structure, like two slowly moving dots of ink. The entire image is imbued with an epic tranquility and the atmosphere of a lonely and determined journey.", + "seed": 634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_15/index_4063_seed_634_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08101495481096208, + "vlm_logprob": -0.08101495481096208, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.08101495481096208, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3073287606239319 + ], + "opposite_logprob_tokens": [ + -1.432328701019287 + ], + "logprob_reward": -0.3073287606239319, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3073287606239319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16717858612537384, + 0.14159080386161804, + 0.09617559611797333, + 0.05821330472826958, + 0.07511654496192932, + 0.1257937252521515, + 0.19845658540725708, + 0.08300444483757019, + 0.04119201749563217, + 0.013278438709676266 + ], + "credit_map": [ + 0.01564059592783451, + 0.030532578006386757, + 0.02794293314218521, + 0.023790420964360237, + 0.012700549326837063, + 0.015734070912003517, + 0.01379417534917593, + 0.027043268084526062, + 0.028915850445628166, + 0.010759019292891026, + 0.04610127955675125, + 0.025038234889507294, + 0.00665762135758996, + 0.006271891295909882, + 0.0050817993469536304, + 0.012765112332999706, + 0.022335289046168327, + 0.013609067536890507, + 0.02505015768110752, + 0.009839293546974659, + 0.005365633871406317, + 0.006645523943006992, + 0.0051951659843325615, + 0.008135463111102581, + 0.009408813901245594, + 0.009034393355250359, + 0.009988349862396717, + 0.005512224044650793, + 0.0053673540242016315, + 0.005258019547909498, + 0.006823176983743906, + 0.006820975337177515, + 0.019999781623482704, + 0.010034050792455673, + 0.02026432938873768, + 0.004977094475179911, + 0.0030052317306399345, + 0.004091820679605007, + 0.005083588417619467, + 0.007660646922886372, + 0.01933187246322632, + 0.02028362639248371, + 0.011587042361497879, + 0.013076425530016422, + 0.015925712883472443, + 0.013704998418688774, + 0.016671642661094666, + 0.015212406404316425, + 0.03304214030504227, + 0.042013999074697495, + 0.01789781264960766, + 0.010706477798521519, + 0.019768591970205307, + 0.018499113619327545, + 0.0349070243537426, + 0.02162140980362892, + 0.0127702746540308, + 0.014650999568402767, + 0.008245313540101051, + 0.007645452860742807, + 0.006660193204879761, + 0.00986248068511486, + 0.011883259750902653, + 0.011286474764347076, + 0.006031637545675039, + 0.0069374144077301025, + 0.004430760163813829, + 0.00413883151486516, + 0.004453123547136784, + 0.004737973678857088, + 0.005398231092840433, + 0.005064045079052448, + 0.0019120597280561924, + 0.0014585467288270593, + 0.0008255285210907459, + 0.0013214697828516364, + 0.0011245874920859933, + 0.0009471472003497183, + 0.004333711229264736, + 0.001355389249511063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.3073287606239319 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003215620992705226 + ], + "opposite_logprob_tokens": [ + -5.753215789794922 + ], + "logprob_reward": -0.003215620992705226, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003215620992705226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21675726771354675, + 0.16019999980926514, + 0.1041993647813797, + 0.11546144634485245, + 0.13936513662338257, + 0.05165042728185654, + 0.08608470112085342, + 0.06423930823802948, + 0.030917832627892494, + 0.031124448403716087 + ], + "credit_map": [ + 0.029812226071953773, + 0.0654095932841301, + 0.03621496260166168, + 0.03381425887346268, + 0.014976299367845058, + 0.010305589064955711, + 0.009553880430758, + 0.016670463606715202, + 0.028819087892770767, + 0.01924252137541771, + 0.03900923579931259, + 0.02473035268485546, + 0.009205872192978859, + 0.009719970636069775, + 0.008230360224843025, + 0.02124260738492012, + 0.023323582485318184, + 0.01551576517522335, + 0.025264956057071686, + 0.009540778584778309, + 0.006138260941952467, + 0.006199232302606106, + 0.005751906428486109, + 0.01246487908065319, + 0.01893668994307518, + 0.021578650921583176, + 0.02095002308487892, + 0.00848517194390297, + 0.011139205656945705, + 0.008267569355666637, + 0.014427486807107925, + 0.011676652356982231, + 0.03679091855883598, + 0.019611293449997902, + 0.0389646552503109, + 0.008310861885547638, + 0.006551061756908894, + 0.007376603316515684, + 0.010615223087370396, + 0.011144527234137058, + 0.008766069076955318, + 0.008620642125606537, + 0.004250059835612774, + 0.0050613367930054665, + 0.00616646371781826, + 0.005872267764061689, + 0.008142364211380482, + 0.004771223291754723, + 0.015148801729083061, + 0.01715562306344509, + 0.007929451763629913, + 0.007122916169464588, + 0.006082682870328426, + 0.009077596478164196, + 0.014621811918914318, + 0.00894581712782383, + 0.010267126373946667, + 0.012463664636015892, + 0.0070510790683329105, + 0.005874948110431433, + 0.005133805330842733, + 0.007185869850218296, + 0.008612772449851036, + 0.007650041487067938, + 0.0046040695160627365, + 0.005440696608275175, + 0.0025344511959701777, + 0.0026930314488708973, + 0.0029347860254347324, + 0.004576872102916241, + 0.00387523858807981, + 0.004258686676621437, + 0.004618892911821604, + 0.006594352889806032, + 0.0020231050439178944, + 0.002293257275596261, + 0.0029395169112831354, + 0.002428795211017132, + 0.0073713879100978374, + 0.002855139784514904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.003215620992705226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1724260300397873, + 0.09820190072059631, + 0.07047794759273529, + 0.10677213966846466, + 0.089964359998703, + 0.09374726563692093, + 0.19098415970802307, + 0.04779618978500366, + 0.045289862900972366, + 0.08434008061885834 + ], + "credit_map": [ + 0.019240878522396088, + 0.03530486673116684, + 0.03834858164191246, + 0.03535349667072296, + 0.011116428300738335, + 0.00890457071363926, + 0.010445911437273026, + 0.013711304403841496, + 0.009955172426998615, + 0.01598377153277397, + 0.015194003470242023, + 0.009596765041351318, + 0.0091434670612216, + 0.009906723164021969, + 0.011421066708862782, + 0.017000935971736908, + 0.015353451482951641, + 0.0083834333345294, + 0.016616711392998695, + 0.007632740307599306, + 0.004399759229272604, + 0.003394848434254527, + 0.004076914396136999, + 0.010620096698403358, + 0.01439413707703352, + 0.013914552517235279, + 0.0211900994181633, + 0.010427229106426239, + 0.01098300889134407, + 0.007613132707774639, + 0.013196755200624466, + 0.01505323313176632, + 0.011492929421365261, + 0.014050479978322983, + 0.013139544054865837, + 0.011023582890629768, + 0.006807776167988777, + 0.01003356371074915, + 0.011625696904957294, + 0.011790791526436806, + 0.014710266143083572, + 0.016731414943933487, + 0.0066909040324389935, + 0.008487950079143047, + 0.013440133072435856, + 0.010996759869158268, + 0.013189315795898438, + 0.009500517509877682, + 0.03069409355521202, + 0.03534546494483948, + 0.017577340826392174, + 0.014127401635050774, + 0.013296717777848244, + 0.01569349318742752, + 0.041848331689834595, + 0.022401336580514908, + 0.008377782069146633, + 0.007977359928190708, + 0.003488993737846613, + 0.0031556421890854836, + 0.005319517571479082, + 0.005154484882950783, + 0.009064460173249245, + 0.005257951561361551, + 0.0072115035727620125, + 0.009493297897279263, + 0.0035779138561338186, + 0.0028626129496842623, + 0.003884523641318083, + 0.004531664773821831, + 0.008341795764863491, + 0.005386556498706341, + 0.011031165719032288, + 0.01192757673561573, + 0.00488171586766839, + 0.006269348319619894, + 0.008216459304094315, + 0.006379426456987858, + 0.02858543023467064, + 0.007048964500427246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.016067897900938988 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980896644294262 + ], + "opposite_logprob_tokens": [ + -3.52980899810791 + ], + "logprob_reward": -0.02980896644294262, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980896644294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.127400740981102, + 0.08975247293710709, + 0.0934654250741005, + 0.06414207071065903, + 0.20565244555473328, + 0.05477446690201759, + 0.1992483139038086, + 0.08871080726385117, + 0.03943656012415886, + 0.03741665184497833 + ], + "credit_map": [ + 0.015208669938147068, + 0.030136408284306526, + 0.021660685539245605, + 0.02508661150932312, + 0.00982586108148098, + 0.00786155927926302, + 0.0069066863507032394, + 0.010714255273342133, + 0.009284903295338154, + 0.012546202167868614, + 0.013515464961528778, + 0.007986965589225292, + 0.008669817820191383, + 0.008600636385381222, + 0.00849431287497282, + 0.02065417543053627, + 0.026442205533385277, + 0.007962450385093689, + 0.03064759075641632, + 0.008924104273319244, + 0.0034927453380078077, + 0.00452102767303586, + 0.0032899128273129463, + 0.00818538572639227, + 0.006576432380825281, + 0.011061171069741249, + 0.00708228861913085, + 0.006309525575488806, + 0.007493634708225727, + 0.006173520348966122, + 0.010606355033814907, + 0.00883914902806282, + 0.05116867274045944, + 0.026981143280863762, + 0.04444507881999016, + 0.01942378096282482, + 0.011903301812708378, + 0.015612548217177391, + 0.018035966902971268, + 0.018081944435834885, + 0.007679148111492395, + 0.008009504526853561, + 0.005006736144423485, + 0.004351056646555662, + 0.007857895456254482, + 0.0059962100349366665, + 0.007737622130662203, + 0.008136295713484287, + 0.030451146885752678, + 0.03204338252544403, + 0.01827215403318405, + 0.014014131389558315, + 0.015613006427884102, + 0.015672285109758377, + 0.048073381185531616, + 0.025108827278017998, + 0.014315782114863396, + 0.016330191865563393, + 0.008852537721395493, + 0.006929095834493637, + 0.008693077601492405, + 0.010907197371125221, + 0.013411542400717735, + 0.00927137490361929, + 0.005750495009124279, + 0.008498182520270348, + 0.002680255798622966, + 0.0026615294627845287, + 0.0049018277786672115, + 0.003587015438824892, + 0.006967894267290831, + 0.00438935961574316, + 0.004898387007415295, + 0.006321514956653118, + 0.0019869094248861074, + 0.002781027927994728, + 0.004501072224229574, + 0.003616360016167164, + 0.008948665112257004, + 0.0043627177365124226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.02980896644294262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865352809429169 + ], + "opposite_logprob_tokens": [ + -3.0486536026000977 + ], + "logprob_reward": -0.04865352809429169, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865352809429169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08975397795438766, + 0.16220949590206146, + 0.05221150815486908, + 0.14558421075344086, + 0.23029819130897522, + 0.18658646941184998, + 0.08036132156848907, + 0.016219863668084145, + 0.018347684293985367, + 0.018427228555083275 + ], + "credit_map": [ + 0.00998593494296074, + 0.019837306812405586, + 0.015019763261079788, + 0.01732831634581089, + 0.005559598095715046, + 0.006631412543356419, + 0.005194545723497868, + 0.01019711047410965, + 0.036384277045726776, + 0.015085002407431602, + 0.05704823508858681, + 0.030297216027975082, + 0.005023371893912554, + 0.005406487733125687, + 0.005510168150067329, + 0.007454744074493647, + 0.007787750102579594, + 0.01075797714293003, + 0.008904079906642437, + 0.005528272595256567, + 0.005351150408387184, + 0.004052689298987389, + 0.0030662831850349903, + 0.006763309706002474, + 0.01669568382203579, + 0.024558378383517265, + 0.014108926057815552, + 0.015962209552526474, + 0.019000805914402008, + 0.014309183694422245, + 0.024243025109171867, + 0.016706008464097977, + 0.03218993917107582, + 0.03726803883910179, + 0.025300070643424988, + 0.029073596000671387, + 0.017997141927480698, + 0.024303406476974487, + 0.032942574471235275, + 0.031223440542817116, + 0.02834688313305378, + 0.029871517792344093, + 0.027505384758114815, + 0.022650131955742836, + 0.015178319066762924, + 0.014974880032241344, + 0.028224267065525055, + 0.019835088402032852, + 0.011809710413217545, + 0.016224922612309456, + 0.007372750900685787, + 0.007604813203215599, + 0.0066019087098538876, + 0.006684435065835714, + 0.015284770168364048, + 0.00877801701426506, + 0.002219161717221141, + 0.0030550372321158648, + 0.0015723740216344595, + 0.0019366109045222402, + 0.0016527760308235884, + 0.0016817576251924038, + 0.002367068314924836, + 0.0017350796842947602, + 0.00296797719784081, + 0.0038041877560317516, + 0.0018858503317460418, + 0.0015804434660822153, + 0.0015025525353848934, + 0.0016266134334728122, + 0.0029522376134991646, + 0.002027824055403471, + 0.0025580739602446556, + 0.002591569907963276, + 0.0013057903852313757, + 0.0016321084694936872, + 0.0018857893301174045, + 0.0011385257821530104, + 0.00587537931278348, + 0.0014399936189875007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.04865352809429169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic long-range shot, captured from a fixed-altitude drone, captures the first rays of golden sunlight piercing the mist, illuminating the majestic Great Wall nestled between towering mountains. The ancient, mottled gray brickwork of the wall winds its way like a giant dragon, its weathered surface richly textured and cast long, pale blue shadows in the slanting morning light. Against this majestic and timeless backdrop, two incredibly tiny figures advance along the broad path atop the wall. The leading figure, draped in a crimson cloak, stands out against the gray landscape, his gait steady. Following closely behind, a second figure, clad in dark clothing, maintains a few paces behind, following steadily and calmly. Their movements appear insignificant against the vast structure, like two slowly moving dots of ink. The entire image is imbued with an epic tranquility and the atmosphere of a lonely and determined journey.", + "seed": 635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_15/index_4063_seed_635_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07514001438394188, + "vlm_logprob": -0.07514001438394188, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.07514001438394188, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2553868293762207 + ], + "opposite_logprob_tokens": [ + -1.6303868293762207 + ], + "logprob_reward": -0.2553868293762207, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2553868293762207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19301699101924896, + 0.04730147495865822, + 0.1541297286748886, + 0.056909628212451935, + 0.06388947367668152, + 0.06439435482025146, + 0.16033293306827545, + 0.05780778080224991, + 0.09540989249944687, + 0.10680776834487915 + ], + "credit_map": [ + 0.01568988896906376, + 0.02059686742722988, + 0.04060943052172661, + 0.02595428191125393, + 0.022065237164497375, + 0.025707393884658813, + 0.022526105865836143, + 0.0198677871376276, + 0.0051961736753582954, + 0.005137919448316097, + 0.009771982207894325, + 0.004838269203901291, + 0.006233238149434328, + 0.010844707489013672, + 0.00272792368195951, + 0.002551260171458125, + 0.02531358040869236, + 0.011910811066627502, + 0.027804892510175705, + 0.012003373354673386, + 0.011824095621705055, + 0.047599196434020996, + 0.009637371636927128, + 0.008036404848098755, + 0.005151677411049604, + 0.010075671598315239, + 0.01060778833925724, + 0.005747453309595585, + 0.006619859021157026, + 0.007803826127201319, + 0.0048455530777573586, + 0.006057798396795988, + 0.014779569581151009, + 0.007050677668303251, + 0.009575326927006245, + 0.003799429628998041, + 0.004866352304816246, + 0.018585646525025368, + 0.0022976926993578672, + 0.0029347767122089863, + 0.008829950354993343, + 0.01043460052460432, + 0.009223445318639278, + 0.006504727527499199, + 0.008726575411856174, + 0.01034010574221611, + 0.004268086049705744, + 0.006066863425076008, + 0.018611352890729904, + 0.022507037967443466, + 0.025209829211235046, + 0.022688334807753563, + 0.016947168856859207, + 0.018977317959070206, + 0.0176455806940794, + 0.01774630881845951, + 0.006206699181348085, + 0.009304191917181015, + 0.011343448422849178, + 0.0054563372395932674, + 0.006089244969189167, + 0.011350588873028755, + 0.003651391714811325, + 0.004405876621603966, + 0.015697073191404343, + 0.012302477844059467, + 0.014555887319147587, + 0.006742123514413834, + 0.007573840674012899, + 0.02640586532652378, + 0.005285212304443121, + 0.006847405340522528, + 0.016261018812656403, + 0.0156470388174057, + 0.01629977487027645, + 0.007108746096491814, + 0.013008929789066315, + 0.026286449283361435, + 0.006853316444903612, + 0.005342494230717421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.2553868293762207 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052775111980736256 + ], + "opposite_logprob_tokens": [ + -5.255277633666992 + ], + "logprob_reward": -0.0052775111980736256, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052775111980736256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1630839854478836, + 0.07400619238615036, + 0.24148041009902954, + 0.12137771397829056, + 0.06351681053638458, + 0.07172632962465286, + 0.14306284487247467, + 0.028229137882590294, + 0.02995089627802372, + 0.06356567144393921 + ], + "credit_map": [ + 0.013015968725085258, + 0.021586645394563675, + 0.02342355065047741, + 0.020524336025118828, + 0.02469387650489807, + 0.022871149703860283, + 0.01663811504840851, + 0.02033034898340702, + 0.007214056793600321, + 0.00959353893995285, + 0.014606280252337456, + 0.007261686958372593, + 0.010077317245304585, + 0.013660049065947533, + 0.005801084451377392, + 0.005792180076241493, + 0.04514774680137634, + 0.014786357060074806, + 0.050331585109233856, + 0.007750332821160555, + 0.015737073495984077, + 0.09247318655252457, + 0.006881712470203638, + 0.008372417651116848, + 0.01820763200521469, + 0.012570612132549286, + 0.01918647065758705, + 0.005984021350741386, + 0.010952998884022236, + 0.03584684431552887, + 0.007513576652854681, + 0.011115544475615025, + 0.012388437986373901, + 0.005830191075801849, + 0.011964863166213036, + 0.002580630360171199, + 0.005046759732067585, + 0.020350312814116478, + 0.0025157704949378967, + 0.002839843975380063, + 0.008893406018614769, + 0.0070615108124911785, + 0.00895181018859148, + 0.009130152873694897, + 0.007923110388219357, + 0.011162940412759781, + 0.00608914066106081, + 0.012514257803559303, + 0.02195865660905838, + 0.015046444721519947, + 0.020117705687880516, + 0.01237497292459011, + 0.021709397435188293, + 0.02489256113767624, + 0.01350595522671938, + 0.01345714833587408, + 0.0037798271514475346, + 0.003829024964943528, + 0.005080134607851505, + 0.0021509923972189426, + 0.004672300536185503, + 0.005084123462438583, + 0.0020740050822496414, + 0.0015587281668558717, + 0.004648275673389435, + 0.002747373888269067, + 0.004256481770426035, + 0.0018985207425430417, + 0.0038392425049096346, + 0.007946694269776344, + 0.002251030644401908, + 0.0023632787633687258, + 0.01336496602743864, + 0.009104671888053417, + 0.007754826452583075, + 0.004419623874127865, + 0.005128215998411179, + 0.01780972070991993, + 0.003185806330293417, + 0.0027978471480309963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0052775111980736256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349563151597977 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349563151597977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349563151597977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23238709568977356, + 0.05582903325557709, + 0.15335220098495483, + 0.059243135154247284, + 0.09026720374822617, + 0.07440146803855896, + 0.12795482575893402, + 0.03587011620402336, + 0.07222921401262283, + 0.09846574068069458 + ], + "credit_map": [ + 0.01256136316806078, + 0.044061172753572464, + 0.038475893437862396, + 0.014356294646859169, + 0.06672336161136627, + 0.03166193142533302, + 0.014431959018111229, + 0.010115093551576138, + 0.008179509080946445, + 0.0071089849807322025, + 0.007701844442635775, + 0.0058038015849888325, + 0.005779072642326355, + 0.012267056852579117, + 0.004597460385411978, + 0.00439129862934351, + 0.023194169625639915, + 0.013641892932355404, + 0.028929350897669792, + 0.00980140920728445, + 0.013945596292614937, + 0.04431477561593056, + 0.007848178967833519, + 0.011676819995045662, + 0.004529857542365789, + 0.008046546950936317, + 0.007594389840960503, + 0.006930902134627104, + 0.007134042680263519, + 0.008602081798017025, + 0.00675064604729414, + 0.009654661640524864, + 0.017742516472935677, + 0.00986167136579752, + 0.014692400582134724, + 0.008095866069197655, + 0.006636306177824736, + 0.023613378405570984, + 0.004986907821148634, + 0.004638160578906536, + 0.011714207008481026, + 0.00888222735375166, + 0.011863877065479755, + 0.007734678685665131, + 0.008223244920372963, + 0.01542247086763382, + 0.004635723773390055, + 0.005925031844526529, + 0.012288305908441544, + 0.01530416402965784, + 0.018823476508259773, + 0.014485983178019524, + 0.02136622928082943, + 0.01831192523241043, + 0.013985048048198223, + 0.013389683328568935, + 0.0036868678871542215, + 0.005158168263733387, + 0.004633014090359211, + 0.0037969236727803946, + 0.00683995196595788, + 0.0058867051266133785, + 0.0026411155704408884, + 0.0032273654360324144, + 0.011679981835186481, + 0.007372260093688965, + 0.010299001820385456, + 0.004822884686291218, + 0.010879270732402802, + 0.018871106207370758, + 0.003833785420283675, + 0.004470914136618376, + 0.019196534529328346, + 0.016457870602607727, + 0.01196589320898056, + 0.005175634752959013, + 0.009971494786441326, + 0.026052331551909447, + 0.004960460122674704, + 0.004685520194470882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.026349563151597977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372656926512718 + ], + "opposite_logprob_tokens": [ + -3.408726453781128 + ], + "logprob_reward": -0.03372656926512718, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372656926512718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1422591507434845, + 0.10859270393848419, + 0.18993961811065674, + 0.10001346468925476, + 0.04047057032585144, + 0.17383845150470734, + 0.09665968269109726, + 0.05286078900098801, + 0.048238880932331085, + 0.04712660610675812 + ], + "credit_map": [ + 0.016159264370799065, + 0.02310190163552761, + 0.020980779081583023, + 0.018789418041706085, + 0.020676875486969948, + 0.018192630261182785, + 0.013174989260733128, + 0.011183290742337704, + 0.010724097490310669, + 0.018916569650173187, + 0.021869666874408722, + 0.009508570656180382, + 0.013347270898520947, + 0.023378795012831688, + 0.0062170978635549545, + 0.004630638752132654, + 0.030587319284677505, + 0.015754669904708862, + 0.03604486212134361, + 0.011923239566385746, + 0.014681736938655376, + 0.06016167625784874, + 0.009314339607954025, + 0.011471790261566639, + 0.015062158927321434, + 0.014620833098888397, + 0.01613982580602169, + 0.008730808272957802, + 0.009380556643009186, + 0.02208627015352249, + 0.005978123284876347, + 0.008014899678528309, + 0.004234598018229008, + 0.006032764911651611, + 0.005626228172332048, + 0.005415864754468203, + 0.0044273873791098595, + 0.0058843279257416725, + 0.004029004834592342, + 0.004820396192371845, + 0.025019332766532898, + 0.019463632255792618, + 0.021299980580806732, + 0.017689798027276993, + 0.01721077784895897, + 0.03449530899524689, + 0.012669178657233715, + 0.025990450754761696, + 0.01780318282544613, + 0.013620925135910511, + 0.013410995714366436, + 0.006741578225046396, + 0.011506902053952217, + 0.02325468696653843, + 0.004786852281540632, + 0.005534567404538393, + 0.0075500826351344585, + 0.008740597404539585, + 0.0069534918293356895, + 0.0039034762885421515, + 0.007750576827675104, + 0.011446340009570122, + 0.0030350012239068747, + 0.0034812227822840214, + 0.004680491518229246, + 0.011173270642757416, + 0.006670414470136166, + 0.0037012656684964895, + 0.0073066772893071175, + 0.007768996525555849, + 0.0034880773164331913, + 0.0034496914595365524, + 0.006939291022717953, + 0.007156061474233866, + 0.007753091864287853, + 0.005158588290214539, + 0.004628842696547508, + 0.008946155197918415, + 0.0036091816145926714, + 0.0029354006983339787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.03372656926512718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495959892868996 + ], + "opposite_logprob_tokens": [ + -2.929959535598755 + ], + "logprob_reward": -0.05495959892868996, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495959892868996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20404674112796783, + 0.14508947730064392, + 0.09978931397199631, + 0.1010480597615242, + 0.07860594242811203, + 0.07161229103803635, + 0.06813803315162659, + 0.03694641962647438, + 0.05495791882276535, + 0.1397658884525299 + ], + "credit_map": [ + 0.021307867020368576, + 0.015758091583848, + 0.05174485966563225, + 0.02241474948823452, + 0.02352924272418022, + 0.03254566341638565, + 0.02187001146376133, + 0.014876244589686394, + 0.01337825134396553, + 0.027235934510827065, + 0.03184094280004501, + 0.01778344064950943, + 0.015119917690753937, + 0.024084774777293205, + 0.007687019649893045, + 0.00795917771756649, + 0.01431796420365572, + 0.007194018457084894, + 0.01891096867620945, + 0.007171601057052612, + 0.0076242657378315926, + 0.03166762366890907, + 0.007085768505930901, + 0.005817094352096319, + 0.012390990741550922, + 0.012656363658607006, + 0.018495650961995125, + 0.008342170156538486, + 0.00976039469242096, + 0.02515125647187233, + 0.0056126900017261505, + 0.00863853469491005, + 0.01828683167695999, + 0.007333294954150915, + 0.012427734211087227, + 0.004079064819961786, + 0.005212839692831039, + 0.02418525144457817, + 0.003299761563539505, + 0.0037811598740518093, + 0.008973627351224422, + 0.010493002831935883, + 0.009204134345054626, + 0.007303919177502394, + 0.006457431707531214, + 0.013755292631685734, + 0.004935161210596561, + 0.010489706881344318, + 0.0080778319388628, + 0.008110057562589645, + 0.010695663280785084, + 0.006697023753076792, + 0.008399642072618008, + 0.014134835451841354, + 0.006066999863833189, + 0.0059559764340519905, + 0.005444540176540613, + 0.005667522549629211, + 0.0047422959469258785, + 0.003332156455144286, + 0.005469682160764933, + 0.006295790430158377, + 0.0032509092707186937, + 0.0027435182128101587, + 0.010523336008191109, + 0.006819516886025667, + 0.007287277840077877, + 0.002622680738568306, + 0.004657880403101444, + 0.017375176772475243, + 0.0021239323541522026, + 0.003548108506947756, + 0.024839237332344055, + 0.019903331995010376, + 0.017838148400187492, + 0.008101235143840313, + 0.018699830397963524, + 0.037742964923381805, + 0.00675672572106123, + 0.005884391721338034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.05495959892868996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic long-range shot, captured from a fixed-altitude drone, captures the first rays of golden sunlight piercing the mist, illuminating the majestic Great Wall nestled between towering mountains. The ancient, mottled gray brickwork of the wall winds its way like a giant dragon, its weathered surface richly textured and cast long, pale blue shadows in the slanting morning light. Against this majestic and timeless backdrop, two incredibly tiny figures advance along the broad path atop the wall. The leading figure, draped in a crimson cloak, stands out against the gray landscape, his gait steady. Following closely behind, a second figure, clad in dark clothing, maintains a few paces behind, following steadily and calmly. Their movements appear insignificant against the vast structure, like two slowly moving dots of ink. The entire image is imbued with an epic tranquility and the atmosphere of a lonely and determined journey.", + "seed": 636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_15/index_4063_seed_636_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13915745820850134, + "vlm_logprob": -0.13915745820850134, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.13915745820850134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39442071318626404 + ], + "opposite_logprob_tokens": [ + -1.2694207429885864 + ], + "logprob_reward": -0.39442071318626404, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.39442071318626404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18920394778251648, + 0.0327354297041893, + 0.038475923240184784, + 0.060196854174137115, + 0.054554592818021774, + 0.030549585819244385, + 0.031321387737989426, + 0.03989715501666069, + 0.06319012492895126, + 0.4598749876022339 + ], + "credit_map": [ + 0.02223304845392704, + 0.06746970862150192, + 0.012934045866131783, + 0.009823989123106003, + 0.04386579245328903, + 0.014484482817351818, + 0.005398983135819435, + 0.0129938879981637, + 0.002898014383390546, + 0.0046177152544260025, + 0.005982662085443735, + 0.0027078804560005665, + 0.002971871756017208, + 0.005434292834252119, + 0.002657528268173337, + 0.0054654632695019245, + 0.003895397065207362, + 0.005853016395121813, + 0.0038615053053945303, + 0.003442129585891962, + 0.005171719007194042, + 0.006217836868017912, + 0.0025162792298942804, + 0.007518039550632238, + 0.006871909834444523, + 0.02103983424603939, + 0.00490980502218008, + 0.004125967156141996, + 0.008989885449409485, + 0.006032979115843773, + 0.003632287960499525, + 0.004594183061271906, + 0.004106689244508743, + 0.004890700802206993, + 0.0030268372502177954, + 0.00621454743668437, + 0.0234205462038517, + 0.006131830159574747, + 0.003772704629227519, + 0.00299073220230639, + 0.002322218846529722, + 0.007082192227244377, + 0.0027581611648201942, + 0.0033920828718692064, + 0.006332492455840111, + 0.0025412850081920624, + 0.002822211943566799, + 0.0032989385072141886, + 0.0024463501758873463, + 0.0036794175393879414, + 0.005382859613746405, + 0.002601852873340249, + 0.003936384338885546, + 0.0071661886759102345, + 0.002431213855743408, + 0.0036771183367818594, + 0.004629232920706272, + 0.004647222813218832, + 0.005824110005050898, + 0.0030532036907970905, + 0.00625381525605917, + 0.006557796616107225, + 0.0033002078998833895, + 0.005631564185023308, + 0.006745276041328907, + 0.016042860224843025, + 0.00940045714378357, + 0.003908345941454172, + 0.006825817748904228, + 0.006844004150480032, + 0.005095662549138069, + 0.008327699266374111, + 0.054039958864450455, + 0.1561741828918457, + 0.01683862693607807, + 0.034375205636024475, + 0.02166627161204815, + 0.023264100775122643, + 0.024272294715046883, + 0.1292443573474884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.39442071318626404 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011105983518064022 + ], + "opposite_logprob_tokens": [ + -4.511106014251709 + ], + "logprob_reward": -0.011105983518064022, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011105983518064022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10832186043262482, + 0.031204329803586006, + 0.08963280916213989, + 0.013563383370637894, + 0.017611196264624596, + 0.02442631870508194, + 0.028087193146348, + 0.028261613100767136, + 0.013426687568426132, + 0.6454645991325378 + ], + "credit_map": [ + 0.014740613289177418, + 0.041654475033283234, + 0.006917956750839949, + 0.005587438587099314, + 0.02187107317149639, + 0.008255798369646072, + 0.002683699131011963, + 0.006610810291022062, + 0.004094711039215326, + 0.010869747027754784, + 0.0026777724269777536, + 0.0021328935399651527, + 0.00311041297391057, + 0.004960745107382536, + 0.0011527696624398232, + 0.0022052829153835773, + 0.007720078341662884, + 0.02117043174803257, + 0.00739964609965682, + 0.005145200528204441, + 0.01592108979821205, + 0.020325463265180588, + 0.0036666844971477985, + 0.008284217678010464, + 0.0016736662946641445, + 0.003171207383275032, + 0.0016983847599476576, + 0.0010768516222015023, + 0.0016069982666522264, + 0.0015639987541362643, + 0.0012714514741674066, + 0.0015008270274847746, + 0.0019891448318958282, + 0.0031505385413765907, + 0.0014764220686629415, + 0.0016920949565246701, + 0.004896844271570444, + 0.0017858825158327818, + 0.001495596137829125, + 0.001124675152823329, + 0.0027560433372855186, + 0.008614481426775455, + 0.0018009936902672052, + 0.001597816590219736, + 0.0036699590273201466, + 0.0019147774437442422, + 0.0012787234736606479, + 0.0027935244143009186, + 0.003535808064043522, + 0.007924356497824192, + 0.0026996361557394266, + 0.0015326638240367174, + 0.002724555553868413, + 0.006077341269701719, + 0.0015837750397622585, + 0.002009060001000762, + 0.0030247808899730444, + 0.005443343427032232, + 0.0021635862067341805, + 0.002684052102267742, + 0.005625618621706963, + 0.004949113819748163, + 0.0019744914025068283, + 0.002396626863628626, + 0.0013024768559262156, + 0.0023242454044520855, + 0.0024474505335092545, + 0.0009380750125274062, + 0.001804567058570683, + 0.001758140278980136, + 0.0012691295705735683, + 0.0015826034359633923, + 0.0758143812417984, + 0.21836186945438385, + 0.023688001558184624, + 0.04859855771064758, + 0.030951321125030518, + 0.032511256635189056, + 0.03371771425008774, + 0.1818215399980545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.011105983518064022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370178863406181 + ], + "opposite_logprob_tokens": [ + -3.4087018966674805 + ], + "logprob_reward": -0.03370178863406181, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370178863406181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3985874652862549, + 0.04344402998685837, + 0.03419279307126999, + 0.029658889397978783, + 0.10678721219301224, + 0.010389565490186214, + 0.00690002366900444, + 0.028967222198843956, + 0.018028629943728447, + 0.32304418087005615 + ], + "credit_map": [ + 0.06352680921554565, + 0.17681802809238434, + 0.03397761285305023, + 0.015974167734384537, + 0.060704752802848816, + 0.02855464443564415, + 0.007518255617469549, + 0.01151319406926632, + 0.006043694447726011, + 0.01637214981019497, + 0.0036922418512403965, + 0.002791588194668293, + 0.003995547071099281, + 0.005866088438779116, + 0.0020493771880865097, + 0.002633339026942849, + 0.0034779137931764126, + 0.008031248115003109, + 0.0037258868105709553, + 0.002477681264281273, + 0.005249397829174995, + 0.00600954657420516, + 0.0019143226090818644, + 0.0033067958429455757, + 0.003131051314994693, + 0.004204906988888979, + 0.002927830209955573, + 0.00226411921903491, + 0.006048806011676788, + 0.004582286346703768, + 0.0023922401014715433, + 0.004107649438083172, + 0.006557611748576164, + 0.024154163897037506, + 0.0047646788880229, + 0.009703034535050392, + 0.04071664810180664, + 0.008770531043410301, + 0.006858809385448694, + 0.0052617331966757774, + 0.0009112185798585415, + 0.0013632902409881353, + 0.0014356637839227915, + 0.001395583269186318, + 0.0014493835624307394, + 0.0010749140055850148, + 0.0013472752179950476, + 0.00141223706305027, + 0.0005439015803858638, + 0.000731015345081687, + 0.0011274205753579736, + 0.0005477464874275029, + 0.0007910886197350919, + 0.0013842704938724637, + 0.0005756810423918068, + 0.0011988991172984242, + 0.0035917593631893396, + 0.0038174365181475878, + 0.0033832453191280365, + 0.0019134884933009744, + 0.006618250161409378, + 0.004810500890016556, + 0.001998655032366514, + 0.0028338858392089605, + 0.0016334888059645891, + 0.00493759848177433, + 0.0022917960304766893, + 0.001268599065952003, + 0.0022518804762512445, + 0.0021522180177271366, + 0.0014040782116353512, + 0.0020889712031930685, + 0.03782428056001663, + 0.10904022306203842, + 0.012211032211780548, + 0.024328090250492096, + 0.015197759494185448, + 0.01629307121038437, + 0.01711486279964447, + 0.09103485196828842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.03370178863406181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499051511287689 + ], + "opposite_logprob_tokens": [ + -2.929990530014038 + ], + "logprob_reward": -0.05499051511287689, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499051511287689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32419148087501526, + 0.04202559217810631, + 0.03700632229447365, + 0.028211714699864388, + 0.08130130916833878, + 0.012944050133228302, + 0.028999246656894684, + 0.029065271839499474, + 0.013985015451908112, + 0.40227001905441284 + ], + "credit_map": [ + 0.05364037677645683, + 0.14589257538318634, + 0.029141224920749664, + 0.01279026735574007, + 0.04108600318431854, + 0.02478506974875927, + 0.005817065015435219, + 0.011038868688046932, + 0.0054736146703362465, + 0.011724713258445263, + 0.004244779236614704, + 0.0030257150065153837, + 0.004045555833727121, + 0.005108085460960865, + 0.0019939872436225414, + 0.006409139838069677, + 0.003912437241524458, + 0.008679413236677647, + 0.0039320774376392365, + 0.0029075131751596928, + 0.005569557659327984, + 0.0057229273952543736, + 0.0021636683959513903, + 0.004118725657463074, + 0.0025391960516572, + 0.004289266187697649, + 0.002379454905167222, + 0.0022236709482967854, + 0.00636251550167799, + 0.00454391585662961, + 0.002190567087382078, + 0.0036831279285252094, + 0.005788914393633604, + 0.02113596349954605, + 0.0031647838186472654, + 0.006865076720714569, + 0.03176349028944969, + 0.005925056524574757, + 0.003748376155272126, + 0.0029096449725329876, + 0.00119139498565346, + 0.0019221638794988394, + 0.0014994946541264653, + 0.0019607713911682367, + 0.0017769914120435715, + 0.0013660233234986663, + 0.0009246733388863504, + 0.002302537439391017, + 0.0029566879384219646, + 0.0058298720978200436, + 0.003671251004561782, + 0.0018794885836541653, + 0.0026559638790786266, + 0.006953986827284098, + 0.002406805520877242, + 0.0026451912708580494, + 0.0032722626347094774, + 0.003271408611908555, + 0.003210801864042878, + 0.0028957659378647804, + 0.004931985400617123, + 0.004666473716497421, + 0.0036432784982025623, + 0.00317329540848732, + 0.001403002766892314, + 0.0025325280148535967, + 0.0025956707540899515, + 0.0011245358036831021, + 0.0016319468850269914, + 0.001698112115263939, + 0.0012850634520873427, + 0.0017141563585028052, + 0.04686907306313515, + 0.13482564687728882, + 0.01509195938706398, + 0.03046124242246151, + 0.019689349457621574, + 0.020421713590621948, + 0.021284036338329315, + 0.11362698674201965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.05499051511287689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156829059123993 + ], + "opposite_logprob_tokens": [ + -1.7015682458877563 + ], + "logprob_reward": -0.20156829059123993, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156829059123993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24595364928245544, + 0.07678740471601486, + 0.0829881951212883, + 0.04545937106013298, + 0.03424186632037163, + 0.02076694741845131, + 0.05911730229854584, + 0.08179832994937897, + 0.037667371332645416, + 0.31521955132484436 + ], + "credit_map": [ + 0.030739890411496162, + 0.09428783506155014, + 0.018231814727187157, + 0.012763072736561298, + 0.04943634942173958, + 0.021457795053720474, + 0.006342583801597357, + 0.012694324366748333, + 0.010389800183475018, + 0.023611435666680336, + 0.006745744496583939, + 0.004541600588709116, + 0.009594804607331753, + 0.012091215699911118, + 0.00412732595577836, + 0.005685488693416119, + 0.005988781340420246, + 0.015892228111624718, + 0.005962715484201908, + 0.006199459079653025, + 0.018531112000346184, + 0.0159620251506567, + 0.005158571060746908, + 0.009293314069509506, + 0.004303105175495148, + 0.0069180261343717575, + 0.004393238108605146, + 0.0035647093318402767, + 0.011114698834717274, + 0.007028509862720966, + 0.0028107482939958572, + 0.005326340440660715, + 0.0036744314711540937, + 0.0061051431111991405, + 0.0034158374182879925, + 0.0031958292238414288, + 0.006367720663547516, + 0.0054559349082410336, + 0.0032628506887704134, + 0.0027641234919428825, + 0.002509730402380228, + 0.006156662944704294, + 0.0013521884102374315, + 0.0025249947793781757, + 0.003055623499676585, + 0.001973771722987294, + 0.0014447427820414305, + 0.0017492348561063409, + 0.007442324887961149, + 0.015953265130519867, + 0.006677540484815836, + 0.003730849362909794, + 0.005777918733656406, + 0.010059925727546215, + 0.003966609947383404, + 0.00550887593999505, + 0.008293027058243752, + 0.01876094937324524, + 0.00537874223664403, + 0.0050748856738209724, + 0.019931679591536522, + 0.010553151369094849, + 0.007025535684078932, + 0.006780372001230717, + 0.00442523043602705, + 0.005933400243520737, + 0.005867485422641039, + 0.002903333632275462, + 0.003813004121184349, + 0.00650591729208827, + 0.0027956117410212755, + 0.005423392169177532, + 0.03645133972167969, + 0.10788265615701675, + 0.011345124803483486, + 0.023529013618826866, + 0.01516931876540184, + 0.01598183438181877, + 0.016488922759890556, + 0.08837136626243591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.20156829059123993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic long-range shot, captured from a fixed-altitude drone, captures the first rays of golden sunlight piercing the mist, illuminating the majestic Great Wall nestled between towering mountains. The ancient, mottled gray brickwork of the wall winds its way like a giant dragon, its weathered surface richly textured and cast long, pale blue shadows in the slanting morning light. Against this majestic and timeless backdrop, two incredibly tiny figures advance along the broad path atop the wall. The leading figure, draped in a crimson cloak, stands out against the gray landscape, his gait steady. Following closely behind, a second figure, clad in dark clothing, maintains a few paces behind, following steadily and calmly. Their movements appear insignificant against the vast structure, like two slowly moving dots of ink. The entire image is imbued with an epic tranquility and the atmosphere of a lonely and determined journey.", + "seed": 637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_15/index_4063_seed_637_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08805906721390784, + "vlm_logprob": -0.08805906721390784, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.08805906721390784, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3112943172454834 + ], + "opposite_logprob_tokens": [ + -1.4362943172454834 + ], + "logprob_reward": -0.3112943172454834, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3112943172454834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1843818873167038, + 0.08798828721046448, + 0.36109259724617004, + 0.07853378355503082, + 0.029486751183867455, + 0.0412413626909256, + 0.08460289239883423, + 0.05412714183330536, + 0.05981341004371643, + 0.01873180828988552 + ], + "credit_map": [ + 0.02014814130961895, + 0.015873583033680916, + 0.03937772661447525, + 0.02655823901295662, + 0.014943117275834084, + 0.016297444701194763, + 0.031239140778779984, + 0.01994452252984047, + 0.009127266705036163, + 0.007503402419388294, + 0.010045508854091167, + 0.010991602204740047, + 0.007382682524621487, + 0.009574060328304768, + 0.018502669408917427, + 0.014861101284623146, + 0.02505546063184738, + 0.019100088626146317, + 0.06788218021392822, + 0.043255388736724854, + 0.03190227970480919, + 0.0517866425216198, + 0.0699184238910675, + 0.05219216272234917, + 0.006540858652442694, + 0.006779496558010578, + 0.01152405608445406, + 0.00737744802609086, + 0.00655350973829627, + 0.012506214901804924, + 0.010827441699802876, + 0.016424769535660744, + 0.0025858047883957624, + 0.0025702877901494503, + 0.004405653104186058, + 0.005443754140287638, + 0.002770828315988183, + 0.0027525071054697037, + 0.004055109340697527, + 0.004902810789644718, + 0.005062873940914869, + 0.005625351797789335, + 0.0071057905443012714, + 0.00487134326249361, + 0.004774394910782576, + 0.004369406029582024, + 0.003637045156210661, + 0.005795166362076998, + 0.005487102549523115, + 0.01380335446447134, + 0.019150087609887123, + 0.00791169237345457, + 0.01615387760102749, + 0.008695165626704693, + 0.006380863022059202, + 0.007020751014351845, + 0.004422887694090605, + 0.008345192298293114, + 0.008895941078662872, + 0.00460837222635746, + 0.007270555943250656, + 0.005591267254203558, + 0.0036099280696362257, + 0.011383002623915672, + 0.008996997028589249, + 0.009354368783533573, + 0.008575632236897945, + 0.010332858189940453, + 0.006793506443500519, + 0.006528747268021107, + 0.003930612467229366, + 0.00530069787055254, + 0.0021653606090694666, + 0.0026762757916003466, + 0.002767436672002077, + 0.0034011031966656446, + 0.0017855055630207062, + 0.002435779431834817, + 0.001482787891291082, + 0.002017561113461852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.3112943172454834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025206005666404963 + ], + "opposite_logprob_tokens": [ + -6.002520561218262 + ], + "logprob_reward": -0.0025206005666404963, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025206005666404963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3139459490776062, + 0.15279676020145416, + 0.07736869156360626, + 0.07197100669145584, + 0.06132638826966286, + 0.041810836642980576, + 0.13732610642910004, + 0.035078469663858414, + 0.04204277694225311, + 0.06633304804563522 + ], + "credit_map": [ + 0.027945689857006073, + 0.03530912846326828, + 0.042486052960157394, + 0.037307579070329666, + 0.03384369984269142, + 0.03482005000114441, + 0.055564865469932556, + 0.046668872237205505, + 0.01403470803052187, + 0.01120242103934288, + 0.01796608604490757, + 0.020841816440224648, + 0.013444870710372925, + 0.013538901694118977, + 0.032029684633016586, + 0.02973826602101326, + 0.004819871857762337, + 0.005948756355792284, + 0.00914741214364767, + 0.011693277396261692, + 0.010203544981777668, + 0.008224844932556152, + 0.01917850598692894, + 0.008152478374540806, + 0.005037452559918165, + 0.004912683740258217, + 0.01659102737903595, + 0.015888836234807968, + 0.00633401982486248, + 0.008485935628414154, + 0.006740127224475145, + 0.00798092596232891, + 0.0033327999990433455, + 0.003644675714895129, + 0.010430984199047089, + 0.009305410087108612, + 0.004611887037754059, + 0.0052847787737846375, + 0.012323901988565922, + 0.012391949072480202, + 0.004018108826130629, + 0.0058489455841481686, + 0.009020401164889336, + 0.004165295977145433, + 0.005331156309694052, + 0.005736648570746183, + 0.0029183218721300364, + 0.004771958105266094, + 0.014788277447223663, + 0.028750784695148468, + 0.025259004905819893, + 0.016587013378739357, + 0.020266905426979065, + 0.012540610507130623, + 0.010108689777553082, + 0.009024822153151035, + 0.004244685173034668, + 0.006457193288952112, + 0.005976324435323477, + 0.0049453843384981155, + 0.004430554341524839, + 0.002490401268005371, + 0.0024334003683179617, + 0.004100524354726076, + 0.004378057084977627, + 0.00697696628049016, + 0.0058521986939013, + 0.009106180630624294, + 0.004680342972278595, + 0.0033128373324871063, + 0.003328577848151326, + 0.004407616797834635, + 0.007640790659934282, + 0.012090486474335194, + 0.011050879955291748, + 0.008123403415083885, + 0.0068313730880618095, + 0.009362596087157726, + 0.00439893314614892, + 0.0068345810286700726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0025206005666404963 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083466000854969 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083466000854969, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083466000854969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25175291299819946, + 0.07619225978851318, + 0.1868974268436432, + 0.10928483307361603, + 0.052484843879938126, + 0.037716906517744064, + 0.10112398862838745, + 0.05605761334300041, + 0.0806109830737114, + 0.0478782057762146 + ], + "credit_map": [ + 0.03023163042962551, + 0.01776391640305519, + 0.07080129534006119, + 0.03152245655655861, + 0.02174147218465805, + 0.03735442832112312, + 0.022364286705851555, + 0.019973469898104668, + 0.005926347337663174, + 0.006237759254872799, + 0.009821860119700432, + 0.010098272934556007, + 0.00623053964227438, + 0.013299578800797462, + 0.01146958488970995, + 0.013108320534229279, + 0.013626840896904469, + 0.009740595705807209, + 0.03932062163949013, + 0.02126268483698368, + 0.018140671774744987, + 0.0211985744535923, + 0.039422448724508286, + 0.024185020476579666, + 0.013067692518234253, + 0.006365652196109295, + 0.02880037948489189, + 0.01245015487074852, + 0.010701064020395279, + 0.008086847141385078, + 0.016017744317650795, + 0.013795306906104088, + 0.0030578593723475933, + 0.0029437970370054245, + 0.00843073707073927, + 0.00664653442800045, + 0.004555861931294203, + 0.005131850019097328, + 0.011216772720217705, + 0.010501439683139324, + 0.0030951809603720903, + 0.004942243453115225, + 0.00901444349437952, + 0.003192536998540163, + 0.005878796335309744, + 0.004405141342431307, + 0.0026672668755054474, + 0.004521301481872797, + 0.012401841580867767, + 0.015292428433895111, + 0.01794082671403885, + 0.010056845843791962, + 0.01625588722527027, + 0.008756025694310665, + 0.01144105102866888, + 0.008979096077382565, + 0.004183783195912838, + 0.009842981584370136, + 0.00886463187634945, + 0.004321580287069082, + 0.009567438624799252, + 0.007743136491626501, + 0.0043051401153206825, + 0.007228929549455643, + 0.008171936497092247, + 0.013347883708775043, + 0.011795975267887115, + 0.014887519180774689, + 0.008859594352543354, + 0.010481428354978561, + 0.005892151966691017, + 0.007174502592533827, + 0.0054575433023273945, + 0.006648354232311249, + 0.006356367375701666, + 0.006757085677236319, + 0.004406047984957695, + 0.0055914814583957195, + 0.005713016726076603, + 0.00694831321015954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011083466000854969 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356181129813194 + ], + "opposite_logprob_tokens": [ + -3.6513562202453613 + ], + "logprob_reward": -0.026356181129813194, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356181129813194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14260879158973694, + 0.18185663223266602, + 0.03407004848122597, + 0.0434701144695282, + 0.05509480461478233, + 0.05859167128801346, + 0.06200729310512543, + 0.22528593242168427, + 0.14813996851444244, + 0.04887469857931137 + ], + "credit_map": [ + 0.01732487604022026, + 0.009514122270047665, + 0.04097505658864975, + 0.01927162893116474, + 0.014068914577364922, + 0.012691658921539783, + 0.016022207215428352, + 0.012740334495902061, + 0.01846267096698284, + 0.013831671327352524, + 0.025120534002780914, + 0.02207375317811966, + 0.015048753470182419, + 0.016749026253819466, + 0.03863827511668205, + 0.03193197026848793, + 0.0030115176923573017, + 0.004095545504242182, + 0.004283499903976917, + 0.004337322432547808, + 0.0036221975460648537, + 0.0035236936528235674, + 0.006044012028723955, + 0.00515225762501359, + 0.004470850341022015, + 0.0032663855236023664, + 0.0070530022494494915, + 0.0054899416863918304, + 0.005140988621860743, + 0.0038210393395274878, + 0.005456698592752218, + 0.00877121090888977, + 0.003713472280651331, + 0.0031942613422870636, + 0.009560247883200645, + 0.0066170282661914825, + 0.005769882816821337, + 0.006642939522862434, + 0.008882489055395126, + 0.010714485310018063, + 0.0054757362231612206, + 0.008167519234120846, + 0.011359207332134247, + 0.007059027906507254, + 0.009418327361345291, + 0.005854653660207987, + 0.004422336351126432, + 0.006834867410361767, + 0.0056046186946332455, + 0.012480841018259525, + 0.008593576028943062, + 0.006519104819744825, + 0.011220322921872139, + 0.006210408639162779, + 0.005860516801476479, + 0.005517905578017235, + 0.021222000941634178, + 0.0444437637925148, + 0.02848988212645054, + 0.015946006402373314, + 0.04340957850217819, + 0.03388965129852295, + 0.015783125534653664, + 0.022101938724517822, + 0.014790753833949566, + 0.02323880046606064, + 0.018323974683880806, + 0.01711118221282959, + 0.03171548619866371, + 0.01974109187722206, + 0.010162493214011192, + 0.013056186959147453, + 0.005837222095578909, + 0.006364116910845041, + 0.006331008393317461, + 0.0077787134796381, + 0.004187919199466705, + 0.004261142108589411, + 0.007300860248506069, + 0.006813716143369675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.026356181129813194 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene captured from a fixed-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall visible nestled between towering mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two tiny figures visible advancing along the path atop the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the leading figure wearing a crimson cloak and the following figure wearing dark clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Great Wall illuminated by golden sunlight piercing through mist with long pale blue shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904077112674713 + ], + "opposite_logprob_tokens": [ + -2.464040756225586 + ], + "logprob_reward": -0.08904077112674713, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904077112674713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1986420750617981, + 0.12446030229330063, + 0.22486285865306854, + 0.08972958475351334, + 0.04366535693407059, + 0.022891269996762276, + 0.06039518117904663, + 0.05842138081789017, + 0.1108262687921524, + 0.06610579788684845 + ], + "credit_map": [ + 0.019835378974676132, + 0.02591925673186779, + 0.0367431603372097, + 0.026466019451618195, + 0.022271031513810158, + 0.021880721673369408, + 0.02834518626332283, + 0.017181318253278732, + 0.012011543847620487, + 0.01061665266752243, + 0.016937479376792908, + 0.014413519762456417, + 0.011377902701497078, + 0.010798496194183826, + 0.027383722364902496, + 0.020920975133776665, + 0.016204386949539185, + 0.013593711890280247, + 0.045026738196611404, + 0.024296261370182037, + 0.020847290754318237, + 0.021663013845682144, + 0.050981082022190094, + 0.03225037083029747, + 0.006842197384685278, + 0.008594891056418419, + 0.018749097362160683, + 0.013985621742904186, + 0.008784370496869087, + 0.013114568777382374, + 0.008157646283507347, + 0.011501200497150421, + 0.0033168289810419083, + 0.004562899004667997, + 0.007325352635234594, + 0.005215308628976345, + 0.005361094605177641, + 0.008322149515151978, + 0.003444425296038389, + 0.006117300596088171, + 0.0021683762315660715, + 0.003217338351532817, + 0.0037394193932414055, + 0.00266966107301414, + 0.003504569409415126, + 0.0025676258374005556, + 0.0018150582909584045, + 0.0032092202454805374, + 0.006883099675178528, + 0.011416725814342499, + 0.008527769707143307, + 0.0063355970196425915, + 0.011269865557551384, + 0.006354090292006731, + 0.005302144680172205, + 0.004305894020944834, + 0.0038671069778501987, + 0.00841484684497118, + 0.008627616800367832, + 0.006181391887366772, + 0.007316785864531994, + 0.005330940242856741, + 0.003866790095344186, + 0.014815895818173885, + 0.010790370404720306, + 0.01721048168838024, + 0.013324345462024212, + 0.012762840837240219, + 0.02430625446140766, + 0.0164942666888237, + 0.006554612424224615, + 0.009383095428347588, + 0.007453758269548416, + 0.013643953949213028, + 0.010715030133724213, + 0.008518315851688385, + 0.005810997448861599, + 0.00821694079786539, + 0.0054925610311329365, + 0.006254240404814482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.08904077112674713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_tensor([2923, 2923])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_tensor([2923, 2923])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..04e3e948830e8876300b50b11e7a0fc30309183f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_tensor([2923, 2923])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle tracking shot follows closely behind a four-horse bronze chariot, placing the emperor at the center of the frame as an unshakable focal point. Atop the chariot, a middle-aged emperor with a stern face and a piercing gaze stands steady. He wears a magnificent robe of dark black silk, densely embroidered with coiled dragons in shimmering gold thread. The dragon's scales gleam in the light, and his wide sleeves billow backward in the gusts of wind. He wears a black and gold imperial crown, its tassels adorned with jade beads that sway violently with the chariot's violent jolts. The bronze chariot beneath him, etched with intricate ancient cloud patterns, has a deep patina from the wear of time. Ahead, four muscular black horses, galloping abreast, paw the ground with a thunderous rhythm, their manes and tails dancing wildly in the wind. The emperor grips the leather reins firmly, maintaining a powerful balance at high speed as he steers his chariot across a vast, desolate wilderness. The massive spoked wheels roll over the parched earth, kicking up clouds of dust. The setting sun's golden afterglow paints the entire scene with a warm and majestic hue, lengthening the shadows of the chariot and horses. The dust in the air shimmers in the light, creating an atmosphere of epic grandeur, solemnity, and power.", + "seed": 21219, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_2923_seed_21219_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.042630831338465215, + "vlm_logprob": -0.042630831338465215, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.042630831338465215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0396118126809597 + ], + "opposite_logprob_tokens": [ + -3.41461181640625 + ], + "logprob_reward": -0.0396118126809597, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0396118126809597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2454017996788025, + 0.13247469067573547, + 0.046398065984249115, + 0.10416066646575928, + 0.062216974794864655, + 0.09017878025770187, + 0.10552069544792175, + 0.07935027778148651, + 0.06522829830646515, + 0.06906967610120773 + ], + "credit_map": [ + 0.0587196983397007, + 0.02592228725552559, + 0.015253502875566483, + 0.024759477004408836, + 0.03251737728714943, + 0.059135667979717255, + 0.014599161222577095, + 0.014494639821350574, + 0.020746055990457535, + 0.029323121532797813, + 0.008982625789940357, + 0.011472102254629135, + 0.015191159211099148, + 0.02672584541141987, + 0.007126932498067617, + 0.012906849384307861, + 0.007537006866186857, + 0.008918050676584244, + 0.004487613681703806, + 0.006103762425482273, + 0.004814651794731617, + 0.008132338523864746, + 0.0031932583078742027, + 0.0032113869674503803, + 0.01230558194220066, + 0.011169987730681896, + 0.010811546817421913, + 0.01011513452976942, + 0.012286882847547531, + 0.014663969166576862, + 0.010798887349665165, + 0.022008681669831276, + 0.008214439265429974, + 0.006786357145756483, + 0.0080791637301445, + 0.009696762077510357, + 0.004526980221271515, + 0.009941398166120052, + 0.005122577305883169, + 0.009849300608038902, + 0.009688064455986023, + 0.010286808013916016, + 0.009312032721936703, + 0.0133701516315341, + 0.015664370730519295, + 0.011535094119608402, + 0.009154479019343853, + 0.011167782358825207, + 0.011423958465456963, + 0.013559569604694843, + 0.010904870927333832, + 0.011109434068202972, + 0.03181660920381546, + 0.010072566568851471, + 0.010204637423157692, + 0.006429044064134359, + 0.010750830173492432, + 0.010284189134836197, + 0.012750435620546341, + 0.009213490411639214, + 0.013418858870863914, + 0.006465297657996416, + 0.008687674067914486, + 0.007779499050229788, + 0.009177963249385357, + 0.008876093663275242, + 0.013372400775551796, + 0.007825469598174095, + 0.011542711406946182, + 0.005342300049960613, + 0.00425943685695529, + 0.004831923171877861, + 0.01018273364752531, + 0.006018761079758406, + 0.017884084954857826, + 0.007846031337976456, + 0.009392944164574146, + 0.007830111309885979, + 0.005836369935423136, + 0.0040786415338516235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0396118126809597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899362593889236 + ], + "opposite_logprob_tokens": [ + -2.578993558883667 + ], + "logprob_reward": -0.07899362593889236, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899362593889236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27692529559135437, + 0.0923973098397255, + 0.09972604364156723, + 0.14974211156368256, + 0.028714321553707123, + 0.07215011864900589, + 0.07313293218612671, + 0.05304174870252609, + 0.07541130483150482, + 0.0787588432431221 + ], + "credit_map": [ + 0.07070112973451614, + 0.030783487483859062, + 0.014988156035542488, + 0.018948128446936607, + 0.024998517706990242, + 0.07926426827907562, + 0.015499320812523365, + 0.02174229919910431, + 0.015195897780358791, + 0.021528197452425957, + 0.007058376912027597, + 0.012517726048827171, + 0.013154110871255398, + 0.010529866442084312, + 0.006656927522271872, + 0.00575620261952281, + 0.014730398543179035, + 0.012919812463223934, + 0.010005012154579163, + 0.013480527326464653, + 0.012247970327734947, + 0.02060585282742977, + 0.00831254106014967, + 0.007423923816531897, + 0.021726658567786217, + 0.015424322336912155, + 0.016529008746147156, + 0.015990568324923515, + 0.038032110780477524, + 0.01701541431248188, + 0.009046168066561222, + 0.01597786135971546, + 0.0041212779469788074, + 0.0033894323278218508, + 0.005120708607137203, + 0.00340936123393476, + 0.0030945532489567995, + 0.0030399030074477196, + 0.0028839786536991596, + 0.0036551051307469606, + 0.010045529343187809, + 0.0069600786082446575, + 0.008691461756825447, + 0.011969187296926975, + 0.01248324103653431, + 0.007811743300408125, + 0.007366097532212734, + 0.006822782568633556, + 0.008372281678020954, + 0.008327611722052097, + 0.01144730020314455, + 0.00948330108076334, + 0.018510200083255768, + 0.005634841043502092, + 0.007373067084699869, + 0.0039843302220106125, + 0.007009412627667189, + 0.006626410875469446, + 0.0076831975020468235, + 0.01066885981708765, + 0.007426938973367214, + 0.00426105223596096, + 0.0047571370378136635, + 0.0046087391674518585, + 0.009000852704048157, + 0.007503426633775234, + 0.008292690850794315, + 0.011601806618273258, + 0.010981942526996136, + 0.011168748140335083, + 0.009921308606863022, + 0.00694052642211318, + 0.011391512118279934, + 0.007224440574645996, + 0.017380094155669212, + 0.011225426569581032, + 0.009396705776453018, + 0.007794877514243126, + 0.008033346384763718, + 0.006312441546469927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07899362593889236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815446585416794 + ], + "opposite_logprob_tokens": [ + -3.529815435409546 + ], + "logprob_reward": -0.029815446585416794, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815446585416794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40913012623786926, + 0.09142632782459259, + 0.05069795995950699, + 0.07741134613752365, + 0.053387172520160675, + 0.05403829365968704, + 0.06106651574373245, + 0.09805302321910858, + 0.033559635281562805, + 0.07122959196567535 + ], + "credit_map": [ + 0.11042356491088867, + 0.04326436296105385, + 0.021433837711811066, + 0.023587666451931, + 0.038805264979600906, + 0.12031952291727066, + 0.02185235731303692, + 0.02944357879459858, + 0.0116519033908844, + 0.01354123093187809, + 0.0077142189256846905, + 0.008384701795876026, + 0.021828269585967064, + 0.011149056255817413, + 0.009816205129027367, + 0.0073407418094575405, + 0.007408598903566599, + 0.005761583801358938, + 0.0053044273518025875, + 0.007765294052660465, + 0.008526305668056011, + 0.005544185638427734, + 0.005996732972562313, + 0.004390828777104616, + 0.01053035818040371, + 0.008925904519855976, + 0.009100420400500298, + 0.00967344082891941, + 0.015918441116809845, + 0.007856989279389381, + 0.007136794272810221, + 0.008268993347883224, + 0.007064135279506445, + 0.007409743964672089, + 0.006152234505861998, + 0.006309780292212963, + 0.008965952321887016, + 0.005184129346162081, + 0.005646551959216595, + 0.006654645781964064, + 0.006446792744100094, + 0.0048253657296299934, + 0.005659535061568022, + 0.008337097242474556, + 0.013210946694016457, + 0.006353139411658049, + 0.004579344764351845, + 0.004626070149242878, + 0.007894708774983883, + 0.005276318173855543, + 0.011401581577956676, + 0.009482401423156261, + 0.008421305567026138, + 0.006856245920062065, + 0.005836703348904848, + 0.005897250026464462, + 0.012139247730374336, + 0.01193076092749834, + 0.01813500188291073, + 0.017799237743020058, + 0.011909914202988148, + 0.008171851746737957, + 0.009075705893337727, + 0.008891310542821884, + 0.005066815763711929, + 0.003706788644194603, + 0.003979469649493694, + 0.0051994649693369865, + 0.003729661926627159, + 0.005466612987220287, + 0.002829537494108081, + 0.003581282449886203, + 0.011320168152451515, + 0.0071080634370446205, + 0.016620159149169922, + 0.01139130536466837, + 0.008408065885305405, + 0.00646517239511013, + 0.005516143515706062, + 0.0044005149975419044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.029815446585416794 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495587736368179 + ], + "opposite_logprob_tokens": [ + -2.9299559593200684 + ], + "logprob_reward": -0.05495587736368179, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495587736368179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40807515382766724, + 0.08135458081960678, + 0.04707849770784378, + 0.09880237281322479, + 0.11085209995508194, + 0.04226512089371681, + 0.047398779541254044, + 0.07750584930181503, + 0.028573352843523026, + 0.058094169944524765 + ], + "credit_map": [ + 0.10709313303232193, + 0.04086574912071228, + 0.022823594510555267, + 0.03310411050915718, + 0.039854828268289566, + 0.11707567423582077, + 0.02162911929190159, + 0.025628939270973206, + 0.011976255103945732, + 0.010082757100462914, + 0.008304839953780174, + 0.0127943130210042, + 0.013814493082463741, + 0.010096004232764244, + 0.0067857238464057446, + 0.007500196807086468, + 0.0065373629331588745, + 0.005030651111155748, + 0.005155441351234913, + 0.0074262372218072414, + 0.008689605630934238, + 0.004610224161297083, + 0.005075463559478521, + 0.004553514998406172, + 0.014535244554281235, + 0.013325883075594902, + 0.009170707315206528, + 0.01306156162172556, + 0.016733389347791672, + 0.013568535447120667, + 0.007989564910531044, + 0.010417490266263485, + 0.01630385033786297, + 0.012309107929468155, + 0.016845442354679108, + 0.015634801238775253, + 0.008998782373964787, + 0.01944420486688614, + 0.00924889836460352, + 0.01206701248884201, + 0.005355055909603834, + 0.005488550756126642, + 0.00589514197781682, + 0.007678251713514328, + 0.004760257434099913, + 0.004739662632346153, + 0.0032838264014571905, + 0.005064375698566437, + 0.005578687880188227, + 0.0061184195801615715, + 0.006402180530130863, + 0.005911772605031729, + 0.009351469576358795, + 0.005210103001445532, + 0.00360085885040462, + 0.005225288216024637, + 0.008748291060328484, + 0.00881329458206892, + 0.01503043994307518, + 0.010390771552920341, + 0.013917486183345318, + 0.0058579654432833195, + 0.006500466261059046, + 0.0082471314817667, + 0.003947534132748842, + 0.003560202196240425, + 0.004548322409391403, + 0.005578142590820789, + 0.0032871728762984276, + 0.0024037680123001337, + 0.0027788663282990456, + 0.0024693459272384644, + 0.009014916606247425, + 0.006291497964411974, + 0.01026626955717802, + 0.009061090648174286, + 0.007405060343444347, + 0.006993215065449476, + 0.004679640755057335, + 0.004382486455142498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.05495587736368179 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777394123375416 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777394123375416, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777394123375416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20219901204109192, + 0.11286772042512894, + 0.02248249761760235, + 0.03908393904566765, + 0.0383131206035614, + 0.03526576608419418, + 0.08655539900064468, + 0.0495990589261055, + 0.19501270353794098, + 0.21862076222896576 + ], + "credit_map": [ + 0.04897935315966606, + 0.021184654906392097, + 0.01312249805778265, + 0.015330163761973381, + 0.021159421652555466, + 0.051465000957250595, + 0.014107557944953442, + 0.016850344836711884, + 0.014644703827798367, + 0.019743699580430984, + 0.007753919810056686, + 0.011018128134310246, + 0.021976683288812637, + 0.018435727804899216, + 0.009074216708540916, + 0.010220642201602459, + 0.003457640763372183, + 0.0029872902669012547, + 0.0022801326122134924, + 0.0037638237699866295, + 0.003217557677999139, + 0.003368847072124481, + 0.001467702561058104, + 0.0019395040581002831, + 0.00518422108143568, + 0.005045262631028891, + 0.004180566407740116, + 0.004245723597705364, + 0.00543171726167202, + 0.004238841589540243, + 0.00409338902682066, + 0.006664218381047249, + 0.00467199319973588, + 0.0048780497163534164, + 0.0046032085083425045, + 0.005276328884065151, + 0.005620126146823168, + 0.005084932316094637, + 0.0034134690649807453, + 0.00476501090452075, + 0.0034294596407562494, + 0.0046608769334852695, + 0.0034941858612000942, + 0.006058019585907459, + 0.004338731989264488, + 0.004039227496832609, + 0.003646549303084612, + 0.005598715506494045, + 0.01103856973350048, + 0.008608603850007057, + 0.016578059643507004, + 0.01570623181760311, + 0.014149514958262444, + 0.007712106220424175, + 0.008092719130218029, + 0.0046696048229932785, + 0.006354826036840677, + 0.007062102202326059, + 0.011177917011082172, + 0.007110427133738995, + 0.006604686379432678, + 0.0029021757654845715, + 0.003686001058667898, + 0.004700921475887299, + 0.0273920726031065, + 0.02723914198577404, + 0.038626737892627716, + 0.02630392089486122, + 0.033057428896427155, + 0.016228266060352325, + 0.012533926405012608, + 0.013631218113005161, + 0.037738144397735596, + 0.019477561116218567, + 0.06815330684185028, + 0.02907555364072323, + 0.018037861213088036, + 0.018339717760682106, + 0.017100563272833824, + 0.010698029771447182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.009777394123375416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot follows closely behind a four-horse bronze chariot, placing the emperor at the center of the frame as an unshakable focal point. Atop the chariot, a middle-aged emperor with a stern face and a piercing gaze stands steady. He wears a magnificent robe of dark black silk, densely embroidered with coiled dragons in shimmering gold thread. The dragon's scales gleam in the light, and his wide sleeves billow backward in the gusts of wind. He wears a black and gold imperial crown, its tassels adorned with jade beads that sway violently with the chariot's violent jolts. The bronze chariot beneath him, etched with intricate ancient cloud patterns, has a deep patina from the wear of time. Ahead, four muscular black horses, galloping abreast, paw the ground with a thunderous rhythm, their manes and tails dancing wildly in the wind. The emperor grips the leather reins firmly, maintaining a powerful balance at high speed as he steers his chariot across a vast, desolate wilderness. The massive spoked wheels roll over the parched earth, kicking up clouds of dust. The setting sun's golden afterglow paints the entire scene with a warm and majestic hue, lengthening the shadows of the chariot and horses. The dust in the air shimmers in the light, creating an atmosphere of epic grandeur, solemnity, and power.", + "seed": 21220, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_2923_seed_21220_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08441250324249268, + "vlm_logprob": -0.08441250324249268, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.08441250324249268, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054557714611291885 + ], + "opposite_logprob_tokens": [ + -3.0545578002929688 + ], + "logprob_reward": -0.054557714611291885, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.054557714611291885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054488830268383026, + 0.6744024157524109, + 0.014471342787146568, + 0.015206568874418736, + 0.05308143422007561, + 0.007140638306736946, + 0.0836121141910553, + 0.02716689370572567, + 0.022852374240756035, + 0.047577351331710815 + ], + "credit_map": [ + 0.01236971840262413, + 0.006184048485010862, + 0.010518409311771393, + 0.005940182134509087, + 0.005392228253185749, + 0.00510067492723465, + 0.004821701906621456, + 0.004161868244409561, + 0.26818186044692993, + 0.03903401270508766, + 0.2337665855884552, + 0.02837253548204899, + 0.010447688400745392, + 0.014097332023084164, + 0.055327434092760086, + 0.02517501637339592, + 0.001439484884031117, + 0.0017529336037114263, + 0.00193311576731503, + 0.002017049351707101, + 0.0017387917032465339, + 0.00146078213583678, + 0.002550108591094613, + 0.0015790779143571854, + 0.0018296490889042616, + 0.0015117361908778548, + 0.0017167944461107254, + 0.002065044129267335, + 0.0015678303316235542, + 0.0016672530910000205, + 0.003299963427707553, + 0.0015482993330806494, + 0.00599302863702178, + 0.004805636592209339, + 0.008178493939340115, + 0.004690984264016151, + 0.006133308634161949, + 0.009603890590369701, + 0.006803880445659161, + 0.006872213911265135, + 0.001102354726754129, + 0.000910949136596173, + 0.0010928750270977616, + 0.0009447348420508206, + 0.0008658075821585953, + 0.0006162729696370661, + 0.0008081233827397227, + 0.000799521105363965, + 0.026387492194771767, + 0.004591413773596287, + 0.005768073722720146, + 0.004128608852624893, + 0.029512258246541023, + 0.004623029846698046, + 0.0028011116664856672, + 0.005800133571028709, + 0.004454785026609898, + 0.002331365132704377, + 0.002881145104765892, + 0.0028458323795348406, + 0.004801504779607058, + 0.002809541067108512, + 0.0033576078712940216, + 0.0036851156037300825, + 0.0023275716230273247, + 0.003104502335190773, + 0.0041894689202308655, + 0.0027303926181048155, + 0.0020889879669994116, + 0.0026776508893817663, + 0.003109769895672798, + 0.00262402999214828, + 0.004576924256980419, + 0.00740758515894413, + 0.009568492881953716, + 0.006888162344694138, + 0.005396425724029541, + 0.004974977113306522, + 0.004361105151474476, + 0.004403684753924608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.054557714611291885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813361883163452 + ], + "opposite_logprob_tokens": [ + -1.4063361883163452 + ], + "logprob_reward": -0.2813361883163452, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813361883163452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17080402374267578, + 0.08597175031900406, + 0.06181420758366585, + 0.16786862909793854, + 0.09421897679567337, + 0.07377449423074722, + 0.08281818777322769, + 0.04194827005267143, + 0.11209449172019958, + 0.10868699103593826 + ], + "credit_map": [ + 0.03873313218355179, + 0.0169671718031168, + 0.03746139630675316, + 0.014731067232787609, + 0.018042970448732376, + 0.015708554536104202, + 0.01491516176611185, + 0.01424454990774393, + 0.014874120242893696, + 0.008771372027695179, + 0.021415244787931442, + 0.00666778115555644, + 0.010070154443383217, + 0.008925938047468662, + 0.007919080555438995, + 0.00732805673032999, + 0.006983539555221796, + 0.008653083816170692, + 0.007083434145897627, + 0.004600341897457838, + 0.009960958734154701, + 0.00851292535662651, + 0.008852669037878513, + 0.007167251314967871, + 0.03259396553039551, + 0.01699572615325451, + 0.03443948179483414, + 0.026274917647242546, + 0.01614622212946415, + 0.009534347802400589, + 0.020237796008586884, + 0.01164616085588932, + 0.010788253508508205, + 0.010643357411026955, + 0.011274861171841621, + 0.009009999223053455, + 0.011780970729887486, + 0.01624787598848343, + 0.012445583008229733, + 0.012028063647449017, + 0.01268785260617733, + 0.006312887649983168, + 0.021185819059610367, + 0.008915801532566547, + 0.007215841207653284, + 0.006031760014593601, + 0.005928538274019957, + 0.0054959882982075214, + 0.019330443814396858, + 0.007745292503386736, + 0.007253915071487427, + 0.006942192558199167, + 0.02391640655696392, + 0.007185021415352821, + 0.0042360033839941025, + 0.006208908278495073, + 0.005202149040997028, + 0.006458757445216179, + 0.00453840009868145, + 0.005149284843355417, + 0.004983503837138414, + 0.006746315397322178, + 0.00396902346983552, + 0.004900835454463959, + 0.010915274731814861, + 0.01174579095095396, + 0.016727060079574585, + 0.01374199241399765, + 0.01067622471600771, + 0.014147976413369179, + 0.019534559920430183, + 0.014605614356696606, + 0.015170891769230366, + 0.009872321970760822, + 0.009383146651089191, + 0.011216594837605953, + 0.01710256189107895, + 0.02155906893312931, + 0.013032786548137665, + 0.011349606327712536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.2813361883163452 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331462688744068 + ], + "opposite_logprob_tokens": [ + -3.7733147144317627 + ], + "logprob_reward": -0.02331462688744068, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331462688744068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1009782999753952, + 0.20360472798347473, + 0.07711999118328094, + 0.17004923522472382, + 0.15872134268283844, + 0.055255692452192307, + 0.04672794044017792, + 0.05431294068694115, + 0.07363440841436386, + 0.05959545820951462 + ], + "credit_map": [ + 0.014008906669914722, + 0.014324255287647247, + 0.015371516346931458, + 0.011204256676137447, + 0.014121020212769508, + 0.011234009638428688, + 0.008882200345396996, + 0.011832131072878838, + 0.06328806281089783, + 0.011185036040842533, + 0.05884494632482529, + 0.008285598829388618, + 0.008546603843569756, + 0.010080973617732525, + 0.015897680073976517, + 0.02747582457959652, + 0.010439127683639526, + 0.008684077300131321, + 0.015120710246264935, + 0.008566383272409439, + 0.01091238297522068, + 0.008656482212245464, + 0.00746161350980401, + 0.007279215846210718, + 0.047607023268938065, + 0.013725025579333305, + 0.05054321885108948, + 0.01907253824174404, + 0.009184868074953556, + 0.008768465369939804, + 0.010706963017582893, + 0.010441124439239502, + 0.025920843705534935, + 0.010472534224390984, + 0.042378321290016174, + 0.014134847559034824, + 0.011933313682675362, + 0.02081954851746559, + 0.018173184245824814, + 0.014888753183186054, + 0.0076443045400083065, + 0.006171075627207756, + 0.011632840149104595, + 0.005405417177826166, + 0.007168913725763559, + 0.005867711268365383, + 0.006419671233743429, + 0.004945756867527962, + 0.011114866472780704, + 0.0037633178289979696, + 0.003581753931939602, + 0.002946121385321021, + 0.00989079475402832, + 0.008494449779391289, + 0.0031327044125646353, + 0.003803928615525365, + 0.00826637540012598, + 0.007551476359367371, + 0.006733091548085213, + 0.006174495909363031, + 0.010120353661477566, + 0.006601494271308184, + 0.0038496116176247597, + 0.005016039125621319, + 0.009414345026016235, + 0.00863677728921175, + 0.012925415299832821, + 0.007176535204052925, + 0.0074898479506373405, + 0.009268008172512054, + 0.009444215334951878, + 0.009279264137148857, + 0.006279672496020794, + 0.010682260617613792, + 0.00804077461361885, + 0.008095274679362774, + 0.006603350397199392, + 0.007511452306061983, + 0.00620111171156168, + 0.006181561853736639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.02331462688744068 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866068437695503 + ], + "opposite_logprob_tokens": [ + -3.0486607551574707 + ], + "logprob_reward": -0.04866068437695503, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866068437695503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04022606462240219, + 0.3092445433139801, + 0.0403551310300827, + 0.09352231025695801, + 0.07970824092626572, + 0.05867759510874748, + 0.21443137526512146, + 0.04835094138979912, + 0.08905889838933945, + 0.02642490155994892 + ], + "credit_map": [ + 0.007980022579431534, + 0.005738692358136177, + 0.004419855307787657, + 0.004002777393907309, + 0.0052648792043328285, + 0.004975473042577505, + 0.0034724243450909853, + 0.004371941089630127, + 0.12027020007371902, + 0.01927832141518593, + 0.10211572796106339, + 0.013660801574587822, + 0.009579057805240154, + 0.009217613376677036, + 0.023796070367097855, + 0.011326773092150688, + 0.005738719832152128, + 0.004970163106918335, + 0.005388691555708647, + 0.00450156582519412, + 0.003894484369084239, + 0.00634142616763711, + 0.005311234388500452, + 0.004208848811686039, + 0.019409483298659325, + 0.009630581364035606, + 0.02148415893316269, + 0.01422905083745718, + 0.006809248588979244, + 0.005748237948864698, + 0.010069738142192364, + 0.006141815334558487, + 0.007989679463207722, + 0.0069260322488844395, + 0.01569536328315735, + 0.00849449448287487, + 0.009870801120996475, + 0.013878374360501766, + 0.009142107330262661, + 0.007711397483944893, + 0.008716349489986897, + 0.005074701737612486, + 0.01444359589368105, + 0.007051017601042986, + 0.00669737346470356, + 0.005307628307491541, + 0.006070446223020554, + 0.005316485650837421, + 0.06648816913366318, + 0.011671937070786953, + 0.017905212938785553, + 0.008794822730123997, + 0.07538798451423645, + 0.012853149324655533, + 0.007336624898016453, + 0.013993517495691776, + 0.007521921303123236, + 0.006209634244441986, + 0.005420504603534937, + 0.005339443683624268, + 0.007297103758901358, + 0.005659311544150114, + 0.006161846220493317, + 0.004741177428513765, + 0.012574720196425915, + 0.010028176940977573, + 0.011942021548748016, + 0.007802707143127918, + 0.011659864336252213, + 0.012998671270906925, + 0.01213997881859541, + 0.009912765584886074, + 0.0027585721109062433, + 0.0034591369330883026, + 0.0034624668769538403, + 0.003880367847159505, + 0.003680072957649827, + 0.004659990780055523, + 0.0018130810931324959, + 0.0027112148236483335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.04866068437695503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193302020430565 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193302020430565, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193302020430565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14959357678890228, + 0.48555049300193787, + 0.025900796055793762, + 0.04440014436841011, + 0.10597001016139984, + 0.026516150683164597, + 0.04760776460170746, + 0.049702972173690796, + 0.04297507181763649, + 0.021782994270324707 + ], + "credit_map": [ + 0.024852972477674484, + 0.015066803433001041, + 0.02148471772670746, + 0.01693129912018776, + 0.020601259544491768, + 0.023394299671053886, + 0.015034232288599014, + 0.012227999046444893, + 0.18970654904842377, + 0.027325492352247238, + 0.16935297846794128, + 0.02027313783764839, + 0.008297495543956757, + 0.010188957676291466, + 0.04129564017057419, + 0.019110236316919327, + 0.0031817362178117037, + 0.0019547403790056705, + 0.0038044366519898176, + 0.001992820529267192, + 0.0032958376687020063, + 0.0040893773548305035, + 0.0040603275410830975, + 0.0035215201787650585, + 0.010612376034259796, + 0.00391421839594841, + 0.01124770287424326, + 0.00639204028993845, + 0.0038595465011894703, + 0.002049861242994666, + 0.0033923580776900053, + 0.0029320416506379843, + 0.016527673229575157, + 0.0065435972064733505, + 0.03431331366300583, + 0.007107933051884174, + 0.009155257605016232, + 0.014769654721021652, + 0.007141226436942816, + 0.010411345399916172, + 0.004605487454682589, + 0.0022878802847117186, + 0.006568897049874067, + 0.003037101589143276, + 0.0026700289454311132, + 0.0024713831953704357, + 0.002870800904929638, + 0.002004570560529828, + 0.014584554359316826, + 0.0027494882233440876, + 0.003709951648488641, + 0.0021266667172312737, + 0.015705445781350136, + 0.003152943914756179, + 0.0028332951478660107, + 0.0027454192750155926, + 0.008167516440153122, + 0.005481707863509655, + 0.005726790055632591, + 0.004364669322967529, + 0.006511075887829065, + 0.0070512318052351475, + 0.005375119857490063, + 0.007024860475212336, + 0.004575340077280998, + 0.004756094887852669, + 0.009015687741339207, + 0.0052438657730817795, + 0.0038795783184468746, + 0.004285398870706558, + 0.006275249179452658, + 0.004943856969475746, + 0.00274762324988842, + 0.003354335203766823, + 0.002721319207921624, + 0.003301449818536639, + 0.0023013262543827295, + 0.003220395417883992, + 0.002066465327516198, + 0.002070079790428281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.014193302020430565 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot follows closely behind a four-horse bronze chariot, placing the emperor at the center of the frame as an unshakable focal point. Atop the chariot, a middle-aged emperor with a stern face and a piercing gaze stands steady. He wears a magnificent robe of dark black silk, densely embroidered with coiled dragons in shimmering gold thread. The dragon's scales gleam in the light, and his wide sleeves billow backward in the gusts of wind. He wears a black and gold imperial crown, its tassels adorned with jade beads that sway violently with the chariot's violent jolts. The bronze chariot beneath him, etched with intricate ancient cloud patterns, has a deep patina from the wear of time. Ahead, four muscular black horses, galloping abreast, paw the ground with a thunderous rhythm, their manes and tails dancing wildly in the wind. The emperor grips the leather reins firmly, maintaining a powerful balance at high speed as he steers his chariot across a vast, desolate wilderness. The massive spoked wheels roll over the parched earth, kicking up clouds of dust. The setting sun's golden afterglow paints the entire scene with a warm and majestic hue, lengthening the shadows of the chariot and horses. The dust in the air shimmers in the light, creating an atmosphere of epic grandeur, solemnity, and power.", + "seed": 21221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_2923_seed_21221_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06727697933092713, + "vlm_logprob": -0.06727697933092713, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.06727697933092713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040396831929683685 + ], + "opposite_logprob_tokens": [ + -3.4153969287872314 + ], + "logprob_reward": -0.040396831929683685, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.040396831929683685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1990860402584076, + 0.11165077239274979, + 0.07394364476203918, + 0.12751726806163788, + 0.040938954800367355, + 0.06275901198387146, + 0.07286813110113144, + 0.15253639221191406, + 0.04270068183541298, + 0.11599909514188766 + ], + "credit_map": [ + 0.025400064885616302, + 0.02224554866552353, + 0.030659833922982216, + 0.031339894980192184, + 0.018081463873386383, + 0.0177933219820261, + 0.0232213344424963, + 0.030344586819410324, + 0.014907999895513058, + 0.013848250731825829, + 0.024634715169668198, + 0.01831931434571743, + 0.011979788541793823, + 0.009880649857223034, + 0.009869721718132496, + 0.008210327476263046, + 0.021868687123060226, + 0.022019026800990105, + 0.005970136262476444, + 0.0057776086032390594, + 0.0050492254085838795, + 0.005490983836352825, + 0.0038767310325056314, + 0.0038912470918148756, + 0.019537797197699547, + 0.017243390902876854, + 0.017790788784623146, + 0.020836330950260162, + 0.012934341095387936, + 0.007521240971982479, + 0.01900586485862732, + 0.012647513300180435, + 0.005004202481359243, + 0.0041818744502961636, + 0.009775634855031967, + 0.005683424416929483, + 0.005047025624662638, + 0.003021647920832038, + 0.004214892629534006, + 0.004010254517197609, + 0.007750311400741339, + 0.0067008123733103275, + 0.010310614481568336, + 0.008205348625779152, + 0.006552836857736111, + 0.006653535179793835, + 0.006970015820115805, + 0.009615540504455566, + 0.0077497814781963825, + 0.007146888412535191, + 0.008853751234710217, + 0.009944085031747818, + 0.013726437464356422, + 0.006473956163972616, + 0.009474409744143486, + 0.009498818777501583, + 0.01795964688062668, + 0.01409684307873249, + 0.03131052851676941, + 0.021966293454170227, + 0.017133202403783798, + 0.02490312047302723, + 0.012309334240853786, + 0.012857416644692421, + 0.004927434492856264, + 0.004012892488390207, + 0.008694512769579887, + 0.005659584421664476, + 0.00566074950620532, + 0.0034456243738532066, + 0.005393587052822113, + 0.004906292539089918, + 0.011477543041110039, + 0.012516210786998272, + 0.021551886573433876, + 0.017372073605656624, + 0.014276816509664059, + 0.011420091614127159, + 0.0165119506418705, + 0.010872522369027138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.040396831929683685 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558540105819702 + ], + "opposite_logprob_tokens": [ + -1.6005854606628418 + ], + "logprob_reward": -0.22558540105819702, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558540105819702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16069267690181732, + 0.08071515709161758, + 0.10054609179496765, + 0.05291745811700821, + 0.0741700604557991, + 0.05702042952179909, + 0.054889701306819916, + 0.17927397787570953, + 0.0830121636390686, + 0.1567622721195221 + ], + "credit_map": [ + 0.018734710291028023, + 0.017461150884628296, + 0.027693644165992737, + 0.03196116164326668, + 0.01838243566453457, + 0.01560438983142376, + 0.013392201624810696, + 0.017462994903326035, + 0.010146095417439938, + 0.010368323884904385, + 0.016917239874601364, + 0.010442092083394527, + 0.009428530931472778, + 0.007234236225485802, + 0.009242148138582706, + 0.006936484482139349, + 0.01703418605029583, + 0.016352644190192223, + 0.01845662109553814, + 0.013366582803428173, + 0.01226275134831667, + 0.00697041442617774, + 0.010002322494983673, + 0.0061005717143416405, + 0.006382488179951906, + 0.007165675051510334, + 0.006642452906817198, + 0.0073197828605771065, + 0.005178610794246197, + 0.0031008096411824226, + 0.011274851858615875, + 0.005852784030139446, + 0.00874248705804348, + 0.006424362771213055, + 0.016422666609287262, + 0.008444219827651978, + 0.010182118974626064, + 0.0068641044199466705, + 0.009381348267197609, + 0.007708745542913675, + 0.006998393218964338, + 0.005455212201923132, + 0.011744421906769276, + 0.00695958873257041, + 0.007630906533449888, + 0.004405099432915449, + 0.007745429873466492, + 0.006081373896449804, + 0.0053131249733269215, + 0.006783731281757355, + 0.005105632357299328, + 0.0100611736997962, + 0.006000002380460501, + 0.007761900778859854, + 0.0049064550548791885, + 0.008957679383456707, + 0.018295487388968468, + 0.012897150591015816, + 0.038607824593782425, + 0.03018934465944767, + 0.023607535287737846, + 0.015691610053181648, + 0.01825951598584652, + 0.021725520491600037, + 0.009622198529541492, + 0.0073220389895141125, + 0.011397851631045341, + 0.01105059590190649, + 0.013761340640485287, + 0.007178931031376123, + 0.01419015508145094, + 0.008489048108458519, + 0.015593096613883972, + 0.01316731609404087, + 0.0152429835870862, + 0.025031697005033493, + 0.017407797276973724, + 0.014032726176083088, + 0.04086964204907417, + 0.015417033806443214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.22558540105819702 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819726943969727 + ], + "opposite_logprob_tokens": [ + -3.5298197269439697 + ], + "logprob_reward": -0.029819726943969727, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819726943969727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15242646634578705, + 0.04878121614456177, + 0.2582511007785797, + 0.05679157003760338, + 0.05237102881073952, + 0.048202019184827805, + 0.10087256878614426, + 0.12453846633434296, + 0.034976471215486526, + 0.12278914451599121 + ], + "credit_map": [ + 0.017525389790534973, + 0.01883220300078392, + 0.035875435918569565, + 0.020389564335346222, + 0.020843209698796272, + 0.012676166370511055, + 0.012284474447369576, + 0.014000020921230316, + 0.00657601049169898, + 0.007225163746625185, + 0.009747964330017567, + 0.0047186496667563915, + 0.00805000588297844, + 0.0037754420191049576, + 0.005150500684976578, + 0.0035374818835407495, + 0.09492345154285431, + 0.08696293085813522, + 0.020071718841791153, + 0.013561232946813107, + 0.015579045750200748, + 0.01058252900838852, + 0.008933812379837036, + 0.007636373396962881, + 0.009455407969653606, + 0.00785911176353693, + 0.011621229350566864, + 0.006738538853824139, + 0.009055204689502716, + 0.0030825138092041016, + 0.0053815883584320545, + 0.003597973147407174, + 0.008498402312397957, + 0.005794392433017492, + 0.014650515280663967, + 0.005379731301218271, + 0.00638313964009285, + 0.0028643179684877396, + 0.005481581203639507, + 0.003318945411592722, + 0.004762404132634401, + 0.0043439678847789764, + 0.006862055044621229, + 0.006192915141582489, + 0.008453917689621449, + 0.005706342402845621, + 0.006017318926751614, + 0.005863097961992025, + 0.010850568301975727, + 0.008234168402850628, + 0.008379227481782436, + 0.01242881827056408, + 0.023083951324224472, + 0.016964688897132874, + 0.008714917115867138, + 0.012216233648359776, + 0.017009634524583817, + 0.01854582317173481, + 0.013233901001513004, + 0.017780937254428864, + 0.01887388527393341, + 0.01194927841424942, + 0.01098914910107851, + 0.016155853867530823, + 0.004582610912621021, + 0.003774716518819332, + 0.004102005157619715, + 0.003970788326114416, + 0.006449041422456503, + 0.003456762060523033, + 0.004654735792428255, + 0.003985808230936527, + 0.011956620961427689, + 0.012580052949488163, + 0.01835845597088337, + 0.01609467901289463, + 0.015525844879448414, + 0.013007859699428082, + 0.022052006796002388, + 0.01321362890303135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.029819726943969727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809741511940956 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809741511940956, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809741511940956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586158722639084, + 0.05760728195309639, + 0.3604526221752167, + 0.06987366825342178, + 0.045670632272958755, + 0.03097390942275524, + 0.05288519710302353, + 0.08066732436418533, + 0.07951537519693375, + 0.0637381449341774 + ], + "credit_map": [ + 0.016382532194256783, + 0.01844104379415512, + 0.026460306718945503, + 0.01721242070198059, + 0.02809782326221466, + 0.015203560702502728, + 0.015547387301921844, + 0.02127077803015709, + 0.007762428838759661, + 0.00785129889845848, + 0.013436170294880867, + 0.007462080102413893, + 0.006872427649796009, + 0.005400007590651512, + 0.005235706921666861, + 0.0035871563013643026, + 0.15834391117095947, + 0.14697721600532532, + 0.009418106637895107, + 0.010849911719560623, + 0.01208583451807499, + 0.00590256555005908, + 0.009199556894600391, + 0.007675500586628914, + 0.008913491852581501, + 0.006188687868416309, + 0.008793230168521404, + 0.007700152229517698, + 0.008004944771528244, + 0.0049620564095675945, + 0.01635519415140152, + 0.008955905213952065, + 0.0055327508598566055, + 0.004367183428257704, + 0.007079038769006729, + 0.005965670570731163, + 0.007083605043590069, + 0.0035777080338448286, + 0.006870281416922808, + 0.00519439484924078, + 0.004043379798531532, + 0.004941582214087248, + 0.005713649094104767, + 0.003672840539366007, + 0.004568673670291901, + 0.002161925658583641, + 0.003658872563391924, + 0.0022129840217530727, + 0.005997533444315195, + 0.005875340197235346, + 0.0051246266812086105, + 0.008078532293438911, + 0.010433568619191647, + 0.006487932056188583, + 0.004874231293797493, + 0.006013429258018732, + 0.010571463033556938, + 0.006129981018602848, + 0.012685658410191536, + 0.00899230595678091, + 0.011796804144978523, + 0.012990607880055904, + 0.008192501962184906, + 0.009308007545769215, + 0.009489038959145546, + 0.006987038999795914, + 0.01441721711307764, + 0.012786859646439552, + 0.009182854555547237, + 0.006224383134394884, + 0.01170068047940731, + 0.008727298118174076, + 0.006375907454639673, + 0.0049842712469398975, + 0.0053657907992601395, + 0.008784838020801544, + 0.00795179232954979, + 0.004479319788515568, + 0.018148474395275116, + 0.007647745776921511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.03809741511940956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024855216033756733 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024855216033756733, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024855216033756733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21522219479084015, + 0.06776025146245956, + 0.31149277091026306, + 0.05459318682551384, + 0.029966099187731743, + 0.03832503780722618, + 0.05309686064720154, + 0.08099997043609619, + 0.06911073625087738, + 0.07943292707204819 + ], + "credit_map": [ + 0.043771419674158096, + 0.03782633692026138, + 0.05667718127369881, + 0.021826811134815216, + 0.01657169871032238, + 0.011546632274985313, + 0.012799333781003952, + 0.014202771708369255, + 0.009814132004976273, + 0.01079725380986929, + 0.014232490211725235, + 0.008084827102720737, + 0.009212917648255825, + 0.0053999354131519794, + 0.005387535318732262, + 0.004831155762076378, + 0.13623285293579102, + 0.12810547649860382, + 0.009965358301997185, + 0.008508472703397274, + 0.011524553410708904, + 0.004277931526303291, + 0.006393261253833771, + 0.006484828423708677, + 0.008227366954088211, + 0.0073739769868552685, + 0.010129095055162907, + 0.009308990091085434, + 0.007516846060752869, + 0.0036255388986319304, + 0.005007777828723192, + 0.003403590526431799, + 0.005312788765877485, + 0.003250738838687539, + 0.009456919506192207, + 0.003443254856392741, + 0.0021736626513302326, + 0.0014247342478483915, + 0.0027443168219178915, + 0.0021596793085336685, + 0.005759083665907383, + 0.007565062493085861, + 0.00624984921887517, + 0.005956796929240227, + 0.004547185264527798, + 0.0024673035368323326, + 0.003489300375804305, + 0.002290450967848301, + 0.0062006632797420025, + 0.005364682991057634, + 0.00666420254856348, + 0.007999572902917862, + 0.008791360072791576, + 0.007809439208358526, + 0.0042654285207390785, + 0.0060015046037733555, + 0.009838305413722992, + 0.00693454360589385, + 0.013790787197649479, + 0.010146893560886383, + 0.013307377696037292, + 0.013527865521609783, + 0.0067619625478982925, + 0.00669222604483366, + 0.008116652257740498, + 0.006224225275218487, + 0.013377584517002106, + 0.011963602155447006, + 0.007453338708728552, + 0.003980384673923254, + 0.010812139138579369, + 0.0071828025393188, + 0.010705454275012016, + 0.010360839776694775, + 0.013796371407806873, + 0.012041156180202961, + 0.008535143919289112, + 0.007329552434384823, + 0.009594826959073544, + 0.007069569546729326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.0024855216033756733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot follows closely behind a four-horse bronze chariot, placing the emperor at the center of the frame as an unshakable focal point. Atop the chariot, a middle-aged emperor with a stern face and a piercing gaze stands steady. He wears a magnificent robe of dark black silk, densely embroidered with coiled dragons in shimmering gold thread. The dragon's scales gleam in the light, and his wide sleeves billow backward in the gusts of wind. He wears a black and gold imperial crown, its tassels adorned with jade beads that sway violently with the chariot's violent jolts. The bronze chariot beneath him, etched with intricate ancient cloud patterns, has a deep patina from the wear of time. Ahead, four muscular black horses, galloping abreast, paw the ground with a thunderous rhythm, their manes and tails dancing wildly in the wind. The emperor grips the leather reins firmly, maintaining a powerful balance at high speed as he steers his chariot across a vast, desolate wilderness. The massive spoked wheels roll over the parched earth, kicking up clouds of dust. The setting sun's golden afterglow paints the entire scene with a warm and majestic hue, lengthening the shadows of the chariot and horses. The dust in the air shimmers in the light, creating an atmosphere of epic grandeur, solemnity, and power.", + "seed": 21222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_3/index_2923_seed_21222_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07674180073663592, + "vlm_logprob": -0.07674180073663592, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.07674180073663592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05277838930487633 + ], + "opposite_logprob_tokens": [ + -3.177778482437134 + ], + "logprob_reward": -0.05277838930487633, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05277838930487633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14646662771701813, + 0.0858364850282669, + 0.13841041922569275, + 0.04974859207868576, + 0.0917387381196022, + 0.11921217292547226, + 0.09054912626743317, + 0.19050131738185883, + 0.03021322749555111, + 0.05732329562306404 + ], + "credit_map": [ + 0.02064746432006359, + 0.014942708425223827, + 0.02170952968299389, + 0.01654009521007538, + 0.02389463596045971, + 0.020616190508008003, + 0.01264032069593668, + 0.015475681982934475, + 0.014948933385312557, + 0.009746325202286243, + 0.017288843169808388, + 0.008240051567554474, + 0.008846428245306015, + 0.009070165455341339, + 0.006459178403019905, + 0.011236564256250858, + 0.022424109280109406, + 0.022627433761954308, + 0.021946704015135765, + 0.012026156298816204, + 0.01709064282476902, + 0.017125744372606277, + 0.01081747468560934, + 0.014352143742144108, + 0.007252184208482504, + 0.00613712053745985, + 0.009261160157620907, + 0.005089360289275646, + 0.005971788894385099, + 0.006554071791470051, + 0.0040160613134503365, + 0.005466845817863941, + 0.011568935588002205, + 0.008769810199737549, + 0.01599525660276413, + 0.012017086148262024, + 0.008112285286188126, + 0.008341343142092228, + 0.007801796775311232, + 0.019132224842905998, + 0.01735510863363743, + 0.01773461513221264, + 0.018117139115929604, + 0.013562369160354137, + 0.014368927106261253, + 0.0176547858864069, + 0.008303389884531498, + 0.012115835212171078, + 0.012147852219641209, + 0.007047484628856182, + 0.012993844226002693, + 0.010105597786605358, + 0.0069474149495363235, + 0.009750066325068474, + 0.012301825918257236, + 0.019255034625530243, + 0.02438242919743061, + 0.018725452944636345, + 0.023856105282902718, + 0.018546035513281822, + 0.05044911056756973, + 0.029954349622130394, + 0.011307967826724052, + 0.013279874809086323, + 0.0024289898574352264, + 0.0031674313358962536, + 0.004487100522965193, + 0.002296022605150938, + 0.0056784311309456825, + 0.006487398874014616, + 0.0020772493444383144, + 0.0035906019620597363, + 0.0073964339680969715, + 0.00643793074414134, + 0.009614191018044949, + 0.009809624403715134, + 0.0071275499649345875, + 0.006646380294114351, + 0.004789474420249462, + 0.005501709412783384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.05277838930487633 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521054446697235 + ], + "opposite_logprob_tokens": [ + -1.502105474472046 + ], + "logprob_reward": -0.2521054446697235, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521054446697235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19161419570446014, + 0.0827191174030304, + 0.09425882995128632, + 0.06426014006137848, + 0.12054812163114548, + 0.08597299456596375, + 0.05165370926260948, + 0.16398146748542786, + 0.07088913768529892, + 0.0741022601723671 + ], + "credit_map": [ + 0.028058823198080063, + 0.01459593791514635, + 0.04061010479927063, + 0.014561563730239868, + 0.018559109419584274, + 0.02700541727244854, + 0.019366588443517685, + 0.028856664896011353, + 0.015342660248279572, + 0.009293539449572563, + 0.018301015719771385, + 0.007985473610460758, + 0.0073005398735404015, + 0.0072169229388237, + 0.0059758699499070644, + 0.01130310446023941, + 0.014720801264047623, + 0.013465975411236286, + 0.011413054540753365, + 0.009141841903328896, + 0.012720593251287937, + 0.016325879842042923, + 0.008153574541211128, + 0.008317122235894203, + 0.010156058706343174, + 0.00589487561956048, + 0.01174018532037735, + 0.009828077629208565, + 0.004898818675428629, + 0.0070257182233035564, + 0.0057211993262171745, + 0.008995207026600838, + 0.016701946035027504, + 0.009007634595036507, + 0.028481028974056244, + 0.01449585985392332, + 0.00739393150433898, + 0.011958244256675243, + 0.011595395393669605, + 0.02091408707201481, + 0.011215081438422203, + 0.012955808080732822, + 0.01315315905958414, + 0.010257660411298275, + 0.012876372784376144, + 0.012236857786774635, + 0.005746265407651663, + 0.007531796582043171, + 0.0060762823559343815, + 0.005547867156565189, + 0.006199065130203962, + 0.006012555677443743, + 0.004957775119692087, + 0.009059303440153599, + 0.006094982381910086, + 0.007705881725996733, + 0.01981421932578087, + 0.015044959262013435, + 0.034724440425634384, + 0.01642928458750248, + 0.0249074287712574, + 0.029972445219755173, + 0.011876218020915985, + 0.011212489567697048, + 0.008319888263940811, + 0.006117578130215406, + 0.013678307645022869, + 0.006194536108523607, + 0.00927885714918375, + 0.0114380968734622, + 0.005048457067459822, + 0.010813423432409763, + 0.010267502628266811, + 0.01333488617092371, + 0.00688683008775115, + 0.010043477639555931, + 0.008208910003304482, + 0.011596151627600193, + 0.007005711551755667, + 0.006758794654160738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.2521054446697235 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821347445249557 + ], + "opposite_logprob_tokens": [ + -3.5298213958740234 + ], + "logprob_reward": -0.029821347445249557, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821347445249557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19687527418136597, + 0.0901469886302948, + 0.13863587379455566, + 0.03529006242752075, + 0.05808348208665848, + 0.10063862055540085, + 0.04289410635828972, + 0.11863023042678833, + 0.12362325936555862, + 0.0951821431517601 + ], + "credit_map": [ + 0.032909322530031204, + 0.0174518134444952, + 0.02884088270366192, + 0.02162635698914528, + 0.018634390085935593, + 0.03882797062397003, + 0.014943938702344894, + 0.023640604689717293, + 0.014035969972610474, + 0.008802738972008228, + 0.015496696345508099, + 0.011365401558578014, + 0.013558633625507355, + 0.00936814397573471, + 0.007895861752331257, + 0.009623536840081215, + 0.022001061588525772, + 0.01909741759300232, + 0.021945443004369736, + 0.012838992290198803, + 0.017961055040359497, + 0.01982608623802662, + 0.011280519887804985, + 0.013685286976397038, + 0.005129305180162191, + 0.0033511428628116846, + 0.005589492619037628, + 0.004675049800425768, + 0.004563992843031883, + 0.003995096776634455, + 0.002716532675549388, + 0.005269447807222605, + 0.007844132371246815, + 0.0071738846600055695, + 0.009338224306702614, + 0.007409350015223026, + 0.009700355120003223, + 0.004662399645894766, + 0.00523364869877696, + 0.006721483077853918, + 0.013849752023816109, + 0.010911475867033005, + 0.020238272845745087, + 0.012954446487128735, + 0.01425892859697342, + 0.012308433651924133, + 0.007146020885556936, + 0.008971292525529861, + 0.005267071537673473, + 0.004584182053804398, + 0.006566328927874565, + 0.006853737868368626, + 0.0036509865894913673, + 0.004881585948169231, + 0.004575618077069521, + 0.006514596752822399, + 0.012452129274606705, + 0.01378589030355215, + 0.020487260073423386, + 0.014230157248675823, + 0.013797448948025703, + 0.019193308427929878, + 0.01375472079962492, + 0.010929322801530361, + 0.009408766403794289, + 0.011182831600308418, + 0.018518205732107162, + 0.0077555919997394085, + 0.021851137280464172, + 0.029704609885811806, + 0.007449140306562185, + 0.01775297522544861, + 0.011422738432884216, + 0.023443598300218582, + 0.012274493463337421, + 0.016042377799749374, + 0.010032926686108112, + 0.008987962268292904, + 0.0069610499776899815, + 0.006016995292156935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.029821347445249557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305974766612053 + ], + "opposite_logprob_tokens": [ + -3.168059825897217 + ], + "logprob_reward": -0.04305974766612053, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305974766612053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1323356181383133, + 0.06312528252601624, + 0.16153475642204285, + 0.019069600850343704, + 0.06641729176044464, + 0.13804340362548828, + 0.04256536439061165, + 0.1754515916109085, + 0.1671493798494339, + 0.034307703375816345 + ], + "credit_map": [ + 0.021342812106013298, + 0.012360985390841961, + 0.021640341728925705, + 0.011642699129879475, + 0.012909014709293842, + 0.018740514293313026, + 0.014030826278030872, + 0.019668424502015114, + 0.008929803967475891, + 0.007254867814481258, + 0.011121568270027637, + 0.009450053796172142, + 0.007118348963558674, + 0.009790193289518356, + 0.0044671930372715, + 0.00499325105920434, + 0.025470895692706108, + 0.024702072143554688, + 0.024624736979603767, + 0.015414140187203884, + 0.021561570465564728, + 0.01881130412220955, + 0.013718483038246632, + 0.017231544479727745, + 0.0026119255926460028, + 0.0015474039828404784, + 0.0026436299085617065, + 0.002904432127252221, + 0.0015130937099456787, + 0.0021659161429852247, + 0.0018846278544515371, + 0.0037985711824148893, + 0.00785976555198431, + 0.0042331768199801445, + 0.012606674805283546, + 0.0070833745412528515, + 0.00580938346683979, + 0.008857754059135914, + 0.007134664803743362, + 0.012832491658627987, + 0.020098194479942322, + 0.018639765679836273, + 0.02071293443441391, + 0.016520535573363304, + 0.016155578196048737, + 0.0242892038077116, + 0.009648547507822514, + 0.011978639289736748, + 0.004991609137505293, + 0.004554831888526678, + 0.007083199918270111, + 0.005919735413044691, + 0.004152372013777494, + 0.005041430238634348, + 0.004804198630154133, + 0.006017984822392464, + 0.020369548350572586, + 0.016334623098373413, + 0.02625267207622528, + 0.017990075051784515, + 0.014912158250808716, + 0.037643592804670334, + 0.025265445932745934, + 0.01668347604572773, + 0.016187207773327827, + 0.020104905590415, + 0.025970542803406715, + 0.013822346925735474, + 0.024263784289360046, + 0.026135580614209175, + 0.014413083903491497, + 0.02625192143023014, + 0.004270649049431086, + 0.009335959330201149, + 0.0035423100925982, + 0.005104051902890205, + 0.0038459443021565676, + 0.003610388608649373, + 0.0026689425576478243, + 0.0019294554367661476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.04305974766612053 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and tracking closely behind the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged emperor with a stern face and piercing gaze standing at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor wear a dark black silk robe embroidered with gold dragons and a black and gold crown with jade tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are four muscular black horses galloping abreast ahead of a bronze chariot with intricate cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a vast, desolate wilderness with a setting sun casting a golden afterglow and kicking up dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944074597209692 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944074597209692, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944074597209692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1630917638540268, + 0.21752604842185974, + 0.05384013429284096, + 0.11122126132249832, + 0.06191705912351608, + 0.23156699538230896, + 0.03283978998661041, + 0.08136656135320663, + 0.025119271129369736, + 0.021511171013116837 + ], + "credit_map": [ + 0.026517683640122414, + 0.013646174222230911, + 0.027180179953575134, + 0.016363391652703285, + 0.01411204319447279, + 0.02438109926879406, + 0.01905526965856552, + 0.021835898980498314, + 0.0392729751765728, + 0.020812127739191055, + 0.05978100374341011, + 0.018100235611200333, + 0.02804790809750557, + 0.022847672924399376, + 0.011974809691309929, + 0.016689306125044823, + 0.008895094506442547, + 0.0060552372597157955, + 0.010902542620897293, + 0.005785558372735977, + 0.006475269794464111, + 0.0060544852167367935, + 0.004312005825340748, + 0.005359934642910957, + 0.016339724883437157, + 0.011240090243518353, + 0.02440943755209446, + 0.01436806470155716, + 0.0143256401643157, + 0.01107867993414402, + 0.00780014181509614, + 0.011659474112093449, + 0.006892126053571701, + 0.006619402207434177, + 0.008425762876868248, + 0.006860701367259026, + 0.008841359056532383, + 0.005099011119455099, + 0.006776954047381878, + 0.012401733547449112, + 0.032920848578214645, + 0.023935195058584213, + 0.046988073736429214, + 0.03217334300279617, + 0.024668337777256966, + 0.03307521343231201, + 0.016946865245699883, + 0.020859085023403168, + 0.004250382073223591, + 0.003253007074818015, + 0.0037431418895721436, + 0.004023969639092684, + 0.002933899639174342, + 0.0033851880580186844, + 0.005336789879947901, + 0.0059134080074727535, + 0.008595513179898262, + 0.006021070759743452, + 0.007127010729163885, + 0.006417716853320599, + 0.021776830777525902, + 0.017872439697384834, + 0.00723511166870594, + 0.006320858374238014, + 0.0023345863446593285, + 0.002714878413826227, + 0.0035757792647928, + 0.0019071148708462715, + 0.002228986006230116, + 0.005135589744895697, + 0.002355668693780899, + 0.004866665229201317, + 0.0024087843485176563, + 0.005502517335116863, + 0.002786230994388461, + 0.002945245010778308, + 0.002514114137738943, + 0.0026123609859496355, + 0.0016024205833673477, + 0.0011394971515983343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.005944074597209692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_4/index_tensor([8269, 8269])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_4/index_tensor([8269, 8269])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bee2416b8223147bc95fba181fd140f70d042c20 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_4/index_tensor([8269, 8269])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene opens with a low-angle tracking shot of a classic light blue convertible gliding smoothly along a winding cobblestone road at a leisurely pace. The car's polished chrome bumpers and wheel covers beautifully reflect the sunlight, and the tires rumble softly over the uneven, historic stone. The car passes a series of charming storybook cottages, their whitewashed, half-timbered walls adorned with climbing ivy and vibrant flower pots filled with geraniums. As the convertible continues its journey through this idyllic setting, the camera begins a steady, majestic aerial crane rise, gracefully ascending into the air. The perspective dramatically expands and pulls back to reveal the entire sun-drenched village nestled in a lush valley. The sequence culminates with a breathtaking aerial view, showcasing an intricate tapestry of rooftop houses, a small stone church with a steeple, and lush gardens, all bathed in the warm, golden light of a perfect afternoon, creating an atmosphere of profound calm and nostalgic beauty.", + "seed": 12689, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_4/index_8269_seed_12689_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.023116341046988963, + "vlm_logprob": -0.023116341046988963, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.023116341046988963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01059054210782051 + ], + "opposite_logprob_tokens": [ + -5.135590553283691 + ], + "logprob_reward": -0.01059054210782051, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01059054210782051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11224895715713501, + 0.10685316473245621, + 0.1080627590417862, + 0.10747086256742477, + 0.07818955928087234, + 0.13288740813732147, + 0.06432019174098969, + 0.10391470789909363, + 0.09744811058044434, + 0.08860423415899277 + ], + "credit_map": [ + 0.011845405213534832, + 0.009093279018998146, + 0.016600394621491432, + 0.01573294587433338, + 0.009839636273682117, + 0.010320117697119713, + 0.025229770690202713, + 0.013587400317192078, + 0.008827144280076027, + 0.00999433919787407, + 0.02087676152586937, + 0.015058238059282303, + 0.008848659694194794, + 0.007552991155534983, + 0.02076692320406437, + 0.014928106218576431, + 0.005937053821980953, + 0.004506048280745745, + 0.015616567805409431, + 0.011252541095018387, + 0.008120210841298103, + 0.03111978806555271, + 0.006903804838657379, + 0.024606741964817047, + 0.0065248990431427956, + 0.007560443133115768, + 0.01254989579319954, + 0.03381747379899025, + 0.012220012955367565, + 0.00934463832527399, + 0.01150290947407484, + 0.013950597494840622, + 0.0072101266123354435, + 0.010188184678554535, + 0.008214125409722328, + 0.014445546083152294, + 0.009004956111311913, + 0.012488255277276039, + 0.008692645467817783, + 0.00794572476297617, + 0.012999066151678562, + 0.028479469940066338, + 0.012408051639795303, + 0.027444783598184586, + 0.014822058379650116, + 0.011595282703638077, + 0.014051012694835663, + 0.011087684892117977, + 0.01044488325715065, + 0.005262361373752356, + 0.007347357925027609, + 0.006384906824678183, + 0.00978637021034956, + 0.006392699666321278, + 0.011382329277694225, + 0.007319282274693251, + 0.012890554033219814, + 0.01223609410226345, + 0.014563243836164474, + 0.011963140219449997, + 0.01183457113802433, + 0.012209709733724594, + 0.01749458536505699, + 0.010722803883254528, + 0.012036187574267387, + 0.01417553797364235, + 0.01105438731610775, + 0.01384874526411295, + 0.009254777804017067, + 0.012307853437960148, + 0.014232940040528774, + 0.010537679307162762, + 0.005459615960717201, + 0.008262346498668194, + 0.023824375122785568, + 0.012866460718214512, + 0.005487267393618822, + 0.00881313718855381, + 0.007784941699355841, + 0.016106097027659416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.01059054210782051 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003971189260483 + ], + "opposite_logprob_tokens": [ + -2.695039749145508 + ], + "logprob_reward": -0.07003971189260483, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003971189260483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08056626468896866, + 0.09416480362415314, + 0.08091670274734497, + 0.09563713520765305, + 0.12307586520910263, + 0.1835748553276062, + 0.05154856666922569, + 0.09307702630758286, + 0.06469094753265381, + 0.13274778425693512 + ], + "credit_map": [ + 0.007541066966950893, + 0.008405080065131187, + 0.011154054664075375, + 0.008700277656316757, + 0.008473328314721584, + 0.009990484453737736, + 0.014948572032153606, + 0.011353402398526669, + 0.008283075876533985, + 0.010244317352771759, + 0.020298322662711143, + 0.012753400020301342, + 0.006726683117449284, + 0.006707740947604179, + 0.015621223486959934, + 0.013530041091144085, + 0.004862804897129536, + 0.004624603316187859, + 0.02285768836736679, + 0.006924942135810852, + 0.005521097220480442, + 0.023559164255857468, + 0.004859361331909895, + 0.007707038428634405, + 0.007697508204728365, + 0.007445801515132189, + 0.007944019511342049, + 0.0322919599711895, + 0.010088273324072361, + 0.006439666263759136, + 0.009924735873937607, + 0.013805166818201542, + 0.007025036960840225, + 0.011468887329101562, + 0.014572794549167156, + 0.023463629186153412, + 0.015347357839345932, + 0.015360969118773937, + 0.024151921272277832, + 0.011685281060636044, + 0.01863769441843033, + 0.017533935606479645, + 0.016291776672005653, + 0.030561266466975212, + 0.018496405333280563, + 0.021877432242035866, + 0.03967095911502838, + 0.020505383610725403, + 0.006652639247477055, + 0.007008304353803396, + 0.010029887780547142, + 0.008669624105095863, + 0.0048805121332407, + 0.004112987779080868, + 0.0054013109765946865, + 0.00479330075904727, + 0.009524963796138763, + 0.009775998070836067, + 0.020982887595891953, + 0.015417434275150299, + 0.009691893123090267, + 0.007714089006185532, + 0.010747320018708706, + 0.009222439490258694, + 0.007163249887526035, + 0.011108036153018475, + 0.008524409495294094, + 0.011857124045491219, + 0.006022871471941471, + 0.005961482413113117, + 0.0077038840390741825, + 0.006349894218146801, + 0.011447210796177387, + 0.008172616362571716, + 0.040080394595861435, + 0.0202743262052536, + 0.008016259409487247, + 0.011774400249123573, + 0.013534467667341232, + 0.019448114559054375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.07003971189260483 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632735185325146 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632735185325146, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632735185325146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14089637994766235, + 0.06894593685865402, + 0.20345795154571533, + 0.02717500925064087, + 0.069559745490551, + 0.1087464839220047, + 0.05171326547861099, + 0.05717473104596138, + 0.05211272090673447, + 0.220217764377594 + ], + "credit_map": [ + 0.021074684336781502, + 0.018929027020931244, + 0.013224226422607899, + 0.008829163387417793, + 0.017479071393609047, + 0.01629849337041378, + 0.023980669677257538, + 0.021081048995256424, + 0.005692432168871164, + 0.005744447931647301, + 0.011964029632508755, + 0.009190989658236504, + 0.0075727058574557304, + 0.004793774336576462, + 0.013868396170437336, + 0.01011915784329176, + 0.013582238927483559, + 0.009391874074935913, + 0.07724133878946304, + 0.015491952188313007, + 0.010334477759897709, + 0.05747626721858978, + 0.007128833793103695, + 0.012810960412025452, + 0.003078036243095994, + 0.002972615184262395, + 0.0026567233726382256, + 0.005363021977245808, + 0.0037137409672141075, + 0.0030984138138592243, + 0.002758481539785862, + 0.003533976851031184, + 0.006795590743422508, + 0.006702398415654898, + 0.008607649244368076, + 0.014591948129236698, + 0.0059050763957202435, + 0.01173343788832426, + 0.008356551639735699, + 0.006867092568427324, + 0.0132950060069561, + 0.019545553252100945, + 0.013492099940776825, + 0.019412996247410774, + 0.011715088039636612, + 0.010047958232462406, + 0.011747404001653194, + 0.009490381926298141, + 0.0063080210238695145, + 0.00711422273889184, + 0.009229003451764584, + 0.007948325015604496, + 0.006909199524670839, + 0.004825086332857609, + 0.004611171782016754, + 0.004768235608935356, + 0.005572723690420389, + 0.006086548324674368, + 0.010663743130862713, + 0.007255560252815485, + 0.007191444747149944, + 0.004652767442166805, + 0.008707237429916859, + 0.007044702768325806, + 0.005883045960217714, + 0.007715539075434208, + 0.007302660495042801, + 0.009451922960579395, + 0.0037474846467375755, + 0.0051081920973956585, + 0.005832022987306118, + 0.007071854546666145, + 0.013389256782829762, + 0.017377665266394615, + 0.06652175635099411, + 0.033387523144483566, + 0.011105414479970932, + 0.019263412803411484, + 0.018337532877922058, + 0.04083521291613579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.007632735185325146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583592355251312 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583592355251312, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583592355251312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03414933383464813, + 0.11204370856285095, + 0.5628239512443542, + 0.04974788427352905, + 0.03265969455242157, + 0.052004165947437286, + 0.053275059908628464, + 0.020418427884578705, + 0.015059600584208965, + 0.06781812757253647 + ], + "credit_map": [ + 0.0028990162536501884, + 0.003941886126995087, + 0.00375383417122066, + 0.004260031972080469, + 0.003154196310788393, + 0.002879006788134575, + 0.008700769394636154, + 0.004560593981295824, + 0.010019108653068542, + 0.008601623587310314, + 0.02549639903008938, + 0.015544574707746506, + 0.007662101183086634, + 0.007370255887508392, + 0.021306132897734642, + 0.016043517738580704, + 0.03683755174279213, + 0.026059934869408607, + 0.21908800303936005, + 0.042148612439632416, + 0.026705510914325714, + 0.15487012267112732, + 0.019301362335681915, + 0.03781282901763916, + 0.004342676606029272, + 0.003567321691662073, + 0.004261445254087448, + 0.015779349952936172, + 0.004944951739162207, + 0.004900298081338406, + 0.0046616424806416035, + 0.007290198467671871, + 0.0036047447938472033, + 0.004127175081521273, + 0.0030800150707364082, + 0.005030413158237934, + 0.003683008486405015, + 0.005676389671862125, + 0.004261217080056667, + 0.0031967326067388058, + 0.005634787026792765, + 0.0072665028274059296, + 0.0040968372486531734, + 0.008956557139754295, + 0.006550638936460018, + 0.006170645356178284, + 0.008683924563229084, + 0.00464427238330245, + 0.009039697237312794, + 0.005792423617094755, + 0.005633667577058077, + 0.00546250818297267, + 0.006957797799259424, + 0.009711723774671555, + 0.004813222214579582, + 0.0058640181086957455, + 0.002494333777576685, + 0.0027100613806396723, + 0.0027039130218327045, + 0.0020465899724513292, + 0.0020603612065315247, + 0.0026855911128222942, + 0.002677211072295904, + 0.0030403640121221542, + 0.0015965185593813658, + 0.0024954569526016712, + 0.0020708669908344746, + 0.0019437980372458696, + 0.0011684418423101306, + 0.0018287625862285495, + 0.0021224061492830515, + 0.0018333496991544962, + 0.004522472154349089, + 0.004345507360994816, + 0.023648159578442574, + 0.011245707049965858, + 0.00317103392444551, + 0.00461853900924325, + 0.004376614466309547, + 0.011890100315213203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.020583592355251312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735123693943024 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735123693943024, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735123693943024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03953530266880989, + 0.09433019161224365, + 0.15888448059558868, + 0.07227005064487457, + 0.04477708786725998, + 0.16637952625751495, + 0.09942508488893509, + 0.03708209469914436, + 0.0891803428530693, + 0.19813576340675354 + ], + "credit_map": [ + 0.004213340580463409, + 0.0038369467947632074, + 0.005455578677356243, + 0.00412477133795619, + 0.0041214823722839355, + 0.005142078269273043, + 0.008174661546945572, + 0.004466446582227945, + 0.007951715029776096, + 0.00880036223679781, + 0.018905015662312508, + 0.012214364483952522, + 0.007420267444103956, + 0.00658990116789937, + 0.01794975996017456, + 0.0144988177344203, + 0.010315481573343277, + 0.0073498995043337345, + 0.05858777463436127, + 0.01285481359809637, + 0.008168725296854973, + 0.047977205365896225, + 0.005496332887560129, + 0.008134271949529648, + 0.004810951184481382, + 0.005082764197140932, + 0.006531359627842903, + 0.024957366287708282, + 0.006253544241189957, + 0.005940137431025505, + 0.007699127309024334, + 0.01099480502307415, + 0.004088050220161676, + 0.006529988255351782, + 0.004565326496958733, + 0.007416645064949989, + 0.004471697378903627, + 0.007431867066770792, + 0.005194808356463909, + 0.005078714806586504, + 0.017663147300481796, + 0.04053835570812225, + 0.01745646446943283, + 0.03232980892062187, + 0.022175246849656105, + 0.014715020544826984, + 0.009605331346392632, + 0.01189616322517395, + 0.01741434633731842, + 0.009315652772784233, + 0.01262105256319046, + 0.012835226953029633, + 0.011050490662455559, + 0.01742449216544628, + 0.008174803107976913, + 0.010589021258056164, + 0.004010021686553955, + 0.004630871582776308, + 0.00682829087600112, + 0.004135174211114645, + 0.004346247296780348, + 0.003705232637003064, + 0.004778238013386726, + 0.004648022353649139, + 0.009058807976543903, + 0.01535109244287014, + 0.009933577850461006, + 0.011829938739538193, + 0.008546228520572186, + 0.012322685681283474, + 0.010540926828980446, + 0.011597095988690853, + 0.014860549941658974, + 0.0205784160643816, + 0.046892281621694565, + 0.02944115549325943, + 0.01647566258907318, + 0.022259218618273735, + 0.02175149880349636, + 0.02587699703872204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.006735123693943024 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene opens with a low-angle tracking shot of a classic light blue convertible gliding smoothly along a winding cobblestone road at a leisurely pace. The car's polished chrome bumpers and wheel covers beautifully reflect the sunlight, and the tires rumble softly over the uneven, historic stone. The car passes a series of charming storybook cottages, their whitewashed, half-timbered walls adorned with climbing ivy and vibrant flower pots filled with geraniums. As the convertible continues its journey through this idyllic setting, the camera begins a steady, majestic aerial crane rise, gracefully ascending into the air. The perspective dramatically expands and pulls back to reveal the entire sun-drenched village nestled in a lush valley. The sequence culminates with a breathtaking aerial view, showcasing an intricate tapestry of rooftop houses, a small stone church with a steeple, and lush gardens, all bathed in the warm, golden light of a perfect afternoon, creating an atmosphere of profound calm and nostalgic beauty.", + "seed": 12690, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_4/index_8269_seed_12690_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03480044929310679, + "vlm_logprob": -0.03480044929310679, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03480044929310679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011125907301902771 + ], + "opposite_logprob_tokens": [ + -6.0111260414123535 + ], + "logprob_reward": -0.011125907301902771, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.011125907301902771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07930821180343628, + 0.10716713964939117, + 0.06339094787836075, + 0.04224899411201477, + 0.02812313102185726, + 0.021735696122050285, + 0.04361451417207718, + 0.5461652278900146, + 0.04292074963450432, + 0.025325359776616096 + ], + "credit_map": [ + 0.007376061286777258, + 0.009860193356871605, + 0.009691819548606873, + 0.008589651435613632, + 0.012145640328526497, + 0.010650618001818657, + 0.01126576866954565, + 0.009728462435305119, + 0.01597956381738186, + 0.007225631270557642, + 0.005561866331845522, + 0.008042093366384506, + 0.017568198963999748, + 0.038263048976659775, + 0.008257422596216202, + 0.006269308738410473, + 0.008793314918875694, + 0.012317261658608913, + 0.011910438537597656, + 0.004416004754602909, + 0.007527718786150217, + 0.009385227225720882, + 0.004217846319079399, + 0.004823141731321812, + 0.008005086332559586, + 0.010548338294029236, + 0.007210127077996731, + 0.004502815194427967, + 0.002904978347942233, + 0.0035364299546927214, + 0.002659604884684086, + 0.0028816137928515673, + 0.003768762806430459, + 0.005483107641339302, + 0.003419863758608699, + 0.0030687921680510044, + 0.0033829272724688053, + 0.0037989134434610605, + 0.00259884144179523, + 0.002601921558380127, + 0.0029325862415134907, + 0.0033456264063715935, + 0.0026573094073683023, + 0.003746441099792719, + 0.0021663822699338198, + 0.002285240450873971, + 0.0026023932732641697, + 0.001999717205762863, + 0.005694426130503416, + 0.004894730634987354, + 0.0043672011233866215, + 0.00680227717384696, + 0.004990549758076668, + 0.004745048470795155, + 0.005785895511507988, + 0.006334385834634304, + 0.028233129531145096, + 0.018614839762449265, + 0.08394908905029297, + 0.2225332111120224, + 0.015790322795510292, + 0.027838842943310738, + 0.01597217656672001, + 0.13323360681533813, + 0.007349884137511253, + 0.0029660018626600504, + 0.005385121796280146, + 0.006320351269096136, + 0.004137068521231413, + 0.007019519340246916, + 0.0038879872299730778, + 0.005854814313352108, + 0.0030039793346077204, + 0.0037077635060995817, + 0.0034887236542999744, + 0.0029325748328119516, + 0.00340793514624238, + 0.002426415914669633, + 0.002444441430270672, + 0.003913526888936758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.011125907301902771 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306888207793236 + ], + "opposite_logprob_tokens": [ + -3.1680688858032227 + ], + "logprob_reward": -0.04306888207793236, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306888207793236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07992185652256012, + 0.08662128448486328, + 0.04959685355424881, + 0.059538595378398895, + 0.04842083156108856, + 0.09495451301336288, + 0.17820721864700317, + 0.24431969225406647, + 0.10067372024059296, + 0.05774539336562157 + ], + "credit_map": [ + 0.0075660827569663525, + 0.011056717485189438, + 0.009895588271319866, + 0.01186211034655571, + 0.006128476932644844, + 0.009795351885259151, + 0.012693512253463268, + 0.010924020782113075, + 0.012703686952590942, + 0.009971613995730877, + 0.005938002374023199, + 0.005798278376460075, + 0.01028376817703247, + 0.026387445628643036, + 0.005143557675182819, + 0.010394933633506298, + 0.00655335932970047, + 0.008661705069243908, + 0.008147659711539745, + 0.0036062526050955057, + 0.006994875613600016, + 0.007809959817677736, + 0.0030973900575190783, + 0.004725651815533638, + 0.010446334257721901, + 0.01438377145677805, + 0.007674310822039843, + 0.004415756557136774, + 0.006562335416674614, + 0.008985724300146103, + 0.00375258456915617, + 0.00331778172403574, + 0.004807118326425552, + 0.004898793995380402, + 0.007895384915173054, + 0.010716944001615047, + 0.0038848009426146746, + 0.0037362666334956884, + 0.006414559204131365, + 0.006066964939236641, + 0.017250992357730865, + 0.021822059527039528, + 0.010692794807255268, + 0.017493024468421936, + 0.00594296446070075, + 0.007785535883158445, + 0.007329933810979128, + 0.006637211423367262, + 0.017077436670660973, + 0.027575183659791946, + 0.02644241973757744, + 0.028499307110905647, + 0.023142175748944283, + 0.01863139495253563, + 0.014422200620174408, + 0.022417102009058, + 0.014527094550430775, + 0.01096365600824356, + 0.035409554839134216, + 0.09412404149770737, + 0.007600853685289621, + 0.013501795008778572, + 0.008906960487365723, + 0.0592857226729393, + 0.01849132776260376, + 0.00876858364790678, + 0.014224687591195107, + 0.015122982673346996, + 0.008555326610803604, + 0.011649873107671738, + 0.008822072297334671, + 0.015038866549730301, + 0.007603656966239214, + 0.007185460068285465, + 0.007981245405972004, + 0.006722540594637394, + 0.007102676667273045, + 0.007502022199332714, + 0.005097626242786646, + 0.008550163358449936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.04306888207793236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738202180713415 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738202180713415, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738202180713415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06601012498140335, + 0.04711122065782547, + 0.06590785831212997, + 0.042452018707990646, + 0.05105515941977501, + 0.04037487879395485, + 0.12653528153896332, + 0.4902523159980774, + 0.05544155836105347, + 0.014859581366181374 + ], + "credit_map": [ + 0.004693895578384399, + 0.006247836630791426, + 0.006100288592278957, + 0.012745277024805546, + 0.004319850355386734, + 0.007163250353187323, + 0.017143860459327698, + 0.00759586738422513, + 0.0038115973584353924, + 0.00656118243932724, + 0.0028576550539582968, + 0.003481066320091486, + 0.006489028222858906, + 0.012895370833575726, + 0.005242643412202597, + 0.0057726772502064705, + 0.009979801252484322, + 0.016257911920547485, + 0.00960381980985403, + 0.0039000927936285734, + 0.009968914091587067, + 0.00771694490686059, + 0.0035396632738411427, + 0.004940709099173546, + 0.0077774557285010815, + 0.01120657380670309, + 0.005064727738499641, + 0.003020567586645484, + 0.00498716626316309, + 0.004628698341548443, + 0.00299481931142509, + 0.0027720131911337376, + 0.00883767195045948, + 0.01156721729785204, + 0.006021615117788315, + 0.008080719970166683, + 0.00454419432207942, + 0.004439209122210741, + 0.00408909423276782, + 0.003475435543805361, + 0.006067133974283934, + 0.006798954680562019, + 0.004369528032839298, + 0.0084531856700778, + 0.003876639762893319, + 0.003455272177234292, + 0.003202160121873021, + 0.004152003675699234, + 0.01884295605123043, + 0.01577823981642723, + 0.014922155998647213, + 0.014034871011972427, + 0.014827742241322994, + 0.01916072703897953, + 0.007651832420378923, + 0.021316763013601303, + 0.026837343350052834, + 0.016846908256411552, + 0.07187467068433762, + 0.19862419366836548, + 0.011976216919720173, + 0.026657380163669586, + 0.014645056799054146, + 0.12279053777456284, + 0.014514066278934479, + 0.0055252485908567905, + 0.007747713476419449, + 0.007094420492649078, + 0.0048979707062244415, + 0.006496863439679146, + 0.0046625747345387936, + 0.004502706695348024, + 0.0023477207869291306, + 0.0017681065946817398, + 0.002027224050834775, + 0.0016380511224269867, + 0.0014941117260605097, + 0.001850050175562501, + 0.0012893329840153456, + 0.002444983460009098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.006738202180713415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306088760495186 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306088760495186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306088760495186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06177690625190735, + 0.1260591298341751, + 0.1038285568356514, + 0.042378343641757965, + 0.07182993739843369, + 0.09188713133335114, + 0.14387169480323792, + 0.17744912207126617, + 0.16669489443302155, + 0.01422432716935873 + ], + "credit_map": [ + 0.005106176715344191, + 0.004117200151085854, + 0.007968769408762455, + 0.01108002569526434, + 0.005439216271042824, + 0.010483293794095516, + 0.009323629550635815, + 0.008258592337369919, + 0.013712571002542973, + 0.007934484630823135, + 0.008040503598749638, + 0.006840287242084742, + 0.02092326059937477, + 0.047364138066768646, + 0.013499503023922443, + 0.007744385860860348, + 0.018010538071393967, + 0.028041567653417587, + 0.013409323059022427, + 0.006110536865890026, + 0.013122218661010265, + 0.011265032924711704, + 0.005571510177105665, + 0.008297821506857872, + 0.0073868390172719955, + 0.010988830588757992, + 0.004758948460221291, + 0.004516847897320986, + 0.004439881071448326, + 0.0044091809540987015, + 0.0027893572114408016, + 0.0030884600710123777, + 0.012731104157865047, + 0.011465436778962612, + 0.009655301459133625, + 0.011807957664132118, + 0.005780466832220554, + 0.006994833704084158, + 0.006463691592216492, + 0.006931142881512642, + 0.008411847054958344, + 0.008989824913442135, + 0.013393580913543701, + 0.01559069100767374, + 0.006893136538565159, + 0.009830893017351627, + 0.011142531409859657, + 0.017634622752666473, + 0.015118238516151905, + 0.01657969504594803, + 0.01516388263553381, + 0.020764419808983803, + 0.022383050993084908, + 0.020741460844874382, + 0.017419373616576195, + 0.01570156216621399, + 0.01192026399075985, + 0.011511707678437233, + 0.02620311640202999, + 0.06134724244475365, + 0.012354735285043716, + 0.008862035349011421, + 0.007914392277598381, + 0.03733561933040619, + 0.02177412249147892, + 0.013868474401533604, + 0.02390200085937977, + 0.02375790663063526, + 0.018936699256300926, + 0.018903439864516258, + 0.01459028571844101, + 0.030961964279413223, + 0.002092170063406229, + 0.001388402539305389, + 0.0018331784522160888, + 0.0019972536247223616, + 0.002171402098610997, + 0.0013421899639070034, + 0.0014650350203737617, + 0.0019346942426636815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.04306088760495186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000836730003357 + ], + "opposite_logprob_tokens": [ + -2.6950082778930664 + ], + "logprob_reward": -0.07000836730003357, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000836730003357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019160496070981026, + 0.02061520703136921, + 0.025927528738975525, + 0.0635501891374588, + 0.03563595190644264, + 0.01794186420738697, + 0.027457285672426224, + 0.6964613795280457, + 0.08188187330961227, + 0.011368242092430592 + ], + "credit_map": [ + 0.0014268317027017474, + 0.0023532193154096603, + 0.0027144793421030045, + 0.002778409980237484, + 0.0017461831448599696, + 0.002276854356750846, + 0.0036591440439224243, + 0.002205374650657177, + 0.0025632057804614305, + 0.004766928032040596, + 0.00209002080373466, + 0.0022840090095996857, + 0.0028144882526248693, + 0.0023151750210672617, + 0.0018275883048772812, + 0.0019537911284714937, + 0.004069614689797163, + 0.005913300905376673, + 0.0038118346128612757, + 0.0016214590286836028, + 0.0029130505863577127, + 0.0036096624098718166, + 0.0015678169438615441, + 0.002420790260657668, + 0.011798267252743244, + 0.018724709749221802, + 0.007526399567723274, + 0.006130104418843985, + 0.0051383706741034985, + 0.0059768869541585445, + 0.004212845582515001, + 0.0040426054038107395, + 0.006724393926560879, + 0.007719697896391153, + 0.003970809280872345, + 0.005488976836204529, + 0.0038556174840778112, + 0.0030105269979685545, + 0.002530675148591399, + 0.002335255267098546, + 0.002202979987487197, + 0.003358439076691866, + 0.0023731791879981756, + 0.00364492810331285, + 0.0015545610804110765, + 0.001849586609750986, + 0.0011780436616390944, + 0.001780146500095725, + 0.004065927118062973, + 0.004083950538188219, + 0.002930290997028351, + 0.004182213917374611, + 0.002697313902899623, + 0.003540661418810487, + 0.0026092934422194958, + 0.00334763340651989, + 0.03747258335351944, + 0.025326920673251152, + 0.1036127358675003, + 0.28270530700683594, + 0.016591135412454605, + 0.03692331537604332, + 0.02021496370434761, + 0.17361444234848022, + 0.021876171231269836, + 0.008113778196275234, + 0.011015770956873894, + 0.009921100921928883, + 0.008320124819874763, + 0.009408115409314632, + 0.006106054875999689, + 0.007120748516172171, + 0.0016288668848574162, + 0.0021748836152255535, + 0.001658148830756545, + 0.0011533069191500545, + 0.0010680685518309474, + 0.001165541703812778, + 0.0010390098905190825, + 0.0014804152306169271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.07000836730003357 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene opens with a low-angle tracking shot of a classic light blue convertible gliding smoothly along a winding cobblestone road at a leisurely pace. The car's polished chrome bumpers and wheel covers beautifully reflect the sunlight, and the tires rumble softly over the uneven, historic stone. The car passes a series of charming storybook cottages, their whitewashed, half-timbered walls adorned with climbing ivy and vibrant flower pots filled with geraniums. As the convertible continues its journey through this idyllic setting, the camera begins a steady, majestic aerial crane rise, gracefully ascending into the air. The perspective dramatically expands and pulls back to reveal the entire sun-drenched village nestled in a lush valley. The sequence culminates with a breathtaking aerial view, showcasing an intricate tapestry of rooftop houses, a small stone church with a steeple, and lush gardens, all bathed in the warm, golden light of a perfect afternoon, creating an atmosphere of profound calm and nostalgic beauty.", + "seed": 12691, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_4/index_8269_seed_12691_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06733976295217872, + "vlm_logprob": -0.06733976295217872, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06733976295217872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009832167066633701 + ], + "opposite_logprob_tokens": [ + -6.134832382202148 + ], + "logprob_reward": -0.009832167066633701, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.009832167066633701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06175075098872185, + 0.04868496581912041, + 0.07065971940755844, + 0.08555182069540024, + 0.07417110353708267, + 0.08838454633951187, + 0.0611356720328331, + 0.16149182617664337, + 0.15257145464420319, + 0.19559817016124725 + ], + "credit_map": [ + 0.006820593029260635, + 0.0058800396509468555, + 0.011240694671869278, + 0.007998703047633171, + 0.010736439377069473, + 0.006857725791633129, + 0.007047389168292284, + 0.0051691592670977116, + 0.005419837310910225, + 0.008348112925887108, + 0.008457846008241177, + 0.005649344529956579, + 0.0037165640387684107, + 0.003862333483994007, + 0.008302316069602966, + 0.0049286046996712685, + 0.010240732692182064, + 0.008765814825892448, + 0.004954044241458178, + 0.006223846226930618, + 0.006119126453995705, + 0.009986110962927341, + 0.008665218017995358, + 0.015704821795225143, + 0.0074215298518538475, + 0.004719414748251438, + 0.004183079581707716, + 0.002803418319672346, + 0.004921093583106995, + 0.02362642250955105, + 0.00431746756657958, + 0.03355938941240311, + 0.004569262731820345, + 0.009459704160690308, + 0.002030349802225828, + 0.0018437228864058852, + 0.03027857281267643, + 0.014100221917033195, + 0.004187621176242828, + 0.007701647002249956, + 0.014178404584527016, + 0.012756175361573696, + 0.009233009070158005, + 0.00669071963056922, + 0.011077603325247765, + 0.012849216349422932, + 0.007034138776361942, + 0.014565272256731987, + 0.006244816817343235, + 0.004767206031829119, + 0.007526107598096132, + 0.0064555928111076355, + 0.007266181521117687, + 0.00880570150911808, + 0.006018604151904583, + 0.01405145414173603, + 0.015394196845591068, + 0.02295825257897377, + 0.0225039254873991, + 0.03167124465107918, + 0.02394537813961506, + 0.015993408858776093, + 0.01221307460218668, + 0.016812339425086975, + 0.008301346562802792, + 0.020262202247977257, + 0.008910158649086952, + 0.04399219900369644, + 0.05323709920048714, + 0.008437965996563435, + 0.004819070920348167, + 0.004611387383192778, + 0.012535173445940018, + 0.04549875110387802, + 0.027150537818670273, + 0.06386081129312515, + 0.01320703886449337, + 0.014981708489358425, + 0.011238505132496357, + 0.007125630043447018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.009832167066633701 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428111046552658 + ], + "opposite_logprob_tokens": [ + -2.0178110599517822 + ], + "logprob_reward": -0.1428111046552658, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428111046552658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18729177117347717, + 0.13519138097763062, + 0.04461776092648506, + 0.08569890260696411, + 0.0845678374171257, + 0.04186270758509636, + 0.04929797351360321, + 0.06692824512720108, + 0.21009483933448792, + 0.09444854408502579 + ], + "credit_map": [ + 0.017151370644569397, + 0.019336873665452003, + 0.029157552868127823, + 0.010204026475548744, + 0.046497248113155365, + 0.032628051936626434, + 0.014200039207935333, + 0.018116600811481476, + 0.02237197570502758, + 0.006407168228179216, + 0.012861008755862713, + 0.006046377122402191, + 0.007348799612373114, + 0.017293835058808327, + 0.013305559754371643, + 0.04955664277076721, + 0.003085627220571041, + 0.004928130190819502, + 0.00429700268432498, + 0.005168837029486895, + 0.0033516951370984316, + 0.007547546178102493, + 0.005301440600305796, + 0.01093748677521944, + 0.00826990231871605, + 0.0036618083249777555, + 0.00412657018750906, + 0.0027986124623566866, + 0.007783613633364439, + 0.022747164592146873, + 0.004636645317077637, + 0.0316745899617672, + 0.00541381910443306, + 0.011629222892224789, + 0.003901443677023053, + 0.0029570129700005054, + 0.029720256105065346, + 0.015805626288056374, + 0.005960649345070124, + 0.009179802611470222, + 0.002928671659901738, + 0.007643566932529211, + 0.008150972425937653, + 0.004168218467384577, + 0.004473445937037468, + 0.0063554891385138035, + 0.003090481972321868, + 0.005051859188824892, + 0.008411800488829613, + 0.005952157080173492, + 0.0050118411891162395, + 0.003017137525603175, + 0.008031736128032207, + 0.006916387472301722, + 0.003780931467190385, + 0.008175979368388653, + 0.005904779303818941, + 0.0075206332840025425, + 0.006648913491517305, + 0.012462085112929344, + 0.01052948646247387, + 0.007318128366023302, + 0.005912126041948795, + 0.010632093995809555, + 0.014736779034137726, + 0.0325065553188324, + 0.018250836059451103, + 0.03474987670779228, + 0.06594721972942352, + 0.022311989217996597, + 0.01104446966201067, + 0.010547123849391937, + 0.01158746238797903, + 0.01146569475531578, + 0.015332753770053387, + 0.022715238854289055, + 0.009448778815567493, + 0.010225579142570496, + 0.007222582120448351, + 0.006450457964092493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.1428111046552658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076298960484564304 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.0076298960484564304, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076298960484564304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14553391933441162, + 0.12531256675720215, + 0.042243991047143936, + 0.17172810435295105, + 0.08235403150320053, + 0.07497727125883102, + 0.06266490370035172, + 0.08622752130031586, + 0.13490575551986694, + 0.07405189424753189 + ], + "credit_map": [ + 0.010934009216725826, + 0.014083231799304485, + 0.02606254070997238, + 0.014678597450256348, + 0.04542059451341629, + 0.008239283226430416, + 0.016047615557909012, + 0.010068060830235481, + 0.012840906158089638, + 0.016331151127815247, + 0.012660782784223557, + 0.008072701282799244, + 0.011201242916285992, + 0.01583930477499962, + 0.017554858699440956, + 0.03081163391470909, + 0.003901056945323944, + 0.003994550555944443, + 0.0034728979226201773, + 0.0055505260825157166, + 0.005338724236935377, + 0.005917145870625973, + 0.004795847460627556, + 0.009273244068026543, + 0.014468863606452942, + 0.007027991581708193, + 0.008882995694875717, + 0.005664516240358353, + 0.011439423076808453, + 0.04767881706357002, + 0.009574873372912407, + 0.06699062883853912, + 0.007434957195073366, + 0.008871261961758137, + 0.004632767755538225, + 0.0038759077433496714, + 0.022918837144970894, + 0.01577042043209076, + 0.007996764965355396, + 0.010853125713765621, + 0.009342570789158344, + 0.006414726842194796, + 0.006543239112943411, + 0.005713712889701128, + 0.0077112712897360325, + 0.011457775719463825, + 0.006812242791056633, + 0.020981743931770325, + 0.011479404754936695, + 0.007670798804610968, + 0.006782221607863903, + 0.0036024071741849184, + 0.011074193753302097, + 0.009029228240251541, + 0.004839091096073389, + 0.008187566883862019, + 0.008624603971838951, + 0.010462336242198944, + 0.011782444082200527, + 0.01666371151804924, + 0.014619352295994759, + 0.008995610289275646, + 0.0064772446639835835, + 0.008602229878306389, + 0.01035032607614994, + 0.019659869372844696, + 0.01874108985066414, + 0.017845025286078453, + 0.046115197241306305, + 0.01119108684360981, + 0.00550020020455122, + 0.005502956919372082, + 0.005517332348972559, + 0.013868940994143486, + 0.011633796617388725, + 0.024253910407423973, + 0.0045789391733706, + 0.006280198227614164, + 0.005183731205761433, + 0.0027350529562681913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0076298960484564304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036079823970795 + ], + "opposite_logprob_tokens": [ + -1.9103608131408691 + ], + "logprob_reward": -0.16036079823970795, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036079823970795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12761250138282776, + 0.033769212663173676, + 0.04104260355234146, + 0.021338261663913727, + 0.05425810441374779, + 0.07135508209466934, + 0.046474557369947433, + 0.195535808801651, + 0.23158764839172363, + 0.17702625691890717 + ], + "credit_map": [ + 0.009284685365855694, + 0.007896422408521175, + 0.034011855721473694, + 0.008031677454710007, + 0.04176744073629379, + 0.007128090597689152, + 0.011363698169589043, + 0.008128621615469456, + 0.003773543518036604, + 0.004617022350430489, + 0.004475652240216732, + 0.003215102944523096, + 0.0024658204056322575, + 0.003345975885167718, + 0.005295116920024157, + 0.006580978166311979, + 0.0038157422095537186, + 0.004433088004589081, + 0.0035969261080026627, + 0.004508340265601873, + 0.002867831150069833, + 0.007731278892606497, + 0.007490241900086403, + 0.006599155720323324, + 0.00411874707788229, + 0.0027009767945855856, + 0.0018879116978496313, + 0.001804239465855062, + 0.00308116152882576, + 0.0020637582056224346, + 0.0029008304700255394, + 0.0027806363068521023, + 0.0031850433442741632, + 0.0056052543222904205, + 0.0023786877281963825, + 0.0025325287133455276, + 0.018398061394691467, + 0.010486719198524952, + 0.005170866381376982, + 0.006500943563878536, + 0.010731928050518036, + 0.007610933389514685, + 0.005837101023644209, + 0.0052696228958666325, + 0.007983125746250153, + 0.010359704494476318, + 0.007261134684085846, + 0.016301525756716728, + 0.005322442390024662, + 0.0040911422111094, + 0.005050872452557087, + 0.004146120510995388, + 0.004895300138741732, + 0.006338202394545078, + 0.004945029970258474, + 0.011685448698699474, + 0.01986268348991871, + 0.029153503477573395, + 0.024712935090065002, + 0.037443164736032486, + 0.028751101344823837, + 0.019990133121609688, + 0.015346795320510864, + 0.02027549408376217, + 0.014390114694833755, + 0.028704410418868065, + 0.011740372516214848, + 0.06495042145252228, + 0.07176234573125839, + 0.016886955127120018, + 0.010335573926568031, + 0.012817451730370522, + 0.010316389612853527, + 0.039274126291275024, + 0.026717979460954666, + 0.05924015864729881, + 0.009457126259803772, + 0.012190214358270168, + 0.013671978376805782, + 0.006158289033919573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.16036079823970795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064848750829697 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064848750829697, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064848750829697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06541699171066284, + 0.030830806121230125, + 0.027107996866106987, + 0.03974982351064682, + 0.055364303290843964, + 0.03711975738406181, + 0.04031935706734657, + 0.0832691639661789, + 0.5334768295288086, + 0.08734484761953354 + ], + "credit_map": [ + 0.008022126741707325, + 0.009947309270501137, + 0.011038471944630146, + 0.0056192162446677685, + 0.011495156213641167, + 0.00813975278288126, + 0.005639379378408194, + 0.005515586584806442, + 0.004295628983527422, + 0.005474108271300793, + 0.005859347525984049, + 0.002383908024057746, + 0.0027913169469684362, + 0.0022686822339892387, + 0.0040489076636731625, + 0.0037089106626808643, + 0.0020726099610328674, + 0.0031486356165260077, + 0.001964076654985547, + 0.0025684896390885115, + 0.002218367997556925, + 0.004072403069585562, + 0.004848669283092022, + 0.0062147481366992, + 0.003522661281749606, + 0.0021250152494758368, + 0.0021313491743057966, + 0.001897029229439795, + 0.003864900441840291, + 0.009364004246890545, + 0.003547292435541749, + 0.013297573663294315, + 0.004192942753434181, + 0.006015208084136248, + 0.0034590759314596653, + 0.003289986401796341, + 0.011623977683484554, + 0.010052000172436237, + 0.006014264654368162, + 0.010716849938035011, + 0.0052715204656124115, + 0.0031914592254906893, + 0.004938027821481228, + 0.00350041501224041, + 0.0037967958487570286, + 0.0053150285966694355, + 0.002909873379394412, + 0.008196639828383923, + 0.005246568936854601, + 0.0035875963512808084, + 0.004198235459625721, + 0.0032956604845821857, + 0.0065332939848303795, + 0.006193848326802254, + 0.0034104606602340937, + 0.007853700779378414, + 0.009230196475982666, + 0.010736733675003052, + 0.012564748525619507, + 0.015341900289058685, + 0.012570948339998722, + 0.008130520582199097, + 0.005947340279817581, + 0.00874678697437048, + 0.026417816057801247, + 0.075217604637146, + 0.03219582140445709, + 0.15001709759235382, + 0.19550572335720062, + 0.027520667761564255, + 0.013960074633359909, + 0.01264207437634468, + 0.005979219451546669, + 0.016568146646022797, + 0.01453198492527008, + 0.029077261686325073, + 0.005427275784313679, + 0.007345895748585463, + 0.005148147698491812, + 0.0032669322099536657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.016064848750829697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene opens with a low-angle tracking shot of a classic light blue convertible gliding smoothly along a winding cobblestone road at a leisurely pace. The car's polished chrome bumpers and wheel covers beautifully reflect the sunlight, and the tires rumble softly over the uneven, historic stone. The car passes a series of charming storybook cottages, their whitewashed, half-timbered walls adorned with climbing ivy and vibrant flower pots filled with geraniums. As the convertible continues its journey through this idyllic setting, the camera begins a steady, majestic aerial crane rise, gracefully ascending into the air. The perspective dramatically expands and pulls back to reveal the entire sun-drenched village nestled in a lush valley. The sequence culminates with a breathtaking aerial view, showcasing an intricate tapestry of rooftop houses, a small stone church with a steeple, and lush gardens, all bathed in the warm, golden light of a perfect afternoon, creating an atmosphere of profound calm and nostalgic beauty.", + "seed": 12692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_4/index_8269_seed_12692_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025909436214715243, + "vlm_logprob": -0.025909436214715243, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.025909436214715243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004755615256726742 + ], + "opposite_logprob_tokens": [ + -6.25475549697876 + ], + "logprob_reward": -0.004755615256726742, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004755615256726742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14211180806159973, + 0.07031317055225372, + 0.025463227182626724, + 0.0974968895316124, + 0.03738148510456085, + 0.060111865401268005, + 0.1771443784236908, + 0.15099184215068817, + 0.1850792020559311, + 0.05390615016222 + ], + "credit_map": [ + 0.028377031907439232, + 0.010152062401175499, + 0.008754423819482327, + 0.009119225665926933, + 0.026148425415158272, + 0.026665056124329567, + 0.012842878699302673, + 0.020052701234817505, + 0.007779897190630436, + 0.008553333580493927, + 0.006810368038713932, + 0.0065289512276649475, + 0.01358130294829607, + 0.00902653206139803, + 0.008922227658331394, + 0.009110553190112114, + 0.0031388464849442244, + 0.0028965529054403305, + 0.002422306453809142, + 0.0027641162741929293, + 0.003560231765732169, + 0.0028296366799622774, + 0.0033326048869639635, + 0.004518931265920401, + 0.017335660755634308, + 0.005461086984723806, + 0.04154784232378006, + 0.005316657479852438, + 0.008420359343290329, + 0.00833571795374155, + 0.005584387574344873, + 0.005495178978890181, + 0.004766372498124838, + 0.003668352495878935, + 0.0038460504729300737, + 0.00833032839000225, + 0.0035788549575954676, + 0.0028406912460923195, + 0.007094362750649452, + 0.0032564718276262283, + 0.006580145098268986, + 0.0058150324039161205, + 0.009369352832436562, + 0.01606650836765766, + 0.00664748577401042, + 0.0047709462232887745, + 0.006310084369033575, + 0.0045523145236074924, + 0.011567812412977219, + 0.01067699771374464, + 0.06062985956668854, + 0.045335493981838226, + 0.009403623640537262, + 0.007965407334268093, + 0.020193589851260185, + 0.011371598578989506, + 0.008046846836805344, + 0.006536770146340132, + 0.006729711778461933, + 0.04956354200839996, + 0.007977833040058613, + 0.00903294701129198, + 0.014446060173213482, + 0.048658132553100586, + 0.03278804197907448, + 0.03934774920344353, + 0.022451339289546013, + 0.01473484467715025, + 0.01845928281545639, + 0.0166385006159544, + 0.01976914517581463, + 0.020890288054943085, + 0.011739886365830898, + 0.005748861934989691, + 0.005462992936372757, + 0.0062429532408714294, + 0.005154048558324575, + 0.008422496728599072, + 0.005641658790409565, + 0.005493245553225279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.004755615256726742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897985726594925 + ], + "opposite_logprob_tokens": [ + -2.578979969024658 + ], + "logprob_reward": -0.07897985726594925, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897985726594925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19637268781661987, + 0.02372269704937935, + 0.035655174404382706, + 0.06676867604255676, + 0.01876709796488285, + 0.059626705944538116, + 0.3996998071670532, + 0.09982849657535553, + 0.02125352807343006, + 0.07830508053302765 + ], + "credit_map": [ + 0.02835506573319435, + 0.013272630050778389, + 0.018393000587821007, + 0.015245135873556137, + 0.023111190646886826, + 0.03598053753376007, + 0.015714526176452637, + 0.04630062356591225, + 0.0029811507556587458, + 0.0030321828089654446, + 0.001801176811568439, + 0.0020717098377645016, + 0.004559574648737907, + 0.0038174446672201157, + 0.0025752882938832045, + 0.0028841716703027487, + 0.005117934197187424, + 0.004497167654335499, + 0.00391675578430295, + 0.004225355572998524, + 0.005549659486860037, + 0.00329038524068892, + 0.004711077082902193, + 0.004346844274550676, + 0.015391901135444641, + 0.0053516561165452, + 0.025174112990498543, + 0.0037795149255543947, + 0.0060949851758778095, + 0.0045914240181446075, + 0.003368838457390666, + 0.0030162509065121412, + 0.0019024325301870704, + 0.0011439960217103362, + 0.0015314052579924464, + 0.0037987607065588236, + 0.001934496103785932, + 0.002915644785389304, + 0.002822955371811986, + 0.0027174088172614574, + 0.006671165116131306, + 0.006006744224578142, + 0.007655859459191561, + 0.013764391653239727, + 0.0062384093180298805, + 0.0048425812274217606, + 0.0071982271037995815, + 0.00724932923913002, + 0.019981438294053078, + 0.014199710451066494, + 0.13563169538974762, + 0.12113091349601746, + 0.015022129751741886, + 0.011362142860889435, + 0.05523556470870972, + 0.027136215940117836, + 0.004459975752979517, + 0.0052239131182432175, + 0.004405268933624029, + 0.02961769327521324, + 0.005087325349450111, + 0.004661384504288435, + 0.00818656850606203, + 0.03818637132644653, + 0.002785084769129753, + 0.0027743512764573097, + 0.002208809833973646, + 0.0023233992978930473, + 0.002395652700215578, + 0.0018942889291793108, + 0.0038927365094423294, + 0.002979208482429385, + 0.00629413640126586, + 0.0052899825386703014, + 0.009467603638768196, + 0.015597955323755741, + 0.011218281462788582, + 0.00881311297416687, + 0.01272167544811964, + 0.008902340196073055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.07897985726594925 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036170552484691143 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036170552484691143, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036170552484691143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07691247761249542, + 0.018314233049750328, + 0.02685144916176796, + 0.04557563364505768, + 0.07992972433567047, + 0.05820083245635033, + 0.3000357747077942, + 0.1767648309469223, + 0.1742057353258133, + 0.04320934787392616 + ], + "credit_map": [ + 0.010419772006571293, + 0.006893687415868044, + 0.0064618997275829315, + 0.005374223459511995, + 0.011421773582696915, + 0.014083374291658401, + 0.008193827234208584, + 0.014063908718526363, + 0.0027384194545447826, + 0.0027266237884759903, + 0.0016596904024481773, + 0.0018964606570079923, + 0.002227455610409379, + 0.0021574762649834156, + 0.0023319488391280174, + 0.002576155588030815, + 0.003141895169392228, + 0.003044197568669915, + 0.0033346188720315695, + 0.0026340014301240444, + 0.003106745658442378, + 0.005556629039347172, + 0.0030127274803817272, + 0.0030206323135644197, + 0.01175979059189558, + 0.007260222919285297, + 0.008281654678285122, + 0.0038466062396764755, + 0.0038512065075337887, + 0.0027595004066824913, + 0.004727636463940144, + 0.0030890093185007572, + 0.011121872812509537, + 0.00740009406581521, + 0.008376795798540115, + 0.017826944589614868, + 0.008611558005213737, + 0.007733685430139303, + 0.010402262210845947, + 0.00845650490373373, + 0.007622077129781246, + 0.006987807806581259, + 0.010900772176682949, + 0.010846049524843693, + 0.005652387160807848, + 0.004260579124093056, + 0.007012379355728626, + 0.004918774124234915, + 0.017114128917455673, + 0.009529816918075085, + 0.1280401349067688, + 0.046755824238061905, + 0.014352654106914997, + 0.0068147145211696625, + 0.0583110973238945, + 0.01911737397313118, + 0.010070351883769035, + 0.010339071974158287, + 0.008058860898017883, + 0.06025078892707825, + 0.006359240040183067, + 0.007470716722309589, + 0.017276739701628685, + 0.056939058005809784, + 0.032387640327215195, + 0.033575333654880524, + 0.02259642444550991, + 0.01692216657102108, + 0.017623914405703545, + 0.014760816469788551, + 0.019502293318510056, + 0.01683713123202324, + 0.0058237831108272076, + 0.004916651640087366, + 0.007139130961149931, + 0.005540939979255199, + 0.005603743717074394, + 0.0039234040305018425, + 0.0065889256075024605, + 0.0036727634724229574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.0036170552484691143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810212016105652 + ], + "opposite_logprob_tokens": [ + -3.288102149963379 + ], + "logprob_reward": -0.03810212016105652, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810212016105652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05331576243042946, + 0.016627801582217216, + 0.03709989786148071, + 0.11488644778728485, + 0.031676094979047775, + 0.08787044882774353, + 0.5716181993484497, + 0.021166492253541946, + 0.027199987322092056, + 0.03853881359100342 + ], + "credit_map": [ + 0.006925265770405531, + 0.005351371597498655, + 0.005468603689223528, + 0.0046634366735816, + 0.008894776925444603, + 0.006310043856501579, + 0.005612872075289488, + 0.01008940115571022, + 0.002079129684716463, + 0.002369633410125971, + 0.0018767531728371978, + 0.001406446099281311, + 0.0022029378451406956, + 0.0025244213175028563, + 0.0019039928447455168, + 0.0022644877899438143, + 0.003778686746954918, + 0.005196771584451199, + 0.00391705147922039, + 0.003815662581473589, + 0.005566637963056564, + 0.007985231466591358, + 0.003044971264898777, + 0.003794889198616147, + 0.031707193702459335, + 0.014425314031541348, + 0.03803138807415962, + 0.0059658559039235115, + 0.0077623785473406315, + 0.0067594824358820915, + 0.005741924047470093, + 0.004492924083024263, + 0.004205088131129742, + 0.003184657311066985, + 0.00252581131644547, + 0.005814297124743462, + 0.0036187677178531885, + 0.003578264731913805, + 0.004492331761866808, + 0.004256879910826683, + 0.012103920802474022, + 0.007974274456501007, + 0.012460424564778805, + 0.017489705234766006, + 0.01113932952284813, + 0.007580381818115711, + 0.010136617347598076, + 0.008985799737274647, + 0.0332251712679863, + 0.018612507730722427, + 0.2229442596435547, + 0.1555861532688141, + 0.020343374460935593, + 0.015737144276499748, + 0.06315391510725021, + 0.04201573505997658, + 0.0027162565384060144, + 0.0028035545255988836, + 0.001918794703669846, + 0.0018297777278348804, + 0.0026705299969762564, + 0.003376623149961233, + 0.002954003633931279, + 0.002896953606978059, + 0.004474543500691652, + 0.0064436146058142185, + 0.004326797090470791, + 0.002239554189145565, + 0.00173346814699471, + 0.0015133533161133528, + 0.0034097363241016865, + 0.0030589241068810225, + 0.004488382488489151, + 0.0033579126466065645, + 0.006808633916079998, + 0.005658107344061136, + 0.004685457795858383, + 0.0060345022939145565, + 0.0040158433839678764, + 0.0034899807069450617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.03810212016105652 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic light blue convertible car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding cobblestone road at a leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car pass by whitewashed half-timbered cottages with climbing ivy and flower pots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle tracking shot and transition into a steady aerial crane rise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final view show a sun-drenched village with a stone church steeple and lush gardens in a valley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092533141374588 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092533141374588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092533141374588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07184575498104095, + 0.013482177630066872, + 0.014024716801941395, + 0.02789568528532982, + 0.04329764470458031, + 0.049943868070840836, + 0.33121398091316223, + 0.24259164929389954, + 0.1255698800086975, + 0.08013465255498886 + ], + "credit_map": [ + 0.0077192531898617744, + 0.008166945539414883, + 0.007259279023855925, + 0.005156958010047674, + 0.011054101400077343, + 0.010471771471202374, + 0.007385781034827232, + 0.01463166531175375, + 0.0016823861515149474, + 0.0015221663052216172, + 0.0012922686291858554, + 0.0012916221749037504, + 0.0021956167183816433, + 0.0014984882436692715, + 0.0015347956214100122, + 0.002464833203703165, + 0.0018192847492173314, + 0.0016727898037061095, + 0.0023870719596743584, + 0.001192500931210816, + 0.0019826488569378853, + 0.0018916625995188951, + 0.001182522508315742, + 0.0018962346948683262, + 0.0032667890191078186, + 0.003979356028139591, + 0.008234100416302681, + 0.0033631979022175074, + 0.0026940698735415936, + 0.002158166142180562, + 0.002230568090453744, + 0.0019694375805556774, + 0.005531817674636841, + 0.004505000542849302, + 0.0036725159734487534, + 0.007723307237029076, + 0.005731309298425913, + 0.006129826419055462, + 0.00476302532479167, + 0.005240841303020716, + 0.009570115245878696, + 0.004927993286401033, + 0.006233167834579945, + 0.008313530124723911, + 0.006519378628581762, + 0.004289201460778713, + 0.005918823182582855, + 0.004171660635620356, + 0.018688436597585678, + 0.009735871106386185, + 0.14139875769615173, + 0.06197698786854744, + 0.015255860053002834, + 0.008414029143750668, + 0.053221769630908966, + 0.022522250190377235, + 0.012154849246144295, + 0.01463492400944233, + 0.013233455829322338, + 0.06162780150771141, + 0.012191861867904663, + 0.01197531633079052, + 0.09455958753824234, + 0.022213835269212723, + 0.027730176225304604, + 0.022026391699910164, + 0.015804030001163483, + 0.01389815378934145, + 0.009802115149796009, + 0.008699373342096806, + 0.015753397718071938, + 0.011856247670948505, + 0.00858890451490879, + 0.007594470866024494, + 0.011974664404988289, + 0.012488953769207, + 0.010126552544534206, + 0.010916857048869133, + 0.009412376210093498, + 0.009031870402395725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.004092533141374588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_tensor([1528, 1528])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_tensor([1528, 1528])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..167c979e0fdcf5f8032d8e397e5053d6b672d293 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_tensor([1528, 1528])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, static, macro-time-lapse shot meticulously documents the accelerated life cycle of a single Ulysses plant, compressing 45 distinct flowering stages over three years into a fluid visual sequence. As a small, green bud first pokes through a patch of dark, fertile soil, the camera remains perfectly still, focused in extreme detail. It rapidly unfolds, revealing a radiant flower with an intricate array of vibrant yellow petals. The time-lapse effect is striking, with the background light rapidly flickering, simulating the passage of countless days and nights as the flower blooms, gracefully withers, and re-blooms in a continuous, hypnotic cycle. In the final stage, the yellow petals shrink, miraculously replaced by a perfect, ethereal sphere of fluffy white seeds, each a delicate parachute with a tiny seed at its base, seemingly poised and ready to be dispersed by the breeze. The entire natural drama is set against a stark, contemplative backdrop: a vertical scroll crafted from ancient black paper, adorned with elegant hand-painted Eastern script, its black ink brushstrokes creating a striking contrast to the plant's organic evolution. The controlled lighting subtly changes as the “seasons” pass, casting dynamic shadows that highlight the delicate textures of pencil and calligraphy, creating a tranquil, philosophical atmosphere that contemplates the beauty of the fleeting life cycle.", + "seed": 25568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_1528_seed_25568_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06316932723857463, + "vlm_logprob": -0.06316932723857463, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.06316932723857463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030465511605143547 + ], + "opposite_logprob_tokens": [ + -3.905465602874756 + ], + "logprob_reward": -0.030465511605143547, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.030465511605143547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08778367191553116, + 0.08372660726308823, + 0.17775858938694, + 0.035978276282548904, + 0.052888672798871994, + 0.12854476273059845, + 0.07169678807258606, + 0.15213236212730408, + 0.10612917691469193, + 0.1033610925078392 + ], + "credit_map": [ + 0.010990869253873825, + 0.013100319541990757, + 0.008819586597383022, + 0.012595929205417633, + 0.014656431041657925, + 0.00826740637421608, + 0.013292377814650536, + 0.006060764659196138, + 0.007020759396255016, + 0.01873846724629402, + 0.006817290559411049, + 0.0053288317285478115, + 0.007966889999806881, + 0.02865787222981453, + 0.004579651169478893, + 0.004616857971996069, + 0.022237764671444893, + 0.06501521915197372, + 0.011563071981072426, + 0.010241106152534485, + 0.007477637846022844, + 0.048062555491924286, + 0.006088068708777428, + 0.007073175627738237, + 0.003868025029078126, + 0.007859136909246445, + 0.0026845172978937626, + 0.003807314671576023, + 0.002356182783842087, + 0.007108630146831274, + 0.004484390839934349, + 0.0038100844249129295, + 0.006268363445997238, + 0.007864355109632015, + 0.004472511354833841, + 0.005132820922881365, + 0.004557895008474588, + 0.009769974276423454, + 0.011636861599981785, + 0.003185891779139638, + 0.011788233183324337, + 0.04369538277387619, + 0.007540938910096884, + 0.005279128439724445, + 0.006822862662374973, + 0.04345853254199028, + 0.004085710272192955, + 0.005873980466276407, + 0.008158651180565357, + 0.02028035745024681, + 0.0064076632261276245, + 0.004291257821023464, + 0.005267088767141104, + 0.018696989864110947, + 0.00448232376947999, + 0.00411246670410037, + 0.04885093867778778, + 0.05919038504362106, + 0.009488826617598534, + 0.005450618453323841, + 0.004878027830272913, + 0.01237958762794733, + 0.007995967753231525, + 0.003898013150319457, + 0.011911125853657722, + 0.02401919476687908, + 0.008534243330359459, + 0.009567022323608398, + 0.011578293517231941, + 0.026807870715856552, + 0.007436017971485853, + 0.0062754168175160885, + 0.016318969428539276, + 0.02650446444749832, + 0.011304458603262901, + 0.012470035813748837, + 0.006489874795079231, + 0.014785898849368095, + 0.005834249313920736, + 0.009653153829276562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.030465511605143547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004668526817113161 + ], + "opposite_logprob_tokens": [ + -5.379668712615967 + ], + "logprob_reward": -0.004668526817113161, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004668526817113161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10447266697883606, + 0.07189081609249115, + 0.23189598321914673, + 0.06713155657052994, + 0.07016751915216446, + 0.06387341767549515, + 0.057826969772577286, + 0.21518240869045258, + 0.062341365963220596, + 0.055217258632183075 + ], + "credit_map": [ + 0.012847875244915485, + 0.016049399971961975, + 0.01198266725987196, + 0.012009904719889164, + 0.015791382640600204, + 0.012368151918053627, + 0.015290814451873302, + 0.00813247635960579, + 0.00664743036031723, + 0.01784835383296013, + 0.007545196916908026, + 0.005350747611373663, + 0.007381293922662735, + 0.018367642536759377, + 0.005086314398795366, + 0.003663835348561406, + 0.017472602427005768, + 0.07486056536436081, + 0.015658410266041756, + 0.010851673781871796, + 0.009838801808655262, + 0.08441536873579025, + 0.005289262626320124, + 0.01350930891931057, + 0.005727922543883324, + 0.02278207615017891, + 0.004450883716344833, + 0.0037831945810467005, + 0.002557244384661317, + 0.02134547010064125, + 0.0022147682029753923, + 0.004269997123628855, + 0.006462969817221165, + 0.013532229699194431, + 0.005301144439727068, + 0.006154241506010294, + 0.005706483963876963, + 0.014811713248491287, + 0.014278335496783257, + 0.00392040004953742, + 0.006956073921173811, + 0.01994599960744381, + 0.0037261631805449724, + 0.005038475152105093, + 0.003913456108421087, + 0.015681536868214607, + 0.006238915957510471, + 0.0023727980442345142, + 0.004907109308987856, + 0.014570633880794048, + 0.006268943659961224, + 0.004786392208188772, + 0.003890062216669321, + 0.0150382649153471, + 0.00459085451439023, + 0.0037747106980532408, + 0.06904364377260208, + 0.08480857312679291, + 0.013215094804763794, + 0.007305994164198637, + 0.007101686205714941, + 0.018079055473208427, + 0.01041245274245739, + 0.005215901415795088, + 0.01429374236613512, + 0.006538389250636101, + 0.005832154769450426, + 0.00309958984144032, + 0.004547053948044777, + 0.02197515033185482, + 0.002633770229294896, + 0.00342151359654963, + 0.008297242224216461, + 0.016373131424188614, + 0.006522375158965588, + 0.006151578854769468, + 0.0034964594524353743, + 0.005247524939477444, + 0.003418552689254284, + 0.005710393190383911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.004668526817113161 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635536715388298 + ], + "opposite_logprob_tokens": [ + -3.651355266571045 + ], + "logprob_reward": -0.02635536715388298, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635536715388298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2146008014678955, + 0.0850299745798111, + 0.20852814614772797, + 0.03105812519788742, + 0.010185911320149899, + 0.05241016671061516, + 0.02064591646194458, + 0.11181167513132095, + 0.09149967133998871, + 0.17422956228256226 + ], + "credit_map": [ + 0.020507633686065674, + 0.036041438579559326, + 0.02059505693614483, + 0.020148679614067078, + 0.060973282903432846, + 0.03180002421140671, + 0.015081646852195263, + 0.009453026577830315, + 0.008562172763049603, + 0.03102163039147854, + 0.005880843847990036, + 0.0050328560173511505, + 0.003199225291609764, + 0.026046287268400192, + 0.0019643446430563927, + 0.003322612028568983, + 0.01637968234717846, + 0.07114730030298233, + 0.013652562163770199, + 0.009831998497247696, + 0.00866719614714384, + 0.07345529645681381, + 0.004658633377403021, + 0.010735477320849895, + 0.0020953170023858547, + 0.008581622503697872, + 0.0022213137708604336, + 0.00235068635083735, + 0.0013144247932359576, + 0.010308288037776947, + 0.001268342835828662, + 0.0029181300196796656, + 0.0015287382993847132, + 0.0018428119365125895, + 0.0015716621419414878, + 0.0011493860511109233, + 0.0011612335219979286, + 0.0012684018583968282, + 0.0009562012855894864, + 0.0007074762252159417, + 0.0044746254570782185, + 0.017179418355226517, + 0.0032761958427727222, + 0.00232783448882401, + 0.0024305423721671104, + 0.01878548413515091, + 0.0016438598977401853, + 0.002292205346748233, + 0.002178995404392481, + 0.005815816577523947, + 0.0020005556289106607, + 0.0013087569968774915, + 0.0012089470401406288, + 0.0052755665965378284, + 0.0015659702476114035, + 0.0012913089012727141, + 0.036267682909965515, + 0.04399840533733368, + 0.0068262857384979725, + 0.003908140584826469, + 0.0033285540994256735, + 0.009471666999161243, + 0.005236013326793909, + 0.002774924971163273, + 0.024090677499771118, + 0.009596750140190125, + 0.006715924944728613, + 0.0038703978061676025, + 0.004728367086499929, + 0.03415483608841896, + 0.0032577570527791977, + 0.005084962118417025, + 0.0271749310195446, + 0.05061110481619835, + 0.01793714612722397, + 0.01857980154454708, + 0.00945435930043459, + 0.026204315945506096, + 0.00894448533654213, + 0.015323424711823463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02635536715388298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215262174606323 + ], + "opposite_logprob_tokens": [ + -1.502152681350708 + ], + "logprob_reward": -0.25215262174606323, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215262174606323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044805604964494705, + 0.17696446180343628, + 0.3189394474029541, + 0.0502287782728672, + 0.025126272812485695, + 0.08856118470430374, + 0.03235895559191704, + 0.18617519736289978, + 0.0336025208234787, + 0.04323757812380791 + ], + "credit_map": [ + 0.0054553356021642685, + 0.0066420757211744785, + 0.006829528138041496, + 0.0065552531741559505, + 0.008530919440090656, + 0.0039007761515676975, + 0.004014512524008751, + 0.0028772091027349234, + 0.0158543698489666, + 0.06351666152477264, + 0.01242250856012106, + 0.01005382277071476, + 0.006748528685420752, + 0.05797717347741127, + 0.003468017326667905, + 0.00692339800298214, + 0.025005828589200974, + 0.10699083656072617, + 0.02125431038439274, + 0.015047372318804264, + 0.013384209014475346, + 0.11401612311601639, + 0.006443276070058346, + 0.016797497868537903, + 0.0040507749654352665, + 0.01764938049018383, + 0.003761530388146639, + 0.0027931530494242907, + 0.0017482733819633722, + 0.01658397726714611, + 0.0011591368820518255, + 0.00248255324549973, + 0.0041339504532516, + 0.005344045348465443, + 0.0030300109647214413, + 0.0033406687434762716, + 0.002225079806521535, + 0.0031078483443707228, + 0.0016656530788168311, + 0.0022790164221078157, + 0.007884059101343155, + 0.02973603643476963, + 0.004759511910378933, + 0.00373985362239182, + 0.004408266860991716, + 0.030928974971175194, + 0.0035488505382090807, + 0.0035556412767618895, + 0.0035583453718572855, + 0.00853525660932064, + 0.00406049657613039, + 0.0040151337161660194, + 0.002617819933220744, + 0.004811042919754982, + 0.0021253409795463085, + 0.0026355222798883915, + 0.05882631987333298, + 0.07296484708786011, + 0.011661847122013569, + 0.007046385668218136, + 0.005897169467061758, + 0.016165077686309814, + 0.008790334686636925, + 0.004823233466595411, + 0.0041782446205616, + 0.009298441000282764, + 0.0054711271077394485, + 0.002860704204067588, + 0.0020999072585254908, + 0.005080605391412973, + 0.001718784449622035, + 0.0028947116807103157, + 0.006511372048407793, + 0.012655822560191154, + 0.006671905983239412, + 0.004718597047030926, + 0.0019768693018704653, + 0.004315584897994995, + 0.002209216821938753, + 0.004178214818239212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.25215262174606323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022046088706701994 + ], + "opposite_logprob_tokens": [ + -6.127204418182373 + ], + "logprob_reward": -0.0022046088706701994, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022046088706701994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10287067294120789, + 0.3515082001686096, + 0.04553065821528435, + 0.02834155224263668, + 0.01220536045730114, + 0.040495939552783966, + 0.019057447090744972, + 0.16081511974334717, + 0.21745648980140686, + 0.021718548610806465 + ], + "credit_map": [ + 0.00989302434027195, + 0.019242003560066223, + 0.010155517607927322, + 0.00926119089126587, + 0.01271597109735012, + 0.017631549388170242, + 0.015538114123046398, + 0.008433304727077484, + 0.029201652854681015, + 0.12663821876049042, + 0.023185206577181816, + 0.01900876685976982, + 0.014619685709476471, + 0.11838874220848083, + 0.006866407580673695, + 0.013599511235952377, + 0.007727606222033501, + 0.015221253037452698, + 0.004156897310167551, + 0.003982695750892162, + 0.0023896710481494665, + 0.006144626997411251, + 0.002955901902168989, + 0.0029520064126700163, + 0.002162944758310914, + 0.008792169392108917, + 0.0019069185946136713, + 0.001677182735875249, + 0.0013813163386657834, + 0.008988809771835804, + 0.0018239927012473345, + 0.0016082163201645017, + 0.0021293226163834333, + 0.001825542887672782, + 0.001569259213283658, + 0.001435950631275773, + 0.0010409276001155376, + 0.0016380669549107552, + 0.0015379569958895445, + 0.0010283337906002998, + 0.00499566039070487, + 0.01300852932035923, + 0.0024461408611387014, + 0.002180274575948715, + 0.003179374849423766, + 0.010403892025351524, + 0.00260297954082489, + 0.001679086359217763, + 0.0020560009870678186, + 0.00520359817892313, + 0.0024430325720459223, + 0.002093741437420249, + 0.0011939851101487875, + 0.002842104993760586, + 0.0016792104579508305, + 0.0015457732370123267, + 0.05140310153365135, + 0.0635436400771141, + 0.009685914032161236, + 0.006001120898872614, + 0.004884954076260328, + 0.013658059760928154, + 0.0076806372962892056, + 0.003957698121666908, + 0.05500726029276848, + 0.02675149217247963, + 0.015935327857732773, + 0.007954730652272701, + 0.013077403418719769, + 0.08001741766929626, + 0.00816277414560318, + 0.010550092905759811, + 0.004687751643359661, + 0.004826503340154886, + 0.0021177753806114197, + 0.0016594150802120566, + 0.002207221230491996, + 0.0033806192222982645, + 0.0017987515311688185, + 0.0010405114153400064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.0022046088706701994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, macro-time-lapse shot meticulously documents the accelerated life cycle of a single Ulysses plant, compressing 45 distinct flowering stages over three years into a fluid visual sequence. As a small, green bud first pokes through a patch of dark, fertile soil, the camera remains perfectly still, focused in extreme detail. It rapidly unfolds, revealing a radiant flower with an intricate array of vibrant yellow petals. The time-lapse effect is striking, with the background light rapidly flickering, simulating the passage of countless days and nights as the flower blooms, gracefully withers, and re-blooms in a continuous, hypnotic cycle. In the final stage, the yellow petals shrink, miraculously replaced by a perfect, ethereal sphere of fluffy white seeds, each a delicate parachute with a tiny seed at its base, seemingly poised and ready to be dispersed by the breeze. The entire natural drama is set against a stark, contemplative backdrop: a vertical scroll crafted from ancient black paper, adorned with elegant hand-painted Eastern script, its black ink brushstrokes creating a striking contrast to the plant's organic evolution. The controlled lighting subtly changes as the “seasons” pass, casting dynamic shadows that highlight the delicate textures of pencil and calligraphy, creating a tranquil, philosophical atmosphere that contemplates the beauty of the fleeting life cycle.", + "seed": 25569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_1528_seed_25569_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07078575047198683, + "vlm_logprob": -0.07078575047198683, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.07078575047198683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026586048305034637 + ], + "opposite_logprob_tokens": [ + -3.9015860557556152 + ], + "logprob_reward": -0.026586048305034637, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026586048305034637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14422626793384552, + 0.06546148657798767, + 0.1015925258398056, + 0.0660020187497139, + 0.06961268931627274, + 0.09843681752681732, + 0.06216207891702652, + 0.04085467755794525, + 0.2721952199935913, + 0.07945618778467178 + ], + "credit_map": [ + 0.014774695970118046, + 0.022616855800151825, + 0.01991840824484825, + 0.022911436855793, + 0.011991825886070728, + 0.01527810376137495, + 0.02016054466366768, + 0.016574399545788765, + 0.008352132514119148, + 0.00793372094631195, + 0.01073798630386591, + 0.010960530489683151, + 0.0049901544116437435, + 0.007621848024427891, + 0.007685405667871237, + 0.007179711014032364, + 0.01746823638677597, + 0.01198651734739542, + 0.013454992324113846, + 0.02252846211194992, + 0.006296394858509302, + 0.01013866439461708, + 0.010809042491018772, + 0.008910213597118855, + 0.010008499026298523, + 0.006621686276048422, + 0.010258764028549194, + 0.008001940324902534, + 0.00537943746894598, + 0.008560430258512497, + 0.004793884232640266, + 0.012377380393445492, + 0.00575444009155035, + 0.010261094197630882, + 0.007373871747404337, + 0.01318247802555561, + 0.0084230313077569, + 0.006625750567764044, + 0.009719736874103546, + 0.008272279053926468, + 0.010539163835346699, + 0.009902654215693474, + 0.01433480903506279, + 0.015155589208006859, + 0.008669858798384666, + 0.009032892994582653, + 0.019614554941654205, + 0.0111872972920537, + 0.0055450438521802425, + 0.007122705224901438, + 0.007473778910934925, + 0.010970095172524452, + 0.006832638289779425, + 0.007737055886536837, + 0.00853681843727827, + 0.007943939417600632, + 0.004557866603136063, + 0.0026908821891993284, + 0.002669976558536291, + 0.004246526397764683, + 0.004770880565047264, + 0.008508577942848206, + 0.00433802604675293, + 0.00907194148749113, + 0.02956039272248745, + 0.021067481487989426, + 0.02019372768700123, + 0.024817971512675285, + 0.027377795428037643, + 0.05963259935379028, + 0.04342009127140045, + 0.04612518101930618, + 0.008118847385048866, + 0.013118259608745575, + 0.004847514908760786, + 0.006929130293428898, + 0.009563091211020947, + 0.013089067302644253, + 0.00828400906175375, + 0.015506275929510593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.026586048305034637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003637247486039996 + ], + "opposite_logprob_tokens": [ + -5.628637313842773 + ], + "logprob_reward": -0.003637247486039996, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003637247486039996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1826915591955185, + 0.09166748076677322, + 0.11847684532403946, + 0.11851987987756729, + 0.07163634151220322, + 0.1034989133477211, + 0.05182866007089615, + 0.06762882322072983, + 0.13480687141418457, + 0.05924471095204353 + ], + "credit_map": [ + 0.026640955358743668, + 0.02726886421442032, + 0.030568979680538177, + 0.02380507066845894, + 0.014790467917919159, + 0.022473080083727837, + 0.01761840097606182, + 0.019525719806551933, + 0.010094398632645607, + 0.014741484075784683, + 0.013472337275743484, + 0.014574910514056683, + 0.007110736332833767, + 0.010222681798040867, + 0.010874059982597828, + 0.010576867498457432, + 0.024666525423526764, + 0.01346631906926632, + 0.014790697023272514, + 0.01762029156088829, + 0.009751883335411549, + 0.014231951907277107, + 0.012286094948649406, + 0.011663071811199188, + 0.013304712250828743, + 0.01751100644469261, + 0.01668257638812065, + 0.018612729385495186, + 0.01364101842045784, + 0.011899551376700401, + 0.012023279443383217, + 0.014845001511275768, + 0.007490017917007208, + 0.006908741779625416, + 0.011576025746762753, + 0.009083777666091919, + 0.007368335034698248, + 0.009103297255933285, + 0.010716662742197514, + 0.009389478713274002, + 0.010178416036069393, + 0.01080743782222271, + 0.014204657636582851, + 0.0159209705889225, + 0.007682075724005699, + 0.012081134133040905, + 0.019357968121767044, + 0.0132662458345294, + 0.004984403029084206, + 0.006354468408972025, + 0.0066239116713404655, + 0.010536817833781242, + 0.004245005548000336, + 0.005757610313594341, + 0.006312937941402197, + 0.007013501599431038, + 0.0077186468988657, + 0.006574440281838179, + 0.005664438009262085, + 0.01106971688568592, + 0.006587868090718985, + 0.010805356316268444, + 0.010936817154288292, + 0.008271532133221626, + 0.01394707802683115, + 0.011680339463055134, + 0.009108657017350197, + 0.011426955461502075, + 0.015507969073951244, + 0.029247377067804337, + 0.020983360707759857, + 0.022905128076672554, + 0.006773216184228659, + 0.00811429787427187, + 0.0050690146163105965, + 0.006132783368229866, + 0.007450404576957226, + 0.011240352876484394, + 0.005951520521193743, + 0.008513111621141434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.003637247486039996 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002059370279312 + ], + "opposite_logprob_tokens": [ + -2.6950206756591797 + ], + "logprob_reward": -0.07002059370279312, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002059370279312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27473556995391846, + 0.20023320615291595, + 0.07041231542825699, + 0.04254744201898575, + 0.04943454638123512, + 0.0645381286740303, + 0.04696998745203018, + 0.0731600672006607, + 0.1448141187429428, + 0.03315458819270134 + ], + "credit_map": [ + 0.03508149832487106, + 0.04427776485681534, + 0.043260443955659866, + 0.01616905815899372, + 0.03201426565647125, + 0.04026380553841591, + 0.03382197394967079, + 0.0298467930406332, + 0.035658273845911026, + 0.01510531734675169, + 0.03458721935749054, + 0.013505624607205391, + 0.021625975146889687, + 0.03545423224568367, + 0.02313467115163803, + 0.021161898970603943, + 0.01093523669987917, + 0.005518224555999041, + 0.011265350505709648, + 0.0059852441772818565, + 0.007743744179606438, + 0.01245298981666565, + 0.0075134821236133575, + 0.008998037315905094, + 0.005544091109186411, + 0.004427156411111355, + 0.006125087384134531, + 0.004442793782800436, + 0.004260287620127201, + 0.00737498514354229, + 0.0042207506485283375, + 0.006152292247861624, + 0.006383979227393866, + 0.003876032307744026, + 0.006610543932765722, + 0.005202526692301035, + 0.004186473321169615, + 0.010670493356883526, + 0.004430828616023064, + 0.008073665201663971, + 0.012192186899483204, + 0.004017617553472519, + 0.0066794403828680515, + 0.005251266062259674, + 0.004721757955849171, + 0.017068423330783844, + 0.006398976314812899, + 0.008208462968468666, + 0.006286830175668001, + 0.0045349556021392345, + 0.006270568817853928, + 0.00638343533501029, + 0.004219449125230312, + 0.007582715712487698, + 0.005247294902801514, + 0.006444737780839205, + 0.00923002976924181, + 0.00564293097704649, + 0.004752709995955229, + 0.004883633926510811, + 0.006752328481525183, + 0.017671840265393257, + 0.006272824015468359, + 0.017953766509890556, + 0.01629839465022087, + 0.01164002250880003, + 0.009550817310810089, + 0.01366486120969057, + 0.016504133120179176, + 0.03484850749373436, + 0.02180544286966324, + 0.020501943305134773, + 0.004718058276921511, + 0.004041918553411961, + 0.003453065874055028, + 0.0029335685539990664, + 0.004144630860537291, + 0.005791602656245232, + 0.0032997566740959883, + 0.004771985579282045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.07002059370279312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25217002630233765 + ], + "opposite_logprob_tokens": [ + -1.5021700859069824 + ], + "logprob_reward": -0.25217002630233765, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25217002630233765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12865296006202698, + 0.050467561930418015, + 0.07079996913671494, + 0.07682894915342331, + 0.041820038110017776, + 0.08183178305625916, + 0.06296753883361816, + 0.36718496680259705, + 0.08960165828466415, + 0.02984461560845375 + ], + "credit_map": [ + 0.01656986214220524, + 0.01830379292368889, + 0.017313746735453606, + 0.017905771732330322, + 0.012284724041819572, + 0.014009999111294746, + 0.01843494363129139, + 0.01383011881262064, + 0.007126776501536369, + 0.004724306054413319, + 0.009015456773340702, + 0.004699832759797573, + 0.006654537748545408, + 0.006108936388045549, + 0.005476794205605984, + 0.006660922430455685, + 0.015562392771244049, + 0.008103109896183014, + 0.00943121500313282, + 0.008796839974820614, + 0.005117764230817556, + 0.008836745284497738, + 0.00760727608576417, + 0.007344622630625963, + 0.008961654268205166, + 0.009419016540050507, + 0.009454622864723206, + 0.014115403406322002, + 0.006697147618979216, + 0.008875511586666107, + 0.007766348775476217, + 0.011539248749613762, + 0.004889014642685652, + 0.004119446966797113, + 0.004590435419231653, + 0.00477159209549427, + 0.0037393055390566587, + 0.009070169180631638, + 0.0039015563670545816, + 0.006738518830388784, + 0.009379979223012924, + 0.010239762254059315, + 0.010361537337303162, + 0.011669452302157879, + 0.008276783861219883, + 0.010373787954449654, + 0.012491910718381405, + 0.009038571268320084, + 0.006660205312073231, + 0.0062572224996984005, + 0.008397402241826057, + 0.00932830199599266, + 0.0059967683628201485, + 0.009061803109943867, + 0.008602093905210495, + 0.008663749322295189, + 0.0355093777179718, + 0.04080977663397789, + 0.03401784971356392, + 0.047040387988090515, + 0.040351759642362595, + 0.06600908190011978, + 0.0550369992852211, + 0.04840973764657974, + 0.009399360977113247, + 0.010155027732253075, + 0.005848005414009094, + 0.008754691109061241, + 0.012684543617069721, + 0.020301612094044685, + 0.011521218344569206, + 0.010937202721834183, + 0.003970500081777573, + 0.0033022158313542604, + 0.0016345131443813443, + 0.00201144115999341, + 0.004584671929478645, + 0.007287371903657913, + 0.002957145217806101, + 0.004096756223589182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.25217002630233765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015148365637287498 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015148365637287498, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015148365637287498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11763429641723633, + 0.029438866302371025, + 0.054183755069971085, + 0.08450564742088318, + 0.06804561614990234, + 0.09585206210613251, + 0.04975153133273125, + 0.05844610184431076, + 0.2853138744831085, + 0.15682823956012726 + ], + "credit_map": [ + 0.013715659268200397, + 0.018009284511208534, + 0.01498623751103878, + 0.01653030887246132, + 0.01121527235955, + 0.013483225367963314, + 0.016658209264278412, + 0.013036112301051617, + 0.003165705595165491, + 0.003284658771008253, + 0.004251588135957718, + 0.003844638355076313, + 0.0026879378128796816, + 0.006176600232720375, + 0.0029986693989485502, + 0.0030290705617517233, + 0.007524593733251095, + 0.007433328311890364, + 0.00744083896279335, + 0.010274513624608517, + 0.003966028802096844, + 0.006982938852161169, + 0.0060180858708918095, + 0.004543432500213385, + 0.008351952768862247, + 0.012029552832245827, + 0.010804818943142891, + 0.01711389049887657, + 0.009736282750964165, + 0.009212613105773926, + 0.009618948213756084, + 0.007637588307261467, + 0.00695387739688158, + 0.008643059991300106, + 0.008278745226562023, + 0.010862750001251698, + 0.007326358463615179, + 0.008059728890657425, + 0.009170269593596458, + 0.00875082891434431, + 0.009937863796949387, + 0.00969194620847702, + 0.011371860280632973, + 0.014814290218055248, + 0.009188448078930378, + 0.008138206787407398, + 0.02146793156862259, + 0.011241518892347813, + 0.0041368938982486725, + 0.005936193745583296, + 0.005636034999042749, + 0.010822576470673084, + 0.005357713904231787, + 0.005155050661414862, + 0.006482949946075678, + 0.006224121432751417, + 0.006288410164415836, + 0.0035659705754369497, + 0.004359189420938492, + 0.008127176202833652, + 0.005134160164743662, + 0.009684748016297817, + 0.008233631029725075, + 0.013052823022007942, + 0.03255091980099678, + 0.02070239745080471, + 0.021940352395176888, + 0.02728460542857647, + 0.028218723833560944, + 0.06217273324728012, + 0.045383159071207047, + 0.0470610037446022, + 0.016273649409413338, + 0.030877644196152687, + 0.006564910523593426, + 0.011570432223379612, + 0.018674882128834724, + 0.02040707692503929, + 0.014438224025070667, + 0.03802142292261124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0015148365637287498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, macro-time-lapse shot meticulously documents the accelerated life cycle of a single Ulysses plant, compressing 45 distinct flowering stages over three years into a fluid visual sequence. As a small, green bud first pokes through a patch of dark, fertile soil, the camera remains perfectly still, focused in extreme detail. It rapidly unfolds, revealing a radiant flower with an intricate array of vibrant yellow petals. The time-lapse effect is striking, with the background light rapidly flickering, simulating the passage of countless days and nights as the flower blooms, gracefully withers, and re-blooms in a continuous, hypnotic cycle. In the final stage, the yellow petals shrink, miraculously replaced by a perfect, ethereal sphere of fluffy white seeds, each a delicate parachute with a tiny seed at its base, seemingly poised and ready to be dispersed by the breeze. The entire natural drama is set against a stark, contemplative backdrop: a vertical scroll crafted from ancient black paper, adorned with elegant hand-painted Eastern script, its black ink brushstrokes creating a striking contrast to the plant's organic evolution. The controlled lighting subtly changes as the “seasons” pass, casting dynamic shadows that highlight the delicate textures of pencil and calligraphy, creating a tranquil, philosophical atmosphere that contemplates the beauty of the fleeting life cycle.", + "seed": 25570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_1528_seed_25570_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05668552778661251, + "vlm_logprob": -0.05668552778661251, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.05668552778661251, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016304366290569305 + ], + "opposite_logprob_tokens": [ + -4.7663044929504395 + ], + "logprob_reward": -0.016304366290569305, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.016304366290569305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21030908823013306, + 0.07572416961193085, + 0.06590403616428375, + 0.08834732323884964, + 0.0854693278670311, + 0.08925549685955048, + 0.13013558089733124, + 0.11919281631708145, + 0.0443318672478199, + 0.09133030474185944 + ], + "credit_map": [ + 0.019726654514670372, + 0.03403744846582413, + 0.028766002506017685, + 0.02380499616265297, + 0.039459872990846634, + 0.025888103991746902, + 0.021463697776198387, + 0.017162291333079338, + 0.0069004627875983715, + 0.009671664796769619, + 0.007310446817427874, + 0.0067697386257350445, + 0.014081616885960102, + 0.011854813434183598, + 0.011873224750161171, + 0.007262200117111206, + 0.006539836525917053, + 0.008711026981472969, + 0.007054295390844345, + 0.008854097686707973, + 0.013192481361329556, + 0.008007226511836052, + 0.007612199988216162, + 0.005932867527008057, + 0.01272420585155487, + 0.011634325608611107, + 0.012903345748782158, + 0.01022946834564209, + 0.008285580202937126, + 0.010203525424003601, + 0.00812217965722084, + 0.014244697988033295, + 0.013484081253409386, + 0.01020190492272377, + 0.0120247146114707, + 0.009390447288751602, + 0.011842945590615273, + 0.010964122600853443, + 0.008685302920639515, + 0.008875813335180283, + 0.017752615734934807, + 0.008963379077613354, + 0.012838221155107021, + 0.0070174820721149445, + 0.0072990781627595425, + 0.010446674190461636, + 0.009825089946389198, + 0.01511295884847641, + 0.01997474953532219, + 0.01247574295848608, + 0.013914707116782665, + 0.022662637755274773, + 0.013525797985494137, + 0.017873356118798256, + 0.01873409003019333, + 0.010974482633173466, + 0.01853228360414505, + 0.011093239299952984, + 0.016636118292808533, + 0.017524078488349915, + 0.009570574387907982, + 0.014420464634895325, + 0.0178169347345829, + 0.013599121943116188, + 0.007194475270807743, + 0.0052618952468037605, + 0.006115516647696495, + 0.004247040953487158, + 0.0038395195733755827, + 0.004829734563827515, + 0.006390080787241459, + 0.0064536044374108315, + 0.01688666082918644, + 0.009292807430028915, + 0.012767201289534569, + 0.009184039197862148, + 0.011125960387289524, + 0.00972831528633833, + 0.011371285654604435, + 0.010974042117595673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.016304366290569305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002841247245669365 + ], + "opposite_logprob_tokens": [ + -5.877841472625732 + ], + "logprob_reward": -0.002841247245669365, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002841247245669365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2530612051486969, + 0.1077055037021637, + 0.15485364198684692, + 0.06513971835374832, + 0.05178873613476753, + 0.12375647574663162, + 0.09110968559980392, + 0.08901238441467285, + 0.02607928030192852, + 0.03749333694577217 + ], + "credit_map": [ + 0.0372384712100029, + 0.045049600303173065, + 0.024235492572188377, + 0.020106690004467964, + 0.04301225394010544, + 0.03847884014248848, + 0.025694390758872032, + 0.019245458766818047, + 0.012816553935408592, + 0.024000409990549088, + 0.011182373389601707, + 0.008554788306355476, + 0.01343611441552639, + 0.01690131053328514, + 0.011938110925257206, + 0.008875847794115543, + 0.019457982853055, + 0.020588768646121025, + 0.02048504166305065, + 0.018885739147663116, + 0.018982958048582077, + 0.019402872771024704, + 0.018305066972970963, + 0.01874522864818573, + 0.008389243856072426, + 0.007267683744430542, + 0.009791899472475052, + 0.006830189842730761, + 0.007895608432590961, + 0.007479853928089142, + 0.006627685856074095, + 0.010857556015253067, + 0.009822464548051357, + 0.00517011433839798, + 0.008869307115674019, + 0.00622053537517786, + 0.005140726920217276, + 0.005781481042504311, + 0.0051101818680763245, + 0.005673928186297417, + 0.016501357778906822, + 0.009686743840575218, + 0.023638593032956123, + 0.014163728803396225, + 0.01224448811262846, + 0.01582094095647335, + 0.010674234479665756, + 0.021026398986577988, + 0.01734965108335018, + 0.006568565033376217, + 0.01037853304296732, + 0.010155368596315384, + 0.00977703370153904, + 0.012295819818973541, + 0.013142731972038746, + 0.011441980488598347, + 0.011197746731340885, + 0.00740728760138154, + 0.01342816837131977, + 0.013968507759273052, + 0.008053857833147049, + 0.010534035041928291, + 0.014416702091693878, + 0.010006091557443142, + 0.003556982148438692, + 0.0029853086452931166, + 0.0032987140584737062, + 0.003457307117059827, + 0.0024140202440321445, + 0.0034598405472934246, + 0.003749306546524167, + 0.0031578033231198788, + 0.006264780182391405, + 0.004133853130042553, + 0.004897479433566332, + 0.003727615112438798, + 0.005671184044331312, + 0.004077594727277756, + 0.004646838176995516, + 0.00407399283722043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.002841247245669365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080282740294933 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080282740294933, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080282740294933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5832068920135498, + 0.06856232136487961, + 0.1379343867301941, + 0.05064234882593155, + 0.039293982088565826, + 0.02643515169620514, + 0.02216177247464657, + 0.03670480474829674, + 0.014027790166437626, + 0.021030591800808907 + ], + "credit_map": [ + 0.08383423835039139, + 0.12653757631778717, + 0.06568550318479538, + 0.04101158306002617, + 0.10145407170057297, + 0.10073596239089966, + 0.03206157684326172, + 0.031886350363492966, + 0.007363990880548954, + 0.014109198935329914, + 0.007121676113456488, + 0.006545572075992823, + 0.009210101328790188, + 0.00907403975725174, + 0.00941462628543377, + 0.005723120179027319, + 0.024802708998322487, + 0.02830197662115097, + 0.016179265454411507, + 0.01142640970647335, + 0.01070667989552021, + 0.028942210599780083, + 0.006837786640971899, + 0.010737357661128044, + 0.009382897056639194, + 0.004998030606657267, + 0.008953546173870564, + 0.005358855705708265, + 0.0038911853916943073, + 0.005959284491837025, + 0.004088452085852623, + 0.00801009964197874, + 0.01178198866546154, + 0.003207329660654068, + 0.00595990102738142, + 0.0030645274091511965, + 0.0021211933344602585, + 0.005192236043512821, + 0.0027504439931362867, + 0.005216359626501799, + 0.006399321369826794, + 0.0026325881481170654, + 0.003322751261293888, + 0.0020909816958010197, + 0.0024572405964136124, + 0.003551530884578824, + 0.0027561327442526817, + 0.0032246061600744724, + 0.002566021401435137, + 0.002584635280072689, + 0.002787540899589658, + 0.003217523219063878, + 0.0026181982830166817, + 0.0028561358340084553, + 0.003282133024185896, + 0.0022495838347822428, + 0.00716344453394413, + 0.0030990815721452236, + 0.006363683845847845, + 0.004022964276373386, + 0.00390365207567811, + 0.004971514455974102, + 0.002910447306931019, + 0.004270017612725496, + 0.0030606803484261036, + 0.00114563072565943, + 0.0015839786501601338, + 0.0010625311406329274, + 0.001246825442649424, + 0.0017421228112652898, + 0.0017492173938080668, + 0.0024368036538362503, + 0.0034053476992994547, + 0.002098631113767624, + 0.002807780634611845, + 0.00218517635948956, + 0.0023164297454059124, + 0.0022744801826775074, + 0.002654116600751877, + 0.003288628999143839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.011080282740294933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215622782707214 + ], + "opposite_logprob_tokens": [ + -1.5021562576293945 + ], + "logprob_reward": -0.25215622782707214, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215622782707214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2943941354751587, + 0.11180419474840164, + 0.13049444556236267, + 0.08542473614215851, + 0.061155471950769424, + 0.0762159675359726, + 0.07871183007955551, + 0.08068908751010895, + 0.026842359453439713, + 0.054267801344394684 + ], + "credit_map": [ + 0.032285433262586594, + 0.05660215765237808, + 0.023585621267557144, + 0.027766192331910133, + 0.07161888480186462, + 0.0384419783949852, + 0.026091773062944412, + 0.018002063035964966, + 0.016574548557400703, + 0.019535500556230545, + 0.013606561347842216, + 0.01041441224515438, + 0.012567232362926006, + 0.01368308812379837, + 0.016526425257325172, + 0.008896423503756523, + 0.015678631141781807, + 0.017347142100334167, + 0.02032085508108139, + 0.014407569542527199, + 0.015430569648742676, + 0.015526728704571724, + 0.014288360252976418, + 0.017494579777121544, + 0.013557223603129387, + 0.015183844603598118, + 0.010637883096933365, + 0.010840523056685925, + 0.007976297289133072, + 0.008761918172240257, + 0.008718885481357574, + 0.009748154319822788, + 0.010200321674346924, + 0.006501308176666498, + 0.012006939388811588, + 0.006506298668682575, + 0.005043880548328161, + 0.007902401499450207, + 0.005708161275833845, + 0.007286156993359327, + 0.013104165904223919, + 0.00627924082800746, + 0.012589709833264351, + 0.005098402034491301, + 0.005290558561682701, + 0.01165925431996584, + 0.008099285885691643, + 0.014095344580709934, + 0.008046257309615612, + 0.009527338668704033, + 0.00867290049791336, + 0.01230290811508894, + 0.008593006059527397, + 0.010253098793327808, + 0.011028960347175598, + 0.01028735563158989, + 0.01393794920295477, + 0.0067853485234081745, + 0.008191428147256374, + 0.014014290645718575, + 0.006301239598542452, + 0.009172649122774601, + 0.01422372367233038, + 0.008062451146543026, + 0.004137988667935133, + 0.002436238806694746, + 0.0042329044081270695, + 0.003041269024834037, + 0.0026638275012373924, + 0.0030664734076708555, + 0.0035012164153158665, + 0.003762439824640751, + 0.004923629574477673, + 0.005706868600100279, + 0.0070992205291986465, + 0.007385555189102888, + 0.006735623814165592, + 0.005746593698859215, + 0.007331802509725094, + 0.009338499046862125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.25215622782707214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010455148294568062 + ], + "opposite_logprob_tokens": [ + -6.8760457038879395 + ], + "logprob_reward": -0.0010455148294568062, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010455148294568062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22439062595367432, + 0.069674551486969, + 0.09674990922212601, + 0.06471075117588043, + 0.08327015489339828, + 0.09409336745738983, + 0.11345949023962021, + 0.1363675594329834, + 0.0432487428188324, + 0.07403479516506195 + ], + "credit_map": [ + 0.025342648848891258, + 0.037984833121299744, + 0.02518613450229168, + 0.025013767182826996, + 0.02776993066072464, + 0.025278910994529724, + 0.03203810006380081, + 0.025776302441954613, + 0.008473883382976055, + 0.008818554691970348, + 0.008335929363965988, + 0.007075244095176458, + 0.009808436036109924, + 0.009507947601377964, + 0.011045979335904121, + 0.006608585827052593, + 0.01570565067231655, + 0.014677596278488636, + 0.014774407260119915, + 0.010123583488166332, + 0.009688073769211769, + 0.011273705400526524, + 0.00899103470146656, + 0.011515862308442593, + 0.011405100114643574, + 0.007567533291876316, + 0.010229820385575294, + 0.006216397043317556, + 0.005856178235262632, + 0.008553982712328434, + 0.0056291613727808, + 0.009252584539353848, + 0.014870163053274155, + 0.009299236349761486, + 0.012381428852677345, + 0.010009368881583214, + 0.008596248924732208, + 0.010332341305911541, + 0.010144553147256374, + 0.007636826019734144, + 0.014620200730860233, + 0.00973187480121851, + 0.01717129535973072, + 0.010966621339321136, + 0.006442144978791475, + 0.012091520242393017, + 0.008594699203968048, + 0.0144750215113163, + 0.011478356085717678, + 0.012596153654158115, + 0.010660449042916298, + 0.021669860929250717, + 0.01373312622308731, + 0.014827130362391472, + 0.018114931881427765, + 0.010379488579928875, + 0.013259109109640121, + 0.01556254830211401, + 0.013435151427984238, + 0.023005353286862373, + 0.016354570165276527, + 0.016746647655963898, + 0.02610100246965885, + 0.011903177946805954, + 0.004138266667723656, + 0.00757800554856658, + 0.00440692063421011, + 0.006187212187796831, + 0.005364602897316217, + 0.004803470801562071, + 0.006416208576411009, + 0.004354060161858797, + 0.010389086790382862, + 0.011638855561614037, + 0.008329696021974087, + 0.007728812284767628, + 0.013407359831035137, + 0.007827920839190483, + 0.007240880746394396, + 0.007472185883671045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0010455148294568062 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, macro-time-lapse shot meticulously documents the accelerated life cycle of a single Ulysses plant, compressing 45 distinct flowering stages over three years into a fluid visual sequence. As a small, green bud first pokes through a patch of dark, fertile soil, the camera remains perfectly still, focused in extreme detail. It rapidly unfolds, revealing a radiant flower with an intricate array of vibrant yellow petals. The time-lapse effect is striking, with the background light rapidly flickering, simulating the passage of countless days and nights as the flower blooms, gracefully withers, and re-blooms in a continuous, hypnotic cycle. In the final stage, the yellow petals shrink, miraculously replaced by a perfect, ethereal sphere of fluffy white seeds, each a delicate parachute with a tiny seed at its base, seemingly poised and ready to be dispersed by the breeze. The entire natural drama is set against a stark, contemplative backdrop: a vertical scroll crafted from ancient black paper, adorned with elegant hand-painted Eastern script, its black ink brushstrokes creating a striking contrast to the plant's organic evolution. The controlled lighting subtly changes as the “seasons” pass, casting dynamic shadows that highlight the delicate textures of pencil and calligraphy, creating a tranquil, philosophical atmosphere that contemplates the beauty of the fleeting life cycle.", + "seed": 25571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_6/index_1528_seed_25571_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06252442883560434, + "vlm_logprob": -0.06252442883560434, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.06252442883560434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01833958923816681 + ], + "opposite_logprob_tokens": [ + -4.26833963394165 + ], + "logprob_reward": -0.01833958923816681, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01833958923816681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5547515749931335, + 0.07249007374048233, + 0.030635301023721695, + 0.03159809857606888, + 0.0362364687025547, + 0.11015544831752777, + 0.030210213735699654, + 0.026386229321360588, + 0.06644121557474136, + 0.04109543561935425 + ], + "credit_map": [ + 0.06418722122907639, + 0.04274201765656471, + 0.053056154400110245, + 0.08818116039037704, + 0.058684103190898895, + 0.0843605026602745, + 0.11888080090284348, + 0.04465959966182709, + 0.011522946879267693, + 0.006717204581946135, + 0.008060702122747898, + 0.007729227188974619, + 0.008590885438024998, + 0.010041228495538235, + 0.011934044770896435, + 0.007893836125731468, + 0.003906849771738052, + 0.003570712171494961, + 0.005954237654805183, + 0.0044270590879023075, + 0.0027607146184891462, + 0.003465905087068677, + 0.0031949502881616354, + 0.003354872576892376, + 0.0032925521954894066, + 0.0037311327178031206, + 0.004629802890121937, + 0.003581276861950755, + 0.003955626394599676, + 0.004189221654087305, + 0.004243301227688789, + 0.003975183703005314, + 0.004345771856606007, + 0.003287755884230137, + 0.008594372309744358, + 0.004802583251148462, + 0.0025971371214836836, + 0.003894064575433731, + 0.004739738069474697, + 0.003975045867264271, + 0.011306462809443474, + 0.02076839469373226, + 0.011836524121463299, + 0.0136205293238163, + 0.02058085985481739, + 0.012006270699203014, + 0.011434548534452915, + 0.008601848036050797, + 0.002801906783133745, + 0.004433506168425083, + 0.0036284036468714476, + 0.004446308594197035, + 0.004780896008014679, + 0.00379812135361135, + 0.0037820073775947094, + 0.002539061475545168, + 0.005123325623571873, + 0.0029530758038163185, + 0.003930266946554184, + 0.003195340046659112, + 0.003109119599685073, + 0.0026177160907536745, + 0.003036248032003641, + 0.0024211383424699306, + 0.009064987301826477, + 0.005632711108773947, + 0.010429935529828072, + 0.009802255779504776, + 0.005811546463519335, + 0.008583305403590202, + 0.008550621569156647, + 0.008565851487219334, + 0.005113824736326933, + 0.005309300031512976, + 0.006779377348721027, + 0.0032721315510571003, + 0.005407185759395361, + 0.005842315964400768, + 0.0033132329117506742, + 0.00605806615203619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01833958923816681 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003633328014984727 + ], + "opposite_logprob_tokens": [ + -5.628633499145508 + ], + "logprob_reward": -0.003633328014984727, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003633328014984727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37769004702568054, + 0.059468239545822144, + 0.026934491470456123, + 0.07567448914051056, + 0.0438249409198761, + 0.2137964963912964, + 0.024048006162047386, + 0.029760729521512985, + 0.11500080674886703, + 0.03380174934864044 + ], + "credit_map": [ + 0.06464697420597076, + 0.03734715282917023, + 0.041120026260614395, + 0.0388098768889904, + 0.05138826370239258, + 0.05258253589272499, + 0.05839134380221367, + 0.03340388834476471, + 0.006956083234399557, + 0.006621385924518108, + 0.00645449198782444, + 0.006105500739067793, + 0.009797113947570324, + 0.008860213682055473, + 0.009888645261526108, + 0.004784803371876478, + 0.00388042232953012, + 0.003311417531222105, + 0.003044016659259796, + 0.0026602691505104303, + 0.0031220074743032455, + 0.005194506607949734, + 0.002937882672995329, + 0.0027839697431772947, + 0.007697300985455513, + 0.00973901804536581, + 0.01042175106704235, + 0.00771970534697175, + 0.01024194248020649, + 0.011338085867464542, + 0.008349704556167126, + 0.010166979394853115, + 0.0046684592962265015, + 0.005908889230340719, + 0.005525785963982344, + 0.005323759280145168, + 0.004681711085140705, + 0.0049513294361531734, + 0.008051886223256588, + 0.004713122732937336, + 0.02012503705918789, + 0.04196666553616524, + 0.023818472400307655, + 0.02726692147552967, + 0.03860611468553543, + 0.024808768182992935, + 0.02111155167222023, + 0.016092993319034576, + 0.0037423139438033104, + 0.002292333170771599, + 0.004307459574192762, + 0.002148000057786703, + 0.002774157328531146, + 0.0038951700553297997, + 0.0025073885917663574, + 0.0023811834398657084, + 0.0056244973093271255, + 0.0026180229615420103, + 0.0038837094325572252, + 0.0031229108572006226, + 0.0033953676465898752, + 0.0032406949903815985, + 0.004851395729929209, + 0.003024129895493388, + 0.015427704900503159, + 0.0083439527079463, + 0.01915259100496769, + 0.009541776031255722, + 0.007636082824319601, + 0.021691488102078438, + 0.010024172253906727, + 0.023183036595582962, + 0.00511199003085494, + 0.003752538003027439, + 0.004525010474026203, + 0.00440586544573307, + 0.0036722731310874224, + 0.0041770413517951965, + 0.0038351467810571194, + 0.004321882501244545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.003633328014984727 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641467429697514 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641467429697514, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641467429697514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8389713168144226, + 0.03333612158894539, + 0.015383228659629822, + 0.015252707526087761, + 0.018575649708509445, + 0.032834507524967194, + 0.008129063062369823, + 0.01027760561555624, + 0.015024639666080475, + 0.012215136550366879 + ], + "credit_map": [ + 0.10219217091798782, + 0.06785370409488678, + 0.08208850771188736, + 0.1268225610256195, + 0.0882711336016655, + 0.13324736058712006, + 0.16612176597118378, + 0.072374127805233, + 0.006139474920928478, + 0.0040988679975271225, + 0.004304512869566679, + 0.0036770079750567675, + 0.004141601733863354, + 0.0045963856391608715, + 0.0034204022958874702, + 0.0029578693211078644, + 0.001887989928945899, + 0.0015127738006412983, + 0.001869659055955708, + 0.0023928850423544645, + 0.001705982256680727, + 0.0018101760651916265, + 0.0025353510864079, + 0.0016684120055288076, + 0.0017284790519624949, + 0.0024248366244137287, + 0.0018525444902479649, + 0.0017248140648007393, + 0.002604742767289281, + 0.0017763079376891255, + 0.0018171933479607105, + 0.0013237879611551762, + 0.002269327873364091, + 0.0019064482767134905, + 0.003794546937569976, + 0.002455294830724597, + 0.0015583831118419766, + 0.0018061097944155335, + 0.0026391148567199707, + 0.002146422630175948, + 0.0026197165716439486, + 0.005690611898899078, + 0.004590586759150028, + 0.004023396410048008, + 0.005480172112584114, + 0.003978097345679998, + 0.003649091348052025, + 0.002802835311740637, + 0.0007915801834315062, + 0.0008163074380718172, + 0.0012256917543709278, + 0.0013278834521770477, + 0.0010446503292769194, + 0.0009385916055180132, + 0.001080565620213747, + 0.0009037918644025922, + 0.0011989938793703914, + 0.0014933878555893898, + 0.001387687516398728, + 0.0015198534820228815, + 0.0012644266244024038, + 0.0010459163459017873, + 0.0014344302471727133, + 0.0009329096064902842, + 0.0017260834574699402, + 0.0012429730268195271, + 0.0029838052578270435, + 0.0021184058859944344, + 0.001340314862318337, + 0.0019662047270685434, + 0.0019567273557186127, + 0.0016901246272027493, + 0.0016958483029156923, + 0.0010793546680361032, + 0.002183075761422515, + 0.0011768522672355175, + 0.0012280158698558807, + 0.0015111388638615608, + 0.0010697852121666074, + 0.002271065255627036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008641467429697514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813706696033478 + ], + "opposite_logprob_tokens": [ + -1.4063706398010254 + ], + "logprob_reward": -0.2813706696033478, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813706696033478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5033018589019775, + 0.06329379975795746, + 0.021335873752832413, + 0.11294516921043396, + 0.0500100739300251, + 0.12806367874145508, + 0.051192756742239, + 0.01851332001388073, + 0.03378184139728546, + 0.01756160333752632 + ], + "credit_map": [ + 0.0809897929430008, + 0.047912511974573135, + 0.053663309663534164, + 0.061784468591213226, + 0.05614825710654259, + 0.07001630961894989, + 0.08692926168441772, + 0.04585795849561691, + 0.010359914042055607, + 0.007407022640109062, + 0.008421620354056358, + 0.005307708401232958, + 0.009015204384922981, + 0.007834823802113533, + 0.007589348591864109, + 0.007358155213296413, + 0.0032212319783866405, + 0.0017639481229707599, + 0.004215669352561235, + 0.002162369666621089, + 0.002041463740170002, + 0.0023383062798529863, + 0.0032677212730050087, + 0.0023251641541719437, + 0.0115247443318367, + 0.017670033499598503, + 0.012414037249982357, + 0.013176467269659042, + 0.01345144771039486, + 0.015887925401329994, + 0.017248891294002533, + 0.011571616865694523, + 0.006236426066607237, + 0.005834961775690317, + 0.0078353900462389, + 0.005838647484779358, + 0.005887365434318781, + 0.005933697801083326, + 0.007575820200145245, + 0.004867765586823225, + 0.010630692355334759, + 0.022585341706871986, + 0.016778388991951942, + 0.018720969557762146, + 0.020500391721725464, + 0.014907469972968102, + 0.01374224666506052, + 0.010198170319199562, + 0.005961963441222906, + 0.006476175505667925, + 0.006719511933624744, + 0.00771901523694396, + 0.007349720224738121, + 0.005535203032195568, + 0.007402513641864061, + 0.00402865232899785, + 0.0020372048020362854, + 0.0014716815203428268, + 0.0033155123237520456, + 0.0021149064414203167, + 0.0015098723815754056, + 0.0025446792133152485, + 0.003112323582172394, + 0.002407138468697667, + 0.0053850142285227776, + 0.0025586630217731, + 0.00464879535138607, + 0.004260506946593523, + 0.0019113573944196105, + 0.005133488215506077, + 0.004177074413746595, + 0.0057069421745836735, + 0.002460318151861429, + 0.001728971372358501, + 0.003800793783739209, + 0.0015149492537602782, + 0.0014945394359529018, + 0.002521489979699254, + 0.0019819834269583225, + 0.0020585583988577127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.2813706696033478 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single Ulysses plant growing from a green bud into a flower and then a sphere of white seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the plant's life cycle compressed into a time-lapse sequence covering multiple stages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vertical scroll made of ancient black paper with hand-painted Eastern script?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background lighting flicker rapidly to simulate the passage of time while the camera remains perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the final seeds depicted as a perfect sphere of fluffy white structures resembling delicate parachutes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006370898918248713 + ], + "opposite_logprob_tokens": [ + -7.375637054443359 + ], + "logprob_reward": -0.0006370898918248713, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006370898918248713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6276331543922424, + 0.04749125987291336, + 0.008659793063998222, + 0.03094770573079586, + 0.015315578319132328, + 0.11974845081567764, + 0.013180534355342388, + 0.01982811838388443, + 0.07166314125061035, + 0.0455322228372097 + ], + "credit_map": [ + 0.05776369571685791, + 0.041686419397592545, + 0.049277640879154205, + 0.11397313326597214, + 0.07462093979120255, + 0.08657395839691162, + 0.1488584578037262, + 0.05487896129488945, + 0.00834297388792038, + 0.005462937988340855, + 0.006492380052804947, + 0.005600667558610439, + 0.0053030154667794704, + 0.006705258972942829, + 0.005295503418892622, + 0.004288519266992807, + 0.0011282408377155662, + 0.0008776101749390364, + 0.001245143823325634, + 0.0009263880783692002, + 0.0007812075200490654, + 0.0013170818565413356, + 0.0012767299776896834, + 0.0011073906207457185, + 0.0028206908609718084, + 0.004452304914593697, + 0.003974684048444033, + 0.0036153756082057953, + 0.00466125039383769, + 0.004167346283793449, + 0.003686962416395545, + 0.003569091437384486, + 0.0015371307963505387, + 0.0017323754727840424, + 0.0021141311153769493, + 0.0025210073217749596, + 0.0018605224322527647, + 0.0016084256349131465, + 0.0022678705863654613, + 0.0016741147264838219, + 0.01009584590792656, + 0.022735904902219772, + 0.01590508036315441, + 0.014493563212454319, + 0.020519044250249863, + 0.01446850597858429, + 0.011477482505142689, + 0.010053024627268314, + 0.0021209632977843285, + 0.0014411777956411242, + 0.001967737916857004, + 0.0015891932416707277, + 0.0012495008995756507, + 0.0015644876984879375, + 0.0016287154285237193, + 0.001618758193217218, + 0.003460559295490384, + 0.0024896704126149416, + 0.003070675767958164, + 0.002260094042867422, + 0.0017967360327020288, + 0.0022214341443032026, + 0.0024425152223557234, + 0.0020864338148385286, + 0.007989010773599148, + 0.00461335014551878, + 0.013284435495734215, + 0.013207908719778061, + 0.003976571373641491, + 0.008978033438324928, + 0.008748753927648067, + 0.010865074582397938, + 0.0045693665742874146, + 0.0057271188125014305, + 0.006529370788484812, + 0.005937374662607908, + 0.004665108863264322, + 0.005812469404190779, + 0.004801533184945583, + 0.007489881478250027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0006370898918248713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_tensor([2423, 2423])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_tensor([2423, 2423])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f3fb7ad0aeef6e057f26e83e6e5617bb44ecd67 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_tensor([2423, 2423])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Amidst a scene rich in detailed CG art, a powerful, tight close-up shot, viewed from a slightly upward angle, captures a mysterious, cloaked Protoss warrior as he strives to complete a powerful sweep from right to left. His exposed right arm is defined by defined muscles, veins bulging, each muscle fiber tense and trembling with the sheer force of his exertion. A golden bracer adorned with ancient runes reflects the cool metallic texture in the dim light. He clutches his signature emerald blade, composed of pure psionic energy. The blade traces a dazzling, fleeting arc of light through the darkness. Its high-speed motion creates a subtle energy spill and particle effect at its edge, stirring up tiny particles suspended in the air to form ripples of light, perfectly recreating the blade's path through the air. His cloak, a heavy, intricately patterned dark blue fabric, fluttered violently. From the deep shadow of his hood, only a glimpse of emerald eyes gleamed with intense psychic energy, revealing unmistakable focus and strength. The setting was the gloomy interior of an alien temple, its cold metal walls carved with intricate geometric patterns. The psychic blade was the sole primary light source, casting a sharp, dynamic green glow, instantly illuminating the warrior's taut muscles, the folds of his cloak, and the metal reliefs on the background walls. This created a striking contrast between light and dark, and a highly dramatic effect of light and shadow. The entire scene, with its deep, cold tones, was broken by the piercing emerald green light, imbued with a tense, dynamic, and mysterious atmosphere of epic proportions, showcasing the finest CG artistry.", + "seed": 21673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21673_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03501222294289619, + "vlm_logprob": -0.03501222294289619, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03501222294289619, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019299466162919998 + ], + "opposite_logprob_tokens": [ + -4.144299507141113 + ], + "logprob_reward": -0.019299466162919998, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019299466162919998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06740779429674149, + 0.13026396930217743, + 0.2309693992137909, + 0.06905388087034225, + 0.09166961163282394, + 0.07485014200210571, + 0.05184606462717056, + 0.09345701336860657, + 0.050356924533843994, + 0.14012518525123596 + ], + "credit_map": [ + 0.006276207976043224, + 0.005251915194094181, + 0.006121600512415171, + 0.00515343900769949, + 0.008627105504274368, + 0.009745043702423573, + 0.012731654569506645, + 0.013500828295946121, + 0.011322328820824623, + 0.009468304924666882, + 0.008955695666372776, + 0.007431067060679197, + 0.015029286965727806, + 0.03743810951709747, + 0.027516063302755356, + 0.013103120028972626, + 0.03040439262986183, + 0.04269865155220032, + 0.014050377532839775, + 0.019906383007764816, + 0.021446604281663895, + 0.03867622837424278, + 0.01565680466592312, + 0.04812996834516525, + 0.007064868230372667, + 0.00606559356674552, + 0.009042472578585148, + 0.007159097585827112, + 0.007242831867188215, + 0.010409604758024216, + 0.010909264907240868, + 0.011160152964293957, + 0.015305008739233017, + 0.019809387624263763, + 0.006539075169712305, + 0.011041353456676006, + 0.009602864272892475, + 0.009320501238107681, + 0.009184238500893116, + 0.010867186821997166, + 0.009793013334274292, + 0.010755077935755253, + 0.007769488263875246, + 0.0073150163516402245, + 0.014154854230582714, + 0.009074585512280464, + 0.007203223649412394, + 0.008784881792962551, + 0.004653529729694128, + 0.007001873571425676, + 0.003838650416582823, + 0.006288222037255764, + 0.009087974205613136, + 0.00605993065983057, + 0.005606601480394602, + 0.009309282526373863, + 0.010878730565309525, + 0.008802500553429127, + 0.01236681081354618, + 0.009701034985482693, + 0.011534111574292183, + 0.01811547577381134, + 0.01350806187838316, + 0.008550290949642658, + 0.006676390301436186, + 0.003231370123103261, + 0.006673584692180157, + 0.009704748168587685, + 0.005525109823793173, + 0.006244813557714224, + 0.006843468174338341, + 0.0054574389941990376, + 0.01694307290017605, + 0.017672503367066383, + 0.022232485935091972, + 0.00896433461457491, + 0.011228743009269238, + 0.03523970767855644, + 0.007932750508189201, + 0.019911592826247215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.019299466162919998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866022989153862 + ], + "opposite_logprob_tokens": [ + -3.0486602783203125 + ], + "logprob_reward": -0.04866022989153862, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866022989153862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10311009734869003, + 0.09396141022443771, + 0.15550631284713745, + 0.03447524830698967, + 0.12941132485866547, + 0.06376020610332489, + 0.07889050990343094, + 0.09178266674280167, + 0.05197774991393089, + 0.19712448120117188 + ], + "credit_map": [ + 0.00751902349293232, + 0.006347429007291794, + 0.009146535769104958, + 0.008058455772697926, + 0.01493746042251587, + 0.013974952511489391, + 0.01781226135790348, + 0.025313986465334892, + 0.007540144491940737, + 0.008284795098006725, + 0.007744098547846079, + 0.007306997198611498, + 0.013341328129172325, + 0.018443413078784943, + 0.016735345125198364, + 0.014565286226570606, + 0.014017932116985321, + 0.015580822713673115, + 0.01430608332157135, + 0.011469443328678608, + 0.013007041998207569, + 0.03848955035209656, + 0.014139696024358273, + 0.03449572995305061, + 0.0066043054684996605, + 0.004539654590189457, + 0.003255542367696762, + 0.0028814419638365507, + 0.003800899488851428, + 0.004477746319025755, + 0.003663028124719858, + 0.005252632778137922, + 0.021928800269961357, + 0.01840086467564106, + 0.008350750431418419, + 0.015514836646616459, + 0.016750873997807503, + 0.013699940405786037, + 0.013052018359303474, + 0.021713240072131157, + 0.006174168549478054, + 0.012041216716170311, + 0.0066134072840213776, + 0.006846137810498476, + 0.008180475793778896, + 0.008416025899350643, + 0.0072166575118899345, + 0.008272114209830761, + 0.007415177766233683, + 0.006835366133600473, + 0.005766072776168585, + 0.010176305659115314, + 0.013954244554042816, + 0.009027490392327309, + 0.010265469551086426, + 0.015450382605195045, + 0.010057670995593071, + 0.009565016254782677, + 0.011850797571241856, + 0.00805007666349411, + 0.015964368358254433, + 0.012074957601726055, + 0.012331681326031685, + 0.011888100765645504, + 0.004090491682291031, + 0.005655339453369379, + 0.008485415950417519, + 0.007973563857376575, + 0.007011210545897484, + 0.006176438182592392, + 0.00789306964725256, + 0.004692223388701677, + 0.020989183336496353, + 0.013564149849116802, + 0.03163167089223862, + 0.01368753146380186, + 0.013516156002879143, + 0.056219469755887985, + 0.0130615858361125, + 0.03445474058389664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.04866022989153862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525664072483778 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.00525664072483778, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525664072483778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05015482008457184, + 0.10652381926774979, + 0.179765984416008, + 0.03757794201374054, + 0.10086175799369812, + 0.0989503487944603, + 0.05963081493973732, + 0.05165832117199898, + 0.06817156076431274, + 0.24670468270778656 + ], + "credit_map": [ + 0.005996122024953365, + 0.0033321997616440058, + 0.007007014937698841, + 0.003629544982686639, + 0.006658522412180901, + 0.0062533277086913586, + 0.009699998423457146, + 0.007578086107969284, + 0.010304797440767288, + 0.009515883401036263, + 0.01020397711545229, + 0.005264281760901213, + 0.010360844433307648, + 0.030183905735611916, + 0.021651864051818848, + 0.00903826579451561, + 0.022722981870174408, + 0.035521239042282104, + 0.013287540525197983, + 0.01463178638368845, + 0.017530210316181183, + 0.03193119540810585, + 0.010005765594542027, + 0.03413526713848114, + 0.005715579260140657, + 0.004860781133174896, + 0.004547356627881527, + 0.003810061840340495, + 0.0037092226557433605, + 0.004450508393347263, + 0.004942991770803928, + 0.005541435442864895, + 0.01624346897006035, + 0.011641503311693668, + 0.010145067237317562, + 0.011393319815397263, + 0.014274735935032368, + 0.010043532587587833, + 0.010343196801841259, + 0.016776930540800095, + 0.010432517156004906, + 0.00970711000263691, + 0.013885464519262314, + 0.007281675934791565, + 0.010647928342223167, + 0.014782977290451527, + 0.007732932455837727, + 0.024479737505316734, + 0.004689602181315422, + 0.00964214839041233, + 0.00433362228795886, + 0.007816080003976822, + 0.010331390425562859, + 0.006861892994493246, + 0.0076422616839408875, + 0.008313811384141445, + 0.006989592220634222, + 0.005938461050391197, + 0.006908508483320475, + 0.005890952423214912, + 0.007247440051287413, + 0.00721675343811512, + 0.00570022314786911, + 0.0057663884945213795, + 0.00918910838663578, + 0.005749252159148455, + 0.010041983798146248, + 0.01237489003688097, + 0.00819934718310833, + 0.008080837316811085, + 0.007583946455270052, + 0.0069521949626505375, + 0.028451530262827873, + 0.0279576126486063, + 0.03701554238796234, + 0.01589156687259674, + 0.01580452360212803, + 0.06798607110977173, + 0.01408673170953989, + 0.03951109200716019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.00525664072483778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032887011766434 + ], + "opposite_logprob_tokens": [ + -2.3503289222717285 + ], + "logprob_reward": -0.10032887011766434, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032887011766434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040106065571308136, + 0.08759915083646774, + 0.16216272115707397, + 0.025349179282784462, + 0.05680263787508011, + 0.03432740643620491, + 0.05613330379128456, + 0.05875561758875847, + 0.060728635638952255, + 0.4180352985858917 + ], + "credit_map": [ + 0.006102721206843853, + 0.002554042497649789, + 0.006598586682230234, + 0.003230457426980138, + 0.0038178442046046257, + 0.005829588510096073, + 0.005837383680045605, + 0.006135443225502968, + 0.00936795398592949, + 0.009792586788535118, + 0.012390007264912128, + 0.00598490983247757, + 0.009244724176824093, + 0.019998235628008842, + 0.013536191545426846, + 0.007284540217369795, + 0.01927877403795719, + 0.023640980944037437, + 0.0168645940721035, + 0.010386285372078419, + 0.015569864772260189, + 0.033638522028923035, + 0.009449055418372154, + 0.033334650099277496, + 0.003642774187028408, + 0.0026859694626182318, + 0.003031495725736022, + 0.0019071793649345636, + 0.002891580108553171, + 0.004277336876839399, + 0.002993216272443533, + 0.00391962518915534, + 0.008073573000729084, + 0.0065470170229673386, + 0.005328489001840353, + 0.006115319672971964, + 0.009131912142038345, + 0.008228404447436333, + 0.006173395551741123, + 0.007204529829323292, + 0.004241735674440861, + 0.0033109395299106836, + 0.004422995261847973, + 0.0031028653029352427, + 0.003797064069658518, + 0.004143903963267803, + 0.004004697315394878, + 0.007303202990442514, + 0.0050894287414848804, + 0.007775365374982357, + 0.005217690486460924, + 0.006294297520071268, + 0.009358758106827736, + 0.006142998579889536, + 0.006550164893269539, + 0.009704597294330597, + 0.007970250211656094, + 0.005023701582103968, + 0.008848610334098339, + 0.0037158867344260216, + 0.006864647846668959, + 0.012029915116727352, + 0.009172247722744942, + 0.005130356177687645, + 0.006473777815699577, + 0.00641739834100008, + 0.008911442942917347, + 0.01118486374616623, + 0.007842699997127056, + 0.007233578711748123, + 0.007438295986503363, + 0.005226577632129192, + 0.04615011066198349, + 0.05252022668719292, + 0.06291480362415314, + 0.02889452502131462, + 0.03346128761768341, + 0.10323817282915115, + 0.02974715828895569, + 0.0611090287566185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.10032887011766434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015159078175202012 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015159078175202012, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015159078175202012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059235259890556335, + 0.04590913653373718, + 0.24224726855754852, + 0.02956191450357437, + 0.051861170679330826, + 0.05842968448996544, + 0.07059309631586075, + 0.1155979186296463, + 0.10767916589975357, + 0.2188853770494461 + ], + "credit_map": [ + 0.009213125333189964, + 0.00513653177767992, + 0.008522083051502705, + 0.004361078608781099, + 0.005677256267517805, + 0.008757125586271286, + 0.00860058143734932, + 0.008967473171651363, + 0.005656299646943808, + 0.007093257736414671, + 0.005845432169735432, + 0.003574024885892868, + 0.005818928591907024, + 0.005409707315266132, + 0.006192910019308329, + 0.006318577565252781, + 0.02687188982963562, + 0.02960733138024807, + 0.031118152663111687, + 0.015722468495368958, + 0.025673383846879005, + 0.04979529604315758, + 0.01850120536983013, + 0.044957537204027176, + 0.002895546844229102, + 0.0031925085932016373, + 0.003589220345020294, + 0.0025341634172946215, + 0.0036685869563370943, + 0.004806167911738157, + 0.00426948256790638, + 0.004606238100677729, + 0.008136183023452759, + 0.005904991179704666, + 0.004573197569698095, + 0.0050513530150055885, + 0.005394089501351118, + 0.006493757478892803, + 0.005084988195449114, + 0.01122260745614767, + 0.00634790351614356, + 0.0076888613402843475, + 0.0070674060843884945, + 0.004438033327460289, + 0.008448446169495583, + 0.008154422976076603, + 0.004528537392616272, + 0.011756074614822865, + 0.006445019040256739, + 0.011202418245375156, + 0.006655085366219282, + 0.009342396631836891, + 0.009743764065206051, + 0.007497720420360565, + 0.00879170186817646, + 0.01091498788446188, + 0.014040948823094368, + 0.006869771517813206, + 0.010902300477027893, + 0.006955706048756838, + 0.007635307963937521, + 0.03826699033379555, + 0.02291281521320343, + 0.008014081045985222, + 0.01535677071660757, + 0.008929123170673847, + 0.015372051857411861, + 0.017575403675436974, + 0.013717574067413807, + 0.013054384849965572, + 0.012508350424468517, + 0.011165506206452847, + 0.024782700464129448, + 0.019147835671901703, + 0.03696746006608009, + 0.014797956682741642, + 0.015117671340703964, + 0.05941477045416832, + 0.013994541950523853, + 0.03466242924332619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.0015159078175202012 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst a scene rich in detailed CG art, a powerful, tight close-up shot, viewed from a slightly upward angle, captures a mysterious, cloaked Protoss warrior as he strives to complete a powerful sweep from right to left. His exposed right arm is defined by defined muscles, veins bulging, each muscle fiber tense and trembling with the sheer force of his exertion. A golden bracer adorned with ancient runes reflects the cool metallic texture in the dim light. He clutches his signature emerald blade, composed of pure psionic energy. The blade traces a dazzling, fleeting arc of light through the darkness. Its high-speed motion creates a subtle energy spill and particle effect at its edge, stirring up tiny particles suspended in the air to form ripples of light, perfectly recreating the blade's path through the air. His cloak, a heavy, intricately patterned dark blue fabric, fluttered violently. From the deep shadow of his hood, only a glimpse of emerald eyes gleamed with intense psychic energy, revealing unmistakable focus and strength. The setting was the gloomy interior of an alien temple, its cold metal walls carved with intricate geometric patterns. The psychic blade was the sole primary light source, casting a sharp, dynamic green glow, instantly illuminating the warrior's taut muscles, the folds of his cloak, and the metal reliefs on the background walls. This created a striking contrast between light and dark, and a highly dramatic effect of light and shadow. The entire scene, with its deep, cold tones, was broken by the piercing emerald green light, imbued with a tense, dynamic, and mysterious atmosphere of epic proportions, showcasing the finest CG artistry.", + "seed": 21674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21674_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05778219827916473, + "vlm_logprob": -0.05778219827916473, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.05778219827916473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04176096245646477 + ], + "opposite_logprob_tokens": [ + -3.2917609214782715 + ], + "logprob_reward": -0.04176096245646477, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04176096245646477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18293629586696625, + 0.09408916532993317, + 0.03900943323969841, + 0.056049250066280365, + 0.36518028378486633, + 0.037759553641080856, + 0.03404572233557701, + 0.05292079225182533, + 0.07579214870929718, + 0.06221744045615196 + ], + "credit_map": [ + 0.020673858001828194, + 0.01892072707414627, + 0.0352836437523365, + 0.02680598758161068, + 0.021185848861932755, + 0.018119003623723984, + 0.021909179165959358, + 0.020038040354847908, + 0.012294982559978962, + 0.006905184593051672, + 0.013901777565479279, + 0.007444276008754969, + 0.019917186349630356, + 0.01124721858650446, + 0.011497307568788528, + 0.010881228372454643, + 0.0036923137959092855, + 0.003774130716919899, + 0.0028103680815547705, + 0.004006051924079657, + 0.005276517476886511, + 0.004325252957642078, + 0.0067546009086072445, + 0.008370192721486092, + 0.0055444068275392056, + 0.008239078335464, + 0.007201679982244968, + 0.005660544149577618, + 0.008930033072829247, + 0.005549916997551918, + 0.008205000311136246, + 0.006718585733324289, + 0.012061521410942078, + 0.009715021587908268, + 0.009475941769778728, + 0.017206452786922455, + 0.014479727484285831, + 0.14852575957775116, + 0.016038093715906143, + 0.13767774403095245, + 0.005179407075047493, + 0.006526981480419636, + 0.005073529202491045, + 0.003276773961260915, + 0.0036334882024675608, + 0.004614822566509247, + 0.003606733400374651, + 0.005847814958542585, + 0.003945418167859316, + 0.005279686767607927, + 0.004094748757779598, + 0.005405440926551819, + 0.0046082353219389915, + 0.0048186639323830605, + 0.002974889473989606, + 0.002918639685958624, + 0.006400094833225012, + 0.005399100948125124, + 0.008444604463875294, + 0.008094624616205692, + 0.005408350843936205, + 0.006898979190737009, + 0.0053467885591089725, + 0.006928245071321726, + 0.007646861020475626, + 0.008139343932271004, + 0.012214711867272854, + 0.007392049767076969, + 0.00794809777289629, + 0.0111489063128829, + 0.00966728013008833, + 0.011634892784059048, + 0.006621185224503279, + 0.006187659688293934, + 0.009457916021347046, + 0.005624422337859869, + 0.008476827293634415, + 0.006233674939721823, + 0.010243282653391361, + 0.009372465312480927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.04176096245646477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495982617139816 + ], + "opposite_logprob_tokens": [ + -2.929959774017334 + ], + "logprob_reward": -0.05495982617139816, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495982617139816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18732307851314545, + 0.10078711807727814, + 0.07122920453548431, + 0.06055944040417671, + 0.21952740848064423, + 0.09439174085855484, + 0.06258214265108109, + 0.08694504201412201, + 0.02724565751850605, + 0.08940919488668442 + ], + "credit_map": [ + 0.030032789334654808, + 0.022116336971521378, + 0.03844422474503517, + 0.03204461559653282, + 0.017649779096245766, + 0.014113926328718662, + 0.019208479672670364, + 0.013712919317185879, + 0.010032564401626587, + 0.008354958146810532, + 0.012246579863131046, + 0.0126058803871274, + 0.012837846763432026, + 0.014243614859879017, + 0.012357202358543873, + 0.01810847595334053, + 0.009298725053668022, + 0.006612615194171667, + 0.00509581109508872, + 0.006607172079384327, + 0.008207363076508045, + 0.008806928992271423, + 0.010991854593157768, + 0.015608732588589191, + 0.006398214027285576, + 0.008455818518996239, + 0.009377972222864628, + 0.00565149262547493, + 0.008093679323792458, + 0.006923513952642679, + 0.00939514022320509, + 0.0062636095099151134, + 0.00856483168900013, + 0.01357402466237545, + 0.006631552241742611, + 0.01482698880136013, + 0.016254212707281113, + 0.07138790935277939, + 0.018277356401085854, + 0.07001054286956787, + 0.008767295628786087, + 0.01532910205423832, + 0.00849911943078041, + 0.008723859675228596, + 0.010930982418358326, + 0.0148212481290102, + 0.012304024770855904, + 0.015016100369393826, + 0.005581106524914503, + 0.008836528286337852, + 0.007994110696017742, + 0.0096385907381773, + 0.00689389742910862, + 0.008493950590491295, + 0.008130203001201153, + 0.0070137521252036095, + 0.013534223660826683, + 0.010039888322353363, + 0.01068190485239029, + 0.016067635267972946, + 0.009730483405292034, + 0.01067529059946537, + 0.0068022520281374454, + 0.009413369931280613, + 0.004178062081336975, + 0.002812563208863139, + 0.003797054523602128, + 0.0027178435120731592, + 0.0023129654582589865, + 0.003634714987128973, + 0.004261916037648916, + 0.00353053817525506, + 0.009022965095937252, + 0.006941864266991615, + 0.013983775861561298, + 0.00797196477651596, + 0.009235996752977371, + 0.007918157614767551, + 0.023492351174354553, + 0.010842111892998219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.05495982617139816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082876473665237 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082876473665237, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082876473665237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18163925409317017, + 0.1033400371670723, + 0.08315441012382507, + 0.08012871444225311, + 0.22165092825889587, + 0.04334656521677971, + 0.03691566735506058, + 0.17197833955287933, + 0.03659110888838768, + 0.041254978626966476 + ], + "credit_map": [ + 0.030389685183763504, + 0.015011796727776527, + 0.035586051642894745, + 0.026195844635367393, + 0.01411294937133789, + 0.017201878130435944, + 0.015460959635674953, + 0.027680085971951485, + 0.010965770110487938, + 0.008477150462567806, + 0.01316021103411913, + 0.013049648143351078, + 0.013575631193816662, + 0.013704714365303516, + 0.011755711399018764, + 0.018651189282536507, + 0.00851824227720499, + 0.008699506521224976, + 0.008056119084358215, + 0.00740882009267807, + 0.012169497087597847, + 0.010336481034755707, + 0.012392958626151085, + 0.015572781674563885, + 0.008492374792695045, + 0.011272244155406952, + 0.013398065231740475, + 0.008734410628676414, + 0.010801317170262337, + 0.007208717055618763, + 0.011773383244872093, + 0.008448204025626183, + 0.010007224045693874, + 0.013793857768177986, + 0.00944924633949995, + 0.012332292273640633, + 0.009458044543862343, + 0.07796549052000046, + 0.01269528828561306, + 0.07594949007034302, + 0.004272758029401302, + 0.010276992805302143, + 0.004887999966740608, + 0.003999960143119097, + 0.004618263337761164, + 0.005746546667069197, + 0.004902302753180265, + 0.004641739651560783, + 0.00534982280805707, + 0.0065182773396372795, + 0.004879338201135397, + 0.005180188920348883, + 0.003513335483148694, + 0.0051537854596972466, + 0.00312856980599463, + 0.0031923495698720217, + 0.02814178355038166, + 0.017728082835674286, + 0.023589469492435455, + 0.03631141036748886, + 0.013665420934557915, + 0.021669065579771996, + 0.014079560525715351, + 0.016793539747595787, + 0.004586504772305489, + 0.0030169140081852674, + 0.007078316528350115, + 0.0036508298944681883, + 0.004528331104665995, + 0.004435161594301462, + 0.0050461734645068645, + 0.004248877055943012, + 0.004815862979739904, + 0.00354041182436049, + 0.005453101359307766, + 0.00499257817864418, + 0.00402130326256156, + 0.0040740338154137135, + 0.008420564234256744, + 0.005937122739851475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.011082876473665237 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992320656776428 + ], + "opposite_logprob_tokens": [ + -1.804923176765442 + ], + "logprob_reward": -0.17992320656776428, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992320656776428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25104743242263794, + 0.14220774173736572, + 0.03466661646962166, + 0.19385161995887756, + 0.16708694398403168, + 0.053473036736249924, + 0.04809551686048508, + 0.034220464527606964, + 0.022993601858615875, + 0.052357085049152374 + ], + "credit_map": [ + 0.04055085778236389, + 0.030036861076951027, + 0.04944154620170593, + 0.04883544147014618, + 0.01716163381934166, + 0.017960557714104652, + 0.019334597513079643, + 0.027725929394364357, + 0.014243456535041332, + 0.012734230607748032, + 0.017818743363022804, + 0.01895875297486782, + 0.0198776014149189, + 0.017072074115276337, + 0.01722003147006035, + 0.024282852187752724, + 0.0032808848191052675, + 0.002677495591342449, + 0.002456477377563715, + 0.002338182646781206, + 0.00705321179702878, + 0.005459981970489025, + 0.005540086422115564, + 0.005860296543687582, + 0.021947767585515976, + 0.02873210236430168, + 0.03254423663020134, + 0.018560923635959625, + 0.025274446234107018, + 0.017708640545606613, + 0.028383567929267883, + 0.02069992385804653, + 0.0071847946383059025, + 0.007435740437358618, + 0.00681152381002903, + 0.011641125194728374, + 0.011070901528000832, + 0.06032712385058403, + 0.009404036216437817, + 0.05321168899536133, + 0.006219496950507164, + 0.010685100220143795, + 0.008140786550939083, + 0.005984945222735405, + 0.00419266102835536, + 0.0074356007389724255, + 0.004382084123790264, + 0.00643236143514514, + 0.004869160708039999, + 0.00859204214066267, + 0.006969626061618328, + 0.008193089626729488, + 0.005398149602115154, + 0.005558548029512167, + 0.004567995667457581, + 0.003946907818317413, + 0.006135857198387384, + 0.004309328272938728, + 0.004434406757354736, + 0.0037292770575731993, + 0.003648435929790139, + 0.004475038032978773, + 0.0034680459648370743, + 0.004020076245069504, + 0.002806541044265032, + 0.0021785099525004625, + 0.004058925434947014, + 0.002290975069627166, + 0.002498994115740061, + 0.002867775969207287, + 0.0023171689826995134, + 0.003974711522459984, + 0.006227413658052683, + 0.004305430222302675, + 0.007780591491609812, + 0.008900600485503674, + 0.004883840214461088, + 0.006081006955355406, + 0.007976185530424118, + 0.006202015560120344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.17992320656776428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011841197265312076 + ], + "opposite_logprob_tokens": [ + -6.751183986663818 + ], + "logprob_reward": -0.0011841197265312076, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011841197265312076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28574591875076294, + 0.09378278255462646, + 0.03268106281757355, + 0.11519492417573929, + 0.20516259968280792, + 0.05483650416135788, + 0.033548176288604736, + 0.07456833869218826, + 0.02719634398818016, + 0.07728338986635208 + ], + "credit_map": [ + 0.02403131313621998, + 0.02416238561272621, + 0.07621583342552185, + 0.08469018340110779, + 0.022953161969780922, + 0.0172096136957407, + 0.022999433800578117, + 0.013483968563377857, + 0.00965038500726223, + 0.00857546366751194, + 0.012988181784749031, + 0.013573302887380123, + 0.01254701055586338, + 0.011667496524751186, + 0.011218864470720291, + 0.013562069274485111, + 0.0031020978931337595, + 0.002690988127142191, + 0.002073446987196803, + 0.0027663882356137037, + 0.0059543270617723465, + 0.004234473221004009, + 0.005841183941811323, + 0.006018155254423618, + 0.011089818552136421, + 0.014673463068902493, + 0.019776906818151474, + 0.010530659928917885, + 0.015263577923178673, + 0.010923669673502445, + 0.018422292545437813, + 0.014514535665512085, + 0.008056088350713253, + 0.00831590685993433, + 0.00714531447738409, + 0.01264269556850195, + 0.011808880604803562, + 0.07699614018201828, + 0.01158628799021244, + 0.06861128658056259, + 0.005993835162371397, + 0.012259331531822681, + 0.008184965699911118, + 0.006071803160011768, + 0.004525552038103342, + 0.008790112100541592, + 0.003456923644989729, + 0.005553980823606253, + 0.0028973466251045465, + 0.004656664561480284, + 0.0035286988131701946, + 0.004354104399681091, + 0.005005779210478067, + 0.005511855706572533, + 0.003982442896813154, + 0.0036112810485064983, + 0.014555376023054123, + 0.00997498445212841, + 0.011305917985737324, + 0.009041747078299522, + 0.0065254648216068745, + 0.009175914339721203, + 0.005769401788711548, + 0.008219531737267971, + 0.0030696901958435774, + 0.00306428549811244, + 0.004790460225194693, + 0.00289525487460196, + 0.0032344709616154432, + 0.003336205380037427, + 0.003664166433736682, + 0.0031418080907315016, + 0.00800631195306778, + 0.007491751573979855, + 0.011529693379998207, + 0.008499104529619217, + 0.009574834257364273, + 0.00756363570690155, + 0.016130607575178146, + 0.00848744809627533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0011841197265312076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst a scene rich in detailed CG art, a powerful, tight close-up shot, viewed from a slightly upward angle, captures a mysterious, cloaked Protoss warrior as he strives to complete a powerful sweep from right to left. His exposed right arm is defined by defined muscles, veins bulging, each muscle fiber tense and trembling with the sheer force of his exertion. A golden bracer adorned with ancient runes reflects the cool metallic texture in the dim light. He clutches his signature emerald blade, composed of pure psionic energy. The blade traces a dazzling, fleeting arc of light through the darkness. Its high-speed motion creates a subtle energy spill and particle effect at its edge, stirring up tiny particles suspended in the air to form ripples of light, perfectly recreating the blade's path through the air. His cloak, a heavy, intricately patterned dark blue fabric, fluttered violently. From the deep shadow of his hood, only a glimpse of emerald eyes gleamed with intense psychic energy, revealing unmistakable focus and strength. The setting was the gloomy interior of an alien temple, its cold metal walls carved with intricate geometric patterns. The psychic blade was the sole primary light source, casting a sharp, dynamic green glow, instantly illuminating the warrior's taut muscles, the folds of his cloak, and the metal reliefs on the background walls. This created a striking contrast between light and dark, and a highly dramatic effect of light and shadow. The entire scene, with its deep, cold tones, was broken by the piercing emerald green light, imbued with a tense, dynamic, and mysterious atmosphere of epic proportions, showcasing the finest CG artistry.", + "seed": 21675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21675_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0295941797317937, + "vlm_logprob": -0.0295941797317937, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.0295941797317937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02422971837222576 + ], + "opposite_logprob_tokens": [ + -3.7742297649383545 + ], + "logprob_reward": -0.02422971837222576, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02422971837222576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05962391570210457, + 0.05391325429081917, + 0.13277378678321838, + 0.16910842061042786, + 0.028088275343179703, + 0.04996116831898689, + 0.14843882620334625, + 0.20976126194000244, + 0.09781666100025177, + 0.05051443353295326 + ], + "credit_map": [ + 0.014592332765460014, + 0.015693267807364464, + 0.004160567186772823, + 0.0055475360713899136, + 0.005615958943963051, + 0.004239169880747795, + 0.004285369999706745, + 0.005489715840667486, + 0.009107568301260471, + 0.0049127875827252865, + 0.0051187751814723015, + 0.008251747116446495, + 0.006664738524705172, + 0.0057703242637217045, + 0.00804317556321621, + 0.006044143810868263, + 0.029110532253980637, + 0.019472550600767136, + 0.016093388199806213, + 0.011047444306313992, + 0.016562022268772125, + 0.010374108329415321, + 0.013320425525307655, + 0.01679331064224243, + 0.056488581001758575, + 0.011146679520606995, + 0.008928635157644749, + 0.010083463974297047, + 0.01386517845094204, + 0.010648668743669987, + 0.01164887472987175, + 0.046298328787088394, + 0.0037689032033085823, + 0.002632518531754613, + 0.003114152466878295, + 0.0034208567813038826, + 0.0033336065243929625, + 0.002859721193090081, + 0.004398340824991465, + 0.004560176748782396, + 0.0048927380703389645, + 0.0042467317543923855, + 0.0069322530180215836, + 0.008335854858160019, + 0.008192724548280239, + 0.0053857071325182915, + 0.007049380335956812, + 0.0049257781356573105, + 0.006074186414480209, + 0.011470865458250046, + 0.00935264490544796, + 0.022239068523049355, + 0.012376874685287476, + 0.014630182646214962, + 0.038420096039772034, + 0.03387492150068283, + 0.0173322893679142, + 0.06574695557355881, + 0.02240029163658619, + 0.024450533092021942, + 0.032271936535835266, + 0.013875382952392101, + 0.02344723977148533, + 0.010236631147563457, + 0.008110925555229187, + 0.03343302756547928, + 0.021585099399089813, + 0.006991730071604252, + 0.005404896102845669, + 0.005899246782064438, + 0.012452845461666584, + 0.003938890062272549, + 0.00749661959707737, + 0.007751736789941788, + 0.008405841886997223, + 0.0075496891513466835, + 0.006234881933778524, + 0.005881778430193663, + 0.004177900496870279, + 0.003015988739207387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.02422971837222576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865000769495964 + ], + "opposite_logprob_tokens": [ + -3.048650026321411 + ], + "logprob_reward": -0.04865000769495964, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865000769495964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1400459110736847, + 0.04311388358473778, + 0.0934511050581932, + 0.053945593535900116, + 0.04828812554478645, + 0.05140647664666176, + 0.08552698791027069, + 0.15546320378780365, + 0.2954554557800293, + 0.03330322727560997 + ], + "credit_map": [ + 0.04121990501880646, + 0.04692479968070984, + 0.007158943917602301, + 0.009266961365938187, + 0.010337188839912415, + 0.007876604795455933, + 0.007764171343296766, + 0.00949734915047884, + 0.008785022422671318, + 0.005700872279703617, + 0.0031242314726114273, + 0.004028549883514643, + 0.008276204578578472, + 0.0038932510651648045, + 0.004333535209298134, + 0.004972220864146948, + 0.02140148915350437, + 0.012890704907476902, + 0.011437743902206421, + 0.00853381771594286, + 0.011177959851920605, + 0.00679090665653348, + 0.009934666566550732, + 0.011283825151622295, + 0.01051598321646452, + 0.004337620921432972, + 0.005260496865957975, + 0.006586914416402578, + 0.005283279810100794, + 0.005456923972815275, + 0.005314015783369541, + 0.011190364137291908, + 0.004114288371056318, + 0.0027139808516949415, + 0.0069022588431835175, + 0.006204761099070311, + 0.0068941256031394005, + 0.005537847522646189, + 0.009541114792227745, + 0.006379749160259962, + 0.007535604760050774, + 0.005025214049965143, + 0.006843317300081253, + 0.008376211859285831, + 0.00661352789029479, + 0.004855763167142868, + 0.006986338645219803, + 0.005170508753508329, + 0.0034915239084511995, + 0.002859471831470728, + 0.004395368974655867, + 0.007309179287403822, + 0.0034016731660813093, + 0.005589780863374472, + 0.030004095286130905, + 0.02847590297460556, + 0.01431338582187891, + 0.043244704604148865, + 0.016994280740618706, + 0.02308736927807331, + 0.016842646524310112, + 0.011502258479595184, + 0.0203133262693882, + 0.009165240451693535, + 0.029430100694298744, + 0.0447939969599247, + 0.04326640069484711, + 0.026107389479875565, + 0.0421275720000267, + 0.04132896289229393, + 0.037625908851623535, + 0.03077513910830021, + 0.003710950957611203, + 0.0054114204831421375, + 0.005015865433961153, + 0.0044937231577932835, + 0.005001632962375879, + 0.004507516510784626, + 0.0029777968302369118, + 0.002184323500841856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.04865000769495964 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191400915384293 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.018191400915384293, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191400915384293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054454728960990906, + 0.0757128968834877, + 0.1484168916940689, + 0.21391256153583527, + 0.027507735416293144, + 0.05963872745633125, + 0.12349674850702286, + 0.1309165060520172, + 0.13771075010299683, + 0.028232481330633163 + ], + "credit_map": [ + 0.011993915773928165, + 0.013192074373364449, + 0.004203674383461475, + 0.006987248081713915, + 0.005029851570725441, + 0.0037984722293913364, + 0.004862539935857058, + 0.004386955872178078, + 0.012362842448055744, + 0.0077594490721821785, + 0.007941683754324913, + 0.01208999939262867, + 0.010139006190001965, + 0.0068761869333684444, + 0.011133544147014618, + 0.0074101840145885944, + 0.03112628497183323, + 0.020867688581347466, + 0.015963111072778702, + 0.012230773456394672, + 0.018807709217071533, + 0.012847448699176311, + 0.015470230020582676, + 0.02110365219414234, + 0.05442577227950096, + 0.014420217834413052, + 0.026812994852662086, + 0.020695088431239128, + 0.01897566020488739, + 0.019175773486495018, + 0.017049742862582207, + 0.04235730692744255, + 0.0024342911783605814, + 0.0018500561127439141, + 0.0023548630997538567, + 0.004893557634204626, + 0.004246640484780073, + 0.004103959538042545, + 0.0040170252323150635, + 0.003607341321185231, + 0.0058309114538133144, + 0.005504096858203411, + 0.007930309511721134, + 0.008880829438567162, + 0.009751846082508564, + 0.008031290955841541, + 0.008362348191440105, + 0.005347094032913446, + 0.0071247778832912445, + 0.0059940628707408905, + 0.010700006037950516, + 0.012309186160564423, + 0.008333826437592506, + 0.009538980200886726, + 0.0349244549870491, + 0.03457144647836685, + 0.010142854414880276, + 0.04071700945496559, + 0.009917883202433586, + 0.013279418461024761, + 0.021750204265117645, + 0.009185083210468292, + 0.01917382888495922, + 0.006750218570232391, + 0.00749242166057229, + 0.05006996542215347, + 0.03177080303430557, + 0.008334160782396793, + 0.00804924126714468, + 0.008054166100919247, + 0.019560860469937325, + 0.00437912717461586, + 0.0036152824759483337, + 0.0039528124034404755, + 0.004409910179674625, + 0.004084952641278505, + 0.0040049185045063496, + 0.003342627314850688, + 0.00290059857070446, + 0.001921377144753933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.018191400915384293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495711788535118 + ], + "opposite_logprob_tokens": [ + -2.929957151412964 + ], + "logprob_reward": -0.05495711788535118, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495711788535118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1179671362042427, + 0.0576479472219944, + 0.13005447387695312, + 0.17207598686218262, + 0.07817664742469788, + 0.14477071166038513, + 0.06517376005649567, + 0.1311044991016388, + 0.06585732102394104, + 0.03717141970992088 + ], + "credit_map": [ + 0.015338114462792873, + 0.018046297132968903, + 0.009782764129340649, + 0.015014168806374073, + 0.016610853374004364, + 0.015461638569831848, + 0.012502139434218407, + 0.015211175195872784, + 0.008656762540340424, + 0.006380307953804731, + 0.005305532831698656, + 0.008418399840593338, + 0.0066071366891264915, + 0.006963134743273258, + 0.008591687306761742, + 0.006724991369992495, + 0.028352346271276474, + 0.01763159967958927, + 0.01494507770985365, + 0.011926239356398582, + 0.0156917292624712, + 0.009681867435574532, + 0.014920745976269245, + 0.016904883086681366, + 0.025397855788469315, + 0.019000550732016563, + 0.01709773577749729, + 0.019181162118911743, + 0.018439607694745064, + 0.02353547140955925, + 0.018655691295862198, + 0.030767930671572685, + 0.007925220765173435, + 0.005644424818456173, + 0.00764688104391098, + 0.008020111359655857, + 0.013655337505042553, + 0.012614879757165909, + 0.010524631477892399, + 0.012145169079303741, + 0.02284267172217369, + 0.015570230782032013, + 0.02324826642870903, + 0.024169443175196648, + 0.011970652267336845, + 0.01390095055103302, + 0.019446320831775665, + 0.013622193597257137, + 0.0035162654239684343, + 0.004467547871172428, + 0.0058716037310659885, + 0.005852459464222193, + 0.004395502619445324, + 0.007631920278072357, + 0.016954774037003517, + 0.016483690589666367, + 0.010240286588668823, + 0.0405719019472599, + 0.017140420153737068, + 0.012129838578402996, + 0.022770032286643982, + 0.0076799034141004086, + 0.014640363864600658, + 0.005931764375418425, + 0.004813618957996368, + 0.02047751285135746, + 0.01936200074851513, + 0.0038707132916897535, + 0.003486690344288945, + 0.004555599298328161, + 0.006186655722558498, + 0.003104535862803459, + 0.005191592499613762, + 0.005341458134353161, + 0.006846978794783354, + 0.004847480915486813, + 0.0057365018874406815, + 0.004001130815595388, + 0.0028544552624225616, + 0.0023518234957009554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.05495711788535118 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019426537910476327 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019426537910476327, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019426537910476327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09219080954790115, + 0.1020696833729744, + 0.16608546674251556, + 0.17818844318389893, + 0.03384002670645714, + 0.06232263520359993, + 0.09813029319047928, + 0.1271505504846573, + 0.10370352864265442, + 0.03631848096847534 + ], + "credit_map": [ + 0.02335171215236187, + 0.025806136429309845, + 0.006239110603928566, + 0.008083410561084747, + 0.008139295503497124, + 0.0066414689645171165, + 0.0068883029744029045, + 0.007041378412395716, + 0.015083043836057186, + 0.008324499242007732, + 0.008941089734435081, + 0.014857576228678226, + 0.013762641698122025, + 0.013189626857638359, + 0.01590629667043686, + 0.012004902586340904, + 0.03546348959207535, + 0.024336354807019234, + 0.01989142782986164, + 0.013631025329232216, + 0.020889660343527794, + 0.012937217950820923, + 0.017250042408704758, + 0.021686255931854248, + 0.033948659896850586, + 0.014929882250726223, + 0.020704450085759163, + 0.022664392367005348, + 0.020458074286580086, + 0.022687243297696114, + 0.015682147815823555, + 0.027113601565361023, + 0.002896802267059684, + 0.0026121214032173157, + 0.003923608921468258, + 0.00474547827616334, + 0.006320514716207981, + 0.004563493188470602, + 0.0038071114104241133, + 0.004970897454768419, + 0.0064844610169529915, + 0.0052925883792340755, + 0.007798363454639912, + 0.011212043464183807, + 0.008623883128166199, + 0.00800948217511177, + 0.009113973937928677, + 0.005787841510027647, + 0.0035186896566301584, + 0.004841217305511236, + 0.005717924330383539, + 0.009248527698218822, + 0.004774641711264849, + 0.007018854841589928, + 0.03234558179974556, + 0.030664855614304543, + 0.010168333537876606, + 0.043664686381816864, + 0.007566260639578104, + 0.010611281730234623, + 0.021067271009087563, + 0.0074653890915215015, + 0.01999286562204361, + 0.006614460609853268, + 0.007448185700923204, + 0.03554750233888626, + 0.01654736138880253, + 0.0075867497362196445, + 0.008020208217203617, + 0.005832942668348551, + 0.01793072372674942, + 0.004789853468537331, + 0.004240875132381916, + 0.0040786899626255035, + 0.00772976316511631, + 0.004496661480516195, + 0.004768021870404482, + 0.005481276195496321, + 0.003022450255230069, + 0.0025007419753819704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.0019426537910476327 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst a scene rich in detailed CG art, a powerful, tight close-up shot, viewed from a slightly upward angle, captures a mysterious, cloaked Protoss warrior as he strives to complete a powerful sweep from right to left. His exposed right arm is defined by defined muscles, veins bulging, each muscle fiber tense and trembling with the sheer force of his exertion. A golden bracer adorned with ancient runes reflects the cool metallic texture in the dim light. He clutches his signature emerald blade, composed of pure psionic energy. The blade traces a dazzling, fleeting arc of light through the darkness. Its high-speed motion creates a subtle energy spill and particle effect at its edge, stirring up tiny particles suspended in the air to form ripples of light, perfectly recreating the blade's path through the air. His cloak, a heavy, intricately patterned dark blue fabric, fluttered violently. From the deep shadow of his hood, only a glimpse of emerald eyes gleamed with intense psychic energy, revealing unmistakable focus and strength. The setting was the gloomy interior of an alien temple, its cold metal walls carved with intricate geometric patterns. The psychic blade was the sole primary light source, casting a sharp, dynamic green glow, instantly illuminating the warrior's taut muscles, the folds of his cloak, and the metal reliefs on the background walls. This created a striking contrast between light and dark, and a highly dramatic effect of light and shadow. The entire scene, with its deep, cold tones, was broken by the piercing emerald green light, imbued with a tense, dynamic, and mysterious atmosphere of epic proportions, showcasing the finest CG artistry.", + "seed": 21676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_7/index_2423_seed_21676_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04672937884461135, + "vlm_logprob": -0.04672937884461135, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04672937884461135, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07223831117153168 + ], + "opposite_logprob_tokens": [ + -2.6972382068634033 + ], + "logprob_reward": -0.07223831117153168, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07223831117153168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.142563596367836, + 0.05795258656144142, + 0.030872197821736336, + 0.08804701268672943, + 0.030942842364311218, + 0.06311188638210297, + 0.08270126581192017, + 0.11052680015563965, + 0.33877620100975037, + 0.054505690932273865 + ], + "credit_map": [ + 0.019282637163996696, + 0.012068893760442734, + 0.02454971894621849, + 0.030614029616117477, + 0.02482418715953827, + 0.011078613810241222, + 0.009875396266579628, + 0.010270125232636929, + 0.004515926353633404, + 0.003364695468917489, + 0.013950612396001816, + 0.020084284245967865, + 0.00290796160697937, + 0.00375960860401392, + 0.0061479550786316395, + 0.0032215372193604708, + 0.003384718671441078, + 0.0026489191222935915, + 0.0035548394080251455, + 0.0031226088758558035, + 0.004293729551136494, + 0.00465493043884635, + 0.0042847939766943455, + 0.004927659407258034, + 0.0032047696877270937, + 0.0030283345840871334, + 0.005012481939047575, + 0.008722996339201927, + 0.003528259228914976, + 0.0036462151911109686, + 0.035112641751766205, + 0.025791309773921967, + 0.004257393069565296, + 0.0034758097026497126, + 0.005266430322080851, + 0.0035371626727283, + 0.003396130632609129, + 0.00532136857509613, + 0.0022331811487674713, + 0.003455367172136903, + 0.0051981499418616295, + 0.008387590758502483, + 0.021225610747933388, + 0.004485822282731533, + 0.004600509535521269, + 0.010847222059965134, + 0.005451463162899017, + 0.0029155216179788113, + 0.028288841247558594, + 0.007142846938222647, + 0.004824726842343807, + 0.006052840035408735, + 0.002812879392877221, + 0.008033332414925098, + 0.004633464850485325, + 0.020912330597639084, + 0.018715469166636467, + 0.011800323612987995, + 0.015295350924134254, + 0.011681209318339825, + 0.012638577260077, + 0.011309381574392319, + 0.015055892989039421, + 0.014030593447387218, + 0.026685066521167755, + 0.019672023132443428, + 0.018969714641571045, + 0.03949996456503868, + 0.059739142656326294, + 0.05960256978869438, + 0.06605757772922516, + 0.04855015501379967, + 0.005502109415829182, + 0.005518284160643816, + 0.0071897953748703, + 0.00653824070468545, + 0.008857755921781063, + 0.006467226892709732, + 0.00797160342335701, + 0.006460676435381174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.07223831117153168 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866238683462143 + ], + "opposite_logprob_tokens": [ + -3.0486624240875244 + ], + "logprob_reward": -0.04866238683462143, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866238683462143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12101726233959198, + 0.08299779891967773, + 0.03950304165482521, + 0.04299429431557655, + 0.027159258723258972, + 0.08373293280601501, + 0.1311274766921997, + 0.09122253954410553, + 0.1447213739156723, + 0.23552393913269043 + ], + "credit_map": [ + 0.01881762593984604, + 0.01045334804803133, + 0.01888810470700264, + 0.029270516708493233, + 0.014433995820581913, + 0.007590990513563156, + 0.010813146829605103, + 0.01074954029172659, + 0.007644419092684984, + 0.006873588543385267, + 0.01894107088446617, + 0.013657745905220509, + 0.007568806875497103, + 0.008420431986451149, + 0.00885014608502388, + 0.011041593737900257, + 0.003447919152677059, + 0.0023103496059775352, + 0.004333208780735731, + 0.005842641927301884, + 0.005559667944908142, + 0.005323885940015316, + 0.008396810851991177, + 0.004288562573492527, + 0.004520250018686056, + 0.004076613113284111, + 0.005605763755738735, + 0.0056844609789550304, + 0.0036875836085528135, + 0.0032791858538985252, + 0.011545151472091675, + 0.0045952871441841125, + 0.004155775066465139, + 0.0037592072039842606, + 0.0031717310193926096, + 0.003697552951052785, + 0.0038542135152965784, + 0.0035134123172611, + 0.001978643238544464, + 0.003028727835044265, + 0.008436091244220734, + 0.012312306091189384, + 0.029779542237520218, + 0.004801678936928511, + 0.005759605206549168, + 0.012498454190790653, + 0.006722099147737026, + 0.003423168556764722, + 0.03679411858320236, + 0.010038328357040882, + 0.007031164597719908, + 0.008496677502989769, + 0.00579436169937253, + 0.018840162083506584, + 0.008386204019188881, + 0.035746484994888306, + 0.017364228144288063, + 0.013423368334770203, + 0.013110755011439323, + 0.0104415537789464, + 0.005988254211843014, + 0.007807221729308367, + 0.01032737921923399, + 0.012759796343743801, + 0.025176463648676872, + 0.009647904895246029, + 0.01335958018898964, + 0.014120077714323997, + 0.018691187724471092, + 0.0165675338357687, + 0.021960852667689323, + 0.025197798386216164, + 0.012257246300578117, + 0.00980378221720457, + 0.014163876883685589, + 0.09212511777877808, + 0.009304584003984928, + 0.007216030731797218, + 0.07867401838302612, + 0.011979312635958195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.04866238683462143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082758195698261 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082758195698261, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082758195698261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31203821301460266, + 0.047921910881996155, + 0.029687494039535522, + 0.1178058385848999, + 0.07608368992805481, + 0.09675516188144684, + 0.07246895879507065, + 0.08037325739860535, + 0.11544804275035858, + 0.05141745135188103 + ], + "credit_map": [ + 0.04777396097779274, + 0.026499804109334946, + 0.048143476247787476, + 0.06940106302499771, + 0.05229140445590019, + 0.021923620253801346, + 0.022174665704369545, + 0.023830216377973557, + 0.005412353202700615, + 0.004627929534763098, + 0.008310764096677303, + 0.007241528481245041, + 0.007319287396967411, + 0.004301133099943399, + 0.006246547680348158, + 0.0044623673893511295, + 0.0027347879949957132, + 0.0021606332156807184, + 0.0032471921294927597, + 0.004930511116981506, + 0.00430422043427825, + 0.0037076992448419333, + 0.004281741101294756, + 0.0043207090348005295, + 0.007696019485592842, + 0.0073682465590536594, + 0.008107365109026432, + 0.014007343910634518, + 0.007442934904247522, + 0.005303156096488237, + 0.041677914559841156, + 0.026202857494354248, + 0.009861409664154053, + 0.008556508459150791, + 0.00913972221314907, + 0.015492322854697704, + 0.010354982689023018, + 0.0092593003064394, + 0.005742110777646303, + 0.007677333429455757, + 0.010251651518046856, + 0.012204883620142937, + 0.02351665124297142, + 0.00868268497288227, + 0.008462999947369099, + 0.015112821012735367, + 0.011064562946557999, + 0.007458893582224846, + 0.022260751575231552, + 0.006920143496245146, + 0.005682006012648344, + 0.006841035559773445, + 0.0024992041289806366, + 0.006423294544219971, + 0.0039002865087240934, + 0.017942234873771667, + 0.012096395716071129, + 0.007656851317733526, + 0.010072850622236729, + 0.010294569656252861, + 0.00619053328409791, + 0.010984892956912518, + 0.012362156063318253, + 0.010715010575950146, + 0.013772676698863506, + 0.0070706927217543125, + 0.00945933535695076, + 0.014701557345688343, + 0.018577173352241516, + 0.015687119215726852, + 0.019548920914530754, + 0.016630567610263824, + 0.0050176712684333324, + 0.0057335360907018185, + 0.006116976961493492, + 0.007684097625315189, + 0.005349999293684959, + 0.006534070707857609, + 0.007475156802684069, + 0.007505943067371845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.011082758195698261 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032466053962708 + ], + "opposite_logprob_tokens": [ + -2.3503246307373047 + ], + "logprob_reward": -0.10032466053962708, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032466053962708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08807358145713806, + 0.12144862860441208, + 0.030913634225726128, + 0.02057260274887085, + 0.030159233137965202, + 0.05490284785628319, + 0.09502644091844559, + 0.1132473275065422, + 0.3141258656978607, + 0.1315297931432724 + ], + "credit_map": [ + 0.01583169400691986, + 0.006637828890234232, + 0.014125889167189598, + 0.019785329699516296, + 0.012807521969079971, + 0.006147370673716068, + 0.006907158996909857, + 0.005830797832459211, + 0.010957915335893631, + 0.007837465032935143, + 0.02951149269938469, + 0.03948254510760307, + 0.008006734773516655, + 0.00938747264444828, + 0.01108098030090332, + 0.005184037145227194, + 0.0018414376536384225, + 0.0017702850745990872, + 0.002601131098344922, + 0.007529486436396837, + 0.0050737932324409485, + 0.002808231394737959, + 0.004648710135370493, + 0.004640561528503895, + 0.001160676358267665, + 0.0011595780961215496, + 0.003663664683699608, + 0.0025132771115750074, + 0.001375306281261146, + 0.0014696536818519235, + 0.0052863797172904015, + 0.003944069147109985, + 0.004321760963648558, + 0.003149530151858926, + 0.0035475664772093296, + 0.0030658359173685312, + 0.003609726205468178, + 0.004434417001903057, + 0.0032559500541538, + 0.004774449393153191, + 0.005108668934553862, + 0.00807756744325161, + 0.020216146484017372, + 0.0036305931862443686, + 0.0029575845692306757, + 0.00852132122963667, + 0.00431788619607687, + 0.0020730914548039436, + 0.017643356695771217, + 0.00990606751292944, + 0.010655662044882774, + 0.010636483319103718, + 0.0035706518683582544, + 0.010346070863306522, + 0.00919174775481224, + 0.023076405748724937, + 0.022945232689380646, + 0.012189606204628944, + 0.01650390960276127, + 0.01072988286614418, + 0.012504500336945057, + 0.013619874604046345, + 0.01038124319165945, + 0.014373078010976315, + 0.06369362771511078, + 0.022558603435754776, + 0.028309032320976257, + 0.02645406313240528, + 0.02359614707529545, + 0.03834562003612518, + 0.050799671560525894, + 0.06036915257573128, + 0.007944994606077671, + 0.00707382382825017, + 0.009977094829082489, + 0.04534337297081947, + 0.009159358218312263, + 0.006796290632337332, + 0.038725510239601135, + 0.006509358994662762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.10032466053962708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cloaked Protoss warrior with an exposed right arm showing defined muscles and veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a heavy, intricately patterned dark blue cloak that is fluttering violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior holding a signature emerald blade composed of pure psionic energy that creates a dazzling arc of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the gloomy interior of an alien temple with cold metal walls carved with intricate geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit primarily by the psychic blade casting a sharp, dynamic green glow that illuminates the warrior and background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013387774815782905 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013387774815782905, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013387774815782905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6071159839630127, + 0.02443176880478859, + 0.012545679695904255, + 0.01590936817228794, + 0.02167484536767006, + 0.04319688677787781, + 0.06198161095380783, + 0.05667727440595627, + 0.07572750002145767, + 0.08073905855417252 + ], + "credit_map": [ + 0.09990998357534409, + 0.049405913800001144, + 0.09572232514619827, + 0.1372615098953247, + 0.09517985582351685, + 0.03960813954472542, + 0.04358794167637825, + 0.046440351754426956, + 0.0025895191356539726, + 0.0019671074114739895, + 0.0037158771883696318, + 0.006568557117134333, + 0.002105658408254385, + 0.001936801359988749, + 0.0031409619841724634, + 0.002407287945970893, + 0.0013235248625278473, + 0.0008618356077931821, + 0.0010908092372119427, + 0.002517673186957836, + 0.0016479099867865443, + 0.001310016610659659, + 0.0017101522535085678, + 0.002083758357912302, + 0.001655648578889668, + 0.00196319492533803, + 0.004117039497941732, + 0.0020506400614976883, + 0.0016815293347463012, + 0.0014121680287644267, + 0.0016850685933604836, + 0.00134407973382622, + 0.0026610700879245996, + 0.0027140334714204073, + 0.003120614681392908, + 0.002439145464450121, + 0.0029240392614156008, + 0.0031557034235447645, + 0.001591469976119697, + 0.003068768884986639, + 0.004559485241770744, + 0.005592070985585451, + 0.014508344233036041, + 0.00352164707146585, + 0.002945312298834324, + 0.005262891296297312, + 0.004855057690292597, + 0.0019520820351317525, + 0.0196990929543972, + 0.003892261767759919, + 0.004039538558572531, + 0.00468783313408494, + 0.002258623018860817, + 0.006414703093469143, + 0.004616043996065855, + 0.016373513266444206, + 0.006714516319334507, + 0.007010824047029018, + 0.0066352542489767075, + 0.0075499978847801685, + 0.00556309474632144, + 0.005606794264167547, + 0.010387808084487915, + 0.007208988070487976, + 0.01415803749114275, + 0.005449584219604731, + 0.006364056374877691, + 0.010367067530751228, + 0.006120934616774321, + 0.009418812580406666, + 0.010702790692448616, + 0.013146218843758106, + 0.004525473341345787, + 0.006476733833551407, + 0.005164270754903555, + 0.02611703984439373, + 0.005970227066427469, + 0.004480937495827675, + 0.023422321304678917, + 0.004582058638334274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0013387774815782905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_9/index_tensor([2766, 2766])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_9/index_tensor([2766, 2766])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e6e822499c1f58290906a9ed7688e169713bb4a7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_9/index_tensor([2766, 2766])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "graceful, slow, 360-degree medium shot rotates smoothly and poetically around an ancient poet. The poet stands alone on a vast black rock by the river. He wears a light, wide-sleeved white silk robe. The night breeze blows, causing the sleeves and hem to flutter violently, as if he is about to ride away. His long, jet-black hair is loosely tied back with a simple wooden hairpin, a few strands of hair blowing across his face. His sharply defined face appears exceptionally thin in the sidelight of the bright moon. His brows are slightly furrowed, his eyes fixed on the vast, shimmering river before him. His lips move in sync with the silent verses of his poem, his expression a fusion of deep sorrow and creative passion. The backdrop is a deep navy blue night sky, a full moon hanging high in the sky. Its cool, silvery glow provides the sole illumination of the scene, stained the poet's white robe with a radiant snow-white brilliance, casting soft shadows on the wet rocks, and spreading a fragmented, flowing silver light across the vast river. On the far bank, rolling hills appear in dark silhouette, like an ink painting. The steady rotation of the camera creates a tranquil atmosphere of grandeur, loneliness, yet a sense of romance and power.", + "seed": 20022, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_9/index_2766_seed_20022_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.036196162551641466, + "vlm_logprob": -0.036196162551641466, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.036196162551641466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03574220463633537 + ], + "opposite_logprob_tokens": [ + -3.5357422828674316 + ], + "logprob_reward": -0.03574220463633537, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03574220463633537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08720938116312027, + 0.027974281460046768, + 0.04816335812211037, + 0.027647659182548523, + 0.03276803717017174, + 0.10391102731227875, + 0.08914139121770859, + 0.3838094174861908, + 0.07924382388591766, + 0.12013159692287445 + ], + "credit_map": [ + 0.009915037080645561, + 0.013218743726611137, + 0.010907403193414211, + 0.0067179384641349316, + 0.00888411607593298, + 0.01691371761262417, + 0.0053251623176038265, + 0.015327266417443752, + 0.003094908082857728, + 0.0035194214433431625, + 0.0030368738807737827, + 0.0017173641826957464, + 0.0032025519758462906, + 0.007191563490778208, + 0.002524737734347582, + 0.003686861600726843, + 0.006680781487375498, + 0.006188591010868549, + 0.005371096543967724, + 0.005362207535654306, + 0.0055268616415560246, + 0.00640453863888979, + 0.006934367120265961, + 0.005694918800145388, + 0.0030307236593216658, + 0.004795357119292021, + 0.005774817895144224, + 0.0027163636405020952, + 0.0025247337762266397, + 0.0033914733212441206, + 0.0021413140930235386, + 0.0032728766091167927, + 0.004063154570758343, + 0.007013005670160055, + 0.003941742703318596, + 0.00258066738024354, + 0.003776074619963765, + 0.005302275996655226, + 0.002831263467669487, + 0.0032598546240478754, + 0.014076956547796726, + 0.02480960451066494, + 0.009962803684175014, + 0.011172431521117687, + 0.009374990127980709, + 0.02337242104113102, + 0.005901044234633446, + 0.005240787286311388, + 0.013057289645075798, + 0.022982535883784294, + 0.006576680578291416, + 0.004509032238274813, + 0.006554280407726765, + 0.025546973571181297, + 0.0037662405520677567, + 0.006148369051516056, + 0.045312680304050446, + 0.07284395396709442, + 0.06958182156085968, + 0.019793793559074402, + 0.03239303082227707, + 0.0626019611954689, + 0.016505351290106773, + 0.06477686017751694, + 0.013570492155849934, + 0.014607875607907772, + 0.005953311920166016, + 0.002693576505407691, + 0.007940392941236496, + 0.02392115257680416, + 0.003936991561204195, + 0.006620040163397789, + 0.010612602345645428, + 0.05016843229532242, + 0.005070607177913189, + 0.0025983676314353943, + 0.005515089724212885, + 0.03586139157414436, + 0.0021925540640950203, + 0.008112568408250809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.03574220463633537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794510900974274 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794510900974274, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794510900974274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08446367084980011, + 0.07774531841278076, + 0.09714973717927933, + 0.045675553381443024, + 0.08928234875202179, + 0.13038036227226257, + 0.1598784476518631, + 0.1309482902288437, + 0.062152836471796036, + 0.12232339382171631 + ], + "credit_map": [ + 0.01054142601788044, + 0.012357600964605808, + 0.008379210717976093, + 0.006200367119163275, + 0.010490133427083492, + 0.016758853569626808, + 0.0056827194057404995, + 0.014053368009626865, + 0.007114375941455364, + 0.025628942996263504, + 0.004505602177232504, + 0.002374232280999422, + 0.007010023575276136, + 0.01081235520541668, + 0.0025274446234107018, + 0.01777234859764576, + 0.010361663065850735, + 0.010634807869791985, + 0.014483047649264336, + 0.005711343139410019, + 0.006492219399660826, + 0.01436594221740961, + 0.010337656363844872, + 0.024763068184256554, + 0.004566786345094442, + 0.009653099812567234, + 0.006902826018631458, + 0.0026108683086931705, + 0.00437750481069088, + 0.0059309727512300014, + 0.003962500020861626, + 0.007670997641980648, + 0.01205984503030777, + 0.014299474656581879, + 0.012902477756142616, + 0.009383920580148697, + 0.008128087967634201, + 0.017132468521595, + 0.007929611951112747, + 0.007446470204740763, + 0.0188722126185894, + 0.027538759633898735, + 0.023316381499171257, + 0.009861040860414505, + 0.008672581054270267, + 0.023590940982103348, + 0.0066047064028680325, + 0.011923746205866337, + 0.027797076851129532, + 0.04342379420995712, + 0.010950954630970955, + 0.00985280517488718, + 0.012530253268778324, + 0.04111538454890251, + 0.007196397054940462, + 0.00701179401949048, + 0.014523752965033054, + 0.02300577610731125, + 0.022984391078352928, + 0.007646470796316862, + 0.011580152437090874, + 0.017949063330888748, + 0.008196073584258556, + 0.025062616914510727, + 0.005291922017931938, + 0.015793468803167343, + 0.0053830440156161785, + 0.0030696175526827574, + 0.007270432077348232, + 0.015743743628263474, + 0.003999968525022268, + 0.0056006405502557755, + 0.014978131279349327, + 0.049980852752923965, + 0.0077803004533052444, + 0.0027620119508355856, + 0.0076097967103123665, + 0.030778972432017326, + 0.002631555777043104, + 0.005801782477647066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.009794510900974274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015193596482276917 + ], + "opposite_logprob_tokens": [ + -6.501519203186035 + ], + "logprob_reward": -0.0015193596482276917, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015193596482276917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19634580612182617, + 0.10063530504703522, + 0.07296113669872284, + 0.027716943994164467, + 0.04529384896159172, + 0.0491795614361763, + 0.1922920048236847, + 0.19902703166007996, + 0.05644562467932701, + 0.06010273098945618 + ], + "credit_map": [ + 0.02536354959011078, + 0.02094186469912529, + 0.018412796780467033, + 0.011745957657694817, + 0.02063160017132759, + 0.05466968193650246, + 0.011288381181657314, + 0.03329196944832802, + 0.011109445244073868, + 0.014311039820313454, + 0.010830066166818142, + 0.005497191101312637, + 0.00990784727036953, + 0.02926270104944706, + 0.005961970891803503, + 0.013755043968558311, + 0.007693619932979345, + 0.009652727283537388, + 0.010821291245520115, + 0.004679796285927296, + 0.005473772529512644, + 0.010149450041353703, + 0.00811520591378212, + 0.01637526974081993, + 0.002366787288337946, + 0.004014225676655769, + 0.003890176070854068, + 0.00228302925825119, + 0.00336873228661716, + 0.0038251159712672234, + 0.003036808455362916, + 0.004932069219648838, + 0.004101225640624762, + 0.0062635550275444984, + 0.006365818902850151, + 0.006157887168228626, + 0.004784048069268465, + 0.0061583137139678, + 0.005169677548110485, + 0.00629332335665822, + 0.0050618951208889484, + 0.010408400557935238, + 0.005595172755420208, + 0.0051221647299826145, + 0.004814461804926395, + 0.008727297186851501, + 0.003366453805938363, + 0.0060837147757411, + 0.01974637247622013, + 0.060476403683423996, + 0.011014057323336601, + 0.008454433642327785, + 0.012209835462272167, + 0.06506063044071198, + 0.00742658507078886, + 0.007903684861958027, + 0.022427957504987717, + 0.03589635342359543, + 0.036027755588293076, + 0.011146609671413898, + 0.018559100106358528, + 0.02986297383904457, + 0.010551304556429386, + 0.03455498442053795, + 0.005139967426657677, + 0.011753548868000507, + 0.006873355247080326, + 0.0024821169208735228, + 0.00352716026827693, + 0.014999213628470898, + 0.0027489918284118176, + 0.008921267464756966, + 0.007052550558000803, + 0.022797739133238792, + 0.0040352134965360165, + 0.0019048359245061874, + 0.004268948920071125, + 0.014289560727775097, + 0.0019094495801255107, + 0.0038444343954324722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0015193596482276917 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495666712522507 + ], + "opposite_logprob_tokens": [ + -2.9299566745758057 + ], + "logprob_reward": -0.05495666712522507, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495666712522507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09877952188253403, + 0.04447405785322189, + 0.07585868239402771, + 0.04829145595431328, + 0.030231596902012825, + 0.04112230986356735, + 0.2438642531633377, + 0.05715339258313179, + 0.03587319329380989, + 0.32435154914855957 + ], + "credit_map": [ + 0.014727645553648472, + 0.014040689915418625, + 0.012522906996309757, + 0.007283560000360012, + 0.009992633014917374, + 0.01815648563206196, + 0.006196913775056601, + 0.015858683735132217, + 0.005679690279066563, + 0.009156432002782822, + 0.0038789049722254276, + 0.0020679659210145473, + 0.00396262900903821, + 0.010000723414123058, + 0.0024414751678705215, + 0.007286238018423319, + 0.008489757776260376, + 0.012609733268618584, + 0.006981366779655218, + 0.007109570782631636, + 0.006891746073961258, + 0.012023327872157097, + 0.00872617494314909, + 0.013027004897594452, + 0.0076054781675338745, + 0.008210326544940472, + 0.006087911780923605, + 0.0041867149993777275, + 0.005098021123558283, + 0.005269639194011688, + 0.0057632531970739365, + 0.006070107687264681, + 0.004085409455001354, + 0.005377626046538353, + 0.003849285189062357, + 0.002720878692343831, + 0.002725004218518734, + 0.004261948633939028, + 0.0028499451000243425, + 0.0043615004979074, + 0.007387869991362095, + 0.0071635060012340546, + 0.005951488856226206, + 0.0033672545105218887, + 0.0034711265470832586, + 0.006824980489909649, + 0.0022372049279510975, + 0.00471887830644846, + 0.05143309384584427, + 0.05262863636016846, + 0.024149399250745773, + 0.014121083542704582, + 0.023496713489294052, + 0.04739633575081825, + 0.007576902396976948, + 0.023062098771333694, + 0.005560935474932194, + 0.012807420454919338, + 0.006163258105516434, + 0.00260677351616323, + 0.005438493564724922, + 0.01344109233468771, + 0.0032441075891256332, + 0.007891313172876835, + 0.00380862969905138, + 0.006279869936406612, + 0.004649643786251545, + 0.0016171857714653015, + 0.004478614777326584, + 0.006878303829580545, + 0.0024681491777300835, + 0.0056927925907075405, + 0.03326709568500519, + 0.1417006552219391, + 0.010657751932740211, + 0.006650135852396488, + 0.01524763461202383, + 0.10178975015878677, + 0.004649623762816191, + 0.010388885624706745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.05495666712522507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896807044744492 + ], + "opposite_logprob_tokens": [ + -2.578968048095703 + ], + "logprob_reward": -0.07896807044744492, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896807044744492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06458240747451782, + 0.0514412485063076, + 0.10182488709688187, + 0.044777799397706985, + 0.0427311472594738, + 0.12724192440509796, + 0.24605123698711395, + 0.11130888760089874, + 0.07772607356309891, + 0.13231435418128967 + ], + "credit_map": [ + 0.009394547902047634, + 0.010220278054475784, + 0.006056994665414095, + 0.00541471503674984, + 0.007734834216535091, + 0.011903663165867329, + 0.0039839111268520355, + 0.00987346563488245, + 0.004608646035194397, + 0.01385441143065691, + 0.003411120967939496, + 0.002375462092459202, + 0.004353450145572424, + 0.00884533952921629, + 0.003265533596277237, + 0.01072729006409645, + 0.011508981697261333, + 0.010131108574569225, + 0.013989933766424656, + 0.005919777322560549, + 0.008262536488473415, + 0.013227900490164757, + 0.010040287859737873, + 0.028744367882609367, + 0.005633410066366196, + 0.008782872930169106, + 0.004082727711647749, + 0.004101613536477089, + 0.006224941927939653, + 0.005498308688402176, + 0.004636185709387064, + 0.005817739758640528, + 0.003628392005339265, + 0.005517043173313141, + 0.004659207537770271, + 0.007143757771700621, + 0.005330304615199566, + 0.006662183906883001, + 0.005888842511922121, + 0.003901421558111906, + 0.01177779957652092, + 0.030162999406456947, + 0.012647052295506, + 0.01145875733345747, + 0.010378682054579258, + 0.03206363320350647, + 0.008486652746796608, + 0.010266353376209736, + 0.04575037583708763, + 0.051494013518095016, + 0.022618766874074936, + 0.01486554741859436, + 0.023869624361395836, + 0.052306242287158966, + 0.00917990505695343, + 0.025966782122850418, + 0.015688957646489143, + 0.01886497624218464, + 0.016917072236537933, + 0.005733471363782883, + 0.01157260313630104, + 0.016927402466535568, + 0.007445129565894604, + 0.018159279599785805, + 0.00918755866587162, + 0.00940178707242012, + 0.01123668160289526, + 0.00413656048476696, + 0.01391412690281868, + 0.011378324590623379, + 0.006771364249289036, + 0.0116996755823493, + 0.019137367606163025, + 0.027165701612830162, + 0.007206573151051998, + 0.0036176645662635565, + 0.009423954412341118, + 0.050511859357357025, + 0.003912557847797871, + 0.011338676325976849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.07896807044744492 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "graceful, slow, 360-degree medium shot rotates smoothly and poetically around an ancient poet. The poet stands alone on a vast black rock by the river. He wears a light, wide-sleeved white silk robe. The night breeze blows, causing the sleeves and hem to flutter violently, as if he is about to ride away. His long, jet-black hair is loosely tied back with a simple wooden hairpin, a few strands of hair blowing across his face. His sharply defined face appears exceptionally thin in the sidelight of the bright moon. His brows are slightly furrowed, his eyes fixed on the vast, shimmering river before him. His lips move in sync with the silent verses of his poem, his expression a fusion of deep sorrow and creative passion. The backdrop is a deep navy blue night sky, a full moon hanging high in the sky. Its cool, silvery glow provides the sole illumination of the scene, stained the poet's white robe with a radiant snow-white brilliance, casting soft shadows on the wet rocks, and spreading a fragmented, flowing silver light across the vast river. On the far bank, rolling hills appear in dark silhouette, like an ink painting. The steady rotation of the camera creates a tranquil atmosphere of grandeur, loneliness, yet a sense of romance and power.", + "seed": 20023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_9/index_2766_seed_20023_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1018898301757872, + "vlm_logprob": -0.1018898301757872, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.1018898301757872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015937793999910355 + ], + "opposite_logprob_tokens": [ + -4.265937805175781 + ], + "logprob_reward": -0.015937793999910355, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015937793999910355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2273963987827301, + 0.3366028964519501, + 0.07537327706813812, + 0.02685580402612686, + 0.06778297573328018, + 0.059040367603302, + 0.07632578909397125, + 0.06377674639225006, + 0.026397988200187683, + 0.04044773802161217 + ], + "credit_map": [ + 0.04111049696803093, + 0.04077111557126045, + 0.03397596627473831, + 0.019575579091906548, + 0.024471241980791092, + 0.032696619629859924, + 0.01982417330145836, + 0.014971195720136166, + 0.05164465308189392, + 0.08006291091442108, + 0.049160197377204895, + 0.023128163069486618, + 0.03689467906951904, + 0.04265213757753372, + 0.022876109927892685, + 0.030184078961610794, + 0.0066342661157250404, + 0.010263746604323387, + 0.009183228947222233, + 0.007973510771989822, + 0.013651163317263126, + 0.011773215606808662, + 0.010145164094865322, + 0.005748985335230827, + 0.0030006980523467064, + 0.0033940968569368124, + 0.004400504287332296, + 0.002237854525446892, + 0.004879445768892765, + 0.0040078298188745975, + 0.0031491019763052464, + 0.0017862730892375112, + 0.007264717947691679, + 0.008260654285550117, + 0.012292315252125263, + 0.004487805999815464, + 0.008576382882893085, + 0.010253384709358215, + 0.005589752458035946, + 0.011057959869503975, + 0.009248543530702591, + 0.005745727103203535, + 0.007801878731697798, + 0.004660980310291052, + 0.009888691827654839, + 0.008748198859393597, + 0.004827930126339197, + 0.008118409663438797, + 0.0069743562489748, + 0.010221448726952076, + 0.009689218364655972, + 0.008136805146932602, + 0.013273744843900204, + 0.010349440388381481, + 0.009654502384364605, + 0.008026272989809513, + 0.0052507370710372925, + 0.007170622702687979, + 0.00879912544041872, + 0.0030895338859409094, + 0.005822451785206795, + 0.011220780201256275, + 0.004262211732566357, + 0.018161287531256676, + 0.0029250490479171276, + 0.0040345084853470325, + 0.0023198635317385197, + 0.0033424950670450926, + 0.003428124589845538, + 0.0034682955592870712, + 0.0037769663613289595, + 0.00310268672183156, + 0.006281854584813118, + 0.006333696190267801, + 0.004094351548701525, + 0.003901405492797494, + 0.0061338781379163265, + 0.005628162994980812, + 0.003483045380562544, + 0.004591341596096754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.015937793999910355 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22593359649181366, + 0.3546096682548523, + 0.09584885090589523, + 0.0347251333296299, + 0.04481656104326248, + 0.046050090342760086, + 0.037458259612321854, + 0.05958694964647293, + 0.03910398110747337, + 0.061866890639066696 + ], + "credit_map": [ + 0.028090786188840866, + 0.024245524778962135, + 0.046035829931497574, + 0.014084101654589176, + 0.036456700414419174, + 0.0349116288125515, + 0.01944340206682682, + 0.022665604948997498, + 0.052378639578819275, + 0.0853273868560791, + 0.04830123484134674, + 0.023932870477437973, + 0.041218169033527374, + 0.044709213078022, + 0.023435594514012337, + 0.035306550562381744, + 0.010009456425905228, + 0.012112549506127834, + 0.01278349943459034, + 0.010341852903366089, + 0.015507912263274193, + 0.011463026516139507, + 0.012626653537154198, + 0.011003904975950718, + 0.003944799769669771, + 0.0038900042418390512, + 0.005471880082041025, + 0.0035460535436868668, + 0.007059486582875252, + 0.004446167964488268, + 0.003883846802636981, + 0.002482892479747534, + 0.004994323011487722, + 0.005591264460235834, + 0.008468163199722767, + 0.004073335789144039, + 0.006360565312206745, + 0.005547882057726383, + 0.004686285275965929, + 0.0050947400741279125, + 0.0049803550355136395, + 0.0060089873149991035, + 0.008393450640141964, + 0.0038542533293366432, + 0.006587512791156769, + 0.00742653664201498, + 0.004070768598467112, + 0.0047282264567911625, + 0.0031192877795547247, + 0.005416005849838257, + 0.005460130982100964, + 0.0032149464823305607, + 0.006561208516359329, + 0.005194355268031359, + 0.004002422094345093, + 0.004489902872592211, + 0.005346374586224556, + 0.0076943691819906235, + 0.008069015108048916, + 0.005925030447542667, + 0.0059629175812006, + 0.008215009234845638, + 0.005718739237636328, + 0.01265549287199974, + 0.003493267111480236, + 0.005360163748264313, + 0.004630671814084053, + 0.0038406301755458117, + 0.004769727122038603, + 0.005172113422304392, + 0.005498660262674093, + 0.006338747683912516, + 0.010787409730255604, + 0.010383316315710545, + 0.010017430409789085, + 0.003986489959061146, + 0.00501971784979105, + 0.009157225489616394, + 0.004088298417627811, + 0.008427003398537636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.014206582680344582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003198272082954645 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.003198272082954645, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003198272082954645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11199592053890228, + 0.5254203677177429, + 0.0654497966170311, + 0.035718146711587906, + 0.07182344049215317, + 0.027981601655483246, + 0.02941315621137619, + 0.06379172205924988, + 0.029153253883123398, + 0.039252620190382004 + ], + "credit_map": [ + 0.01571574993431568, + 0.015370928682386875, + 0.02494506724178791, + 0.007010109722614288, + 0.014935864135622978, + 0.016266725957393646, + 0.008390586823225021, + 0.009360874071717262, + 0.07920800149440765, + 0.12613503634929657, + 0.07136809080839157, + 0.035000015050172806, + 0.05850560590624809, + 0.0639900267124176, + 0.03468135744333267, + 0.05653219670057297, + 0.006625386420637369, + 0.007544166408479214, + 0.014059748500585556, + 0.005068514030426741, + 0.011042725294828415, + 0.009284522384405136, + 0.007003896404057741, + 0.004820835776627064, + 0.0033718033228069544, + 0.004509677179157734, + 0.008042200468480587, + 0.0031034843996167183, + 0.005098937079310417, + 0.0049779294058680534, + 0.0037606321275234222, + 0.0028534785378724337, + 0.006935012061148882, + 0.008554276078939438, + 0.014337267726659775, + 0.006787351798266172, + 0.011005543172359467, + 0.008981163613498211, + 0.006290217861533165, + 0.008932609111070633, + 0.0023353700526058674, + 0.0037231864407658577, + 0.004588474053889513, + 0.0023897048085927963, + 0.005722615402191877, + 0.0037962733767926693, + 0.002909652655944228, + 0.0025163216050714254, + 0.0029379031620919704, + 0.003689445788040757, + 0.0050023989751935005, + 0.0024989235680550337, + 0.005335482303053141, + 0.004086662549525499, + 0.0029049310833215714, + 0.0029574043583124876, + 0.00526046846061945, + 0.007212749682366848, + 0.008400911465287209, + 0.007002062629908323, + 0.006583993323147297, + 0.008937952108681202, + 0.006669140420854092, + 0.013724436052143574, + 0.0027791352476924658, + 0.0036410053726285696, + 0.004543646704405546, + 0.0038926051929593086, + 0.003227712819352746, + 0.0033481158316135406, + 0.0031626734416931868, + 0.004558356944471598, + 0.009870057925581932, + 0.006228094920516014, + 0.004261527210474014, + 0.002723779296502471, + 0.0031725503504276276, + 0.005268556997179985, + 0.002795893233269453, + 0.004932155832648277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.003198272082954645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890248492360115 + ], + "opposite_logprob_tokens": [ + -2.464024782180786 + ], + "logprob_reward": -0.0890248492360115, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890248492360115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21609282493591309, + 0.33803069591522217, + 0.05819167569279671, + 0.0317634679377079, + 0.13925360143184662, + 0.04269490763545036, + 0.047337114810943604, + 0.04479103535413742, + 0.04787337779998779, + 0.03397130221128464 + ], + "credit_map": [ + 0.03842714801430702, + 0.026580601930618286, + 0.043143924325704575, + 0.013672254048287868, + 0.02413860149681568, + 0.03551341965794563, + 0.011887991800904274, + 0.02272888831794262, + 0.04676811397075653, + 0.0799286961555481, + 0.04908332973718643, + 0.022071342915296555, + 0.037160757929086685, + 0.041668277233839035, + 0.02071000076830387, + 0.04064018651843071, + 0.005606016144156456, + 0.007494195364415646, + 0.012164759449660778, + 0.004055328667163849, + 0.010304523631930351, + 0.007798954378813505, + 0.0053696297109127045, + 0.005398266017436981, + 0.004103559534996748, + 0.0041336421854794025, + 0.006088891997933388, + 0.002781488234177232, + 0.003717781975865364, + 0.0045530772767961025, + 0.003741663880646229, + 0.002643364714458585, + 0.011344310827553272, + 0.01573162153363228, + 0.019000962376594543, + 0.012512361630797386, + 0.022746700793504715, + 0.017355453222990036, + 0.014405591413378716, + 0.026156602427363396, + 0.0037376300897449255, + 0.0059418813325464725, + 0.006585466675460339, + 0.002810818376019597, + 0.00742584653198719, + 0.006573278922587633, + 0.00298703252337873, + 0.0066329557448625565, + 0.004820335656404495, + 0.005330933723598719, + 0.006506112404167652, + 0.003493617055937648, + 0.008581440895795822, + 0.008581437170505524, + 0.0038227904587984085, + 0.006200446747243404, + 0.003389587625861168, + 0.007170841563493013, + 0.004786576144397259, + 0.0027633204590529203, + 0.00813281536102295, + 0.006729918997734785, + 0.004003831651061773, + 0.007814142853021622, + 0.005507160909473896, + 0.007190664764493704, + 0.00509144552052021, + 0.005767471622675657, + 0.006035149563103914, + 0.0063220662996172905, + 0.00599881075322628, + 0.005960609298199415, + 0.003992970567196608, + 0.005644186865538359, + 0.0031512377317994833, + 0.0044427234679460526, + 0.005386350676417351, + 0.004425820894539356, + 0.0035874973982572556, + 0.0033405134454369545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0890248492360115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38708165287971497 + ], + "opposite_logprob_tokens": [ + -1.1370816230773926 + ], + "logprob_reward": -0.38708165287971497, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38708165287971497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11977546662092209, + 0.5146321654319763, + 0.07975500822067261, + 0.027876755222678185, + 0.0634748786687851, + 0.03658391907811165, + 0.01975022815167904, + 0.07430445402860641, + 0.047273773699998856, + 0.0165733452886343 + ], + "credit_map": [ + 0.017718547955155373, + 0.019746948033571243, + 0.01960109733045101, + 0.009671899490058422, + 0.017513412982225418, + 0.01501764077693224, + 0.01087602972984314, + 0.00962989591062069, + 0.07394616305828094, + 0.12444023787975311, + 0.072907455265522, + 0.034244369715452194, + 0.05668533965945244, + 0.06341876834630966, + 0.032036684453487396, + 0.05695316940546036, + 0.00878187920898199, + 0.011430898681282997, + 0.010491698049008846, + 0.006496582645922899, + 0.01626393012702465, + 0.00925702415406704, + 0.009887411259114742, + 0.007145596668124199, + 0.002601227257400751, + 0.002904508961364627, + 0.004313708748668432, + 0.0025367436464875937, + 0.005108697339892387, + 0.0045064883306622505, + 0.003951536025851965, + 0.0019538486376404762, + 0.004973051603883505, + 0.007177043240517378, + 0.008199483156204224, + 0.006485719699412584, + 0.011791042983531952, + 0.007366695906966925, + 0.007828176952898502, + 0.009653672575950623, + 0.0032905542757362127, + 0.00520501472055912, + 0.005640560295432806, + 0.0027857727836817503, + 0.006925721187144518, + 0.005148000083863735, + 0.0026955674402415752, + 0.004892729222774506, + 0.002266291296109557, + 0.0025726528838276863, + 0.0032125222496688366, + 0.0015604806831106544, + 0.0026357239112257957, + 0.0030407309532165527, + 0.0016251110937446356, + 0.0028367170598357916, + 0.005045449361205101, + 0.007422635797411203, + 0.01109586376696825, + 0.005642660427838564, + 0.005544062703847885, + 0.012354901991784573, + 0.005705197807401419, + 0.021493686363101006, + 0.005397978704422712, + 0.006420683115720749, + 0.004569046199321747, + 0.006161897908896208, + 0.007159677799791098, + 0.00600948091596365, + 0.007631492801010609, + 0.003923520445823669, + 0.0025147367268800735, + 0.002286213915795088, + 0.001426219823770225, + 0.0022899506147950888, + 0.0024496123660355806, + 0.00201002461835742, + 0.001924355048686266, + 0.0016722324071452022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.38708165287971497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "graceful, slow, 360-degree medium shot rotates smoothly and poetically around an ancient poet. The poet stands alone on a vast black rock by the river. He wears a light, wide-sleeved white silk robe. The night breeze blows, causing the sleeves and hem to flutter violently, as if he is about to ride away. His long, jet-black hair is loosely tied back with a simple wooden hairpin, a few strands of hair blowing across his face. His sharply defined face appears exceptionally thin in the sidelight of the bright moon. His brows are slightly furrowed, his eyes fixed on the vast, shimmering river before him. His lips move in sync with the silent verses of his poem, his expression a fusion of deep sorrow and creative passion. The backdrop is a deep navy blue night sky, a full moon hanging high in the sky. Its cool, silvery glow provides the sole illumination of the scene, stained the poet's white robe with a radiant snow-white brilliance, casting soft shadows on the wet rocks, and spreading a fragmented, flowing silver light across the vast river. On the far bank, rolling hills appear in dark silhouette, like an ink painting. The steady rotation of the camera creates a tranquil atmosphere of grandeur, loneliness, yet a sense of romance and power.", + "seed": 20024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_9/index_2766_seed_20024_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10925556016154588, + "vlm_logprob": -0.10925556016154588, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.10925556016154588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02552008256316185 + ], + "opposite_logprob_tokens": [ + -3.775520086288452 + ], + "logprob_reward": -0.02552008256316185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02552008256316185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13478778302669525, + 0.10738512128591537, + 0.16447974741458893, + 0.28932520747184753, + 0.08051463961601257, + 0.0329192690551281, + 0.06413133442401886, + 0.032266225665807724, + 0.03461190313100815, + 0.059578701853752136 + ], + "credit_map": [ + 0.013374959118664265, + 0.013478005304932594, + 0.01032934058457613, + 0.012811985798180103, + 0.0300912968814373, + 0.02318589761853218, + 0.011416967026889324, + 0.020099326968193054, + 0.00967300496995449, + 0.012855852954089642, + 0.005876290611922741, + 0.015332967974245548, + 0.01455402560532093, + 0.02229771390557289, + 0.010483818128705025, + 0.01631144806742668, + 0.009975248947739601, + 0.008160538040101528, + 0.006218938156962395, + 0.004035855643451214, + 0.008241604082286358, + 0.06612102687358856, + 0.00546521320939064, + 0.0562613308429718, + 0.030240993946790695, + 0.04594060406088829, + 0.029145538806915283, + 0.03195818141102791, + 0.05960291996598244, + 0.034096311777830124, + 0.03284597769379616, + 0.025494664907455444, + 0.010136071592569351, + 0.010965205729007721, + 0.008977871388196945, + 0.005716368090361357, + 0.011196495033800602, + 0.01030108705163002, + 0.007120505906641483, + 0.016101041808724403, + 0.0071663144044578075, + 0.004218627233058214, + 0.004654633812606335, + 0.003167622024193406, + 0.003316988702863455, + 0.005648827645927668, + 0.0017452242318540812, + 0.003001031232997775, + 0.014078947715461254, + 0.007213147357106209, + 0.011914162896573544, + 0.00614892179146409, + 0.005168423987925053, + 0.010130073875188828, + 0.004745807033032179, + 0.004731850698590279, + 0.0052598146721720695, + 0.0068998136557638645, + 0.004857847932726145, + 0.0022703497670590878, + 0.004701399244368076, + 0.00421794131398201, + 0.0017143174773082137, + 0.002344743348658085, + 0.004860223270952702, + 0.004233208950608969, + 0.005839906632900238, + 0.0033612409606575966, + 0.0047072856687009335, + 0.004888234194368124, + 0.0026229016948491335, + 0.004098900128155947, + 0.012607552111148834, + 0.007185847498476505, + 0.018168261274695396, + 0.00472815427929163, + 0.0042030964978039265, + 0.006757676601409912, + 0.002885224297642708, + 0.0030428890604525805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.02552008256316185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199361860752106 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.018199361860752106, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199361860752106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2572753131389618, + 0.21133902668952942, + 0.13773222267627716, + 0.06613530218601227, + 0.05101766809821129, + 0.08142393827438354, + 0.0360051728785038, + 0.028585927560925484, + 0.035175736993551254, + 0.09530967473983765 + ], + "credit_map": [ + 0.02349400334060192, + 0.04062439873814583, + 0.016295377165079117, + 0.013530372641980648, + 0.027722209692001343, + 0.0703810602426529, + 0.011628631502389908, + 0.053599245846271515, + 0.028634661808609962, + 0.03830672800540924, + 0.01604369282722473, + 0.020541777834296227, + 0.02580926939845085, + 0.0343368761241436, + 0.015562405809760094, + 0.03210362046957016, + 0.01598004437983036, + 0.01740768551826477, + 0.006681039929389954, + 0.005594410002231598, + 0.010266591794788837, + 0.03581344336271286, + 0.0060544996522367, + 0.03993450477719307, + 0.010955339297652245, + 0.008092801086604595, + 0.0069450922310352325, + 0.0057433792389929295, + 0.009775867685675621, + 0.010802339762449265, + 0.0061906566843390465, + 0.007629822939634323, + 0.007952342741191387, + 0.005217867903411388, + 0.007707676850259304, + 0.0032818508334457874, + 0.003933596890419722, + 0.008971056900918484, + 0.003621803130954504, + 0.010331472381949425, + 0.020104924216866493, + 0.0074275461956858635, + 0.012140902690589428, + 0.006445267703384161, + 0.00575708644464612, + 0.018362483009696007, + 0.003840148216113448, + 0.00734558142721653, + 0.0038484849501401186, + 0.005365665070712566, + 0.0058074588887393475, + 0.004465480335056782, + 0.0037273848429322243, + 0.005164331756532192, + 0.003490147180855274, + 0.004136220086365938, + 0.004022619221359491, + 0.004364293534308672, + 0.006106924265623093, + 0.0023540814872831106, + 0.0034983118530362844, + 0.003497336059808731, + 0.001950518344528973, + 0.0027918426785618067, + 0.006070278584957123, + 0.00439911475405097, + 0.00880578625947237, + 0.0022417623549699783, + 0.003597051603719592, + 0.004930451512336731, + 0.0020241490565240383, + 0.003107143798843026, + 0.01631106436252594, + 0.008720730431377888, + 0.03295202553272247, + 0.0042811366729438305, + 0.00446604797616601, + 0.019873088225722313, + 0.0033731823787093163, + 0.005332410801202059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018199361860752106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036235882434993982 + ], + "opposite_logprob_tokens": [ + -5.6286234855651855 + ], + "logprob_reward": -0.0036235882434993982, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036235882434993982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20499806106090546, + 0.15780843794345856, + 0.34649765491485596, + 0.04177793115377426, + 0.038616422563791275, + 0.0334198996424675, + 0.027078205719590187, + 0.0238980483263731, + 0.02523374743759632, + 0.10067154467105865 + ], + "credit_map": [ + 0.024634376168251038, + 0.032672252506017685, + 0.023512685671448708, + 0.014220476150512695, + 0.029736144468188286, + 0.03437231853604317, + 0.009773611091077328, + 0.03607621416449547, + 0.014711955562233925, + 0.019169939681887627, + 0.007045740261673927, + 0.01927313208580017, + 0.03576735407114029, + 0.018450846895575523, + 0.02113560400903225, + 0.022253867238759995, + 0.01694922149181366, + 0.017312100157141685, + 0.0064033521339297295, + 0.00735859340056777, + 0.009717757813632488, + 0.1472441852092743, + 0.00811946764588356, + 0.13339298963546753, + 0.00461166724562645, + 0.003777846461161971, + 0.0053490507416427135, + 0.004012129735201597, + 0.0054158358834683895, + 0.006046141963452101, + 0.004128450993448496, + 0.008436815813183784, + 0.0054375184699893, + 0.004454141948372126, + 0.004240969195961952, + 0.003100678091868758, + 0.003364994190633297, + 0.00524235051125288, + 0.0029734959825873375, + 0.009802273474633694, + 0.006604742258787155, + 0.004368698690086603, + 0.003788466565310955, + 0.003339728806167841, + 0.004195856396108866, + 0.004313624929636717, + 0.0025131956208497286, + 0.004295585677027702, + 0.004257524851709604, + 0.0041060419753193855, + 0.004575903527438641, + 0.002935656113550067, + 0.0029458273202180862, + 0.0032078507356345654, + 0.0024587889201939106, + 0.0025906141381710768, + 0.0032217635307461023, + 0.0037935019936412573, + 0.003558224532753229, + 0.0023675274569541216, + 0.003143856069073081, + 0.004463877994567156, + 0.0018248207634314895, + 0.0015244776150211692, + 0.004529471043497324, + 0.0034907031804323196, + 0.004858674015849829, + 0.0021508345380425453, + 0.0033124240580946207, + 0.003396679414436221, + 0.0017804143717512488, + 0.0017145497258752584, + 0.01919439621269703, + 0.011598469689488411, + 0.02801898866891861, + 0.008289763703942299, + 0.009126095101237297, + 0.014118282124400139, + 0.005478952080011368, + 0.004846606869250536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0036235882434993982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001214474439621 + ], + "opposite_logprob_tokens": [ + -2.695012092590332 + ], + "logprob_reward": -0.07001214474439621, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001214474439621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16077609360218048, + 0.10325971245765686, + 0.37013354897499084, + 0.05448117479681969, + 0.07586085051298141, + 0.05086558684706688, + 0.03136022761464119, + 0.031246012076735497, + 0.035403113812208176, + 0.08661366254091263 + ], + "credit_map": [ + 0.025651857256889343, + 0.026683298870921135, + 0.012930132448673248, + 0.010667277500033379, + 0.02000281773507595, + 0.030746223405003548, + 0.009518814273178577, + 0.024575689807534218, + 0.009049844928085804, + 0.012979584746062756, + 0.004606546368449926, + 0.012593674473464489, + 0.02540285512804985, + 0.011722144670784473, + 0.014087253250181675, + 0.01281780656427145, + 0.01832498610019684, + 0.017013605684041977, + 0.0066748070530593395, + 0.0059318686835467815, + 0.00850333459675312, + 0.16508984565734863, + 0.007690146565437317, + 0.14090494811534882, + 0.0075959376990795135, + 0.003859075950458646, + 0.005522248800843954, + 0.0038278624415397644, + 0.0053605251014232635, + 0.014069399796426296, + 0.004753224551677704, + 0.009492903016507626, + 0.019062398001551628, + 0.007953353226184845, + 0.00867003109306097, + 0.004299151245504618, + 0.007824556902050972, + 0.013727699406445026, + 0.004475693684071302, + 0.009847967885434628, + 0.007878107018768787, + 0.006761714816093445, + 0.008184854872524738, + 0.004922314081341028, + 0.006720328703522682, + 0.006284083239734173, + 0.004637662786990404, + 0.005476519465446472, + 0.005819862708449364, + 0.003239039331674576, + 0.004750609863549471, + 0.003339387010782957, + 0.002847274299710989, + 0.005100542213767767, + 0.0023407444823533297, + 0.0039227670058608055, + 0.004599032457917929, + 0.004364857915788889, + 0.007208713795989752, + 0.002281227381899953, + 0.003337254747748375, + 0.004693648312240839, + 0.0017569741467013955, + 0.003004303202033043, + 0.0069245644845068455, + 0.0041586244478821754, + 0.006650942377746105, + 0.00350433262065053, + 0.0038954089395701885, + 0.005510025192052126, + 0.0024376679211854935, + 0.002321545034646988, + 0.014205040410161018, + 0.0069062188267707825, + 0.03053039126098156, + 0.004656990524381399, + 0.0035321053583174944, + 0.020345298573374748, + 0.0033419979736208916, + 0.0030956233385950327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07001214474439621 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289226233959198 + ], + "opposite_logprob_tokens": [ + -1.0539226531982422 + ], + "logprob_reward": -0.4289226233959198, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289226233959198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2488791048526764, + 0.24760331213474274, + 0.333507239818573, + 0.05228680744767189, + 0.015522989444434643, + 0.030808385461568832, + 0.018837563693523407, + 0.015946952626109123, + 0.01858767680823803, + 0.018019970506429672 + ], + "credit_map": [ + 0.027921529486775398, + 0.03701581060886383, + 0.01789724826812744, + 0.011953751556575298, + 0.0257762148976326, + 0.06474926322698593, + 0.01651894859969616, + 0.04704633727669716, + 0.022524308413267136, + 0.03097730502486229, + 0.0100177600979805, + 0.03352288529276848, + 0.05559723079204559, + 0.03031689114868641, + 0.03673525154590607, + 0.02791166491806507, + 0.01535753533244133, + 0.017701372504234314, + 0.006082907784730196, + 0.006004825700074434, + 0.011862524785101414, + 0.14015385508537292, + 0.007610605098307133, + 0.12873363494873047, + 0.007913368754088879, + 0.004358367994427681, + 0.005879047326743603, + 0.0033265198580920696, + 0.007233554497361183, + 0.010281091555953026, + 0.005139505490660667, + 0.008155349642038345, + 0.002367987995967269, + 0.0021898008417338133, + 0.0019847708754241467, + 0.0016016997396945953, + 0.0017862265231087804, + 0.002525560325011611, + 0.0013216618681326509, + 0.0017452816246077418, + 0.0075248549692332745, + 0.003250709269195795, + 0.003526990534737706, + 0.00214839493855834, + 0.0031326699536293745, + 0.0061138831079006195, + 0.0016665310831740499, + 0.0034443517215549946, + 0.0033857638482004404, + 0.0027939050924032927, + 0.0030897886026650667, + 0.0017610006034374237, + 0.001859224634245038, + 0.0029449150897562504, + 0.0015570507384836674, + 0.0014459153171628714, + 0.0022763083688914776, + 0.0015073383692651987, + 0.004118307493627071, + 0.0009678495698608458, + 0.0011413979809731245, + 0.0026306367944926023, + 0.0009013046510517597, + 0.0024038099218159914, + 0.0037014929112046957, + 0.002066567772999406, + 0.0047412351705133915, + 0.0014438400976359844, + 0.001390078803524375, + 0.0025628777220845222, + 0.0010461171623319387, + 0.0016354675171896815, + 0.00394404074177146, + 0.0023317676968872547, + 0.0035952189937233925, + 0.0018289998406544328, + 0.0018874808447435498, + 0.0020802034996449947, + 0.0011444658739492297, + 0.0012077928986400366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.4289226233959198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "graceful, slow, 360-degree medium shot rotates smoothly and poetically around an ancient poet. The poet stands alone on a vast black rock by the river. He wears a light, wide-sleeved white silk robe. The night breeze blows, causing the sleeves and hem to flutter violently, as if he is about to ride away. His long, jet-black hair is loosely tied back with a simple wooden hairpin, a few strands of hair blowing across his face. His sharply defined face appears exceptionally thin in the sidelight of the bright moon. His brows are slightly furrowed, his eyes fixed on the vast, shimmering river before him. His lips move in sync with the silent verses of his poem, his expression a fusion of deep sorrow and creative passion. The backdrop is a deep navy blue night sky, a full moon hanging high in the sky. Its cool, silvery glow provides the sole illumination of the scene, stained the poet's white robe with a radiant snow-white brilliance, casting soft shadows on the wet rocks, and spreading a fragmented, flowing silver light across the vast river. On the far bank, rolling hills appear in dark silhouette, like an ink painting. The steady rotation of the camera creates a tranquil atmosphere of grandeur, loneliness, yet a sense of romance and power.", + "seed": 20025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000127/group_9/index_2766_seed_20025_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04900023431982845, + "vlm_logprob": -0.04900023431982845, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.04900023431982845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009154489263892174 + ], + "opposite_logprob_tokens": [ + -4.884154319763184 + ], + "logprob_reward": -0.009154489263892174, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009154489263892174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20689500868320465, + 0.05946454405784607, + 0.036070045083761215, + 0.0724269300699234, + 0.3834359645843506, + 0.03993169963359833, + 0.10218623280525208, + 0.028728188946843147, + 0.04031403362751007, + 0.03054741956293583 + ], + "credit_map": [ + 0.038173820823431015, + 0.036440201103687286, + 0.011611598543822765, + 0.014558031223714352, + 0.027759412303566933, + 0.029371846467256546, + 0.018518242985010147, + 0.030461855232715607, + 0.00832868367433548, + 0.010882746428251266, + 0.0050233108922839165, + 0.006624546833336353, + 0.008827353827655315, + 0.006953422445803881, + 0.006281018722802401, + 0.006543457508087158, + 0.003758680773898959, + 0.005082548130303621, + 0.002475712215527892, + 0.004319674335420132, + 0.006682243663817644, + 0.004541488830000162, + 0.005879167467355728, + 0.003330531530082226, + 0.00878257304430008, + 0.009099915623664856, + 0.0047013601288199425, + 0.009079868905246258, + 0.011095396243035793, + 0.00870558712631464, + 0.012803159654140472, + 0.008159065619111061, + 0.039094697684049606, + 0.042607780545949936, + 0.03314068168401718, + 0.059854865074157715, + 0.054607950150966644, + 0.04154055565595627, + 0.07758302986621857, + 0.035006411373615265, + 0.005434003192931414, + 0.00416647968813777, + 0.005410448648035526, + 0.004088804591447115, + 0.005464782472699881, + 0.0048421830870211124, + 0.006238732486963272, + 0.0042862617410719395, + 0.014138060621917248, + 0.010643525049090385, + 0.016377786174416542, + 0.01025120448321104, + 0.013548009097576141, + 0.013784816488623619, + 0.014206385239958763, + 0.009236443787813187, + 0.004803407471626997, + 0.0032659966964274645, + 0.0026228781789541245, + 0.002653994597494602, + 0.00468993978574872, + 0.004677696153521538, + 0.0032537588849663734, + 0.002760516479611397, + 0.0049218470230698586, + 0.0038985719438642263, + 0.00899418629705906, + 0.0037564660888165236, + 0.0039884853176772594, + 0.006863907445222139, + 0.0031318641267716885, + 0.004758705850690603, + 0.00425694091245532, + 0.003855440765619278, + 0.006055676843971014, + 0.0028427434153854847, + 0.0037316521629691124, + 0.0037483801133930683, + 0.0031708923634141684, + 0.002885691821575165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.009154489263892174 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545713223516941 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545713223516941, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545713223516941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24566185474395752, + 0.07519680261611938, + 0.04323225095868111, + 0.0414026714861393, + 0.27044808864593506, + 0.05861883610486984, + 0.07911323010921478, + 0.06488659977912903, + 0.0633678063750267, + 0.05807189270853996 + ], + "credit_map": [ + 0.04274135082960129, + 0.045318812131881714, + 0.016418125480413437, + 0.025208953768014908, + 0.0322391614317894, + 0.032092783600091934, + 0.025266407057642937, + 0.026376252993941307, + 0.008124560117721558, + 0.013427789323031902, + 0.004584409296512604, + 0.010617531836032867, + 0.0108901746571064, + 0.008513074368238449, + 0.008963007479906082, + 0.010076255537569523, + 0.006463693920522928, + 0.009995832107961178, + 0.0034129635896533728, + 0.0031111370772123337, + 0.005762419197708368, + 0.006491528358310461, + 0.0036431329790502787, + 0.004351546987891197, + 0.00469503179192543, + 0.004829863552004099, + 0.003660470712929964, + 0.003316594520583749, + 0.006029666867107153, + 0.005678584799170494, + 0.004609075374901295, + 0.008583382703363895, + 0.02495400235056877, + 0.03167460858821869, + 0.02882186323404312, + 0.0385855995118618, + 0.037780191749334335, + 0.03028222545981407, + 0.04987039417028427, + 0.028479216620326042, + 0.007318413816392422, + 0.0050737266428768635, + 0.010136980563402176, + 0.0054678041487932205, + 0.007930454798042774, + 0.006676421035081148, + 0.006838874891400337, + 0.009176157414913177, + 0.006839786656200886, + 0.009678126312792301, + 0.015080277808010578, + 0.007973956875503063, + 0.011511438526213169, + 0.009618785232305527, + 0.008678867481648922, + 0.009731990285217762, + 0.009459437802433968, + 0.009036152623593807, + 0.006909005343914032, + 0.005509538576006889, + 0.009983371011912823, + 0.010480199940502644, + 0.0060349018312990665, + 0.007473994977772236, + 0.008416621945798397, + 0.006045487243682146, + 0.012433135882019997, + 0.006750814151018858, + 0.007216658443212509, + 0.010134807787835598, + 0.0074376980774104595, + 0.004932584706693888, + 0.009241095744073391, + 0.004998169373720884, + 0.0104752816259861, + 0.0035990001633763313, + 0.0056060319766402245, + 0.015576609410345554, + 0.004739850759506226, + 0.0038358578458428383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.012545713223516941 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011834054021164775 + ], + "opposite_logprob_tokens": [ + -6.75118350982666 + ], + "logprob_reward": -0.0011834054021164775, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011834054021164775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3210313022136688, + 0.10750740021467209, + 0.05166716128587723, + 0.041712354868650436, + 0.11923936009407043, + 0.08422166854143143, + 0.08631659299135208, + 0.0791708305478096, + 0.05541316792368889, + 0.0537201389670372 + ], + "credit_map": [ + 0.07071562111377716, + 0.06519752740859985, + 0.01639077253639698, + 0.026086872443556786, + 0.03744562342762947, + 0.04606407880783081, + 0.02359524928033352, + 0.035535577684640884, + 0.014601167291402817, + 0.023849518969655037, + 0.008452720940113068, + 0.012364673428237438, + 0.014238957315683365, + 0.013176572509109974, + 0.009254617616534233, + 0.011569169349968433, + 0.008186384104192257, + 0.010151675902307034, + 0.00614621676504612, + 0.003788605099543929, + 0.00588446669280529, + 0.006969190202653408, + 0.0046653444878757, + 0.005875276867300272, + 0.006293944548815489, + 0.007439248729497194, + 0.003570508910343051, + 0.0034677237272262573, + 0.005664058495312929, + 0.005606476217508316, + 0.004533869680017233, + 0.005136523395776749, + 0.020743632689118385, + 0.02263863943517208, + 0.009010863490402699, + 0.011511104181408882, + 0.015445366501808167, + 0.013908055610954762, + 0.016091354191303253, + 0.009890350513160229, + 0.009867879562079906, + 0.01288681011646986, + 0.008939402177929878, + 0.008843517862260342, + 0.013057093136012554, + 0.00920527707785368, + 0.011416560038924217, + 0.01000512856990099, + 0.01411377266049385, + 0.017995204776525497, + 0.012303760275244713, + 0.006724374834448099, + 0.008536283858120441, + 0.01190891582518816, + 0.005169229581952095, + 0.009565046988427639, + 0.016372298821806908, + 0.00733990641310811, + 0.008231216110289097, + 0.004674812778830528, + 0.008363587781786919, + 0.019836904481053352, + 0.005542530212551355, + 0.008809566497802734, + 0.008268304169178009, + 0.006050418131053448, + 0.010143714025616646, + 0.004778671078383923, + 0.0058665224350988865, + 0.008897785097360611, + 0.004735192283987999, + 0.006672559306025505, + 0.008118893951177597, + 0.00533066364005208, + 0.007810654584318399, + 0.0035757792647928, + 0.0068302941508591175, + 0.013624702580273151, + 0.005710891913622618, + 0.002718262141570449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0011834054021164775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576467737555504 + ], + "opposite_logprob_tokens": [ + -3.8955764770507812 + ], + "logprob_reward": -0.020576467737555504, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576467737555504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17081473767757416, + 0.06320830434560776, + 0.031222863122820854, + 0.025306060910224915, + 0.3984275460243225, + 0.10044959932565689, + 0.0521986298263073, + 0.06995116174221039, + 0.07011087238788605, + 0.018310192972421646 + ], + "credit_map": [ + 0.038582943379879, + 0.03199833631515503, + 0.0076896874234080315, + 0.014812914654612541, + 0.019584426656365395, + 0.02386924810707569, + 0.013718031346797943, + 0.0205591581761837, + 0.009043714962899685, + 0.011688876897096634, + 0.004895900841802359, + 0.00802325177937746, + 0.008014426566660404, + 0.0064673093147575855, + 0.005932082422077656, + 0.009142743423581123, + 0.0038906855043023825, + 0.004894507583230734, + 0.0038279113359749317, + 0.0026222451124340296, + 0.004379871767014265, + 0.0033652586862444878, + 0.003689051140099764, + 0.004553330596536398, + 0.0035156887024641037, + 0.004537110682576895, + 0.00237202737480402, + 0.0023588603362441063, + 0.0036824808921664953, + 0.0031125701498240232, + 0.0030582903418689966, + 0.0026690305676311255, + 0.03915708139538765, + 0.042101211845874786, + 0.03323778510093689, + 0.06446381658315659, + 0.05529570206999779, + 0.04380558058619499, + 0.08336130529642105, + 0.037005092948675156, + 0.016631076112389565, + 0.010138008743524551, + 0.015102854929864407, + 0.012257590889930725, + 0.013943791389465332, + 0.011066954582929611, + 0.01083160936832428, + 0.010477714240550995, + 0.0073735942132771015, + 0.008105427026748657, + 0.0058190226554870605, + 0.004695896524935961, + 0.00763697549700737, + 0.006736184004694223, + 0.007656585890799761, + 0.004174943082034588, + 0.007560511585325003, + 0.005088580306619406, + 0.011469614692032337, + 0.007977939210832119, + 0.012756885029375553, + 0.008845334872603416, + 0.008883838541805744, + 0.007368454709649086, + 0.007443166803568602, + 0.006597431376576424, + 0.009483215399086475, + 0.007879338227212429, + 0.010230318643152714, + 0.010459992103278637, + 0.00924849882721901, + 0.008768907748162746, + 0.002330105984583497, + 0.002569526433944702, + 0.0035114299971610308, + 0.0011406706180423498, + 0.0023705237545073032, + 0.0024922501761466265, + 0.0017732864944264293, + 0.002122400561347604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.020576467737555504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient poet wearing a light, wide-sleeved white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the poet standing alone on a vast black rock by a river with rolling hills in dark silhouette on the far bank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a bright full moon in a deep navy blue night sky, casting a silvery glow on the white robe and wet rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the poet's sleeves and hem flutter violently in the night breeze, and is his long jet-black hair loosely tied back with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth 360-degree medium shot rotation around the poet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154109597206116 + ], + "opposite_logprob_tokens": [ + -1.7015410661697388 + ], + "logprob_reward": -0.20154109597206116, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154109597206116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25854626297950745, + 0.0818629264831543, + 0.04419713839888573, + 0.03985385596752167, + 0.19897383451461792, + 0.11237745732069016, + 0.04757145047187805, + 0.040703464299440384, + 0.13253816962242126, + 0.04337539151310921 + ], + "credit_map": [ + 0.055984485894441605, + 0.046611856669187546, + 0.013677352108061314, + 0.02061806619167328, + 0.028912493959069252, + 0.034357618540525436, + 0.024840189144015312, + 0.03354421630501747, + 0.010083625093102455, + 0.015930550172924995, + 0.005304188001900911, + 0.01027948409318924, + 0.010803600773215294, + 0.008538220077753067, + 0.00796155259013176, + 0.012961710803210735, + 0.005575280636548996, + 0.007140726316720247, + 0.00420343317091465, + 0.0038885953836143017, + 0.006075880955904722, + 0.004491192288696766, + 0.00586027093231678, + 0.006961756385862827, + 0.005311205983161926, + 0.00489322142675519, + 0.004159664269536734, + 0.004126370884478092, + 0.005747126415371895, + 0.004957288503646851, + 0.006494802888482809, + 0.004164176993072033, + 0.018849074840545654, + 0.02252747304737568, + 0.017649367451667786, + 0.031061554327607155, + 0.028957568109035492, + 0.02197657898068428, + 0.03988656401634216, + 0.018065672367811203, + 0.017718791961669922, + 0.009430625475943089, + 0.012262149713933468, + 0.015002640895545483, + 0.017593728378415108, + 0.011478746309876442, + 0.016292527318000793, + 0.01259825099259615, + 0.006028288975358009, + 0.0062248604372143745, + 0.004036158788949251, + 0.005767130292952061, + 0.007826636545360088, + 0.007943760603666306, + 0.005902459844946861, + 0.0038421533536165953, + 0.004988839849829674, + 0.004072254057973623, + 0.0043335650116205215, + 0.0036744659300893545, + 0.007222169078886509, + 0.005815968848764896, + 0.005628371145576239, + 0.004967830143868923, + 0.016478093340992928, + 0.01326152216643095, + 0.019088465720415115, + 0.012957330793142319, + 0.016188597306609154, + 0.02825711853802204, + 0.01538086961954832, + 0.01092616468667984, + 0.006225123070180416, + 0.004411234986037016, + 0.005729137919843197, + 0.0033054049126803875, + 0.005089054815471172, + 0.011267609894275665, + 0.0051960120908916, + 0.002151813358068466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.20154109597206116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_0/index_tensor([4044, 4044])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_0/index_tensor([4044, 4044])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3162a17c043d5028280b09dd3e4666d345dddd6a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_0/index_tensor([4044, 4044])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking, static, low-angle shot, shot from above, lingers on the wet asphalt, clearly reflecting the shimmering cyan, magenta, and electric blue glow of a giant holographic billboard atop a futuristic skyscraper. On this rainy night steeped in cyberpunk, fine raindrops trace brilliant trails against the vibrant neon lights. At the center of the frame, a suspended mass of dark brown liquid and finely pulsated herbal granules—the visual embodiment of Huoxiang Zhengqi Powder—swirls slowly. Suddenly, the flowing mixture abruptly stops in mid-air. A jet of frost, as black as obsidian, spreads outward at a visible speed, instantly solidifying into a sharp, multifaceted black ice crystal. Simultaneously, countless delicate circuit patterns, emitting a faint blue glow, are instantly etched into the crystal's surface, interweaving and pulsating like a complex neural network. This interplay with the cold neon lights of the surrounding environment creates a mysterious, high-tech, and tense atmosphere.", + "seed": 18127, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_0/index_4044_seed_18127_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3531876377761364, + "vlm_logprob": -0.3531876377761364, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.3531876377761364, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.70572829246521 + ], + "opposite_logprob_tokens": [ + -0.20572823286056519 + ], + "logprob_reward": -1.70572829246521, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.70572829246521, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1013740673661232, + 0.03314236178994179, + 0.0362367182970047, + 0.3886527121067047, + 0.08822977542877197, + 0.10425832122564316, + 0.04989129677414894, + 0.05928428843617439, + 0.0864141434431076, + 0.05251626670360565 + ], + "credit_map": [ + 0.020956821739673615, + 0.008855337277054787, + 0.010336196050047874, + 0.013865395449101925, + 0.010024058632552624, + 0.0069748410023748875, + 0.011348065920174122, + 0.019013356417417526, + 0.0025946935638785362, + 0.001522766542620957, + 0.003929721657186747, + 0.00808414164930582, + 0.002433235989883542, + 0.0025372968520969152, + 0.0063115982338786125, + 0.005728909280151129, + 0.003595806658267975, + 0.0020694707054644823, + 0.0046031419187784195, + 0.008863089606165886, + 0.002540551358833909, + 0.00200709723867476, + 0.007322365418076515, + 0.005235196091234684, + 0.025991035625338554, + 0.027530252933502197, + 0.038893479853868484, + 0.10292503237724304, + 0.027492642402648926, + 0.031083544716238976, + 0.07645434886217117, + 0.058282408863306046, + 0.003996925428509712, + 0.004329645540565252, + 0.009982702322304249, + 0.027279632166028023, + 0.006769952364265919, + 0.0050015817396342754, + 0.015783466398715973, + 0.015085877850651741, + 0.0073318504728376865, + 0.00684932479634881, + 0.011603831313550472, + 0.020201530307531357, + 0.014843931421637535, + 0.006857326254248619, + 0.021420812234282494, + 0.015149718150496483, + 0.0037578896153718233, + 0.00303007778711617, + 0.006326773669570684, + 0.007829153910279274, + 0.007172431796789169, + 0.004063823726028204, + 0.01294194906949997, + 0.00476920185610652, + 0.0040126158855855465, + 0.004014767240732908, + 0.004919605329632759, + 0.005442660767585039, + 0.004557092674076557, + 0.010412203148007393, + 0.005352427251636982, + 0.02057291753590107, + 0.004616698250174522, + 0.004826921503990889, + 0.005543158855289221, + 0.004374748095870018, + 0.0069129797630012035, + 0.02708864025771618, + 0.005088111385703087, + 0.027962882071733475, + 0.006148516200482845, + 0.005605040583759546, + 0.0065274047665297985, + 0.006870717275887728, + 0.006885523442178965, + 0.006828043609857559, + 0.007274122908711433, + 0.0063769021071493626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -1.70572829246521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233156755566597 + ], + "opposite_logprob_tokens": [ + -3.773315668106079 + ], + "logprob_reward": -0.0233156755566597, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233156755566597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14129407703876495, + 0.12250152975320816, + 0.0659933090209961, + 0.22122979164123535, + 0.11002405732870102, + 0.10278427600860596, + 0.067320317029953, + 0.03820764273405075, + 0.042097121477127075, + 0.08854788541793823 + ], + "credit_map": [ + 0.026882410049438477, + 0.011497638188302517, + 0.018775539472699165, + 0.02266170084476471, + 0.00717385858297348, + 0.007646283134818077, + 0.01662716269493103, + 0.03002949059009552, + 0.00958489440381527, + 0.006390037015080452, + 0.011644351296126842, + 0.0315173976123333, + 0.008660634979605675, + 0.005919700022786856, + 0.028719140216708183, + 0.02006538398563862, + 0.0031231469474732876, + 0.0025108461268246174, + 0.0056180511601269245, + 0.0194252822548151, + 0.0029214194510132074, + 0.002270575612783432, + 0.017862636595964432, + 0.01226135529577732, + 0.014852136373519897, + 0.018514322116971016, + 0.023134589195251465, + 0.05146804824471474, + 0.020015466958284378, + 0.01845535822212696, + 0.041944172233343124, + 0.03284570202231407, + 0.005166187882423401, + 0.005337993148714304, + 0.013559220358729362, + 0.032950885593891144, + 0.009015779010951519, + 0.005941847339272499, + 0.01930801011621952, + 0.01874414086341858, + 0.010174225084483624, + 0.007404108997434378, + 0.012001306749880314, + 0.017739390954375267, + 0.014711794443428516, + 0.007731924299150705, + 0.017798785120248795, + 0.015222746878862381, + 0.004241776652634144, + 0.002166078193113208, + 0.02226240560412407, + 0.0054214950650930405, + 0.004193709697574377, + 0.0046426719054579735, + 0.007588012609630823, + 0.016804171726107597, + 0.004735134541988373, + 0.0028252210468053818, + 0.003730905707925558, + 0.005533300805836916, + 0.004142311401665211, + 0.003292844397947192, + 0.004537221044301987, + 0.009410704486072063, + 0.003469897899776697, + 0.003427612828090787, + 0.003382748458534479, + 0.003379011293873191, + 0.005553260445594788, + 0.010660906322300434, + 0.003083773422986269, + 0.009139916859567165, + 0.009766954928636551, + 0.009603545069694519, + 0.010691204108297825, + 0.01036185771226883, + 0.008815901353955269, + 0.012049318291246891, + 0.010058216750621796, + 0.01720089092850685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0233156755566597 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634272798895836 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.007634272798895836, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634272798895836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10279130190610886, + 0.09236044436693192, + 0.197538360953331, + 0.1561613827943802, + 0.06497698277235031, + 0.09456945955753326, + 0.08734609186649323, + 0.058874357491731644, + 0.0695326179265976, + 0.0758490040898323 + ], + "credit_map": [ + 0.015256058424711227, + 0.010889659635722637, + 0.009411072358489037, + 0.01872885413467884, + 0.010654406622052193, + 0.008860927075147629, + 0.015022809617221355, + 0.013967500068247318, + 0.007766857277601957, + 0.004475677851587534, + 0.011780442669987679, + 0.020755089819431305, + 0.006644637323915958, + 0.009021736681461334, + 0.014975498430430889, + 0.016940496861934662, + 0.008376019075512886, + 0.005927664693444967, + 0.013509953394532204, + 0.05660279467701912, + 0.008123350329697132, + 0.006893723737448454, + 0.05996204540133476, + 0.038142792880535126, + 0.01236727461218834, + 0.010101163759827614, + 0.01687881350517273, + 0.040038466453552246, + 0.011161650530993938, + 0.01364189013838768, + 0.029966870322823524, + 0.022005237638950348, + 0.003092910395935178, + 0.003200193867087364, + 0.008800373412668705, + 0.021400371566414833, + 0.0038091186434030533, + 0.0038246733602136374, + 0.010877929627895355, + 0.009971404448151588, + 0.009729703888297081, + 0.006572216749191284, + 0.011040519922971725, + 0.016997454687952995, + 0.014228484593331814, + 0.007497308310121298, + 0.015135575085878372, + 0.013368187472224236, + 0.004341569263488054, + 0.0029911615420132875, + 0.029440751299262047, + 0.007041826378554106, + 0.005701760295778513, + 0.004844968672841787, + 0.010777701623737812, + 0.022206343710422516, + 0.0078972689807415, + 0.0032934967894107103, + 0.004296237602829933, + 0.004081251565366983, + 0.005729323718696833, + 0.01754208281636238, + 0.00503804674372077, + 0.010996646247804165, + 0.003880375763401389, + 0.004118487238883972, + 0.005192893091589212, + 0.006064365617930889, + 0.009830656461417675, + 0.01264763344079256, + 0.00866097304970026, + 0.019137222319841385, + 0.008903398178517818, + 0.007785622961819172, + 0.008276009932160378, + 0.0073691001161932945, + 0.007521319668740034, + 0.01171761006116867, + 0.008524368517100811, + 0.015751570463180542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.007634272798895836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076156049966812 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011076156049966812, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076156049966812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1907038539648056, + 0.020566482096910477, + 0.04226064309477806, + 0.17239123582839966, + 0.05857885628938675, + 0.18521732091903687, + 0.06400729715824127, + 0.13935044407844543, + 0.04376252740621567, + 0.08316132426261902 + ], + "credit_map": [ + 0.026975762099027634, + 0.0144912488758564, + 0.018795065581798553, + 0.043545108288526535, + 0.016169240698218346, + 0.008881527930498123, + 0.028828918933868408, + 0.033016983419656754, + 0.002780120586976409, + 0.0017062487313523889, + 0.0024574080016463995, + 0.003621378680691123, + 0.002562119159847498, + 0.001458986778743565, + 0.0027906165923923254, + 0.00318960496224463, + 0.002189562190324068, + 0.0015765109565109015, + 0.0031395789701491594, + 0.011516700498759747, + 0.0017966630402952433, + 0.0015547540970146656, + 0.012042820453643799, + 0.00844405684620142, + 0.015289178118109703, + 0.00875160563737154, + 0.01533362828195095, + 0.04604560136795044, + 0.014423795975744724, + 0.01618339493870735, + 0.031863827258348465, + 0.024500209838151932, + 0.0028225521091371775, + 0.0025379431899636984, + 0.0069220843724906445, + 0.02033696509897709, + 0.003242063568904996, + 0.002964052138850093, + 0.010171836242079735, + 0.009581362828612328, + 0.01475605834275484, + 0.014250717125833035, + 0.02242734096944332, + 0.03188450261950493, + 0.028803955763578415, + 0.01336737908422947, + 0.035500623285770416, + 0.02422674372792244, + 0.0038885721005499363, + 0.003981375601142645, + 0.014118485152721405, + 0.008986670523881912, + 0.006424883380532265, + 0.00524058286100626, + 0.009660584852099419, + 0.011706145480275154, + 0.026729797944426537, + 0.01064441166818142, + 0.013396024703979492, + 0.011020560748875141, + 0.013440532609820366, + 0.039610959589481354, + 0.010929899290204048, + 0.013578270561993122, + 0.0039037892129272223, + 0.004613013006746769, + 0.004248255863785744, + 0.006082423031330109, + 0.00952192209661007, + 0.00447117630392313, + 0.006702114827930927, + 0.004219833295792341, + 0.007706653792411089, + 0.008588720113039017, + 0.009184405207633972, + 0.00919704232364893, + 0.011296072974801064, + 0.010207983665168285, + 0.011571492068469524, + 0.015408964827656746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.011076156049966812 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183792009949684 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183792009949684, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183792009949684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07230635732412338, + 0.04427598789334297, + 0.04453067481517792, + 0.4239938259124756, + 0.05813787877559662, + 0.12217321246862411, + 0.0545056089758873, + 0.04288141801953316, + 0.09208743274211884, + 0.04510751739144325 + ], + "credit_map": [ + 0.012638849206268787, + 0.006615934427827597, + 0.007338236086070538, + 0.014806156046688557, + 0.0039346362464129925, + 0.003679658053442836, + 0.009711520746350288, + 0.013581370003521442, + 0.003974881488829851, + 0.002538064494729042, + 0.0051828427240252495, + 0.010058117099106312, + 0.003090320387855172, + 0.0025558071210980415, + 0.008714176714420319, + 0.008161776699125767, + 0.003959834575653076, + 0.0019612398464232683, + 0.0045040566474199295, + 0.011244806461036205, + 0.0033250381238758564, + 0.0028018152806907892, + 0.010345406830310822, + 0.006388479843735695, + 0.031377773731946945, + 0.023662401363253593, + 0.03843769058585167, + 0.11562830954790115, + 0.030989164486527443, + 0.034498587250709534, + 0.08427701145410538, + 0.06512288749217987, + 0.0037301534321159124, + 0.0026928044389933348, + 0.007082899566739798, + 0.01752503216266632, + 0.004426183179020882, + 0.0037729397881776094, + 0.008951001800596714, + 0.009956860914826393, + 0.016104185953736305, + 0.007348814979195595, + 0.018257122486829758, + 0.018931357190012932, + 0.01112664956599474, + 0.011363008059561253, + 0.020847303792834282, + 0.01819477789103985, + 0.0036626383662223816, + 0.0033734641037881374, + 0.011815798468887806, + 0.009182967245578766, + 0.0054177045822143555, + 0.004881995730102062, + 0.008823796175420284, + 0.0073472377844154835, + 0.004776203539222479, + 0.004241210408508778, + 0.0046191103756427765, + 0.005516203586012125, + 0.006556316744536161, + 0.00757699366658926, + 0.005553616210818291, + 0.004041764885187149, + 0.0042491150088608265, + 0.0038320806343108416, + 0.00598726375028491, + 0.004160483833402395, + 0.006630962248891592, + 0.028838718309998512, + 0.005958991125226021, + 0.03242982178926468, + 0.006225030403584242, + 0.004859876818954945, + 0.005007066298276186, + 0.005247976165264845, + 0.005127732641994953, + 0.006379200145602226, + 0.00656574871391058, + 0.005694884806871414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.018183792009949684 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, static, low-angle shot, shot from above, lingers on the wet asphalt, clearly reflecting the shimmering cyan, magenta, and electric blue glow of a giant holographic billboard atop a futuristic skyscraper. On this rainy night steeped in cyberpunk, fine raindrops trace brilliant trails against the vibrant neon lights. At the center of the frame, a suspended mass of dark brown liquid and finely pulsated herbal granules—the visual embodiment of Huoxiang Zhengqi Powder—swirls slowly. Suddenly, the flowing mixture abruptly stops in mid-air. A jet of frost, as black as obsidian, spreads outward at a visible speed, instantly solidifying into a sharp, multifaceted black ice crystal. Simultaneously, countless delicate circuit patterns, emitting a faint blue glow, are instantly etched into the crystal's surface, interweaving and pulsating like a complex neural network. This interplay with the cold neon lights of the surrounding environment creates a mysterious, high-tech, and tense atmosphere.", + "seed": 18128, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_0/index_4044_seed_18128_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.8351955119520426, + "vlm_logprob": -0.8351955119520426, + "vlm_yes_no_margin": 1.8500000476837157, + "vlm_token_credit": -0.8351955119520426, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -4.019805431365967 + ], + "opposite_logprob_tokens": [ + -0.019805647432804108 + ], + "logprob_reward": -4.019805431365967, + "yes_no_margin_reward": -3.999999761581421, + "token_credit_reward": -4.019805431365967, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19507886469364166, + 0.07635416090488434, + 0.05047091841697693, + 0.10923248529434204, + 0.05855795368552208, + 0.13217411935329437, + 0.15921562910079956, + 0.11331851780414581, + 0.05745363608002663, + 0.048143740743398666 + ], + "credit_map": [ + 0.011352616362273693, + 0.021386802196502686, + 0.027929482981562614, + 0.023309320211410522, + 0.016596946865320206, + 0.013763483613729477, + 0.056374795734882355, + 0.024365436285734177, + 0.007580322213470936, + 0.008033808320760727, + 0.00946662575006485, + 0.013795270584523678, + 0.010579786263406277, + 0.006600148975849152, + 0.011020400561392307, + 0.009277796372771263, + 0.003807701403275132, + 0.007403816562145948, + 0.007704763673245907, + 0.006222176365554333, + 0.0077331336215138435, + 0.0034586868714541197, + 0.006841536611318588, + 0.007299100514501333, + 0.008289240300655365, + 0.013862492516636848, + 0.01329924538731575, + 0.02073783054947853, + 0.012599365785717964, + 0.011365355923771858, + 0.017248166725039482, + 0.011830790899693966, + 0.004333644639700651, + 0.0057677654549479485, + 0.007415318861603737, + 0.010670569725334644, + 0.007040683180093765, + 0.0037867166101932526, + 0.011110454797744751, + 0.00843279529362917, + 0.015062084421515465, + 0.010763821192085743, + 0.01889270357787609, + 0.025768738240003586, + 0.017114022746682167, + 0.00913314986974001, + 0.023488260805606842, + 0.011951340362429619, + 0.022250544279813766, + 0.018191277980804443, + 0.02632180042564869, + 0.021207846701145172, + 0.014067955315113068, + 0.01695278100669384, + 0.021683918312191963, + 0.018539506942033768, + 0.012074030935764313, + 0.013404601253569126, + 0.014467449858784676, + 0.01636139489710331, + 0.016688134521245956, + 0.014849593862891197, + 0.01316697895526886, + 0.012306330725550652, + 0.004350544419139624, + 0.007978077046573162, + 0.006161017809063196, + 0.00862631481140852, + 0.00634982343763113, + 0.010708153247833252, + 0.008491216227412224, + 0.004788488149642944, + 0.004671864211559296, + 0.006628991570323706, + 0.005349877290427685, + 0.003855522722005844, + 0.0059968275018036366, + 0.012044276110827923, + 0.006176596041768789, + 0.0034197834320366383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -4.019805431365967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07015185058116913 + ], + "opposite_logprob_tokens": [ + -2.6951518058776855 + ], + "logprob_reward": -0.07015185058116913, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07015185058116913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36137625575065613, + 0.07571917772293091, + 0.10551367700099945, + 0.0668097510933876, + 0.059511713683605194, + 0.08145260810852051, + 0.08717850595712662, + 0.09751817584037781, + 0.04556351155042648, + 0.01935659535229206 + ], + "credit_map": [ + 0.018865916877985, + 0.039594922214746475, + 0.05175304412841797, + 0.03814519941806793, + 0.03448856621980667, + 0.0223504900932312, + 0.11010432243347168, + 0.046073783189058304, + 0.007409426383674145, + 0.01795572228729725, + 0.008141107857227325, + 0.009214132092893124, + 0.009078803472220898, + 0.0080244280397892, + 0.007810349576175213, + 0.00808520708233118, + 0.010561308823525906, + 0.009742953814566135, + 0.019725942984223366, + 0.01164861861616373, + 0.011331800371408463, + 0.011552379466593266, + 0.011819524690508842, + 0.019131140783429146, + 0.0045787254348397255, + 0.009798507206141949, + 0.010606791824102402, + 0.009790110401809216, + 0.006569567136466503, + 0.007595587521791458, + 0.009606986306607723, + 0.008263478055596352, + 0.0036624143831431866, + 0.008583178743720055, + 0.00929258018732071, + 0.01109323464334011, + 0.007060582283884287, + 0.00404007313773036, + 0.010164774022996426, + 0.005614875815808773, + 0.009433610364794731, + 0.007490002084523439, + 0.011303571984171867, + 0.01440759003162384, + 0.01245184801518917, + 0.005485380534082651, + 0.013946043327450752, + 0.006934567354619503, + 0.008739055134356022, + 0.011087240651249886, + 0.012158459983766079, + 0.017441559582948685, + 0.007721628062427044, + 0.007419515401124954, + 0.01326744630932808, + 0.009343600831925869, + 0.011858879588544369, + 0.014217937365174294, + 0.012445117346942425, + 0.012155349366366863, + 0.011639712378382683, + 0.014495030976831913, + 0.01061243750154972, + 0.010093710385262966, + 0.0050494251772761345, + 0.006181823555380106, + 0.004826594144105911, + 0.00649142125621438, + 0.004482580814510584, + 0.008920464664697647, + 0.004444187507033348, + 0.005167014896869659, + 0.0019260505214333534, + 0.0031355558894574642, + 0.0025529528502374887, + 0.001985491719096899, + 0.0021556902211159468, + 0.002506235148757696, + 0.0031763880979269743, + 0.0019182312535122037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.07015185058116913 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026388458907604218 + ], + "opposite_logprob_tokens": [ + -3.65138840675354 + ], + "logprob_reward": -0.026388458907604218, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026388458907604218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21818314492702484, + 0.14848016202449799, + 0.12133101373910904, + 0.11457540094852448, + 0.05166429281234741, + 0.06599803268909454, + 0.06442032009363174, + 0.11564084887504578, + 0.043648865073919296, + 0.0560578852891922 + ], + "credit_map": [ + 0.01105379406362772, + 0.018444206565618515, + 0.03342258185148239, + 0.028713848441839218, + 0.029217813163995743, + 0.012588628567755222, + 0.05907611548900604, + 0.02566615678369999, + 0.012197053991258144, + 0.03792296722531319, + 0.0157019030302763, + 0.021409913897514343, + 0.016264384612441063, + 0.008703107014298439, + 0.023932727053761482, + 0.012348117306828499, + 0.008538209833204746, + 0.011190085671842098, + 0.020402763038873672, + 0.02208719030022621, + 0.017838269472122192, + 0.007690011989325285, + 0.017258591949939728, + 0.01632588915526867, + 0.0084394421428442, + 0.013084636069834232, + 0.022450800985097885, + 0.015697836875915527, + 0.010191284120082855, + 0.01163642667233944, + 0.015101412311196327, + 0.01797354966402054, + 0.0029214967507869005, + 0.0069515397772192955, + 0.006374045740813017, + 0.008889663964509964, + 0.008821197785437107, + 0.0031931025441735983, + 0.009695366956293583, + 0.0048178816214203835, + 0.0076560815796256065, + 0.0058255624026060104, + 0.009886486455798149, + 0.010955557227134705, + 0.008607824333012104, + 0.005686160642653704, + 0.010687052272260189, + 0.0066933101043105125, + 0.008606410585343838, + 0.008291970007121563, + 0.011556765995919704, + 0.007740961387753487, + 0.005347442347556353, + 0.0061848219484090805, + 0.009072035551071167, + 0.007619915530085564, + 0.013137043453752995, + 0.013897210359573364, + 0.014043482951819897, + 0.01605835370719433, + 0.010633804835379124, + 0.02098066359758377, + 0.012120204977691174, + 0.014770089648663998, + 0.004118768498301506, + 0.0044052377343177795, + 0.004741630516946316, + 0.005500578787177801, + 0.009344052523374557, + 0.005883769132196903, + 0.0056143817491829395, + 0.00404044333845377, + 0.004449791274964809, + 0.004904733505100012, + 0.00974283181130886, + 0.0051386416889727116, + 0.004942082334309816, + 0.017093732953071594, + 0.006043616682291031, + 0.003742454806342721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026388458907604218 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819034039974213 + ], + "opposite_logprob_tokens": [ + -3.5298190116882324 + ], + "logprob_reward": -0.029819034039974213, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819034039974213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09659198671579361, + 0.06672115623950958, + 0.06922458857297897, + 0.066800557076931, + 0.05415227264165878, + 0.03482198715209961, + 0.0951467826962471, + 0.23026451468467712, + 0.19233356416225433, + 0.09394259750843048 + ], + "credit_map": [ + 0.006868297699838877, + 0.011356715112924576, + 0.013369161635637283, + 0.013692736625671387, + 0.012104290537536144, + 0.005433749873191118, + 0.025095872581005096, + 0.00867115706205368, + 0.005638149566948414, + 0.015346490778028965, + 0.006924471817910671, + 0.010376148857176304, + 0.00683691306039691, + 0.004249648656696081, + 0.01102067157626152, + 0.0063286591321229935, + 0.006473019253462553, + 0.007837679237127304, + 0.010150661692023277, + 0.013288659043610096, + 0.008927267976105213, + 0.004997795447707176, + 0.010156692005693913, + 0.007392812520265579, + 0.006735918577760458, + 0.00763237988576293, + 0.012126595713198185, + 0.008772586472332478, + 0.0061330716125667095, + 0.00786631740629673, + 0.007717031519860029, + 0.009816656820476055, + 0.004503884818404913, + 0.008070678450167179, + 0.007268659304827452, + 0.009164322167634964, + 0.005809722933918238, + 0.005772653501480818, + 0.0074797761626541615, + 0.0060825711116194725, + 0.003523631952702999, + 0.004727014806121588, + 0.005154523998498917, + 0.005532585084438324, + 0.0033154403790831566, + 0.002741527045145631, + 0.006347925867885351, + 0.003479336854070425, + 0.013301031664013863, + 0.008512965403497219, + 0.016273258253932, + 0.0148068368434906, + 0.008151606656610966, + 0.009670063853263855, + 0.01295522227883339, + 0.011475798673927784, + 0.02490714006125927, + 0.03727170452475548, + 0.025791890919208527, + 0.02568429708480835, + 0.031972017139196396, + 0.034278176724910736, + 0.02434968762099743, + 0.026009611785411835, + 0.014591304585337639, + 0.027624517679214478, + 0.02070193737745285, + 0.02367299795150757, + 0.03480956330895424, + 0.032710153609514236, + 0.02091044932603836, + 0.01731262169778347, + 0.007233592215925455, + 0.009136173874139786, + 0.015275506302714348, + 0.008449136279523373, + 0.01009413506835699, + 0.029571112245321274, + 0.008642400614917278, + 0.005540542770177126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.029819034039974213 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812784865498543 + ], + "opposite_logprob_tokens": [ + -3.529812812805176 + ], + "logprob_reward": -0.029812784865498543, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812784865498543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10743077099323273, + 0.06043040007352829, + 0.09720194339752197, + 0.1283552348613739, + 0.04665415361523628, + 0.08668328821659088, + 0.08786247670650482, + 0.18712501227855682, + 0.12397632002830505, + 0.07428032904863358 + ], + "credit_map": [ + 0.0077846460044384, + 0.008911089040338993, + 0.013143532909452915, + 0.01387300156056881, + 0.014910818077623844, + 0.008190164342522621, + 0.0309547558426857, + 0.009662769734859467, + 0.005003698635846376, + 0.009533185511827469, + 0.007641597650945187, + 0.009840616025030613, + 0.006668920628726482, + 0.004644639790058136, + 0.00822053849697113, + 0.008877206593751907, + 0.00850200280547142, + 0.010626012459397316, + 0.01268914993852377, + 0.018344998359680176, + 0.014246118254959583, + 0.007433148566633463, + 0.013688521459698677, + 0.011671988293528557, + 0.010433921590447426, + 0.017779409885406494, + 0.01990857720375061, + 0.017779558897018433, + 0.013494056649506092, + 0.015539661049842834, + 0.015657318755984306, + 0.01776273176074028, + 0.003359828842803836, + 0.0045187934301793575, + 0.0050352695398032665, + 0.009222437627613544, + 0.006085011176764965, + 0.004299344029277563, + 0.00748972874134779, + 0.0066437385976314545, + 0.0107077332213521, + 0.007283376529812813, + 0.011946487240493298, + 0.016213655471801758, + 0.01169166062027216, + 0.00588335283100605, + 0.01484835147857666, + 0.008108665235340595, + 0.011422649957239628, + 0.007644939236342907, + 0.01381695456802845, + 0.01553873810917139, + 0.007573637180030346, + 0.008660477586090565, + 0.012929756194353104, + 0.01027532760053873, + 0.018809041008353233, + 0.02980237454175949, + 0.02334839478135109, + 0.018974000588059425, + 0.027887247502803802, + 0.02732795663177967, + 0.019830577075481415, + 0.021145418286323547, + 0.01021676417440176, + 0.011470841243863106, + 0.015394449234008789, + 0.016889505088329315, + 0.024069635197520256, + 0.012518390081822872, + 0.019696205854415894, + 0.013720525428652763, + 0.006212909705936909, + 0.008737575262784958, + 0.007637643255293369, + 0.008904845453798771, + 0.009308706037700176, + 0.009843901731073856, + 0.017514465376734734, + 0.00612028269097209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.029812784865498543 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, static, low-angle shot, shot from above, lingers on the wet asphalt, clearly reflecting the shimmering cyan, magenta, and electric blue glow of a giant holographic billboard atop a futuristic skyscraper. On this rainy night steeped in cyberpunk, fine raindrops trace brilliant trails against the vibrant neon lights. At the center of the frame, a suspended mass of dark brown liquid and finely pulsated herbal granules—the visual embodiment of Huoxiang Zhengqi Powder—swirls slowly. Suddenly, the flowing mixture abruptly stops in mid-air. A jet of frost, as black as obsidian, spreads outward at a visible speed, instantly solidifying into a sharp, multifaceted black ice crystal. Simultaneously, countless delicate circuit patterns, emitting a faint blue glow, are instantly etched into the crystal's surface, interweaving and pulsating like a complex neural network. This interplay with the cold neon lights of the surrounding environment creates a mysterious, high-tech, and tense atmosphere.", + "seed": 18129, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_0/index_4044_seed_18129_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.31627590658608823, + "vlm_logprob": -0.31627590658608823, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.31627590658608823, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5041512250900269 + ], + "opposite_logprob_tokens": [ + -0.25415119528770447 + ], + "logprob_reward": -1.5041512250900269, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5041512250900269, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1096605509519577, + 0.1090228483080864, + 0.08112136274576187, + 0.3084487318992615, + 0.029737789183855057, + 0.24983733892440796, + 0.04673236235976219, + 0.025557933375239372, + 0.02342977188527584, + 0.01645132713019848 + ], + "credit_map": [ + 0.009869653731584549, + 0.0084628751501441, + 0.025413217023015022, + 0.00879481341689825, + 0.00959219690412283, + 0.013677438721060753, + 0.010272675193846226, + 0.023577673360705376, + 0.010085908696055412, + 0.010159456171095371, + 0.019929936155676842, + 0.009428242221474648, + 0.011088758707046509, + 0.01904195547103882, + 0.011618049815297127, + 0.017670532688498497, + 0.010026616044342518, + 0.005740379448980093, + 0.006756526883691549, + 0.007286288309842348, + 0.007595517206937075, + 0.01162786316126585, + 0.012456509284675121, + 0.019631652161478996, + 0.013517439365386963, + 0.006708455737680197, + 0.11607234925031662, + 0.03213352710008621, + 0.009952272288501263, + 0.0625196322798729, + 0.04101473465561867, + 0.026530295610427856, + 0.002002731431275606, + 0.0032447082921862602, + 0.003548621665686369, + 0.006109976675361395, + 0.0034470499958842993, + 0.004265668336302042, + 0.0043637715280056, + 0.0027552570682018995, + 0.02146725542843342, + 0.00706843426451087, + 0.08379160612821579, + 0.014956047758460045, + 0.006623984780162573, + 0.01992477849125862, + 0.021031340584158897, + 0.07497385889291763, + 0.004537858534604311, + 0.0036007575690746307, + 0.006883648689836264, + 0.0034275075886398554, + 0.0061050476506352425, + 0.006581255700439215, + 0.0082448935136199, + 0.007351391948759556, + 0.0028298916295170784, + 0.0034003015607595444, + 0.0024836105294525623, + 0.0018005031161010265, + 0.003305862657725811, + 0.004452845547348261, + 0.00304290559142828, + 0.00424200901761651, + 0.002142960438504815, + 0.0019470305414870381, + 0.003947739023715258, + 0.0019323258893564343, + 0.0018550099339336157, + 0.004926718771457672, + 0.0029547149315476418, + 0.0037232700269669294, + 0.0012873138766735792, + 0.0017681950703263283, + 0.0011707233497872949, + 0.0016914111329242587, + 0.002310799667611718, + 0.0036716621834784746, + 0.0026095297653228045, + 0.0019416901050135493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -1.5041512250900269 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054997172206640244 + ], + "opposite_logprob_tokens": [ + -2.929997205734253 + ], + "logprob_reward": -0.054997172206640244, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054997172206640244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16154102981090546, + 0.13404330611228943, + 0.11660531908273697, + 0.08370357751846313, + 0.026084892451763153, + 0.3690537214279175, + 0.032684653997421265, + 0.02105601876974106, + 0.03529922664165497, + 0.01992824487388134 + ], + "credit_map": [ + 0.014472932554781437, + 0.013539426028728485, + 0.03812408074736595, + 0.009593113325536251, + 0.011393998749554157, + 0.024466097354888916, + 0.013534446246922016, + 0.03641694039106369, + 0.012663912028074265, + 0.005841662175953388, + 0.02270095981657505, + 0.010127249173820019, + 0.00925051886588335, + 0.031964413821697235, + 0.015365377999842167, + 0.0261292215436697, + 0.021833542734384537, + 0.009128721430897713, + 0.010621120221912861, + 0.00996299460530281, + 0.007109928410500288, + 0.017850207164883614, + 0.01442286092787981, + 0.0256759412586689, + 0.0030972317326813936, + 0.0021637333557009697, + 0.028581522405147552, + 0.006972523406147957, + 0.00314721348695457, + 0.004591572098433971, + 0.02361666038632393, + 0.011533115059137344, + 0.0025212978944182396, + 0.0023043120745569468, + 0.0026827207766473293, + 0.00522457342594862, + 0.0036014714278280735, + 0.002968620043247938, + 0.003929697908461094, + 0.002852198202162981, + 0.03035215474665165, + 0.008271805942058563, + 0.137818843126297, + 0.016905277967453003, + 0.006720180623233318, + 0.024144122377038002, + 0.021358540281653404, + 0.12348278611898422, + 0.005255191586911678, + 0.002037599217146635, + 0.004630459938198328, + 0.002343333326280117, + 0.0031195380724966526, + 0.004541769158095121, + 0.005486911628395319, + 0.0052698529325425625, + 0.0021209779661148787, + 0.002181683899834752, + 0.0012988083763048053, + 0.0011286967201158404, + 0.0018042532028630376, + 0.006572724785655737, + 0.0013551147421821952, + 0.004593757912516594, + 0.0031804260797798634, + 0.004496436566114426, + 0.0054105366580188274, + 0.0039648800157010555, + 0.003602445125579834, + 0.005799056962132454, + 0.003825428197160363, + 0.00502001540735364, + 0.001439304556697607, + 0.0017841198714450002, + 0.0019755959510803223, + 0.0015530168311670423, + 0.0028111636638641357, + 0.005142693407833576, + 0.002476821653544903, + 0.002745529869571328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.054997172206640244 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084088012576103 + ], + "opposite_logprob_tokens": [ + -4.14108419418335 + ], + "logprob_reward": -0.016084088012576103, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084088012576103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10967694222927094, + 0.16212081909179688, + 0.039672743529081345, + 0.3576880395412445, + 0.03212398663163185, + 0.1729145497083664, + 0.026729116216301918, + 0.0393010713160038, + 0.04207335412502289, + 0.017699386924505234 + ], + "credit_map": [ + 0.008835899643599987, + 0.013686689548194408, + 0.02029922790825367, + 0.009871184825897217, + 0.009927995502948761, + 0.015104765072464943, + 0.009033387526869774, + 0.022917799651622772, + 0.011431709863245487, + 0.0168344397097826, + 0.017759060487151146, + 0.013844640925526619, + 0.02676188014447689, + 0.02561846934258938, + 0.016519658267498016, + 0.03335098177194595, + 0.0038375770673155785, + 0.0031871492974460125, + 0.00477327685803175, + 0.005607012193650007, + 0.005117970053106546, + 0.005577586125582457, + 0.0054074437357485294, + 0.006164730526506901, + 0.013627655804157257, + 0.0057925996370613575, + 0.13357840478420258, + 0.030497591942548752, + 0.010142186656594276, + 0.042029716074466705, + 0.0968281701207161, + 0.02519175037741661, + 0.0030201140325516462, + 0.0026639115530997515, + 0.004585302900522947, + 0.006821326445788145, + 0.0027281297370791435, + 0.004997087176889181, + 0.004430105909705162, + 0.0028780100401490927, + 0.013141934759914875, + 0.003943708725273609, + 0.06261662393808365, + 0.008687837049365044, + 0.003308299696072936, + 0.013601279817521572, + 0.010986861772835255, + 0.05662801116704941, + 0.0021783376578241587, + 0.0023966676089912653, + 0.0037853324320167303, + 0.003041751217097044, + 0.004809857811778784, + 0.003227783367037773, + 0.004340425133705139, + 0.00294896331615746, + 0.0038325837813317776, + 0.007069618906825781, + 0.0029146878514438868, + 0.002979699056595564, + 0.00843304954469204, + 0.005564420484006405, + 0.0044588204473257065, + 0.004048194270581007, + 0.00487164594233036, + 0.007814567536115646, + 0.0063988990150392056, + 0.004291636403650045, + 0.004813379142433405, + 0.005800905637443066, + 0.0035072152968496084, + 0.004575106780976057, + 0.0011678403243422508, + 0.0034693668130785227, + 0.0012225713580846786, + 0.00216332427226007, + 0.0021335508208721876, + 0.002737445058301091, + 0.002667167689651251, + 0.0021381231490522623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.016084088012576103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015125750796869397 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015125750796869397, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015125750796869397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11974801123142242, + 0.14023320376873016, + 0.09304751455783844, + 0.18406161665916443, + 0.05829780176281929, + 0.13754253089427948, + 0.06570696085691452, + 0.07886812090873718, + 0.08304405957460403, + 0.03945020213723183 + ], + "credit_map": [ + 0.01076592318713665, + 0.00914300512522459, + 0.028697412461042404, + 0.010210569016635418, + 0.00975880865007639, + 0.014641789719462395, + 0.012213398702442646, + 0.024317102506756783, + 0.010970554314553738, + 0.017871413379907608, + 0.02050928957760334, + 0.01592324674129486, + 0.01861266791820526, + 0.02251722477376461, + 0.015399422496557236, + 0.018429381772875786, + 0.008535013534128666, + 0.007592903915792704, + 0.00921060610562563, + 0.010088242590427399, + 0.01234336756169796, + 0.015708958730101585, + 0.013715983368456364, + 0.01585243083536625, + 0.008022541180253029, + 0.009306197054684162, + 0.04030243679881096, + 0.026573816314339638, + 0.011608834378421307, + 0.046101853251457214, + 0.01883552223443985, + 0.023310421034693718, + 0.0038033409509807825, + 0.004855025094002485, + 0.006436707451939583, + 0.010786810889840126, + 0.0076937805861234665, + 0.007437050808221102, + 0.011358566582202911, + 0.005926522426307201, + 0.010722958482801914, + 0.009365883655846119, + 0.021756742149591446, + 0.01638791337609291, + 0.009647768922150135, + 0.019421027973294258, + 0.026453601196408272, + 0.0237866323441267, + 0.006060481071472168, + 0.006022193469107151, + 0.007956343702971935, + 0.007802492473274469, + 0.009596183896064758, + 0.007912865839898586, + 0.009928897023200989, + 0.010427508503198624, + 0.00954973790794611, + 0.009098044596612453, + 0.006676412187516689, + 0.00509913032874465, + 0.00815103854984045, + 0.01430613361299038, + 0.008072135969996452, + 0.017915479838848114, + 0.011558782309293747, + 0.018601011484861374, + 0.00843051541596651, + 0.006601554807275534, + 0.009920707903802395, + 0.010570535436272621, + 0.00748106325045228, + 0.009879888035356998, + 0.004173912573605776, + 0.006757830269634724, + 0.002970863366499543, + 0.0037351399660110474, + 0.0045732008293271065, + 0.007306436076760292, + 0.0043596443720161915, + 0.005573173053562641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.0015125750796869397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634472541511059 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634472541511059, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634472541511059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0836912989616394, + 0.11344272643327713, + 0.05876278877258301, + 0.29082584381103516, + 0.04020226001739502, + 0.056494809687137604, + 0.07126590609550476, + 0.062027543783187866, + 0.055642105638980865, + 0.167644664645195 + ], + "credit_map": [ + 0.007220475003123283, + 0.00685003399848938, + 0.019117945805191994, + 0.007828576490283012, + 0.007499882020056248, + 0.011384041048586369, + 0.008205387741327286, + 0.015584966167807579, + 0.008323533460497856, + 0.010050653479993343, + 0.012805722653865814, + 0.010430729947984219, + 0.011797132901847363, + 0.023064276203513145, + 0.010867261327803135, + 0.026103418320417404, + 0.005638399161398411, + 0.0045353928580880165, + 0.007046273909509182, + 0.007485492154955864, + 0.005123514216393232, + 0.00795003678649664, + 0.008330647833645344, + 0.012653036043047905, + 0.010730745270848274, + 0.006672183983027935, + 0.015434535220265388, + 0.04458901286125183, + 0.013623746111989021, + 0.04877842962741852, + 0.11306219547986984, + 0.037935011088848114, + 0.0032730293460190296, + 0.003977711778134108, + 0.004338039550930262, + 0.006388980429619551, + 0.005377814639359713, + 0.0065944925881922245, + 0.005940679926425219, + 0.004311517346650362, + 0.004143532365560532, + 0.002376770367845893, + 0.01606553979218006, + 0.004512361250817776, + 0.0027092064265161753, + 0.006501774303615093, + 0.006547309923917055, + 0.013638320378959179, + 0.005668943282216787, + 0.0048314654268324375, + 0.016973985359072685, + 0.0054054856300354, + 0.005993303377181292, + 0.012259862385690212, + 0.010660786181688309, + 0.009472081437706947, + 0.006569244433194399, + 0.007868608459830284, + 0.004439469892531633, + 0.0035749671515077353, + 0.006062292493879795, + 0.012919112108647823, + 0.00583771662786603, + 0.014756138436496258, + 0.005620288196951151, + 0.008666874840855598, + 0.010142953135073185, + 0.0060779317282140255, + 0.0062577929347753525, + 0.006011148914694786, + 0.00503546092659235, + 0.007829660549759865, + 0.010127104818820953, + 0.028556784614920616, + 0.011502022854983807, + 0.019512999802827835, + 0.03279340639710426, + 0.01975567825138569, + 0.02057131938636303, + 0.024825364351272583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.004634472541511059 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, static, low-angle shot, shot from above, lingers on the wet asphalt, clearly reflecting the shimmering cyan, magenta, and electric blue glow of a giant holographic billboard atop a futuristic skyscraper. On this rainy night steeped in cyberpunk, fine raindrops trace brilliant trails against the vibrant neon lights. At the center of the frame, a suspended mass of dark brown liquid and finely pulsated herbal granules—the visual embodiment of Huoxiang Zhengqi Powder—swirls slowly. Suddenly, the flowing mixture abruptly stops in mid-air. A jet of frost, as black as obsidian, spreads outward at a visible speed, instantly solidifying into a sharp, multifaceted black ice crystal. Simultaneously, countless delicate circuit patterns, emitting a faint blue glow, are instantly etched into the crystal's surface, interweaving and pulsating like a complex neural network. This interplay with the cold neon lights of the surrounding environment creates a mysterious, high-tech, and tense atmosphere.", + "seed": 18130, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_0/index_4044_seed_18130_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.6957317279651761, + "vlm_logprob": -0.6957317279651761, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.6957317279651761, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.2902414798736572 + ], + "opposite_logprob_tokens": [ + -0.040241457521915436 + ], + "logprob_reward": -3.2902414798736572, + "yes_no_margin_reward": -3.25, + "token_credit_reward": -3.2902414798736572, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.787539005279541, + 0.02413075417280197, + 0.02835816703736782, + 0.011216878890991211, + 0.02015785314142704, + 0.07939473539590836, + 0.006776774767786264, + 0.013522392138838768, + 0.009588922373950481, + 0.01931449957191944 + ], + "credit_map": [ + 0.04319174215197563, + 0.04562455788254738, + 0.2583657205104828, + 0.16044412553310394, + 0.028168877586722374, + 0.16213074326515198, + 0.044390223920345306, + 0.045223038643598557, + 0.0022296865936368704, + 0.002041294239461422, + 0.003628355450928211, + 0.0028904799837619066, + 0.0027142618782818317, + 0.003307974198833108, + 0.0029948658775538206, + 0.004323835950344801, + 0.0031744211446493864, + 0.0013667554594576359, + 0.009345129132270813, + 0.003338847542181611, + 0.0011594496900215745, + 0.006042709574103355, + 0.0021222366485744715, + 0.001808619243092835, + 0.0019267983734607697, + 0.001478063641116023, + 0.002084780251607299, + 0.0012056090636178851, + 0.001583308563567698, + 0.001023992896080017, + 0.0009337778319604695, + 0.0009805484442040324, + 0.001661893678829074, + 0.002829550299793482, + 0.0024793031625449657, + 0.0022249219473451376, + 0.0036024798173457384, + 0.002264489885419607, + 0.003040140960365534, + 0.0020550733897835016, + 0.005556813906878233, + 0.01365108136087656, + 0.009292692877352238, + 0.018050530925393105, + 0.010028720833361149, + 0.006411876995116472, + 0.009464002214372158, + 0.00693902000784874, + 0.000870680611114949, + 0.0009989449754357338, + 0.0007436913438141346, + 0.0007215329096652567, + 0.0009284354164265096, + 0.0010025020455941558, + 0.0007316952105611563, + 0.0007792927208356559, + 0.0014775670133531094, + 0.0010847320081666112, + 0.0033509954810142517, + 0.0009272253955714405, + 0.001006059581413865, + 0.0015883861342445016, + 0.0012306947028264403, + 0.0028567330446094275, + 0.000965341052506119, + 0.0010273281950503588, + 0.0015945114428177476, + 0.0012339194072410464, + 0.000985497492365539, + 0.0013995774788782, + 0.0013387061189860106, + 0.0010440419428050518, + 0.005397203378379345, + 0.0014120419509708881, + 0.003769888309761882, + 0.0011567872716113925, + 0.0009680254734121263, + 0.002534512896090746, + 0.0015114680863916874, + 0.002564573660492897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -3.2902414798736572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12721696496009827 + ], + "opposite_logprob_tokens": [ + -2.1272170543670654 + ], + "logprob_reward": -0.12721696496009827, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12721696496009827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8521561026573181, + 0.020057762041687965, + 0.017296331003308296, + 0.01597638614475727, + 0.019282814115285873, + 0.026129262521862984, + 0.008647270500659943, + 0.012903238646686077, + 0.00796419009566307, + 0.019586632028222084 + ], + "credit_map": [ + 0.046356428414583206, + 0.04951649531722069, + 0.2789344787597656, + 0.17306725680828094, + 0.030385266989469528, + 0.17555324733257294, + 0.04852263629436493, + 0.04982027783989906, + 0.001749772229231894, + 0.001711733639240265, + 0.0029838206246495247, + 0.002816802356392145, + 0.001852679532021284, + 0.002567616291344166, + 0.0025467711966484785, + 0.0038285651244223118, + 0.001971662510186434, + 0.0010393582051619887, + 0.0050733559764921665, + 0.0023602028377354145, + 0.0008789494750089943, + 0.0033236900344491005, + 0.0013091269647702575, + 0.0013399841263890266, + 0.002162642776966095, + 0.001866052858531475, + 0.002454211702570319, + 0.002451618667691946, + 0.002051956718787551, + 0.0012133184354752302, + 0.001811388647183776, + 0.0019651949405670166, + 0.0012223436497151852, + 0.0038805692456662655, + 0.002190743573009968, + 0.00148063525557518, + 0.004170013591647148, + 0.002242035698145628, + 0.002506360411643982, + 0.0015901117585599422, + 0.0030250598210841417, + 0.002322620013728738, + 0.00446448614820838, + 0.002034675097092986, + 0.0018943047616630793, + 0.002549012890085578, + 0.0031159284990280867, + 0.006723173428326845, + 0.001222733873873949, + 0.001205605105496943, + 0.001733470824547112, + 0.0007655197405256331, + 0.0008403834654018283, + 0.001405674614943564, + 0.0006497796275652945, + 0.0008241023169830441, + 0.0016701092245057225, + 0.0007400998147204518, + 0.004774569533765316, + 0.0007734447717666626, + 0.0006728390580974519, + 0.000759621150791645, + 0.0007789622177369893, + 0.0027335917111486197, + 0.0007933665183372796, + 0.0011110127670690417, + 0.0011105743469670415, + 0.000887701113242656, + 0.0009557373123243451, + 0.0012250171275809407, + 0.0009629892883822322, + 0.0009177917381748557, + 0.005245236214250326, + 0.001124887727200985, + 0.0011677616275846958, + 0.0007889497210271657, + 0.0008217694121412933, + 0.0024841788690537214, + 0.0011066125007346272, + 0.006847235839813948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.12721696496009827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026392173022031784 + ], + "opposite_logprob_tokens": [ + -3.6513922214508057 + ], + "logprob_reward": -0.026392173022031784, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026392173022031784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5019742250442505, + 0.030710164457559586, + 0.04950189217925072, + 0.023684551939368248, + 0.026663273572921753, + 0.1717357635498047, + 0.010423506610095501, + 0.009809982031583786, + 0.013490465469658375, + 0.16200615465641022 + ], + "credit_map": [ + 0.027176963165402412, + 0.028851380571722984, + 0.1668984293937683, + 0.10433636605739594, + 0.017944762483239174, + 0.10118436068296432, + 0.027576817199587822, + 0.028005167841911316, + 0.002964307786896825, + 0.00260105449706316, + 0.0057286289520561695, + 0.003612537868320942, + 0.002507611410692334, + 0.002273530699312687, + 0.006008237134665251, + 0.005014259368181229, + 0.003912457264959812, + 0.004498352762311697, + 0.01068554911762476, + 0.006350224371999502, + 0.004760521464049816, + 0.005652798805385828, + 0.008198395371437073, + 0.005443598609417677, + 0.0027346364222466946, + 0.0037231124006211758, + 0.0032155406661331654, + 0.002385964384302497, + 0.003645618911832571, + 0.002819913672283292, + 0.001990665215998888, + 0.003169103991240263, + 0.002271926961839199, + 0.004456572700291872, + 0.0032352085690945387, + 0.0026601434219628572, + 0.004869505297392607, + 0.003187609137967229, + 0.003461712272837758, + 0.0025206003338098526, + 0.013647310435771942, + 0.028480160981416702, + 0.020679257810115814, + 0.04031127318739891, + 0.02020280994474888, + 0.013742930255830288, + 0.02029281295835972, + 0.014379223808646202, + 0.0014697457663714886, + 0.0011574019445106387, + 0.0024889609776437283, + 0.001024151686578989, + 0.0009286567801609635, + 0.0012818571412935853, + 0.0011713295243680477, + 0.0009014037786982954, + 0.000895904318895191, + 0.0009455181425437331, + 0.002350345253944397, + 0.0008765956154093146, + 0.0009271949529647827, + 0.0009189317934215069, + 0.0011496186489239335, + 0.0017458739457651973, + 0.0011067033046856523, + 0.0024009502958506346, + 0.0016196012729778886, + 0.0017930256435647607, + 0.0015790731413289905, + 0.0019419984892010689, + 0.0017085459548979998, + 0.0013405687641352415, + 0.04984647408127785, + 0.008090481162071228, + 0.018772481009364128, + 0.005033445078879595, + 0.005394347477704287, + 0.017942003905773163, + 0.006263952236622572, + 0.05066297948360443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.026392173022031784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020604494959115982 + ], + "opposite_logprob_tokens": [ + -3.8956046104431152 + ], + "logprob_reward": -0.020604494959115982, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020604494959115982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5210662484169006, + 0.01564093492925167, + 0.01746848039329052, + 0.01262530218809843, + 0.020195871591567993, + 0.10200589895248413, + 0.015001908876001835, + 0.016162723302841187, + 0.04714290797710419, + 0.23268967866897583 + ], + "credit_map": [ + 0.027849145233631134, + 0.02967703342437744, + 0.17298448085784912, + 0.1069682389497757, + 0.018609529361128807, + 0.10680875182151794, + 0.028990386053919792, + 0.029178684577345848, + 0.0014050541212782264, + 0.0010508010163903236, + 0.0027462223079055548, + 0.0019759933929890394, + 0.0013027368113398552, + 0.0011175817344337702, + 0.0030471356585621834, + 0.0029954100027680397, + 0.0032384786754846573, + 0.0015157441375777125, + 0.002983426908031106, + 0.002125687897205353, + 0.0013516308972612023, + 0.00200645555742085, + 0.0016090493882074952, + 0.0026380063500255346, + 0.0016414872370660305, + 0.0016797115094959736, + 0.0019919213373214006, + 0.0013036163290962577, + 0.0018293707398697734, + 0.0014086486771702766, + 0.0014326181262731552, + 0.001337928813882172, + 0.0018576545407995582, + 0.0033242383506149054, + 0.0025173919275403023, + 0.0017018122598528862, + 0.003913384862244129, + 0.002589528216049075, + 0.002631960203871131, + 0.0016599004156887531, + 0.006960168015211821, + 0.016385702416300774, + 0.012713962234556675, + 0.02246711775660515, + 0.01375002320855856, + 0.008358698338270187, + 0.012295630760490894, + 0.00907459668815136, + 0.002458876231685281, + 0.0017556893872097135, + 0.003292273497208953, + 0.0013089077547192574, + 0.0020202717278152704, + 0.0014510349137708545, + 0.001479997648857534, + 0.0012348568998277187, + 0.00167581788264215, + 0.0013554096221923828, + 0.0044508096762001514, + 0.0012000934220850468, + 0.0013451010454446077, + 0.0011774349259212613, + 0.0015696536283940077, + 0.003388403682038188, + 0.004919509869068861, + 0.007219208404421806, + 0.004330069292336702, + 0.005794971249997616, + 0.005837589502334595, + 0.006719614379107952, + 0.005721316207200289, + 0.006600630469620228, + 0.0739770457148552, + 0.012801773846149445, + 0.0309619028121233, + 0.007859629578888416, + 0.008383670821785927, + 0.02275417372584343, + 0.008444799110293388, + 0.06750667840242386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.020604494959115982 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot looking up at a skyscraper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rainy night with wet asphalt reflecting cyan, magenta, and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a suspended mass of dark brown liquid and herbal granules visible swirling in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mixture abruptly stop in mid-air and transform into a sharp, multifaceted black ice crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate circuit patterns emitting a faint blue glow etched onto the surface of the black ice crystal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203527010977268 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203527010977268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203527010977268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17446351051330566, + 0.043090641498565674, + 0.056669656187295914, + 0.02566586434841156, + 0.04092247784137726, + 0.39582571387290955, + 0.027130411937832832, + 0.026149064302444458, + 0.036364708095788956, + 0.17371797561645508 + ], + "credit_map": [ + 0.009575244039297104, + 0.010682526975870132, + 0.054599422961473465, + 0.03577537462115288, + 0.00866742804646492, + 0.034065622836351395, + 0.0109480582177639, + 0.010149842128157616, + 0.004301695618778467, + 0.0030950463842600584, + 0.0071408916264772415, + 0.004930045455694199, + 0.004230552352964878, + 0.004500184208154678, + 0.007139426656067371, + 0.007752798032015562, + 0.007190293166786432, + 0.0032665403559803963, + 0.016321875154972076, + 0.006981526035815477, + 0.002971033798530698, + 0.010534754022955894, + 0.004182502161711454, + 0.0052211349830031395, + 0.005165716167539358, + 0.0026914544869214296, + 0.005472317337989807, + 0.0032853223383426666, + 0.0022487142123281956, + 0.001604576944373548, + 0.0024588010273873806, + 0.0027389642782509327, + 0.0029036281630396843, + 0.00680328207090497, + 0.004405118525028229, + 0.004506540484726429, + 0.007393424864858389, + 0.004523977171629667, + 0.0060765910893678665, + 0.004309917334467173, + 0.02911665476858616, + 0.0684414952993393, + 0.04689071699976921, + 0.09296749532222748, + 0.047603853046894073, + 0.031613096594810486, + 0.04693571850657463, + 0.03225666284561157, + 0.003396956017240882, + 0.003139529377222061, + 0.004550097975879908, + 0.0030389016028493643, + 0.0030877392273396254, + 0.0031271190382540226, + 0.003697795793414116, + 0.0030922715086489916, + 0.002652285620570183, + 0.002845770912244916, + 0.0030381062533706427, + 0.002553042955696583, + 0.003078076522797346, + 0.005519366357475519, + 0.0030437218956649303, + 0.0034186949487775564, + 0.003600488882511854, + 0.004935517441481352, + 0.004175860900431871, + 0.004014961887151003, + 0.005645092576742172, + 0.004390780348330736, + 0.00583307258784771, + 0.003768936498090625, + 0.05144684389233589, + 0.008971797302365303, + 0.018753185868263245, + 0.0062567186541855335, + 0.00634760083630681, + 0.02063322626054287, + 0.007171312812715769, + 0.05413730442523956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.014203527010977268 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_tensor([4542, 4542])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_tensor([4542, 4542])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3f712c81b9873a7653b5b8eaa57555408853e383 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_tensor([4542, 4542])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, deliberate camera pans from a slightly low angle to capture a medium close-up of a young Hispanic woman, her face a mask of intense concentration. Her dark, damp hair is slicked back, revealing high cheekbones and a resolute jawline, all partially illuminated by the scene's ambient light. Intricate, glowing cyan cybernetic circuits, like a luminous tattoo, lie embedded beneath the skin of her right arm, shimmering with a soft, cool light, tracing the lines of muscle from forearm to fingertips. She wears a worn black synthetic jacket, its surface shimmering with rainwater beneath the perpetual twilight of a dystopian city. The scene is saturated with the cold, vibrant glow of the towering neon signs lining the rain-slicked alleyway, casting long, distorted reflections of purplish red and electric blue on the grimy, corroded metal wall before her. Steadily and deliberately, she extends her arm and fingers, each tipped with a glowing node of a cybernetic pattern, to grasp a recessed data port in the wall. As her fingertips make precise contact with the port, the cybernetic tattoo flashes brightly for a moment, casting sharp shadows on her focused face and illuminating the fine raindrops on the wall's surface, immersing the entire composition in a gritty dystopian sci-fi style, creating a tense and secretive atmosphere.", + "seed": 16155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_4542_seed_16155_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016783644212409854, + "vlm_logprob": -0.016783644212409854, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.016783644212409854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04902328923344612 + ], + "opposite_logprob_tokens": [ + -3.049023389816284 + ], + "logprob_reward": -0.04902328923344612, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04902328923344612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31753772497177124, + 0.08667375147342682, + 0.07319098711013794, + 0.07034283876419067, + 0.1443735957145691, + 0.10339316725730896, + 0.08566207438707352, + 0.0371682271361351, + 0.060682542622089386, + 0.02097512036561966 + ], + "credit_map": [ + 0.02464423142373562, + 0.024173589423298836, + 0.04944086819887161, + 0.048354826867580414, + 0.06668595969676971, + 0.0259547121822834, + 0.045698992908000946, + 0.03258451446890831, + 0.011249072849750519, + 0.006529456470161676, + 0.013397066853940487, + 0.00941634550690651, + 0.0128090288490057, + 0.008663774468004704, + 0.011485991068184376, + 0.013123011216521263, + 0.010073313489556313, + 0.013505203649401665, + 0.008283170871436596, + 0.011973919346928596, + 0.006907126400619745, + 0.006566519849002361, + 0.007976086810231209, + 0.007905637845396996, + 0.006668826099485159, + 0.013009503483772278, + 0.00757832545787096, + 0.008774375542998314, + 0.014345833100378513, + 0.006371477153152227, + 0.00754927285015583, + 0.006045224145054817, + 0.019572855904698372, + 0.02408084273338318, + 0.016544276848435402, + 0.02111431397497654, + 0.02086726576089859, + 0.011090399697422981, + 0.015510309487581253, + 0.015593315474689007, + 0.008626529015600681, + 0.01196043286472559, + 0.009964073076844215, + 0.012586405500769615, + 0.014027608558535576, + 0.020053554326295853, + 0.011535318568348885, + 0.014639238826930523, + 0.005573604721575975, + 0.01669534668326378, + 0.009547773748636246, + 0.008080211468040943, + 0.014561038464307785, + 0.00748943816870451, + 0.00661127595230937, + 0.01710337959229946, + 0.002925593638792634, + 0.0031587081030011177, + 0.006733374670147896, + 0.005941472947597504, + 0.003381385002285242, + 0.0034013781696558, + 0.0055737365037202835, + 0.0060525741428136826, + 0.0072633023373782635, + 0.011888913810253143, + 0.008638382889330387, + 0.004969108384102583, + 0.006730066146701574, + 0.008580310270190239, + 0.004190634936094284, + 0.008421818725764751, + 0.0020105415023863316, + 0.00352868577465415, + 0.002317512407898903, + 0.0019547429401427507, + 0.0031800330616533756, + 0.002405044622719288, + 0.0018841720884665847, + 0.003694384591653943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.04902328923344612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024975319392979145 + ], + "opposite_logprob_tokens": [ + -6.002497673034668 + ], + "logprob_reward": -0.0024975319392979145, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024975319392979145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1068204715847969, + 0.08964680135250092, + 0.09028209745883942, + 0.15717482566833496, + 0.07408561557531357, + 0.055748965591192245, + 0.07603120058774948, + 0.17974719405174255, + 0.08870140463113785, + 0.0817614272236824 + ], + "credit_map": [ + 0.009591182693839073, + 0.009199277497828007, + 0.025033781304955482, + 0.014432373456656933, + 0.011858034878969193, + 0.010919302701950073, + 0.013063464313745499, + 0.012723052874207497, + 0.015035659074783325, + 0.008804245851933956, + 0.014329471625387669, + 0.011059029959142208, + 0.012968242168426514, + 0.008606665767729282, + 0.009111600928008556, + 0.009731881320476532, + 0.010416111908853054, + 0.011815986596047878, + 0.013119371607899666, + 0.017792919650673866, + 0.009563732892274857, + 0.008388664573431015, + 0.01114467903971672, + 0.00804063305258751, + 0.018571412190794945, + 0.018108483403921127, + 0.015907401219010353, + 0.025349071249365807, + 0.02575955167412758, + 0.015043163672089577, + 0.023614052683115005, + 0.014821686781942844, + 0.007122754119336605, + 0.015500493347644806, + 0.005646631121635437, + 0.009519414976239204, + 0.015817852690815926, + 0.0069520799443125725, + 0.006408464163541794, + 0.00711792754009366, + 0.004874115344136953, + 0.009270916692912579, + 0.005865280982106924, + 0.005123070441186428, + 0.008829633705317974, + 0.007732002064585686, + 0.006551691330969334, + 0.007502253632992506, + 0.005017234478145838, + 0.014462270773947239, + 0.008167075924575329, + 0.006122218444943428, + 0.013214502483606339, + 0.00923042744398117, + 0.006306214723736048, + 0.01351125817745924, + 0.011603137478232384, + 0.020815979689359665, + 0.02581833116710186, + 0.033647265285253525, + 0.019523469731211662, + 0.017841368913650513, + 0.02467256598174572, + 0.025825070217251778, + 0.007458471227437258, + 0.00991368480026722, + 0.010185741819441319, + 0.014340035617351532, + 0.008514179848134518, + 0.012225596234202385, + 0.013640334829688072, + 0.01242335606366396, + 0.007024682126939297, + 0.010359062813222408, + 0.010749660432338715, + 0.011439451016485691, + 0.010518282651901245, + 0.008346978574991226, + 0.00912909023463726, + 0.014194218441843987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.0024975319392979145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637567982077599 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637567982077599, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637567982077599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2210969477891922, + 0.20125272870063782, + 0.08547502011060715, + 0.1611286699771881, + 0.05195681005716324, + 0.05185442045331001, + 0.04956916347146034, + 0.07922998070716858, + 0.05930495634675026, + 0.039131369441747665 + ], + "credit_map": [ + 0.017099987715482712, + 0.013588403351604939, + 0.04805595055222511, + 0.03548310697078705, + 0.02891344577074051, + 0.020392365753650665, + 0.033972833305597305, + 0.023590823635458946, + 0.03782922774553299, + 0.01675841026008129, + 0.0300059262663126, + 0.022815659642219543, + 0.028946978971362114, + 0.018976520746946335, + 0.021609129384160042, + 0.02431086264550686, + 0.012622369453310966, + 0.009757274761795998, + 0.013131803832948208, + 0.011505342088639736, + 0.009451585821807384, + 0.011143695563077927, + 0.0071290237829089165, + 0.010733922012150288, + 0.013787181116640568, + 0.023249907419085503, + 0.0221991129219532, + 0.018455125391483307, + 0.023723304271697998, + 0.015405534766614437, + 0.024905500933527946, + 0.019402993842959404, + 0.006836752872914076, + 0.01079630944877863, + 0.0040826634503901005, + 0.004136084113270044, + 0.008013741113245487, + 0.007155838422477245, + 0.0037520367186516523, + 0.0071833753027021885, + 0.004449324682354927, + 0.00727431382983923, + 0.006556387525051832, + 0.004841295070946217, + 0.00938139483332634, + 0.008229483850300312, + 0.004663683474063873, + 0.006458532530814409, + 0.0035552848130464554, + 0.008423566818237305, + 0.005587248131632805, + 0.003884760895743966, + 0.011677381582558155, + 0.0044118864461779594, + 0.004558701999485493, + 0.007470325101166964, + 0.0075233313255012035, + 0.011193647980690002, + 0.012591082602739334, + 0.00753721734508872, + 0.0077156140469014645, + 0.006986573338508606, + 0.007396241184324026, + 0.018286269158124924, + 0.007540654391050339, + 0.012268544174730778, + 0.005700057838112116, + 0.005867870524525642, + 0.008051920682191849, + 0.006802817340940237, + 0.004158337134867907, + 0.008914750069379807, + 0.003792398376390338, + 0.00641725305467844, + 0.005474984645843506, + 0.004342847969383001, + 0.005573304835706949, + 0.004660318605601788, + 0.003308961633592844, + 0.0055612968280911446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008637567982077599 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572848618030548 + ], + "opposite_logprob_tokens": [ + -3.8955729007720947 + ], + "logprob_reward": -0.020572848618030548, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572848618030548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15407180786132812, + 0.06793881207704544, + 0.079937644302845, + 0.04732612892985344, + 0.09991171211004257, + 0.09254268556833267, + 0.16638687252998352, + 0.18988037109375, + 0.07122422009706497, + 0.030779769644141197 + ], + "credit_map": [ + 0.013409879989922047, + 0.015693219378590584, + 0.02927306666970253, + 0.02482525072991848, + 0.020738372579216957, + 0.012412856332957745, + 0.02050553448498249, + 0.017213603481650352, + 0.012362560257315636, + 0.006851673126220703, + 0.008767428807914257, + 0.006211118772625923, + 0.009858368895947933, + 0.006785548757761717, + 0.0053666625171899796, + 0.01173543930053711, + 0.007307716179639101, + 0.008947761729359627, + 0.011770825833082199, + 0.017166392877697945, + 0.008798502385616302, + 0.008231235668063164, + 0.01101716235280037, + 0.0066980356350541115, + 0.004891288001090288, + 0.004186235833913088, + 0.004750877618789673, + 0.008134743198752403, + 0.007281797472387552, + 0.004685316700488329, + 0.007548329886049032, + 0.005847535561770201, + 0.012252619490027428, + 0.019502218812704086, + 0.007984050549566746, + 0.012016898952424526, + 0.0221052635461092, + 0.00858097244054079, + 0.008530568331480026, + 0.008939109742641449, + 0.006464924663305283, + 0.013559546321630478, + 0.00919854361563921, + 0.006149906199425459, + 0.01752069592475891, + 0.021300002932548523, + 0.006405226420611143, + 0.011943827383220196, + 0.01235515158623457, + 0.023079045116901398, + 0.024595823138952255, + 0.015701981261372566, + 0.01679862104356289, + 0.021197609603405, + 0.019814904779195786, + 0.03284372016787529, + 0.01149334292858839, + 0.01972048729658127, + 0.038773395121097565, + 0.02340431697666645, + 0.013352260924875736, + 0.019090311601758003, + 0.019881755113601685, + 0.04416448622941971, + 0.008396399207413197, + 0.013029401190578938, + 0.006665762048214674, + 0.008444939740002155, + 0.009291964583098888, + 0.008352173492312431, + 0.006428868509829044, + 0.010614700615406036, + 0.003325551049783826, + 0.004094681236892939, + 0.0030902971047908068, + 0.003318816889077425, + 0.004553265403956175, + 0.003904018085449934, + 0.0027128332294523716, + 0.005780302453786135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.020572848618030548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031869832891970873 + ], + "opposite_logprob_tokens": [ + -5.75318717956543 + ], + "logprob_reward": -0.0031869832891970873, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031869832891970873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18866989016532898, + 0.19669441878795624, + 0.09892357140779495, + 0.0719781219959259, + 0.07177970558404922, + 0.056258365511894226, + 0.09285344928503036, + 0.15548579394817352, + 0.03897806629538536, + 0.028378605842590332 + ], + "credit_map": [ + 0.017222104594111443, + 0.015071659348905087, + 0.031370941549539566, + 0.03124048560857773, + 0.02912423573434353, + 0.016345441341400146, + 0.026135172694921494, + 0.022159842774271965, + 0.034827813506126404, + 0.02480892837047577, + 0.025461262091994286, + 0.021482013165950775, + 0.029722176492214203, + 0.01784786768257618, + 0.01807507686316967, + 0.024469271302223206, + 0.013460399582982063, + 0.013362128287553787, + 0.012702002190053463, + 0.015109475702047348, + 0.010242492891848087, + 0.0090090436860919, + 0.010516947135329247, + 0.014521078206598759, + 0.0056357975117862225, + 0.00777633348479867, + 0.008459586650133133, + 0.00931025855243206, + 0.010624421760439873, + 0.006953405681997538, + 0.013726209290325642, + 0.00949210487306118, + 0.012822361662983894, + 0.00948527455329895, + 0.007858305238187313, + 0.00859715137630701, + 0.013709528371691704, + 0.006464850157499313, + 0.005297004710882902, + 0.007545223459601402, + 0.003595438553020358, + 0.009270456619560719, + 0.004235421307384968, + 0.003988854121416807, + 0.01339526753872633, + 0.009230556897819042, + 0.003570514963939786, + 0.00897185504436493, + 0.00558199780061841, + 0.015355699695646763, + 0.007857567630708218, + 0.010084749199450016, + 0.01498982310295105, + 0.013888003304600716, + 0.012741082347929478, + 0.01235452014952898, + 0.013250591233372688, + 0.014817104674875736, + 0.033171024173498154, + 0.02009858563542366, + 0.010870691388845444, + 0.01281492505222559, + 0.014682033099234104, + 0.03578082472085953, + 0.004038872197270393, + 0.006802268326282501, + 0.004097099415957928, + 0.004824058152735233, + 0.005489893723279238, + 0.004089091904461384, + 0.0034171510487794876, + 0.006219629198312759, + 0.00361234019510448, + 0.004680979531258345, + 0.0038290880620479584, + 0.0025770419742912054, + 0.003720010630786419, + 0.0032753380946815014, + 0.0022997779306024313, + 0.0043840291909873486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0031869832891970873 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate camera pans from a slightly low angle to capture a medium close-up of a young Hispanic woman, her face a mask of intense concentration. Her dark, damp hair is slicked back, revealing high cheekbones and a resolute jawline, all partially illuminated by the scene's ambient light. Intricate, glowing cyan cybernetic circuits, like a luminous tattoo, lie embedded beneath the skin of her right arm, shimmering with a soft, cool light, tracing the lines of muscle from forearm to fingertips. She wears a worn black synthetic jacket, its surface shimmering with rainwater beneath the perpetual twilight of a dystopian city. The scene is saturated with the cold, vibrant glow of the towering neon signs lining the rain-slicked alleyway, casting long, distorted reflections of purplish red and electric blue on the grimy, corroded metal wall before her. Steadily and deliberately, she extends her arm and fingers, each tipped with a glowing node of a cybernetic pattern, to grasp a recessed data port in the wall. As her fingertips make precise contact with the port, the cybernetic tattoo flashes brightly for a moment, casting sharp shadows on her focused face and illuminating the fine raindrops on the wall's surface, immersing the entire composition in a gritty dystopian sci-fi style, creating a tense and secretive atmosphere.", + "seed": 16156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_4542_seed_16156_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.004366785031743347, + "vlm_logprob": -0.004366785031743347, + "vlm_yes_no_margin": 5.625, + "vlm_token_credit": -0.004366785031743347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007781785447150469 + ], + "opposite_logprob_tokens": [ + -5.007781982421875 + ], + "logprob_reward": -0.007781785447150469, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007781785447150469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032918889075517654, + 0.03186376020312309, + 0.009877888485789299, + 0.012974198907613754, + 0.023852413520216942, + 0.011409488506615162, + 0.015917129814624786, + 0.019327739253640175, + 0.00784880667924881, + 0.8340097665786743 + ], + "credit_map": [ + 0.0020732502453029156, + 0.0021101664751768112, + 0.0046862647868692875, + 0.0023549741599708796, + 0.008044042624533176, + 0.0026536560617387295, + 0.005933671724051237, + 0.005062857177108526, + 0.0018655206076800823, + 0.0013637267984449863, + 0.008252832107245922, + 0.0029977497179061174, + 0.004526324570178986, + 0.0014179847203195095, + 0.008144231513142586, + 0.0032953876070678234, + 0.0006460867589339614, + 0.001108204829506576, + 0.0009613399743102491, + 0.0012103361077606678, + 0.0018958906875923276, + 0.0008518535760231316, + 0.001552367932163179, + 0.00165180757176131, + 0.001341236406005919, + 0.0009278350044041872, + 0.0013368797954171896, + 0.001611244399100542, + 0.001165981637313962, + 0.0036574993282556534, + 0.0013421078911051154, + 0.0015914130490273237, + 0.0016842834884300828, + 0.005882448051124811, + 0.0021385035943239927, + 0.0034700436517596245, + 0.0033238225150853395, + 0.0024258450139313936, + 0.002395785879343748, + 0.00253167818300426, + 0.00106631510425359, + 0.001863341429270804, + 0.0010582831455394626, + 0.0015810603508725762, + 0.0020690755918622017, + 0.0014407335547730327, + 0.0012470564106479287, + 0.0010836217552423477, + 0.0022305166348814964, + 0.0019353694515302777, + 0.001914509921334684, + 0.002666270127519965, + 0.001825669314712286, + 0.0014798708725720644, + 0.002133494010195136, + 0.0017314271535724401, + 0.0020413482561707497, + 0.00204578903503716, + 0.002144008642062545, + 0.0037545920349657536, + 0.0025335540995001793, + 0.002204917138442397, + 0.0019398059230297804, + 0.002663722727447748, + 0.0008393515599891543, + 0.0011017505312338471, + 0.0007313240203075111, + 0.0013304263120517135, + 0.0007341821328736842, + 0.000887920381501317, + 0.0010323062306270003, + 0.0011915449285879731, + 0.06960444152355194, + 0.3450561463832855, + 0.023821761831641197, + 0.01590668223798275, + 0.04055611044168472, + 0.033572133630514145, + 0.021414287388324738, + 0.2840781509876251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.007781785447150469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001347706071101129 + ], + "opposite_logprob_tokens": [ + -6.626347541809082 + ], + "logprob_reward": -0.001347706071101129, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001347706071101129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.008278803899884224, + 0.030262811109423637, + 0.005015028174966574, + 0.006439967080950737, + 0.015495923347771168, + 0.01198867242783308, + 0.01491983886808157, + 0.07267849892377853, + 0.019427800551056862, + 0.8154926300048828 + ], + "credit_map": [ + 0.0007462604553438723, + 0.0005913747590966523, + 0.0010441630147397518, + 0.0006108379457145929, + 0.0015510283410549164, + 0.0011066944571211934, + 0.0014318389585241675, + 0.0011966060847043991, + 0.0021884075831621885, + 0.0011503242421895266, + 0.009030100889503956, + 0.0014360088389366865, + 0.00344135076738894, + 0.000936270400416106, + 0.009878083132207394, + 0.0022022677585482597, + 0.00034499133471399546, + 0.0005183687899261713, + 0.0008416555356234312, + 0.00044400375918485224, + 0.001004182267934084, + 0.00040486472425982356, + 0.0007336473208852112, + 0.0007233151118271053, + 0.0005369110731408, + 0.0006008149357512593, + 0.000793582119513303, + 0.000885427522007376, + 0.00081959820818156, + 0.0014677859144285321, + 0.0006445955368690193, + 0.0006912521203048527, + 0.0009605875820852816, + 0.0027388555463403463, + 0.0010233353823423386, + 0.0022107691038399935, + 0.002817095024511218, + 0.0032269249204546213, + 0.0015484113246202469, + 0.0009699459769763052, + 0.0012161650229245424, + 0.0012260496150702238, + 0.001342418254353106, + 0.00149300426710397, + 0.0026517114602029324, + 0.0014079707907512784, + 0.0010480321943759918, + 0.0016033207066357136, + 0.002526495372876525, + 0.0015835921512916684, + 0.0020645989570766687, + 0.0016403257614001632, + 0.0029306404758244753, + 0.0016000440809875727, + 0.0010479491902515292, + 0.0015261927619576454, + 0.008299704641103745, + 0.00757629843428731, + 0.00957929901778698, + 0.015538197010755539, + 0.008147332817316055, + 0.006067195441573858, + 0.007028679829090834, + 0.010441794991493225, + 0.0021704297978430986, + 0.003350027371197939, + 0.0015955734997987747, + 0.0032497679349035025, + 0.001539526623673737, + 0.0023861974477767944, + 0.0016748670022934675, + 0.003461412852630019, + 0.06864967197179794, + 0.335893452167511, + 0.024020105600357056, + 0.015983667224645615, + 0.03958828002214432, + 0.03331463038921356, + 0.022104311734437943, + 0.27593857049942017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.001347706071101129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05903688445687294, + 0.0521208681166172, + 0.1130995899438858, + 0.15487709641456604, + 0.10236220061779022, + 0.04882001504302025, + 0.0469888374209404, + 0.035397108644247055, + 0.021362004801630974, + 0.36593538522720337 + ], + "credit_map": [ + 0.005274602677673101, + 0.004693496972322464, + 0.00863100215792656, + 0.00507266353815794, + 0.01090250350534916, + 0.005340175703167915, + 0.012283677235245705, + 0.006838764995336533, + 0.005751056130975485, + 0.004512174520641565, + 0.0058534941636025906, + 0.005886495113372803, + 0.008818669244647026, + 0.004381126258522272, + 0.008785904385149479, + 0.008131951093673706, + 0.00946794357150793, + 0.017147259786725044, + 0.014046596363186836, + 0.014123152941465378, + 0.016400733962655067, + 0.01111807394772768, + 0.012673527002334595, + 0.01812230609357357, + 0.010768212378025055, + 0.013785320334136486, + 0.03171112388372421, + 0.031661972403526306, + 0.012401615269482136, + 0.016626788303256035, + 0.020103363320231438, + 0.017818711698055267, + 0.005879747215658426, + 0.020499717444181442, + 0.007636548951268196, + 0.01322978362441063, + 0.01518460176885128, + 0.02023603953421116, + 0.010402025654911995, + 0.009293745271861553, + 0.00503303948789835, + 0.008151921443641186, + 0.005504587199538946, + 0.00552896223962307, + 0.009753823280334473, + 0.005941981449723244, + 0.0041754054836928844, + 0.0047302995808422565, + 0.007270806934684515, + 0.004971885588020086, + 0.008958140388131142, + 0.005924063269048929, + 0.005115760024636984, + 0.004768709652125835, + 0.004778077360242605, + 0.005201398395001888, + 0.0056403991766273975, + 0.0047542802058160305, + 0.0036701089702546597, + 0.009495381265878677, + 0.0026969134341925383, + 0.001972822705283761, + 0.0028503567446023226, + 0.00431685009971261, + 0.0022141479421406984, + 0.0019782327581197023, + 0.0028100176714360714, + 0.0036185896024107933, + 0.001958196284249425, + 0.0025767628103494644, + 0.0027181566692888737, + 0.003487902693450451, + 0.02980928122997284, + 0.15050223469734192, + 0.010853718034923077, + 0.007862886413931847, + 0.01763419434428215, + 0.014570319093763828, + 0.009805439971387386, + 0.12489735335111618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.004639219027012587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249762907624245 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249762907624245, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249762907624245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006544741336256266, + 0.0181466955691576, + 0.0060768937692046165, + 0.0081553366035223, + 0.06001073867082596, + 0.03953779488801956, + 0.016130810603499413, + 0.013007104396820068, + 0.016565170139074326, + 0.8158248066902161 + ], + "credit_map": [ + 0.0007224934524856508, + 0.0007783410837873816, + 0.0007095156470313668, + 0.0008172856760211289, + 0.0011815980542451143, + 0.0006068538059480488, + 0.0010463633807376027, + 0.0006822898867540061, + 0.0014955158112570643, + 0.0006906311027705669, + 0.00516771525144577, + 0.0008118186378851533, + 0.002120484132319689, + 0.0006032622186467052, + 0.005955059081315994, + 0.0013022085186094046, + 0.00039954634848982096, + 0.0008622623863629997, + 0.0009071974200196564, + 0.0007004460785537958, + 0.0012176361633464694, + 0.0004934461321681738, + 0.0008117633406072855, + 0.0006845958414487541, + 0.0007029303815215826, + 0.0008765837410464883, + 0.0009116364526562393, + 0.000764929223805666, + 0.00121013552416116, + 0.001491759205237031, + 0.0011710517574101686, + 0.0010263100266456604, + 0.004109556321054697, + 0.00901291985064745, + 0.004520985297858715, + 0.008776647970080376, + 0.011797919869422913, + 0.008388232439756393, + 0.007708130404353142, + 0.00569634884595871, + 0.00361074972897768, + 0.005809490568935871, + 0.0027805818244814873, + 0.005055571906268597, + 0.007949410006403923, + 0.006082816980779171, + 0.003769179806113243, + 0.004479989875108004, + 0.002258104272186756, + 0.0020896075293421745, + 0.0016500084893777966, + 0.002129477681592107, + 0.0015122913755476475, + 0.002102752448990941, + 0.002717874012887478, + 0.0016706952592357993, + 0.0019007184309884906, + 0.0018052227096632123, + 0.0012714604381471872, + 0.0016903466312214732, + 0.0021767239086329937, + 0.0014497953234240413, + 0.0012613122817128897, + 0.0014515240909531713, + 0.001978805521503091, + 0.001479386817663908, + 0.0021203020587563515, + 0.001843007979914546, + 0.0015116329304873943, + 0.003371089929714799, + 0.001589863677509129, + 0.002671079942956567, + 0.06839001923799515, + 0.3379201591014862, + 0.023324929177761078, + 0.015384407714009285, + 0.040083177387714386, + 0.03279337286949158, + 0.020547941327095032, + 0.27738073468208313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.005249762907624245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815451705828309 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815451705828309, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815451705828309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14317606389522552, + 0.11422116309404373, + 0.16289089620113373, + 0.149910107254982, + 0.08173352479934692, + 0.1100853905081749, + 0.04029718413949013, + 0.07355879247188568, + 0.04751332476735115, + 0.07661355286836624 + ], + "credit_map": [ + 0.016265928745269775, + 0.011950604617595673, + 0.01767815463244915, + 0.012727849185466766, + 0.025504279881715775, + 0.013841697946190834, + 0.02546035870909691, + 0.01974719949066639, + 0.010589351877570152, + 0.00649974774569273, + 0.031996458768844604, + 0.006798821501433849, + 0.012183986604213715, + 0.00569108547642827, + 0.03132730722427368, + 0.009134404361248016, + 0.02385091222822666, + 0.020952770486474037, + 0.01802944950759411, + 0.012442118488252163, + 0.016930174082517624, + 0.025341499596834183, + 0.012559318915009499, + 0.03278464451432228, + 0.01310322992503643, + 0.010388162918388844, + 0.011557843536138535, + 0.011171153746545315, + 0.014660276472568512, + 0.04740671440958977, + 0.015898171812295914, + 0.02572455070912838, + 0.004418696742504835, + 0.012782737612724304, + 0.005330835934728384, + 0.008717551827430725, + 0.017695534974336624, + 0.01485816203057766, + 0.01065338309854269, + 0.007276622578501701, + 0.010125251486897469, + 0.015237806364893913, + 0.010448926128447056, + 0.014886927790939808, + 0.02248464524745941, + 0.01447472907602787, + 0.011093730106949806, + 0.011333368718624115, + 0.005709705874323845, + 0.00502250250428915, + 0.003796257311478257, + 0.003670901758596301, + 0.0033208285458385944, + 0.0067132520489394665, + 0.006877089850604534, + 0.005186646245419979, + 0.008197722025215626, + 0.00832153856754303, + 0.008585667237639427, + 0.013916505500674248, + 0.011375418864190578, + 0.006243519019335508, + 0.008494586683809757, + 0.008423831313848495, + 0.007254339754581451, + 0.005751254968345165, + 0.006174576934427023, + 0.007251573260873556, + 0.004228228237479925, + 0.00589282950386405, + 0.004713291767984629, + 0.006247229874134064, + 0.009853200986981392, + 0.021905114874243736, + 0.007308628875762224, + 0.005777215119451284, + 0.005855747032910585, + 0.0053934319876134396, + 0.0048055038787424564, + 0.01571470871567726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.002815451705828309 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate camera pans from a slightly low angle to capture a medium close-up of a young Hispanic woman, her face a mask of intense concentration. Her dark, damp hair is slicked back, revealing high cheekbones and a resolute jawline, all partially illuminated by the scene's ambient light. Intricate, glowing cyan cybernetic circuits, like a luminous tattoo, lie embedded beneath the skin of her right arm, shimmering with a soft, cool light, tracing the lines of muscle from forearm to fingertips. She wears a worn black synthetic jacket, its surface shimmering with rainwater beneath the perpetual twilight of a dystopian city. The scene is saturated with the cold, vibrant glow of the towering neon signs lining the rain-slicked alleyway, casting long, distorted reflections of purplish red and electric blue on the grimy, corroded metal wall before her. Steadily and deliberately, she extends her arm and fingers, each tipped with a glowing node of a cybernetic pattern, to grasp a recessed data port in the wall. As her fingertips make precise contact with the port, the cybernetic tattoo flashes brightly for a moment, casting sharp shadows on her focused face and illuminating the fine raindrops on the wall's surface, immersing the entire composition in a gritty dystopian sci-fi style, creating a tense and secretive atmosphere.", + "seed": 16157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_4542_seed_16157_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013914789352566004, + "vlm_logprob": -0.013914789352566004, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.013914789352566004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007784506306052208 + ], + "opposite_logprob_tokens": [ + -5.007784366607666 + ], + "logprob_reward": -0.007784506306052208, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007784506306052208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2489548921585083, + 0.10290469229221344, + 0.06920354068279266, + 0.1487051546573639, + 0.1444055587053299, + 0.08693721890449524, + 0.04848122224211693, + 0.06554318964481354, + 0.023337068036198616, + 0.06152745336294174 + ], + "credit_map": [ + 0.018379220739006996, + 0.0202211644500494, + 0.029472894966602325, + 0.02305002138018608, + 0.048878345638513565, + 0.029791470617055893, + 0.05089614912867546, + 0.028265628963708878, + 0.008836593478918076, + 0.007751965895295143, + 0.011148334480822086, + 0.012058623135089874, + 0.01230113860219717, + 0.015858344733715057, + 0.01468716561794281, + 0.020262526348233223, + 0.009990830905735493, + 0.007367794401943684, + 0.012606996111571789, + 0.010927525348961353, + 0.007873293943703175, + 0.006784843746572733, + 0.008361772634088993, + 0.005290487315505743, + 0.01617743819952011, + 0.014440350234508514, + 0.016469858586788177, + 0.01197835337370634, + 0.03725115582346916, + 0.021264614537358284, + 0.015299012884497643, + 0.01582436077296734, + 0.014774800278246403, + 0.011045947670936584, + 0.017680449411273003, + 0.027234308421611786, + 0.011260781437158585, + 0.018169386312365532, + 0.024653304368257523, + 0.019586576148867607, + 0.011673716828227043, + 0.013916848227381706, + 0.00911269523203373, + 0.007798526901751757, + 0.010975523851811886, + 0.01428314857184887, + 0.006258673034608364, + 0.012918084859848022, + 0.00476114684715867, + 0.005506941117346287, + 0.005063872784376144, + 0.004227970261126757, + 0.006407975219190121, + 0.009477932006120682, + 0.00678227748721838, + 0.006253108847886324, + 0.006004576571285725, + 0.006829084362834692, + 0.007571945898234844, + 0.008100764825940132, + 0.011501472443342209, + 0.006913249380886555, + 0.007953315041959286, + 0.010668780654668808, + 0.002496604574844241, + 0.0019335559336468577, + 0.0038131207693368196, + 0.0021377685479819775, + 0.002790053142234683, + 0.0041435942985117435, + 0.0025306541938334703, + 0.0034917162265628576, + 0.00923848245292902, + 0.008148275315761566, + 0.006452048663049936, + 0.005836411379277706, + 0.005614107009023428, + 0.008749709464609623, + 0.008003388531506062, + 0.009485033340752125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.007784506306052208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266245920211077 + ], + "opposite_logprob_tokens": [ + -5.255266189575195 + ], + "logprob_reward": -0.005266245920211077, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266245920211077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15400569140911102, + 0.08486547321081161, + 0.04847338795661926, + 0.12583880126476288, + 0.06972290575504303, + 0.057405781000852585, + 0.1521725058555603, + 0.16548322141170502, + 0.08349098265171051, + 0.05854128673672676 + ], + "credit_map": [ + 0.018200276419520378, + 0.013100501149892807, + 0.02367679961025715, + 0.013858415186405182, + 0.02217351831495762, + 0.01817839965224266, + 0.024768508970737457, + 0.020049268379807472, + 0.011569551192224026, + 0.005533176939934492, + 0.01146203838288784, + 0.008662175387144089, + 0.00947240088135004, + 0.012019585818052292, + 0.009770910255610943, + 0.01637563481926918, + 0.007699368055909872, + 0.005093513056635857, + 0.009309167042374611, + 0.005955518689006567, + 0.005905674304813147, + 0.004492552485316992, + 0.005947049241513014, + 0.0040705460123717785, + 0.01445268839597702, + 0.01172536052763462, + 0.01976022869348526, + 0.012136179953813553, + 0.019704507663846016, + 0.021025072783231735, + 0.01099624764174223, + 0.016038523986935616, + 0.008784707635641098, + 0.004656283184885979, + 0.006804789416491985, + 0.0049435412511229515, + 0.005223836749792099, + 0.017230462282896042, + 0.006128240376710892, + 0.015951044857501984, + 0.007438787259161472, + 0.006413394119590521, + 0.0071038114838302135, + 0.0043521844781935215, + 0.008322927169501781, + 0.010322179645299911, + 0.005624060984700918, + 0.00782843492925167, + 0.017698906362056732, + 0.02103019505739212, + 0.015842627733945847, + 0.012059762142598629, + 0.013433905318379402, + 0.026295043528079987, + 0.019250374287366867, + 0.026561686769127846, + 0.012567345052957535, + 0.015807928517460823, + 0.012096996419131756, + 0.0217036884278059, + 0.03186327591538429, + 0.0182028878480196, + 0.025437986478209496, + 0.02780311554670334, + 0.01133785117417574, + 0.007044391706585884, + 0.00965541135519743, + 0.009184485301375389, + 0.0071435896679759026, + 0.015200139954686165, + 0.012014762498438358, + 0.011910350993275642, + 0.005844608880579472, + 0.004523852374404669, + 0.00470971642062068, + 0.005657994654029608, + 0.005752468015998602, + 0.010859237983822823, + 0.01073706429451704, + 0.010456346906721592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.005266245920211077 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196710661053658 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014196710661053658, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196710661053658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24022679030895233, + 0.08782216161489487, + 0.0776946023106575, + 0.16046185791492462, + 0.12674349546432495, + 0.11881530284881592, + 0.05549202114343643, + 0.06410441547632217, + 0.04120657220482826, + 0.027432838454842567 + ], + "credit_map": [ + 0.024801211431622505, + 0.021162457764148712, + 0.035975344479084015, + 0.02441653609275818, + 0.03301744908094406, + 0.03057035617530346, + 0.039883535355329514, + 0.03039988875389099, + 0.009227055124938488, + 0.005417374894022942, + 0.00957831833511591, + 0.009220577776432037, + 0.010407986119389534, + 0.012757590971887112, + 0.013033321127295494, + 0.01817992702126503, + 0.005861036479473114, + 0.007110121659934521, + 0.009296603500843048, + 0.012074713595211506, + 0.013347933068871498, + 0.007061748765408993, + 0.015346340835094452, + 0.007596092764288187, + 0.008905040100216866, + 0.01799360103905201, + 0.013693055137991905, + 0.029934844002127647, + 0.03254687413573265, + 0.013782832771539688, + 0.02723868191242218, + 0.016366906464099884, + 0.011895043775439262, + 0.012216699309647083, + 0.014437499456107616, + 0.020506860688328743, + 0.016265107318758965, + 0.018280722200870514, + 0.01669379137456417, + 0.01644776575267315, + 0.015418235212564468, + 0.015513724647462368, + 0.012796268798410892, + 0.009136374108493328, + 0.014840193092823029, + 0.021777816116809845, + 0.013705200515687466, + 0.015627488493919373, + 0.00521264411509037, + 0.00882065761834383, + 0.004032598342746496, + 0.0034795436076819897, + 0.005022179801017046, + 0.010035267099738121, + 0.00813657883554697, + 0.010752546600997448, + 0.006558711640536785, + 0.006158328615128994, + 0.0077183921821415424, + 0.007077230606228113, + 0.011536476202309132, + 0.0071620820090174675, + 0.009232323616743088, + 0.008660866878926754, + 0.0036464377772063017, + 0.0027186921797692776, + 0.005628740414977074, + 0.0037964798975735903, + 0.004630548879504204, + 0.006881672888994217, + 0.007549295201897621, + 0.006354702170938253, + 0.004368866328150034, + 0.004218976013362408, + 0.001735091907903552, + 0.0021234480664134026, + 0.0030015555676072836, + 0.0028441448230296373, + 0.004313122946768999, + 0.0048276325687766075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.014196710661053658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369210660457611 + ], + "opposite_logprob_tokens": [ + -3.4086921215057373 + ], + "logprob_reward": -0.03369210660457611, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369210660457611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1566164642572403, + 0.060434550046920776, + 0.01918601430952549, + 0.22016401588916779, + 0.0861603394150734, + 0.10465949028730392, + 0.09313076734542847, + 0.11036400496959686, + 0.07267339527606964, + 0.0766109749674797 + ], + "credit_map": [ + 0.01952640525996685, + 0.01389809139072895, + 0.025837786495685577, + 0.014884128235280514, + 0.019384384155273438, + 0.019770577549934387, + 0.02364232763648033, + 0.01967277005314827, + 0.006741734221577644, + 0.004013011697679758, + 0.008132543414831161, + 0.0057147652842104435, + 0.006780194118618965, + 0.009861301630735397, + 0.006989174988120794, + 0.012201827950775623, + 0.0016623296542093158, + 0.0020965971052646637, + 0.0023253310937434435, + 0.0027571669779717922, + 0.0033856723457574844, + 0.0021526534110307693, + 0.0027920145075768232, + 0.0020142493303865194, + 0.026174649596214294, + 0.022856920957565308, + 0.03592254966497421, + 0.018575847148895264, + 0.03375827521085739, + 0.037777308374643326, + 0.017537688836455345, + 0.027560777962207794, + 0.011158671230077744, + 0.0073958346620202065, + 0.009628108702600002, + 0.01335124485194683, + 0.008519959636032581, + 0.009144900366663933, + 0.01568276435136795, + 0.011278852820396423, + 0.010751960799098015, + 0.02097463607788086, + 0.009231128729879856, + 0.010205126367509365, + 0.011842495761811733, + 0.015653913840651512, + 0.010930289514362812, + 0.015069951303303242, + 0.007376840338110924, + 0.011153592728078365, + 0.009743090718984604, + 0.00843824166804552, + 0.011589937843382359, + 0.015613366849720478, + 0.013191811740398407, + 0.016023879870772362, + 0.008525921031832695, + 0.009334877133369446, + 0.017577901482582092, + 0.011957861483097076, + 0.018425041809678078, + 0.012929212301969528, + 0.01427548285573721, + 0.017337707802653313, + 0.014295995235443115, + 0.01042132917791605, + 0.009563268162310123, + 0.00513219740241766, + 0.006294950842857361, + 0.007436669897288084, + 0.009692075662314892, + 0.009836912155151367, + 0.008111868053674698, + 0.007012825924903154, + 0.007402445655316114, + 0.006969762500375509, + 0.007604256272315979, + 0.009149660356342793, + 0.015776917338371277, + 0.01458323560655117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.03369210660457611 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634377270936966 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634377270936966, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634377270936966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2507193982601166, + 0.18081134557724, + 0.02498149685561657, + 0.12506496906280518, + 0.145595520734787, + 0.0964694395661354, + 0.04489435628056526, + 0.04175831004977226, + 0.04560449346899986, + 0.044100649654865265 + ], + "credit_map": [ + 0.028723690658807755, + 0.0236037690192461, + 0.04287005215883255, + 0.024570805951952934, + 0.03003309853374958, + 0.03171771764755249, + 0.03442006930708885, + 0.03478020802140236, + 0.021412059664726257, + 0.01988351158797741, + 0.024796146899461746, + 0.015705423429608345, + 0.022404445335268974, + 0.030351513996720314, + 0.020067423582077026, + 0.026190834119915962, + 0.001984599744901061, + 0.0028155636973679066, + 0.0036404933780431747, + 0.003424424910917878, + 0.004636271856725216, + 0.002504387404769659, + 0.0033376843202859163, + 0.0026380743365734816, + 0.013819390907883644, + 0.014301286078989506, + 0.014963478781282902, + 0.010812424123287201, + 0.023240793496370316, + 0.01865650899708271, + 0.012447469867765903, + 0.016823628917336464, + 0.016786964610219002, + 0.010273718275129795, + 0.017066171392798424, + 0.023116640746593475, + 0.010811320506036282, + 0.020305922254920006, + 0.02193382754921913, + 0.02530096471309662, + 0.011348024941980839, + 0.015905259177088737, + 0.008625064045190811, + 0.007024882361292839, + 0.01123046688735485, + 0.019001515582203865, + 0.010118188336491585, + 0.013216051273047924, + 0.0057742237113416195, + 0.005920874886214733, + 0.0033020488917827606, + 0.002894761273637414, + 0.006083682179450989, + 0.007199387066066265, + 0.006296835839748383, + 0.007422541733831167, + 0.004864486865699291, + 0.006101268343627453, + 0.0036024460569024086, + 0.004751730244606733, + 0.005307513754814863, + 0.004221531096845865, + 0.00723291514441371, + 0.005676422268152237, + 0.009005493484437466, + 0.004903154913336039, + 0.004346443340182304, + 0.004074146505445242, + 0.0036178971640765667, + 0.004175773821771145, + 0.008802404627203941, + 0.006679183803498745, + 0.006398557219654322, + 0.005154919810593128, + 0.0040259514935314655, + 0.003831615438684821, + 0.005256024654954672, + 0.006980231963098049, + 0.005344979930669069, + 0.007108369376510382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.008634377270936966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate camera pans from a slightly low angle to capture a medium close-up of a young Hispanic woman, her face a mask of intense concentration. Her dark, damp hair is slicked back, revealing high cheekbones and a resolute jawline, all partially illuminated by the scene's ambient light. Intricate, glowing cyan cybernetic circuits, like a luminous tattoo, lie embedded beneath the skin of her right arm, shimmering with a soft, cool light, tracing the lines of muscle from forearm to fingertips. She wears a worn black synthetic jacket, its surface shimmering with rainwater beneath the perpetual twilight of a dystopian city. The scene is saturated with the cold, vibrant glow of the towering neon signs lining the rain-slicked alleyway, casting long, distorted reflections of purplish red and electric blue on the grimy, corroded metal wall before her. Steadily and deliberately, she extends her arm and fingers, each tipped with a glowing node of a cybernetic pattern, to grasp a recessed data port in the wall. As her fingertips make precise contact with the port, the cybernetic tattoo flashes brightly for a moment, casting sharp shadows on her focused face and illuminating the fine raindrops on the wall's surface, immersing the entire composition in a gritty dystopian sci-fi style, creating a tense and secretive atmosphere.", + "seed": 16158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_1/index_4542_seed_16158_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01330040036700666, + "vlm_logprob": -0.01330040036700666, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.01330040036700666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030432671308517456 + ], + "opposite_logprob_tokens": [ + -3.53043270111084 + ], + "logprob_reward": -0.030432671308517456, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.030432671308517456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0843179002404213, + 0.3321595788002014, + 0.1022612676024437, + 0.1510585993528366, + 0.062089305371046066, + 0.05372331663966179, + 0.04647299647331238, + 0.025650229305028915, + 0.09011764079332352, + 0.05214918777346611 + ], + "credit_map": [ + 0.00792282447218895, + 0.009767946787178516, + 0.006470124702900648, + 0.010599455796182156, + 0.014014476910233498, + 0.016706785187125206, + 0.010897972621023655, + 0.00793831143528223, + 0.02197927050292492, + 0.05976758897304535, + 0.02933894470334053, + 0.02887503243982792, + 0.016580063849687576, + 0.1121901273727417, + 0.014821061864495277, + 0.04860749840736389, + 0.007409784477204084, + 0.016105525195598602, + 0.007243300788104534, + 0.021820832043886185, + 0.015208694152534008, + 0.010077569633722305, + 0.012749250046908855, + 0.011646308936178684, + 0.006829178426414728, + 0.009161136113107204, + 0.010193669237196445, + 0.011034351773560047, + 0.041567400097846985, + 0.02008521370589733, + 0.03986886516213417, + 0.01231878250837326, + 0.00849137082695961, + 0.005239163525402546, + 0.007040555123239756, + 0.006197900976985693, + 0.011379624716937542, + 0.0063825505785644054, + 0.009016349911689758, + 0.00834178552031517, + 0.004895069170743227, + 0.009157313965260983, + 0.011944224126636982, + 0.0052979933097958565, + 0.005682770628482103, + 0.005964722950011492, + 0.006470746360719204, + 0.004310475662350655, + 0.0034592864103615284, + 0.006343260873109102, + 0.007506159134209156, + 0.004875400103628635, + 0.004607982002198696, + 0.008880696259438992, + 0.0045796288177371025, + 0.006220580544322729, + 0.0019190621096640825, + 0.0033498676493763924, + 0.0038807678502053022, + 0.002283194102346897, + 0.00380380405113101, + 0.004829812794923782, + 0.0026099798269569874, + 0.0029737423174083233, + 0.00662203086540103, + 0.006746211089193821, + 0.013722659088671207, + 0.005462888162583113, + 0.01960894651710987, + 0.015054735355079174, + 0.010889691300690174, + 0.012010481208562851, + 0.004343001637607813, + 0.004840422421693802, + 0.005315222777426243, + 0.006851199083030224, + 0.007633806671947241, + 0.00775620061904192, + 0.00807436928153038, + 0.00733496667817235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.030432671308517456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651796538382769 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004651796538382769, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651796538382769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06211958825588226, + 0.16175511479377747, + 0.1951666623353958, + 0.20333537459373474, + 0.07732157409191132, + 0.09012192487716675, + 0.040585629642009735, + 0.038383848965168, + 0.09769361466169357, + 0.03351667895913124 + ], + "credit_map": [ + 0.005384813994169235, + 0.00776507006958127, + 0.006604594178497791, + 0.010329648852348328, + 0.008354784920811653, + 0.009077227674424648, + 0.006975560449063778, + 0.007627883926033974, + 0.009728922508656979, + 0.026627397164702415, + 0.013078037649393082, + 0.015215441584587097, + 0.0076309810392558575, + 0.05468127131462097, + 0.009198833256959915, + 0.025594230741262436, + 0.02306360937654972, + 0.031443845480680466, + 0.016915859654545784, + 0.04251154512166977, + 0.02721364051103592, + 0.01921846717596054, + 0.01638115383684635, + 0.01841854117810726, + 0.012318368069827557, + 0.012317083775997162, + 0.014455880969762802, + 0.01780121400952339, + 0.05204480141401291, + 0.027395982295274734, + 0.04940803349018097, + 0.01759401150047779, + 0.009592408314347267, + 0.006697101518511772, + 0.010379632003605366, + 0.00946606881916523, + 0.011711154133081436, + 0.009407399222254753, + 0.010141773149371147, + 0.009926041588187218, + 0.006730410736054182, + 0.020313721150159836, + 0.02194806933403015, + 0.0077933273278176785, + 0.00848102755844593, + 0.00856765080243349, + 0.00988987646996975, + 0.006397835910320282, + 0.00404036371037364, + 0.0032500128727406263, + 0.008242463693022728, + 0.003192421980202198, + 0.004167261999100447, + 0.008018294349312782, + 0.00442815525457263, + 0.005246657412499189, + 0.002667817519977689, + 0.0036448624450713396, + 0.005039238836616278, + 0.00311252661049366, + 0.006329767405986786, + 0.009135340340435505, + 0.003575848415493965, + 0.00487844692543149, + 0.006633783224970102, + 0.00842366088181734, + 0.01295718178153038, + 0.006370292976498604, + 0.017781415954232216, + 0.011393525637686253, + 0.01960975117981434, + 0.01452400628477335, + 0.004440501797944307, + 0.006229723803699017, + 0.004359803628176451, + 0.0036411392502486706, + 0.0037674640771001577, + 0.0031256263609975576, + 0.00249853846617043, + 0.0054538799449801445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.004651796538382769 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540061958134174 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012540061958134174, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540061958134174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11183248460292816, + 0.042010772973299026, + 0.08250418305397034, + 0.3033926784992218, + 0.06535157561302185, + 0.11969836056232452, + 0.11399172246456146, + 0.022717997431755066, + 0.06594273447990417, + 0.0725574940443039 + ], + "credit_map": [ + 0.00987129844725132, + 0.015363817103207111, + 0.01059204526245594, + 0.013418757356703281, + 0.019098196178674698, + 0.014843693934381008, + 0.014433471485972404, + 0.014211216010153294, + 0.0032036437187343836, + 0.0040475851856172085, + 0.003177470760419965, + 0.003135769395157695, + 0.002962736878544092, + 0.01372566819190979, + 0.0035189264453947544, + 0.008238972164690495, + 0.006536086089909077, + 0.013470600359141827, + 0.00612120283767581, + 0.016360852867364883, + 0.011692608706653118, + 0.008023956790566444, + 0.010698183439671993, + 0.009600694291293621, + 0.018953781574964523, + 0.017166730016469955, + 0.019986461848020554, + 0.026985229924321175, + 0.07706597447395325, + 0.040036290884017944, + 0.07686284184455872, + 0.026335345581173897, + 0.007727649062871933, + 0.007204218301922083, + 0.009061096236109734, + 0.00782569870352745, + 0.012150618247687817, + 0.006503095850348473, + 0.009132280945777893, + 0.0057469140738248825, + 0.007975593209266663, + 0.03161674365401268, + 0.030598385259509087, + 0.006164655555039644, + 0.014227189123630524, + 0.009323665872216225, + 0.01390495803207159, + 0.005887177772819996, + 0.009917590767145157, + 0.012249230407178402, + 0.027721432968974113, + 0.016451895236968994, + 0.007949522696435452, + 0.00959067139774561, + 0.010920286178588867, + 0.019191086292266846, + 0.00178020098246634, + 0.003878785064443946, + 0.0031806204933673143, + 0.002826203592121601, + 0.002524606417864561, + 0.002449325518682599, + 0.002878293627873063, + 0.003199961269274354, + 0.00491848262026906, + 0.006977551616728306, + 0.008441582322120667, + 0.005809267517179251, + 0.011122407391667366, + 0.007778827100992203, + 0.011624956503510475, + 0.009269656613469124, + 0.006862454582005739, + 0.0074894665740430355, + 0.007479654159396887, + 0.012503785081207752, + 0.011610598303377628, + 0.007976191118359566, + 0.0094617223367095, + 0.009173624217510223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.012540061958134174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062971204519272 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.016062971204519272, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062971204519272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08127270638942719, + 0.08054187893867493, + 0.04876253008842468, + 0.28885841369628906, + 0.0483100526034832, + 0.26551884412765503, + 0.04272807762026787, + 0.05950538069009781, + 0.057180583477020264, + 0.027321621775627136 + ], + "credit_map": [ + 0.007765031885355711, + 0.010890944860875607, + 0.005275651812553406, + 0.009679227136075497, + 0.015187745913863182, + 0.010233679786324501, + 0.011890497989952564, + 0.010349925607442856, + 0.006099861580878496, + 0.013026189990341663, + 0.006091471295803785, + 0.006246375385671854, + 0.00442890590056777, + 0.01744917407631874, + 0.007387116551399231, + 0.019812781363725662, + 0.004484031815081835, + 0.005660893861204386, + 0.0035283772740513086, + 0.0064011262729763985, + 0.007480969652533531, + 0.0061528319492936134, + 0.00856232363730669, + 0.0064919739961624146, + 0.017403515055775642, + 0.016662200912833214, + 0.01896362006664276, + 0.02640116587281227, + 0.07250584661960602, + 0.037430111318826675, + 0.0753212422132492, + 0.02417067438364029, + 0.006478135474026203, + 0.005820342805236578, + 0.005873676389455795, + 0.00914583820849657, + 0.005826298613101244, + 0.0033476774115115404, + 0.004465841222554445, + 0.007352241314947605, + 0.015534122474491596, + 0.07077544182538986, + 0.07056420296430588, + 0.013136615045368671, + 0.031667180359363556, + 0.019856417551636696, + 0.031149085611104965, + 0.012835736386477947, + 0.003154357895255089, + 0.0042536500841379166, + 0.0033142834436148405, + 0.0035866182297468185, + 0.006051677744835615, + 0.011363937519490719, + 0.003639423521235585, + 0.007364124525338411, + 0.004341991618275642, + 0.006080263759940863, + 0.00771491089835763, + 0.005040355492383242, + 0.011770818382501602, + 0.011256285943090916, + 0.005791641306132078, + 0.007509106304496527, + 0.004905423615127802, + 0.005556762218475342, + 0.00693103251978755, + 0.0060214996337890625, + 0.00985951628535986, + 0.008571610786020756, + 0.007949749007821083, + 0.007384986151009798, + 0.002455912297591567, + 0.003997807390987873, + 0.0030224714428186417, + 0.004045331850647926, + 0.003909536171704531, + 0.0034088853280991316, + 0.0023858940694481134, + 0.00409577926620841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.016062971204519272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Hispanic woman with dark, damp hair slicked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate, glowing cyan cybernetic circuits visible embedded beneath the skin of her right arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a worn black synthetic jacket that appears wet with rainwater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman extend her arm to grasp a recessed data port in a grimy, corroded metal wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a rain-slicked alleyway with towering neon signs casting purplish red and electric blue reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028145008254796267 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.0028145008254796267, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028145008254796267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13441255688667297, + 0.06748668849468231, + 0.06882601231336594, + 0.13203631341457367, + 0.15345582365989685, + 0.12673096358776093, + 0.040820568799972534, + 0.03840159997344017, + 0.11861934512853622, + 0.1192101463675499 + ], + "credit_map": [ + 0.016007423400878906, + 0.01521751657128334, + 0.011747478507459164, + 0.01650063879787922, + 0.02319176122546196, + 0.016008276492357254, + 0.020190270617604256, + 0.015549181029200554, + 0.006277396809309721, + 0.011888708919286728, + 0.00490310974419117, + 0.006553310435265303, + 0.0055346619337797165, + 0.01691655069589615, + 0.0051022921688854694, + 0.01031065359711647, + 0.0061395540833473206, + 0.010218051262199879, + 0.005607379134744406, + 0.01150420680642128, + 0.007263224106281996, + 0.009054106660187244, + 0.009832537733018398, + 0.009206945076584816, + 0.01074754074215889, + 0.014554666355252266, + 0.014333125203847885, + 0.015473948791623116, + 0.024876626208424568, + 0.01383927371352911, + 0.022777218371629715, + 0.015433910302817822, + 0.02130981720983982, + 0.017305532470345497, + 0.017905516549944878, + 0.01960054785013199, + 0.028588244691491127, + 0.01579936593770981, + 0.018217114731669426, + 0.01472965907305479, + 0.010893224738538265, + 0.020171180367469788, + 0.01660442352294922, + 0.016233833506703377, + 0.020229190587997437, + 0.014018705114722252, + 0.016232620924711227, + 0.012347775511443615, + 0.003078347770497203, + 0.003779354505240917, + 0.0056968508288264275, + 0.0035560261458158493, + 0.00450667180120945, + 0.011332386173307896, + 0.003495557000860572, + 0.005375374108552933, + 0.003787852358072996, + 0.004802551120519638, + 0.006103235296905041, + 0.004153763875365257, + 0.005164189264178276, + 0.006172123830765486, + 0.004303013905882835, + 0.003914868459105492, + 0.009897956624627113, + 0.01439109817147255, + 0.013947047293186188, + 0.00945931114256382, + 0.01976575329899788, + 0.01603018492460251, + 0.019759833812713623, + 0.015368149615824223, + 0.012585362419486046, + 0.020957008004188538, + 0.01203985046595335, + 0.01476698275655508, + 0.012339969165623188, + 0.016997236758470535, + 0.010710766538977623, + 0.018812961876392365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0028145008254796267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_tensor([8994, 8994])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_tensor([8994, 8994])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bdd378deecb8d46940bc39814c09bc13588b0bb9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_tensor([8994, 8994])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level perspective shot from inside a car captures the experience of driving in a pouring downpour at night. The windshield is almost completely obscured by a thick layer of rain, countless streams distorting the outside world into an abstract blur of water. Black rubber windshield wiper blades enter the frame, rhythmically sweeping across the glass, methodically pushing the water away and creating a momentary arc of perfect clarity. In this fleeting moment, the smooth, wet asphalt of the road ahead is revealed, gleaming in the bright cones of the car's headlights. Light reflects off the surface of the stagnant water, creating long, shimmering streaks that guide the eye forward. As the wipers complete their motion, rain instantly re-covers the glass, obscuring the view once again, establishing a hypnotic and relentless cycle that creates a moody, cocoon-like atmosphere of focused focus as one navigates the storm.", + "seed": 31945, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_8994_seed_31945_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05784537177532911, + "vlm_logprob": -0.05784537177532911, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.05784537177532911, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08764631301164627 + ], + "opposite_logprob_tokens": [ + -3.712646245956421 + ], + "logprob_reward": -0.08764631301164627, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.08764631301164627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1762012541294098, + 0.0976860374212265, + 0.16025947034358978, + 0.048440638929605484, + 0.047742947936058044, + 0.03693819418549538, + 0.1754838526248932, + 0.14840541779994965, + 0.0727391391992569, + 0.0361030250787735 + ], + "credit_map": [ + 0.015265795402228832, + 0.017246883362531662, + 0.022575311362743378, + 0.013639572076499462, + 0.03110087662935257, + 0.02825654111802578, + 0.027559304609894753, + 0.020556973293423653, + 0.010206952691078186, + 0.015380334109067917, + 0.015949252992868423, + 0.009216278791427612, + 0.014672376215457916, + 0.008850538171827793, + 0.012074140831828117, + 0.011336157098412514, + 0.01966285891830921, + 0.02442205511033535, + 0.023789362981915474, + 0.01341583114117384, + 0.013663695193827152, + 0.020770041272044182, + 0.012407503090798855, + 0.03212812542915344, + 0.005078513640910387, + 0.0038296892307698727, + 0.011633596383035183, + 0.0033163561020046473, + 0.0032625210005789995, + 0.006791795138269663, + 0.004887801129370928, + 0.009640367701649666, + 0.004533540923148394, + 0.006319080945104361, + 0.005366175901144743, + 0.003926203120499849, + 0.006302816327661276, + 0.0058050318621098995, + 0.009406093508005142, + 0.00608400534838438, + 0.005601011216640472, + 0.005339930299669504, + 0.004408581648021936, + 0.0024322974495589733, + 0.0051671466790139675, + 0.004650981165468693, + 0.002954797586426139, + 0.006383451633155346, + 0.014013675972819328, + 0.02384185418486595, + 0.011119954288005829, + 0.012222202494740486, + 0.0327213779091835, + 0.011530173942446709, + 0.049250148236751556, + 0.02078445442020893, + 0.04097072780132294, + 0.008447243832051754, + 0.00543561764061451, + 0.004519500769674778, + 0.05866505578160286, + 0.012788360938429832, + 0.0052407304756343365, + 0.012338177300989628, + 0.014672410674393177, + 0.010903329588472843, + 0.006138295866549015, + 0.00511456374078989, + 0.008888700976967812, + 0.010035501793026924, + 0.009000658057630062, + 0.007985683158040047, + 0.005851319991052151, + 0.004500955808907747, + 0.0019506164826452732, + 0.00260272272862494, + 0.006260851863771677, + 0.006223270203918219, + 0.005142016801983118, + 0.003571269568055868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.08764631301164627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07012884318828583 + ], + "opposite_logprob_tokens": [ + -2.695128917694092 + ], + "logprob_reward": -0.07012884318828583, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07012884318828583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14091888070106506, + 0.22063857316970825, + 0.10391721874475479, + 0.07273662835359573, + 0.07307456433773041, + 0.043487194925546646, + 0.08554429560899734, + 0.11854929476976395, + 0.043470244854688644, + 0.09766309708356857 + ], + "credit_map": [ + 0.01231719832867384, + 0.01649700105190277, + 0.01514708437025547, + 0.010003351606428623, + 0.026267794892191887, + 0.014386294409632683, + 0.026711443439126015, + 0.019588712602853775, + 0.028630295768380165, + 0.03726040944457054, + 0.033696942031383514, + 0.012793252244591713, + 0.03715330362319946, + 0.02263619191944599, + 0.024793554097414017, + 0.02367461286485195, + 0.013898405246436596, + 0.010645786300301552, + 0.014586547389626503, + 0.010743524879217148, + 0.011388243176043034, + 0.01258047018200159, + 0.009543577209115028, + 0.020530667155981064, + 0.007989690639078617, + 0.009249168448150158, + 0.011746933683753014, + 0.007628277875483036, + 0.009122882969677448, + 0.009354943409562111, + 0.00790228508412838, + 0.009742447175085545, + 0.0064508672803640366, + 0.010203500278294086, + 0.01133518386632204, + 0.007674395572394133, + 0.006955355405807495, + 0.009794591926038265, + 0.01039400789886713, + 0.01026665698736906, + 0.004498325753957033, + 0.003834884613752365, + 0.005711684934794903, + 0.0036090260837227106, + 0.008355741389095783, + 0.006323718000203371, + 0.0054193660616874695, + 0.005734448786824942, + 0.012751064263284206, + 0.010130125097930431, + 0.006749632768332958, + 0.005997336935251951, + 0.017000913619995117, + 0.00820143148303032, + 0.01707368716597557, + 0.007640103343874216, + 0.031142011284828186, + 0.0067048100754618645, + 0.0047025009989738464, + 0.003112392034381628, + 0.04641697183251381, + 0.010447688400745392, + 0.004784652031958103, + 0.01123827788978815, + 0.007433949038386345, + 0.005245215725153685, + 0.004604320507496595, + 0.005795806646347046, + 0.005495272111147642, + 0.004903956316411495, + 0.004099549259990454, + 0.0058921752497553825, + 0.011904770508408546, + 0.023889727890491486, + 0.006567907053977251, + 0.007152804173529148, + 0.014440855011343956, + 0.01620972901582718, + 0.007715597283095121, + 0.009781711734831333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.07012884318828583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372357040643692 + ], + "opposite_logprob_tokens": [ + -3.4087235927581787 + ], + "logprob_reward": -0.03372357040643692, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372357040643692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08812493830919266, + 0.1764882206916809, + 0.0788971409201622, + 0.1466691642999649, + 0.05221506208181381, + 0.030072331428527832, + 0.15089452266693115, + 0.11001978069543839, + 0.10970175266265869, + 0.056917157024145126 + ], + "credit_map": [ + 0.012818221934139729, + 0.008778965100646019, + 0.0129316421225667, + 0.006515294313430786, + 0.013844567351043224, + 0.01155194640159607, + 0.011015966534614563, + 0.010668342001736164, + 0.018730297684669495, + 0.015698157250881195, + 0.031229186803102493, + 0.01721199043095112, + 0.021514927968382835, + 0.03164078667759895, + 0.020795265212655067, + 0.019667601212859154, + 0.008798259310424328, + 0.011672046966850758, + 0.011402697302401066, + 0.0077784135937690735, + 0.011717814952135086, + 0.008838550187647343, + 0.010451973415911198, + 0.008237380534410477, + 0.012829950079321861, + 0.013154109008610249, + 0.03367858752608299, + 0.0098049845546484, + 0.017330944538116455, + 0.013949112966656685, + 0.020469360053539276, + 0.02545212022960186, + 0.006243168376386166, + 0.0035359372850507498, + 0.008410485461354256, + 0.003942250739783049, + 0.004902546294033527, + 0.009243499487638474, + 0.007910608313977718, + 0.008026559837162495, + 0.00269739399664104, + 0.0025702936109155416, + 0.0038270154036581516, + 0.0034948980901390314, + 0.005933517124503851, + 0.0032738728914409876, + 0.004182660486549139, + 0.004092678893357515, + 0.021664543077349663, + 0.022848054766654968, + 0.012805224396288395, + 0.008341275155544281, + 0.04058082401752472, + 0.015653518959879875, + 0.015260704793035984, + 0.013740377500653267, + 0.0326763391494751, + 0.006092929746955633, + 0.004121389705687761, + 0.003318910486996174, + 0.04176495224237442, + 0.009528722614049911, + 0.006334617733955383, + 0.006181914359331131, + 0.016292551532387733, + 0.017309367656707764, + 0.010693032294511795, + 0.007463195361196995, + 0.01952766627073288, + 0.010245648212730885, + 0.01787250116467476, + 0.010297791101038456, + 0.007666047662496567, + 0.008176066912710667, + 0.004820061847567558, + 0.004509368911385536, + 0.010040422901511192, + 0.00829987321048975, + 0.006521780043840408, + 0.0068835364654660225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.03372357040643692 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907697349786758 + ], + "opposite_logprob_tokens": [ + -2.4640769958496094 + ], + "logprob_reward": -0.08907697349786758, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907697349786758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1509075164794922, + 0.15071897208690643, + 0.061652034521102905, + 0.1578471064567566, + 0.05291545391082764, + 0.03217816725373268, + 0.16568943858146667, + 0.08434578776359558, + 0.10731405019760132, + 0.03643139451742172 + ], + "credit_map": [ + 0.011923711746931076, + 0.018553094938397408, + 0.012210430577397346, + 0.009655578061938286, + 0.035096243023872375, + 0.01470806635916233, + 0.028112521395087242, + 0.02064787596464157, + 0.023590268567204475, + 0.0256633423268795, + 0.02319181151688099, + 0.010682005435228348, + 0.021457869559526443, + 0.013822280801832676, + 0.015497728250920773, + 0.01681368611752987, + 0.005219035781919956, + 0.007613686844706535, + 0.009120973758399487, + 0.004693415015935898, + 0.011815119534730911, + 0.006919876206666231, + 0.008047066628932953, + 0.008222866803407669, + 0.015097838826477528, + 0.01415302511304617, + 0.037578385323286057, + 0.015138198621571064, + 0.011845503933727741, + 0.016597477719187737, + 0.017260950058698654, + 0.030175743624567986, + 0.007971889339387417, + 0.006555226631462574, + 0.00789458118379116, + 0.005437532439827919, + 0.006920801941305399, + 0.0059901038184762, + 0.006080782040953636, + 0.006064537446945906, + 0.0035022760275751352, + 0.0025208157021552324, + 0.004802883602678776, + 0.003971614874899387, + 0.0033569978550076485, + 0.004046146292239428, + 0.00272387801669538, + 0.007253559306263924, + 0.01953577622771263, + 0.01686331443488598, + 0.013101410120725632, + 0.01089058630168438, + 0.020617086440324783, + 0.015203120186924934, + 0.050293780863285065, + 0.019184377044439316, + 0.01654808782041073, + 0.005625136662274599, + 0.0041404771618545055, + 0.0028438218869268894, + 0.02566279098391533, + 0.015064462088048458, + 0.004902083892375231, + 0.009558933787047863, + 0.021244551986455917, + 0.02171248570084572, + 0.016018372029066086, + 0.013812495395541191, + 0.008349795825779438, + 0.010023018345236778, + 0.007228349335491657, + 0.008924988098442554, + 0.0037272891495376825, + 0.004415800329297781, + 0.0024796503130346537, + 0.0029495637863874435, + 0.008555680513381958, + 0.004186014644801617, + 0.006416663993149996, + 0.0037007357459515333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.08907697349786758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651158772408962 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651158772408962, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651158772408962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1992974728345871, + 0.11760921031236649, + 0.10872414708137512, + 0.02965155802667141, + 0.08720508217811584, + 0.03212948143482208, + 0.19964714348316193, + 0.07664718478918076, + 0.09004826843738556, + 0.05904044210910797 + ], + "credit_map": [ + 0.022155027836561203, + 0.018080832436680794, + 0.017494238913059235, + 0.01418580673635006, + 0.05191555246710777, + 0.0131772980093956, + 0.03731564059853554, + 0.024973079562187195, + 0.012150513008236885, + 0.010753012262284756, + 0.02111969329416752, + 0.013584530912339687, + 0.01450219564139843, + 0.01546939555555582, + 0.01457990612834692, + 0.015449963510036469, + 0.012796632945537567, + 0.013675067573785782, + 0.014318958856165409, + 0.011419416405260563, + 0.01327596791088581, + 0.012870817445218563, + 0.013398830778896809, + 0.016968458890914917, + 0.003263598307967186, + 0.0029996514786034822, + 0.004125221166759729, + 0.003193323966115713, + 0.0031607174314558506, + 0.004241290967911482, + 0.004871297162026167, + 0.0037964589428156614, + 0.009903447702527046, + 0.01082725077867508, + 0.014489565044641495, + 0.007534730713814497, + 0.011894452385604382, + 0.010927572846412659, + 0.010104387067258358, + 0.011523673310875893, + 0.003346079960465431, + 0.0032515411730855703, + 0.003846593899652362, + 0.0032378616742789745, + 0.004918958060443401, + 0.003493055235594511, + 0.004775758367031813, + 0.005259633995592594, + 0.021784938871860504, + 0.022772345691919327, + 0.015779299661517143, + 0.012965249828994274, + 0.04418203979730606, + 0.01813773810863495, + 0.04234304651618004, + 0.02168247289955616, + 0.015949636697769165, + 0.004568168893456459, + 0.0038454518653452396, + 0.0028383657336235046, + 0.02441701479256153, + 0.009654350578784943, + 0.004529330413788557, + 0.010844862088561058, + 0.017163312062621117, + 0.016336165368556976, + 0.012841697782278061, + 0.009088595397770405, + 0.007970547303557396, + 0.009085752069950104, + 0.00705130398273468, + 0.010510892607271671, + 0.009304501116275787, + 0.008149806410074234, + 0.004392356611788273, + 0.0035385724622756243, + 0.008854102343320847, + 0.011173132807016373, + 0.00762016698718071, + 0.00600780313834548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.008651158772408962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level perspective shot from inside a car captures the experience of driving in a pouring downpour at night. The windshield is almost completely obscured by a thick layer of rain, countless streams distorting the outside world into an abstract blur of water. Black rubber windshield wiper blades enter the frame, rhythmically sweeping across the glass, methodically pushing the water away and creating a momentary arc of perfect clarity. In this fleeting moment, the smooth, wet asphalt of the road ahead is revealed, gleaming in the bright cones of the car's headlights. Light reflects off the surface of the stagnant water, creating long, shimmering streaks that guide the eye forward. As the wipers complete their motion, rain instantly re-covers the glass, obscuring the view once again, establishing a hypnotic and relentless cycle that creates a moody, cocoon-like atmosphere of focused focus as one navigates the storm.", + "seed": 31946, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_8994_seed_31946_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03313964381814003, + "vlm_logprob": -0.03313964381814003, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03313964381814003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0712893158197403 + ], + "opposite_logprob_tokens": [ + -3.321289300918579 + ], + "logprob_reward": -0.0712893158197403, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0712893158197403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06734348088502884, + 0.1921224296092987, + 0.06956692039966583, + 0.06478304415941238, + 0.052817076444625854, + 0.06556762754917145, + 0.05563700571656227, + 0.1569615751504898, + 0.20822635293006897, + 0.06697442382574081 + ], + "credit_map": [ + 0.010703367181122303, + 0.009032109752297401, + 0.013879380188882351, + 0.009256797842681408, + 0.005529204849153757, + 0.008178284391760826, + 0.004785777535289526, + 0.005978566128760576, + 0.01805972121655941, + 0.020282475277781487, + 0.01414632797241211, + 0.01723676733672619, + 0.05523615702986717, + 0.04056306183338165, + 0.007760099601000547, + 0.01883784867823124, + 0.008406639099121094, + 0.009163297712802887, + 0.006312047131359577, + 0.008487609215080738, + 0.011122826486825943, + 0.009014196693897247, + 0.00621194951236248, + 0.01084836758673191, + 0.00943690910935402, + 0.010400205850601196, + 0.005728748627007008, + 0.007145453244447708, + 0.010957314632833004, + 0.009308435022830963, + 0.005499938502907753, + 0.006306052207946777, + 0.006443211808800697, + 0.007236469071358442, + 0.008305709809064865, + 0.004009950440376997, + 0.00710226921364665, + 0.008383671753108501, + 0.006952895317226648, + 0.004382909741252661, + 0.007593052461743355, + 0.006883473135530949, + 0.006570525001734495, + 0.008726026862859726, + 0.007024669554084539, + 0.010103178210556507, + 0.010058967396616936, + 0.008607747033238411, + 0.006120620295405388, + 0.004837874788790941, + 0.004393316339701414, + 0.005755014717578888, + 0.010965808294713497, + 0.00912386178970337, + 0.007469686213880777, + 0.0069708311930298805, + 0.012210686691105366, + 0.02076946757733822, + 0.011905486695468426, + 0.012685926631093025, + 0.02475235052406788, + 0.029326599091291428, + 0.013659563846886158, + 0.03165150806307793, + 0.023684095591306686, + 0.023153819143772125, + 0.018010945990681648, + 0.01504150778055191, + 0.030811632052063942, + 0.03027924709022045, + 0.020813189446926117, + 0.046431947499513626, + 0.00551505945622921, + 0.011510036885738373, + 0.01040157862007618, + 0.005649244412779808, + 0.011091094464063644, + 0.010425853542983532, + 0.006097879726439714, + 0.0062836832366883755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.0712893158197403 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311934858560562 + ], + "opposite_logprob_tokens": [ + -3.168119430541992 + ], + "logprob_reward": -0.04311934858560562, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311934858560562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12367525696754456, + 0.2963384687900543, + 0.06485246121883392, + 0.09532877802848816, + 0.05196981877088547, + 0.06753392517566681, + 0.03251948207616806, + 0.13709023594856262, + 0.07262616604566574, + 0.05806545913219452 + ], + "credit_map": [ + 0.024694258347153664, + 0.019315147772431374, + 0.03469359502196312, + 0.010872093960642815, + 0.008118843659758568, + 0.013974511064589024, + 0.006956275552511215, + 0.00505051901564002, + 0.022020122036337852, + 0.034241873770952225, + 0.02080622874200344, + 0.018953554332256317, + 0.11131636053323746, + 0.06935136765241623, + 0.00828430987894535, + 0.011364645324647427, + 0.009542285464704037, + 0.012510070577263832, + 0.005214211530983448, + 0.010882745496928692, + 0.009716983884572983, + 0.007042663171887398, + 0.005955192726105452, + 0.003988306969404221, + 0.010533579625189304, + 0.013973557390272617, + 0.007880717515945435, + 0.011747663840651512, + 0.020963750779628754, + 0.0149752888828516, + 0.006597527302801609, + 0.008656682446599007, + 0.007380919065326452, + 0.007352725602686405, + 0.0042367903515696526, + 0.0037189307622611523, + 0.010474562644958496, + 0.00840666051954031, + 0.007216289173811674, + 0.0031829383224248886, + 0.007026055362075567, + 0.009000546298921108, + 0.0033916092943400145, + 0.006489221937954426, + 0.012184242717921734, + 0.009967180900275707, + 0.009054997004568577, + 0.010420069098472595, + 0.0028158915229141712, + 0.0037109286058694124, + 0.0023497280199080706, + 0.0027566966600716114, + 0.007024403661489487, + 0.006148568820208311, + 0.003997762221843004, + 0.0037155013997107744, + 0.015028354711830616, + 0.012726487591862679, + 0.008654681034386158, + 0.007997846230864525, + 0.020746301859617233, + 0.019695572555065155, + 0.012027015909552574, + 0.04021397605538368, + 0.007049921900033951, + 0.00807122141122818, + 0.006632036063820124, + 0.005434846971184015, + 0.012878207489848137, + 0.010148785077035427, + 0.007785551715642214, + 0.014625592157244682, + 0.005596500355750322, + 0.010321064852178097, + 0.006887177936732769, + 0.007493515964597464, + 0.007902367040514946, + 0.008245084434747696, + 0.006732420064508915, + 0.004887324757874012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.04311934858560562 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302748799324036 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302748799324036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302748799324036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.351521760225296, + 0.16870369017124176, + 0.029576022177934647, + 0.030139164999127388, + 0.049822717905044556, + 0.1240728422999382, + 0.016085652634501457, + 0.15879446268081665, + 0.039894554764032364, + 0.031389154493808746 + ], + "credit_map": [ + 0.11050993204116821, + 0.026928145438432693, + 0.1404549479484558, + 0.016819994896650314, + 0.014350292272865772, + 0.030081825330853462, + 0.00659540481865406, + 0.005781234707683325, + 0.014779172837734222, + 0.019996555522084236, + 0.008891038596630096, + 0.008635484613478184, + 0.06804922223091125, + 0.040869079530239105, + 0.0032183537259697914, + 0.004264770075678825, + 0.00299576367251575, + 0.003468597074970603, + 0.0025089646223932505, + 0.006272696889936924, + 0.0042397091165184975, + 0.003518983954563737, + 0.002567027462646365, + 0.004004280082881451, + 0.003140120767056942, + 0.006379279308021069, + 0.002879789797589183, + 0.004309793934226036, + 0.005221146158874035, + 0.0044543067924678326, + 0.0017308879178017378, + 0.0020238407887518406, + 0.004428110085427761, + 0.006221707910299301, + 0.004782012198120356, + 0.0033324165269732475, + 0.00973562803119421, + 0.010455340147018433, + 0.005319461226463318, + 0.005548039451241493, + 0.014454045332968235, + 0.017177550122141838, + 0.007840456441044807, + 0.013278217054903507, + 0.020902341231703758, + 0.018022386357188225, + 0.01858968660235405, + 0.01380816102027893, + 0.0016415166901424527, + 0.001455982099287212, + 0.0011407398851588368, + 0.0016655055806040764, + 0.003924660384654999, + 0.003018263028934598, + 0.0021603303030133247, + 0.001078655244782567, + 0.014546673744916916, + 0.016156349331140518, + 0.009501341730356216, + 0.006275012157857418, + 0.024075224995613098, + 0.02558009885251522, + 0.011903381906449795, + 0.05075638368725777, + 0.0032960474491119385, + 0.006177262868732214, + 0.0031290582846850157, + 0.0020691771060228348, + 0.007882240228354931, + 0.0057670907117426395, + 0.0035512319300323725, + 0.008022446185350418, + 0.0023094431962817907, + 0.004540080204606056, + 0.0036388770677149296, + 0.003141710301861167, + 0.006464624311774969, + 0.005425816867500544, + 0.0035280550364404917, + 0.002340547274798155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.023302748799324036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197370693087578 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197370693087578, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197370693087578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10364704579114914, + 0.11576729267835617, + 0.04111696034669876, + 0.03342995047569275, + 0.03807913884520531, + 0.0361383892595768, + 0.04680107906460762, + 0.20227332413196564, + 0.3080707788467407, + 0.07467609643936157 + ], + "credit_map": [ + 0.019915500655770302, + 0.011015785858035088, + 0.03178925812244415, + 0.008496375754475594, + 0.007603478152304888, + 0.014651014469563961, + 0.004631073679775, + 0.005544548854231834, + 0.008540639653801918, + 0.01434839516878128, + 0.005796703044325113, + 0.006098504178225994, + 0.03896424546837807, + 0.02849556878209114, + 0.004399253521114588, + 0.009123973548412323, + 0.007103291805833578, + 0.006851364392787218, + 0.003941627219319344, + 0.006994821596890688, + 0.0059012072160840034, + 0.004515832755714655, + 0.0029111052863299847, + 0.0028977079782634974, + 0.003189308801665902, + 0.006603502202779055, + 0.003008411265909672, + 0.003919415175914764, + 0.00637789024040103, + 0.004958630073815584, + 0.0023446253035217524, + 0.0030281662475317717, + 0.005894555244594812, + 0.005114843603223562, + 0.0036013242788612843, + 0.0031967880204319954, + 0.005752562079578638, + 0.007214785553514957, + 0.0045412457548081875, + 0.0027630331460386515, + 0.0032794438302516937, + 0.005062502343207598, + 0.0029140838887542486, + 0.003820940852165222, + 0.008058777078986168, + 0.005290755070745945, + 0.005306507460772991, + 0.0024053785018622875, + 0.00439875666052103, + 0.003653676249086857, + 0.0020816479809582233, + 0.0026586619205772877, + 0.013696148060262203, + 0.008047450333833694, + 0.006123641971498728, + 0.006141094025224447, + 0.019279561936855316, + 0.021736793220043182, + 0.01346997544169426, + 0.008911903016269207, + 0.03028852678835392, + 0.035980235785245895, + 0.01553674042224884, + 0.0570695735514164, + 0.03329778090119362, + 0.032734788954257965, + 0.028465960174798965, + 0.02340407483279705, + 0.03900357708334923, + 0.05046170949935913, + 0.02972789667546749, + 0.07097495347261429, + 0.007604503072798252, + 0.00926599558442831, + 0.00844613928347826, + 0.01131194643676281, + 0.013965227641165257, + 0.011846237815916538, + 0.007620855700224638, + 0.004615187179297209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.018197370693087578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978943519294262 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.00978943519294262, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978943519294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4132578372955322, + 0.1337471902370453, + 0.03800288587808609, + 0.02041403017938137, + 0.018464254215359688, + 0.05110568180680275, + 0.049566879868507385, + 0.16762621700763702, + 0.03856900334358215, + 0.06924601644277573 + ], + "credit_map": [ + 0.13632526993751526, + 0.032170072197914124, + 0.1619790941476822, + 0.020341014489531517, + 0.01590995118021965, + 0.03105955757200718, + 0.007644821889698505, + 0.007828052155673504, + 0.011661848053336143, + 0.017438093200325966, + 0.009378484450280666, + 0.012322677299380302, + 0.04633684828877449, + 0.02796843647956848, + 0.0034753084182739258, + 0.005165498238056898, + 0.007297396659851074, + 0.005512906238436699, + 0.003396647283807397, + 0.005466154310852289, + 0.00581212667748332, + 0.004386937245726585, + 0.0032434931490570307, + 0.0028872236143797636, + 0.00290803425014019, + 0.003448555478826165, + 0.0014083755668252707, + 0.002275144448503852, + 0.003627765690907836, + 0.002294189529493451, + 0.002297945087775588, + 0.002154019195586443, + 0.0022694352082908154, + 0.0026422233786433935, + 0.002310631563887, + 0.0019166485872119665, + 0.002467959653586149, + 0.002545667113736272, + 0.0023307809606194496, + 0.0019809079822152853, + 0.005237269680947065, + 0.007943839766085148, + 0.0031668932642787695, + 0.004547626245766878, + 0.00917524378746748, + 0.00672026164829731, + 0.007264906074851751, + 0.007049639243632555, + 0.0044620828703045845, + 0.004249335266649723, + 0.003078074660152197, + 0.004328642040491104, + 0.008545997552573681, + 0.009407536126673222, + 0.00784391164779663, + 0.00765130203217268, + 0.0145610636100173, + 0.01588120311498642, + 0.009228886105120182, + 0.005938549060374498, + 0.026066916063427925, + 0.02450086921453476, + 0.013598558492958546, + 0.0578501857817173, + 0.0033811971079558134, + 0.004679367411881685, + 0.0040352921932935715, + 0.002560282591730356, + 0.008149980567395687, + 0.004875173792243004, + 0.004088072571903467, + 0.006799635011702776, + 0.00781908631324768, + 0.009281044825911522, + 0.013405073434114456, + 0.00790352001786232, + 0.007574856746941805, + 0.00939739029854536, + 0.007615469861775637, + 0.006249571684747934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.00978943519294262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level perspective shot from inside a car captures the experience of driving in a pouring downpour at night. The windshield is almost completely obscured by a thick layer of rain, countless streams distorting the outside world into an abstract blur of water. Black rubber windshield wiper blades enter the frame, rhythmically sweeping across the glass, methodically pushing the water away and creating a momentary arc of perfect clarity. In this fleeting moment, the smooth, wet asphalt of the road ahead is revealed, gleaming in the bright cones of the car's headlights. Light reflects off the surface of the stagnant water, creating long, shimmering streaks that guide the eye forward. As the wipers complete their motion, rain instantly re-covers the glass, obscuring the view once again, establishing a hypnotic and relentless cycle that creates a moody, cocoon-like atmosphere of focused focus as one navigates the storm.", + "seed": 31947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_8994_seed_31947_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03581058476120234, + "vlm_logprob": -0.03581058476120234, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.03581058476120234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07797142118215561 + ], + "opposite_logprob_tokens": [ + -3.0779714584350586 + ], + "logprob_reward": -0.07797142118215561, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.07797142118215561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08894676715135574, + 0.04658781364560127, + 0.2912946045398712, + 0.07598865777254105, + 0.06092222407460213, + 0.07167552411556244, + 0.11327961087226868, + 0.038342591375112534, + 0.09731701016426086, + 0.11564518511295319 + ], + "credit_map": [ + 0.012105334550142288, + 0.0161391943693161, + 0.010360372252762318, + 0.00991151388734579, + 0.01101603452116251, + 0.011069429107010365, + 0.008738095872104168, + 0.00960678793489933, + 0.008070863783359528, + 0.006550688296556473, + 0.004053542390465736, + 0.005055593326687813, + 0.004073980264365673, + 0.00694890646263957, + 0.0063132853247225285, + 0.005520953796803951, + 0.03171175345778465, + 0.03600330650806427, + 0.035479865968227386, + 0.02191312238574028, + 0.04501741752028465, + 0.04198158532381058, + 0.03625944256782532, + 0.042928121984004974, + 0.013175352476537228, + 0.007161617744714022, + 0.008537747897207737, + 0.007027873303741217, + 0.012580506503582, + 0.009943531826138496, + 0.008659448474645615, + 0.008902576752007008, + 0.00871877372264862, + 0.006467313971370459, + 0.0041601937264204025, + 0.0027146830689162016, + 0.006706401240080595, + 0.011726466938853264, + 0.005289857275784016, + 0.0151385348290205, + 0.007453968282788992, + 0.007031239569187164, + 0.006784906145185232, + 0.005359916016459465, + 0.01106216199696064, + 0.012436102144420147, + 0.0074867368675768375, + 0.014060487039387226, + 0.008547123521566391, + 0.009076432324945927, + 0.014259325340390205, + 0.016837989911437035, + 0.013880768790841103, + 0.0136770224198699, + 0.02737281657755375, + 0.009628127329051495, + 0.004715608898550272, + 0.004703132901340723, + 0.0042477939277887344, + 0.0032142470590770245, + 0.00550422677770257, + 0.0057604038156569, + 0.005887242965400219, + 0.00430993502959609, + 0.011988646350800991, + 0.012316005304455757, + 0.01224361639469862, + 0.007105457130819559, + 0.008027075789868832, + 0.02042529545724392, + 0.010268931277096272, + 0.014941984787583351, + 0.011305764317512512, + 0.017503896728157997, + 0.013347935862839222, + 0.008551218546926975, + 0.020264141261577606, + 0.015404126606881618, + 0.01716512441635132, + 0.012102973647415638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.07797142118215561 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02986924909055233 + ], + "opposite_logprob_tokens": [ + -3.529869318008423 + ], + "logprob_reward": -0.02986924909055233, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02986924909055233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05798274278640747, + 0.17110641300678253, + 0.21677467226982117, + 0.1498413383960724, + 0.047937873750925064, + 0.0383567214012146, + 0.05858299508690834, + 0.08882196992635727, + 0.11283592134714127, + 0.05775931850075722 + ], + "credit_map": [ + 0.006032237317413092, + 0.005069046746939421, + 0.0063500977121293545, + 0.009311738424003124, + 0.009478856809437275, + 0.007753703277558088, + 0.006699275691062212, + 0.007287788204848766, + 0.03030407428741455, + 0.01929096318781376, + 0.015798775479197502, + 0.015285362489521503, + 0.01765294000506401, + 0.02774667553603649, + 0.02120533026754856, + 0.023822292685508728, + 0.02475719526410103, + 0.023065822198987007, + 0.025269653648138046, + 0.015965143218636513, + 0.03855372965335846, + 0.03295721858739853, + 0.02935769408941269, + 0.026848211884498596, + 0.02169696055352688, + 0.01600356213748455, + 0.01746571622788906, + 0.011630801483988762, + 0.028602300211787224, + 0.02054452709853649, + 0.018668193370103836, + 0.015229285694658756, + 0.006619388237595558, + 0.0037182862870395184, + 0.0029988810420036316, + 0.0025200278032571077, + 0.007974987849593163, + 0.007294314913451672, + 0.005321038421243429, + 0.011490949429571629, + 0.003748401300981641, + 0.009461194276809692, + 0.0028134442400187254, + 0.0031916641164571047, + 0.003918458707630634, + 0.005598919000476599, + 0.004395547788590193, + 0.005229088943451643, + 0.0036365652922540903, + 0.005523735657334328, + 0.006077886093407869, + 0.007104621268808842, + 0.00740272318944335, + 0.006767826154828072, + 0.015878921374678612, + 0.006190717685967684, + 0.011140095070004463, + 0.011281452141702175, + 0.0130642494186759, + 0.009685487486422062, + 0.010745521634817123, + 0.011685360223054886, + 0.011271380819380283, + 0.009948422200977802, + 0.01570889540016651, + 0.008408813737332821, + 0.008280592039227486, + 0.004328551702201366, + 0.008329934440553188, + 0.03815225511789322, + 0.009517495520412922, + 0.02010938711464405, + 0.005919214803725481, + 0.009119206108152866, + 0.005996611900627613, + 0.005119789391756058, + 0.0076975226402282715, + 0.007999002002179623, + 0.009567410685122013, + 0.006340560968965292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.02986924909055233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306358605623245 + ], + "opposite_logprob_tokens": [ + -3.773306369781494 + ], + "logprob_reward": -0.023306358605623245, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306358605623245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06212740018963814, + 0.04505027085542679, + 0.47574225068092346, + 0.10457682609558105, + 0.023972153663635254, + 0.0389171838760376, + 0.036722030490636826, + 0.06726084649562836, + 0.10429307073354721, + 0.0413379892706871 + ], + "credit_map": [ + 0.006550034508109093, + 0.006960572209209204, + 0.0070803058333694935, + 0.008630987256765366, + 0.007477473933249712, + 0.008169896900653839, + 0.00860567670315504, + 0.008652452379465103, + 0.0064358883537352085, + 0.006502043455839157, + 0.00543741462752223, + 0.0035914124455302954, + 0.004330452531576157, + 0.007844224572181702, + 0.005499218590557575, + 0.005409615114331245, + 0.04596906155347824, + 0.0534399189054966, + 0.058160800486803055, + 0.04258641228079796, + 0.07345525920391083, + 0.06908826529979706, + 0.0692623108625412, + 0.06378023326396942, + 0.012666678987443447, + 0.009701031260192394, + 0.008921205997467041, + 0.006821286398917437, + 0.024083593860268593, + 0.014942953363060951, + 0.01642759144306183, + 0.011012477800250053, + 0.002368004759773612, + 0.0024561623577028513, + 0.0019308999180793762, + 0.002372806193307042, + 0.003983944188803434, + 0.0032012523151934147, + 0.0026165880262851715, + 0.005042496602982283, + 0.0034433838445693254, + 0.006151656620204449, + 0.0033952933736145496, + 0.0025762715376913548, + 0.00356009928509593, + 0.006319695618003607, + 0.0048125809989869595, + 0.008658202365040779, + 0.0035932855680584908, + 0.003240540623664856, + 0.0029096591752022505, + 0.005048333667218685, + 0.006512543186545372, + 0.0047734081745147705, + 0.007087435107678175, + 0.0035568270832300186, + 0.005679554771631956, + 0.005807817447930574, + 0.004207958932965994, + 0.0038067083805799484, + 0.01774158887565136, + 0.010405507870018482, + 0.014231275767087936, + 0.005380433984100819, + 0.010744957253336906, + 0.012018680572509766, + 0.009989844635128975, + 0.005427929572761059, + 0.007905475795269012, + 0.025881560519337654, + 0.008314836770296097, + 0.024009786546230316, + 0.0036564457695931196, + 0.004480763804167509, + 0.00372113985940814, + 0.0029598120599985123, + 0.007856776006519794, + 0.005834338255226612, + 0.008907485753297806, + 0.003921228460967541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.023306358605623245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811657801270485 + ], + "opposite_logprob_tokens": [ + -3.288116693496704 + ], + "logprob_reward": -0.03811657801270485, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811657801270485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049296870827674866, + 0.09942130744457245, + 0.09048756957054138, + 0.09246991574764252, + 0.11405734717845917, + 0.04852396994829178, + 0.07072579115629196, + 0.0828477293252945, + 0.25306448340415955, + 0.09910499304533005 + ], + "credit_map": [ + 0.005191820207983255, + 0.004785794299095869, + 0.00427760137245059, + 0.01013285480439663, + 0.006917996797710657, + 0.006338975392282009, + 0.005636835470795631, + 0.006014995276927948, + 0.01070423610508442, + 0.009052509441971779, + 0.011118556372821331, + 0.0076371366158127785, + 0.009678478352725506, + 0.016970572993159294, + 0.016933364793658257, + 0.01732644997537136, + 0.016699982807040215, + 0.008765585720539093, + 0.011819682084023952, + 0.007622106000781059, + 0.00825236365199089, + 0.01311521790921688, + 0.01213537435978651, + 0.01207725703716278, + 0.014610295183956623, + 0.0131012462079525, + 0.010676603764295578, + 0.009590002708137035, + 0.012198499403893948, + 0.01353359967470169, + 0.009599566459655762, + 0.009160094894468784, + 0.01846098154783249, + 0.01629818044602871, + 0.010625517927110195, + 0.008501476608216763, + 0.014631694182753563, + 0.018328581005334854, + 0.014810152351856232, + 0.012400762178003788, + 0.004754459485411644, + 0.00762744527310133, + 0.004691768903285265, + 0.0055659180507063866, + 0.005728008225560188, + 0.006942950654774904, + 0.004987938329577446, + 0.008225486613810062, + 0.0073017883114516735, + 0.016124607995152473, + 0.00546308234333992, + 0.00776283023878932, + 0.007830668240785599, + 0.00869671255350113, + 0.010308855213224888, + 0.007237243000417948, + 0.008350824937224388, + 0.010559918358922005, + 0.009417477995157242, + 0.010023313574492931, + 0.009947030805051327, + 0.010667604394257069, + 0.012035243213176727, + 0.011846312321722507, + 0.035994354635477066, + 0.014655696228146553, + 0.021827643737196922, + 0.01024715043604374, + 0.01807847060263157, + 0.08846840262413025, + 0.023336052894592285, + 0.04045671597123146, + 0.014398830011487007, + 0.012621646746993065, + 0.013476449064910412, + 0.013674227520823479, + 0.008003177121281624, + 0.013600789941847324, + 0.014857337810099125, + 0.008472539484500885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.03811657801270485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789316914975643 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789316914975643, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789316914975643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15879105031490326, + 0.07963506877422333, + 0.087596595287323, + 0.12908625602722168, + 0.07161105424165726, + 0.03053787350654602, + 0.07156343013048172, + 0.1293107569217682, + 0.15909656882286072, + 0.08277132362127304 + ], + "credit_map": [ + 0.024120375514030457, + 0.02207772247493267, + 0.019086971879005432, + 0.024120552465319633, + 0.016397634521126747, + 0.023233406245708466, + 0.01271244790405035, + 0.017041955143213272, + 0.012782486155629158, + 0.01223651971668005, + 0.008706636726856232, + 0.007120718713849783, + 0.006219475995749235, + 0.013656511902809143, + 0.009559119120240211, + 0.00935360323637724, + 0.009758200496435165, + 0.010016205720603466, + 0.010950619354844093, + 0.01048124860972166, + 0.009707432240247726, + 0.012349192053079605, + 0.011346987448632717, + 0.012986721470952034, + 0.019429711624979973, + 0.015346120111644268, + 0.014418743550777435, + 0.015674011781811714, + 0.0170256607234478, + 0.019956687465310097, + 0.012141316197812557, + 0.01509400736540556, + 0.008999601006507874, + 0.0047617945820093155, + 0.005294828210026026, + 0.0032412176951766014, + 0.012236019596457481, + 0.010758448392152786, + 0.003980740904808044, + 0.02233840525150299, + 0.003160956082865596, + 0.003000464290380478, + 0.002686489140614867, + 0.0019357549026608467, + 0.0060590398497879505, + 0.004580062348395586, + 0.0037286330480128527, + 0.005386475473642349, + 0.007533924654126167, + 0.004218292888253927, + 0.010827059857547283, + 0.0068609751760959625, + 0.006367467809468508, + 0.009517381899058819, + 0.016856331378221512, + 0.00938200019299984, + 0.014580007642507553, + 0.014526749029755592, + 0.018287453800439835, + 0.015119871124625206, + 0.01823079213500023, + 0.019340036436915398, + 0.01654776744544506, + 0.012678093276917934, + 0.01760724000632763, + 0.013499628752470016, + 0.013369066640734673, + 0.008131656795740128, + 0.009940685704350471, + 0.05591677129268646, + 0.008977341465651989, + 0.031654179096221924, + 0.008305033668875694, + 0.021048933267593384, + 0.0073086051270365715, + 0.007342337165027857, + 0.006509581580758095, + 0.011347648687660694, + 0.012207613326609135, + 0.008701572194695473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.009789316914975643 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level perspective shot from inside a car captures the experience of driving in a pouring downpour at night. The windshield is almost completely obscured by a thick layer of rain, countless streams distorting the outside world into an abstract blur of water. Black rubber windshield wiper blades enter the frame, rhythmically sweeping across the glass, methodically pushing the water away and creating a momentary arc of perfect clarity. In this fleeting moment, the smooth, wet asphalt of the road ahead is revealed, gleaming in the bright cones of the car's headlights. Light reflects off the surface of the stagnant water, creating long, shimmering streaks that guide the eye forward. As the wipers complete their motion, rain instantly re-covers the glass, obscuring the view once again, establishing a hypnotic and relentless cycle that creates a moody, cocoon-like atmosphere of focused focus as one navigates the storm.", + "seed": 31948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_10/index_8994_seed_31948_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029302932135760785, + "vlm_logprob": -0.029302932135760785, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.029302932135760785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04944600909948349 + ], + "opposite_logprob_tokens": [ + -3.7994461059570312 + ], + "logprob_reward": -0.04944600909948349, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04944600909948349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36800435185432434, + 0.04002891480922699, + 0.04619341716170311, + 0.3559219539165497, + 0.05725175514817238, + 0.02693209797143936, + 0.0430082343518734, + 0.030972516164183617, + 0.013370418921113014, + 0.0183163620531559 + ], + "credit_map": [ + 0.12275641411542892, + 0.026206593960523605, + 0.040877897292375565, + 0.014034007675945759, + 0.10582947731018066, + 0.0263754241168499, + 0.013807878829538822, + 0.018116649240255356, + 0.005631354637444019, + 0.0035946290008723736, + 0.005279733333736658, + 0.0032224543392658234, + 0.007184970658272505, + 0.004903627093881369, + 0.004975210875272751, + 0.005236937198787928, + 0.007493781857192516, + 0.005677774548530579, + 0.005335783585906029, + 0.005069310311228037, + 0.005476240534335375, + 0.004080611746758223, + 0.00847563985735178, + 0.004584276117384434, + 0.049575645476579666, + 0.055338054895401, + 0.04152621701359749, + 0.04575030133128166, + 0.02786875143647194, + 0.048368122428655624, + 0.04037034884095192, + 0.047124508768320084, + 0.008934387937188148, + 0.0076834168285131454, + 0.007826047949492931, + 0.006239875219762325, + 0.011022240854799747, + 0.006101601291447878, + 0.0056332689709961414, + 0.0038109091110527515, + 0.003740908345207572, + 0.0028056027367711067, + 0.003971555735915899, + 0.002302819164469838, + 0.002034449251368642, + 0.007634532637894154, + 0.0015483208699151874, + 0.002893909579142928, + 0.007685429882258177, + 0.005626159254461527, + 0.007891162298619747, + 0.007091922219842672, + 0.003853922476992011, + 0.0034711286425590515, + 0.0037719933316111565, + 0.0036165164783596992, + 0.004465493839234114, + 0.004142885562032461, + 0.005786839406937361, + 0.0039368364959955215, + 0.002625643275678158, + 0.0038104720879346132, + 0.002105855382978916, + 0.004098488949239254, + 0.00155532022472471, + 0.001879451097920537, + 0.0016288176411762834, + 0.0019457944435998797, + 0.0014221323654055595, + 0.0014496867079287767, + 0.002092731185257435, + 0.001396485255099833, + 0.0016147340647876263, + 0.0023541245609521866, + 0.0024854913353919983, + 0.002142480108886957, + 0.0039089093916118145, + 0.0023335162550210953, + 0.001860221498645842, + 0.0016168850706890225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.04944600909948349 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05504072457551956 + ], + "opposite_logprob_tokens": [ + -2.9300408363342285 + ], + "logprob_reward": -0.05504072457551956, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05504072457551956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4287298023700714, + 0.1247839480638504, + 0.036415427923202515, + 0.1349938064813614, + 0.0427255742251873, + 0.03612532839179039, + 0.050557129085063934, + 0.01723466068506241, + 0.07609755545854568, + 0.052336737513542175 + ], + "credit_map": [ + 0.14606425166130066, + 0.028006242588162422, + 0.05482463166117668, + 0.016846459358930588, + 0.11100547015666962, + 0.033968497067689896, + 0.016744783148169518, + 0.021269479766488075, + 0.01867941953241825, + 0.010854366235435009, + 0.013657419010996819, + 0.010492525063455105, + 0.022947074845433235, + 0.014723114669322968, + 0.016746414825320244, + 0.016683606430888176, + 0.004049976821988821, + 0.004029461182653904, + 0.003825188148766756, + 0.0036955911200493574, + 0.005945941433310509, + 0.0034886703360825777, + 0.006396049167960882, + 0.004984551575034857, + 0.018103623762726784, + 0.02031603828072548, + 0.01513680350035429, + 0.01685836724936962, + 0.01066870242357254, + 0.020835518836975098, + 0.01450346689671278, + 0.018571283668279648, + 0.00955081544816494, + 0.00945149827748537, + 0.006383564323186874, + 0.003029653336852789, + 0.0033557331189513206, + 0.004828483331948519, + 0.0028114758897572756, + 0.003314349800348282, + 0.003422374837100506, + 0.0027689323760569096, + 0.00918728206306696, + 0.004752453416585922, + 0.0027288885321468115, + 0.00947536714375019, + 0.0014468617737293243, + 0.0023431668523699045, + 0.011624746955931187, + 0.003929154947400093, + 0.018553609028458595, + 0.009427385404706001, + 0.0017616951372474432, + 0.0013522524386644363, + 0.0018730969168245792, + 0.0020351908169686794, + 0.002126171253621578, + 0.002332826144993305, + 0.002177030546590686, + 0.003003303427249193, + 0.0016976515762507915, + 0.0018988364608958364, + 0.0016282782889902592, + 0.002370562870055437, + 0.013152946718037128, + 0.007442406844347715, + 0.026575079187750816, + 0.007230098824948072, + 0.006522989831864834, + 0.004397660959511995, + 0.00320182996802032, + 0.007574544288218021, + 0.0032134356442838907, + 0.003205927787348628, + 0.013373341411352158, + 0.003271081717684865, + 0.021480437368154526, + 0.004889308009296656, + 0.0017068290617316961, + 0.0011963756987825036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.05504072457551956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330298162996769 + ], + "opposite_logprob_tokens": [ + -3.7733030319213867 + ], + "logprob_reward": -0.02330298162996769, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330298162996769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21683281660079956, + 0.1292608231306076, + 0.03739320859313011, + 0.31176310777664185, + 0.11089920997619629, + 0.048588190227746964, + 0.04677445814013481, + 0.039649687707424164, + 0.027087457478046417, + 0.03175105154514313 + ], + "credit_map": [ + 0.05671989917755127, + 0.016237426549196243, + 0.035086870193481445, + 0.01012919470667839, + 0.03656145930290222, + 0.01841823011636734, + 0.01687888614833355, + 0.0268008504062891, + 0.018558967858552933, + 0.010509083047509193, + 0.01289148535579443, + 0.011442908085882664, + 0.02303103543817997, + 0.014360166154801846, + 0.020072411745786667, + 0.018394766375422478, + 0.00526859937235713, + 0.005888262297958136, + 0.004567854572087526, + 0.003671170910820365, + 0.004807103890925646, + 0.002600601641461253, + 0.004392531234771013, + 0.0061970860697329044, + 0.04163852334022522, + 0.03940875828266144, + 0.035041045397520065, + 0.03283879533410072, + 0.03425680473446846, + 0.04993690550327301, + 0.03801463544368744, + 0.04062764346599579, + 0.014557160437107086, + 0.015128593891859055, + 0.013502521440386772, + 0.010100970976054668, + 0.02038993127644062, + 0.013777178712189198, + 0.015506533905863762, + 0.00793631561100483, + 0.004407756496220827, + 0.004756603389978409, + 0.007468388881534338, + 0.006780372932553291, + 0.0066821277141571045, + 0.009539748542010784, + 0.00417670700699091, + 0.004776485729962587, + 0.008441902697086334, + 0.003544010454788804, + 0.016689011827111244, + 0.0091185811907053, + 0.0021727392449975014, + 0.0024634574074298143, + 0.0021994789130985737, + 0.00214527640491724, + 0.005145208444446325, + 0.006322204135358334, + 0.008256726898252964, + 0.005303641781210899, + 0.003482663305476308, + 0.004991141147911549, + 0.002752633998170495, + 0.0033954689279198647, + 0.0035995058715343475, + 0.0032105424907058477, + 0.006287052761763334, + 0.0027670450508594513, + 0.003540509846061468, + 0.00298312003724277, + 0.002027098787948489, + 0.002672582631930709, + 0.0036855742800980806, + 0.0030036140233278275, + 0.005972534883767366, + 0.003195442259311676, + 0.010369859635829926, + 0.003070902545005083, + 0.0012867187615484, + 0.0011664057383313775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.02330298162996769 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638650014996529 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.007638650014996529, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638650014996529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2732310891151428, + 0.05137007683515549, + 0.0759652853012085, + 0.2109544724225998, + 0.05804692208766937, + 0.07182193547487259, + 0.0465967059135437, + 0.04364635422825813, + 0.020506324246525764, + 0.14786085486412048 + ], + "credit_map": [ + 0.0954982191324234, + 0.016752388328313828, + 0.043836161494255066, + 0.01170946005731821, + 0.05910533666610718, + 0.02085922472178936, + 0.011158006265759468, + 0.014312307350337505, + 0.0070472657680511475, + 0.005594164133071899, + 0.0074938442558050156, + 0.0028394439723342657, + 0.007228496018797159, + 0.007497011683881283, + 0.00469921575859189, + 0.008970633149147034, + 0.010720986872911453, + 0.007969873026013374, + 0.008738486096262932, + 0.0075491066090762615, + 0.014541403390467167, + 0.007150232791900635, + 0.010973948054015636, + 0.008321242406964302, + 0.02773752622306347, + 0.032899800688028336, + 0.027019741013646126, + 0.025566816329956055, + 0.017121031880378723, + 0.029684077948331833, + 0.024344805628061295, + 0.02658066339790821, + 0.010110300034284592, + 0.007429412100464106, + 0.005933683831244707, + 0.007758664432913065, + 0.007910522632300854, + 0.00622189324349165, + 0.006827172357589006, + 0.005855277646332979, + 0.007518875412642956, + 0.009180872701108456, + 0.012660871259868145, + 0.011161310598254204, + 0.007582161109894514, + 0.015252495184540749, + 0.0037352144718170166, + 0.004730137065052986, + 0.005461293272674084, + 0.0035319970920681953, + 0.01411222480237484, + 0.009271693415939808, + 0.0032503600232303143, + 0.004725584760308266, + 0.0033067604526877403, + 0.0029367879033088684, + 0.007684716954827309, + 0.006245817057788372, + 0.004978635348379612, + 0.006567202042788267, + 0.003773142583668232, + 0.00465444615110755, + 0.004205279517918825, + 0.005537110380828381, + 0.002482644049450755, + 0.002998107811436057, + 0.0030939895659685135, + 0.0022075169254094362, + 0.0023308214731514454, + 0.0030600058380514383, + 0.0013985937694087625, + 0.00293464376591146, + 0.012559543363749981, + 0.010732380673289299, + 0.03392663970589638, + 0.00999679509550333, + 0.04942931607365608, + 0.014960624277591705, + 0.008152661845088005, + 0.00810288731008768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.007638650014996529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an eye-level perspective inside a car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the windshield almost completely obscured by a thick layer of rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are black rubber windshield wiper blades visible sweeping across the glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the road ahead visible as smooth, wet asphalt gleaming in the car's headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show rain instantly re-covering the glass after the wipers clear it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108629535883665 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.01108629535883665, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108629535883665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2854223847389221, + 0.13954401016235352, + 0.026167333126068115, + 0.2783307731151581, + 0.044909194111824036, + 0.04188830032944679, + 0.03828845173120499, + 0.03176835924386978, + 0.019558493047952652, + 0.09412268549203873 + ], + "credit_map": [ + 0.09403721243143082, + 0.01958600804209709, + 0.02037280611693859, + 0.011207755655050278, + 0.08606721460819244, + 0.02607034705579281, + 0.012547076679766178, + 0.015533964149653912, + 0.022932128980755806, + 0.011206942610442638, + 0.01718752831220627, + 0.01112681906670332, + 0.02091677114367485, + 0.016411732882261276, + 0.01641792431473732, + 0.02334415726363659, + 0.0036299473140388727, + 0.002953653922304511, + 0.0029801588971167803, + 0.0022269657347351313, + 0.0038544239941984415, + 0.002287016250193119, + 0.004087650682777166, + 0.004147515632212162, + 0.033951543271541595, + 0.03585629165172577, + 0.03596033528447151, + 0.030045682564377785, + 0.029303954914212227, + 0.04517647624015808, + 0.034533604979515076, + 0.03350288048386574, + 0.0037965320516377687, + 0.01003353763371706, + 0.005510051269084215, + 0.004360242281109095, + 0.008970145136117935, + 0.006762132979929447, + 0.0025765933096408844, + 0.002899960847571492, + 0.004467822145670652, + 0.0032206785399466753, + 0.010159062221646309, + 0.00510986614972353, + 0.002712483052164316, + 0.011745273135602474, + 0.001909812563098967, + 0.0025633019395172596, + 0.0070747812278568745, + 0.002838950837031007, + 0.013245309703052044, + 0.007152304984629154, + 0.001868313062004745, + 0.002449866384267807, + 0.001760874642059207, + 0.0018980520544573665, + 0.004090582020580769, + 0.004244029521942139, + 0.006610923912376165, + 0.0049230968579649925, + 0.0025496522430330515, + 0.004473976790904999, + 0.0021776417270302773, + 0.0026984571013599634, + 0.002909138100221753, + 0.0027640897314995527, + 0.0027167184744030237, + 0.0024680898059159517, + 0.0020345114171504974, + 0.0024734053295105696, + 0.001994161633774638, + 0.0021983773913234472, + 0.007721940521150827, + 0.0068462989293038845, + 0.015380154363811016, + 0.009812108241021633, + 0.03700825572013855, + 0.010064692236483097, + 0.0041772606782615185, + 0.0031119759660214186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.01108629535883665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_11/index_tensor([5587, 5587])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_11/index_tensor([5587, 5587])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..288bd22dd891410e3e99184e5f5b38537b4f4046 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_11/index_tensor([5587, 5587])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady close-up, tilting slowly downward from a slightly downward angle, focuses squarely on a pair of calloused, rough-skinned hands. These hands, with prominent knuckles, are covered in deep veins and scarred by age. They firmly and deftly grip a wooden chisel, its dark brown handle slick with ageless wear. Against the backdrop of a dimly lit workshop, a warm yellow light streams down from above and sideways, precisely illuminating both the hands and the ancient, weathered, dark gray stone tablet beneath. The artisan's hands move steadily and deliberately, the sharp tip of the knife carefully tracing the surface. With each precise thrust and turn, tiny stone chips are peeled away, revealing fluid, intricate curved lines amidst the stark contrast of light and dark. The entire scene exudes an atmosphere of intense concentration and tranquility, imbued with a reverence for traditional craftsmanship and the ritual of the creative moment.", + "seed": 26668, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_11/index_5587_seed_26668_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22221201807260513, + "vlm_logprob": -0.22221201807260513, + "vlm_yes_no_margin": 1.875, + "vlm_token_credit": -0.22221201807260513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03216928988695145 + ], + "opposite_logprob_tokens": [ + -4.157169342041016 + ], + "logprob_reward": -0.03216928988695145, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03216928988695145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07396318763494492, + 0.09713845700025558, + 0.06381102651357651, + 0.12429770082235336, + 0.051672399044036865, + 0.17958536744117737, + 0.061377134174108505, + 0.18963623046875, + 0.07743793725967407, + 0.08108057826757431 + ], + "credit_map": [ + 0.007441381458193064, + 0.008721397258341312, + 0.012417248450219631, + 0.011463350616395473, + 0.009819227270781994, + 0.009248778223991394, + 0.007850943133234978, + 0.007000853773206472, + 0.01052554976195097, + 0.013160389848053455, + 0.013715908862650394, + 0.009495234116911888, + 0.01383588183671236, + 0.012083121575415134, + 0.014673253521323204, + 0.00964909978210926, + 0.007983668707311153, + 0.0074559482745826244, + 0.008333515375852585, + 0.00909793097525835, + 0.00755934277549386, + 0.008680904284119606, + 0.006282547023147345, + 0.008417165838181973, + 0.013078811578452587, + 0.019371025264263153, + 0.02199453115463257, + 0.010108563117682934, + 0.014780282974243164, + 0.01580154336988926, + 0.014575954526662827, + 0.0145869804546237, + 0.004218874964863062, + 0.004210262093693018, + 0.007333348970860243, + 0.008429571986198425, + 0.004924219101667404, + 0.006257448345422745, + 0.008400640450417995, + 0.007898030802607536, + 0.02174556627869606, + 0.023096295073628426, + 0.022749237716197968, + 0.01099937129765749, + 0.029640430584549904, + 0.029893379658460617, + 0.024635575711727142, + 0.016825491562485695, + 0.00635650847107172, + 0.005268009379506111, + 0.007671421859413385, + 0.010045357048511505, + 0.004474361892789602, + 0.008524389006197453, + 0.00838343147188425, + 0.010653656907379627, + 0.009830656461417675, + 0.05195510759949684, + 0.07057495415210724, + 0.008335459977388382, + 0.011683233082294464, + 0.014636647887527943, + 0.013732852414250374, + 0.00888731423765421, + 0.008323494344949722, + 0.008648126386106014, + 0.00983942486345768, + 0.007341739255934954, + 0.010257241316139698, + 0.009970678016543388, + 0.005961614660918713, + 0.01709561236202717, + 0.009512829594314098, + 0.008126121945679188, + 0.010344214737415314, + 0.007577822543680668, + 0.013799157924950123, + 0.012783000245690346, + 0.007658378221094608, + 0.011279051192104816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.03216928988695145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1606316864490509 + ], + "opposite_logprob_tokens": [ + -1.9106316566467285 + ], + "logprob_reward": -0.1606316864490509, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1606316864490509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0928792953491211, + 0.06413266807794571, + 0.08997425436973572, + 0.2636408805847168, + 0.08734116703271866, + 0.19564726948738098, + 0.0435781292617321, + 0.07734932005405426, + 0.031788185238838196, + 0.05366874858736992 + ], + "credit_map": [ + 0.01111292839050293, + 0.010684935376048088, + 0.013689612038433552, + 0.009646780788898468, + 0.008575951680541039, + 0.014670233242213726, + 0.009540584869682789, + 0.014958272688090801, + 0.007628006860613823, + 0.010071906261146069, + 0.008129800669848919, + 0.006773387081921101, + 0.007253210060298443, + 0.008142882958054543, + 0.007886801846325397, + 0.00824667327105999, + 0.010768802836537361, + 0.009663628414273262, + 0.01194845512509346, + 0.011527187190949917, + 0.009076469577848911, + 0.008813356049358845, + 0.015023274347186089, + 0.013153083622455597, + 0.029582498595118523, + 0.03711336851119995, + 0.03964057192206383, + 0.020843498408794403, + 0.034741081297397614, + 0.03492981567978859, + 0.034193381667137146, + 0.032596688717603683, + 0.008525294251739979, + 0.006739155389368534, + 0.011304948478937149, + 0.013376457616686821, + 0.01017756387591362, + 0.0105618080124259, + 0.014832763932645321, + 0.011823175475001335, + 0.02893875725567341, + 0.021909331902861595, + 0.02616770938038826, + 0.01276707649230957, + 0.0320109948515892, + 0.03095574676990509, + 0.02879144623875618, + 0.014106202870607376, + 0.005831179674714804, + 0.004183715209364891, + 0.005252175964415073, + 0.004993508569896221, + 0.005582025740295649, + 0.006990379188209772, + 0.0038634082302451134, + 0.006881736684590578, + 0.0075890617445111275, + 0.013244935311377048, + 0.016257327049970627, + 0.004951094277203083, + 0.01012480165809393, + 0.0072305272333323956, + 0.010449539870023727, + 0.007502037100493908, + 0.003768257098272443, + 0.004678050987422466, + 0.003878531511873007, + 0.0028387242928147316, + 0.0055725653655827045, + 0.003435424529016018, + 0.004112532828003168, + 0.003504097694531083, + 0.006407920271158218, + 0.004741447512060404, + 0.007106601260602474, + 0.004174693953245878, + 0.008449491113424301, + 0.008889320306479931, + 0.004959194455295801, + 0.0089400764554739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.1606316864490509 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211426615715027 + ], + "opposite_logprob_tokens": [ + -1.5021142959594727 + ], + "logprob_reward": -0.25211426615715027, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211426615715027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05443071573972702, + 0.13096876442432404, + 0.052329808473587036, + 0.18163028359413147, + 0.07112757116556168, + 0.1251479834318161, + 0.04960979148745537, + 0.1821127086877823, + 0.042076561599969864, + 0.11056583374738693 + ], + "credit_map": [ + 0.006144045386463404, + 0.008469023741781712, + 0.00927724689245224, + 0.007733118254691362, + 0.004819605499505997, + 0.005799556616693735, + 0.006927854381501675, + 0.005260259844362736, + 0.01767408661544323, + 0.018826955929398537, + 0.020144633948802948, + 0.01577254757285118, + 0.013230212964117527, + 0.01217127125710249, + 0.01695122942328453, + 0.016197815537452698, + 0.00553491385653615, + 0.006395092699676752, + 0.007880410179495811, + 0.0074280062690377235, + 0.00681658461689949, + 0.005295953713357449, + 0.005911705549806356, + 0.0070671397261321545, + 0.018897604197263718, + 0.033740974962711334, + 0.03419548273086548, + 0.01074662059545517, + 0.0209866501390934, + 0.025741836056113243, + 0.02179747074842453, + 0.015523644164204597, + 0.010170374065637589, + 0.00446963170543313, + 0.010266631841659546, + 0.008400618098676205, + 0.010315588675439358, + 0.006594439502805471, + 0.010964690707623959, + 0.009945595636963844, + 0.012019979767501354, + 0.012058847583830357, + 0.0163810346275568, + 0.015173381194472313, + 0.013887456618249416, + 0.016448909416794777, + 0.018354322761297226, + 0.02082403376698494, + 0.006785714067518711, + 0.0055227819830179214, + 0.006357207428663969, + 0.007909563370049, + 0.005453062243759632, + 0.006307628471404314, + 0.006386432331055403, + 0.004887401591986418, + 0.012108979746699333, + 0.04521312192082405, + 0.04762183502316475, + 0.010703187435865402, + 0.014750117436051369, + 0.016277076676487923, + 0.02021145075559616, + 0.015226916410028934, + 0.004743976052850485, + 0.006201345007866621, + 0.005202192347496748, + 0.004591050557792187, + 0.005743284244090319, + 0.004766934085637331, + 0.003838198259472847, + 0.006989579182118177, + 0.012797783128917217, + 0.008195152506232262, + 0.013804118148982525, + 0.0066949063912034035, + 0.016239525750279427, + 0.018870387226343155, + 0.01283520832657814, + 0.02112875133752823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.25211426615715027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233559608459473 + ], + "opposite_logprob_tokens": [ + -0.8983559608459473 + ], + "logprob_reward": -0.5233559608459473, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233559608459473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05965995416045189, + 0.08198107779026031, + 0.08568837493658066, + 0.07345189899206161, + 0.07600308209657669, + 0.1212967187166214, + 0.058909252285957336, + 0.37647825479507446, + 0.029169974848628044, + 0.03736145794391632 + ], + "credit_map": [ + 0.005518690682947636, + 0.007356379646807909, + 0.00946500338613987, + 0.006991945672780275, + 0.008644591085612774, + 0.010009139776229858, + 0.004986410029232502, + 0.006687789224088192, + 0.010530224069952965, + 0.012482360936701298, + 0.010835038498044014, + 0.009053140878677368, + 0.008837309665977955, + 0.008340775966644287, + 0.010039244778454304, + 0.01186298206448555, + 0.012440026737749577, + 0.010782095603644848, + 0.012756665237247944, + 0.012968861497938633, + 0.008715261705219746, + 0.010066439397633076, + 0.008890515193343163, + 0.009068506769835949, + 0.008220626041293144, + 0.011967255733907223, + 0.013525763526558876, + 0.004457684233784676, + 0.007837457582354546, + 0.012532941065728664, + 0.008395668119192123, + 0.006514497101306915, + 0.007267597131431103, + 0.0053832209669053555, + 0.010131682269275188, + 0.010877598077058792, + 0.008428425528109074, + 0.009414434432983398, + 0.011950445361435413, + 0.01254967786371708, + 0.009558289311826229, + 0.011472636833786964, + 0.02005917765200138, + 0.017815938219428062, + 0.010487345978617668, + 0.010471584275364876, + 0.0178073737770319, + 0.023624362424016, + 0.006995582487434149, + 0.005179092288017273, + 0.0068735708482563496, + 0.005471933167427778, + 0.006083070300519466, + 0.010774568654596806, + 0.00582479452714324, + 0.011706635355949402, + 0.022018833085894585, + 0.09720546752214432, + 0.12212319672107697, + 0.02071920968592167, + 0.02492523193359375, + 0.02715265564620495, + 0.031181341037154198, + 0.03115229308605194, + 0.003285413607954979, + 0.002751097548753023, + 0.003566998289898038, + 0.003621264360845089, + 0.0037726613227277994, + 0.0032505716662853956, + 0.004533025436103344, + 0.004388941917568445, + 0.004795538727194071, + 0.0033719672355800867, + 0.004175073001533747, + 0.0033050025813281536, + 0.007387744262814522, + 0.005638724192976952, + 0.003625923302024603, + 0.005061480216681957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.5233559608459473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278888702392578 + ], + "opposite_logprob_tokens": [ + -2.017788887023926 + ], + "logprob_reward": -0.14278888702392578, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278888702392578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07778136432170868, + 0.07319681346416473, + 0.050993118435144424, + 0.11854185909032822, + 0.14942656457424164, + 0.06659157574176788, + 0.07422097027301788, + 0.08208173513412476, + 0.07172417640686035, + 0.23544181883335114 + ], + "credit_map": [ + 0.008552336134016514, + 0.010462417267262936, + 0.012081582099199295, + 0.010149442590773106, + 0.01021315436810255, + 0.007821944542229176, + 0.007582768797874451, + 0.010917725041508675, + 0.011405454948544502, + 0.009571855887770653, + 0.010577852837741375, + 0.007546049077063799, + 0.007690623868256807, + 0.006758896168321371, + 0.01036239042878151, + 0.009283690713346004, + 0.0055999755859375, + 0.006857805419713259, + 0.0062994868494570255, + 0.006236634217202663, + 0.006269284524023533, + 0.005592658184468746, + 0.007359132170677185, + 0.006778140086680651, + 0.00990308728069067, + 0.017399094998836517, + 0.028695104643702507, + 0.01112393569201231, + 0.00955835822969675, + 0.011929675936698914, + 0.01455163024365902, + 0.01538096833974123, + 0.014546815305948257, + 0.011869453825056553, + 0.022663656622171402, + 0.023556312546133995, + 0.016556676477193832, + 0.017809787765145302, + 0.023848524317145348, + 0.01857532188296318, + 0.006419350393116474, + 0.006191030610352755, + 0.007974137552082539, + 0.007147411350160837, + 0.0073856511153280735, + 0.007891394197940826, + 0.010870847851037979, + 0.012711752206087112, + 0.0083448002114892, + 0.007407509256154299, + 0.010237937793135643, + 0.01107616163790226, + 0.0077389613725245, + 0.011003067716956139, + 0.009587914682924747, + 0.008824612945318222, + 0.006388518493622541, + 0.014829671010375023, + 0.015487774275243282, + 0.008968926966190338, + 0.008355551399290562, + 0.008068603463470936, + 0.00950231309980154, + 0.01048037689179182, + 0.008696899749338627, + 0.012282420881092548, + 0.011404878459870815, + 0.004730658605694771, + 0.008031051605939865, + 0.00848765205591917, + 0.006594640202820301, + 0.011495972983539104, + 0.015293538570404053, + 0.01887843757867813, + 0.030925553292036057, + 0.02463226392865181, + 0.036862194538116455, + 0.04131663218140602, + 0.02508598193526268, + 0.04244722053408623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.14278888702392578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady close-up, tilting slowly downward from a slightly downward angle, focuses squarely on a pair of calloused, rough-skinned hands. These hands, with prominent knuckles, are covered in deep veins and scarred by age. They firmly and deftly grip a wooden chisel, its dark brown handle slick with ageless wear. Against the backdrop of a dimly lit workshop, a warm yellow light streams down from above and sideways, precisely illuminating both the hands and the ancient, weathered, dark gray stone tablet beneath. The artisan's hands move steadily and deliberately, the sharp tip of the knife carefully tracing the surface. With each precise thrust and turn, tiny stone chips are peeled away, revealing fluid, intricate curved lines amidst the stark contrast of light and dark. The entire scene exudes an atmosphere of intense concentration and tranquility, imbued with a reverence for traditional craftsmanship and the ritual of the creative moment.", + "seed": 26669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_11/index_5587_seed_26669_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1749817155301571, + "vlm_logprob": -0.1749817155301571, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.1749817155301571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032478030771017075 + ], + "opposite_logprob_tokens": [ + -4.282477855682373 + ], + "logprob_reward": -0.032478030771017075, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.032478030771017075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13047529757022858, + 0.06358004361391068, + 0.06844747811555862, + 0.039852917194366455, + 0.12362296134233475, + 0.06496177613735199, + 0.16670742630958557, + 0.08556180447340012, + 0.2145056575536728, + 0.04228459671139717 + ], + "credit_map": [ + 0.011976226232945919, + 0.01247764565050602, + 0.031094323843717575, + 0.027956046164035797, + 0.009661279618740082, + 0.01116220373660326, + 0.01196723710745573, + 0.014180335216224194, + 0.008045563474297523, + 0.007102267816662788, + 0.010550006292760372, + 0.007518311031162739, + 0.0073680696077644825, + 0.009444467723369598, + 0.008202025666832924, + 0.005349328275769949, + 0.008472933433949947, + 0.005212335381656885, + 0.018086658790707588, + 0.0068410467356443405, + 0.007200365420430899, + 0.008086971938610077, + 0.004576141480356455, + 0.009971022605895996, + 0.008177461102604866, + 0.0036270476412028074, + 0.007504994980990887, + 0.004893569741398096, + 0.003822537837550044, + 0.0036100102588534355, + 0.004724716767668724, + 0.003492581658065319, + 0.02357301488518715, + 0.0068458556197583675, + 0.02596439979970455, + 0.0373242013156414, + 0.006416970398277044, + 0.005935488734394312, + 0.0075614056549966335, + 0.010001628659665585, + 0.010711920447647572, + 0.00537920230999589, + 0.017345653846859932, + 0.011240351013839245, + 0.005331534892320633, + 0.006033712532371283, + 0.004138368181884289, + 0.004781036637723446, + 0.05184592306613922, + 0.008205404505133629, + 0.010956480167806149, + 0.012150902301073074, + 0.01156845223158598, + 0.03962312638759613, + 0.02527407370507717, + 0.007083062548190355, + 0.010243762284517288, + 0.00534432427957654, + 0.029014185070991516, + 0.007619809824973345, + 0.005473677068948746, + 0.018793169409036636, + 0.004463432356715202, + 0.004609445109963417, + 0.025032902136445045, + 0.030621374025940895, + 0.020443640649318695, + 0.03843126818537712, + 0.031125301495194435, + 0.031361162662506104, + 0.018356211483478546, + 0.0191337987780571, + 0.00507277762517333, + 0.005272252485156059, + 0.005031026434153318, + 0.005389621946960688, + 0.006122708786278963, + 0.004678559489548206, + 0.004162678960710764, + 0.00655496958643198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.032478030771017075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07013095915317535 + ], + "opposite_logprob_tokens": [ + -2.6951310634613037 + ], + "logprob_reward": -0.07013095915317535, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07013095915317535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09428651630878448, + 0.033959925174713135, + 0.14618219435214996, + 0.1395934671163559, + 0.03928333520889282, + 0.05678170919418335, + 0.3420496881008148, + 0.07187490165233612, + 0.043576233088970184, + 0.03241198882460594 + ], + "credit_map": [ + 0.010405599139630795, + 0.011586670763790607, + 0.017025263980031013, + 0.01470346748828888, + 0.00799105130136013, + 0.010638640262186527, + 0.010171941481530666, + 0.01176388468593359, + 0.004552939906716347, + 0.0028525772504508495, + 0.007756707724183798, + 0.0032447350677102804, + 0.0032111620530486107, + 0.00421507190912962, + 0.004145483486354351, + 0.00398124847561121, + 0.020016027614474297, + 0.013477777130901814, + 0.01976822130382061, + 0.014461405575275421, + 0.0218172837048769, + 0.02145472541451454, + 0.01596643216907978, + 0.019220320507884026, + 0.03443901613354683, + 0.010743699036538601, + 0.02527601271867752, + 0.01512551587074995, + 0.01328941062092781, + 0.010795867070555687, + 0.01839878410100937, + 0.01152515783905983, + 0.005237029865384102, + 0.0031515334267169237, + 0.006367552559822798, + 0.007280740886926651, + 0.004836553242057562, + 0.004630116280168295, + 0.003659056033939123, + 0.004120752681046724, + 0.00849131029099226, + 0.004380357451736927, + 0.014505975879728794, + 0.009836051613092422, + 0.005589085165411234, + 0.005840390454977751, + 0.003905380144715309, + 0.004233161453157663, + 0.13692331314086914, + 0.011646623723208904, + 0.02185266651213169, + 0.0237464290112257, + 0.018452614545822144, + 0.10294631123542786, + 0.015767866745591164, + 0.010713877156376839, + 0.007734741549938917, + 0.004289936739951372, + 0.022072693333029747, + 0.006624521687626839, + 0.007392249070107937, + 0.013689744286239147, + 0.00584449851885438, + 0.004226521123200655, + 0.0034741638228297234, + 0.006856579799205065, + 0.003844846971333027, + 0.004496516659855843, + 0.005561200901865959, + 0.007696392014622688, + 0.007269392255693674, + 0.004377138800919056, + 0.005212945863604546, + 0.004123914986848831, + 0.004103636834770441, + 0.0034974936861544847, + 0.0044967723079025745, + 0.004729432985186577, + 0.0037734683137387037, + 0.0024743222165852785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.07013095915317535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813374400138855 + ], + "opposite_logprob_tokens": [ + -1.4063374996185303 + ], + "logprob_reward": -0.2813374400138855, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813374400138855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09844470769166946, + 0.020839475095272064, + 0.044674113392829895, + 0.04965231940150261, + 0.16002157330513, + 0.07183023542165756, + 0.41467875242233276, + 0.020981820300221443, + 0.052328694611787796, + 0.06654831022024155 + ], + "credit_map": [ + 0.008950593881309032, + 0.008616711013019085, + 0.02560855634510517, + 0.025017598643898964, + 0.005088054109364748, + 0.008302983827888966, + 0.008621880784630775, + 0.008238324895501137, + 0.002291357610374689, + 0.002325155772268772, + 0.0026290512178093195, + 0.0030465968884527683, + 0.003041704883798957, + 0.002828123513609171, + 0.002457545604556799, + 0.0022199400700628757, + 0.0044923210516572, + 0.004530586302280426, + 0.007530428934842348, + 0.007740206550806761, + 0.006520384922623634, + 0.0056235939264297485, + 0.004031075164675713, + 0.004205517936497927, + 0.008981338702142239, + 0.004544758703559637, + 0.009663665667176247, + 0.0050048534758389, + 0.0058209579437971115, + 0.004344646818935871, + 0.005731894634664059, + 0.005560203921049833, + 0.023475179448723793, + 0.00875904317945242, + 0.03464188426733017, + 0.04476012662053108, + 0.011672895401716232, + 0.016517460346221924, + 0.008994821459054947, + 0.01120017934590578, + 0.01156594604253769, + 0.00475742993876338, + 0.01952046900987625, + 0.013595816679298878, + 0.004810049198567867, + 0.007232209667563438, + 0.0055456929840147495, + 0.004802614916115999, + 0.16684553027153015, + 0.014049927704036236, + 0.025600185617804527, + 0.028245043009519577, + 0.02311050519347191, + 0.12653355300426483, + 0.017440974712371826, + 0.012853031978011131, + 0.0023614291567355394, + 0.0022611822932958603, + 0.0034453340340405703, + 0.002260908717289567, + 0.0031441091559827328, + 0.0034813573583960533, + 0.0017271747346967459, + 0.0023003260139375925, + 0.005129111465066671, + 0.007495393976569176, + 0.004966735374182463, + 0.0063613587990403175, + 0.006378002464771271, + 0.0073033967055380344, + 0.009404826909303665, + 0.0052898721769452095, + 0.00574361952021718, + 0.006155151873826981, + 0.011177649721503258, + 0.011296163313090801, + 0.006608603056520224, + 0.006770074833184481, + 0.010986926965415478, + 0.007810126058757305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.2813374400138855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892876267433167 + ], + "opposite_logprob_tokens": [ + -1.0539287328720093 + ], + "logprob_reward": -0.42892876267433167, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892876267433167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23074662685394287, + 0.02679695188999176, + 0.07306971400976181, + 0.1330592781305313, + 0.14174668490886688, + 0.10418400168418884, + 0.17105011641979218, + 0.012037270702421665, + 0.06910059601068497, + 0.03820870816707611 + ], + "credit_map": [ + 0.021455001085996628, + 0.014244993217289448, + 0.07207280397415161, + 0.08061753213405609, + 0.009063784964382648, + 0.011906620115041733, + 0.01255942415446043, + 0.00882647279649973, + 0.0038710618391633034, + 0.0033633827697485685, + 0.004227901808917522, + 0.0029707730282098055, + 0.0026727167423814535, + 0.0033461672719568014, + 0.0024964644107967615, + 0.003848486812785268, + 0.015961626544594765, + 0.0025579105131328106, + 0.02821366861462593, + 0.004452293273061514, + 0.005141065921634436, + 0.005036423448473215, + 0.002414809772744775, + 0.009291918016970158, + 0.039597805589437485, + 0.008477475494146347, + 0.029184455052018166, + 0.010651065036654472, + 0.01331719383597374, + 0.009122458286583424, + 0.011489308439195156, + 0.011219526641070843, + 0.021431243047118187, + 0.010540468618273735, + 0.0220916997641325, + 0.037455033510923386, + 0.014733212999999523, + 0.016295665875077248, + 0.009123961441218853, + 0.010075407102704048, + 0.015132519416511059, + 0.008517809212207794, + 0.025211140513420105, + 0.015583129599690437, + 0.010229907929897308, + 0.011890876106917858, + 0.008257748559117317, + 0.009360870346426964, + 0.053469665348529816, + 0.008310050703585148, + 0.015757186338305473, + 0.012445759028196335, + 0.011665897443890572, + 0.03954344615340233, + 0.02312532812356949, + 0.006732778623700142, + 0.0024334744084626436, + 0.000943448452744633, + 0.0013343498576432467, + 0.0009665268007665873, + 0.0015595415607094765, + 0.0016260562697425485, + 0.001329543418250978, + 0.0018443302251398563, + 0.007473761215806007, + 0.008881717920303345, + 0.006943537387996912, + 0.013008751906454563, + 0.010448405519127846, + 0.009721004404127598, + 0.0069322530180215836, + 0.0056911613792181015, + 0.00472556846216321, + 0.002968589775264263, + 0.008404184132814407, + 0.004296676721423864, + 0.0038759757298976183, + 0.0038346729706972837, + 0.006274611223489046, + 0.003828432410955429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.42892876267433167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062033385038375854 + ], + "opposite_logprob_tokens": [ + -2.8120334148406982 + ], + "logprob_reward": -0.062033385038375854, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062033385038375854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2501596212387085, + 0.08120441436767578, + 0.04718611761927605, + 0.06647713482379913, + 0.08665280044078827, + 0.05426454544067383, + 0.23544971644878387, + 0.09446008503437042, + 0.0573575384914875, + 0.02678801864385605 + ], + "credit_map": [ + 0.024382440373301506, + 0.015449781902134418, + 0.08021898567676544, + 0.08128836750984192, + 0.010370463132858276, + 0.012357527390122414, + 0.01538284495472908, + 0.010709214955568314, + 0.010415208525955677, + 0.00992512609809637, + 0.012540409341454506, + 0.012117126025259495, + 0.008163389749825, + 0.011264714412391186, + 0.010210133157670498, + 0.0065683103166520596, + 0.007643264718353748, + 0.0036130102816969156, + 0.011827104724943638, + 0.005889615975320339, + 0.004772586282342672, + 0.005331412889063358, + 0.003935121465474367, + 0.004174003377556801, + 0.012494022957980633, + 0.003939580172300339, + 0.021286750212311745, + 0.0057200961746275425, + 0.009638664312660694, + 0.003696760395541787, + 0.004750245716422796, + 0.004951015580445528, + 0.01635313779115677, + 0.007975229062139988, + 0.018119478598237038, + 0.01117042638361454, + 0.011059883050620556, + 0.010083957575261593, + 0.004762507509440184, + 0.007128178142011166, + 0.0087337177246809, + 0.004475892521440983, + 0.012018389999866486, + 0.008824890479445457, + 0.005422354210168123, + 0.0064915381371974945, + 0.003922328818589449, + 0.004375434946268797, + 0.04869040846824646, + 0.015380826778709888, + 0.02024715021252632, + 0.016123343259096146, + 0.021146075800061226, + 0.03464900702238083, + 0.06337267905473709, + 0.015840236097574234, + 0.012015074491500854, + 0.005820973310619593, + 0.025281602516770363, + 0.007255705073475838, + 0.00887403916567564, + 0.016827000305056572, + 0.006955068092793226, + 0.01143062487244606, + 0.005324740428477526, + 0.006095059681683779, + 0.008968617767095566, + 0.0076904818415641785, + 0.008068236522376537, + 0.008247090503573418, + 0.007604042999446392, + 0.00535927340388298, + 0.0027231841813772917, + 0.0032348025124520063, + 0.003140509594231844, + 0.0039702593348920345, + 0.0033758063800632954, + 0.0034466530196368694, + 0.004166554659605026, + 0.0027302480302751064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.062033385038375854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady close-up, tilting slowly downward from a slightly downward angle, focuses squarely on a pair of calloused, rough-skinned hands. These hands, with prominent knuckles, are covered in deep veins and scarred by age. They firmly and deftly grip a wooden chisel, its dark brown handle slick with ageless wear. Against the backdrop of a dimly lit workshop, a warm yellow light streams down from above and sideways, precisely illuminating both the hands and the ancient, weathered, dark gray stone tablet beneath. The artisan's hands move steadily and deliberately, the sharp tip of the knife carefully tracing the surface. With each precise thrust and turn, tiny stone chips are peeled away, revealing fluid, intricate curved lines amidst the stark contrast of light and dark. The entire scene exudes an atmosphere of intense concentration and tranquility, imbued with a reverence for traditional craftsmanship and the ritual of the creative moment.", + "seed": 26670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_11/index_5587_seed_26670_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11959807686507702, + "vlm_logprob": -0.11959807686507702, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.11959807686507702, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023984061554074287 + ], + "opposite_logprob_tokens": [ + -4.648983955383301 + ], + "logprob_reward": -0.023984061554074287, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.023984061554074287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.503287136554718, + 0.04574177786707878, + 0.05069824308156967, + 0.01484594214707613, + 0.045313648879528046, + 0.06562280654907227, + 0.09198056161403656, + 0.0498521514236927, + 0.09255335479974747, + 0.04010432958602905 + ], + "credit_map": [ + 0.03534765914082527, + 0.01783817820250988, + 0.21068918704986572, + 0.016399957239627838, + 0.017609013244509697, + 0.16723783314228058, + 0.01861683465540409, + 0.019548485055565834, + 0.009190896525979042, + 0.0052648368291556835, + 0.007368697784841061, + 0.0076914275996387005, + 0.0026125365402549505, + 0.0049240305088460445, + 0.00391283119097352, + 0.004776519723236561, + 0.004822995979338884, + 0.006859061773866415, + 0.004344465211033821, + 0.007531436160206795, + 0.006667186040431261, + 0.007187733892351389, + 0.008991078473627567, + 0.004294287413358688, + 0.001416667364537716, + 0.0017105068545788527, + 0.0024280447978526354, + 0.002357030287384987, + 0.002201895462349057, + 0.0014121703570708632, + 0.0021705972030758858, + 0.0011490302858874202, + 0.004895010031759739, + 0.004336503800004721, + 0.008020046167075634, + 0.004392170812934637, + 0.0044814604334533215, + 0.008188717067241669, + 0.006037393119186163, + 0.004962347913533449, + 0.008690059185028076, + 0.0058885738253593445, + 0.010037777945399284, + 0.009365915320813656, + 0.009697302244603634, + 0.00584307499229908, + 0.007905544713139534, + 0.008194555528461933, + 0.0161993820220232, + 0.010213944129645824, + 0.017721720039844513, + 0.009094742126762867, + 0.012377822771668434, + 0.007565986830741167, + 0.009523391723632812, + 0.009283577091991901, + 0.007910719141364098, + 0.007032410241663456, + 0.005200209561735392, + 0.00428436603397131, + 0.007099250331521034, + 0.007786663249135017, + 0.006799883209168911, + 0.0037386512849479914, + 0.008077261969447136, + 0.00672816252335906, + 0.028366850689053535, + 0.022185932844877243, + 0.007273936178535223, + 0.006999562028795481, + 0.006772102788090706, + 0.006149550434201956, + 0.004122376907616854, + 0.010730933398008347, + 0.004573165904730558, + 0.00566765945404768, + 0.004345273599028587, + 0.003758228849619627, + 0.003748074173927307, + 0.0031586186960339546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.023984061554074287 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08916202932596207 + ], + "opposite_logprob_tokens": [ + -2.4641621112823486 + ], + "logprob_reward": -0.08916202932596207, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08916202932596207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2128022462129593, + 0.05300780013203621, + 0.11065885424613953, + 0.16806603968143463, + 0.07001391798257828, + 0.04411149024963379, + 0.14418256282806396, + 0.04326577112078667, + 0.08087973296642303, + 0.07301155477762222 + ], + "credit_map": [ + 0.015954934060573578, + 0.011560851708054543, + 0.07271793484687805, + 0.00934482179582119, + 0.012660596519708633, + 0.06290328502655029, + 0.01495834905654192, + 0.012701489962637424, + 0.008494122885167599, + 0.00805006641894579, + 0.00873680878430605, + 0.008761314675211906, + 0.0032732761465013027, + 0.005945782642811537, + 0.005121887661516666, + 0.004624543245881796, + 0.013312212191522121, + 0.01890527829527855, + 0.01484671887010336, + 0.013286652974784374, + 0.013568823225796223, + 0.014394693076610565, + 0.011790035292506218, + 0.010554456152021885, + 0.015207149088382721, + 0.021900003775954247, + 0.015295364893972874, + 0.024015236645936966, + 0.03338943421840668, + 0.0221197959035635, + 0.022883901372551918, + 0.013255163095891476, + 0.007184069138020277, + 0.005857014562934637, + 0.016324331983923912, + 0.007450974080711603, + 0.007406890857964754, + 0.006499889772385359, + 0.009363924153149128, + 0.00992683507502079, + 0.005480161868035793, + 0.003097194479778409, + 0.0060323369689285755, + 0.003865822684019804, + 0.007098131347447634, + 0.0032776931766420603, + 0.005810091737657785, + 0.009450063109397888, + 0.017299605533480644, + 0.015736930072307587, + 0.027335355058312416, + 0.015304047614336014, + 0.015523303300142288, + 0.015686430037021637, + 0.01994321309030056, + 0.01735369861125946, + 0.005607918370515108, + 0.006792970467358828, + 0.0047444566152989864, + 0.0034623872488737106, + 0.005863612052053213, + 0.006574839353561401, + 0.0052292426116764545, + 0.004990347661077976, + 0.006383928935974836, + 0.004825008567422628, + 0.01512789074331522, + 0.009497052989900112, + 0.005899058189243078, + 0.0035203523002564907, + 0.031215021386742592, + 0.004411428701132536, + 0.011786055751144886, + 0.013456903398036957, + 0.008461467921733856, + 0.010466545820236206, + 0.007457889150828123, + 0.005897953640669584, + 0.00813718419522047, + 0.007347559556365013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.08916202932596207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557589411735535 + ], + "opposite_logprob_tokens": [ + -1.6005759239196777 + ], + "logprob_reward": -0.22557589411735535, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557589411735535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3808125853538513, + 0.03044094331562519, + 0.05732262134552002, + 0.0948747918009758, + 0.04822544753551483, + 0.10051880776882172, + 0.10260991752147675, + 0.07001589238643646, + 0.06807954609394073, + 0.047099519520998 + ], + "credit_map": [ + 0.028399547562003136, + 0.014897365123033524, + 0.16170579195022583, + 0.011456042528152466, + 0.012329170480370522, + 0.12101234495639801, + 0.013432644307613373, + 0.01757964864373207, + 0.006237039342522621, + 0.0040482040494680405, + 0.004077925346791744, + 0.005377829074859619, + 0.0018737229984253645, + 0.0029254253022372723, + 0.0025860450696200132, + 0.0033147521317005157, + 0.005586792249232531, + 0.006374511402100325, + 0.005992078222334385, + 0.0095591489225626, + 0.004670548718422651, + 0.008215103298425674, + 0.012188481166958809, + 0.004735957365483046, + 0.009112540632486343, + 0.011750615201890469, + 0.008192352019250393, + 0.012422936037182808, + 0.019450265914201736, + 0.01283950824290514, + 0.014194446615874767, + 0.006912125740200281, + 0.005695942789316177, + 0.005811692681163549, + 0.005078545771539211, + 0.005607555154711008, + 0.006190738640725613, + 0.00718059204518795, + 0.0067001027055084705, + 0.0059602768160402775, + 0.0122549282386899, + 0.009854104369878769, + 0.01303581241518259, + 0.01699991710484028, + 0.016299627721309662, + 0.007600830867886543, + 0.01085624098777771, + 0.01361734326928854, + 0.014777783304452896, + 0.00799411628395319, + 0.018786275759339333, + 0.014669539406895638, + 0.01128372922539711, + 0.00808769091963768, + 0.012456689961254597, + 0.014554102905094624, + 0.008261168375611305, + 0.01271572895348072, + 0.006872645579278469, + 0.005691359750926495, + 0.01051960326731205, + 0.010212517343461514, + 0.008519311435520649, + 0.007223553955554962, + 0.00605774624273181, + 0.00843360461294651, + 0.011305004358291626, + 0.014116565696895123, + 0.007623330224305391, + 0.00987766869366169, + 0.0042077405378222466, + 0.006457884330302477, + 0.005558916367590427, + 0.008549721911549568, + 0.005909724626690149, + 0.006780360359698534, + 0.0065140193328261375, + 0.0037119430489838123, + 0.005343600641936064, + 0.004731231369078159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.22557589411735535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557836771011353 + ], + "opposite_logprob_tokens": [ + -1.6005783081054688 + ], + "logprob_reward": -0.22557836771011353, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557836771011353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0850699320435524, + 0.02072351798415184, + 0.05665530636906624, + 0.05203981325030327, + 0.07341501861810684, + 0.30440476536750793, + 0.15876054763793945, + 0.04572078585624695, + 0.118852399289608, + 0.08435795456171036 + ], + "credit_map": [ + 0.006891593337059021, + 0.005398126319050789, + 0.019965702667832375, + 0.00558084761723876, + 0.004827036056667566, + 0.02661014162003994, + 0.006648377049714327, + 0.009148101322352886, + 0.0030096210539340973, + 0.0035111228935420513, + 0.0031380250584334135, + 0.00350668141618371, + 0.0013946114340797067, + 0.0018356991931796074, + 0.002109826309606433, + 0.0022179291117936373, + 0.005390969570726156, + 0.007271645125001669, + 0.006300304085016251, + 0.008569774217903614, + 0.004283359739929438, + 0.008135946467518806, + 0.011222682893276215, + 0.005480624735355377, + 0.004998940043151379, + 0.005925352685153484, + 0.008814121596515179, + 0.006614722311496735, + 0.008634114637970924, + 0.005691382102668285, + 0.007300830911844969, + 0.004060348030179739, + 0.008146630600094795, + 0.004028964322060347, + 0.027903782203793526, + 0.006371317431330681, + 0.005227094981819391, + 0.0050392746925354, + 0.009842567145824432, + 0.006855382118374109, + 0.04228148236870766, + 0.02713131718337536, + 0.0498335100710392, + 0.03681602329015732, + 0.054743826389312744, + 0.025480400770902634, + 0.03128877282142639, + 0.03682941943407059, + 0.01993262767791748, + 0.017041070386767387, + 0.023855159059166908, + 0.02069355919957161, + 0.0168436411768198, + 0.014119959436357021, + 0.025367923080921173, + 0.020906586199998856, + 0.005176458973437548, + 0.006287286523729563, + 0.0054627968929708, + 0.005903915036469698, + 0.005408876109868288, + 0.005113554652780294, + 0.006695291958749294, + 0.005672600120306015, + 0.008252868428826332, + 0.009417637251317501, + 0.020893020555377007, + 0.022953355684876442, + 0.01006226148456335, + 0.009000197052955627, + 0.020170534029603004, + 0.01810251548886299, + 0.011364727281033993, + 0.01841791905462742, + 0.009436702355742455, + 0.00760993268340826, + 0.008542333729565144, + 0.006829403340816498, + 0.010088157840073109, + 0.012068770825862885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.22557836771011353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369003161787987 + ], + "opposite_logprob_tokens": [ + -3.4086899757385254 + ], + "logprob_reward": -0.03369003161787987, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369003161787987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21580152213573456, + 0.027852743864059448, + 0.019685611128807068, + 0.04201699420809746, + 0.1058279275894165, + 0.06431137025356293, + 0.10922717303037643, + 0.025309601798653603, + 0.3022063374519348, + 0.08776070177555084 + ], + "credit_map": [ + 0.014714010991156101, + 0.0077918036840856075, + 0.08738557994365692, + 0.008121196180582047, + 0.008111580274999142, + 0.07197736948728561, + 0.009738059714436531, + 0.007961933501064777, + 0.005124950315803289, + 0.0037549997214227915, + 0.00414617545902729, + 0.005349942948669195, + 0.0016698079416528344, + 0.0025282606948167086, + 0.002507504541426897, + 0.0027711004950106144, + 0.0019510292913764715, + 0.0023568233009427786, + 0.0025290425401180983, + 0.002543899230659008, + 0.002161009470000863, + 0.0024884019512683153, + 0.0037142811343073845, + 0.0019411237444728613, + 0.005225284490734339, + 0.0039703939110040665, + 0.006311566103249788, + 0.006574858911335468, + 0.003996974788606167, + 0.005472434218972921, + 0.0068710194900631905, + 0.003594461362808943, + 0.010145360603928566, + 0.011895987205207348, + 0.021680567413568497, + 0.012597057037055492, + 0.012392234988510609, + 0.013089928776025772, + 0.013789558783173561, + 0.010237231850624084, + 0.006734870839864016, + 0.007010679692029953, + 0.0087392907589674, + 0.009237195365130901, + 0.008041620254516602, + 0.005316139664500952, + 0.010679543949663639, + 0.008552031591534615, + 0.018265660852193832, + 0.010943830944597721, + 0.015280982479453087, + 0.00851208996027708, + 0.014991478994488716, + 0.01068549882620573, + 0.015505735762417316, + 0.015041890554130077, + 0.0029442748054862022, + 0.0027166360523551702, + 0.003950247075408697, + 0.0029027035925537348, + 0.003182320622727275, + 0.0019642263650894165, + 0.004495875909924507, + 0.003153315046802163, + 0.021683908998966217, + 0.024898355826735497, + 0.04965914040803909, + 0.01767689920961857, + 0.0247405543923378, + 0.023660337552428246, + 0.10040591657161713, + 0.03948122262954712, + 0.012668889947235584, + 0.01705886796116829, + 0.011432623490691185, + 0.009815429337322712, + 0.012166178785264492, + 0.007100006565451622, + 0.008461094461381435, + 0.009057608433067799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.03369003161787987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady close-up, tilting slowly downward from a slightly downward angle, focuses squarely on a pair of calloused, rough-skinned hands. These hands, with prominent knuckles, are covered in deep veins and scarred by age. They firmly and deftly grip a wooden chisel, its dark brown handle slick with ageless wear. Against the backdrop of a dimly lit workshop, a warm yellow light streams down from above and sideways, precisely illuminating both the hands and the ancient, weathered, dark gray stone tablet beneath. The artisan's hands move steadily and deliberately, the sharp tip of the knife carefully tracing the surface. With each precise thrust and turn, tiny stone chips are peeled away, revealing fluid, intricate curved lines amidst the stark contrast of light and dark. The entire scene exudes an atmosphere of intense concentration and tranquility, imbued with a reverence for traditional craftsmanship and the ritual of the creative moment.", + "seed": 26671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_11/index_5587_seed_26671_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2135280951857567, + "vlm_logprob": -0.2135280951857567, + "vlm_yes_no_margin": 1.825, + "vlm_token_credit": -0.2135280951857567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05072341114282608 + ], + "opposite_logprob_tokens": [ + -3.3007233142852783 + ], + "logprob_reward": -0.05072341114282608, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05072341114282608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11095220595598221, + 0.07994583994150162, + 0.0877484381198883, + 0.09961459785699844, + 0.10131514817476273, + 0.09436482191085815, + 0.2617245018482208, + 0.04336496815085411, + 0.02817951701581478, + 0.0927899032831192 + ], + "credit_map": [ + 0.01730070821940899, + 0.009931125678122044, + 0.012285755015909672, + 0.01348049659281969, + 0.017201563343405724, + 0.015327472239732742, + 0.014066532254219055, + 0.011358569376170635, + 0.007255325093865395, + 0.01803532987833023, + 0.009740532375872135, + 0.0061260005459189415, + 0.0200964268296957, + 0.0061175646260380745, + 0.008836000226438046, + 0.0037386652547866106, + 0.006839602254331112, + 0.005881824996322393, + 0.012319568544626236, + 0.007701822556555271, + 0.009787301532924175, + 0.01007121428847313, + 0.01083669438958168, + 0.024310413748025894, + 0.00933179073035717, + 0.014923881739377975, + 0.014444751664996147, + 0.008572500199079514, + 0.015791155397892, + 0.011334141716361046, + 0.009918681345880032, + 0.015297699719667435, + 0.011296248063445091, + 0.01276494562625885, + 0.014841381460428238, + 0.009175071492791176, + 0.015288140624761581, + 0.01021964754909277, + 0.012045935727655888, + 0.015683775767683983, + 0.008638907223939896, + 0.007879452779889107, + 0.014831717126071453, + 0.008143434301018715, + 0.013534614816308022, + 0.014421375468373299, + 0.010048240423202515, + 0.016867080703377724, + 0.015956265851855278, + 0.010070953518152237, + 0.09911070764064789, + 0.01084268931299448, + 0.0740065947175026, + 0.012770329602062702, + 0.02800118736922741, + 0.010965793393552303, + 0.00456817913800478, + 0.004277389496564865, + 0.0074096196331083775, + 0.005418687127530575, + 0.0030799852684140205, + 0.009543510153889656, + 0.005527626257389784, + 0.003539973171427846, + 0.004505731165409088, + 0.005032941699028015, + 0.0028861179016530514, + 0.00301685300655663, + 0.0037810939829796553, + 0.0031882477924227715, + 0.0031838619615882635, + 0.002584672998636961, + 0.012450400739908218, + 0.012250976637005806, + 0.009988057427108288, + 0.013076565228402615, + 0.011884829960763454, + 0.012403261847794056, + 0.009734881110489368, + 0.011000936850905418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.05072341114282608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16054102778434753 + ], + "opposite_logprob_tokens": [ + -1.91054105758667 + ], + "logprob_reward": -0.16054102778434753, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16054102778434753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523755043745041, + 0.06413871794939041, + 0.07265704870223999, + 0.0380188412964344, + 0.0943579375743866, + 0.08482228219509125, + 0.2236221879720688, + 0.06467884033918381, + 0.1424812525510788, + 0.06284743547439575 + ], + "credit_map": [ + 0.025789353996515274, + 0.01552657876163721, + 0.01695232465863228, + 0.01772565022110939, + 0.02201416715979576, + 0.018549885600805283, + 0.01880686543881893, + 0.017010660842061043, + 0.006831291131675243, + 0.006139089353382587, + 0.010010208003222942, + 0.007280155085027218, + 0.008451546542346478, + 0.00995846651494503, + 0.009504870511591434, + 0.0059630912728607655, + 0.005901219788938761, + 0.00813983753323555, + 0.010830000974237919, + 0.008346683345735073, + 0.011298338882625103, + 0.010566764511168003, + 0.008172111585736275, + 0.00940208975225687, + 0.0035437243059277534, + 0.004845107439905405, + 0.005306274630129337, + 0.0034013877157121897, + 0.005695132538676262, + 0.004079312086105347, + 0.004981148988008499, + 0.006166749633848667, + 0.007903827354311943, + 0.012674931436777115, + 0.015756890177726746, + 0.007306503131985664, + 0.019192568957805634, + 0.012105475179851055, + 0.008117939345538616, + 0.011299793608486652, + 0.005197556223720312, + 0.007965879514813423, + 0.012266204692423344, + 0.009593673050403595, + 0.013537263497710228, + 0.013593767769634724, + 0.008236504159867764, + 0.014431430958211422, + 0.013781133107841015, + 0.008378084748983383, + 0.08655520528554916, + 0.0080769844353199, + 0.05692720413208008, + 0.01180802471935749, + 0.023231327533721924, + 0.01486421748995781, + 0.007817748934030533, + 0.008798648603260517, + 0.008213640190660954, + 0.008903725072741508, + 0.005777006968855858, + 0.0127967344596982, + 0.007225640583783388, + 0.005145692266523838, + 0.017046133056282997, + 0.03887755051255226, + 0.010173316113650799, + 0.012613791041076183, + 0.016104545444250107, + 0.013549999333918095, + 0.016577230766415596, + 0.01753867045044899, + 0.010401209816336632, + 0.006982802413403988, + 0.0075743542984128, + 0.00817556120455265, + 0.006522636394947767, + 0.00872822105884552, + 0.007112066261470318, + 0.007350576110184193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.16054102778434753 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990051209926605 + ], + "opposite_logprob_tokens": [ + -1.8049005270004272 + ], + "logprob_reward": -0.17990051209926605, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990051209926605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04903259873390198, + 0.08023228496313095, + 0.04406989738345146, + 0.09711045771837234, + 0.11028041690587997, + 0.09355372190475464, + 0.18082600831985474, + 0.03602449223399162, + 0.04616525024175644, + 0.26270484924316406 + ], + "credit_map": [ + 0.007935636676847935, + 0.006662247236818075, + 0.0039008199237287045, + 0.006323755253106356, + 0.008537642657756805, + 0.005362347699701786, + 0.005346561316400766, + 0.004963591694831848, + 0.008571763522922993, + 0.017984600737690926, + 0.009538916870951653, + 0.006187590770423412, + 0.018548814579844475, + 0.005798342637717724, + 0.01030946709215641, + 0.0032927889842540026, + 0.004435562063008547, + 0.00574926333501935, + 0.006194147281348705, + 0.004577270708978176, + 0.005674964748322964, + 0.005895869806408882, + 0.004650670103728771, + 0.006892152596265078, + 0.006912732031196356, + 0.018771857023239136, + 0.013973268680274487, + 0.008786865510046482, + 0.013537370599806309, + 0.006517026107758284, + 0.012493034824728966, + 0.016118302941322327, + 0.008382179774343967, + 0.012243772856891155, + 0.018586302176117897, + 0.008513819426298141, + 0.022153953090310097, + 0.014254191890358925, + 0.012876539491117, + 0.013269660994410515, + 0.006890927907079458, + 0.009637871757149696, + 0.013960768468677998, + 0.008972792886197567, + 0.01785188540816307, + 0.01785270869731903, + 0.00745487492531538, + 0.010931898839771748, + 0.012043272145092487, + 0.009848296642303467, + 0.030020616948604584, + 0.006264001131057739, + 0.06276549398899078, + 0.01110044028609991, + 0.019641468301415443, + 0.029142439365386963, + 0.006798450369387865, + 0.0036269056145101786, + 0.006086388137191534, + 0.005290476139634848, + 0.002477064495906234, + 0.004344835877418518, + 0.005048318766057491, + 0.00235205446369946, + 0.006421081256121397, + 0.00739958556368947, + 0.005543523468077183, + 0.004451600834727287, + 0.007329707499593496, + 0.006389537360519171, + 0.0051992726512253284, + 0.00343094184063375, + 0.02582269348204136, + 0.033798716962337494, + 0.02211465872824192, + 0.03094670921564102, + 0.02936387062072754, + 0.042260654270648956, + 0.04756758734583855, + 0.03082997165620327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.17990051209926605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761618614196777 + ], + "opposite_logprob_tokens": [ + -0.8261618614196777 + ], + "logprob_reward": -0.5761618614196777, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761618614196777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0651634931564331, + 0.14994557201862335, + 0.03244900703430176, + 0.06712116301059723, + 0.17717693746089935, + 0.08956924825906754, + 0.18125899136066437, + 0.04895391687750816, + 0.05786674842238426, + 0.13049492239952087 + ], + "credit_map": [ + 0.009487153962254524, + 0.010695954784750938, + 0.004987883847206831, + 0.007789694704115391, + 0.011720623821020126, + 0.006509474478662014, + 0.008869016543030739, + 0.0051036919467151165, + 0.015879066661000252, + 0.02233327366411686, + 0.023409215733408928, + 0.013207546435296535, + 0.02501191757619381, + 0.018412860110402107, + 0.020748833194375038, + 0.010942856781184673, + 0.0023360494524240494, + 0.002983585000038147, + 0.004906385205686092, + 0.0033030076883733273, + 0.0043053943663835526, + 0.004616935737431049, + 0.0033129663206636906, + 0.006684682331979275, + 0.004826510790735483, + 0.01273765042424202, + 0.010686948895454407, + 0.006039723288267851, + 0.011640039272606373, + 0.005642722360789776, + 0.0050226603634655476, + 0.010524912737309933, + 0.014884622767567635, + 0.022749582305550575, + 0.029635252431035042, + 0.014395579695701599, + 0.03677701577544212, + 0.020072242245078087, + 0.015673931688070297, + 0.022988708689808846, + 0.007378573063760996, + 0.007939238101243973, + 0.013902494683861732, + 0.011772437021136284, + 0.01529760006815195, + 0.009596629068255424, + 0.009439456276595592, + 0.014242822304368019, + 0.011859667487442493, + 0.006779845338314772, + 0.06764461100101471, + 0.007073718123137951, + 0.05422394722700119, + 0.010027093812823296, + 0.019065571948885918, + 0.0045845420099794865, + 0.008276021108031273, + 0.007612365763634443, + 0.006348620634526014, + 0.003812741953879595, + 0.003880293108522892, + 0.007616471033543348, + 0.007524378132075071, + 0.003883026773110032, + 0.007899691350758076, + 0.010260529816150665, + 0.007792332675307989, + 0.006812013685703278, + 0.006015211809426546, + 0.007019544020295143, + 0.0067278072237968445, + 0.005339618772268295, + 0.014284173958003521, + 0.01969519443809986, + 0.011679410934448242, + 0.014376416802406311, + 0.017696982249617577, + 0.022499576210975647, + 0.016581257805228233, + 0.013681906275451183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.5761618614196777 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands shown in a steady close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands described as calloused, rough-skinned, with prominent knuckles, deep veins, and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hands gripping a wooden chisel with a dark brown handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hands carving an ancient, weathered, dark gray stone tablet with a knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031366348266602 + ], + "opposite_logprob_tokens": [ + -2.350313663482666 + ], + "logprob_reward": -0.10031366348266602, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031366348266602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056663475930690765, + 0.06355790793895721, + 0.05934220179915428, + 0.06380002945661545, + 0.3351622521877289, + 0.06720232963562012, + 0.15121030807495117, + 0.033435553312301636, + 0.10474663227796555, + 0.06487933546304703 + ], + "credit_map": [ + 0.0064868261106312275, + 0.006759993731975555, + 0.0063145579770207405, + 0.00727713480591774, + 0.009241152554750443, + 0.007329599466174841, + 0.0073647769168019295, + 0.005889429710805416, + 0.00590257765725255, + 0.015530971810221672, + 0.0062795053236186504, + 0.004894225858151913, + 0.016149351373314857, + 0.0046370928175747395, + 0.007017645984888077, + 0.0031465317588299513, + 0.004379624966531992, + 0.004780360031872988, + 0.008494626730680466, + 0.007298736367374659, + 0.006616231985390186, + 0.007250034250319004, + 0.007100438699126244, + 0.013422149233520031, + 0.0053112399764359, + 0.008135898038744926, + 0.008519536815583706, + 0.0052881669253110886, + 0.01062464714050293, + 0.0069229197688400745, + 0.008213385939598083, + 0.010784230194985867, + 0.02894928865134716, + 0.044866207987070084, + 0.05225281044840813, + 0.02872045338153839, + 0.0692218616604805, + 0.03698136657476425, + 0.031168535351753235, + 0.04300171509385109, + 0.006153131369501352, + 0.007146000396460295, + 0.009899311698973179, + 0.00907648541033268, + 0.01215845625847578, + 0.009501307271420956, + 0.005555455107241869, + 0.007712181191891432, + 0.010996750555932522, + 0.011447039432823658, + 0.024053068831562996, + 0.005919900257140398, + 0.04741109535098076, + 0.012805451638996601, + 0.01655997522175312, + 0.022017033770680428, + 0.004840463399887085, + 0.003889268497005105, + 0.0030360324308276176, + 0.0049008834175765514, + 0.0033459754195064306, + 0.005713112652301788, + 0.00406400253996253, + 0.003645813325420022, + 0.014511278830468655, + 0.014998343773186207, + 0.01221552211791277, + 0.008359195664525032, + 0.01363667193800211, + 0.01598600670695305, + 0.010155759751796722, + 0.014883851632475853, + 0.0062929196283221245, + 0.009729765355587006, + 0.006818078923970461, + 0.008542061783373356, + 0.010651206597685814, + 0.009171226061880589, + 0.0058022430166602135, + 0.007871839217841625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.10031366348266602 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_12/index_tensor([6441, 6441])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_12/index_tensor([6441, 6441])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2518ccdbfe60e3fbd8801ecbb5ef3189b0031146 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_12/index_tensor([6441, 6441])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Static close-ups, framed as physics-based observational macro shots, capture the mesmerizing process within the clear, cylindrical glass shutters. Initially, the water is a turbulent, opaque cloud of rich, dark brown, studded with fine coffee grounds, swirling like a suspended miniature galaxy. Soft, even kitchen light, diffused from an unseen source, illuminates the scene without harsh shadows, capturing the delicate glint of individual microscopic particles. The invisible yet persistent force of gravity takes over, and as each particle steadily drifts downward, a slow, deliberate descent begins. As they cautiously accumulate at the bottom of the glass, they form a distinct, dense deposit of dark amber. A sharp, ever-clarifying separating line emerges, separating solid from liquid, creating a stark visual boundary that rises as the water above gradually transforms from a murky brown to a translucent, warm amber, creating an atmosphere of tranquility and concentration. The details of observed natural phenomena are captured.", + "seed": 1666, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_12/index_6441_seed_1666_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02484555037226528, + "vlm_logprob": -0.02484555037226528, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.02484555037226528, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02019043080508709 + ], + "opposite_logprob_tokens": [ + -5.520190238952637 + ], + "logprob_reward": -0.02019043080508709, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.02019043080508709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26975831389427185, + 0.03941691294312477, + 0.02307884767651558, + 0.025045156478881836, + 0.4992644488811493, + 0.02688588574528694, + 0.01765441708266735, + 0.04663902148604393, + 0.04198062792420387, + 0.010276366025209427 + ], + "credit_map": [ + 0.012585136108100414, + 0.007298920303583145, + 0.010857454501092434, + 0.02362161874771118, + 0.025165367871522903, + 0.018636150285601616, + 0.09691577404737473, + 0.07467789947986603, + 0.004108313005417585, + 0.0032589028123766184, + 0.002494706306606531, + 0.0033063762821257114, + 0.012291774153709412, + 0.003100800095126033, + 0.0036880364641547203, + 0.007168002892285585, + 0.0019494645530357957, + 0.006496625021100044, + 0.0018001688877120614, + 0.0026702615432441235, + 0.004337233025580645, + 0.0018444538582116365, + 0.0026660659350454807, + 0.0013145763659849763, + 0.0026781796477735043, + 0.0035898422356694937, + 0.0026189489290118217, + 0.00358632137067616, + 0.004974272567778826, + 0.0018739675870165229, + 0.003625919809564948, + 0.002097703516483307, + 0.0233071930706501, + 0.05925443395972252, + 0.07976167649030685, + 0.04819954186677933, + 0.18735100328922272, + 0.01899227872490883, + 0.06488879770040512, + 0.017509490251541138, + 0.0023398783523589373, + 0.0031940448097884655, + 0.0042112176306545734, + 0.003747663227841258, + 0.0028407659847289324, + 0.0028444938361644745, + 0.005717095918953419, + 0.0019907252863049507, + 0.0020877839997410774, + 0.0015495928237214684, + 0.0023446856066584587, + 0.0030519147403538227, + 0.0015741030219942331, + 0.0020478866063058376, + 0.0029567463789135218, + 0.002041705185547471, + 0.005609448533505201, + 0.011029206216335297, + 0.002461883472278714, + 0.005238221026957035, + 0.006527075543999672, + 0.005747729912400246, + 0.0066764336079359055, + 0.003349023638293147, + 0.002751322230324149, + 0.011202621273696423, + 0.0024175874423235655, + 0.011375241912901402, + 0.0025111769791692495, + 0.003761656815186143, + 0.006516986060887575, + 0.0014440339291468263, + 0.0008071461925283074, + 0.001383297611027956, + 0.000745985540561378, + 0.002199823036789894, + 0.0015503749018535018, + 0.0009981479961425066, + 0.0018198585603386164, + 0.0007717317203059793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.02019043080508709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052833217196166515 + ], + "opposite_logprob_tokens": [ + -5.255283355712891 + ], + "logprob_reward": -0.0052833217196166515, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052833217196166515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16584806144237518, + 0.03732810169458389, + 0.11754978448152542, + 0.03404614329338074, + 0.4190486967563629, + 0.010581604205071926, + 0.03087461180984974, + 0.15457072854042053, + 0.024587048217654228, + 0.005565257743000984 + ], + "credit_map": [ + 0.009221151471138, + 0.0052003939636051655, + 0.005924735683947802, + 0.010252724401652813, + 0.02557527832686901, + 0.010747057385742664, + 0.037622299045324326, + 0.06130440905690193, + 0.002631294308230281, + 0.002216224791482091, + 0.004569253884255886, + 0.0046780710108578205, + 0.006466289982199669, + 0.002603463362902403, + 0.005138383712619543, + 0.009025119245052338, + 0.005555376410484314, + 0.00926470011472702, + 0.0038471303414553404, + 0.03465105965733528, + 0.014202550053596497, + 0.006260708440095186, + 0.0378958098590374, + 0.005872434005141258, + 0.0024821991100907326, + 0.005088575650006533, + 0.0030713339801877737, + 0.00453668087720871, + 0.007451818324625492, + 0.00296345679089427, + 0.005643411073833704, + 0.0028086663223803043, + 0.01841908134520054, + 0.050730425864458084, + 0.06460363417863846, + 0.03623884171247482, + 0.16185247898101807, + 0.015850994735956192, + 0.058030229061841965, + 0.013322962448000908, + 0.001048122183419764, + 0.0013961612712591887, + 0.001277437317185104, + 0.0014116971287876368, + 0.0012927623465657234, + 0.0013111792504787445, + 0.0019402207108214498, + 0.0009040224249474704, + 0.00223542470484972, + 0.0026043409015983343, + 0.0024070865474641323, + 0.007460202556103468, + 0.0026008221320807934, + 0.0029668351635336876, + 0.008492754772305489, + 0.0021071427036076784, + 0.01708787865936756, + 0.022874660789966583, + 0.00956797320395708, + 0.02781107649207115, + 0.02122518979012966, + 0.017516111955046654, + 0.026395078748464584, + 0.012092738412320614, + 0.0015396851813420653, + 0.006286529824137688, + 0.0013424636563286185, + 0.0063415938057005405, + 0.0019921218045055866, + 0.0019690373446792364, + 0.004063078667968512, + 0.001052536303177476, + 0.0005093345534987748, + 0.0005444416892714798, + 0.0006423586164601147, + 0.0010881844209507108, + 0.0008114376687444746, + 0.0005161736626178026, + 0.0010002615163102746, + 0.0004530650330707431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.0052833217196166515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010521834483370185 + ], + "opposite_logprob_tokens": [ + -6.876052379608154 + ], + "logprob_reward": -0.0010521834483370185, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010521834483370185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15664251148700714, + 0.03016630746424198, + 0.05295279249548912, + 0.017024001106619835, + 0.2823401987552643, + 0.09900838136672974, + 0.05836034193634987, + 0.11716138571500778, + 0.1556505560874939, + 0.030693527311086655 + ], + "credit_map": [ + 0.006129100918769836, + 0.0040711755864322186, + 0.006694549694657326, + 0.013676170259714127, + 0.02026115357875824, + 0.011605904437601566, + 0.056827057152986526, + 0.037377405911684036, + 0.0019868214149028063, + 0.0028156081680208445, + 0.00167462311219424, + 0.0035851462744176388, + 0.009878970682621002, + 0.0023693928960710764, + 0.0032415352761745453, + 0.004614211153239012, + 0.003787136869505048, + 0.01457343902438879, + 0.002730214037001133, + 0.006694698240607977, + 0.009559846483170986, + 0.003513241885229945, + 0.009351827204227448, + 0.0027423850260674953, + 0.0013548423303291202, + 0.002551287179812789, + 0.0018642791546881199, + 0.0030802825931459665, + 0.0022442759945988655, + 0.0015962154138833284, + 0.002825103932991624, + 0.0015077157877385616, + 0.011563219130039215, + 0.035961657762527466, + 0.03776507079601288, + 0.02224041149020195, + 0.11200451105833054, + 0.010967819951474667, + 0.0430481992661953, + 0.008789291605353355, + 0.008079315535724163, + 0.0115983160212636, + 0.01242715772241354, + 0.01633412390947342, + 0.01015356183052063, + 0.01107377465814352, + 0.02002900280058384, + 0.009313127025961876, + 0.004606717731803656, + 0.004102024249732494, + 0.003992672543972731, + 0.015141710638999939, + 0.0060652270913124084, + 0.004661949817091227, + 0.016666460782289505, + 0.0031235809437930584, + 0.01363685168325901, + 0.02226216346025467, + 0.007093330845236778, + 0.015844101086258888, + 0.016629787161946297, + 0.014603248797357082, + 0.017547450959682465, + 0.009544449858367443, + 0.009991365484893322, + 0.04138828068971634, + 0.007824384607374668, + 0.04315862059593201, + 0.009922203607857227, + 0.013188553974032402, + 0.025228222832083702, + 0.004948926158249378, + 0.0016251569613814354, + 0.003724752925336361, + 0.0016534721944481134, + 0.007628600113093853, + 0.0043458337895572186, + 0.0015668530249968171, + 0.008492589928209782, + 0.0016562704695388675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0010521834483370185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647849783301353 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647849783301353, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647849783301353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12768982350826263, + 0.014102771878242493, + 0.0618838369846344, + 0.015154844149947166, + 0.6020923256874084, + 0.03281071409583092, + 0.03881560638546944, + 0.024929730221629143, + 0.06536289304494858, + 0.017157427966594696 + ], + "credit_map": [ + 0.007835456170141697, + 0.003982414025813341, + 0.005240901373326778, + 0.006785224191844463, + 0.03001239337027073, + 0.008532709442079067, + 0.02075599692761898, + 0.04454473778605461, + 0.0011549446498975158, + 0.0011102990247309208, + 0.0007690454367548227, + 0.002246737014502287, + 0.003878276329487562, + 0.0009404594311490655, + 0.002886617323383689, + 0.0011163937160745263, + 0.004346943460404873, + 0.016487954184412956, + 0.004518390167504549, + 0.008746063336730003, + 0.011956162750720978, + 0.004542483016848564, + 0.008328470401465893, + 0.002957370597869158, + 0.0010550129227340221, + 0.0022003690246492624, + 0.001209354610182345, + 0.0023026224225759506, + 0.002864267211407423, + 0.0011295005679130554, + 0.0027014531660825014, + 0.001692265272140503, + 0.026688244193792343, + 0.07255763560533524, + 0.0922323688864708, + 0.05280470848083496, + 0.2316466122865677, + 0.023486847057938576, + 0.08325043320655823, + 0.019425498321652412, + 0.00238371011801064, + 0.0020959412213414907, + 0.0062546078115701675, + 0.005711815319955349, + 0.002288166666403413, + 0.0015386117156594992, + 0.010374242439866066, + 0.002163621364161372, + 0.003564093727618456, + 0.0032891046721488237, + 0.0030339749064296484, + 0.008514287881553173, + 0.004467834252864122, + 0.003720279084518552, + 0.009496848098933697, + 0.0027291837614029646, + 0.002591383643448353, + 0.007038929499685764, + 0.0013482562499120831, + 0.0034150450956076384, + 0.00275496207177639, + 0.0028578045312315226, + 0.003631833242252469, + 0.0012915180996060371, + 0.004017141181975603, + 0.016527734696865082, + 0.0036547628697007895, + 0.018933361396193504, + 0.0031655568163841963, + 0.005411723628640175, + 0.011312922462821007, + 0.002339694881811738, + 0.001224722363986075, + 0.0019907623063772917, + 0.001163374981842935, + 0.004105659667402506, + 0.002161441370844841, + 0.0009576426818966866, + 0.004471896216273308, + 0.001081929192878306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.008647849783301353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905396610498428 + ], + "opposite_logprob_tokens": [ + -2.4640538692474365 + ], + "logprob_reward": -0.08905396610498428, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905396610498428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17379894852638245, + 0.021816547960042953, + 0.051278382539749146, + 0.030546654015779495, + 0.5738767385482788, + 0.05982294678688049, + 0.026638610288500786, + 0.040096528828144073, + 0.013922112062573433, + 0.008202635683119297 + ], + "credit_map": [ + 0.011012249626219273, + 0.006085935048758984, + 0.00624591251835227, + 0.00787096843123436, + 0.04560096561908722, + 0.011838284321129322, + 0.022621944546699524, + 0.06252267211675644, + 0.006359807215631008, + 0.0013546705013141036, + 0.001959799090400338, + 0.001444052322767675, + 0.003038566093891859, + 0.001666327123530209, + 0.0018844279693439603, + 0.004108896013349295, + 0.004168789368122816, + 0.0136561905965209, + 0.0028250617906451225, + 0.007147424854338169, + 0.008288945071399212, + 0.0038469440769404173, + 0.008589861914515495, + 0.002755162538960576, + 0.0023332126438617706, + 0.005245101638138294, + 0.002097511198371649, + 0.0048597645945847034, + 0.005842843092978001, + 0.002972773741930723, + 0.005006982944905758, + 0.002188462531194091, + 0.022569699212908745, + 0.0701192170381546, + 0.08028972148895264, + 0.047465138137340546, + 0.2254565805196762, + 0.022665157914161682, + 0.08837441354990005, + 0.016936693340539932, + 0.0051016719080507755, + 0.0074364664033055305, + 0.005836922209709883, + 0.011034932918846607, + 0.007072432432323694, + 0.006941565312445164, + 0.010948119685053825, + 0.005450830794870853, + 0.0027219061739742756, + 0.004559243563562632, + 0.0018089916557073593, + 0.005084374453872442, + 0.0024342609103769064, + 0.002436137991026044, + 0.0058241114020347595, + 0.0017695812275633216, + 0.004160746466368437, + 0.006289775483310223, + 0.0025011051911860704, + 0.007066105958074331, + 0.005217260681092739, + 0.005049018654972315, + 0.006547506898641586, + 0.0032650032080709934, + 0.001206536078825593, + 0.0029826543759554625, + 0.00124660087749362, + 0.0038574470672756433, + 0.0005672439583577216, + 0.0011755909072235227, + 0.0021456636022776365, + 0.0007403745548799634, + 0.0007997474167495966, + 0.0012562507763504982, + 0.0012034953106194735, + 0.0011337058385834098, + 0.0009582144557498395, + 0.0007095968467183411, + 0.0014299281174317002, + 0.0007116959313862026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.08905396610498428 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static close-ups, framed as physics-based observational macro shots, capture the mesmerizing process within the clear, cylindrical glass shutters. Initially, the water is a turbulent, opaque cloud of rich, dark brown, studded with fine coffee grounds, swirling like a suspended miniature galaxy. Soft, even kitchen light, diffused from an unseen source, illuminates the scene without harsh shadows, capturing the delicate glint of individual microscopic particles. The invisible yet persistent force of gravity takes over, and as each particle steadily drifts downward, a slow, deliberate descent begins. As they cautiously accumulate at the bottom of the glass, they form a distinct, dense deposit of dark amber. A sharp, ever-clarifying separating line emerges, separating solid from liquid, creating a stark visual boundary that rises as the water above gradually transforms from a murky brown to a translucent, warm amber, creating an atmosphere of tranquility and concentration. The details of observed natural phenomena are captured.", + "seed": 1667, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_12/index_6441_seed_1667_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02059371683280915, + "vlm_logprob": -0.02059371683280915, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.02059371683280915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0295708104968071 + ], + "opposite_logprob_tokens": [ + -5.779570579528809 + ], + "logprob_reward": -0.0295708104968071, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0295708104968071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2679122984409332, + 0.1694285124540329, + 0.14266827702522278, + 0.027229025959968567, + 0.0316116064786911, + 0.054404448717832565, + 0.07349292188882828, + 0.12833398580551147, + 0.04509393125772476, + 0.059824999421834946 + ], + "credit_map": [ + 0.02576150745153427, + 0.03900879621505737, + 0.04307638481259346, + 0.03806234896183014, + 0.032343365252017975, + 0.028846990317106247, + 0.041621752083301544, + 0.019191164523363113, + 0.011991185136139393, + 0.0331912487745285, + 0.009851617738604546, + 0.02767217345535755, + 0.021335307508707047, + 0.015112250111997128, + 0.03532653674483299, + 0.014948176220059395, + 0.006641268730163574, + 0.017405081540346146, + 0.0062894378788769245, + 0.03019225411117077, + 0.018760716542601585, + 0.008100518025457859, + 0.04327448457479477, + 0.012004523538053036, + 0.0028405236080288887, + 0.004011000040918589, + 0.002902344334870577, + 0.0049940356984734535, + 0.002614579861983657, + 0.002280023880302906, + 0.004537057597190142, + 0.0030494602397084236, + 0.0029885016847401857, + 0.003307582577690482, + 0.005440453998744488, + 0.0041809361428022385, + 0.003049373859539628, + 0.0034711207263171673, + 0.003959340043365955, + 0.0052142976783216, + 0.004409940913319588, + 0.006590256467461586, + 0.005672085098922253, + 0.01127604953944683, + 0.0048042587004601955, + 0.006576515734195709, + 0.008656065911054611, + 0.006419277749955654, + 0.007594511844217777, + 0.011201715096831322, + 0.009960230439901352, + 0.00947369821369648, + 0.010779245756566525, + 0.005792251322418451, + 0.00871670339256525, + 0.00997456256300211, + 0.009043648838996887, + 0.02496297098696232, + 0.008081925101578236, + 0.02060864306986332, + 0.02894260175526142, + 0.005232592578977346, + 0.020916422829031944, + 0.010545188561081886, + 0.004286486655473709, + 0.005307959392666817, + 0.004759296774864197, + 0.008521541953086853, + 0.006548809353262186, + 0.0036372088361531496, + 0.008167851716279984, + 0.0038647782057523727, + 0.004729587584733963, + 0.009309190325438976, + 0.004798651672899723, + 0.011200995184481144, + 0.011347946710884571, + 0.004424592945724726, + 0.009348887018859386, + 0.004665148910135031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0295708104968071 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012581500224769115 + ], + "opposite_logprob_tokens": [ + -4.3875813484191895 + ], + "logprob_reward": -0.012581500224769115, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012581500224769115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09051354974508286, + 0.2847636342048645, + 0.33487778902053833, + 0.02989940531551838, + 0.05841802433133125, + 0.04049275442957878, + 0.03778638318181038, + 0.03543449565768242, + 0.03451218456029892, + 0.05330178141593933 + ], + "credit_map": [ + 0.006838376168161631, + 0.010569668374955654, + 0.013185878284275532, + 0.015322287566959858, + 0.010018430650234222, + 0.007383563555777073, + 0.016501611098647118, + 0.010693741030991077, + 0.01357371173799038, + 0.03225476294755936, + 0.017739901319146156, + 0.07384189963340759, + 0.027390632778406143, + 0.01989818550646305, + 0.0796724334359169, + 0.020392149686813354, + 0.01586020365357399, + 0.03467268496751785, + 0.013315358199179173, + 0.07209045439958572, + 0.038917481899261475, + 0.02183644473552704, + 0.10606172680854797, + 0.03212343156337738, + 0.002151338616386056, + 0.003839643904939294, + 0.002292257733643055, + 0.0055605401284992695, + 0.004119005985558033, + 0.003858376294374466, + 0.005051438231021166, + 0.003026807913556695, + 0.004462852142751217, + 0.007320781238377094, + 0.005711676552891731, + 0.012883181683719158, + 0.008096824400126934, + 0.0050763925537467, + 0.01019328087568283, + 0.004673038609325886, + 0.0026283778715878725, + 0.0037559138145297766, + 0.003524686209857464, + 0.012378474697470665, + 0.0033608735539019108, + 0.004355947952717543, + 0.007310173008590937, + 0.0031783077865839005, + 0.004216090310364962, + 0.005842122714966536, + 0.0037300700787454844, + 0.006396964192390442, + 0.006240033078938723, + 0.003223913023248315, + 0.005118072964251041, + 0.0030191191472113132, + 0.00226625450886786, + 0.006517179775983095, + 0.002672667847946286, + 0.006094388663768768, + 0.006752578541636467, + 0.0017075072973966599, + 0.00555406603962183, + 0.003869857871904969, + 0.001995101571083069, + 0.0042883227579295635, + 0.0024813851341605186, + 0.008125768974423409, + 0.004616068210452795, + 0.00179827690590173, + 0.0070076328702270985, + 0.004199630115181208, + 0.0021952746901661158, + 0.006207118276506662, + 0.003817165270447731, + 0.01371006853878498, + 0.007206437643617392, + 0.002465563826262951, + 0.011645297519862652, + 0.006054860074073076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.012581500224769115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011928117601200938 + ], + "opposite_logprob_tokens": [ + -6.751193046569824 + ], + "logprob_reward": -0.0011928117601200938, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011928117601200938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04640720412135124, + 0.12934435904026031, + 0.2738010883331299, + 0.021799786016345024, + 0.045206211507320404, + 0.0625913143157959, + 0.08018140494823456, + 0.07634985446929932, + 0.04700472205877304, + 0.21731403470039368 + ], + "credit_map": [ + 0.005321632605046034, + 0.006330534815788269, + 0.007751933764666319, + 0.004978765733540058, + 0.0057544712908566, + 0.005796168465167284, + 0.0055386656895279884, + 0.004935034085065126, + 0.02340546064078808, + 0.012391512282192707, + 0.03050805814564228, + 0.016999507322907448, + 0.009409690275788307, + 0.00894182175397873, + 0.020498059689998627, + 0.007190254051238298, + 0.013492568396031857, + 0.03706645220518112, + 0.009917187504470348, + 0.050397634506225586, + 0.040115177631378174, + 0.02105807140469551, + 0.07209312170743942, + 0.029660886153578758, + 0.0024273409508168697, + 0.002116955118253827, + 0.0024760460946708918, + 0.0045114546082913876, + 0.0019030501134693623, + 0.0020003970712423325, + 0.0037926994264125824, + 0.0025718428660184145, + 0.003511794377118349, + 0.005576177965849638, + 0.004276622086763382, + 0.009466052055358887, + 0.006728133652359247, + 0.00411729933694005, + 0.00711007509380579, + 0.004420061130076647, + 0.005513459909707308, + 0.014436068944633007, + 0.004372163210064173, + 0.009268640540540218, + 0.01085147075355053, + 0.006366166286170483, + 0.006829644553363323, + 0.004953693598508835, + 0.006846616044640541, + 0.015451353043317795, + 0.006035027094185352, + 0.009552082978188992, + 0.018080154433846474, + 0.006082610692828894, + 0.008711612783372402, + 0.009421948343515396, + 0.004720110911875963, + 0.013094586320221424, + 0.004251518286764622, + 0.014230066910386086, + 0.014802233316004276, + 0.003605971112847328, + 0.015500808134675026, + 0.0061445580795407295, + 0.005395989399403334, + 0.004630059003829956, + 0.007499696221202612, + 0.00981038436293602, + 0.005148471333086491, + 0.0019182867836207151, + 0.007808903232216835, + 0.004792928230017424, + 0.011278122663497925, + 0.027202283963561058, + 0.015532882884144783, + 0.049146126955747604, + 0.031278543174266815, + 0.01185129676014185, + 0.05044512823224068, + 0.020579643547534943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.0011928117601200938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647050518542528 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647050518542528, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647050518542528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0737917348742485, + 0.08355412632226944, + 0.16975857317447662, + 0.03122163936495781, + 0.04474995657801628, + 0.07897292822599411, + 0.149358332157135, + 0.14490906894207, + 0.03853927552700043, + 0.18514440953731537 + ], + "credit_map": [ + 0.005594910588115454, + 0.007848377339541912, + 0.009909703396260738, + 0.0139625770971179, + 0.007143492344766855, + 0.0075879269279539585, + 0.012762295082211494, + 0.008982446976006031, + 0.009962721727788448, + 0.007908599451184273, + 0.01287410594522953, + 0.01560690812766552, + 0.005858954042196274, + 0.00601982744410634, + 0.019064651802182198, + 0.006258357781916857, + 0.00929549615830183, + 0.02004525624215603, + 0.00957509595900774, + 0.03343002125620842, + 0.01943209208548069, + 0.010855608619749546, + 0.049129847437143326, + 0.017995154485106468, + 0.003024438861757517, + 0.0033972375094890594, + 0.002990785287693143, + 0.005228585563600063, + 0.00275025493465364, + 0.0029613065999001265, + 0.005689093377441168, + 0.005179939325898886, + 0.0037965700030326843, + 0.005691584665328264, + 0.0050294469110667706, + 0.008657450787723064, + 0.0068181175738573074, + 0.003691501682624221, + 0.0068087805993855, + 0.004256502725183964, + 0.005872040521353483, + 0.014051019214093685, + 0.006593008991330862, + 0.014798137359321117, + 0.010931930504739285, + 0.008397014811635017, + 0.011420668102800846, + 0.006909109652042389, + 0.01464990433305502, + 0.031033942475914955, + 0.011673151515424252, + 0.017678827047348022, + 0.03727862611413002, + 0.008434729650616646, + 0.017852190881967545, + 0.01075696013867855, + 0.008012734353542328, + 0.02424239180982113, + 0.009271547198295593, + 0.02856878563761711, + 0.025744983926415443, + 0.005763279274106026, + 0.02980182133615017, + 0.01350353378802538, + 0.0029559065587818623, + 0.0043633985333144665, + 0.005287775304168463, + 0.009139180183410645, + 0.004696462769061327, + 0.0017087777378037572, + 0.007518154568970203, + 0.0028696157969534397, + 0.00805194117128849, + 0.022674011066555977, + 0.011466839350759983, + 0.04801998287439346, + 0.025825046002864838, + 0.009275750257074833, + 0.038448285311460495, + 0.021382559090852737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.004647050518542528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054976411163806915 + ], + "opposite_logprob_tokens": [ + -2.929976463317871 + ], + "logprob_reward": -0.054976411163806915, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054976411163806915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2262152135372162, + 0.17176665365695953, + 0.10179342329502106, + 0.0509001724421978, + 0.08124380558729172, + 0.056396085768938065, + 0.03627704456448555, + 0.1426043063402176, + 0.05498760566115379, + 0.07781565934419632 + ], + "credit_map": [ + 0.02597218006849289, + 0.02624712511897087, + 0.042589373886585236, + 0.03573402762413025, + 0.021027516573667526, + 0.020497793331742287, + 0.03549565002322197, + 0.01865154691040516, + 0.011487050913274288, + 0.017397336661815643, + 0.015540930442512035, + 0.03819604590535164, + 0.018069390207529068, + 0.016733882948756218, + 0.038535065948963165, + 0.015806952491402626, + 0.007098236586898565, + 0.018048327416181564, + 0.007303124293684959, + 0.013226441107690334, + 0.013086897321045399, + 0.01012315321713686, + 0.01687966100871563, + 0.01602758653461933, + 0.003595296060666442, + 0.007027459796518087, + 0.00332624651491642, + 0.011286494322121143, + 0.005929275415837765, + 0.005119467619806528, + 0.008265347220003605, + 0.006350586656481028, + 0.006273570470511913, + 0.010281989350914955, + 0.008043280802667141, + 0.016481660306453705, + 0.01090661995112896, + 0.008544052951037884, + 0.014121489599347115, + 0.0065911454148590565, + 0.0037223680410534143, + 0.010142428800463676, + 0.004894760902971029, + 0.01149181742221117, + 0.007677941117435694, + 0.004917648620903492, + 0.008579926565289497, + 0.004969197325408459, + 0.002186054829508066, + 0.005391772370785475, + 0.0040024397894740105, + 0.004739188123494387, + 0.005946195684373379, + 0.004853478632867336, + 0.0036853442434221506, + 0.005472566932439804, + 0.007369486149400473, + 0.016193881630897522, + 0.011585075408220291, + 0.033027902245521545, + 0.01687321998178959, + 0.007064634934067726, + 0.03587410971522331, + 0.014615981839597225, + 0.0029034644830971956, + 0.008100620470941067, + 0.003734139259904623, + 0.012021731585264206, + 0.00933192577213049, + 0.003314262256026268, + 0.010306441225111485, + 0.0052750203758478165, + 0.004694386385381222, + 0.011025871150195599, + 0.0045301103964447975, + 0.014731843955814838, + 0.01040936540812254, + 0.005083893891423941, + 0.01672586053609848, + 0.010614335536956787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.054976411163806915 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static close-ups, framed as physics-based observational macro shots, capture the mesmerizing process within the clear, cylindrical glass shutters. Initially, the water is a turbulent, opaque cloud of rich, dark brown, studded with fine coffee grounds, swirling like a suspended miniature galaxy. Soft, even kitchen light, diffused from an unseen source, illuminates the scene without harsh shadows, capturing the delicate glint of individual microscopic particles. The invisible yet persistent force of gravity takes over, and as each particle steadily drifts downward, a slow, deliberate descent begins. As they cautiously accumulate at the bottom of the glass, they form a distinct, dense deposit of dark amber. A sharp, ever-clarifying separating line emerges, separating solid from liquid, creating a stark visual boundary that rises as the water above gradually transforms from a murky brown to a translucent, warm amber, creating an atmosphere of tranquility and concentration. The details of observed natural phenomena are captured.", + "seed": 1668, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_12/index_6441_seed_1668_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02565528154373169, + "vlm_logprob": -0.02565528154373169, + "vlm_yes_no_margin": 4.224999952316284, + "vlm_token_credit": -0.02565528154373169, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02372029982507229 + ], + "opposite_logprob_tokens": [ + -4.8987202644348145 + ], + "logprob_reward": -0.02372029982507229, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.02372029982507229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08450419455766678, + 0.03655431792140007, + 0.05452863126993179, + 0.026327911764383316, + 0.023632409051060677, + 0.024730976670980453, + 0.025951974093914032, + 0.07139285653829575, + 0.04553008824586868, + 0.6068466901779175 + ], + "credit_map": [ + 0.008449656888842583, + 0.003904770128428936, + 0.010600745677947998, + 0.023373868316411972, + 0.007093219086527824, + 0.004201273899525404, + 0.01961548626422882, + 0.007265172898769379, + 0.003344481810927391, + 0.004514469765126705, + 0.0038545485585927963, + 0.006217877380549908, + 0.006597782019525766, + 0.002530591795220971, + 0.006369993556290865, + 0.003124573267996311, + 0.0028653366025537252, + 0.004731499589979649, + 0.0038228794001042843, + 0.016707561910152435, + 0.005718712694942951, + 0.003494038013741374, + 0.014017164707183838, + 0.0031714360229671, + 0.004966031294316053, + 0.007067814003676176, + 0.0022861841134727, + 0.004688667133450508, + 0.0030692387372255325, + 0.0009442392620258033, + 0.0023589464835822582, + 0.0009467917261645198, + 0.0033753272145986557, + 0.0037624745164066553, + 0.003687998978421092, + 0.0035509292501956224, + 0.002692640293389559, + 0.0016517549520358443, + 0.0030479407869279385, + 0.0018633445724844933, + 0.0015866939211264253, + 0.005588998552411795, + 0.001931328559294343, + 0.002053310163319111, + 0.008964231237769127, + 0.0012860988499596715, + 0.0024290618021041155, + 0.0008912526536732912, + 0.002396830590441823, + 0.0044166408479213715, + 0.0028864056803286076, + 0.003649052930995822, + 0.0026582996360957623, + 0.002478739945217967, + 0.004355308599770069, + 0.0031106972601264715, + 0.005087088327854872, + 0.010084308683872223, + 0.007969165220856667, + 0.010204502381384373, + 0.01126339752227068, + 0.006298705004155636, + 0.011724882759153843, + 0.008760806173086166, + 0.0034517396707087755, + 0.0078496802598238, + 0.005293871276080608, + 0.00603749742731452, + 0.008055499754846096, + 0.003596703987568617, + 0.007267430890351534, + 0.003977664280682802, + 0.009801666252315044, + 0.2826090455055237, + 0.004282149951905012, + 0.004730432294309139, + 0.28390130400657654, + 0.011511112563312054, + 0.00643613887950778, + 0.0035748013760894537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.02372029982507229 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018222538754343987 + ], + "opposite_logprob_tokens": [ + -4.018222332000732 + ], + "logprob_reward": -0.018222538754343987, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018222538754343987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1779329925775528, + 0.12917208671569824, + 0.10599013417959213, + 0.025674968957901, + 0.058362603187561035, + 0.020532628521323204, + 0.05879945307970047, + 0.01786837726831436, + 0.012789707630872726, + 0.39287710189819336 + ], + "credit_map": [ + 0.012205203995108604, + 0.00943074468523264, + 0.028204621747136116, + 0.038269396871328354, + 0.018261637538671494, + 0.016063934192061424, + 0.04150265082716942, + 0.013994792476296425, + 0.013190596364438534, + 0.015663841739296913, + 0.01185121200978756, + 0.02763226628303528, + 0.02004951797425747, + 0.00809362344443798, + 0.02473686821758747, + 0.00795414112508297, + 0.00593981659039855, + 0.018773753196001053, + 0.009424794465303421, + 0.027302728965878487, + 0.015448481775820255, + 0.005430896766483784, + 0.0185526255518198, + 0.005117020569741726, + 0.0029776867013424635, + 0.005429643206298351, + 0.002761821961030364, + 0.00411336962133646, + 0.003612686414271593, + 0.001318821799941361, + 0.004191019106656313, + 0.0012699157232418656, + 0.0039884368889033794, + 0.010030128993093967, + 0.005741463974118233, + 0.0052591594867408276, + 0.0127977654337883, + 0.006258679553866386, + 0.008195994421839714, + 0.006090969778597355, + 0.001935229985974729, + 0.003421947592869401, + 0.0029031347949057817, + 0.002407341729849577, + 0.005062946118414402, + 0.0013758705463260412, + 0.001682815607637167, + 0.0017433414468541741, + 0.005798744037747383, + 0.010755647905170918, + 0.006354661658406258, + 0.00865701213479042, + 0.008636407554149628, + 0.0043755569495260715, + 0.00749633926898241, + 0.006725077051669359, + 0.0016914334846660495, + 0.002409586915746331, + 0.0022457973100245, + 0.002918123733252287, + 0.002646917710080743, + 0.0018270692089572549, + 0.002501734532415867, + 0.001627713325433433, + 0.001173327211290598, + 0.002068157307803631, + 0.001340461545623839, + 0.0017222703900188208, + 0.0023694601841270924, + 0.0008949613547883928, + 0.0016525641549378633, + 0.001568503794260323, + 0.006654667668044567, + 0.18162761628627777, + 0.0031109836418181658, + 0.00431777024641633, + 0.18216310441493988, + 0.007379875052720308, + 0.0049472288228571415, + 0.002675815485417843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.018222538754343987 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656476624310017 + ], + "opposite_logprob_tokens": [ + -4.7586565017700195 + ], + "logprob_reward": -0.008656476624310017, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656476624310017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05119436979293823, + 0.022400973364710808, + 0.04774454981088638, + 0.024617953225970268, + 0.04305270314216614, + 0.018962541595101357, + 0.019283251836895943, + 0.015588302165269852, + 0.009380077011883259, + 0.7477752566337585 + ], + "credit_map": [ + 0.00434200931340456, + 0.003624965902417898, + 0.007150894030928612, + 0.010325786657631397, + 0.005249750334769487, + 0.004411238245666027, + 0.011917060241103172, + 0.004172667395323515, + 0.0017260204767808318, + 0.002356383018195629, + 0.003300265409052372, + 0.004039174411445856, + 0.0031417522113770247, + 0.0017649972578510642, + 0.004209793638437986, + 0.0018625862430781126, + 0.0035975680220872164, + 0.007545094937086105, + 0.003339453600347042, + 0.012569529004395008, + 0.007757872808724642, + 0.0016470596892759204, + 0.00928976759314537, + 0.001998203108087182, + 0.0021980362944304943, + 0.004783439449965954, + 0.0019070322159677744, + 0.0048113372176885605, + 0.004060437437146902, + 0.0014818388735875487, + 0.004057426005601883, + 0.0013184061972424388, + 0.0034224344417452812, + 0.007634087465703487, + 0.004277562722563744, + 0.004462506156414747, + 0.010824241675436497, + 0.0029876679182052612, + 0.006334453821182251, + 0.003109752433374524, + 0.0016991840675473213, + 0.003459553699940443, + 0.002739023184403777, + 0.003362636547535658, + 0.0035839632619172335, + 0.0009297431097365916, + 0.0018115326529368758, + 0.0013769067591056228, + 0.001957360887899995, + 0.004141216166317463, + 0.0019071983406320214, + 0.002643172862008214, + 0.0037078317254781723, + 0.00135419808793813, + 0.0019906347151845694, + 0.0015816381201148033, + 0.0015946150524541736, + 0.0021773853804916143, + 0.0027915998362004757, + 0.0021111899986863136, + 0.001567694591358304, + 0.0015619541518390179, + 0.0015970366075634956, + 0.0021868273615837097, + 0.0011778062907978892, + 0.0015516027342528105, + 0.0013847177615389228, + 0.0012643731897696853, + 0.0017162967706099153, + 0.0008004516712389886, + 0.0008445593994110823, + 0.0006402693688869476, + 0.011772248893976212, + 0.3499179482460022, + 0.004824730567634106, + 0.004784226417541504, + 0.35134223103523254, + 0.014087038114666939, + 0.0070999967865645885, + 0.003946850076317787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.008656476624310017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763415452092886 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.00763415452092886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763415452092886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10519672185182571, + 0.12280609458684921, + 0.08544342964887619, + 0.05117360129952431, + 0.05148687958717346, + 0.0787505954504013, + 0.08057938516139984, + 0.03936520591378212, + 0.021781396120786667, + 0.3634167015552521 + ], + "credit_map": [ + 0.00902605801820755, + 0.00852543581277132, + 0.014530510641634464, + 0.018356680870056152, + 0.008843944407999516, + 0.010485004633665085, + 0.024439789354801178, + 0.010989310219883919, + 0.011523095890879631, + 0.016635367646813393, + 0.015012603253126144, + 0.02189180999994278, + 0.018628176301717758, + 0.01027351338416338, + 0.0216729324311018, + 0.007168598473072052, + 0.007021530531346798, + 0.015868574380874634, + 0.007382628507912159, + 0.01615515537559986, + 0.012520676478743553, + 0.003699531080201268, + 0.01756570301949978, + 0.005229630041867495, + 0.00804081466048956, + 0.009263093583285809, + 0.00505194952711463, + 0.009529332630336285, + 0.006317651830613613, + 0.0030269718263298273, + 0.007631428074091673, + 0.0023123647551983595, + 0.004784205928444862, + 0.008542193099856377, + 0.0053663128055632114, + 0.006121928803622723, + 0.010201547294855118, + 0.003457778599113226, + 0.0072743757627904415, + 0.005738541949540377, + 0.0046753054484725, + 0.021391291171312332, + 0.006143444217741489, + 0.004915619269013405, + 0.029615825042128563, + 0.00440338347107172, + 0.0043955519795417786, + 0.0032101781107485294, + 0.008988657034933567, + 0.013246672227978706, + 0.008741972967982292, + 0.014949050731956959, + 0.00964328832924366, + 0.004953931551426649, + 0.013691835105419159, + 0.006363983731716871, + 0.004029836505651474, + 0.005413692910224199, + 0.006847233511507511, + 0.0035477213095873594, + 0.004511648323386908, + 0.0038649754133075476, + 0.00417142640799284, + 0.006978674791753292, + 0.0028250424657016993, + 0.002877676859498024, + 0.0031485012732446194, + 0.0023573630023747683, + 0.003376537235453725, + 0.0020023419056087732, + 0.002842199755832553, + 0.0023517359513789415, + 0.007643284276127815, + 0.15987899899482727, + 0.005762225482612848, + 0.008945691399276257, + 0.16058798134326935, + 0.00773604167625308, + 0.007944440469145775, + 0.004918062128126621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.00763415452092886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700429379940033 + ], + "opposite_logprob_tokens": [ + -2.695042848587036 + ], + "logprob_reward": -0.0700429379940033, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700429379940033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03398222103714943, + 0.06317268311977386, + 0.04855438694357872, + 0.028623512014746666, + 0.028657155111432076, + 0.015415815636515617, + 0.022874461486935616, + 0.012093770317733288, + 0.017158817499876022, + 0.7294672131538391 + ], + "credit_map": [ + 0.002216661348938942, + 0.0032828121911734343, + 0.004062772262841463, + 0.007522616535425186, + 0.003266480751335621, + 0.0021053622476756573, + 0.008966191671788692, + 0.002559323562309146, + 0.0063542150892317295, + 0.007156962528824806, + 0.006883980706334114, + 0.013774263672530651, + 0.008989214897155762, + 0.004513368476182222, + 0.01229328103363514, + 0.003207394853234291, + 0.0023005332332104445, + 0.0063485330902040005, + 0.005028801038861275, + 0.012467073276638985, + 0.007754656486213207, + 0.00247307401150465, + 0.009922086261212826, + 0.0022596288472414017, + 0.0036538965068757534, + 0.004692403599619865, + 0.0029750806279480457, + 0.004484336823225021, + 0.005925507750362158, + 0.0019504916854202747, + 0.0037636812776327133, + 0.0011781122302636504, + 0.0030507706105709076, + 0.004328573122620583, + 0.0031917549204081297, + 0.003863794496282935, + 0.0044536832720041275, + 0.0031094448640942574, + 0.004253308288753033, + 0.0024058257695287466, + 0.002351538510993123, + 0.0020393712911754847, + 0.0014016417553648353, + 0.002164612291380763, + 0.004012865945696831, + 0.0009640147327445447, + 0.001548898289911449, + 0.0009328730520792305, + 0.00200512888841331, + 0.0041193412616848946, + 0.002537302440032363, + 0.003415856044739485, + 0.0034395072143524885, + 0.0015545905334874988, + 0.0034105617087334394, + 0.002392174443230033, + 0.0008704382344149053, + 0.0015619875630363822, + 0.0017091576009988785, + 0.001496904413215816, + 0.00152507983148098, + 0.0009346730657853186, + 0.001818385673686862, + 0.00217714486643672, + 0.001531541463918984, + 0.0029565845616161823, + 0.0022076258901506662, + 0.001657912158407271, + 0.003102475544437766, + 0.0013574431650340557, + 0.0019865306094288826, + 0.0023587059695273638, + 0.011565607972443104, + 0.341139018535614, + 0.004683824721723795, + 0.0051927161403000355, + 0.3426694869995117, + 0.013675444759428501, + 0.006695915013551712, + 0.0038452004082500935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.0700429379940033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static close-ups, framed as physics-based observational macro shots, capture the mesmerizing process within the clear, cylindrical glass shutters. Initially, the water is a turbulent, opaque cloud of rich, dark brown, studded with fine coffee grounds, swirling like a suspended miniature galaxy. Soft, even kitchen light, diffused from an unseen source, illuminates the scene without harsh shadows, capturing the delicate glint of individual microscopic particles. The invisible yet persistent force of gravity takes over, and as each particle steadily drifts downward, a slow, deliberate descent begins. As they cautiously accumulate at the bottom of the glass, they form a distinct, dense deposit of dark amber. A sharp, ever-clarifying separating line emerges, separating solid from liquid, creating a stark visual boundary that rises as the water above gradually transforms from a murky brown to a translucent, warm amber, creating an atmosphere of tranquility and concentration. The details of observed natural phenomena are captured.", + "seed": 1669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_12/index_6441_seed_1669_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2428981613367796, + "vlm_logprob": -0.2428981613367796, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.2428981613367796, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05033852532505989 + ], + "opposite_logprob_tokens": [ + -3.9253385066986084 + ], + "logprob_reward": -0.05033852532505989, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.05033852532505989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20218254625797272, + 0.16407254338264465, + 0.05192766711115837, + 0.08582339435815811, + 0.05626799538731575, + 0.18046191334724426, + 0.09806855022907257, + 0.09303736686706543, + 0.0356811061501503, + 0.03247688710689545 + ], + "credit_map": [ + 0.022117650136351585, + 0.017140686511993408, + 0.014494283124804497, + 0.016554152593016624, + 0.03912802040576935, + 0.042151231318712234, + 0.022333841770887375, + 0.02826268970966339, + 0.015976646915078163, + 0.03657055273652077, + 0.016227619722485542, + 0.0158875472843647, + 0.018879979848861694, + 0.023016372695565224, + 0.014293908141553402, + 0.02321993187069893, + 0.00357775273732841, + 0.009920547716319561, + 0.005949878599494696, + 0.007501447107642889, + 0.009442050009965897, + 0.004485644865781069, + 0.0064855096861720085, + 0.004564837086945772, + 0.00443267822265625, + 0.011317932046949863, + 0.006901192013174295, + 0.01243804395198822, + 0.017888091504573822, + 0.013722541742026806, + 0.012574969790875912, + 0.006547949742525816, + 0.004105599131435156, + 0.013121712021529675, + 0.005165466107428074, + 0.007426140829920769, + 0.006715530529618263, + 0.006440737284719944, + 0.007914253510534763, + 0.005378558300435543, + 0.006044464185833931, + 0.06763645261526108, + 0.007720550522208214, + 0.009720426052808762, + 0.060070525854825974, + 0.016311153769493103, + 0.008404535241425037, + 0.00455381628125906, + 0.005155855789780617, + 0.0349615216255188, + 0.006246986333280802, + 0.006173037923872471, + 0.02957146055996418, + 0.007948624901473522, + 0.0037010631058365107, + 0.004310008604079485, + 0.005834253970533609, + 0.024667581543326378, + 0.00710259797051549, + 0.007654092274606228, + 0.022911176085472107, + 0.006774640176445246, + 0.005853551905602217, + 0.012239479459822178, + 0.005142526235431433, + 0.006976688746362925, + 0.004135088995099068, + 0.0030016452074050903, + 0.005994576495140791, + 0.003457556711509824, + 0.0027151727117598057, + 0.004257852677255869, + 0.0034386978950351477, + 0.0040892306715250015, + 0.0053006382659077644, + 0.004519195761531591, + 0.0043297722004354, + 0.004174347035586834, + 0.00242197816260159, + 0.004203028976917267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.05033852532505989 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539950132369995 + ], + "opposite_logprob_tokens": [ + -0.4289950430393219 + ], + "logprob_reward": -1.0539950132369995, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539950132369995, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24148301780223846, + 0.26258644461631775, + 0.05848300829529762, + 0.0843682512640953, + 0.09752252697944641, + 0.03357778117060661, + 0.07597698271274567, + 0.10281563550233841, + 0.026250477880239487, + 0.016935920342803 + ], + "credit_map": [ + 0.018680213019251823, + 0.021988697350025177, + 0.025817858055233955, + 0.03757121413946152, + 0.033272381871938705, + 0.03874264657497406, + 0.03583082556724548, + 0.029579194262623787, + 0.015412162989377975, + 0.06509371846914291, + 0.027391303330659866, + 0.018745247274637222, + 0.03190768510103226, + 0.03797822818160057, + 0.02760206162929535, + 0.03845604136586189, + 0.003524045692756772, + 0.005855846218764782, + 0.006133528426289558, + 0.007906496524810791, + 0.012900570407509804, + 0.006810743827372789, + 0.00944247655570507, + 0.005909300874918699, + 0.00718333013355732, + 0.014500933699309826, + 0.008308357559144497, + 0.009062027558684349, + 0.01878930814564228, + 0.008942931890487671, + 0.010620716959238052, + 0.006960644386708736, + 0.007545906119048595, + 0.02361445687711239, + 0.008175412192940712, + 0.011760121211409569, + 0.015858784317970276, + 0.012669187970459461, + 0.010718903504312038, + 0.007179753854870796, + 0.002489549107849598, + 0.008062689565122128, + 0.0030777063220739365, + 0.002856750041246414, + 0.0068804980255663395, + 0.005395296961069107, + 0.002441954333335161, + 0.0023733361158519983, + 0.00531611917540431, + 0.023864256218075752, + 0.006165902595967054, + 0.005784367676824331, + 0.0207301527261734, + 0.007047771476209164, + 0.0035138223320245743, + 0.003554592840373516, + 0.0057793715968728065, + 0.03551309555768967, + 0.006694064941257238, + 0.006786477752029896, + 0.030561598017811775, + 0.00788917951285839, + 0.004242226015776396, + 0.005349610932171345, + 0.0023868365678936243, + 0.005114174913614988, + 0.002445995109155774, + 0.0024814584758132696, + 0.005065341480076313, + 0.004433969967067242, + 0.001683996059000492, + 0.002638705540448427, + 0.0015958752483129501, + 0.003180993255227804, + 0.0014342146459966898, + 0.0015677149640396237, + 0.003315544920042157, + 0.0033757591154426336, + 0.0011261263862252235, + 0.001339691923931241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -1.0539950132369995 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03817165642976761 + ], + "opposite_logprob_tokens": [ + -3.2881717681884766 + ], + "logprob_reward": -0.03817165642976761, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03817165642976761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15823876857757568, + 0.41347575187683105, + 0.048357971012592316, + 0.07379595190286636, + 0.16351033747196198, + 0.03542771935462952, + 0.026933593675494194, + 0.04987785592675209, + 0.018408862873911858, + 0.011973193846642971 + ], + "credit_map": [ + 0.012835124507546425, + 0.01443804707378149, + 0.011856871657073498, + 0.01932261325418949, + 0.026139726862311363, + 0.035763975232839584, + 0.021326163783669472, + 0.01655624434351921, + 0.025580832734704018, + 0.08283182233572006, + 0.03891732543706894, + 0.04300973191857338, + 0.049359600991010666, + 0.05402996018528938, + 0.05503270402550697, + 0.06471381336450577, + 0.003238534554839134, + 0.0050310296937823296, + 0.004105069674551487, + 0.007276464253664017, + 0.012062051333487034, + 0.005610787775367498, + 0.006145149003714323, + 0.004888886120170355, + 0.006232536863535643, + 0.012412512674927711, + 0.0068734860979020596, + 0.007340859156101942, + 0.01782621443271637, + 0.008125356398522854, + 0.009107013233006, + 0.0058779725804924965, + 0.013929803855717182, + 0.041281040757894516, + 0.015222413465380669, + 0.02018563263118267, + 0.021988824009895325, + 0.020626919344067574, + 0.018552517518401146, + 0.011723201721906662, + 0.002991195535287261, + 0.007851404137909412, + 0.003022594377398491, + 0.0033512089867144823, + 0.006473590154200792, + 0.006204182282090187, + 0.0031834139954298735, + 0.0023501296527683735, + 0.002231455408036709, + 0.007626911625266075, + 0.002226430457085371, + 0.0022409886587411165, + 0.006432500667870045, + 0.0031610222067683935, + 0.0013951220316812396, + 0.0016191622707992792, + 0.0026515345089137554, + 0.017705796286463737, + 0.0027704578824341297, + 0.0029740193858742714, + 0.014799079857766628, + 0.003988716751337051, + 0.0019381997408345342, + 0.003050051163882017, + 0.001970254350453615, + 0.0034261182881891727, + 0.0018979457672685385, + 0.0013606989523395896, + 0.002965246792882681, + 0.003218927653506398, + 0.0013960327487438917, + 0.0021736398339271545, + 0.001243723090738058, + 0.0020643589086830616, + 0.0012269840808585286, + 0.001357351546175778, + 0.001939829788170755, + 0.0022199058439582586, + 0.0008128916961140931, + 0.0011081480188295245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.03817165642976761 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308804258704185 + ], + "opposite_logprob_tokens": [ + -3.773308753967285 + ], + "logprob_reward": -0.023308804258704185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308804258704185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15861058235168457, + 0.34363123774528503, + 0.08358361572027206, + 0.14758463203907013, + 0.15564283728599548, + 0.02902321331202984, + 0.023329628631472588, + 0.018016371876001358, + 0.02464359812438488, + 0.015934288501739502 + ], + "credit_map": [ + 0.011692470870912075, + 0.016802996397018433, + 0.013522069901227951, + 0.017327571287751198, + 0.02878660522401333, + 0.025619810447096825, + 0.023135608062148094, + 0.021723449230194092, + 0.02422059141099453, + 0.06218434497714043, + 0.03828871250152588, + 0.027056735008955002, + 0.04034833610057831, + 0.048848655074834824, + 0.03797627240419388, + 0.06470759958028793, + 0.005390968173742294, + 0.015265108086168766, + 0.00942969974130392, + 0.011966554448008537, + 0.014800737611949444, + 0.009852600283920765, + 0.008925911039113998, + 0.00795204471796751, + 0.01059349998831749, + 0.017140550538897514, + 0.01225996669381857, + 0.019840994849801064, + 0.02882763370871544, + 0.02457679808139801, + 0.015400964766740799, + 0.018944229930639267, + 0.011948255822062492, + 0.039807893335819244, + 0.014471722766757011, + 0.018786365166306496, + 0.019383862614631653, + 0.02109576016664505, + 0.018103694543242455, + 0.012045270763337612, + 0.0021702710073441267, + 0.007910015061497688, + 0.002926622284576297, + 0.0024525548797100782, + 0.005796722136437893, + 0.002935323864221573, + 0.002131304470822215, + 0.002700396813452244, + 0.00231185182929039, + 0.00617872504517436, + 0.0024887239560484886, + 0.0017879053484648466, + 0.004884612746536732, + 0.003021832788363099, + 0.0012498272117227316, + 0.0014061499387025833, + 0.0016044428339228034, + 0.0038577725645154715, + 0.001783875748515129, + 0.0020388562697917223, + 0.0032235225662589073, + 0.0024144172202795744, + 0.0015869763446971774, + 0.0015065090265125036, + 0.0021429178304970264, + 0.007382956799119711, + 0.002125297673046589, + 0.0019541585352271795, + 0.006388979498296976, + 0.002216542139649391, + 0.001056363107636571, + 0.0013763833558186889, + 0.0014978108229115605, + 0.0030013457871973515, + 0.0013513920130208135, + 0.0016514372546225786, + 0.002534437458962202, + 0.0021729369182139635, + 0.0011426410637795925, + 0.0025822867173701525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.023308804258704185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a close-up macro shot inside a clear cylindrical glass container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the initial water appear as a turbulent, opaque cloud of dark brown studded with fine coffee grounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the coffee grounds drift downward and accumulate at the bottom to form a dense deposit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp separating line emerge between the solid deposit and the liquid above as the water clarifies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even without harsh shadows, illuminating the microscopic particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048676807433366776 + ], + "opposite_logprob_tokens": [ + -3.0486767292022705 + ], + "logprob_reward": -0.048676807433366776, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048676807433366776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26971593499183655, + 0.0757092535495758, + 0.09205646812915802, + 0.1266028881072998, + 0.07125643640756607, + 0.07213075459003448, + 0.06169712916016579, + 0.15301544964313507, + 0.038481708616018295, + 0.0393340177834034 + ], + "credit_map": [ + 0.02916909195482731, + 0.027266113087534904, + 0.025947079062461853, + 0.027176687493920326, + 0.05221525579690933, + 0.04556664451956749, + 0.030787451192736626, + 0.031587615609169006, + 0.00559943588450551, + 0.01255455706268549, + 0.011416890658438206, + 0.00922273751348257, + 0.00899131502956152, + 0.008646990172564983, + 0.008385911583900452, + 0.010891415178775787, + 0.004655410069972277, + 0.010505612008273602, + 0.008167400024831295, + 0.015268770977854729, + 0.024295883253216743, + 0.009202694520354271, + 0.011679112911224365, + 0.00828158762305975, + 0.0086424071341753, + 0.015701573342084885, + 0.011558882892131805, + 0.02001621387898922, + 0.02169261872768402, + 0.012516538612544537, + 0.024540407583117485, + 0.01193423755466938, + 0.008091759867966175, + 0.01789158396422863, + 0.006182648241519928, + 0.006847611162811518, + 0.009714760817587376, + 0.009064718149602413, + 0.007429146207869053, + 0.006034206133335829, + 0.003871988970786333, + 0.02477153018116951, + 0.003777892328798771, + 0.004757720977067947, + 0.017827002331614494, + 0.010103804990649223, + 0.00455932691693306, + 0.002461484167724848, + 0.002661796985194087, + 0.022715047001838684, + 0.002761462004855275, + 0.0031419163569808006, + 0.021866660565137863, + 0.0036714812740683556, + 0.002818266861140728, + 0.002060498343780637, + 0.009170141071081161, + 0.052489083260297775, + 0.009345294907689095, + 0.009597256779670715, + 0.04223889112472534, + 0.013314873911440372, + 0.0066936747170984745, + 0.010166232474148273, + 0.005417721811681986, + 0.006271387450397015, + 0.004689459223300219, + 0.0034871238749474287, + 0.005170449614524841, + 0.005893437657505274, + 0.0031871763058006763, + 0.004364952445030212, + 0.002819494577124715, + 0.012387053109705448, + 0.0023707367945462465, + 0.0026676023844629526, + 0.012732372619211674, + 0.0023501524701714516, + 0.001755802659317851, + 0.0022508068941533566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.048676807433366776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_tensor([8369, 8369])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_tensor([8369, 8369])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5bfc84b88f527c34613dad04e7e2b7c5ef9e6096 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_tensor([8369, 8369])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro-probe shot dynamically follows a perfectly spherical drop of viscous black engine oil as it travels through the complex, roaring heart of an engine. The camera tracks smoothly, keeping the glistening droplet centered as it pushes through the narrow, machined steel passageway. The oil's mirror-like surface reflects twisting, elongated images of interconnected gears and the fiery orange glow of a distant burn. It slides steadily along the metal wall, its path illuminated by the harsh, specular highlights of polished surfaces and the oil itself. The environment is a claustrophobic maze of metallic grays and blacks, punctuated by the crimson of overheated components, creating a visible heat haze. As the droplet deftly navigates sharp corners, it briefly stretches and elongates before surface tension pulls it back into a sphere, revealing its viscosity. The entire scene vibrates with the immense power of the engine, creating an atmosphere of tension, mechanical power, and power. The camera follows the droplet's relentless, fluid movement through this complex, high-pressure world.", + "seed": 19152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_8369_seed_19152_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04888880364596844, + "vlm_logprob": -0.04888880364596844, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04888880364596844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08491045981645584 + ], + "opposite_logprob_tokens": [ + -2.5849103927612305 + ], + "logprob_reward": -0.08491045981645584, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08491045981645584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17039918899536133, + 0.21405677497386932, + 0.3190109133720398, + 0.03770036622881889, + 0.07529141008853912, + 0.025310855358839035, + 0.03661634773015976, + 0.03893196955323219, + 0.04899870604276657, + 0.03368350490927696 + ], + "credit_map": [ + 0.01704849675297737, + 0.010249066166579723, + 0.0363444909453392, + 0.009964552707970142, + 0.010144311934709549, + 0.05049988999962807, + 0.010513613931834698, + 0.025634752586483955, + 0.02580253593623638, + 0.03703169897198677, + 0.036063265055418015, + 0.020966317504644394, + 0.02001015469431877, + 0.02380610629916191, + 0.021102460101246834, + 0.029274217784404755, + 0.022123878821730614, + 0.017090870067477226, + 0.033063653856515884, + 0.016019510105252266, + 0.01255074329674244, + 0.11783403903245926, + 0.08316150307655334, + 0.01716667413711548, + 0.005345087498426437, + 0.004239704459905624, + 0.003918228670954704, + 0.004731202498078346, + 0.004276596009731293, + 0.005828673020005226, + 0.005718035623431206, + 0.0036428316961973906, + 0.006264358758926392, + 0.009350264444947243, + 0.0110700074583292, + 0.00966352503746748, + 0.011375918053090572, + 0.010581089183688164, + 0.0051722745411098, + 0.011813962832093239, + 0.003341623116284609, + 0.003880346193909645, + 0.003777759149670601, + 0.0039726137183606625, + 0.001861703465692699, + 0.0026358182076364756, + 0.0026162131689488888, + 0.003224776592105627, + 0.0049080322496593, + 0.005329867359250784, + 0.005216546822339296, + 0.0043116058222949505, + 0.004134255927056074, + 0.005706964060664177, + 0.004855612758547068, + 0.00215346016921103, + 0.00618226733058691, + 0.004398603457957506, + 0.0045172483660280704, + 0.00970105454325676, + 0.004624576773494482, + 0.0030533315148204565, + 0.0036751111038029194, + 0.002779773436486721, + 0.010408792644739151, + 0.002782934345304966, + 0.005592070519924164, + 0.005091994535177946, + 0.013851981610059738, + 0.005633044987916946, + 0.0030547326896339655, + 0.002583148656412959, + 0.004542930983006954, + 0.004654715768992901, + 0.004047786816954613, + 0.003936767112463713, + 0.0032946153078228235, + 0.003079297486692667, + 0.0067175342701375484, + 0.0034098532050848007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.08491045981645584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299889534711838 + ], + "opposite_logprob_tokens": [ + -2.2379989624023438 + ], + "logprob_reward": -0.11299889534711838, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299889534711838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2776334583759308, + 0.1605233997106552, + 0.1434270292520523, + 0.04501728713512421, + 0.1250392198562622, + 0.04157228022813797, + 0.05714578554034233, + 0.050423722714185715, + 0.07357943058013916, + 0.025638442486524582 + ], + "credit_map": [ + 0.026285018771886826, + 0.016904274001717567, + 0.059908971190452576, + 0.020192334428429604, + 0.016336310654878616, + 0.09372065961360931, + 0.015302401967346668, + 0.02898346818983555, + 0.018282050266861916, + 0.02324102260172367, + 0.02932767942547798, + 0.013177625834941864, + 0.01547246240079403, + 0.022261083126068115, + 0.017337936908006668, + 0.02142352983355522, + 0.018008198589086533, + 0.010454371571540833, + 0.024630440399050713, + 0.009455484338104725, + 0.013639401644468307, + 0.031610000878572464, + 0.022016307339072227, + 0.013612810522317886, + 0.004930428229272366, + 0.0035565681755542755, + 0.005293740890920162, + 0.004778524860739708, + 0.007586501073092222, + 0.007762247696518898, + 0.007105157244950533, + 0.004004117101430893, + 0.009507709182798862, + 0.014259891584515572, + 0.020797457545995712, + 0.015485339798033237, + 0.02281327359378338, + 0.014757397584617138, + 0.01061253622174263, + 0.016805605962872505, + 0.0055358209647238255, + 0.00611210847273469, + 0.004266505595296621, + 0.006738914176821709, + 0.0032582117710262537, + 0.004702993668615818, + 0.006418274715542793, + 0.004539450164884329, + 0.008953452110290527, + 0.005177810322493315, + 0.011631540954113007, + 0.00625870656222105, + 0.004640219267457724, + 0.008202512748539448, + 0.007290818262845278, + 0.004990722052752972, + 0.009309436194598675, + 0.0042133815586566925, + 0.004062306135892868, + 0.01114318985491991, + 0.00518068578094244, + 0.006251509767025709, + 0.0059957122430205345, + 0.004267499316483736, + 0.012039968743920326, + 0.004461292177438736, + 0.007108885329216719, + 0.003971440717577934, + 0.023170106112957, + 0.014019754715263844, + 0.005737251602113247, + 0.003070725128054619, + 0.0038210228085517883, + 0.0033109928481280804, + 0.003270302200689912, + 0.0022446129005402327, + 0.0031944673974066973, + 0.004193469882011414, + 0.0027103524189442396, + 0.00289321830496192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11299889534711838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647049613296986 + ], + "opposite_logprob_tokens": [ + -4.8826470375061035 + ], + "logprob_reward": -0.007647049613296986, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647049613296986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.50046306848526, + 0.059368424117565155, + 0.04604864865541458, + 0.0309724360704422, + 0.10239008069038391, + 0.06695687770843506, + 0.06604334712028503, + 0.03552967309951782, + 0.05020257085561752, + 0.042024798691272736 + ], + "credit_map": [ + 0.039871200919151306, + 0.016168871894478798, + 0.1496322900056839, + 0.021803921088576317, + 0.021782241761684418, + 0.19624781608581543, + 0.016663722693920135, + 0.038293011486530304, + 0.007566041778773069, + 0.011182435788214207, + 0.008642513304948807, + 0.004395973868668079, + 0.005939983297139406, + 0.008024495095014572, + 0.0057366653345525265, + 0.007880313321948051, + 0.011985884979367256, + 0.0030446588061749935, + 0.007423793897032738, + 0.0027332929894328117, + 0.0028814435936510563, + 0.00635318411514163, + 0.008259298279881477, + 0.003367091529071331, + 0.00488173495978117, + 0.002209900878369808, + 0.004981434438377619, + 0.002668229164555669, + 0.003161993809044361, + 0.004944866523146629, + 0.0047200825065374374, + 0.0034041947219520807, + 0.00814551766961813, + 0.013345016166567802, + 0.016146233305335045, + 0.014547040686011314, + 0.015508170239627361, + 0.010298540815711021, + 0.007620642893016338, + 0.016778916120529175, + 0.011274508200585842, + 0.009211585856974125, + 0.010581986978650093, + 0.009487621486186981, + 0.004513172898441553, + 0.006472518667578697, + 0.00644394988194108, + 0.008971537463366985, + 0.010842995718121529, + 0.005838796496391296, + 0.009648389182984829, + 0.005300676450133324, + 0.004876626189798117, + 0.008382054977118969, + 0.013257930986583233, + 0.007895881310105324, + 0.005459832027554512, + 0.003453721757978201, + 0.003913309890776873, + 0.004209951963275671, + 0.0033942207228392363, + 0.007559522055089474, + 0.003819651436060667, + 0.003719465807080269, + 0.0065163434483110905, + 0.0034750811755657196, + 0.004211295861750841, + 0.005000275559723377, + 0.00669230567291379, + 0.013151580467820168, + 0.006406509783118963, + 0.004749181680381298, + 0.005455336533486843, + 0.005841231904923916, + 0.004392925649881363, + 0.0033761218655854464, + 0.005847782827913761, + 0.006914262659847736, + 0.007064956706017256, + 0.003132180543616414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.007647049613296986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545830570161343 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545830570161343, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545830570161343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2485489696264267, + 0.13583408296108246, + 0.16555367410182953, + 0.1402599960565567, + 0.04035681486129761, + 0.07628192752599716, + 0.050653230398893356, + 0.031908463686704636, + 0.08169510215520859, + 0.02890770323574543 + ], + "credit_map": [ + 0.024444306269288063, + 0.011708548292517662, + 0.06218285858631134, + 0.011762710288167, + 0.01166753750294447, + 0.08283255994319916, + 0.013529159128665924, + 0.030421296134591103, + 0.016453493386507034, + 0.01777905784547329, + 0.018581802025437355, + 0.012319212779402733, + 0.014922037720680237, + 0.022414740175008774, + 0.014946892857551575, + 0.018416836857795715, + 0.04714135453104973, + 0.011775370687246323, + 0.02952912636101246, + 0.009668098762631416, + 0.01435908768326044, + 0.021127047017216682, + 0.02022220380604267, + 0.011731397360563278, + 0.028461076319217682, + 0.010114013217389584, + 0.02733580768108368, + 0.010295148938894272, + 0.009966021403670311, + 0.02675987407565117, + 0.016197843477129936, + 0.011130208149552345, + 0.0039722491055727005, + 0.003887425409629941, + 0.007272033486515284, + 0.003838634816929698, + 0.006030130200088024, + 0.0077989608980715275, + 0.0031853914260864258, + 0.004371988121420145, + 0.012397353537380695, + 0.013762292452156544, + 0.013266418129205704, + 0.014340579509735107, + 0.0046592773869633675, + 0.005183948669582605, + 0.0068954783491790295, + 0.005776580423116684, + 0.006703886669129133, + 0.003474294673651457, + 0.008744582533836365, + 0.0073187085799872875, + 0.005118532571941614, + 0.007907264865934849, + 0.006881304085254669, + 0.0045046559534966946, + 0.004628318361938, + 0.003488800721243024, + 0.004248417913913727, + 0.004445506259799004, + 0.0028911177068948746, + 0.005734555888921022, + 0.0037388054188340902, + 0.002732940949499607, + 0.010649862699210644, + 0.004154645837843418, + 0.0055376202799379826, + 0.009419312700629234, + 0.023189261555671692, + 0.016270175576210022, + 0.00797759834676981, + 0.004496615380048752, + 0.004769702907651663, + 0.004291644785553217, + 0.0033800541423261166, + 0.002351576229557395, + 0.0036679881159216166, + 0.0039725881069898605, + 0.0036725718528032303, + 0.002801578026264906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.012545830570161343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634178288280964 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.02634178288280964, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634178288280964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12862783670425415, + 0.09666759520769119, + 0.1379445642232895, + 0.1107686460018158, + 0.21626155078411102, + 0.08505077660083771, + 0.07046902924776077, + 0.060115035623311996, + 0.06630906462669373, + 0.02778587117791176 + ], + "credit_map": [ + 0.011105719022452831, + 0.023331770673394203, + 0.028568612411618233, + 0.009270543232560158, + 0.012448042631149292, + 0.029887743294239044, + 0.005504949949681759, + 0.00851045735180378, + 0.011921045370399952, + 0.01965157873928547, + 0.012632060796022415, + 0.011362602934241295, + 0.010123883374035358, + 0.008478820323944092, + 0.008844234980642796, + 0.013653360307216644, + 0.022310761734843254, + 0.009411769919097424, + 0.005819942802190781, + 0.007289397530257702, + 0.011162209324538708, + 0.04590883478522301, + 0.02771744877099991, + 0.00832420028746128, + 0.021155428141355515, + 0.009471590630710125, + 0.02241651341319084, + 0.007101036608219147, + 0.01073435042053461, + 0.01787126064300537, + 0.010695850476622581, + 0.011322615668177605, + 0.016796806827187538, + 0.02889900654554367, + 0.03461317718029022, + 0.03154435381293297, + 0.03193332254886627, + 0.0247286818921566, + 0.01571821980178356, + 0.032027989625930786, + 0.011293534189462662, + 0.011636071838438511, + 0.01111944206058979, + 0.008443787693977356, + 0.0076157404109835625, + 0.011006138287484646, + 0.01136067695915699, + 0.012575384229421616, + 0.010673750191926956, + 0.007227556314319372, + 0.010545711033046246, + 0.006200219504535198, + 0.007769336923956871, + 0.013247826136648655, + 0.007829330861568451, + 0.006975298747420311, + 0.009496735408902168, + 0.005784669425338507, + 0.0058585358783602715, + 0.011352893896400928, + 0.007330393418669701, + 0.009028123691678047, + 0.006195859517902136, + 0.0050678253173828125, + 0.008915362879633904, + 0.006936550606042147, + 0.006652900483459234, + 0.01340489648282528, + 0.013502079993486404, + 0.007170403376221657, + 0.006020107306540012, + 0.0037067681550979614, + 0.0028844166081398726, + 0.0038536638021469116, + 0.00309968669898808, + 0.003144112415611744, + 0.003642607480287552, + 0.0029683352913707495, + 0.005797237623482943, + 0.0023958133533596992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.02634178288280964 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro-probe shot dynamically follows a perfectly spherical drop of viscous black engine oil as it travels through the complex, roaring heart of an engine. The camera tracks smoothly, keeping the glistening droplet centered as it pushes through the narrow, machined steel passageway. The oil's mirror-like surface reflects twisting, elongated images of interconnected gears and the fiery orange glow of a distant burn. It slides steadily along the metal wall, its path illuminated by the harsh, specular highlights of polished surfaces and the oil itself. The environment is a claustrophobic maze of metallic grays and blacks, punctuated by the crimson of overheated components, creating a visible heat haze. As the droplet deftly navigates sharp corners, it briefly stretches and elongates before surface tension pulls it back into a sphere, revealing its viscosity. The entire scene vibrates with the immense power of the engine, creating an atmosphere of tension, mechanical power, and power. The camera follows the droplet's relentless, fluid movement through this complex, high-pressure world.", + "seed": 19153, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_8369_seed_19153_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03823832683265209, + "vlm_logprob": -0.03823832683265209, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03823832683265209, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11954472213983536 + ], + "opposite_logprob_tokens": [ + -2.244544744491577 + ], + "logprob_reward": -0.11954472213983536, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11954472213983536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14616939425468445, + 0.07707637548446655, + 0.04318796098232269, + 0.15650004148483276, + 0.11293185502290726, + 0.0453580878674984, + 0.09642051160335541, + 0.10474614053964615, + 0.1704307645559311, + 0.047178782522678375 + ], + "credit_map": [ + 0.011103524826467037, + 0.013619055040180683, + 0.024030335247516632, + 0.016273804008960724, + 0.010659275576472282, + 0.02596278488636017, + 0.013831784948706627, + 0.030688831582665443, + 0.014261693693697453, + 0.013863962143659592, + 0.010325456969439983, + 0.011468893848359585, + 0.00801780540496111, + 0.005584699101746082, + 0.0063254269771277905, + 0.007228437345474958, + 0.004620379768311977, + 0.005960887763649225, + 0.006082952953875065, + 0.008165608160197735, + 0.0046036625280976295, + 0.006153950933367014, + 0.004176134709268808, + 0.003424383932724595, + 0.020455483347177505, + 0.032281626015901566, + 0.02306843362748623, + 0.01878829300403595, + 0.029693596065044403, + 0.01051294431090355, + 0.01109851524233818, + 0.010601144284009933, + 0.011731758713722229, + 0.015603393316268921, + 0.024721970781683922, + 0.01627179980278015, + 0.0084259994328022, + 0.009919088333845139, + 0.012584230862557888, + 0.01367362029850483, + 0.006576936691999435, + 0.007071373984217644, + 0.006078996695578098, + 0.0036783688701689243, + 0.005528229288756847, + 0.004856261890381575, + 0.007554095238447189, + 0.004013827536255121, + 0.01985500194132328, + 0.011489033699035645, + 0.011922425590455532, + 0.00827511865645647, + 0.010398486629128456, + 0.009669387713074684, + 0.009401516057550907, + 0.01540953479707241, + 0.021376047283411026, + 0.007326582912355661, + 0.017344828695058823, + 0.014401141554117203, + 0.014958714134991169, + 0.008916775695979595, + 0.008827551268041134, + 0.011594491079449654, + 0.009664052166044712, + 0.0068243881687521935, + 0.007811170537024736, + 0.05127352848649025, + 0.015029439702630043, + 0.01700298860669136, + 0.04950601980090141, + 0.013319194316864014, + 0.0046423752792179585, + 0.006151576526463032, + 0.005482472945004702, + 0.006509555969387293, + 0.005765808746218681, + 0.0062981280498206615, + 0.0058357687667012215, + 0.006493096239864826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.11954472213983536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018212705850601196 + ], + "opposite_logprob_tokens": [ + -4.018212795257568 + ], + "logprob_reward": -0.018212705850601196, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018212705850601196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08466137945652008, + 0.07449816167354584, + 0.04613664373755455, + 0.0994676947593689, + 0.10652882605791092, + 0.04957357421517372, + 0.060603704303503036, + 0.19360952079296112, + 0.25422507524490356, + 0.03069543093442917 + ], + "credit_map": [ + 0.007571041584014893, + 0.008667072281241417, + 0.00891589093953371, + 0.01253377553075552, + 0.010689772665500641, + 0.01771046593785286, + 0.01111526321619749, + 0.007458089850842953, + 0.0061803595162928104, + 0.008041652850806713, + 0.021109726279973984, + 0.009431972168385983, + 0.005508796311914921, + 0.0046033733524382114, + 0.0038656156975775957, + 0.015756666660308838, + 0.006311305332928896, + 0.009549817070364952, + 0.00589995039626956, + 0.007127676159143448, + 0.004125541541725397, + 0.005835739430040121, + 0.003878576448187232, + 0.0034080378245562315, + 0.013402472250163555, + 0.01984954997897148, + 0.014214915223419666, + 0.015304089523851871, + 0.018620144575834274, + 0.005332294385880232, + 0.007178018800914288, + 0.005566210951656103, + 0.014008238911628723, + 0.025892194360494614, + 0.013978921808302402, + 0.010865877382457256, + 0.008795413188636303, + 0.006072906777262688, + 0.00708720413967967, + 0.0198280680924654, + 0.009737239219248295, + 0.006855279207229614, + 0.009542656131088734, + 0.00353039987385273, + 0.004128636792302132, + 0.006847272627055645, + 0.005345365032553673, + 0.0035867225378751755, + 0.009641042910516262, + 0.00608632480725646, + 0.009079496376216412, + 0.007601275108754635, + 0.006186049431562424, + 0.006491452921181917, + 0.00923234224319458, + 0.006285723298788071, + 0.04056607186794281, + 0.014012103900313377, + 0.033446867018938065, + 0.023859616369009018, + 0.02719833329319954, + 0.01706935279071331, + 0.015052272006869316, + 0.02240491844713688, + 0.014120715670287609, + 0.009310607798397541, + 0.00892857275903225, + 0.06253958493471146, + 0.016803644597530365, + 0.023631010204553604, + 0.09202172607183456, + 0.02686922624707222, + 0.005817448254674673, + 0.0029566676821559668, + 0.002655021846294403, + 0.003645932534709573, + 0.002965701976791024, + 0.006030916702002287, + 0.0029164340812712908, + 0.0037073087878525257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.018212705850601196 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011095844209194183 + ], + "opposite_logprob_tokens": [ + -4.511096000671387 + ], + "logprob_reward": -0.011095844209194183, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011095844209194183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06457043439149857, + 0.13738223910331726, + 0.028467563912272453, + 0.055059764534235, + 0.13720445334911346, + 0.09934234619140625, + 0.13402217626571655, + 0.0892571434378624, + 0.19619685411453247, + 0.05849703028798103 + ], + "credit_map": [ + 0.00869990885257721, + 0.0053001753985881805, + 0.008748145774006844, + 0.011345157399773598, + 0.0069472165778279305, + 0.008848349563777447, + 0.009090080857276917, + 0.0055913967080414295, + 0.009135492146015167, + 0.014309562742710114, + 0.045555174350738525, + 0.017377009615302086, + 0.010993708856403828, + 0.007420921698212624, + 0.006886489223688841, + 0.025703897699713707, + 0.002945710439234972, + 0.00287467148154974, + 0.004160911310464144, + 0.004522736184298992, + 0.0033337813802063465, + 0.004094062373042107, + 0.00392149155959487, + 0.0026141994167119265, + 0.005227680783718824, + 0.01119469478726387, + 0.006995991338044405, + 0.006962381768971682, + 0.011805616319179535, + 0.004333819262683392, + 0.005635626148432493, + 0.0029039541259407997, + 0.013444365002214909, + 0.02309533767402172, + 0.028841326013207436, + 0.011052272282540798, + 0.013985102996230125, + 0.00621554721146822, + 0.009250583127140999, + 0.03131992742419243, + 0.01686907559633255, + 0.016983116045594215, + 0.019879095256328583, + 0.008215196430683136, + 0.010956558398902416, + 0.00795247033238411, + 0.012096294201910496, + 0.006390536669641733, + 0.024984309449791908, + 0.019106347113847733, + 0.017168749123811722, + 0.014443127438426018, + 0.009200453758239746, + 0.014457043260335922, + 0.01876671053469181, + 0.015895437449216843, + 0.016795223578810692, + 0.006480729207396507, + 0.015718284994363785, + 0.009457763284444809, + 0.01639300212264061, + 0.006401641294360161, + 0.006736982613801956, + 0.011273515410721302, + 0.0111331045627594, + 0.006756441667675972, + 0.006564913783222437, + 0.04445667937397957, + 0.017901025712490082, + 0.016968615353107452, + 0.06865376234054565, + 0.02376229502260685, + 0.007583848666399717, + 0.005584455095231533, + 0.004703805781900883, + 0.008661758154630661, + 0.008810753934085369, + 0.008391447365283966, + 0.009769035503268242, + 0.0049919243901968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.011095844209194183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545594945549965 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545594945549965, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545594945549965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387212574481964, + 0.11663178354501724, + 0.051461923867464066, + 0.11441466212272644, + 0.06653446704149246, + 0.10492779314517975, + 0.20215392112731934, + 0.05571002513170242, + 0.09769284725189209, + 0.05175134912133217 + ], + "credit_map": [ + 0.009669572114944458, + 0.012223863042891026, + 0.018163442611694336, + 0.026298649609088898, + 0.01380862481892109, + 0.029302924871444702, + 0.016199495643377304, + 0.013054677285254002, + 0.00790907721966505, + 0.013485413044691086, + 0.03877575322985649, + 0.013678174465894699, + 0.007026775740087032, + 0.0060891322791576385, + 0.0063771456480026245, + 0.02329031191766262, + 0.005485291127115488, + 0.010057824663817883, + 0.007639061193913221, + 0.00682786712422967, + 0.006083357613533735, + 0.0063987658359110355, + 0.004380395170301199, + 0.00458936020731926, + 0.009103024378418922, + 0.019283471629023552, + 0.01373364683240652, + 0.013946255668997765, + 0.02892872877418995, + 0.008034908212721348, + 0.013677442446351051, + 0.007707186974585056, + 0.008224728517234325, + 0.010648690164089203, + 0.012858770787715912, + 0.007152736186981201, + 0.005444645881652832, + 0.004569240380078554, + 0.005050608888268471, + 0.0125850485637784, + 0.02650264836847782, + 0.01231420785188675, + 0.0180830005556345, + 0.008893390186131, + 0.006068861111998558, + 0.012705264613032341, + 0.014357485808432102, + 0.006002930458635092, + 0.025626804679632187, + 0.021146832033991814, + 0.029259735718369484, + 0.022634001448750496, + 0.031517162919044495, + 0.02080661617219448, + 0.027881328016519547, + 0.02328144758939743, + 0.009858190082013607, + 0.006369261536747217, + 0.0077427406795322895, + 0.008986866101622581, + 0.006088364403694868, + 0.0038934883195906878, + 0.005236908793449402, + 0.007534205447882414, + 0.006873669568449259, + 0.004664918873459101, + 0.005534568335860968, + 0.016998913139104843, + 0.012037497945129871, + 0.009740709327161312, + 0.03007618337869644, + 0.01176638063043356, + 0.007563794497400522, + 0.00428994745016098, + 0.0062729353085160255, + 0.006829145364463329, + 0.006706081330776215, + 0.006339744199067354, + 0.009342593140900135, + 0.004407105967402458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.012545594945549965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792767018079758 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.029792767018079758, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792767018079758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1297682374715805, + 0.1293376386165619, + 0.02493089623749256, + 0.09429185092449188, + 0.06977447122335434, + 0.04776306450366974, + 0.09707336127758026, + 0.07566583901643753, + 0.23208782076835632, + 0.09930682927370071 + ], + "credit_map": [ + 0.014266064390540123, + 0.015044102445244789, + 0.014414645731449127, + 0.019977251067757607, + 0.015229237265884876, + 0.01951199397444725, + 0.019167978316545486, + 0.012156959623098373, + 0.008085994981229305, + 0.012599676847457886, + 0.044089194387197495, + 0.01285377237945795, + 0.00828587170690298, + 0.006888102740049362, + 0.005254718009382486, + 0.03128030523657799, + 0.002823969116434455, + 0.004085963126271963, + 0.0038246929179877043, + 0.003264432307332754, + 0.0028893209528177977, + 0.0026876407209783792, + 0.002533532911911607, + 0.002821342321112752, + 0.007830213755369186, + 0.02225434221327305, + 0.012042409740388393, + 0.012048489414155483, + 0.01886628195643425, + 0.0075174602679908276, + 0.008581786416471004, + 0.005150866229087114, + 0.009118199348449707, + 0.008014371618628502, + 0.014830195344984531, + 0.010842936113476753, + 0.005626153200864792, + 0.006125277373939753, + 0.007505054119974375, + 0.007712284568697214, + 0.005884677171707153, + 0.007479535415768623, + 0.007720834109932184, + 0.005796170327812433, + 0.006449347361922264, + 0.004936574958264828, + 0.005866545252501965, + 0.003629382001236081, + 0.01391543447971344, + 0.011469695717096329, + 0.01279249507933855, + 0.0110006183385849, + 0.011995484121143818, + 0.010076811537146568, + 0.014718027785420418, + 0.01110479049384594, + 0.013221555389463902, + 0.007592278067022562, + 0.01088939793407917, + 0.010771064087748528, + 0.007912050001323223, + 0.004947158973664045, + 0.0073798117227852345, + 0.012952519580721855, + 0.014055483974516392, + 0.008246141485869884, + 0.009044433012604713, + 0.06860994547605515, + 0.012301137670874596, + 0.018771592527627945, + 0.07977709174156189, + 0.021282006055116653, + 0.01238013431429863, + 0.011603625491261482, + 0.011670577339828014, + 0.01197804044932127, + 0.01116032525897026, + 0.014483964070677757, + 0.012103380635380745, + 0.013926778919994831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.029792767018079758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro-probe shot dynamically follows a perfectly spherical drop of viscous black engine oil as it travels through the complex, roaring heart of an engine. The camera tracks smoothly, keeping the glistening droplet centered as it pushes through the narrow, machined steel passageway. The oil's mirror-like surface reflects twisting, elongated images of interconnected gears and the fiery orange glow of a distant burn. It slides steadily along the metal wall, its path illuminated by the harsh, specular highlights of polished surfaces and the oil itself. The environment is a claustrophobic maze of metallic grays and blacks, punctuated by the crimson of overheated components, creating a visible heat haze. As the droplet deftly navigates sharp corners, it briefly stretches and elongates before surface tension pulls it back into a sphere, revealing its viscosity. The entire scene vibrates with the immense power of the engine, creating an atmosphere of tension, mechanical power, and power. The camera follows the droplet's relentless, fluid movement through this complex, high-pressure world.", + "seed": 19154, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_8369_seed_19154_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07446433566510677, + "vlm_logprob": -0.07446433566510677, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.07446433566510677, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1346079409122467 + ], + "opposite_logprob_tokens": [ + -2.134608030319214 + ], + "logprob_reward": -0.1346079409122467, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1346079409122467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09728474915027618, + 0.018737556412816048, + 0.033650804311037064, + 0.020634252578020096, + 0.07187929004430771, + 0.09279444813728333, + 0.016807062551379204, + 0.012343322858214378, + 0.1131906807422638, + 0.5226778388023376 + ], + "credit_map": [ + 0.004671868402510881, + 0.01257525198161602, + 0.004120133351534605, + 0.017817605286836624, + 0.019170530140399933, + 0.014200720936059952, + 0.007105379365384579, + 0.017623260617256165, + 0.0014236519346013665, + 0.003297129413112998, + 0.0019740930292755365, + 0.00290068332105875, + 0.0029174955561757088, + 0.0026081427931785583, + 0.0016267906175926328, + 0.0019895685836672783, + 0.007691273000091314, + 0.004482846707105637, + 0.0036534566897898912, + 0.0021431134082376957, + 0.00313530582934618, + 0.003323498647660017, + 0.004322926979511976, + 0.004898378625512123, + 0.001814420917071402, + 0.004920610226690769, + 0.0012763435952365398, + 0.0031983598601073027, + 0.0023590202908962965, + 0.002360487589612603, + 0.0025070509873330593, + 0.0021979575976729393, + 0.007029418833553791, + 0.01232218462973833, + 0.010428654961287975, + 0.007461024913936853, + 0.009123858995735645, + 0.010316932573914528, + 0.007316862232983112, + 0.007880354300141335, + 0.024728382006287575, + 0.004428096581250429, + 0.009472289122641087, + 0.012831652536988258, + 0.004417280200868845, + 0.02727958932518959, + 0.0033181284088641405, + 0.006319036241620779, + 0.0026723549235612154, + 0.0017608245834708214, + 0.001991813536733389, + 0.0014650913653895259, + 0.0016124062240123749, + 0.0016413646517321467, + 0.0011411227751523256, + 0.004522084258496761, + 0.0013255438534542918, + 0.0021218552719801664, + 0.0007885824306868017, + 0.0013286336325109005, + 0.002038293518126011, + 0.0017627759370952845, + 0.0010886050295084715, + 0.001889032544568181, + 0.0068003470078110695, + 0.009517132304608822, + 0.009035887196660042, + 0.008103260770440102, + 0.02359452098608017, + 0.010383539833128452, + 0.013055358082056046, + 0.032700635492801666, + 0.019960617646574974, + 0.019887631759047508, + 0.02643710747361183, + 0.016869772225618362, + 0.021481363102793694, + 0.21357259154319763, + 0.025919195264577866, + 0.17854957282543182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.1346079409122467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006482779979706 + ], + "opposite_logprob_tokens": [ + -2.6950647830963135 + ], + "logprob_reward": -0.07006482779979706, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006482779979706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08831548690795898, + 0.05836763232946396, + 0.08058781921863556, + 0.028442950919270515, + 0.07192127406597137, + 0.1070842519402504, + 0.017200419679284096, + 0.028453616425395012, + 0.12056562304496765, + 0.3990609049797058 + ], + "credit_map": [ + 0.00593749713152647, + 0.011863844469189644, + 0.0064294710755348206, + 0.012659634463489056, + 0.009432155638933182, + 0.015210197307169437, + 0.009399732574820518, + 0.017382964491844177, + 0.0077013177797198296, + 0.010548634454607964, + 0.00668207136914134, + 0.00552820460870862, + 0.005183191038668156, + 0.009396682493388653, + 0.004608752205967903, + 0.008718781173229218, + 0.016065577045083046, + 0.00794120505452156, + 0.007948556914925575, + 0.006798162125051022, + 0.006522198207676411, + 0.0077628446742892265, + 0.008141512051224709, + 0.01940777152776718, + 0.0026065984275192022, + 0.006903588306158781, + 0.0021953508257865906, + 0.0032661997247487307, + 0.002165905199944973, + 0.005561691243201494, + 0.002436908194795251, + 0.0033067092299461365, + 0.00880690198391676, + 0.007894877344369888, + 0.017673753201961517, + 0.0065970588475465775, + 0.006398504134267569, + 0.01069076918065548, + 0.005491036456078291, + 0.008368374779820442, + 0.0173842441290617, + 0.0047253151424229145, + 0.006655583158135414, + 0.028651101514697075, + 0.00583904841914773, + 0.033030685037374496, + 0.004666063003242016, + 0.0061322166584432125, + 0.003137481166049838, + 0.0028033091221004725, + 0.001292831264436245, + 0.0015173513675108552, + 0.0023031006567180157, + 0.0016248542815446854, + 0.0012355619110167027, + 0.003285932121798396, + 0.002128690481185913, + 0.005868026055395603, + 0.0014743782812729478, + 0.003231876529753208, + 0.004825879819691181, + 0.004188557155430317, + 0.0030376899521797895, + 0.003698520129546523, + 0.008540702052414417, + 0.012177610769867897, + 0.008694472722709179, + 0.008418458513915539, + 0.02518950216472149, + 0.010728535242378712, + 0.010478719137609005, + 0.036337628960609436, + 0.019570961594581604, + 0.014482728205621243, + 0.014561791904270649, + 0.012992064468562603, + 0.017225485295057297, + 0.14422433078289032, + 0.0221434086561203, + 0.1538601666688919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.07006482779979706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710002064704895 + ], + "opposite_logprob_tokens": [ + -2.1270999908447266 + ], + "logprob_reward": -0.12710002064704895, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710002064704895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011933849193155766, + 0.006836813874542713, + 0.03204837813973427, + 0.005208359565585852, + 0.01316855289041996, + 0.02917318418622017, + 0.004048631060868502, + 0.006643444299697876, + 0.11645594239234924, + 0.7744828462600708 + ], + "credit_map": [ + 0.0009059884469024837, + 0.0015312342438846827, + 0.0009368149912916124, + 0.002377118216827512, + 0.0014958849642425776, + 0.001668876619078219, + 0.0012116909492760897, + 0.0018062419258058071, + 0.0003425387549214065, + 0.0018158318707719445, + 0.00049241614760831, + 0.0009429650963284075, + 0.0009842336876317859, + 0.000973440008237958, + 0.0008514579967595637, + 0.0004339306033216417, + 0.006493634544312954, + 0.005873287562280893, + 0.003992429003119469, + 0.001654481515288353, + 0.0032146386802196503, + 0.003034287365153432, + 0.002827099058777094, + 0.004958520643413067, + 0.0004267244949005544, + 0.0013401450123637915, + 0.0005966657190583646, + 0.000607624591793865, + 0.0004533307219389826, + 0.0007780130254104733, + 0.0004690478672273457, + 0.0005368079291656613, + 0.001715474994853139, + 0.0010679877595975995, + 0.002429843647405505, + 0.00146994658280164, + 0.0012855413369834423, + 0.0027194679714739323, + 0.0011523208813741803, + 0.0013279691338539124, + 0.008413976989686489, + 0.0014697135193273425, + 0.0023063491098582745, + 0.006221903953701258, + 0.0019844300113618374, + 0.004611451644450426, + 0.001860956777818501, + 0.0023044003173708916, + 0.0006108686211518943, + 0.00039799275691621006, + 0.0004911086289212108, + 0.0005098787369206548, + 0.0005133998347446322, + 0.00048642049659974873, + 0.0003478934522718191, + 0.0006910685915499926, + 0.0005992138758301735, + 0.0015381801640614867, + 0.0005573149537667632, + 0.0006399535341188312, + 0.0011439906666055322, + 0.0009087364887818694, + 0.0006527198129333556, + 0.0006033353274688125, + 0.008876502513885498, + 0.012352987192571163, + 0.005724131595343351, + 0.007432273589074612, + 0.022913508117198944, + 0.010177281685173512, + 0.010618571192026138, + 0.03836068883538246, + 0.028940921649336815, + 0.029224714264273643, + 0.0706285759806633, + 0.027333807200193405, + 0.031002338975667953, + 0.31969669461250305, + 0.035370636731386185, + 0.23228517174720764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.12710002064704895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201881363987923 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014201881363987923, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201881363987923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08086708933115005, + 0.023778505623340607, + 0.08114814758300781, + 0.042135581374168396, + 0.02943011373281479, + 0.06998846679925919, + 0.008945695124566555, + 0.045286331325769424, + 0.11108088493347168, + 0.5073392987251282 + ], + "credit_map": [ + 0.0034774376545101404, + 0.011674551293253899, + 0.0058441804721951485, + 0.014977216720581055, + 0.016037629917263985, + 0.011507310904562473, + 0.00626334547996521, + 0.01108541339635849, + 0.0011602932354435325, + 0.007429220248013735, + 0.0018699932843446732, + 0.0029370870906859636, + 0.003526530461385846, + 0.0034011700190603733, + 0.002077375538647175, + 0.0013768353965133429, + 0.01673416793346405, + 0.013083258643746376, + 0.010653873905539513, + 0.004499274771660566, + 0.007478067185729742, + 0.007263291161507368, + 0.008081973530352116, + 0.013354233466088772, + 0.004126309882849455, + 0.01304408349096775, + 0.003168182447552681, + 0.004836671520024538, + 0.002869764342904091, + 0.006826409604400396, + 0.003198296297341585, + 0.004065860994160175, + 0.002413000911474228, + 0.004333050921559334, + 0.0037139926571398973, + 0.003451660042628646, + 0.004207433667033911, + 0.004000589717179537, + 0.004268050659447908, + 0.003042330499738455, + 0.007605207618325949, + 0.0035253583919256926, + 0.007399581838399172, + 0.020340239629149437, + 0.004421703051775694, + 0.01943046972155571, + 0.003217922756448388, + 0.0040479726158082485, + 0.001649515237659216, + 0.0008039832464419305, + 0.0009079200099222362, + 0.0009156096493825316, + 0.0010216630762442946, + 0.0006805713637731969, + 0.0011334255104884505, + 0.0018330055754631758, + 0.0033654191065579653, + 0.01073917280882597, + 0.0032387145329266787, + 0.004971436224877834, + 0.007083947770297527, + 0.0068875630386173725, + 0.004038758110255003, + 0.004961317405104637, + 0.008655262179672718, + 0.014452669769525528, + 0.008129501715302467, + 0.0072889551520347595, + 0.015932997688651085, + 0.008559021167457104, + 0.012948235496878624, + 0.03511423617601395, + 0.020366555079817772, + 0.019258713349699974, + 0.014818006195127964, + 0.01595417968928814, + 0.021068688482046127, + 0.20544975996017456, + 0.025611363351345062, + 0.1848120242357254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.014201881363987923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347007602453232 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.026347007602453232, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347007602453232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08552094548940659, + 0.035982657223939896, + 0.03793781250715256, + 0.04427063837647438, + 0.043684571981430054, + 0.31977221369743347, + 0.06935090571641922, + 0.023935819044709206, + 0.167766734957695, + 0.1717776656150818 + ], + "credit_map": [ + 0.009534222073853016, + 0.007491602096706629, + 0.008050680160522461, + 0.018704915419220924, + 0.007691436912864447, + 0.008877002634108067, + 0.00611441396176815, + 0.019056672230362892, + 0.0033055106177926064, + 0.01015942171216011, + 0.0028460160829126835, + 0.0036913412623107433, + 0.004961196333169937, + 0.00419987877830863, + 0.0031285174190998077, + 0.00369077711366117, + 0.006992923095822334, + 0.006760253105312586, + 0.0034769545309245586, + 0.0033094841055572033, + 0.004028692375868559, + 0.00395972328260541, + 0.0032020530197769403, + 0.006207730621099472, + 0.0046464381739497185, + 0.012270811013877392, + 0.00458320789039135, + 0.004613772965967655, + 0.0034872894175350666, + 0.007217386271804571, + 0.0031721126288175583, + 0.004279621876776218, + 0.003525904845446348, + 0.006850962992757559, + 0.006062687840312719, + 0.005195489153265953, + 0.005699302535504103, + 0.007386110723018646, + 0.0035394111182540655, + 0.005424699280411005, + 0.03141029551625252, + 0.01577145792543888, + 0.03207556903362274, + 0.09664113074541092, + 0.01970072090625763, + 0.08991839736700058, + 0.016735585406422615, + 0.017519084736704826, + 0.01472705788910389, + 0.0060439822264015675, + 0.007783678360283375, + 0.006979113444685936, + 0.005518822465091944, + 0.007818521931767464, + 0.005829859524965286, + 0.014649865217506886, + 0.002680846955627203, + 0.003199451370164752, + 0.002057089004665613, + 0.002618626458570361, + 0.002769863000139594, + 0.0036150997038930655, + 0.0020851921290159225, + 0.004909649025648832, + 0.009197333827614784, + 0.028424449265003204, + 0.016504531726241112, + 0.018874941393733025, + 0.03828023746609688, + 0.018059680238366127, + 0.018840225413441658, + 0.01958533003926277, + 0.014614498242735863, + 0.005569913424551487, + 0.03509404882788658, + 0.008844837546348572, + 0.010051919147372246, + 0.0605405755341053, + 0.009725657291710377, + 0.027336210012435913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.026347007602453232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro-probe shot dynamically follows a perfectly spherical drop of viscous black engine oil as it travels through the complex, roaring heart of an engine. The camera tracks smoothly, keeping the glistening droplet centered as it pushes through the narrow, machined steel passageway. The oil's mirror-like surface reflects twisting, elongated images of interconnected gears and the fiery orange glow of a distant burn. It slides steadily along the metal wall, its path illuminated by the harsh, specular highlights of polished surfaces and the oil itself. The environment is a claustrophobic maze of metallic grays and blacks, punctuated by the crimson of overheated components, creating a visible heat haze. As the droplet deftly navigates sharp corners, it briefly stretches and elongates before surface tension pulls it back into a sphere, revealing its viscosity. The entire scene vibrates with the immense power of the engine, creating an atmosphere of tension, mechanical power, and power. The camera follows the droplet's relentless, fluid movement through this complex, high-pressure world.", + "seed": 19155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_13/index_8369_seed_19155_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.391607678309083, + "vlm_logprob": -0.391607678309083, + "vlm_yes_no_margin": 1.85, + "vlm_token_credit": -0.391607678309083, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31826385855674744 + ], + "opposite_logprob_tokens": [ + -1.3182638883590698 + ], + "logprob_reward": -0.31826385855674744, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31826385855674744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062472134828567505, + 0.08186142146587372, + 0.05870171636343002, + 0.1324877142906189, + 0.08304722607135773, + 0.12164425849914551, + 0.11836829781532288, + 0.19291207194328308, + 0.12232162803411484, + 0.02618349716067314 + ], + "credit_map": [ + 0.006940644234418869, + 0.007283193524926901, + 0.0065538035705685616, + 0.008400795981287956, + 0.008573316968977451, + 0.010740404948592186, + 0.008049950003623962, + 0.005930021870881319, + 0.014085538685321808, + 0.009797457605600357, + 0.014199239201843739, + 0.011001257225871086, + 0.008998050354421139, + 0.007412972394376993, + 0.010296043939888477, + 0.006070862989872694, + 0.012615539133548737, + 0.0063913557678461075, + 0.00480409013107419, + 0.004176134243607521, + 0.01147629227489233, + 0.0058518145233392715, + 0.00890330970287323, + 0.004483177792280912, + 0.01368191372603178, + 0.014143484644591808, + 0.017048457637429237, + 0.015527633018791676, + 0.028935818001627922, + 0.014241155236959457, + 0.012451408430933952, + 0.01645784080028534, + 0.018018420785665512, + 0.013461610302329063, + 0.007505079265683889, + 0.009727600030601025, + 0.013580136001110077, + 0.0062095411121845245, + 0.0053917705081403255, + 0.00915306806564331, + 0.014155816286802292, + 0.010858535766601562, + 0.014838374219834805, + 0.020619286224246025, + 0.018530478700995445, + 0.010665541514754295, + 0.015781914815306664, + 0.016194313764572144, + 0.008234371431171894, + 0.01568903774023056, + 0.017325518652796745, + 0.03215411677956581, + 0.011559979058802128, + 0.008613488636910915, + 0.009990384802222252, + 0.014801396057009697, + 0.010386400856077671, + 0.007938435301184654, + 0.058915868401527405, + 0.01033700443804264, + 0.004521574359387159, + 0.02082885429263115, + 0.005057378672063351, + 0.07492655515670776, + 0.023003170266747475, + 0.011255921795964241, + 0.025432327762246132, + 0.011945335194468498, + 0.005695259664207697, + 0.013689566403627396, + 0.011776789091527462, + 0.019523246213793755, + 0.0036406591534614563, + 0.0027976450510323048, + 0.0035875614266842604, + 0.002920429687947035, + 0.0032914327457547188, + 0.0030306235421448946, + 0.0028001335449516773, + 0.004115011543035507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.31826385855674744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062087271362543106 + ], + "opposite_logprob_tokens": [ + -2.812087297439575 + ], + "logprob_reward": -0.062087271362543106, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062087271362543106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08344028145074844, + 0.1211981400847435, + 0.042271967977285385, + 0.050499290227890015, + 0.03143538907170296, + 0.2336641252040863, + 0.10104524344205856, + 0.24206997454166412, + 0.04655769094824791, + 0.04781787842512131 + ], + "credit_map": [ + 0.009528455324470997, + 0.009935867972671986, + 0.010700566694140434, + 0.008682440966367722, + 0.010819034650921822, + 0.01709224469959736, + 0.01088889129459858, + 0.00579279288649559, + 0.019475016742944717, + 0.011710083112120628, + 0.02457277476787567, + 0.01968265511095524, + 0.015388729982078075, + 0.009073949418962002, + 0.012283504009246826, + 0.009011439979076385, + 0.004688256420195103, + 0.007027398329228163, + 0.0056819552555680275, + 0.004561249632388353, + 0.004865850321948528, + 0.006107760593295097, + 0.004889203701168299, + 0.004450296983122826, + 0.0053366986103355885, + 0.0047096144407987595, + 0.006354648619890213, + 0.006897950079292059, + 0.009936727583408356, + 0.005918156821280718, + 0.004788532387465239, + 0.006556968670338392, + 0.007680267095565796, + 0.004379387013614178, + 0.0032381643541157246, + 0.0041936407797038555, + 0.004744570702314377, + 0.001982730580493808, + 0.0024099627044051886, + 0.0028066711965948343, + 0.02634616382420063, + 0.02178717963397503, + 0.02844231389462948, + 0.03936835378408432, + 0.03202611207962036, + 0.018956810235977173, + 0.03486034274101257, + 0.03187686949968338, + 0.007287580985575914, + 0.014271237887442112, + 0.013383425772190094, + 0.02433007024228573, + 0.009404275566339493, + 0.009110739454627037, + 0.011436321772634983, + 0.011821595951914787, + 0.015402836725115776, + 0.009843091480433941, + 0.0742032453417778, + 0.01243172213435173, + 0.0054438053630292416, + 0.025796953588724136, + 0.005788153503090143, + 0.09316018968820572, + 0.00855872593820095, + 0.004912738688290119, + 0.008808963000774384, + 0.005112078506499529, + 0.0031850773375481367, + 0.00526778306812048, + 0.004768949002027512, + 0.0059433807618916035, + 0.013138415291905403, + 0.0041992999613285065, + 0.0029778066091239452, + 0.0026689087972044945, + 0.0027229017578065395, + 0.00449691666290164, + 0.0026359562762081623, + 0.014977678656578064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.062087271362543106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502109169960022 + ], + "opposite_logprob_tokens": [ + -0.252109169960022 + ], + "logprob_reward": -1.502109169960022, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502109169960022, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09073041379451752, + 0.1071525514125824, + 0.055327724665403366, + 0.15223030745983124, + 0.04115832969546318, + 0.07407782971858978, + 0.15330909192562103, + 0.12993744015693665, + 0.15945585072040558, + 0.036620449274778366 + ], + "credit_map": [ + 0.010609256103634834, + 0.011448659002780914, + 0.012582242488861084, + 0.01275313924998045, + 0.013929166831076145, + 0.01186315342783928, + 0.01084686815738678, + 0.006697928998619318, + 0.010237306356430054, + 0.016111226752400398, + 0.017841773107647896, + 0.00892846379429102, + 0.023243363946676254, + 0.009904395788908005, + 0.012337962165474892, + 0.008548056706786156, + 0.005483932327479124, + 0.009267479181289673, + 0.0053218756802380085, + 0.009363858960568905, + 0.007976524531841278, + 0.005301521625369787, + 0.008298520930111408, + 0.004314013756811619, + 0.016412906348705292, + 0.01611252874135971, + 0.018642347306013107, + 0.019454680383205414, + 0.027802204713225365, + 0.018843891099095345, + 0.01421512570232153, + 0.020746607333421707, + 0.007394629530608654, + 0.006238441914319992, + 0.0036268362309783697, + 0.006644859444350004, + 0.006382608320564032, + 0.0033008656464517117, + 0.003101138398051262, + 0.004468947183340788, + 0.008465884253382683, + 0.006589309778064489, + 0.00924634002149105, + 0.011726063676178455, + 0.00939438957720995, + 0.006605932489037514, + 0.010865841060876846, + 0.01118407305330038, + 0.014047283679246902, + 0.02087598852813244, + 0.020863385871052742, + 0.039093922823667526, + 0.013593742623925209, + 0.013178537599742413, + 0.014200734905898571, + 0.01745550148189068, + 0.00712190754711628, + 0.005269746761769056, + 0.04003158584237099, + 0.006301210727542639, + 0.002823897870257497, + 0.014248847030103207, + 0.00353448954410851, + 0.05060576647520065, + 0.02559337392449379, + 0.016077188774943352, + 0.027372311800718307, + 0.012709991075098515, + 0.008565760217607021, + 0.01666080765426159, + 0.014327922835946083, + 0.03814850002527237, + 0.0051154750399291515, + 0.004359905142337084, + 0.0071376352570950985, + 0.003976437263190746, + 0.0027910121716558933, + 0.005463040433824062, + 0.0033825146965682507, + 0.0043944320641458035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -1.502109169960022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607998594641685 + ], + "opposite_logprob_tokens": [ + -3.8956079483032227 + ], + "logprob_reward": -0.020607998594641685, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607998594641685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11035453528165817, + 0.07353883236646652, + 0.06251327693462372, + 0.041204094886779785, + 0.04563155770301819, + 0.09471961110830307, + 0.04598245769739151, + 0.3676340579986572, + 0.08748985826969147, + 0.07093176245689392 + ], + "credit_map": [ + 0.016284845769405365, + 0.021619098260998726, + 0.012623325921595097, + 0.01334378868341446, + 0.018863167613744736, + 0.010050796903669834, + 0.011065150611102581, + 0.006504358258098364, + 0.008091570809483528, + 0.009945917874574661, + 0.009663557633757591, + 0.008616366423666477, + 0.012376807630062103, + 0.009630507789552212, + 0.010364214889705181, + 0.004849892109632492, + 0.010812553577125072, + 0.00795846525579691, + 0.004831382539123297, + 0.0052871801890432835, + 0.01168505847454071, + 0.008919171057641506, + 0.007336295675486326, + 0.005683168303221464, + 0.004751103930175304, + 0.005948803387582302, + 0.005524173844605684, + 0.004575117025524378, + 0.007412536535412073, + 0.005135546904057264, + 0.003418279578909278, + 0.0044385334476828575, + 0.011506330221891403, + 0.00526565732434392, + 0.00430734409019351, + 0.006505157332867384, + 0.0054448451846838, + 0.003486029803752899, + 0.00387482438236475, + 0.005241366568952799, + 0.010013976134359837, + 0.00885842926800251, + 0.011508905328810215, + 0.014511330053210258, + 0.012959271669387817, + 0.008323400281369686, + 0.015248887240886688, + 0.013295415788888931, + 0.0057295868173241615, + 0.0076324655674397945, + 0.004743762314319611, + 0.008460325188934803, + 0.004581490531563759, + 0.003355399938300252, + 0.005025979597121477, + 0.00645345076918602, + 0.01971355825662613, + 0.014005448669195175, + 0.1114140972495079, + 0.020449712872505188, + 0.009024410508573055, + 0.0400567464530468, + 0.010507267899811268, + 0.142462819814682, + 0.008987796492874622, + 0.009651235304772854, + 0.008793995715677738, + 0.006968030706048012, + 0.005143365357071161, + 0.012924594804644585, + 0.008168879896402359, + 0.026851952075958252, + 0.012869694270193577, + 0.0072058397345244884, + 0.01020366232842207, + 0.005258150864392519, + 0.00388344400562346, + 0.01052887737751007, + 0.006255711894482374, + 0.014726383611559868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.020607998594641685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a perfectly spherical drop of viscous black engine oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a complex engine interior with machined steel passageways and gears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature fiery orange glow and crimson overheated components?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil droplet stretch and elongate when navigating corners before returning to a sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly, keeping the droplet centered throughout the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054970093071460724 + ], + "opposite_logprob_tokens": [ + -2.9299700260162354 + ], + "logprob_reward": -0.054970093071460724, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054970093071460724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06068359687924385, + 0.0468776673078537, + 0.10417439043521881, + 0.021937690675258636, + 0.025716213509440422, + 0.07053835690021515, + 0.1390564739704132, + 0.43566176295280457, + 0.046450428664684296, + 0.04890340939164162 + ], + "credit_map": [ + 0.006790099199861288, + 0.010441030375659466, + 0.00739265838637948, + 0.008260111324489117, + 0.008244428783655167, + 0.006700947415083647, + 0.008569092489778996, + 0.004285228438675404, + 0.007690638769418001, + 0.0056810034438967705, + 0.00873799528926611, + 0.005165935959666967, + 0.007187620736658573, + 0.004257575608789921, + 0.004587934818118811, + 0.0035689612850546837, + 0.016422003507614136, + 0.014256307855248451, + 0.009871633723378181, + 0.011030133813619614, + 0.015477250330150127, + 0.015345510095357895, + 0.010464890860021114, + 0.01130666770040989, + 0.002783045871183276, + 0.002611612668260932, + 0.0028528166003525257, + 0.0019652259070426226, + 0.0036655531730502844, + 0.0026776946615427732, + 0.0029632996302098036, + 0.0024184435606002808, + 0.0033008211757987738, + 0.003746945410966873, + 0.002825184492394328, + 0.0043836855329573154, + 0.003555671777576208, + 0.00240509957075119, + 0.002227083081379533, + 0.003271721536293626, + 0.006305317394435406, + 0.007135762367397547, + 0.008720566518604755, + 0.01351669430732727, + 0.009952708147466183, + 0.005729181692004204, + 0.0098961116746068, + 0.00928201712667942, + 0.010812024585902691, + 0.018360113725066185, + 0.018018051981925964, + 0.034076277166604996, + 0.013723745010793209, + 0.014230224303901196, + 0.01456213928759098, + 0.015273887664079666, + 0.024834437295794487, + 0.018816493451595306, + 0.13306903839111328, + 0.023695534095168114, + 0.009417861700057983, + 0.04714377596974373, + 0.010457348078489304, + 0.16822732985019684, + 0.009295287542045116, + 0.004264540504664183, + 0.01006657537072897, + 0.004882731940597296, + 0.0030274130403995514, + 0.004583851434290409, + 0.0045355320908129215, + 0.0057944948785007, + 0.012944437563419342, + 0.004673836287111044, + 0.004188947845250368, + 0.0027613926213234663, + 0.002965263556689024, + 0.005575047340244055, + 0.0020970902405679226, + 0.013697395101189613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.054970093071460724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_14/index_tensor([8759, 8759])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_14/index_tensor([8759, 8759])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a8a03113cbe8ef328d9b210e33a6baddfe71bc0d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_14/index_tensor([8759, 8759])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "documentary-style medium close-up shot, shot from a level angle and with a slight handheld shakiness, focuses on a young boy, approximately six or seven years old, standing alone amidst a ruined landscape. His cheeks are stained with a mixture of dust and dried earth, his short, messy black hair matted with grayish-white dust. He wears a worn, oversized T-shirt, faded to a gray. He hunches slightly, his lips pressed together as if trying to stifle sobs. But large, glistening tears silently fall from his reddened eyes, leaving two distinct wet streaks along his stained cheeks. His vacant eyes stare at a point beyond the camera, filled with a sadness and confusion beyond his years. Behind him, out-of-focus rubble clung to the ground, twisted rebar protruding from shattered concrete blocks, and rubble and debris littered the ground. A cool, soft beam of natural light, streaming through a sky obscured by smoke and dust, obliquely illuminates half of his face, casting deep, soft shadows that highlight the moist sheen of tear marks and the rough texture of the earth. The overall color palette of the painting tends toward cool grays and earthy yellows, creating an atmosphere of oppression, heaviness, and profound realism.", + "seed": 7458, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_14/index_8759_seed_7458_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03791235298849642, + "vlm_logprob": -0.03791235298849642, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03791235298849642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018532924354076385 + ], + "opposite_logprob_tokens": [ + -5.143532752990723 + ], + "logprob_reward": -0.018532924354076385, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.018532924354076385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044397685676813126, + 0.01339490246027708, + 0.009706425480544567, + 0.017827872186899185, + 0.020730143412947655, + 0.037352174520492554, + 0.011888880282640457, + 0.014019181951880455, + 0.4242781102657318, + 0.40640467405319214 + ], + "credit_map": [ + 0.0047149271704256535, + 0.004043560940772295, + 0.00513914879411459, + 0.006198009010404348, + 0.007349174469709396, + 0.0041817896999418736, + 0.005597677081823349, + 0.007173402234911919, + 0.0011716533917933702, + 0.001276492141187191, + 0.0014233263209462166, + 0.002122285310178995, + 0.0012455264804884791, + 0.0016096471808850765, + 0.001453407690860331, + 0.003092565108090639, + 0.000653675349894911, + 0.000761093629989773, + 0.0011797109618782997, + 0.0012662530643865466, + 0.0019483704818412662, + 0.0013228277675807476, + 0.0014680547174066305, + 0.0011064392747357488, + 0.0013054332230240107, + 0.0015394343063235283, + 0.0015297627542167902, + 0.0023643379099667072, + 0.003342187497764826, + 0.0030977509450167418, + 0.0031548086553812027, + 0.0014941568952053785, + 0.0020743361674249172, + 0.0016434104181826115, + 0.0025595524348318577, + 0.0023517021909356117, + 0.00258164806291461, + 0.002121909987181425, + 0.0030954298563301563, + 0.004302157089114189, + 0.004005591385066509, + 0.004658176098018885, + 0.006164737045764923, + 0.006051020696759224, + 0.002772952662780881, + 0.00685590784996748, + 0.0033115146216005087, + 0.003532277885824442, + 0.0010598680237308145, + 0.0011311942944303155, + 0.0020122092682868242, + 0.0022351096849888563, + 0.0016032833373174071, + 0.00137324386741966, + 0.0011201795423403382, + 0.0013537936611101031, + 0.001502145198173821, + 0.0019268791656941175, + 0.0022522606886923313, + 0.002022850327193737, + 0.0015418494585901499, + 0.0014671023236587644, + 0.0016516244504600763, + 0.001654472085647285, + 0.08971550315618515, + 0.011567487381398678, + 0.02528846636414528, + 0.03632614016532898, + 0.030881427228450775, + 0.18268346786499023, + 0.0165215153247118, + 0.031294114887714386, + 0.09486562013626099, + 0.010910269804298878, + 0.020443545654416084, + 0.033686231821775436, + 0.024528395384550095, + 0.17852617800235748, + 0.014980961568653584, + 0.02846345119178295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.018532924354076385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014235729351639748 + ], + "opposite_logprob_tokens": [ + -4.264235496520996 + ], + "logprob_reward": -0.014235729351639748, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014235729351639748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02169066108763218, + 0.010684859938919544, + 0.006242775358259678, + 0.01993020623922348, + 0.008899245411157608, + 0.008428951725363731, + 0.008218432776629925, + 0.005489427596330643, + 0.1287883222103119, + 0.7816270589828491 + ], + "credit_map": [ + 0.0018905355827882886, + 0.0018916859989985824, + 0.0038440313655883074, + 0.0036644998472183943, + 0.002034669043496251, + 0.0029068742878735065, + 0.0023161766584962606, + 0.003142188536003232, + 0.001069274265319109, + 0.0009777280502021313, + 0.0014073011698201299, + 0.001269343658350408, + 0.0018632275750860572, + 0.0012931213714182377, + 0.001367555116303265, + 0.0014373089652508497, + 0.0005305000813677907, + 0.0005278392927721143, + 0.0011225943453609943, + 0.0009357378003187478, + 0.000565794762223959, + 0.0012060601729899645, + 0.0005541063146665692, + 0.0008001427631825209, + 0.003150500124320388, + 0.0014519765973091125, + 0.0016335936961695552, + 0.00234981719404459, + 0.0014574454398825765, + 0.006385870277881622, + 0.001787838526070118, + 0.0017131641507148743, + 0.00048705533845350146, + 0.0007085675024427474, + 0.0009499155567027628, + 0.0009360240655951202, + 0.0006754297064617276, + 0.002563273534178734, + 0.0008517422247678041, + 0.0017272380646318197, + 0.0008692978881299496, + 0.0010428568348288536, + 0.0012197276810184121, + 0.0011831732699647546, + 0.0012912573292851448, + 0.0011292448034510016, + 0.0006877764244563878, + 0.0010056179016828537, + 0.0009739201050251722, + 0.001057382789440453, + 0.0014009700389578938, + 0.0016470198752358556, + 0.0008790969732217491, + 0.0007270355126820505, + 0.0007075794856064022, + 0.0008254276472143829, + 0.0005017634248360991, + 0.0007634646608494222, + 0.0007343580364249647, + 0.0007962023373693228, + 0.0009735540370456874, + 0.0004886381211690605, + 0.0005817713099531829, + 0.000649675726890564, + 0.025812868028879166, + 0.010324698872864246, + 0.008954990655183792, + 0.010842680931091309, + 0.01070843730121851, + 0.0463302880525589, + 0.006540309172123671, + 0.009274047799408436, + 0.1802210956811905, + 0.02490103803575039, + 0.04275844618678093, + 0.0651213601231575, + 0.047111790627241135, + 0.3383980095386505, + 0.027735905721783638, + 0.05537942796945572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.014235729351639748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497494712471962 + ], + "opposite_logprob_tokens": [ + -2.9299750328063965 + ], + "logprob_reward": -0.05497494712471962, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497494712471962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13088615238666534, + 0.0511045902967453, + 0.027661509811878204, + 0.04045489430427551, + 0.0791848748922348, + 0.081509530544281, + 0.035895440727472305, + 0.020081646740436554, + 0.24341876804828644, + 0.28980258107185364 + ], + "credit_map": [ + 0.011230891570448875, + 0.010690978728234768, + 0.01604907400906086, + 0.020181383937597275, + 0.021046491339802742, + 0.01291348971426487, + 0.014458988793194294, + 0.02431485801935196, + 0.005843297578394413, + 0.004797397647053003, + 0.005681550595909357, + 0.0066603487357497215, + 0.010207148268818855, + 0.004547568969428539, + 0.005921348929405212, + 0.007445927243679762, + 0.0023101717233657837, + 0.0029859922360628843, + 0.0029756459407508373, + 0.0040025413036346436, + 0.003546219551935792, + 0.0045262412168085575, + 0.0032379820477217436, + 0.004076716955751181, + 0.003999293781816959, + 0.005069744307547808, + 0.0038086548447608948, + 0.0042623369954526424, + 0.0053993393667042255, + 0.009198443032801151, + 0.0036648425739258528, + 0.005052241031080484, + 0.005339405033737421, + 0.0045007881708443165, + 0.005413794424384832, + 0.006739876698702574, + 0.006170734763145447, + 0.028716206550598145, + 0.0035389296244829893, + 0.018765147775411606, + 0.009692360647022724, + 0.01448539737612009, + 0.010575762949883938, + 0.009455418214201927, + 0.008643629029393196, + 0.01132491696625948, + 0.0056134117767214775, + 0.01171863079071045, + 0.004258906003087759, + 0.003928735852241516, + 0.006800828035920858, + 0.007744636852294207, + 0.0030817026272416115, + 0.003448977367952466, + 0.0033361902460455894, + 0.00329546257853508, + 0.0022722589783370495, + 0.0025276977103203535, + 0.0027415973600000143, + 0.002906878711655736, + 0.0040552010759711266, + 0.0016885367222130299, + 0.002150213113054633, + 0.0017392621375620365, + 0.03662574663758278, + 0.013162451796233654, + 0.014161917380988598, + 0.02005855180323124, + 0.023187393322587013, + 0.09229595959186554, + 0.017703494057059288, + 0.026223259046673775, + 0.05579704791307449, + 0.014212182722985744, + 0.018137173727154732, + 0.02126343548297882, + 0.02377512864768505, + 0.10638201981782913, + 0.031404200941324234, + 0.018831392750144005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.05497494712471962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015161458868533373 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015161458868533373, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015161458868533373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09268449246883392, + 0.04608859121799469, + 0.02814600057899952, + 0.027700042352080345, + 0.04351220652461052, + 0.02435239776968956, + 0.03046063333749771, + 0.018465587869286537, + 0.14265218377113342, + 0.5459378361701965 + ], + "credit_map": [ + 0.009401082992553711, + 0.010305027477443218, + 0.013504218310117722, + 0.016345910727977753, + 0.009087019599974155, + 0.009550584480166435, + 0.011513764038681984, + 0.012976891361176968, + 0.005278730299323797, + 0.004837123677134514, + 0.007782510481774807, + 0.007881141267716885, + 0.003945112694054842, + 0.004858560860157013, + 0.00518620852380991, + 0.0063192034140229225, + 0.00196011527441442, + 0.0027373034972697496, + 0.003902320982888341, + 0.004646648652851582, + 0.006110344082117081, + 0.003124609589576721, + 0.0031172307208180428, + 0.002547427313402295, + 0.0032098826486617327, + 0.002865525195375085, + 0.002571661490947008, + 0.002524221781641245, + 0.0030726813711225986, + 0.006755817215889692, + 0.003950260579586029, + 0.002749990439042449, + 0.003624582663178444, + 0.003383536823093891, + 0.0047055077739059925, + 0.004269355908036232, + 0.002736875554546714, + 0.014092303812503815, + 0.002325371839106083, + 0.008374673314392567, + 0.002814464271068573, + 0.004571577999740839, + 0.002693929011002183, + 0.0027377947699278593, + 0.0030428562313318253, + 0.004057851154357195, + 0.0019771477673202753, + 0.002456777961924672, + 0.003506388748064637, + 0.004094928968697786, + 0.005029700696468353, + 0.00485300412401557, + 0.0022702678106725216, + 0.005180297419428825, + 0.0025785150937736034, + 0.0029475309420377016, + 0.001799098332412541, + 0.0023688841611146927, + 0.00289500062353909, + 0.003046975936740637, + 0.0024464568123221397, + 0.0017536485102027655, + 0.0018660902278497815, + 0.002289434662088752, + 0.029329869896173477, + 0.005044835153967142, + 0.008742664940655231, + 0.0120591651648283, + 0.010949013754725456, + 0.057905830442905426, + 0.00870015099644661, + 0.00992064829915762, + 0.12839411199092865, + 0.015240826644003391, + 0.02786744013428688, + 0.045773960649967194, + 0.03298620134592056, + 0.2392827421426773, + 0.01821994222700596, + 0.03817262127995491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0015161458868533373 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030201822519302 + ], + "opposite_logprob_tokens": [ + -2.35030198097229 + ], + "logprob_reward": -0.10030201822519302, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030201822519302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17960922420024872, + 0.056260041892528534, + 0.05041227117180824, + 0.0478341318666935, + 0.03208806738257408, + 0.01636364497244358, + 0.029287921264767647, + 0.032268960028886795, + 0.09187507629394531, + 0.46400073170661926 + ], + "credit_map": [ + 0.017690014094114304, + 0.01828121580183506, + 0.022447166964411736, + 0.023265646770596504, + 0.026365024968981743, + 0.019323810935020447, + 0.02138763666152954, + 0.03084869682788849, + 0.005050464998930693, + 0.005053607281297445, + 0.008132364600896835, + 0.008951068855822086, + 0.005165766924619675, + 0.004209808073937893, + 0.006552817765623331, + 0.013144140131771564, + 0.004270670469850302, + 0.005620093550533056, + 0.0069457655772566795, + 0.006210752762854099, + 0.00861056987196207, + 0.006195950321853161, + 0.007134467829018831, + 0.005424000322818756, + 0.006579346023499966, + 0.007016770541667938, + 0.0044920193031430244, + 0.005532200448215008, + 0.006602889858186245, + 0.007173073478043079, + 0.004825822077691555, + 0.005612009204924107, + 0.0025218739174306393, + 0.002365710912272334, + 0.004067691508680582, + 0.002843838417902589, + 0.002355562523007393, + 0.00819313433021307, + 0.0026661711744964123, + 0.007074084598571062, + 0.001687024487182498, + 0.0019273009384050965, + 0.0019853580743074417, + 0.0021314164623618126, + 0.0020859132055193186, + 0.0034078932367265224, + 0.001563087454997003, + 0.001575650298036635, + 0.002982598263770342, + 0.004198581445962191, + 0.002897711703553796, + 0.004042114596813917, + 0.0039124698378145695, + 0.003183302702382207, + 0.005149229895323515, + 0.002921910723671317, + 0.002641508588567376, + 0.0033678768668323755, + 0.0036148731596767902, + 0.004608616698533297, + 0.006818754132837057, + 0.0026257100980728865, + 0.005316810216754675, + 0.0032748058438301086, + 0.007298208773136139, + 0.0060788425616919994, + 0.009235674515366554, + 0.01075858622789383, + 0.01640581712126732, + 0.015908462926745415, + 0.012817570008337498, + 0.013371909037232399, + 0.10872853547334671, + 0.013132931664586067, + 0.023700183257460594, + 0.03860224783420563, + 0.027714386582374573, + 0.20373040437698364, + 0.01582906022667885, + 0.03256295993924141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.10030201822519302 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "documentary-style medium close-up shot, shot from a level angle and with a slight handheld shakiness, focuses on a young boy, approximately six or seven years old, standing alone amidst a ruined landscape. His cheeks are stained with a mixture of dust and dried earth, his short, messy black hair matted with grayish-white dust. He wears a worn, oversized T-shirt, faded to a gray. He hunches slightly, his lips pressed together as if trying to stifle sobs. But large, glistening tears silently fall from his reddened eyes, leaving two distinct wet streaks along his stained cheeks. His vacant eyes stare at a point beyond the camera, filled with a sadness and confusion beyond his years. Behind him, out-of-focus rubble clung to the ground, twisted rebar protruding from shattered concrete blocks, and rubble and debris littered the ground. A cool, soft beam of natural light, streaming through a sky obscured by smoke and dust, obliquely illuminates half of his face, casting deep, soft shadows that highlight the moist sheen of tear marks and the rough texture of the earth. The overall color palette of the painting tends toward cool grays and earthy yellows, creating an atmosphere of oppression, heaviness, and profound realism.", + "seed": 7459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_14/index_8759_seed_7459_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03295752846170217, + "vlm_logprob": -0.03295752846170217, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03295752846170217, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025878744199872017 + ], + "opposite_logprob_tokens": [ + -4.65087890625 + ], + "logprob_reward": -0.025878744199872017, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.025878744199872017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15514758229255676, + 0.07906044274568558, + 0.11837778240442276, + 0.08091815561056137, + 0.05038945749402046, + 0.039919208735227585, + 0.02445804327726364, + 0.02226814441382885, + 0.3385423719882965, + 0.09091874212026596 + ], + "credit_map": [ + 0.01654812879860401, + 0.026147082448005676, + 0.028272533789277077, + 0.024224186316132545, + 0.012937833555042744, + 0.016727827489376068, + 0.014970549382269382, + 0.0153194610029459, + 0.0060061197727918625, + 0.008928157389163971, + 0.007881315425038338, + 0.006362394895404577, + 0.007808569353073835, + 0.024406304582953453, + 0.011517253704369068, + 0.0061503383331000805, + 0.012411251664161682, + 0.015613174997270107, + 0.023440571501851082, + 0.020887713879346848, + 0.011637008748948574, + 0.010790864937007427, + 0.009816576726734638, + 0.013780636712908745, + 0.007038196548819542, + 0.012700016610324383, + 0.008089382201433182, + 0.0068909586407244205, + 0.01564791426062584, + 0.011703434400260448, + 0.010492640547454357, + 0.00835562963038683, + 0.004744496662169695, + 0.011545535176992416, + 0.005412156227976084, + 0.004482430871576071, + 0.004655388183891773, + 0.012415766716003418, + 0.003749263472855091, + 0.003384428098797798, + 0.0030060706194490194, + 0.007172220852226019, + 0.006988156586885452, + 0.0070058926939964294, + 0.005704110022634268, + 0.002894658362492919, + 0.003989502787590027, + 0.0031586005352437496, + 0.0025353305973112583, + 0.003533750306814909, + 0.0038290286902338266, + 0.0029646200127899647, + 0.003160571912303567, + 0.004087742418050766, + 0.002185674151405692, + 0.002161329844966531, + 0.0018883049488067627, + 0.0023065966088324785, + 0.004038816783577204, + 0.0037606500554829836, + 0.003012766595929861, + 0.0024327582214027643, + 0.002346635330468416, + 0.0024816193617880344, + 0.061320461332798004, + 0.0138316098600626, + 0.13254012167453766, + 0.017276782542467117, + 0.06786748766899109, + 0.015736687928438187, + 0.01958438940346241, + 0.010384844616055489, + 0.019828375428915024, + 0.007371360436081886, + 0.019511185586452484, + 0.008976808749139309, + 0.010558033362030983, + 0.010864447802305222, + 0.007510667201131582, + 0.0062978751957416534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.025878744199872017 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01257431972771883 + ], + "opposite_logprob_tokens": [ + -4.387574195861816 + ], + "logprob_reward": -0.01257431972771883, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01257431972771883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03786041587591171, + 0.018729543313384056, + 0.035405080765485764, + 0.019687747582793236, + 0.01581427827477455, + 0.28394848108291626, + 0.03783009201288223, + 0.051753848791122437, + 0.37227731943130493, + 0.12669320404529572 + ], + "credit_map": [ + 0.0032798724714666605, + 0.0059168641455471516, + 0.006589308846741915, + 0.00592905655503273, + 0.002839162014424801, + 0.003975679632276297, + 0.004503500182181597, + 0.004826971795409918, + 0.0012665302492678165, + 0.0023816723842173815, + 0.001979922177270055, + 0.00234796735458076, + 0.002454672707244754, + 0.0037426394410431385, + 0.0028131010476499796, + 0.0017430396983399987, + 0.0038447475526481867, + 0.00421172333881259, + 0.007482616230845451, + 0.00485490495339036, + 0.004069959744811058, + 0.003259232733398676, + 0.0032195663079619408, + 0.00446232920512557, + 0.001827736385166645, + 0.002365176100283861, + 0.0027839948888868093, + 0.0016330910148099065, + 0.002985844388604164, + 0.0038924526888877153, + 0.0021327147260308266, + 0.0020667368080466986, + 0.0017941913101822138, + 0.0029180541168898344, + 0.0018191338749602437, + 0.001938844216056168, + 0.0017407159321010113, + 0.0027521103620529175, + 0.0012447087792679667, + 0.0016065174713730812, + 0.017936984077095985, + 0.033450037240982056, + 0.06151660159230232, + 0.043008480221033096, + 0.0655781701207161, + 0.02053835429251194, + 0.019286569207906723, + 0.022633260115981102, + 0.0031969265546649694, + 0.004205194767564535, + 0.011868255212903023, + 0.004282189998775721, + 0.00613131420686841, + 0.0034259508829563856, + 0.0023565071169286966, + 0.0023637553676962852, + 0.0055442373268306255, + 0.006802774965763092, + 0.006347096525132656, + 0.007047225721180439, + 0.005591066088527441, + 0.008339866064488888, + 0.004906604066491127, + 0.007174978964030743, + 0.06327616423368454, + 0.014639121480286121, + 0.14485016465187073, + 0.022128066048026085, + 0.07624737918376923, + 0.015367502346634865, + 0.023964280262589455, + 0.011804644949734211, + 0.01909736357629299, + 0.010657881386578083, + 0.02305212803184986, + 0.008435078896582127, + 0.03139159828424454, + 0.017434503883123398, + 0.009559379890561104, + 0.00706527242437005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.01257431972771883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498284474015236 + ], + "opposite_logprob_tokens": [ + -2.929982900619507 + ], + "logprob_reward": -0.05498284474015236, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498284474015236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13876241445541382, + 0.0766032263636589, + 0.10475873947143555, + 0.06559877842664719, + 0.14878612756729126, + 0.0689498707652092, + 0.030947651714086533, + 0.022909516468644142, + 0.28372833132743835, + 0.05895540490746498 + ], + "credit_map": [ + 0.018067900091409683, + 0.01651768386363983, + 0.027178553864359856, + 0.02110583521425724, + 0.012498092837631702, + 0.012569628655910492, + 0.017063049599528313, + 0.01376164797693491, + 0.007501800544559956, + 0.009708642959594727, + 0.01552384439855814, + 0.009532221592962742, + 0.008253772743046284, + 0.009644662961363792, + 0.010800819844007492, + 0.005637452472001314, + 0.015337490476667881, + 0.012775231152772903, + 0.01225231122225523, + 0.027505220845341682, + 0.00839696079492569, + 0.010080724023282528, + 0.007721313741058111, + 0.010689474642276764, + 0.004774665925651789, + 0.00777288805693388, + 0.009587431326508522, + 0.007307655178010464, + 0.010280805639922619, + 0.010684906505048275, + 0.008066017180681229, + 0.007124399766325951, + 0.015747040510177612, + 0.03073415905237198, + 0.018008483573794365, + 0.01705246977508068, + 0.014411164447665215, + 0.032372601330280304, + 0.01207044254988432, + 0.008389749564230442, + 0.006005113013088703, + 0.011749929748475552, + 0.011052567511796951, + 0.01337100937962532, + 0.008206339552998543, + 0.0049997298046946526, + 0.006684217602014542, + 0.006880954839289188, + 0.0026403148658573627, + 0.002746989717707038, + 0.004003642126917839, + 0.0035679268185049295, + 0.005689736921340227, + 0.0049750120379030704, + 0.005013785790652037, + 0.002310238778591156, + 0.0018432735232636333, + 0.002577339531853795, + 0.005541421007364988, + 0.002852553967386484, + 0.0035128442104905844, + 0.0018497036071494222, + 0.0022118305787444115, + 0.002520548878237605, + 0.04847954213619232, + 0.012437594123184681, + 0.10963404178619385, + 0.015112286433577538, + 0.0571238249540329, + 0.01423325389623642, + 0.017021477222442627, + 0.009686288423836231, + 0.005467343144118786, + 0.004221326671540737, + 0.020130308344960213, + 0.004300386179238558, + 0.013164106756448746, + 0.004194875713437796, + 0.003745834343135357, + 0.003731220494955778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.05498284474015236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013392536202445626 + ], + "opposite_logprob_tokens": [ + -6.626339435577393 + ], + "logprob_reward": -0.0013392536202445626, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013392536202445626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14822761714458466, + 0.058559611439704895, + 0.11001340299844742, + 0.10541725903749466, + 0.03195473179221153, + 0.15330538153648376, + 0.09571853280067444, + 0.019585374742746353, + 0.24734359979629517, + 0.029874511063098907 + ], + "credit_map": [ + 0.016794590279459953, + 0.0244479700922966, + 0.028922228142619133, + 0.020130855962634087, + 0.011177003383636475, + 0.01749032735824585, + 0.01679055579006672, + 0.01247408241033554, + 0.006032393779605627, + 0.009027348831295967, + 0.007612441200762987, + 0.007129347417503595, + 0.006514151114970446, + 0.010621064342558384, + 0.0077537717297673225, + 0.0038690862711519003, + 0.010646362788975239, + 0.01113571971654892, + 0.025144701823592186, + 0.021568629890680313, + 0.012962575070559978, + 0.009741896763443947, + 0.010264897719025612, + 0.008548622019588947, + 0.011701923795044422, + 0.016224881634116173, + 0.020224053412675858, + 0.017077675089240074, + 0.010965357534587383, + 0.014794287271797657, + 0.008243645541369915, + 0.006185426842421293, + 0.0029586683958768845, + 0.006802401039749384, + 0.0040998258627951145, + 0.004015497863292694, + 0.002721752505749464, + 0.00631612166762352, + 0.002939563477411866, + 0.0021008963230997324, + 0.010089165531098843, + 0.017466792836785316, + 0.03040585294365883, + 0.024873698130249977, + 0.031746041029691696, + 0.01063673198223114, + 0.014709233306348324, + 0.013377874158322811, + 0.007629115600138903, + 0.01011150423437357, + 0.030005915090441704, + 0.009211026132106781, + 0.01667647436261177, + 0.010378737933933735, + 0.005755683872848749, + 0.005950074642896652, + 0.0017129479674622416, + 0.0021660563070327044, + 0.004243308678269386, + 0.0027428509201854467, + 0.003063718555495143, + 0.0020435380283743143, + 0.002069974783807993, + 0.0015429777558892965, + 0.04305506497621536, + 0.009661104530096054, + 0.09758022427558899, + 0.012476883828639984, + 0.05126616731286049, + 0.010584085248410702, + 0.014904678799211979, + 0.00781538151204586, + 0.004420246928930283, + 0.002513309707865119, + 0.009579584933817387, + 0.002033964265137911, + 0.005111889448016882, + 0.0022820632439106703, + 0.0023403209634125233, + 0.0015931278467178345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.0013392536202445626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001248002052307 + ], + "opposite_logprob_tokens": [ + -2.6950125694274902 + ], + "logprob_reward": -0.07001248002052307, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001248002052307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13606715202331543, + 0.10624033212661743, + 0.1306867152452469, + 0.07903248816728592, + 0.07443660497665405, + 0.15206195414066315, + 0.07476191967725754, + 0.03157944232225418, + 0.15785397589206696, + 0.057279426604509354 + ], + "credit_map": [ + 0.013490738347172737, + 0.019357921555638313, + 0.023419324308633804, + 0.02275189757347107, + 0.01392530556768179, + 0.010760141536593437, + 0.019809601828455925, + 0.012552214786410332, + 0.009473255835473537, + 0.01122959889471531, + 0.0143814105540514, + 0.007184754591435194, + 0.010069445706903934, + 0.030852600932121277, + 0.015734869986772537, + 0.0073143960908055305, + 0.015834840014576912, + 0.016101697459816933, + 0.030027516186237335, + 0.016542231664061546, + 0.012165505439043045, + 0.012466516345739365, + 0.010921136476099491, + 0.01662726327776909, + 0.00816337950527668, + 0.009947516955435276, + 0.018977446481585503, + 0.00933602824807167, + 0.0075767613016068935, + 0.011075860820710659, + 0.007054389454424381, + 0.006901101674884558, + 0.008340718224644661, + 0.014739743433892727, + 0.007826108485460281, + 0.006614480633288622, + 0.008814269676804543, + 0.013786924071609974, + 0.007806642446666956, + 0.00650771614164114, + 0.008297943510115147, + 0.009347295388579369, + 0.040515825152397156, + 0.008941440843045712, + 0.05060898885130882, + 0.01353520154953003, + 0.014427270740270615, + 0.006387983914464712, + 0.006056567188352346, + 0.010096900165081024, + 0.010266751982271671, + 0.011372880078852177, + 0.010949072428047657, + 0.012002029456198215, + 0.009214510209858418, + 0.004803201649338007, + 0.0029765961226075888, + 0.004270099103450775, + 0.004452915396541357, + 0.00562710128724575, + 0.003730553900822997, + 0.0030101826414465904, + 0.0035544568672776222, + 0.0039575365372002125, + 0.027389932423830032, + 0.006523173302412033, + 0.059794727712869644, + 0.008652592077851295, + 0.03099275566637516, + 0.00959090981632471, + 0.00935396645218134, + 0.005555922165513039, + 0.004881545435637236, + 0.00479854503646493, + 0.009463002905249596, + 0.0061658225022256374, + 0.013774220831692219, + 0.009262165986001492, + 0.004727433901280165, + 0.00420669000595808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.07001248002052307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "documentary-style medium close-up shot, shot from a level angle and with a slight handheld shakiness, focuses on a young boy, approximately six or seven years old, standing alone amidst a ruined landscape. His cheeks are stained with a mixture of dust and dried earth, his short, messy black hair matted with grayish-white dust. He wears a worn, oversized T-shirt, faded to a gray. He hunches slightly, his lips pressed together as if trying to stifle sobs. But large, glistening tears silently fall from his reddened eyes, leaving two distinct wet streaks along his stained cheeks. His vacant eyes stare at a point beyond the camera, filled with a sadness and confusion beyond his years. Behind him, out-of-focus rubble clung to the ground, twisted rebar protruding from shattered concrete blocks, and rubble and debris littered the ground. A cool, soft beam of natural light, streaming through a sky obscured by smoke and dust, obliquely illuminates half of his face, casting deep, soft shadows that highlight the moist sheen of tear marks and the rough texture of the earth. The overall color palette of the painting tends toward cool grays and earthy yellows, creating an atmosphere of oppression, heaviness, and profound realism.", + "seed": 7460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_14/index_8759_seed_7460_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02408235869370401, + "vlm_logprob": -0.02408235869370401, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.02408235869370401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017070623114705086 + ], + "opposite_logprob_tokens": [ + -5.142070770263672 + ], + "logprob_reward": -0.017070623114705086, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.017070623114705086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20982825756072998, + 0.1538536548614502, + 0.08879639953374863, + 0.07600487023591995, + 0.048382196575403214, + 0.03835323825478554, + 0.04866577684879303, + 0.23640207946300507, + 0.033458542078733444, + 0.06625498086214066 + ], + "credit_map": [ + 0.02711203694343567, + 0.03091355413198471, + 0.04959464818239212, + 0.02906123176217079, + 0.019452715292572975, + 0.016025586053729057, + 0.01908624917268753, + 0.018582239747047424, + 0.024041064083576202, + 0.009155632928013802, + 0.05795951560139656, + 0.02393306978046894, + 0.011209838092327118, + 0.006832837127149105, + 0.011723200790584087, + 0.00899849459528923, + 0.013288173824548721, + 0.0038379079196602106, + 0.03315967693924904, + 0.010688720270991325, + 0.005083835683763027, + 0.0070085711777210236, + 0.00961111206561327, + 0.006118412595242262, + 0.006068205460906029, + 0.006533822510391474, + 0.009816513396799564, + 0.010837574489414692, + 0.013101237826049328, + 0.009165991097688675, + 0.011597644537687302, + 0.008883873000741005, + 0.007116733584553003, + 0.003444388508796692, + 0.010216576978564262, + 0.005240670405328274, + 0.004545958247035742, + 0.005746936891227961, + 0.005675429943948984, + 0.006395505275577307, + 0.0054639531299471855, + 0.0025511716958135366, + 0.009152461774647236, + 0.0044848741963505745, + 0.0043779644183814526, + 0.004652191419154406, + 0.003995161969214678, + 0.0036754636093974113, + 0.006718190386891365, + 0.0028097834438085556, + 0.008208454586565495, + 0.007529504597187042, + 0.005198292434215546, + 0.006925752852112055, + 0.005639183800667524, + 0.005636617075651884, + 0.04067068174481392, + 0.006028094328939915, + 0.10031360387802124, + 0.022822069004178047, + 0.013856067322194576, + 0.009727753698825836, + 0.025287918746471405, + 0.017695896327495575, + 0.004578469321131706, + 0.0017729769460856915, + 0.004548891447484493, + 0.005272167734801769, + 0.003238821169361472, + 0.005246007349342108, + 0.00362029648385942, + 0.005180912092328072, + 0.008253323845565319, + 0.004915888421237469, + 0.010124183259904385, + 0.010047516785562038, + 0.0071412683464586735, + 0.004864938091486692, + 0.011044003069400787, + 0.009863851591944695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.017070623114705086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008672904223203659 + ], + "opposite_logprob_tokens": [ + -4.758672714233398 + ], + "logprob_reward": -0.008672904223203659, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008672904223203659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23805910348892212, + 0.18007837235927582, + 0.22810935974121094, + 0.209459587931633, + 0.027154728770256042, + 0.031913839280605316, + 0.018636412918567657, + 0.02138000912964344, + 0.010905248112976551, + 0.034303296357393265 + ], + "credit_map": [ + 0.035329241305589676, + 0.015820665284991264, + 0.08427639305591583, + 0.04069989547133446, + 0.012394263409078121, + 0.014332400634884834, + 0.015273278579115868, + 0.01993297040462494, + 0.02191021665930748, + 0.021819958463311195, + 0.050389450043439865, + 0.03148747235536575, + 0.011461659334599972, + 0.014646426774561405, + 0.014437583275139332, + 0.013925634324550629, + 0.03272704780101776, + 0.009355255402624607, + 0.08093024045228958, + 0.03178992494940758, + 0.015240770764648914, + 0.018867727369070053, + 0.02218686416745186, + 0.017011543735861778, + 0.031693462282419205, + 0.011699232272803783, + 0.06369152665138245, + 0.029778238385915756, + 0.008756760507822037, + 0.019712477922439575, + 0.025200625881552696, + 0.018927281722426414, + 0.00413662800565362, + 0.0015329618472605944, + 0.007298522163182497, + 0.003284781938418746, + 0.001897240523248911, + 0.0029141807463020086, + 0.003783455351367593, + 0.0023069579619914293, + 0.00531897833570838, + 0.00161725003272295, + 0.00680727930739522, + 0.00315649202093482, + 0.0023708061780780554, + 0.004908716771751642, + 0.0036433509085327387, + 0.004090967122465372, + 0.003255450865253806, + 0.0010366940405219793, + 0.0021339580416679382, + 0.0029227507766336203, + 0.0019003088818863034, + 0.0025943280197679996, + 0.002355970209464431, + 0.002436953131109476, + 0.0023389230482280254, + 0.0010817016009241343, + 0.00497226370498538, + 0.003222511615604162, + 0.0019409164087846875, + 0.0022150613367557526, + 0.0027379130478948355, + 0.0028707205783575773, + 0.0017781425267457962, + 0.0005987820331938565, + 0.0013905125670135021, + 0.001173881464637816, + 0.001114077866077423, + 0.0013590033631771803, + 0.0014535423833876848, + 0.002037307247519493, + 0.0056997500360012054, + 0.0027919893618673086, + 0.005409534089267254, + 0.00426532793790102, + 0.003619302762672305, + 0.00412034522742033, + 0.004005860537290573, + 0.004391191061586142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008672904223203659 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331031858921051 + ], + "opposite_logprob_tokens": [ + -3.773310422897339 + ], + "logprob_reward": -0.02331031858921051, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331031858921051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3173271119594574, + 0.060555290430784225, + 0.05641746520996094, + 0.18587759137153625, + 0.06780770421028137, + 0.03514639660716057, + 0.057129472494125366, + 0.1133975088596344, + 0.03131389617919922, + 0.07502757757902145 + ], + "credit_map": [ + 0.03377853333950043, + 0.029844943434000015, + 0.08985576778650284, + 0.05280958116054535, + 0.026457473635673523, + 0.02977686934173107, + 0.026776276528835297, + 0.02802763506770134, + 0.006299235858023167, + 0.008053191006183624, + 0.010020886547863483, + 0.007579957600682974, + 0.007078448310494423, + 0.007818582467734814, + 0.007376931142061949, + 0.006328051909804344, + 0.004622620530426502, + 0.00413146149367094, + 0.006545865908265114, + 0.006928622722625732, + 0.009555475786328316, + 0.009945867583155632, + 0.00921948254108429, + 0.005468064919114113, + 0.025048620998859406, + 0.010855364613234997, + 0.04421459883451462, + 0.02339956909418106, + 0.014107254333794117, + 0.017834613099694252, + 0.030879061669111252, + 0.019538499414920807, + 0.010228394530713558, + 0.0060692159458994865, + 0.009994814172387123, + 0.008095763623714447, + 0.007578135933727026, + 0.009353486821055412, + 0.009591775946319103, + 0.006896111182868481, + 0.004843303933739662, + 0.0022126089315861464, + 0.004360576625913382, + 0.005459363106638193, + 0.003105630399659276, + 0.006069928873330355, + 0.00421402882784605, + 0.0048809535801410675, + 0.010638154111802578, + 0.0039022888522595167, + 0.015197565779089928, + 0.005268777254968882, + 0.0043287118896842, + 0.007056118454784155, + 0.005639778915792704, + 0.005098070949316025, + 0.01861201971769333, + 0.004256703890860081, + 0.04194542393088341, + 0.014984381385147572, + 0.007677383255213499, + 0.0057432642206549644, + 0.01156407780945301, + 0.008614245802164078, + 0.0044675241224467754, + 0.001675248029641807, + 0.005711535457521677, + 0.00458886194974184, + 0.0031733019277453423, + 0.0032634970266371965, + 0.003535363357514143, + 0.004898563027381897, + 0.010109714232385159, + 0.004306084476411343, + 0.01983623392879963, + 0.00784002523869276, + 0.007059023249894381, + 0.00752620305866003, + 0.009885209612548351, + 0.00846507865935564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.02331031858921051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013391345273703337 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013391345273703337, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013391345273703337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.254928320646286, + 0.08100766688585281, + 0.025990638881921768, + 0.3591465353965759, + 0.03490497171878815, + 0.027103804051876068, + 0.031412363052368164, + 0.04635415971279144, + 0.05308130756020546, + 0.086070217192173 + ], + "credit_map": [ + 0.040223781019449234, + 0.022013459354639053, + 0.08725745975971222, + 0.04403715208172798, + 0.010659436695277691, + 0.013800566084682941, + 0.016015125438570976, + 0.020921343937516212, + 0.011988659389317036, + 0.013172591105103493, + 0.02044045366346836, + 0.011905798688530922, + 0.005934457294642925, + 0.005948442034423351, + 0.005791158881038427, + 0.005826102569699287, + 0.003689895384013653, + 0.0010645936708897352, + 0.008519969880580902, + 0.003692443948239088, + 0.0020255697891116142, + 0.0019585455302149057, + 0.003067773301154375, + 0.0019718443509191275, + 0.06441345065832138, + 0.01714443787932396, + 0.13836756348609924, + 0.028834214434027672, + 0.020359667018055916, + 0.020111769437789917, + 0.03958361968398094, + 0.030331801623106003, + 0.006080060265958309, + 0.0015587331727147102, + 0.01128164678812027, + 0.003266983898356557, + 0.002503708703443408, + 0.003182711312547326, + 0.0035245362669229507, + 0.0035065901465713978, + 0.004832169972360134, + 0.0013231293996796012, + 0.0068332478404045105, + 0.0027523012831807137, + 0.0017666533822193742, + 0.003419733140617609, + 0.002858924912288785, + 0.003317641094326973, + 0.00550715159624815, + 0.0019458197057247162, + 0.008050800301134586, + 0.0025204576086252928, + 0.0018852929351851344, + 0.005437500774860382, + 0.002784638898447156, + 0.0032806990202516317, + 0.00819840282201767, + 0.0013207198353484273, + 0.01960587501525879, + 0.00429666880518198, + 0.0029483726248145103, + 0.0018822200363501906, + 0.004772714804857969, + 0.0033291801810264587, + 0.010044563561677933, + 0.002794666914269328, + 0.008480857126414776, + 0.006232754327356815, + 0.004748203791677952, + 0.0063835615292191505, + 0.005396157503128052, + 0.009000538848340511, + 0.01556497160345316, + 0.00560956122353673, + 0.017073458060622215, + 0.010879768058657646, + 0.00589096499606967, + 0.011820269748568535, + 0.0075721596367657185, + 0.011659055016934872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.0013391345273703337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001881301403046 + ], + "opposite_logprob_tokens": [ + -2.695018768310547 + ], + "logprob_reward": -0.07001881301403046, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001881301403046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2557734549045563, + 0.08659475296735764, + 0.10490574687719345, + 0.23852801322937012, + 0.07372493296861649, + 0.03520714119076729, + 0.024408072233200073, + 0.07395640760660172, + 0.0494055338203907, + 0.05749591067433357 + ], + "credit_map": [ + 0.03337906301021576, + 0.02948853000998497, + 0.07528747618198395, + 0.042468249797821045, + 0.016314197331666946, + 0.018375061452388763, + 0.021158836781978607, + 0.019302058964967728, + 0.011695809662342072, + 0.009391145780682564, + 0.029769212007522583, + 0.014040744863450527, + 0.006057235412299633, + 0.004082704894244671, + 0.007158143445849419, + 0.004399758297950029, + 0.015467078424990177, + 0.004527775105088949, + 0.03808167949318886, + 0.013346509076654911, + 0.00613191444426775, + 0.008007883094251156, + 0.011931898072361946, + 0.007411010563373566, + 0.040988728404045105, + 0.012689530849456787, + 0.08574765920639038, + 0.01916375569999218, + 0.01706969179213047, + 0.016305092722177505, + 0.02718285843729973, + 0.019380686804652214, + 0.011361822485923767, + 0.004145814571529627, + 0.019845962524414062, + 0.008576306514441967, + 0.008739963173866272, + 0.00604640319943428, + 0.008724820800125599, + 0.006283846218138933, + 0.004742118995636702, + 0.001872723805718124, + 0.010708669200539589, + 0.004036701750010252, + 0.0035164556466042995, + 0.0036590974777936935, + 0.0037920689210295677, + 0.002879305509850383, + 0.0030665143858641386, + 0.0019321965519338846, + 0.0043590920977294445, + 0.0026700859889388084, + 0.0032837865874171257, + 0.003566396189853549, + 0.0025959142949432135, + 0.002934084739536047, + 0.010611538775265217, + 0.00254194182343781, + 0.027155769988894463, + 0.008474678732454777, + 0.005173643119633198, + 0.006064437795430422, + 0.008156374096870422, + 0.005778023973107338, + 0.009795531630516052, + 0.0027828309684991837, + 0.00828088354319334, + 0.005487360060214996, + 0.004486647434532642, + 0.006402067840099335, + 0.005105409771203995, + 0.007064804434776306, + 0.011019417084753513, + 0.0023402052465826273, + 0.010347793810069561, + 0.005262589547783136, + 0.004214071668684483, + 0.005738489795476198, + 0.0056628515012562275, + 0.012910488992929459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07001881301403046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "documentary-style medium close-up shot, shot from a level angle and with a slight handheld shakiness, focuses on a young boy, approximately six or seven years old, standing alone amidst a ruined landscape. His cheeks are stained with a mixture of dust and dried earth, his short, messy black hair matted with grayish-white dust. He wears a worn, oversized T-shirt, faded to a gray. He hunches slightly, his lips pressed together as if trying to stifle sobs. But large, glistening tears silently fall from his reddened eyes, leaving two distinct wet streaks along his stained cheeks. His vacant eyes stare at a point beyond the camera, filled with a sadness and confusion beyond his years. Behind him, out-of-focus rubble clung to the ground, twisted rebar protruding from shattered concrete blocks, and rubble and debris littered the ground. A cool, soft beam of natural light, streaming through a sky obscured by smoke and dust, obliquely illuminates half of his face, casting deep, soft shadows that highlight the moist sheen of tear marks and the rough texture of the earth. The overall color palette of the painting tends toward cool grays and earthy yellows, creating an atmosphere of oppression, heaviness, and profound realism.", + "seed": 7461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_14/index_8759_seed_7461_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03388692550361157, + "vlm_logprob": -0.03388692550361157, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03388692550361157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021839478984475136 + ], + "opposite_logprob_tokens": [ + -4.896839618682861 + ], + "logprob_reward": -0.021839478984475136, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.021839478984475136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3610247075557709, + 0.14808151125907898, + 0.12485341727733612, + 0.10573463141918182, + 0.05095529556274414, + 0.0599600225687027, + 0.02689533680677414, + 0.03685418888926506, + 0.03496694564819336, + 0.0506739504635334 + ], + "credit_map": [ + 0.0405132882297039, + 0.0826483815908432, + 0.08172723650932312, + 0.04153144732117653, + 0.023106928914785385, + 0.0337710827589035, + 0.02992594800889492, + 0.027800388634204865, + 0.017213044688105583, + 0.028780866414308548, + 0.028214946389198303, + 0.014241340570151806, + 0.007089812308549881, + 0.02324100397527218, + 0.014031456783413887, + 0.015269043855369091, + 0.012769609689712524, + 0.015395059250295162, + 0.042863234877586365, + 0.013777553103864193, + 0.005030348896980286, + 0.011007707566022873, + 0.013330631889402866, + 0.010679274797439575, + 0.014193377457559109, + 0.01684856414794922, + 0.015369672328233719, + 0.013249286450445652, + 0.007794499397277832, + 0.013022170402109623, + 0.00948867667466402, + 0.0157683864235878, + 0.00401349039748311, + 0.0048363483510911465, + 0.011874149553477764, + 0.005899782758206129, + 0.0046519069001078606, + 0.005549285095185041, + 0.006836389657109976, + 0.007293942384421825, + 0.005033218301832676, + 0.006893986836075783, + 0.010935904458165169, + 0.006481940858066082, + 0.004049517214298248, + 0.011570557951927185, + 0.006350915413349867, + 0.0086439810693264, + 0.0025566869881004095, + 0.0020416530314832926, + 0.004378750454634428, + 0.003740622429177165, + 0.002427133498713374, + 0.0038691917434334755, + 0.002853617537766695, + 0.0050276825204491615, + 0.003508106805384159, + 0.002248586155474186, + 0.00861261785030365, + 0.004976856522262096, + 0.0029312411788851023, + 0.004831254016608, + 0.004150147549808025, + 0.005595377646386623, + 0.0034623632673174143, + 0.002964643994346261, + 0.005456898827105761, + 0.004422217607498169, + 0.0038992036134004593, + 0.0036601887550204992, + 0.00391046516597271, + 0.007190967444330454, + 0.004417932126671076, + 0.0043374537490308285, + 0.011908897198736668, + 0.007224737200886011, + 0.00604834733530879, + 0.00552863534539938, + 0.00601234333589673, + 0.005195602308958769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.021839478984475136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822850853204727 + ], + "opposite_logprob_tokens": [ + -4.018228530883789 + ], + "logprob_reward": -0.01822850853204727, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822850853204727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36325931549072266, + 0.08205030113458633, + 0.04360109195113182, + 0.08041543513536453, + 0.085869200527668, + 0.04122096672654152, + 0.032946933060884476, + 0.09755592793226242, + 0.07626931369304657, + 0.09681156277656555 + ], + "credit_map": [ + 0.05128685384988785, + 0.03820981830358505, + 0.08689405769109726, + 0.0444287434220314, + 0.024013953283429146, + 0.027958692982792854, + 0.05372156947851181, + 0.036745619028806686, + 0.008455637842416763, + 0.010934654623270035, + 0.018101148307323456, + 0.01293065957725048, + 0.0055886683985590935, + 0.008184992708265781, + 0.008897858671844006, + 0.008956681936979294, + 0.00411154143512249, + 0.005263995844870806, + 0.009341930039227009, + 0.0062089888378977776, + 0.0037600367795675993, + 0.004197570960968733, + 0.005990037694573402, + 0.004726992454379797, + 0.006765604484826326, + 0.007222773972898722, + 0.012398330494761467, + 0.012830285355448723, + 0.011822927743196487, + 0.0075281644240021706, + 0.012052560225129128, + 0.009794785641133785, + 0.00828535109758377, + 0.009967499412596226, + 0.013527843169867992, + 0.014663481153547764, + 0.008000652305781841, + 0.010244693607091904, + 0.010918275453150272, + 0.010261408984661102, + 0.004035694990307093, + 0.0053690276108682156, + 0.0059463102370500565, + 0.004691047593951225, + 0.002913298085331917, + 0.0080646313726902, + 0.0039548794738948345, + 0.006246078293770552, + 0.002355049131438136, + 0.0028402984607964754, + 0.006933317054063082, + 0.003881370183080435, + 0.0035570254549384117, + 0.0037576532922685146, + 0.0034748443868011236, + 0.006147373467683792, + 0.008929736912250519, + 0.00460770633071661, + 0.020785892382264137, + 0.015578631311655045, + 0.009600095450878143, + 0.012402581050992012, + 0.011737676337361336, + 0.013913613744080067, + 0.006370539776980877, + 0.0051608821377158165, + 0.012181096710264683, + 0.012461875565350056, + 0.009550320915877819, + 0.011666255071759224, + 0.009903629310429096, + 0.008974705822765827, + 0.00652465270832181, + 0.006748073734343052, + 0.01878313533961773, + 0.019724935293197632, + 0.011831413954496384, + 0.011187979020178318, + 0.01199966762214899, + 0.010011706501245499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.01822850853204727 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048677489161491394 + ], + "opposite_logprob_tokens": [ + -3.048677444458008 + ], + "logprob_reward": -0.048677489161491394, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048677489161491394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5595749616622925, + 0.09095551818609238, + 0.0849662721157074, + 0.05116411671042442, + 0.05473002791404724, + 0.03578437864780426, + 0.03783921152353287, + 0.03972894698381424, + 0.029709866270422935, + 0.015546767972409725 + ], + "credit_map": [ + 0.07739648967981339, + 0.11467090249061584, + 0.13772013783454895, + 0.06122329458594322, + 0.03542757406830788, + 0.04015088081359863, + 0.05805201083421707, + 0.03493363410234451, + 0.013392733410000801, + 0.010661107487976551, + 0.01652388833463192, + 0.009234422817826271, + 0.010100294835865498, + 0.009144500829279423, + 0.010072999633848667, + 0.011825578287243843, + 0.008270841091871262, + 0.008791383355855942, + 0.0307210274040699, + 0.009686518460512161, + 0.0044380673207342625, + 0.006818265188485384, + 0.01016307808458805, + 0.006077087018638849, + 0.00656855059787631, + 0.0075760274194180965, + 0.007367785554379225, + 0.006492293439805508, + 0.005197999533265829, + 0.00658643851056695, + 0.005868411157280207, + 0.005506611429154873, + 0.005999299697577953, + 0.007381202187389135, + 0.008051218464970589, + 0.0072106774896383286, + 0.0042478241957724094, + 0.009357784874737263, + 0.005369648803025484, + 0.007112370803952217, + 0.0037211794406175613, + 0.004183955956250429, + 0.006816853769123554, + 0.004279801622033119, + 0.002658841433003545, + 0.00511845713481307, + 0.004167269449681044, + 0.0048380219377577305, + 0.004499707370996475, + 0.004312282893806696, + 0.004956687334924936, + 0.0038643309380859137, + 0.004127101972699165, + 0.006201679352670908, + 0.003668233985081315, + 0.006209188606590033, + 0.0032479900401085615, + 0.004386696498841047, + 0.006151888519525528, + 0.0038307334762066603, + 0.004090463276952505, + 0.004707858432084322, + 0.004355950281023979, + 0.00895736925303936, + 0.0021781516261398792, + 0.0023206877522170544, + 0.00542274909093976, + 0.004962355829775333, + 0.004482460208237171, + 0.004165996331721544, + 0.0032748105004429817, + 0.0029026558622717857, + 0.0014846783597022295, + 0.0011469986056908965, + 0.0026061981916427612, + 0.0022123854141682386, + 0.002409046282991767, + 0.0018863793229684234, + 0.0021375257056206465, + 0.0016635562060400844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.048677489161491394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017150230705738068 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.0017150230705738068, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017150230705738068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34419187903404236, + 0.17683175206184387, + 0.05340857803821564, + 0.08099447190761566, + 0.14063644409179688, + 0.0668732300400734, + 0.03878721594810486, + 0.039424095302820206, + 0.02070271037518978, + 0.03814958781003952 + ], + "credit_map": [ + 0.049330487847328186, + 0.07958009839057922, + 0.08658292144536972, + 0.03325764462351799, + 0.02030584029853344, + 0.02543380670249462, + 0.032453153282403946, + 0.017247943207621574, + 0.028224579989910126, + 0.024131864309310913, + 0.043535266071558, + 0.018871279433369637, + 0.01122464518994093, + 0.016591016203165054, + 0.015197942964732647, + 0.019055167213082314, + 0.005790947005152702, + 0.006937936879694462, + 0.016783487051725388, + 0.006317409686744213, + 0.0031562030781060457, + 0.004503318108618259, + 0.00586001155897975, + 0.004059273283928633, + 0.012450695037841797, + 0.010864333249628544, + 0.01133033074438572, + 0.010536798276007175, + 0.004775482229888439, + 0.013658784329891205, + 0.006160940509289503, + 0.011217107065021992, + 0.019229784607887268, + 0.025210276246070862, + 0.0183009821921587, + 0.017550675198435783, + 0.009901165962219238, + 0.02433258853852749, + 0.010062211193144321, + 0.016048766672611237, + 0.007169563323259354, + 0.008617199957370758, + 0.010736580938100815, + 0.007838883437216282, + 0.004304117988795042, + 0.0123680979013443, + 0.005641923751682043, + 0.010196869261562824, + 0.004095424897968769, + 0.0048061213456094265, + 0.007609587162733078, + 0.00483492249622941, + 0.0026613238733261824, + 0.005076443776488304, + 0.0037307264283299446, + 0.005972668994218111, + 0.003972339443862438, + 0.005022046156227589, + 0.006985805928707123, + 0.005122964736074209, + 0.004224006552249193, + 0.005684352479875088, + 0.004832138307392597, + 0.0035804470535367727, + 0.0019072780851274729, + 0.0012689469149336219, + 0.004371942020952702, + 0.0023609024938195944, + 0.0020180046558380127, + 0.003541720798239112, + 0.0022908204700797796, + 0.0029430969152599573, + 0.004239318892359734, + 0.002534543862566352, + 0.005696020554751158, + 0.003559969598427415, + 0.0036107399500906467, + 0.008167047053575516, + 0.003805573796853423, + 0.006536377593874931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0017150230705738068 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young boy (approximately six or seven years old)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a worn, oversized T-shirt that is faded to gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there large, glistening tears falling from the boy's reddened eyes, leaving wet streaks on his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an out-of-focus ruined landscape with rubble, twisted rebar, and shattered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with slight shakiness from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897412776947021 + ], + "opposite_logprob_tokens": [ + -2.5789742469787598 + ], + "logprob_reward": -0.07897412776947021, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897412776947021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38791683316230774, + 0.16863788664340973, + 0.09452997148036957, + 0.10892758518457413, + 0.05286883935332298, + 0.025608191266655922, + 0.04068916290998459, + 0.036070674657821655, + 0.07142568379640579, + 0.013325109146535397 + ], + "credit_map": [ + 0.057635147124528885, + 0.03798912093043327, + 0.09295044094324112, + 0.049864549189805984, + 0.028002088889479637, + 0.04188765585422516, + 0.04147913306951523, + 0.03810872882604599, + 0.02181589975953102, + 0.03748707473278046, + 0.02983182854950428, + 0.015495703555643559, + 0.007626980543136597, + 0.02680668607354164, + 0.011202641762793064, + 0.01837107352912426, + 0.011915646493434906, + 0.009611794725060463, + 0.022122051566839218, + 0.011525833979249, + 0.005613669753074646, + 0.012399240396916866, + 0.008687421679496765, + 0.012654315680265427, + 0.014457829296588898, + 0.016474248841404915, + 0.015722202137112617, + 0.013502822257578373, + 0.007669555488973856, + 0.01460173074156046, + 0.009212937206029892, + 0.017286254093050957, + 0.004620351828634739, + 0.004734514746814966, + 0.012042583897709846, + 0.0077418372966349125, + 0.004325650166720152, + 0.005339621566236019, + 0.006761019583791494, + 0.007303258404135704, + 0.00323475175537169, + 0.0022289578337222338, + 0.004005299415439367, + 0.0037464178167283535, + 0.0028470936231315136, + 0.003509455593302846, + 0.0027728050481528044, + 0.003263411344960332, + 0.003063164185732603, + 0.005568745080381632, + 0.006774194538593292, + 0.004655295982956886, + 0.0035766535438597202, + 0.008017892017960548, + 0.003267935710027814, + 0.005765281151980162, + 0.002773342654109001, + 0.0028222238179296255, + 0.006358952261507511, + 0.005396994296461344, + 0.004446034319698811, + 0.004531750921159983, + 0.0037501861806958914, + 0.005991190671920776, + 0.007448785472661257, + 0.004136231262236834, + 0.010304846800863743, + 0.010641572065651417, + 0.0073827337473630905, + 0.011919520795345306, + 0.008729408495128155, + 0.010862583294510841, + 0.0012055806582793593, + 0.0010349344229325652, + 0.0023433761671185493, + 0.0015955414855852723, + 0.001920349313877523, + 0.0022265303414314985, + 0.0018817855743691325, + 0.001117010717280209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07897412776947021 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_15/index_tensor([9891, 9891])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_15/index_tensor([9891, 9891])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a5effc35286448760cb6ce1a19893b726ac39b2c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_15/index_tensor([9891, 9891])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly high-angle extreme close-up slowly pans to the right, meticulously tracking the deliberate movement of the artist's brush across the canvas. The brush itself is a classic hazelnut with a slender wooden handle stained with ghosts of previous colors, its fine sable bristles thick with a mixture of thick, lustrous cadmium yellow and sky blue paint. With a steady, reverent glide, the bristles bend slightly, leaving vibrant, lifeless brushstrokes across the rough, latticed texture of the canvas, already bearing the weight of a painterly, impressionistic composition. The entire scene is bathed in dramatic, ethereal light, emanating from a vast, out-of-focus Gothic stained glass window in the historic church setting. A kaleidoscope of pale and deep sapphire, ruby red, and brilliant gold penetrates the cool, dim air, casting a divine mosaic of colorful patterns across the stone floor and the artist's wooden easel. A ray of golden light catches the polished silver ring of the moving brush. The atmosphere is one of profound tranquility and focused creativity, as the act of painting becomes a direct translation of the sacred, vibrant light that fills the sacred space.", + "seed": 17561, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_15/index_9891_seed_17561_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06662552133202553, + "vlm_logprob": -0.06662552133202553, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.06662552133202553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07300744950771332 + ], + "opposite_logprob_tokens": [ + -2.948007345199585 + ], + "logprob_reward": -0.07300744950771332, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07300744950771332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1778896450996399, + 0.0599086657166481, + 0.051351431757211685, + 0.12420093268156052, + 0.034386102110147476, + 0.09688964486122131, + 0.10824832320213318, + 0.11110621690750122, + 0.13174466788768768, + 0.10427437722682953 + ], + "credit_map": [ + 0.040137141942977905, + 0.020115742459893227, + 0.022161489352583885, + 0.025243811309337616, + 0.013483840972185135, + 0.019936835393309593, + 0.01899065263569355, + 0.01782013103365898, + 0.011098328977823257, + 0.007056142669171095, + 0.00841794814914465, + 0.00659654987975955, + 0.006823874078691006, + 0.010090586729347706, + 0.004700544290244579, + 0.005124690942466259, + 0.012933013960719109, + 0.006854318082332611, + 0.006833723280578852, + 0.006248026620596647, + 0.003726231399923563, + 0.0048383385874331, + 0.004978431388735771, + 0.004939347039908171, + 0.017525942996144295, + 0.016357626765966415, + 0.017400335520505905, + 0.013032662682235241, + 0.01742416061460972, + 0.014384313486516476, + 0.014310098253190517, + 0.013765789568424225, + 0.0051569221541285515, + 0.005526177119463682, + 0.0038540565874427557, + 0.0036683878861367702, + 0.0036462945863604546, + 0.0031227602157741785, + 0.0037674482446163893, + 0.005644059274345636, + 0.00997967179864645, + 0.011840303428471088, + 0.010062196291983128, + 0.010273212566971779, + 0.008513531647622585, + 0.018873995169997215, + 0.014564215205609798, + 0.012782517820596695, + 0.014611145481467247, + 0.01890779659152031, + 0.011820404790341854, + 0.010487821884453297, + 0.014016511850059032, + 0.01269733626395464, + 0.012423194944858551, + 0.013284112326800823, + 0.01713278517127037, + 0.009477083571255207, + 0.010629488155245781, + 0.011286023072898388, + 0.015858056023716927, + 0.012919330969452858, + 0.020537778735160828, + 0.01326567493379116, + 0.011776365339756012, + 0.009226324036717415, + 0.020038573071360588, + 0.01253554318100214, + 0.02410740964114666, + 0.015371602959930897, + 0.024554889649152756, + 0.014133957214653492, + 0.014105445705354214, + 0.007004037965089083, + 0.018265090882778168, + 0.010101931169629097, + 0.009669376537203789, + 0.013028574176132679, + 0.014459862373769283, + 0.017640063539147377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.07300744950771332 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054948654025793076 + ], + "opposite_logprob_tokens": [ + -2.929948568344116 + ], + "logprob_reward": -0.054948654025793076, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054948654025793076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2908175587654114, + 0.09336177259683609, + 0.042762573808431625, + 0.05426397547125816, + 0.03281020000576973, + 0.10676930844783783, + 0.03386383503675461, + 0.08820495754480362, + 0.08923031389713287, + 0.16791550815105438 + ], + "credit_map": [ + 0.05818016827106476, + 0.020902467891573906, + 0.027423877269029617, + 0.11461734771728516, + 0.010706635192036629, + 0.021528644487261772, + 0.0217895545065403, + 0.01566886529326439, + 0.008146834559738636, + 0.007779213134199381, + 0.013295057229697704, + 0.01754075102508068, + 0.01182650402188301, + 0.009020843543112278, + 0.01289242785423994, + 0.01286014262586832, + 0.007515207398682833, + 0.004587450530380011, + 0.0075975097715854645, + 0.005270775407552719, + 0.003048742888495326, + 0.006257039029151201, + 0.0037391299847513437, + 0.004746715072542429, + 0.00842658244073391, + 0.008355283178389072, + 0.006595729850232601, + 0.005537362769246101, + 0.006210732273757458, + 0.006077437661588192, + 0.006142552010715008, + 0.006918291095644236, + 0.006003604270517826, + 0.004081402439624071, + 0.004488575272262096, + 0.003941583447158337, + 0.0042734001763165, + 0.0028746325988322496, + 0.00413768645375967, + 0.0030093141831457615, + 0.01413875725120306, + 0.016024427488446236, + 0.010355059057474136, + 0.02082769200205803, + 0.010720171965658665, + 0.013293411582708359, + 0.011831051670014858, + 0.009578741155564785, + 0.004170100204646587, + 0.0030555492267012596, + 0.0035760407336056232, + 0.0031438253354281187, + 0.007271893788129091, + 0.004317762330174446, + 0.004436095245182514, + 0.0038925670087337494, + 0.01415970642119646, + 0.007647596299648285, + 0.010045738890767097, + 0.008990858681499958, + 0.012895421124994755, + 0.010098785161972046, + 0.013958478346467018, + 0.010408374480903149, + 0.008870439603924751, + 0.006719849072396755, + 0.01086968183517456, + 0.008159173652529716, + 0.014302906580269337, + 0.010813470929861069, + 0.01893533393740654, + 0.010559454560279846, + 0.01686934567987919, + 0.012308359146118164, + 0.026386821642518044, + 0.021066833287477493, + 0.019865015521645546, + 0.022075029090046883, + 0.02598760277032852, + 0.02335650473833084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.054948654025793076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205455958843231 + ], + "opposite_logprob_tokens": [ + -2.8120546340942383 + ], + "logprob_reward": -0.06205455958843231, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205455958843231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28343909978866577, + 0.08910945802927017, + 0.07536885887384415, + 0.04293500632047653, + 0.038716863840818405, + 0.10846933722496033, + 0.09542366862297058, + 0.12528228759765625, + 0.10960540920495987, + 0.03164992481470108 + ], + "credit_map": [ + 0.066447913646698, + 0.023067202419042587, + 0.026152150705456734, + 0.1044861301779747, + 0.01115364022552967, + 0.020333975553512573, + 0.018328644335269928, + 0.013469474390149117, + 0.008058622479438782, + 0.007645001634955406, + 0.009816317819058895, + 0.015443709678947926, + 0.01102595217525959, + 0.011435892432928085, + 0.012210122309625149, + 0.01347384788095951, + 0.016106950119137764, + 0.011202595196664333, + 0.01162394043058157, + 0.009051048196852207, + 0.004328399430960417, + 0.008868873119354248, + 0.00621303915977478, + 0.007974023930728436, + 0.004852112382650375, + 0.006202234420925379, + 0.0055630989372730255, + 0.004358576610684395, + 0.004693598952144384, + 0.007016152609139681, + 0.005260291043668985, + 0.00498894602060318, + 0.004698057193309069, + 0.0067293825559318066, + 0.003962961491197348, + 0.003980813082307577, + 0.003975602798163891, + 0.0032466857228428125, + 0.008123146370053291, + 0.0040002185851335526, + 0.011791425757110119, + 0.012438713572919369, + 0.008474734611809254, + 0.018216637894511223, + 0.011428807862102985, + 0.015076257288455963, + 0.0186761487275362, + 0.012366626411676407, + 0.0075992257334291935, + 0.011925894767045975, + 0.00854160264134407, + 0.008190762251615524, + 0.024327289313077927, + 0.010690402239561081, + 0.013132159598171711, + 0.011016342788934708, + 0.016680173575878143, + 0.014797721989452839, + 0.011054396629333496, + 0.013319703750312328, + 0.03341474011540413, + 0.014276432804763317, + 0.01234685629606247, + 0.009392288513481617, + 0.010498777031898499, + 0.008146311156451702, + 0.015840250998735428, + 0.00940045714378357, + 0.021327583119273186, + 0.012781125493347645, + 0.018045706674456596, + 0.013565211556851864, + 0.003524534171447158, + 0.0032935175113379955, + 0.004977299366146326, + 0.0030929583590477705, + 0.003960215020924807, + 0.004370579496026039, + 0.003956390079110861, + 0.004474432673305273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.06205455958843231 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22128647565841675, + 0.06479554623365402, + 0.04588582366704941, + 0.0690360739827156, + 0.03474707528948784, + 0.19689401984214783, + 0.06105382367968559, + 0.12128429114818573, + 0.06557542085647583, + 0.11944141983985901 + ], + "credit_map": [ + 0.038668956607580185, + 0.020100777968764305, + 0.02484135702252388, + 0.06702941656112671, + 0.016253719106316566, + 0.015823496505618095, + 0.020177535712718964, + 0.018391236662864685, + 0.008320444263517857, + 0.0060005090199410915, + 0.007606774102896452, + 0.008847475051879883, + 0.005687297787517309, + 0.008293994702398777, + 0.00725201889872551, + 0.012787039391696453, + 0.005030324216932058, + 0.004179340321570635, + 0.006790379993617535, + 0.0066732014529407024, + 0.003325073281303048, + 0.005836918950080872, + 0.00676001375541091, + 0.007290579378604889, + 0.007955847308039665, + 0.010867737233638763, + 0.00880490429699421, + 0.0069728814996778965, + 0.011357557959854603, + 0.007401793263852596, + 0.008834737353026867, + 0.006840631365776062, + 0.005441558081656694, + 0.007662511430680752, + 0.0039961026050150394, + 0.0030062824953347445, + 0.002923156600445509, + 0.003484379267320037, + 0.004470848012715578, + 0.0037622409872710705, + 0.025924351066350937, + 0.02673020213842392, + 0.01917419768869877, + 0.031111950054764748, + 0.019227933138608932, + 0.024534868076443672, + 0.03248392790555954, + 0.017706604674458504, + 0.008262453600764275, + 0.011230501346290112, + 0.006469605024904013, + 0.007287684362381697, + 0.006847287528216839, + 0.007515507284551859, + 0.00536077655851841, + 0.008080019615590572, + 0.02047589235007763, + 0.00994866993278265, + 0.018800562247633934, + 0.013384765945374966, + 0.013736149296164513, + 0.014305122196674347, + 0.013570388779044151, + 0.017062745988368988, + 0.013602344319224358, + 0.006237793248146772, + 0.009231814183294773, + 0.006801501847803593, + 0.007993715815246105, + 0.008287854492664337, + 0.007081933785229921, + 0.006338465493172407, + 0.009789199568331242, + 0.009192613884806633, + 0.016189249232411385, + 0.011206741444766521, + 0.01623334176838398, + 0.017209690064191818, + 0.018020397052168846, + 0.021600209176540375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.01607329398393631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704364955425262 + ], + "opposite_logprob_tokens": [ + -2.1270437240600586 + ], + "logprob_reward": -0.12704364955425262, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704364955425262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.377797931432724, + 0.23761288821697235, + 0.0841340646147728, + 0.1223662793636322, + 0.027630962431430817, + 0.05510010942816734, + 0.012366151437163353, + 0.03235916420817375, + 0.023311037570238113, + 0.02732139453291893 + ], + "credit_map": [ + 0.0897001400589943, + 0.027447426691651344, + 0.034629542380571365, + 0.14756515622138977, + 0.011918236501514912, + 0.027225753292441368, + 0.02376343496143818, + 0.015548241324722767, + 0.021770671010017395, + 0.025504322722554207, + 0.027227263897657394, + 0.05060124769806862, + 0.02237255871295929, + 0.027705760672688484, + 0.032002367079257965, + 0.030428720638155937, + 0.019879892468452454, + 0.014634019695222378, + 0.009811476804316044, + 0.010908429510891438, + 0.006169304251670837, + 0.008641989901661873, + 0.005703388713300228, + 0.008385566994547844, + 0.021998846903443336, + 0.029168471693992615, + 0.024398187175393105, + 0.010529262945055962, + 0.007330820895731449, + 0.008378071710467339, + 0.011049928143620491, + 0.009512693621218204, + 0.003716003382578492, + 0.006519821472465992, + 0.0036186561919748783, + 0.0021066581830382347, + 0.002537683118134737, + 0.002281410386785865, + 0.004485070239752531, + 0.002365661785006523, + 0.007946625351905823, + 0.006784750614315271, + 0.005607613828033209, + 0.01073527429252863, + 0.00479660602286458, + 0.007907556369900703, + 0.006051512900739908, + 0.0052701737731695175, + 0.0019165227422490716, + 0.0012432824587449431, + 0.0013797549763694406, + 0.0016944799572229385, + 0.0010805570054799318, + 0.00165002781432122, + 0.002150279935449362, + 0.0012512480607256293, + 0.005010535009205341, + 0.003946377895772457, + 0.0032612537033855915, + 0.0028528235852718353, + 0.007217092905193567, + 0.0037623431999236345, + 0.002875897102057934, + 0.0034328405745327473, + 0.003653634339570999, + 0.00176721322350204, + 0.0024879577104002237, + 0.002449421677738428, + 0.003250991925597191, + 0.0026983399875462055, + 0.0042334385216236115, + 0.002770040649920702, + 0.002899406012147665, + 0.0025532697327435017, + 0.0038940412923693657, + 0.0037037490401417017, + 0.004017520695924759, + 0.0036643571220338345, + 0.003551725996658206, + 0.003037327202036977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.12704364955425262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle extreme close-up slowly pans to the right, meticulously tracking the deliberate movement of the artist's brush across the canvas. The brush itself is a classic hazelnut with a slender wooden handle stained with ghosts of previous colors, its fine sable bristles thick with a mixture of thick, lustrous cadmium yellow and sky blue paint. With a steady, reverent glide, the bristles bend slightly, leaving vibrant, lifeless brushstrokes across the rough, latticed texture of the canvas, already bearing the weight of a painterly, impressionistic composition. The entire scene is bathed in dramatic, ethereal light, emanating from a vast, out-of-focus Gothic stained glass window in the historic church setting. A kaleidoscope of pale and deep sapphire, ruby red, and brilliant gold penetrates the cool, dim air, casting a divine mosaic of colorful patterns across the stone floor and the artist's wooden easel. A ray of golden light catches the polished silver ring of the moving brush. The atmosphere is one of profound tranquility and focused creativity, as the act of painting becomes a direct translation of the sacred, vibrant light that fills the sacred space.", + "seed": 17562, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_15/index_9891_seed_17562_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09046348780393601, + "vlm_logprob": -0.09046348780393601, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.09046348780393601, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056729756295681 + ], + "opposite_logprob_tokens": [ + -3.431729793548584 + ], + "logprob_reward": -0.056729756295681, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.056729756295681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17844310402870178, + 0.05155874043703079, + 0.10966792702674866, + 0.06526742875576019, + 0.022670429199934006, + 0.11442919820547104, + 0.1270376741886139, + 0.050956569612026215, + 0.13108526170253754, + 0.14888370037078857 + ], + "credit_map": [ + 0.03770594671368599, + 0.027090562507510185, + 0.01624426804482937, + 0.020181432366371155, + 0.018982918933033943, + 0.02068260870873928, + 0.019954107701778412, + 0.01760125532746315, + 0.008569102734327316, + 0.0059400638565421104, + 0.006648074369877577, + 0.006517116911709309, + 0.005374185740947723, + 0.008683533407747746, + 0.004748720210045576, + 0.005077942740172148, + 0.032734356820583344, + 0.016306966543197632, + 0.009063133038580418, + 0.01007652934640646, + 0.010026143863797188, + 0.014419429004192352, + 0.0090638417750597, + 0.007977534085512161, + 0.009862334467470646, + 0.015214802697300911, + 0.005242879502475262, + 0.0048644691705703735, + 0.01304010022431612, + 0.00646174093708396, + 0.005915071349591017, + 0.004666029009968042, + 0.005708979442715645, + 0.006036566104739904, + 0.0015407341998070478, + 0.002032667165622115, + 0.002187376609072089, + 0.002049598377197981, + 0.0014278937596827745, + 0.0016866130754351616, + 0.0351373665034771, + 0.025006473064422607, + 0.00939379632472992, + 0.009594442322850227, + 0.013143670745193958, + 0.007691601291298866, + 0.008882232010364532, + 0.005579618737101555, + 0.021532537415623665, + 0.02312185801565647, + 0.01254245825111866, + 0.012276838533580303, + 0.02802165225148201, + 0.01056098472326994, + 0.011647296138107777, + 0.007334050722420216, + 0.006120339501649141, + 0.005974194034934044, + 0.009698956273496151, + 0.005769972689449787, + 0.005868036299943924, + 0.007262634113430977, + 0.005982539150863886, + 0.004279895685613155, + 0.00944256130605936, + 0.014379573985934258, + 0.02259003557264805, + 0.016055410727858543, + 0.01407527830451727, + 0.01426479872316122, + 0.030433814972639084, + 0.009843777865171432, + 0.008240221999585629, + 0.009329511784017086, + 0.04194105416536331, + 0.023044098168611526, + 0.009700840339064598, + 0.016219783574342728, + 0.017463210970163345, + 0.02294498309493065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.056729756295681 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279869198799133 + ], + "opposite_logprob_tokens": [ + -2.017798662185669 + ], + "logprob_reward": -0.14279869198799133, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279869198799133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12411375343799591, + 0.04431881383061409, + 0.04384631663560867, + 0.0701930820941925, + 0.11363609880208969, + 0.11097711324691772, + 0.12705069780349731, + 0.09995768219232559, + 0.17330309748649597, + 0.09260331839323044 + ], + "credit_map": [ + 0.01720629818737507, + 0.015955060720443726, + 0.010876834392547607, + 0.008541678078472614, + 0.0175077673047781, + 0.01761983335018158, + 0.023427801206707954, + 0.01297848578542471, + 0.005805482156574726, + 0.007235884200781584, + 0.004096100106835365, + 0.005044034216552973, + 0.004474414046853781, + 0.008228194899857044, + 0.0053256298415362835, + 0.004109078086912632, + 0.0062835547141730785, + 0.004841684829443693, + 0.004481079988181591, + 0.0046941619366407394, + 0.004069215152412653, + 0.007083055563271046, + 0.006689771078526974, + 0.005703797563910484, + 0.01139067206531763, + 0.008176390081644058, + 0.0065399715676903725, + 0.004606078378856182, + 0.012286162003874779, + 0.012440858408808708, + 0.00813068263232708, + 0.006622275337576866, + 0.04198738932609558, + 0.042852070182561874, + 0.005038395524024963, + 0.004393512848764658, + 0.005056815687566996, + 0.006648611277341843, + 0.003022977616637945, + 0.004636324476450682, + 0.024324025958776474, + 0.014784724451601505, + 0.008916137740015984, + 0.01326759997755289, + 0.011080958880484104, + 0.017165474593639374, + 0.013388711027801037, + 0.008049502968788147, + 0.019118618220090866, + 0.02387601137161255, + 0.011719350703060627, + 0.011616377159953117, + 0.02531188540160656, + 0.013491458259522915, + 0.013541112653911114, + 0.00837588869035244, + 0.013928315602242947, + 0.012931169010698795, + 0.018340758979320526, + 0.010216202586889267, + 0.01174153108149767, + 0.01515545230358839, + 0.009684809483587742, + 0.007959451526403427, + 0.01172658335417509, + 0.018496036529541016, + 0.028358716517686844, + 0.020509742200374603, + 0.01789035089313984, + 0.021317513659596443, + 0.041286613792181015, + 0.013717565685510635, + 0.0050686439499258995, + 0.006456938572227955, + 0.022879064083099365, + 0.02283322997391224, + 0.005785328336060047, + 0.009883365593850613, + 0.01185187604278326, + 0.007844878360629082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.14279869198799133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307093471288681 + ], + "opposite_logprob_tokens": [ + -3.1680710315704346 + ], + "logprob_reward": -0.04307093471288681, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307093471288681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08686884492635727, + 0.07123708724975586, + 0.06219468265771866, + 0.0783528983592987, + 0.2974461615085602, + 0.09727166593074799, + 0.06261342763900757, + 0.11010981351137161, + 0.07217203825712204, + 0.06173339858651161 + ], + "credit_map": [ + 0.012804797850549221, + 0.01157329697161913, + 0.008755696937441826, + 0.007960749790072441, + 0.010086441412568092, + 0.011013466864824295, + 0.01342235691845417, + 0.011252028867602348, + 0.008312594145536423, + 0.011199383065104485, + 0.005746172741055489, + 0.006932260002940893, + 0.008145490661263466, + 0.012164200656116009, + 0.012241585180163383, + 0.006495395675301552, + 0.008153301663696766, + 0.006606068927794695, + 0.004394229035824537, + 0.008487945422530174, + 0.006200439762324095, + 0.012816772796213627, + 0.009838827885687351, + 0.005697096232324839, + 0.011564246378839016, + 0.011456877924501896, + 0.007807726040482521, + 0.006087254732847214, + 0.014625534415245056, + 0.009597938507795334, + 0.010666069574654102, + 0.006547247525304556, + 0.11365530639886856, + 0.11310338973999023, + 0.009242656640708447, + 0.012027934193611145, + 0.013621753081679344, + 0.014287943951785564, + 0.009431599639356136, + 0.012075531296432018, + 0.026652798056602478, + 0.019433896988630295, + 0.00527004711329937, + 0.008113313466310501, + 0.008495282381772995, + 0.011365323327481747, + 0.01160688977688551, + 0.0063341096974909306, + 0.010369875468313694, + 0.015993131324648857, + 0.006065155379474163, + 0.008410830050706863, + 0.00556989898905158, + 0.00588372303172946, + 0.005715708713978529, + 0.004605099093168974, + 0.013942881487309933, + 0.017358727753162384, + 0.02101288177073002, + 0.012394077144563198, + 0.011095838621258736, + 0.01215899083763361, + 0.011602262035012245, + 0.01054414827376604, + 0.005269419867545366, + 0.005519988480955362, + 0.01944870874285698, + 0.013853811658918858, + 0.005018531810492277, + 0.004945812746882439, + 0.013002507388591766, + 0.005113247316330671, + 0.0033311480656266212, + 0.004030180163681507, + 0.013367990031838417, + 0.015479428693652153, + 0.0047620306722819805, + 0.0066591654904186726, + 0.00921263825148344, + 0.004890807904303074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.04307093471288681 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815562069416046 + ], + "opposite_logprob_tokens": [ + -3.529815673828125 + ], + "logprob_reward": -0.029815562069416046, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815562069416046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10762457549571991, + 0.06407467275857925, + 0.11655978113412857, + 0.11817044764757156, + 0.18407338857650757, + 0.10782209783792496, + 0.07592308521270752, + 0.09267671406269073, + 0.08921284973621368, + 0.043862368911504745 + ], + "credit_map": [ + 0.02059108205139637, + 0.02225005254149437, + 0.008077054284512997, + 0.00851522944867611, + 0.01416098140180111, + 0.009735756553709507, + 0.014556773006916046, + 0.009737655520439148, + 0.00827729981392622, + 0.007677192334085703, + 0.005818458274006844, + 0.012411407195031643, + 0.005913169588893652, + 0.007699406705796719, + 0.009428059682250023, + 0.006849680095911026, + 0.034394413232803345, + 0.022109873592853546, + 0.008816632442176342, + 0.009990713559091091, + 0.009536375291645527, + 0.009230813942849636, + 0.013190308585762978, + 0.009290672838687897, + 0.021143103018403053, + 0.024580955505371094, + 0.009075312875211239, + 0.009765462018549442, + 0.021011430770158768, + 0.008223503828048706, + 0.01842040754854679, + 0.005950281396508217, + 0.06648742407560349, + 0.06747226417064667, + 0.008726147934794426, + 0.007755286525934935, + 0.008260831236839294, + 0.012167579494416714, + 0.005888170562684536, + 0.007315720431506634, + 0.025185372680425644, + 0.02045929618179798, + 0.006779246497899294, + 0.009327162988483906, + 0.008932975120842457, + 0.016223490238189697, + 0.012352528050541878, + 0.008562029339373112, + 0.010646006092429161, + 0.018712198361754417, + 0.009377909824252129, + 0.007242812309414148, + 0.010127747431397438, + 0.007603061385452747, + 0.006772201973944902, + 0.005441154818981886, + 0.01280271913856268, + 0.014571093954145908, + 0.01724197342991829, + 0.009446635842323303, + 0.009637142531573772, + 0.00960357766598463, + 0.009578010067343712, + 0.009795570746064186, + 0.006629754323512316, + 0.00846020970493555, + 0.0199756920337677, + 0.015277263708412647, + 0.006590257398784161, + 0.006843505892902613, + 0.01806188002228737, + 0.007374288514256477, + 0.0033267103135585785, + 0.0047640083357691765, + 0.005302630830556154, + 0.009490606375038624, + 0.005106136668473482, + 0.005415929015725851, + 0.0064492253586649895, + 0.004007122945040464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.029815562069416046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990249395370483 + ], + "opposite_logprob_tokens": [ + -1.8049025535583496 + ], + "logprob_reward": -0.17990249395370483, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990249395370483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.327591210603714, + 0.04855957627296448, + 0.23603211343288422, + 0.0664772242307663, + 0.05150894820690155, + 0.08581790328025818, + 0.08540192246437073, + 0.029511483386158943, + 0.03344915062189102, + 0.03565051779150963 + ], + "credit_map": [ + 0.09242111444473267, + 0.07027491927146912, + 0.02233254909515381, + 0.025176849216222763, + 0.04213808849453926, + 0.025648323819041252, + 0.019863871857523918, + 0.029735460877418518, + 0.008470913395285606, + 0.005312113557010889, + 0.004289065487682819, + 0.008344519883394241, + 0.005016688257455826, + 0.007694525644183159, + 0.005225369706749916, + 0.004206376150250435, + 0.046982407569885254, + 0.03792962431907654, + 0.01813076250255108, + 0.0230193380266428, + 0.02308913879096508, + 0.028440862894058228, + 0.03183365613222122, + 0.026606278494000435, + 0.010230787098407745, + 0.008181256242096424, + 0.0064226556569337845, + 0.005914174020290375, + 0.007141788490116596, + 0.009767410345375538, + 0.013144700787961483, + 0.005674443673342466, + 0.013691909611225128, + 0.011447473429143429, + 0.005243084859102964, + 0.00465718237683177, + 0.0038575029466301203, + 0.005812715273350477, + 0.00339799583889544, + 0.003401077352464199, + 0.011930732987821102, + 0.008846123702824116, + 0.007124108728021383, + 0.01215725764632225, + 0.008674843236804008, + 0.01446934137493372, + 0.014862451702356339, + 0.007753035984933376, + 0.014254417270421982, + 0.016262609511613846, + 0.009672495536506176, + 0.010423114523291588, + 0.01315244846045971, + 0.007371810264885426, + 0.008360335603356361, + 0.005904683843255043, + 0.003100861329585314, + 0.005039580166339874, + 0.005007115658372641, + 0.002888640621677041, + 0.0034160169307142496, + 0.003653252264484763, + 0.003388783195987344, + 0.003017227165400982, + 0.002852371893823147, + 0.0026234262622892857, + 0.007185435853898525, + 0.005474414676427841, + 0.003958092071115971, + 0.003954919055104256, + 0.00460817851126194, + 0.0027923111338168383, + 0.0028834803961217403, + 0.004463336430490017, + 0.00455428846180439, + 0.007767373230308294, + 0.0033989015500992537, + 0.004945378750562668, + 0.004892295692116022, + 0.002745460020378232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.17990249395370483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle extreme close-up slowly pans to the right, meticulously tracking the deliberate movement of the artist's brush across the canvas. The brush itself is a classic hazelnut with a slender wooden handle stained with ghosts of previous colors, its fine sable bristles thick with a mixture of thick, lustrous cadmium yellow and sky blue paint. With a steady, reverent glide, the bristles bend slightly, leaving vibrant, lifeless brushstrokes across the rough, latticed texture of the canvas, already bearing the weight of a painterly, impressionistic composition. The entire scene is bathed in dramatic, ethereal light, emanating from a vast, out-of-focus Gothic stained glass window in the historic church setting. A kaleidoscope of pale and deep sapphire, ruby red, and brilliant gold penetrates the cool, dim air, casting a divine mosaic of colorful patterns across the stone floor and the artist's wooden easel. A ray of golden light catches the polished silver ring of the moving brush. The atmosphere is one of profound tranquility and focused creativity, as the act of painting becomes a direct translation of the sacred, vibrant light that fills the sacred space.", + "seed": 17563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_15/index_9891_seed_17563_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09837833791971207, + "vlm_logprob": -0.09837833791971207, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.09837833791971207, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0716920793056488 + ], + "opposite_logprob_tokens": [ + -3.0716919898986816 + ], + "logprob_reward": -0.0716920793056488, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0716920793056488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1898684948682785, + 0.09832604229450226, + 0.08216873556375504, + 0.16174115240573883, + 0.20059217512607574, + 0.0639486238360405, + 0.03283447399735451, + 0.045405417680740356, + 0.08046837151050568, + 0.0446464978158474 + ], + "credit_map": [ + 0.014937379397451878, + 0.04040129482746124, + 0.028666868805885315, + 0.02578178234398365, + 0.02525501511991024, + 0.017471924424171448, + 0.023076968267560005, + 0.014277276583015919, + 0.010401262901723385, + 0.01329214870929718, + 0.012340444140136242, + 0.01357270684093237, + 0.01410853210836649, + 0.014230334199965, + 0.011002954095602036, + 0.009377662092447281, + 0.011761474423110485, + 0.015026876702904701, + 0.01614132896065712, + 0.00791926495730877, + 0.008203374221920967, + 0.011315741576254368, + 0.006770498119294643, + 0.005030169617384672, + 0.013705641962587833, + 0.018513232469558716, + 0.014215197414159775, + 0.019603634253144264, + 0.041742220520973206, + 0.030120301991701126, + 0.011627081781625748, + 0.012213848531246185, + 0.027319585904479027, + 0.057759758085012436, + 0.06595373153686523, + 0.008253137581050396, + 0.010496252216398716, + 0.01041078008711338, + 0.007139556109905243, + 0.013259385712444782, + 0.008586421608924866, + 0.011495430953800678, + 0.010008737444877625, + 0.005392985884100199, + 0.006485476158559322, + 0.010482889600098133, + 0.0034758022520691156, + 0.00802087876945734, + 0.0035484149120748043, + 0.005343719385564327, + 0.0031835990957915783, + 0.0027603083290159702, + 0.006261879578232765, + 0.005907686892896891, + 0.002306097885593772, + 0.0035227707121521235, + 0.00492676580324769, + 0.008282207883894444, + 0.004182828590273857, + 0.005171989090740681, + 0.0038651919458061457, + 0.008451824076473713, + 0.003629982005804777, + 0.006894630379974842, + 0.008661946281790733, + 0.009841639548540115, + 0.006741865538060665, + 0.0058949775993824005, + 0.007807253859937191, + 0.0156638752669096, + 0.006703306920826435, + 0.01915350928902626, + 0.005303421523422003, + 0.0023705149069428444, + 0.0031971775460988283, + 0.0027658117469400167, + 0.007236543111503124, + 0.012558779679238796, + 0.0022890588734298944, + 0.008925190195441246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0716920793056488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001670449972153 + ], + "opposite_logprob_tokens": [ + -2.695016622543335 + ], + "logprob_reward": -0.07001670449972153, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001670449972153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07124421745538712, + 0.06925316900014877, + 0.06350651383399963, + 0.13323412835597992, + 0.12071540206670761, + 0.0674934983253479, + 0.14713867008686066, + 0.03191649913787842, + 0.15322892367839813, + 0.14226897060871124 + ], + "credit_map": [ + 0.005446471273899078, + 0.010124405845999718, + 0.010137917473912239, + 0.011820774525403976, + 0.009197575971484184, + 0.008388826623558998, + 0.010968133807182312, + 0.005160113796591759, + 0.00782699603587389, + 0.011230076663196087, + 0.009105633944272995, + 0.0110723115503788, + 0.008524567820131779, + 0.008166009560227394, + 0.008310324512422085, + 0.0050172447226941586, + 0.00973224826157093, + 0.008034576661884785, + 0.010702154599130154, + 0.007283562794327736, + 0.005789157934486866, + 0.006409512832760811, + 0.00495678186416626, + 0.010598516091704369, + 0.014284146018326283, + 0.015875056385993958, + 0.012046901509165764, + 0.04189276322722435, + 0.014333912171423435, + 0.013801829889416695, + 0.012916664592921734, + 0.008082852698862553, + 0.015597994439303875, + 0.03350536525249481, + 0.03310620039701462, + 0.006266676355153322, + 0.0074017164297401905, + 0.008282321505248547, + 0.005970729049295187, + 0.010584387928247452, + 0.013241576962172985, + 0.00905725546181202, + 0.012713742442429066, + 0.00516419718042016, + 0.0057293083518743515, + 0.009400255046784878, + 0.004057020414620638, + 0.00813014805316925, + 0.01585901528596878, + 0.019365621730685234, + 0.014717478305101395, + 0.01440514251589775, + 0.02097141556441784, + 0.034443922340869904, + 0.009767993353307247, + 0.017608080059289932, + 0.0027606601361185312, + 0.004566579591482878, + 0.0027013637591153383, + 0.0033710612915456295, + 0.008035341277718544, + 0.005421337205916643, + 0.002470148028805852, + 0.002590007847175002, + 0.011854543350636959, + 0.012622972950339317, + 0.011976286768913269, + 0.01510061975568533, + 0.045833759009838104, + 0.027660829946398735, + 0.01106164138764143, + 0.017118269577622414, + 0.025386817753314972, + 0.008326823823153973, + 0.009313827380537987, + 0.009591805748641491, + 0.013661582954227924, + 0.018750641494989395, + 0.005911964923143387, + 0.05132550746202469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.07001670449972153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497799068689346 + ], + "opposite_logprob_tokens": [ + -2.9299778938293457 + ], + "logprob_reward": -0.05497799068689346, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497799068689346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15408828854560852, + 0.12635831534862518, + 0.07907623797655106, + 0.10003568232059479, + 0.19269947707653046, + 0.08779237419366837, + 0.0788111761212349, + 0.06885722279548645, + 0.040472231805324554, + 0.07180900871753693 + ], + "credit_map": [ + 0.012692625634372234, + 0.04147251695394516, + 0.029353179037570953, + 0.014725224114954472, + 0.016803480684757233, + 0.01181910838931799, + 0.0176690723747015, + 0.009553080424666405, + 0.015208076685667038, + 0.02298484556376934, + 0.014340264722704887, + 0.013974764384329319, + 0.018848014995455742, + 0.017324350774288177, + 0.014716305769979954, + 0.008961686864495277, + 0.011176357977092266, + 0.010560273192822933, + 0.01211201585829258, + 0.007445655297487974, + 0.009232880547642708, + 0.016685165464878082, + 0.007774559315294027, + 0.004089332651346922, + 0.010622966103255749, + 0.010422661900520325, + 0.008835733868181705, + 0.02500877156853676, + 0.014231878332793713, + 0.012439713813364506, + 0.011187180876731873, + 0.007286777254194021, + 0.028395304456353188, + 0.05079163610935211, + 0.06524704396724701, + 0.008437997661530972, + 0.011715165339410305, + 0.00991636049002409, + 0.006988186854869127, + 0.011207790113985538, + 0.013790248893201351, + 0.019682930782437325, + 0.010102946311235428, + 0.006320158485323191, + 0.009626023471355438, + 0.012875880114734173, + 0.0067686825059354305, + 0.008625500835478306, + 0.007137457840144634, + 0.006515654269605875, + 0.008258500136435032, + 0.008712269365787506, + 0.01849689893424511, + 0.018312247470021248, + 0.004731765016913414, + 0.006646382622420788, + 0.011120105162262917, + 0.0071214041672647, + 0.007036369293928146, + 0.0044667525216937065, + 0.010026042349636555, + 0.012197651900351048, + 0.004511185921728611, + 0.012377713806927204, + 0.007176592480391264, + 0.004139603581279516, + 0.004068125505000353, + 0.005166844464838505, + 0.005345340818166733, + 0.0064821806736290455, + 0.003052406245842576, + 0.005041135940700769, + 0.01136142760515213, + 0.004389198496937752, + 0.004260989371687174, + 0.00371554191224277, + 0.01067475788295269, + 0.012096473947167397, + 0.004227263852953911, + 0.021083349362015724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.05497799068689346 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043080754578113556 + ], + "opposite_logprob_tokens": [ + -3.1680808067321777 + ], + "logprob_reward": -0.043080754578113556, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043080754578113556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1178777888417244, + 0.1417769491672516, + 0.05155801773071289, + 0.22205808758735657, + 0.13745909929275513, + 0.06664050370454788, + 0.04559605196118355, + 0.07269477844238281, + 0.036185070872306824, + 0.10815364122390747 + ], + "credit_map": [ + 0.008825818076729774, + 0.024182917550206184, + 0.01509231049567461, + 0.013359655626118183, + 0.01467988733202219, + 0.014034977182745934, + 0.016190897673368454, + 0.011511317454278469, + 0.012931304052472115, + 0.03223341330885887, + 0.017101379111409187, + 0.011964811943471432, + 0.02958536520600319, + 0.016629865393042564, + 0.012434526346623898, + 0.008896277286112309, + 0.00713723711669445, + 0.006379069294780493, + 0.008728759363293648, + 0.004952350631356239, + 0.005360748153179884, + 0.007423904724419117, + 0.007787908893078566, + 0.003788033500313759, + 0.020938709378242493, + 0.02385859005153179, + 0.01960241235792637, + 0.04693801701068878, + 0.039507459849119186, + 0.037418220192193985, + 0.015183890238404274, + 0.018610790371894836, + 0.02004048228263855, + 0.03738996013998985, + 0.04756418615579605, + 0.005224606953561306, + 0.00760879460722208, + 0.006966066546738148, + 0.004293117206543684, + 0.008371871896088123, + 0.01163469534367323, + 0.008396516554057598, + 0.011525478214025497, + 0.005397319793701172, + 0.005779664497822523, + 0.009989764541387558, + 0.004050474613904953, + 0.009866585955023766, + 0.005000284872949123, + 0.006104744039475918, + 0.0053378683514893055, + 0.0044228252954781055, + 0.007509960327297449, + 0.009335181675851345, + 0.003329378319904208, + 0.0045558069832623005, + 0.00879833847284317, + 0.012366976588964462, + 0.007443178445100784, + 0.011346233077347279, + 0.009766052477061749, + 0.009204593487083912, + 0.006259536370635033, + 0.0075098685920238495, + 0.004403191152960062, + 0.002511246595531702, + 0.0030112769454717636, + 0.0031172065064311028, + 0.008960908278822899, + 0.006414645817130804, + 0.002894094679504633, + 0.004872496705502272, + 0.020905474200844765, + 0.006415750831365585, + 0.005861536134034395, + 0.005812954623252153, + 0.009988454170525074, + 0.012359650805592537, + 0.004422300960868597, + 0.04238750785589218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.043080754578113556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252124160528183 + ], + "opposite_logprob_tokens": [ + -1.5021241903305054 + ], + "logprob_reward": -0.252124160528183, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252124160528183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12901116907596588, + 0.08385638147592545, + 0.2540088891983032, + 0.11083368211984634, + 0.16188932955265045, + 0.04295472428202629, + 0.015592336654663086, + 0.04135198891162872, + 0.06542576104402542, + 0.09507570415735245 + ], + "credit_map": [ + 0.01003283355385065, + 0.023685496300458908, + 0.0199789647012949, + 0.020567888393998146, + 0.01708625815808773, + 0.013988695107400417, + 0.012959614396095276, + 0.010711426846683025, + 0.008929219096899033, + 0.0108738848939538, + 0.01074602548032999, + 0.011271431110799313, + 0.009648767299950123, + 0.010925819166004658, + 0.011835822835564613, + 0.009625409729778767, + 0.039635587483644485, + 0.03953571617603302, + 0.06100470945239067, + 0.023599889129400253, + 0.025224696844816208, + 0.03543885797262192, + 0.014408452436327934, + 0.015160977840423584, + 0.009479847736656666, + 0.015510939992964268, + 0.007650896441191435, + 0.025550710037350655, + 0.02705451287329197, + 0.008279512636363506, + 0.011020384728908539, + 0.006286880001425743, + 0.01980479620397091, + 0.05326857045292854, + 0.052108097821474075, + 0.006219062488526106, + 0.007360120303928852, + 0.009373310953378677, + 0.005553715396672487, + 0.008201655000448227, + 0.007499493192881346, + 0.007500725798308849, + 0.006760008633136749, + 0.0030304479878395796, + 0.003998005297034979, + 0.006201514042913914, + 0.002049422590062022, + 0.005915109068155289, + 0.0017079473473131657, + 0.0027974951080977917, + 0.0020424099639058113, + 0.00192107108887285, + 0.001964987488463521, + 0.0020813732407987118, + 0.0014299688627943397, + 0.0016470832051709294, + 0.0037566220853477716, + 0.00598929962143302, + 0.0028951496351510286, + 0.00534402672201395, + 0.009740748442709446, + 0.007459814660251141, + 0.0027641423512250185, + 0.0034021830651909113, + 0.009826462715864182, + 0.008020168170332909, + 0.00686822272837162, + 0.0055780671536922455, + 0.005625145509839058, + 0.011171940714120865, + 0.00418282113969326, + 0.014152932912111282, + 0.016836371272802353, + 0.006367465015500784, + 0.007418936118483543, + 0.007457968778908253, + 0.009849656373262405, + 0.011598442681133747, + 0.004069095943123102, + 0.03147776797413826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.252124160528183 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle extreme close-up slowly pans to the right, meticulously tracking the deliberate movement of the artist's brush across the canvas. The brush itself is a classic hazelnut with a slender wooden handle stained with ghosts of previous colors, its fine sable bristles thick with a mixture of thick, lustrous cadmium yellow and sky blue paint. With a steady, reverent glide, the bristles bend slightly, leaving vibrant, lifeless brushstrokes across the rough, latticed texture of the canvas, already bearing the weight of a painterly, impressionistic composition. The entire scene is bathed in dramatic, ethereal light, emanating from a vast, out-of-focus Gothic stained glass window in the historic church setting. A kaleidoscope of pale and deep sapphire, ruby red, and brilliant gold penetrates the cool, dim air, casting a divine mosaic of colorful patterns across the stone floor and the artist's wooden easel. A ray of golden light catches the polished silver ring of the moving brush. The atmosphere is one of profound tranquility and focused creativity, as the act of painting becomes a direct translation of the sacred, vibrant light that fills the sacred space.", + "seed": 17564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_15/index_9891_seed_17564_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04995787031948566, + "vlm_logprob": -0.04995787031948566, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.04995787031948566, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032356392592191696 + ], + "opposite_logprob_tokens": [ + -4.282356262207031 + ], + "logprob_reward": -0.032356392592191696, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.032356392592191696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24165090918540955, + 0.2046380192041397, + 0.04926448315382004, + 0.021924128755927086, + 0.022120434790849686, + 0.1305079311132431, + 0.06999187916517258, + 0.08487517386674881, + 0.05572561174631119, + 0.1193014606833458 + ], + "credit_map": [ + 0.03270408883690834, + 0.023181308060884476, + 0.048054151237010956, + 0.03761205077171326, + 0.016578806564211845, + 0.027521951124072075, + 0.018673868849873543, + 0.037324659526348114, + 0.025321688503026962, + 0.059286992996931076, + 0.013833791948854923, + 0.016577554866671562, + 0.017801757901906967, + 0.018987935036420822, + 0.008147628977894783, + 0.04468066245317459, + 0.003928558435291052, + 0.006180174648761749, + 0.005983396898955107, + 0.006374056916683912, + 0.006147260311990976, + 0.0074720256961882114, + 0.004489150829613209, + 0.00868985615670681, + 0.004197418689727783, + 0.0030244807712733746, + 0.002648379886522889, + 0.0026851429138332605, + 0.002230623736977577, + 0.002897289115935564, + 0.0016611310420557857, + 0.002579661086201668, + 0.002211374929174781, + 0.0026670682709664106, + 0.002109147608280182, + 0.0018277589697390795, + 0.002275610575452447, + 0.0033999066799879074, + 0.002842118265107274, + 0.004787448327988386, + 0.009107287041842937, + 0.03633195161819458, + 0.00937028881162405, + 0.00475652888417244, + 0.01292492263019085, + 0.01769767887890339, + 0.007056942209601402, + 0.033262331038713455, + 0.004634313751012087, + 0.004446197301149368, + 0.00569148501381278, + 0.003149457508698106, + 0.006140712648630142, + 0.020900573581457138, + 0.004069250542670488, + 0.020959889516234398, + 0.01053126435726881, + 0.012616564519703388, + 0.010567732155323029, + 0.006798013113439083, + 0.00903986394405365, + 0.01448129117488861, + 0.011063696816563606, + 0.00977674126625061, + 0.006035804282873869, + 0.01656615547835827, + 0.00974764209240675, + 0.008260371163487434, + 0.005007426720112562, + 0.0030698871705681086, + 0.0030130911618471146, + 0.0040252297185361385, + 0.011242830194532871, + 0.007536482065916061, + 0.03672916442155838, + 0.009512326680123806, + 0.015617010183632374, + 0.009296588599681854, + 0.0062608434818685055, + 0.02310621552169323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.032356392592191696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902539312839508 + ], + "opposite_logprob_tokens": [ + -2.4640254974365234 + ], + "logprob_reward": -0.08902539312839508, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902539312839508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17586266994476318, + 0.11460860818624496, + 0.02329094335436821, + 0.023005807772278786, + 0.01574653945863247, + 0.08195101469755173, + 0.15645480155944824, + 0.06273967027664185, + 0.061500184237957, + 0.2848397493362427 + ], + "credit_map": [ + 0.023266009986400604, + 0.012613371014595032, + 0.04348692670464516, + 0.023176126182079315, + 0.011022655293345451, + 0.021644840016961098, + 0.008516817353665829, + 0.03213593363761902, + 0.009549654088914394, + 0.03903121501207352, + 0.0068168337456882, + 0.0071840607561171055, + 0.011839473620057106, + 0.009803088381886482, + 0.004405864980071783, + 0.025978414341807365, + 0.002583943074569106, + 0.0027651493437588215, + 0.0031525499653071165, + 0.002877606078982353, + 0.0028857081197202206, + 0.003919695038348436, + 0.0018927857745438814, + 0.003213505493476987, + 0.0032952027395367622, + 0.0025176636409014463, + 0.0020802642684429884, + 0.002088108565658331, + 0.004719282500445843, + 0.003185070352628827, + 0.002230067504569888, + 0.0028901465702801943, + 0.0017340100603178144, + 0.0023760783951729536, + 0.0017213857499882579, + 0.0016205403953790665, + 0.0013739516725763679, + 0.002344822743907571, + 0.0022096247412264347, + 0.002366125350818038, + 0.005550782661885023, + 0.022254183888435364, + 0.005860971286892891, + 0.0035594988148659468, + 0.00773948198184371, + 0.011921313591301441, + 0.005400426685810089, + 0.019664350897073746, + 0.007047009188681841, + 0.004695484880357981, + 0.011299197562038898, + 0.0043257856741547585, + 0.010280156508088112, + 0.05709005147218704, + 0.005995933897793293, + 0.055721186101436615, + 0.007706033997237682, + 0.007608877960592508, + 0.008712398819625378, + 0.003106358926743269, + 0.005210415925830603, + 0.005381373688578606, + 0.004571218974888325, + 0.02044299617409706, + 0.006629223935306072, + 0.016053343191742897, + 0.008514221757650375, + 0.010596639476716518, + 0.006200910545885563, + 0.004170156549662352, + 0.004723463673144579, + 0.004612227436155081, + 0.023477064445614815, + 0.0368611216545105, + 0.036532603204250336, + 0.02539638616144657, + 0.07515089213848114, + 0.03360792249441147, + 0.019924409687519073, + 0.03388931602239609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.08902539312839508 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329576015472412 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.02329576015472412, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329576015472412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13077862560749054, + 0.23761065304279327, + 0.07272590696811676, + 0.02206634357571602, + 0.032155271619558334, + 0.14578746259212494, + 0.11569514870643616, + 0.02359212562441826, + 0.04937688261270523, + 0.170211523771286 + ], + "credit_map": [ + 0.019234545528888702, + 0.013018835335969925, + 0.013867342844605446, + 0.017864754423499107, + 0.007862536236643791, + 0.026847509667277336, + 0.009240297600626945, + 0.02284281887114048, + 0.036538224667310715, + 0.08737993240356445, + 0.02641274407505989, + 0.013727712444961071, + 0.024396363645792007, + 0.014320469461381435, + 0.006213665008544922, + 0.028621552512049675, + 0.003779742168262601, + 0.007412951439619064, + 0.008616133593022823, + 0.009368800558149815, + 0.01051014568656683, + 0.012300661765038967, + 0.0064812819473445415, + 0.014256198890507221, + 0.002532502869144082, + 0.003295237896963954, + 0.0027598415035754442, + 0.0029962293338030577, + 0.002136149210855365, + 0.002707607578486204, + 0.002959802048280835, + 0.0026789745315909386, + 0.005794166587293148, + 0.003149755299091339, + 0.0026950722094625235, + 0.003328354097902775, + 0.003322961274534464, + 0.0045710280537605286, + 0.003910842351615429, + 0.005383094307035208, + 0.011325453408062458, + 0.03902098163962364, + 0.012573394924402237, + 0.00626844959333539, + 0.014765949919819832, + 0.01890091598033905, + 0.00955306738615036, + 0.033379264175891876, + 0.00518418475985527, + 0.002454460598528385, + 0.009136307053267956, + 0.00263719679787755, + 0.0057402257807552814, + 0.04406661167740822, + 0.0037949439138174057, + 0.042681220918893814, + 0.002759661292657256, + 0.0055818031542003155, + 0.0021495213732123375, + 0.0019907832611352205, + 0.0019332761876285076, + 0.0021894569508731365, + 0.001479178899899125, + 0.005508447997272015, + 0.00969131849706173, + 0.0087328115478158, + 0.00633588619530201, + 0.00856029987335205, + 0.004350312054157257, + 0.0035379535984247923, + 0.004419857170432806, + 0.003748449496924877, + 0.01509226206690073, + 0.02182077243924141, + 0.027468634769320488, + 0.017094256356358528, + 0.04003024473786354, + 0.018251677975058556, + 0.011739708483219147, + 0.018713979050517082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.02329576015472412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072122380137444 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072122380137444, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072122380137444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10451454669237137, + 0.26837804913520813, + 0.04336683079600334, + 0.016279522329568863, + 0.03298657760024071, + 0.16634400188922882, + 0.12963831424713135, + 0.04136447235941887, + 0.15533848106861115, + 0.0417892150580883 + ], + "credit_map": [ + 0.011985095217823982, + 0.008209031075239182, + 0.02486455999314785, + 0.015432802028954029, + 0.009278439916670322, + 0.014105002395808697, + 0.00687706982716918, + 0.013762546703219414, + 0.06599941849708557, + 0.0324724018573761, + 0.03068406879901886, + 0.0341063030064106, + 0.026416217908263206, + 0.02740490436553955, + 0.01860894076526165, + 0.03268580511212349, + 0.0033669450785964727, + 0.005472737830132246, + 0.004392297472804785, + 0.004221712239086628, + 0.005632740445435047, + 0.007548025343567133, + 0.0033401271793991327, + 0.009392247535288334, + 0.00211990624666214, + 0.0019306510221213102, + 0.001453167642466724, + 0.001751860836520791, + 0.002250834135338664, + 0.0029452312737703323, + 0.0013426963705569506, + 0.0024851751513779163, + 0.005103959236294031, + 0.0038519841618835926, + 0.0029245801270008087, + 0.0035259744618088007, + 0.0035701915621757507, + 0.00431652320548892, + 0.004640106111764908, + 0.005053262691944838, + 0.015316304750740528, + 0.04168727621436119, + 0.017117340117692947, + 0.008103915490210056, + 0.016725342720746994, + 0.020661761984229088, + 0.012490333057940006, + 0.03424173966050148, + 0.007323031313717365, + 0.0033350761514157057, + 0.01062131766229868, + 0.00300047779455781, + 0.0071233119815588, + 0.047730665653944016, + 0.004447333514690399, + 0.04605710878968239, + 0.004629233852028847, + 0.006758878938853741, + 0.0037654475308954716, + 0.004144167993217707, + 0.004255060572177172, + 0.003983793780207634, + 0.00406144093722105, + 0.00976645015180111, + 0.025486784055829048, + 0.018993305042386055, + 0.01897992007434368, + 0.03364468738436699, + 0.014711576513946056, + 0.013851793482899666, + 0.014226547442376614, + 0.015443866141140461, + 0.003426335984840989, + 0.006239299196749926, + 0.00634318171069026, + 0.0029175677336752415, + 0.011018045246601105, + 0.004891781602054834, + 0.002386069158092141, + 0.004566938616335392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.016072122380137444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right while tracking the brush movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brush a classic hazelnut shape with a slender wooden handle stained with previous colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brush bristles thick with a mixture of cadmium yellow and sky blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by dramatic light from a vast, out-of-focus Gothic stained glass window in a church setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903968334197998 + ], + "opposite_logprob_tokens": [ + -2.4640398025512695 + ], + "logprob_reward": -0.08903968334197998, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903968334197998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2426440268754959, + 0.18570591509342194, + 0.02279466763138771, + 0.013045614585280418, + 0.023164162412285805, + 0.28359276056289673, + 0.059598490595817566, + 0.016473548486828804, + 0.07970380038022995, + 0.073277048766613 + ], + "credit_map": [ + 0.03582333028316498, + 0.024556806311011314, + 0.06286102533340454, + 0.032436929643154144, + 0.01045555155724287, + 0.023047229275107384, + 0.009645071811974049, + 0.04381809011101723, + 0.03447429835796356, + 0.06358537077903748, + 0.02241191640496254, + 0.011092734523117542, + 0.01754925586283207, + 0.010944909416139126, + 0.004904300905764103, + 0.02074313722550869, + 0.0023820155765861273, + 0.002584934700280428, + 0.0027684594970196486, + 0.0023337453603744507, + 0.002564740367233753, + 0.0037960316985845566, + 0.002009463496506214, + 0.004355277866125107, + 0.002226303331553936, + 0.0028585053514689207, + 0.00132105709053576, + 0.0010337616549804807, + 0.001267167623154819, + 0.0014598442940041423, + 0.0008140822756104171, + 0.00206489278934896, + 0.005469612777233124, + 0.0032963366247713566, + 0.0020675156265497208, + 0.0015776912914589047, + 0.0024954036343842745, + 0.0020421170629560947, + 0.001825341023504734, + 0.0043901437893509865, + 0.031156867742538452, + 0.0666184201836586, + 0.03029189631342888, + 0.014799785800278187, + 0.02263440005481243, + 0.0365767702460289, + 0.01784113608300686, + 0.0636734813451767, + 0.002787622855976224, + 0.00133542122785002, + 0.004625244066119194, + 0.001357787405140698, + 0.0029619017150253057, + 0.022651541978120804, + 0.002001303480938077, + 0.021877670660614967, + 0.002226546872407198, + 0.0027036245446652174, + 0.002664003986865282, + 0.0012357862433418632, + 0.0015270252479240298, + 0.001968494849279523, + 0.0018667875556275249, + 0.002281280467286706, + 0.01592697575688362, + 0.015513574704527855, + 0.013187691569328308, + 0.012275315821170807, + 0.005504684057086706, + 0.004959513898938894, + 0.006655734032392502, + 0.00568030821159482, + 0.005830513779073954, + 0.0037500751204788685, + 0.02153991162776947, + 0.0057066213339567184, + 0.011393624357879162, + 0.007035405375063419, + 0.003541309619322419, + 0.014479588717222214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.08903968334197998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_tensor([5404, 5404])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_tensor([5404, 5404])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..691caa21ad13b6a6aeb863865a1243a230907030 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_tensor([5404, 5404])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning slow-motion close-up, captured from a slightly upward angle, captures the mythical creature, the \"Brooktail,\" in a moment of solemn resoluteness. Its scales, shimmering with a cyan-blue metallic sheen, appear faintly amidst a soft, mysterious halo. With the sharp tips of its fins, it slowly but surely traces the bases of its wings, composed of translucent, glass-like feathers. With the tearing motion, each feather begins to disintegrate the moment it is released from its body. Through meticulous particle simulation, these fragments transform into countless golden powders shimmering with a warm glow. Like a shower of miniature, luminous stardust, they descend with a light and sacred grace, ultimately landing unerringly in a pristine brown pottery jar reverently held by a pair of calloused, weathered villagers. The camera quickly cuts to a close-up of the interior of the medicine jar. The viscous dark liquid in the jar begins to boil violently as soon as the golden powder is incorporated. What is tumbling around are not ordinary bubbles, but a texture that flows brilliantly like liquid metal. All the golden light quickly converges, rotates, and condenses towards the center of the jar, eventually forming a perfectly round golden pill that exudes a soft yet powerful breath of life. It quietly floats above the still churning liquid surface, and the light from within illuminates the rough inner wall of the pottery jar, creating a magical, hopeful atmosphere of alchemical transformation.", + "seed": 32847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_5404_seed_32847_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1013539125211537, + "vlm_logprob": -0.1013539125211537, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.1013539125211537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.284603476524353 + ], + "opposite_logprob_tokens": [ + -1.409603476524353 + ], + "logprob_reward": -0.284603476524353, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.284603476524353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06374756991863251, + 0.23063620924949646, + 0.09342730045318604, + 0.017602913081645966, + 0.3792199492454529, + 0.02038772962987423, + 0.013296699151396751, + 0.06868812441825867, + 0.09091103821992874, + 0.022082418203353882 + ], + "credit_map": [ + 0.006842497270554304, + 0.006069807335734367, + 0.01075458899140358, + 0.010518639348447323, + 0.00627335486933589, + 0.007777721155434847, + 0.007535399869084358, + 0.007975565269589424, + 0.030885396525263786, + 0.018921926617622375, + 0.020191410556435585, + 0.020197317004203796, + 0.03677327185869217, + 0.025844017043709755, + 0.04441387578845024, + 0.03340902179479599, + 0.007688573561608791, + 0.006413542199879885, + 0.008058485575020313, + 0.0074158720672130585, + 0.011639270931482315, + 0.01977728120982647, + 0.013609608635306358, + 0.01882467046380043, + 0.0018737241625785828, + 0.0019524968229234219, + 0.002112752990797162, + 0.0016749257920309901, + 0.0023452439345419407, + 0.0023319979663938284, + 0.002742849290370941, + 0.0025689243339002132, + 0.03151728957891464, + 0.11748723685741425, + 0.012553499080240726, + 0.009776380844414234, + 0.15366198122501373, + 0.03134157508611679, + 0.012791099958121777, + 0.010090925730764866, + 0.0017152026994153857, + 0.0028173865284770727, + 0.0019187342841178179, + 0.002306530252099037, + 0.0041055092588067055, + 0.0025857265572994947, + 0.002763738390058279, + 0.0021749043371528387, + 0.0013802433386445045, + 0.0025365236215293407, + 0.0011515150545164943, + 0.0017924632411450148, + 0.0018032153602689505, + 0.0014862040989100933, + 0.002175253350287676, + 0.0009712826577015221, + 0.024118730798363686, + 0.005376627668738365, + 0.0033871892374008894, + 0.007322354707866907, + 0.01032322458922863, + 0.01132351253181696, + 0.003940966445952654, + 0.00289552821777761, + 0.026170779019594193, + 0.012055770494043827, + 0.006240165792405605, + 0.007449483964592218, + 0.016750479117035866, + 0.00745069095864892, + 0.008574903942644596, + 0.006218777969479561, + 0.00671111885458231, + 0.002512506674975157, + 0.0011091650230810046, + 0.0018473782110959291, + 0.005343499127775431, + 0.0013730924110859632, + 0.0019162773387506604, + 0.0012693817261606455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.284603476524353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309491440653801 + ], + "opposite_logprob_tokens": [ + -3.1680948734283447 + ], + "logprob_reward": -0.04309491440653801, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309491440653801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043300408869981766, + 0.4459903836250305, + 0.03727179393172264, + 0.04852835834026337, + 0.26958975195884705, + 0.020606694743037224, + 0.03806501254439354, + 0.03224294260144234, + 0.04004726931452751, + 0.024357357993721962 + ], + "credit_map": [ + 0.004160561133176088, + 0.005206592381000519, + 0.005656823981553316, + 0.006044087465852499, + 0.008365542627871037, + 0.005646202713251114, + 0.004644911270588636, + 0.003575689159333706, + 0.06134594604372978, + 0.03656376153230667, + 0.037066634744405746, + 0.0372331403195858, + 0.07484398782253265, + 0.05166906490921974, + 0.08093392103910446, + 0.06633393466472626, + 0.003217230783775449, + 0.004341728985309601, + 0.00306932395324111, + 0.004138745367527008, + 0.005739007145166397, + 0.006742921192198992, + 0.0035638182889670134, + 0.006459018215537071, + 0.005632870830595493, + 0.005802905652672052, + 0.004147917963564396, + 0.004682506900280714, + 0.00797143392264843, + 0.005473813973367214, + 0.008954481221735477, + 0.0058624292723834515, + 0.02104438655078411, + 0.08795047551393509, + 0.00821992289274931, + 0.006860274355858564, + 0.10914386808872223, + 0.020327169448137283, + 0.008968050591647625, + 0.007075612898916006, + 0.0014687846414744854, + 0.003168694209307432, + 0.002127053216099739, + 0.001786421169526875, + 0.005561525002121925, + 0.002252897247672081, + 0.002256163861602545, + 0.001985155278816819, + 0.0022805333137512207, + 0.004397079814225435, + 0.0020144495647400618, + 0.005104613956063986, + 0.003335952991619706, + 0.009866772219538689, + 0.008901070803403854, + 0.0021645412780344486, + 0.012215168215334415, + 0.0022537419572472572, + 0.0018515388946980238, + 0.0015131470281630754, + 0.004979090765118599, + 0.006734439637511969, + 0.0012125305365771055, + 0.001483284984715283, + 0.01223821472376585, + 0.00374111020937562, + 0.0029118137899786234, + 0.0035959321539849043, + 0.007449353113770485, + 0.003743010340258479, + 0.003452239092439413, + 0.002915597753599286, + 0.008034221827983856, + 0.0025286320596933365, + 0.001726249698549509, + 0.0013956172624602914, + 0.006094538141041994, + 0.0016575921326875687, + 0.0016601427923887968, + 0.0012603631475940347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.04309491440653801 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076275295577943325 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.0076275295577943325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076275295577943325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048822443932294846, + 0.21168434619903564, + 0.05470087379217148, + 0.024895085021853447, + 0.3441845774650574, + 0.06168302521109581, + 0.05072326585650444, + 0.08265130966901779, + 0.11260979622602463, + 0.008045230060815811 + ], + "credit_map": [ + 0.0043907733634114265, + 0.004249069839715958, + 0.006556967739015818, + 0.008678902871906757, + 0.006766629405319691, + 0.006371716968715191, + 0.006163333542644978, + 0.0056450520642101765, + 0.02371191792190075, + 0.01587044447660446, + 0.018647868186235428, + 0.021280068904161453, + 0.034772541373968124, + 0.022919028997421265, + 0.04234400764107704, + 0.032138459384441376, + 0.004470976535230875, + 0.0051460471004247665, + 0.0041641672141849995, + 0.004865686409175396, + 0.01533594075590372, + 0.012012140825390816, + 0.004479506518691778, + 0.004226413089782, + 0.002625472377985716, + 0.002513126004487276, + 0.002158508636057377, + 0.0023482718970626593, + 0.00433588121086359, + 0.003050072118639946, + 0.00408979831263423, + 0.003773954464122653, + 0.027892565354704857, + 0.10933411121368408, + 0.010952702723443508, + 0.008076608180999756, + 0.14002138376235962, + 0.02719816379249096, + 0.01133382972329855, + 0.00937521830201149, + 0.005812981631606817, + 0.007871153764426708, + 0.0064880624413490295, + 0.006686321925371885, + 0.011464065872132778, + 0.007776312530040741, + 0.008908655494451523, + 0.006675466429442167, + 0.003351782914251089, + 0.0036526864860206842, + 0.00207721465267241, + 0.004102532751858234, + 0.004294870421290398, + 0.018973492085933685, + 0.011974717490375042, + 0.0022959690541028976, + 0.02814509905874729, + 0.006753832101821899, + 0.004643990192562342, + 0.006911365780979395, + 0.013328395783901215, + 0.014686192385852337, + 0.0043748305179178715, + 0.0038076024502515793, + 0.03607996180653572, + 0.010470258072018623, + 0.0069259763695299625, + 0.008976420387625694, + 0.024113059043884277, + 0.009356998838484287, + 0.009274780750274658, + 0.007412335369735956, + 0.0018468716880306602, + 0.0009016856201924384, + 0.0005340989446267486, + 0.000999781652353704, + 0.0015625301748514175, + 0.0006108500529080629, + 0.0009682095260359347, + 0.0006212024600245059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.0076275295577943325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072383262217045 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072383262217045, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072383262217045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040050312876701355, + 0.1846921592950821, + 0.033403873443603516, + 0.04427475854754448, + 0.24559152126312256, + 0.02651398442685604, + 0.05583295226097107, + 0.2601839005947113, + 0.07812231034040451, + 0.03133419156074524 + ], + "credit_map": [ + 0.004509522579610348, + 0.003593866480514407, + 0.006346615497022867, + 0.006481475196778774, + 0.004295079968869686, + 0.005018250085413456, + 0.0037042503245174885, + 0.006101255305111408, + 0.02450443245470524, + 0.01443956233561039, + 0.0163254514336586, + 0.015844665467739105, + 0.028710119426250458, + 0.02042742259800434, + 0.037019744515419006, + 0.027420774102211, + 0.0012167644454166293, + 0.0021688558626919985, + 0.0017998027615249157, + 0.0018557996954768896, + 0.012619784101843834, + 0.005700171459466219, + 0.0027593194972723722, + 0.005283378064632416, + 0.00644092308357358, + 0.006056474056094885, + 0.004136775154620409, + 0.004862906876951456, + 0.006966954097151756, + 0.004730550572276115, + 0.00637547904625535, + 0.004704698920249939, + 0.01940636709332466, + 0.07855260372161865, + 0.007629740051925182, + 0.006066750269383192, + 0.09936217963695526, + 0.019695211201906204, + 0.00838171225041151, + 0.006496991030871868, + 0.0022312712389975786, + 0.003994032274931669, + 0.002295534359291196, + 0.0034287848975509405, + 0.006128593813627958, + 0.0029096663929522038, + 0.003477900754660368, + 0.0020482027903199196, + 0.0032868674024939537, + 0.005445078480988741, + 0.001967713236808777, + 0.004690253641456366, + 0.004631234798580408, + 0.019293369725346565, + 0.014613138511776924, + 0.0019053065916523337, + 0.08252976834774017, + 0.019114965572953224, + 0.012129065580666065, + 0.051383793354034424, + 0.029984863474965096, + 0.03677571192383766, + 0.019613800570368767, + 0.00865193922072649, + 0.02366538718342781, + 0.008599200285971165, + 0.005648538004606962, + 0.006981182377785444, + 0.013661480508744717, + 0.006807347759604454, + 0.007041217293590307, + 0.0057179587893188, + 0.010591511614620686, + 0.002493803622201085, + 0.0020354411099106073, + 0.0021144652273505926, + 0.008347180671989918, + 0.0022362417075783014, + 0.0020781743805855513, + 0.001437375321984291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.011072383262217045 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037125885486603 + ], + "opposite_logprob_tokens": [ + -1.9103713035583496 + ], + "logprob_reward": -0.16037125885486603, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037125885486603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06431495398283005, + 0.20984970033168793, + 0.03231764957308769, + 0.025530284270644188, + 0.3503707945346832, + 0.0725402683019638, + 0.030074959620833397, + 0.16789808869361877, + 0.03039112687110901, + 0.01671212539076805 + ], + "credit_map": [ + 0.007047089282423258, + 0.0046116444282233715, + 0.009957492351531982, + 0.009509542025625706, + 0.0053224521689116955, + 0.00887263286858797, + 0.00639233598485589, + 0.012601765803992748, + 0.027759995311498642, + 0.016749652102589607, + 0.017168067395687103, + 0.01862374320626259, + 0.035086680203676224, + 0.02361910790205002, + 0.03997673839330673, + 0.030865687876939774, + 0.001934482017531991, + 0.0017741761403158307, + 0.001640629954636097, + 0.0017289023380726576, + 0.0074600037187337875, + 0.010355341248214245, + 0.0021974658593535423, + 0.005226647015661001, + 0.003447585739195347, + 0.002596798585727811, + 0.002679517725482583, + 0.0034976969473063946, + 0.004083534702658653, + 0.00322741293348372, + 0.003256148425862193, + 0.002741588745266199, + 0.027148809283971786, + 0.11091908812522888, + 0.0114608658477664, + 0.009292819537222385, + 0.1415632963180542, + 0.02759741246700287, + 0.013021656312048435, + 0.009366837330162525, + 0.006894090212881565, + 0.011444094590842724, + 0.006843289826065302, + 0.006686835084110498, + 0.01649276353418827, + 0.00938460323959589, + 0.0076856124214828014, + 0.007108980789780617, + 0.0038605688605457544, + 0.003362560411915183, + 0.002085167681798339, + 0.0029609024059027433, + 0.0037057045847177505, + 0.00759554049000144, + 0.004413892049342394, + 0.002090621739625931, + 0.05377760902047157, + 0.012327684089541435, + 0.009063972160220146, + 0.03627067059278488, + 0.024673955515027046, + 0.014609645120799541, + 0.012445323169231415, + 0.0047292266972362995, + 0.005024840123951435, + 0.00406148936599493, + 0.004164861980825663, + 0.003529838053509593, + 0.0045306007377803326, + 0.002876415615901351, + 0.0029868094716221094, + 0.0032162698917090893, + 0.0038995591457933187, + 0.0023138842079788446, + 0.001212884671986103, + 0.0017230055527761579, + 0.004238949157297611, + 0.0011682737385854125, + 0.0011302726343274117, + 0.0010252975625917315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.16037125885486603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion close-up, captured from a slightly upward angle, captures the mythical creature, the \"Brooktail,\" in a moment of solemn resoluteness. Its scales, shimmering with a cyan-blue metallic sheen, appear faintly amidst a soft, mysterious halo. With the sharp tips of its fins, it slowly but surely traces the bases of its wings, composed of translucent, glass-like feathers. With the tearing motion, each feather begins to disintegrate the moment it is released from its body. Through meticulous particle simulation, these fragments transform into countless golden powders shimmering with a warm glow. Like a shower of miniature, luminous stardust, they descend with a light and sacred grace, ultimately landing unerringly in a pristine brown pottery jar reverently held by a pair of calloused, weathered villagers. The camera quickly cuts to a close-up of the interior of the medicine jar. The viscous dark liquid in the jar begins to boil violently as soon as the golden powder is incorporated. What is tumbling around are not ordinary bubbles, but a texture that flows brilliantly like liquid metal. All the golden light quickly converges, rotates, and condenses towards the center of the jar, eventually forming a perfectly round golden pill that exudes a soft yet powerful breath of life. It quietly floats above the still churning liquid surface, and the light from within illuminates the rough inner wall of the pottery jar, creating a magical, hopeful atmosphere of alchemical transformation.", + "seed": 32848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_5404_seed_32848_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07514911834150553, + "vlm_logprob": -0.07514911834150553, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.07514911834150553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2051118165254593 + ], + "opposite_logprob_tokens": [ + -1.7051118612289429 + ], + "logprob_reward": -0.2051118165254593, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2051118165254593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1363169550895691, + 0.20210102200508118, + 0.09948774427175522, + 0.10212685167789459, + 0.12561938166618347, + 0.025090375915169716, + 0.030682936310768127, + 0.06925095617771149, + 0.14884363114833832, + 0.060480158776044846 + ], + "credit_map": [ + 0.024797985330224037, + 0.012101612985134125, + 0.018957367166876793, + 0.020861638709902763, + 0.00895676389336586, + 0.02380266599357128, + 0.008034509606659412, + 0.018804408609867096, + 0.02445409446954727, + 0.047290582209825516, + 0.0267234668135643, + 0.01969490572810173, + 0.02545923739671707, + 0.024587534368038177, + 0.019781678915023804, + 0.014109513722360134, + 0.016856001690030098, + 0.014595010317862034, + 0.011562900617718697, + 0.006692146882414818, + 0.014434496872127056, + 0.012671212665736675, + 0.016203805804252625, + 0.006472171749919653, + 0.014688625931739807, + 0.011826367117464542, + 0.016905324533581734, + 0.012516140937805176, + 0.011507843621075153, + 0.015818113461136818, + 0.010940519161522388, + 0.007923915050923824, + 0.016067011281847954, + 0.009711110033094883, + 0.015009233728051186, + 0.011190938763320446, + 0.02746370993554592, + 0.021061185747385025, + 0.012684834189713001, + 0.012431345880031586, + 0.003118169726803899, + 0.002846511546522379, + 0.0031589565332978964, + 0.004366864915937185, + 0.0024910548236221075, + 0.0031431084498763084, + 0.0033171814866364002, + 0.0026485281996428967, + 0.0034359670244157314, + 0.005231785587966442, + 0.0038636408280581236, + 0.0038848314434289932, + 0.005384073592722416, + 0.0029499134980142117, + 0.0033225107472389936, + 0.0026102145202457905, + 0.006705134641379118, + 0.013487493619322777, + 0.00575889041647315, + 0.005246118176728487, + 0.01933661662042141, + 0.006093338597565889, + 0.005566457752138376, + 0.00705690635368228, + 0.051446884870529175, + 0.007454417645931244, + 0.00766841322183609, + 0.003768292488530278, + 0.010146425105631351, + 0.055364008992910385, + 0.004606000613421202, + 0.008389182388782501, + 0.017757512629032135, + 0.003099964465945959, + 0.004194668028503656, + 0.0031344906892627478, + 0.003006864571943879, + 0.019222434610128403, + 0.004491889849305153, + 0.0055723343975842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.2051118165254593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038133103400468826 + ], + "opposite_logprob_tokens": [ + -3.288133144378662 + ], + "logprob_reward": -0.038133103400468826, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038133103400468826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10353045165538788, + 0.10806183516979218, + 0.10152946412563324, + 0.11999239027500153, + 0.23349034786224365, + 0.04523388668894768, + 0.029830327257514, + 0.04802028462290764, + 0.15738129615783691, + 0.05292965844273567 + ], + "credit_map": [ + 0.011235617101192474, + 0.00983584113419056, + 0.014130367897450924, + 0.016590213403105736, + 0.009256336838006973, + 0.01983433961868286, + 0.008717252872884274, + 0.013930505141615868, + 0.014007002115249634, + 0.021494485437870026, + 0.015589997172355652, + 0.011692184954881668, + 0.013542802073061466, + 0.011667399667203426, + 0.011620752513408661, + 0.008447223342955112, + 0.016610531136393547, + 0.01212639082223177, + 0.015966173261404037, + 0.00874848011881113, + 0.013813702389597893, + 0.013153101317584515, + 0.012172064743936062, + 0.008939031511545181, + 0.016568263992667198, + 0.019164467230439186, + 0.007116945926100016, + 0.009291311725974083, + 0.024286529049277306, + 0.020594386383891106, + 0.01474559772759676, + 0.00822490081191063, + 0.023481011390686035, + 0.015881747007369995, + 0.016800934448838234, + 0.025885367766022682, + 0.026576900854706764, + 0.0796874538064003, + 0.01940438523888588, + 0.0257725827395916, + 0.006057020742446184, + 0.004922398366034031, + 0.005709902383387089, + 0.007198392879217863, + 0.00712937256321311, + 0.004308158066123724, + 0.004308004397898912, + 0.005600642412900925, + 0.0025244373828172684, + 0.005654874257743359, + 0.003449413925409317, + 0.005012251902371645, + 0.004885212052613497, + 0.0024182803463190794, + 0.0028281814884394407, + 0.0030576824210584164, + 0.004507892299443483, + 0.01164977066218853, + 0.006069681141525507, + 0.002272562123835087, + 0.014547222293913364, + 0.0034947190433740616, + 0.002548195654526353, + 0.002930251881480217, + 0.05449340492486954, + 0.010468896478414536, + 0.00876598246395588, + 0.0036814736668020487, + 0.010841261595487595, + 0.055516015738248825, + 0.0042461189441382885, + 0.009368160739541054, + 0.01707024686038494, + 0.003659839043393731, + 0.0032997336238622665, + 0.002053305972367525, + 0.002877619815990329, + 0.01648068055510521, + 0.003786893794313073, + 0.003701347392052412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.038133103400468826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784595109522343 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784595109522343, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784595109522343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09018077701330185, + 0.08802162110805511, + 0.06065235286951065, + 0.05668817460536957, + 0.0990154817700386, + 0.046702172607183456, + 0.01972329616546631, + 0.055799130350351334, + 0.4544689357280731, + 0.028748104348778725 + ], + "credit_map": [ + 0.020007556304335594, + 0.007387280929833651, + 0.010667997412383556, + 0.012130091898143291, + 0.005608675070106983, + 0.011000079102814198, + 0.006525952368974686, + 0.016853133216500282, + 0.014188934117555618, + 0.012149290181696415, + 0.01741231046617031, + 0.009863080456852913, + 0.009366530925035477, + 0.010091213509440422, + 0.007047941442579031, + 0.007902313023805618, + 0.006645109970122576, + 0.007773894816637039, + 0.010707274079322815, + 0.007080338429659605, + 0.008211872540414333, + 0.009101293049752712, + 0.00640504714101553, + 0.004727513529360294, + 0.006810104474425316, + 0.00598512776196003, + 0.0059647392481565475, + 0.008693553507328033, + 0.008771117776632309, + 0.005935202352702618, + 0.009442675858736038, + 0.005085651762783527, + 0.011964486911892891, + 0.006782046519219875, + 0.011629643850028515, + 0.017800716683268547, + 0.019112365320324898, + 0.009500490501523018, + 0.0121196573600173, + 0.010106069035828114, + 0.011220192536711693, + 0.0050403885543346405, + 0.006882542744278908, + 0.005506153684109449, + 0.005380235146731138, + 0.004025178961455822, + 0.003820146666839719, + 0.00482732942327857, + 0.0026932007167488337, + 0.0032783187925815582, + 0.002817284082993865, + 0.002766144461929798, + 0.002897869795560837, + 0.0018468754133209586, + 0.0018161303596571088, + 0.0016074697487056255, + 0.004789391066879034, + 0.01282412651926279, + 0.004198580514639616, + 0.0037009587977081537, + 0.016764571890234947, + 0.00465296721085906, + 0.004102581180632114, + 0.004765948746353388, + 0.15668421983718872, + 0.027523793280124664, + 0.022716544568538666, + 0.008996439166367054, + 0.028774021193385124, + 0.16948220133781433, + 0.012837150134146214, + 0.02745453082025051, + 0.007944365963339806, + 0.0019288271432742476, + 0.0025441923644393682, + 0.0022033206187188625, + 0.0016628430457785726, + 0.007358136586844921, + 0.001960140187293291, + 0.0031462768092751503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.009784595109522343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14306223392486572, + 0.11689963191747665, + 0.10857412964105606, + 0.1403256207704544, + 0.16066665947437286, + 0.06998913735151291, + 0.03689851239323616, + 0.026052162051200867, + 0.15966789424419403, + 0.03786398097872734 + ], + "credit_map": [ + 0.019709451124072075, + 0.013758675195276737, + 0.01644592359662056, + 0.01946260593831539, + 0.01338664535433054, + 0.02268269658088684, + 0.012459982186555862, + 0.025156259536743164, + 0.01882900483906269, + 0.02105068415403366, + 0.013045676983892918, + 0.012178618460893631, + 0.016254203394055367, + 0.015824612230062485, + 0.010735711082816124, + 0.0089811310172081, + 0.019403871148824692, + 0.007811297662556171, + 0.026942169293761253, + 0.011528718285262585, + 0.007981848903000355, + 0.015061972662806511, + 0.010682267136871815, + 0.009161989204585552, + 0.023228641599416733, + 0.015809519216418266, + 0.019366050139069557, + 0.014726701192557812, + 0.025694966316223145, + 0.016742700710892677, + 0.012895883060991764, + 0.011861169710755348, + 0.01874377764761448, + 0.010389568284153938, + 0.010674449615180492, + 0.01691138558089733, + 0.01582229696214199, + 0.05449496582150459, + 0.014146710745990276, + 0.019483506679534912, + 0.007254349999129772, + 0.006132305134087801, + 0.007915624417364597, + 0.014121134765446186, + 0.009672490879893303, + 0.005390730686485767, + 0.012413921765983105, + 0.0070885783061385155, + 0.002952846232801676, + 0.0051361555233597755, + 0.00564128952100873, + 0.004732134751975536, + 0.00610155938193202, + 0.0023895849008113146, + 0.006445983424782753, + 0.003498961916193366, + 0.002151430118829012, + 0.005064093973487616, + 0.0034542963840067387, + 0.002599353203549981, + 0.005139388144016266, + 0.0023834125604480505, + 0.0027901476714760065, + 0.002470041159540415, + 0.024989420548081398, + 0.013196329586207867, + 0.01707375794649124, + 0.023257311433553696, + 0.012042627669870853, + 0.0323563814163208, + 0.0200837142765522, + 0.01666836068034172, + 0.004525266587734222, + 0.0027195438742637634, + 0.0032394793815910816, + 0.005437721032649279, + 0.0032457285560667515, + 0.005465340800583363, + 0.007309249602258205, + 0.005921653006225824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293596774339676 + ], + "opposite_logprob_tokens": [ + -2.237936019897461 + ], + "logprob_reward": -0.11293596774339676, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293596774339676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08741723001003265, + 0.10609222203493118, + 0.02755391038954258, + 0.04343791306018829, + 0.35604870319366455, + 0.10005255043506622, + 0.04468254745006561, + 0.045672833919525146, + 0.14254504442214966, + 0.04649708420038223 + ], + "credit_map": [ + 0.014979616738855839, + 0.007110013626515865, + 0.013687541708350182, + 0.010873076505959034, + 0.005812003277242184, + 0.019297996535897255, + 0.0062538632191717625, + 0.009403113275766373, + 0.014642390422523022, + 0.01916838064789772, + 0.01095390971750021, + 0.01095658354461193, + 0.016453737393021584, + 0.014809966087341309, + 0.009476386941969395, + 0.009630875661969185, + 0.003975364845246077, + 0.0032524142879992723, + 0.003484893124550581, + 0.00289923301897943, + 0.004305108916014433, + 0.0037072282284498215, + 0.0032539451494812965, + 0.002675725379958749, + 0.005398506298661232, + 0.007013955619186163, + 0.003928136080503464, + 0.004911736585199833, + 0.007898727431893349, + 0.004589117597788572, + 0.005199609789997339, + 0.004498125519603491, + 0.03770983964204788, + 0.020531609654426575, + 0.0263051837682724, + 0.026899801567196846, + 0.049226123839616776, + 0.11946525424718857, + 0.03491661325097084, + 0.04099425673484802, + 0.010728421621024609, + 0.013015835545957088, + 0.014713156037032604, + 0.015803512185811996, + 0.011075037531554699, + 0.01310665812343359, + 0.00982788298279047, + 0.011782047338783741, + 0.005183435045182705, + 0.0072261495515704155, + 0.0052256276831030846, + 0.0058573042042553425, + 0.007181013934314251, + 0.004298223182559013, + 0.005274166353046894, + 0.0044366284273564816, + 0.004153169225901365, + 0.010287771001458168, + 0.006654832977801561, + 0.003526291809976101, + 0.009254288859665394, + 0.0038962229155004025, + 0.0038887420669198036, + 0.004011513665318489, + 0.017081160098314285, + 0.025964070111513138, + 0.01701578125357628, + 0.015584397129714489, + 0.013759789057075977, + 0.019760362803936005, + 0.014853015542030334, + 0.018526459112763405, + 0.01245712861418724, + 0.005281055811792612, + 0.00342015759088099, + 0.0028309221379458904, + 0.003891276428475976, + 0.011246517300605774, + 0.003088585799559951, + 0.004281442612409592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.11293596774339676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion close-up, captured from a slightly upward angle, captures the mythical creature, the \"Brooktail,\" in a moment of solemn resoluteness. Its scales, shimmering with a cyan-blue metallic sheen, appear faintly amidst a soft, mysterious halo. With the sharp tips of its fins, it slowly but surely traces the bases of its wings, composed of translucent, glass-like feathers. With the tearing motion, each feather begins to disintegrate the moment it is released from its body. Through meticulous particle simulation, these fragments transform into countless golden powders shimmering with a warm glow. Like a shower of miniature, luminous stardust, they descend with a light and sacred grace, ultimately landing unerringly in a pristine brown pottery jar reverently held by a pair of calloused, weathered villagers. The camera quickly cuts to a close-up of the interior of the medicine jar. The viscous dark liquid in the jar begins to boil violently as soon as the golden powder is incorporated. What is tumbling around are not ordinary bubbles, but a texture that flows brilliantly like liquid metal. All the golden light quickly converges, rotates, and condenses towards the center of the jar, eventually forming a perfectly round golden pill that exudes a soft yet powerful breath of life. It quietly floats above the still churning liquid surface, and the light from within illuminates the rough inner wall of the pottery jar, creating a magical, hopeful atmosphere of alchemical transformation.", + "seed": 32849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_5404_seed_32849_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08719030851498247, + "vlm_logprob": -0.08719030851498247, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.08719030851498247, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14494773745536804 + ], + "opposite_logprob_tokens": [ + -2.0199477672576904 + ], + "logprob_reward": -0.14494773745536804, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14494773745536804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16042809188365936, + 0.05653570964932442, + 0.030603667721152306, + 0.050637099891901016, + 0.05408931523561478, + 0.0674363449215889, + 0.18115222454071045, + 0.08843465149402618, + 0.24526970088481903, + 0.06541313976049423 + ], + "credit_map": [ + 0.01273068692535162, + 0.013984549790620804, + 0.01496333722025156, + 0.029782649129629135, + 0.025448383763432503, + 0.024473095312714577, + 0.013564973138272762, + 0.025480415672063828, + 0.006737256422638893, + 0.005786850582808256, + 0.006028671283274889, + 0.009526892565190792, + 0.005431735888123512, + 0.006930038332939148, + 0.007614564150571823, + 0.008479701355099678, + 0.003820535959675908, + 0.0036138782743364573, + 0.004615556448698044, + 0.004411439877003431, + 0.0033815347123891115, + 0.003473005024716258, + 0.004321124404668808, + 0.0029665955808013678, + 0.0062729837372899055, + 0.003476908430457115, + 0.003880737815052271, + 0.005840889643877745, + 0.007060332223773003, + 0.007388183381408453, + 0.010683644562959671, + 0.006033421028405428, + 0.004002742003649473, + 0.004957377910614014, + 0.0055282581597566605, + 0.01044756080955267, + 0.00950440764427185, + 0.005124748684465885, + 0.007041816134005785, + 0.007482409011572599, + 0.0054657431319355965, + 0.016602667048573494, + 0.004755394533276558, + 0.008544995449483395, + 0.011610414832830429, + 0.00741319265216589, + 0.00813327170908451, + 0.004910667892545462, + 0.020864112302660942, + 0.017687486484646797, + 0.01181873306632042, + 0.020711690187454224, + 0.023682309314608574, + 0.053564514964818954, + 0.017511583864688873, + 0.015311812050640583, + 0.009398488327860832, + 0.010592146776616573, + 0.008851727470755577, + 0.014956883154809475, + 0.012489261105656624, + 0.009389064274728298, + 0.011324268765747547, + 0.011432825587689877, + 0.06411031633615494, + 0.013392961584031582, + 0.013396842405200005, + 0.01732533611357212, + 0.019033849239349365, + 0.08812379837036133, + 0.012170289643108845, + 0.017716309055685997, + 0.01328366156667471, + 0.007176815997809172, + 0.008388583548367023, + 0.0157033558934927, + 0.004762906115502119, + 0.005017673131078482, + 0.0068818144500255585, + 0.004198328126221895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.14494773745536804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209735572338104 + ], + "opposite_logprob_tokens": [ + -2.8120973110198975 + ], + "logprob_reward": -0.06209735572338104, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209735572338104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14205509424209595, + 0.15720410645008087, + 0.04535767808556557, + 0.08707310259342194, + 0.07891570776700974, + 0.0526777058839798, + 0.11423321813344955, + 0.1131351962685585, + 0.1600416749715805, + 0.04930652678012848 + ], + "credit_map": [ + 0.012084721587598324, + 0.011593698523938656, + 0.015296339057385921, + 0.024803627282381058, + 0.021536916494369507, + 0.02033027820289135, + 0.013681932352483273, + 0.022727573290467262, + 0.013950448483228683, + 0.01518265437334776, + 0.01631033420562744, + 0.02262759953737259, + 0.02413124591112137, + 0.018002985045313835, + 0.02470654435455799, + 0.022292284294962883, + 0.007231724448502064, + 0.005304904654622078, + 0.0075279478915035725, + 0.0037062871269881725, + 0.005403715651482344, + 0.009167625568807125, + 0.003066086443141103, + 0.003949381411075592, + 0.012211102060973644, + 0.005062222480773926, + 0.00855683721601963, + 0.013409498147666454, + 0.007482938468456268, + 0.007117421366274357, + 0.021870261058211327, + 0.011362815275788307, + 0.007140481844544411, + 0.006585976574569941, + 0.009017013013362885, + 0.014418303966522217, + 0.013363728299736977, + 0.006479647941887379, + 0.011200116947293282, + 0.010710434056818485, + 0.003946101292967796, + 0.011136021465063095, + 0.0038733375258743763, + 0.00699771149083972, + 0.009280860424041748, + 0.00545919593423605, + 0.007356217596679926, + 0.00462825782597065, + 0.018372422084212303, + 0.011777828447520733, + 0.007229491602629423, + 0.008458795957267284, + 0.015465336851775646, + 0.03510979562997818, + 0.009232096374034882, + 0.008587444201111794, + 0.01151489932090044, + 0.011617419309914112, + 0.008264604024589062, + 0.020488198846578598, + 0.015239288099110126, + 0.02241431549191475, + 0.014958685263991356, + 0.008637785911560059, + 0.039251819252967834, + 0.011981368064880371, + 0.012554099783301353, + 0.01531167607754469, + 0.01252342201769352, + 0.047244843095541, + 0.01226722914725542, + 0.00890721008181572, + 0.00790777150541544, + 0.004404863808304071, + 0.005532954819500446, + 0.009747326374053955, + 0.003877615788951516, + 0.008971028961241245, + 0.004544624127447605, + 0.004320343025028706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.06209735572338104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067374915815889835 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.0067374915815889835, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067374915815889835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07499851286411285, + 0.09163061529397964, + 0.026729701086878777, + 0.11098910123109818, + 0.0991758182644844, + 0.04270879924297333, + 0.18647976219654083, + 0.10313450545072556, + 0.22492781281471252, + 0.03922535851597786 + ], + "credit_map": [ + 0.006691998336464167, + 0.005551675334572792, + 0.006435123272240162, + 0.008013746701180935, + 0.009230823256075382, + 0.011560806073248386, + 0.005717013496905565, + 0.02179732546210289, + 0.016023803502321243, + 0.007347472012042999, + 0.007273316849023104, + 0.012245423160493374, + 0.011178987100720406, + 0.011250638402998447, + 0.01157113816589117, + 0.014739842154085636, + 0.003477800404652953, + 0.0027917998377233744, + 0.003585848491638899, + 0.0037261375691741705, + 0.0030500018037855625, + 0.003269974607974291, + 0.003948322031646967, + 0.002879818668588996, + 0.017389493063092232, + 0.007062870543450117, + 0.012156689539551735, + 0.014566682279109955, + 0.009350446052849293, + 0.01188532542437315, + 0.02053600549697876, + 0.018041599541902542, + 0.011059587821364403, + 0.009824487380683422, + 0.014115405268967152, + 0.014416174031794071, + 0.013441557995975018, + 0.011525736190378666, + 0.011862587183713913, + 0.012930289842188358, + 0.002997445175424218, + 0.007368454709649086, + 0.0053712548688054085, + 0.00695304200053215, + 0.007558635901659727, + 0.004526984877884388, + 0.004045083187520504, + 0.0038879034109413624, + 0.017490630969405174, + 0.01848720945417881, + 0.011848488822579384, + 0.024036413058638573, + 0.023619908839464188, + 0.048925016075372696, + 0.02514154464006424, + 0.016930557787418365, + 0.011331108398735523, + 0.012263921089470387, + 0.008162591606378555, + 0.020144931972026825, + 0.01476766262203455, + 0.013792119920253754, + 0.013195420615375042, + 0.009476758539676666, + 0.05260017514228821, + 0.014437559060752392, + 0.01380744855850935, + 0.02201787754893303, + 0.019336393103003502, + 0.07063574343919754, + 0.017981844022870064, + 0.014110797084867954, + 0.005176611244678497, + 0.0038633637595921755, + 0.003975966013967991, + 0.008639478124678135, + 0.003076198510825634, + 0.005017242860049009, + 0.005105569493025541, + 0.004370932932943106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.0067374915815889835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058078907430172 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.02058078907430172, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058078907430172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06254709511995316, + 0.11167090386152267, + 0.05761313810944557, + 0.12417212873697281, + 0.11403403431177139, + 0.04142434522509575, + 0.11859007179737091, + 0.10198678076267242, + 0.19370107352733612, + 0.07426043599843979 + ], + "credit_map": [ + 0.006203947123140097, + 0.004711620043963194, + 0.006415657699108124, + 0.006275627296417952, + 0.015209158882498741, + 0.007295530289411545, + 0.006655402481555939, + 0.009780148975551128, + 0.013862465508282185, + 0.012898405082523823, + 0.020284857600927353, + 0.01719362661242485, + 0.012505119666457176, + 0.011829194612801075, + 0.011472705751657486, + 0.011624529026448727, + 0.009318927302956581, + 0.0070715732872486115, + 0.009236221201717854, + 0.005276618991047144, + 0.006861853413283825, + 0.009305000305175781, + 0.0053040687926113605, + 0.0052388738840818405, + 0.01886236108839512, + 0.014608466997742653, + 0.015707386657595634, + 0.015491263940930367, + 0.01546940952539444, + 0.010712560266256332, + 0.01744181290268898, + 0.015878863632678986, + 0.00945407897233963, + 0.012215252965688705, + 0.013938376680016518, + 0.015347743406891823, + 0.022156916558742523, + 0.012822415679693222, + 0.013332389295101166, + 0.014766860753297806, + 0.0025080281775444746, + 0.007125277072191238, + 0.003424755297601223, + 0.008204751648008823, + 0.006641932297497988, + 0.003817792749032378, + 0.006960321217775345, + 0.0027414876967668533, + 0.011066167615354061, + 0.011707707308232784, + 0.0070103248581290245, + 0.016782034188508987, + 0.014721940271556377, + 0.03390410542488098, + 0.013210142031311989, + 0.010187639854848385, + 0.014723744243383408, + 0.010698072612285614, + 0.007921033538877964, + 0.013032650575041771, + 0.01310973521322012, + 0.02311575971543789, + 0.009953976608812809, + 0.00943180825561285, + 0.03927898406982422, + 0.016732392832636833, + 0.00991600938141346, + 0.02541148103773594, + 0.016824232414364815, + 0.05945931747555733, + 0.013568992726504803, + 0.012509661726653576, + 0.011091513559222221, + 0.00418531009927392, + 0.005548376590013504, + 0.006992895156145096, + 0.007532653398811817, + 0.022216681391000748, + 0.006349034607410431, + 0.01034396979957819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.02058078907430172 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158816874027252 + ], + "opposite_logprob_tokens": [ + -1.7015881538391113 + ], + "logprob_reward": -0.20158816874027252, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158816874027252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05616195499897003, + 0.029041415080428123, + 0.034710776060819626, + 0.07302180677652359, + 0.09097052365541458, + 0.07002207636833191, + 0.18213844299316406, + 0.16325564682483673, + 0.2573481500148773, + 0.0433291494846344 + ], + "credit_map": [ + 0.004738898482173681, + 0.003689851611852646, + 0.005406682379543781, + 0.008183793164789677, + 0.012102152220904827, + 0.007168412674218416, + 0.006226230878382921, + 0.008645928464829922, + 0.004097234923392534, + 0.002999056363478303, + 0.0034580142237246037, + 0.0035328571684658527, + 0.0036555782426148653, + 0.003302294062450528, + 0.003951424267143011, + 0.0040449537336826324, + 0.003936489578336477, + 0.0032095678616315126, + 0.0043328916653990746, + 0.004979297518730164, + 0.002684083068743348, + 0.004262750968337059, + 0.005016273818910122, + 0.0062894211150705814, + 0.011947079561650753, + 0.009306338615715504, + 0.008248575031757355, + 0.009913674555718899, + 0.005527091212570667, + 0.0057189189828932285, + 0.012091554701328278, + 0.010268574580550194, + 0.00842683482915163, + 0.01018704567104578, + 0.012222287245094776, + 0.012993204407393932, + 0.014558898285031319, + 0.010456949472427368, + 0.009737432934343815, + 0.012387868016958237, + 0.005551806651055813, + 0.016637593507766724, + 0.006435486488044262, + 0.01035100407898426, + 0.010665950365364552, + 0.007423940114676952, + 0.008163523860275745, + 0.0047927736304700375, + 0.015524787828326225, + 0.016902120783925056, + 0.011531542055308819, + 0.02317170985043049, + 0.02616116590797901, + 0.05157836154103279, + 0.021230289712548256, + 0.016038481146097183, + 0.02275317907333374, + 0.01846439205110073, + 0.015443104319274426, + 0.0223897323012352, + 0.020238276571035385, + 0.03575031831860542, + 0.015038411132991314, + 0.013178219087421894, + 0.0644957423210144, + 0.015281078405678272, + 0.013348875567317009, + 0.021154433488845825, + 0.02126004733145237, + 0.08822160959243774, + 0.016444915905594826, + 0.017141476273536682, + 0.00783248245716095, + 0.0029596034437417984, + 0.0034689048770815134, + 0.004207869991660118, + 0.004012619610875845, + 0.01358330249786377, + 0.0035696893464773893, + 0.0036946777254343033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.20158816874027252 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion close-up, captured from a slightly upward angle, captures the mythical creature, the \"Brooktail,\" in a moment of solemn resoluteness. Its scales, shimmering with a cyan-blue metallic sheen, appear faintly amidst a soft, mysterious halo. With the sharp tips of its fins, it slowly but surely traces the bases of its wings, composed of translucent, glass-like feathers. With the tearing motion, each feather begins to disintegrate the moment it is released from its body. Through meticulous particle simulation, these fragments transform into countless golden powders shimmering with a warm glow. Like a shower of miniature, luminous stardust, they descend with a light and sacred grace, ultimately landing unerringly in a pristine brown pottery jar reverently held by a pair of calloused, weathered villagers. The camera quickly cuts to a close-up of the interior of the medicine jar. The viscous dark liquid in the jar begins to boil violently as soon as the golden powder is incorporated. What is tumbling around are not ordinary bubbles, but a texture that flows brilliantly like liquid metal. All the golden light quickly converges, rotates, and condenses towards the center of the jar, eventually forming a perfectly round golden pill that exudes a soft yet powerful breath of life. It quietly floats above the still churning liquid surface, and the light from within illuminates the rough inner wall of the pottery jar, creating a magical, hopeful atmosphere of alchemical transformation.", + "seed": 32850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_2/index_5404_seed_32850_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08460734086111188, + "vlm_logprob": -0.08460734086111188, + "vlm_yes_no_margin": 3.0749999284744263, + "vlm_token_credit": -0.08460734086111188, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14961084723472595 + ], + "opposite_logprob_tokens": [ + -2.024610757827759 + ], + "logprob_reward": -0.14961084723472595, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14961084723472595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10402513295412064, + 0.061234250664711, + 0.04192286357283592, + 0.10673154145479202, + 0.10528489202260971, + 0.42040738463401794, + 0.049710992723703384, + 0.027761399745941162, + 0.06371232122182846, + 0.019209209829568863 + ], + "credit_map": [ + 0.009726952761411667, + 0.01484600454568863, + 0.013827241957187653, + 0.013155090622603893, + 0.01348033919930458, + 0.014228816144168377, + 0.01334866788238287, + 0.011412025429308414, + 0.005006661172956228, + 0.00479960348457098, + 0.01219082809984684, + 0.007270868867635727, + 0.005111096892505884, + 0.012535293586552143, + 0.005841745063662529, + 0.008478149771690369, + 0.005084725096821785, + 0.0036979259457439184, + 0.006307780742645264, + 0.005648871883749962, + 0.0037523636128753424, + 0.005127216689288616, + 0.005269790533930063, + 0.007034188602119684, + 0.010327641852200031, + 0.012211312539875507, + 0.026229238137602806, + 0.006734689697623253, + 0.012227049097418785, + 0.023670684546232224, + 0.008324692025780678, + 0.0070062242448329926, + 0.04308090731501579, + 0.011316828429698944, + 0.005884458776563406, + 0.0059121171943843365, + 0.007648721802979708, + 0.005506652407348156, + 0.004006495699286461, + 0.021928712725639343, + 0.13654635846614838, + 0.05017971619963646, + 0.02096956968307495, + 0.02306280843913555, + 0.11277708411216736, + 0.02746092528104782, + 0.015049993991851807, + 0.034360963851213455, + 0.0051359436474740505, + 0.007485009729862213, + 0.0041015055030584335, + 0.008877885527908802, + 0.0039616115391254425, + 0.006507466547191143, + 0.0043153357692062855, + 0.00932623352855444, + 0.002156255068257451, + 0.0037199733778834343, + 0.002273272955790162, + 0.004446255508810282, + 0.0020354788284748793, + 0.004357683006674051, + 0.0028963855002075434, + 0.005876095499843359, + 0.007144808303564787, + 0.016181377694010735, + 0.004947950132191181, + 0.008657634258270264, + 0.006375864148139954, + 0.005944966804236174, + 0.00684153288602829, + 0.0076181828044354916, + 0.001742162392474711, + 0.00387425790540874, + 0.001578527968376875, + 0.0026330736000090837, + 0.0021790119353681803, + 0.0024243087973445654, + 0.0021035014651715755, + 0.0026743675116449594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.14961084723472595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008461654186249 + ], + "opposite_logprob_tokens": [ + -2.695084571838379 + ], + "logprob_reward": -0.07008461654186249, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008461654186249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05343450605869293, + 0.08971530199050903, + 0.07179898768663406, + 0.06318769603967667, + 0.12384943664073944, + 0.3946775794029236, + 0.060228604823350906, + 0.036155425012111664, + 0.0859190970659256, + 0.021033350378274918 + ], + "credit_map": [ + 0.004378105979412794, + 0.006327454466372728, + 0.006434175185859203, + 0.007267616223543882, + 0.007756062317639589, + 0.006773480214178562, + 0.007352802902460098, + 0.007144812494516373, + 0.006697469856590033, + 0.007343165576457977, + 0.020162785425782204, + 0.009174342267215252, + 0.007739244494587183, + 0.02124062180519104, + 0.007966212928295135, + 0.009391466155648232, + 0.007372734602540731, + 0.005556272342801094, + 0.011898234486579895, + 0.01207537017762661, + 0.006105301436036825, + 0.009135687723755836, + 0.0071655819192528725, + 0.012489816173911095, + 0.009889664128422737, + 0.010058668442070484, + 0.009514975361526012, + 0.005392623599618673, + 0.005386631935834885, + 0.012966020032763481, + 0.005184356123209, + 0.004794757813215256, + 0.037006527185440063, + 0.020016711205244064, + 0.006001160480082035, + 0.013705221936106682, + 0.008892870508134365, + 0.008001337759196758, + 0.006317066494375467, + 0.023908548057079315, + 0.1327119916677475, + 0.047201357781887054, + 0.020312434062361717, + 0.02640000730752945, + 0.09739625453948975, + 0.02403225190937519, + 0.014152618125081062, + 0.03247072920203209, + 0.005521596875041723, + 0.008817377500236034, + 0.004450594540685415, + 0.01134694553911686, + 0.004648473579436541, + 0.009229893796145916, + 0.006206422578543425, + 0.010007306933403015, + 0.0036495451349765062, + 0.004531269893050194, + 0.005114296451210976, + 0.004831879399716854, + 0.0026939241215586662, + 0.006031327415257692, + 0.0036772966850548983, + 0.0056258877739310265, + 0.007991003803908825, + 0.024358050897717476, + 0.007094673812389374, + 0.01034627202898264, + 0.007265140768140554, + 0.009505067951977253, + 0.008076200261712074, + 0.01128269825130701, + 0.0015872351359575987, + 0.004823638126254082, + 0.0019486336968839169, + 0.002480672672390938, + 0.0024180978070944548, + 0.0020850575529038906, + 0.0018170507391914725, + 0.003872967790812254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.07008461654186249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052567594684660435 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052567594684660435, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052567594684660435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05740350857377052, + 0.066753089427948, + 0.10322757065296173, + 0.2313237190246582, + 0.06977146863937378, + 0.3007938861846924, + 0.07184583693742752, + 0.021146422252058983, + 0.03911321610212326, + 0.03862130269408226 + ], + "credit_map": [ + 0.005419138818979263, + 0.006793640553951263, + 0.009135771542787552, + 0.0061789280734956264, + 0.007366123143583536, + 0.008585497736930847, + 0.006087301764637232, + 0.007837106473743916, + 0.006000669207423925, + 0.0047797998413443565, + 0.012725556269288063, + 0.008999722078442574, + 0.004684514366090298, + 0.013641223311424255, + 0.007889599539339542, + 0.00803199503570795, + 0.012315556406974792, + 0.008382917381823063, + 0.017926929518580437, + 0.01605907641351223, + 0.008962249383330345, + 0.013128465041518211, + 0.010535841807723045, + 0.01591653563082218, + 0.022563260048627853, + 0.027277734130620956, + 0.05235537141561508, + 0.01297193206846714, + 0.021642450243234634, + 0.06066739559173584, + 0.01923055574297905, + 0.014614993706345558, + 0.008940580300986767, + 0.00753811514005065, + 0.004385791718959808, + 0.005618334282189608, + 0.013547576032578945, + 0.007924262434244156, + 0.0029518422670662403, + 0.018864966928958893, + 0.1032930463552475, + 0.03398408740758896, + 0.01592451147735119, + 0.017000196501612663, + 0.07583179324865341, + 0.018956191837787628, + 0.010697416961193085, + 0.025106623768806458, + 0.005247470457106829, + 0.011315845884382725, + 0.003828717628493905, + 0.014993511140346527, + 0.009574910625815392, + 0.006258340552449226, + 0.006916648242622614, + 0.0137103833258152, + 0.002163961995393038, + 0.0027390120085328817, + 0.002923771273344755, + 0.002850343007594347, + 0.001575295813381672, + 0.003400979097932577, + 0.0023407211992889643, + 0.0031523355282843113, + 0.0038701987359672785, + 0.009863710030913353, + 0.0032038316130638123, + 0.004322070628404617, + 0.003243885235860944, + 0.005494613200426102, + 0.004102708771824837, + 0.0050121936947107315, + 0.004303475841879845, + 0.0050868503749370575, + 0.004049019422382116, + 0.0038080166559666395, + 0.0035983314737677574, + 0.008058939129114151, + 0.004661148879677057, + 0.0050555141642689705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.0052567594684660435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189646303653717 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189646303653717, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018189646303653717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02925819158554077, + 0.03141028806567192, + 0.0347474180161953, + 0.1295095533132553, + 0.09840753674507141, + 0.49843674898147583, + 0.03762831166386604, + 0.024148955941200256, + 0.03775645047426224, + 0.07869663834571838 + ], + "credit_map": [ + 0.0024584743659943342, + 0.004609831143170595, + 0.00316065875813365, + 0.004499611910432577, + 0.0033496865071356297, + 0.0034079679753631353, + 0.002782948547974229, + 0.004989007953554392, + 0.0024206663947552443, + 0.0023221455048769712, + 0.006860204041004181, + 0.0039144656620919704, + 0.002845701528713107, + 0.005936451256275177, + 0.0031906466465443373, + 0.003920004237443209, + 0.00567511236295104, + 0.0038001886568963528, + 0.005319354590028524, + 0.004538483917713165, + 0.003486770438030362, + 0.004260878078639507, + 0.0037474159616976976, + 0.003919207490980625, + 0.014495549723505974, + 0.010167542845010757, + 0.03361676633358002, + 0.006698654033243656, + 0.007810166571289301, + 0.04521302133798599, + 0.004785975906997919, + 0.006721873767673969, + 0.037047892808914185, + 0.013488228432834148, + 0.007743773981928825, + 0.0057022543624043465, + 0.016089802607893944, + 0.007239712402224541, + 0.004436006303876638, + 0.006659853737801313, + 0.163703054189682, + 0.05814234912395477, + 0.025501415133476257, + 0.026510681957006454, + 0.13383987545967102, + 0.032344426959753036, + 0.018211370334029198, + 0.04018351808190346, + 0.0030181612819433212, + 0.005474352277815342, + 0.00256868707947433, + 0.007513027638196945, + 0.004348327871412039, + 0.004250396974384785, + 0.004697511438280344, + 0.00575784407556057, + 0.0021465534809976816, + 0.0039627342484891415, + 0.002331249648705125, + 0.00435924855992198, + 0.002048342954367399, + 0.00363193079829216, + 0.002285258611664176, + 0.003383633680641651, + 0.0036105765029788017, + 0.0077510313130915165, + 0.0031589725986123085, + 0.00417078472673893, + 0.0027896601241081953, + 0.005238828714936972, + 0.0039680227637290955, + 0.007068565580993891, + 0.005154767073690891, + 0.022579215466976166, + 0.005147589836269617, + 0.0059958817437291145, + 0.010586226359009743, + 0.006451004184782505, + 0.0050314562395215034, + 0.017750486731529236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.018189646303653717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical creature named 'Brooktail' with cyan-blue metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature use its sharp fin tips to trace and disintegrate translucent, glass-like wings into golden powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slow-motion close-up from a slightly upward angle before cutting to the interior of a brown pottery jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the golden powder land in a jar held by villagers and transform into a round golden pill floating above dark, boiling liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid inside the jar boil violently with a texture resembling liquid metal before the golden pill forms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798948347568512 + ], + "opposite_logprob_tokens": [ + -1.8048948049545288 + ], + "logprob_reward": -0.1798948347568512, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798948347568512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024929508566856384, + 0.03686416521668434, + 0.0283004529774189, + 0.023282798007130623, + 0.08148559927940369, + 0.5458965301513672, + 0.17119131982326508, + 0.03383224830031395, + 0.019836749881505966, + 0.03438061103224754 + ], + "credit_map": [ + 0.0017476186621934175, + 0.0034698909148573875, + 0.0032756312284618616, + 0.0028851148672401905, + 0.0038548463489860296, + 0.003204901469871402, + 0.0033113083336502314, + 0.0031801965087652206, + 0.002116825897246599, + 0.002240662695840001, + 0.00920165330171585, + 0.0025078372564166784, + 0.0024191129487007856, + 0.01312818843871355, + 0.0019212233601137996, + 0.003328661434352398, + 0.003941019997000694, + 0.0025111623108386993, + 0.004957812372595072, + 0.0039131916128098965, + 0.0027123866602778435, + 0.003957740496844053, + 0.0024813381023705006, + 0.003825799096375704, + 0.002375137060880661, + 0.00396076962351799, + 0.0035221257712692022, + 0.001500784303061664, + 0.002898676320910454, + 0.004403494298458099, + 0.0029413551092147827, + 0.0016804533079266548, + 0.030742047354578972, + 0.010456832125782967, + 0.006697408854961395, + 0.005921799223870039, + 0.010987713932991028, + 0.007665617391467094, + 0.003000813303515315, + 0.006013370584696531, + 0.18181706964969635, + 0.06355048716068268, + 0.027882922440767288, + 0.027363717555999756, + 0.14853213727474213, + 0.03616199642419815, + 0.017882876098155975, + 0.04270531237125397, + 0.012256155721843243, + 0.02364923246204853, + 0.008722340688109398, + 0.025729479268193245, + 0.02742726355791092, + 0.016132637858390808, + 0.011682345531880856, + 0.04559186100959778, + 0.0023149861954152584, + 0.006060226354748011, + 0.003040096489712596, + 0.007126411888748407, + 0.0033086289186030626, + 0.003676059190183878, + 0.004141763783991337, + 0.004164078272879124, + 0.0024739233776926994, + 0.003894721856340766, + 0.0017405245453119278, + 0.0033900225535035133, + 0.0015802334528416395, + 0.0018982414621859789, + 0.0020827508997172117, + 0.0027763317339122295, + 0.003558998228982091, + 0.006007770542055368, + 0.0035361284390091896, + 0.005121220368891954, + 0.003782836953178048, + 0.005022196564823389, + 0.0035536300856620073, + 0.003797826822847128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.1798948347568512 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_tensor([2286, 2286])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_tensor([2286, 2286])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..52fb84069920663debea5869b885a5962f43171d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_tensor([2286, 2286])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, shot from a low ground angle, slowly pans right, capturing a fierce confrontation on the main thoroughfare of an abandoned city. On the left, a heavy main battle tank, clad in dark gray and heavily damaged, scratched, and rusted, its angular turret and tracks caked with dirt and gravel make its massive metal bulk appear particularly heavy. Opposite it, another, more streamlined tank, painted in desert camouflage, rapidly turns its turret toward it, locking onto its target. The surrounding landscape is a dead, apocalyptic scene: crumbling concrete skeletons line its sides, shattered windows like empty eye sockets, walls pockmarked with bullet holes and black scorch marks, and the streets are littered with the wreckage of burned-out cars and twisted rebar. At that moment, a massive ball of orange-red flame erupted from the left tank's muzzle. A shell, trailing a fiery trail, whizzed out at astonishing speed, slamming fiercely into the rubble of the building in front of the right tank. A deafening explosion instantly ignited, sending countless concrete fragments and sparks flying in all directions. The blast's bright light briefly illuminated the entire street, casting dancing, high-contrast shadows on the tank's metal hull. The sky, stained an ominous dark orange by the thick smoke and dust raised by the battle, filled the entire scene with the tension, chaos, and destructive power of war.", + "seed": 27838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_2286_seed_27838_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08955323807895184, + "vlm_logprob": -0.08955323807895184, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.08955323807895184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.259848952293396 + ], + "opposite_logprob_tokens": [ + -1.509848952293396 + ], + "logprob_reward": -0.259848952293396, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.259848952293396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5470918416976929, + 0.06620413064956665, + 0.022141164168715477, + 0.02357184700667858, + 0.020519427955150604, + 0.05797085911035538, + 0.0176727082580328, + 0.10940808802843094, + 0.1021040603518486, + 0.03331587836146355 + ], + "credit_map": [ + 0.037623416632413864, + 0.025960899889469147, + 0.029642809182405472, + 0.028553728014230728, + 0.17490826547145844, + 0.11908161640167236, + 0.10659133642911911, + 0.024729754775762558, + 0.007976323366165161, + 0.015445330180227757, + 0.006246989592909813, + 0.005833855830132961, + 0.006962703540921211, + 0.010742098093032837, + 0.007298745680600405, + 0.005698081571608782, + 0.002678765682503581, + 0.003809134243056178, + 0.003756656078621745, + 0.0018346300348639488, + 0.004059403203427792, + 0.002747909864410758, + 0.0014148455811664462, + 0.001839818200096488, + 0.002426046645268798, + 0.003249170957133174, + 0.0026272968389093876, + 0.0025860723108053207, + 0.004340021405369043, + 0.0032290243543684483, + 0.0031681465916335583, + 0.0019460672046989202, + 0.002279529580846429, + 0.0017448625294491649, + 0.002927035093307495, + 0.002663611201569438, + 0.0040590460412204266, + 0.0021020073909312487, + 0.002498723566532135, + 0.0022446131333708763, + 0.003492027521133423, + 0.006163463462144136, + 0.0038888007402420044, + 0.013816357590258121, + 0.007798446342349052, + 0.013232866302132607, + 0.0050076330080628395, + 0.0045712608844041824, + 0.0017742087366059422, + 0.00160090287681669, + 0.0027748183347284794, + 0.0031824016477912664, + 0.0020713580306619406, + 0.0015061545418575406, + 0.001953671220690012, + 0.0028091934509575367, + 0.009488100185990334, + 0.0205575842410326, + 0.01090438012033701, + 0.020818907767534256, + 0.010551970452070236, + 0.013301779516041279, + 0.010452221147716045, + 0.013333141803741455, + 0.0028370716609060764, + 0.00289271492511034, + 0.002818502951413393, + 0.001993647078052163, + 0.005552430171519518, + 0.03983338177204132, + 0.0031595618929713964, + 0.04301674664020538, + 0.0033829030580818653, + 0.003146765287965536, + 0.0031195301562547684, + 0.0034078184980899096, + 0.0036048449110239744, + 0.004353461321443319, + 0.00568463746458292, + 0.006615913473069668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.259848952293396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004125181119889021 + ], + "opposite_logprob_tokens": [ + -5.504125118255615 + ], + "logprob_reward": -0.004125181119889021, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004125181119889021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19922296702861786, + 0.0641418918967247, + 0.036675579845905304, + 0.026081444695591927, + 0.05545860901474953, + 0.06417985260486603, + 0.041333623230457306, + 0.0337904691696167, + 0.42712846398353577, + 0.05198712274432182 + ], + "credit_map": [ + 0.011765730567276478, + 0.00734400749206543, + 0.04552162438631058, + 0.014388272538781166, + 0.02593175694346428, + 0.00884680263698101, + 0.0700690895318985, + 0.015355685725808144, + 0.00810270942747593, + 0.010189445689320564, + 0.00781453587114811, + 0.004314585588872433, + 0.007240633014589548, + 0.006572183221578598, + 0.012874173000454903, + 0.0070336186327040195, + 0.0027910731732845306, + 0.0028357950504869223, + 0.008754356764256954, + 0.0034984226804226637, + 0.01025344617664814, + 0.003554523689672351, + 0.002829508390277624, + 0.0021584543865174055, + 0.001778722507879138, + 0.002445588819682598, + 0.0035257015842944384, + 0.0036065129097551107, + 0.004373654723167419, + 0.0027984781190752983, + 0.004533993545919657, + 0.003018793649971485, + 0.005098066292703152, + 0.005076311994343996, + 0.008314897306263447, + 0.007820685394108295, + 0.009436434134840965, + 0.008078874088823795, + 0.005372840911149979, + 0.006260501686483622, + 0.0034325956366956234, + 0.0036877335514873266, + 0.003327511018142104, + 0.020680395886301994, + 0.005147786345332861, + 0.02127671428024769, + 0.002712030429393053, + 0.003915085457265377, + 0.004945696331560612, + 0.0037740757688879967, + 0.005988462828099728, + 0.007221526931971312, + 0.004409892484545708, + 0.004109160043299198, + 0.004752512089908123, + 0.0061322953552007675, + 0.00223138346336782, + 0.004077879246324301, + 0.003103274153545499, + 0.010380118153989315, + 0.0022421793546527624, + 0.0036288786213845015, + 0.0031006657518446445, + 0.005026089958846569, + 0.011582329869270325, + 0.009100102819502354, + 0.01013591606169939, + 0.00575146684423089, + 0.0244065560400486, + 0.17004436254501343, + 0.012113695964217186, + 0.18399403989315033, + 0.004371271003037691, + 0.005828088615089655, + 0.004877058323472738, + 0.006143214181065559, + 0.0051921941339969635, + 0.005762607324868441, + 0.009281530044972897, + 0.010531158186495304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.004125181119889021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195733428001404 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195733428001404, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195733428001404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1296110600233078, + 0.0429571233689785, + 0.05199778825044632, + 0.025580719113349915, + 0.07441814988851547, + 0.22838285565376282, + 0.027419129386544228, + 0.06130285561084747, + 0.33446183800697327, + 0.023868480697274208 + ], + "credit_map": [ + 0.007510301191359758, + 0.005040326155722141, + 0.011404085904359818, + 0.006913160905241966, + 0.032745324075222015, + 0.023732412606477737, + 0.03377800062298775, + 0.008487439714372158, + 0.004298029001802206, + 0.006096182391047478, + 0.0032112153712660074, + 0.0033516385592520237, + 0.005058471579104662, + 0.008278927765786648, + 0.00593569315969944, + 0.006726963445544243, + 0.003828653134405613, + 0.004579135216772556, + 0.01015228871256113, + 0.0049879830330610275, + 0.013699991628527641, + 0.0070332675240933895, + 0.004284972790628672, + 0.0034314950462430716, + 0.0021911722142249346, + 0.003236451419070363, + 0.002897591795772314, + 0.002871329430490732, + 0.005016404669731855, + 0.00281564355827868, + 0.004540866240859032, + 0.002011260250583291, + 0.005642760545015335, + 0.00746432039886713, + 0.013666179962456226, + 0.009705238044261932, + 0.018626317381858826, + 0.007956374436616898, + 0.007163709960877895, + 0.004193246830254793, + 0.011654380708932877, + 0.01687631569802761, + 0.009114635176956654, + 0.07038825005292892, + 0.022910313680768013, + 0.07286355644464493, + 0.010495935566723347, + 0.014079450629651546, + 0.002725783037021756, + 0.0022129786666482687, + 0.003235375043004751, + 0.0051152510568499565, + 0.003312051296234131, + 0.0030097791459411383, + 0.0032005056273192167, + 0.004607406444847584, + 0.006423508748412132, + 0.00677196541801095, + 0.0060815284959971905, + 0.007356094196438789, + 0.007550350856035948, + 0.009412072598934174, + 0.007827031426131725, + 0.009880303405225277, + 0.009178944863379002, + 0.007098162546753883, + 0.007530669681727886, + 0.005009019747376442, + 0.015541955828666687, + 0.13249358534812927, + 0.009759497828781605, + 0.14785000681877136, + 0.0021555509883910418, + 0.0028353980742394924, + 0.0019183853873983026, + 0.0026260721497237682, + 0.004011490382254124, + 0.0028060146141797304, + 0.003369085257872939, + 0.004146483726799488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.018195733428001404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052567594684660435 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052567594684660435, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052567594684660435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30533793568611145, + 0.031544364988803864, + 0.03850829228758812, + 0.017605025321245193, + 0.07436559349298477, + 0.1017218679189682, + 0.045755784958601, + 0.035409118980169296, + 0.2506575286388397, + 0.09909453243017197 + ], + "credit_map": [ + 0.02275908552110195, + 0.015914978459477425, + 0.044011346995830536, + 0.015651289373636246, + 0.09061948955059052, + 0.08389998227357864, + 0.01907818391919136, + 0.013403600081801414, + 0.005048866383731365, + 0.0038486430421471596, + 0.004174720961600542, + 0.0019860935863107443, + 0.0038753931876271963, + 0.0067581906914711, + 0.0031679938547313213, + 0.0026844630483537912, + 0.0033861417323350906, + 0.004505946300923824, + 0.006278811488300562, + 0.004138591233640909, + 0.007562749087810516, + 0.006104529369622469, + 0.0036541575100272894, + 0.0028773669619113207, + 0.0017367423279210925, + 0.0017158478731289506, + 0.0019482058705762029, + 0.0020725519862025976, + 0.003266648156568408, + 0.0028393487446010113, + 0.0021838564425706863, + 0.0018418250838294625, + 0.006722065154463053, + 0.008162347599864006, + 0.01081253495067358, + 0.009055111557245255, + 0.01718224212527275, + 0.00872378796339035, + 0.009012682363390923, + 0.00469482596963644, + 0.004203921649605036, + 0.007172036450356245, + 0.0059123351238667965, + 0.031007396057248116, + 0.009186347015202045, + 0.03200014308094978, + 0.0060536800883710384, + 0.006186004262417555, + 0.00417307810857892, + 0.00510523421689868, + 0.006212146021425724, + 0.007481862790882587, + 0.005487898364663124, + 0.004758722148835659, + 0.005031492095440626, + 0.007505349349230528, + 0.0034768555779010057, + 0.0065759094431996346, + 0.002829506993293762, + 0.003660058369860053, + 0.00567430118098855, + 0.005816696211695671, + 0.0027767247520387173, + 0.004599069245159626, + 0.00693330867215991, + 0.007102331146597862, + 0.007253715302795172, + 0.004788939841091633, + 0.013501550070941448, + 0.09527164697647095, + 0.00842262152582407, + 0.10738342255353928, + 0.005836180876940489, + 0.008190144784748554, + 0.007336075883358717, + 0.00969021487981081, + 0.012670343741774559, + 0.010932612232863903, + 0.022797398269176483, + 0.02164156176149845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0052567594684660435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603395640850067 + ], + "opposite_logprob_tokens": [ + -1.910339593887329 + ], + "logprob_reward": -0.1603395640850067, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603395640850067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36993956565856934, + 0.03541366010904312, + 0.03730519115924835, + 0.027618516236543655, + 0.07278211414813995, + 0.1933673620223999, + 0.024081336334347725, + 0.03147245571017265, + 0.15955960750579834, + 0.04846012964844704 + ], + "credit_map": [ + 0.03385743126273155, + 0.022709868848323822, + 0.028178498148918152, + 0.019662799313664436, + 0.10309673845767975, + 0.1190672516822815, + 0.02272344008088112, + 0.020643576979637146, + 0.003772533731535077, + 0.004255527164787054, + 0.003421441651880741, + 0.00299436179921031, + 0.003877798328176141, + 0.0039030699990689754, + 0.004810404498130083, + 0.008378523401916027, + 0.0029510408639907837, + 0.002670014975592494, + 0.007343258708715439, + 0.0038931414019316435, + 0.009108011610805988, + 0.005512711126357317, + 0.002911485731601715, + 0.0029155330266803503, + 0.0030624838545918465, + 0.0025668975431472063, + 0.00410756329074502, + 0.00290496414527297, + 0.0031595593318343163, + 0.0040426841005682945, + 0.004479382652789354, + 0.003294984810054302, + 0.0063591403886675835, + 0.005917557515203953, + 0.015935344621539116, + 0.012821177020668983, + 0.011949502862989902, + 0.006978197954595089, + 0.007527471985667944, + 0.005293726921081543, + 0.008043002337217331, + 0.012180084362626076, + 0.008477699011564255, + 0.06115012615919113, + 0.01793869026005268, + 0.06200559437274933, + 0.010386807844042778, + 0.01318538561463356, + 0.0024912222288548946, + 0.0026387597899883986, + 0.0031281798146665096, + 0.003484017215669155, + 0.003411596640944481, + 0.0022999539505690336, + 0.002508878242224455, + 0.004118730314075947, + 0.0028173287864774466, + 0.004166715312749147, + 0.0036580003798007965, + 0.0048918272368609905, + 0.005549049004912376, + 0.00393588887527585, + 0.0028414728585630655, + 0.0036121755838394165, + 0.005991218611598015, + 0.006728680804371834, + 0.011037923395633698, + 0.009006528183817863, + 0.009384864009916782, + 0.05178641527891159, + 0.007453656289726496, + 0.058170340955257416, + 0.006093237549066544, + 0.005293655674904585, + 0.005929237697273493, + 0.0070176320150494576, + 0.004631661809980869, + 0.005998911336064339, + 0.0061050355434417725, + 0.007390763610601425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.1603395640850067 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a low ground angle, slowly pans right, capturing a fierce confrontation on the main thoroughfare of an abandoned city. On the left, a heavy main battle tank, clad in dark gray and heavily damaged, scratched, and rusted, its angular turret and tracks caked with dirt and gravel make its massive metal bulk appear particularly heavy. Opposite it, another, more streamlined tank, painted in desert camouflage, rapidly turns its turret toward it, locking onto its target. The surrounding landscape is a dead, apocalyptic scene: crumbling concrete skeletons line its sides, shattered windows like empty eye sockets, walls pockmarked with bullet holes and black scorch marks, and the streets are littered with the wreckage of burned-out cars and twisted rebar. At that moment, a massive ball of orange-red flame erupted from the left tank's muzzle. A shell, trailing a fiery trail, whizzed out at astonishing speed, slamming fiercely into the rubble of the building in front of the right tank. A deafening explosion instantly ignited, sending countless concrete fragments and sparks flying in all directions. The blast's bright light briefly illuminated the entire street, casting dancing, high-contrast shadows on the tank's metal hull. The sky, stained an ominous dark orange by the thick smoke and dust raised by the battle, filled the entire scene with the tension, chaos, and destructive power of war.", + "seed": 27839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_2286_seed_27839_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.053651193249970676, + "vlm_logprob": -0.053651193249970676, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.053651193249970676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02085728757083416 + ], + "opposite_logprob_tokens": [ + -4.645857334136963 + ], + "logprob_reward": -0.02085728757083416, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.02085728757083416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1547560691833496, + 0.04594345763325691, + 0.07457882165908813, + 0.047141849994659424, + 0.1898413449525833, + 0.08297527581453323, + 0.12469898164272308, + 0.20301753282546997, + 0.029802950099110603, + 0.04724368080496788 + ], + "credit_map": [ + 0.015317747369408607, + 0.023264462128281593, + 0.027643006294965744, + 0.02070697769522667, + 0.020038293674588203, + 0.017992280423641205, + 0.014348804950714111, + 0.01544449757784605, + 0.005735794082283974, + 0.005960420705378056, + 0.006647899281233549, + 0.008418241515755653, + 0.005232538562268019, + 0.004267433192580938, + 0.005514970514923334, + 0.004166162572801113, + 0.00853519793599844, + 0.0076842838898301125, + 0.008973204530775547, + 0.009908169507980347, + 0.010098104365170002, + 0.012773558497428894, + 0.00896216370165348, + 0.007644138764590025, + 0.0061953249387443066, + 0.0048814937472343445, + 0.00733910221606493, + 0.004763420205563307, + 0.004387131426483393, + 0.00826273113489151, + 0.005599444266408682, + 0.005713201593607664, + 0.011091355234384537, + 0.022524382919073105, + 0.05584179610013962, + 0.027654465287923813, + 0.02092074230313301, + 0.015908004716038704, + 0.01630479283630848, + 0.019595803692936897, + 0.009698878042399883, + 0.012622498907148838, + 0.010273953899741173, + 0.011345003731548786, + 0.00742329191416502, + 0.00745737599208951, + 0.015609120950102806, + 0.00854515191167593, + 0.008907771669328213, + 0.008157558739185333, + 0.027245130389928818, + 0.021195948123931885, + 0.004909918177872896, + 0.00822397693991661, + 0.004642019979655743, + 0.04141665995121002, + 0.01756369136273861, + 0.01015490386635065, + 0.04410510137677193, + 0.011436940170824528, + 0.009870219975709915, + 0.019813356921076775, + 0.012469714507460594, + 0.07760361582040787, + 0.002594424644485116, + 0.002668926492333412, + 0.0056093367747962475, + 0.0025066398084163666, + 0.005981978960335255, + 0.003799869678914547, + 0.0036922637373209, + 0.002949509071186185, + 0.006869105622172356, + 0.0031840545125305653, + 0.006065861321985722, + 0.0034683470148593187, + 0.005850387271493673, + 0.0041540684178471565, + 0.00892450287938118, + 0.008727357722818851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.02085728757083416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005971685517579317 + ], + "opposite_logprob_tokens": [ + -5.130971908569336 + ], + "logprob_reward": -0.005971685517579317, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005971685517579317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19819553196430206, + 0.0961528941988945, + 0.1497901976108551, + 0.12371093779802322, + 0.08199264854192734, + 0.06661327183246613, + 0.06447392702102661, + 0.1107783392071724, + 0.04309602826833725, + 0.0651962012052536 + ], + "credit_map": [ + 0.01808886229991913, + 0.03735954314470291, + 0.02603665553033352, + 0.03496671840548515, + 0.015064939856529236, + 0.03082539513707161, + 0.01942008174955845, + 0.0164333488792181, + 0.01501419022679329, + 0.010134623385965824, + 0.012966958805918694, + 0.01379399187862873, + 0.01181594654917717, + 0.014025592245161533, + 0.007855439558625221, + 0.010546162724494934, + 0.023622622713446617, + 0.028559373691678047, + 0.023286402225494385, + 0.01998141221702099, + 0.013750249519944191, + 0.01226647850126028, + 0.012327865697443485, + 0.015995819121599197, + 0.015882223844528198, + 0.013059784658253193, + 0.022531524300575256, + 0.015410710126161575, + 0.008871358819305897, + 0.013727640733122826, + 0.01516649965196848, + 0.01906120777130127, + 0.006212096661329269, + 0.007691576145589352, + 0.014388599433004856, + 0.01134431641548872, + 0.013468464836478233, + 0.010776865296065807, + 0.0074019040912389755, + 0.01070883497595787, + 0.007602146361023188, + 0.008933314122259617, + 0.008326364681124687, + 0.008121953345835209, + 0.00636420538648963, + 0.011175801046192646, + 0.008559068664908409, + 0.00753041822463274, + 0.006676923483610153, + 0.006097337231040001, + 0.009301942773163319, + 0.013426737859845161, + 0.006246507167816162, + 0.007130816578865051, + 0.0048551298677921295, + 0.010738539509475231, + 0.013778026215732098, + 0.005625884979963303, + 0.03141274303197861, + 0.008177999407052994, + 0.007752737496048212, + 0.009610144421458244, + 0.008885481394827366, + 0.025535333901643753, + 0.0026371225249022245, + 0.004185555502772331, + 0.005008831154555082, + 0.0036730102729052305, + 0.008343059569597244, + 0.006034961435943842, + 0.004272897262126207, + 0.00894059520214796, + 0.007950053550302982, + 0.006003915332257748, + 0.00880905706435442, + 0.005942688789218664, + 0.0073837777599692345, + 0.006613654550164938, + 0.00946547370404005, + 0.013027585111558437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.005971685517579317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002870738506317 + ], + "opposite_logprob_tokens": [ + -2.695028781890869 + ], + "logprob_reward": -0.07002870738506317, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002870738506317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17669017612934113, + 0.12965935468673706, + 0.07218306511640549, + 0.07775572687387466, + 0.2356707751750946, + 0.035713933408260345, + 0.04411204904317856, + 0.08551201969385147, + 0.10137779265642166, + 0.041325151920318604 + ], + "credit_map": [ + 0.02067972905933857, + 0.030602335929870605, + 0.020846107974648476, + 0.027012910693883896, + 0.015373884700238705, + 0.019635183736681938, + 0.01628192700445652, + 0.026258086785674095, + 0.01866663433611393, + 0.013465900905430317, + 0.018096670508384705, + 0.01947042904794216, + 0.018839437514543533, + 0.017183097079396248, + 0.011968011036515236, + 0.011969160288572311, + 0.0067265830002725124, + 0.007202297914773226, + 0.009040703065693378, + 0.010347703471779823, + 0.010929401032626629, + 0.011466382071375847, + 0.009487097151577473, + 0.006982891820371151, + 0.011815088801085949, + 0.009451012127101421, + 0.01265525259077549, + 0.010274208150804043, + 0.005404794588685036, + 0.009018493816256523, + 0.008695671334862709, + 0.010441197082400322, + 0.014961985871195793, + 0.028441429138183594, + 0.05162367224693298, + 0.039611075073480606, + 0.032557349652051926, + 0.024792395532131195, + 0.022112920880317688, + 0.021569928154349327, + 0.004138528369367123, + 0.0035702751483768225, + 0.003858749521896243, + 0.004857244435697794, + 0.0038894074968993664, + 0.00504739535972476, + 0.0066981748677790165, + 0.003654156345874071, + 0.004378197249025106, + 0.005348958540707827, + 0.005428137723356485, + 0.007476157043129206, + 0.004169691354036331, + 0.0077575515024363995, + 0.003294230904430151, + 0.00625911820679903, + 0.009162398986518383, + 0.006164319347590208, + 0.021614298224449158, + 0.007069514133036137, + 0.007541737984865904, + 0.006807928439229727, + 0.00815991684794426, + 0.018991895020008087, + 0.008319866843521595, + 0.011008736677467823, + 0.02173667401075363, + 0.01201171800494194, + 0.013856125995516777, + 0.009206229820847511, + 0.010732162743806839, + 0.014506264589726925, + 0.006168213672935963, + 0.004580809269100428, + 0.006072220858186483, + 0.003511931514367461, + 0.003512105206027627, + 0.0036440303083509207, + 0.007005434483289719, + 0.0068304031156003475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.07002870738506317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074859648942947 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074859648942947, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074859648942947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0924108624458313, + 0.05105193704366684, + 0.1308056265115738, + 0.06451327353715897, + 0.11283544450998306, + 0.08376152813434601, + 0.1525477170944214, + 0.10536088794469833, + 0.03427334129810333, + 0.17243938148021698 + ], + "credit_map": [ + 0.007868777960538864, + 0.01699874922633171, + 0.01820506900548935, + 0.014568385668098927, + 0.00884249247610569, + 0.010363585315644741, + 0.006364541593939066, + 0.009199262596666813, + 0.007963521406054497, + 0.006480694282799959, + 0.006294493563473225, + 0.0074296677485108376, + 0.005893218796700239, + 0.006277845241129398, + 0.0059876954182982445, + 0.004724795930087566, + 0.012689725495874882, + 0.029304329305887222, + 0.016580617055296898, + 0.017994895577430725, + 0.012871296145021915, + 0.019371433183550835, + 0.011085105128586292, + 0.01090823020786047, + 0.01024165004491806, + 0.008638222701847553, + 0.010711501352488995, + 0.008428676053881645, + 0.004003357607871294, + 0.0060980296693742275, + 0.006844308692961931, + 0.009547526948153973, + 0.00759342173114419, + 0.012130824849009514, + 0.02968812920153141, + 0.019423585385084152, + 0.012181824073195457, + 0.010338107123970985, + 0.011079290881752968, + 0.010400259867310524, + 0.008900167420506477, + 0.013679525814950466, + 0.009874640963971615, + 0.010011718608438969, + 0.009330084547400475, + 0.012462785467505455, + 0.00800841860473156, + 0.011494184844195843, + 0.01006945502012968, + 0.01026087999343872, + 0.03147302567958832, + 0.018708553165197372, + 0.011595715768635273, + 0.016392510384321213, + 0.00846569612622261, + 0.0455818697810173, + 0.01048780232667923, + 0.006007349584251642, + 0.028418907895684242, + 0.007697070017457008, + 0.008576040156185627, + 0.007638696581125259, + 0.00829002633690834, + 0.028244996443390846, + 0.0028946504462510347, + 0.004070911090821028, + 0.005749119911342859, + 0.0035038269124925137, + 0.005368377082049847, + 0.005234801676124334, + 0.003027072874829173, + 0.00442458176985383, + 0.011643815785646439, + 0.011828197166323662, + 0.01328795962035656, + 0.009569293819367886, + 0.01426882017403841, + 0.009900842793285847, + 0.050544992089271545, + 0.0513954684138298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.011074859648942947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032342612743378 + ], + "opposite_logprob_tokens": [ + -1.9103233814239502 + ], + "logprob_reward": -0.16032342612743378, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032342612743378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1016513779759407, + 0.09446278214454651, + 0.15428932011127472, + 0.04809492826461792, + 0.28578615188598633, + 0.04655042290687561, + 0.11035726964473724, + 0.05400743708014488, + 0.035822730511426926, + 0.06897758692502975 + ], + "credit_map": [ + 0.010896844789385796, + 0.020823882892727852, + 0.016196174547076225, + 0.01684102788567543, + 0.009946257807314396, + 0.010895062237977982, + 0.007715979591012001, + 0.008336148224771023, + 0.016171816736459732, + 0.011132513172924519, + 0.010666709393262863, + 0.013707210309803486, + 0.010387849994003773, + 0.007844900712370872, + 0.012981402687728405, + 0.011570384725928307, + 0.0155485849827528, + 0.019936541095376015, + 0.016706546768546104, + 0.02075178362429142, + 0.022489897906780243, + 0.032517123967409134, + 0.013209027238190174, + 0.013129818253219128, + 0.00488841999322176, + 0.005918179173022509, + 0.007396166678518057, + 0.006307271774858236, + 0.003885838435962796, + 0.005413727834820747, + 0.00673955213278532, + 0.007545772008597851, + 0.017367953434586525, + 0.03423817828297615, + 0.07850857824087143, + 0.04554373398423195, + 0.033261988312006, + 0.024820946156978607, + 0.02578134462237358, + 0.02626344934105873, + 0.004851389676332474, + 0.004375246819108725, + 0.006034171674400568, + 0.005727341864258051, + 0.0052925460040569305, + 0.006027281284332275, + 0.008603719994425774, + 0.005638728383928537, + 0.0094657177105546, + 0.01003320049494505, + 0.015961039811372757, + 0.02035527490079403, + 0.012997491285204887, + 0.012733682990074158, + 0.008285395801067352, + 0.020525464788079262, + 0.006207101512700319, + 0.003073866246268153, + 0.00808302965015173, + 0.0030808241572231054, + 0.004788060672581196, + 0.007384851109236479, + 0.004637939855456352, + 0.016751762479543686, + 0.002460024319589138, + 0.004071698058396578, + 0.003909151069819927, + 0.0027252694126218557, + 0.00854816846549511, + 0.0055116708390414715, + 0.0036027731839567423, + 0.00499397236853838, + 0.004901356529444456, + 0.005442210007458925, + 0.005212098825722933, + 0.00376504915766418, + 0.006350341252982616, + 0.005794837139546871, + 0.021291909739375114, + 0.016219789162278175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.16032342612743378 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a low ground angle, slowly pans right, capturing a fierce confrontation on the main thoroughfare of an abandoned city. On the left, a heavy main battle tank, clad in dark gray and heavily damaged, scratched, and rusted, its angular turret and tracks caked with dirt and gravel make its massive metal bulk appear particularly heavy. Opposite it, another, more streamlined tank, painted in desert camouflage, rapidly turns its turret toward it, locking onto its target. The surrounding landscape is a dead, apocalyptic scene: crumbling concrete skeletons line its sides, shattered windows like empty eye sockets, walls pockmarked with bullet holes and black scorch marks, and the streets are littered with the wreckage of burned-out cars and twisted rebar. At that moment, a massive ball of orange-red flame erupted from the left tank's muzzle. A shell, trailing a fiery trail, whizzed out at astonishing speed, slamming fiercely into the rubble of the building in front of the right tank. A deafening explosion instantly ignited, sending countless concrete fragments and sparks flying in all directions. The blast's bright light briefly illuminated the entire street, casting dancing, high-contrast shadows on the tank's metal hull. The sky, stained an ominous dark orange by the thick smoke and dust raised by the battle, filled the entire scene with the tension, chaos, and destructive power of war.", + "seed": 27840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_2286_seed_27840_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07854745578952134, + "vlm_logprob": -0.07854745578952134, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.07854745578952134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1752127707004547 + ], + "opposite_logprob_tokens": [ + -1.9252127408981323 + ], + "logprob_reward": -0.1752127707004547, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1752127707004547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05481015890836716, + 0.04528215900063515, + 0.11244082450866699, + 0.10049954056739807, + 0.0276698749512434, + 0.25394174456596375, + 0.0352102667093277, + 0.05939660966396332, + 0.09245605021715164, + 0.2182927280664444 + ], + "credit_map": [ + 0.010673481971025467, + 0.005274688825011253, + 0.005999098997563124, + 0.009665937162935734, + 0.005575431045144796, + 0.007044900208711624, + 0.004552402999252081, + 0.006024216301739216, + 0.0065850550308823586, + 0.004018903709948063, + 0.014454868622124195, + 0.0035303838085383177, + 0.0038726760540157557, + 0.004652947653084993, + 0.003560965647920966, + 0.0046063573099672794, + 0.03692637011408806, + 0.03622114658355713, + 0.008402004837989807, + 0.005738930311053991, + 0.006730237044394016, + 0.006721848621964455, + 0.005389309022575617, + 0.0063109933398664, + 0.01836528815329075, + 0.010657571256160736, + 0.011212139390408993, + 0.012102332897484303, + 0.009144539944827557, + 0.01601302996277809, + 0.011553822085261345, + 0.011450818739831448, + 0.003033950924873352, + 0.0038682741578668356, + 0.0023679272271692753, + 0.0031699142418801785, + 0.0046958113089203835, + 0.0029833875596523285, + 0.0032635650131851435, + 0.0042870440520346165, + 0.02636752277612686, + 0.027593379840254784, + 0.03826604038476944, + 0.022879747673869133, + 0.02686513029038906, + 0.05236855149269104, + 0.026142852380871773, + 0.033458516001701355, + 0.005475760903209448, + 0.004705361090600491, + 0.003579994197934866, + 0.0033578649163246155, + 0.004261988215148449, + 0.004024396650493145, + 0.0052085802890360355, + 0.004596319515258074, + 0.011051803827285767, + 0.007910232059657574, + 0.008773786947131157, + 0.0068579176440835, + 0.006756818853318691, + 0.004344673827290535, + 0.007667492143809795, + 0.006033886689692736, + 0.007567237131297588, + 0.014975824393332005, + 0.005653353873640299, + 0.006045986898243427, + 0.028838085010647774, + 0.011022388935089111, + 0.010067177005112171, + 0.00828599464148283, + 0.02101939730346203, + 0.03044251538813114, + 0.014855368994176388, + 0.018443554639816284, + 0.045687440782785416, + 0.022035088390111923, + 0.0436166487634182, + 0.02219269797205925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.1752127707004547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003659577341750264 + ], + "opposite_logprob_tokens": [ + -5.628659725189209 + ], + "logprob_reward": -0.003659577341750264, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003659577341750264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16334590315818787, + 0.03508288040757179, + 0.03575410321354866, + 0.04446275532245636, + 0.27635663747787476, + 0.21079669892787933, + 0.04474731162190437, + 0.03662094846367836, + 0.08394578099250793, + 0.06888691335916519 + ], + "credit_map": [ + 0.04978947713971138, + 0.013724839314818382, + 0.012678460218012333, + 0.012504165060818195, + 0.016160743311047554, + 0.012463943101465702, + 0.0110678905621171, + 0.034956399351358414, + 0.006092058960348368, + 0.0046562072820961475, + 0.003991601523011923, + 0.003630691673606634, + 0.004702459089457989, + 0.0036266655661165714, + 0.0038500374648720026, + 0.004533159546554089, + 0.011109035462141037, + 0.011331355199217796, + 0.001974313287064433, + 0.0024234424345195293, + 0.0030324351973831654, + 0.0026001327205449343, + 0.0016855726717039943, + 0.0015978135634213686, + 0.006538309622555971, + 0.004506977275013924, + 0.0048632496036589146, + 0.005975807551294565, + 0.0047643450088799, + 0.007002204190939665, + 0.0051874336786568165, + 0.005624430254101753, + 0.019896285608410835, + 0.013743245042860508, + 0.0981898307800293, + 0.031608834862709045, + 0.05780819058418274, + 0.02070060931146145, + 0.015370490029454231, + 0.019039135426282883, + 0.0205539520829916, + 0.02971191145479679, + 0.027567926794290543, + 0.017574816942214966, + 0.0208269152790308, + 0.04363695904612541, + 0.02457442879676819, + 0.026349792256951332, + 0.007352289743721485, + 0.006829860620200634, + 0.004826026037335396, + 0.004159057978540659, + 0.005383757874369621, + 0.004514401312917471, + 0.006649060174822807, + 0.005032860208302736, + 0.005493178963661194, + 0.003021510550752282, + 0.008564197458326817, + 0.002793100429698825, + 0.007642406038939953, + 0.0029908001888543367, + 0.004354008007794619, + 0.0017617458943277597, + 0.005820850841701031, + 0.015828130766749382, + 0.006627654656767845, + 0.009695792570710182, + 0.0230275709182024, + 0.007604336831718683, + 0.010023054666817188, + 0.005318388808518648, + 0.007308297790586948, + 0.00788129959255457, + 0.004124845378100872, + 0.005762810353189707, + 0.016173692420125008, + 0.011194265447556973, + 0.010865936987102032, + 0.00557576771825552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.003659577341750264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359431445598602 + ], + "opposite_logprob_tokens": [ + -3.6513593196868896 + ], + "logprob_reward": -0.026359431445598602, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359431445598602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18172308802604675, + 0.05924359709024429, + 0.03712684288620949, + 0.05406475439667702, + 0.25079527497291565, + 0.12200428545475006, + 0.06239407882094383, + 0.05070464313030243, + 0.06067730486392975, + 0.12126621603965759 + ], + "credit_map": [ + 0.03884243965148926, + 0.014875025488436222, + 0.017208823934197426, + 0.014355963096022606, + 0.02034449577331543, + 0.016301248222589493, + 0.01728261075913906, + 0.04251246154308319, + 0.008336170576512814, + 0.0077927918173372746, + 0.005395662970840931, + 0.005059795919805765, + 0.01000149454921484, + 0.005796539131551981, + 0.009457066655158997, + 0.007404068950563669, + 0.007840155623853207, + 0.008716809563338757, + 0.004004251677542925, + 0.0038192563224583864, + 0.003774592187255621, + 0.002848037052899599, + 0.003559759585186839, + 0.002563976915553212, + 0.008622744120657444, + 0.005487136542797089, + 0.005214537028223276, + 0.011436296626925468, + 0.00750687113031745, + 0.004408278968185186, + 0.006344629917293787, + 0.005044252146035433, + 0.01678365282714367, + 0.01788036711513996, + 0.07268642634153366, + 0.026651155203580856, + 0.05349988117814064, + 0.028325067833065987, + 0.01808990351855755, + 0.016878793016076088, + 0.012863321229815483, + 0.02304060012102127, + 0.00925307348370552, + 0.010274355299770832, + 0.011572460643947124, + 0.027185987681150436, + 0.013069726526737213, + 0.014744739979505539, + 0.012261277996003628, + 0.008835096843540668, + 0.0069312285631895065, + 0.0058432347141206264, + 0.006356833036988974, + 0.006800211034715176, + 0.009039199911057949, + 0.006326989736407995, + 0.006756608374416828, + 0.007300466764718294, + 0.005175708793103695, + 0.005390863865613937, + 0.007878370583057404, + 0.0067102545872330666, + 0.006376191973686218, + 0.005116170272231102, + 0.005506305024027824, + 0.010863850824534893, + 0.0060954918153584, + 0.0041781323961913586, + 0.01672249473631382, + 0.005565405357629061, + 0.007996576838195324, + 0.0037490411195904016, + 0.014526959508657455, + 0.01647684909403324, + 0.01001505833119154, + 0.009793412871658802, + 0.018413782119750977, + 0.013787109404802322, + 0.022369995713233948, + 0.015883028507232666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.026359431445598602 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632971741259098 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007632971741259098, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632971741259098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1595800817012787, + 0.03843783587217331, + 0.04144413769245148, + 0.03790004551410675, + 0.08634237200021744, + 0.12381885945796967, + 0.145910382270813, + 0.08264733850955963, + 0.20621097087860107, + 0.0777079239487648 + ], + "credit_map": [ + 0.047695815563201904, + 0.014362773858010769, + 0.01069787796586752, + 0.012655694968998432, + 0.01435853261500597, + 0.012624915689229965, + 0.009223065339028835, + 0.03796139732003212, + 0.006288825534284115, + 0.0039470079354941845, + 0.006795505993068218, + 0.003873394802212715, + 0.005745519418269396, + 0.004526214208453894, + 0.004072003066539764, + 0.0031893663108348846, + 0.006137410644441843, + 0.005292276851832867, + 0.005674518644809723, + 0.0048548015765845776, + 0.004915805999189615, + 0.006014897953718901, + 0.004382444079965353, + 0.004171984735876322, + 0.0065564606338739395, + 0.003217071760445833, + 0.004258270841091871, + 0.0072022052481770515, + 0.004052239004522562, + 0.004496022127568722, + 0.003895155619829893, + 0.004222620278596878, + 0.007413908839225769, + 0.006091170012950897, + 0.020914340391755104, + 0.007901608012616634, + 0.0177371297031641, + 0.007360970135778189, + 0.010438872501254082, + 0.00848437286913395, + 0.013650554232299328, + 0.019966794177889824, + 0.008648822084069252, + 0.010423307307064533, + 0.012437085621058941, + 0.029925324022769928, + 0.012883514165878296, + 0.015883460640907288, + 0.026064705103635788, + 0.01684139482676983, + 0.014300940558314323, + 0.013777106069028378, + 0.016886534169316292, + 0.01581902988255024, + 0.02459774725139141, + 0.017622925341129303, + 0.013430539518594742, + 0.010799582116305828, + 0.008077604696154594, + 0.006616729777306318, + 0.01730014942586422, + 0.010237835347652435, + 0.00830168928951025, + 0.007883205078542233, + 0.015882840380072594, + 0.03375032916665077, + 0.0204265508800745, + 0.02893463335931301, + 0.043487608432769775, + 0.024401197209954262, + 0.022757703438401222, + 0.016570111736655235, + 0.009837773628532887, + 0.010911460965871811, + 0.006250796839594841, + 0.006234940607100725, + 0.01595502719283104, + 0.0074797021225094795, + 0.013788716867566109, + 0.007249502930790186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.007632971741259098 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179872527718544 + ], + "opposite_logprob_tokens": [ + -1.8048725128173828 + ], + "logprob_reward": -0.179872527718544, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179872527718544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15168210864067078, + 0.03281637281179428, + 0.09608807414770126, + 0.034636203199625015, + 0.055273961275815964, + 0.12825152277946472, + 0.10843531787395477, + 0.10013977438211441, + 0.1879628747701645, + 0.10471384227275848 + ], + "credit_map": [ + 0.05553821846842766, + 0.008892661891877651, + 0.010270251892507076, + 0.00973240565508604, + 0.009957830421626568, + 0.007177331484854221, + 0.007737948093563318, + 0.04237545281648636, + 0.004192472435534, + 0.0028035349678248167, + 0.009185444563627243, + 0.003518752520903945, + 0.004391943570226431, + 0.003218844998627901, + 0.003102231305092573, + 0.002403149614110589, + 0.03371559828519821, + 0.03431772068142891, + 0.005474676378071308, + 0.00499130180105567, + 0.004964105784893036, + 0.004557294771075249, + 0.003687480231747031, + 0.004379900172352791, + 0.006848807446658611, + 0.003762998152524233, + 0.002471391810104251, + 0.007080586161464453, + 0.0045431009493768215, + 0.003328202059492469, + 0.0038738902658224106, + 0.0027272289153188467, + 0.00525628449395299, + 0.0066858502104878426, + 0.00525387329980731, + 0.00490625761449337, + 0.008218154311180115, + 0.010901437141001225, + 0.006688762456178665, + 0.007363340817391872, + 0.012231064029037952, + 0.02032092772424221, + 0.009407161734998226, + 0.010612870566546917, + 0.011575473472476006, + 0.03237622231245041, + 0.014802523888647556, + 0.016925279051065445, + 0.020828409120440483, + 0.017219098284840584, + 0.010308204218745232, + 0.009930514730513096, + 0.012906374409794807, + 0.012168936431407928, + 0.015698567032814026, + 0.009375212714076042, + 0.0145050548017025, + 0.015570342540740967, + 0.013611240312457085, + 0.012470517307519913, + 0.014227677136659622, + 0.01125517301261425, + 0.012003732845187187, + 0.00649603083729744, + 0.021498244255781174, + 0.026790885254740715, + 0.021138595417141914, + 0.02374967746436596, + 0.03779006749391556, + 0.02072988636791706, + 0.020353049039840698, + 0.015912458300590515, + 0.013268266804516315, + 0.013273082673549652, + 0.006786158308386803, + 0.010683275759220123, + 0.020427236333489418, + 0.01113375648856163, + 0.018917325884103775, + 0.010224738158285618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.179872527718544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a low ground angle, slowly pans right, capturing a fierce confrontation on the main thoroughfare of an abandoned city. On the left, a heavy main battle tank, clad in dark gray and heavily damaged, scratched, and rusted, its angular turret and tracks caked with dirt and gravel make its massive metal bulk appear particularly heavy. Opposite it, another, more streamlined tank, painted in desert camouflage, rapidly turns its turret toward it, locking onto its target. The surrounding landscape is a dead, apocalyptic scene: crumbling concrete skeletons line its sides, shattered windows like empty eye sockets, walls pockmarked with bullet holes and black scorch marks, and the streets are littered with the wreckage of burned-out cars and twisted rebar. At that moment, a massive ball of orange-red flame erupted from the left tank's muzzle. A shell, trailing a fiery trail, whizzed out at astonishing speed, slamming fiercely into the rubble of the building in front of the right tank. A deafening explosion instantly ignited, sending countless concrete fragments and sparks flying in all directions. The blast's bright light briefly illuminated the entire street, casting dancing, high-contrast shadows on the tank's metal hull. The sky, stained an ominous dark orange by the thick smoke and dust raised by the battle, filled the entire scene with the tension, chaos, and destructive power of war.", + "seed": 27841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_3/index_2286_seed_27841_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05585438897833228, + "vlm_logprob": -0.05585438897833228, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.05585438897833228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08082085847854614 + ], + "opposite_logprob_tokens": [ + -2.7058207988739014 + ], + "logprob_reward": -0.08082085847854614, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08082085847854614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16395992040634155, + 0.11802130937576294, + 0.10197200626134872, + 0.03154299780726433, + 0.052251558750867844, + 0.09989096969366074, + 0.029170963913202286, + 0.05854470282793045, + 0.09269107133150101, + 0.25195449590682983 + ], + "credit_map": [ + 0.03453463315963745, + 0.019060632213950157, + 0.012774938717484474, + 0.01518959179520607, + 0.01721099764108658, + 0.024234874173998833, + 0.02222057804465294, + 0.018733667209744453, + 0.017389798536896706, + 0.02234940603375435, + 0.01073986105620861, + 0.015201795846223831, + 0.01410393975675106, + 0.011307856999337673, + 0.01810939982533455, + 0.008819238282740116, + 0.024928538128733635, + 0.01411957386881113, + 0.00683090602979064, + 0.009827966801822186, + 0.008905184455215931, + 0.013316743075847626, + 0.0161618459969759, + 0.007881240919232368, + 0.0050638350658118725, + 0.004411057103425264, + 0.003245231229811907, + 0.0031696006190031767, + 0.0027826237492263317, + 0.0038884549867361784, + 0.004174140747636557, + 0.004808049649000168, + 0.009173539467155933, + 0.006833871826529503, + 0.005962111987173557, + 0.005863247439265251, + 0.005328889936208725, + 0.007134974468499422, + 0.0061435578390955925, + 0.0058113643899559975, + 0.01823321543633938, + 0.011679341085255146, + 0.010165520012378693, + 0.009433104656636715, + 0.011449359357357025, + 0.011121782474219799, + 0.00975322350859642, + 0.01805541291832924, + 0.0039694239385426044, + 0.005988846067339182, + 0.0022712573409080505, + 0.0026317546144127846, + 0.0035032157320529222, + 0.0034984343219548464, + 0.004466980695724487, + 0.002841049339622259, + 0.004637066274881363, + 0.004048389382660389, + 0.01291175652295351, + 0.004110387992113829, + 0.00728069432079792, + 0.004971833899617195, + 0.010147575289011002, + 0.010436990298330784, + 0.012365425936877728, + 0.010324336588382721, + 0.008962685242295265, + 0.007755473721772432, + 0.017821019515395164, + 0.00941550824791193, + 0.012613202445209026, + 0.013433408923447132, + 0.01961445063352585, + 0.016445636749267578, + 0.019696170464158058, + 0.019284427165985107, + 0.040243037045001984, + 0.0639716163277626, + 0.03268887475132942, + 0.0400102473795414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.08082085847854614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004129336215555668 + ], + "opposite_logprob_tokens": [ + -5.504129409790039 + ], + "logprob_reward": -0.004129336215555668, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004129336215555668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14623521268367767, + 0.100010946393013, + 0.0709541067481041, + 0.0830787643790245, + 0.257390558719635, + 0.105953149497509, + 0.06958623975515366, + 0.055148106068372726, + 0.02932269498705864, + 0.08232024312019348 + ], + "credit_map": [ + 0.022121140733361244, + 0.024897901341319084, + 0.014992598444223404, + 0.015318132005631924, + 0.016717981547117233, + 0.019747411832213402, + 0.017554229125380516, + 0.014885813929140568, + 0.013925964944064617, + 0.011507079005241394, + 0.012739698402583599, + 0.024355722591280937, + 0.006025200244039297, + 0.011748027056455612, + 0.010884416289627552, + 0.008824838325381279, + 0.01060392428189516, + 0.012776919640600681, + 0.00538681261241436, + 0.010301410220563412, + 0.007945721969008446, + 0.009474554099142551, + 0.007350733038038015, + 0.007114033680409193, + 0.011735929176211357, + 0.014301473274827003, + 0.005974899977445602, + 0.008407632820308208, + 0.008445683866739273, + 0.008913274854421616, + 0.010144392028450966, + 0.015155477449297905, + 0.08056368678808212, + 0.016769040375947952, + 0.012261845171451569, + 0.011315752752125263, + 0.015073100104928017, + 0.07825479656457901, + 0.0256353672593832, + 0.01751697063446045, + 0.019787291064858437, + 0.016535459086298943, + 0.008975091390311718, + 0.007619543932378292, + 0.01549196895211935, + 0.01365218311548233, + 0.011231455020606518, + 0.012660147622227669, + 0.010011847130954266, + 0.008352276869118214, + 0.004942900966852903, + 0.0060647898353636265, + 0.008989791385829449, + 0.007000073324888945, + 0.016550220549106598, + 0.00767434062436223, + 0.005726428236812353, + 0.006065444555133581, + 0.012035957537591457, + 0.00457283528521657, + 0.005412801168859005, + 0.00382164865732193, + 0.011159813962876797, + 0.006353179458528757, + 0.0027094678953289986, + 0.0030912195798009634, + 0.00439077103510499, + 0.004412548616528511, + 0.0036378828808665276, + 0.0034699109382927418, + 0.00421782536432147, + 0.0033930675126612186, + 0.008832225576043129, + 0.007910159416496754, + 0.008911392651498318, + 0.007276292424649, + 0.0098886638879776, + 0.015164204873144627, + 0.012179034762084484, + 0.012158269993960857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.004129336215555668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635699324309826 + ], + "opposite_logprob_tokens": [ + -3.6513569355010986 + ], + "logprob_reward": -0.02635699324309826, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635699324309826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10919312387704849, + 0.07668305933475494, + 0.12940636277198792, + 0.05430033430457115, + 0.200128436088562, + 0.1404440999031067, + 0.07933811843395233, + 0.10535941272974014, + 0.03892474249005318, + 0.06622228026390076 + ], + "credit_map": [ + 0.027018476277589798, + 0.015237796120345592, + 0.01130936574190855, + 0.008094889111816883, + 0.00795060209929943, + 0.010239808820188046, + 0.018126370385289192, + 0.011215812526643276, + 0.015377254225313663, + 0.010214749723672867, + 0.010007916018366814, + 0.008614365942776203, + 0.006258904002606869, + 0.012443115003407001, + 0.006774473935365677, + 0.006992282811552286, + 0.02869712933897972, + 0.01598885841667652, + 0.009874158538877964, + 0.013112514279782772, + 0.012353158555924892, + 0.019043676555156708, + 0.019577739760279655, + 0.01075912918895483, + 0.008802607655525208, + 0.014593207277357578, + 0.004936572164297104, + 0.005100376438349485, + 0.00350672984495759, + 0.005848069675266743, + 0.0053796726278960705, + 0.006133098620921373, + 0.0640307292342186, + 0.013278917409479618, + 0.010660219937562943, + 0.0070507884956896305, + 0.018834227696061134, + 0.05248020216822624, + 0.020432747900485992, + 0.013360602781176567, + 0.023164978250861168, + 0.018545472994446754, + 0.013903269544243813, + 0.01568497344851494, + 0.017619164660573006, + 0.01746942102909088, + 0.014142771251499653, + 0.019914042204618454, + 0.01575729064643383, + 0.010346274822950363, + 0.0040204175747931, + 0.004556253552436829, + 0.011816717684268951, + 0.0089528514072299, + 0.016108313575387, + 0.007779994048178196, + 0.01086913701146841, + 0.007677809800952673, + 0.01577930897474289, + 0.00759061798453331, + 0.011732959188520908, + 0.009348045103251934, + 0.028371168300509453, + 0.01399037055671215, + 0.004305838141590357, + 0.0038847248069941998, + 0.00470758369192481, + 0.0034993221051990986, + 0.006341992877423763, + 0.0034929949324578047, + 0.007655794266611338, + 0.005036490503698587, + 0.007419292349368334, + 0.008282742463052273, + 0.005585344973951578, + 0.008141576312482357, + 0.008479340933263302, + 0.010549915954470634, + 0.008935700170695782, + 0.008828366175293922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.02635699324309826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09952579438686371, + 0.08326227217912674, + 0.04395506903529167, + 0.0597095713019371, + 0.18213219940662384, + 0.12829674780368805, + 0.22643807530403137, + 0.024212365970015526, + 0.06635209918022156, + 0.08611583709716797 + ], + "credit_map": [ + 0.017917947843670845, + 0.013332696631550789, + 0.008622937835752964, + 0.007715315092355013, + 0.008093212731182575, + 0.013785242103040218, + 0.01843980886042118, + 0.011618629097938538, + 0.014067408628761768, + 0.009382877498865128, + 0.007822724059224129, + 0.019595490768551826, + 0.0048813652247190475, + 0.0064023383893072605, + 0.013572454452514648, + 0.007537610363215208, + 0.007046780548989773, + 0.006664786487817764, + 0.004591825883835554, + 0.005608361680060625, + 0.004174763336777687, + 0.006370057351887226, + 0.0057793790474534035, + 0.0037191116716712713, + 0.009430368430912495, + 0.015576035715639591, + 0.004664170555770397, + 0.004501987714320421, + 0.0044861542992293835, + 0.006261537317186594, + 0.005494052544236183, + 0.00929525587707758, + 0.05874365195631981, + 0.010530240833759308, + 0.009317911230027676, + 0.006876284722238779, + 0.01297922246158123, + 0.05267832428216934, + 0.017837397754192352, + 0.013169152662158012, + 0.02414766512811184, + 0.014159216545522213, + 0.011524499393999577, + 0.010651549324393272, + 0.01852693222463131, + 0.014448931440711021, + 0.013726145960390568, + 0.02111179567873478, + 0.03835345432162285, + 0.03213798627257347, + 0.014736390672624111, + 0.01902472786605358, + 0.027908263728022575, + 0.019921857863664627, + 0.05160313844680786, + 0.022752247750759125, + 0.002345304237678647, + 0.0024231569841504097, + 0.0035168342292308807, + 0.0027255096938461065, + 0.002575815422460437, + 0.002967953449115157, + 0.0048311264254152775, + 0.0028266645967960358, + 0.006927040405571461, + 0.007897777482867241, + 0.007838516496121883, + 0.006116664968430996, + 0.008215907961130142, + 0.006855062674731016, + 0.013927794061601162, + 0.008573327213525772, + 0.0057057542726397514, + 0.009370517916977406, + 0.006566183641552925, + 0.0069936541840434074, + 0.016441652551293373, + 0.019276510924100876, + 0.010578067973256111, + 0.011183494701981544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.007631078828126192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the main thoroughfare of an abandoned city with crumbling concrete structures and wrecked cars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a heavy, dark gray battle tank on the left that appears damaged, scratched, rusted, and caked with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a streamlined tank painted in desert camouflage on the right that rapidly turns its turret to lock onto the left tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive orange-red flame erupt from the left tank's muzzle, followed by a shell slamming into rubble and causing a large explosion with flying debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a wide-angle view from a low ground angle that slowly pans right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033367812633514 + ], + "opposite_logprob_tokens": [ + -1.9103336334228516 + ], + "logprob_reward": -0.16033367812633514, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033367812633514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19475947320461273, + 0.09052877873182297, + 0.14326490461826324, + 0.06179170310497284, + 0.18064053356647491, + 0.05802445113658905, + 0.07176604866981506, + 0.06144644320011139, + 0.05520298704504967, + 0.08257462829351425 + ], + "credit_map": [ + 0.04363732039928436, + 0.02451195754110813, + 0.01650887168943882, + 0.016511226072907448, + 0.01269114762544632, + 0.02467118576169014, + 0.028948256745934486, + 0.02727951854467392, + 0.015326730906963348, + 0.011245193891227245, + 0.010260934941470623, + 0.01467939279973507, + 0.007391206920146942, + 0.01454954408109188, + 0.009747752919793129, + 0.007328021805733442, + 0.04062933474779129, + 0.014739871956408024, + 0.01641061343252659, + 0.012043491005897522, + 0.010143358260393143, + 0.01857270859181881, + 0.0207953080534935, + 0.009930221363902092, + 0.009341729804873466, + 0.009973317384719849, + 0.005342285614460707, + 0.0060862163081765175, + 0.005793148186057806, + 0.006300915032625198, + 0.00852073822170496, + 0.010433354414999485, + 0.0258279237896204, + 0.019191879779100418, + 0.01846349611878395, + 0.013581009581685066, + 0.027521967887878418, + 0.039763934910297394, + 0.02001609280705452, + 0.016274234279990196, + 0.01033856626600027, + 0.0075766402296721935, + 0.006010481622070074, + 0.005803467705845833, + 0.006234109401702881, + 0.0074538602493703365, + 0.0069046118296682835, + 0.007702711503952742, + 0.013561919331550598, + 0.013237462379038334, + 0.004346711561083794, + 0.005425835959613323, + 0.00808393582701683, + 0.00749664381146431, + 0.011698771268129349, + 0.007914765737950802, + 0.006063379812985659, + 0.006276087835431099, + 0.007710092701017857, + 0.005028350278735161, + 0.006204447243362665, + 0.006442881189286709, + 0.01609225757420063, + 0.007628951221704483, + 0.005982274655252695, + 0.005755605176091194, + 0.004829145036637783, + 0.005794369615614414, + 0.0077587575651705265, + 0.006314074154943228, + 0.01009751670062542, + 0.008671246469020844, + 0.005687319207936525, + 0.005916752386838198, + 0.007006929721683264, + 0.0066892351023852825, + 0.00939592532813549, + 0.021612314507365227, + 0.011720961891114712, + 0.014545188285410404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.16033367812633514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_4/index_tensor([1759, 1759])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_4/index_tensor([1759, 1759])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d3d2ff236ae9977b589ea77b70994c460aa529ba --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_4/index_tensor([1759, 1759])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Rendered in a warm and vibrant Pixar style, this static, eye-level close-up captures a charming small potted plant nestled on a sun-drenched white wooden windowsill. The plant itself is quite expressive, with two large, heart-shaped, lush green leaves and a slender, flexible stem, all housed in a simple, earthy clay pot with a soft, matte finish. A bright shaft of sunlight passes through the scene, illuminating the plant and casting soft, shifting shadows. Visible motes of dust dance magically in the golden light. In a gentle time-lapse sequence, the plant's entire form slowly and gracefully follows the sunlight's path across the windowsill; the stem deliberately bends into smooth arcs, and the leaves lean eagerly towards warmth, a silent, graceful dance of life tracing its source of energy across the room, creating a tranquil and whimsical atmosphere.", + "seed": 29134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_4/index_1759_seed_29134_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010418748715892434, + "vlm_logprob": -0.010418748715892434, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.010418748715892434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01944398134946823 + ], + "opposite_logprob_tokens": [ + -5.269443988800049 + ], + "logprob_reward": -0.01944398134946823, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01944398134946823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09613017737865448, + 0.03433538228273392, + 0.036304816603660583, + 0.08202951401472092, + 0.0586024671792984, + 0.08292035758495331, + 0.11371053755283356, + 0.22450873255729675, + 0.2487981617450714, + 0.022659827023744583 + ], + "credit_map": [ + 0.012818373739719391, + 0.007157814688980579, + 0.022605979815125465, + 0.010146918706595898, + 0.005634536501020193, + 0.013674112036824226, + 0.00919972825795412, + 0.014892712235450745, + 0.003668416989967227, + 0.003050458151847124, + 0.006749527063220739, + 0.004678438417613506, + 0.0025597393978387117, + 0.004355727694928646, + 0.004891842603683472, + 0.00438123382627964, + 0.0034086552914232016, + 0.004106370732188225, + 0.0058522592298686504, + 0.0045639583840966225, + 0.003662098664790392, + 0.006523365154862404, + 0.0038818654138594866, + 0.004306241869926453, + 0.012734740041196346, + 0.011392622254788876, + 0.013249710202217102, + 0.006514195818454027, + 0.010724819265305996, + 0.012596032582223415, + 0.006969014648348093, + 0.00784838106483221, + 0.009418814443051815, + 0.00822996161878109, + 0.009089011698961258, + 0.004839726258069277, + 0.009361947886645794, + 0.007358325645327568, + 0.004206190351396799, + 0.006098492071032524, + 0.00855953898280859, + 0.01674286462366581, + 0.01586064137518406, + 0.006312321871519089, + 0.013265731744468212, + 0.00765443267300725, + 0.00890746247023344, + 0.005617375951260328, + 0.0255568977445364, + 0.006098277401179075, + 0.02146846428513527, + 0.00601522671058774, + 0.008680537343025208, + 0.020757272839546204, + 0.008627746254205704, + 0.016506126150488853, + 0.036703258752822876, + 0.025334030389785767, + 0.03857780992984772, + 0.014190695248544216, + 0.028866315260529518, + 0.038168154656887054, + 0.02365409955382347, + 0.019014397636055946, + 0.04161093384027481, + 0.021524248644709587, + 0.05038103088736534, + 0.016942163929343224, + 0.0303140077739954, + 0.04302823171019554, + 0.02224905788898468, + 0.022748494520783424, + 0.003898821072652936, + 0.0017495009815320373, + 0.0043979366309940815, + 0.0014112620847299695, + 0.002332499483600259, + 0.0043713487684726715, + 0.0020676569547504187, + 0.002430801745504141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01944398134946823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790024720132351 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009790024720132351, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790024720132351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06913077086210251, + 0.047540001571178436, + 0.05618098005652428, + 0.08654677867889404, + 0.05559153109788895, + 0.060084324330091476, + 0.42828232049942017, + 0.14371106028556824, + 0.027174046263098717, + 0.025758245959877968 + ], + "credit_map": [ + 0.010170710273087025, + 0.006041666958481073, + 0.010363847948610783, + 0.01030706986784935, + 0.005098345689475536, + 0.009415199980139732, + 0.008743798360228539, + 0.008990132249891758, + 0.00890276674181223, + 0.002909769769757986, + 0.009126598946750164, + 0.0033803642727434635, + 0.003672450315207243, + 0.0066729942336678505, + 0.0039181336760520935, + 0.008956924080848694, + 0.00877428986132145, + 0.002704010810703039, + 0.011903984472155571, + 0.005295397713780403, + 0.0036053925286978483, + 0.007576006464660168, + 0.00460635544732213, + 0.011715542525053024, + 0.01327551156282425, + 0.00716682942584157, + 0.011151487939059734, + 0.009893583133816719, + 0.00960645079612732, + 0.012291541323065758, + 0.01199526246637106, + 0.011166106909513474, + 0.011438362300395966, + 0.00605440977960825, + 0.008907451294362545, + 0.0033546278718858957, + 0.006447197403758764, + 0.010718575678765774, + 0.0037271033506840467, + 0.004943810403347015, + 0.007113486994057894, + 0.008054595440626144, + 0.008495384827256203, + 0.007541398983448744, + 0.00693431356921792, + 0.007345228455960751, + 0.007312330882996321, + 0.007287583313882351, + 0.09282846748828888, + 0.01886162720620632, + 0.08584192395210266, + 0.01926395297050476, + 0.03218219429254532, + 0.08545207977294922, + 0.032576777040958405, + 0.061275284737348557, + 0.02694535441696644, + 0.007826314307749271, + 0.018877249211072922, + 0.014756388030946255, + 0.012998963706195354, + 0.024039123207330704, + 0.017779236659407616, + 0.020488422363996506, + 0.0030607422813773155, + 0.002064584754407406, + 0.00406019389629364, + 0.0026122184935957193, + 0.001990528078749776, + 0.003512751078233123, + 0.0033559047151356936, + 0.006517123430967331, + 0.003984380047768354, + 0.0024151289835572243, + 0.0050976695492863655, + 0.001907434780150652, + 0.0026314114220440388, + 0.0036516794934868813, + 0.002529418794438243, + 0.0035411224234849215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.009790024720132351 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024915861431509256 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024915861431509256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024915861431509256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07327508926391602, + 0.04326678439974785, + 0.057241808623075485, + 0.11655101925134659, + 0.08199500292539597, + 0.11089768260717392, + 0.08291326463222504, + 0.19057711958885193, + 0.1638627052307129, + 0.07941953837871552 + ], + "credit_map": [ + 0.011530978605151176, + 0.0057931384071707726, + 0.01060565933585167, + 0.00963530968874693, + 0.006914732977747917, + 0.010490195825695992, + 0.009435280226171017, + 0.008869794197380543, + 0.006777690723538399, + 0.005194562487304211, + 0.0048170858062803745, + 0.004640595521777868, + 0.006947760004550219, + 0.004388181027024984, + 0.0053234463557600975, + 0.005177463870495558, + 0.006723694037646055, + 0.004768728744238615, + 0.008030561730265617, + 0.005913745611906052, + 0.006923688109964132, + 0.008064334280788898, + 0.009006245993077755, + 0.007810811512172222, + 0.014447558671236038, + 0.007486230693757534, + 0.021383559331297874, + 0.011118743568658829, + 0.010095889680087566, + 0.02048656716942787, + 0.011715080589056015, + 0.019817393273115158, + 0.011003871448338032, + 0.010115488432347775, + 0.013794904574751854, + 0.009445673786103725, + 0.01123163290321827, + 0.01025357749313116, + 0.007723063696175814, + 0.00842679850757122, + 0.01264095213264227, + 0.019742079079151154, + 0.01823224499821663, + 0.010253623127937317, + 0.015241737477481365, + 0.01336440909653902, + 0.010150149464607239, + 0.011272484436631203, + 0.016367999836802483, + 0.007187042385339737, + 0.015403833240270615, + 0.0047508892603218555, + 0.007147317286580801, + 0.014513755217194557, + 0.007533395662903786, + 0.01000902894884348, + 0.03269562870264053, + 0.022582119330763817, + 0.03245947137475014, + 0.011633509770035744, + 0.026131419464945793, + 0.030978232622146606, + 0.0173733439296484, + 0.0167233943939209, + 0.026030655950307846, + 0.017010916024446487, + 0.029267821460962296, + 0.01318388432264328, + 0.02227172814309597, + 0.024853955954313278, + 0.01694527268409729, + 0.014298477210104465, + 0.01567673496901989, + 0.007043061312288046, + 0.01095220074057579, + 0.009068272076547146, + 0.009074977599084377, + 0.011229136027395725, + 0.009175625629723072, + 0.007199530489742756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.0024915861431509256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09886132925748825, + 0.027198124676942825, + 0.04762954264879227, + 0.19056721031665802, + 0.046337615698575974, + 0.1571040004491806, + 0.0968572199344635, + 0.17434990406036377, + 0.13159723579883575, + 0.029497768729925156 + ], + "credit_map": [ + 0.013402056880295277, + 0.008406492881476879, + 0.013037191703915596, + 0.014638570137321949, + 0.00793064758181572, + 0.013418668881058693, + 0.01803484745323658, + 0.009992865845561028, + 0.00404792744666338, + 0.0020902121905237436, + 0.005205772817134857, + 0.003674472449347377, + 0.0025001224130392075, + 0.003220881102606654, + 0.0028581363148987293, + 0.0036006025038659573, + 0.006256334949284792, + 0.0036106896586716175, + 0.007771789561957121, + 0.004461098928004503, + 0.0044548846781253815, + 0.006638178601861, + 0.006526149809360504, + 0.007910422049462795, + 0.028130289167165756, + 0.016164585947990417, + 0.028480468317866325, + 0.020269252359867096, + 0.019343402236700058, + 0.026248440146446228, + 0.018882684409618378, + 0.03304808959364891, + 0.00914897583425045, + 0.0031639947555959225, + 0.00803380087018013, + 0.0035805662628263235, + 0.0049849278293550014, + 0.008060498163104057, + 0.0031037279404699802, + 0.006261127069592476, + 0.016277238726615906, + 0.026037583127617836, + 0.02942788228392601, + 0.013865355402231216, + 0.021246112883090973, + 0.01628572680056095, + 0.015666838735342026, + 0.018297255039215088, + 0.019803332164883614, + 0.007151065859943628, + 0.02033044770359993, + 0.005839311517775059, + 0.007642343174666166, + 0.01749218814074993, + 0.007520710118114948, + 0.011077824048697948, + 0.029221046715974808, + 0.02223646640777588, + 0.029133016243577003, + 0.011099249124526978, + 0.02448984794318676, + 0.02697283774614334, + 0.017039451748132706, + 0.014158015139400959, + 0.022432686761021614, + 0.013456563465297222, + 0.02250310406088829, + 0.010733428411185741, + 0.018231527879834175, + 0.019765451550483704, + 0.011973056942224503, + 0.012501433491706848, + 0.0064764684066176414, + 0.00231185439042747, + 0.004129223525524139, + 0.002495827153325081, + 0.0024959025904536247, + 0.004737799055874348, + 0.002853749319911003, + 0.00399694824591279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.002197353169322014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817079819738865 + ], + "opposite_logprob_tokens": [ + -4.0181708335876465 + ], + "logprob_reward": -0.01817079819738865, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817079819738865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0626237690448761, + 0.026228157803416252, + 0.07972507178783417, + 0.21524390578269958, + 0.06093538552522659, + 0.2154405266046524, + 0.09919632971286774, + 0.058973278850317, + 0.1049484834074974, + 0.07668504118919373 + ], + "credit_map": [ + 0.009498209692537785, + 0.006794528104364872, + 0.006752586457878351, + 0.008530310355126858, + 0.007633513770997524, + 0.008942918851971626, + 0.007489881943911314, + 0.006981819402426481, + 0.003686748445034027, + 0.002528125885874033, + 0.0031603702809661627, + 0.0028170377481728792, + 0.0037226909771561623, + 0.003068734658882022, + 0.0030722555238753557, + 0.004172196611762047, + 0.009534458629786968, + 0.006134992931038141, + 0.015527665615081787, + 0.008317883126437664, + 0.006451766472309828, + 0.0143373291939497, + 0.008223086595535278, + 0.011197899468243122, + 0.023907838389277458, + 0.020073195919394493, + 0.037133194506168365, + 0.024778170511126518, + 0.025686204433441162, + 0.03230992332100868, + 0.022328881546854973, + 0.02902652882039547, + 0.009940104559063911, + 0.007466540206223726, + 0.00910938810557127, + 0.00689099682494998, + 0.011262165382504463, + 0.006965592037886381, + 0.004225786309689283, + 0.005074811168015003, + 0.02470448613166809, + 0.038311418145895004, + 0.03242964297533035, + 0.01819792576134205, + 0.02874196693301201, + 0.02482122741639614, + 0.02544303983449936, + 0.02279084362089634, + 0.019389847293496132, + 0.011149458587169647, + 0.013950088061392307, + 0.006529590580612421, + 0.010799788869917393, + 0.014405397698283195, + 0.010146135464310646, + 0.012826027348637581, + 0.006490566302090883, + 0.0059234038926661015, + 0.006875135935842991, + 0.006329579744488001, + 0.008958847261965275, + 0.007992586120963097, + 0.009115064516663551, + 0.007288097869604826, + 0.020279934629797935, + 0.011437789537012577, + 0.016002126038074493, + 0.0073688137345016, + 0.010166386142373085, + 0.01807541772723198, + 0.008473648689687252, + 0.01314436737447977, + 0.01682024635374546, + 0.006724062841385603, + 0.011855250224471092, + 0.006326600443571806, + 0.008522484451532364, + 0.012885711155831814, + 0.0065645189024508, + 0.006986171472817659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.01817079819738865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a warm and vibrant Pixar style, this static, eye-level close-up captures a charming small potted plant nestled on a sun-drenched white wooden windowsill. The plant itself is quite expressive, with two large, heart-shaped, lush green leaves and a slender, flexible stem, all housed in a simple, earthy clay pot with a soft, matte finish. A bright shaft of sunlight passes through the scene, illuminating the plant and casting soft, shifting shadows. Visible motes of dust dance magically in the golden light. In a gentle time-lapse sequence, the plant's entire form slowly and gracefully follows the sunlight's path across the windowsill; the stem deliberately bends into smooth arcs, and the leaves lean eagerly towards warmth, a silent, graceful dance of life tracing its source of energy across the room, creating a tranquil and whimsical atmosphere.", + "seed": 29135, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_4/index_1759_seed_29135_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008764861547388137, + "vlm_logprob": -0.008764861547388137, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.008764861547388137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005324112717062235 + ], + "opposite_logprob_tokens": [ + -6.380323886871338 + ], + "logprob_reward": -0.005324112717062235, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.005324112717062235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2076391726732254, + 0.09425701200962067, + 0.10441756248474121, + 0.026578256860375404, + 0.029444748535752296, + 0.2721790075302124, + 0.11113445460796356, + 0.06878933310508728, + 0.06022297590970993, + 0.025337494909763336 + ], + "credit_map": [ + 0.024481389671564102, + 0.026078933849930763, + 0.026849862188100815, + 0.010554053820669651, + 0.04383373633027077, + 0.02297259122133255, + 0.02512708306312561, + 0.02774151973426342, + 0.01283062994480133, + 0.01629488915205002, + 0.005619360599666834, + 0.007970458827912807, + 0.015017452649772167, + 0.013117791153490543, + 0.016596172004938126, + 0.006810259073972702, + 0.012995224446058273, + 0.010917291976511478, + 0.013786214403808117, + 0.008634481579065323, + 0.014760879799723625, + 0.012731542810797691, + 0.009487219154834747, + 0.021104712039232254, + 0.0034165417309850454, + 0.003300759708508849, + 0.0031129305716603994, + 0.002371191745623946, + 0.002751935739070177, + 0.003364955075085163, + 0.005438288673758507, + 0.0028216547798365355, + 0.005756198428571224, + 0.003832595655694604, + 0.0022248427849262953, + 0.002044140128418803, + 0.004428037907928228, + 0.0039732432924211025, + 0.003942946903407574, + 0.0032427445985376835, + 0.03855328634381294, + 0.037554871290922165, + 0.018624700605869293, + 0.017479950562119484, + 0.05432518944144249, + 0.03452982008457184, + 0.03306964412331581, + 0.038041554391384125, + 0.015705781057476997, + 0.014787680469453335, + 0.011938588693737984, + 0.007833288982510567, + 0.01275577675551176, + 0.014289045706391335, + 0.011739627458155155, + 0.022084666416049004, + 0.009297274984419346, + 0.009999698959290981, + 0.004107931163161993, + 0.00581471249461174, + 0.013702047988772392, + 0.010822374373674393, + 0.009085321798920631, + 0.005959971342235804, + 0.009926198050379753, + 0.0062662758864462376, + 0.005708744283765554, + 0.004791054408997297, + 0.007539345882833004, + 0.010954450815916061, + 0.008897187188267708, + 0.006139715667814016, + 0.004118101205676794, + 0.00451138848438859, + 0.0023619888816028833, + 0.002208019606769085, + 0.003092017490416765, + 0.003998489584773779, + 0.0030617539305239916, + 0.0019857354927808046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.005324112717062235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645604364573956 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645604364573956, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645604364573956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12888678908348083, + 0.11675207316875458, + 0.23554059863090515, + 0.04994061961770058, + 0.04634135961532593, + 0.17561037838459015, + 0.15294678509235382, + 0.039477068930864334, + 0.023859292268753052, + 0.030645018443465233 + ], + "credit_map": [ + 0.018084771931171417, + 0.015437768772244453, + 0.011553112417459488, + 0.010561004281044006, + 0.024666568264365196, + 0.015951942652463913, + 0.020483920350670815, + 0.012147699482738972, + 0.01662488654255867, + 0.020137103274464607, + 0.00749542098492384, + 0.009745814837515354, + 0.018796641379594803, + 0.015530680306255817, + 0.018432417884469032, + 0.009989100508391857, + 0.03449924290180206, + 0.021309947595000267, + 0.023238319903612137, + 0.01880311220884323, + 0.04116837680339813, + 0.030549677088856697, + 0.035472821444272995, + 0.03049909882247448, + 0.007822026498615742, + 0.007117451634258032, + 0.005321127828210592, + 0.003962927497923374, + 0.006538801360875368, + 0.007029638160020113, + 0.006499167066067457, + 0.005649477243423462, + 0.008777488023042679, + 0.00472300685942173, + 0.0041798874735832214, + 0.003244402352720499, + 0.003955696243792772, + 0.006945973727852106, + 0.008245822973549366, + 0.006269080098718405, + 0.02463233843445778, + 0.02571779303252697, + 0.012162701226770878, + 0.010817456990480423, + 0.035092104226350784, + 0.022688740864396095, + 0.019952965900301933, + 0.024546287953853607, + 0.01625991240143776, + 0.02395392395555973, + 0.007489703595638275, + 0.013167471624910831, + 0.03385554999113083, + 0.025176309049129486, + 0.022870440036058426, + 0.010173476301133633, + 0.005722173023968935, + 0.004886130336672068, + 0.002960687503218651, + 0.0033814632333815098, + 0.004617741331458092, + 0.005907286889851093, + 0.00808743666857481, + 0.0039141494780778885, + 0.0031675559002906084, + 0.0035125308204442263, + 0.0017196746775880456, + 0.0016764008905738592, + 0.0030536502599716187, + 0.0034473161213099957, + 0.0042311931028962135, + 0.003050972241908312, + 0.005854126065969467, + 0.0025412542745471, + 0.0041829925030469894, + 0.0022364058531820774, + 0.002284357091411948, + 0.0039554862305521965, + 0.0035963424015790224, + 0.005994052160531282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.008645604364573956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021977098658680916 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021977098658680916, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021977098658680916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11578033119440079, + 0.21565471589565277, + 0.17143864929676056, + 0.03612665832042694, + 0.034666482359170914, + 0.20596227049827576, + 0.09184886515140533, + 0.07480719685554504, + 0.015249147079885006, + 0.038465715944767 + ], + "credit_map": [ + 0.020648721605539322, + 0.009417335502803326, + 0.015245539136230946, + 0.007598863448947668, + 0.016998276114463806, + 0.014949973672628403, + 0.010272005572915077, + 0.020649611949920654, + 0.03059040755033493, + 0.03221682086586952, + 0.01834305189549923, + 0.01947481743991375, + 0.02956313267350197, + 0.02769087441265583, + 0.033650897443294525, + 0.024124709889292717, + 0.02121225744485855, + 0.01927247829735279, + 0.010671459138393402, + 0.016907937824726105, + 0.03928612917661667, + 0.023871131241321564, + 0.02808237634599209, + 0.012134856544435024, + 0.004433447495102882, + 0.004855731502175331, + 0.003845058847218752, + 0.002976137213408947, + 0.005359791684895754, + 0.005022463388741016, + 0.005282049532979727, + 0.004351973999291658, + 0.004314641933888197, + 0.004659966100007296, + 0.0023439291398972273, + 0.002910877112299204, + 0.004464291501790285, + 0.004685983061790466, + 0.008412613533437252, + 0.0028741764836013317, + 0.027495916932821274, + 0.03741378337144852, + 0.012090621516108513, + 0.013044276274740696, + 0.04306848719716072, + 0.027790460735559464, + 0.022170599550008774, + 0.02288810722529888, + 0.01119279209524393, + 0.011532248929142952, + 0.0074482327327132225, + 0.006178839597851038, + 0.016737567260861397, + 0.014114920981228352, + 0.010906374081969261, + 0.013737880624830723, + 0.011594036594033241, + 0.009158127009868622, + 0.004263156559318304, + 0.006842040456831455, + 0.013268829323351383, + 0.012035202234983444, + 0.010519876144826412, + 0.007125921081751585, + 0.002166661899536848, + 0.0016525008250027895, + 0.0013508506817743182, + 0.0012910618679597974, + 0.002657815581187606, + 0.002242576563730836, + 0.0019275470403954387, + 0.001960130874067545, + 0.005608388688415289, + 0.007633542641997337, + 0.0034008133225142956, + 0.002978837350383401, + 0.004671631380915642, + 0.006022760644555092, + 0.00494763907045126, + 0.003202099585905671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.0021977098658680916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334372558631003 + ], + "opposite_logprob_tokens": [ + -6.626334190368652 + ], + "logprob_reward": -0.001334372558631003, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334372558631003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1572057604789734, + 0.024031145498156548, + 0.13186176121234894, + 0.0326344259083271, + 0.10025639832019806, + 0.4215242266654968, + 0.04152100905776024, + 0.034115880727767944, + 0.026188252493739128, + 0.03066117689013481 + ], + "credit_map": [ + 0.023018963634967804, + 0.01795976795256138, + 0.01205651555210352, + 0.012960890308022499, + 0.028113698586821556, + 0.023243937641382217, + 0.02556563727557659, + 0.0142863430082798, + 0.003514823503792286, + 0.003245594445616007, + 0.0020063866395503283, + 0.0026241105515509844, + 0.0033259254414588213, + 0.0031450719106942415, + 0.00417400011792779, + 0.0019952307920902967, + 0.01658630184829235, + 0.016647813841700554, + 0.008619261905550957, + 0.012222007848322392, + 0.03093656152486801, + 0.019176507368683815, + 0.019347073510289192, + 0.008326233364641666, + 0.00434183469042182, + 0.004123656079173088, + 0.003239299403503537, + 0.002710987813770771, + 0.004412515088915825, + 0.004255109932273626, + 0.006132436450570822, + 0.0034185852855443954, + 0.013681503012776375, + 0.015560552477836609, + 0.008211895823478699, + 0.00816203374415636, + 0.012970879673957825, + 0.013048545457422733, + 0.0188481155782938, + 0.009772876277565956, + 0.05756460130214691, + 0.06188337877392769, + 0.0289485864341259, + 0.026963574811816216, + 0.08647783100605011, + 0.05357661843299866, + 0.04786724969744682, + 0.058242376893758774, + 0.005112355574965477, + 0.007236001547425985, + 0.003183940192684531, + 0.003310112515464425, + 0.007125479634851217, + 0.006541529670357704, + 0.005581269972026348, + 0.0034303213469684124, + 0.005174931604415178, + 0.004280399531126022, + 0.002653384581208229, + 0.002953220857307315, + 0.00518606835976243, + 0.004541357513517141, + 0.004755335859954357, + 0.0045711807906627655, + 0.003377399640157819, + 0.004859211854636669, + 0.0020310129038989544, + 0.0023238046560436487, + 0.003497462021186948, + 0.0037412140518426895, + 0.003649434307590127, + 0.0027087146881967783, + 0.0037287231534719467, + 0.007297206670045853, + 0.001847487292252481, + 0.0024915318936109543, + 0.00415824307128787, + 0.004598742350935936, + 0.004631057847291231, + 0.0019081821665167809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.001334372558631003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026322508230805397 + ], + "opposite_logprob_tokens": [ + -3.651322603225708 + ], + "logprob_reward": -0.026322508230805397, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026322508230805397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20952171087265015, + 0.33418506383895874, + 0.04771646484732628, + 0.02485354244709015, + 0.02401064522564411, + 0.13966669142246246, + 0.055179961025714874, + 0.0390719398856163, + 0.07350373268127441, + 0.05229026451706886 + ], + "credit_map": [ + 0.025009486824274063, + 0.03478560224175453, + 0.012634438462555408, + 0.017963280901312828, + 0.047166693955659866, + 0.02846381813287735, + 0.029652196913957596, + 0.013846202753484249, + 0.047664493322372437, + 0.05328558757901192, + 0.027863113209605217, + 0.03167356923222542, + 0.049903541803359985, + 0.04298892244696617, + 0.04484134167432785, + 0.0359644740819931, + 0.006096523255109787, + 0.007158198859542608, + 0.0029870704747736454, + 0.004472596570849419, + 0.007003682665526867, + 0.00700259767472744, + 0.00888855941593647, + 0.004107239656150341, + 0.0036977974232286215, + 0.003906347788870335, + 0.0012635962339118123, + 0.002457442693412304, + 0.0039689624682068825, + 0.00376152410171926, + 0.004071233328431845, + 0.0017266396898776293, + 0.003262838814407587, + 0.00319877453148365, + 0.0017730953404679894, + 0.0023452285677194595, + 0.0038768220692873, + 0.0034885709173977375, + 0.0039847251027822495, + 0.0020805911626666784, + 0.018422957509756088, + 0.022606566548347473, + 0.010312581434845924, + 0.00841356161981821, + 0.025746244937181473, + 0.01864713802933693, + 0.015603563748300076, + 0.019914088770747185, + 0.008541727438569069, + 0.008480488322675228, + 0.00502641499042511, + 0.0033629611134529114, + 0.007800902705639601, + 0.007941738702356815, + 0.005763541907072067, + 0.008262190036475658, + 0.005932731553912163, + 0.004647482186555862, + 0.0029395066667348146, + 0.0036810673773288727, + 0.006224814802408218, + 0.005224544554948807, + 0.005756382364779711, + 0.0046654134057462215, + 0.009760920889675617, + 0.014531003311276436, + 0.005314106121659279, + 0.006420735735446215, + 0.009744065813720226, + 0.011141124181449413, + 0.010252011008560658, + 0.006339767016470432, + 0.006193866021931171, + 0.012602894566953182, + 0.002701524645090103, + 0.004289623349905014, + 0.007252557668834925, + 0.008136370219290257, + 0.007596824783831835, + 0.0035166058223694563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.026322508230805397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a warm and vibrant Pixar style, this static, eye-level close-up captures a charming small potted plant nestled on a sun-drenched white wooden windowsill. The plant itself is quite expressive, with two large, heart-shaped, lush green leaves and a slender, flexible stem, all housed in a simple, earthy clay pot with a soft, matte finish. A bright shaft of sunlight passes through the scene, illuminating the plant and casting soft, shifting shadows. Visible motes of dust dance magically in the golden light. In a gentle time-lapse sequence, the plant's entire form slowly and gracefully follows the sunlight's path across the windowsill; the stem deliberately bends into smooth arcs, and the leaves lean eagerly towards warmth, a silent, graceful dance of life tracing its source of energy across the room, creating a tranquil and whimsical atmosphere.", + "seed": 29136, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_4/index_1759_seed_29136_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01637766098137945, + "vlm_logprob": -0.01637766098137945, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.01637766098137945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01441987231373787 + ], + "opposite_logprob_tokens": [ + -5.389420032501221 + ], + "logprob_reward": -0.01441987231373787, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01441987231373787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21680252254009247, + 0.07008256018161774, + 0.06785563379526138, + 0.1739102303981781, + 0.09971656650304794, + 0.049963727593421936, + 0.04456927627325058, + 0.04496293142437935, + 0.130342036485672, + 0.1017945408821106 + ], + "credit_map": [ + 0.02027776837348938, + 0.013802463188767433, + 0.056173790246248245, + 0.019710326567292213, + 0.013554828241467476, + 0.018212970346212387, + 0.020558154210448265, + 0.05451222136616707, + 0.009448965080082417, + 0.007452524267137051, + 0.007991628721356392, + 0.009156165644526482, + 0.008274474181234837, + 0.008872446604073048, + 0.008492400869727135, + 0.010393951088190079, + 0.005725964903831482, + 0.007252111565321684, + 0.00839650072157383, + 0.007727236486971378, + 0.011469346471130848, + 0.008102674037218094, + 0.009106656536459923, + 0.010075145401060581, + 0.025942090898752213, + 0.014126165769994259, + 0.029761161655187607, + 0.017785657197237015, + 0.022655371576547623, + 0.01702454872429371, + 0.02109220251441002, + 0.025523023679852486, + 0.012828825041651726, + 0.00850022304803133, + 0.012188533321022987, + 0.012636691331863403, + 0.010393503122031689, + 0.012119313701987267, + 0.012670954689383507, + 0.018378525972366333, + 0.004716240335255861, + 0.003984901588410139, + 0.00812966376543045, + 0.006235754117369652, + 0.0049331821501255035, + 0.00743707362562418, + 0.006331876385957003, + 0.008195038884878159, + 0.005789660383015871, + 0.004558920860290527, + 0.0050218114629387856, + 0.004969804547727108, + 0.005312433931976557, + 0.007190136704593897, + 0.005997168365865946, + 0.005729337688535452, + 0.006599822547286749, + 0.003658575937151909, + 0.006881658919155598, + 0.0038910761941224337, + 0.005430877208709717, + 0.0061413017101585865, + 0.004824490752071142, + 0.007535128388553858, + 0.02503211423754692, + 0.010562059469521046, + 0.018241707235574722, + 0.011122161522507668, + 0.013836211524903774, + 0.023260312154889107, + 0.012107934802770615, + 0.016179531812667847, + 0.009149891324341297, + 0.013142530806362629, + 0.02008332498371601, + 0.008882247842848301, + 0.017272230237722397, + 0.006816309876739979, + 0.010913162492215633, + 0.01553483959287405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.01441987231373787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076378220692276955 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.0076378220692276955, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076378220692276955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15732161700725555, + 0.09391392767429352, + 0.06835789978504181, + 0.057629890739917755, + 0.08225230872631073, + 0.03576897457242012, + 0.056476444005966187, + 0.07912664115428925, + 0.12569040060043335, + 0.24346189200878143 + ], + "credit_map": [ + 0.03315074369311333, + 0.010915782302618027, + 0.025807036086916924, + 0.018383299931883812, + 0.008935236372053623, + 0.014553068205714226, + 0.018020955845713615, + 0.02755548804998398, + 0.012957989238202572, + 0.00824419129639864, + 0.016315490007400513, + 0.011675357818603516, + 0.009005439467728138, + 0.010559950023889542, + 0.010720858350396156, + 0.014434651471674442, + 0.007049220614135265, + 0.006208047270774841, + 0.010525422170758247, + 0.007975658401846886, + 0.00795674230903387, + 0.009866531938314438, + 0.007521658204495907, + 0.01125461794435978, + 0.006832598242908716, + 0.006264315452426672, + 0.008906089700758457, + 0.00716708367690444, + 0.005446832627058029, + 0.00744806881994009, + 0.0065100244246423244, + 0.009054882451891899, + 0.016928832978010178, + 0.0041091120801866055, + 0.013881256803870201, + 0.010246718302369118, + 0.005545950494706631, + 0.005007067229598761, + 0.008741681464016438, + 0.017791682854294777, + 0.007113305851817131, + 0.002290412550792098, + 0.005934091284871101, + 0.00400887057185173, + 0.003306324128061533, + 0.0025703648570924997, + 0.0037356431130319834, + 0.006809963844716549, + 0.009069335646927357, + 0.005678720772266388, + 0.007274959702044725, + 0.005101032555103302, + 0.005875920411199331, + 0.00873174425214529, + 0.006536796689033508, + 0.008207934908568859, + 0.014219984412193298, + 0.006177305243909359, + 0.011599499732255936, + 0.0076116095297038555, + 0.007331499829888344, + 0.0096324747428298, + 0.007516431622207165, + 0.01503783743828535, + 0.021947773173451424, + 0.010640640743076801, + 0.01967880129814148, + 0.009101619943976402, + 0.012414769269526005, + 0.0240665003657341, + 0.009513992816209793, + 0.018326302990317345, + 0.030797287821769714, + 0.02007051184773445, + 0.04936273768544197, + 0.019412774592638016, + 0.024032317101955414, + 0.03643712028861046, + 0.024930773302912712, + 0.03841838240623474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0076378220692276955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017157370457425714 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.0017157370457425714, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017157370457425714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08259209245443344, + 0.05106508731842041, + 0.1638840138912201, + 0.11830353736877441, + 0.08273494243621826, + 0.026876935735344887, + 0.03499385714530945, + 0.09891220182180405, + 0.2766491770744324, + 0.06398814171552658 + ], + "credit_map": [ + 0.010797698982059956, + 0.007317947689443827, + 0.013777815736830235, + 0.008220583200454712, + 0.009226650930941105, + 0.008970815688371658, + 0.010054253041744232, + 0.014226331375539303, + 0.005644756834954023, + 0.00449506938457489, + 0.007576045114547014, + 0.005754171404987574, + 0.007193232420831919, + 0.0054248603992164135, + 0.007556277792900801, + 0.007420679088681936, + 0.01815597154200077, + 0.019828643649816513, + 0.018996059894561768, + 0.016075745224952698, + 0.029911762103438377, + 0.02033473178744316, + 0.018621066585183144, + 0.021960046142339706, + 0.01835034415125847, + 0.012825566343963146, + 0.01626972295343876, + 0.014271879568696022, + 0.010618235915899277, + 0.014464336447417736, + 0.012433021329343319, + 0.01907043531537056, + 0.011020754463970661, + 0.007374017499387264, + 0.014726165682077408, + 0.007859131321310997, + 0.010034458711743355, + 0.008154460228979588, + 0.009761295281350613, + 0.013804662972688675, + 0.002806943841278553, + 0.002611083909869194, + 0.004071540664881468, + 0.0024342145770788193, + 0.002244128379970789, + 0.005047506652772427, + 0.0025861035101115704, + 0.005075415596365929, + 0.004808501340448856, + 0.002965394174680114, + 0.005193447694182396, + 0.0036125893238931894, + 0.00396031653508544, + 0.005184030160307884, + 0.004433629103004932, + 0.004835954401642084, + 0.013929953798651695, + 0.008277338929474354, + 0.012992721982300282, + 0.008979463018476963, + 0.011228286661207676, + 0.01983468420803547, + 0.010382319800555706, + 0.013287450186908245, + 0.0510733388364315, + 0.02695315144956112, + 0.04351380839943886, + 0.014756808057427406, + 0.02744903229176998, + 0.05412597581744194, + 0.019102001562714577, + 0.039675116539001465, + 0.005802197381854057, + 0.007501024752855301, + 0.013168955221772194, + 0.004863635636866093, + 0.011993151158094406, + 0.004220413044095039, + 0.006089299451559782, + 0.010349476709961891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.0017157370457425714 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191380063071847 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.003191380063071847, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191380063071847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22772017121315002, + 0.0942930281162262, + 0.0642191469669342, + 0.1674610823392868, + 0.08369753509759903, + 0.036161623895168304, + 0.0670323520898819, + 0.044127125293016434, + 0.1680556684732437, + 0.047232307493686676 + ], + "credit_map": [ + 0.028072526678442955, + 0.02233663760125637, + 0.03699968382716179, + 0.026300271973013878, + 0.019090717658400536, + 0.03187739476561546, + 0.029641272500157356, + 0.033401668071746826, + 0.009714290499687195, + 0.009542617946863174, + 0.013004744425415993, + 0.011281573213636875, + 0.009564929641783237, + 0.01831943541765213, + 0.010838397778570652, + 0.01202703919261694, + 0.007006227970123291, + 0.007736952975392342, + 0.009103978052735329, + 0.007091641891747713, + 0.008286760188639164, + 0.009987544268369675, + 0.0065040686167776585, + 0.008501973003149033, + 0.024024251848459244, + 0.015068982727825642, + 0.024496030062437057, + 0.017006289213895798, + 0.024530746042728424, + 0.018437614664435387, + 0.019768662750720978, + 0.024128513410687447, + 0.01229892484843731, + 0.007715029641985893, + 0.012699658051133156, + 0.008979346603155136, + 0.010622546076774597, + 0.007808954454958439, + 0.010297107510268688, + 0.013275965116918087, + 0.0038370504043996334, + 0.002253724727779627, + 0.008686922490596771, + 0.002435076516121626, + 0.004090661182999611, + 0.0036279575433582067, + 0.003183045657351613, + 0.008047182112932205, + 0.009554091840982437, + 0.005422655493021011, + 0.008175854571163654, + 0.008724438957870007, + 0.008041968569159508, + 0.008745516650378704, + 0.009122217074036598, + 0.009245605207979679, + 0.005856059025973082, + 0.0040177893824875355, + 0.006456159520894289, + 0.0034508861135691404, + 0.004698212258517742, + 0.008572662249207497, + 0.004462824668735266, + 0.00661253510043025, + 0.02101782336831093, + 0.030987704172730446, + 0.024654146283864975, + 0.014034399762749672, + 0.022816604003310204, + 0.0214220080524683, + 0.01410534419119358, + 0.019017646089196205, + 0.0038391791749745607, + 0.007912088185548782, + 0.007366654928773642, + 0.005969969555735588, + 0.005837532691657543, + 0.004358251579105854, + 0.005136911291629076, + 0.006811721250414848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.003191380063071847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054923493415117264 + ], + "opposite_logprob_tokens": [ + -2.9299235343933105 + ], + "logprob_reward": -0.054923493415117264, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054923493415117264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13468235731124878, + 0.09515725821256638, + 0.07449579238891602, + 0.10332843661308289, + 0.052262403070926666, + 0.04666512459516525, + 0.10453289747238159, + 0.1153101697564125, + 0.21328124403953552, + 0.06028429791331291 + ], + "credit_map": [ + 0.018225423991680145, + 0.013947521336376667, + 0.020533496513962746, + 0.012208660133183002, + 0.012832163833081722, + 0.022670214995741844, + 0.015499236062169075, + 0.018765637651085854, + 0.010068733245134354, + 0.008828166872262955, + 0.014720450155436993, + 0.012436426244676113, + 0.008917462080717087, + 0.014366204850375652, + 0.012255046516656876, + 0.0135647626593709, + 0.005843088496476412, + 0.008358209393918514, + 0.009302773512899876, + 0.006619150284677744, + 0.017816076055169106, + 0.008448323234915733, + 0.009102261625230312, + 0.009005911648273468, + 0.012540568597614765, + 0.0111623490229249, + 0.01770642399787903, + 0.011324095539748669, + 0.009935936890542507, + 0.012612308375537395, + 0.010409446433186531, + 0.017637301236391068, + 0.007435075007379055, + 0.005919693969190121, + 0.006119154393672943, + 0.0048110997304320335, + 0.006700385361909866, + 0.008013159967958927, + 0.0068597677163779736, + 0.006404071114957333, + 0.004928839858621359, + 0.004569531884044409, + 0.00708329863846302, + 0.004196557216346264, + 0.005091462749987841, + 0.008576713502407074, + 0.004561718553304672, + 0.007657003588974476, + 0.01694144681096077, + 0.008185157552361488, + 0.010963019914925098, + 0.011555854231119156, + 0.014290116727352142, + 0.01477798167616129, + 0.01298753172159195, + 0.0148317851126194, + 0.016476023942232132, + 0.010020343586802483, + 0.015970200300216675, + 0.010610361583530903, + 0.012456582859158516, + 0.02157268300652504, + 0.011655920185148716, + 0.016548050567507744, + 0.026850255206227303, + 0.04617966711521149, + 0.02633589133620262, + 0.020607080310583115, + 0.03079202026128769, + 0.017017222940921783, + 0.02427293173968792, + 0.0212261825799942, + 0.006154920905828476, + 0.01293193083256483, + 0.008072538301348686, + 0.00563377421349287, + 0.009927691891789436, + 0.004610639065504074, + 0.006295704748481512, + 0.0066570984199643135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.054923493415117264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a warm and vibrant Pixar style, this static, eye-level close-up captures a charming small potted plant nestled on a sun-drenched white wooden windowsill. The plant itself is quite expressive, with two large, heart-shaped, lush green leaves and a slender, flexible stem, all housed in a simple, earthy clay pot with a soft, matte finish. A bright shaft of sunlight passes through the scene, illuminating the plant and casting soft, shifting shadows. Visible motes of dust dance magically in the golden light. In a gentle time-lapse sequence, the plant's entire form slowly and gracefully follows the sunlight's path across the windowsill; the stem deliberately bends into smooth arcs, and the leaves lean eagerly towards warmth, a silent, graceful dance of life tracing its source of energy across the room, creating a tranquil and whimsical atmosphere.", + "seed": 29137, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_4/index_1759_seed_29137_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05454064980149269, + "vlm_logprob": -0.05454064980149269, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.05454064980149269, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015241818502545357 + ], + "opposite_logprob_tokens": [ + -4.890241622924805 + ], + "logprob_reward": -0.015241818502545357, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015241818502545357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09859662503004074, + 0.2788228392601013, + 0.11739352345466614, + 0.09196080267429352, + 0.04101335629820824, + 0.05958271771669388, + 0.05502913519740105, + 0.04491490125656128, + 0.061520397663116455, + 0.15116572380065918 + ], + "credit_map": [ + 0.012406318448483944, + 0.017292344942688942, + 0.009506971575319767, + 0.009031831286847591, + 0.009408154524862766, + 0.015651149675250053, + 0.007882202044129372, + 0.01741763949394226, + 0.03977178782224655, + 0.017549598589539528, + 0.041672639548778534, + 0.01815900392830372, + 0.013794715516269207, + 0.07216930389404297, + 0.017784198746085167, + 0.05792155861854553, + 0.018513016402721405, + 0.007515802048146725, + 0.007962035946547985, + 0.005118388682603836, + 0.00637459009885788, + 0.03979915753006935, + 0.006170559208840132, + 0.02593996748328209, + 0.012296297587454319, + 0.007775159552693367, + 0.004827411845326424, + 0.005351436324417591, + 0.007888051681220531, + 0.028211036697030067, + 0.006820362992584705, + 0.01879103295505047, + 0.004365888424217701, + 0.0033405639696866274, + 0.004128428641706705, + 0.004501206800341606, + 0.004225472453981638, + 0.007478876505047083, + 0.004580516368150711, + 0.008392399176955223, + 0.008123045787215233, + 0.010204190388321877, + 0.0035682865418493748, + 0.005774443503469229, + 0.006036353297531605, + 0.011623193509876728, + 0.005699006374925375, + 0.008554191328585148, + 0.00533789349719882, + 0.007619643583893776, + 0.00465879961848259, + 0.005233950912952423, + 0.005852667614817619, + 0.00963553600013256, + 0.006351117044687271, + 0.010339518077671528, + 0.005313836503773928, + 0.005970516242086887, + 0.0064927274361252785, + 0.003580496646463871, + 0.00437210826203227, + 0.0071165477856993675, + 0.004219868220388889, + 0.00784879270941019, + 0.008261671289801598, + 0.004268905147910118, + 0.0054858773946762085, + 0.004332869313657284, + 0.005424147937446833, + 0.014851442538201809, + 0.006414434872567654, + 0.012481046840548515, + 0.020764393731951714, + 0.017517587170004845, + 0.011745710857212543, + 0.012125995941460133, + 0.017046524211764336, + 0.02141355350613594, + 0.015583539381623268, + 0.03496841341257095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.015241818502545357 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08545877784490585, + 0.38291701674461365, + 0.14945131540298462, + 0.05813375487923622, + 0.03690720722079277, + 0.02706390619277954, + 0.04182100296020508, + 0.07029268890619278, + 0.020040206611156464, + 0.12791411578655243 + ], + "credit_map": [ + 0.01023637130856514, + 0.010312307626008987, + 0.010260873474180698, + 0.0057031442411243916, + 0.009056742303073406, + 0.018609602004289627, + 0.005317972972989082, + 0.01596176251769066, + 0.06791839748620987, + 0.01691722683608532, + 0.024846244603395462, + 0.01183321513235569, + 0.014614630490541458, + 0.14297112822532654, + 0.01304426696151495, + 0.09077192842960358, + 0.01866135187447071, + 0.010147222317755222, + 0.03517972677946091, + 0.012508894316852093, + 0.007288196589797735, + 0.02208668738603592, + 0.012675573118031025, + 0.030903665348887444, + 0.007571115158498287, + 0.005884573329240084, + 0.007400488015264273, + 0.004681834019720554, + 0.005080083850771189, + 0.011626158840954304, + 0.0053065805695950985, + 0.010582922957837582, + 0.004823279567062855, + 0.003400040790438652, + 0.0041305869817733765, + 0.0031198987271636724, + 0.0020612021908164024, + 0.008694684132933617, + 0.003671192331239581, + 0.007006323430687189, + 0.003185867564752698, + 0.004221189301460981, + 0.003460818901658058, + 0.00285960896871984, + 0.0025522378273308277, + 0.0035310343373566866, + 0.0033442021813243628, + 0.003908948507159948, + 0.006037853192538023, + 0.004073258023709059, + 0.004437621217221022, + 0.0026854602620005608, + 0.002440118696540594, + 0.0076253945007920265, + 0.003917296417057514, + 0.010604003444314003, + 0.006876012776046991, + 0.008831987157464027, + 0.008581864647567272, + 0.00587594136595726, + 0.008059090934693813, + 0.011152723804116249, + 0.007742429617792368, + 0.013172640465199947, + 0.002415819326415658, + 0.002963100792840123, + 0.0029575591906905174, + 0.0017455826746299863, + 0.003316671121865511, + 0.0023257448337972164, + 0.001580814248882234, + 0.0027349155861884356, + 0.018180502578616142, + 0.014211559668183327, + 0.009483112022280693, + 0.007806664798408747, + 0.014674548991024494, + 0.019134877249598503, + 0.012254019267857075, + 0.032168835401535034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978341419249773 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.00978341419249773, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978341419249773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11022017151117325, + 0.11995313316583633, + 0.053120970726013184, + 0.05547613650560379, + 0.025669744238257408, + 0.04276357963681221, + 0.024666214361786842, + 0.09407662600278854, + 0.049559444189071655, + 0.4244939684867859 + ], + "credit_map": [ + 0.014260462485253811, + 0.02418225072324276, + 0.008679232560098171, + 0.011494351550936699, + 0.01768188364803791, + 0.011325542815029621, + 0.010362444445490837, + 0.012234007939696312, + 0.014956721104681492, + 0.009813743643462658, + 0.016003182157874107, + 0.010809311643242836, + 0.010427921079099178, + 0.024134071543812752, + 0.009473134763538837, + 0.024335039779543877, + 0.0063864924013614655, + 0.00505429320037365, + 0.007805400528013706, + 0.003307358594611287, + 0.004477458540350199, + 0.012281849049031734, + 0.004488770850002766, + 0.009319349192082882, + 0.007964719086885452, + 0.003518418874591589, + 0.0035759592428803444, + 0.002731271553784609, + 0.004034717101603746, + 0.017824210226535797, + 0.0034449524246156216, + 0.012381887063384056, + 0.004517023917287588, + 0.002495173364877701, + 0.002082883147522807, + 0.002343802247196436, + 0.0021119422744959593, + 0.005314227659255266, + 0.0021814738865941763, + 0.0046232170425355434, + 0.006475506350398064, + 0.005301309749484062, + 0.0033504636958241463, + 0.004289868753403425, + 0.00420785928145051, + 0.007647274993360043, + 0.0040886555798351765, + 0.007402641698718071, + 0.003849508473649621, + 0.0032323484774678946, + 0.0026220609433948994, + 0.0018089318182319403, + 0.002066622953861952, + 0.0036602921318262815, + 0.00268703605979681, + 0.004739414434880018, + 0.012540715746581554, + 0.009824768640100956, + 0.011011839844286442, + 0.0064216661266982555, + 0.013057397678494453, + 0.013881477527320385, + 0.009464076720178127, + 0.017874689772725105, + 0.005486201029270887, + 0.0101547846570611, + 0.0034795834217220545, + 0.005083135329186916, + 0.00889067817479372, + 0.005583202466368675, + 0.004857613239437342, + 0.0060242475010454655, + 0.059111982583999634, + 0.04851527884602547, + 0.030935324728488922, + 0.02998390421271324, + 0.04797704517841339, + 0.06191810220479965, + 0.04529755935072899, + 0.10075479000806808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.00978341419249773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07934711873531342, + 0.22318996489048004, + 0.04937407374382019, + 0.09703710675239563, + 0.03568355366587639, + 0.07050890475511551, + 0.024762624874711037, + 0.09055513143539429, + 0.08263825625181198, + 0.24690327048301697 + ], + "credit_map": [ + 0.009678700007498264, + 0.017139190807938576, + 0.007235875353217125, + 0.006970855873078108, + 0.011846933513879776, + 0.009771142154932022, + 0.00635199248790741, + 0.010352426208555698, + 0.037162624299526215, + 0.012783117592334747, + 0.01032527256757021, + 0.008392926305532455, + 0.008982833474874496, + 0.08330192416906357, + 0.00847640261054039, + 0.053764861077070236, + 0.006075131241232157, + 0.003943260293453932, + 0.007132915314286947, + 0.003898942843079567, + 0.003957095555961132, + 0.01049093808978796, + 0.0034168395213782787, + 0.010458949021995068, + 0.012964571826159954, + 0.007486524526029825, + 0.006825686898082495, + 0.005295774899423122, + 0.007897752337157726, + 0.02957027032971382, + 0.006000704597681761, + 0.02099582552909851, + 0.004761997144669294, + 0.0025245253928005695, + 0.0018266187980771065, + 0.0017765517113730311, + 0.002684048842638731, + 0.011835215613245964, + 0.002606127643957734, + 0.007668467238545418, + 0.009281842038035393, + 0.010056301020085812, + 0.0035620650742202997, + 0.0057451450265944, + 0.007525524124503136, + 0.016789229586720467, + 0.005811326205730438, + 0.01173747330904007, + 0.002689709886908531, + 0.0034979945048689842, + 0.0017177227418869734, + 0.0027189813554286957, + 0.003257718402892351, + 0.003439490683376789, + 0.004086130298674107, + 0.0033548781648278236, + 0.009434753097593784, + 0.014726734720170498, + 0.008755890652537346, + 0.008797576650977135, + 0.011746066622436047, + 0.012401479296386242, + 0.010967309586703777, + 0.013725320808589458, + 0.008888665586709976, + 0.015239641070365906, + 0.007121371570974588, + 0.007876718416810036, + 0.012124700471758842, + 0.011446529999375343, + 0.008015012368559837, + 0.01192561723291874, + 0.03266437351703644, + 0.030097637325525284, + 0.01947794482111931, + 0.018837783485651016, + 0.030216913670301437, + 0.03437691926956177, + 0.02650204859673977, + 0.054729659110307693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.011071558110415936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with a warm and vibrant Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a static, eye-level close-up of a small potted plant on a white wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have two large, heart-shaped, lush green leaves and a slender stem in an earthy clay pot with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright shaft of sunlight illuminating the plant and casting soft, shifting shadows with visible motes of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a time-lapse sequence where the plant's stem bends into smooth arcs and leaves lean towards the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22552546858787537 + ], + "opposite_logprob_tokens": [ + -1.6005254983901978 + ], + "logprob_reward": -0.22552546858787537, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22552546858787537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07176070660352707, + 0.16180622577667236, + 0.0878087729215622, + 0.03429710492491722, + 0.030865689739584923, + 0.05973087623715401, + 0.04406830668449402, + 0.11538883298635483, + 0.12000256776809692, + 0.2742708623409271 + ], + "credit_map": [ + 0.008081764914095402, + 0.012274390086531639, + 0.006427640561014414, + 0.005300260614603758, + 0.009616777300834656, + 0.011260195635259151, + 0.0059997644275426865, + 0.01279991865158081, + 0.027034055441617966, + 0.01020875759422779, + 0.014865029603242874, + 0.007235791999846697, + 0.007993986830115318, + 0.052571576088666916, + 0.007575065828859806, + 0.03432198241353035, + 0.013905985280871391, + 0.005647665821015835, + 0.004517636727541685, + 0.003750055795535445, + 0.003945061471313238, + 0.03191211074590683, + 0.0038085756823420525, + 0.020321689546108246, + 0.0032085590064525604, + 0.00478384830057621, + 0.0031177974306046963, + 0.004019324667751789, + 0.004705814644694328, + 0.004866372793912888, + 0.004224677104502916, + 0.0053707147017121315, + 0.0031724930740892887, + 0.0038105021230876446, + 0.0025064197834581137, + 0.003009756561368704, + 0.004040411673486233, + 0.005714999046176672, + 0.003530095797032118, + 0.005081014707684517, + 0.008406241424381733, + 0.008387351408600807, + 0.0032315205316990614, + 0.005305470898747444, + 0.006758987437933683, + 0.012861906550824642, + 0.005297919735312462, + 0.009481479413807392, + 0.004159403033554554, + 0.006291235331445932, + 0.002994389273226261, + 0.004406416323035955, + 0.00579060846939683, + 0.006374699994921684, + 0.006420710124075413, + 0.007630851585417986, + 0.011840769089758396, + 0.019389139488339424, + 0.010759673081338406, + 0.010950771160423756, + 0.015023784711956978, + 0.01649494841694832, + 0.013420579954981804, + 0.017509177327156067, + 0.012636376544833183, + 0.02200605347752571, + 0.00990816205739975, + 0.01126734260469675, + 0.018089987337589264, + 0.016862276941537857, + 0.01128495205193758, + 0.017947429791092873, + 0.03776956722140312, + 0.03387904167175293, + 0.02021481655538082, + 0.021339556202292442, + 0.0342760868370533, + 0.03742172569036484, + 0.03042346052825451, + 0.05894660949707031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.22552546858787537 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_5/index_tensor([2256, 2256])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_5/index_tensor([2256, 2256])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e72e849fc1f4f644cdfb874f689331094a53711a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_5/index_tensor([2256, 2256])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fixed, low-angle, overhead shot captures the rich, freehand style of ink painting. Deep within a tranquil bamboo forest, bright moonlight filters through the gaps between the towering bamboo leaves, casting a dappled, cool light and shadow. A poet, his face free and uninhibited, his eyes radiating uninhibited energy, stands amidst the light and shadow. He wears a long, flowing, dark cyan Tang Dynasty robe. The silken sleeves flutter in the night breeze, like smudges of ink on rice paper. His long, dark hair is held loosely together with a simple headband, a few strands flowing in the breeze. In an instant, his previously calm demeanor shifts to one of intense focus. With lightning speed, his right hand grasps the hilt of the ancient longsword at his waist. With a sudden thrust, a gentle metallic clang is heard, and a flash of cold light pierces the silence of the night. The highly polished sword blade perfectly captured and reflected the cold moonlight the moment it was drawn out of the scabbard, forming a brief and dazzling white flash. The entire sword-drawing action was done in one go, full of explosive power and decisive beauty, perfectly blending the dynamics of the characters with the static environment, creating a poetic martial arts atmosphere that is both tranquil and full of tension.", + "seed": 7916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_5/index_2256_seed_7916_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3867181334644556, + "vlm_logprob": -0.3867181334644556, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.3867181334644556, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2564862072467804 + ], + "opposite_logprob_tokens": [ + -1.506486177444458 + ], + "logprob_reward": -0.2564862072467804, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2564862072467804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23118627071380615, + 0.07506092637777328, + 0.035874143242836, + 0.03458757698535919, + 0.10885349661111832, + 0.03612298518419266, + 0.1733577996492386, + 0.08521915227174759, + 0.13125912845134735, + 0.08847848325967789 + ], + "credit_map": [ + 0.03242575749754906, + 0.024485858157277107, + 0.03558184951543808, + 0.029646435752511024, + 0.03669998422265053, + 0.01636262610554695, + 0.02796345017850399, + 0.02802029252052307, + 0.008505800738930702, + 0.00813360046595335, + 0.013472629711031914, + 0.0082296347245574, + 0.009438203647732735, + 0.006578077562153339, + 0.01072140783071518, + 0.009981575421988964, + 0.0040999893099069595, + 0.005549008492380381, + 0.004344994202256203, + 0.003228641115128994, + 0.004935781937092543, + 0.004494641907513142, + 0.0043768552131950855, + 0.004844230134040117, + 0.004497508052736521, + 0.004248816054314375, + 0.005639244336634874, + 0.002763770753517747, + 0.0034585115499794483, + 0.004181441850960255, + 0.00423682713881135, + 0.005561459809541702, + 0.019811423495411873, + 0.01374563854187727, + 0.017170531675219536, + 0.009690058417618275, + 0.010184063576161861, + 0.011701189912855625, + 0.00981574784964323, + 0.01673484407365322, + 0.004646007902920246, + 0.004916916135698557, + 0.0062504783272743225, + 0.0035449762362986803, + 0.004035599995404482, + 0.003288573119789362, + 0.003207810688763857, + 0.006232621613889933, + 0.0234080757945776, + 0.02956588938832283, + 0.01747623272240162, + 0.013331661000847816, + 0.013005145825445652, + 0.02249196544289589, + 0.019937830045819283, + 0.03414100408554077, + 0.019921721890568733, + 0.012278527021408081, + 0.009042621590197086, + 0.005722669884562492, + 0.007237492129206657, + 0.0078389011323452, + 0.006550503429025412, + 0.016626711934804916, + 0.0077696144580841064, + 0.010298425331711769, + 0.01692601852118969, + 0.038471125066280365, + 0.009547479450702667, + 0.024424750357866287, + 0.012207048013806343, + 0.011614678427577019, + 0.009258939884603024, + 0.00869886577129364, + 0.017870550975203514, + 0.013473263010382652, + 0.006052791606634855, + 0.012661371380090714, + 0.006411610171198845, + 0.01405109278857708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.2564862072467804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421281136572361 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.01421281136572361, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421281136572361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16603617370128632, + 0.06708328425884247, + 0.06597911566495895, + 0.049400828778743744, + 0.21120759844779968, + 0.03217633068561554, + 0.033418264240026474, + 0.04943760856986046, + 0.223292738199234, + 0.10196810215711594 + ], + "credit_map": [ + 0.016856597736477852, + 0.021732831373810768, + 0.016058940440416336, + 0.03345726802945137, + 0.020707983523607254, + 0.01417621411383152, + 0.021466711536049843, + 0.021579625084996223, + 0.008381976746022701, + 0.009090406820178032, + 0.01113035250455141, + 0.0073502808809280396, + 0.0070499274879693985, + 0.007513323333114386, + 0.008554385975003242, + 0.008012630976736546, + 0.008871842175722122, + 0.012426445260643959, + 0.008505502715706825, + 0.005261286161839962, + 0.005872944835573435, + 0.011936265975236893, + 0.005022922996431589, + 0.008081905543804169, + 0.0077641811221838, + 0.005328630097210407, + 0.008101062849164009, + 0.0043010832741856575, + 0.0046643600799143314, + 0.006839755456894636, + 0.005486641079187393, + 0.006915112026035786, + 0.05085897073149681, + 0.017053375020623207, + 0.02768038772046566, + 0.01370079256594181, + 0.016420964151620865, + 0.02798522636294365, + 0.01535453274846077, + 0.04215332865715027, + 0.0044098603539168835, + 0.003714347258210182, + 0.005581391043961048, + 0.0023707777727395296, + 0.0026932654436677694, + 0.0028576396871358156, + 0.0029416801407933235, + 0.0076073696836829185, + 0.005085526499897242, + 0.0035693536046892405, + 0.004590386524796486, + 0.0027683440130203962, + 0.00426524318754673, + 0.0030917911790311337, + 0.004097617696970701, + 0.005950002931058407, + 0.0114346444606781, + 0.006412031129002571, + 0.0055082920007407665, + 0.0035773280542343855, + 0.003356682136654854, + 0.00516343442723155, + 0.0037859820295125246, + 0.010199216194450855, + 0.014155863784253597, + 0.009499557316303253, + 0.04336266592144966, + 0.019326236099004745, + 0.00923345610499382, + 0.04457707330584526, + 0.010333915241062641, + 0.07280395179986954, + 0.00847393274307251, + 0.0094212107360363, + 0.018452094867825508, + 0.011813466437160969, + 0.011819648556411266, + 0.018121257424354553, + 0.008210373111069202, + 0.0156561192125082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.01421281136572361 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371366113424301 + ], + "opposite_logprob_tokens": [ + -3.4087135791778564 + ], + "logprob_reward": -0.03371366113424301, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371366113424301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10249323397874832, + 0.060845013707876205, + 0.03326750174164772, + 0.017734689638018608, + 0.05616829916834831, + 0.028254415839910507, + 0.13017146289348602, + 0.10601924359798431, + 0.3512994647026062, + 0.11374662816524506 + ], + "credit_map": [ + 0.010593983344733715, + 0.00949386227875948, + 0.008266458287835121, + 0.013110335916280746, + 0.020645800977945328, + 0.009021657519042492, + 0.01864706538617611, + 0.012714073061943054, + 0.00817167479544878, + 0.006155565846711397, + 0.005395571235567331, + 0.010299657471477985, + 0.008942495100200176, + 0.006078892387449741, + 0.007964669726788998, + 0.00783649180084467, + 0.005951195023953915, + 0.003047388279810548, + 0.0037124990485608578, + 0.0029023857787251472, + 0.005301969591528177, + 0.003176794620230794, + 0.005212612450122833, + 0.003962657880038023, + 0.0023040669038891792, + 0.0015233265003189445, + 0.0020392006263136864, + 0.0016425886424258351, + 0.003503846237435937, + 0.001673989463597536, + 0.0032126824371516705, + 0.0018349894089624286, + 0.011698714457452297, + 0.005778937600553036, + 0.007876699790358543, + 0.004823863506317139, + 0.004981344565749168, + 0.004665127489715815, + 0.005852408707141876, + 0.010491201654076576, + 0.0030159996822476387, + 0.0029070363380014896, + 0.002351510338485241, + 0.002651546150445938, + 0.0053914799354970455, + 0.0026621599681675434, + 0.004958093166351318, + 0.004316591192036867, + 0.016879785805940628, + 0.01845199055969715, + 0.010772909969091415, + 0.010054856538772583, + 0.015106619335711002, + 0.01581980288028717, + 0.012942807748913765, + 0.03014269657433033, + 0.015376080758869648, + 0.009260378777980804, + 0.013193298131227493, + 0.008015255443751812, + 0.013965061865746975, + 0.010868911631405354, + 0.01817273534834385, + 0.017167530953884125, + 0.029070943593978882, + 0.020538778975605965, + 0.057055409997701645, + 0.023489898070693016, + 0.02249343879520893, + 0.06088314950466156, + 0.02350427955389023, + 0.11426357924938202, + 0.010436570271849632, + 0.01616617850959301, + 0.017517685890197754, + 0.016006331890821457, + 0.01283702440559864, + 0.01541091501712799, + 0.010026790201663971, + 0.0153451357036829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.03371366113424301 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706369161605835 + ], + "opposite_logprob_tokens": [ + -2.127063751220703 + ], + "logprob_reward": -0.12706369161605835, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706369161605835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04879161715507507, + 0.024003375321626663, + 0.0199708454310894, + 0.02044481784105301, + 0.023678723722696304, + 0.025707120075821877, + 0.10717931389808655, + 0.044763144105672836, + 0.30742254853248596, + 0.3780384957790375 + ], + "credit_map": [ + 0.005414016079157591, + 0.005122475791722536, + 0.004341958090662956, + 0.0049754297360777855, + 0.009458986110985279, + 0.004252677317708731, + 0.007588053587824106, + 0.0076380218379199505, + 0.0029373872093856335, + 0.0035895872861146927, + 0.0028676590882241726, + 0.0024377324152737856, + 0.003651520935818553, + 0.0027618135791271925, + 0.0030916347168385983, + 0.0026660412549972534, + 0.0019679770339280367, + 0.0024667333345860243, + 0.0022951916325837374, + 0.002319670282304287, + 0.0035846219398081303, + 0.0017909450689330697, + 0.003212926909327507, + 0.0023327800445258617, + 0.002072889357805252, + 0.00204116920940578, + 0.0021640965715050697, + 0.0019925362430512905, + 0.002619845326989889, + 0.00249689444899559, + 0.0044181328266859055, + 0.002639253856614232, + 0.002116439864039421, + 0.0023024296388030052, + 0.002514168620109558, + 0.0053882477805018425, + 0.0032676728442311287, + 0.0022103306837379932, + 0.0030224777292460203, + 0.0028569577261805534, + 0.003157064551487565, + 0.0034700187388807535, + 0.004425767343491316, + 0.0019924703519791365, + 0.002649955917149782, + 0.002493055071681738, + 0.0031239534728229046, + 0.004394834861159325, + 0.014289140701293945, + 0.01565714180469513, + 0.01061272993683815, + 0.010573381558060646, + 0.009267300367355347, + 0.011444717645645142, + 0.009814577177166939, + 0.025520330294966698, + 0.006209224462509155, + 0.005001267418265343, + 0.006655560806393623, + 0.004995363298803568, + 0.00568292336538434, + 0.004038736689835787, + 0.006709248758852482, + 0.0054708197712898254, + 0.017961256206035614, + 0.02760818600654602, + 0.049113284796476364, + 0.03557706996798515, + 0.013886276632547379, + 0.06966310739517212, + 0.014614108949899673, + 0.07899928092956543, + 0.03210049867630005, + 0.06946059316396713, + 0.057602155953645706, + 0.06045018509030342, + 0.035531967878341675, + 0.04543522745370865, + 0.029624084010720253, + 0.04783378168940544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.12706369161605835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5021142959594727 + ], + "opposite_logprob_tokens": [ + -0.25211426615715027 + ], + "logprob_reward": -1.5021142959594727, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5021142959594727, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17245185375213623, + 0.0816938504576683, + 0.04844456911087036, + 0.07903288304805756, + 0.05856095626950264, + 0.05259858816862106, + 0.09002279490232468, + 0.07751812785863876, + 0.09770587086677551, + 0.24197052419185638 + ], + "credit_map": [ + 0.026298195123672485, + 0.019689086824655533, + 0.020773934200406075, + 0.017972024157643318, + 0.027585884556174278, + 0.016204550862312317, + 0.0245109423995018, + 0.01941721886396408, + 0.008958780206739902, + 0.009654045104980469, + 0.010807697661221027, + 0.009372501634061337, + 0.011231734417378902, + 0.009569860063493252, + 0.013181110844016075, + 0.008918121457099915, + 0.007005301769822836, + 0.008707532659173012, + 0.005311124958097935, + 0.004307140130549669, + 0.005122348666191101, + 0.006546089891344309, + 0.005801913794130087, + 0.005643119104206562, + 0.01235482469201088, + 0.008252875879406929, + 0.012182356789708138, + 0.006028714589774609, + 0.007071393076330423, + 0.013664104044437408, + 0.008894807659089565, + 0.010583802126348019, + 0.006456905044615269, + 0.0061650765128433704, + 0.008098756894469261, + 0.008323893882334232, + 0.008163968101143837, + 0.004938795696943998, + 0.009447205811738968, + 0.006966352462768555, + 0.006042430642992258, + 0.008152937516570091, + 0.007467460352927446, + 0.006040394771844149, + 0.006598858628422022, + 0.004169028252363205, + 0.006384274922311306, + 0.007743199355900288, + 0.007873845286667347, + 0.011937937699258327, + 0.023118671029806137, + 0.011180750094354153, + 0.007403749972581863, + 0.008547484874725342, + 0.010830205865204334, + 0.009130151942372322, + 0.011195543222129345, + 0.013245075941085815, + 0.009951984509825706, + 0.006998253054916859, + 0.008366934023797512, + 0.0071484399959445, + 0.008022034540772438, + 0.012589864432811737, + 0.009428980760276318, + 0.010067015886306763, + 0.011090030893683434, + 0.011447697877883911, + 0.010522033087909222, + 0.023109566420316696, + 0.009986377321183681, + 0.012054158374667168, + 0.023563455790281296, + 0.0372123122215271, + 0.03794224560260773, + 0.03475312143564224, + 0.0269028227776289, + 0.027147680521011353, + 0.02238718420267105, + 0.03206167742609978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -1.5021142959594727 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "fixed, low-angle, overhead shot captures the rich, freehand style of ink painting. Deep within a tranquil bamboo forest, bright moonlight filters through the gaps between the towering bamboo leaves, casting a dappled, cool light and shadow. A poet, his face free and uninhibited, his eyes radiating uninhibited energy, stands amidst the light and shadow. He wears a long, flowing, dark cyan Tang Dynasty robe. The silken sleeves flutter in the night breeze, like smudges of ink on rice paper. His long, dark hair is held loosely together with a simple headband, a few strands flowing in the breeze. In an instant, his previously calm demeanor shifts to one of intense focus. With lightning speed, his right hand grasps the hilt of the ancient longsword at his waist. With a sudden thrust, a gentle metallic clang is heard, and a flash of cold light pierces the silence of the night. The highly polished sword blade perfectly captured and reflected the cold moonlight the moment it was drawn out of the scabbard, forming a brief and dazzling white flash. The entire sword-drawing action was done in one go, full of explosive power and decisive beauty, perfectly blending the dynamics of the characters with the static environment, creating a poetic martial arts atmosphere that is both tranquil and full of tension.", + "seed": 7917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_5/index_2256_seed_7917_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.33740109838545324, + "vlm_logprob": -0.33740109838545324, + "vlm_yes_no_margin": 1.775000011920929, + "vlm_token_credit": -0.33740109838545324, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23093903064727783 + ], + "opposite_logprob_tokens": [ + -1.6059390306472778 + ], + "logprob_reward": -0.23093903064727783, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23093903064727783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24079705774784088, + 0.17449797689914703, + 0.11256784200668335, + 0.11511306464672089, + 0.09512794017791748, + 0.05980302393436432, + 0.05724982172250748, + 0.07762657105922699, + 0.031401779502630234, + 0.03581500053405762 + ], + "credit_map": [ + 0.04167464002966881, + 0.02053503878414631, + 0.0464436337351799, + 0.016547933220863342, + 0.027967946603894234, + 0.02353896014392376, + 0.03640007972717285, + 0.027688803151249886, + 0.023975182324647903, + 0.02190348319709301, + 0.02611706219613552, + 0.027295583859086037, + 0.019489910453557968, + 0.020431648939847946, + 0.015614217147231102, + 0.019670873880386353, + 0.011850006878376007, + 0.012048611417412758, + 0.026793528348207474, + 0.01426294632256031, + 0.010727273300290108, + 0.011674015782773495, + 0.0138306999579072, + 0.01138075441122055, + 0.010193385183811188, + 0.013604709878563881, + 0.01380117516964674, + 0.012964111752808094, + 0.02334980107843876, + 0.01626429334282875, + 0.012618757784366608, + 0.012316818349063396, + 0.007079428061842918, + 0.008891509845852852, + 0.018913887441158295, + 0.01519923098385334, + 0.009988796897232533, + 0.010311605408787727, + 0.014563556760549545, + 0.010179917328059673, + 0.008525799959897995, + 0.008413314819335938, + 0.008443420752882957, + 0.006878807675093412, + 0.010108516551554203, + 0.007625849451869726, + 0.004365249071270227, + 0.005442060995846987, + 0.004921286832541227, + 0.0085472222417593, + 0.006206337362527847, + 0.0065254042856395245, + 0.006496076006442308, + 0.008742192760109901, + 0.00611520791426301, + 0.009696092456579208, + 0.008561650291085243, + 0.011120372451841831, + 0.015900736674666405, + 0.005805006716400385, + 0.008154219016432762, + 0.016245584934949875, + 0.005461015738546848, + 0.006377972662448883, + 0.004559192340821028, + 0.00619996152818203, + 0.004283621441572905, + 0.0033478359691798687, + 0.004224932752549648, + 0.003502635518088937, + 0.0027540158480405807, + 0.002529582940042019, + 0.0072504677809774876, + 0.0073012360371649265, + 0.005126594100147486, + 0.0029454834293574095, + 0.003302631201222539, + 0.004605086520314217, + 0.0028801264706999063, + 0.002403371036052704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.23093903064727783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209779635071754 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018209779635071754, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209779635071754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36447882652282715, + 0.10847768932580948, + 0.12454640120267868, + 0.11160985380411148, + 0.057045284658670425, + 0.030856432393193245, + 0.047198329120874405, + 0.09157834202051163, + 0.040675945580005646, + 0.023532846942543983 + ], + "credit_map": [ + 0.05652320757508278, + 0.027453789487481117, + 0.06968913972377777, + 0.03125075250864029, + 0.03358801081776619, + 0.05567866191267967, + 0.03175315633416176, + 0.05854212865233421, + 0.015312152914702892, + 0.012785598635673523, + 0.013477586209774017, + 0.01509526465088129, + 0.007721500471234322, + 0.018923798575997353, + 0.010378418490290642, + 0.014783375896513462, + 0.007724315859377384, + 0.010193811729550362, + 0.017092814669013023, + 0.007236653473228216, + 0.007147455122321844, + 0.03453748673200607, + 0.009015819989144802, + 0.031598057597875595, + 0.009734843857586384, + 0.015157913789153099, + 0.01468308549374342, + 0.017352677881717682, + 0.010009288787841797, + 0.020629271864891052, + 0.010989319533109665, + 0.013053463771939278, + 0.0047620669938623905, + 0.0068297004327178, + 0.008787931874394417, + 0.007341845892369747, + 0.006430380512028933, + 0.008283626288175583, + 0.008033939637243748, + 0.006575797684490681, + 0.0040417923592031, + 0.005156887229532003, + 0.003582016797736287, + 0.0035470116417855024, + 0.003736025420948863, + 0.00490988977253437, + 0.0024557451251894236, + 0.0034270635806024075, + 0.00642463518306613, + 0.00871267169713974, + 0.0055548385716974735, + 0.005898747593164444, + 0.006291259080171585, + 0.005407129414379597, + 0.003966381773352623, + 0.004942665807902813, + 0.010024008341133595, + 0.013096224516630173, + 0.02731746807694435, + 0.006591533310711384, + 0.006240657065063715, + 0.016854524612426758, + 0.00776407727971673, + 0.003689850214868784, + 0.005632749758660793, + 0.006713067181408405, + 0.009105728007853031, + 0.004014474805444479, + 0.0022249515168368816, + 0.007675122935324907, + 0.0022648880258202553, + 0.003044965909793973, + 0.003141198307275772, + 0.006793691776692867, + 0.002729381900280714, + 0.0025397075805813074, + 0.001688493648543954, + 0.0026737095322459936, + 0.001244436833076179, + 0.002722229575738311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.018209779635071754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866442829370499 + ], + "opposite_logprob_tokens": [ + -3.0486643314361572 + ], + "logprob_reward": -0.04866442829370499, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866442829370499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25622180104255676, + 0.11050662398338318, + 0.08541424572467804, + 0.067019023001194, + 0.24265961349010468, + 0.02886221371591091, + 0.03148801624774933, + 0.06323902308940887, + 0.09067974984645844, + 0.023909704759716988 + ], + "credit_map": [ + 0.03996798023581505, + 0.022048428654670715, + 0.04861766844987869, + 0.035446956753730774, + 0.024421658366918564, + 0.03733929246664047, + 0.018648773431777954, + 0.029731057584285736, + 0.015353722497820854, + 0.011986854486167431, + 0.01611309126019478, + 0.015429078601300716, + 0.011522498913109303, + 0.012279687449336052, + 0.011014177463948727, + 0.01680750586092472, + 0.005074369255453348, + 0.00886534620076418, + 0.008550782687962055, + 0.010419956408441067, + 0.011429673060774803, + 0.014606769196689129, + 0.012526201084256172, + 0.013941144570708275, + 0.005399084184318781, + 0.006140394601970911, + 0.007083520758897066, + 0.008569108322262764, + 0.017509792000055313, + 0.008543665520846844, + 0.005574591923505068, + 0.008198867551982403, + 0.013363860547542572, + 0.021413065493106842, + 0.03079456463456154, + 0.027568574994802475, + 0.022175172343850136, + 0.03719249740242958, + 0.04272979125380516, + 0.04742206633090973, + 0.0034185436088591814, + 0.004283619578927755, + 0.0040498534217476845, + 0.003316562855616212, + 0.004461443051695824, + 0.003696477971971035, + 0.002575361169874668, + 0.0030603529885411263, + 0.002765643410384655, + 0.003715262748301029, + 0.003193207085132599, + 0.0037007827777415514, + 0.0044059292413294315, + 0.004617281258106232, + 0.003523655002936721, + 0.005566252861171961, + 0.0068672471679747105, + 0.009720198810100555, + 0.006374096963554621, + 0.010260650888085365, + 0.009689167141914368, + 0.010214014910161495, + 0.00641942024230957, + 0.0036942288279533386, + 0.008337647654116154, + 0.02122175134718418, + 0.024607961997389793, + 0.006357081700116396, + 0.0065719252452254295, + 0.012248419225215912, + 0.004628192633390427, + 0.006706772372126579, + 0.0033328470308333635, + 0.007013667840510607, + 0.0026154760271310806, + 0.0025201463140547276, + 0.0020630296785384417, + 0.0023295420687645674, + 0.0013225978473201394, + 0.002712397836148739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04866442829370499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210896134376526 + ], + "opposite_logprob_tokens": [ + -1.5021089315414429 + ], + "logprob_reward": -0.25210896134376526, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210896134376526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09242044389247894, + 0.041160330176353455, + 0.05982675775885582, + 0.10834935307502747, + 0.09264253079891205, + 0.063450887799263, + 0.052179962396621704, + 0.17776955664157867, + 0.27113863825798035, + 0.04106149449944496 + ], + "credit_map": [ + 0.012475117109715939, + 0.008313657715916634, + 0.015447103418409824, + 0.01503715105354786, + 0.008204551413655281, + 0.008466209284961224, + 0.01294840220361948, + 0.011528261937201023, + 0.0045965141616761684, + 0.0030236972961574793, + 0.0038450579158961773, + 0.004406663589179516, + 0.0034239268861711025, + 0.008403911255300045, + 0.006096282508224249, + 0.00736427353695035, + 0.004400511737912893, + 0.004360557533800602, + 0.006565434858202934, + 0.004192492924630642, + 0.005469838157296181, + 0.014299474656581879, + 0.007505303248763084, + 0.013033145107328892, + 0.0078077916987240314, + 0.009334426373243332, + 0.013883241452276707, + 0.019979000091552734, + 0.012974360957741737, + 0.01663544774055481, + 0.01456419751048088, + 0.013170882128179073, + 0.0052696834318339825, + 0.007886862382292747, + 0.02004067972302437, + 0.014724405482411385, + 0.00863089133054018, + 0.009303523227572441, + 0.01714128628373146, + 0.009645192883908749, + 0.007345930207520723, + 0.009250047616660595, + 0.00777531648054719, + 0.006997772958129644, + 0.008729212917387486, + 0.011202334426343441, + 0.005828960333019495, + 0.006321310065686703, + 0.004300006665289402, + 0.0074123721569776535, + 0.005935145076364279, + 0.006830132100731134, + 0.009179577231407166, + 0.006920069456100464, + 0.006392238661646843, + 0.005210422910749912, + 0.010680658742785454, + 0.047884777188301086, + 0.05677931383252144, + 0.014974691905081272, + 0.013722961768507957, + 0.012855149805545807, + 0.011364125646650791, + 0.009507877752184868, + 0.02354104071855545, + 0.07751855254173279, + 0.076849564909935, + 0.019736725836992264, + 0.020299488678574562, + 0.022382764145731926, + 0.013923170045018196, + 0.01688733696937561, + 0.006292787380516529, + 0.011889849789440632, + 0.005775759927928448, + 0.0040009357035160065, + 0.0037068736273795366, + 0.0036838469095528126, + 0.0024356029462069273, + 0.0032758379820734262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.25210896134376526 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370832920074463 + ], + "opposite_logprob_tokens": [ + -0.38708335161209106 + ], + "logprob_reward": -1.1370832920074463, + "yes_no_margin_reward": -0.7499999403953552, + "token_credit_reward": -1.1370832920074463, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21854864060878754, + 0.06692540645599365, + 0.06564533710479736, + 0.20288097858428955, + 0.10409869998693466, + 0.0885777547955513, + 0.0520952083170414, + 0.08872389793395996, + 0.08119905740022659, + 0.03130502998828888 + ], + "credit_map": [ + 0.04510372877120972, + 0.0172458253800869, + 0.04415331035852432, + 0.022576894611120224, + 0.018672842532396317, + 0.02573394775390625, + 0.0170334093272686, + 0.028028689324855804, + 0.008180687204003334, + 0.006638910621404648, + 0.008864110335707664, + 0.005260650999844074, + 0.011263735592365265, + 0.0067422883585095406, + 0.011610248126089573, + 0.008364775218069553, + 0.0054796920157969, + 0.006587832700461149, + 0.00929284282028675, + 0.005884120706468821, + 0.00681940745562315, + 0.01167561300098896, + 0.00776679627597332, + 0.012139028869569302, + 0.016160327941179276, + 0.01811857335269451, + 0.02456234022974968, + 0.02513396553695202, + 0.03076895698904991, + 0.037137195467948914, + 0.03202826902270317, + 0.018971355631947517, + 0.009385868906974792, + 0.008804764598608017, + 0.013554244302213192, + 0.011845765635371208, + 0.015835482627153397, + 0.014665299095213413, + 0.014331065118312836, + 0.015676213428378105, + 0.0090331444516778, + 0.011277386918663979, + 0.012204957194626331, + 0.010540437884628773, + 0.017929309979081154, + 0.012614676728844643, + 0.006246195174753666, + 0.008731652982532978, + 0.004016876220703125, + 0.007584688253700733, + 0.0061696721240878105, + 0.007132167462259531, + 0.00902697816491127, + 0.00480527151376009, + 0.006788546219468117, + 0.0065710111521184444, + 0.006852387450635433, + 0.018956830725073814, + 0.031285155564546585, + 0.008849130012094975, + 0.003844592487439513, + 0.00906587578356266, + 0.004012437537312508, + 0.00585749140009284, + 0.008338349871337414, + 0.012521177530288696, + 0.00955679826438427, + 0.00785066094249487, + 0.007297642063349485, + 0.02167285792529583, + 0.005803109612315893, + 0.008158468641340733, + 0.0042283302173018456, + 0.006865667179226875, + 0.005595645867288113, + 0.002940241014584899, + 0.003105378011241555, + 0.004293186590075493, + 0.0018795594805851579, + 0.0023970231413841248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -1.1370832920074463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "fixed, low-angle, overhead shot captures the rich, freehand style of ink painting. Deep within a tranquil bamboo forest, bright moonlight filters through the gaps between the towering bamboo leaves, casting a dappled, cool light and shadow. A poet, his face free and uninhibited, his eyes radiating uninhibited energy, stands amidst the light and shadow. He wears a long, flowing, dark cyan Tang Dynasty robe. The silken sleeves flutter in the night breeze, like smudges of ink on rice paper. His long, dark hair is held loosely together with a simple headband, a few strands flowing in the breeze. In an instant, his previously calm demeanor shifts to one of intense focus. With lightning speed, his right hand grasps the hilt of the ancient longsword at his waist. With a sudden thrust, a gentle metallic clang is heard, and a flash of cold light pierces the silence of the night. The highly polished sword blade perfectly captured and reflected the cold moonlight the moment it was drawn out of the scabbard, forming a brief and dazzling white flash. The entire sword-drawing action was done in one go, full of explosive power and decisive beauty, perfectly blending the dynamics of the characters with the static environment, creating a poetic martial arts atmosphere that is both tranquil and full of tension.", + "seed": 7918, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_5/index_2256_seed_7918_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.4325601620599627, + "vlm_logprob": -0.4325601620599627, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.4325601620599627, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25530704855918884 + ], + "opposite_logprob_tokens": [ + -1.5053070783615112 + ], + "logprob_reward": -0.25530704855918884, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25530704855918884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16939187049865723, + 0.03580643609166145, + 0.026515381410717964, + 0.07888074219226837, + 0.09791528433561325, + 0.11213362216949463, + 0.07057373970746994, + 0.07725690305233002, + 0.24133895337581635, + 0.09018709510564804 + ], + "credit_map": [ + 0.026341304183006287, + 0.021342365071177483, + 0.024769369512796402, + 0.0203029066324234, + 0.019198568537831306, + 0.01935455948114395, + 0.019294030964374542, + 0.018788764253258705, + 0.004305094014853239, + 0.004124749451875687, + 0.0057130358181893826, + 0.004763337783515453, + 0.003627439495176077, + 0.005330739077180624, + 0.0038448385894298553, + 0.004097198601812124, + 0.004252044018357992, + 0.0032903035171329975, + 0.0032023333478718996, + 0.004348549526184797, + 0.0023162502329796553, + 0.003004672471433878, + 0.002993518253788352, + 0.0031077100429683924, + 0.009570312686264515, + 0.009448026306927204, + 0.00795978493988514, + 0.012474809773266315, + 0.006926002446562052, + 0.015735793858766556, + 0.008997716009616852, + 0.007768295705318451, + 0.011116656474769115, + 0.00922603439539671, + 0.015220612287521362, + 0.011856215074658394, + 0.01118498481810093, + 0.011017507873475552, + 0.0130442064255476, + 0.015249068848788738, + 0.011604223400354385, + 0.011722416616976261, + 0.021699609234929085, + 0.011065288446843624, + 0.009359006769955158, + 0.014690185897052288, + 0.020007960498332977, + 0.011984924785792828, + 0.010378253646194935, + 0.005749547854065895, + 0.007915262132883072, + 0.006603879854083061, + 0.005852602422237396, + 0.009898480959236622, + 0.014559177681803703, + 0.00961652584373951, + 0.014145619235932827, + 0.006699476856738329, + 0.00833324994891882, + 0.01033408846706152, + 0.00649454677477479, + 0.010677034966647625, + 0.008007561787962914, + 0.012565329670906067, + 0.050162672996520996, + 0.023020310327410698, + 0.029984423890709877, + 0.032597560435533524, + 0.024550283327698708, + 0.02745148353278637, + 0.02789248526096344, + 0.02567973919212818, + 0.005530227441340685, + 0.0083226403221488, + 0.005730492062866688, + 0.008160341531038284, + 0.006521911825984716, + 0.023241417482495308, + 0.005732360761612654, + 0.02694769762456417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.25530704855918884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090303771197796 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090303771197796, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090303771197796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22713370621204376, + 0.10626386851072311, + 0.07517780363559723, + 0.053800202906131744, + 0.06691619753837585, + 0.062349557876586914, + 0.04443575069308281, + 0.23323220014572144, + 0.05376053228974342, + 0.07693024724721909 + ], + "credit_map": [ + 0.06067528575658798, + 0.019882647320628166, + 0.03964477777481079, + 0.020045947283506393, + 0.013711708597838879, + 0.030849946662783623, + 0.020190149545669556, + 0.022133242338895798, + 0.02244202420115471, + 0.009691091254353523, + 0.01990208402276039, + 0.01075500063598156, + 0.007456908468157053, + 0.01455939095467329, + 0.009730079211294651, + 0.011727280914783478, + 0.013939248397946358, + 0.006092487834393978, + 0.01297474093735218, + 0.008892286568880081, + 0.0054072169587016106, + 0.008818472735583782, + 0.010726895183324814, + 0.008326450362801552, + 0.006080344319343567, + 0.005310543347150087, + 0.0058228494599461555, + 0.011442773975431919, + 0.003981187008321285, + 0.010707891546189785, + 0.006681653205305338, + 0.003772957716137171, + 0.007105352357029915, + 0.005838026292622089, + 0.008168120868504047, + 0.004792079329490662, + 0.007849708199501038, + 0.009353085421025753, + 0.009912176989018917, + 0.013897641561925411, + 0.005538579076528549, + 0.006123701110482216, + 0.013254838064312935, + 0.005787436850368977, + 0.0047226594761013985, + 0.008288338780403137, + 0.010828228667378426, + 0.007805774454027414, + 0.004830141551792622, + 0.004104449413716793, + 0.004717271309345961, + 0.005261596292257309, + 0.003380663925781846, + 0.005938326939940453, + 0.009099787101149559, + 0.007103513926267624, + 0.04615841433405876, + 0.02351662516593933, + 0.02050088904798031, + 0.053549427539110184, + 0.01728423684835434, + 0.028479207307100296, + 0.018915096297860146, + 0.024828309193253517, + 0.010362316854298115, + 0.005971068982034922, + 0.005296334624290466, + 0.004942038562148809, + 0.00872693583369255, + 0.006448037456721067, + 0.006773278582841158, + 0.005240525119006634, + 0.010637514293193817, + 0.007223672233521938, + 0.008399248123168945, + 0.007470397278666496, + 0.006995143368840218, + 0.012626192532479763, + 0.009384827688336372, + 0.014193248935043812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.011090303771197796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544889003038406 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012544889003038406, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544889003038406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15870900452136993, + 0.07941790670156479, + 0.05484674125909805, + 0.04730619490146637, + 0.14403831958770752, + 0.1514919549226761, + 0.10328010469675064, + 0.05822376534342766, + 0.1048068180680275, + 0.09787921607494354 + ], + "credit_map": [ + 0.01986919902265072, + 0.02178115025162697, + 0.021046441048383713, + 0.018912719562649727, + 0.01862042024731636, + 0.02018778957426548, + 0.02433040551841259, + 0.01396086998283863, + 0.007838863879442215, + 0.011370522901415825, + 0.006460567470639944, + 0.01061243936419487, + 0.008757255040109158, + 0.009927961975336075, + 0.01670105569064617, + 0.007749236188828945, + 0.007941344752907753, + 0.00590521888807416, + 0.006144996732473373, + 0.0066128987818956375, + 0.005406522192060947, + 0.008681799285113811, + 0.008764130994677544, + 0.005389830097556114, + 0.007531417068094015, + 0.005311938468366861, + 0.0050054253078997135, + 0.009638589806854725, + 0.004419013857841492, + 0.006608219817280769, + 0.005144794471561909, + 0.003646794706583023, + 0.013172844424843788, + 0.015926575288176537, + 0.02133418619632721, + 0.019763043150305748, + 0.018834028393030167, + 0.020503874868154526, + 0.022881053388118744, + 0.011622724123299122, + 0.015291860327124596, + 0.01613045111298561, + 0.03156683221459389, + 0.0184638611972332, + 0.011415817774832249, + 0.019094644114375114, + 0.027070971205830574, + 0.012457506731152534, + 0.006966549903154373, + 0.00913060363382101, + 0.008371992968022823, + 0.01309084240347147, + 0.013672074303030968, + 0.01405106857419014, + 0.025556737557053566, + 0.012440234422683716, + 0.00817282497882843, + 0.006164805497974157, + 0.006201986689120531, + 0.006590873468667269, + 0.006763020996004343, + 0.007659887429326773, + 0.007281559053808451, + 0.009388807229697704, + 0.020928209647536278, + 0.010935341008007526, + 0.012036126106977463, + 0.013954329304397106, + 0.00956748053431511, + 0.012185177765786648, + 0.011472764424979687, + 0.01372738927602768, + 0.0059292372316122055, + 0.007011419627815485, + 0.005728929303586483, + 0.004995853174477816, + 0.006396036595106125, + 0.029286585748195648, + 0.008212032727897167, + 0.03031912073493004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.012544889003038406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898426800966263 + ], + "opposite_logprob_tokens": [ + -2.578984260559082 + ], + "logprob_reward": -0.07898426800966263, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898426800966263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0679352879524231, + 0.04118117690086365, + 0.05347044765949249, + 0.052586961537599564, + 0.10001388192176819, + 0.03948641195893288, + 0.2665923833847046, + 0.2518313527107239, + 0.045427460223436356, + 0.08147457242012024 + ], + "credit_map": [ + 0.008944366127252579, + 0.006732674781233072, + 0.009888870641589165, + 0.010820241644978523, + 0.005778908263891935, + 0.01006351225078106, + 0.008618234656751156, + 0.007088481914252043, + 0.00416575325652957, + 0.003782349405810237, + 0.006854889914393425, + 0.007710042409598827, + 0.003130524419248104, + 0.005779418163001537, + 0.004980535712093115, + 0.004777665250003338, + 0.005587300751358271, + 0.006892012897878885, + 0.006668602116405964, + 0.009158041328191757, + 0.006187091581523418, + 0.006943054497241974, + 0.006793719716370106, + 0.0052406275644898415, + 0.00876111164689064, + 0.0064987982623279095, + 0.004482762888073921, + 0.009829158894717693, + 0.005389257334172726, + 0.006610165350139141, + 0.007067548111081123, + 0.003948160447180271, + 0.009942295961081982, + 0.0105466702952981, + 0.014424175024032593, + 0.007633129600435495, + 0.011909660883247852, + 0.012084057554602623, + 0.014764414168894291, + 0.018709467723965645, + 0.0035826938692480326, + 0.004637398757040501, + 0.005790473893284798, + 0.00471713300794363, + 0.0049201371148228645, + 0.004062680993229151, + 0.005368350073695183, + 0.006407541688531637, + 0.014168608002364635, + 0.01640644669532776, + 0.025429919362068176, + 0.028224406763911247, + 0.02414843626320362, + 0.0369168184697628, + 0.08880342543125153, + 0.03249433636665344, + 0.0413561649620533, + 0.025004597380757332, + 0.02170673944056034, + 0.056705985218286514, + 0.024402786046266556, + 0.03148684650659561, + 0.02428489178419113, + 0.026883352547883987, + 0.006402153521776199, + 0.004282713867723942, + 0.005197424441576004, + 0.0053731962107121944, + 0.0032376404851675034, + 0.00594349205493927, + 0.007388476748019457, + 0.00760236382484436, + 0.005490539595484734, + 0.0041718147695064545, + 0.004455554764717817, + 0.0031744216103106737, + 0.005788171663880348, + 0.026760373264551163, + 0.004003773909062147, + 0.027629924938082695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.07898426800966263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.804874300956726 + ], + "opposite_logprob_tokens": [ + -0.17987431585788727 + ], + "logprob_reward": -1.804874300956726, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.804874300956726, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08985098451375961, + 0.04272518306970596, + 0.04537622258067131, + 0.05772646889090538, + 0.05798112228512764, + 0.08433892577886581, + 0.10639016330242157, + 0.30019277334213257, + 0.1462850123643875, + 0.06913318485021591 + ], + "credit_map": [ + 0.012645658105611801, + 0.01246631983667612, + 0.013199150562286377, + 0.01139808725565672, + 0.008718608878552914, + 0.010340026579797268, + 0.011776581406593323, + 0.009306549094617367, + 0.005279358942061663, + 0.0033355150371789932, + 0.006336317863315344, + 0.0052635991014540195, + 0.0033453512005507946, + 0.004837305285036564, + 0.007069020066410303, + 0.007258713245391846, + 0.00866724643856287, + 0.003947109449654818, + 0.00579862529411912, + 0.005351829808205366, + 0.004150354769080877, + 0.006594072096049786, + 0.005527378525584936, + 0.005339606199413538, + 0.005449879448860884, + 0.006671533454209566, + 0.006935549434274435, + 0.00834069773554802, + 0.0055082556791603565, + 0.012931091710925102, + 0.008932402357459068, + 0.0029570595361292362, + 0.005463752429932356, + 0.007565946783870459, + 0.00905346218496561, + 0.005384569521993399, + 0.006108930334448814, + 0.006836577318608761, + 0.008519587107002735, + 0.009048300795257092, + 0.014312958344817162, + 0.014501402154564857, + 0.011513852514326572, + 0.00920422375202179, + 0.00915165152400732, + 0.009124217554926872, + 0.009303026832640171, + 0.007227592635899782, + 0.009799020364880562, + 0.00727312546223402, + 0.009821921586990356, + 0.010821180418133736, + 0.008189814165234566, + 0.013767125084996223, + 0.027804061770439148, + 0.01891391910612583, + 0.04629584029316902, + 0.029597334563732147, + 0.026188915595412254, + 0.07203482836484909, + 0.025910064578056335, + 0.03885974362492561, + 0.029133979231119156, + 0.032172057777643204, + 0.014197329990565777, + 0.016611043363809586, + 0.01517339888960123, + 0.01975114643573761, + 0.011508380994200706, + 0.018605757504701614, + 0.0328800342977047, + 0.01755792275071144, + 0.005986820440739393, + 0.00641518784686923, + 0.005943612661212683, + 0.008630486205220222, + 0.00479530356824398, + 0.014214787632226944, + 0.007324686739593744, + 0.01582229696214199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -1.804874300956726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "fixed, low-angle, overhead shot captures the rich, freehand style of ink painting. Deep within a tranquil bamboo forest, bright moonlight filters through the gaps between the towering bamboo leaves, casting a dappled, cool light and shadow. A poet, his face free and uninhibited, his eyes radiating uninhibited energy, stands amidst the light and shadow. He wears a long, flowing, dark cyan Tang Dynasty robe. The silken sleeves flutter in the night breeze, like smudges of ink on rice paper. His long, dark hair is held loosely together with a simple headband, a few strands flowing in the breeze. In an instant, his previously calm demeanor shifts to one of intense focus. With lightning speed, his right hand grasps the hilt of the ancient longsword at his waist. With a sudden thrust, a gentle metallic clang is heard, and a flash of cold light pierces the silence of the night. The highly polished sword blade perfectly captured and reflected the cold moonlight the moment it was drawn out of the scabbard, forming a brief and dazzling white flash. The entire sword-drawing action was done in one go, full of explosive power and decisive beauty, perfectly blending the dynamics of the characters with the static environment, creating a poetic martial arts atmosphere that is both tranquil and full of tension.", + "seed": 7919, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_5/index_2256_seed_7919_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24883626364171504, + "vlm_logprob": -0.24883626364171504, + "vlm_yes_no_margin": 2.324999952316284, + "vlm_token_credit": -0.24883626364171504, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2556888461112976 + ], + "opposite_logprob_tokens": [ + -1.5056889057159424 + ], + "logprob_reward": -0.2556888461112976, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2556888461112976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25465497374534607, + 0.03924690932035446, + 0.3217026889324188, + 0.061392564326524734, + 0.07290000468492508, + 0.06043413281440735, + 0.03964892029762268, + 0.045554738491773605, + 0.039288006722927094, + 0.0651770532131195 + ], + "credit_map": [ + 0.040188632905483246, + 0.010606666095554829, + 0.08990224450826645, + 0.009587232954800129, + 0.009625422768294811, + 0.02399977669119835, + 0.011020762845873833, + 0.059724222868680954, + 0.00672769732773304, + 0.004128712695091963, + 0.004283769987523556, + 0.005247758701443672, + 0.004237479064613581, + 0.007634651847183704, + 0.004081630147993565, + 0.0029052086174488068, + 0.11133444309234619, + 0.009026242420077324, + 0.014725635759532452, + 0.021918701007962227, + 0.017642494291067123, + 0.11931690573692322, + 0.012603589333593845, + 0.015134691260755062, + 0.016173476353287697, + 0.003618597984313965, + 0.006026472896337509, + 0.006842467933893204, + 0.005105080083012581, + 0.016361894086003304, + 0.00362989935092628, + 0.0036346770357340574, + 0.016560323536396027, + 0.0049554104916751385, + 0.005418666638433933, + 0.006241136696189642, + 0.008154401555657387, + 0.023667575791478157, + 0.003812691429629922, + 0.004089800640940666, + 0.019389169290661812, + 0.002719172975048423, + 0.002452549757435918, + 0.002182872500270605, + 0.0030148986261337996, + 0.023939721286296844, + 0.0038904035463929176, + 0.002845345763489604, + 0.005099940113723278, + 0.002827439922839403, + 0.004208110272884369, + 0.005016301292926073, + 0.0025614998303353786, + 0.0061561716720461845, + 0.009884396567940712, + 0.0038950585294514894, + 0.01207646168768406, + 0.00370741612277925, + 0.003453932935371995, + 0.0034950145054608583, + 0.002663449617102742, + 0.013456814922392368, + 0.0032997089438140392, + 0.003401941154152155, + 0.012651407159864902, + 0.003190138842910528, + 0.0031490474939346313, + 0.001965037314221263, + 0.002010243944823742, + 0.004743572790175676, + 0.0018059266731142998, + 0.009772634133696556, + 0.008572718128561974, + 0.00306258094497025, + 0.005174038000404835, + 0.0030288000125437975, + 0.0053457235917449, + 0.02308279648423195, + 0.002984273713082075, + 0.013926125131547451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.2556888461112976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820533163845539 + ], + "opposite_logprob_tokens": [ + -4.018205165863037 + ], + "logprob_reward": -0.01820533163845539, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01820533163845539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26932579278945923, + 0.07498262077569962, + 0.1018303707242012, + 0.06769805401563644, + 0.07119736075401306, + 0.2090054303407669, + 0.044260695576667786, + 0.0739288404583931, + 0.05878628045320511, + 0.028984561562538147 + ], + "credit_map": [ + 0.032596439123153687, + 0.00976738054305315, + 0.08224717527627945, + 0.010768728330731392, + 0.01047710794955492, + 0.029902298003435135, + 0.013666738756000996, + 0.07989992946386337, + 0.022084878757596016, + 0.004058277700096369, + 0.007689533289521933, + 0.003506893990561366, + 0.004523149691522121, + 0.025093091651797295, + 0.004217593465000391, + 0.0038091992028057575, + 0.029612774029374123, + 0.0049828109331429005, + 0.007425377145409584, + 0.006948084570467472, + 0.006222591735422611, + 0.03077969327569008, + 0.006866143550723791, + 0.008992896415293217, + 0.013971766456961632, + 0.004962547216564417, + 0.010281790047883987, + 0.00853925384581089, + 0.005799959413707256, + 0.01415148563683033, + 0.003932836465537548, + 0.006058417726308107, + 0.009871101006865501, + 0.006927090231329203, + 0.00973076093941927, + 0.006394064519554377, + 0.006689147092401981, + 0.02082081139087677, + 0.0067055318504571915, + 0.0040588523261249065, + 0.07035595923662186, + 0.008728030137717724, + 0.008225490339100361, + 0.004965673666447401, + 0.009249619208276272, + 0.08613348752260208, + 0.011997098103165627, + 0.009350082837045193, + 0.010306785814464092, + 0.0032982141710817814, + 0.004758186172693968, + 0.0028228305745869875, + 0.004099368583410978, + 0.011047041974961758, + 0.00426369858905673, + 0.0036645676009356976, + 0.01060251984745264, + 0.0057101561687886715, + 0.014258639886975288, + 0.007092350162565708, + 0.007518393453210592, + 0.01077272929251194, + 0.006434575654566288, + 0.011539475060999393, + 0.02142093889415264, + 0.0032540294341742992, + 0.016154319047927856, + 0.0029947832226753235, + 0.003358960384503007, + 0.006275922525674105, + 0.0016236534575000405, + 0.0037036717403680086, + 0.002510755555704236, + 0.0027018424589186907, + 0.0028034120332449675, + 0.00185753358528018, + 0.0038020419888198376, + 0.006340471561998129, + 0.0017587057081982493, + 0.0072097983211278915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.01820533163845539 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329517900943756 + ], + "opposite_logprob_tokens": [ + -3.7732951641082764 + ], + "logprob_reward": -0.02329517900943756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329517900943756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6097073554992676, + 0.08639677613973618, + 0.12655098736286163, + 0.018500078469514847, + 0.016098108142614365, + 0.020113004371523857, + 0.0249482449144125, + 0.019187090918421745, + 0.054724350571632385, + 0.02377402037382126 + ], + "credit_map": [ + 0.11347383260726929, + 0.02189878188073635, + 0.23084793984889984, + 0.022690825164318085, + 0.018409613519906998, + 0.05164032801985741, + 0.019118657335639, + 0.1316273808479309, + 0.020469535142183304, + 0.004148606210947037, + 0.007859276607632637, + 0.006574202328920364, + 0.010281655006110668, + 0.022857964038848877, + 0.00790628045797348, + 0.006299246568232775, + 0.03811078146100044, + 0.005530616734176874, + 0.008532592095434666, + 0.009527911432087421, + 0.006802609656006098, + 0.04144271835684776, + 0.010151519440114498, + 0.006452224217355251, + 0.0022576760966330767, + 0.0019251145422458649, + 0.00266958586871624, + 0.0036995236296206713, + 0.0016525917453691363, + 0.002603127621114254, + 0.0016635560896247625, + 0.002028901129961014, + 0.0030645057559013367, + 0.0010876207379624248, + 0.0020045912824571133, + 0.0020962839480489492, + 0.0016377746360376477, + 0.003866118611767888, + 0.00110488454811275, + 0.0012363274581730366, + 0.002988226944580674, + 0.0029814820736646652, + 0.002076271455734968, + 0.0017526950687170029, + 0.0022307890467345715, + 0.0037549000699073076, + 0.0027269371785223484, + 0.0016017023008316755, + 0.00396466301754117, + 0.0018277911003679037, + 0.003126859897747636, + 0.003758843056857586, + 0.0016818174626678228, + 0.0041642324067652225, + 0.0043508922681212425, + 0.002073145005851984, + 0.003231822047382593, + 0.0022899240721017122, + 0.001434622798115015, + 0.003088277531787753, + 0.0018560688477009535, + 0.004044360015541315, + 0.0019435506546869874, + 0.0012984650675207376, + 0.01429671049118042, + 0.0032779700122773647, + 0.018986456096172333, + 0.0034522509668022394, + 0.0026367316022515297, + 0.0063030896708369255, + 0.001855130074545741, + 0.0039160107262432575, + 0.003121374873444438, + 0.0014591808430850506, + 0.0032576865050941706, + 0.002606123685836792, + 0.002363999141380191, + 0.002852904610335827, + 0.002132852328941226, + 0.005979896988719702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.02329517900943756 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865841194987297 + ], + "opposite_logprob_tokens": [ + -3.0486583709716797 + ], + "logprob_reward": -0.04865841194987297, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865841194987297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07632862776517868, + 0.02976279705762863, + 0.08553112298250198, + 0.10916449874639511, + 0.07392360270023346, + 0.20016208291053772, + 0.02354533225297928, + 0.015624653548002243, + 0.16277842223644257, + 0.22317878901958466 + ], + "credit_map": [ + 0.014447759836912155, + 0.008471385575830936, + 0.018266554921865463, + 0.006340181455016136, + 0.005131840240210295, + 0.010306811891496181, + 0.006175340618938208, + 0.00718875415623188, + 0.004677650518715382, + 0.003122800961136818, + 0.0033588127698749304, + 0.0028763783629983664, + 0.0037882309406995773, + 0.004998866468667984, + 0.003361306618899107, + 0.0035787508822977543, + 0.014221911318600178, + 0.005728140473365784, + 0.008894816040992737, + 0.00753405736759305, + 0.006121465936303139, + 0.018567640334367752, + 0.010059233754873276, + 0.01440386101603508, + 0.027445176616311073, + 0.006387821398675442, + 0.012578481808304787, + 0.011071774177253246, + 0.008706641383469105, + 0.027901453897356987, + 0.006359218619763851, + 0.008713923394680023, + 0.012006795965135098, + 0.010968975722789764, + 0.006523407530039549, + 0.008441326208412647, + 0.0090405298396945, + 0.012865612283349037, + 0.009676968678832054, + 0.00439998647198081, + 0.06251675635576248, + 0.009444786235690117, + 0.010036258026957512, + 0.007916468195617199, + 0.00963183306157589, + 0.07711590826511383, + 0.012914249673485756, + 0.010585827752947807, + 0.0031529366970062256, + 0.002122453646734357, + 0.003880961798131466, + 0.0028285279404371977, + 0.002406437648460269, + 0.004185589030385017, + 0.0027780018281191587, + 0.002190423896536231, + 0.0020617127884179354, + 0.0015404431615024805, + 0.001714156474918127, + 0.0028288511093705893, + 0.002035704208537936, + 0.0022899608593434095, + 0.0017524013528600335, + 0.0014014234766364098, + 0.04286833852529526, + 0.006387405563145876, + 0.06043362244963646, + 0.010271535255014896, + 0.0053264484740793705, + 0.015410277992486954, + 0.0046815574169158936, + 0.01739923469722271, + 0.024690279737114906, + 0.017043860629200935, + 0.01633172295987606, + 0.01743561029434204, + 0.019144875928759575, + 0.06866537779569626, + 0.017995618283748627, + 0.041871439665555954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.04865841194987297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a rich, freehand ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a bamboo forest with moonlight filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a long, flowing, dark cyan Tang Dynasty robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character draw an ancient longsword from its scabbard with a sudden thrust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983335494995117 + ], + "opposite_logprob_tokens": [ + -0.5233335494995117 + ], + "logprob_reward": -0.8983335494995117, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983335494995117, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3090342581272125, + 0.029820704832673073, + 0.2897801399230957, + 0.10849906504154205, + 0.054363448172807693, + 0.0414322130382061, + 0.02652101218700409, + 0.018435923382639885, + 0.06369049847126007, + 0.0584227629005909 + ], + "credit_map": [ + 0.03649992123246193, + 0.00991947203874588, + 0.0965726375579834, + 0.0095287524163723, + 0.012731485068798065, + 0.03081885352730751, + 0.016113251447677612, + 0.09684987366199493, + 0.004732821136713028, + 0.00251809973269701, + 0.00323630403727293, + 0.0031758423428982496, + 0.004555969964712858, + 0.004933028481900692, + 0.003491842420771718, + 0.0031767943874001503, + 0.09992846101522446, + 0.009056119248270988, + 0.013639149256050587, + 0.019800176844000816, + 0.0164771880954504, + 0.10757669806480408, + 0.01171836443245411, + 0.011583957821130753, + 0.022457540035247803, + 0.00832396186888218, + 0.015352785587310791, + 0.014402459375560284, + 0.009959557093679905, + 0.02162533439695835, + 0.008743765763938427, + 0.007633658591657877, + 0.007998070679605007, + 0.004817657172679901, + 0.00768212741240859, + 0.004910240415483713, + 0.004313590005040169, + 0.015955768525600433, + 0.004709013272076845, + 0.003976978827267885, + 0.009826677851378918, + 0.0030321525409817696, + 0.0031822596210986376, + 0.003033451270312071, + 0.003014747751876712, + 0.011704945005476475, + 0.004576321691274643, + 0.003061656840145588, + 0.005495857447385788, + 0.0018126486102119088, + 0.002082271035760641, + 0.003125343006104231, + 0.0020391882862895727, + 0.005720899906009436, + 0.004615099634975195, + 0.0016297006513923407, + 0.002218638313934207, + 0.0015947030624374747, + 0.0025202338583767414, + 0.003972770646214485, + 0.0018705107504501939, + 0.002346066990867257, + 0.0018323453841730952, + 0.002080653328448534, + 0.019358156248927116, + 0.00501906918361783, + 0.010001133196055889, + 0.005597306881099939, + 0.004298788961023092, + 0.00662900460883975, + 0.0035746616777032614, + 0.0092123718932271, + 0.0072756013832986355, + 0.0036893868818879128, + 0.004917448852211237, + 0.004499912261962891, + 0.005606560967862606, + 0.01684998720884323, + 0.005542821250855923, + 0.010041041299700737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.8983335494995117 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_tensor([5857, 5857])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_tensor([5857, 5857])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ea55e39c1560dd3640093bb614b53c7647f32ddf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_tensor([5857, 5857])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, solemn camera tilts upward, first sweeping across a cracked asphalt surface streaked with spiderweb-like cracks, before slowly revealing a full-length portrait of a young man standing tall. In the extreme black and white HDR lighting, a sharp light source shines from above and to the side, outlining his body with a dazzling edge of brightness, while the other side is completely immersed in a bottomless, inky shadow. He sports a meticulously styled, now messy, black crew cut, with a few strands of hair, damp with sweat, plastered to his forehead. He clutches a cold, metal baseball bat tightly in his hands. The shaft gleams harshly in the bright sunlight, its surface dented and scratched. He wears a frayed black leather jacket, half-zipped, revealing a stained white T-shirt underneath. His knuckles are swollen and red from raw skin. His jaw is clenched, his lips pressed into a resolute line as he grits his teeth in an effort of endurance. Sweat mixed with dust trickles down his forehead, streaking across his young face, which is covered in tiny scratches and bruises. Though his body sways slightly from exhaustion, his legs remain firmly rooted to the ground, his back straight. The high-contrast light and shadow vividly depict his scarred yet unyielding posture, creating an atmosphere of intense tension, a sense of tragic toughness, and resilience.", + "seed": 1371, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_5857_seed_1371_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04010159336030483, + "vlm_logprob": -0.04010159336030483, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04010159336030483, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09673362970352173 + ], + "opposite_logprob_tokens": [ + -2.596733570098877 + ], + "logprob_reward": -0.09673362970352173, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09673362970352173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18373052775859833, + 0.2116139978170395, + 0.1404232382774353, + 0.11340118944644928, + 0.07173645496368408, + 0.07312346249818802, + 0.04914732649922371, + 0.10969504714012146, + 0.028162917122244835, + 0.018965858966112137 + ], + "credit_map": [ + 0.022718781605362892, + 0.017930908128619194, + 0.013163871131837368, + 0.019440218806266785, + 0.033617470413446426, + 0.025598466396331787, + 0.018827760592103004, + 0.03243304416537285, + 0.01725723408162594, + 0.013355963863432407, + 0.016306571662425995, + 0.01276136003434658, + 0.03383220359683037, + 0.03577977046370506, + 0.02728523313999176, + 0.05503564700484276, + 0.009604225866496563, + 0.010315967723727226, + 0.011101635172963142, + 0.010307767428457737, + 0.021987617015838623, + 0.024455377832055092, + 0.032524947077035904, + 0.02012571133673191, + 0.015800882130861282, + 0.015119240619242191, + 0.013045270927250385, + 0.016116973012685776, + 0.012748831883072853, + 0.01587967574596405, + 0.011934187263250351, + 0.012756130658090115, + 0.009365551173686981, + 0.007883355021476746, + 0.004988776054233313, + 0.008008267730474472, + 0.011254771612584591, + 0.00836989376693964, + 0.010698456317186356, + 0.011167379096150398, + 0.008413402363657951, + 0.005893267225474119, + 0.009404013864696026, + 0.009308802895247936, + 0.009188788942992687, + 0.010967414826154709, + 0.008475717157125473, + 0.01147205289453268, + 0.006127259694039822, + 0.005260095931589603, + 0.0054649608209729195, + 0.00320449355058372, + 0.007454930804669857, + 0.007567671127617359, + 0.0067847599275410175, + 0.007283158600330353, + 0.0073387278243899345, + 0.013832168653607368, + 0.008076117374002934, + 0.01293964870274067, + 0.01902376115322113, + 0.016314541921019554, + 0.01644248515367508, + 0.01572759449481964, + 0.0034856258425861597, + 0.003320986870676279, + 0.0025948849506676197, + 0.002154305577278137, + 0.005005992017686367, + 0.003959649708122015, + 0.003987959120422602, + 0.0036535144317895174, + 0.0015256208134815097, + 0.0021023605950176716, + 0.0013930911663919687, + 0.0014984364388510585, + 0.0036486706230789423, + 0.002688766922801733, + 0.00293720792979002, + 0.0031717021483927965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.09673362970352173 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029842983931303024 + ], + "opposite_logprob_tokens": [ + -3.5298430919647217 + ], + "logprob_reward": -0.029842983931303024, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029842983931303024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13424994051456451, + 0.16779197752475739, + 0.19293472170829773, + 0.07772593200206757, + 0.11386293172836304, + 0.09980902075767517, + 0.03350047767162323, + 0.08607988804578781, + 0.03116106614470482, + 0.06288401037454605 + ], + "credit_map": [ + 0.01957588456571102, + 0.015560024417936802, + 0.013559486716985703, + 0.013999862596392632, + 0.019610000774264336, + 0.01783844828605652, + 0.01027454249560833, + 0.023831689730286598, + 0.010758524760603905, + 0.014923883602023125, + 0.011763626709580421, + 0.019291182979941368, + 0.03221961110830307, + 0.02531259134411812, + 0.025187015533447266, + 0.028335541486740112, + 0.013536977581679821, + 0.017839020118117332, + 0.020098568871617317, + 0.016174744814634323, + 0.024770675227046013, + 0.046682681888341904, + 0.022822512313723564, + 0.031009558588266373, + 0.0065590133890509605, + 0.005201366264373064, + 0.00569132948294282, + 0.008533060550689697, + 0.012667059898376465, + 0.013878950849175453, + 0.01625974290072918, + 0.008935412392020226, + 0.013316630385816097, + 0.00862718652933836, + 0.010287303477525711, + 0.008934765122830868, + 0.016833754256367683, + 0.01965208910405636, + 0.014199560508131981, + 0.022011641412973404, + 0.01412131916731596, + 0.010841225273907185, + 0.01112725306302309, + 0.012740528210997581, + 0.012258772738277912, + 0.014201389625668526, + 0.012043346650898457, + 0.012475186958909035, + 0.004275967832654715, + 0.003937442786991596, + 0.0031370557844638824, + 0.0020348697435110807, + 0.004748697392642498, + 0.005152438301593065, + 0.005155711900442839, + 0.005058294627815485, + 0.006613396108150482, + 0.010830448940396309, + 0.00629411730915308, + 0.010008682496845722, + 0.01575278304517269, + 0.012306016869843006, + 0.011775116436183453, + 0.012499328702688217, + 0.006646555382758379, + 0.002608791459351778, + 0.003233428345993161, + 0.0016996765043586493, + 0.003635927801951766, + 0.00495328800752759, + 0.0033832951448857784, + 0.005000103265047073, + 0.013038055039942265, + 0.003919308539479971, + 0.007309812121093273, + 0.0034502779599279165, + 0.007994870655238628, + 0.011073898524045944, + 0.005701356567442417, + 0.01039643120020628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.029842983931303024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298904299736023 + ], + "opposite_logprob_tokens": [ + -3.773298978805542 + ], + "logprob_reward": -0.023298904299736023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298904299736023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15620380640029907, + 0.06708411127328873, + 0.035542942583560944, + 0.05761902034282684, + 0.04181633144617081, + 0.14478585124015808, + 0.12441509962081909, + 0.2903216779232025, + 0.03511647880077362, + 0.04709466174244881 + ], + "credit_map": [ + 0.01894243434071541, + 0.013118312694132328, + 0.010489080101251602, + 0.025035342201590538, + 0.02503209002315998, + 0.019373074173927307, + 0.019603228196501732, + 0.024610238149762154, + 0.00628397474065423, + 0.005495354067534208, + 0.005143535323441029, + 0.004891378339380026, + 0.015341606922447681, + 0.01146424375474453, + 0.008301043882966042, + 0.01016297284513712, + 0.002437551971524954, + 0.002633601427078247, + 0.00513193616643548, + 0.0026387188117951155, + 0.006828921381384134, + 0.005562304984778166, + 0.006293866783380508, + 0.004016042221337557, + 0.007255136501044035, + 0.0054462063126266, + 0.007150650955736637, + 0.007515266537666321, + 0.006999106612056494, + 0.0077439951710402966, + 0.009134515188634396, + 0.0063741435296833515, + 0.005559714511036873, + 0.0043130600824952126, + 0.004298749379813671, + 0.003930100705474615, + 0.007158136460930109, + 0.006008945871144533, + 0.005287110805511475, + 0.005260510835796595, + 0.01814734935760498, + 0.014856834895908833, + 0.018314270302653313, + 0.019788844510912895, + 0.015117364935576916, + 0.02212575450539589, + 0.014599624089896679, + 0.021835822612047195, + 0.010277635417878628, + 0.014241649769246578, + 0.012031257152557373, + 0.008626568131148815, + 0.02384689450263977, + 0.015724193304777145, + 0.02149488590657711, + 0.018172018229961395, + 0.020676109939813614, + 0.03673327714204788, + 0.01991608552634716, + 0.03451525792479515, + 0.05094408243894577, + 0.04212290793657303, + 0.040643200278282166, + 0.0447707399725914, + 0.0030945201870054007, + 0.0031985784880816936, + 0.0028548859991133213, + 0.0030554833356291056, + 0.005176765378564596, + 0.005114106461405754, + 0.00671090604737401, + 0.005911233369261026, + 0.0038482756353914738, + 0.0036066146567463875, + 0.0037673581391572952, + 0.0036765988916158676, + 0.009898512624204159, + 0.006522372364997864, + 0.008483110927045345, + 0.0072918180376291275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.023298904299736023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809156268835068 + ], + "opposite_logprob_tokens": [ + -3.2880916595458984 + ], + "logprob_reward": -0.03809156268835068, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809156268835068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13723716139793396, + 0.0780671164393425, + 0.040903523564338684, + 0.0888795554637909, + 0.21997369825839996, + 0.08240588009357452, + 0.028360212221741676, + 0.16019530594348907, + 0.09416916221380234, + 0.06980826705694199 + ], + "credit_map": [ + 0.014102010056376457, + 0.010694541968405247, + 0.01117353793233633, + 0.0173024982213974, + 0.02733691781759262, + 0.017605625092983246, + 0.01667265221476555, + 0.022349389269948006, + 0.008039407432079315, + 0.006665457971394062, + 0.005595923867076635, + 0.0040098801255226135, + 0.016711730509996414, + 0.015095612965524197, + 0.009844924323260784, + 0.012104181572794914, + 0.003984838258475065, + 0.004263033624738455, + 0.005865642335265875, + 0.004165003541857004, + 0.005534839350730181, + 0.005900454241782427, + 0.006125574931502342, + 0.005064139608293772, + 0.007971437647938728, + 0.00653350492939353, + 0.006772476248443127, + 0.00953991711139679, + 0.01465182937681675, + 0.013793112710118294, + 0.018608948215842247, + 0.011008336208760738, + 0.023802941665053368, + 0.019496865570545197, + 0.02129269763827324, + 0.020894166082143784, + 0.03302112966775894, + 0.03647738695144653, + 0.031533461064100266, + 0.03345508500933647, + 0.008982169441878796, + 0.008081930689513683, + 0.01037543173879385, + 0.007529049646109343, + 0.011489608325064182, + 0.01147783175110817, + 0.009272506460547447, + 0.015197359025478363, + 0.003416888415813446, + 0.002872623037546873, + 0.003036012640222907, + 0.001973407343029976, + 0.004779001697897911, + 0.0038469061255455017, + 0.004459842573851347, + 0.003975534811615944, + 0.011968151666224003, + 0.018120644614100456, + 0.012198763899505138, + 0.01761787384748459, + 0.02773721143603325, + 0.024203842505812645, + 0.022107847034931183, + 0.026240991428494453, + 0.006803648546338081, + 0.005780254490673542, + 0.010212303139269352, + 0.003747982671484351, + 0.015576953999698162, + 0.016518868505954742, + 0.01663588173687458, + 0.018893275409936905, + 0.003959724213927984, + 0.0052865045145154, + 0.004402101505547762, + 0.006521059665828943, + 0.014395219273865223, + 0.010367701761424541, + 0.016254030168056488, + 0.008621936663985252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.03809156268835068 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540886178612709 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540886178612709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540886178612709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.563724160194397, + 0.0365319661796093, + 0.07196459174156189, + 0.026840487495064735, + 0.04570150747895241, + 0.07482516020536423, + 0.03491009771823883, + 0.07858475297689438, + 0.03486115485429764, + 0.032056115567684174 + ], + "credit_map": [ + 0.05467262119054794, + 0.05861367657780647, + 0.03388786315917969, + 0.11825329810380936, + 0.07959070056676865, + 0.09255951642990112, + 0.07942502945661545, + 0.04672151431441307, + 0.002762954216450453, + 0.0030387849546968937, + 0.00256236526183784, + 0.002603523898869753, + 0.004751540720462799, + 0.006749667227268219, + 0.004520298447459936, + 0.009542835876345634, + 0.004716895055025816, + 0.004817307461053133, + 0.006066438741981983, + 0.0064244517125189304, + 0.011932777240872383, + 0.01368030160665512, + 0.012492385692894459, + 0.011834044009447098, + 0.002059561898931861, + 0.0020144975278526545, + 0.0019959439523518085, + 0.002005869522690773, + 0.004061536397784948, + 0.004659184254705906, + 0.006718226708471775, + 0.0033256688620895147, + 0.005197442136704922, + 0.0034315946977585554, + 0.00438519474118948, + 0.0031964906957000494, + 0.005763221066445112, + 0.006934399250894785, + 0.005537388846278191, + 0.011255782097578049, + 0.007676105014979839, + 0.007896637544035912, + 0.008266338147222996, + 0.007946684025228024, + 0.010170918889343739, + 0.011043308302760124, + 0.008447886444628239, + 0.013377293944358826, + 0.003671601414680481, + 0.0034975502640008926, + 0.003333751345053315, + 0.0037231154274195433, + 0.00525366235524416, + 0.00532532949000597, + 0.005901659373193979, + 0.004203430376946926, + 0.006115061230957508, + 0.012332434765994549, + 0.00464587053284049, + 0.009181635454297066, + 0.014372067525982857, + 0.010847891680896282, + 0.009560737758874893, + 0.01152905821800232, + 0.003248833818361163, + 0.0030253457371145487, + 0.004826415795832872, + 0.002358649857342243, + 0.005992391612380743, + 0.005899648182094097, + 0.00480281189084053, + 0.004707057494670153, + 0.002245127921923995, + 0.002384319668635726, + 0.002732747932896018, + 0.0024927931372076273, + 0.006286601535975933, + 0.0052083805203437805, + 0.00525325583294034, + 0.00545289134606719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.012540886178612709 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, solemn camera tilts upward, first sweeping across a cracked asphalt surface streaked with spiderweb-like cracks, before slowly revealing a full-length portrait of a young man standing tall. In the extreme black and white HDR lighting, a sharp light source shines from above and to the side, outlining his body with a dazzling edge of brightness, while the other side is completely immersed in a bottomless, inky shadow. He sports a meticulously styled, now messy, black crew cut, with a few strands of hair, damp with sweat, plastered to his forehead. He clutches a cold, metal baseball bat tightly in his hands. The shaft gleams harshly in the bright sunlight, its surface dented and scratched. He wears a frayed black leather jacket, half-zipped, revealing a stained white T-shirt underneath. His knuckles are swollen and red from raw skin. His jaw is clenched, his lips pressed into a resolute line as he grits his teeth in an effort of endurance. Sweat mixed with dust trickles down his forehead, streaking across his young face, which is covered in tiny scratches and bruises. Though his body sways slightly from exhaustion, his legs remain firmly rooted to the ground, his back straight. The high-contrast light and shadow vividly depict his scarred yet unyielding posture, creating an atmosphere of intense tension, a sense of tragic toughness, and resilience.", + "seed": 1372, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_5857_seed_1372_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0377425080165267, + "vlm_logprob": -0.0377425080165267, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0377425080165267, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013586431741714478 + ], + "opposite_logprob_tokens": [ + -5.138586521148682 + ], + "logprob_reward": -0.013586431741714478, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.013586431741714478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15198029577732086, + 0.23216792941093445, + 0.08907096832990646, + 0.06544860452413559, + 0.07365680485963821, + 0.08416072279214859, + 0.07033254951238632, + 0.08284133672714233, + 0.11859476566314697, + 0.031745947897434235 + ], + "credit_map": [ + 0.0205170176923275, + 0.017132218927145004, + 0.020288128405809402, + 0.013285097666084766, + 0.01600244827568531, + 0.03220440074801445, + 0.011604586616158485, + 0.020946405827999115, + 0.030064396560192108, + 0.04768034815788269, + 0.022057171911001205, + 0.01733091101050377, + 0.05059152469038963, + 0.031148996204137802, + 0.012912026606500149, + 0.02038257010281086, + 0.012874322943389416, + 0.010599632747471333, + 0.007227140478789806, + 0.008640921674668789, + 0.007237883750349283, + 0.013498791493475437, + 0.013740983791649342, + 0.015251300297677517, + 0.007128940429538488, + 0.009198866784572601, + 0.008037039078772068, + 0.005549320951104164, + 0.011860061436891556, + 0.008885649032890797, + 0.008204178884625435, + 0.006584559101611376, + 0.008184939622879028, + 0.013705898076295853, + 0.006467990577220917, + 0.008374075405299664, + 0.009826096706092358, + 0.009579101577401161, + 0.008768700063228607, + 0.008750011213123798, + 0.010762698948383331, + 0.008505329489707947, + 0.01046478096395731, + 0.018204685300588608, + 0.0074921175837516785, + 0.009654429741203785, + 0.012191938236355782, + 0.006884745322167873, + 0.007226873654872179, + 0.00770034221932292, + 0.012464849278330803, + 0.00841132178902626, + 0.008295158855617046, + 0.008769950829446316, + 0.009371785447001457, + 0.008092275820672512, + 0.012431344948709011, + 0.004974149167537689, + 0.022226668894290924, + 0.004906571935862303, + 0.010419194586575031, + 0.012969298288226128, + 0.006705469451844692, + 0.008208639919757843, + 0.01717025600373745, + 0.007352097425609827, + 0.041739773005247116, + 0.0059880707412958145, + 0.009401191957294941, + 0.023542916402220726, + 0.005723563022911549, + 0.00767690921202302, + 0.003970273304730654, + 0.0032345459330826998, + 0.009758330881595612, + 0.0016932801809161901, + 0.0037725064903497696, + 0.005289069376885891, + 0.0012822935823351145, + 0.0027456542011350393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.013586431741714478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906073123216629 + ], + "opposite_logprob_tokens": [ + -2.4640607833862305 + ], + "logprob_reward": -0.08906073123216629, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906073123216629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059732481837272644, + 0.42296290397644043, + 0.27436792850494385, + 0.04755657538771629, + 0.029488740488886833, + 0.0250003132969141, + 0.08086566627025604, + 0.03573817387223244, + 0.015309745445847511, + 0.00897746067494154 + ], + "credit_map": [ + 0.006819215603172779, + 0.011030631139874458, + 0.008456159383058548, + 0.0055993604473769665, + 0.007612229324877262, + 0.006964551750570536, + 0.007243502419441938, + 0.006006830837577581, + 0.0553940087556839, + 0.07959525287151337, + 0.040056027472019196, + 0.031059030443429947, + 0.0947079136967659, + 0.05823899805545807, + 0.02396509423851967, + 0.03994658961892128, + 0.04421154037117958, + 0.023195119574666023, + 0.024167733266949654, + 0.025850635021924973, + 0.0242869034409523, + 0.04282033070921898, + 0.04096988961100578, + 0.04886578395962715, + 0.005094812251627445, + 0.008361033163964748, + 0.006129502784460783, + 0.004001491237431765, + 0.006829853169620037, + 0.006587791256606579, + 0.0045244209468364716, + 0.0060276733711361885, + 0.0043779173865914345, + 0.004359832499176264, + 0.002994012786075473, + 0.0036647850647568703, + 0.0034799983259290457, + 0.0037531680427491665, + 0.003918475937098265, + 0.0029405499808490276, + 0.003910708241164684, + 0.0026765232905745506, + 0.004346036817878485, + 0.002225905656814575, + 0.002069719135761261, + 0.003700234228745103, + 0.0028156759217381477, + 0.003255510935559869, + 0.014889604412019253, + 0.0036634865682572126, + 0.027818698436021805, + 0.004558749962598085, + 0.0059052021242678165, + 0.01323765143752098, + 0.004472901578992605, + 0.006319369189441204, + 0.004192707594484091, + 0.002688011387363076, + 0.007165741641074419, + 0.0028564571402966976, + 0.005403687246143818, + 0.005581929814070463, + 0.0034827839117497206, + 0.004366857931017876, + 0.0021615037694573402, + 0.0014978796243667603, + 0.004200139082968235, + 0.0008715247386135161, + 0.0018464078893885016, + 0.00247706426307559, + 0.0009901084704324603, + 0.0012651184806600213, + 0.0012382040731608868, + 0.0007915490423329175, + 0.002461525145918131, + 0.0005585175240412354, + 0.0009800699772313237, + 0.001470171962864697, + 0.0006272023310884833, + 0.0008502207929268479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.08906073123216629 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865920543670654 + ], + "opposite_logprob_tokens": [ + -3.048659324645996 + ], + "logprob_reward": -0.04865920543670654, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865920543670654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0974263846874237, + 0.3049188554286957, + 0.11031705141067505, + 0.10694360733032227, + 0.06751333922147751, + 0.06867922097444534, + 0.10764845460653305, + 0.057905275374650955, + 0.044201821088790894, + 0.03444604203104973 + ], + "credit_map": [ + 0.015120310708880424, + 0.009728181175887585, + 0.011862363666296005, + 0.01028972677886486, + 0.00733998091891408, + 0.017309455201029778, + 0.008310636505484581, + 0.01746571622788906, + 0.044690437614917755, + 0.05610944330692291, + 0.029314780607819557, + 0.0245413426309824, + 0.06488081812858582, + 0.04136183112859726, + 0.016760418191552162, + 0.027259759604930878, + 0.0174204483628273, + 0.009707203134894371, + 0.009958324022591114, + 0.009342428296804428, + 0.010737162083387375, + 0.01688600704073906, + 0.015237036161124706, + 0.021028432995080948, + 0.010688753798604012, + 0.012013589963316917, + 0.01511464174836874, + 0.009447450749576092, + 0.015360933728516102, + 0.01628437452018261, + 0.014667309820652008, + 0.013366535305976868, + 0.006459767930209637, + 0.008982611820101738, + 0.007690649013966322, + 0.008776914328336716, + 0.00981717836111784, + 0.008155560120940208, + 0.010196295566856861, + 0.007434357423335314, + 0.00653432309627533, + 0.006543340627104044, + 0.00918645691126585, + 0.01784609630703926, + 0.005349085666239262, + 0.007931118831038475, + 0.00815058033913374, + 0.00713821267709136, + 0.011626801453530788, + 0.005968164652585983, + 0.01634603925049305, + 0.007268686778843403, + 0.006793946027755737, + 0.019508063793182373, + 0.012028399854898453, + 0.028108341619372368, + 0.007523294538259506, + 0.004167646635323763, + 0.012837275862693787, + 0.004604442045092583, + 0.006942306645214558, + 0.00782851967960596, + 0.00518176332116127, + 0.008820018731057644, + 0.00559933902695775, + 0.0032633899245411158, + 0.014678562991321087, + 0.00281970901414752, + 0.004093517549335957, + 0.008040904998779297, + 0.0024510466028004885, + 0.0032553446944803, + 0.003864649450406432, + 0.004370685201138258, + 0.006459007039666176, + 0.0026503480039536953, + 0.0057049039751291275, + 0.004933777265250683, + 0.002770966850221157, + 0.0036917030811309814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.04865920543670654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026332726702094078 + ], + "opposite_logprob_tokens": [ + -3.6513326168060303 + ], + "logprob_reward": -0.026332726702094078, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332726702094078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16303782165050507, + 0.31510186195373535, + 0.12066475301980972, + 0.03666817024350166, + 0.06331437081098557, + 0.05298066511750221, + 0.1409260332584381, + 0.06595078855752945, + 0.024821797385811806, + 0.016533762216567993 + ], + "credit_map": [ + 0.025214705616235733, + 0.01804906688630581, + 0.0187299195677042, + 0.016099343076348305, + 0.014490830712020397, + 0.027787912636995316, + 0.014004159718751907, + 0.028661897405982018, + 0.0423574335873127, + 0.059005193412303925, + 0.029390528798103333, + 0.025169288739562035, + 0.07012685388326645, + 0.04308364912867546, + 0.017558583989739418, + 0.028410330414772034, + 0.019814923405647278, + 0.009835018776357174, + 0.011241341009736061, + 0.009610429406166077, + 0.011828744783997536, + 0.01881973259150982, + 0.017394468188285828, + 0.022120090201497078, + 0.003817860735580325, + 0.003383012255653739, + 0.0049124062061309814, + 0.003718589199706912, + 0.004852812271565199, + 0.005614063236862421, + 0.0052958279848098755, + 0.005073598586022854, + 0.008503533899784088, + 0.01399692241102457, + 0.005501378793269396, + 0.007024212274700403, + 0.008920623920857906, + 0.007256065495312214, + 0.00753021938726306, + 0.004581419285386801, + 0.008162730373442173, + 0.004067409783601761, + 0.006818963214755058, + 0.004682276863604784, + 0.006928051821887493, + 0.007920853793621063, + 0.00725494883954525, + 0.007145430892705917, + 0.013892482034862041, + 0.009317224845290184, + 0.02816656231880188, + 0.010708755813539028, + 0.014473645016551018, + 0.0247337743639946, + 0.01407705619931221, + 0.0255565345287323, + 0.005662693176418543, + 0.005600656382739544, + 0.010630019940435886, + 0.006650907918810844, + 0.012583302333950996, + 0.0077895489521324635, + 0.010110361501574516, + 0.006923298351466656, + 0.0032749322708696127, + 0.0021529921796172857, + 0.007328995503485203, + 0.0016260880511254072, + 0.0024237227626144886, + 0.004418023396283388, + 0.0013067356776446104, + 0.0022903059143573046, + 0.002087518572807312, + 0.0016341746086254716, + 0.0044205584563314915, + 0.0010857766028493643, + 0.0017908766167238355, + 0.002575302729383111, + 0.0011833321768790483, + 0.0017562219873070717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.026332726702094078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073444969952106 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073444969952106, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073444969952106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0777636468410492, + 0.12774749100208282, + 0.06140818074345589, + 0.11682862788438797, + 0.16032059490680695, + 0.0577743835747242, + 0.23074042797088623, + 0.09789454936981201, + 0.034214086830616, + 0.035307977348566055 + ], + "credit_map": [ + 0.01150705385953188, + 0.008801733143627644, + 0.012024748139083385, + 0.007257874123752117, + 0.00782613456249237, + 0.013240626081824303, + 0.005120390560477972, + 0.011985084041953087, + 0.025648631155490875, + 0.012740683741867542, + 0.017304854467511177, + 0.010323169641196728, + 0.019045496359467506, + 0.015445535071194172, + 0.007302945014089346, + 0.019936176016926765, + 0.011222610250115395, + 0.00659687677398324, + 0.00528735714033246, + 0.0043723611161112785, + 0.0054568275809288025, + 0.009356564842164516, + 0.009734340943396091, + 0.009381241165101528, + 0.01223501656204462, + 0.009212395176291466, + 0.01772150956094265, + 0.01064034178853035, + 0.01913619600236416, + 0.01793333888053894, + 0.01564924791455269, + 0.014300577342510223, + 0.0177454836666584, + 0.02916761301457882, + 0.017303701490163803, + 0.017415033653378487, + 0.021497510373592377, + 0.01968497410416603, + 0.023252032697200775, + 0.014254248701035976, + 0.007619383279234171, + 0.006175753194838762, + 0.006957530975341797, + 0.012366271577775478, + 0.00721854018047452, + 0.006721597630530596, + 0.005538304802030325, + 0.005177001468837261, + 0.02424292452633381, + 0.01568111591041088, + 0.03990032523870468, + 0.016333479434251785, + 0.020396722480654716, + 0.04128372296690941, + 0.025669388473033905, + 0.047232769429683685, + 0.010908037424087524, + 0.008230234496295452, + 0.024407723918557167, + 0.007076771464198828, + 0.013573991134762764, + 0.012335925363004208, + 0.011236458085477352, + 0.010125408880412579, + 0.004517514258623123, + 0.0028677687514573336, + 0.009945814497768879, + 0.0022078948095440865, + 0.0038837832398712635, + 0.005826106294989586, + 0.0020769645925611258, + 0.0028882403858006, + 0.004195286426693201, + 0.003384646959602833, + 0.009015374816954136, + 0.0023289667442440987, + 0.004381309729069471, + 0.005579074379056692, + 0.0028202261310070753, + 0.0036030940245836973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.011073444969952106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, solemn camera tilts upward, first sweeping across a cracked asphalt surface streaked with spiderweb-like cracks, before slowly revealing a full-length portrait of a young man standing tall. In the extreme black and white HDR lighting, a sharp light source shines from above and to the side, outlining his body with a dazzling edge of brightness, while the other side is completely immersed in a bottomless, inky shadow. He sports a meticulously styled, now messy, black crew cut, with a few strands of hair, damp with sweat, plastered to his forehead. He clutches a cold, metal baseball bat tightly in his hands. The shaft gleams harshly in the bright sunlight, its surface dented and scratched. He wears a frayed black leather jacket, half-zipped, revealing a stained white T-shirt underneath. His knuckles are swollen and red from raw skin. His jaw is clenched, his lips pressed into a resolute line as he grits his teeth in an effort of endurance. Sweat mixed with dust trickles down his forehead, streaking across his young face, which is covered in tiny scratches and bruises. Though his body sways slightly from exhaustion, his legs remain firmly rooted to the ground, his back straight. The high-contrast light and shadow vividly depict his scarred yet unyielding posture, creating an atmosphere of intense tension, a sense of tragic toughness, and resilience.", + "seed": 1373, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_5857_seed_1373_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03328490685671568, + "vlm_logprob": -0.03328490685671568, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03328490685671568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031634680926799774 + ], + "opposite_logprob_tokens": [ + -3.906634569168091 + ], + "logprob_reward": -0.031634680926799774, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.031634680926799774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2553583085536957, + 0.14898324012756348, + 0.14483104646205902, + 0.12945230305194855, + 0.05910315364599228, + 0.07229966670274734, + 0.024745231494307518, + 0.0816371887922287, + 0.034658920019865036, + 0.048930902034044266 + ], + "credit_map": [ + 0.03360390290617943, + 0.05725416913628578, + 0.02174781821668148, + 0.02894606441259384, + 0.03427647054195404, + 0.03475964441895485, + 0.026069898158311844, + 0.01870035007596016, + 0.018006902188062668, + 0.02745639532804489, + 0.02092183381319046, + 0.020616864785552025, + 0.013627966865897179, + 0.008948144502937794, + 0.017558403313159943, + 0.021846747025847435, + 0.01962020806968212, + 0.028816644102334976, + 0.02356291376054287, + 0.022312473505735397, + 0.007583911530673504, + 0.009718319401144981, + 0.021013084799051285, + 0.012203510850667953, + 0.019226310774683952, + 0.01667037606239319, + 0.023530909791588783, + 0.013413447886705399, + 0.006179468240588903, + 0.030422667041420937, + 0.007151241414248943, + 0.012857900932431221, + 0.0060521746054291725, + 0.006429968401789665, + 0.011899827048182487, + 0.008928895927965641, + 0.005673480220139027, + 0.005665686912834644, + 0.006290555465966463, + 0.008162569254636765, + 0.014605989679694176, + 0.011342852376401424, + 0.010600263252854347, + 0.004873137921094894, + 0.005467014852911234, + 0.009061574004590511, + 0.004178172443062067, + 0.012170675210654736, + 0.002260992769151926, + 0.002263695700094104, + 0.0038615113589912653, + 0.002352279145270586, + 0.005139682441949844, + 0.003450083779171109, + 0.002600340638309717, + 0.002816644497215748, + 0.004231177736073732, + 0.005719945300370455, + 0.006409280933439732, + 0.0058560981415212154, + 0.006795815192162991, + 0.024539364501833916, + 0.004953979980200529, + 0.023131534457206726, + 0.00365551863797009, + 0.004270448815077543, + 0.005609078798443079, + 0.0035245914477854967, + 0.004620817024260759, + 0.005180651787668467, + 0.0037380221765488386, + 0.0040597934275865555, + 0.004624294117093086, + 0.005066660698503256, + 0.00897673238068819, + 0.006273181643337011, + 0.0054330588318407536, + 0.007452623452991247, + 0.004756364971399307, + 0.006347988732159138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.031634680926799774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870576038956642 + ], + "opposite_logprob_tokens": [ + -3.048705816268921 + ], + "logprob_reward": -0.04870576038956642, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870576038956642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10739844292402267, + 0.07565717399120331, + 0.16414423286914825, + 0.045683395117521286, + 0.28753823041915894, + 0.06795614957809448, + 0.021445972844958305, + 0.03409359231591225, + 0.032574139535427094, + 0.16350869834423065 + ], + "credit_map": [ + 0.008870543912053108, + 0.018477624282240868, + 0.01096061710268259, + 0.014142464846372604, + 0.01596631295979023, + 0.01718231849372387, + 0.012804463505744934, + 0.008994094096124172, + 0.01269283052533865, + 0.010210657492280006, + 0.010221252217888832, + 0.008606958203017712, + 0.004763152450323105, + 0.006132388487458229, + 0.012404069304466248, + 0.010625859722495079, + 0.016996867954730988, + 0.03514481708407402, + 0.023683831095695496, + 0.024996735155582428, + 0.012269003316760063, + 0.015860211104154587, + 0.016721783205866814, + 0.018470976501703262, + 0.005844808649271727, + 0.005525773856788874, + 0.00841102097183466, + 0.00413903221487999, + 0.004223904572427273, + 0.008589432574808598, + 0.004327254369854927, + 0.004622164648026228, + 0.022400539368391037, + 0.018896469846367836, + 0.066607765853405, + 0.03148490935564041, + 0.026344062760472298, + 0.03018067218363285, + 0.03092223033308983, + 0.06070154160261154, + 0.00794228632003069, + 0.008586125448346138, + 0.01921788603067398, + 0.0055071511305868626, + 0.005178662016987801, + 0.00695641478523612, + 0.005074921064078808, + 0.009492699056863785, + 0.0019774851389229298, + 0.0020387812983244658, + 0.004684250336140394, + 0.0016787783242762089, + 0.0029133714269846678, + 0.00348382699303329, + 0.0024507397320121527, + 0.0022187356371432543, + 0.0023571462370455265, + 0.003741590538993478, + 0.005146993324160576, + 0.003622588934376836, + 0.004367748275399208, + 0.006495596840977669, + 0.0022932998836040497, + 0.006068629678338766, + 0.0048159328289330006, + 0.0027878154069185257, + 0.0059153069742023945, + 0.003262662561610341, + 0.004079227801412344, + 0.004986284300684929, + 0.0024320681113749743, + 0.004294840153306723, + 0.017198197543621063, + 0.01744694449007511, + 0.030307387933135033, + 0.012607462704181671, + 0.01709224097430706, + 0.027284957468509674, + 0.015453091822564602, + 0.026118403300642967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.04870576038956642 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867010563611984 + ], + "opposite_logprob_tokens": [ + -3.0486700534820557 + ], + "logprob_reward": -0.04867010563611984, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867010563611984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13078653812408447, + 0.08088015019893646, + 0.15647061169147491, + 0.06346730142831802, + 0.18212346732616425, + 0.1287921518087387, + 0.12031643837690353, + 0.0335315503180027, + 0.02688770182430744, + 0.07674410194158554 + ], + "credit_map": [ + 0.013806314207613468, + 0.019925059750676155, + 0.010672601871192455, + 0.014951571822166443, + 0.020155612379312515, + 0.01896660402417183, + 0.01764596626162529, + 0.014662805013358593, + 0.009942039847373962, + 0.011934551410377026, + 0.010367807932198048, + 0.012374168261885643, + 0.005818030331283808, + 0.007453056052327156, + 0.010626934468746185, + 0.012363565154373646, + 0.022455472499132156, + 0.02993050403892994, + 0.021505707874894142, + 0.02501310221850872, + 0.011544991284608841, + 0.012508532963693142, + 0.018177075311541557, + 0.015335229225456715, + 0.00925281923264265, + 0.008138637989759445, + 0.008175627328455448, + 0.008099219761788845, + 0.006576104089617729, + 0.0097461873665452, + 0.004876264370977879, + 0.008602436631917953, + 0.017268158495426178, + 0.020542604848742485, + 0.025640586391091347, + 0.026860661804676056, + 0.01784425415098667, + 0.0275855902582407, + 0.020184924826025963, + 0.026196690276265144, + 0.017686424776911736, + 0.015337472781538963, + 0.020848941057920456, + 0.011793709360063076, + 0.015994878485798836, + 0.014587558805942535, + 0.013898658566176891, + 0.018644509837031364, + 0.008391034789383411, + 0.009398217312991619, + 0.022330837324261665, + 0.010179949924349785, + 0.016924899071455002, + 0.020193293690681458, + 0.017738882452249527, + 0.015159327536821365, + 0.0032998668029904366, + 0.004479699768126011, + 0.0069688456133008, + 0.004051767289638519, + 0.0032306837383657694, + 0.0046554626896977425, + 0.0030732981394976377, + 0.0037719253450632095, + 0.0023529562167823315, + 0.003450801596045494, + 0.0034899485763162374, + 0.0029796527232974768, + 0.003577148774638772, + 0.004532588645815849, + 0.0027008079923689365, + 0.003803797997534275, + 0.007119732443243265, + 0.009356400929391384, + 0.010115033015608788, + 0.006334446370601654, + 0.009817137382924557, + 0.012154275551438332, + 0.009145925752818584, + 0.012701151892542839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.04867010563611984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338066905736923 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338066905736923, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338066905736923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09237392246723175, + 0.0911015048623085, + 0.08217044174671173, + 0.0830681324005127, + 0.1403205543756485, + 0.04038854315876961, + 0.03449714928865433, + 0.131738543510437, + 0.10017481446266174, + 0.20416630804538727 + ], + "credit_map": [ + 0.010143241845071316, + 0.017114663496613503, + 0.009207007475197315, + 0.012757365591824055, + 0.012330206111073494, + 0.012085075490176678, + 0.010378507897257805, + 0.00835785549134016, + 0.015805531293153763, + 0.030413540080189705, + 0.006748725660145283, + 0.008128196932375431, + 0.005453366320580244, + 0.004168112762272358, + 0.008087612688541412, + 0.012296423316001892, + 0.011361352168023586, + 0.02118990384042263, + 0.014102950692176819, + 0.012644736096262932, + 0.0047578285448253155, + 0.0046996562741696835, + 0.007637710776180029, + 0.005776308011263609, + 0.00810289103537798, + 0.011709200218319893, + 0.011586206033825874, + 0.011851543560624123, + 0.006535724736750126, + 0.013600763864815235, + 0.007920721545815468, + 0.011761080473661423, + 0.010586610995233059, + 0.009072922170162201, + 0.035833049565553665, + 0.013564839027822018, + 0.011468200013041496, + 0.016207871958613396, + 0.01198994554579258, + 0.03159712255001068, + 0.004843830596655607, + 0.003279584227129817, + 0.005812725517898798, + 0.005721980705857277, + 0.005612469278275967, + 0.005087832920253277, + 0.0034472644329071045, + 0.006582855712622404, + 0.002782468916848302, + 0.002935518976300955, + 0.0058791679330170155, + 0.003036219161003828, + 0.005946971010416746, + 0.005304621532559395, + 0.0035544277634471655, + 0.005057754926383495, + 0.008955340832471848, + 0.017131134867668152, + 0.02024206519126892, + 0.019078146666288376, + 0.014821910299360752, + 0.02206571400165558, + 0.010020511224865913, + 0.019423723220825195, + 0.008366398513317108, + 0.011675028130412102, + 0.01865273527801037, + 0.010774113237857819, + 0.01422915980219841, + 0.015616214834153652, + 0.007483265828341246, + 0.013377895578742027, + 0.016338076442480087, + 0.01961689256131649, + 0.029719557613134384, + 0.014819414354860783, + 0.025422442704439163, + 0.037890128791332245, + 0.019194694235920906, + 0.04116509482264519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.026338066905736923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075920425355434 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075920425355434, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075920425355434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13311168551445007, + 0.09672760218381882, + 0.1369248330593109, + 0.19895590841770172, + 0.058774128556251526, + 0.08372924476861954, + 0.049566153436899185, + 0.09156975150108337, + 0.11548840999603271, + 0.03515227884054184 + ], + "credit_map": [ + 0.017345067113637924, + 0.030065370723605156, + 0.011233975179493427, + 0.01541027519851923, + 0.01591433398425579, + 0.017910931259393692, + 0.013826449401676655, + 0.011405298486351967, + 0.016349295154213905, + 0.024024762213230133, + 0.010175103321671486, + 0.01008381973952055, + 0.009405544959008694, + 0.006781819276511669, + 0.008928367868065834, + 0.01097889244556427, + 0.018894080072641373, + 0.027679257094860077, + 0.020474983379244804, + 0.022720925509929657, + 0.010853667743504047, + 0.009972031228244305, + 0.015134137123823166, + 0.011195757426321507, + 0.014976336620748043, + 0.02590784803032875, + 0.027773983776569366, + 0.024824216961860657, + 0.017521778121590614, + 0.035108473151922226, + 0.02338980697095394, + 0.02945348434150219, + 0.00554238585755229, + 0.005079358350485563, + 0.010491219349205494, + 0.008435413241386414, + 0.006430742330849171, + 0.007697048597037792, + 0.007693338207900524, + 0.007404625881463289, + 0.010352032259106636, + 0.008958389982581139, + 0.013560249470174313, + 0.0086033521220088, + 0.009188062511384487, + 0.011246117763221264, + 0.007744660601019859, + 0.014076385647058487, + 0.003515202784910798, + 0.003396790474653244, + 0.008153780363500118, + 0.005690412130206823, + 0.007091414649039507, + 0.008319209329783916, + 0.006386857479810715, + 0.007012491580098867, + 0.006712764967232943, + 0.011395814828574657, + 0.007775070611387491, + 0.009960847906768322, + 0.007442406378686428, + 0.0219156201928854, + 0.005446170922368765, + 0.020921064540743828, + 0.00892068911343813, + 0.00929593201726675, + 0.026094980537891388, + 0.013167796656489372, + 0.013074164278805256, + 0.01670517399907112, + 0.014602203853428364, + 0.01362747885286808, + 0.004335590172559023, + 0.004338168539106846, + 0.00573569955304265, + 0.00356009928509593, + 0.005124562419950962, + 0.0040682228282094, + 0.003737492486834526, + 0.004252446349710226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.011075920425355434 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, solemn camera tilts upward, first sweeping across a cracked asphalt surface streaked with spiderweb-like cracks, before slowly revealing a full-length portrait of a young man standing tall. In the extreme black and white HDR lighting, a sharp light source shines from above and to the side, outlining his body with a dazzling edge of brightness, while the other side is completely immersed in a bottomless, inky shadow. He sports a meticulously styled, now messy, black crew cut, with a few strands of hair, damp with sweat, plastered to his forehead. He clutches a cold, metal baseball bat tightly in his hands. The shaft gleams harshly in the bright sunlight, its surface dented and scratched. He wears a frayed black leather jacket, half-zipped, revealing a stained white T-shirt underneath. His knuckles are swollen and red from raw skin. His jaw is clenched, his lips pressed into a resolute line as he grits his teeth in an effort of endurance. Sweat mixed with dust trickles down his forehead, streaking across his young face, which is covered in tiny scratches and bruises. Though his body sways slightly from exhaustion, his legs remain firmly rooted to the ground, his back straight. The high-contrast light and shadow vividly depict his scarred yet unyielding posture, creating an atmosphere of intense tension, a sense of tragic toughness, and resilience.", + "seed": 1374, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_6/index_5857_seed_1374_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03284468185156584, + "vlm_logprob": -0.03284468185156584, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.03284468185156584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03503616899251938 + ], + "opposite_logprob_tokens": [ + -3.535036087036133 + ], + "logprob_reward": -0.03503616899251938, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03503616899251938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6946961879730225, + 0.0947430357336998, + 0.054182276129722595, + 0.0583927147090435, + 0.01941111497581005, + 0.008864888921380043, + 0.009011175483465195, + 0.02090076170861721, + 0.02735229767858982, + 0.01244552992284298 + ], + "credit_map": [ + 0.048844870179891586, + 0.044229380786418915, + 0.029838666319847107, + 0.022250983864068985, + 0.23116938769817352, + 0.21578170359134674, + 0.03675202652812004, + 0.06582923233509064, + 0.010024107061326504, + 0.0039697797037661076, + 0.0202792938798666, + 0.006011072546243668, + 0.0054299416951835155, + 0.01671086996793747, + 0.02120389975607395, + 0.011114073917269707, + 0.007218116894364357, + 0.0033224131911993027, + 0.009234289638698101, + 0.00499296048656106, + 0.004006294067949057, + 0.007315218448638916, + 0.010651661083102226, + 0.007441327441483736, + 0.006528588477522135, + 0.005102659109979868, + 0.0088499179109931, + 0.005944501142948866, + 0.006145058199763298, + 0.011499958112835884, + 0.00756998872384429, + 0.006752046290785074, + 0.0020808158442378044, + 0.002249038079753518, + 0.002768806414678693, + 0.002442216034978628, + 0.002075496595352888, + 0.003667559241876006, + 0.0016574460314586759, + 0.0024697380140423775, + 0.0013103543315082788, + 0.0007856348529458046, + 0.0019445221405476332, + 0.000732607499230653, + 0.001412337995134294, + 0.0010278117842972279, + 0.0006840563728474081, + 0.0009675645851530135, + 0.001344493473879993, + 0.0005196601268835366, + 0.002612041076645255, + 0.0005067163147032261, + 0.0008458330412395298, + 0.0015031789662316442, + 0.0005772192380391061, + 0.0011020335368812084, + 0.003323906334117055, + 0.0014513011556118727, + 0.005230424460023642, + 0.0012172004207968712, + 0.0026032663881778717, + 0.003953489474952221, + 0.0012643665540963411, + 0.001856807037256658, + 0.002882112981751561, + 0.003367349971085787, + 0.008973647840321064, + 0.0015934174880385399, + 0.0015320861712098122, + 0.005165041424334049, + 0.001683831331320107, + 0.002154812216758728, + 0.0015034760581329465, + 0.0020830314606428146, + 0.002575727179646492, + 0.0012713616015389562, + 0.0010194695787504315, + 0.0019435846479609609, + 0.0005261203041300178, + 0.0015227595577016473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03503616899251938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310404881834984 + ], + "opposite_logprob_tokens": [ + -3.1681039333343506 + ], + "logprob_reward": -0.04310404881834984, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310404881834984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35084956884384155, + 0.09022077918052673, + 0.07347195595502853, + 0.06361311674118042, + 0.06965300440788269, + 0.035109322518110275, + 0.1727226972579956, + 0.024772780016064644, + 0.0823730081319809, + 0.03721372038125992 + ], + "credit_map": [ + 0.017886439338326454, + 0.01629895716905594, + 0.012953728437423706, + 0.014841035939753056, + 0.12613637745380402, + 0.12755708396434784, + 0.015846822410821915, + 0.019329147413372993, + 0.011048899963498116, + 0.003130401484668255, + 0.017543531954288483, + 0.004971453920006752, + 0.005606631748378277, + 0.016165660694241524, + 0.021443042904138565, + 0.010311153717339039, + 0.008842197246849537, + 0.005696621257811785, + 0.012441940605640411, + 0.008808106184005737, + 0.004517784342169762, + 0.014341156929731369, + 0.00907745212316513, + 0.00974669586867094, + 0.01020616665482521, + 0.004776506219059229, + 0.010373678989708424, + 0.006381546147167683, + 0.004957219585776329, + 0.013017106801271439, + 0.005535859614610672, + 0.008365032263100147, + 0.012319408357143402, + 0.006668140180408955, + 0.011506859213113785, + 0.012921355664730072, + 0.0064757936634123325, + 0.008027187548577785, + 0.00486659724265337, + 0.006867663469165564, + 0.004944880958646536, + 0.0032466198317706585, + 0.008181584998965263, + 0.002370257396250963, + 0.004183588549494743, + 0.005103047471493483, + 0.003116080304607749, + 0.003963262774050236, + 0.030272241681814194, + 0.010903663001954556, + 0.053692273795604706, + 0.009817457757890224, + 0.013500894419848919, + 0.030050450935959816, + 0.0074168420396745205, + 0.017068883404135704, + 0.004015834536403418, + 0.0018409221665933728, + 0.004860091954469681, + 0.0016971779987215996, + 0.00377445830963552, + 0.004213674459606409, + 0.0019174168119207025, + 0.002453202847391367, + 0.008555962704122066, + 0.009901050478219986, + 0.023980818688869476, + 0.006963186897337437, + 0.005164069589227438, + 0.016202276572585106, + 0.004364983644336462, + 0.007240652106702328, + 0.005793411750346422, + 0.004518854897469282, + 0.009584756568074226, + 0.002776987152174115, + 0.0034705940634012222, + 0.005902834702283144, + 0.002383728511631489, + 0.002782556228339672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04310404881834984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048665452748537064 + ], + "opposite_logprob_tokens": [ + -3.0486655235290527 + ], + "logprob_reward": -0.048665452748537064, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048665452748537064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14667098224163055, + 0.09562698006629944, + 0.10210633277893066, + 0.06762979924678802, + 0.09440780431032181, + 0.24514317512512207, + 0.04311588034033775, + 0.0723506361246109, + 0.09566863626241684, + 0.03727982938289642 + ], + "credit_map": [ + 0.015054858289659023, + 0.016472911462187767, + 0.011448066681623459, + 0.009439630433917046, + 0.03481684625148773, + 0.030272699892520905, + 0.010496455244719982, + 0.018669510260224342, + 0.010132133029401302, + 0.004546412266790867, + 0.01423158124089241, + 0.006648253183811903, + 0.00895331334322691, + 0.013604597188532352, + 0.02695062756538391, + 0.010560070164501667, + 0.010526922531425953, + 0.009753775782883167, + 0.01830228418111801, + 0.010865628719329834, + 0.009752048179507256, + 0.015355756506323814, + 0.013662922196090221, + 0.013886986300349236, + 0.008661959320306778, + 0.0060955495573580265, + 0.009399421513080597, + 0.007766322232782841, + 0.007839350029826164, + 0.01051940768957138, + 0.007717749569565058, + 0.009630038402974606, + 0.015546689741313457, + 0.007562755141407251, + 0.018079550936818123, + 0.009252918884158134, + 0.007650509011000395, + 0.016135655343532562, + 0.0070867580361664295, + 0.013092967681586742, + 0.025193894281983376, + 0.018118228763341904, + 0.038227375596761703, + 0.02629612199962139, + 0.025800390169024467, + 0.03021753579378128, + 0.02666494995355606, + 0.05462469160556793, + 0.008039258420467377, + 0.002519764006137848, + 0.0125558041036129, + 0.0021019401028752327, + 0.0035960045643150806, + 0.006867581512778997, + 0.0024707550182938576, + 0.0049647679552435875, + 0.009734343737363815, + 0.004752700682729483, + 0.014743153005838394, + 0.005277932621538639, + 0.008612977340817451, + 0.011111104860901833, + 0.007606382016092539, + 0.01051204465329647, + 0.01145919319242239, + 0.013048287481069565, + 0.02548590674996376, + 0.007771684788167477, + 0.006844605319201946, + 0.01343744620680809, + 0.006104808300733566, + 0.011516704224050045, + 0.004183122422546148, + 0.006024997681379318, + 0.00735638989135623, + 0.004017543978989124, + 0.003276896197348833, + 0.005129777826368809, + 0.002207092707976699, + 0.00508400984108448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.048665452748537064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339460164308548 + ], + "opposite_logprob_tokens": [ + -3.651339530944824 + ], + "logprob_reward": -0.026339460164308548, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339460164308548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29446351528167725, + 0.07734666764736176, + 0.11707647889852524, + 0.11739619821310043, + 0.05730034038424492, + 0.07550565898418427, + 0.034837763756513596, + 0.12032580375671387, + 0.06907708197832108, + 0.03667047247290611 + ], + "credit_map": [ + 0.0304393470287323, + 0.02071436680853367, + 0.01837148331105709, + 0.015067785046994686, + 0.07136590778827667, + 0.07907756417989731, + 0.0193440243601799, + 0.04008302837610245, + 0.007273911964148283, + 0.0036995606496930122, + 0.021097496151924133, + 0.007095256354659796, + 0.004520777612924576, + 0.013335170224308968, + 0.013340418227016926, + 0.006984084378927946, + 0.011251463554799557, + 0.01112646795809269, + 0.018363917246460915, + 0.01258514728397131, + 0.011856966651976109, + 0.015459823422133923, + 0.016062594950199127, + 0.02037009410560131, + 0.01675848662853241, + 0.009697549976408482, + 0.01665523089468479, + 0.013884516432881355, + 0.010863608680665493, + 0.020709868520498276, + 0.01101219654083252, + 0.01781473495066166, + 0.007163285743445158, + 0.005343835335224867, + 0.008333624340593815, + 0.008306288160383701, + 0.00434463145211339, + 0.008567805401980877, + 0.005652454681694508, + 0.009588413871824741, + 0.00967708881944418, + 0.005896085407584906, + 0.01576070860028267, + 0.009847081266343594, + 0.006473845336586237, + 0.009254438802599907, + 0.007181844208389521, + 0.01141456887125969, + 0.004803977441042662, + 0.0023835899773985147, + 0.009102893993258476, + 0.0031248738523572683, + 0.0037031988613307476, + 0.004430302884429693, + 0.0031985642854124308, + 0.0040903594344854355, + 0.016410883516073227, + 0.00783061794936657, + 0.02424534037709236, + 0.0074094925075769424, + 0.015953803434967995, + 0.019586531445384026, + 0.012386665679514408, + 0.016502466052770615, + 0.0075677004642784595, + 0.009470979683101177, + 0.019060399383306503, + 0.006136308424174786, + 0.004751330707222223, + 0.011207890696823597, + 0.003626795718446374, + 0.0072556836530566216, + 0.0037184932734817266, + 0.005603172816336155, + 0.007215204183012247, + 0.004421195015311241, + 0.0036536483094096184, + 0.004815091844648123, + 0.0027569022495299578, + 0.004486763384193182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.026339460164308548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start by tilting upward across a cracked asphalt surface before revealing a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in extreme black and white HDR lighting with a sharp light source creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man holding a dented and scratched metal baseball bat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character have a messy black crew cut with sweat-dampened strands plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character wearing a frayed black leather jacket over a stained white T-shirt with swollen red knuckles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107827853411436 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.01107827853411436, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107827853411436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18005403876304626, + 0.12639306485652924, + 0.11105333268642426, + 0.119413822889328, + 0.09866098314523697, + 0.12405576556921005, + 0.0384737066924572, + 0.07406570017337799, + 0.08665405958890915, + 0.041175566613674164 + ], + "credit_map": [ + 0.020111678168177605, + 0.0183118786662817, + 0.01613362692296505, + 0.011193324811756611, + 0.0404837392270565, + 0.03373301401734352, + 0.014207655563950539, + 0.025879109278321266, + 0.012984136119484901, + 0.0093429796397686, + 0.025025051087141037, + 0.008640408515930176, + 0.008306540548801422, + 0.017293015494942665, + 0.033503301441669464, + 0.011297623626887798, + 0.011954549700021744, + 0.010236556641757488, + 0.019330637529492378, + 0.008381045423448086, + 0.01150900311768055, + 0.017666440457105637, + 0.017047090455889702, + 0.014928010292351246, + 0.019589485600590706, + 0.008208748884499073, + 0.019137948751449585, + 0.013888504356145859, + 0.010136622935533524, + 0.021836118772625923, + 0.010044368915259838, + 0.016572028398513794, + 0.015913963317871094, + 0.012235135771334171, + 0.013343501836061478, + 0.01093258336186409, + 0.010262098163366318, + 0.01648901402950287, + 0.006758848205208778, + 0.012725833803415298, + 0.01377774216234684, + 0.008923862129449844, + 0.02329743094742298, + 0.01575387828052044, + 0.011304212734103203, + 0.015254287980496883, + 0.01254535187035799, + 0.023199006915092468, + 0.005513436626642942, + 0.002659248188138008, + 0.010877246968448162, + 0.0018662558868527412, + 0.0038727130740880966, + 0.00601690448820591, + 0.003106061602011323, + 0.004561842884868383, + 0.007609248626977205, + 0.005783361382782459, + 0.020049432292580605, + 0.006140104960650206, + 0.006722705438733101, + 0.014008253812789917, + 0.006821249146014452, + 0.006931336596608162, + 0.009582070633769035, + 0.010606308467686176, + 0.025852909311652184, + 0.0064359745010733604, + 0.0049615646712481976, + 0.014552464708685875, + 0.005186757538467646, + 0.0094760125502944, + 0.004520634654909372, + 0.006421110592782497, + 0.00690457783639431, + 0.005145728122442961, + 0.003921210765838623, + 0.005350741557776928, + 0.003110612742602825, + 0.005800952203571796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.01107827853411436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_tensor([9935, 9935])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_tensor([9935, 9935])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a32805ed442fd28891a6518a0b03f30af10a43a1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_tensor([9935, 9935])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Ultra-high-speed macro photography captured from a static, high-angle perspective documents the mesmerizing process of crystallization on a perfectly smooth, matte black surface. The scene is illuminated by intense, rhythmic flashes of cold, bright white light, creating a striking strobe effect. The action begins with a few tiny, individual grains of white sugar, their translucent facets capturing the initial flashes. Then, in a series of sharp, stuttering advances synchronized with each pulse of light, complex, interlocking geometric structures explosively sprout from these nucleation points. Needle-like fragments and massive, prismatic shapes jerk forward, methodically stacking upon one another to form a dense, intricate grid that spans the entire frame. The harsh, clinical light casts deep, fleeting shadows and creates dazzling, mirrored highlights along the crystal's edges, making the entire structure shimmer in cold fire. Growth is not a smooth, uninterrupted process, but a hypnotic, abrasive sequence that conveys a powerful and captivating atmosphere of fundamental natural processes revealed in discrete, frozen moments of time.", + "seed": 20263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20263_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.597603633441031, + "vlm_logprob": -0.597603633441031, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.597603633441031, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3360956907272339 + ], + "opposite_logprob_tokens": [ + -1.3360956907272339 + ], + "logprob_reward": -0.3360956907272339, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3360956907272339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16123895347118378, + 0.05492241680622101, + 0.11375352740287781, + 0.10293515026569366, + 0.06784788519144058, + 0.23778527975082397, + 0.082373708486557, + 0.07781676203012466, + 0.025715092197060585, + 0.0756111815571785 + ], + "credit_map": [ + 0.01738225482404232, + 0.019332649186253548, + 0.02800501137971878, + 0.018427327275276184, + 0.019654063507914543, + 0.0153209762647748, + 0.026013219729065895, + 0.017103441059589386, + 0.005815006326884031, + 0.009223081171512604, + 0.00788677018135786, + 0.00602794298902154, + 0.0052594151347875595, + 0.006141135469079018, + 0.007435448933392763, + 0.007133616600185633, + 0.009518084116280079, + 0.028835909441113472, + 0.02951889857649803, + 0.008062819018959999, + 0.0056010400876402855, + 0.007250295020639896, + 0.008987938985228539, + 0.01597854495048523, + 0.012823296710848808, + 0.013133472763001919, + 0.013755284249782562, + 0.013141925446689129, + 0.009875688701868057, + 0.012766905128955841, + 0.015970980748534203, + 0.011467603035271168, + 0.006775861606001854, + 0.010312492027878761, + 0.0055501023307442665, + 0.006917225196957588, + 0.007089593913406134, + 0.005389408208429813, + 0.02047305554151535, + 0.005340144503861666, + 0.021264374256134033, + 0.013547500595450401, + 0.0167926587164402, + 0.0872366800904274, + 0.014750605449080467, + 0.056754715740680695, + 0.018193664029240608, + 0.009245086461305618, + 0.007743922993540764, + 0.009841806255280972, + 0.0073312572203576565, + 0.009463254362344742, + 0.009236781857907772, + 0.012405840680003166, + 0.014282445423305035, + 0.012068403884768486, + 0.0050771236419677734, + 0.008416914381086826, + 0.015808427706360817, + 0.005608974955976009, + 0.01004815474152565, + 0.0068962424993515015, + 0.009829124435782433, + 0.016131803393363953, + 0.0023836949840188026, + 0.0034305613953620195, + 0.0038948887959122658, + 0.003401093650609255, + 0.0018425938906148076, + 0.002533512655645609, + 0.005040880758315325, + 0.0031878661829978228, + 0.00831057969480753, + 0.00908302329480648, + 0.007571261841803789, + 0.009210816584527493, + 0.00618124520406127, + 0.010783826000988483, + 0.010251933708786964, + 0.014218498952686787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.3360956907272339 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.5789566040039062 + ], + "opposite_logprob_tokens": [ + -0.07895661890506744 + ], + "logprob_reward": -2.5789566040039062, + "yes_no_margin_reward": -2.5, + "token_credit_reward": -2.5789566040039062, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14588119089603424, + 0.0494748130440712, + 0.2274167537689209, + 0.08576428145170212, + 0.08476851135492325, + 0.12795959413051605, + 0.051971156150102615, + 0.04238269478082657, + 0.03765637055039406, + 0.14672458171844482 + ], + "credit_map": [ + 0.013507821597158909, + 0.019016442820429802, + 0.021769655868411064, + 0.016175134107470512, + 0.017666220664978027, + 0.020088322460651398, + 0.02080524154007435, + 0.016852350905537605, + 0.005907845683395863, + 0.007559654768556356, + 0.008580099791288376, + 0.00597477937117219, + 0.005378995556384325, + 0.004073790740221739, + 0.0070196958258748055, + 0.004979947581887245, + 0.012724886648356915, + 0.08826557546854019, + 0.07769470661878586, + 0.009754197672009468, + 0.008048228919506073, + 0.007221672683954239, + 0.014367606490850449, + 0.009339899756014347, + 0.01069351751357317, + 0.013179196044802666, + 0.01388560887426138, + 0.010022815316915512, + 0.006646835245192051, + 0.009531600400805473, + 0.011094694957137108, + 0.010710010305047035, + 0.008392758667469025, + 0.010445134714245796, + 0.005704969167709351, + 0.00570610212162137, + 0.014112056232988834, + 0.006407639943063259, + 0.02890544943511486, + 0.005094404332339764, + 0.02086559124290943, + 0.009700710885226727, + 0.01072665024548769, + 0.04401683807373047, + 0.006066582631319761, + 0.014135983772575855, + 0.014395751059055328, + 0.008051494136452675, + 0.003529520705342293, + 0.0043475390411913395, + 0.0039049929473549128, + 0.00407858332619071, + 0.004669661168009043, + 0.011066063307225704, + 0.01563417539000511, + 0.004740620497614145, + 0.0032614527735859156, + 0.00643348740413785, + 0.008403983898460865, + 0.00399320013821125, + 0.004785984288901091, + 0.0044449009001255035, + 0.004467697814106941, + 0.00659199059009552, + 0.0036861910484731197, + 0.00574649590998888, + 0.005831568036228418, + 0.0034810402430593967, + 0.003531582886353135, + 0.005848935339599848, + 0.004712052643299103, + 0.004818507935851812, + 0.015845051035284996, + 0.020851759240031242, + 0.01328454352915287, + 0.01902928203344345, + 0.010291924700140953, + 0.017044980078935623, + 0.026118269190192223, + 0.024258771911263466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -2.5789566040039062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014236669056117535 + ], + "opposite_logprob_tokens": [ + -4.2642364501953125 + ], + "logprob_reward": -0.014236669056117535, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014236669056117535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12677288055419922, + 0.050872329622507095, + 0.10801070183515549, + 0.03412362560629845, + 0.0539017878472805, + 0.44038307666778564, + 0.04282763972878456, + 0.02111233025789261, + 0.03626301884651184, + 0.08573268353939056 + ], + "credit_map": [ + 0.011227509006857872, + 0.01622593030333519, + 0.022419750690460205, + 0.01395275816321373, + 0.021942850202322006, + 0.010315937921404839, + 0.017422961071133614, + 0.013265186920762062, + 0.0058709098957479, + 0.0077187842689454556, + 0.00933279562741518, + 0.00546514056622982, + 0.005727921146899462, + 0.003757529892027378, + 0.00856646616011858, + 0.004432782996445894, + 0.010330739431083202, + 0.022142203524708748, + 0.02597690373659134, + 0.008874570950865746, + 0.006605274509638548, + 0.008119266480207443, + 0.011140171438455582, + 0.014821565710008144, + 0.004439853131771088, + 0.00489333039149642, + 0.0055902679450809956, + 0.0050353300757706165, + 0.0031780297867953777, + 0.003038042690604925, + 0.005040067248046398, + 0.0029087059665471315, + 0.004851286765187979, + 0.007366602309048176, + 0.003828330198302865, + 0.00584616931155324, + 0.007137287873774767, + 0.004173165652900934, + 0.01779097504913807, + 0.002907970454543829, + 0.036732062697410583, + 0.021904511377215385, + 0.019212577491998672, + 0.17524541914463043, + 0.019739175215363503, + 0.11321475356817245, + 0.03668533265590668, + 0.017649240791797638, + 0.003358218353241682, + 0.00402520876377821, + 0.0038075835909694433, + 0.0036481337156146765, + 0.0031804346945136786, + 0.006064474582672119, + 0.014157766476273537, + 0.004585815127938986, + 0.002149311825633049, + 0.003070021513849497, + 0.0027243865188211203, + 0.001982139889150858, + 0.002354973228648305, + 0.0021457558032125235, + 0.0028360479045659304, + 0.0038496917113661766, + 0.0034862461034208536, + 0.007116980850696564, + 0.005008529406040907, + 0.003129869233816862, + 0.0036137267015874386, + 0.00551368435844779, + 0.004264297895133495, + 0.004129681270569563, + 0.009942987002432346, + 0.011858407407999039, + 0.0073711141012609005, + 0.010878405533730984, + 0.005701902322471142, + 0.011232966557145119, + 0.014880234375596046, + 0.013866668567061424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.014236669056117535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813907131552696 + ], + "opposite_logprob_tokens": [ + -3.2881391048431396 + ], + "logprob_reward": -0.03813907131552696, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813907131552696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18021272122859955, + 0.056426748633384705, + 0.3507040739059448, + 0.07384658604860306, + 0.05860305204987526, + 0.09896153956651688, + 0.050355393439531326, + 0.023356836289167404, + 0.03468821570277214, + 0.07284476608037949 + ], + "credit_map": [ + 0.02322341874241829, + 0.023200033232569695, + 0.030602887272834778, + 0.017876775935292244, + 0.02159680426120758, + 0.018491819500923157, + 0.028075380250811577, + 0.01714559830725193, + 0.00773982610553503, + 0.009347674436867237, + 0.008111071772873402, + 0.006886995863169432, + 0.008229037746787071, + 0.0038624682929366827, + 0.006463858764618635, + 0.005785817746073008, + 0.019113143905997276, + 0.13680699467658997, + 0.12127595394849777, + 0.014546526595950127, + 0.011479315347969532, + 0.010359994135797024, + 0.025327226147055626, + 0.011794932186603546, + 0.008808010257780552, + 0.013128811493515968, + 0.007587702013552189, + 0.00701784435659647, + 0.009244710206985474, + 0.009572213515639305, + 0.008616248145699501, + 0.009871050715446472, + 0.005489766132086515, + 0.007452328223735094, + 0.0058844308368861675, + 0.00794768426567316, + 0.004795008338987827, + 0.005622610915452242, + 0.014765571802854538, + 0.006645656190812588, + 0.011677197180688381, + 0.011478638276457787, + 0.012786625884473324, + 0.024396121501922607, + 0.0133778415620327, + 0.009498391300439835, + 0.008830221369862556, + 0.006916506215929985, + 0.003893514396622777, + 0.0053779613226652145, + 0.005404639523476362, + 0.00401898892596364, + 0.005004595033824444, + 0.006657998077571392, + 0.015366727486252785, + 0.004630966577678919, + 0.0022090114653110504, + 0.0032423625234514475, + 0.00292872148565948, + 0.002243307651951909, + 0.004281508270651102, + 0.002559017390012741, + 0.0023606810718774796, + 0.0035322271287441254, + 0.003404092276468873, + 0.00714116869494319, + 0.0036185430362820625, + 0.003244884544983506, + 0.004122493322938681, + 0.006083169020712376, + 0.0034597711637616158, + 0.0036140948068350554, + 0.008608048781752586, + 0.010016621090471745, + 0.006436547264456749, + 0.009530030190944672, + 0.004807164892554283, + 0.010086597874760628, + 0.01198731828480959, + 0.011372435837984085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.03813907131552696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590132102370262 + ], + "opposite_logprob_tokens": [ + -3.89559006690979 + ], + "logprob_reward": -0.020590132102370262, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590132102370262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09590430557727814, + 0.034484345465898514, + 0.29111307859420776, + 0.04201323539018631, + 0.0641428679227829, + 0.29001232981681824, + 0.042287278920412064, + 0.03146570920944214, + 0.0409584604203701, + 0.06761836260557175 + ], + "credit_map": [ + 0.010623663663864136, + 0.014144676737487316, + 0.017302392050623894, + 0.00925732683390379, + 0.010547908954322338, + 0.010652948170900345, + 0.013822998851537704, + 0.009552385658025742, + 0.0039209057576954365, + 0.005507572554051876, + 0.005843647290021181, + 0.004249050747603178, + 0.004565659910440445, + 0.0027080627623945475, + 0.0047163632698357105, + 0.0029730850365012884, + 0.015603773295879364, + 0.11771349608898163, + 0.08873473107814789, + 0.01129263173788786, + 0.011170207522809505, + 0.008215902373194695, + 0.027175914496183395, + 0.011206434108316898, + 0.004952915012836456, + 0.00699596805498004, + 0.004446965176612139, + 0.004802957642823458, + 0.0035484828986227512, + 0.005135505925863981, + 0.0062969401478767395, + 0.0058334991335868835, + 0.00571797601878643, + 0.008718693628907204, + 0.005672724451869726, + 0.005807698704302311, + 0.008957336656749249, + 0.004900557920336723, + 0.01970634236931801, + 0.004661536775529385, + 0.023820552974939346, + 0.012452254071831703, + 0.01687263324856758, + 0.11178554594516754, + 0.013540603220462799, + 0.07624499499797821, + 0.02303985320031643, + 0.012255921959877014, + 0.0036144116893410683, + 0.003941047471016645, + 0.0033505011815577745, + 0.0035749145317822695, + 0.0033717129845172167, + 0.008843489922583103, + 0.011635931208729744, + 0.00395527109503746, + 0.0022187174763530493, + 0.0038917637430131435, + 0.0063367425464093685, + 0.003098890883848071, + 0.005087555386126041, + 0.0035331575199961662, + 0.002718317788094282, + 0.004580566193908453, + 0.003887784667313099, + 0.00939426850527525, + 0.003732172306627035, + 0.0035862980876117945, + 0.004795814398676157, + 0.007776226848363876, + 0.0038139226380735636, + 0.003971972037106752, + 0.007771249860525131, + 0.00891182105988264, + 0.005465409252792597, + 0.009484415873885155, + 0.004663045518100262, + 0.008475751616060734, + 0.010567300021648407, + 0.012279369868338108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.020590132102370262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Ultra-high-speed macro photography captured from a static, high-angle perspective documents the mesmerizing process of crystallization on a perfectly smooth, matte black surface. The scene is illuminated by intense, rhythmic flashes of cold, bright white light, creating a striking strobe effect. The action begins with a few tiny, individual grains of white sugar, their translucent facets capturing the initial flashes. Then, in a series of sharp, stuttering advances synchronized with each pulse of light, complex, interlocking geometric structures explosively sprout from these nucleation points. Needle-like fragments and massive, prismatic shapes jerk forward, methodically stacking upon one another to form a dense, intricate grid that spans the entire frame. The harsh, clinical light casts deep, fleeting shadows and creates dazzling, mirrored highlights along the crystal's edges, making the entire structure shimmer in cold fire. Growth is not a smooth, uninterrupted process, but a hypnotic, abrasive sequence that conveys a powerful and captivating atmosphere of fundamental natural processes revealed in discrete, frozen moments of time.", + "seed": 20264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20264_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.421382100135088, + "vlm_logprob": -0.421382100135088, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.421382100135088, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09102880209684372 + ], + "opposite_logprob_tokens": [ + -2.8410286903381348 + ], + "logprob_reward": -0.09102880209684372, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09102880209684372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22594954073429108, + 0.2071773111820221, + 0.0689389780163765, + 0.06694301217794418, + 0.11220061033964157, + 0.06957437843084335, + 0.10994288325309753, + 0.06922920793294907, + 0.04038604348897934, + 0.029658079147338867 + ], + "credit_map": [ + 0.020634770393371582, + 0.026874713599681854, + 0.03447861969470978, + 0.03758229315280914, + 0.011993586085736752, + 0.060964152216911316, + 0.014566090889275074, + 0.018855294212698936, + 0.041452519595623016, + 0.014421425759792328, + 0.012466457672417164, + 0.006822689436376095, + 0.01264969352632761, + 0.07341194897890091, + 0.00987637136131525, + 0.03607619181275368, + 0.009068205952644348, + 0.008148711174726486, + 0.006325557827949524, + 0.0070433830842375755, + 0.006421884521842003, + 0.01011382695287466, + 0.007027964573353529, + 0.014789442531764507, + 0.009652690030634403, + 0.006598429288715124, + 0.007817136123776436, + 0.0073510087095201015, + 0.010967032052576542, + 0.007322198711335659, + 0.008931643329560757, + 0.008302869275212288, + 0.020108789205551147, + 0.021655380725860596, + 0.012662126682698727, + 0.0070079006254673, + 0.009890034794807434, + 0.01307322084903717, + 0.015512008219957352, + 0.012291138991713524, + 0.014491410925984383, + 0.00846833549439907, + 0.009482555091381073, + 0.004987466149032116, + 0.005826570093631744, + 0.010978495702147484, + 0.008549212478101254, + 0.006790325976908207, + 0.017178373411297798, + 0.009822090156376362, + 0.00891303550451994, + 0.02888506092131138, + 0.007621203549206257, + 0.014525539241731167, + 0.00963516067713499, + 0.013362422585487366, + 0.008116992190480232, + 0.008296902291476727, + 0.009565354324877262, + 0.00898920837789774, + 0.010637764818966389, + 0.006316265091300011, + 0.011231104843318462, + 0.006075612735003233, + 0.005675066262483597, + 0.004622860811650753, + 0.0043472484685480595, + 0.004071023315191269, + 0.00792168639600277, + 0.004084703512489796, + 0.004928054288029671, + 0.004735392984002829, + 0.003148613963276148, + 0.0038497282657772303, + 0.005024573765695095, + 0.0033454573713243008, + 0.005494813434779644, + 0.002228308701887727, + 0.0027926634065806866, + 0.0037739186082035303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.09102880209684372 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103082418441772 + ], + "opposite_logprob_tokens": [ + -0.16030828654766083 + ], + "logprob_reward": -1.9103082418441772, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103082418441772, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.254206120967865, + 0.29135560989379883, + 0.14470332860946655, + 0.03390203043818474, + 0.032006409019231796, + 0.03996380418539047, + 0.06261677294969559, + 0.054684944450855255, + 0.023978112265467644, + 0.06258287280797958 + ], + "credit_map": [ + 0.024898633360862732, + 0.02865997888147831, + 0.026232002303004265, + 0.03657706081867218, + 0.015136152505874634, + 0.06188809126615524, + 0.020824840292334557, + 0.03998937830328941, + 0.0851985514163971, + 0.017852570861577988, + 0.013640984892845154, + 0.00736773107200861, + 0.014620267786085606, + 0.11859139800071716, + 0.010665624402463436, + 0.023418502882122993, + 0.017627974972128868, + 0.018369406461715698, + 0.0177792739123106, + 0.012032736092805862, + 0.020325396209955215, + 0.030021827667951584, + 0.011433294042944908, + 0.01711340621113777, + 0.0065508028492331505, + 0.004297653213143349, + 0.006114444229751825, + 0.0032281074672937393, + 0.002746592042967677, + 0.0038492302410304546, + 0.003503154031932354, + 0.003612048691138625, + 0.004242682363837957, + 0.0031613826286047697, + 0.005334933288395405, + 0.0024217767640948296, + 0.005341082811355591, + 0.004134920425713062, + 0.004080794285982847, + 0.003288836218416691, + 0.008330689743161201, + 0.005511046387255192, + 0.004316616337746382, + 0.0029748696833848953, + 0.004140892066061497, + 0.006661486811935902, + 0.003979916218668222, + 0.004048284143209457, + 0.0073790522292256355, + 0.005205392371863127, + 0.0046824016608297825, + 0.02010156773030758, + 0.0035710749216377735, + 0.0050277928821742535, + 0.00624810392037034, + 0.010401385836303234, + 0.006020328029990196, + 0.005900138523429632, + 0.004578479565680027, + 0.004821448121219873, + 0.015562910586595535, + 0.0056894076988101006, + 0.007847480475902557, + 0.0042647505179047585, + 0.0035695857368409634, + 0.00236708321608603, + 0.002957000397145748, + 0.0036160435993224382, + 0.002856154926121235, + 0.00335224368609488, + 0.0030089530628174543, + 0.002251045312732458, + 0.006541151087731123, + 0.007952043786644936, + 0.009405205026268959, + 0.006572050973773003, + 0.010744292289018631, + 0.005987176671624184, + 0.005140167661011219, + 0.010240783914923668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -1.9103082418441772 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813149780035019 + ], + "opposite_logprob_tokens": [ + -3.2881314754486084 + ], + "logprob_reward": -0.03813149780035019, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813149780035019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3313768804073334, + 0.37683728337287903, + 0.021614782512187958, + 0.021118536591529846, + 0.04863165691494942, + 0.0976782813668251, + 0.04630051180720329, + 0.011188306845724583, + 0.022585701197385788, + 0.022668050602078438 + ], + "credit_map": [ + 0.03388539329171181, + 0.03574183210730553, + 0.03433104231953621, + 0.05149156227707863, + 0.018236972391605377, + 0.07891051471233368, + 0.02498859167098999, + 0.05379099398851395, + 0.10563012957572937, + 0.02135237492620945, + 0.01779114082455635, + 0.00662212073802948, + 0.01634872704744339, + 0.15545468032360077, + 0.011531094089150429, + 0.042107000946998596, + 0.003530171001330018, + 0.002932175761088729, + 0.002477043541148305, + 0.0015499580185860395, + 0.0022023413330316544, + 0.003292627865448594, + 0.0021472154185175896, + 0.0034832481760531664, + 0.002722481032833457, + 0.0018961193272843957, + 0.0026252062525600195, + 0.0023095211945474148, + 0.003593639237806201, + 0.002219889312982559, + 0.002876897109672427, + 0.002874783705919981, + 0.01029276940971613, + 0.005011124536395073, + 0.010129488073289394, + 0.003493138123303652, + 0.0036174089182168245, + 0.00521808909252286, + 0.0053426166996359825, + 0.005527024157345295, + 0.02414528839290142, + 0.009930908679962158, + 0.014754562638700008, + 0.007603800855576992, + 0.006621495820581913, + 0.017074046656489372, + 0.010159933939576149, + 0.007388248108327389, + 0.015518639236688614, + 0.005826071370393038, + 0.003294296795502305, + 0.010365131311118603, + 0.0021647876128554344, + 0.0021780270617455244, + 0.0038335418794304132, + 0.0031200158409774303, + 0.0013633666094392538, + 0.0020000035874545574, + 0.001421330962330103, + 0.001322415773756802, + 0.001794473035261035, + 0.001012606779113412, + 0.0014741327613592148, + 0.0007999764056876302, + 0.0022282821591943502, + 0.0030655325390398502, + 0.002474348759278655, + 0.0019386103376746178, + 0.005871870554983616, + 0.002626467030495405, + 0.0032209078781306744, + 0.0011596819385886192, + 0.0027360005769878626, + 0.0029842110816389322, + 0.0021529817022383213, + 0.002075992291793227, + 0.005954457446932793, + 0.0019762516021728516, + 0.0025432626716792583, + 0.00224489439278841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.03813149780035019 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373025730252266 + ], + "opposite_logprob_tokens": [ + -3.4087302684783936 + ], + "logprob_reward": -0.03373025730252266, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373025730252266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19157887995243073, + 0.16348594427108765, + 0.14875799417495728, + 0.08714444935321808, + 0.07831136137247086, + 0.04935818165540695, + 0.0798225924372673, + 0.03877090662717819, + 0.053223226219415665, + 0.10954652726650238 + ], + "credit_map": [ + 0.016262508928775787, + 0.026007631793618202, + 0.023815972730517387, + 0.04102075845003128, + 0.011103085242211819, + 0.046537939459085464, + 0.012661157175898552, + 0.014169838279485703, + 0.038375817239284515, + 0.011085388250648975, + 0.012198646552860737, + 0.008965426124632359, + 0.008634260855615139, + 0.05684783682227135, + 0.007737454492598772, + 0.019641108810901642, + 0.017745833843946457, + 0.017341019585728645, + 0.01656903512775898, + 0.018567116931080818, + 0.020776808261871338, + 0.02677202597260475, + 0.014803087338805199, + 0.016183072701096535, + 0.014354496262967587, + 0.007921911776065826, + 0.008241835050284863, + 0.009387453086674213, + 0.011097753420472145, + 0.009735330007970333, + 0.013044508174061775, + 0.013361158780753613, + 0.01010953914374113, + 0.008279037661850452, + 0.011140820570290089, + 0.007171447388827801, + 0.015081800520420074, + 0.011130059137940407, + 0.00740864546969533, + 0.00799001008272171, + 0.009753700345754623, + 0.004570007789880037, + 0.005017245654016733, + 0.004546733107417822, + 0.00393584743142128, + 0.006145373918116093, + 0.006530628073960543, + 0.008858643472194672, + 0.024338049814105034, + 0.009964647702872753, + 0.005576467607170343, + 0.02106972225010395, + 0.003762675914913416, + 0.005040501244366169, + 0.004917202051728964, + 0.005153321661055088, + 0.00403651874512434, + 0.004913174547255039, + 0.004796134773641825, + 0.00425734044983983, + 0.006946233566850424, + 0.0042008222080767155, + 0.006244410295039415, + 0.0033762750681489706, + 0.004960689228028059, + 0.00667792372405529, + 0.00810295157134533, + 0.005163269583135843, + 0.011428875848650932, + 0.005224448163062334, + 0.00793374702334404, + 0.0037313203793019056, + 0.008526074700057507, + 0.013395688496530056, + 0.022068627178668976, + 0.010075743310153484, + 0.011502953246235847, + 0.00933188758790493, + 0.00954144261777401, + 0.0251041017472744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03373025730252266 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371170163154602 + ], + "opposite_logprob_tokens": [ + -3.4087116718292236 + ], + "logprob_reward": -0.03371170163154602, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371170163154602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3580981194972992, + 0.11251623928546906, + 0.0615430548787117, + 0.017447030171751976, + 0.0443250797688961, + 0.04143219441175461, + 0.07605774700641632, + 0.05769357830286026, + 0.05979021638631821, + 0.17109669744968414 + ], + "credit_map": [ + 0.03534403443336487, + 0.03380311280488968, + 0.036752521991729736, + 0.06751707941293716, + 0.024464497342705727, + 0.05275269225239754, + 0.045354947447776794, + 0.06210922449827194, + 0.011451300233602524, + 0.009051814675331116, + 0.01502428948879242, + 0.005104895681142807, + 0.012179679237306118, + 0.020611392334103584, + 0.015234175138175488, + 0.023858696222305298, + 0.008514954708516598, + 0.00826145987957716, + 0.00770149240270257, + 0.00579131068661809, + 0.008651770651340485, + 0.009792424738407135, + 0.006749869789928198, + 0.006079770624637604, + 0.0020070341415703297, + 0.002128315158188343, + 0.003192284144461155, + 0.0014647250063717365, + 0.001830139895901084, + 0.0021574804559350014, + 0.00194061198271811, + 0.002726438222452998, + 0.0074831778183579445, + 0.0037675194907933474, + 0.007347550708800554, + 0.004765736870467663, + 0.0051021883264184, + 0.0055237109772861, + 0.0054045310243964195, + 0.004930665250867605, + 0.0050819129683077335, + 0.006007286254316568, + 0.004522860515862703, + 0.004338585771620274, + 0.004493957385420799, + 0.00867555197328329, + 0.0034214265178889036, + 0.004890612792223692, + 0.023630531504750252, + 0.010678160935640335, + 0.006077948026359081, + 0.017265083268284798, + 0.003687872551381588, + 0.005580119322985411, + 0.004757025744765997, + 0.004381000995635986, + 0.005576941184699535, + 0.007374826353043318, + 0.006481992546468973, + 0.006546780001372099, + 0.012254225090146065, + 0.006057222373783588, + 0.008868280798196793, + 0.004533306695520878, + 0.008115274831652641, + 0.00585784250870347, + 0.0072751292027533054, + 0.007737726904451847, + 0.009694399312138557, + 0.007326430641114712, + 0.006713797803968191, + 0.007069614715874195, + 0.023970237001776695, + 0.02104230783879757, + 0.024785194545984268, + 0.0167806688696146, + 0.039592936635017395, + 0.012157032266259193, + 0.01529459934681654, + 0.01747373305261135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.03371170163154602 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Ultra-high-speed macro photography captured from a static, high-angle perspective documents the mesmerizing process of crystallization on a perfectly smooth, matte black surface. The scene is illuminated by intense, rhythmic flashes of cold, bright white light, creating a striking strobe effect. The action begins with a few tiny, individual grains of white sugar, their translucent facets capturing the initial flashes. Then, in a series of sharp, stuttering advances synchronized with each pulse of light, complex, interlocking geometric structures explosively sprout from these nucleation points. Needle-like fragments and massive, prismatic shapes jerk forward, methodically stacking upon one another to form a dense, intricate grid that spans the entire frame. The harsh, clinical light casts deep, fleeting shadows and creates dazzling, mirrored highlights along the crystal's edges, making the entire structure shimmer in cold fire. Growth is not a smooth, uninterrupted process, but a hypnotic, abrasive sequence that conveys a powerful and captivating atmosphere of fundamental natural processes revealed in discrete, frozen moments of time.", + "seed": 20265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20265_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5100912198424339, + "vlm_logprob": -0.5100912198424339, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.5100912198424339, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25190773606300354 + ], + "opposite_logprob_tokens": [ + -1.6269077062606812 + ], + "logprob_reward": -0.25190773606300354, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.25190773606300354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1744045615196228, + 0.09859561920166016, + 0.054382506757974625, + 0.16169771552085876, + 0.036046769469976425, + 0.12208138406276703, + 0.051697514951229095, + 0.025425096973776817, + 0.028634991496801376, + 0.24703390896320343 + ], + "credit_map": [ + 0.017394540831446648, + 0.01743020862340927, + 0.02531777136027813, + 0.018811337649822235, + 0.011937704868614674, + 0.026518022641539574, + 0.03327174857258797, + 0.02372319996356964, + 0.0076774838380515575, + 0.00740366755053401, + 0.014223922975361347, + 0.007400131784379482, + 0.007776857353746891, + 0.03194247931241989, + 0.00867999903857708, + 0.013491056859493256, + 0.00882522203028202, + 0.005627773702144623, + 0.004640867933630943, + 0.009433574974536896, + 0.006418706849217415, + 0.005219617392867804, + 0.006133010145276785, + 0.008083727210760117, + 0.0362461656332016, + 0.01176441553980112, + 0.01299986895173788, + 0.014086766168475151, + 0.009814045391976833, + 0.04503124952316284, + 0.009713570587337017, + 0.022041616961359978, + 0.006391932722181082, + 0.004046193324029446, + 0.005052563734352589, + 0.003224332118406892, + 0.003946893848478794, + 0.004298086278140545, + 0.0045510646887123585, + 0.004535699263215065, + 0.01828344725072384, + 0.01586015149950981, + 0.012976848520338535, + 0.010132251307368279, + 0.013343569822609425, + 0.028792019933462143, + 0.00957741029560566, + 0.013115674257278442, + 0.005911718122661114, + 0.004807823803275824, + 0.005422727670520544, + 0.005228251218795776, + 0.005685899406671524, + 0.005925582721829414, + 0.01244668010622263, + 0.00626883190125227, + 0.002958083525300026, + 0.0025609727017581463, + 0.003411775454878807, + 0.0031222200486809015, + 0.003397720167413354, + 0.002889785682782531, + 0.004681209102272987, + 0.002403328660875559, + 0.0026419381611049175, + 0.0036751399748027325, + 0.0031443301122635603, + 0.0031721347477287054, + 0.005410265643149614, + 0.00532450620085001, + 0.0025424056220799685, + 0.002724271034821868, + 0.012522424571216106, + 0.09198407828807831, + 0.007608553394675255, + 0.006966732908040285, + 0.013946973718702793, + 0.09554475545883179, + 0.0073707373812794685, + 0.011089634150266647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.25190773606300354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1270272731781006 + ], + "opposite_logprob_tokens": [ + -0.12702716886997223 + ], + "logprob_reward": -2.1270272731781006, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1270272731781006, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.35421833395957947, + 0.136582151055336, + 0.03966302052140236, + 0.0839252844452858, + 0.039109230041503906, + 0.11351478099822998, + 0.023083360865712166, + 0.019638221710920334, + 0.02552019990980625, + 0.16474539041519165 + ], + "credit_map": [ + 0.029762789607048035, + 0.02218605950474739, + 0.04465551674365997, + 0.03261370211839676, + 0.021044688299298286, + 0.06958700716495514, + 0.07148858159780502, + 0.0628800168633461, + 0.009584481827914715, + 0.011234959587454796, + 0.01760799065232277, + 0.014965289272367954, + 0.008222037926316261, + 0.03841334208846092, + 0.017115052789449692, + 0.019439000636339188, + 0.006163690704852343, + 0.004185680765658617, + 0.005277332849800587, + 0.005581262521445751, + 0.0031621975358575583, + 0.005403456743806601, + 0.003900212235748768, + 0.005989191122353077, + 0.015178577974438667, + 0.006216989830136299, + 0.006498691625893116, + 0.005961574148386717, + 0.020793315023183823, + 0.006246041040867567, + 0.009692388586699963, + 0.013337705284357071, + 0.005212344229221344, + 0.003922815900295973, + 0.006088480353355408, + 0.0063468087464571, + 0.0028696965891867876, + 0.0033839265815913677, + 0.006597144529223442, + 0.0046880124136805534, + 0.03143271058797836, + 0.022423366084694862, + 0.012624429538846016, + 0.007719515357166529, + 0.010380392894148827, + 0.013111358508467674, + 0.0066039105877280235, + 0.009219094179570675, + 0.0017227454809471965, + 0.0069187106564641, + 0.0018683858215808868, + 0.0013305417960509658, + 0.0016132794553413987, + 0.00678311288356781, + 0.0010641736444085836, + 0.0017824119422584772, + 0.0028704912401735783, + 0.001873960718512535, + 0.0024661384522914886, + 0.0027473587542772293, + 0.0026393632870167494, + 0.0023569937329739332, + 0.0028089385014027357, + 0.0018749781884253025, + 0.002242883201688528, + 0.005800670478492975, + 0.0018398150568827987, + 0.0025498210452497005, + 0.004003468900918961, + 0.0053994320333004, + 0.0015833117067813873, + 0.002100797602906823, + 0.007454820442944765, + 0.06366309523582458, + 0.005074262619018555, + 0.003630426013842225, + 0.010616855695843697, + 0.06291641294956207, + 0.005197432823479176, + 0.0061920927837491035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -2.1270272731781006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033741552382707596 + ], + "opposite_logprob_tokens": [ + -3.4087414741516113 + ], + "logprob_reward": -0.033741552382707596, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033741552382707596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06953902542591095, + 0.4949205219745636, + 0.07852786034345627, + 0.1685059815645218, + 0.02361568622291088, + 0.06970838457345963, + 0.01100272499024868, + 0.013714903965592384, + 0.028048990294337273, + 0.04241590201854706 + ], + "credit_map": [ + 0.00490380497649312, + 0.006498144939541817, + 0.010128483176231384, + 0.005373373627662659, + 0.005456696730107069, + 0.012408631853759289, + 0.012744931504130363, + 0.012024953030049801, + 0.034283485263586044, + 0.04436120390892029, + 0.056815773248672485, + 0.06246810033917427, + 0.02932681329548359, + 0.12265210598707199, + 0.08360458165407181, + 0.06140846386551857, + 0.006446335464715958, + 0.007467525079846382, + 0.008960670791566372, + 0.016374334692955017, + 0.00632051844149828, + 0.006220492534339428, + 0.012912425212562084, + 0.013825559988617897, + 0.020981676876544952, + 0.007021473255008459, + 0.006923818960785866, + 0.00750009436160326, + 0.06779211759567261, + 0.040697064250707626, + 0.008467796258628368, + 0.009121953509747982, + 0.0026207759510725737, + 0.0025341880973428488, + 0.0032553328201174736, + 0.0022726920433342457, + 0.002185724675655365, + 0.0024893158115446568, + 0.004184551537036896, + 0.0040731062181293964, + 0.01553067471832037, + 0.013825863599777222, + 0.009171278215944767, + 0.004193384665995836, + 0.008691424503922462, + 0.0071310424245893955, + 0.005223482847213745, + 0.005941231735050678, + 0.0018711966695263982, + 0.0013685959856957197, + 0.0008800287032499909, + 0.0013496760511770844, + 0.00104424636811018, + 0.0018325578421354294, + 0.0011513178469613194, + 0.0015051052905619144, + 0.0018715072656050324, + 0.0013860873878002167, + 0.0016810611123219132, + 0.0017463340191170573, + 0.0014597716508433223, + 0.0015081648016348481, + 0.002634132979437709, + 0.0014278445160016418, + 0.002354338299483061, + 0.004282989539206028, + 0.0036545887123793364, + 0.0026969395112246275, + 0.00501334248110652, + 0.005094647407531738, + 0.0028743408620357513, + 0.0020778030157089233, + 0.002186676487326622, + 0.012696940451860428, + 0.002331152558326721, + 0.0017466965364292264, + 0.004449538420885801, + 0.014263259246945381, + 0.0027454523369669914, + 0.0019961881916970015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.033741552382707596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909943699836731 + ], + "opposite_logprob_tokens": [ + -2.464099407196045 + ], + "logprob_reward": -0.08909943699836731, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909943699836731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0475117564201355, + 0.1540297269821167, + 0.07202712446451187, + 0.0311442781239748, + 0.01589162088930607, + 0.022647390142083168, + 0.03144010528922081, + 0.010707680135965347, + 0.03493301197886467, + 0.579667329788208 + ], + "credit_map": [ + 0.004442671313881874, + 0.003611894091591239, + 0.0070106759667396545, + 0.004435990937054157, + 0.002900186227634549, + 0.009016494266688824, + 0.007603977806866169, + 0.008489866741001606, + 0.010691315867006779, + 0.012582208961248398, + 0.017919303849339485, + 0.017848001793026924, + 0.009883079677820206, + 0.04168824478983879, + 0.022342370823025703, + 0.02107519470155239, + 0.006781356409192085, + 0.007414643187075853, + 0.009028168395161629, + 0.014396288432180882, + 0.006213068962097168, + 0.005779439117759466, + 0.010591525584459305, + 0.011822633445262909, + 0.006916023790836334, + 0.0022718135733157396, + 0.002665360691025853, + 0.0037086401134729385, + 0.006562323775142431, + 0.00249281944707036, + 0.0019902486819773912, + 0.004537047818303108, + 0.002401546575129032, + 0.0017306836089119315, + 0.001708166440948844, + 0.001448430703021586, + 0.001544708153232932, + 0.0015597781166434288, + 0.002475850749760866, + 0.0030224567744880915, + 0.003608351806178689, + 0.0036660281475633383, + 0.0035426681861281395, + 0.0017558910185471177, + 0.002869401825591922, + 0.0034678576048463583, + 0.0015400726115331054, + 0.0021971198730170727, + 0.002386280335485935, + 0.007735737133771181, + 0.0022941280622035265, + 0.001805163687095046, + 0.0021104428451508284, + 0.00994519330561161, + 0.0027790062595158815, + 0.0023841538932174444, + 0.001911723637022078, + 0.0012441184371709824, + 0.0012547995429486036, + 0.0015690976288169622, + 0.0012573051499202847, + 0.0011752748396247625, + 0.0015440522693097591, + 0.0007513081654906273, + 0.0028177157510071993, + 0.0046561602503061295, + 0.0034626973792910576, + 0.007166375871747732, + 0.004369063302874565, + 0.0040346309542655945, + 0.004939641337841749, + 0.0034867257345467806, + 0.027872610837221146, + 0.22810009121894836, + 0.014897330664098263, + 0.01052386499941349, + 0.029382437467575073, + 0.2330397665500641, + 0.014653853140771389, + 0.0211973674595356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.08909943699836731 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048680100589990616 + ], + "opposite_logprob_tokens": [ + -3.048680067062378 + ], + "logprob_reward": -0.048680100589990616, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048680100589990616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07631266117095947, + 0.023391835391521454, + 0.028292065486311913, + 0.03164110332727432, + 0.019443605095148087, + 0.020260384306311607, + 0.03043507970869541, + 0.010357814840972424, + 0.021291129291057587, + 0.7385742664337158 + ], + "credit_map": [ + 0.0066768089309334755, + 0.006047221831977367, + 0.009813759475946426, + 0.007586220279335976, + 0.0050062621012330055, + 0.013794451020658016, + 0.01371023803949356, + 0.013677695766091347, + 0.0015194296138361096, + 0.0024220896884799004, + 0.002987149404361844, + 0.0029840474016964436, + 0.0018543562619015574, + 0.004639406688511372, + 0.004455015528947115, + 0.002530338941141963, + 0.0024997987784445286, + 0.0032024485990405083, + 0.003214776050299406, + 0.00536939175799489, + 0.00253759091719985, + 0.002866131253540516, + 0.004413947928696871, + 0.004187980201095343, + 0.006759421434253454, + 0.001994784688577056, + 0.0018317915964871645, + 0.002537759952247143, + 0.010280050337314606, + 0.003057698952034116, + 0.00241087400354445, + 0.002768725622445345, + 0.004657632205635309, + 0.001866076374426484, + 0.0016807463252916932, + 0.002095986157655716, + 0.0011944278376176953, + 0.0015999894822016358, + 0.003068363294005394, + 0.003280382603406906, + 0.0037599357310682535, + 0.0035801429767161608, + 0.0028645105194300413, + 0.0019174405606463552, + 0.0019062100909650326, + 0.0022994026076048613, + 0.0020641942974179983, + 0.001868546474725008, + 0.004384416621178389, + 0.00427916320040822, + 0.0013838657177984715, + 0.0013542856322601438, + 0.0023961651604622602, + 0.011210980825126171, + 0.003791735740378499, + 0.0016344659961760044, + 0.0012666797265410423, + 0.001109987497329712, + 0.001621172414161265, + 0.0014000701485201716, + 0.0010356982238590717, + 0.0012341573601588607, + 0.0017281771870329976, + 0.00096187210874632, + 0.0016887368401512504, + 0.003180142492055893, + 0.0026069441810250282, + 0.0028561085928231478, + 0.002333571668714285, + 0.003505690023303032, + 0.00288358423858881, + 0.0022363511379808187, + 0.035745322704315186, + 0.28966668248176575, + 0.018908796831965446, + 0.014165109023451805, + 0.03780672326683998, + 0.2970171272754669, + 0.018200451508164406, + 0.02706412971019745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.048680100589990616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Ultra-high-speed macro photography captured from a static, high-angle perspective documents the mesmerizing process of crystallization on a perfectly smooth, matte black surface. The scene is illuminated by intense, rhythmic flashes of cold, bright white light, creating a striking strobe effect. The action begins with a few tiny, individual grains of white sugar, their translucent facets capturing the initial flashes. Then, in a series of sharp, stuttering advances synchronized with each pulse of light, complex, interlocking geometric structures explosively sprout from these nucleation points. Needle-like fragments and massive, prismatic shapes jerk forward, methodically stacking upon one another to form a dense, intricate grid that spans the entire frame. The harsh, clinical light casts deep, fleeting shadows and creates dazzling, mirrored highlights along the crystal's edges, making the entire structure shimmer in cold fire. Growth is not a smooth, uninterrupted process, but a hypnotic, abrasive sequence that conveys a powerful and captivating atmosphere of fundamental natural processes revealed in discrete, frozen moments of time.", + "seed": 20266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_7/index_9935_seed_20266_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.8449229955673218, + "vlm_logprob": -0.8449229955673218, + "vlm_yes_no_margin": 1.225, + "vlm_token_credit": -0.8449229955673218, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5517035722732544 + ], + "opposite_logprob_tokens": [ + -0.9267035722732544 + ], + "logprob_reward": -0.5517035722732544, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5517035722732544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10797649621963501, + 0.11425813287496567, + 0.12993715703487396, + 0.10582908987998962, + 0.2918860614299774, + 0.0444895438849926, + 0.018003586679697037, + 0.046875108033418655, + 0.075576551258564, + 0.0651683434844017 + ], + "credit_map": [ + 0.010591177269816399, + 0.011939121410250664, + 0.01403817255049944, + 0.013542319647967815, + 0.015467021614313126, + 0.013906675390899181, + 0.011255297809839249, + 0.017236707732081413, + 0.013479435816407204, + 0.015644049271941185, + 0.006576552055776119, + 0.009782172739505768, + 0.019758479669690132, + 0.01842758245766163, + 0.015625260770320892, + 0.014964592643082142, + 0.01128684263676405, + 0.022012272849678993, + 0.006903328932821751, + 0.011390233412384987, + 0.015906566753983498, + 0.018810024484992027, + 0.020113207399845123, + 0.02351466566324234, + 0.008951766416430473, + 0.013319876044988632, + 0.010438564233481884, + 0.012101169675588608, + 0.014930544421076775, + 0.01390952430665493, + 0.01950598880648613, + 0.01267164759337902, + 0.02730065956711769, + 0.0460766963660717, + 0.026295850053429604, + 0.026948275044560432, + 0.0479869470000267, + 0.041496459394693375, + 0.04077098146080971, + 0.035010192543268204, + 0.004929602146148682, + 0.004853144753724337, + 0.006845775526016951, + 0.005455496720969677, + 0.006650728173553944, + 0.005622092168778181, + 0.005650739185512066, + 0.004481963813304901, + 0.002127000130712986, + 0.002281184308230877, + 0.0016788466600701213, + 0.0021786107681691647, + 0.0031494933646172285, + 0.001965148141607642, + 0.0020165937021374702, + 0.002606709022074938, + 0.005053091328591108, + 0.0056766546331346035, + 0.004747288767248392, + 0.005916023161262274, + 0.005868471693247557, + 0.006372965406626463, + 0.008892696350812912, + 0.004347912967205048, + 0.012756185606122017, + 0.015984851866960526, + 0.0035672092344611883, + 0.0031538442708551884, + 0.027918968349695206, + 0.004046199843287468, + 0.005002796184271574, + 0.0031464826315641403, + 0.004226634744554758, + 0.0034517236053943634, + 0.02212105132639408, + 0.02299860119819641, + 0.0035358378663659096, + 0.0024722330272197723, + 0.003472525393590331, + 0.0028897279407829046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.5517035722732544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.5297980308532715 + ], + "opposite_logprob_tokens": [ + -0.02979809045791626 + ], + "logprob_reward": -3.5297980308532715, + "yes_no_margin_reward": -3.5, + "token_credit_reward": -3.5297980308532715, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12759947776794434, + 0.08710844814777374, + 0.10076414048671722, + 0.05462285503745079, + 0.13188478350639343, + 0.03572787344455719, + 0.07188208401203156, + 0.06655757874250412, + 0.07075043022632599, + 0.2531023919582367 + ], + "credit_map": [ + 0.01651267148554325, + 0.018255935981869698, + 0.015335685573518276, + 0.01670987345278263, + 0.01392290834337473, + 0.015385820530354977, + 0.011039013043045998, + 0.02043755166232586, + 0.010282597504556179, + 0.01273719035089016, + 0.0059636440128088, + 0.008317811414599419, + 0.013401324860751629, + 0.009885052219033241, + 0.015178720466792583, + 0.011342101730406284, + 0.009303106926381588, + 0.01607557199895382, + 0.009927977807819843, + 0.012557046487927437, + 0.014127116650342941, + 0.010850105434656143, + 0.01494051143527031, + 0.012982693500816822, + 0.004706310108304024, + 0.0074640903621912, + 0.0057320185005664825, + 0.005747254006564617, + 0.008303945884108543, + 0.007604509126394987, + 0.009022120386362076, + 0.006042602472007275, + 0.011471879668533802, + 0.021342119202017784, + 0.011758505366742611, + 0.015604772605001926, + 0.021054301410913467, + 0.014107463881373405, + 0.021099895238876343, + 0.015445838682353497, + 0.003960090223699808, + 0.004096019547432661, + 0.004806871991604567, + 0.004449129104614258, + 0.0052533927373588085, + 0.003842257894575596, + 0.004278575070202351, + 0.005041534546762705, + 0.008279334753751755, + 0.011163534596562386, + 0.008433898910880089, + 0.0104405852034688, + 0.008801370859146118, + 0.00807330384850502, + 0.008756838738918304, + 0.007933209650218487, + 0.007082068361341953, + 0.010415663942694664, + 0.00911574624478817, + 0.007078952621668577, + 0.008984808810055256, + 0.007752660661935806, + 0.008538166992366314, + 0.007589507382363081, + 0.008206890895962715, + 0.011850817129015923, + 0.007078972645103931, + 0.00624639680609107, + 0.01121454406529665, + 0.007144484203308821, + 0.011152773164212704, + 0.007855543866753578, + 0.01731603965163231, + 0.013008542358875275, + 0.09002970904111862, + 0.0901719480752945, + 0.013855715282261372, + 0.00711067346855998, + 0.012560555711388588, + 0.009049165062606335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -3.5297980308532715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07016785442829132 + ], + "opposite_logprob_tokens": [ + -2.6951677799224854 + ], + "logprob_reward": -0.07016785442829132, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07016785442829132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11002826690673828, + 0.08621206879615784, + 0.0683477595448494, + 0.05316933989524841, + 0.08820454776287079, + 0.0279252827167511, + 0.133518248796463, + 0.03598643094301224, + 0.31558293104171753, + 0.08102508634328842 + ], + "credit_map": [ + 0.010657192207872868, + 0.012233086861670017, + 0.013502527959644794, + 0.01582256332039833, + 0.02033381722867489, + 0.012599115259945393, + 0.013715647161006927, + 0.011164320632815361, + 0.008242476731538773, + 0.01593632623553276, + 0.00561008183285594, + 0.007113782688975334, + 0.009818170219659805, + 0.01460390817373991, + 0.010796267539262772, + 0.01409105770289898, + 0.007517364341765642, + 0.01017365138977766, + 0.003529279725626111, + 0.005106687545776367, + 0.007397955749183893, + 0.010834624990820885, + 0.012732798233628273, + 0.01105540432035923, + 0.005263808649033308, + 0.0068864282220602036, + 0.00506589887663722, + 0.00703191664069891, + 0.007873045280575752, + 0.006372442934662104, + 0.009319846518337727, + 0.005355955101549625, + 0.0065995086915791035, + 0.015437081456184387, + 0.00788834411650896, + 0.009652343578636646, + 0.016284499317407608, + 0.010213570669293404, + 0.013233812525868416, + 0.008895392529666424, + 0.002574273617938161, + 0.0034370762296020985, + 0.0035515122581273317, + 0.0030888088513165712, + 0.005842430517077446, + 0.00325802736915648, + 0.0037742750719189644, + 0.0023988820612430573, + 0.017378129065036774, + 0.03194553777575493, + 0.014441740699112415, + 0.010798133909702301, + 0.01499140914529562, + 0.01875295303761959, + 0.01567329280078411, + 0.00953706819564104, + 0.0035486945416778326, + 0.004110071808099747, + 0.0037130771670490503, + 0.004001581110060215, + 0.005104305688291788, + 0.005907104816287756, + 0.00662439689040184, + 0.00297720218077302, + 0.03661985695362091, + 0.09408973157405853, + 0.009582294151186943, + 0.01097572036087513, + 0.1234397366642952, + 0.013735909014940262, + 0.017697615548968315, + 0.009442076086997986, + 0.00534469448029995, + 0.005021718796342611, + 0.025924744084477425, + 0.027541231364011765, + 0.004303231835365295, + 0.0044443560764193535, + 0.004448396153748035, + 0.003996716346591711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.07016785442829132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043121859431266785 + ], + "opposite_logprob_tokens": [ + -3.168121814727783 + ], + "logprob_reward": -0.043121859431266785, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043121859431266785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09810816496610641, + 0.13067640364170074, + 0.08902891725301743, + 0.05675213411450386, + 0.05157293751835823, + 0.033177558332681656, + 0.06941007822751999, + 0.053061991930007935, + 0.24709653854370117, + 0.1711152344942093 + ], + "credit_map": [ + 0.009211828000843525, + 0.015245256945490837, + 0.01331514585763216, + 0.013938086107373238, + 0.01664951629936695, + 0.009323004633188248, + 0.009492280893027782, + 0.010933060199022293, + 0.014881813898682594, + 0.023330628871917725, + 0.008106710389256477, + 0.012140468694269657, + 0.0174835417419672, + 0.018794633448123932, + 0.019160611554980278, + 0.01677800342440605, + 0.009666457772254944, + 0.012121876701712608, + 0.004460941068828106, + 0.007122864481061697, + 0.011428874917328358, + 0.013703538104891777, + 0.018647443503141403, + 0.011876922100782394, + 0.005619474686682224, + 0.009181279689073563, + 0.004251581151038408, + 0.006184145342558622, + 0.007096289191395044, + 0.007046943996101618, + 0.011759061366319656, + 0.005613364744931459, + 0.006399893667548895, + 0.007570510730147362, + 0.0068985698744654655, + 0.00664365803822875, + 0.007537615019828081, + 0.004570705816149712, + 0.0068352110683918, + 0.005116776563227177, + 0.003937693312764168, + 0.004559644032269716, + 0.004084767773747444, + 0.004858415573835373, + 0.004324529320001602, + 0.003980504814535379, + 0.003990220837295055, + 0.0034417847637087107, + 0.009741595014929771, + 0.016673995181918144, + 0.0070555731654167175, + 0.006717266980558634, + 0.008731632493436337, + 0.006733902730047703, + 0.0077625843696296215, + 0.0059935362078249454, + 0.004503758158534765, + 0.005977917928248644, + 0.006383783649653196, + 0.005952139385044575, + 0.009613461792469025, + 0.005881104152649641, + 0.0090084969997406, + 0.005741329398006201, + 0.0208881925791502, + 0.08188647031784058, + 0.007477653678506613, + 0.009864620864391327, + 0.09461980313062668, + 0.01062975823879242, + 0.014285109005868435, + 0.007444965653121471, + 0.011513750068843365, + 0.008524321019649506, + 0.06205921992659569, + 0.06168553605675697, + 0.00723812123760581, + 0.004678403027355671, + 0.00877016969025135, + 0.006645736750215292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.043121859431266785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective static and high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface perfectly smooth and matte black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the light sources intense, rhythmic flashes of cold, bright white light creating a strobe effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do complex, interlocking geometric structures sprout from tiny white sugar grains in sharp, stuttering advances synchronized with the light pulses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystals form a dense, intricate grid spanning the entire frame with deep shadows and dazzling mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029823660850524902 + ], + "opposite_logprob_tokens": [ + -3.5298237800598145 + ], + "logprob_reward": -0.029823660850524902, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029823660850524902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0782792717218399, + 0.07095278799533844, + 0.08107148855924606, + 0.038191597908735275, + 0.03891320526599884, + 0.03128977119922638, + 0.04690845310688019, + 0.05676037818193436, + 0.4368686378002167, + 0.12076443433761597 + ], + "credit_map": [ + 0.006743101868778467, + 0.008399185724556446, + 0.009700547903776169, + 0.014083916321396828, + 0.012118971906602383, + 0.008459295146167278, + 0.010167277418076992, + 0.008606974966824055, + 0.00676138699054718, + 0.00970473513007164, + 0.004224844742566347, + 0.005369917955249548, + 0.009087425656616688, + 0.011412274092435837, + 0.009862092323601246, + 0.014530113898217678, + 0.006906561553478241, + 0.012487439438700676, + 0.005975448060780764, + 0.008312872610986233, + 0.012424057349562645, + 0.009543300606310368, + 0.014902730472385883, + 0.010519076138734818, + 0.0038007409311830997, + 0.006609043572098017, + 0.0034520456101745367, + 0.004264173563569784, + 0.004820005502551794, + 0.004850404802709818, + 0.006718868389725685, + 0.0036763164680451155, + 0.004377223551273346, + 0.006357763893902302, + 0.004632413387298584, + 0.005015120375901461, + 0.006049601826816797, + 0.003906298894435167, + 0.0048707835376262665, + 0.003704003058373928, + 0.003315524896606803, + 0.0035395845770835876, + 0.004465966019779444, + 0.0039240773767232895, + 0.0043196240440011024, + 0.0039012732449918985, + 0.0049416604451835155, + 0.0028820601291954517, + 0.007140877190977335, + 0.010700264014303684, + 0.004745290149003267, + 0.0036425823345780373, + 0.00559223210439086, + 0.004943410400301218, + 0.0055970726534724236, + 0.004546725191175938, + 0.006055610254406929, + 0.006467567291110754, + 0.005653316155076027, + 0.006554430350661278, + 0.007811867631971836, + 0.00773246306926012, + 0.010656981728971004, + 0.005828135646879673, + 0.08178223669528961, + 0.0933753252029419, + 0.015575387515127659, + 0.016726898029446602, + 0.17097732424736023, + 0.01913408190011978, + 0.025653328746557236, + 0.013644056394696236, + 0.008846094831824303, + 0.006803900469094515, + 0.04253190755844116, + 0.04152752086520195, + 0.0051513547077775, + 0.0036757246125489473, + 0.0067509617656469345, + 0.005476971622556448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.029823660850524902 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_8/index_tensor([6806, 6806])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_8/index_tensor([6806, 6806])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e750973f6514c0b26c698e4ed8e72cba9e5dc860 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_8/index_tensor([6806, 6806])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping drone shot, slowly zooming out, reveals a surreal scene: a massive rectangular dining table, crafted from weathered, dark oak. Its polished surface reflects the vast, cloudless sky above, revealing distinct wood grain and subtle signs of wear. Like a majestic ancient ship, it steadily sails across a boundless, sapphire-like sea. The table glides forward with a steady, deliberate pace, its prow-like edge gracefully slicing the nearly glassy surface, leaving a faint V-shaped wake in its wake before swiftly dissipating. Bathed in the bright, clear light of midday, the entire scene dances on the wooden surface and gently ripples, contrasting starkly with the deep blue waters stretching to the perfect horizon in all directions, creating a sense of ultimate tranquility, solitude, and a dreamlike journey.", + "seed": 17271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_8/index_6806_seed_17271_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3674787857569754, + "vlm_logprob": -0.3674787857569754, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.3674787857569754, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7020373344421387 + ], + "opposite_logprob_tokens": [ + -0.2020372897386551 + ], + "logprob_reward": -1.7020373344421387, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7020373344421387, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1586717665195465, + 0.06531492620706558, + 0.06150854006409645, + 0.35377076268196106, + 0.07321863621473312, + 0.03383463993668556, + 0.08710083365440369, + 0.08769368380308151, + 0.05353384464979172, + 0.025352373719215393 + ], + "credit_map": [ + 0.01606663689017296, + 0.01850864663720131, + 0.023426620289683342, + 0.023379141464829445, + 0.025094570592045784, + 0.017315726727247238, + 0.01928706094622612, + 0.015593360178172588, + 0.00760950893163681, + 0.006500810384750366, + 0.008376621641218662, + 0.011219964362680912, + 0.01051892526447773, + 0.007905151695013046, + 0.006676018703728914, + 0.006507919169962406, + 0.0065740253776311874, + 0.005278836935758591, + 0.008575741201639175, + 0.012294741347432137, + 0.008285986259579659, + 0.00613747164607048, + 0.008268529549241066, + 0.006093211472034454, + 0.052179787307977676, + 0.022594653069972992, + 0.05191413313150406, + 0.05597155913710594, + 0.03380340337753296, + 0.03504391387104988, + 0.038515605032444, + 0.06374771147966385, + 0.013645276427268982, + 0.009874965064227581, + 0.007890765555202961, + 0.007734687998890877, + 0.008073647506535053, + 0.011027365922927856, + 0.00766650540754199, + 0.007305425591766834, + 0.003001562785357237, + 0.0022412361577153206, + 0.0038982052356004715, + 0.006659813225269318, + 0.005189960356801748, + 0.003593009663745761, + 0.006107439752668142, + 0.003143410896882415, + 0.009783817455172539, + 0.006366704124957323, + 0.009045640006661415, + 0.014521944336593151, + 0.013791009783744812, + 0.00944240391254425, + 0.015353585593402386, + 0.008795728906989098, + 0.009132442995905876, + 0.008214051835238934, + 0.009356502443552017, + 0.01369445025920868, + 0.01320815458893776, + 0.010813447646796703, + 0.013232012279331684, + 0.010042618960142136, + 0.010402185842394829, + 0.005448373965919018, + 0.005407467018812895, + 0.006268145050853491, + 0.007722078822553158, + 0.00808684341609478, + 0.006415128707885742, + 0.003783621359616518, + 0.0028358905110508204, + 0.002539880108088255, + 0.0033672982826828957, + 0.0025638092774897814, + 0.004658146295696497, + 0.0030674913432449102, + 0.004272908437997103, + 0.0020469510927796364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -1.7020373344421387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014245836064219475 + ], + "opposite_logprob_tokens": [ + -4.264245986938477 + ], + "logprob_reward": -0.014245836064219475, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014245836064219475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1600848138332367, + 0.06416493654251099, + 0.058039791882038116, + 0.33351269364356995, + 0.08571440726518631, + 0.03980187326669693, + 0.1018417477607727, + 0.07385656982660294, + 0.05803343653678894, + 0.02494972012937069 + ], + "credit_map": [ + 0.019739367067813873, + 0.018422968685626984, + 0.024165287613868713, + 0.022013938054442406, + 0.02401193417608738, + 0.018413059413433075, + 0.017733076587319374, + 0.015585185028612614, + 0.008671018294990063, + 0.006231969688087702, + 0.007543973624706268, + 0.010308007709681988, + 0.010342280380427837, + 0.008433887735009193, + 0.005992893595248461, + 0.006640910170972347, + 0.006849630735814571, + 0.004628910217434168, + 0.008318386971950531, + 0.012193243950605392, + 0.007080412469804287, + 0.005498972721397877, + 0.00741920992732048, + 0.006051022559404373, + 0.04964815080165863, + 0.02143005095422268, + 0.048584405332803726, + 0.05051436275243759, + 0.033266063779592514, + 0.0339418388903141, + 0.0356585867702961, + 0.06046925112605095, + 0.014908120036125183, + 0.011491304263472557, + 0.009191137738525867, + 0.009510908275842667, + 0.01040156651288271, + 0.012291344814002514, + 0.00984545424580574, + 0.008074571378529072, + 0.005221948958933353, + 0.003574896836653352, + 0.004827914293855429, + 0.006386211607605219, + 0.004992691334336996, + 0.004320486914366484, + 0.0066555519588291645, + 0.00382216926664114, + 0.011976314708590508, + 0.008951563388109207, + 0.013216866180300713, + 0.014016483910381794, + 0.01616853103041649, + 0.010087700560688972, + 0.020157890394330025, + 0.007266400847584009, + 0.0071434141136705875, + 0.0065506333485245705, + 0.009919441305100918, + 0.012313189916312695, + 0.011914126574993134, + 0.00899494718760252, + 0.010554150678217411, + 0.006466665305197239, + 0.011176030151546001, + 0.005749797914177179, + 0.007133940234780312, + 0.0067164902575314045, + 0.007547191344201565, + 0.008418407291173935, + 0.006771308835595846, + 0.004520266316831112, + 0.003071567742154002, + 0.0030235550366342068, + 0.0028944984078407288, + 0.002441043732687831, + 0.004740120377391577, + 0.003057852853089571, + 0.003836349816992879, + 0.0018847316969186068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.014245836064219475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095738288015127 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095738288015127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095738288015127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17927613854408264, + 0.08123356103897095, + 0.08565057814121246, + 0.2375650554895401, + 0.09876812994480133, + 0.055053092539310455, + 0.09751245379447937, + 0.05417858809232712, + 0.05460856109857559, + 0.05615377798676491 + ], + "credit_map": [ + 0.020687609910964966, + 0.02004975825548172, + 0.026318732649087906, + 0.02547304704785347, + 0.030984245240688324, + 0.020753875374794006, + 0.01912393607199192, + 0.015884926542639732, + 0.01042886357754469, + 0.008363376371562481, + 0.00841891672462225, + 0.012197449803352356, + 0.012954444624483585, + 0.010655205696821213, + 0.007555290125310421, + 0.010660011321306229, + 0.007302554324269295, + 0.006681832484900951, + 0.012780438177287579, + 0.02071498893201351, + 0.01021557953208685, + 0.008290877565741539, + 0.010618511587381363, + 0.009045790880918503, + 0.0330236554145813, + 0.014833219349384308, + 0.03690878301858902, + 0.037337303161621094, + 0.020923562347888947, + 0.023493826389312744, + 0.026249194517731667, + 0.04479550942778587, + 0.017856037244200706, + 0.012003413401544094, + 0.010259594768285751, + 0.011187797412276268, + 0.01322131510823965, + 0.013949028216302395, + 0.011357638984918594, + 0.00893330480903387, + 0.008794961497187614, + 0.0052496520802378654, + 0.0053479610942304134, + 0.007586432620882988, + 0.008608033880591393, + 0.006855107843875885, + 0.007558907847851515, + 0.005052036140114069, + 0.008782915771007538, + 0.0083821602165699, + 0.012149480171501637, + 0.012882495298981667, + 0.018236923962831497, + 0.01003240142017603, + 0.01948348991572857, + 0.0075625889003276825, + 0.004848681390285492, + 0.004599963780492544, + 0.008941654115915298, + 0.00928336288779974, + 0.00913354940712452, + 0.006904358975589275, + 0.007009667810052633, + 0.00345735065639019, + 0.009909143671393394, + 0.004726856015622616, + 0.0074522956274449825, + 0.005209764465689659, + 0.006760882213711739, + 0.007864962331950665, + 0.007063900586217642, + 0.0056207552552223206, + 0.007665181066840887, + 0.005917870905250311, + 0.006447463296353817, + 0.007292319554835558, + 0.010547100566327572, + 0.007871277630329132, + 0.00795417558401823, + 0.002458393108099699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004095738288015127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291999369859695 + ], + "opposite_logprob_tokens": [ + -2.237920045852661 + ], + "logprob_reward": -0.11291999369859695, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291999369859695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10607043653726578, + 0.04283088073134422, + 0.07178504765033722, + 0.09181515872478485, + 0.1012926772236824, + 0.11512594670057297, + 0.30477476119995117, + 0.1297963261604309, + 0.025242919102311134, + 0.011265819892287254 + ], + "credit_map": [ + 0.014062345027923584, + 0.013140297494828701, + 0.011652221903204918, + 0.015041066333651543, + 0.01816144399344921, + 0.013995946384966373, + 0.010784000158309937, + 0.009233115240931511, + 0.003999223466962576, + 0.0034375411923974752, + 0.006829553283751011, + 0.0075128041207790375, + 0.005165800917893648, + 0.004580721724778414, + 0.005321270786225796, + 0.0059839654713869095, + 0.004689602646976709, + 0.005504393018782139, + 0.012499582022428513, + 0.018111206591129303, + 0.007389117032289505, + 0.005880360025912523, + 0.009372596628963947, + 0.008338192477822304, + 0.009911131113767624, + 0.006365431006997824, + 0.016470305621623993, + 0.0117844482883811, + 0.013195537962019444, + 0.009676150046288967, + 0.011885249987244606, + 0.012526907958090305, + 0.022726144641637802, + 0.012573729269206524, + 0.009673860855400562, + 0.008720171637833118, + 0.012530519627034664, + 0.014472771435976028, + 0.008649700321257114, + 0.011945772916078568, + 0.012551914900541306, + 0.01013738289475441, + 0.0131739042699337, + 0.016285425052046776, + 0.01935424469411373, + 0.01148873008787632, + 0.01829020492732525, + 0.013844132423400879, + 0.024456707760691643, + 0.028759481385350227, + 0.0372413694858551, + 0.043349605053663254, + 0.05312462896108627, + 0.03454948589205742, + 0.054911233484745026, + 0.028382280841469765, + 0.009856372140347958, + 0.011169351637363434, + 0.019824882969260216, + 0.015757760033011436, + 0.019276246428489685, + 0.017773540690541267, + 0.017091620713472366, + 0.01904655434191227, + 0.0030900982674211264, + 0.002834138460457325, + 0.003363455645740032, + 0.0031264221761375666, + 0.0028688302263617516, + 0.0034453573171049356, + 0.0037286917213350534, + 0.0027859266847372055, + 0.002243748866021633, + 0.0011802545050159097, + 0.0013189109740778804, + 0.0011620422592386603, + 0.0015614059520885348, + 0.0014987245667725801, + 0.0016967877745628357, + 0.0006039452273398638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.11291999369859695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095026291906834 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095026291906834, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095026291906834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16624614596366882, + 0.08153606951236725, + 0.07175619900226593, + 0.2543882131576538, + 0.11163626611232758, + 0.06206792965531349, + 0.10118994116783142, + 0.04546581208705902, + 0.05683106184005737, + 0.0488823726773262 + ], + "credit_map": [ + 0.02060273475944996, + 0.01677355356514454, + 0.0230413731187582, + 0.022867564111948013, + 0.02779580093920231, + 0.019278496503829956, + 0.018209587782621384, + 0.01767703704535961, + 0.010778488591313362, + 0.008332313969731331, + 0.009039540775120258, + 0.012854568660259247, + 0.012838601134717464, + 0.010457032360136509, + 0.007983959279954433, + 0.009251565672457218, + 0.006131446920335293, + 0.005830345209687948, + 0.010615010745823383, + 0.016329241916537285, + 0.008779987692832947, + 0.007597958669066429, + 0.008801605552434921, + 0.0076706064864993095, + 0.040724318474531174, + 0.017696434631943703, + 0.036990851163864136, + 0.03632409870624542, + 0.02555977739393711, + 0.027699768543243408, + 0.02452797070145607, + 0.044865015894174576, + 0.0216608215123415, + 0.012775580398738384, + 0.012035801075398922, + 0.012172235175967216, + 0.014060183428227901, + 0.016091424971818924, + 0.01098794024437666, + 0.011852283030748367, + 0.007212968543171883, + 0.004823003895580769, + 0.006967863533645868, + 0.011042580008506775, + 0.008356173522770405, + 0.006806155201047659, + 0.010909116826951504, + 0.005950066726654768, + 0.008356763049960136, + 0.008996928110718727, + 0.012765360064804554, + 0.013518906198441982, + 0.01839471608400345, + 0.010117707774043083, + 0.019277213141322136, + 0.009762347675859928, + 0.00473418366163969, + 0.003933749161660671, + 0.0077402470633387566, + 0.007384258788079023, + 0.007217541802674532, + 0.005679699592292309, + 0.005759743042290211, + 0.0030163866467773914, + 0.010608377866446972, + 0.005387152079492807, + 0.006550140213221312, + 0.006100555881857872, + 0.007677264045923948, + 0.008474272675812244, + 0.00721505144611001, + 0.004818246699869633, + 0.008356005884706974, + 0.005575007759034634, + 0.006120126228779554, + 0.005583926569670439, + 0.007023044861853123, + 0.0068861027248203754, + 0.006589449476450682, + 0.002748708473518491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004095026291906834 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot, slowly zooming out, reveals a surreal scene: a massive rectangular dining table, crafted from weathered, dark oak. Its polished surface reflects the vast, cloudless sky above, revealing distinct wood grain and subtle signs of wear. Like a majestic ancient ship, it steadily sails across a boundless, sapphire-like sea. The table glides forward with a steady, deliberate pace, its prow-like edge gracefully slicing the nearly glassy surface, leaving a faint V-shaped wake in its wake before swiftly dissipating. Bathed in the bright, clear light of midday, the entire scene dances on the wooden surface and gently ripples, contrasting starkly with the deep blue waters stretching to the perfect horizon in all directions, creating a sense of ultimate tranquility, solitude, and a dreamlike journey.", + "seed": 17272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_8/index_6806_seed_17272_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12016696045175194, + "vlm_logprob": -0.12016696045175194, + "vlm_yes_no_margin": 3.349999988079071, + "vlm_token_credit": -0.12016696045175194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43028169870376587 + ], + "opposite_logprob_tokens": [ + -1.055281639099121 + ], + "logprob_reward": -0.43028169870376587, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.43028169870376587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1847752332687378, + 0.22774235904216766, + 0.04792863875627518, + 0.09063377231359482, + 0.024244798347353935, + 0.04384886845946312, + 0.09107296913862228, + 0.08884340524673462, + 0.15249522030353546, + 0.04841470345854759 + ], + "credit_map": [ + 0.034828491508960724, + 0.0200552586466074, + 0.014832858927547932, + 0.013913508504629135, + 0.01940293051302433, + 0.05653306096792221, + 0.013002745807170868, + 0.012206378392875195, + 0.03147858753800392, + 0.0314009003341198, + 0.013757938519120216, + 0.03101050667464733, + 0.03625165671110153, + 0.030848048627376556, + 0.02925049141049385, + 0.023744242265820503, + 0.00620741443708539, + 0.005948837846517563, + 0.003933762200176716, + 0.008259781636297703, + 0.006858813110738993, + 0.006259304005652666, + 0.006302081048488617, + 0.004158644936978817, + 0.007121057715266943, + 0.00703650526702404, + 0.005368806887418032, + 0.021549411118030548, + 0.008146033622324467, + 0.0060903411358594894, + 0.027883682399988174, + 0.007437932305037975, + 0.003899328876286745, + 0.0031821588054299355, + 0.0029934148769825697, + 0.003100466914474964, + 0.002828072290867567, + 0.003229196649044752, + 0.002556642983108759, + 0.0024555157870054245, + 0.004566614981740713, + 0.007700389716774225, + 0.003866911632940173, + 0.006341068539768457, + 0.007250186521559954, + 0.004855687264353037, + 0.005616037640720606, + 0.0036519733257591724, + 0.013185788877308369, + 0.01526318397372961, + 0.005910041742026806, + 0.006968188565224409, + 0.030257461592555046, + 0.008955491706728935, + 0.006672453135251999, + 0.0038603590801358223, + 0.01049906387925148, + 0.013831817544996738, + 0.010422623716294765, + 0.014974200166761875, + 0.014970581978559494, + 0.010106978937983513, + 0.009386475197970867, + 0.004651661496609449, + 0.012502572499215603, + 0.013232086785137653, + 0.010905250906944275, + 0.0379507802426815, + 0.022252945229411125, + 0.007565481588244438, + 0.037804849445819855, + 0.010281260125339031, + 0.005920934025198221, + 0.009395099245011806, + 0.003043355653062463, + 0.006660444661974907, + 0.010947021655738354, + 0.004477282054722309, + 0.004896074067801237, + 0.0030744916293770075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.43028169870376587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214104041457176 + ], + "opposite_logprob_tokens": [ + -4.264214038848877 + ], + "logprob_reward": -0.014214104041457176, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214104041457176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32481905817985535, + 0.15157297253608704, + 0.2311408966779709, + 0.04028139263391495, + 0.060428593307733536, + 0.04342349246144295, + 0.016670463606715202, + 0.014924576506018639, + 0.08710262924432755, + 0.029635969549417496 + ], + "credit_map": [ + 0.06352534145116806, + 0.04661601409316063, + 0.02285980060696602, + 0.019081156700849533, + 0.0270862877368927, + 0.10919833928346634, + 0.01564813032746315, + 0.020803989842534065, + 0.025104708969593048, + 0.02105049602687359, + 0.010169608518481255, + 0.012306625954806805, + 0.018353719264268875, + 0.041748397052288055, + 0.013487558811903, + 0.009351853281259537, + 0.03632152080535889, + 0.018723290413618088, + 0.0146369319409132, + 0.03276669606566429, + 0.025117894634604454, + 0.04069306328892708, + 0.044011782854795456, + 0.01886969804763794, + 0.0050249481573700905, + 0.00450581731274724, + 0.0023704813793301582, + 0.005867230240255594, + 0.005736603867262602, + 0.005198498256504536, + 0.007622285280376673, + 0.003955526277422905, + 0.009197358042001724, + 0.007667469792068005, + 0.003217784920707345, + 0.009879296645522118, + 0.010683629661798477, + 0.006351173855364323, + 0.010395503602921963, + 0.00303637539036572, + 0.006369598209857941, + 0.00783150177448988, + 0.002088394947350025, + 0.0025405294727534056, + 0.012428972870111465, + 0.007057617884129286, + 0.0027819029055535793, + 0.002324972301721573, + 0.0027609814424067736, + 0.0021993257105350494, + 0.0009832977084442973, + 0.0013935398310422897, + 0.00501621887087822, + 0.002328369067981839, + 0.0011907503940165043, + 0.0007979812799021602, + 0.001855812151916325, + 0.0023070580791682005, + 0.0014689108356833458, + 0.0024128337390720844, + 0.002582744462415576, + 0.001781131373718381, + 0.0016216338844969869, + 0.0008944512810558081, + 0.010698148980736732, + 0.010613960213959217, + 0.004690483212471008, + 0.012524301186203957, + 0.025332290679216385, + 0.006293746642768383, + 0.012536928988993168, + 0.004412774927914143, + 0.0034012969117611647, + 0.004659942351281643, + 0.0021117867436259985, + 0.0035516598727554083, + 0.005941271781921387, + 0.004065117333084345, + 0.003702207701280713, + 0.002202688716351986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014214104041457176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947748199105263 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947748199105263, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947748199105263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2467307299375534, + 0.11919128894805908, + 0.09738022089004517, + 0.08509606122970581, + 0.10138373076915741, + 0.06693228334188461, + 0.05674893781542778, + 0.04508338123559952, + 0.12111876904964447, + 0.06033458188176155 + ], + "credit_map": [ + 0.047584034502506256, + 0.030360911041498184, + 0.020649636164307594, + 0.017572008073329926, + 0.022119969129562378, + 0.07592837512493134, + 0.01691950485110283, + 0.015596271492540836, + 0.018663203343749046, + 0.017591651529073715, + 0.007022568956017494, + 0.01238084677606821, + 0.011764150112867355, + 0.03207430616021156, + 0.00890018418431282, + 0.010794383473694324, + 0.009465004317462444, + 0.009194063022732735, + 0.005291462875902653, + 0.020684467628598213, + 0.010193414054811, + 0.00702345697209239, + 0.028313368558883667, + 0.007214989047497511, + 0.013212443329393864, + 0.009941297583281994, + 0.003991106990724802, + 0.008371878415346146, + 0.009152283892035484, + 0.023545537143945694, + 0.010517068207263947, + 0.006364441476762295, + 0.014361441135406494, + 0.012661374174058437, + 0.007753714453428984, + 0.008246197365224361, + 0.011963872238993645, + 0.026450347155332565, + 0.009261666797101498, + 0.010685118846595287, + 0.010017859749495983, + 0.011414896696805954, + 0.0032877521589398384, + 0.004371825605630875, + 0.01811692863702774, + 0.012489553540945053, + 0.004091197624802589, + 0.003142269793897867, + 0.007317584473639727, + 0.008985312655568123, + 0.004390629008412361, + 0.00572851812466979, + 0.0144052654504776, + 0.006853253114968538, + 0.0060428413562476635, + 0.0030255320016294718, + 0.006950086914002895, + 0.006597473286092281, + 0.004972823429852724, + 0.005677226465195417, + 0.009457241743803024, + 0.004395423922687769, + 0.004082916304469109, + 0.002950188471004367, + 0.011535895057022572, + 0.011622481048107147, + 0.007971592247486115, + 0.024571849033236504, + 0.024123823270201683, + 0.007964321412146091, + 0.025683680549263954, + 0.007645132951438427, + 0.00882476195693016, + 0.009080017916858196, + 0.003242708742618561, + 0.004138092044740915, + 0.01781213842332363, + 0.0082470141351223, + 0.0057903858833014965, + 0.003199462778866291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005947748199105263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14276407659053802 + ], + "opposite_logprob_tokens": [ + -2.017764091491699 + ], + "logprob_reward": -0.14276407659053802, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14276407659053802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2790921926498413, + 0.1989220827817917, + 0.28500890731811523, + 0.06473588198423386, + 0.05099713057279587, + 0.01921248622238636, + 0.017061619088053703, + 0.013232341036200523, + 0.03454086184501648, + 0.0371965654194355 + ], + "credit_map": [ + 0.03535210341215134, + 0.0525694265961647, + 0.02859278954565525, + 0.03136489540338516, + 0.03610042482614517, + 0.047973956912755966, + 0.022197848185896873, + 0.024940745905041695, + 0.02323157526552677, + 0.03768923133611679, + 0.01749304123222828, + 0.026978274807333946, + 0.02954125590622425, + 0.020036280155181885, + 0.026396632194519043, + 0.017555775120854378, + 0.030804922804236412, + 0.029812250286340714, + 0.0248777586966753, + 0.046009864658117294, + 0.040100809186697006, + 0.029120491817593575, + 0.05906127020716667, + 0.025221532210707664, + 0.010074550285935402, + 0.006637308280915022, + 0.0048398771323263645, + 0.009839276783168316, + 0.006010160781443119, + 0.009586070664227009, + 0.011700589209794998, + 0.006048050243407488, + 0.007481415756046772, + 0.007036581169813871, + 0.0026748511008918285, + 0.007550571113824844, + 0.01018450129777193, + 0.004150110762566328, + 0.008695146068930626, + 0.0032239537686109543, + 0.0025456121657043695, + 0.0030925748869776726, + 0.001360792899504304, + 0.0013948031701147556, + 0.004784252494573593, + 0.003066113917157054, + 0.001621528179384768, + 0.0013468086253851652, + 0.0032447082921862602, + 0.0024361582472920418, + 0.0010404248023405671, + 0.0013528084382414818, + 0.0058345915749669075, + 0.001339090638794005, + 0.0010579578811302781, + 0.0007558789802715182, + 0.0017990581691265106, + 0.0014790687710046768, + 0.0007858305471017957, + 0.0020887679420411587, + 0.002716535236686468, + 0.0020623847376555204, + 0.0015012006042525172, + 0.0007994956686161458, + 0.00446682795882225, + 0.003318679053336382, + 0.003212342504411936, + 0.0060777487233281136, + 0.005481742322444916, + 0.002729984698817134, + 0.005957598332315683, + 0.0032959398813545704, + 0.005440150387585163, + 0.00523347407579422, + 0.0026017555501312017, + 0.0027702604420483112, + 0.01050938107073307, + 0.004438462667167187, + 0.004024911671876907, + 0.002178170718252659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.14276407659053802 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627174723893404 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627174723893404, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627174723893404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1258840709924698, + 0.14261175692081451, + 0.08072679489850998, + 0.10201361775398254, + 0.15476106107234955, + 0.028902508318424225, + 0.12719710171222687, + 0.07841285318136215, + 0.10533484071493149, + 0.054155364632606506 + ], + "credit_map": [ + 0.014792685396969318, + 0.021578779444098473, + 0.01674017496407032, + 0.013609287329018116, + 0.01767631806433201, + 0.015042847022414207, + 0.013884775340557098, + 0.012559198774397373, + 0.018136076629161835, + 0.022084437310695648, + 0.009846863336861134, + 0.019941333681344986, + 0.02384006977081299, + 0.015674475580453873, + 0.018862305209040642, + 0.014226194471120834, + 0.015615158714354038, + 0.007511190138757229, + 0.004668865352869034, + 0.008273748680949211, + 0.009590942412614822, + 0.019973205402493477, + 0.009699099697172642, + 0.00539458729326725, + 0.0076027619652450085, + 0.006549471523612738, + 0.0035390444099903107, + 0.0253301914781332, + 0.006933185737580061, + 0.01268795970827341, + 0.032394178211688995, + 0.00697683310136199, + 0.02132275700569153, + 0.01752740517258644, + 0.007658700924366713, + 0.02157367579638958, + 0.01706050895154476, + 0.0321841835975647, + 0.026216698810458183, + 0.011217125691473484, + 0.004044305998831987, + 0.004570291377604008, + 0.0016207420267164707, + 0.0027897770050913095, + 0.004212673753499985, + 0.007476224564015865, + 0.0023388282861560583, + 0.0018496671691536903, + 0.02388928271830082, + 0.016262752935290337, + 0.0075354729779064655, + 0.011950933374464512, + 0.028082042932510376, + 0.02348029986023903, + 0.009880030527710915, + 0.0061162905767560005, + 0.01619274728000164, + 0.010472040623426437, + 0.006462600082159042, + 0.0073608397506177425, + 0.0181332565844059, + 0.011221101507544518, + 0.004780896008014679, + 0.0037893683183938265, + 0.008824720978736877, + 0.009463902562856674, + 0.007564314175397158, + 0.024447644129395485, + 0.017330117523670197, + 0.007433961145579815, + 0.023688960820436478, + 0.006581215187907219, + 0.006471977569162846, + 0.007217048201709986, + 0.003530096961185336, + 0.0033301464281976223, + 0.016324855387210846, + 0.007863465696573257, + 0.005952924024313688, + 0.0034648492000997066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.007627174723893404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot, slowly zooming out, reveals a surreal scene: a massive rectangular dining table, crafted from weathered, dark oak. Its polished surface reflects the vast, cloudless sky above, revealing distinct wood grain and subtle signs of wear. Like a majestic ancient ship, it steadily sails across a boundless, sapphire-like sea. The table glides forward with a steady, deliberate pace, its prow-like edge gracefully slicing the nearly glassy surface, leaving a faint V-shaped wake in its wake before swiftly dissipating. Bathed in the bright, clear light of midday, the entire scene dances on the wooden surface and gently ripples, contrasting starkly with the deep blue waters stretching to the perfect horizon in all directions, creating a sense of ultimate tranquility, solitude, and a dreamlike journey.", + "seed": 17273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_8/index_6806_seed_17273_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13555179857648908, + "vlm_logprob": -0.13555179857648908, + "vlm_yes_no_margin": 3.550000011920929, + "vlm_token_credit": -0.13555179857648908, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22813433408737183 + ], + "opposite_logprob_tokens": [ + -1.6031343936920166 + ], + "logprob_reward": -0.22813433408737183, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22813433408737183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15102003514766693, + 0.06589510291814804, + 0.1604224145412445, + 0.038909878581762314, + 0.089351586997509, + 0.09662813693284988, + 0.1766178458929062, + 0.11596320569515228, + 0.041774775832891464, + 0.06341700255870819 + ], + "credit_map": [ + 0.01557538378983736, + 0.01448044553399086, + 0.014836163260042667, + 0.017637914046645164, + 0.02527747116982937, + 0.0218893364071846, + 0.027863172814249992, + 0.013460138812661171, + 0.007150521036237478, + 0.0066628544591367245, + 0.007409243378788233, + 0.00842210277915001, + 0.0067519862204790115, + 0.00918679591268301, + 0.01191251166164875, + 0.00839909166097641, + 0.024633198976516724, + 0.020348886027932167, + 0.01380150020122528, + 0.0219744760543108, + 0.01795073226094246, + 0.025557970628142357, + 0.018462853506207466, + 0.017692802473902702, + 0.0039408267475664616, + 0.005982763133943081, + 0.003747570561245084, + 0.00537968659773469, + 0.005353694315999746, + 0.005494238343089819, + 0.004778515547513962, + 0.004232584964483976, + 0.007588578853756189, + 0.011062132194638252, + 0.01141894981265068, + 0.012156815268099308, + 0.013800415210425854, + 0.012173696421086788, + 0.014248187653720379, + 0.006902806926518679, + 0.013300688937306404, + 0.012600676156580448, + 0.00917135551571846, + 0.01092254463583231, + 0.017427876591682434, + 0.014849930070340633, + 0.010894724167883396, + 0.007460340391844511, + 0.03229544311761856, + 0.018828153610229492, + 0.01622508279979229, + 0.021195903420448303, + 0.02275518886744976, + 0.027030954137444496, + 0.02397419698536396, + 0.01431291364133358, + 0.022244565188884735, + 0.01589359901845455, + 0.010987363755702972, + 0.01229746825993061, + 0.013229500502347946, + 0.01622091792523861, + 0.01480675395578146, + 0.0102830296382308, + 0.006974960211664438, + 0.006716838572174311, + 0.004374669399112463, + 0.003968256525695324, + 0.005360888782888651, + 0.005744153168052435, + 0.004153577145189047, + 0.0044814287684857845, + 0.006128438748419285, + 0.00797530822455883, + 0.005023738835006952, + 0.0061629475094377995, + 0.014199387282133102, + 0.010976896621286869, + 0.00753497751429677, + 0.005415311083197594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.22813433408737183 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420411467552185 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.01420411467552185, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420411467552185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0652993768453598, + 0.0443284772336483, + 0.08280496299266815, + 0.1738821417093277, + 0.09119304269552231, + 0.07384934276342392, + 0.033470261842012405, + 0.020093102008104324, + 0.04378892108798027, + 0.37129032611846924 + ], + "credit_map": [ + 0.009418312460184097, + 0.00895565003156662, + 0.008508685044944286, + 0.004680562764406204, + 0.007909484207630157, + 0.011281438171863556, + 0.007187090814113617, + 0.0073581598699092865, + 0.00563851511105895, + 0.006257406435906887, + 0.0046937717124819756, + 0.0055818455293774605, + 0.006185621954500675, + 0.004671713337302208, + 0.006449700333178043, + 0.004849908407777548, + 0.014320294372737408, + 0.011662201024591923, + 0.00682859867811203, + 0.009557113982737064, + 0.008778471499681473, + 0.014108022674918175, + 0.009177098982036114, + 0.00837316457182169, + 0.027101004496216774, + 0.01903267204761505, + 0.011544703505933285, + 0.02207517996430397, + 0.01779528521001339, + 0.024127088487148285, + 0.028916871175169945, + 0.023289348930120468, + 0.01654634065926075, + 0.010064831003546715, + 0.008396349847316742, + 0.011484780348837376, + 0.011895678006112576, + 0.012011932209134102, + 0.011372951790690422, + 0.009420184418559074, + 0.012739242985844612, + 0.008847616612911224, + 0.006131685338914394, + 0.00928837712854147, + 0.008261196315288544, + 0.010796859860420227, + 0.007598354946821928, + 0.010186023078858852, + 0.00561299966648221, + 0.003587661776691675, + 0.0035155608784407377, + 0.003391945967450738, + 0.005432207603007555, + 0.004853771533817053, + 0.004271560814231634, + 0.002804558025673032, + 0.003434301819652319, + 0.002551454585045576, + 0.0017752847634255886, + 0.002013931982219219, + 0.003178061917424202, + 0.002882130676880479, + 0.0024723231326788664, + 0.0017856151098385453, + 0.006664820481091738, + 0.0050992718897759914, + 0.005245694890618324, + 0.004593415651470423, + 0.0045363097451627254, + 0.005267010536044836, + 0.00578631367534399, + 0.006596091203391552, + 0.06006937101483345, + 0.03873132914304733, + 0.032836467027664185, + 0.039292603731155396, + 0.06002453342080116, + 0.05253155529499054, + 0.04324522241950035, + 0.04455926641821861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.01420411467552185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090751986950636 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.004090751986950636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090751986950636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22577127814292908, + 0.078864187002182, + 0.06367619335651398, + 0.10271178930997849, + 0.08271915465593338, + 0.06504987180233002, + 0.05905040726065636, + 0.046396560966968536, + 0.017039882019162178, + 0.2587205767631531 + ], + "credit_map": [ + 0.04238591715693474, + 0.028553225100040436, + 0.021323656663298607, + 0.019947027787566185, + 0.03364311903715134, + 0.0357041098177433, + 0.022714763879776, + 0.02149948477745056, + 0.014556129463016987, + 0.010322843678295612, + 0.007958547212183475, + 0.006792300380766392, + 0.008313339203596115, + 0.011511478573083878, + 0.008980563841760159, + 0.010428992100059986, + 0.007823530584573746, + 0.006213005166500807, + 0.0053064413368701935, + 0.00877834390848875, + 0.00639486825093627, + 0.008573963306844234, + 0.009837158024311066, + 0.010748893022537231, + 0.017039205878973007, + 0.01077860314399004, + 0.006835080683231354, + 0.011923150159418583, + 0.010242028161883354, + 0.016322920098900795, + 0.016512107104063034, + 0.013058705255389214, + 0.013828730210661888, + 0.009298616088926792, + 0.005983519833534956, + 0.008739442564547062, + 0.012766974978148937, + 0.01408375520259142, + 0.009645534679293633, + 0.008372584357857704, + 0.009150932542979717, + 0.00988681335002184, + 0.005282416008412838, + 0.005189362447708845, + 0.011253113858401775, + 0.012089289724826813, + 0.005825827829539776, + 0.0063721234910190105, + 0.011937122792005539, + 0.0067880223505198956, + 0.005319482181221247, + 0.006317393388599157, + 0.007721424102783203, + 0.008637386374175549, + 0.006645819637924433, + 0.005683764349669218, + 0.008067416958510876, + 0.005832258146256208, + 0.004532294813543558, + 0.004770167637616396, + 0.00558365136384964, + 0.006697038654237986, + 0.006073468364775181, + 0.004840270150452852, + 0.0035422523505985737, + 0.001978454180061817, + 0.0013134534237906337, + 0.001410843338817358, + 0.002451523207128048, + 0.00285839824937284, + 0.002126622013747692, + 0.001358335604891181, + 0.04600298032164574, + 0.027293702587485313, + 0.02153851091861725, + 0.02578187920153141, + 0.04080664739012718, + 0.037516169250011444, + 0.02864266000688076, + 0.031138041988015175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.004090751986950636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288432002067566 + ], + "opposite_logprob_tokens": [ + -1.0538432598114014 + ], + "logprob_reward": -0.4288432002067566, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4288432002067566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.070253387093544, + 0.0395636186003685, + 0.13159121572971344, + 0.15531648695468903, + 0.09802503138780594, + 0.20074641704559326, + 0.037863362580537796, + 0.011869278736412525, + 0.12572328746318817, + 0.1290479153394699 + ], + "credit_map": [ + 0.013368499465286732, + 0.008685946464538574, + 0.008505593053996563, + 0.006068676244467497, + 0.008672554977238178, + 0.009305967949330807, + 0.00931970402598381, + 0.006326444912701845, + 0.00605732062831521, + 0.004520190414041281, + 0.0039641959592700005, + 0.0043371086940169334, + 0.00412381999194622, + 0.005695700645446777, + 0.006649460177868605, + 0.004215824417769909, + 0.027966775000095367, + 0.012487796135246754, + 0.009401609189808369, + 0.012644208036363125, + 0.011947045102715492, + 0.023739945143461227, + 0.01299364399164915, + 0.020410196855664253, + 0.01621716096997261, + 0.018448298797011375, + 0.0124058211222291, + 0.024200839921832085, + 0.018530333414673805, + 0.02147449180483818, + 0.02754649519920349, + 0.0164930559694767, + 0.016243133693933487, + 0.010126519948244095, + 0.009295527823269367, + 0.014006922021508217, + 0.010524305514991283, + 0.0140214953571558, + 0.013634586706757545, + 0.010172545909881592, + 0.038224250078201294, + 0.024717871099710464, + 0.015261451713740826, + 0.021472785621881485, + 0.02520734630525112, + 0.03076876699924469, + 0.019194504246115685, + 0.025899454951286316, + 0.006648164242506027, + 0.00555081432685256, + 0.003710381453856826, + 0.0033473975490778685, + 0.005619169678539038, + 0.005628268234431744, + 0.0037332759238779545, + 0.0036258972249925137, + 0.0017843132372945547, + 0.0015170754631981254, + 0.0012643018271774054, + 0.0012317872606217861, + 0.0017311625415459275, + 0.0016664686845615506, + 0.001529611530713737, + 0.0011445586569607258, + 0.024194130674004555, + 0.021702460944652557, + 0.008416003547608852, + 0.011352283880114555, + 0.017420796677470207, + 0.019245142117142677, + 0.010719308629631996, + 0.01267317682504654, + 0.016903366893529892, + 0.014161431230604649, + 0.013501444831490517, + 0.012485023587942123, + 0.022615889087319374, + 0.019288908690214157, + 0.01527081336826086, + 0.014821060001850128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.4288432002067566 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248659192584455 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.00248659192584455, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248659192584455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26696622371673584, + 0.08608946949243546, + 0.0801268145442009, + 0.1498543918132782, + 0.05093250423669815, + 0.0884191170334816, + 0.08674640953540802, + 0.06888917833566666, + 0.03417528048157692, + 0.08780063688755035 + ], + "credit_map": [ + 0.048780832439661026, + 0.03421100229024887, + 0.03418497368693352, + 0.020853042602539062, + 0.03159266710281372, + 0.03949500247836113, + 0.030238239094614983, + 0.02761044166982174, + 0.015956535935401917, + 0.008318600244820118, + 0.0077991425059735775, + 0.006947345566004515, + 0.009661016054451466, + 0.014875398948788643, + 0.009367916733026505, + 0.013163517229259014, + 0.0077803232707083225, + 0.008981382474303246, + 0.008262218907475471, + 0.012530273757874966, + 0.009682969190180302, + 0.010043689049780369, + 0.012304606847465038, + 0.010541355237364769, + 0.026062915101647377, + 0.01587754301726818, + 0.009943215176463127, + 0.017991289496421814, + 0.014594263397157192, + 0.021124349907040596, + 0.02450968325138092, + 0.019751137122511864, + 0.00544534157961607, + 0.005432714242488146, + 0.005125794559717178, + 0.007877134718000889, + 0.006645160261541605, + 0.0069337342865765095, + 0.008961125276982784, + 0.0045114969834685326, + 0.015749668702483177, + 0.01244952343404293, + 0.007302966900169849, + 0.005714294034987688, + 0.014793884009122849, + 0.01510754693299532, + 0.007407650351524353, + 0.009893578477203846, + 0.016645465046167374, + 0.009629182517528534, + 0.007931393571197987, + 0.009169369004666805, + 0.011159948073327541, + 0.013503441587090492, + 0.01011647004634142, + 0.008591133169829845, + 0.012423466891050339, + 0.00903473049402237, + 0.006578186526894569, + 0.007200304418802261, + 0.008415712043642998, + 0.009743827395141125, + 0.008623959496617317, + 0.00686899246647954, + 0.006955334916710854, + 0.005330208223313093, + 0.002636551158502698, + 0.0032542983535677195, + 0.003936317749321461, + 0.0049505215138196945, + 0.0025833353865891695, + 0.004528712946921587, + 0.012688844464719296, + 0.008834291249513626, + 0.007075755391269922, + 0.009373082779347897, + 0.014534160494804382, + 0.013541667722165585, + 0.00997543428093195, + 0.011777402833104134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.00248659192584455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot, slowly zooming out, reveals a surreal scene: a massive rectangular dining table, crafted from weathered, dark oak. Its polished surface reflects the vast, cloudless sky above, revealing distinct wood grain and subtle signs of wear. Like a majestic ancient ship, it steadily sails across a boundless, sapphire-like sea. The table glides forward with a steady, deliberate pace, its prow-like edge gracefully slicing the nearly glassy surface, leaving a faint V-shaped wake in its wake before swiftly dissipating. Bathed in the bright, clear light of midday, the entire scene dances on the wooden surface and gently ripples, contrasting starkly with the deep blue waters stretching to the perfect horizon in all directions, creating a sense of ultimate tranquility, solitude, and a dreamlike journey.", + "seed": 17274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_8/index_6806_seed_17274_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07215890237130225, + "vlm_logprob": -0.07215890237130225, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.07215890237130225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31453636288642883 + ], + "opposite_logprob_tokens": [ + -1.3145363330841064 + ], + "logprob_reward": -0.31453636288642883, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31453636288642883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11587226390838623, + 0.031703874468803406, + 0.03272650018334389, + 0.12814076244831085, + 0.3722885251045227, + 0.10351055860519409, + 0.09670334309339523, + 0.01682809740304947, + 0.05535326898097992, + 0.04687277227640152 + ], + "credit_map": [ + 0.019720442593097687, + 0.008672663010656834, + 0.04047025367617607, + 0.015599232167005539, + 0.007476212922483683, + 0.005416501313447952, + 0.005644531454890966, + 0.012872427701950073, + 0.0030192492995411158, + 0.007953675463795662, + 0.0020347845274955034, + 0.0030504867900162935, + 0.0075082783587276936, + 0.001979311928153038, + 0.004387537948787212, + 0.0017705501522868872, + 0.0018878643168136477, + 0.004464874044060707, + 0.0022466799709945917, + 0.005948491860181093, + 0.004429602064192295, + 0.0018039498245343566, + 0.009133625775575638, + 0.0028114118613302708, + 0.006814316380769014, + 0.016767514869570732, + 0.031378187239170074, + 0.017964089289307594, + 0.031803421676158905, + 0.00701961200684309, + 0.00913565419614315, + 0.007257963065057993, + 0.011556820012629032, + 0.015066521242260933, + 0.038137275725603104, + 0.15698669850826263, + 0.01005766075104475, + 0.009946867823600769, + 0.02445509098470211, + 0.10608159005641937, + 0.004799515474587679, + 0.004072058014571667, + 0.0258448775857687, + 0.04162220656871796, + 0.0030334454495459795, + 0.002290119882673025, + 0.01559398602694273, + 0.006254349369555712, + 0.0035382925998419523, + 0.0030754348263144493, + 0.011664868332445621, + 0.008415407501161098, + 0.003348772646859288, + 0.00394937489181757, + 0.026200078427791595, + 0.036511119455099106, + 0.0014173610834404826, + 0.0018132554832845926, + 0.002997669856995344, + 0.0028141115326434374, + 0.0014239216689020395, + 0.0014659062726423144, + 0.0031206572894006968, + 0.0017752139829099178, + 0.0027790351305156946, + 0.002678438788279891, + 0.0032583943102508783, + 0.014130077324807644, + 0.004299392458051443, + 0.001415356295183301, + 0.021442245692014694, + 0.005350329913198948, + 0.004041561856865883, + 0.008637603372335434, + 0.002756782341748476, + 0.0026424492243677378, + 0.017600826919078827, + 0.004809324163943529, + 0.004197029396891594, + 0.0021871975623071194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.31453636288642883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086884886026382 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.011086884886026382, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086884886026382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24009495973587036, + 0.01566096395254135, + 0.008673394098877907, + 0.14675487577915192, + 0.2744544744491577, + 0.03845915198326111, + 0.16646146774291992, + 0.019000371918082237, + 0.07169321924448013, + 0.018747132271528244 + ], + "credit_map": [ + 0.02780642732977867, + 0.011804129928350449, + 0.09899882227182388, + 0.040144845843315125, + 0.012764519080519676, + 0.008890846744179726, + 0.007822250947356224, + 0.03186310455203056, + 0.0017121961573138833, + 0.0031424551270902157, + 0.00144151970744133, + 0.0020759112667292356, + 0.0030318191275000572, + 0.0010733020026236773, + 0.0016609405865892768, + 0.0015228183474391699, + 0.0006703289691358805, + 0.0010437966557219625, + 0.0007955879555083811, + 0.0014759034384042025, + 0.0011225987691432238, + 0.0005064653814770281, + 0.002325710142031312, + 0.0007330013904720545, + 0.006049842573702335, + 0.014036843553185463, + 0.049476537853479385, + 0.027601925656199455, + 0.028180841356515884, + 0.0057405815459787846, + 0.009214814752340317, + 0.006453467067331076, + 0.008518815971910954, + 0.011555511504411697, + 0.028972184285521507, + 0.11574387550354004, + 0.007402853108942509, + 0.00683217216283083, + 0.01723349280655384, + 0.07819552719593048, + 0.0020953495986759663, + 0.0027623276691883802, + 0.01066395454108715, + 0.013137525878846645, + 0.001865042606368661, + 0.001359263900667429, + 0.0042564054019749165, + 0.0023192772641777992, + 0.004356205929070711, + 0.0032481434755027294, + 0.026536138728260994, + 0.015589078888297081, + 0.004794050939381123, + 0.005328444764018059, + 0.04564597085118294, + 0.06096343696117401, + 0.0015867387410253286, + 0.001992381177842617, + 0.0019923585932701826, + 0.002629987895488739, + 0.0025604881811887026, + 0.0020846796687692404, + 0.0035307584330439568, + 0.002622978063300252, + 0.003731650998815894, + 0.004141809418797493, + 0.004374196752905846, + 0.018204515799880028, + 0.005316355265676975, + 0.002047010464593768, + 0.027229977771639824, + 0.006647700909525156, + 0.0013638560194522142, + 0.002156211296096444, + 0.0036831467878073454, + 0.003252410562708974, + 0.0018354068743065, + 0.0010825269855558872, + 0.0029429104179143906, + 0.002430662279948592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.011086884886026382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191023599356413 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.003191023599356413, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191023599356413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3533022403717041, + 0.021098855882883072, + 0.030387787148356438, + 0.20889237523078918, + 0.06953205168247223, + 0.04815955087542534, + 0.07250290364027023, + 0.022422993555665016, + 0.1328207105398178, + 0.040880460292100906 + ], + "credit_map": [ + 0.04231730476021767, + 0.016539903357625008, + 0.1465010941028595, + 0.05910784378647804, + 0.016772428527474403, + 0.013274861499667168, + 0.01197846233844757, + 0.04681038483977318, + 0.001968629425391555, + 0.00472624646499753, + 0.0015172576531767845, + 0.002408271422609687, + 0.004579242318868637, + 0.0015990490792319179, + 0.002585848094895482, + 0.001714313868433237, + 0.004634475335478783, + 0.003360585542395711, + 0.00274976110085845, + 0.0051948679611086845, + 0.0030143032781779766, + 0.002259331289678812, + 0.006360096391290426, + 0.002814370207488537, + 0.009142598137259483, + 0.02125256136059761, + 0.06925918906927109, + 0.03846669942140579, + 0.040328167378902435, + 0.008244876749813557, + 0.012618724256753922, + 0.009579586796462536, + 0.00280550355091691, + 0.0038708040956407785, + 0.006252498831599951, + 0.024506734684109688, + 0.004032614175230265, + 0.003476229263469577, + 0.005552925635129213, + 0.019034750759601593, + 0.0028621985111385584, + 0.004380031488835812, + 0.010252960957586765, + 0.014383099041879177, + 0.00473710848018527, + 0.0024851614143699408, + 0.006038090214133263, + 0.003020906588062644, + 0.0033914113882929087, + 0.0027431794442236423, + 0.013385784812271595, + 0.007586343679577112, + 0.0029339042957872152, + 0.003148687770590186, + 0.016962412744760513, + 0.022351186722517014, + 0.0018375676590949297, + 0.0018923155730590224, + 0.002691615605726838, + 0.003712415462359786, + 0.0020802710205316544, + 0.0020030781161040068, + 0.004646044224500656, + 0.0035596885718405247, + 0.007378565147519112, + 0.008893237449228764, + 0.007804966531693935, + 0.031969983130693436, + 0.011586010456085205, + 0.004350986331701279, + 0.04865419492125511, + 0.012182765640318394, + 0.005399941932410002, + 0.0034726576413959265, + 0.0032735620625317097, + 0.0035978069063276052, + 0.009158414788544178, + 0.0032866005785763264, + 0.005276178475469351, + 0.00741530442610383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.003191023599356413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978455275297165 + ], + "opposite_logprob_tokens": [ + -3.5297844409942627 + ], + "logprob_reward": -0.02978455275297165, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978455275297165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24157854914665222, + 0.00872073881328106, + 0.007694370578974485, + 0.14946703612804413, + 0.14919185638427734, + 0.13016360998153687, + 0.25273391604423523, + 0.03543974831700325, + 0.021575508639216423, + 0.003434540703892708 + ], + "credit_map": [ + 0.029490388929843903, + 0.012359402142465115, + 0.09786331653594971, + 0.04158991575241089, + 0.013300146907567978, + 0.008410359732806683, + 0.008280616253614426, + 0.030284428969025612, + 0.0007411225815303624, + 0.0018005212768912315, + 0.0008017694926820695, + 0.0010575555497780442, + 0.001981620443984866, + 0.0006196665926836431, + 0.001041663927026093, + 0.000676819880027324, + 0.0006386398454196751, + 0.0010042940266430378, + 0.0006386993336491287, + 0.001775003271177411, + 0.0010957354679703712, + 0.00042688666144385934, + 0.0015932200476527214, + 0.0005218932637944818, + 0.0061578163877129555, + 0.01614724099636078, + 0.04774303734302521, + 0.0267203189432621, + 0.029618732631206512, + 0.005971797741949558, + 0.01027381606400013, + 0.006834301631897688, + 0.004810523707419634, + 0.006601189728826284, + 0.016011372208595276, + 0.06258229911327362, + 0.004849109798669815, + 0.0037714382633566856, + 0.010382380336523056, + 0.040183551609516144, + 0.005916818045079708, + 0.007014929316937923, + 0.02873110957443714, + 0.05168516933917999, + 0.004773829597979784, + 0.003786781569942832, + 0.021037138998508453, + 0.0072178468108177185, + 0.0067136227153241634, + 0.004930126015096903, + 0.038949452340602875, + 0.023537468165159225, + 0.008038565516471863, + 0.008144155144691467, + 0.07038865983486176, + 0.09203188866376877, + 0.00277029094286263, + 0.0037924766074866056, + 0.003318994538858533, + 0.005222964100539684, + 0.005126406904309988, + 0.0034338708501309156, + 0.006565173156559467, + 0.005209573544561863, + 0.0021122563630342484, + 0.002577604027464986, + 0.0018442493164911866, + 0.0025353017263114452, + 0.0044998000375926495, + 0.0023764469660818577, + 0.003060854272916913, + 0.002568997675552964, + 0.00016026775119826198, + 0.0006288596778176725, + 0.0003481002931948751, + 0.00029877701308578253, + 0.000982722151093185, + 0.0002960177371278405, + 0.0002360912476433441, + 0.0004837050219066441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.02978455275297165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive rectangular dining table made of weathered dark oak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table moving forward like a ship, leaving a faint V-shaped wake in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a boundless sapphire-like sea with a clear horizon under a cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a sweeping shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, clear midday light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021956877317279577 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.0021956877317279577, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021956877317279577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21393699944019318, + 0.027828026562929153, + 0.05241085961461067, + 0.2168348729610443, + 0.06005621701478958, + 0.14630021154880524, + 0.10973015427589417, + 0.027004778385162354, + 0.04504737630486488, + 0.10085051506757736 + ], + "credit_map": [ + 0.032498497515916824, + 0.01329787727445364, + 0.07739675790071487, + 0.03649015352129936, + 0.0140951257199049, + 0.007940738461911678, + 0.008955424651503563, + 0.02326241508126259, + 0.002898453501984477, + 0.008310300298035145, + 0.001630355603992939, + 0.002857262035831809, + 0.006018753629177809, + 0.0014771296409890056, + 0.002519384492188692, + 0.0021163872443139553, + 0.0053905886597931385, + 0.006335054524242878, + 0.00461245235055685, + 0.009208731353282928, + 0.0062793525867164135, + 0.00286918506026268, + 0.013490726239979267, + 0.004224766511470079, + 0.010256980545818806, + 0.020412800833582878, + 0.07222824543714523, + 0.039732638746500015, + 0.040651001036167145, + 0.009124656207859516, + 0.013942226767539978, + 0.010486334562301636, + 0.005445148330181837, + 0.0054602413438260555, + 0.00545844528824091, + 0.01568196713924408, + 0.008030630648136139, + 0.004212367348372936, + 0.008167831227183342, + 0.007599583826959133, + 0.007058051880449057, + 0.007424049079418182, + 0.029999865218997, + 0.058558136224746704, + 0.006242751143872738, + 0.004098991863429546, + 0.024391641840338707, + 0.008526735007762909, + 0.0030432504136115313, + 0.003904913319274783, + 0.013164816424250603, + 0.009709788486361504, + 0.004133948590606451, + 0.005603903438895941, + 0.031362134963274, + 0.038807403296232224, + 0.002802183385938406, + 0.002989869099110365, + 0.0031138507183641195, + 0.0035763296764343977, + 0.0025202776305377483, + 0.0037721600383520126, + 0.004738486837595701, + 0.0034916196018457413, + 0.0034751666244119406, + 0.0037380261346697807, + 0.002756088273599744, + 0.009815339930355549, + 0.004105579573661089, + 0.004082425497472286, + 0.013327807188034058, + 0.003746944712474942, + 0.014219313859939575, + 0.007424655836075544, + 0.010310187004506588, + 0.015415803529322147, + 0.013643218204379082, + 0.007293135393410921, + 0.014183733612298965, + 0.01836046576499939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0021956877317279577 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_9/index_tensor([7770, 7770])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_9/index_tensor([7770, 7770])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b670b5a5db35a2d0fb42417232cfeab62acac1c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_9/index_tensor([7770, 7770])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an unbroken single shot, the camera begins in a mid-level shot, capturing a delicate, small Asian rattan cart. Crafted entirely from hand-woven, light honey-colored rattan strips, the cart displays the intricate weave and the warm, natural luster of the material. Its four small, solid wooden wheels rest against the polished, light-colored wooden floor. As the cart gracefully spins like a top, the camera orbits it in a slow, smooth, clockwise orbit, perfectly synchronizing its movements. The cart pivots on its central axis, steadily and deliberately rotating, its fluid motion carving graceful arcs across the frame. The scene is bathed in soft afternoon sunlight streaming through a large side window. The light, filtered by the gauzy white curtains, is incredibly soft, casting subtle, almost edgeless shadows on the floor. The background is a minimalist interior, scarcely furnished save for a vague pot of greenery in the far corner. The main colors of the picture are cream white, light brown of wood and warm yellow of rattan, creating an extremely tranquil, peaceful and healing atmosphere. The overall visual effect is like a dynamic, airy Japanese healing illustration.", + "seed": 6211, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_9/index_7770_seed_6211_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019348927959799766, + "vlm_logprob": -0.019348927959799766, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.019348927959799766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027356883510947227 + ], + "opposite_logprob_tokens": [ + -3.7773568630218506 + ], + "logprob_reward": -0.027356883510947227, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027356883510947227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1508924514055252, + 0.055955298244953156, + 0.052329666912555695, + 0.061542000621557236, + 0.08940956741571426, + 0.07742834836244583, + 0.06173279136419296, + 0.20922626554965973, + 0.14400054514408112, + 0.09748300909996033 + ], + "credit_map": [ + 0.017350368201732635, + 0.019635001197457314, + 0.0198985505849123, + 0.018429899588227272, + 0.0230740737169981, + 0.015434619039297104, + 0.015912916511297226, + 0.021157026290893555, + 0.010097898542881012, + 0.005063023418188095, + 0.006295970641076565, + 0.005766438785940409, + 0.005966105964034796, + 0.011432379484176636, + 0.0050518689677119255, + 0.00628161383792758, + 0.003154834732413292, + 0.006463579833507538, + 0.009655525907874107, + 0.006660400424152613, + 0.0059343501925468445, + 0.005303063429892063, + 0.004557504318654537, + 0.010600406676530838, + 0.004739339929074049, + 0.007031461223959923, + 0.01050544623285532, + 0.0077309985645115376, + 0.007445128168910742, + 0.006580607034265995, + 0.007800062652677298, + 0.009708954952657223, + 0.0038011125288903713, + 0.014483378268778324, + 0.010802742093801498, + 0.006329402793198824, + 0.028398403897881508, + 0.010096274316310883, + 0.008220295421779156, + 0.0072779543697834015, + 0.0076168919913470745, + 0.008187085390090942, + 0.009721511974930763, + 0.010556579567492008, + 0.01602303422987461, + 0.007274535018950701, + 0.009821043349802494, + 0.00822766125202179, + 0.011434114538133144, + 0.005921560805290937, + 0.008689994923770428, + 0.006230777129530907, + 0.005647714715451002, + 0.007854845374822617, + 0.0076454863883554935, + 0.008308305405080318, + 0.02520471066236496, + 0.01964857243001461, + 0.02797786518931389, + 0.021050894632935524, + 0.03566041216254234, + 0.02580154873430729, + 0.03721807524561882, + 0.01666419394314289, + 0.01272881031036377, + 0.020795349031686783, + 0.016139496117830276, + 0.015842363238334656, + 0.024001266807317734, + 0.0180649533867836, + 0.022536493837833405, + 0.013891808688640594, + 0.0061651659198105335, + 0.012558544054627419, + 0.01554113533347845, + 0.01454539317637682, + 0.012410974130034447, + 0.011691803112626076, + 0.01292247325181961, + 0.011647517792880535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.027356883510947227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001189596951007843 + ], + "opposite_logprob_tokens": [ + -6.751189708709717 + ], + "logprob_reward": -0.001189596951007843, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001189596951007843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07798243314027786, + 0.02336832694709301, + 0.026842091232538223, + 0.0261659137904644, + 0.12131273746490479, + 0.16728754341602325, + 0.04177653789520264, + 0.2142036408185959, + 0.24226105213165283, + 0.05879971385002136 + ], + "credit_map": [ + 0.006211359519511461, + 0.01037731021642685, + 0.008821751922369003, + 0.011999676004052162, + 0.013275626115500927, + 0.008261775597929955, + 0.008049425669014454, + 0.010985509492456913, + 0.0021096477285027504, + 0.003441175213083625, + 0.0031142323277890682, + 0.0019528494449332356, + 0.0031828235369175673, + 0.003168331226333976, + 0.0029710887465626, + 0.003428178606554866, + 0.0027486199978739023, + 0.003668301273137331, + 0.003908874932676554, + 0.003063684096559882, + 0.003323429496958852, + 0.003523255931213498, + 0.0026829105336219072, + 0.00392301706597209, + 0.0019931814167648554, + 0.0029484417755156755, + 0.003705053124576807, + 0.0024713671300560236, + 0.0037976603489369154, + 0.005232839845120907, + 0.0027741240337491035, + 0.0032432458829134703, + 0.007030392996966839, + 0.019334515556693077, + 0.01722792722284794, + 0.004844199400395155, + 0.0398455448448658, + 0.01364304032176733, + 0.010125440545380116, + 0.009261672385036945, + 0.01683260314166546, + 0.018907325342297554, + 0.027141332626342773, + 0.016217201948165894, + 0.027630863711237907, + 0.029658203944563866, + 0.017287183552980423, + 0.013612820766866207, + 0.005849892273545265, + 0.004417878109961748, + 0.0063641928136348724, + 0.004918436054140329, + 0.004449231084436178, + 0.004667420871555805, + 0.007474612444639206, + 0.003634875174611807, + 0.032170116901397705, + 0.01931094005703926, + 0.038063425570726395, + 0.018464423716068268, + 0.023640284314751625, + 0.0337311290204525, + 0.023436488583683968, + 0.025386828929185867, + 0.028090961277484894, + 0.03822864219546318, + 0.03240383416414261, + 0.022487269714474678, + 0.05040539801120758, + 0.020216697826981544, + 0.03195350617170334, + 0.018474750220775604, + 0.004186459816992283, + 0.0063830045983195305, + 0.009835784323513508, + 0.008057938888669014, + 0.005901344586163759, + 0.008256428875029087, + 0.009403560310602188, + 0.006775189191102982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.001189596951007843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06970357894897461, + 0.03735533356666565, + 0.057631488889455795, + 0.064652219414711, + 0.3454681932926178, + 0.04151598736643791, + 0.03746342658996582, + 0.26407214999198914, + 0.04430375248193741, + 0.03783388063311577 + ], + "credit_map": [ + 0.006049147341400385, + 0.01053164154291153, + 0.006712133996188641, + 0.007767985109239817, + 0.014938505366444588, + 0.009080971591174603, + 0.005986632313579321, + 0.00863655749708414, + 0.006336430087685585, + 0.0033307154662907124, + 0.004175832029432058, + 0.0035149632021784782, + 0.003309040330350399, + 0.007194744423031807, + 0.005025907885283232, + 0.004467699211090803, + 0.0033341082744300365, + 0.003671354381367564, + 0.01809922233223915, + 0.004086528904736042, + 0.0035833215806633234, + 0.002852741628885269, + 0.00323706422932446, + 0.01876714825630188, + 0.005171087570488453, + 0.007623116951435804, + 0.00949940551072359, + 0.004495925735682249, + 0.006241935305297375, + 0.01817234419286251, + 0.005203630309551954, + 0.008244771510362625, + 0.024569431319832802, + 0.04656248912215233, + 0.051114749163389206, + 0.014701992273330688, + 0.10835051536560059, + 0.03994442895054817, + 0.030712222680449486, + 0.029512332752346992, + 0.0044429851695895195, + 0.004507525358349085, + 0.00616772286593914, + 0.004359920509159565, + 0.006472649984061718, + 0.007871059700846672, + 0.004721852485090494, + 0.002972274087369442, + 0.00712232943624258, + 0.004600513726472855, + 0.004144606180489063, + 0.004793831612914801, + 0.004375071730464697, + 0.0036330739967525005, + 0.0059760198928415775, + 0.00281797768548131, + 0.048046864569187164, + 0.022989094257354736, + 0.058184318244457245, + 0.021082483232021332, + 0.021603938192129135, + 0.037341341376304626, + 0.017240362241864204, + 0.03758372738957405, + 0.004421323537826538, + 0.007121725473552942, + 0.006752961315214634, + 0.003609713399782777, + 0.009519523940980434, + 0.005004234611988068, + 0.004720565862953663, + 0.0031537008471786976, + 0.003164358902722597, + 0.0038045349065214396, + 0.006037186365574598, + 0.0050503844395279884, + 0.004102157428860664, + 0.00584222748875618, + 0.005616661626845598, + 0.004216370638459921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.008636740036308765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054930150508880615 + ], + "opposite_logprob_tokens": [ + -2.9299302101135254 + ], + "logprob_reward": -0.054930150508880615, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054930150508880615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04825650155544281, + 0.04819754138588905, + 0.048180341720581055, + 0.10563357174396515, + 0.21608316898345947, + 0.15575695037841797, + 0.0960979238152504, + 0.2327827364206314, + 0.030635641887784004, + 0.01837565191090107 + ], + "credit_map": [ + 0.004314242396503687, + 0.006558137014508247, + 0.005076459143310785, + 0.005820034071803093, + 0.007183885667473078, + 0.006681229453533888, + 0.006246600300073624, + 0.006375916302204132, + 0.003212637035176158, + 0.007675232365727425, + 0.005947538651525974, + 0.006196726579219103, + 0.007576140109449625, + 0.005296545103192329, + 0.0066948565654456615, + 0.005597858689725399, + 0.0030108194332569838, + 0.0062657673843204975, + 0.007256920915096998, + 0.006738412659615278, + 0.007919618859887123, + 0.0035124600399285555, + 0.0056778728030622005, + 0.0077984691597521305, + 0.006901389919221401, + 0.013279447332024574, + 0.011826393194496632, + 0.012189186178147793, + 0.010557204484939575, + 0.02734077163040638, + 0.010659345425665379, + 0.012879831716418266, + 0.024978984147310257, + 0.03533853963017464, + 0.02997327595949173, + 0.021304793655872345, + 0.029244771227240562, + 0.01946813054382801, + 0.034619107842445374, + 0.021155575290322304, + 0.013553625904023647, + 0.017391059547662735, + 0.027041811496019363, + 0.013303624466061592, + 0.02480866014957428, + 0.029923323541879654, + 0.015011653304100037, + 0.014723190106451511, + 0.013481984846293926, + 0.010981101542711258, + 0.012568310834467411, + 0.01102756429463625, + 0.011538418009877205, + 0.01203436590731144, + 0.01476098783314228, + 0.009705194272100925, + 0.039701398462057114, + 0.021687252447009087, + 0.047979265451431274, + 0.018343117088079453, + 0.022214993834495544, + 0.032845381647348404, + 0.019213877618312836, + 0.03079746663570404, + 0.002780380193144083, + 0.0047190021723508835, + 0.004145697690546513, + 0.003936434164643288, + 0.0036861575208604336, + 0.005032817833125591, + 0.003718057181686163, + 0.0026170932687819004, + 0.0017589317867532372, + 0.001670615398325026, + 0.003482565050944686, + 0.0018496618140488863, + 0.0018985670758411288, + 0.002211599610745907, + 0.0026520108804106712, + 0.0028517006430774927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.054930150508880615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046312687918543816 + ], + "opposite_logprob_tokens": [ + -5.379631042480469 + ], + "logprob_reward": -0.0046312687918543816, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046312687918543816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0695272758603096, + 0.03034793958067894, + 0.04726168140769005, + 0.02004835568368435, + 0.08711179345846176, + 0.11641408503055573, + 0.06952831894159317, + 0.3471173048019409, + 0.18503274023532867, + 0.02761041559278965 + ], + "credit_map": [ + 0.009376443922519684, + 0.006004910450428724, + 0.010676377452909946, + 0.009299328550696373, + 0.006636383477598429, + 0.008075577206909657, + 0.00901344045996666, + 0.010444819927215576, + 0.00523139676079154, + 0.0030758411157876253, + 0.003807220607995987, + 0.0023956832010298967, + 0.0030495543032884598, + 0.006373938173055649, + 0.0026810576673597097, + 0.0037332498468458652, + 0.0033339867368340492, + 0.0032260813750326633, + 0.01376604288816452, + 0.0034967558458447456, + 0.003004743717610836, + 0.0035595993977040052, + 0.002537565538659692, + 0.014336911961436272, + 0.0026220877189189196, + 0.0025069615803658962, + 0.0030430001206696033, + 0.002107589738443494, + 0.00219230935908854, + 0.003860329743474722, + 0.001728772185742855, + 0.0019873075652867556, + 0.005852000322192907, + 0.01010980736464262, + 0.013328787870705128, + 0.005625034682452679, + 0.02437654510140419, + 0.011270646937191486, + 0.007088022772222757, + 0.009460954926908016, + 0.011885344050824642, + 0.012833046726882458, + 0.020664861425757408, + 0.009333212859928608, + 0.01780695840716362, + 0.023481179028749466, + 0.009810883551836014, + 0.01059861108660698, + 0.014334773644804955, + 0.0053438604809343815, + 0.009256306104362011, + 0.006908657029271126, + 0.006403984967619181, + 0.012735226191580296, + 0.007314343936741352, + 0.007231174036860466, + 0.06069716811180115, + 0.02994149550795555, + 0.07564414292573929, + 0.027580788359045982, + 0.029064571484923363, + 0.05065952613949776, + 0.0249797236174345, + 0.048549946397542953, + 0.021535901352763176, + 0.028631817549467087, + 0.02500736340880394, + 0.015522460453212261, + 0.03684885427355766, + 0.020128650590777397, + 0.022773703560233116, + 0.01458400022238493, + 0.0026374340523034334, + 0.0027476693503558636, + 0.005730418022722006, + 0.0028315929230302572, + 0.002717768307775259, + 0.0037003292236477137, + 0.003393187653273344, + 0.0038520172238349915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0046312687918543816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an unbroken single shot, the camera begins in a mid-level shot, capturing a delicate, small Asian rattan cart. Crafted entirely from hand-woven, light honey-colored rattan strips, the cart displays the intricate weave and the warm, natural luster of the material. Its four small, solid wooden wheels rest against the polished, light-colored wooden floor. As the cart gracefully spins like a top, the camera orbits it in a slow, smooth, clockwise orbit, perfectly synchronizing its movements. The cart pivots on its central axis, steadily and deliberately rotating, its fluid motion carving graceful arcs across the frame. The scene is bathed in soft afternoon sunlight streaming through a large side window. The light, filtered by the gauzy white curtains, is incredibly soft, casting subtle, almost edgeless shadows on the floor. The background is a minimalist interior, scarcely furnished save for a vague pot of greenery in the far corner. The main colors of the picture are cream white, light brown of wood and warm yellow of rattan, creating an extremely tranquil, peaceful and healing atmosphere. The overall visual effect is like a dynamic, airy Japanese healing illustration.", + "seed": 6212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_9/index_7770_seed_6212_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016624494618736208, + "vlm_logprob": -0.016624494618736208, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.016624494618736208, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030420759692788124 + ], + "opposite_logprob_tokens": [ + -3.6554207801818848 + ], + "logprob_reward": -0.030420759692788124, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030420759692788124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05981360748410225, + 0.05535260960459709, + 0.01098564825952053, + 0.009544029831886292, + 0.03035610541701317, + 0.03246090188622475, + 0.6787741184234619, + 0.05690420791506767, + 0.044907599687576294, + 0.02090117707848549 + ], + "credit_map": [ + 0.007298417389392853, + 0.006451508961617947, + 0.005960023496299982, + 0.00662273121997714, + 0.01038604136556387, + 0.006578095722943544, + 0.010228900238871574, + 0.006287890486419201, + 0.004129614681005478, + 0.00888819433748722, + 0.006338018923997879, + 0.009627062827348709, + 0.010788656771183014, + 0.005097521934658289, + 0.005483720917254686, + 0.0049998159520328045, + 0.0019434276036918163, + 0.001741173560731113, + 0.0016164227854460478, + 0.0011921464465558529, + 0.0010372288525104523, + 0.0012468022760003805, + 0.0010993087198585272, + 0.0011091375490650535, + 0.0008777775801718235, + 0.0011810867581516504, + 0.0009988582460209727, + 0.0009894979884847999, + 0.0013569226721301675, + 0.0014854395994916558, + 0.0012986970832571387, + 0.0013557500205934048, + 0.0029590558260679245, + 0.0045417845249176025, + 0.0033848497550934553, + 0.0035260505974292755, + 0.004686935339123011, + 0.004281206056475639, + 0.0040224757976830006, + 0.00295374752022326, + 0.004334411583840847, + 0.002585410140454769, + 0.004221199546009302, + 0.0029409495182335377, + 0.0053143189288675785, + 0.004826657939702272, + 0.004302609246224165, + 0.003935347311198711, + 0.17811447381973267, + 0.057012107223272324, + 0.10673906654119492, + 0.03477786108851433, + 0.05186893790960312, + 0.1575564444065094, + 0.03524002805352211, + 0.05746520310640335, + 0.007363655138760805, + 0.005042682867497206, + 0.007749613840132952, + 0.005524724256247282, + 0.004302459768950939, + 0.011290342546999454, + 0.0046877129934728146, + 0.010943019762635231, + 0.007949264720082283, + 0.003104456467553973, + 0.0077004870399832726, + 0.004311712458729744, + 0.0031856419518589973, + 0.005883524194359779, + 0.006849840749055147, + 0.00592267420142889, + 0.0016327181365340948, + 0.0030568167567253113, + 0.0034051917027682066, + 0.002403387101367116, + 0.0019935809541493654, + 0.002289582509547472, + 0.003012893721461296, + 0.0031070054974406958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.030420759692788124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000931663322262466 + ], + "opposite_logprob_tokens": [ + -7.000931739807129 + ], + "logprob_reward": -0.000931663322262466, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000931663322262466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1691165268421173, + 0.0798591747879982, + 0.016735145822167397, + 0.019757544621825218, + 0.023092424497008324, + 0.03052203170955181, + 0.4256025552749634, + 0.04215420410037041, + 0.10451243072748184, + 0.08864791691303253 + ], + "credit_map": [ + 0.01700446382164955, + 0.018454335629940033, + 0.01754087768495083, + 0.021356018260121346, + 0.023972565308213234, + 0.020462162792682648, + 0.033029768615961075, + 0.0172963235527277, + 0.009389135055243969, + 0.012115396559238434, + 0.009826220571994781, + 0.013449462130665779, + 0.007451160810887814, + 0.006457467097789049, + 0.013232463039457798, + 0.007937868125736713, + 0.0017976408125832677, + 0.0021921321749687195, + 0.0019664925057440996, + 0.0020658185239881277, + 0.0022182681132107973, + 0.0022418322041630745, + 0.001916181412525475, + 0.0023367819376289845, + 0.0018982517067342997, + 0.00234472774900496, + 0.0024832156486809254, + 0.0025479758623987436, + 0.003049129154533148, + 0.0026362398639321327, + 0.0022402829490602016, + 0.0025577209889888763, + 0.002663963008671999, + 0.0024479783605784178, + 0.0025318192783743143, + 0.0027015437372028828, + 0.0030179712921380997, + 0.0040560318157076836, + 0.0030549487564712763, + 0.002618168480694294, + 0.003965318668633699, + 0.0028576492331922054, + 0.0036926246248185635, + 0.0024116658605635166, + 0.005040799733251333, + 0.0036532289814203978, + 0.005068332422524691, + 0.00383241125382483, + 0.11143457144498825, + 0.034485988318920135, + 0.06632518768310547, + 0.021701030433177948, + 0.031969960778951645, + 0.09912585467100143, + 0.024053245782852173, + 0.036506716161966324, + 0.005344209726899862, + 0.00364018976688385, + 0.005015572998672724, + 0.005647303536534309, + 0.005803800653666258, + 0.0071016959846019745, + 0.004515476059168577, + 0.005085954908281565, + 0.010226171463727951, + 0.006257156375795603, + 0.018382076174020767, + 0.01352772954851389, + 0.009556367062032223, + 0.013269851915538311, + 0.010926524177193642, + 0.02236655354499817, + 0.010023780167102814, + 0.006077342666685581, + 0.02023433707654476, + 0.007055988069623709, + 0.007786946836858988, + 0.011267944239079952, + 0.008141553960740566, + 0.018060019239783287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.000931663322262466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05202091485261917, + 0.025620073080062866, + 0.010164237581193447, + 0.009921271353960037, + 0.022559430450201035, + 0.009650067426264286, + 0.7474452257156372, + 0.036506686359643936, + 0.038830094039440155, + 0.047281984239816666 + ], + "credit_map": [ + 0.006705488543957472, + 0.006177261471748352, + 0.005786545109003782, + 0.004961041733622551, + 0.007293886970728636, + 0.006021328270435333, + 0.009817913174629211, + 0.005257454235106707, + 0.002537428867071867, + 0.0033488040789961815, + 0.0022210590541362762, + 0.0028782347217202187, + 0.004484787583351135, + 0.00326144858263433, + 0.004547071177512407, + 0.0023412383161485195, + 0.0019232244230806828, + 0.0013870781986042857, + 0.0014081571716815233, + 0.0012238587951287627, + 0.0012138067977502942, + 0.0010686193127185106, + 0.0010551317827776074, + 0.0008843601681292057, + 0.0008377179619856179, + 0.0009734281920827925, + 0.0010956666665151715, + 0.001257473835721612, + 0.0013825370697304606, + 0.0012693842872977257, + 0.0014020224334672093, + 0.0017030410235747695, + 0.003755225334316492, + 0.002949213143438101, + 0.0027745957486331463, + 0.0019162038806825876, + 0.003096184227615595, + 0.0036052558571100235, + 0.0020665363408625126, + 0.002396213822066784, + 0.001072283717803657, + 0.0011741522466763854, + 0.001244598301127553, + 0.0010084250243380666, + 0.0013009494869038463, + 0.0009474236867390573, + 0.00176624592859298, + 0.0011359890922904015, + 0.1984344720840454, + 0.06062615290284157, + 0.11431068927049637, + 0.038571447134017944, + 0.0558578297495842, + 0.17552323639392853, + 0.042801354080438614, + 0.06132006645202637, + 0.004545893520116806, + 0.0045329732820391655, + 0.0045935227535665035, + 0.003766866633668542, + 0.0037030144594609737, + 0.006157331168651581, + 0.0033459786791354418, + 0.00586110632866621, + 0.007492380682379007, + 0.00296768918633461, + 0.006895957514643669, + 0.0033919105771929026, + 0.0025506895035505295, + 0.0053179822862148285, + 0.003314733039587736, + 0.0068987528793513775, + 0.006364012602716684, + 0.003113193204626441, + 0.011123529635369778, + 0.0037236702628433704, + 0.002776071662083268, + 0.007610626984387636, + 0.004310180898755789, + 0.008260700851678848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304375872015953 + ], + "opposite_logprob_tokens": [ + -3.168043851852417 + ], + "logprob_reward": -0.04304375872015953, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304375872015953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11694810539484024, + 0.05319507420063019, + 0.0576779879629612, + 0.03852039575576782, + 0.03512685000896454, + 0.08627410233020782, + 0.3218976855278015, + 0.11803025007247925, + 0.10474425554275513, + 0.0675852969288826 + ], + "credit_map": [ + 0.012309043668210506, + 0.015017645433545113, + 0.0119748180732131, + 0.014008600264787674, + 0.01753571256995201, + 0.010620632208883762, + 0.023923132568597794, + 0.01155850850045681, + 0.00432467320933938, + 0.006276723463088274, + 0.00667871069163084, + 0.009511064738035202, + 0.006971146911382675, + 0.006216008216142654, + 0.007580472622066736, + 0.005636272020637989, + 0.006958339363336563, + 0.006844089832156897, + 0.007417530287057161, + 0.008335230872035027, + 0.007632754277437925, + 0.006534500513225794, + 0.007997008040547371, + 0.00595853291451931, + 0.0049884263426065445, + 0.003939344082027674, + 0.004368708468973637, + 0.003998545929789543, + 0.004876375198364258, + 0.006488426588475704, + 0.00482975784689188, + 0.00503080990165472, + 0.0035240757279098034, + 0.004355412442237139, + 0.004313231445848942, + 0.00269950064830482, + 0.007356259040534496, + 0.0036940027493983507, + 0.005638244096189737, + 0.0035461201332509518, + 0.007571342866867781, + 0.0076913186348974705, + 0.010659350082278252, + 0.010054162703454494, + 0.013821440748870373, + 0.009489409625530243, + 0.016968408599495888, + 0.010018662549555302, + 0.08027641475200653, + 0.026175551116466522, + 0.05114157497882843, + 0.01817120611667633, + 0.02476770430803299, + 0.07420806586742401, + 0.019231172278523445, + 0.027925986796617508, + 0.02011505514383316, + 0.011060620658099651, + 0.015067927539348602, + 0.01947890967130661, + 0.011790511198341846, + 0.014803898520767689, + 0.011730204336345196, + 0.01398311834782362, + 0.01045906636863947, + 0.008438512682914734, + 0.019369902089238167, + 0.013515752740204334, + 0.006490119732916355, + 0.009995798580348492, + 0.018251268193125725, + 0.018223837018013, + 0.006056784652173519, + 0.006307366769760847, + 0.011405855417251587, + 0.01001005806028843, + 0.006562161725014448, + 0.005984237417578697, + 0.010633856058120728, + 0.010624981485307217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.04304375872015953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088377580046654 + ], + "opposite_logprob_tokens": [ + -5.504088401794434 + ], + "logprob_reward": -0.004088377580046654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088377580046654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1119571253657341, + 0.08461792021989822, + 0.05920534208416939, + 0.03842669352889061, + 0.09744490683078766, + 0.10029271245002747, + 0.3420141041278839, + 0.04894304275512695, + 0.07340554893016815, + 0.04369264841079712 + ], + "credit_map": [ + 0.01749357022345066, + 0.015005073510110378, + 0.01272718608379364, + 0.010560809634625912, + 0.013146505691111088, + 0.01373287569731474, + 0.01680057868361473, + 0.012490513734519482, + 0.009817247278988361, + 0.009046023711562157, + 0.011523250490427017, + 0.01688300259411335, + 0.011610174551606178, + 0.0069382344372570515, + 0.010389411821961403, + 0.008410559967160225, + 0.00744252884760499, + 0.008106193505227566, + 0.007311705499887466, + 0.009686595760285854, + 0.006088884547352791, + 0.008962026797235012, + 0.005718511529266834, + 0.005888890940696001, + 0.00614018552005291, + 0.004005835857242346, + 0.005287960171699524, + 0.0033047785982489586, + 0.004496442619711161, + 0.006574888247996569, + 0.003560543293133378, + 0.005056057125329971, + 0.012841613031923771, + 0.005980870686471462, + 0.008343316614627838, + 0.006925965193659067, + 0.014397842809557915, + 0.027753246948122978, + 0.00722389854490757, + 0.013978150673210621, + 0.02351118065416813, + 0.00679294066503644, + 0.014191022142767906, + 0.008238324895501137, + 0.009016432799398899, + 0.010263695381581783, + 0.011869982816278934, + 0.01640912890434265, + 0.08850368857383728, + 0.02880556881427765, + 0.05510544776916504, + 0.017322277650237083, + 0.026508746668696404, + 0.0771188884973526, + 0.018961308524012566, + 0.029688144102692604, + 0.01200100313872099, + 0.005821483675390482, + 0.006436379160732031, + 0.003843923332169652, + 0.004226248245686293, + 0.006323909852653742, + 0.004015329759567976, + 0.006274764891713858, + 0.0109379468485713, + 0.006217745132744312, + 0.012708411552011967, + 0.007646187208592892, + 0.004961550701409578, + 0.010207615792751312, + 0.006110782735049725, + 0.014615303836762905, + 0.004823720082640648, + 0.004687323234975338, + 0.007796961348503828, + 0.0053744069300591946, + 0.004461674485355616, + 0.004480583127588034, + 0.006203195080161095, + 0.005864775739610195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.004088377580046654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an unbroken single shot, the camera begins in a mid-level shot, capturing a delicate, small Asian rattan cart. Crafted entirely from hand-woven, light honey-colored rattan strips, the cart displays the intricate weave and the warm, natural luster of the material. Its four small, solid wooden wheels rest against the polished, light-colored wooden floor. As the cart gracefully spins like a top, the camera orbits it in a slow, smooth, clockwise orbit, perfectly synchronizing its movements. The cart pivots on its central axis, steadily and deliberately rotating, its fluid motion carving graceful arcs across the frame. The scene is bathed in soft afternoon sunlight streaming through a large side window. The light, filtered by the gauzy white curtains, is incredibly soft, casting subtle, almost edgeless shadows on the floor. The background is a minimalist interior, scarcely furnished save for a vague pot of greenery in the far corner. The main colors of the picture are cream white, light brown of wood and warm yellow of rattan, creating an extremely tranquil, peaceful and healing atmosphere. The overall visual effect is like a dynamic, airy Japanese healing illustration.", + "seed": 6213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_9/index_7770_seed_6213_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03186970045790076, + "vlm_logprob": -0.03186970045790076, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03186970045790076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0814008042216301 + ], + "opposite_logprob_tokens": [ + -2.5814008712768555 + ], + "logprob_reward": -0.0814008042216301, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0814008042216301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07643214613199234, + 0.060910265892744064, + 0.213433638215065, + 0.04594648629426956, + 0.2634661793708801, + 0.05127270519733429, + 0.02679344080388546, + 0.021904800087213516, + 0.20994317531585693, + 0.029897136613726616 + ], + "credit_map": [ + 0.010159183293581009, + 0.008719290606677532, + 0.010337800718843937, + 0.008788717910647392, + 0.008364072069525719, + 0.009153361432254314, + 0.0063817379996180534, + 0.01452798955142498, + 0.010685984045267105, + 0.007275775074958801, + 0.016345636919140816, + 0.005241378676146269, + 0.006101092789322138, + 0.006104674190282822, + 0.004077454563230276, + 0.0050782710313797, + 0.02900373749434948, + 0.035057876259088516, + 0.04779941216111183, + 0.01801699586212635, + 0.026651114225387573, + 0.02017262391746044, + 0.019122576341032982, + 0.01760931871831417, + 0.0057315705344080925, + 0.003981905989348888, + 0.003708360018208623, + 0.005392659921199083, + 0.007114863488823175, + 0.008068803697824478, + 0.006657759193331003, + 0.00529056740924716, + 0.013102400116622448, + 0.015488095581531525, + 0.06688540428876877, + 0.010911114513874054, + 0.10089316964149475, + 0.0329940989613533, + 0.01230614073574543, + 0.010885761119425297, + 0.00478809978812933, + 0.014264301396906376, + 0.009315375238656998, + 0.003973214887082577, + 0.005335064139217138, + 0.004395380616188049, + 0.005741749424487352, + 0.003459524130448699, + 0.002639030572026968, + 0.003982714377343655, + 0.002487217541784048, + 0.004419637378305197, + 0.002555896993726492, + 0.005002544727176428, + 0.0030264630913734436, + 0.002679936122149229, + 0.002007481176406145, + 0.0023186078760772943, + 0.0034107486717402935, + 0.0025815258268266916, + 0.0028628818690776825, + 0.004504806362092495, + 0.002094063675031066, + 0.002124684862792492, + 0.011800957843661308, + 0.029906028881669044, + 0.0263530183583498, + 0.03293962776660919, + 0.03215285763144493, + 0.02094682678580284, + 0.034956615418195724, + 0.020887259393930435, + 0.002111515961587429, + 0.004221267066895962, + 0.002976904157549143, + 0.005891493055969477, + 0.005169504322111607, + 0.0017410198925063014, + 0.005361448973417282, + 0.00242398539558053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.0814008042216301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024978886358439922 + ], + "opposite_logprob_tokens": [ + -6.002497673034668 + ], + "logprob_reward": -0.0024978886358439922, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024978886358439922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06998111307621002, + 0.05378315597772598, + 0.15861119329929352, + 0.0657060518860817, + 0.2388555109500885, + 0.0862317830324173, + 0.03267008438706398, + 0.036017678678035736, + 0.17714263498783112, + 0.08100082725286484 + ], + "credit_map": [ + 0.008653287775814533, + 0.008412745781242847, + 0.011846811510622501, + 0.00918011087924242, + 0.008953926153481007, + 0.009105108678340912, + 0.007411787752062082, + 0.006417330354452133, + 0.006592216901481152, + 0.01128148939460516, + 0.005270053632557392, + 0.007307784166187048, + 0.007267187349498272, + 0.008469238877296448, + 0.002955035539343953, + 0.004640147089958191, + 0.02245396561920643, + 0.018482668325304985, + 0.04116296023130417, + 0.014589585363864899, + 0.01693001016974449, + 0.01454981043934822, + 0.015256300568580627, + 0.01518588699400425, + 0.007378722075372934, + 0.006784360855817795, + 0.006236308254301548, + 0.010536841116845608, + 0.009512201882898808, + 0.010377131402492523, + 0.00657001044601202, + 0.008310476318001747, + 0.014533904381096363, + 0.014795769937336445, + 0.06210947036743164, + 0.009245876222848892, + 0.09108822792768478, + 0.02730085328221321, + 0.009117621928453445, + 0.01066378504037857, + 0.006991160102188587, + 0.02154565416276455, + 0.013957166112959385, + 0.006781876552850008, + 0.009214959107339382, + 0.00828493945300579, + 0.011939872987568378, + 0.007516155019402504, + 0.00595439225435257, + 0.0038430069107562304, + 0.002734122099354863, + 0.0034906575456261635, + 0.004466983024030924, + 0.004934874828904867, + 0.003953577950596809, + 0.0032924700062721968, + 0.0037464883644133806, + 0.008531725965440273, + 0.0050954995676875114, + 0.003338856855407357, + 0.003222654340788722, + 0.005045183934271336, + 0.0031629703007638454, + 0.003874300979077816, + 0.010737535543739796, + 0.02510247193276882, + 0.02512834221124649, + 0.026261301711201668, + 0.026533208787441254, + 0.017429184168577194, + 0.02903437614440918, + 0.01691620796918869, + 0.006689336150884628, + 0.014138936065137386, + 0.010808607563376427, + 0.011276792734861374, + 0.010770595632493496, + 0.004790097009390593, + 0.011163905262947083, + 0.011362554505467415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.0024978886358439922 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637094870209694 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008637094870209694, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637094870209694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2013077586889267, + 0.06722938269376755, + 0.22819916903972626, + 0.0372682623565197, + 0.12135443836450577, + 0.08078718930482864, + 0.08461631834506989, + 0.03153020143508911, + 0.07116402685642242, + 0.07654323428869247 + ], + "credit_map": [ + 0.023937515914440155, + 0.02806938625872135, + 0.02222270518541336, + 0.02314632385969162, + 0.035347823053598404, + 0.02648797072470188, + 0.021132059395313263, + 0.020963983610272408, + 0.01073000393807888, + 0.010033836588263512, + 0.015652569010853767, + 0.006100438069552183, + 0.00628279522061348, + 0.009123270399868488, + 0.0030710508581250906, + 0.006235412787646055, + 0.0306932982057333, + 0.031073564663529396, + 0.053020332008600235, + 0.019473841413855553, + 0.027520323172211647, + 0.022142931818962097, + 0.020987920463085175, + 0.023286955431103706, + 0.00544445076957345, + 0.004060657694935799, + 0.004173826891928911, + 0.005050635896623135, + 0.004856606014072895, + 0.005260571837425232, + 0.0035488291177898645, + 0.004872684832662344, + 0.008079088293015957, + 0.007568172179162502, + 0.029814446344971657, + 0.004981046076864004, + 0.04356381297111511, + 0.015451001934707165, + 0.006191507447510958, + 0.005705373827368021, + 0.00642020720988512, + 0.016637979075312614, + 0.011720500886440277, + 0.006940217223018408, + 0.008449376560747623, + 0.01159617118537426, + 0.010313349775969982, + 0.008709385991096497, + 0.016112977638840675, + 0.010633141733705997, + 0.00959024578332901, + 0.007932265289127827, + 0.009453813545405865, + 0.01483960822224617, + 0.007136063650250435, + 0.008918201550841331, + 0.0027723440434783697, + 0.007516564801335335, + 0.0046464828774333, + 0.0035605395678430796, + 0.00256901397369802, + 0.004504058975726366, + 0.003288606181740761, + 0.002672589849680662, + 0.003623406169936061, + 0.010302226059138775, + 0.006352291442453861, + 0.010712315328419209, + 0.013828326016664505, + 0.006420281250029802, + 0.013728542253375053, + 0.0061966353096067905, + 0.006207102909684181, + 0.01352836936712265, + 0.011729502119123936, + 0.009885338135063648, + 0.00883565004914999, + 0.004689249210059643, + 0.009898222982883453, + 0.011769796721637249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.008637094870209694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863535985350609 + ], + "opposite_logprob_tokens": [ + -3.048635244369507 + ], + "logprob_reward": -0.04863535985350609, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863535985350609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11515295505523682, + 0.08266239613294601, + 0.11923377215862274, + 0.0560610294342041, + 0.19342085719108582, + 0.16141079366207123, + 0.07595066726207733, + 0.11069082468748093, + 0.03722604364156723, + 0.0481906533241272 + ], + "credit_map": [ + 0.014138392172753811, + 0.014238516800105572, + 0.011516681872308254, + 0.01409442350268364, + 0.02426716312766075, + 0.018099086359143257, + 0.00968075543642044, + 0.009117928333580494, + 0.010635587386786938, + 0.011624054051935673, + 0.016437314450740814, + 0.004938580561429262, + 0.016278769820928574, + 0.012712178751826286, + 0.005595508497208357, + 0.004440407734364271, + 0.02005269005894661, + 0.013871394097805023, + 0.037606362253427505, + 0.006817924324423075, + 0.01717056892812252, + 0.007533124182373285, + 0.01081082969903946, + 0.005370876286178827, + 0.007270645350217819, + 0.005956439301371574, + 0.005522767547518015, + 0.008881445974111557, + 0.006984969601035118, + 0.008486554026603699, + 0.006423762999475002, + 0.006534444633871317, + 0.014647020027041435, + 0.014256909489631653, + 0.047623153775930405, + 0.008990387432277203, + 0.07196314632892609, + 0.021473074331879616, + 0.007377726025879383, + 0.007089447230100632, + 0.01123808789998293, + 0.04705416038632393, + 0.03284527733922005, + 0.01721692457795143, + 0.012787614949047565, + 0.015597718767821789, + 0.015377339906990528, + 0.009293674491345882, + 0.00586781557649374, + 0.009171715006232262, + 0.005925996229052544, + 0.008868636563420296, + 0.011405548080801964, + 0.015357710421085358, + 0.011303875595331192, + 0.008049370720982552, + 0.007343057077378035, + 0.03202926367521286, + 0.015172340907156467, + 0.01065201498568058, + 0.012093329802155495, + 0.014740249142050743, + 0.011450428515672684, + 0.007210131268948317, + 0.0032121180556714535, + 0.004891602322459221, + 0.004319118335843086, + 0.003558807773515582, + 0.006839855574071407, + 0.003403971903026104, + 0.006935915444046259, + 0.004064654000103474, + 0.0034593911841511726, + 0.00624378165230155, + 0.0052607739344239235, + 0.01130523718893528, + 0.0054022567346692085, + 0.002432808745652437, + 0.007565823383629322, + 0.006520580966025591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.04863535985350609 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177354708313942 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177354708313942, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177354708313942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2853146493434906, + 0.038609810173511505, + 0.0420268177986145, + 0.12475048750638962, + 0.10856888443231583, + 0.04896989092230797, + 0.08167210966348648, + 0.023635294288396835, + 0.12659664452075958, + 0.11985539644956589 + ], + "credit_map": [ + 0.03200005739927292, + 0.04434765875339508, + 0.029407983645796776, + 0.02272723987698555, + 0.060792457312345505, + 0.04694550484418869, + 0.020250020548701286, + 0.028843728825449944, + 0.008186179213225842, + 0.003420476568862796, + 0.01257804874330759, + 0.0024086448829621077, + 0.0037228723522275686, + 0.003774773795157671, + 0.0023782995995134115, + 0.002140514086931944, + 0.004194383509457111, + 0.010509287007153034, + 0.0035103200934827328, + 0.003840781282633543, + 0.007589790970087051, + 0.005022760014981031, + 0.003691914724186063, + 0.003667576937004924, + 0.018016504123806953, + 0.02174294926226139, + 0.010266188532114029, + 0.01742270402610302, + 0.01591627672314644, + 0.014094634912908077, + 0.014015767723321915, + 0.013275462202727795, + 0.009544305503368378, + 0.007313696201890707, + 0.024977920576930046, + 0.005246253684163094, + 0.03703215718269348, + 0.010803296230733395, + 0.006220665294677019, + 0.007430591154843569, + 0.004167960956692696, + 0.010405152104794979, + 0.007518555968999863, + 0.004308376461267471, + 0.00524486368522048, + 0.00476045161485672, + 0.0075581432320177555, + 0.0050063892267644405, + 0.013129604049026966, + 0.013624098151922226, + 0.007657050620764494, + 0.010774902068078518, + 0.006833473686128855, + 0.01452995277941227, + 0.006844209972769022, + 0.008278820663690567, + 0.0018426404567435384, + 0.005744277033954859, + 0.0023908568546175957, + 0.0026101393159478903, + 0.002500941976904869, + 0.004082856699824333, + 0.00223891856148839, + 0.0022246630396693945, + 0.011143816635012627, + 0.019063755869865417, + 0.021931204944849014, + 0.016584714874625206, + 0.011540780775249004, + 0.013826786540448666, + 0.016369378194212914, + 0.016136208549141884, + 0.011563926935195923, + 0.020395787432789803, + 0.01960025168955326, + 0.014046629890799522, + 0.01022964995354414, + 0.011296555399894714, + 0.013014293275773525, + 0.0197082981467247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.018177354708313942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an unbroken single shot, the camera begins in a mid-level shot, capturing a delicate, small Asian rattan cart. Crafted entirely from hand-woven, light honey-colored rattan strips, the cart displays the intricate weave and the warm, natural luster of the material. Its four small, solid wooden wheels rest against the polished, light-colored wooden floor. As the cart gracefully spins like a top, the camera orbits it in a slow, smooth, clockwise orbit, perfectly synchronizing its movements. The cart pivots on its central axis, steadily and deliberately rotating, its fluid motion carving graceful arcs across the frame. The scene is bathed in soft afternoon sunlight streaming through a large side window. The light, filtered by the gauzy white curtains, is incredibly soft, casting subtle, almost edgeless shadows on the floor. The background is a minimalist interior, scarcely furnished save for a vague pot of greenery in the far corner. The main colors of the picture are cream white, light brown of wood and warm yellow of rattan, creating an extremely tranquil, peaceful and healing atmosphere. The overall visual effect is like a dynamic, airy Japanese healing illustration.", + "seed": 6214, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000150/group_9/index_7770_seed_6214_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012207211775239557, + "vlm_logprob": -0.012207211775239557, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.012207211775239557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017395777627825737 + ], + "opposite_logprob_tokens": [ + -4.267395973205566 + ], + "logprob_reward": -0.017395777627825737, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017395777627825737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18236641585826874, + 0.0455414243042469, + 0.09146706014871597, + 0.10824574530124664, + 0.04323301464319229, + 0.21121254563331604, + 0.19580157101154327, + 0.04162850230932236, + 0.04147740826010704, + 0.03902635723352432 + ], + "credit_map": [ + 0.015809400007128716, + 0.026766082271933556, + 0.014206923544406891, + 0.022126339375972748, + 0.026931438595056534, + 0.02755747362971306, + 0.02794012986123562, + 0.021028641611337662, + 0.0054939561523497105, + 0.008489157073199749, + 0.00401622848585248, + 0.005210486706346273, + 0.00690501369535923, + 0.0077733928337693214, + 0.00401117792353034, + 0.0036420091055333614, + 0.007738495245575905, + 0.01016103383153677, + 0.0075333802960813046, + 0.012388894334435463, + 0.019925961270928383, + 0.015311126597225666, + 0.008105617016553879, + 0.010302545502781868, + 0.01311240904033184, + 0.019152432680130005, + 0.01100903283804655, + 0.012647894211113453, + 0.013495638966560364, + 0.013000828213989735, + 0.012061175890266895, + 0.013766332529485226, + 0.004456713330000639, + 0.006200561765581369, + 0.003408326068893075, + 0.005380446091294289, + 0.005434335675090551, + 0.009456997737288475, + 0.0054619996808469296, + 0.003433634527027607, + 0.030644796788692474, + 0.04816802963614464, + 0.01857774518430233, + 0.022332964465022087, + 0.025592053309082985, + 0.03405873104929924, + 0.02062341757118702, + 0.011214795522391796, + 0.022235943004488945, + 0.03969479352235794, + 0.016576359048485756, + 0.03596687689423561, + 0.024621514603495598, + 0.023264266550540924, + 0.023998107761144638, + 0.00944368913769722, + 0.007698190398514271, + 0.00546602439135313, + 0.004247721750289202, + 0.005477484315633774, + 0.003945860080420971, + 0.005873748101294041, + 0.006129395216703415, + 0.0027900778222829103, + 0.005854938644915819, + 0.004890223499387503, + 0.0038037076592445374, + 0.005917706526815891, + 0.004502337891608477, + 0.004929741378873587, + 0.0072908950969576836, + 0.0042878552339971066, + 0.005859821569174528, + 0.007039415650069714, + 0.0040176790207624435, + 0.003511616727337241, + 0.00547909177839756, + 0.005812318995594978, + 0.002730355830863118, + 0.004576057195663452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.017395777627825737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008219677838496864 + ], + "opposite_logprob_tokens": [ + -7.125822067260742 + ], + "logprob_reward": -0.0008219677838496864, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008219677838496864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12238907068967819, + 0.11802099645137787, + 0.14747628569602966, + 0.06621260195970535, + 0.04443785175681114, + 0.0814119502902031, + 0.22135257720947266, + 0.05563775449991226, + 0.048336636275053024, + 0.09472427517175674 + ], + "credit_map": [ + 0.015401653945446014, + 0.015925850719213486, + 0.00862058438360691, + 0.010088796727359295, + 0.02730466052889824, + 0.019376115873456, + 0.014785155653953552, + 0.010886258445680141, + 0.015595241449773312, + 0.018456000834703445, + 0.01244262047111988, + 0.01642674207687378, + 0.015746628865599632, + 0.016548480838537216, + 0.011542558670043945, + 0.011262714862823486, + 0.012257633730769157, + 0.01433784794062376, + 0.01024781633168459, + 0.013623470440506935, + 0.03283662348985672, + 0.03612698242068291, + 0.011074290610849857, + 0.01697162538766861, + 0.007192795630544424, + 0.010325020179152489, + 0.006598367355763912, + 0.01096668653190136, + 0.005759034305810928, + 0.008297329768538475, + 0.009702497161924839, + 0.007370867300778627, + 0.006082585547119379, + 0.006033899262547493, + 0.0046929712407290936, + 0.005796143319457769, + 0.004767984617501497, + 0.005726323928683996, + 0.006873768754303455, + 0.004464174620807171, + 0.0166401956230402, + 0.010192287154495716, + 0.005660803988575935, + 0.0078033218160271645, + 0.0086180055513978, + 0.019808964803814888, + 0.00770206144079566, + 0.004986307118088007, + 0.02355547435581684, + 0.03842170163989067, + 0.019014567136764526, + 0.05256656929850578, + 0.021723756566643715, + 0.024745797738432884, + 0.029844406992197037, + 0.01148030161857605, + 0.007141555193811655, + 0.006092218682169914, + 0.006004233844578266, + 0.013290820643305779, + 0.005683744326233864, + 0.007144625298678875, + 0.0070665376260876656, + 0.003214022610336542, + 0.005174597725272179, + 0.005454829894006252, + 0.005957308225333691, + 0.006037966348230839, + 0.006198153831064701, + 0.006096289958804846, + 0.00856329221278429, + 0.004854198079556227, + 0.009753292426466942, + 0.0193273164331913, + 0.0075105708092451096, + 0.018025079742074013, + 0.010490043088793755, + 0.013099825009703636, + 0.006849299184978008, + 0.009668855927884579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.0008219677838496864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031925681978464127 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031925681978464127, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031925681978464127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09746701270341873, + 0.03997877240180969, + 0.1053764596581459, + 0.08258463442325592, + 0.041204724460840225, + 0.18198035657405853, + 0.2192496359348297, + 0.07786756008863449, + 0.08128204196691513, + 0.07300880551338196 + ], + "credit_map": [ + 0.013410956598818302, + 0.011241092346608639, + 0.012175790034234524, + 0.008369631133973598, + 0.011832017451524734, + 0.013963798061013222, + 0.013191822916269302, + 0.013281897641718388, + 0.00347515637986362, + 0.004699323326349258, + 0.004034620244055986, + 0.004870469216257334, + 0.006083178333938122, + 0.007748774718493223, + 0.004851006902754307, + 0.004216240253299475, + 0.009888813830912113, + 0.014415204524993896, + 0.009670927189290524, + 0.013513279147446156, + 0.013829217292368412, + 0.02094833180308342, + 0.010733164846897125, + 0.01237751916050911, + 0.007775145582854748, + 0.01146306935697794, + 0.008423338644206524, + 0.014420552179217339, + 0.010888403281569481, + 0.011341134086251259, + 0.007611747831106186, + 0.010661240667104721, + 0.005135794635862112, + 0.004905989393591881, + 0.0031197492498904467, + 0.004851032048463821, + 0.007583849597722292, + 0.006498289760202169, + 0.005464312620460987, + 0.0036457080859690905, + 0.02850537747144699, + 0.0252663716673851, + 0.014229643158614635, + 0.022166535258293152, + 0.016891786828637123, + 0.03783543407917023, + 0.024345271289348602, + 0.01273992471396923, + 0.023096250370144844, + 0.039415664970874786, + 0.0176002886146307, + 0.05149712786078453, + 0.022160697728395462, + 0.02524712309241295, + 0.029732368886470795, + 0.010500116273760796, + 0.01017057616263628, + 0.009180969558656216, + 0.009105886332690716, + 0.016652336344122887, + 0.007880300283432007, + 0.00935585517436266, + 0.01029879879206419, + 0.0052228402346372604, + 0.010339480824768543, + 0.00851740874350071, + 0.011579166166484356, + 0.010689864866435528, + 0.008785326965153217, + 0.009390258230268955, + 0.014434742741286755, + 0.007545797619968653, + 0.007936769165098667, + 0.01804313063621521, + 0.004471842665225267, + 0.011175769381225109, + 0.009514670819044113, + 0.011066914536058903, + 0.004942149389535189, + 0.005857560317963362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.0031925681978464127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368392214179039 + ], + "opposite_logprob_tokens": [ + -3.408684015274048 + ], + "logprob_reward": -0.03368392214179039, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368392214179039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10833866894245148, + 0.040996626019477844, + 0.2243156135082245, + 0.08895552903413773, + 0.062312670052051544, + 0.08981996029615402, + 0.21176312863826752, + 0.06822491437196732, + 0.06022068113088608, + 0.0450521744787693 + ], + "credit_map": [ + 0.010234858840703964, + 0.021826473996043205, + 0.007826053537428379, + 0.010580804198980331, + 0.015667222440242767, + 0.01672174222767353, + 0.012477315962314606, + 0.013004195876419544, + 0.004711329471319914, + 0.005155626684427261, + 0.0037464704364538193, + 0.004840172361582518, + 0.004695079755038023, + 0.008032245561480522, + 0.0050417352467775345, + 0.004773966036736965, + 0.015164717100560665, + 0.03413620591163635, + 0.012766756117343903, + 0.030352091416716576, + 0.06122775375843048, + 0.04471132531762123, + 0.013498506508767605, + 0.01245828252285719, + 0.010027812793850899, + 0.011369076557457447, + 0.006799495313316584, + 0.019117947667837143, + 0.008446008898317814, + 0.013413063250482082, + 0.011170345358550549, + 0.008611776866018772, + 0.0076287416741251945, + 0.010020062327384949, + 0.005214812699705362, + 0.008455798961222172, + 0.009138248860836029, + 0.009766890667378902, + 0.007656116038560867, + 0.004432000685483217, + 0.01285729743540287, + 0.016110245138406754, + 0.005742932669818401, + 0.011207357048988342, + 0.00851497519761324, + 0.0153869753703475, + 0.012724287807941437, + 0.00727588776499033, + 0.021809037774801254, + 0.04211071878671646, + 0.017223859205842018, + 0.04560801759362221, + 0.023652901872992516, + 0.024594036862254143, + 0.02709021046757698, + 0.009674363769590855, + 0.009250900708138943, + 0.007167219650000334, + 0.007830320857465267, + 0.014335842803120613, + 0.0064062271267175674, + 0.010004759766161442, + 0.00841326080262661, + 0.004816377069801092, + 0.010341038927435875, + 0.00826755166053772, + 0.004956426098942757, + 0.00705537386238575, + 0.006757717579603195, + 0.008042271248996258, + 0.008012857288122177, + 0.006787443067878485, + 0.007222939282655716, + 0.006520075257867575, + 0.004351852461695671, + 0.006211227271705866, + 0.005720073357224464, + 0.006516391411423683, + 0.0035209173802286386, + 0.004988699220120907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.03368392214179039 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small Asian rattan cart crafted from hand-woven strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cart colored light honey and does it display an intricate weave with a warm natural luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cart have four small solid wooden wheels resting on a polished light-colored wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cart spin like a top while the camera orbits it in a slow smooth clockwise motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through gauzy white curtains, casting subtle edgeless shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005941823124885559 + ], + "opposite_logprob_tokens": [ + -5.130941867828369 + ], + "logprob_reward": -0.005941823124885559, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005941823124885559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10796555876731873, + 0.06809934973716736, + 0.24218891561031342, + 0.07588033378124237, + 0.023746352642774582, + 0.08600830286741257, + 0.16667144000530243, + 0.07520724087953568, + 0.059451792389154434, + 0.09478078037500381 + ], + "credit_map": [ + 0.013367763720452785, + 0.012629667297005653, + 0.013998731970787048, + 0.01068187691271305, + 0.015446601435542107, + 0.017767664045095444, + 0.010756888426840305, + 0.013316351920366287, + 0.008380024693906307, + 0.01105772890150547, + 0.00943195354193449, + 0.008746818639338017, + 0.008534461259841919, + 0.010223816148936749, + 0.005865058396011591, + 0.005859481170773506, + 0.026957115158438683, + 0.026926079764962196, + 0.01864677108824253, + 0.0329597182571888, + 0.03169575706124306, + 0.054403480142354965, + 0.021876713261008263, + 0.028723249211907387, + 0.014522142708301544, + 0.011351664550602436, + 0.007269817870110273, + 0.01204679161310196, + 0.0070188045501708984, + 0.00913042202591896, + 0.00691877119243145, + 0.007621905766427517, + 0.0032215998508036137, + 0.0032997475937008858, + 0.0017321956111118197, + 0.0030772823374718428, + 0.0025632206816226244, + 0.004963897168636322, + 0.002322555286809802, + 0.002565850969403982, + 0.012705470435321331, + 0.018977010622620583, + 0.007135590072721243, + 0.009128229692578316, + 0.009941639378666878, + 0.014889481477439404, + 0.007704042363911867, + 0.005526831839233637, + 0.01753525622189045, + 0.033418141305446625, + 0.01624620333313942, + 0.03341638296842575, + 0.018351862207055092, + 0.019491570070385933, + 0.020488178357481956, + 0.007723814807832241, + 0.013857142999768257, + 0.012396245263516903, + 0.010211525484919548, + 0.007691198494285345, + 0.0082197654992342, + 0.010020501911640167, + 0.0070190029218792915, + 0.005791844334453344, + 0.008613292127847672, + 0.009603380225598812, + 0.0044786930084228516, + 0.005406910087913275, + 0.012154832482337952, + 0.007257941644638777, + 0.007181012071669102, + 0.004755723290145397, + 0.012246344238519669, + 0.02287270687520504, + 0.006193708162754774, + 0.012195263057947159, + 0.014364203438162804, + 0.014561357907950878, + 0.005183371249586344, + 0.00716381287202239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.005941823124885559 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_0/index_tensor([3190, 3190])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_0/index_tensor([3190, 3190])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..218bca85818bacbda4dfe878a11af9569856b15c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_0/index_tensor([3190, 3190])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Inspired by traditional Japanese aesthetics, a low-angle, full-shot shot slowly pans forward, focusing on a lone figure standing defiantly amidst a catastrophic urban flood. The figure is adorned with the ornate Sacra Umbra Armor, a masterpiece of darkly painted metal plates intricately connected by deep blue silk thread. The helmet, a formidable kabuto, features a menacing, full-face man-yorai mask that conceals all features, its metallic scowl fixed against the floodwaters. Rain pours down like a thick downpour, flowing over the armor's polished, obsidian-like surface, beading like tiny gems before cascading into the churning waters below. The figure stands waist-deep in raging, murky water, which swirls powerfully around their legs. In the background, the submerged cityscape is a chaotic mix of modern high-rises and traditional buildings, their lower halves engulfed by the floodwaters. The entire scene is lit almost entirely from below by the lights of underwater buildings and neon signs, which flicker ominously beneath the turbulent surface, casting a ghostly, refracted glow of electric blues and garish reds. This otherworldly underwater light creates deep, dramatic shadows that stretch and twist across the rising waters, painting a scene of somber, apocalyptic grandeur, while the figure remains completely motionless, a stoic fortress of shadow and steel against the raging storm.", + "seed": 19671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_0/index_3190_seed_19671_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19945380948483943, + "vlm_logprob": -0.19945380948483943, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.19945380948483943, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19335079193115234 + ], + "opposite_logprob_tokens": [ + -1.8183507919311523 + ], + "logprob_reward": -0.19335079193115234, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19335079193115234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2715263366699219, + 0.38963112235069275, + 0.032483890652656555, + 0.04212860018014908, + 0.060092225670814514, + 0.05486328527331352, + 0.035333383828401566, + 0.04857981577515602, + 0.023774895817041397, + 0.041586488485336304 + ], + "credit_map": [ + 0.02940063551068306, + 0.029289519414305687, + 0.03260160610079765, + 0.03541511297225952, + 0.03617784008383751, + 0.041937656700611115, + 0.039552755653858185, + 0.0271511971950531, + 0.09359525144100189, + 0.017986370250582695, + 0.01749306172132492, + 0.14541757106781006, + 0.024634381756186485, + 0.008981463499367237, + 0.07011875510215759, + 0.011404241435229778, + 0.00428058672696352, + 0.004691945854574442, + 0.003327394602820277, + 0.0032713983673602343, + 0.004272468388080597, + 0.003683100687339902, + 0.004202762618660927, + 0.004754231311380863, + 0.006639615632593632, + 0.0032379908952862024, + 0.004245966672897339, + 0.004748916253447533, + 0.003945678472518921, + 0.00923166610300541, + 0.004212509375065565, + 0.005866253282874823, + 0.008358785882592201, + 0.005025280173867941, + 0.009405543096363544, + 0.0063852877356112, + 0.010666809044778347, + 0.0062262993305921555, + 0.0065907700918614864, + 0.007433444261550903, + 0.006923439912497997, + 0.008569743484258652, + 0.007183014880865812, + 0.008069236762821674, + 0.0070479088462889194, + 0.004968944471329451, + 0.004779329057782888, + 0.0073216622695326805, + 0.00414737593382597, + 0.002899451181292534, + 0.0029707560315728188, + 0.008009369485080242, + 0.003912243060767651, + 0.003025622572749853, + 0.004252828657627106, + 0.006115732714533806, + 0.007882566191256046, + 0.004625731147825718, + 0.005047550890594721, + 0.0068176728673279285, + 0.00596193689852953, + 0.0066379704512655735, + 0.0053172665648162365, + 0.006289117969572544, + 0.0028929514810442924, + 0.0028023701161146164, + 0.00386500614695251, + 0.0029318728484213352, + 0.0028989154379814863, + 0.002907703397795558, + 0.0029247147031128407, + 0.0025513572618365288, + 0.0066180359572172165, + 0.002601912012323737, + 0.008822833187878132, + 0.007527399808168411, + 0.005108751822263002, + 0.002797768684104085, + 0.0037400147411972284, + 0.004369767382740974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.19335079193115234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004938274621964 + ], + "opposite_logprob_tokens": [ + -2.695049285888672 + ], + "logprob_reward": -0.07004938274621964, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004938274621964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13403791189193726, + 0.2834862172603607, + 0.027611540630459785, + 0.06481736153364182, + 0.23174333572387695, + 0.06441688537597656, + 0.039762116968631744, + 0.032745759934186935, + 0.0350513830780983, + 0.08632756024599075 + ], + "credit_map": [ + 0.02015206217765808, + 0.014814326539635658, + 0.018925126641988754, + 0.015439321286976337, + 0.015379149466753006, + 0.014016208238899708, + 0.02385823428630829, + 0.0114534767344594, + 0.06675632297992706, + 0.014004016295075417, + 0.01250719465315342, + 0.10477609187364578, + 0.018406694754958153, + 0.007183866575360298, + 0.04965861514210701, + 0.01019339356571436, + 0.003591301152482629, + 0.003117953659966588, + 0.0029593799263238907, + 0.00270824134349823, + 0.004046075977385044, + 0.005105301737785339, + 0.0027619493193924427, + 0.003321337979286909, + 0.01017259806394577, + 0.004817300010472536, + 0.0057149408385157585, + 0.00655926251783967, + 0.005599783733487129, + 0.016315540298819542, + 0.007096287794411182, + 0.008541649207472801, + 0.04263170436024666, + 0.019386056810617447, + 0.0310690775513649, + 0.02456316351890564, + 0.030158428475260735, + 0.026114411652088165, + 0.02836134098470211, + 0.029459139332175255, + 0.009436988271772861, + 0.007257184479385614, + 0.0072005935944616795, + 0.007805972825735807, + 0.010485111735761166, + 0.008098212070763111, + 0.006117908749729395, + 0.008014914579689503, + 0.0044999029487371445, + 0.0058248452842235565, + 0.0028562808874994516, + 0.007891758345067501, + 0.005951583385467529, + 0.00271217105910182, + 0.005314943380653858, + 0.00471062958240509, + 0.0036470494233071804, + 0.0030213561840355396, + 0.0036180284805595875, + 0.006511426530778408, + 0.0031382963061332703, + 0.0027217925526201725, + 0.004689775872975588, + 0.0053980364464223385, + 0.003860487835481763, + 0.00411073025316, + 0.00685137091204524, + 0.004421956837177277, + 0.005686663556843996, + 0.002513255923986435, + 0.004629071801900864, + 0.002977845724672079, + 0.007469918113201857, + 0.005044837482273579, + 0.0043889558874070644, + 0.033965688198804855, + 0.01487884484231472, + 0.0028087443206459284, + 0.012003054842352867, + 0.005767523311078548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.07004938274621964 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337129682302475 + ], + "opposite_logprob_tokens": [ + -3.408712863922119 + ], + "logprob_reward": -0.0337129682302475, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337129682302475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2259509265422821, + 0.0712079182267189, + 0.08644255995750427, + 0.06583147495985031, + 0.15700529515743256, + 0.09611856937408447, + 0.14337708055973053, + 0.031127041205763817, + 0.03083653375506401, + 0.09210260212421417 + ], + "credit_map": [ + 0.0254252627491951, + 0.021536827087402344, + 0.027097778394818306, + 0.03797828406095505, + 0.026552259922027588, + 0.03138464316725731, + 0.028451308608055115, + 0.027524558827280998, + 0.014765030704438686, + 0.00543289165943861, + 0.0057662539184093475, + 0.01996181346476078, + 0.006926081143319607, + 0.003292517038062215, + 0.010132383555173874, + 0.004930948372930288, + 0.010609500110149384, + 0.013518097810447216, + 0.008148129098117352, + 0.010209104977548122, + 0.011581198312342167, + 0.0076189530082046986, + 0.011598784476518631, + 0.013158786110579967, + 0.008989227004349232, + 0.005139337386935949, + 0.007149877957999706, + 0.01017299760133028, + 0.008275967091321945, + 0.007906473241746426, + 0.00883991178125143, + 0.009357677772641182, + 0.028683334589004517, + 0.013602882623672485, + 0.020521916449069977, + 0.015287780202925205, + 0.02181326039135456, + 0.01758377067744732, + 0.022060392424464226, + 0.017451971769332886, + 0.010858045890927315, + 0.01419179979711771, + 0.011116049252450466, + 0.017368754372000694, + 0.016062825918197632, + 0.005555297248065472, + 0.01070952508598566, + 0.010256266221404076, + 0.01775738224387169, + 0.014462542720139027, + 0.007982471957802773, + 0.031515054404735565, + 0.014186367392539978, + 0.0088003845885396, + 0.015289624221622944, + 0.033383265137672424, + 0.005035012494772673, + 0.0031559860799461603, + 0.003507315181195736, + 0.004703226033598185, + 0.003146024886518717, + 0.004552898462861776, + 0.002873962512239814, + 0.004152616485953331, + 0.003456725273281336, + 0.003205787157639861, + 0.004312777891755104, + 0.004798592068254948, + 0.0032248271163553, + 0.0033010405022650957, + 0.003959351219236851, + 0.004577432759106159, + 0.005226238630712032, + 0.004873081110417843, + 0.007787011098116636, + 0.03041543997824192, + 0.01706693135201931, + 0.0032004062086343765, + 0.016199495643377304, + 0.007333994377404451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0337129682302475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741991266608238 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.006741991266608238, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741991266608238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06998667865991592, + 0.260796457529068, + 0.06378121674060822, + 0.14354413747787476, + 0.0696391761302948, + 0.0549100860953331, + 0.2435389906167984, + 0.04767175018787384, + 0.020553672686219215, + 0.025577859953045845 + ], + "credit_map": [ + 0.010601104237139225, + 0.007961198687553406, + 0.009015148505568504, + 0.010295328684151173, + 0.008045696653425694, + 0.008307525888085365, + 0.0088382912799716, + 0.006922387517988682, + 0.055699631571769714, + 0.022587841376662254, + 0.01892801746726036, + 0.07917699217796326, + 0.022053075954318047, + 0.011732056736946106, + 0.037138257175683975, + 0.013480598106980324, + 0.008731490932404995, + 0.009130429476499557, + 0.006156644318252802, + 0.00634445296600461, + 0.006165951956063509, + 0.009332260116934776, + 0.007238493300974369, + 0.01068149320781231, + 0.022077668458223343, + 0.009249485097825527, + 0.011970067396759987, + 0.01045301090925932, + 0.011302302591502666, + 0.04573585093021393, + 0.011788638308644295, + 0.020967114716768265, + 0.014045918360352516, + 0.004979133605957031, + 0.00922184158116579, + 0.005832433234900236, + 0.009845425374805927, + 0.009388677775859833, + 0.008212340995669365, + 0.00811340194195509, + 0.013766089454293251, + 0.0055085099302232265, + 0.006307868752628565, + 0.005397528409957886, + 0.003813702380284667, + 0.006659446284174919, + 0.004045426845550537, + 0.009411513805389404, + 0.03243625536561012, + 0.01270426157861948, + 0.015962565317749977, + 0.04245220497250557, + 0.0157731082290411, + 0.025340881198644638, + 0.041642360389232635, + 0.057227350771427155, + 0.004924909211695194, + 0.0035345996730029583, + 0.0057091680355370045, + 0.006214749533683062, + 0.004552439786493778, + 0.012549977749586105, + 0.004698937758803368, + 0.005486966110765934, + 0.0027899874839931726, + 0.002137652365490794, + 0.002485950943082571, + 0.002925815060734749, + 0.00186650559771806, + 0.0032214804086834192, + 0.002402943093329668, + 0.0027233369182795286, + 0.005022953264415264, + 0.001327955862507224, + 0.008614066988229752, + 0.0020993323996663094, + 0.0026480916421860456, + 0.0013121800730004907, + 0.0018779506208375096, + 0.0026753272395581007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.006741991266608238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934139132499695 + ], + "opposite_logprob_tokens": [ + -0.6934139132499695 + ], + "logprob_reward": -0.6934139132499695, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934139132499695, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14006122946739197, + 0.08292490243911743, + 0.058164555579423904, + 0.05578334629535675, + 0.07447338849306107, + 0.12858900427818298, + 0.07304716110229492, + 0.08114994317293167, + 0.05658956617116928, + 0.2492169737815857 + ], + "credit_map": [ + 0.017263462767004967, + 0.012884920462965965, + 0.015280663035809994, + 0.02297361008822918, + 0.017895901575684547, + 0.019336923956871033, + 0.021016636863350868, + 0.013409091159701347, + 0.013501963578164577, + 0.00682216277346015, + 0.012630200013518333, + 0.020692242309451103, + 0.00690712733194232, + 0.005616433918476105, + 0.011293139308691025, + 0.005461629014462233, + 0.006236471701413393, + 0.006820496637374163, + 0.006908444222062826, + 0.008177491836249828, + 0.008897879160940647, + 0.004318436607718468, + 0.008243507705628872, + 0.008561820723116398, + 0.006188637111335993, + 0.004118978977203369, + 0.008356931619346142, + 0.008045678026974201, + 0.004675743635743856, + 0.0104049788787961, + 0.006734419614076614, + 0.007257969118654728, + 0.013737747445702553, + 0.0051127527840435505, + 0.008966703899204731, + 0.008111073635518551, + 0.00903463363647461, + 0.00871029868721962, + 0.010763640515506268, + 0.010036531835794449, + 0.014731082133948803, + 0.014605767093598843, + 0.015525945462286472, + 0.01727643795311451, + 0.01952100358903408, + 0.01130360085517168, + 0.013591746799647808, + 0.02203340269625187, + 0.007955971173942089, + 0.006409831810742617, + 0.005638667847961187, + 0.01150595210492611, + 0.00723749166354537, + 0.006335547659546137, + 0.01994837261736393, + 0.008015322498977184, + 0.012959949672222137, + 0.007797217927873135, + 0.0070113264955580235, + 0.012107380665838718, + 0.009691801853477955, + 0.012127258814871311, + 0.008800690993666649, + 0.010654307901859283, + 0.007046559359878302, + 0.0061928085051476955, + 0.005889963824301958, + 0.007519074250012636, + 0.00648893229663372, + 0.004552409052848816, + 0.008424931205809116, + 0.01047488022595644, + 0.013918286189436913, + 0.009846300818026066, + 0.04090707004070282, + 0.1018235832452774, + 0.040275413542985916, + 0.004366961307823658, + 0.029702289029955864, + 0.00837703887373209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.6934139132499695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "Inspired by traditional Japanese aesthetics, a low-angle, full-shot shot slowly pans forward, focusing on a lone figure standing defiantly amidst a catastrophic urban flood. The figure is adorned with the ornate Sacra Umbra Armor, a masterpiece of darkly painted metal plates intricately connected by deep blue silk thread. The helmet, a formidable kabuto, features a menacing, full-face man-yorai mask that conceals all features, its metallic scowl fixed against the floodwaters. Rain pours down like a thick downpour, flowing over the armor's polished, obsidian-like surface, beading like tiny gems before cascading into the churning waters below. The figure stands waist-deep in raging, murky water, which swirls powerfully around their legs. In the background, the submerged cityscape is a chaotic mix of modern high-rises and traditional buildings, their lower halves engulfed by the floodwaters. The entire scene is lit almost entirely from below by the lights of underwater buildings and neon signs, which flicker ominously beneath the turbulent surface, casting a ghostly, refracted glow of electric blues and garish reds. This otherworldly underwater light creates deep, dramatic shadows that stretch and twist across the rising waters, painting a scene of somber, apocalyptic grandeur, while the figure remains completely motionless, a stoic fortress of shadow and steel against the raging storm.", + "seed": 19672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_0/index_3190_seed_19672_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17244424363598226, + "vlm_logprob": -0.17244424363598226, + "vlm_yes_no_margin": 2.674999976158142, + "vlm_token_credit": -0.17244424363598226, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1364504098892212 + ], + "opposite_logprob_tokens": [ + -2.1364502906799316 + ], + "logprob_reward": -0.1364504098892212, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1364504098892212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13712044060230255, + 0.0821613147854805, + 0.14290675520896912, + 0.12820512056350708, + 0.04783663526177406, + 0.10311128944158554, + 0.04416988790035248, + 0.12989644706249237, + 0.06684884428977966, + 0.11774332076311111 + ], + "credit_map": [ + 0.021252818405628204, + 0.012333903461694717, + 0.021517492830753326, + 0.012072530575096607, + 0.016156243160367012, + 0.01731872372329235, + 0.01961352489888668, + 0.016855191439390182, + 0.009801346808671951, + 0.004974946845322847, + 0.00884955283254385, + 0.012386015616357327, + 0.007587448228150606, + 0.014636322855949402, + 0.008844959549605846, + 0.01508072204887867, + 0.01773805171251297, + 0.019839294254779816, + 0.013648652471601963, + 0.01894981786608696, + 0.02178766019642353, + 0.012921063229441643, + 0.02289375104010105, + 0.015128460712730885, + 0.012758591212332249, + 0.013528553768992424, + 0.01741180196404457, + 0.019295670092105865, + 0.01945587433874607, + 0.012698240578174591, + 0.016984185203909874, + 0.01607220061123371, + 0.009838564321398735, + 0.004747208673506975, + 0.004191357642412186, + 0.003716409206390381, + 0.004672161769121885, + 0.008427618071436882, + 0.005367717240005732, + 0.0068755969405174255, + 0.014265860430896282, + 0.014588533900678158, + 0.010557430796325207, + 0.016349975019693375, + 0.01300422940403223, + 0.010064219124615192, + 0.013855816796422005, + 0.010425207205116749, + 0.008025540970265865, + 0.0054151201620697975, + 0.004780481103807688, + 0.004550690297037363, + 0.007311611901968718, + 0.0035911318846046925, + 0.006480752490460873, + 0.004014556296169758, + 0.01376430969685316, + 0.017882132902741432, + 0.016882434487342834, + 0.015228517353534698, + 0.017458468675613403, + 0.022178931161761284, + 0.008792771957814693, + 0.017708860337734222, + 0.014665425755083561, + 0.005139736458659172, + 0.0076659866608679295, + 0.008498442359268665, + 0.008635209873318672, + 0.005721856374293566, + 0.009309976361691952, + 0.007212200202047825, + 0.014706437475979328, + 0.01969187892973423, + 0.011379698291420937, + 0.018579082563519478, + 0.019438011571764946, + 0.00912775844335556, + 0.01423819363117218, + 0.010582252405583858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.1364504098892212 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308418184518814 + ], + "opposite_logprob_tokens": [ + -3.168084144592285 + ], + "logprob_reward": -0.04308418184518814, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308418184518814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08317072689533234, + 0.08324843645095825, + 0.050967734307050705, + 0.09979330003261566, + 0.06308208405971527, + 0.08004686236381531, + 0.06587062776088715, + 0.24922534823417664, + 0.15885910391807556, + 0.06573574244976044 + ], + "credit_map": [ + 0.012245351448655128, + 0.007413109298795462, + 0.012863637879490852, + 0.010799732059240341, + 0.00934768095612526, + 0.010434870608150959, + 0.010714533738791943, + 0.009351814165711403, + 0.013285253196954727, + 0.007944407872855663, + 0.010545086115598679, + 0.009596762247383595, + 0.011307599022984505, + 0.012055633589625359, + 0.008216781541705132, + 0.010296914726495743, + 0.0042867292650043964, + 0.00840162206441164, + 0.00518173910677433, + 0.006230814848095179, + 0.008586511015892029, + 0.0047391075640916824, + 0.0065464358776807785, + 0.006994780618697405, + 0.012506994418799877, + 0.010778145864605904, + 0.012145969085395336, + 0.011715246364474297, + 0.012793431058526039, + 0.009093060158193111, + 0.013836388476192951, + 0.016924073919653893, + 0.00768365990370512, + 0.004971526097506285, + 0.009857699275016785, + 0.004924208391457796, + 0.005428513512015343, + 0.014094271697103977, + 0.006778208073228598, + 0.009344001300632954, + 0.010205020196735859, + 0.010684440843760967, + 0.011756493709981441, + 0.007868828251957893, + 0.012356514111161232, + 0.008773659355938435, + 0.008496662601828575, + 0.009905246086418629, + 0.00610458105802536, + 0.008193830959498882, + 0.004458037205040455, + 0.0062771206721663475, + 0.010443217121064663, + 0.011298500932753086, + 0.014595207758247852, + 0.004500140901654959, + 0.0615507997572422, + 0.018679067492485046, + 0.02728969044983387, + 0.010527275502681732, + 0.012734642252326012, + 0.07756051421165466, + 0.011381598189473152, + 0.029501786455512047, + 0.04060228168964386, + 0.015492157079279423, + 0.014243154786527157, + 0.015782145783305168, + 0.016796784475445747, + 0.011130855418741703, + 0.03263888508081436, + 0.01217285729944706, + 0.00730448542162776, + 0.009054522961378098, + 0.009135515429079533, + 0.008340723812580109, + 0.006646633613854647, + 0.0068481555208563805, + 0.008448846638202667, + 0.009956860914826393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.04308418184518814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430704802274704 + ], + "opposite_logprob_tokens": [ + -3.1680705547332764 + ], + "logprob_reward": -0.0430704802274704, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430704802274704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08495673537254333, + 0.07799581438302994, + 0.1677151322364807, + 0.1263033002614975, + 0.09530951827764511, + 0.10114774107933044, + 0.03279661387205124, + 0.06586917489767075, + 0.08107344806194305, + 0.1668325513601303 + ], + "credit_map": [ + 0.013804321177303791, + 0.00888330489397049, + 0.011303246952593327, + 0.013922336511313915, + 0.010051809251308441, + 0.010097366757690907, + 0.008106356486678123, + 0.008787993341684341, + 0.011947103776037693, + 0.005981617607176304, + 0.007189286407083273, + 0.0070119453594088554, + 0.011604053899645805, + 0.013093280605971813, + 0.006493404507637024, + 0.014675122685730457, + 0.021840276196599007, + 0.02324678562581539, + 0.016132373362779617, + 0.022468404844403267, + 0.021940836682915688, + 0.012555700726807117, + 0.026063980534672737, + 0.023466791957616806, + 0.016600819304585457, + 0.011866722255945206, + 0.019551843404769897, + 0.013569067232310772, + 0.015231085941195488, + 0.015438762493431568, + 0.02108049765229225, + 0.01296451035887003, + 0.011643987149000168, + 0.009734878316521645, + 0.011159145273268223, + 0.008673091419041157, + 0.011128636077046394, + 0.015593807213008404, + 0.015697427093982697, + 0.011678547598421574, + 0.01242996659129858, + 0.010945126414299011, + 0.013936375267803669, + 0.011709644459187984, + 0.01287081092596054, + 0.014565219171345234, + 0.011901271529495716, + 0.012789326719939709, + 0.007378090638667345, + 0.003423322457820177, + 0.0036130305379629135, + 0.002657586243003607, + 0.003135895123705268, + 0.006175057962536812, + 0.0032204610761255026, + 0.0031931702978909016, + 0.008305884897708893, + 0.010086474008858204, + 0.007616992574185133, + 0.006075893994420767, + 0.006639698054641485, + 0.012383109889924526, + 0.004835288971662521, + 0.009925834834575653, + 0.023402009159326553, + 0.007065145764499903, + 0.006489974446594715, + 0.005548988934606314, + 0.00922451727092266, + 0.006151922978460789, + 0.014253156259655952, + 0.008937730453908443, + 0.01654445379972458, + 0.027802858501672745, + 0.01601160131394863, + 0.019124334678053856, + 0.02146170102059841, + 0.020836789160966873, + 0.02141469344496727, + 0.023636111989617348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.0430704802274704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067384387366473675 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.0067384387366473675, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067384387366473675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16192498803138733, + 0.14385417103767395, + 0.11727343499660492, + 0.054955702275037766, + 0.09341365844011307, + 0.11306897550821304, + 0.05947819724678993, + 0.10984566062688828, + 0.09760779142379761, + 0.04857737943530083 + ], + "credit_map": [ + 0.017336560413241386, + 0.015733689069747925, + 0.028500786051154137, + 0.01670290157198906, + 0.021599939092993736, + 0.020970286801457405, + 0.023276712745428085, + 0.01780412346124649, + 0.020475368946790695, + 0.011036721058189869, + 0.020986726507544518, + 0.016688887029886246, + 0.016812138259410858, + 0.01945267990231514, + 0.015413250774145126, + 0.02298840321600437, + 0.019312281161546707, + 0.017418893054127693, + 0.01234408002346754, + 0.015145375393331051, + 0.017324673011898994, + 0.00935089960694313, + 0.016312899067997932, + 0.010064343921840191, + 0.009884851984679699, + 0.005624859593808651, + 0.006847307085990906, + 0.00665305694565177, + 0.006662784609943628, + 0.004795391578227282, + 0.0075195543467998505, + 0.006967898458242416, + 0.014895683154463768, + 0.008504150435328484, + 0.01600276120007038, + 0.009634542278945446, + 0.00956636667251587, + 0.010988032445311546, + 0.012096086516976357, + 0.01172604225575924, + 0.014482151716947556, + 0.013978359289467335, + 0.018506186082959175, + 0.01112359669059515, + 0.015757005661725998, + 0.012205341830849648, + 0.012291434220969677, + 0.014724903739988804, + 0.008377416990697384, + 0.0068832868710160255, + 0.005873516201972961, + 0.005007016472518444, + 0.008810136467218399, + 0.00739444512873888, + 0.01032885629683733, + 0.006803527940064669, + 0.014176055788993835, + 0.010749961249530315, + 0.02568119391798973, + 0.012915881350636482, + 0.008374202996492386, + 0.01360640861093998, + 0.010273149237036705, + 0.014068815857172012, + 0.013084722682833672, + 0.008761171251535416, + 0.010747280903160572, + 0.013409078121185303, + 0.01153541263192892, + 0.014925767667591572, + 0.013708505779504776, + 0.01143586728721857, + 0.009579779580235481, + 0.005293631926178932, + 0.007389215286821127, + 0.005834538955241442, + 0.004054339602589607, + 0.004840536508709192, + 0.005995639134198427, + 0.00558969983831048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0067384387366473675 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328777074813843 + ], + "opposite_logprob_tokens": [ + -0.7578777074813843 + ], + "logprob_reward": -0.6328777074813843, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328777074813843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20750553905963898, + 0.0879141241312027, + 0.07967286556959152, + 0.09130626171827316, + 0.0845155194401741, + 0.1589432656764984, + 0.0762714073061943, + 0.13742771744728088, + 0.02996893599629402, + 0.046474311500787735 + ], + "credit_map": [ + 0.03065815381705761, + 0.012169887311756611, + 0.052904434502124786, + 0.022884629666805267, + 0.015306773595511913, + 0.025898749008774757, + 0.02471351809799671, + 0.02296941913664341, + 0.010916654020547867, + 0.008054346777498722, + 0.009601950645446777, + 0.008050639182329178, + 0.016514567658305168, + 0.016237271949648857, + 0.007696707267314196, + 0.01084198895841837, + 0.017487116158008575, + 0.008351136930286884, + 0.00802492257207632, + 0.00718305166810751, + 0.010213657282292843, + 0.006870407145470381, + 0.01208735816180706, + 0.009455208666622639, + 0.010889533907175064, + 0.009617653675377369, + 0.009431323036551476, + 0.0103837875649333, + 0.01111980713903904, + 0.01009750459343195, + 0.016062818467617035, + 0.01370383519679308, + 0.016189567744731903, + 0.0077059464529156685, + 0.013983377255499363, + 0.00909700058400631, + 0.006769397296011448, + 0.012779437005519867, + 0.009915915317833424, + 0.008074874058365822, + 0.02121574617922306, + 0.019923076033592224, + 0.01920315995812416, + 0.01980554684996605, + 0.022507688030600548, + 0.019018463790416718, + 0.0177958644926548, + 0.01947372779250145, + 0.017237836495041847, + 0.00665629468858242, + 0.007031054235994816, + 0.004555838648229837, + 0.006575349252671003, + 0.018756521865725517, + 0.007317500188946724, + 0.008141008205711842, + 0.022786201909184456, + 0.01570400968194008, + 0.020898737013339996, + 0.012718011625111103, + 0.012424646876752377, + 0.021370012313127518, + 0.011014856398105621, + 0.020511241629719734, + 0.00707087991759181, + 0.002454602625221014, + 0.0036428472958505154, + 0.002820380963385105, + 0.003590718377381563, + 0.00299058691598475, + 0.003945638425648212, + 0.003453281708061695, + 0.005357939749956131, + 0.005116621032357216, + 0.004328114911913872, + 0.006984181236475706, + 0.004336016718298197, + 0.00718720443546772, + 0.005198221188038588, + 0.007966017350554466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.6328777074813843 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Inspired by traditional Japanese aesthetics, a low-angle, full-shot shot slowly pans forward, focusing on a lone figure standing defiantly amidst a catastrophic urban flood. The figure is adorned with the ornate Sacra Umbra Armor, a masterpiece of darkly painted metal plates intricately connected by deep blue silk thread. The helmet, a formidable kabuto, features a menacing, full-face man-yorai mask that conceals all features, its metallic scowl fixed against the floodwaters. Rain pours down like a thick downpour, flowing over the armor's polished, obsidian-like surface, beading like tiny gems before cascading into the churning waters below. The figure stands waist-deep in raging, murky water, which swirls powerfully around their legs. In the background, the submerged cityscape is a chaotic mix of modern high-rises and traditional buildings, their lower halves engulfed by the floodwaters. The entire scene is lit almost entirely from below by the lights of underwater buildings and neon signs, which flicker ominously beneath the turbulent surface, casting a ghostly, refracted glow of electric blues and garish reds. This otherworldly underwater light creates deep, dramatic shadows that stretch and twist across the rising waters, painting a scene of somber, apocalyptic grandeur, while the figure remains completely motionless, a stoic fortress of shadow and steel against the raging storm.", + "seed": 19673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_0/index_3190_seed_19673_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11716921441257, + "vlm_logprob": -0.11716921441257, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.11716921441257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058394916355609894 + ], + "opposite_logprob_tokens": [ + -3.0583949089050293 + ], + "logprob_reward": -0.058394916355609894, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.058394916355609894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10803531110286713, + 0.05139068514108658, + 0.05272592604160309, + 0.230647012591362, + 0.037651754915714264, + 0.10273610800504684, + 0.06363756209611893, + 0.22129730880260468, + 0.04300645366311073, + 0.08887185156345367 + ], + "credit_map": [ + 0.011685525998473167, + 0.008687138557434082, + 0.017131414264440536, + 0.01852591522037983, + 0.009532410651445389, + 0.01500170398503542, + 0.015021265484392643, + 0.012449948117136955, + 0.004835417959839106, + 0.0034165605902671814, + 0.00890465546399355, + 0.006489371415227652, + 0.004175580572336912, + 0.006317907944321632, + 0.013720115646719933, + 0.003531081136316061, + 0.0055288635194301605, + 0.0041052429005503654, + 0.006971665192395449, + 0.005769488401710987, + 0.005853711161762476, + 0.009661546908318996, + 0.00704985624179244, + 0.007785556372255087, + 0.02077925205230713, + 0.009501914493739605, + 0.056673455983400345, + 0.016454942524433136, + 0.016827724874019623, + 0.019664371386170387, + 0.07144293189048767, + 0.01930244080722332, + 0.00299642700701952, + 0.003217488992959261, + 0.004139505326747894, + 0.005294838920235634, + 0.006070347502827644, + 0.0064492071978747845, + 0.005708533339202404, + 0.0037754138465970755, + 0.00929636787623167, + 0.01275451760739088, + 0.013067175634205341, + 0.011610573157668114, + 0.018251478672027588, + 0.013248085975646973, + 0.014108126051723957, + 0.010399787686765194, + 0.01018394436687231, + 0.008703191764652729, + 0.013978248462080956, + 0.005702832248061895, + 0.006140639074146748, + 0.005623589735478163, + 0.008480011485517025, + 0.004825108218938112, + 0.054626598954200745, + 0.016046635806560516, + 0.09386030584573746, + 0.015375103801488876, + 0.005555425770580769, + 0.01369545329362154, + 0.014154261909425259, + 0.007983554154634476, + 0.005718292202800512, + 0.005910045001655817, + 0.004865836352109909, + 0.004167688079178333, + 0.006362785119563341, + 0.006874800659716129, + 0.0061423154547810555, + 0.0029646968469023705, + 0.02427665889263153, + 0.012310267426073551, + 0.02364167757332325, + 0.006921431981027126, + 0.0041507696732878685, + 0.004219117574393749, + 0.008508951403200626, + 0.004842990078032017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.058394916355609894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867760092020035 + ], + "opposite_logprob_tokens": [ + -3.048677682876587 + ], + "logprob_reward": -0.04867760092020035, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867760092020035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06611207872629166, + 0.033735327422618866, + 0.034267034381628036, + 0.030337316915392876, + 0.014431031420826912, + 0.03850472718477249, + 0.0742570236325264, + 0.6109097003936768, + 0.05877313017845154, + 0.03867260366678238 + ], + "credit_map": [ + 0.006597302388399839, + 0.006187438033521175, + 0.010966804809868336, + 0.009162689559161663, + 0.005203113425523043, + 0.008264956064522266, + 0.012768933549523354, + 0.006960835307836533, + 0.0029427630361169577, + 0.0028481981717050076, + 0.006566149182617664, + 0.006788535509258509, + 0.0035270238295197487, + 0.0036749974824488163, + 0.004550279583781958, + 0.002837381325662136, + 0.004407026804983616, + 0.0032755271531641483, + 0.004201493225991726, + 0.002890565898269415, + 0.002995629096403718, + 0.006637231912463903, + 0.006129487417638302, + 0.0037300754338502884, + 0.002998908283188939, + 0.002574794227257371, + 0.0036403711419552565, + 0.0027633572462946177, + 0.004209561739116907, + 0.004629916045814753, + 0.005199081730097532, + 0.004321326036006212, + 0.0014360934728756547, + 0.001429967232979834, + 0.0012999046593904495, + 0.0014321926282718778, + 0.0011867082212120295, + 0.0033964074682444334, + 0.002453566761687398, + 0.0017961907433345914, + 0.0035128507297486067, + 0.004324404522776604, + 0.005145132541656494, + 0.003945507109165192, + 0.006247567944228649, + 0.005753025412559509, + 0.005559392273426056, + 0.00401684595271945, + 0.008281267248094082, + 0.005542519502341747, + 0.026378067210316658, + 0.0050096516497433186, + 0.0066389404237270355, + 0.004626826848834753, + 0.009579310193657875, + 0.008200442418456078, + 0.15068167448043823, + 0.047694943845272064, + 0.2614636719226837, + 0.044900596141815186, + 0.013628079555928707, + 0.0358014740049839, + 0.038473740220069885, + 0.01826554536819458, + 0.007577891461551189, + 0.008648536168038845, + 0.0030210812110453844, + 0.002558139618486166, + 0.003170114941895008, + 0.02427830547094345, + 0.005643490701913834, + 0.003875574329867959, + 0.005171970464289188, + 0.005015971139073372, + 0.00753692677244544, + 0.004999437835067511, + 0.002863741246983409, + 0.005024903919547796, + 0.004553520120680332, + 0.0035061321686953306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.04867760092020035 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307687282562256 + ], + "opposite_logprob_tokens": [ + -3.168076992034912 + ], + "logprob_reward": -0.04307687282562256, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307687282562256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1323806196451187, + 0.038750939071178436, + 0.05518680810928345, + 0.3216300308704376, + 0.04274732246994972, + 0.0592164620757103, + 0.1084698811173439, + 0.06711586564779282, + 0.0599212646484375, + 0.11458076536655426 + ], + "credit_map": [ + 0.009602006524801254, + 0.01148990262299776, + 0.021874353289604187, + 0.02318224497139454, + 0.013121142983436584, + 0.014736322686076164, + 0.01921280287206173, + 0.019161850214004517, + 0.0033256825990974903, + 0.0032186105381697416, + 0.007320135831832886, + 0.006708652712404728, + 0.004889691714197397, + 0.004758008290082216, + 0.0052603366784751415, + 0.003269826527684927, + 0.007238252088427544, + 0.004601937718689442, + 0.005860954523086548, + 0.005266967229545116, + 0.0056837210431694984, + 0.011399980634450912, + 0.008800112642347813, + 0.006334884092211723, + 0.03267212212085724, + 0.013495689257979393, + 0.07205484807491302, + 0.018127892166376114, + 0.02318701334297657, + 0.024753142148256302, + 0.11240214854478836, + 0.024937167763710022, + 0.0039281658828258514, + 0.0029512585606426, + 0.0038512898609042168, + 0.004767179023474455, + 0.006636025384068489, + 0.008378850296139717, + 0.008263559080660343, + 0.003970995545387268, + 0.007054971065372229, + 0.005324156489223242, + 0.006333702243864536, + 0.004215438850224018, + 0.008833340369164944, + 0.010229753330349922, + 0.010502085089683533, + 0.006723020225763321, + 0.00920046679675579, + 0.00962380412966013, + 0.019304297864437103, + 0.012417175807058811, + 0.015189279802143574, + 0.008765598759055138, + 0.017627066001296043, + 0.016342194750905037, + 0.00806462112814188, + 0.007006797473877668, + 0.00919065997004509, + 0.009822018444538116, + 0.006668430287390947, + 0.008439375087618828, + 0.01148346159607172, + 0.006440506782382727, + 0.006775532383471727, + 0.007064979523420334, + 0.0062507884576916695, + 0.004623257089406252, + 0.010148770175874233, + 0.012715413235127926, + 0.0055518909357488155, + 0.006790634710341692, + 0.016276109963655472, + 0.010972479358315468, + 0.016578039154410362, + 0.012817643582820892, + 0.009174616076052189, + 0.01921180449426174, + 0.01626361347734928, + 0.013286462053656578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.04307687282562256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741991266608238 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.006741991266608238, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741991266608238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0620453916490078, + 0.03588113188743591, + 0.016805564984679222, + 0.13035771250724792, + 0.07911920547485352, + 0.0558907575905323, + 0.03497472032904625, + 0.502388060092926, + 0.05556418001651764, + 0.026973221451044083 + ], + "credit_map": [ + 0.0058536408469080925, + 0.006295232567936182, + 0.009088771417737007, + 0.010160611011087894, + 0.007763609755784273, + 0.007289322093129158, + 0.007797906640917063, + 0.007796303369104862, + 0.004787002224475145, + 0.0031374660320580006, + 0.006143278907984495, + 0.0036971657536923885, + 0.0030889171175658703, + 0.005511647555977106, + 0.0064626517705619335, + 0.0030530074145644903, + 0.0019857280422002077, + 0.0014896824723109603, + 0.002542198868468404, + 0.0017624336760491133, + 0.0023212756495922804, + 0.0025602965615689754, + 0.0020778325852006674, + 0.0020661167800426483, + 0.01210477389395237, + 0.005393407307565212, + 0.030701184645295143, + 0.007591493893414736, + 0.009302198886871338, + 0.011008378118276596, + 0.043742064386606216, + 0.010514228604733944, + 0.006753482390195131, + 0.008442942053079605, + 0.008642909117043018, + 0.006039412226527929, + 0.007850497029721737, + 0.019512450322508812, + 0.013913137838244438, + 0.007964382879436016, + 0.00905068963766098, + 0.011040245182812214, + 0.0053283148445189, + 0.0062518250197172165, + 0.0067579736933112144, + 0.0050132847391068935, + 0.007728817407041788, + 0.004719613119959831, + 0.005749604664742947, + 0.004424239508807659, + 0.008101378567516804, + 0.003497937461361289, + 0.0025832010433077812, + 0.0023260118905454874, + 0.004762803670018911, + 0.0035295458510518074, + 0.12990447878837585, + 0.036805350333452225, + 0.21211810410022736, + 0.03545761853456497, + 0.011634180322289467, + 0.031242046505212784, + 0.030920548364520073, + 0.014305749908089638, + 0.007716923486441374, + 0.006956532597541809, + 0.0061684479005634785, + 0.002659614197909832, + 0.0030927243642508984, + 0.020150717347860336, + 0.00487037468701601, + 0.003948845434933901, + 0.0037575673777610064, + 0.0038735568523406982, + 0.00359613006003201, + 0.0033446482848376036, + 0.0028892718255519867, + 0.0037317578680813313, + 0.0030862221028655767, + 0.002694070106372237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.006741991266608238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42895469069480896 + ], + "opposite_logprob_tokens": [ + -1.0539547204971313 + ], + "logprob_reward": -0.42895469069480896, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42895469069480896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15775205194950104, + 0.0582563653588295, + 0.04870845749974251, + 0.20625504851341248, + 0.039952170103788376, + 0.08568526804447174, + 0.057785533368587494, + 0.18461741507053375, + 0.07336485385894775, + 0.08762279897928238 + ], + "credit_map": [ + 0.01464101206511259, + 0.015288322232663631, + 0.020905110985040665, + 0.02748715877532959, + 0.015827948227524757, + 0.01823587715625763, + 0.027079585939645767, + 0.01828702725470066, + 0.00536577682942152, + 0.004704564344137907, + 0.011290821246802807, + 0.010994595475494862, + 0.0068950955756008625, + 0.007205953821539879, + 0.00650096544995904, + 0.005298591684550047, + 0.007062958087772131, + 0.005576012656092644, + 0.00533366808667779, + 0.004326959606260061, + 0.008036387152969837, + 0.00748244347050786, + 0.006616348400712013, + 0.004273678176105022, + 0.021197959780693054, + 0.008720997720956802, + 0.05025235563516617, + 0.010868010111153126, + 0.014543730765581131, + 0.013420440256595612, + 0.07117688655853271, + 0.01607467792928219, + 0.004005488008260727, + 0.0047026388347148895, + 0.006943896878510714, + 0.003788198344409466, + 0.005821154452860355, + 0.005090219900012016, + 0.00724793178960681, + 0.002352640498429537, + 0.008258067071437836, + 0.006664122920483351, + 0.011993884108960629, + 0.008036583662033081, + 0.015713414177298546, + 0.011432832106947899, + 0.01558234728872776, + 0.008004019036889076, + 0.00850771926343441, + 0.008482798933982849, + 0.012235126458108425, + 0.003978894557803869, + 0.0062956735491752625, + 0.006091231480240822, + 0.006246906705200672, + 0.0059471819549798965, + 0.04680527746677399, + 0.01289017777889967, + 0.07152566313743591, + 0.012672359123826027, + 0.007031619548797607, + 0.013219250366091728, + 0.012639779597520828, + 0.007833285257220268, + 0.011556337587535381, + 0.0094783129170537, + 0.006560799200087786, + 0.00467647984623909, + 0.010064090602099895, + 0.01858065277338028, + 0.006837877910584211, + 0.005610301624983549, + 0.026450857520103455, + 0.00839581061154604, + 0.02508488856256008, + 0.004406529013067484, + 0.0040922160260379314, + 0.00611129542812705, + 0.008384696207940578, + 0.004696506075561047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.42895469069480896 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Inspired by traditional Japanese aesthetics, a low-angle, full-shot shot slowly pans forward, focusing on a lone figure standing defiantly amidst a catastrophic urban flood. The figure is adorned with the ornate Sacra Umbra Armor, a masterpiece of darkly painted metal plates intricately connected by deep blue silk thread. The helmet, a formidable kabuto, features a menacing, full-face man-yorai mask that conceals all features, its metallic scowl fixed against the floodwaters. Rain pours down like a thick downpour, flowing over the armor's polished, obsidian-like surface, beading like tiny gems before cascading into the churning waters below. The figure stands waist-deep in raging, murky water, which swirls powerfully around their legs. In the background, the submerged cityscape is a chaotic mix of modern high-rises and traditional buildings, their lower halves engulfed by the floodwaters. The entire scene is lit almost entirely from below by the lights of underwater buildings and neon signs, which flicker ominously beneath the turbulent surface, casting a ghostly, refracted glow of electric blues and garish reds. This otherworldly underwater light creates deep, dramatic shadows that stretch and twist across the rising waters, painting a scene of somber, apocalyptic grandeur, while the figure remains completely motionless, a stoic fortress of shadow and steel against the raging storm.", + "seed": 19674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_0/index_3190_seed_19674_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16019341703504325, + "vlm_logprob": -0.16019341703504325, + "vlm_yes_no_margin": 2.700000011920929, + "vlm_token_credit": -0.16019341703504325, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3714328706264496 + ], + "opposite_logprob_tokens": [ + -1.246432900428772 + ], + "logprob_reward": -0.3714328706264496, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3714328706264496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24672113358974457, + 0.09094924479722977, + 0.046955808997154236, + 0.318923681974411, + 0.045012474060058594, + 0.05831429734826088, + 0.08871853351593018, + 0.044350456446409225, + 0.04211300611495972, + 0.017941342666745186 + ], + "credit_map": [ + 0.01847456954419613, + 0.01789298839867115, + 0.056654512882232666, + 0.025739701464772224, + 0.025486120954155922, + 0.013955526053905487, + 0.06728462874889374, + 0.02123306132853031, + 0.013630039989948273, + 0.009289385750889778, + 0.009768097661435604, + 0.009856536984443665, + 0.01576278917491436, + 0.0082708690315485, + 0.011960445903241634, + 0.01241107564419508, + 0.004431562032550573, + 0.005779898259788752, + 0.005968961399048567, + 0.00795671809464693, + 0.007079415488988161, + 0.005813338328152895, + 0.006622567307204008, + 0.0033033457584679127, + 0.05292109027504921, + 0.010982152074575424, + 0.03595660626888275, + 0.03515511006116867, + 0.03054642304778099, + 0.010429552756249905, + 0.1237964928150177, + 0.01913623698055744, + 0.005657189525663853, + 0.00505770044401288, + 0.0052078948356211185, + 0.0065362900495529175, + 0.0074631101451814175, + 0.0028497448656708, + 0.006136764772236347, + 0.0061037736013531685, + 0.012100103311240673, + 0.003017596434801817, + 0.005738799925893545, + 0.013476432301104069, + 0.0031074974685907364, + 0.005313380155712366, + 0.008067451417446136, + 0.007493028417229652, + 0.006913335062563419, + 0.007131906226277351, + 0.005952071864157915, + 0.023274628445506096, + 0.008194554597139359, + 0.004881375469267368, + 0.01432632002979517, + 0.018044333904981613, + 0.00614801375195384, + 0.0042740823701024055, + 0.0052465577609837055, + 0.006186481565237045, + 0.005140156950801611, + 0.004818316549062729, + 0.00688396580517292, + 0.005652876105159521, + 0.003805832238867879, + 0.003538725432008505, + 0.003971187863498926, + 0.009427236393094063, + 0.005632041022181511, + 0.00337773934006691, + 0.008729720488190651, + 0.003630518214777112, + 0.0015347623266279697, + 0.001338177127763629, + 0.0019064850639551878, + 0.002691970206797123, + 0.0026504367124289274, + 0.0015217335894703865, + 0.0032160459086298943, + 0.0030817296355962753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.3714328706264496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996270686388016 + ], + "opposite_logprob_tokens": [ + -2.9299962520599365 + ], + "logprob_reward": -0.054996270686388016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996270686388016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2831633388996124, + 0.11971218883991241, + 0.08986882865428925, + 0.25269436836242676, + 0.03528991714119911, + 0.03701987862586975, + 0.040858056396245956, + 0.034845635294914246, + 0.08040646463632584, + 0.026141304522752762 + ], + "credit_map": [ + 0.03473073989152908, + 0.03326695039868355, + 0.05324571952223778, + 0.026370074599981308, + 0.028700273483991623, + 0.030202200636267662, + 0.040760453790426254, + 0.03588693588972092, + 0.018758414313197136, + 0.01281032245606184, + 0.01610851101577282, + 0.012677628546953201, + 0.019041629508137703, + 0.011310417205095291, + 0.014574479311704636, + 0.014430788345634937, + 0.012599458917975426, + 0.006839555688202381, + 0.011282230727374554, + 0.010994570329785347, + 0.010668805800378323, + 0.01340014860033989, + 0.012253730557858944, + 0.011830325238406658, + 0.05183882638812065, + 0.016157792881131172, + 0.08083410561084747, + 0.019463125616312027, + 0.015596036799252033, + 0.016245003789663315, + 0.02752586454153061, + 0.0250336155295372, + 0.006631292402744293, + 0.002467224607244134, + 0.003601487958803773, + 0.0027589977253228426, + 0.0030621143523603678, + 0.0024125482887029648, + 0.004212840460240841, + 0.010143411345779896, + 0.007719854358583689, + 0.003548723179847002, + 0.0038799901958554983, + 0.004755549132823944, + 0.0036071117501705885, + 0.0033442568965256214, + 0.004657691344618797, + 0.005506702233105898, + 0.005174489226192236, + 0.003175623482093215, + 0.003058187197893858, + 0.008171114139258862, + 0.0032065678387880325, + 0.0032062458340078592, + 0.005998515523970127, + 0.008867315948009491, + 0.00436379062011838, + 0.0031123575754463673, + 0.0030656661838293076, + 0.006903072819113731, + 0.0051459018141031265, + 0.0024876457173377275, + 0.006104838103055954, + 0.003662361763417721, + 0.003872133558616042, + 0.005324612837284803, + 0.006078760139644146, + 0.024807997047901154, + 0.004240597132593393, + 0.004138058051466942, + 0.012556355446577072, + 0.019387952983379364, + 0.0034796427935361862, + 0.002170367632061243, + 0.004377457778900862, + 0.0038330331444740295, + 0.002056818688288331, + 0.0034773272927850485, + 0.003172046970576048, + 0.003574610687792301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.054996270686388016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198542296886444 + ], + "opposite_logprob_tokens": [ + -4.018198490142822 + ], + "logprob_reward": -0.018198542296886444, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198542296886444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18865832686424255, + 0.11547794193029404, + 0.023263918235898018, + 0.11675091087818146, + 0.17393028736114502, + 0.16951484978199005, + 0.022154245525598526, + 0.12150833010673523, + 0.03940001502633095, + 0.029341107234358788 + ], + "credit_map": [ + 0.016902394592761993, + 0.007693345192819834, + 0.03901445493102074, + 0.022649750113487244, + 0.018849076703190804, + 0.006002154666930437, + 0.0630950927734375, + 0.014452074654400349, + 0.020282559096813202, + 0.013503532856702805, + 0.013079380616545677, + 0.011104107834398746, + 0.015135582536458969, + 0.013222766108810902, + 0.01345203910022974, + 0.015697984024882317, + 0.0031221159733831882, + 0.0021938951686024666, + 0.0030000205151736736, + 0.002515882020816207, + 0.0029815947636961937, + 0.003321785945445299, + 0.003925309516489506, + 0.0022033152636140585, + 0.01972045935690403, + 0.010314028710126877, + 0.02260744571685791, + 0.011605641804635525, + 0.012862010858952999, + 0.013143572025001049, + 0.015080765821039677, + 0.011416995897889137, + 0.043648749589920044, + 0.011123908683657646, + 0.01375033613294363, + 0.011212090961635113, + 0.009246879257261753, + 0.015314485877752304, + 0.020208710804581642, + 0.049425117671489716, + 0.008283118717372417, + 0.011058792471885681, + 0.013828095979988575, + 0.04983934387564659, + 0.01165910717099905, + 0.0060121226124465466, + 0.023892175406217575, + 0.04494210332632065, + 0.003998869564384222, + 0.001872198423370719, + 0.0022278332617133856, + 0.002345740096643567, + 0.0024635058362036943, + 0.002274604281410575, + 0.0031428721267729998, + 0.0038286237977445126, + 0.019695302471518517, + 0.008771493099629879, + 0.008478375151753426, + 0.02990015782415867, + 0.010454639792442322, + 0.013186133466660976, + 0.020109647884964943, + 0.010912591591477394, + 0.003596061607822776, + 0.0039036329835653305, + 0.0037908609956502914, + 0.006542394869029522, + 0.00513212475925684, + 0.002738828770816326, + 0.007618971634656191, + 0.006077139172703028, + 0.0023604680318385363, + 0.0034792895894497633, + 0.00244185421615839, + 0.004994254559278488, + 0.004476544912904501, + 0.003780271392315626, + 0.005363325588405132, + 0.002445102669298649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.018198542296886444 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632262073457241 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632262073457241, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632262073457241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22450058162212372, + 0.034913066774606705, + 0.12041881680488586, + 0.3263668417930603, + 0.11322816461324692, + 0.04056942090392113, + 0.021704068407416344, + 0.03922674059867859, + 0.03807136043906212, + 0.04100095108151436 + ], + "credit_map": [ + 0.021704213693737984, + 0.015880988910794258, + 0.0569167397916317, + 0.01733914390206337, + 0.017108598724007607, + 0.019527681171894073, + 0.04817011579871178, + 0.02785308100283146, + 0.005876596085727215, + 0.0030098468996584415, + 0.0039010283071547747, + 0.0036338872741907835, + 0.004380933940410614, + 0.005628312937915325, + 0.004758563823997974, + 0.003723893780261278, + 0.01566905714571476, + 0.008857623673975468, + 0.01594163477420807, + 0.014194266870617867, + 0.01518709771335125, + 0.01990293711423874, + 0.017512336373329163, + 0.013153848238289356, + 0.02326400764286518, + 0.018178023397922516, + 0.09031186997890472, + 0.032469503581523895, + 0.02639641799032688, + 0.01745879091322422, + 0.09271026402711868, + 0.025577915832400322, + 0.027235055342316628, + 0.006246310658752918, + 0.010251578874886036, + 0.0068992446176707745, + 0.00607409980148077, + 0.008856798522174358, + 0.013331218622624874, + 0.03433384746313095, + 0.0052489545196294785, + 0.002859761007130146, + 0.0033396314829587936, + 0.00919416919350624, + 0.0032382220961153507, + 0.004120801109820604, + 0.005312420427799225, + 0.007255454547703266, + 0.003457004204392433, + 0.0017563813598826528, + 0.0030843208078294992, + 0.0038067924324423075, + 0.0018435517558827996, + 0.0025061427149921656, + 0.00316091300919652, + 0.0020889616571366787, + 0.0049910349771380424, + 0.003099546069279313, + 0.0031315123196691275, + 0.007316492032259703, + 0.0039119962602853775, + 0.0045028263702988625, + 0.006637408398091793, + 0.005635919980704784, + 0.004396536853164434, + 0.003306304570287466, + 0.003633284941315651, + 0.00837861280888319, + 0.003697820007801056, + 0.002908470807597041, + 0.0053375596180558205, + 0.006412767339497805, + 0.005603955592960119, + 0.0032539740204811096, + 0.006627267692238092, + 0.0057448274455964565, + 0.0022643886040896177, + 0.0047067939303815365, + 0.004927522037178278, + 0.007872220128774643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.007632262073457241 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle full-shot that slowly pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing darkly painted metal armor plates connected by deep blue silk thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure wear a kabuto helmet with a full-face man-yorai mask that conceals all facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily from below by underwater building lights and neon signs casting electric blue and red glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure standing completely motionless in waist-deep, churning, murky floodwater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487071394920349 + ], + "opposite_logprob_tokens": [ + -1.2237071990966797 + ], + "logprob_reward": -0.3487071394920349, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3487071394920349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10242513567209244, + 0.08504761010408401, + 0.05066583678126335, + 0.3711424767971039, + 0.03770945966243744, + 0.04044268652796745, + 0.08673033118247986, + 0.19254381954669952, + 0.021432247012853622, + 0.011860350146889687 + ], + "credit_map": [ + 0.012840645387768745, + 0.010711375623941422, + 0.017282141372561455, + 0.012348138727247715, + 0.008826182223856449, + 0.011440346948802471, + 0.012290917336940765, + 0.016685398295521736, + 0.01535053551197052, + 0.009469646029174328, + 0.012584259733557701, + 0.008491412736475468, + 0.0085728969424963, + 0.009544079191982746, + 0.01007227785885334, + 0.010962512344121933, + 0.007039541378617287, + 0.004762757103890181, + 0.006264189723879099, + 0.005925383418798447, + 0.0071832831017673016, + 0.008198227733373642, + 0.007462370675057173, + 0.0038300901651382446, + 0.08688001334667206, + 0.024625731632113457, + 0.09818744659423828, + 0.026985546573996544, + 0.0278334803879261, + 0.0242635328322649, + 0.04471280053257942, + 0.03765396401286125, + 0.006610935553908348, + 0.003480280749499798, + 0.0043014478869736195, + 0.00401667645201087, + 0.005079915747046471, + 0.005410605575889349, + 0.0048581501469016075, + 0.003951449878513813, + 0.0031265041325241327, + 0.0030216830782592297, + 0.0033664877992123365, + 0.010829444043338299, + 0.002926774090155959, + 0.002084985375404358, + 0.00536569906398654, + 0.009721113368868828, + 0.005507767666131258, + 0.010682349093258381, + 0.00836287159472704, + 0.01666525937616825, + 0.01423761248588562, + 0.005548149812966585, + 0.018503783270716667, + 0.007222550921142101, + 0.0272695142775774, + 0.013961841352283955, + 0.012892113998532295, + 0.04891187697649002, + 0.01686280220746994, + 0.02025778777897358, + 0.031582850962877274, + 0.020805057138204575, + 0.0022653101477771997, + 0.0021296697668731213, + 0.002948023146018386, + 0.0034867171198129654, + 0.0024787667207419872, + 0.0021714612375944853, + 0.004145663697272539, + 0.0018066393677145243, + 0.0011654881527647376, + 0.0013082526857033372, + 0.0010911098215728998, + 0.0020562296267598867, + 0.001973235048353672, + 0.0009559700265526772, + 0.002244290430098772, + 0.0010657758684828877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.3487071394920349 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_tensor([7552, 7552])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_tensor([7552, 7552])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ad289121e6b0f9d4e27ae599d04f24635082c720 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_tensor([7552, 7552])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an elegantly decorated Chinese restaurant, a static, eye-level wide-angle shot captures a dynamic social scene, unfolding around a large round table. The warm light from ornate overhead lanterns casts a soft glow over the gathering. To the right, a woman in a crisp white silk blouse engages in conversation, smiles slightly, turns her head, and elegantly smooths back her tousled black hair. Next to her, a woman in a vibrant red dress holds a lit cigarette between her fingers, a wisp of smoke curling toward the ceiling. She leans over and gives a gentle, encouraging pat on the arm of a man in a simple black sweater. Across the table, a man in a forest green jacket deliberately bends down, his upper body disappearing from the frame. He searches the floor for something before slowly returning to his seat. On the far left, a woman with hair in a neat bun, absorbed in her meal, steadily picks up dark wooden chopsticks and reaches forward to select a piece of food from the steaming plate in the center. Meanwhile, a woman in a black turtleneck quietly raises a clear glass to her lips to take a sip of water, then subtly shifts her weight to lean forward across the table, her posture signaling growing interest in the conversation, all contributing to a rich, layered, and naturally casual, friendly dinner atmosphere.", + "seed": 18204, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_7552_seed_18204_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09223234541714191, + "vlm_logprob": -0.09223234541714191, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.09223234541714191, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05609609931707382 + ], + "opposite_logprob_tokens": [ + -3.306096076965332 + ], + "logprob_reward": -0.05609609931707382, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05609609931707382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10431502759456635, + 0.08287391066551208, + 0.056567348539829254, + 0.04403680935502052, + 0.06214764341711998, + 0.09831278771162033, + 0.13791757822036743, + 0.10969951748847961, + 0.04328381270170212, + 0.2608456313610077 + ], + "credit_map": [ + 0.016055557876825333, + 0.012835875153541565, + 0.01715167984366417, + 0.0099205132573843, + 0.010211655870079994, + 0.011329947970807552, + 0.015160870738327503, + 0.011648927815258503, + 0.008921484462916851, + 0.01585656777024269, + 0.012353969737887383, + 0.007470295764505863, + 0.011193201877176762, + 0.008417966775596142, + 0.011025195010006428, + 0.007635228801518679, + 0.005970855709165335, + 0.007137211039662361, + 0.010095221921801567, + 0.005683000199496746, + 0.008633056655526161, + 0.00816353689879179, + 0.004912978038191795, + 0.005971488077193499, + 0.004810772370547056, + 0.0058629815466701984, + 0.005658496171236038, + 0.006660368759185076, + 0.006600421387702227, + 0.005664586555212736, + 0.004559400491416454, + 0.004219782073050737, + 0.0072751217521727085, + 0.008302680216729641, + 0.007263075560331345, + 0.010631927289068699, + 0.00989154726266861, + 0.005686165299266577, + 0.00728785851970315, + 0.005809264723211527, + 0.0076632509008049965, + 0.010978429578244686, + 0.02729746885597706, + 0.012825575657188892, + 0.011279282160103321, + 0.006028232630342245, + 0.01612870581448078, + 0.006111843977123499, + 0.02077685482800007, + 0.014065887778997421, + 0.006313599180430174, + 0.00697757862508297, + 0.050243016332387924, + 0.010786749422550201, + 0.020420365035533905, + 0.008333525620400906, + 0.005888706538826227, + 0.007264357525855303, + 0.0060269394889473915, + 0.005935017019510269, + 0.038672350347042084, + 0.007957753725349903, + 0.00971174705773592, + 0.02824264205992222, + 0.0033845168072730303, + 0.004720423370599747, + 0.0049477219581604, + 0.00842348113656044, + 0.006264422554522753, + 0.003360149683430791, + 0.007654725108295679, + 0.004528372082859278, + 0.0239774901419878, + 0.03284144774079323, + 0.024174120277166367, + 0.06366437673568726, + 0.032002080231904984, + 0.020466022193431854, + 0.0466989129781723, + 0.017021169885993004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.05609609931707382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562174499034882 + ], + "opposite_logprob_tokens": [ + -1.6006217002868652 + ], + "logprob_reward": -0.22562174499034882, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562174499034882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0725925862789154, + 0.06286751478910446, + 0.06812045723199844, + 0.0664980560541153, + 0.09979826211929321, + 0.20173285901546478, + 0.08863784372806549, + 0.10469849407672882, + 0.06621818244457245, + 0.16883575916290283 + ], + "credit_map": [ + 0.016620608046650887, + 0.006561503745615482, + 0.010971109382808208, + 0.005907899234443903, + 0.008629878051578999, + 0.0068049649707973, + 0.00984970387071371, + 0.0072469208389520645, + 0.006858700420707464, + 0.006036511622369289, + 0.010913794860243797, + 0.006567878182977438, + 0.00991580355912447, + 0.005899193696677685, + 0.011132276616990566, + 0.005543351639062166, + 0.00845009833574295, + 0.009820776991546154, + 0.009677233174443245, + 0.007647674065083265, + 0.008221221156418324, + 0.008778184652328491, + 0.009943674318492413, + 0.005581590346992016, + 0.008303509093821049, + 0.008450469933450222, + 0.008642973378300667, + 0.010561729781329632, + 0.010417008772492409, + 0.00844833254814148, + 0.006636952981352806, + 0.005037080030888319, + 0.011321614496409893, + 0.013213184662163258, + 0.012724582105875015, + 0.01483435370028019, + 0.017291123047471046, + 0.010168642736971378, + 0.01029572170227766, + 0.00994904525578022, + 0.01643799990415573, + 0.025249721482396126, + 0.052305568009614944, + 0.02069472335278988, + 0.024266067892313004, + 0.016092395409941673, + 0.03203929588198662, + 0.014647086150944233, + 0.009855304844677448, + 0.00975549966096878, + 0.005683883558958769, + 0.010198105126619339, + 0.02829153835773468, + 0.006594582460820675, + 0.012347569689154625, + 0.005911358632147312, + 0.010344000533223152, + 0.008783201687037945, + 0.011259695515036583, + 0.009795472025871277, + 0.030009254813194275, + 0.008082234300673008, + 0.011406711302697659, + 0.015017922967672348, + 0.0034184204414486885, + 0.008737408556044102, + 0.006363782566040754, + 0.017005322501063347, + 0.012038416229188442, + 0.0038716031704097986, + 0.010697000660002232, + 0.004086225759238005, + 0.008545370772480965, + 0.01625674031674862, + 0.009480479173362255, + 0.017063019797205925, + 0.057750750333070755, + 0.016096064820885658, + 0.034890465438365936, + 0.008752863854169846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.22562174499034882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607716642320156 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.01607716642320156, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607716642320156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06630086153745651, + 0.08949421346187592, + 0.06572065502405167, + 0.104640893638134, + 0.06622399389743805, + 0.1769438534975052, + 0.07501687109470367, + 0.20819850265979767, + 0.10213573276996613, + 0.04532444849610329 + ], + "credit_map": [ + 0.011842304840683937, + 0.009681951254606247, + 0.01112379040569067, + 0.00542439054697752, + 0.007660685572773218, + 0.005696979351341724, + 0.007171972189098597, + 0.007698787841945887, + 0.00885796919465065, + 0.015986531972885132, + 0.012887661345303059, + 0.009172935038805008, + 0.01208366546779871, + 0.009012131951749325, + 0.015600663609802723, + 0.005892663262784481, + 0.006495930720120668, + 0.008749537169933319, + 0.009591817855834961, + 0.006616619415581226, + 0.00911723729223013, + 0.00834653340280056, + 0.011797231622040272, + 0.005005750805139542, + 0.010549603030085564, + 0.021514788269996643, + 0.013187535107135773, + 0.014724738895893097, + 0.014047086238861084, + 0.012299774214625359, + 0.009297549724578857, + 0.009019818156957626, + 0.0064099980518221855, + 0.009166707284748554, + 0.00677880086004734, + 0.010083098895847797, + 0.013342952355742455, + 0.007034960202872753, + 0.0068088555708527565, + 0.006598624866455793, + 0.014472012408077717, + 0.01692803204059601, + 0.058100625872612, + 0.013655715622007847, + 0.02130616270005703, + 0.012003958225250244, + 0.031575337052345276, + 0.008901987224817276, + 0.007448659744113684, + 0.015619772486388683, + 0.00874805822968483, + 0.009881077334284782, + 0.013210387900471687, + 0.005406612530350685, + 0.008638623170554638, + 0.0060636792331933975, + 0.015161585062742233, + 0.03459811210632324, + 0.024777185171842575, + 0.019675394520163536, + 0.05122316628694534, + 0.012415628880262375, + 0.013867916539311409, + 0.03647952899336815, + 0.006199789233505726, + 0.014853542670607567, + 0.01095516886562109, + 0.018894586712121964, + 0.01521088182926178, + 0.008415285497903824, + 0.019413989037275314, + 0.008192485198378563, + 0.005846664309501648, + 0.005245744716376066, + 0.003629429265856743, + 0.00438821641728282, + 0.01065634936094284, + 0.0044547077268362045, + 0.007950683124363422, + 0.0031526521779596806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.01607716642320156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588379353284836 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588379353284836, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588379353284836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27523157000541687, + 0.06514044851064682, + 0.021473808214068413, + 0.07232445478439331, + 0.08847314864397049, + 0.09459369629621506, + 0.05165582522749901, + 0.08864983171224594, + 0.10619790107011795, + 0.1362592875957489 + ], + "credit_map": [ + 0.04378783330321312, + 0.0713324248790741, + 0.027578910812735558, + 0.02108807861804962, + 0.034612953662872314, + 0.027972916141152382, + 0.024264933541417122, + 0.024593517184257507, + 0.00751299737021327, + 0.008637447841465473, + 0.010675409808754921, + 0.00542220426723361, + 0.007958659902215004, + 0.007742304354906082, + 0.012731779366731644, + 0.0044596451334655285, + 0.0025782522279769182, + 0.0033588092774152756, + 0.0029881231021136045, + 0.002297045197337866, + 0.003158335341140628, + 0.0029436987824738026, + 0.002345840446650982, + 0.001803703373298049, + 0.008650586940348148, + 0.011992988176643848, + 0.009743125177919865, + 0.012865813449025154, + 0.009544110856950283, + 0.006317206658422947, + 0.00698797358199954, + 0.006222647149115801, + 0.008123423904180527, + 0.008354546502232552, + 0.013437669724225998, + 0.013830062933266163, + 0.014229956082999706, + 0.009899294935166836, + 0.011750190518796444, + 0.008848002180457115, + 0.007168713491410017, + 0.008826476521790028, + 0.02907312475144863, + 0.007410028018057346, + 0.012656799517571926, + 0.007556662429124117, + 0.01642725057899952, + 0.005474640987813473, + 0.004565898794680834, + 0.00998631026595831, + 0.004126891028136015, + 0.005262312944978476, + 0.009640217758715153, + 0.006527575198560953, + 0.006891397759318352, + 0.004655223339796066, + 0.006330736912786961, + 0.0065910471603274345, + 0.00894846674054861, + 0.008168534375727177, + 0.02815832942724228, + 0.006090657785534859, + 0.008344698697328568, + 0.01601736806333065, + 0.005285089369863272, + 0.012746582739055157, + 0.010532013140618801, + 0.025621412321925163, + 0.018443167209625244, + 0.006118250545114279, + 0.019666405394673347, + 0.007784977555274963, + 0.01759796403348446, + 0.011965467594563961, + 0.012595760636031628, + 0.01740770973265171, + 0.025573857128620148, + 0.01941227912902832, + 0.021732864901423454, + 0.009973389096558094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.020588379353284836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277833700180054 + ], + "opposite_logprob_tokens": [ + -2.0177783966064453 + ], + "logprob_reward": -0.14277833700180054, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277833700180054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12349040061235428, + 0.0697237104177475, + 0.043172597885131836, + 0.05853717029094696, + 0.0513378381729126, + 0.09681232273578644, + 0.09382671862840652, + 0.09505414217710495, + 0.07790198922157288, + 0.29014310240745544 + ], + "credit_map": [ + 0.022662291303277016, + 0.0236160010099411, + 0.013126534409821033, + 0.009839856065809727, + 0.01093590259552002, + 0.01285600010305643, + 0.018848098814487457, + 0.01160571351647377, + 0.007618717849254608, + 0.009263886138796806, + 0.010755923576653004, + 0.007084354758262634, + 0.008468073792755604, + 0.007549195550382137, + 0.012355056591331959, + 0.006628503557294607, + 0.004108026158064604, + 0.0066519021056592464, + 0.004721736069768667, + 0.00463920459151268, + 0.006376106292009354, + 0.00530816987156868, + 0.0073347375728189945, + 0.004032714758068323, + 0.005107082426548004, + 0.006552745122462511, + 0.006826986093074083, + 0.00604311004281044, + 0.007936397567391396, + 0.00877209473401308, + 0.011938725598156452, + 0.005360023584216833, + 0.004982108250260353, + 0.0063512916676700115, + 0.007134618703275919, + 0.011941132135689259, + 0.005486525595188141, + 0.0042115915566682816, + 0.006768431980162859, + 0.00446213549003005, + 0.008298678323626518, + 0.007763902191072702, + 0.027941541746258736, + 0.009434589184820652, + 0.011205241084098816, + 0.007418802473694086, + 0.0182335264980793, + 0.006516046356409788, + 0.008818828500807285, + 0.01609603501856327, + 0.011943851597607136, + 0.012133852578699589, + 0.01097829919308424, + 0.01107602845877409, + 0.011301771737635136, + 0.011478050611913204, + 0.006050008349120617, + 0.005161701235920191, + 0.005509018898010254, + 0.007061431650072336, + 0.031146647408604622, + 0.007612160872668028, + 0.011494513601064682, + 0.021018654108047485, + 0.005044538993388414, + 0.010528376325964928, + 0.006943550426512957, + 0.01655714586377144, + 0.015384864993393421, + 0.006264449562877417, + 0.009737280197441578, + 0.007441789843142033, + 0.026280999183654785, + 0.03345322981476784, + 0.01687086746096611, + 0.027768194675445557, + 0.08061908185482025, + 0.02587435394525528, + 0.05989406630396843, + 0.019382329657673836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.14277833700180054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an elegantly decorated Chinese restaurant, a static, eye-level wide-angle shot captures a dynamic social scene, unfolding around a large round table. The warm light from ornate overhead lanterns casts a soft glow over the gathering. To the right, a woman in a crisp white silk blouse engages in conversation, smiles slightly, turns her head, and elegantly smooths back her tousled black hair. Next to her, a woman in a vibrant red dress holds a lit cigarette between her fingers, a wisp of smoke curling toward the ceiling. She leans over and gives a gentle, encouraging pat on the arm of a man in a simple black sweater. Across the table, a man in a forest green jacket deliberately bends down, his upper body disappearing from the frame. He searches the floor for something before slowly returning to his seat. On the far left, a woman with hair in a neat bun, absorbed in her meal, steadily picks up dark wooden chopsticks and reaches forward to select a piece of food from the steaming plate in the center. Meanwhile, a woman in a black turtleneck quietly raises a clear glass to her lips to take a sip of water, then subtly shifts her weight to lean forward across the table, her posture signaling growing interest in the conversation, all contributing to a rich, layered, and naturally casual, friendly dinner atmosphere.", + "seed": 18205, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_7552_seed_18205_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10390157047659158, + "vlm_logprob": -0.10390157047659158, + "vlm_yes_no_margin": 3.025000011920929, + "vlm_token_credit": -0.10390157047659158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03472740948200226 + ], + "opposite_logprob_tokens": [ + -3.9097273349761963 + ], + "logprob_reward": -0.03472740948200226, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03472740948200226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.210793599486351, + 0.0888158306479454, + 0.05166545510292053, + 0.035792771726846695, + 0.060232143849134445, + 0.09664737433195114, + 0.14082860946655273, + 0.13861609995365143, + 0.06676715612411499, + 0.109840989112854 + ], + "credit_map": [ + 0.02546410635113716, + 0.028959259390830994, + 0.02664833329617977, + 0.0333891324698925, + 0.020372115075588226, + 0.026896897703409195, + 0.024926237761974335, + 0.024137509986758232, + 0.009040113538503647, + 0.011219893582165241, + 0.010878807865083218, + 0.015305200591683388, + 0.010193930938839912, + 0.011944284662604332, + 0.010419770143926144, + 0.009813829325139523, + 0.00797345768660307, + 0.0035512270405888557, + 0.004944088868796825, + 0.008059283718466759, + 0.006644509732723236, + 0.008614101447165012, + 0.005518356338143349, + 0.006360429804772139, + 0.0041444506496191025, + 0.007126835640519857, + 0.001954857259988785, + 0.00574732618406415, + 0.006020613946020603, + 0.003586420789361, + 0.0038528924342244864, + 0.0033593731932342052, + 0.006117188837379217, + 0.009061608463525772, + 0.007154871243983507, + 0.00833805464208126, + 0.008077707141637802, + 0.008170899003744125, + 0.008550372906029224, + 0.004761441610753536, + 0.012216774746775627, + 0.012351851910352707, + 0.010316682048141956, + 0.021148689091205597, + 0.009791102260351181, + 0.012423841282725334, + 0.01304616779088974, + 0.005352261010557413, + 0.02135707437992096, + 0.019191863015294075, + 0.013343806378543377, + 0.014830904081463814, + 0.01600050926208496, + 0.017622118815779686, + 0.027004970237612724, + 0.01147736981511116, + 0.012716024182736874, + 0.0214935764670372, + 0.014670616947114468, + 0.009655443951487541, + 0.020928358659148216, + 0.026176132261753082, + 0.019026031717658043, + 0.013949915766716003, + 0.00791859906166792, + 0.010340084321796894, + 0.010968697257339954, + 0.005106804426759481, + 0.008513450622558594, + 0.007844903506338596, + 0.008362325839698315, + 0.007712285965681076, + 0.011621635407209396, + 0.014061476103961468, + 0.01761309988796711, + 0.010958811268210411, + 0.014025464653968811, + 0.011374001391232014, + 0.0167480930685997, + 0.01343840453773737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.03472740948200226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486599326133728 + ], + "opposite_logprob_tokens": [ + -1.2236599922180176 + ], + "logprob_reward": -0.3486599326133728, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3486599326133728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09356916695833206, + 0.028457988053560257, + 0.033573757857084274, + 0.04986283555626869, + 0.23832377791404724, + 0.13768330216407776, + 0.1426110863685608, + 0.05009569227695465, + 0.136171355843544, + 0.08965100347995758 + ], + "credit_map": [ + 0.013724311254918575, + 0.012957880273461342, + 0.007891763001680374, + 0.015372099354863167, + 0.009241867810487747, + 0.017982717603445053, + 0.008729133754968643, + 0.007669389713555574, + 0.0025916113518178463, + 0.0031893907580524683, + 0.002888565417379141, + 0.005673293489962816, + 0.0029701769817620516, + 0.0029755528084933758, + 0.003751859301701188, + 0.0044175381772220135, + 0.004941870924085379, + 0.003964841831475496, + 0.0033677357714623213, + 0.005246744956821203, + 0.004421907011419535, + 0.005968498066067696, + 0.003227108856663108, + 0.0024350490421056747, + 0.0067517114803195, + 0.008845380507409573, + 0.0033958961721509695, + 0.005503139458596706, + 0.0069435774348676205, + 0.008685370907187462, + 0.004700718447566032, + 0.005037043243646622, + 0.016567358747124672, + 0.06950785219669342, + 0.011401566676795483, + 0.011857710778713226, + 0.07192663848400116, + 0.034463394433259964, + 0.012042625807225704, + 0.01055663451552391, + 0.017476025968790054, + 0.020122665911912918, + 0.013830523937940598, + 0.02056729793548584, + 0.014627814292907715, + 0.029951417818665504, + 0.011467663571238518, + 0.009639889933168888, + 0.01866091601550579, + 0.019700517877936363, + 0.011305074207484722, + 0.013126827776432037, + 0.02178146131336689, + 0.016624582931399345, + 0.030839329585433006, + 0.010572375729680061, + 0.0066328272223472595, + 0.00860339030623436, + 0.006080565042793751, + 0.004100028891116381, + 0.0080278180539608, + 0.007583048194646835, + 0.005257371347397566, + 0.0038106467109173536, + 0.014324158430099487, + 0.012222839519381523, + 0.02177419327199459, + 0.011591730639338493, + 0.022131597623229027, + 0.024986637756228447, + 0.015466348268091679, + 0.013673835434019566, + 0.00879393145442009, + 0.008782926015555859, + 0.013533094897866249, + 0.007606822531670332, + 0.01618228852748871, + 0.015305631794035435, + 0.009163135662674904, + 0.01028317492455244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.3486599326133728 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329133450984955 + ], + "opposite_logprob_tokens": [ + -3.7732913494110107 + ], + "logprob_reward": -0.02329133450984955, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329133450984955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17599204182624817, + 0.08056832104921341, + 0.11534930020570755, + 0.06165003031492233, + 0.14594054222106934, + 0.0685834139585495, + 0.0715453252196312, + 0.1215752437710762, + 0.029188185930252075, + 0.12960760295391083 + ], + "credit_map": [ + 0.02376975677907467, + 0.013978495262563229, + 0.025416122749447823, + 0.019988924264907837, + 0.017245052382349968, + 0.033322207629680634, + 0.021064545959234238, + 0.021206926554441452, + 0.010388921946287155, + 0.005718053784221411, + 0.015447625890374184, + 0.008624754846096039, + 0.009948530234396458, + 0.010122528299689293, + 0.010202078148722649, + 0.01011582650244236, + 0.02623179741203785, + 0.009983206167817116, + 0.01124973688274622, + 0.009461277164518833, + 0.013799495995044708, + 0.01763269491493702, + 0.010715188458561897, + 0.016275905072689056, + 0.008064286783337593, + 0.009837749414145947, + 0.0054540932178497314, + 0.009616999886929989, + 0.00815087091177702, + 0.00808360893279314, + 0.006113037932664156, + 0.00632938276976347, + 0.008591639809310436, + 0.0435219369828701, + 0.0075217075645923615, + 0.010989156551659107, + 0.0465976744890213, + 0.015880383551120758, + 0.007370441220700741, + 0.005467596929520369, + 0.005274895112961531, + 0.008415375836193562, + 0.007810748182237148, + 0.01908227615058422, + 0.007335437461733818, + 0.007168642710894346, + 0.00970318354666233, + 0.003792856354266405, + 0.008929657749831676, + 0.010157928802073002, + 0.006614329293370247, + 0.007472113240510225, + 0.00763333635404706, + 0.008973335847258568, + 0.013971653766930103, + 0.007792966905981302, + 0.012169291265308857, + 0.01168109942227602, + 0.01235395111143589, + 0.00785881094634533, + 0.018831675872206688, + 0.031291984021663666, + 0.017175307497382164, + 0.010213118977844715, + 0.0036860366817563772, + 0.002968265675008297, + 0.004445954691618681, + 0.0023926885332912207, + 0.003990819212049246, + 0.005166477058082819, + 0.003277337644249201, + 0.003260608296841383, + 0.013865522108972073, + 0.009446882642805576, + 0.019863590598106384, + 0.011485734954476357, + 0.019729623571038246, + 0.01328345574438572, + 0.025974981486797333, + 0.015957802534103394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.02329133450984955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254300493746996 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.01254300493746996, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254300493746996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20846985280513763, + 0.09693453460931778, + 0.05886266753077507, + 0.0623386986553669, + 0.06619894504547119, + 0.050055064260959625, + 0.07511293143033981, + 0.16049076616764069, + 0.04454580694437027, + 0.17699070274829865 + ], + "credit_map": [ + 0.034675780683755875, + 0.01609182357788086, + 0.02802782505750656, + 0.017698921263217926, + 0.019212450832128525, + 0.04100219905376434, + 0.02250729128718376, + 0.029253587126731873, + 0.010401664301753044, + 0.005917681381106377, + 0.016314927488565445, + 0.009293773211538792, + 0.011891661211848259, + 0.013535678386688232, + 0.01008573267608881, + 0.019493425264954567, + 0.010338555090129375, + 0.004663687199354172, + 0.00491672195494175, + 0.0048878127709031105, + 0.010897423140704632, + 0.008742674253880978, + 0.006496401038020849, + 0.007919402793049812, + 0.00826863944530487, + 0.007228079251945019, + 0.0062407078221440315, + 0.008697472512722015, + 0.006952356547117233, + 0.010209447704255581, + 0.005740476306527853, + 0.009001524187624454, + 0.005550104659050703, + 0.01849035546183586, + 0.0034606563858687878, + 0.004159376956522465, + 0.019080378115177155, + 0.007871956564486027, + 0.0037385423202067614, + 0.003847582498565316, + 0.0068589430302381516, + 0.008816673420369625, + 0.0054994612000882626, + 0.00614160904660821, + 0.005410017911344767, + 0.0070920973084867, + 0.006798659451305866, + 0.00343760522082448, + 0.009820719249546528, + 0.00919353123754263, + 0.006800011731684208, + 0.007160562090575695, + 0.010115784592926502, + 0.008583690039813519, + 0.015770336613059044, + 0.0076682991348207, + 0.016050290316343307, + 0.015891751274466515, + 0.016739780083298683, + 0.009423313662409782, + 0.01799706555902958, + 0.0424935482442379, + 0.02758423052728176, + 0.014310797676444054, + 0.00652888510376215, + 0.004795498214662075, + 0.00737258605659008, + 0.00375780719332397, + 0.006325798574835062, + 0.007101823575794697, + 0.005126900039613247, + 0.0035365100484341383, + 0.0165566336363554, + 0.016849638894200325, + 0.026819288730621338, + 0.01739971525967121, + 0.026064397767186165, + 0.01811276562511921, + 0.03483085706830025, + 0.020357418805360794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.01254300493746996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028617084026337 + ], + "opposite_logprob_tokens": [ + -2.3502862453460693 + ], + "logprob_reward": -0.10028617084026337, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028617084026337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04849714785814285, + 0.01928701438009739, + 0.038608986884355545, + 0.04584958776831627, + 0.09129122644662857, + 0.03139739856123924, + 0.046339210122823715, + 0.1274118721485138, + 0.22224189341068268, + 0.32907554507255554 + ], + "credit_map": [ + 0.007145368959754705, + 0.0051438105292618275, + 0.005484506022185087, + 0.006286739371716976, + 0.0047304993495345116, + 0.008240999653935432, + 0.005605572834610939, + 0.005859656259417534, + 0.002281910739839077, + 0.001397548126988113, + 0.0029120193794369698, + 0.0019626321736723185, + 0.002325597917661071, + 0.003504708642140031, + 0.0018389347242191434, + 0.0030636668670922518, + 0.005426792893558741, + 0.0029975385405123234, + 0.003473476506769657, + 0.003820107551291585, + 0.005940972361713648, + 0.006306045223027468, + 0.004380027297884226, + 0.00626402860507369, + 0.005541115067899227, + 0.005685772746801376, + 0.003328506136313081, + 0.006937326863408089, + 0.0055878255516290665, + 0.009616215713322163, + 0.004823051858693361, + 0.004329778719693422, + 0.006657160818576813, + 0.02197389118373394, + 0.006586521398276091, + 0.007969479076564312, + 0.024584507569670677, + 0.008289984427392483, + 0.009711923077702522, + 0.005517763085663319, + 0.004506842233240604, + 0.006028938107192516, + 0.003006314393132925, + 0.002772553125396371, + 0.0038781992625445127, + 0.005457274150103331, + 0.0031894873827695847, + 0.002557795960456133, + 0.0049575138837099075, + 0.008947549387812614, + 0.003950796090066433, + 0.00402063550427556, + 0.005293919239193201, + 0.005156784318387508, + 0.009583602659404278, + 0.0044284118339419365, + 0.013630001805722713, + 0.012476854957640171, + 0.01366775669157505, + 0.008015760220587254, + 0.021222511306405067, + 0.02748803421854973, + 0.019651688635349274, + 0.011259284801781178, + 0.02948976494371891, + 0.01833765022456646, + 0.03711649402976036, + 0.01379761379212141, + 0.0421079620718956, + 0.04071177914738655, + 0.024264780804514885, + 0.016415879130363464, + 0.0314151830971241, + 0.02952871099114418, + 0.049406737089157104, + 0.03011263720691204, + 0.04761066287755966, + 0.034195661544799805, + 0.06727727502584457, + 0.03952871263027191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.10028617084026337 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an elegantly decorated Chinese restaurant, a static, eye-level wide-angle shot captures a dynamic social scene, unfolding around a large round table. The warm light from ornate overhead lanterns casts a soft glow over the gathering. To the right, a woman in a crisp white silk blouse engages in conversation, smiles slightly, turns her head, and elegantly smooths back her tousled black hair. Next to her, a woman in a vibrant red dress holds a lit cigarette between her fingers, a wisp of smoke curling toward the ceiling. She leans over and gives a gentle, encouraging pat on the arm of a man in a simple black sweater. Across the table, a man in a forest green jacket deliberately bends down, his upper body disappearing from the frame. He searches the floor for something before slowly returning to his seat. On the far left, a woman with hair in a neat bun, absorbed in her meal, steadily picks up dark wooden chopsticks and reaches forward to select a piece of food from the steaming plate in the center. Meanwhile, a woman in a black turtleneck quietly raises a clear glass to her lips to take a sip of water, then subtly shifts her weight to lean forward across the table, her posture signaling growing interest in the conversation, all contributing to a rich, layered, and naturally casual, friendly dinner atmosphere.", + "seed": 18206, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_7552_seed_18206_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.159987460821867, + "vlm_logprob": -0.159987460821867, + "vlm_yes_no_margin": 2.599999988079071, + "vlm_token_credit": -0.159987460821867, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0302345659583807 + ], + "opposite_logprob_tokens": [ + -4.155234336853027 + ], + "logprob_reward": -0.0302345659583807, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0302345659583807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18387924134731293, + 0.21853354573249817, + 0.10649842768907547, + 0.12802459299564362, + 0.060474321246147156, + 0.09918851405382156, + 0.06025085225701332, + 0.05528343468904495, + 0.04431445151567459, + 0.043552596122026443 + ], + "credit_map": [ + 0.014718207530677319, + 0.014223558828234673, + 0.023479942232370377, + 0.046320002526044846, + 0.016412347555160522, + 0.012569183483719826, + 0.03444705903530121, + 0.021708939224481583, + 0.017126811668276787, + 0.02818886749446392, + 0.026964839547872543, + 0.053726524114608765, + 0.018073854967951775, + 0.015215537510812283, + 0.0369526669383049, + 0.022284455597400665, + 0.010439188219606876, + 0.011150994338095188, + 0.019210882484912872, + 0.012792940251529217, + 0.007893353700637817, + 0.012173563241958618, + 0.013345060870051384, + 0.019492443650960922, + 0.01003563217818737, + 0.011555875651538372, + 0.030047068372368813, + 0.012354161590337753, + 0.0142169538885355, + 0.007355980575084686, + 0.011682434007525444, + 0.030776485800743103, + 0.005271321162581444, + 0.004854974336922169, + 0.01641892082989216, + 0.005422650370746851, + 0.004911419935524464, + 0.004781932104378939, + 0.005322244483977556, + 0.013490855693817139, + 0.013526666909456253, + 0.012696444988250732, + 0.014209482818841934, + 0.01129115093499422, + 0.011691011488437653, + 0.014326528646051884, + 0.010851360857486725, + 0.010595863685011864, + 0.008837870322167873, + 0.0048573315143585205, + 0.01150404755026102, + 0.00807536393404007, + 0.0058251372538506985, + 0.01066630519926548, + 0.005254034418612719, + 0.005230762995779514, + 0.005608117673546076, + 0.005864625331014395, + 0.004780807998031378, + 0.00533853517845273, + 0.005391444079577923, + 0.00868211966007948, + 0.014120768755674362, + 0.005497021600604057, + 0.0034275464713573456, + 0.00458854716271162, + 0.0046173762530088425, + 0.009414860978722572, + 0.0040357052348554134, + 0.004654200281947851, + 0.00858262088149786, + 0.004993595648556948, + 0.0038604140281677246, + 0.006560036446899176, + 0.0065338932909071445, + 0.004820836707949638, + 0.0043168566189706326, + 0.0051405285485088825, + 0.006675080396234989, + 0.0056449477560818195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0302345659583807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709527254104614 + ], + "opposite_logprob_tokens": [ + -1.1370952129364014 + ], + "logprob_reward": -0.38709527254104614, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38709527254104614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12003161758184433, + 0.05205076187849045, + 0.33237284421920776, + 0.27308881282806396, + 0.04803220182657242, + 0.07898612320423126, + 0.018106358125805855, + 0.02339724451303482, + 0.033850040286779404, + 0.020083945244550705 + ], + "credit_map": [ + 0.006223056931048632, + 0.007901504635810852, + 0.009975941851735115, + 0.04256008565425873, + 0.005100339651107788, + 0.006951376795768738, + 0.022391995415091515, + 0.0189273152500391, + 0.0037299415562301874, + 0.005134791601449251, + 0.005693041253834963, + 0.010237877257168293, + 0.0037701884284615517, + 0.004681472200900316, + 0.01168972160667181, + 0.007113731931895018, + 0.029664015397429466, + 0.016362030059099197, + 0.1041964665055275, + 0.027721839025616646, + 0.010585574433207512, + 0.019389301538467407, + 0.018697164952754974, + 0.10575643181800842, + 0.01929674856364727, + 0.015222555957734585, + 0.07181838899850845, + 0.030239757150411606, + 0.014041339978575706, + 0.02376651205122471, + 0.020585300400853157, + 0.0781182125210762, + 0.0048036957159638405, + 0.004040662664920092, + 0.01015249453485012, + 0.0056925006210803986, + 0.00408465601503849, + 0.005473758094012737, + 0.0074530597776174545, + 0.006331371143460274, + 0.006143189966678619, + 0.007065975572913885, + 0.018817946314811707, + 0.008198054507374763, + 0.008271992206573486, + 0.005555277690291405, + 0.0074375951662659645, + 0.017496099695563316, + 0.001991167664527893, + 0.0018779102247208357, + 0.003613373264670372, + 0.0027182819321751595, + 0.002037595957517624, + 0.002052450552582741, + 0.002030288567766547, + 0.0017852912424132228, + 0.002485832665115595, + 0.0027118169236928225, + 0.0027769168373197317, + 0.0027334534097462893, + 0.002419486176222563, + 0.003602053504437208, + 0.0027847776655107737, + 0.0038829080294817686, + 0.001900840550661087, + 0.00333083001896739, + 0.002201068215072155, + 0.009972249157726765, + 0.0028589239809662104, + 0.003237072378396988, + 0.007953840307891369, + 0.002395216841250658, + 0.002032162854447961, + 0.0023610617499798536, + 0.0027575218118727207, + 0.0029491777531802654, + 0.002064344473183155, + 0.001743005821481347, + 0.0032241828739643097, + 0.002952488139271736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.38709527254104614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496241897344589 + ], + "opposite_logprob_tokens": [ + -2.929962396621704 + ], + "logprob_reward": -0.05496241897344589, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496241897344589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22114677727222443, + 0.3614407479763031, + 0.15483491122722626, + 0.06675024330615997, + 0.027784548699855804, + 0.04144314303994179, + 0.05698655545711517, + 0.019692299887537956, + 0.015172325074672699, + 0.03474844992160797 + ], + "credit_map": [ + 0.014152147807180882, + 0.014187832362949848, + 0.016860736533999443, + 0.0699029341340065, + 0.010824124328792095, + 0.011862176470458508, + 0.04003165289759636, + 0.04332518205046654, + 0.02663644216954708, + 0.043675024062395096, + 0.05028686299920082, + 0.08892416208982468, + 0.0219325739890337, + 0.023947052657604218, + 0.06470336765050888, + 0.041335251182317734, + 0.013131901621818542, + 0.00754893384873867, + 0.04763743653893471, + 0.011763771995902061, + 0.0050142197869718075, + 0.010181997902691364, + 0.007992937229573727, + 0.051563702523708344, + 0.00528669124469161, + 0.005360104609280825, + 0.012747158296406269, + 0.007956478744745255, + 0.006914868950843811, + 0.006720522418618202, + 0.008107520639896393, + 0.01365689653903246, + 0.0017904958222061396, + 0.002287409035488963, + 0.008459154516458511, + 0.00253899279050529, + 0.002249618759378791, + 0.0019253274658694863, + 0.0025393550749868155, + 0.005994195118546486, + 0.002720806747674942, + 0.005064547061920166, + 0.010248114354908466, + 0.004939521662890911, + 0.002597302198410034, + 0.0023636072874069214, + 0.004820574540644884, + 0.008688668720424175, + 0.007353578228503466, + 0.0031708755996078253, + 0.014625412411987782, + 0.010595758445560932, + 0.003377301152795553, + 0.006368223577737808, + 0.006560998503118753, + 0.004934410098940134, + 0.0019019499886780977, + 0.0020891136955469847, + 0.001780757331289351, + 0.002412875648587942, + 0.0026825556997209787, + 0.002379703801125288, + 0.0033964794129133224, + 0.003048865357413888, + 0.0013006995432078838, + 0.001514214905910194, + 0.001417367602698505, + 0.002589586190879345, + 0.0015531782992184162, + 0.0018852694192901254, + 0.0033103639725595713, + 0.001601646188646555, + 0.0028390763327479362, + 0.004521109629422426, + 0.005232708062976599, + 0.005191444884985685, + 0.0036853274796158075, + 0.0032984386198222637, + 0.0062884078361094, + 0.003691935446113348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.05496241897344589 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203174039721489 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203174039721489, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203174039721489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24094301462173462, + 0.1781539022922516, + 0.08542059361934662, + 0.14589804410934448, + 0.05503787845373154, + 0.02255793660879135, + 0.06910663843154907, + 0.0505729615688324, + 0.07002639025449753, + 0.08228258788585663 + ], + "credit_map": [ + 0.021826302632689476, + 0.024604812264442444, + 0.03528166934847832, + 0.05021637678146362, + 0.016626786440610886, + 0.017481205984950066, + 0.03354218602180481, + 0.04136369749903679, + 0.014230863191187382, + 0.022377360612154007, + 0.023633090779185295, + 0.042607344686985016, + 0.012040011584758759, + 0.01561546977609396, + 0.030284492298960686, + 0.017365267500281334, + 0.010572210885584354, + 0.0063981409184634686, + 0.01970510557293892, + 0.009591419249773026, + 0.006361865438520908, + 0.010773962363600731, + 0.008774175308644772, + 0.0132437227293849, + 0.012857766821980476, + 0.011792967095971107, + 0.035899315029382706, + 0.014648942276835442, + 0.01829610951244831, + 0.01267626415938139, + 0.012671400792896748, + 0.027055291458964348, + 0.004658363293856382, + 0.007175762671977282, + 0.004766543861478567, + 0.006350007839500904, + 0.008130768314003944, + 0.007888886146247387, + 0.008370744064450264, + 0.00769680505618453, + 0.0021220906637609005, + 0.002159423427656293, + 0.004276153165847063, + 0.0031702856067568064, + 0.002052578842267394, + 0.0018184221116825938, + 0.0030107933562248945, + 0.0039481897838413715, + 0.007516156416386366, + 0.0041109067387878895, + 0.016672952100634575, + 0.013772628270089626, + 0.004336892161518335, + 0.009332719258964062, + 0.007458427455276251, + 0.0059059602208435535, + 0.0042021069675683975, + 0.004931270610541105, + 0.006156102288514376, + 0.013174866326153278, + 0.003543026512488723, + 0.004886723589152098, + 0.00877296831458807, + 0.004905900452286005, + 0.003929901868104935, + 0.0057290093973279, + 0.0050865220837295055, + 0.01544330082833767, + 0.00546792009845376, + 0.007579348981380463, + 0.017325878143310547, + 0.009464520029723644, + 0.007055634167045355, + 0.007166699506342411, + 0.010934247635304928, + 0.006709798704832792, + 0.007132767699658871, + 0.015326432883739471, + 0.012537277303636074, + 0.015419743955135345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014203174039721489 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134418725967407 + ], + "opposite_logprob_tokens": [ + -1.3134418725967407 + ], + "logprob_reward": -0.3134418725967407, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134418725967407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322912871837616, + 0.07032749056816101, + 0.35410287976264954, + 0.18619272112846375, + 0.08307823538780212, + 0.048104897141456604, + 0.026366189122200012, + 0.029597986489534378, + 0.04431523010134697, + 0.02562311477959156 + ], + "credit_map": [ + 0.008808715268969536, + 0.008742299862205982, + 0.008762788027524948, + 0.040072232484817505, + 0.008177867159247398, + 0.009606599807739258, + 0.021831467747688293, + 0.026289325207471848, + 0.006558025721460581, + 0.009150659665465355, + 0.008594364859163761, + 0.014616776257753372, + 0.006931268610060215, + 0.008737698197364807, + 0.007388228550553322, + 0.008350471965968609, + 0.03021213226020336, + 0.015395625494420528, + 0.11662204563617706, + 0.028808685019612312, + 0.006426215171813965, + 0.020700911059975624, + 0.014948912896215916, + 0.1209883764386177, + 0.01796668954193592, + 0.020282503217458725, + 0.02419428527355194, + 0.017794685438275337, + 0.036336980760097504, + 0.02020871266722679, + 0.020601600408554077, + 0.028807274997234344, + 0.005861771758645773, + 0.0052268668077886105, + 0.024499967694282532, + 0.008508604019880295, + 0.005912236403673887, + 0.006114581599831581, + 0.009229891933500767, + 0.01772432215511799, + 0.00340292532928288, + 0.004374133422970772, + 0.008649430237710476, + 0.006326525006443262, + 0.005493559874594212, + 0.00557329785078764, + 0.006457032170146704, + 0.007827996276319027, + 0.003558581229299307, + 0.0024942616000771523, + 0.0027918629348278046, + 0.004893825389444828, + 0.00240912102162838, + 0.0019402209436520934, + 0.005751804914325476, + 0.0025265126023441553, + 0.002888793358579278, + 0.0023294282145798206, + 0.0028893977869302034, + 0.006625297013670206, + 0.002420034259557724, + 0.003015722380951047, + 0.005955867003649473, + 0.003473449032753706, + 0.005399841815233231, + 0.0034013804979622364, + 0.004706373438239098, + 0.005963941570371389, + 0.005059246439486742, + 0.006044942419975996, + 0.007462064269930124, + 0.0062774415127933025, + 0.0018244239035993814, + 0.0028329568449407816, + 0.0025310134515166283, + 0.0037056482397019863, + 0.003392399288713932, + 0.004522988572716713, + 0.003072769846767187, + 0.0037409167271107435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.3134418725967407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an elegantly decorated Chinese restaurant, a static, eye-level wide-angle shot captures a dynamic social scene, unfolding around a large round table. The warm light from ornate overhead lanterns casts a soft glow over the gathering. To the right, a woman in a crisp white silk blouse engages in conversation, smiles slightly, turns her head, and elegantly smooths back her tousled black hair. Next to her, a woman in a vibrant red dress holds a lit cigarette between her fingers, a wisp of smoke curling toward the ceiling. She leans over and gives a gentle, encouraging pat on the arm of a man in a simple black sweater. Across the table, a man in a forest green jacket deliberately bends down, his upper body disappearing from the frame. He searches the floor for something before slowly returning to his seat. On the far left, a woman with hair in a neat bun, absorbed in her meal, steadily picks up dark wooden chopsticks and reaches forward to select a piece of food from the steaming plate in the center. Meanwhile, a woman in a black turtleneck quietly raises a clear glass to her lips to take a sip of water, then subtly shifts her weight to lean forward across the table, her posture signaling growing interest in the conversation, all contributing to a rich, layered, and naturally casual, friendly dinner atmosphere.", + "seed": 18207, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_1/index_7552_seed_18207_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15521160289645194, + "vlm_logprob": -0.15521160289645194, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.15521160289645194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03922491893172264 + ], + "opposite_logprob_tokens": [ + -3.789224863052368 + ], + "logprob_reward": -0.03922491893172264, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03922491893172264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36569005250930786, + 0.02657991275191307, + 0.0477500781416893, + 0.03146496042609215, + 0.02655060775578022, + 0.046562932431697845, + 0.1792941391468048, + 0.15321847796440125, + 0.07048054039478302, + 0.05240831896662712 + ], + "credit_map": [ + 0.02636258862912655, + 0.05089889094233513, + 0.019830966368317604, + 0.02535703219473362, + 0.07074501365423203, + 0.026392601430416107, + 0.028622707352042198, + 0.11748024821281433, + 0.002859341213479638, + 0.0030996135901659727, + 0.0025895321741700172, + 0.004577870946377516, + 0.005169089883565903, + 0.003033675719052553, + 0.0027257902547717094, + 0.0025249989703297615, + 0.004248269367963076, + 0.005575934890657663, + 0.0025202184915542603, + 0.011206150986254215, + 0.004931643605232239, + 0.006928372196853161, + 0.006867704913020134, + 0.005471780896186829, + 0.003597741248086095, + 0.0034092299174517393, + 0.0032017717603594065, + 0.003611526917666197, + 0.0049048010259866714, + 0.004225073382258415, + 0.004592770244926214, + 0.003922044299542904, + 0.0024704099632799625, + 0.0019503935473039746, + 0.002899087266996503, + 0.002027226146310568, + 0.005180520936846733, + 0.002323488937690854, + 0.0032616041135042906, + 0.0064378781244158745, + 0.0034028273075819016, + 0.0029060898814350367, + 0.0052367704920470715, + 0.007914627902209759, + 0.005286311265081167, + 0.0045486451126635075, + 0.010428681038320065, + 0.006838980596512556, + 0.019223017618060112, + 0.010953129269182682, + 0.046239327639341354, + 0.018427545204758644, + 0.012508315034210682, + 0.01283251028507948, + 0.033023249357938766, + 0.026087043806910515, + 0.006752189248800278, + 0.012027484364807606, + 0.010132632218301296, + 0.014509658329188824, + 0.03426460176706314, + 0.026379944756627083, + 0.026804637163877487, + 0.022347329184412956, + 0.00590342003852129, + 0.007994763553142548, + 0.007110154256224632, + 0.008719570003449917, + 0.012681784108281136, + 0.008826167322695255, + 0.008945531211793423, + 0.01029914990067482, + 0.003877675626426935, + 0.0029439805075526237, + 0.005268493667244911, + 0.006265072617679834, + 0.003978782333433628, + 0.010380481369793415, + 0.005433875601738691, + 0.014259958639740944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.03922491893172264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289310872554779 + ], + "opposite_logprob_tokens": [ + -1.0539311170578003 + ], + "logprob_reward": -0.4289310872554779, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289310872554779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12203410267829895, + 0.03997788950800896, + 0.028711266815662384, + 0.048046745359897614, + 0.08847982436418533, + 0.05587198585271835, + 0.15036486089229584, + 0.2606387138366699, + 0.1538645476102829, + 0.052010077983140945 + ], + "credit_map": [ + 0.009217066690325737, + 0.01394228357821703, + 0.0051871067844331264, + 0.006909562740474939, + 0.028250902891159058, + 0.01286925096064806, + 0.009162701666355133, + 0.03649522736668587, + 0.003911667037755251, + 0.003319225739687681, + 0.005222618114203215, + 0.004364163614809513, + 0.010265643708407879, + 0.0053911833092570305, + 0.004914195742458105, + 0.0025891910772770643, + 0.0036687469109892845, + 0.002967446343973279, + 0.003028597915545106, + 0.003390963887795806, + 0.004860249347984791, + 0.005804417189210653, + 0.00207161926664412, + 0.002919226884841919, + 0.006649558898061514, + 0.00799595471471548, + 0.005605045706033707, + 0.005079262424260378, + 0.007020661141723394, + 0.005630316212773323, + 0.0044548772275447845, + 0.005611072760075331, + 0.0074831824749708176, + 0.005667153745889664, + 0.014364064671099186, + 0.00592642230913043, + 0.0082622105255723, + 0.010357324033975601, + 0.0051956200040876865, + 0.031223848462104797, + 0.005755763500928879, + 0.004288701806217432, + 0.007491271011531353, + 0.008317016065120697, + 0.005927947349846363, + 0.00668216310441494, + 0.010194774717092514, + 0.007214346434921026, + 0.013372381217777729, + 0.006755576469004154, + 0.029635170474648476, + 0.009503785520792007, + 0.018482595682144165, + 0.015187930315732956, + 0.035942092537879944, + 0.021485334262251854, + 0.016866551712155342, + 0.025595087558031082, + 0.019850438460707664, + 0.028054606169462204, + 0.041971709579229355, + 0.03982308506965637, + 0.04701874032616615, + 0.04145848751068115, + 0.012079480104148388, + 0.016856934875249863, + 0.014962123706936836, + 0.0203293114900589, + 0.026407551020383835, + 0.022051706910133362, + 0.015365030616521835, + 0.02581242099404335, + 0.005455461330711842, + 0.004950038623064756, + 0.008337373845279217, + 0.005183526314795017, + 0.006882151123136282, + 0.008113907650113106, + 0.005122218281030655, + 0.007965397089719772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.4289310872554779 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336986780166626 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.0336986780166626, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336986780166626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2305270880460739, + 0.07648789137601852, + 0.07889678329229355, + 0.16611775755882263, + 0.10928872972726822, + 0.05257635563611984, + 0.11073554307222366, + 0.05674051493406296, + 0.05477745085954666, + 0.06385196000337601 + ], + "credit_map": [ + 0.02450377680361271, + 0.021220294758677483, + 0.020972657948732376, + 0.017395401373505592, + 0.04398771747946739, + 0.02033228799700737, + 0.01832839660346508, + 0.06378653645515442, + 0.008454480208456516, + 0.006896328646689653, + 0.005769132170826197, + 0.007267859298735857, + 0.025813868269324303, + 0.01140207052230835, + 0.006142826285213232, + 0.004741317126899958, + 0.01085811946541071, + 0.012020016089081764, + 0.011289287358522415, + 0.010474033653736115, + 0.011752813123166561, + 0.009213244542479515, + 0.006870533339679241, + 0.006418723613023758, + 0.015816764906048775, + 0.0195147767663002, + 0.016296271234750748, + 0.019109122455120087, + 0.03996530920267105, + 0.02009633742272854, + 0.019324956461787224, + 0.015994198620319366, + 0.010032149963080883, + 0.010612537153065205, + 0.008452901616692543, + 0.00756025779992342, + 0.03318508341908455, + 0.011242534965276718, + 0.012332478538155556, + 0.01587078906595707, + 0.00455271964892745, + 0.0038075107149779797, + 0.005796432960778475, + 0.00972716137766838, + 0.0077502853237092495, + 0.006076163612306118, + 0.008680171333253384, + 0.0061859083361923695, + 0.00837637484073639, + 0.005998591426759958, + 0.014289644546806812, + 0.008045683614909649, + 0.02115396410226822, + 0.016664555296301842, + 0.020413881167769432, + 0.015792835503816605, + 0.004292956553399563, + 0.005935687571763992, + 0.008507667109370232, + 0.0076469238847494125, + 0.006479727104306221, + 0.011031568050384521, + 0.004968041088432074, + 0.007877943105995655, + 0.004389883019030094, + 0.0037500134203583, + 0.007710843347012997, + 0.00793307926505804, + 0.0066852872259914875, + 0.008319174870848656, + 0.007032160647213459, + 0.00895700417459011, + 0.004551994148641825, + 0.004512763116508722, + 0.004556129686534405, + 0.00499792629852891, + 0.007283720187842846, + 0.013146301731467247, + 0.007473064586520195, + 0.017330050468444824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0336986780166626 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866102337837219 + ], + "opposite_logprob_tokens": [ + -3.04866099357605 + ], + "logprob_reward": -0.04866102337837219, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866102337837219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34107306599617004, + 0.0670357197523117, + 0.04539443179965019, + 0.1420988291501999, + 0.04862057417631149, + 0.09500659257173538, + 0.10596457868814468, + 0.0435047447681427, + 0.06888744235038757, + 0.04241406172513962 + ], + "credit_map": [ + 0.04258841276168823, + 0.0566663034260273, + 0.034971173852682114, + 0.03245560824871063, + 0.040935833007097244, + 0.06046751141548157, + 0.021468596532940865, + 0.051519643515348434, + 0.0058930725790560246, + 0.006268484517931938, + 0.004501172807067633, + 0.006956118158996105, + 0.02273673377931118, + 0.010134179145097733, + 0.0058469926007092, + 0.004698967561125755, + 0.006365625187754631, + 0.007485663518309593, + 0.006155365612357855, + 0.006865309551358223, + 0.008150716312229633, + 0.004094688221812248, + 0.002790311351418495, + 0.003486750414595008, + 0.011087183840572834, + 0.017146049067378044, + 0.012805577367544174, + 0.013775424100458622, + 0.03341325372457504, + 0.017331067472696304, + 0.013424207456409931, + 0.02311606891453266, + 0.004107330460101366, + 0.004357160534709692, + 0.0025181425735354424, + 0.0022849293891340494, + 0.012187156826257706, + 0.005128700286149979, + 0.004473139066249132, + 0.013564012013375759, + 0.007439821492880583, + 0.0059260958805680275, + 0.011218318715691566, + 0.01592121459543705, + 0.01077344361692667, + 0.009288609027862549, + 0.02199818193912506, + 0.012440904974937439, + 0.011474138125777245, + 0.005249373614788055, + 0.02198408544063568, + 0.009334697388112545, + 0.011998942121863365, + 0.007922275923192501, + 0.02266133576631546, + 0.015339733101427555, + 0.0023228947538882494, + 0.0027407046873122454, + 0.003506253706291318, + 0.0046700602397322655, + 0.008403046056628227, + 0.009296282194554806, + 0.0059160590171813965, + 0.00664944713935256, + 0.003735933220013976, + 0.006129818968474865, + 0.004435008391737938, + 0.006866432260721922, + 0.018158921971917152, + 0.01188874151557684, + 0.006838258355855942, + 0.010834330692887306, + 0.004824620205909014, + 0.003623611992225051, + 0.0032588737085461617, + 0.004899360239505768, + 0.00466458685696125, + 0.0068179890513420105, + 0.005361067596822977, + 0.008963951840996742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.04866102337837219 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an elegantly decorated Chinese restaurant with a large round table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static, eye-level, and wide-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in a crisp white silk blouse visible who smooths back her tousled black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman in a vibrant red dress holding a lit cigarette with smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a man in a forest green jacket visible bending down to search the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22554230690002441 + ], + "opposite_logprob_tokens": [ + -1.6005423069000244 + ], + "logprob_reward": -0.22554230690002441, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554230690002441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10375185310840607, + 0.04362493008375168, + 0.045624736696481705, + 0.1016264408826828, + 0.04253518581390381, + 0.38832712173461914, + 0.06863299012184143, + 0.04509499669075012, + 0.10934045910835266, + 0.0514412559568882 + ], + "credit_map": [ + 0.012828155420720577, + 0.01878604292869568, + 0.01033058576285839, + 0.009984182193875313, + 0.014428660273551941, + 0.01634271815419197, + 0.007948308251798153, + 0.013103200122714043, + 0.007100079674273729, + 0.004330534487962723, + 0.0038218023255467415, + 0.006272267084568739, + 0.0057959905825555325, + 0.005188752431422472, + 0.00740013225004077, + 0.0037153735756874084, + 0.007447624113410711, + 0.003877229755744338, + 0.004270289558917284, + 0.006116112228482962, + 0.007055866997689009, + 0.004896641708910465, + 0.006738698575645685, + 0.005222272593528032, + 0.009738311171531677, + 0.009162972681224346, + 0.012103545479476452, + 0.01329744327813387, + 0.02048550732433796, + 0.011050176806747913, + 0.015047695487737656, + 0.01074078120291233, + 0.005397787783294916, + 0.004067385569214821, + 0.003139438107609749, + 0.0033640279434621334, + 0.008267662487924099, + 0.00510721979662776, + 0.005826439242810011, + 0.007365223485976458, + 0.028105786070227623, + 0.02935761585831642, + 0.029393009841442108, + 0.05325893312692642, + 0.046793706715106964, + 0.03946441039443016, + 0.11146412044763565, + 0.050489552319049835, + 0.00673008244484663, + 0.0034180781804025173, + 0.012139606289565563, + 0.005612393841147423, + 0.007256826385855675, + 0.011179214343428612, + 0.0135577991604805, + 0.00873898807913065, + 0.0025338861159980297, + 0.004441625904291868, + 0.003653738647699356, + 0.006366430781781673, + 0.007699479814618826, + 0.007632547523826361, + 0.006463825702667236, + 0.006303465459495783, + 0.008588924072682858, + 0.009483199566602707, + 0.014132039621472359, + 0.013327431865036488, + 0.019005946815013885, + 0.015790551900863647, + 0.012452788650989532, + 0.016559578478336334, + 0.006839517969638109, + 0.004915062338113785, + 0.00615678820759058, + 0.00598045764490962, + 0.0052050454542040825, + 0.0073637100867927074, + 0.006186886224895716, + 0.008793791756033897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.22554230690002441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_tensor([5617, 5617])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_tensor([5617, 5617])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cac3a419145e86828cc92fae937126e1c6516516 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_tensor([5617, 5617])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely close-up, static shot focuses with razor-sharp clarity on the angled tip of a sterile, stainless-steel hypodermic needle. The polished metal gleams in bright, direct light, revealing microscopic striations from its manufacturing process. Suddenly, a clear, viscous liquid powerfully erupts from the narrow gap in a swift and decisive eruption. The syrupy liquid catches the light, shimmering and forming a perfect arc against the cool, softly blurred, sterile blue-white background. The initial, cohesive spray of liquid lingers in the air for a moment before gracefully breaking into shimmering, individual droplets that follow the same trajectory. The entire scene is illuminated by cool, high-key lighting, creating sharp, mirror-like highlights on the needle and making each droplet shimmer, creating an atmosphere of intense focus, clinical precision, and controlled force.", + "seed": 225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_5617_seed_225_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22584435776807368, + "vlm_logprob": -0.22584435776807368, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.22584435776807368, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01801755093038082 + ], + "opposite_logprob_tokens": [ + -4.143017768859863 + ], + "logprob_reward": -0.01801755093038082, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01801755093038082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26666757464408875, + 0.1512652337551117, + 0.05714016780257225, + 0.08177342265844345, + 0.13830456137657166, + 0.04422534257173538, + 0.08095355331897736, + 0.10268202424049377, + 0.047680795192718506, + 0.02930736169219017 + ], + "credit_map": [ + 0.04774165526032448, + 0.028965501114726067, + 0.022924095392227173, + 0.024988938122987747, + 0.033687494695186615, + 0.04059475660324097, + 0.03484907001256943, + 0.03291603922843933, + 0.02265329100191593, + 0.0163570549339056, + 0.0071073658764362335, + 0.009163557551801205, + 0.02824159897863865, + 0.02170672081410885, + 0.028715843334794044, + 0.01731978915631771, + 0.009144187904894352, + 0.007622487843036652, + 0.004720859695225954, + 0.006118260324001312, + 0.00619182363152504, + 0.008670436218380928, + 0.006804953329265118, + 0.007867158390581608, + 0.016115201637148857, + 0.00973968394100666, + 0.005315200891345739, + 0.008705941028892994, + 0.004719629418104887, + 0.012174389325082302, + 0.01515294425189495, + 0.009850428439676762, + 0.022570384666323662, + 0.01968388631939888, + 0.009485180489718914, + 0.023550555109977722, + 0.0123665202409029, + 0.012968813069164753, + 0.01978035271167755, + 0.01789885014295578, + 0.00946776382625103, + 0.004891865886747837, + 0.0030427551828324795, + 0.005505169741809368, + 0.0029484767001122236, + 0.005562038626521826, + 0.007345675025135279, + 0.005461595486849546, + 0.015178539790213108, + 0.012691676616668701, + 0.004494204185903072, + 0.01139301061630249, + 0.00630867900326848, + 0.008339045569300652, + 0.011463195085525513, + 0.01108519360423088, + 0.014672668650746346, + 0.010948670096695423, + 0.006928533315658569, + 0.013162298128008842, + 0.013629208318889141, + 0.013898038305342197, + 0.018010012805461884, + 0.011432582512497902, + 0.008584356866776943, + 0.005103518720716238, + 0.004063697997480631, + 0.006581447087228298, + 0.004118496086448431, + 0.005536227487027645, + 0.007193920202553272, + 0.006499127484858036, + 0.006625236477702856, + 0.0028173031751066446, + 0.0020802582148462534, + 0.003529428504407406, + 0.0030599497258663177, + 0.00310877594165504, + 0.004419778008013964, + 0.0036666307132691145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01801755093038082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578552961349487 + ], + "opposite_logprob_tokens": [ + -0.6328552961349487 + ], + "logprob_reward": -0.7578552961349487, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578552961349487, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4238535761833191, + 0.06900613754987717, + 0.04792025312781334, + 0.09940441697835922, + 0.19137372076511383, + 0.04187753051519394, + 0.04534153640270233, + 0.024218516424298286, + 0.034212030470371246, + 0.02279219776391983 + ], + "credit_map": [ + 0.0648062601685524, + 0.049428924918174744, + 0.031390342861413956, + 0.04182947799563408, + 0.05967153608798981, + 0.07043357938528061, + 0.05739828944206238, + 0.04889519512653351, + 0.01112311240285635, + 0.008133675903081894, + 0.0038426921237260103, + 0.00746398838236928, + 0.011073209345340729, + 0.008382122963666916, + 0.008919925428926945, + 0.010067412629723549, + 0.0068598417565226555, + 0.005976326763629913, + 0.002722467528656125, + 0.004595979582518339, + 0.005608174949884415, + 0.009764856658875942, + 0.0063171545043587685, + 0.006075458135455847, + 0.015046466141939163, + 0.010899248532950878, + 0.005280929617583752, + 0.012819400988519192, + 0.007426262833178043, + 0.01808561384677887, + 0.01679048128426075, + 0.013056023046374321, + 0.033386558294296265, + 0.026945842429995537, + 0.011441146954894066, + 0.02907106652855873, + 0.016945382580161095, + 0.021184947341680527, + 0.025404423475265503, + 0.0269943717867136, + 0.00975196622312069, + 0.005169052630662918, + 0.002532124985009432, + 0.0049646152183413506, + 0.0019656450022011995, + 0.005725343711674213, + 0.006406558211892843, + 0.005362225230783224, + 0.00989446323364973, + 0.00664727995172143, + 0.003863064805045724, + 0.00501293083652854, + 0.003463460598140955, + 0.004659123718738556, + 0.006026450078934431, + 0.0057747685350477695, + 0.004848180338740349, + 0.0026710445526987314, + 0.0018792084883898497, + 0.0027237553149461746, + 0.002190382219851017, + 0.0031463506165891886, + 0.003273722715675831, + 0.0034858740400522947, + 0.008026441559195518, + 0.003296921495348215, + 0.0026839023921638727, + 0.002948571927845478, + 0.0019312890945002437, + 0.003706573508679867, + 0.006737354211509228, + 0.004880979657173157, + 0.003546101739630103, + 0.002344421809539199, + 0.0015053710667416453, + 0.0026836644392460585, + 0.0025409667287021875, + 0.002842354355379939, + 0.004588528070598841, + 0.0027407906018197536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.7578552961349487 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136661648750305 + ], + "opposite_logprob_tokens": [ + -1.4063665866851807 + ], + "logprob_reward": -0.28136661648750305, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136661648750305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24034009873867035, + 0.27069589495658875, + 0.05680359527468681, + 0.05439292639493942, + 0.20325496792793274, + 0.02420402690768242, + 0.05882358178496361, + 0.020980123430490494, + 0.04159987345337868, + 0.02890482358634472 + ], + "credit_map": [ + 0.03274054452776909, + 0.027866870164871216, + 0.018701404333114624, + 0.02135676145553589, + 0.042568691074848175, + 0.04042603448033333, + 0.02934250980615616, + 0.02733728475868702, + 0.04103398323059082, + 0.025310609489679337, + 0.010217014700174332, + 0.019101593643426895, + 0.04484355449676514, + 0.03297479450702667, + 0.06704837828874588, + 0.03016599640250206, + 0.006849108729511499, + 0.006667097564786673, + 0.0033826103899627924, + 0.004947499837726355, + 0.007872853428125381, + 0.011861637234687805, + 0.007862349040806293, + 0.0073604462668299675, + 0.008125171065330505, + 0.006417136173695326, + 0.004201599862426519, + 0.007297905161976814, + 0.004246735945343971, + 0.010112975724041462, + 0.007310789544135332, + 0.006680616643279791, + 0.03645121678709984, + 0.02841872163116932, + 0.012044001370668411, + 0.03257323056459427, + 0.01653158664703369, + 0.02060052752494812, + 0.02759638987481594, + 0.029039308428764343, + 0.004026153590530157, + 0.0025336628314107656, + 0.001972984289750457, + 0.0035206901375204325, + 0.002087753964588046, + 0.003941533621400595, + 0.0027891083154827356, + 0.003332142485305667, + 0.01279974915087223, + 0.009653052315115929, + 0.004211300518363714, + 0.006871600169688463, + 0.003937559202313423, + 0.005429120268672705, + 0.008193686604499817, + 0.007727518677711487, + 0.004237338434904814, + 0.002686124760657549, + 0.0016744796885177493, + 0.002614695578813553, + 0.002190452069044113, + 0.002382236998528242, + 0.0024945924524217844, + 0.0027002061251550913, + 0.008350948803126812, + 0.0047677503898739815, + 0.003536720061674714, + 0.003904435783624649, + 0.0027665370143949986, + 0.005010312423110008, + 0.007618690375238657, + 0.005644483026117086, + 0.004948069807142019, + 0.0030322622042149305, + 0.0016146170673891902, + 0.003388670738786459, + 0.0032869651913642883, + 0.0037583124358206987, + 0.005428109783679247, + 0.0034478185698390007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.28136661648750305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019403931219130754 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.0019403931219130754, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019403931219130754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13438911736011505, + 0.13031314313411713, + 0.12917934358119965, + 0.0352717787027359, + 0.13599620759487152, + 0.055480655282735825, + 0.06593158841133118, + 0.09820736199617386, + 0.028956467285752296, + 0.18627436459064484 + ], + "credit_map": [ + 0.02766299992799759, + 0.020039625465869904, + 0.011305911466479301, + 0.013917500153183937, + 0.013353918679058552, + 0.015323443338274956, + 0.016202200204133987, + 0.01658349484205246, + 0.02097924053668976, + 0.013402726501226425, + 0.008497260510921478, + 0.01009458489716053, + 0.019649317488074303, + 0.017470398917794228, + 0.02203226648271084, + 0.018187347799539566, + 0.026950653642416, + 0.015462118200957775, + 0.008120476268231869, + 0.016010725870728493, + 0.007522611413151026, + 0.01598585955798626, + 0.022832471877336502, + 0.016294419765472412, + 0.006551924627274275, + 0.0056525543332099915, + 0.002137151313945651, + 0.0044409167021512985, + 0.0023488595616072416, + 0.0045237173326313496, + 0.005247382912784815, + 0.00436927005648613, + 0.024200862273573875, + 0.017393410205841064, + 0.00800362043082714, + 0.020935915410518646, + 0.010480408556759357, + 0.014187691733241081, + 0.02312474511563778, + 0.017669541761279106, + 0.013130247592926025, + 0.006061687134206295, + 0.004020385909825563, + 0.008817829191684723, + 0.004145370330661535, + 0.005418493412435055, + 0.006117550190538168, + 0.00776908453553915, + 0.016080759465694427, + 0.010215332731604576, + 0.005261403042823076, + 0.007184696849435568, + 0.0036199172027409077, + 0.005977519787847996, + 0.00915667973458767, + 0.00843527540564537, + 0.022978613153100014, + 0.016263283789157867, + 0.010760584846138954, + 0.010332095436751842, + 0.0072458176873624325, + 0.008996431715786457, + 0.010118973441421986, + 0.011511551216244698, + 0.0072013684548437595, + 0.004338810686022043, + 0.0021658989135175943, + 0.003366622840985656, + 0.001980474218726158, + 0.002627384616062045, + 0.002945623593404889, + 0.004330281168222427, + 0.041469067335128784, + 0.02472822740674019, + 0.008826730772852898, + 0.015433979220688343, + 0.01822715625166893, + 0.016489095985889435, + 0.035295095294713974, + 0.02580498717725277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.0019403931219130754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004193216562271 + ], + "opposite_logprob_tokens": [ + -2.6950418949127197 + ], + "logprob_reward": -0.07004193216562271, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004193216562271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2258967012166977, + 0.22885961830615997, + 0.05499265715479851, + 0.09130409359931946, + 0.14698754251003265, + 0.05389276146888733, + 0.06487593799829483, + 0.03327960520982742, + 0.07030181586742401, + 0.029609231278300285 + ], + "credit_map": [ + 0.03115835227072239, + 0.0257951058447361, + 0.013628820888698101, + 0.01740661822259426, + 0.04258360713720322, + 0.03948076069355011, + 0.030749309808015823, + 0.02509414218366146, + 0.04225343093276024, + 0.023805558681488037, + 0.008514338172972202, + 0.011793006211519241, + 0.039792101830244064, + 0.027753619477152824, + 0.051602624356746674, + 0.023344963788986206, + 0.0111000407487154, + 0.0061771320179104805, + 0.003793990006670356, + 0.00615379773080349, + 0.004286560695618391, + 0.008831650018692017, + 0.008316225372254848, + 0.0063332607969641685, + 0.018155081197619438, + 0.011418940499424934, + 0.005989577155560255, + 0.009704398922622204, + 0.005839415825903416, + 0.01462186686694622, + 0.01560069527477026, + 0.00997412670403719, + 0.026365414261817932, + 0.02171003445982933, + 0.008390353992581367, + 0.02423851750791073, + 0.012024709023535252, + 0.013867857865989208, + 0.02046358399093151, + 0.019927097484469414, + 0.007023318205028772, + 0.00640319287776947, + 0.004942797590047121, + 0.008829157799482346, + 0.004533009137958288, + 0.007956478744745255, + 0.006990375928580761, + 0.007214439567178488, + 0.013540408574044704, + 0.010686984285712242, + 0.004523471929132938, + 0.006528740283101797, + 0.0050928471609950066, + 0.006992679554969072, + 0.008717896416783333, + 0.008792918175458908, + 0.004386858083307743, + 0.003496905555948615, + 0.0025159416254609823, + 0.005345284007489681, + 0.004575422499328852, + 0.004357220139354467, + 0.0044860546477139, + 0.004115920513868332, + 0.014738420024514198, + 0.009003980085253716, + 0.005516080651432276, + 0.0063851564191281796, + 0.004904675297439098, + 0.008626826107501984, + 0.011918281204998493, + 0.009208405390381813, + 0.004437931813299656, + 0.002902918728068471, + 0.002554580569267273, + 0.004005421418696642, + 0.0042225089855492115, + 0.0031745980959385633, + 0.004326910246163607, + 0.003984365146607161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.07004193216562271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close-up, static shot focuses with razor-sharp clarity on the angled tip of a sterile, stainless-steel hypodermic needle. The polished metal gleams in bright, direct light, revealing microscopic striations from its manufacturing process. Suddenly, a clear, viscous liquid powerfully erupts from the narrow gap in a swift and decisive eruption. The syrupy liquid catches the light, shimmering and forming a perfect arc against the cool, softly blurred, sterile blue-white background. The initial, cohesive spray of liquid lingers in the air for a moment before gracefully breaking into shimmering, individual droplets that follow the same trajectory. The entire scene is illuminated by cool, high-key lighting, creating sharp, mirror-like highlights on the needle and making each droplet shimmer, creating an atmosphere of intense focus, clinical precision, and controlled force.", + "seed": 226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_5617_seed_226_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3350769271841273, + "vlm_logprob": -0.3350769271841273, + "vlm_yes_no_margin": 2.624999976158142, + "vlm_token_credit": -0.3350769271841273, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015761205926537514 + ], + "opposite_logprob_tokens": [ + -4.640761375427246 + ], + "logprob_reward": -0.015761205926537514, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015761205926537514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20052604377269745, + 0.15024703741073608, + 0.09374552220106125, + 0.03027142584323883, + 0.08019509166479111, + 0.06008538603782654, + 0.05087319016456604, + 0.2215648889541626, + 0.049317024648189545, + 0.06317437440156937 + ], + "credit_map": [ + 0.016418706625699997, + 0.028398912400007248, + 0.03833521902561188, + 0.021969789639115334, + 0.030796576291322708, + 0.028746243566274643, + 0.022471698001027107, + 0.013388905674219131, + 0.018391741439700127, + 0.01661381870508194, + 0.020246250554919243, + 0.030342038720846176, + 0.01324296835809946, + 0.02505507320165634, + 0.012382036074995995, + 0.013973117806017399, + 0.012924781069159508, + 0.012077564373612404, + 0.014828378334641457, + 0.015524311922490597, + 0.008679640479385853, + 0.014406534843146801, + 0.009840192273259163, + 0.005464108660817146, + 0.005258006509393454, + 0.002844903152436018, + 0.00517005380243063, + 0.002893095836043358, + 0.0028158060740679502, + 0.004816009663045406, + 0.002938095945864916, + 0.0035354571882635355, + 0.010757611133158207, + 0.007727551739662886, + 0.013515650294721127, + 0.009015442803502083, + 0.00871981680393219, + 0.011594215407967567, + 0.009415275417268276, + 0.009449525736272335, + 0.009079904295504093, + 0.004412283189594746, + 0.012586671859025955, + 0.0050164987333118916, + 0.0050084274262189865, + 0.009922954253852367, + 0.006467616185545921, + 0.007591035217046738, + 0.007089938968420029, + 0.006259474437683821, + 0.008012917824089527, + 0.0056947460398077965, + 0.006714052986353636, + 0.006278689019382, + 0.00586582999676466, + 0.004957541357725859, + 0.036631785333156586, + 0.03252600505948067, + 0.028807049617171288, + 0.02430271916091442, + 0.022991228848695755, + 0.02482125721871853, + 0.0348464772105217, + 0.016638364642858505, + 0.011402533389627934, + 0.005853619892150164, + 0.0072187623009085655, + 0.004085996188223362, + 0.005788599606603384, + 0.00605569826439023, + 0.0059695178642869, + 0.0029423004016280174, + 0.007535828743129969, + 0.008578063920140266, + 0.008202094584703445, + 0.006384801585227251, + 0.006917377933859825, + 0.009257860481739044, + 0.012226521968841553, + 0.004071823321282864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.015761205926537514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329367756843567 + ], + "opposite_logprob_tokens": [ + -0.7579367756843567 + ], + "logprob_reward": -0.6329367756843567, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329367756843567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10892603546380997, + 0.07159093767404556, + 0.08424749970436096, + 0.10114730894565582, + 0.040736276656389236, + 0.07524263858795166, + 0.04908929020166397, + 0.0495305173099041, + 0.09561984986066818, + 0.32386961579322815 + ], + "credit_map": [ + 0.0104976212605834, + 0.009726488031446934, + 0.013741249218583107, + 0.016094202175736427, + 0.011714117601513863, + 0.01625015027821064, + 0.01999199204146862, + 0.010910216718912125, + 0.009608149528503418, + 0.0078867357224226, + 0.013552538119256496, + 0.010339269414544106, + 0.007048861589282751, + 0.009338094852864742, + 0.007667729165405035, + 0.0061495606787502766, + 0.012244652956724167, + 0.006224018521606922, + 0.017046010121703148, + 0.010155778378248215, + 0.005712500307708979, + 0.012943943962454796, + 0.009564923122525215, + 0.010355670005083084, + 0.021621637046337128, + 0.007724049035459757, + 0.02276514284312725, + 0.009175053797662258, + 0.008061904460191727, + 0.01473279669880867, + 0.010060977190732956, + 0.00700575252994895, + 0.005519682541489601, + 0.0037596169859170914, + 0.006925949361175299, + 0.00465222354978323, + 0.00425812928006053, + 0.0054145678877830505, + 0.006529170088469982, + 0.0036769367288798094, + 0.015542431734502316, + 0.006380363833159208, + 0.01062828954309225, + 0.00853279884904623, + 0.005101968068629503, + 0.008808190003037453, + 0.01265603769570589, + 0.007592560723423958, + 0.008749491535127163, + 0.004832619801163673, + 0.0072237285785377026, + 0.005256482865661383, + 0.005236979108303785, + 0.0055096810683608055, + 0.006936672609299421, + 0.0053436351008713245, + 0.00888407975435257, + 0.005822160746902227, + 0.006853243336081505, + 0.004818516317754984, + 0.005420542787760496, + 0.006621532142162323, + 0.007633717730641365, + 0.003476723562926054, + 0.019211139529943466, + 0.01130636315792799, + 0.014202119782567024, + 0.008742527104914188, + 0.008475271984934807, + 0.014833624474704266, + 0.012289654463529587, + 0.0065591479651629925, + 0.03249645233154297, + 0.03586621955037117, + 0.048024047166109085, + 0.033111874014139175, + 0.03656763210892677, + 0.05773976817727089, + 0.05753764510154724, + 0.022525981068611145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.6329367756843567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983877897262573 + ], + "opposite_logprob_tokens": [ + -0.5233877897262573 + ], + "logprob_reward": -0.8983877897262573, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983877897262573, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11103222519159317, + 0.06882321089506149, + 0.07208604365587234, + 0.0913221463561058, + 0.03958677873015404, + 0.09929975867271423, + 0.0612679198384285, + 0.04805827885866165, + 0.07391361892223358, + 0.3346100449562073 + ], + "credit_map": [ + 0.017532652243971825, + 0.013809053227305412, + 0.015978291630744934, + 0.01587732881307602, + 0.01362849771976471, + 0.014555206522345543, + 0.009641779586672783, + 0.010009420104324818, + 0.007939163595438004, + 0.007248690351843834, + 0.007185709197074175, + 0.009533336386084557, + 0.007851304486393929, + 0.009747699834406376, + 0.012639581225812435, + 0.006677725818008184, + 0.010256808251142502, + 0.004651710856705904, + 0.016155006363987923, + 0.007151422090828419, + 0.004809004720300436, + 0.011185981333255768, + 0.007649346254765987, + 0.010226760990917683, + 0.020598649978637695, + 0.006425174418836832, + 0.020222382619976997, + 0.008177559822797775, + 0.006903430912643671, + 0.013326671905815601, + 0.009182363748550415, + 0.006485918536782265, + 0.006657992489635944, + 0.003456647042185068, + 0.007297044154256582, + 0.004505213350057602, + 0.004137812182307243, + 0.005910578183829784, + 0.004479947034269571, + 0.003141543595120311, + 0.02500738948583603, + 0.008767199702560902, + 0.0163528211414814, + 0.010331415571272373, + 0.006847058888524771, + 0.01183792669326067, + 0.013047107495367527, + 0.007108832709491253, + 0.009788202121853828, + 0.008384445682168007, + 0.009189841337502003, + 0.006791583728045225, + 0.00632914574816823, + 0.006801172625273466, + 0.008190291002392769, + 0.005793238058686256, + 0.006062354426831007, + 0.0054802740924060345, + 0.0077183861285448074, + 0.005094953812658787, + 0.004839797038584948, + 0.006945336703211069, + 0.008236679248511791, + 0.0036804943811148405, + 0.014571928419172764, + 0.008045872673392296, + 0.012593516148626804, + 0.005904129706323147, + 0.005828252527862787, + 0.012155559845268726, + 0.009965759702026844, + 0.004848599433898926, + 0.03318905457854271, + 0.037686556577682495, + 0.05225629732012749, + 0.032366055995225906, + 0.0381200835108757, + 0.06410141289234161, + 0.05605711042881012, + 0.0208334531635046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.8983877897262573 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011825718684121966 + ], + "opposite_logprob_tokens": [ + -6.751182556152344 + ], + "logprob_reward": -0.0011825718684121966, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011825718684121966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10272755473852158, + 0.05533166974782944, + 0.06048347428441048, + 0.09902428090572357, + 0.1449289470911026, + 0.11436229199171066, + 0.13573762774467468, + 0.06640983372926712, + 0.10145435482263565, + 0.1195400059223175 + ], + "credit_map": [ + 0.01115191075950861, + 0.012081225402653217, + 0.020675119012594223, + 0.012347831390798092, + 0.013410579413175583, + 0.014860759489238262, + 0.011045868508517742, + 0.007154260296374559, + 0.005626322235912085, + 0.006677249912172556, + 0.007705646567046642, + 0.008093328215181828, + 0.006586767733097076, + 0.008180155418813229, + 0.006757332943379879, + 0.005704864859580994, + 0.010463387705385685, + 0.004195317625999451, + 0.01382741704583168, + 0.004955874290317297, + 0.0047801584005355835, + 0.00913532916456461, + 0.006839274428784847, + 0.006286713294684887, + 0.02190457098186016, + 0.0073450724594295025, + 0.02224593050777912, + 0.009214424528181553, + 0.0076678842306137085, + 0.014935829676687717, + 0.009653842076659203, + 0.006056727841496468, + 0.014642097055912018, + 0.011330303736031055, + 0.02254326082766056, + 0.016205886378884315, + 0.013657080009579659, + 0.021010322496294975, + 0.02362985536456108, + 0.02191014215350151, + 0.016225384548306465, + 0.013131380081176758, + 0.019053015857934952, + 0.016192177310585976, + 0.009720728732645512, + 0.01539528276771307, + 0.014626813121140003, + 0.010017506778240204, + 0.016870513558387756, + 0.017519669607281685, + 0.02417529746890068, + 0.012763901613652706, + 0.014924352988600731, + 0.019665708765387535, + 0.01635243371129036, + 0.013465756550431252, + 0.015478783287107944, + 0.007430868223309517, + 0.010048912838101387, + 0.005914817098528147, + 0.004910188727080822, + 0.009832004085183144, + 0.007306339684873819, + 0.005487917922437191, + 0.020883938297629356, + 0.013202369213104248, + 0.017085080966353416, + 0.007418334949761629, + 0.009231165051460266, + 0.012447320856153965, + 0.014535832218825817, + 0.006650312338024378, + 0.012605657801032066, + 0.013381161727011204, + 0.016217239201068878, + 0.013011998496949673, + 0.014453078620135784, + 0.01872197911143303, + 0.022189229726791382, + 0.00895965751260519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.0011825718684121966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12711629271507263 + ], + "opposite_logprob_tokens": [ + -2.1271162033081055 + ], + "logprob_reward": -0.12711629271507263, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12711629271507263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16397333145141602, + 0.05713461712002754, + 0.06174523010849953, + 0.11208415776491165, + 0.04530349373817444, + 0.06713202595710754, + 0.08004886656999588, + 0.0397348590195179, + 0.07792907953262329, + 0.29491427540779114 + ], + "credit_map": [ + 0.027234520763158798, + 0.01846817322075367, + 0.028889277949929237, + 0.025813326239585876, + 0.01596020720899105, + 0.023370087146759033, + 0.010682218708097935, + 0.01355554349720478, + 0.006869771052151918, + 0.006580909714102745, + 0.0066886525601148605, + 0.010668465867638588, + 0.005608442705124617, + 0.008668052963912487, + 0.00697146775200963, + 0.005078858230262995, + 0.008988985791802406, + 0.004814798943698406, + 0.01282021589577198, + 0.006490142550319433, + 0.00416686711832881, + 0.009117675945162773, + 0.0069861700758337975, + 0.008360378444194794, + 0.024701906368136406, + 0.007993596605956554, + 0.02551644667983055, + 0.010027342475950718, + 0.008327019400894642, + 0.016936400905251503, + 0.01087147556245327, + 0.007709973491728306, + 0.007575297728180885, + 0.003800152102485299, + 0.0073995646089315414, + 0.005423466209322214, + 0.0037488453090190887, + 0.0066603138111531734, + 0.004861258901655674, + 0.0058345976285636425, + 0.0161177646368742, + 0.005077378358691931, + 0.009156553074717522, + 0.006560884416103363, + 0.0041061085648834705, + 0.007889957167208195, + 0.012002179399132729, + 0.0062212045304477215, + 0.014902953058481216, + 0.008669758215546608, + 0.011749286204576492, + 0.008525904268026352, + 0.009217821061611176, + 0.00979103147983551, + 0.00977017916738987, + 0.007421932648867369, + 0.007143322378396988, + 0.004615204408764839, + 0.005674582906067371, + 0.003915309906005859, + 0.0034976061433553696, + 0.005014757625758648, + 0.007348256651312113, + 0.0025258215609937906, + 0.013836429454386234, + 0.008915955200791359, + 0.0127632524818182, + 0.006877000909298658, + 0.0073367818258702755, + 0.013040132820606232, + 0.009864693507552147, + 0.0052948398515582085, + 0.028867237269878387, + 0.03091253899037838, + 0.047562725841999054, + 0.027776798233389854, + 0.03314666077494621, + 0.05903095379471779, + 0.049938030540943146, + 0.017679337412118912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.12711629271507263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close-up, static shot focuses with razor-sharp clarity on the angled tip of a sterile, stainless-steel hypodermic needle. The polished metal gleams in bright, direct light, revealing microscopic striations from its manufacturing process. Suddenly, a clear, viscous liquid powerfully erupts from the narrow gap in a swift and decisive eruption. The syrupy liquid catches the light, shimmering and forming a perfect arc against the cool, softly blurred, sterile blue-white background. The initial, cohesive spray of liquid lingers in the air for a moment before gracefully breaking into shimmering, individual droplets that follow the same trajectory. The entire scene is illuminated by cool, high-key lighting, creating sharp, mirror-like highlights on the needle and making each droplet shimmer, creating an atmosphere of intense focus, clinical precision, and controlled force.", + "seed": 227, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_5617_seed_227_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.24362965393811464, + "vlm_logprob": -0.24362965393811464, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.24362965393811464, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11568497866392136 + ], + "opposite_logprob_tokens": [ + -2.240684986114502 + ], + "logprob_reward": -0.11568497866392136, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11568497866392136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16711044311523438, + 0.03878249600529671, + 0.020612582564353943, + 0.08779920637607574, + 0.15066923201084137, + 0.06940177083015442, + 0.14990751445293427, + 0.10256022214889526, + 0.14047741889953613, + 0.07267911732196808 + ], + "credit_map": [ + 0.023821471258997917, + 0.025358019396662712, + 0.01548155490309, + 0.015701433643698692, + 0.02493749000132084, + 0.01769627071917057, + 0.016608957201242447, + 0.02750524878501892, + 0.004549440462142229, + 0.005463536828756332, + 0.004551879595965147, + 0.003941464237868786, + 0.005667116492986679, + 0.006136811804026365, + 0.0030043742153793573, + 0.005467873532325029, + 0.0028663615230470896, + 0.0028179639484733343, + 0.002483043819665909, + 0.0015493546379730105, + 0.003035715315490961, + 0.002844065660610795, + 0.0015530093805864453, + 0.00346306711435318, + 0.010414536111056805, + 0.013362876139581203, + 0.01024012640118599, + 0.01059053000062704, + 0.010220159776508808, + 0.01379707083106041, + 0.008377199992537498, + 0.01079670712351799, + 0.01966259442269802, + 0.02392663061618805, + 0.01611746847629547, + 0.01699247770011425, + 0.02362212911248207, + 0.017765533179044724, + 0.013841537758708, + 0.018740853294730186, + 0.006977434735745192, + 0.009898530319333076, + 0.008298877626657486, + 0.006264849565923214, + 0.012575558386743069, + 0.011665982194244862, + 0.005805519409477711, + 0.007915016263723373, + 0.017058707773685455, + 0.020050998777151108, + 0.019959064200520515, + 0.02274341695010662, + 0.01444303710013628, + 0.017822418361902237, + 0.019616927951574326, + 0.018212944269180298, + 0.007945307530462742, + 0.011791770346462727, + 0.013699664734303951, + 0.0120313111692667, + 0.0169067420065403, + 0.017875974997878075, + 0.012524958699941635, + 0.009784489870071411, + 0.014738999307155609, + 0.02170596644282341, + 0.02116136997938156, + 0.017203563824295998, + 0.015817947685718536, + 0.019193289801478386, + 0.011562082916498184, + 0.019094210118055344, + 0.0072148460894823074, + 0.011386667378246784, + 0.008718345314264297, + 0.013146510347723961, + 0.006698706652969122, + 0.011227251030504704, + 0.00661778450012207, + 0.007669006008654833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.11568497866392136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983858823776245 + ], + "opposite_logprob_tokens": [ + -0.5233858823776245 + ], + "logprob_reward": -0.8983858823776245, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983858823776245, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.070377878844738, + 0.06666475534439087, + 0.07114577293395996, + 0.099357008934021, + 0.11026237159967422, + 0.08785492926836014, + 0.09322311729192734, + 0.16151192784309387, + 0.2002401500940323, + 0.03936203569173813 + ], + "credit_map": [ + 0.009026430547237396, + 0.009741662070155144, + 0.005923384800553322, + 0.008059592917561531, + 0.008966761641204357, + 0.01197901926934719, + 0.006902804598212242, + 0.009778227657079697, + 0.010715781711041927, + 0.009276159107685089, + 0.00719827227294445, + 0.004773977678269148, + 0.009690199978649616, + 0.007918380200862885, + 0.007869674824178219, + 0.009222308173775673, + 0.007742878515273333, + 0.009007103741168976, + 0.008800667710602283, + 0.009038099087774754, + 0.014179127290844917, + 0.009059255011379719, + 0.005656378343701363, + 0.007662265095859766, + 0.014405143447220325, + 0.014955342747271061, + 0.00878935493528843, + 0.011585948057472706, + 0.009404288604855537, + 0.016083160415291786, + 0.010001125745475292, + 0.014132642187178135, + 0.016148585826158524, + 0.01599748060107231, + 0.009232189506292343, + 0.008776875212788582, + 0.014980176463723183, + 0.015109830535948277, + 0.011961786076426506, + 0.01805545575916767, + 0.01242038607597351, + 0.011422787792980671, + 0.008985103107988834, + 0.01194045227020979, + 0.010898196138441563, + 0.010766921564936638, + 0.01079979445785284, + 0.010621288791298866, + 0.013593902811408043, + 0.012909137643873692, + 0.010645871981978416, + 0.01336166076362133, + 0.008914203383028507, + 0.010155484080314636, + 0.011708674021065235, + 0.0119341891258955, + 0.025830065831542015, + 0.020289255306124687, + 0.023779818788170815, + 0.02841080166399479, + 0.0142107168212533, + 0.01784648932516575, + 0.0137988505885005, + 0.017345940694212914, + 0.02277502790093422, + 0.02795945294201374, + 0.02449275739490986, + 0.02611534483730793, + 0.022712163627147675, + 0.025864161550998688, + 0.020895037800073624, + 0.029426192864775658, + 0.005306095350533724, + 0.007391742430627346, + 0.005341873038560152, + 0.004191487096250057, + 0.004747865255922079, + 0.004026248585432768, + 0.002805570838972926, + 0.0055511509999632835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.8983858823776245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284726977348328 + ], + "opposite_logprob_tokens": [ + -2.0178472995758057 + ], + "logprob_reward": -0.14284726977348328, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284726977348328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12000647187232971, + 0.058322973549366, + 0.08183849602937698, + 0.0916021317243576, + 0.09402555227279663, + 0.09849603474140167, + 0.0486186258494854, + 0.18144699931144714, + 0.16337496042251587, + 0.06226780638098717 + ], + "credit_map": [ + 0.014944062568247318, + 0.018780482932925224, + 0.010260311886668205, + 0.014251958578824997, + 0.01242753304541111, + 0.014132032170891762, + 0.0148188890889287, + 0.020391197875142097, + 0.009056433103978634, + 0.0075177717953920364, + 0.0055070179514586926, + 0.004544613882899284, + 0.009295414201915264, + 0.007577253505587578, + 0.006823855917900801, + 0.008000610396265984, + 0.0077518862672150135, + 0.009521438740193844, + 0.009596416726708412, + 0.010324726812541485, + 0.01606457680463791, + 0.011920565739274025, + 0.0076432297937572, + 0.009015657939016819, + 0.012669061310589314, + 0.014055135659873486, + 0.009265149012207985, + 0.009032176807522774, + 0.009911539033055305, + 0.01453615166246891, + 0.007809692528098822, + 0.014323225244879723, + 0.011026945896446705, + 0.011980894021689892, + 0.009249866008758545, + 0.010408444330096245, + 0.013074892573058605, + 0.013426381163299084, + 0.012187108397483826, + 0.012671018950641155, + 0.011845182627439499, + 0.01248960942029953, + 0.010435933247208595, + 0.01522647961974144, + 0.014336531981825829, + 0.013520496897399426, + 0.011326679959893227, + 0.009315116330981255, + 0.00803192239254713, + 0.007295114919543266, + 0.006437667645514011, + 0.0052396091632544994, + 0.003988116979598999, + 0.005557836499065161, + 0.0037892917171120644, + 0.00827906746417284, + 0.021704696118831635, + 0.02413557469844818, + 0.027665631845593452, + 0.032706521451473236, + 0.01891408860683441, + 0.019884666427969933, + 0.016108250245451927, + 0.020327571779489517, + 0.017897674813866615, + 0.024965770542621613, + 0.019729463383555412, + 0.02217830717563629, + 0.01789071410894394, + 0.01835796795785427, + 0.01757068932056427, + 0.02478436380624771, + 0.0062219505198299885, + 0.008059806190431118, + 0.00885288417339325, + 0.010026971809566021, + 0.007230453658849001, + 0.00868046935647726, + 0.005108596757054329, + 0.008086674846708775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.14284726977348328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549244798719883 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549244798719883, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549244798719883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24369387328624725, + 0.1224309429526329, + 0.04463845118880272, + 0.06417618691921234, + 0.08184517920017242, + 0.09323974698781967, + 0.08976773172616959, + 0.08660856634378433, + 0.10367777198553085, + 0.06992153823375702 + ], + "credit_map": [ + 0.04278117045760155, + 0.03858661279082298, + 0.019736457616090775, + 0.021441344171762466, + 0.020998580381274223, + 0.029295427724719048, + 0.025548920035362244, + 0.04530534893274307, + 0.023453671485185623, + 0.019609728828072548, + 0.0115789994597435, + 0.010227272287011147, + 0.011068162508308887, + 0.013785402290523052, + 0.010612055659294128, + 0.02209564857184887, + 0.008574599400162697, + 0.008015195839107037, + 0.003441387787461281, + 0.0033279452472925186, + 0.004970005713403225, + 0.004218777641654015, + 0.0024764591362327337, + 0.009614079259335995, + 0.010678794234991074, + 0.010754321701824665, + 0.006590195931494236, + 0.006523814518004656, + 0.005456532817333937, + 0.009456542320549488, + 0.006035501137375832, + 0.008680486120283604, + 0.010472430847585201, + 0.012149720452725887, + 0.009352163411676884, + 0.008441697806119919, + 0.011616036295890808, + 0.0130738765001297, + 0.007129731122404337, + 0.009609528817236423, + 0.014274444431066513, + 0.013716486282646656, + 0.013257971964776516, + 0.010509620420634747, + 0.011040976271033287, + 0.012783423066139221, + 0.007040776778012514, + 0.010616039857268333, + 0.011688808910548687, + 0.012700806371867657, + 0.013520770706236362, + 0.015571124851703644, + 0.007569197565317154, + 0.009707778692245483, + 0.008051691576838493, + 0.010957551188766956, + 0.019826075062155724, + 0.012733502313494682, + 0.011598488315939903, + 0.006166606675833464, + 0.0055696554481983185, + 0.0060363695956766605, + 0.006966121960431337, + 0.01771174743771553, + 0.013424627482891083, + 0.012977033853530884, + 0.013262343592941761, + 0.01995265856385231, + 0.009376022033393383, + 0.008860376663506031, + 0.009128149598836899, + 0.016696564853191376, + 0.013992621563374996, + 0.00966105330735445, + 0.009254591539502144, + 0.010411346331238747, + 0.006928650196641684, + 0.005810547154396772, + 0.004793738946318626, + 0.009068993851542473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.012549244798719883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868089407682419 + ], + "opposite_logprob_tokens": [ + -3.0486807823181152 + ], + "logprob_reward": -0.04868089407682419, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868089407682419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08861637115478516, + 0.08614779263734818, + 0.1005500853061676, + 0.12525704503059387, + 0.11640630662441254, + 0.12739905714988708, + 0.059628866612911224, + 0.10171400010585785, + 0.1348685324192047, + 0.05941187962889671 + ], + "credit_map": [ + 0.014063217677175999, + 0.013437081128358841, + 0.006887051276862621, + 0.008634225465357304, + 0.008445600979030132, + 0.01035996712744236, + 0.008079073391854763, + 0.01871015876531601, + 0.012395482510328293, + 0.010886983014643192, + 0.008633865974843502, + 0.007233437616378069, + 0.015053101815283298, + 0.01108834519982338, + 0.009839016944169998, + 0.011017559096217155, + 0.011418001726269722, + 0.01210851315408945, + 0.010033348575234413, + 0.008905354887247086, + 0.019624436274170876, + 0.016357913613319397, + 0.007448327261954546, + 0.014654191210865974, + 0.018442798405885696, + 0.020733235403895378, + 0.013357177376747131, + 0.013190069235861301, + 0.012659608386456966, + 0.017726385965943336, + 0.011553240939974785, + 0.017594529315829277, + 0.011333352886140347, + 0.014971778728067875, + 0.010617169551551342, + 0.011041647754609585, + 0.019863611087203026, + 0.021090753376483917, + 0.01332788448780775, + 0.014160115271806717, + 0.012304924428462982, + 0.01625726744532585, + 0.012591509148478508, + 0.017223050817847252, + 0.021460119634866714, + 0.019416147843003273, + 0.01586768962442875, + 0.012278352864086628, + 0.008780724368989468, + 0.009280615486204624, + 0.006797350477427244, + 0.0065338206477463245, + 0.00528223579749465, + 0.005436568986624479, + 0.005567582324147224, + 0.011949973180890083, + 0.013245154172182083, + 0.013640381395816803, + 0.013279395177960396, + 0.015390107408165932, + 0.011412683874368668, + 0.013147708028554916, + 0.008867633529007435, + 0.012730933725833893, + 0.013579354621469975, + 0.018740037456154823, + 0.01917838677763939, + 0.02070995233952999, + 0.013662244193255901, + 0.018461648374795914, + 0.013787684962153435, + 0.016749218106269836, + 0.008060780353844166, + 0.009312652051448822, + 0.007992236874997616, + 0.0077025569044053555, + 0.006720192730426788, + 0.005818320903927088, + 0.004351494368165731, + 0.009453648701310158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.04868089407682419 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close-up, static shot focuses with razor-sharp clarity on the angled tip of a sterile, stainless-steel hypodermic needle. The polished metal gleams in bright, direct light, revealing microscopic striations from its manufacturing process. Suddenly, a clear, viscous liquid powerfully erupts from the narrow gap in a swift and decisive eruption. The syrupy liquid catches the light, shimmering and forming a perfect arc against the cool, softly blurred, sterile blue-white background. The initial, cohesive spray of liquid lingers in the air for a moment before gracefully breaking into shimmering, individual droplets that follow the same trajectory. The entire scene is illuminated by cool, high-key lighting, creating sharp, mirror-like highlights on the needle and making each droplet shimmer, creating an atmosphere of intense focus, clinical precision, and controlled force.", + "seed": 228, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_10/index_5617_seed_228_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.31328070536255836, + "vlm_logprob": -0.31328070536255836, + "vlm_yes_no_margin": 1.425, + "vlm_token_credit": -0.31328070536255836, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16332733631134033 + ], + "opposite_logprob_tokens": [ + -1.9133273363113403 + ], + "logprob_reward": -0.16332733631134033, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16332733631134033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2941979765892029, + 0.2991243600845337, + 0.0576925054192543, + 0.055825214833021164, + 0.04519518092274666, + 0.13479334115982056, + 0.04051850736141205, + 0.011167526245117188, + 0.03529985249042511, + 0.02618543989956379 + ], + "credit_map": [ + 0.019588837400078773, + 0.08706504851579666, + 0.10423041880130768, + 0.03896502032876015, + 0.0076720030046999454, + 0.018578868359327316, + 0.007852206006646156, + 0.01024560071527958, + 0.01827976666390896, + 0.09273984283208847, + 0.08032546192407608, + 0.05230672284960747, + 0.018895050510764122, + 0.016758810728788376, + 0.009156085550785065, + 0.010662647895514965, + 0.0071325600147247314, + 0.007418075110763311, + 0.012816010043025017, + 0.01071246713399887, + 0.004277192987501621, + 0.005347926635295153, + 0.003401055233553052, + 0.006587222684174776, + 0.00488095311447978, + 0.008052995428442955, + 0.015134179964661598, + 0.010169354267418385, + 0.0041882796213030815, + 0.004371989518404007, + 0.003526661777868867, + 0.005500804632902145, + 0.005048577673733234, + 0.005486776120960712, + 0.009935928508639336, + 0.008610167540609837, + 0.0035873905289918184, + 0.003412594785913825, + 0.0031077489256858826, + 0.006005997769534588, + 0.0189485102891922, + 0.015343772247433662, + 0.02070615626871586, + 0.02272426337003708, + 0.013794335536658764, + 0.013424722477793694, + 0.009365583769977093, + 0.0204860121011734, + 0.005654962733387947, + 0.0055295671336352825, + 0.004863721318542957, + 0.008737745694816113, + 0.004600347951054573, + 0.0033167842775583267, + 0.0026004151441156864, + 0.0052149672992527485, + 0.001848104759119451, + 0.001456038560718298, + 0.0015949398512020707, + 0.0018815792864188552, + 0.0015582734486088157, + 0.0009755999781191349, + 0.0005967826000414789, + 0.0012562088668346405, + 0.003970918711274862, + 0.004288623575121164, + 0.005010278895497322, + 0.006664501968771219, + 0.004861067049205303, + 0.003293178277090192, + 0.0024598364252597094, + 0.004751450847834349, + 0.002413342008367181, + 0.005109220743179321, + 0.0045743174850940704, + 0.004195875953882933, + 0.0027313437312841415, + 0.0019487394019961357, + 0.0015559974126517773, + 0.0036566080525517464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.16332733631134033 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934148669242859 + ], + "opposite_logprob_tokens": [ + -0.6934148669242859 + ], + "logprob_reward": -0.6934148669242859, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934148669242859, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3853878676891327, + 0.038851577788591385, + 0.03625575453042984, + 0.17208373546600342, + 0.03469986096024513, + 0.0924789309501648, + 0.042020533233881, + 0.08721744269132614, + 0.09352120012044907, + 0.01748310588300228 + ], + "credit_map": [ + 0.02692122757434845, + 0.11879697442054749, + 0.1514829695224762, + 0.03314589336514473, + 0.010737044736742973, + 0.024109655991196632, + 0.008879791013896465, + 0.011314316652715206, + 0.0038042382802814245, + 0.006658692844212055, + 0.0063898297958076, + 0.006892816629260778, + 0.005735489539802074, + 0.0038280407898128033, + 0.0021861987188458443, + 0.003356274915859103, + 0.0029156594537198544, + 0.0075386036187410355, + 0.006912305951118469, + 0.006834839936345816, + 0.0032518557272851467, + 0.0034737358801066875, + 0.0016449547838419676, + 0.003683799412101507, + 0.01168925128877163, + 0.036451734602451324, + 0.049265969544649124, + 0.021274227648973465, + 0.012801429256796837, + 0.015545187518000603, + 0.010092681273818016, + 0.014963249675929546, + 0.004130121320486069, + 0.004275025799870491, + 0.003628429491072893, + 0.005080392584204674, + 0.004611830692738295, + 0.004501827526837587, + 0.00335435732267797, + 0.005117874126881361, + 0.01165295485407114, + 0.01230746228247881, + 0.02572922594845295, + 0.013076234608888626, + 0.006848524324595928, + 0.005948737263679504, + 0.006101184990257025, + 0.010814608074724674, + 0.005926641635596752, + 0.005676696076989174, + 0.00716322660446167, + 0.008097640238702297, + 0.004777323920279741, + 0.0033712477888911963, + 0.002688400447368622, + 0.004319357220083475, + 0.010515657253563404, + 0.015728944912552834, + 0.010066046379506588, + 0.020703325048089027, + 0.0074923280626535416, + 0.006687396205961704, + 0.005704077426344156, + 0.010319667868316174, + 0.011344912461936474, + 0.012798716314136982, + 0.011636362411081791, + 0.0188018549233675, + 0.012950392439961433, + 0.008240682072937489, + 0.0058439685963094234, + 0.011904311366379261, + 0.0022037348244339228, + 0.003152953926473856, + 0.0022147460840642452, + 0.0036196094006299973, + 0.001551759778521955, + 0.0013871751725673676, + 0.001199298189021647, + 0.0021538282744586468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.6934148669242859 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743145704269409 + ], + "opposite_logprob_tokens": [ + -0.9743145704269409 + ], + "logprob_reward": -0.4743145704269409, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743145704269409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4347984790802002, + 0.0741945430636406, + 0.03360050544142723, + 0.1480254977941513, + 0.05003613978624344, + 0.16269630193710327, + 0.02437625825405121, + 0.027233073487877846, + 0.013789469376206398, + 0.03124965727329254 + ], + "credit_map": [ + 0.02350127510726452, + 0.1530117243528366, + 0.17563840746879578, + 0.03285203129053116, + 0.012956993654370308, + 0.020133471116423607, + 0.008670500479638577, + 0.008034099824726582, + 0.006318189203739166, + 0.014866409823298454, + 0.018772223964333534, + 0.009245832450687885, + 0.008255846798419952, + 0.007323349826037884, + 0.0028711457271128893, + 0.006541544105857611, + 0.005396686960011721, + 0.0042546591721475124, + 0.0050514754839241505, + 0.008195317350327969, + 0.003070063656195998, + 0.0032775590661913157, + 0.0014561605639755726, + 0.002898585982620716, + 0.011242861859500408, + 0.018228506669402122, + 0.020005803555250168, + 0.02763596549630165, + 0.022590359672904015, + 0.020772311836481094, + 0.011686737649142742, + 0.015862951055169106, + 0.004190488252788782, + 0.006995139643549919, + 0.005772987846285105, + 0.006001736503094435, + 0.00661276001483202, + 0.007742916699498892, + 0.004891600459814072, + 0.00782850757241249, + 0.020141668617725372, + 0.02424084208905697, + 0.05076102539896965, + 0.02754722163081169, + 0.010156198404729366, + 0.007966539822518826, + 0.008508406579494476, + 0.013374394737184048, + 0.003201994812116027, + 0.0028590220026671886, + 0.003242060076445341, + 0.004716072231531143, + 0.002997092204168439, + 0.001942819682881236, + 0.001981810899451375, + 0.0034353865776211023, + 0.002916855737566948, + 0.004286372102797031, + 0.003106100484728813, + 0.005556206218898296, + 0.0028869907837361097, + 0.0024328071158379316, + 0.0021336760837584734, + 0.003914067521691322, + 0.0015963922487571836, + 0.0016434428980574012, + 0.0018123335903510451, + 0.002540058922022581, + 0.0020729193929582834, + 0.0015173701103776693, + 0.0010002957424148917, + 0.0016066569369286299, + 0.002985527040436864, + 0.004633964039385319, + 0.004949514754116535, + 0.004705418366938829, + 0.003869953565299511, + 0.0029279757291078568, + 0.0025872045662254095, + 0.004590097814798355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.4743145704269409 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.201620414853096 + ], + "opposite_logprob_tokens": [ + -1.7016204595565796 + ], + "logprob_reward": -0.201620414853096, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.201620414853096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34438440203666687, + 0.22177985310554504, + 0.05456685647368431, + 0.039479613304138184, + 0.07019224762916565, + 0.09702043235301971, + 0.04552256315946579, + 0.021184341982007027, + 0.040734488517045975, + 0.06513521075248718 + ], + "credit_map": [ + 0.02759295143187046, + 0.09185463935136795, + 0.12943531572818756, + 0.038832686841487885, + 0.009717052802443504, + 0.024473346769809723, + 0.009730595164000988, + 0.01274779625236988, + 0.013440622016787529, + 0.06867964565753937, + 0.06495250761508942, + 0.03543134778738022, + 0.012012427672743797, + 0.012429950758814812, + 0.0063936407677829266, + 0.008439714089035988, + 0.006479286123067141, + 0.00827140174806118, + 0.014009158127009869, + 0.00979380588978529, + 0.0033510811626911163, + 0.004649252165108919, + 0.0028071454726159573, + 0.005205726251006126, + 0.004125455394387245, + 0.0061065079644322395, + 0.007966632954776287, + 0.00688102375715971, + 0.0026208737399429083, + 0.004606958013027906, + 0.002873738994821906, + 0.004298422019928694, + 0.008423625491559505, + 0.008473088033497334, + 0.015949653461575508, + 0.01591852866113186, + 0.0049270144663751125, + 0.0046265460550785065, + 0.004052777774631977, + 0.007821017876267433, + 0.011924323625862598, + 0.01258396077901125, + 0.013662105426192284, + 0.0166912991553545, + 0.010699211619794369, + 0.009454156272113323, + 0.007083773147314787, + 0.01492160465568304, + 0.004718816373497248, + 0.008185619488358498, + 0.007373119238764048, + 0.008754643611609936, + 0.004825723357498646, + 0.0032433413434773684, + 0.0025958220940083265, + 0.005825480446219444, + 0.0029733851552009583, + 0.0022162541281431913, + 0.004465954843908548, + 0.004345628432929516, + 0.0020692560356110334, + 0.0014727767556905746, + 0.0011084689758718014, + 0.002532618585973978, + 0.004952531773597002, + 0.004830901511013508, + 0.007288562599569559, + 0.007649850565940142, + 0.004823163151741028, + 0.0036200687754899263, + 0.002690752735361457, + 0.004878653679043055, + 0.0061994860880076885, + 0.014742731116712093, + 0.012285999022424221, + 0.00984874926507473, + 0.006449128035455942, + 0.004879514686763287, + 0.0033431898336857557, + 0.007386414799839258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.201620414853096 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of a sterile, stainless-steel hypodermic needle tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a clear, viscous liquid erupt powerfully from the needle gap in a swift motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cool, softly blurred, sterile blue-white environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the liquid form a perfect arc before breaking into individual shimmering droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372633829712868 + ], + "opposite_logprob_tokens": [ + -3.408726453781128 + ], + "logprob_reward": -0.03372633829712868, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372633829712868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31338948011398315, + 0.2692507803440094, + 0.0437040738761425, + 0.08304285258054733, + 0.04256155341863632, + 0.1673562377691269, + 0.02192358672618866, + 0.017864516004920006, + 0.022724730893969536, + 0.018182141706347466 + ], + "credit_map": [ + 0.017228063195943832, + 0.10823649913072586, + 0.1269531100988388, + 0.017633838579058647, + 0.010860743001103401, + 0.01539513748139143, + 0.008437317796051502, + 0.00864476803690195, + 0.01495213434100151, + 0.08230479806661606, + 0.062309086322784424, + 0.04736107587814331, + 0.025192106142640114, + 0.01875036023557186, + 0.009299596771597862, + 0.009081660769879818, + 0.004987826570868492, + 0.0052447011694312096, + 0.009549767710268497, + 0.00792399700731039, + 0.004124074708670378, + 0.00466213608160615, + 0.0026182879228144884, + 0.004593287128955126, + 0.006240814458578825, + 0.00996838416904211, + 0.010589642450213432, + 0.014425630681216717, + 0.014455619268119335, + 0.010833200067281723, + 0.007693144027143717, + 0.008836425840854645, + 0.004459715448319912, + 0.005368341691792011, + 0.005557809490710497, + 0.006323206704109907, + 0.004924002103507519, + 0.005654243752360344, + 0.004164867103099823, + 0.0061093708500266075, + 0.01824413798749447, + 0.02117285132408142, + 0.037591733038425446, + 0.027465976774692535, + 0.013392258435487747, + 0.016970399767160416, + 0.012216350063681602, + 0.020302532240748405, + 0.0024803022388368845, + 0.002747057005763054, + 0.0047482275404036045, + 0.003139146137982607, + 0.002403455786406994, + 0.0020347756799310446, + 0.0017957995878532529, + 0.0025748249609023333, + 0.0016487636603415012, + 0.0030421859119087458, + 0.0024664781522005796, + 0.0031244568526744843, + 0.001934910542331636, + 0.0015537458239123225, + 0.0016980188665911555, + 0.00239595677703619, + 0.0023818572517484426, + 0.004103211686015129, + 0.0034215557388961315, + 0.004796496592462063, + 0.0024455441161990166, + 0.0016994349425658584, + 0.00105537089984864, + 0.0028212617617100477, + 0.0014002014650031924, + 0.0032950465101748705, + 0.0033802520483732224, + 0.002475869609043002, + 0.0018880943534895778, + 0.00201203441247344, + 0.0018219721969217062, + 0.0019086739048361778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03372633829712868 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_11/index_tensor([2992, 2992])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_11/index_tensor([2992, 2992])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bc31db4846bbcfd625097189e958dfb81664a958 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_11/index_tensor([2992, 2992])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, low-angle, static wide shot captures the towering silhouettes of a modern city. These sharp, geometric skyscrapers stand silently in pure black in the lower half of the frame. A strong desaturated color grading is employed throughout, creating a calm yet striking black-and-white quality. The sky is a deep, almost black canvas. Amidst this silent grandeur, a total lunar eclipse unfolds: Earth's umbra, like an invisible curtain, steadily and relentlessly engulfs the moon's luminous silver disk. As the last shimmer of silver light is swallowed by darkness, the moon, far from disappearing, glows a deep, saturated, blood-red. In this almost entirely monochrome world, the crimson moon provides the sole, breathtaking source of color, its ominous and majestic halo standing out against the stark black-and-white contrast. The entire dynamic unfolds silently, creating an apocalyptic, awe-inspiring stillness and visual tension.", + "seed": 7555, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_11/index_2992_seed_7555_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18852101238444446, + "vlm_logprob": -0.18852101238444446, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.18852101238444446, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024877607822418213 + ], + "opposite_logprob_tokens": [ + -5.024877548217773 + ], + "logprob_reward": -0.024877607822418213, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.024877607822418213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07033222168684006, + 0.08154741674661636, + 0.02078361250460148, + 0.029837848618626595, + 0.02079249545931816, + 0.25543031096458435, + 0.02250347100198269, + 0.03090672194957733, + 0.11646371334791183, + 0.3514021635055542 + ], + "credit_map": [ + 0.01265514176338911, + 0.004842787049710751, + 0.011202652007341385, + 0.0037071898113936186, + 0.002382103353738785, + 0.010202111676335335, + 0.01102369837462902, + 0.014316538348793983, + 0.01200578548014164, + 0.005996231455355883, + 0.01543075405061245, + 0.0051984237506985664, + 0.003754486795514822, + 0.011818661354482174, + 0.012368866242468357, + 0.014974205754697323, + 0.003042694879695773, + 0.0011511793127283454, + 0.0047579980455338955, + 0.0019925532396882772, + 0.001089594210498035, + 0.003294742899015546, + 0.001725411624647677, + 0.0037294391077011824, + 0.004047272726893425, + 0.001295587164349854, + 0.005872524343430996, + 0.0019426168873906136, + 0.0014745773514732718, + 0.0047432719729840755, + 0.0017595886019989848, + 0.008702408522367477, + 0.002428436651825905, + 0.0022725972812622786, + 0.0034122501965612173, + 0.002550586825236678, + 0.0015568361850455403, + 0.0027442490682005882, + 0.0014826311962679029, + 0.004344908054918051, + 0.027830732986330986, + 0.00926959328353405, + 0.04911746829748154, + 0.0122566232457757, + 0.01317807286977768, + 0.04017072916030884, + 0.015758490189909935, + 0.08784861117601395, + 0.0023276936262845993, + 0.0016069184057414532, + 0.003802191698923707, + 0.002125506056472659, + 0.0012024882016703486, + 0.0033731982111930847, + 0.0017862641252577305, + 0.006279211025685072, + 0.0031465464271605015, + 0.0023327094968408346, + 0.0057593705132603645, + 0.0029539354145526886, + 0.0019975092727690935, + 0.004177724476903677, + 0.0017657401040196419, + 0.008773187175393105, + 0.013023488223552704, + 0.008019167929887772, + 0.022911496460437775, + 0.012238478288054466, + 0.006164657883346081, + 0.016146289184689522, + 0.006278056651353836, + 0.03168207406997681, + 0.04105053097009659, + 0.026665614917874336, + 0.07741829007863998, + 0.046687494963407516, + 0.020093630999326706, + 0.049531590193510056, + 0.0176213588565588, + 0.0723336711525917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.024877607822418213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003629883285611868 + ], + "opposite_logprob_tokens": [ + -5.628629684448242 + ], + "logprob_reward": -0.003629883285611868, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003629883285611868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10248908400535583, + 0.17592507600784302, + 0.023630529642105103, + 0.06709399074316025, + 0.028474383056163788, + 0.159043550491333, + 0.06438560783863068, + 0.04886623099446297, + 0.07789132744073868, + 0.2522001564502716 + ], + "credit_map": [ + 0.015331298112869263, + 0.006665988825261593, + 0.01669580489397049, + 0.006754744332283735, + 0.004291655495762825, + 0.01446316484361887, + 0.016840321943163872, + 0.021446099504828453, + 0.0274577084928751, + 0.010592482052743435, + 0.032869357615709305, + 0.010940258391201496, + 0.007529321126639843, + 0.02489095740020275, + 0.02890360727906227, + 0.032741378992795944, + 0.0036530515644699335, + 0.0032356074079871178, + 0.0036885153967887163, + 0.0022581429220736027, + 0.002524995943531394, + 0.0027443119324743748, + 0.0022573780734091997, + 0.0032685273326933384, + 0.008589507080614567, + 0.0030475242529064417, + 0.01442047767341137, + 0.0037785659078508615, + 0.003473467892035842, + 0.010819830000400543, + 0.00442863255739212, + 0.018535982817411423, + 0.0038391160778701305, + 0.002297820756211877, + 0.005612079054117203, + 0.002371871843934059, + 0.002029748633503914, + 0.00414220429956913, + 0.0020488991867750883, + 0.006132643204182386, + 0.016426460817456245, + 0.006797930225729942, + 0.029680222272872925, + 0.010105068795382977, + 0.006964007392525673, + 0.02470746450126171, + 0.010331555269658566, + 0.054030824452638626, + 0.006674438714981079, + 0.004301128443330526, + 0.012001518160104752, + 0.005334365181624889, + 0.002981865080073476, + 0.009493664838373661, + 0.0033988982904702425, + 0.0201997272670269, + 0.0060840779915452, + 0.002323165303096175, + 0.009711159393191338, + 0.003676755353808403, + 0.002000278327614069, + 0.007534724194556475, + 0.003251468064263463, + 0.01428460143506527, + 0.009522957727313042, + 0.004938557278364897, + 0.015479227527976036, + 0.0065083494409918785, + 0.004634978249669075, + 0.010997666977345943, + 0.005384835880249739, + 0.02042475901544094, + 0.030545782297849655, + 0.01822383515536785, + 0.05766094848513603, + 0.033255524933338165, + 0.014955046586692333, + 0.035498470067977905, + 0.013126819394528866, + 0.048933736979961395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.003629883285611868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983577489852905 + ], + "opposite_logprob_tokens": [ + -0.5233577489852905 + ], + "logprob_reward": -0.8983577489852905, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983577489852905, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04820806905627251, + 0.18640165030956268, + 0.05171816796064377, + 0.04678342118859291, + 0.058851879090070724, + 0.13896188139915466, + 0.0928068533539772, + 0.026716765016317368, + 0.2186846286058426, + 0.1308666616678238 + ], + "credit_map": [ + 0.007917918264865875, + 0.003506647190079093, + 0.010262427851557732, + 0.0042435224168002605, + 0.00263371504843235, + 0.007586389780044556, + 0.003364801639690995, + 0.008692644536495209, + 0.025507699698209763, + 0.01724294200539589, + 0.035320624709129333, + 0.012262212112545967, + 0.009747767820954323, + 0.02803920954465866, + 0.025429733097553253, + 0.03285146504640579, + 0.007135441061109304, + 0.006611356046050787, + 0.008257714100182056, + 0.007014167495071888, + 0.004866424482315779, + 0.006669510155916214, + 0.003870778251439333, + 0.007292779162526131, + 0.006029499229043722, + 0.0021639536134898663, + 0.009068785235285759, + 0.0027328303549438715, + 0.0023285814095288515, + 0.0075718373991549015, + 0.003535164287313819, + 0.013352765701711178, + 0.007573229260742664, + 0.0025621489621698856, + 0.012084084562957287, + 0.00444161519408226, + 0.003081584582105279, + 0.009033111855387688, + 0.004574693739414215, + 0.015501407906413078, + 0.014880108647048473, + 0.005590246059000492, + 0.025970663875341415, + 0.00855310633778572, + 0.005591440014541149, + 0.022049276158213615, + 0.009027974680066109, + 0.047299064695835114, + 0.010785051621496677, + 0.00655854819342494, + 0.01880333386361599, + 0.008080612868070602, + 0.005519012920558453, + 0.014151734299957752, + 0.005831350106745958, + 0.023077208548784256, + 0.003689634846523404, + 0.002823185408487916, + 0.004613895900547504, + 0.003531794063746929, + 0.002422393299639225, + 0.0036630777176469564, + 0.0018891635118052363, + 0.004083620384335518, + 0.02484237775206566, + 0.013140477240085602, + 0.0425606295466423, + 0.020297933369874954, + 0.013878805562853813, + 0.03051760606467724, + 0.012422087602317333, + 0.061024703085422516, + 0.014723136089742184, + 0.01097728218883276, + 0.029584746807813644, + 0.018491853028535843, + 0.008044120855629444, + 0.018107356503605843, + 0.007169578690081835, + 0.02376859448850155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.8983577489852905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958531983196735 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.005958531983196735, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958531983196735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08004865050315857, + 0.056656695902347565, + 0.033132120966911316, + 0.020712563768029213, + 0.037723660469055176, + 0.10155521333217621, + 0.09366178512573242, + 0.04459896683692932, + 0.35262414813041687, + 0.17928625643253326 + ], + "credit_map": [ + 0.01184763852506876, + 0.007744447328150272, + 0.015227477997541428, + 0.006794900167733431, + 0.0051055848598480225, + 0.012676496058702469, + 0.0076385075226426125, + 0.013013592921197414, + 0.007070578634738922, + 0.0090370774269104, + 0.00970512442290783, + 0.0046925013884902, + 0.004095512442290783, + 0.008964632637798786, + 0.0034310552291572094, + 0.009660214185714722, + 0.00502348318696022, + 0.0035136956721544266, + 0.005061722360551357, + 0.004167987499386072, + 0.002910178853198886, + 0.00452284375205636, + 0.0027325169648975134, + 0.005199691280722618, + 0.0027471331413835287, + 0.0011915587820112705, + 0.003756451653316617, + 0.0013072788715362549, + 0.0012705557746812701, + 0.003175610676407814, + 0.002132567809894681, + 0.005131406709551811, + 0.004491851665079594, + 0.002483807737007737, + 0.006874802988022566, + 0.0032685415353626013, + 0.0022063390351831913, + 0.0054635219275951385, + 0.0025811842642724514, + 0.01035361085087061, + 0.012821614742279053, + 0.005033394321799278, + 0.01855374500155449, + 0.005999723449349403, + 0.004910034593194723, + 0.015609090216457844, + 0.00694951182231307, + 0.03167809545993805, + 0.010763434693217278, + 0.005829075817018747, + 0.017894534394145012, + 0.008545510470867157, + 0.005170500837266445, + 0.013728095218539238, + 0.005742302630096674, + 0.02598833106458187, + 0.0043789600022137165, + 0.0048395246267318726, + 0.007410532329231501, + 0.0038443300873041153, + 0.003946173470467329, + 0.005979766137897968, + 0.0022966095712035894, + 0.011903068982064724, + 0.039772529155015945, + 0.02201482281088829, + 0.06942536681890488, + 0.03512196987867355, + 0.02024327777326107, + 0.0492032989859581, + 0.019549161195755005, + 0.09729370474815369, + 0.019974658265709877, + 0.01474155392497778, + 0.03902185335755348, + 0.02652784436941147, + 0.010420060716569424, + 0.02490963600575924, + 0.009013124741613865, + 0.03467751666903496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.005958531983196735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781289845705032 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781289845705032, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781289845705032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07590220123529434, + 0.4643903076648712, + 0.035788822919130325, + 0.02104640007019043, + 0.03681951388716698, + 0.03640135005116463, + 0.09106069058179855, + 0.0333511084318161, + 0.15449301898479462, + 0.05074654147028923 + ], + "credit_map": [ + 0.010677906684577465, + 0.004389974754303694, + 0.01472797617316246, + 0.005327291786670685, + 0.003194105112925172, + 0.011428090743720531, + 0.009199565276503563, + 0.01695728674530983, + 0.06435824930667877, + 0.04119756072759628, + 0.0894131138920784, + 0.0323416031897068, + 0.023881100118160248, + 0.06853950768709183, + 0.06756703555583954, + 0.07709214091300964, + 0.0049337707459926605, + 0.004873049911111593, + 0.004922305699437857, + 0.004655783995985985, + 0.00341721810400486, + 0.004689556080847979, + 0.0029539994429796934, + 0.0053431387059390545, + 0.0027966334018856287, + 0.00129945098888129, + 0.004512283485382795, + 0.0013899770565330982, + 0.0013009781250730157, + 0.00320409401319921, + 0.002385071013122797, + 0.004157912451773882, + 0.005109338089823723, + 0.0017728840466588736, + 0.0077597578056156635, + 0.0037064135540276766, + 0.002077877288684249, + 0.005610206164419651, + 0.003818943165242672, + 0.006964090745896101, + 0.004499274305999279, + 0.004343125503510237, + 0.005761634558439255, + 0.00591724319383502, + 0.002874387428164482, + 0.005295704118907452, + 0.002951570088043809, + 0.004758412018418312, + 0.010247832164168358, + 0.0057990592904388905, + 0.018246568739414215, + 0.007993072271347046, + 0.004640586674213409, + 0.013635058887302876, + 0.005700519774109125, + 0.024797987192869186, + 0.0041628130711615086, + 0.002984782448038459, + 0.0067338114604353905, + 0.0021554373670369387, + 0.0025623529218137264, + 0.004742445424199104, + 0.0021298418287187815, + 0.007879623211920261, + 0.01737821288406849, + 0.00947206187993288, + 0.0301548782736063, + 0.014506548643112183, + 0.009197104722261429, + 0.02164064534008503, + 0.008908555842936039, + 0.04323502257466316, + 0.00521291047334671, + 0.005009887740015984, + 0.011651345528662205, + 0.006552491337060928, + 0.003472513984888792, + 0.006995042786002159, + 0.0037384217139333487, + 0.008113926276564598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.009781289845705032 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle, static wide shot captures the towering silhouettes of a modern city. These sharp, geometric skyscrapers stand silently in pure black in the lower half of the frame. A strong desaturated color grading is employed throughout, creating a calm yet striking black-and-white quality. The sky is a deep, almost black canvas. Amidst this silent grandeur, a total lunar eclipse unfolds: Earth's umbra, like an invisible curtain, steadily and relentlessly engulfs the moon's luminous silver disk. As the last shimmer of silver light is swallowed by darkness, the moon, far from disappearing, glows a deep, saturated, blood-red. In this almost entirely monochrome world, the crimson moon provides the sole, breathtaking source of color, its ominous and majestic halo standing out against the stark black-and-white contrast. The entire dynamic unfolds silently, creating an apocalyptic, awe-inspiring stillness and visual tension.", + "seed": 7556, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_11/index_2992_seed_7556_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02854962518904358, + "vlm_logprob": -0.02854962518904358, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.02854962518904358, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018470199778676033 + ], + "opposite_logprob_tokens": [ + -4.393470287322998 + ], + "logprob_reward": -0.018470199778676033, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.018470199778676033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11178778111934662, + 0.08050862699747086, + 0.03755816072225571, + 0.0687163695693016, + 0.07192220538854599, + 0.031761303544044495, + 0.15234163403511047, + 0.12884697318077087, + 0.2930147647857666, + 0.02354210801422596 + ], + "credit_map": [ + 0.016616009175777435, + 0.016898389905691147, + 0.010431571863591671, + 0.01313671376556158, + 0.012873684987425804, + 0.01588711328804493, + 0.015477916225790977, + 0.010466397739946842, + 0.008593440987169743, + 0.01520268153399229, + 0.007129746489226818, + 0.00826916005462408, + 0.010750134475529194, + 0.014507106505334377, + 0.007089105434715748, + 0.008967263624072075, + 0.004419276956468821, + 0.0033992740791291, + 0.008272608742117882, + 0.006281109526753426, + 0.004236468579620123, + 0.0048272148706018925, + 0.002663609804585576, + 0.0034586042165756226, + 0.009984606876969337, + 0.008512482978403568, + 0.009450484998524189, + 0.006630655378103256, + 0.005750559736043215, + 0.010637449100613594, + 0.008089464157819748, + 0.009660678915679455, + 0.008214428089559078, + 0.008431900292634964, + 0.010234879329800606, + 0.005513523239642382, + 0.005248658824712038, + 0.010466979816555977, + 0.01153504941612482, + 0.012276793830096722, + 0.004662323743104935, + 0.0029839822091162205, + 0.006170876324176788, + 0.0022600956726819277, + 0.001855484675616026, + 0.004354838747531176, + 0.0030524597968906164, + 0.006421247031539679, + 0.00866114441305399, + 0.003017376409843564, + 0.06604141741991043, + 0.00406275037676096, + 0.004505260847508907, + 0.05641612410545349, + 0.0021643624641001225, + 0.007473217789083719, + 0.006599383894354105, + 0.0023545536678284407, + 0.053154028952121735, + 0.003217503661289811, + 0.0040975091978907585, + 0.050644904375076294, + 0.001777975237928331, + 0.007001135498285294, + 0.046283748000860214, + 0.012114313431084156, + 0.045406658202409744, + 0.01778656244277954, + 0.010535922832787037, + 0.061429910361766815, + 0.017361672595143318, + 0.08209600299596786, + 0.002842654474079609, + 0.002750843996182084, + 0.004511023405939341, + 0.0023164518643170595, + 0.0030138124711811543, + 0.002657316392287612, + 0.002717240247875452, + 0.0027327686548233032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.018470199778676033 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013465156080201268 + ], + "opposite_logprob_tokens": [ + -6.626346588134766 + ], + "logprob_reward": -0.0013465156080201268, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013465156080201268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07873885333538055, + 0.05616278573870659, + 0.030043097212910652, + 0.07197888195514679, + 0.17691873013973236, + 0.010966100730001926, + 0.31091806292533875, + 0.08246468752622604, + 0.16166572272777557, + 0.02014302834868431 + ], + "credit_map": [ + 0.008673373609781265, + 0.01056024432182312, + 0.008472273126244545, + 0.009290758520364761, + 0.008042504079639912, + 0.008164794184267521, + 0.016569243744015694, + 0.00896566454321146, + 0.0065856208093464375, + 0.007730840239673853, + 0.006697827018797398, + 0.005595455411821604, + 0.007174091879278421, + 0.009827319532632828, + 0.005377449560910463, + 0.0071741859428584576, + 0.0038379542529582977, + 0.005319061689078808, + 0.0043514021672308445, + 0.003468960290774703, + 0.0045732432045042515, + 0.004004086833447218, + 0.0018812406342476606, + 0.0026071479078382254, + 0.007874418050050735, + 0.00770787475630641, + 0.01082505565136671, + 0.006790479179471731, + 0.0070036728866398335, + 0.009315947070717812, + 0.00969174224883318, + 0.01276969350874424, + 0.023579062893986702, + 0.010742975398898125, + 0.0232892744243145, + 0.012704165652394295, + 0.00853808131068945, + 0.03332102671265602, + 0.03237750381231308, + 0.032366637140512466, + 0.001556857954710722, + 0.0016625181306153536, + 0.0014481915859505534, + 0.0009643880766816437, + 0.000715048867277801, + 0.001692345947958529, + 0.0009803564753383398, + 0.0019463934004306793, + 0.01720525696873665, + 0.0038153573404997587, + 0.1376420259475708, + 0.007826363667845726, + 0.00890263170003891, + 0.12402738630771637, + 0.0027114260010421276, + 0.008787629194557667, + 0.00424657529219985, + 0.0021128556691110134, + 0.0313306488096714, + 0.0023176008835434914, + 0.0030178178567439318, + 0.03295810893177986, + 0.0021392228081822395, + 0.004341853782534599, + 0.025740278884768486, + 0.006353145930916071, + 0.02449849061667919, + 0.009860984981060028, + 0.005267968866974115, + 0.033947065472602844, + 0.00920156016945839, + 0.04679623618721962, + 0.002190571976825595, + 0.002705217571929097, + 0.0035760439932346344, + 0.001560092088766396, + 0.002052250551059842, + 0.0032256238628178835, + 0.001860627788119018, + 0.0029726005159318447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.0013465156080201268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129983589053154 + ], + "opposite_logprob_tokens": [ + -2.2379982471466064 + ], + "logprob_reward": -0.1129983589053154, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129983589053154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12520477175712585, + 0.01614401862025261, + 0.02655772864818573, + 0.02903902344405651, + 0.18940812349319458, + 0.03830358386039734, + 0.3422425091266632, + 0.0274922177195549, + 0.10709664970636368, + 0.09851133078336716 + ], + "credit_map": [ + 0.029424207285046577, + 0.014026343822479248, + 0.012876036576926708, + 0.015274868346750736, + 0.01397988386452198, + 0.014080645516514778, + 0.008860002271831036, + 0.016682783141732216, + 0.0019453655695542693, + 0.003181833541020751, + 0.0013592143077403307, + 0.001598637318238616, + 0.003296083537861705, + 0.0020942208357155323, + 0.0010918095940724015, + 0.0015768537996336818, + 0.003428930416703224, + 0.004412779584527016, + 0.0020427422132343054, + 0.003332423744723201, + 0.005217732395976782, + 0.0034957940224558115, + 0.00199465942569077, + 0.002632667776197195, + 0.004278211854398251, + 0.003294059308245778, + 0.0033921387512236834, + 0.0029604011215269566, + 0.0029523943085223436, + 0.0045552863739430904, + 0.0032245819456875324, + 0.004381948616355658, + 0.02457529492676258, + 0.011610040441155434, + 0.025301922112703323, + 0.013912850990891457, + 0.009492401964962482, + 0.03508185222744942, + 0.03389311209321022, + 0.03554064780473709, + 0.00509853707626462, + 0.005157154519110918, + 0.006645394489169121, + 0.0028121538925915956, + 0.0024533115793019533, + 0.005494385026395321, + 0.003567431354895234, + 0.007075218018144369, + 0.018856657668948174, + 0.00398265989497304, + 0.15202659368515015, + 0.00851986650377512, + 0.009804732166230679, + 0.13610412180423737, + 0.00285246386192739, + 0.010095430538058281, + 0.002178189577534795, + 0.0019360932055860758, + 0.006382873747497797, + 0.0025837586726993322, + 0.001704024849459529, + 0.006772404070943594, + 0.0018387692980468273, + 0.004096103832125664, + 0.015600821934640408, + 0.0058083017356693745, + 0.019028840586543083, + 0.007511508651077747, + 0.005566463805735111, + 0.022320840507745743, + 0.006709694396704435, + 0.0245501808822155, + 0.01609017513692379, + 0.01005492452532053, + 0.01632623001933098, + 0.008988048881292343, + 0.010938617400825024, + 0.012956302613019943, + 0.007704729679971933, + 0.015452311374247074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.1129983589053154 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032014804892241955 + ], + "opposite_logprob_tokens": [ + -5.753201484680176 + ], + "logprob_reward": -0.0032014804892241955, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032014804892241955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10076544433832169, + 0.03203359618782997, + 0.030178550630807877, + 0.0365106537938118, + 0.522396981716156, + 0.05380416661500931, + 0.056867606937885284, + 0.03673505410552025, + 0.06172547861933708, + 0.06898246705532074 + ], + "credit_map": [ + 0.011096839793026447, + 0.013160348869860172, + 0.008578619919717312, + 0.017181305214762688, + 0.01588209718465805, + 0.011392325162887573, + 0.012742209248244762, + 0.010731703601777554, + 0.0038223653100430965, + 0.005634843371808529, + 0.003139349864795804, + 0.002254169201478362, + 0.003042488358914852, + 0.00915867555886507, + 0.002258414402604103, + 0.0027232905849814415, + 0.0036501691211014986, + 0.004248236306011677, + 0.002953135874122381, + 0.004861005581915379, + 0.004765176214277744, + 0.003381977090612054, + 0.002616373123601079, + 0.0037024803459644318, + 0.0038015912286937237, + 0.004526171367615461, + 0.005089661106467247, + 0.003692298661917448, + 0.004014196805655956, + 0.004375757183879614, + 0.005190008319914341, + 0.005820970982313156, + 0.0683864876627922, + 0.031026078388094902, + 0.07022549957036972, + 0.037971071898937225, + 0.02581321820616722, + 0.0934501588344574, + 0.09388618171215057, + 0.10163827240467072, + 0.008770913816988468, + 0.009823578409850597, + 0.006842365954071283, + 0.0048779998905956745, + 0.004648968577384949, + 0.007473469246178865, + 0.003992245066910982, + 0.007374627981334925, + 0.003770876443013549, + 0.0028708523605018854, + 0.018906397745013237, + 0.003024503355845809, + 0.002970175351947546, + 0.020032059401273727, + 0.002102220430970192, + 0.0031905262731015682, + 0.004427047446370125, + 0.0038364618085324764, + 0.006950368173420429, + 0.005777457728981972, + 0.0030943849124014378, + 0.004365730099380016, + 0.0033242572098970413, + 0.004959345329552889, + 0.0066224136389791965, + 0.007379603572189808, + 0.00803727563470602, + 0.007875802926719189, + 0.006206208374351263, + 0.010014673694968224, + 0.005850648041814566, + 0.009738852269947529, + 0.007919522933661938, + 0.008496250957250595, + 0.012216197326779366, + 0.0085526704788208, + 0.010255156084895134, + 0.008901393041014671, + 0.006129926070570946, + 0.006511349231004715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0032014804892241955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731571163982153 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731571163982153, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731571163982153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06482496857643127, + 0.030857911333441734, + 0.029938437044620514, + 0.03813963755965233, + 0.40459519624710083, + 0.10306656360626221, + 0.09821366518735886, + 0.10624292492866516, + 0.07082249224185944, + 0.05329817533493042 + ], + "credit_map": [ + 0.012563156895339489, + 0.009064856916666031, + 0.006490292493253946, + 0.0075686657801270485, + 0.0063517470844089985, + 0.006018053740262985, + 0.008650449104607105, + 0.008117737248539925, + 0.003340755123645067, + 0.007257347460836172, + 0.0029708778019994497, + 0.0031254179775714874, + 0.004113486036658287, + 0.0049320547841489315, + 0.0022488196846097708, + 0.002869150135666132, + 0.0035176463425159454, + 0.003491730196401477, + 0.0029707206413149834, + 0.004185773432254791, + 0.005103233270347118, + 0.0034175824839621782, + 0.0029510604217648506, + 0.004300687462091446, + 0.004526098724454641, + 0.004548486787825823, + 0.004665026441216469, + 0.004277137573808432, + 0.004123545251786709, + 0.005135098472237587, + 0.0052136923186481, + 0.005650548730045557, + 0.05207477882504463, + 0.022391846403479576, + 0.052491866052150726, + 0.030345918610692024, + 0.021575327962636948, + 0.07517366856336594, + 0.07587449252605438, + 0.0746673122048378, + 0.014845674857497215, + 0.008179182186722755, + 0.019237974658608437, + 0.007253032177686691, + 0.006403422448784113, + 0.01508645061403513, + 0.008783039636909962, + 0.023277781903743744, + 0.008070754818618298, + 0.004055471625179052, + 0.03289146348834038, + 0.0036904383450746536, + 0.004154012072831392, + 0.030280346050858498, + 0.004459479358047247, + 0.010611693374812603, + 0.007962406612932682, + 0.0045168097130954266, + 0.03242671489715576, + 0.003809130983427167, + 0.004068213049322367, + 0.03472704440355301, + 0.003994717728346586, + 0.014737886376678944, + 0.013589856214821339, + 0.0031353861559182405, + 0.009814100340008736, + 0.004587843548506498, + 0.0030128243379294872, + 0.012298790737986565, + 0.004384227562695742, + 0.019999457523226738, + 0.008128363639116287, + 0.005712447687983513, + 0.009482023306190968, + 0.004776678513735533, + 0.004763839300721884, + 0.0077800932340323925, + 0.003951811697334051, + 0.008702912367880344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.006731571163982153 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle, static wide shot captures the towering silhouettes of a modern city. These sharp, geometric skyscrapers stand silently in pure black in the lower half of the frame. A strong desaturated color grading is employed throughout, creating a calm yet striking black-and-white quality. The sky is a deep, almost black canvas. Amidst this silent grandeur, a total lunar eclipse unfolds: Earth's umbra, like an invisible curtain, steadily and relentlessly engulfs the moon's luminous silver disk. As the last shimmer of silver light is swallowed by darkness, the moon, far from disappearing, glows a deep, saturated, blood-red. In this almost entirely monochrome world, the crimson moon provides the sole, breathtaking source of color, its ominous and majestic halo standing out against the stark black-and-white contrast. The entire dynamic unfolds silently, creating an apocalyptic, awe-inspiring stillness and visual tension.", + "seed": 7557, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_11/index_2992_seed_7557_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.29508465388789773, + "vlm_logprob": -0.29508465388789773, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.29508465388789773, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044367339462041855 + ], + "opposite_logprob_tokens": [ + -3.6693673133850098 + ], + "logprob_reward": -0.044367339462041855, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.044367339462041855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1373630315065384, + 0.10052400082349777, + 0.044024404138326645, + 0.20398545265197754, + 0.32591623067855835, + 0.0855337455868721, + 0.02804812230169773, + 0.03303387016057968, + 0.035432327538728714, + 0.006138855125755072 + ], + "credit_map": [ + 0.011307691223919392, + 0.01525858137756586, + 0.014831461943686008, + 0.027929233387112617, + 0.021358557045459747, + 0.01622115448117256, + 0.013514066115021706, + 0.016942281275987625, + 0.005484149791300297, + 0.037972621619701385, + 0.0057726358063519, + 0.008866771124303341, + 0.029918989166617393, + 0.007548321038484573, + 0.0022684060968458652, + 0.002692091977223754, + 0.0024810947943478823, + 0.00651304516941309, + 0.005559761077165604, + 0.0033816969953477383, + 0.008939189836382866, + 0.012286546640098095, + 0.0023432502057403326, + 0.002519818488508463, + 0.006492889020591974, + 0.037226662039756775, + 0.00723193259909749, + 0.006715748459100723, + 0.08838648349046707, + 0.04929467663168907, + 0.004345398396253586, + 0.0042916457168757915, + 0.0138733871281147, + 0.05321751907467842, + 0.02102779783308506, + 0.00937593448907137, + 0.13385994732379913, + 0.07471344619989395, + 0.00958876870572567, + 0.01025941502302885, + 0.0033911026548594236, + 0.0038480651564896107, + 0.0076978495344519615, + 0.0028087086975574493, + 0.030219927430152893, + 0.03172808885574341, + 0.0025232958141714334, + 0.003316702554002404, + 0.0023622671142220497, + 0.0016500810161232948, + 0.006943690124899149, + 0.0025014362763613462, + 0.0035385331138968468, + 0.004869956523180008, + 0.0020892068278044462, + 0.004092947579920292, + 0.0032937335781753063, + 0.003577049123123288, + 0.005082790739834309, + 0.007013176567852497, + 0.003839861834421754, + 0.005151340737938881, + 0.0017389025306329131, + 0.003337013302370906, + 0.003439495572820306, + 0.002568054711446166, + 0.008997519500553608, + 0.0028789301868528128, + 0.003910643979907036, + 0.006567820906639099, + 0.002253524260595441, + 0.0048163351602852345, + 0.0006378396064974368, + 0.0008941279374994338, + 0.0012812605127692223, + 0.0006723541300743818, + 0.0007504134555347264, + 0.0009419427369721234, + 0.0003483263135422021, + 0.0006125903455540538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.044367339462041855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011102918535470963 + ], + "opposite_logprob_tokens": [ + -4.51110315322876 + ], + "logprob_reward": -0.011102918535470963, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011102918535470963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16827699542045593, + 0.09258425980806351, + 0.09941918402910233, + 0.34527766704559326, + 0.14646659791469574, + 0.03629154711961746, + 0.02541782520711422, + 0.029640501365065575, + 0.0333210714161396, + 0.023304326459765434 + ], + "credit_map": [ + 0.01306707039475441, + 0.02685767225921154, + 0.017759907990694046, + 0.030442560091614723, + 0.025480376556515694, + 0.01976044662296772, + 0.015438382513821125, + 0.019470572471618652, + 0.00519946776330471, + 0.03308200463652611, + 0.006284626200795174, + 0.011205692775547504, + 0.024766679853200912, + 0.007294093258678913, + 0.002386713633313775, + 0.0023649795912206173, + 0.004556470084935427, + 0.01788952946662903, + 0.010122154839336872, + 0.004601231310516596, + 0.034974295645952225, + 0.016790024936199188, + 0.004922208376228809, + 0.005563271231949329, + 0.015598288737237453, + 0.048765067011117935, + 0.011392180807888508, + 0.0030251415446400642, + 0.15663200616836548, + 0.0989217460155487, + 0.006365961395204067, + 0.004577274434268475, + 0.006560545414686203, + 0.027690857648849487, + 0.00926615484058857, + 0.0023347700480371714, + 0.06182039901614189, + 0.031752604991197586, + 0.0035899353679269552, + 0.0034513308200985193, + 0.001965000992640853, + 0.0018544492777436972, + 0.0024113324470818043, + 0.0016216437797993422, + 0.01212122943252325, + 0.013913214206695557, + 0.0012656394392251968, + 0.0011390401050448418, + 0.002178874332457781, + 0.0024705962277948856, + 0.0040224227122962475, + 0.004921442829072475, + 0.0039206380024552345, + 0.00487889489158988, + 0.0017823149682953954, + 0.001242640195414424, + 0.0031104995869100094, + 0.0037506369408220053, + 0.004658171441406012, + 0.006694351322948933, + 0.003546176478266716, + 0.003867363091558218, + 0.0017972192727029324, + 0.002216083463281393, + 0.003377350978553295, + 0.0024176305159926414, + 0.007632269989699125, + 0.002328383270651102, + 0.004007854498922825, + 0.006862524431198835, + 0.002214248524978757, + 0.0044808099046349525, + 0.001964618219062686, + 0.0014388872077688575, + 0.006236892193555832, + 0.0011757100000977516, + 0.002112980931997299, + 0.003812965005636215, + 0.0019175250781700015, + 0.004644749220460653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.011102918535470963 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063537120819092 + ], + "opposite_logprob_tokens": [ + -0.2813536524772644 + ], + "logprob_reward": -1.4063537120819092, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063537120819092, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17232419550418854, + 0.17206527292728424, + 0.18494541943073273, + 0.26630347967147827, + 0.04998849704861641, + 0.0426335446536541, + 0.037604108452796936, + 0.022796200588345528, + 0.04159023240208626, + 0.00974906887859106 + ], + "credit_map": [ + 0.012523033656179905, + 0.023338329046964645, + 0.02268935739994049, + 0.041058409959077835, + 0.023949481546878815, + 0.021780356764793396, + 0.012126252986490726, + 0.014858949929475784, + 0.011240553110837936, + 0.06837527453899384, + 0.008505340665578842, + 0.012645289301872253, + 0.04945271834731102, + 0.014219744130969048, + 0.0034532598219811916, + 0.0041730813682079315, + 0.009673879481852055, + 0.03000245988368988, + 0.018234405666589737, + 0.008435243740677834, + 0.06830623000860214, + 0.03614283353090286, + 0.007123860530555248, + 0.007026501931250095, + 0.008100801147520542, + 0.02751370146870613, + 0.009456000290811062, + 0.00779646635055542, + 0.11265214532613754, + 0.08962786197662354, + 0.005222440231591463, + 0.005934057757258415, + 0.004550789017230272, + 0.005821344442665577, + 0.00906310509890318, + 0.00422688527032733, + 0.009886039420962334, + 0.009892940521240234, + 0.002987536136060953, + 0.0035598520189523697, + 0.003105136100202799, + 0.0028921153862029314, + 0.006506324280053377, + 0.005050648469477892, + 0.010765990242362022, + 0.010264184325933456, + 0.0021883505396544933, + 0.0018607905367389321, + 0.0037716568913310766, + 0.004463340155780315, + 0.0065429019741714, + 0.008017786778509617, + 0.004796185530722141, + 0.00528955040499568, + 0.0025825395714491606, + 0.0021401450503617525, + 0.0020330501720309258, + 0.002062972402200103, + 0.00566883897408843, + 0.0023892109747976065, + 0.0032088470179587603, + 0.0035622844006866217, + 0.0015514877159148455, + 0.002319508697837591, + 0.0037775670643895864, + 0.004419153090566397, + 0.009387708269059658, + 0.004383524879813194, + 0.0057849218137562275, + 0.006305749528110027, + 0.0026776117738336325, + 0.004853994119912386, + 0.0008897013030946255, + 0.0008941479027271271, + 0.002929177600890398, + 0.0006778482347726822, + 0.000981125165708363, + 0.0016295935492962599, + 0.0004897589096799493, + 0.0012577153975144029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -1.4063537120819092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647877559065819 + ], + "opposite_logprob_tokens": [ + -4.88264799118042 + ], + "logprob_reward": -0.007647877559065819, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647877559065819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25466564297676086, + 0.1251010149717331, + 0.0631294921040535, + 0.23132531344890594, + 0.11565645784139633, + 0.11675019562244415, + 0.0357351116836071, + 0.02559746243059635, + 0.024306846782565117, + 0.007732470519840717 + ], + "credit_map": [ + 0.012153098359704018, + 0.07179181277751923, + 0.022572671994566917, + 0.030959054827690125, + 0.07135738432407379, + 0.02436271868646145, + 0.010156730189919472, + 0.011312163434922695, + 0.008210637606680393, + 0.0501292385160923, + 0.005547508131712675, + 0.008838030509650707, + 0.03652491047978401, + 0.010258258320391178, + 0.0025151295121759176, + 0.0030772965401411057, + 0.0049567921087145805, + 0.008150691166520119, + 0.008675454184412956, + 0.004580941051244736, + 0.013103899545967579, + 0.011144101619720459, + 0.004936679266393185, + 0.007580931764096022, + 0.006273036822676659, + 0.027482949197292328, + 0.00718239601701498, + 0.004996019881218672, + 0.10085275024175644, + 0.07559017091989517, + 0.0041894870810210705, + 0.004758509341627359, + 0.007043227553367615, + 0.013943853788077831, + 0.01495524775236845, + 0.00785993505269289, + 0.036404918879270554, + 0.024927468970417976, + 0.005207855254411697, + 0.005313952453434467, + 0.004968492314219475, + 0.0064456588588654995, + 0.013787716627120972, + 0.004714089445769787, + 0.0373707115650177, + 0.042137015610933304, + 0.0032545244321227074, + 0.0040719881653785706, + 0.004330093506723642, + 0.0030710804276168346, + 0.007849026471376419, + 0.0054335035383701324, + 0.004098596982657909, + 0.005088285077363253, + 0.002281217835843563, + 0.0035833041183650494, + 0.0025429364759474993, + 0.001994667574763298, + 0.006382668856531382, + 0.0029821028001606464, + 0.003095172578468919, + 0.004260366782546043, + 0.0014500587712973356, + 0.0028894899878650904, + 0.0026610572822391987, + 0.0024856373202055693, + 0.004506452474743128, + 0.0020575132220983505, + 0.003427454736083746, + 0.004755014553666115, + 0.0015608950052410364, + 0.0028528221882879734, + 0.0008265324868261814, + 0.0009969259845092893, + 0.001827710191719234, + 0.0007395711727440357, + 0.0010568569414317608, + 0.0012412382056936622, + 0.00048489467008039355, + 0.0005587407504208386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007647877559065819 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059514218010008335 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.0059514218010008335, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059514218010008335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5875972509384155, + 0.06810212135314941, + 0.09095702320337296, + 0.1410306692123413, + 0.04677596688270569, + 0.02108103409409523, + 0.012351397424936295, + 0.0074024442583322525, + 0.019077405333518982, + 0.00562459509819746 + ], + "credit_map": [ + 0.03714878112077713, + 0.15450234711170197, + 0.0455683097243309, + 0.05449913442134857, + 0.16052192449569702, + 0.05343206226825714, + 0.036345746368169785, + 0.045579053461551666, + 0.004217527341097593, + 0.023676887154579163, + 0.00341067835688591, + 0.010063426569104195, + 0.01725073903799057, + 0.0057698809541761875, + 0.00183020846452564, + 0.0018827819731086493, + 0.0045260959304869175, + 0.01614289917051792, + 0.007980135269463062, + 0.0035642306320369244, + 0.034545931965112686, + 0.017625093460083008, + 0.0030924221500754356, + 0.0034802246373146772, + 0.0038925455883145332, + 0.019902287051081657, + 0.0043988884426653385, + 0.0029769146349281073, + 0.06263674050569534, + 0.04210025072097778, + 0.0025180140510201454, + 0.0026050396263599396, + 0.002856903476640582, + 0.0059536718763411045, + 0.006313206162303686, + 0.00429653050377965, + 0.013778528198599815, + 0.009365358389914036, + 0.002090746769681573, + 0.0021210231352597475, + 0.0014734796714037657, + 0.0019999004434794188, + 0.00517411669716239, + 0.0015225831884890795, + 0.0037741190753877163, + 0.004752781242132187, + 0.0011693229898810387, + 0.001214733812958002, + 0.0013387642102316022, + 0.001476201810874045, + 0.00239308993332088, + 0.0022456056904047728, + 0.0016643207054585218, + 0.001671510748565197, + 0.000714954046998173, + 0.0008469524909742177, + 0.0006906304624862969, + 0.000666732550598681, + 0.0014142637373879552, + 0.0007494893507100642, + 0.0013354390393942595, + 0.0012904386967420578, + 0.0005426107090897858, + 0.0007128402357921004, + 0.0019380254670977592, + 0.0023290044628083706, + 0.003968023229390383, + 0.0021364069543778896, + 0.002686478663235903, + 0.002830800600349903, + 0.001218122779391706, + 0.0019705453887581825, + 0.0005856615607626736, + 0.0004805994685739279, + 0.0016450142720714211, + 0.00046829748316667974, + 0.0006581418565474451, + 0.0009586148080416024, + 0.000355474156094715, + 0.00047279195860028267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0059514218010008335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle, static wide shot captures the towering silhouettes of a modern city. These sharp, geometric skyscrapers stand silently in pure black in the lower half of the frame. A strong desaturated color grading is employed throughout, creating a calm yet striking black-and-white quality. The sky is a deep, almost black canvas. Amidst this silent grandeur, a total lunar eclipse unfolds: Earth's umbra, like an invisible curtain, steadily and relentlessly engulfs the moon's luminous silver disk. As the last shimmer of silver light is swallowed by darkness, the moon, far from disappearing, glows a deep, saturated, blood-red. In this almost entirely monochrome world, the crimson moon provides the sole, breathtaking source of color, its ominous and majestic halo standing out against the stark black-and-white contrast. The entire dynamic unfolds silently, creating an apocalyptic, awe-inspiring stillness and visual tension.", + "seed": 7558, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_11/index_2992_seed_7558_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.044808465987443924, + "vlm_logprob": -0.044808465987443924, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.044808465987443924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05403438210487366 + ], + "opposite_logprob_tokens": [ + -3.429034471511841 + ], + "logprob_reward": -0.05403438210487366, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.05403438210487366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2370651513338089, + 0.10808754712343216, + 0.336775541305542, + 0.0466499887406826, + 0.03131387010216713, + 0.04579351842403412, + 0.04823170229792595, + 0.05450547859072685, + 0.0692267045378685, + 0.022350501269102097 + ], + "credit_map": [ + 0.02175872027873993, + 0.025113577023148537, + 0.0288758035749197, + 0.035400185734033585, + 0.03649959713220596, + 0.030736461281776428, + 0.03990807384252548, + 0.018772736191749573, + 0.010468457825481892, + 0.014628840610384941, + 0.010392370633780956, + 0.01574622094631195, + 0.02465406060218811, + 0.015482914634048939, + 0.009995857253670692, + 0.006718827877193689, + 0.046385809779167175, + 0.041092585772275925, + 0.036802567541599274, + 0.05577922239899635, + 0.0506005734205246, + 0.04759733006358147, + 0.0372602604329586, + 0.02125720866024494, + 0.005375112406909466, + 0.004198302980512381, + 0.007295054383575916, + 0.006414293311536312, + 0.0068530067801475525, + 0.006703838240355253, + 0.00618803221732378, + 0.0036223484203219414, + 0.003424136433750391, + 0.0032892245799303055, + 0.005520965903997421, + 0.004309695679694414, + 0.003630327060818672, + 0.0043548280373215675, + 0.004303921014070511, + 0.002480772091075778, + 0.005568549502640963, + 0.006132848560810089, + 0.005316452588886023, + 0.006151605397462845, + 0.00794563814997673, + 0.007102202624082565, + 0.005086091347038746, + 0.002490127691999078, + 0.006912477780133486, + 0.005037014372646809, + 0.005058562383055687, + 0.008738873526453972, + 0.004348783288151026, + 0.008454744704067707, + 0.005998166278004646, + 0.003683075774461031, + 0.008734413422644138, + 0.005556024145334959, + 0.0053084371611475945, + 0.009269827045500278, + 0.005460073705762625, + 0.009899063035845757, + 0.0064622340723872185, + 0.003815410193055868, + 0.009882914833724499, + 0.005915350746363401, + 0.010032158344984055, + 0.011593667790293694, + 0.005531026981770992, + 0.01131339743733406, + 0.009624104015529156, + 0.005334087181836367, + 0.0028368320781737566, + 0.002050819806754589, + 0.0024111291859298944, + 0.0037583166267722845, + 0.0032564273569732904, + 0.0033359811641275883, + 0.0030938927084207535, + 0.0016071017598733306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.05403438210487366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869622364640236 + ], + "opposite_logprob_tokens": [ + -3.048696279525757 + ], + "logprob_reward": -0.04869622364640236, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869622364640236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32933881878852844, + 0.0888877660036087, + 0.09117131680250168, + 0.04955810308456421, + 0.07113218307495117, + 0.10203470289707184, + 0.06606801599264145, + 0.13910230994224548, + 0.026742147281765938, + 0.03596462309360504 + ], + "credit_map": [ + 0.031123919412493706, + 0.041859425604343414, + 0.037714120000600815, + 0.0479736290872097, + 0.07258056849241257, + 0.04645051807165146, + 0.03400533273816109, + 0.01763131469488144, + 0.011491881683468819, + 0.012644501402974129, + 0.009249518625438213, + 0.014482932165265083, + 0.01211722381412983, + 0.010625887662172318, + 0.01029268465936184, + 0.0079831313341856, + 0.008722801692783833, + 0.011865835636854172, + 0.011277546174824238, + 0.011985164135694504, + 0.01779276505112648, + 0.011062528938055038, + 0.01271111611276865, + 0.005753566510975361, + 0.005184859968721867, + 0.006794595159590244, + 0.005042207892984152, + 0.006315639708191156, + 0.00915905088186264, + 0.007805341389030218, + 0.006609515752643347, + 0.0026468930300325155, + 0.007379807997494936, + 0.008968246169388294, + 0.00875631533563137, + 0.01038694940507412, + 0.014046264812350273, + 0.00928480178117752, + 0.008004826493561268, + 0.004304967820644379, + 0.009348648600280285, + 0.013509728945791721, + 0.008318733423948288, + 0.015618186444044113, + 0.020577264949679375, + 0.014730824157595634, + 0.013897967524826527, + 0.006033350247889757, + 0.008504188619554043, + 0.009272822178900242, + 0.00564637640491128, + 0.009558528661727905, + 0.009492688812315464, + 0.009913611225783825, + 0.009580019861459732, + 0.004099781159311533, + 0.016088031232357025, + 0.020021861419081688, + 0.010909013450145721, + 0.019897637888789177, + 0.026868583634495735, + 0.021611932665109634, + 0.01618080399930477, + 0.007524445187300444, + 0.0029495544731616974, + 0.0018523915205150843, + 0.00504258181899786, + 0.005540232174098492, + 0.0017686575883999467, + 0.004004353191703558, + 0.0037760059349238873, + 0.0018083713948726654, + 0.005033715162426233, + 0.0038939632941037416, + 0.004604066722095013, + 0.0070826224982738495, + 0.003792688949033618, + 0.0053297774866223335, + 0.0038309681694954634, + 0.002396821277216077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.04869622364640236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036002844572067 + ], + "opposite_logprob_tokens": [ + -2.3503599166870117 + ], + "logprob_reward": -0.10036002844572067, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036002844572067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2550666928291321, + 0.16063828766345978, + 0.19580909609794617, + 0.06710328161716461, + 0.045695073902606964, + 0.09295625984668732, + 0.049638811498880386, + 0.08312872797250748, + 0.02551381103694439, + 0.024449966847896576 + ], + "credit_map": [ + 0.02628451958298683, + 0.03766704350709915, + 0.02392234094440937, + 0.0362292155623436, + 0.04107973352074623, + 0.034811023622751236, + 0.037270739674568176, + 0.01780206523835659, + 0.01780962385237217, + 0.031283557415008545, + 0.011557075195014477, + 0.01634383760392666, + 0.038557831197977066, + 0.02208462730050087, + 0.01623356156051159, + 0.006768177263438702, + 0.019101951271295547, + 0.023780884221196175, + 0.020223775878548622, + 0.02749691531062126, + 0.025818102061748505, + 0.025387806817889214, + 0.042081013321876526, + 0.011918644420802593, + 0.007147998083382845, + 0.010276801884174347, + 0.005314996466040611, + 0.010801317170262337, + 0.013292322866618633, + 0.009491733275353909, + 0.006718209013342857, + 0.004059897735714912, + 0.005182676948606968, + 0.006024724803864956, + 0.005472069606184959, + 0.006831508129835129, + 0.007361655589193106, + 0.006345315836369991, + 0.0053598699159920216, + 0.003117254702374339, + 0.009152628481388092, + 0.013063659891486168, + 0.00811680220067501, + 0.013258546590805054, + 0.020771292969584465, + 0.013690156862139702, + 0.009727118536829948, + 0.00517605384811759, + 0.00432874308899045, + 0.007618548348546028, + 0.004374834708869457, + 0.007817832753062248, + 0.008908560499548912, + 0.008148600347340107, + 0.005364846903830767, + 0.003076844150200486, + 0.009478423744440079, + 0.012202643789350986, + 0.005646119359880686, + 0.011115534231066704, + 0.016653839498758316, + 0.012959550134837627, + 0.010360331274569035, + 0.004712287802249193, + 0.003868637140840292, + 0.002070165006443858, + 0.003217455465346575, + 0.0050564599223434925, + 0.0020252002868801355, + 0.004277615807950497, + 0.0031471040565520525, + 0.0018511731177568436, + 0.0030018524266779423, + 0.0022913378197699785, + 0.0026577957905828953, + 0.004453577566891909, + 0.004137348383665085, + 0.0035217509139329195, + 0.0025422817561775446, + 0.0018440219573676586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.10036002844572067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216572046279907 + ], + "opposite_logprob_tokens": [ + -4.264216423034668 + ], + "logprob_reward": -0.014216572046279907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216572046279907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16332733631134033, + 0.07742875814437866, + 0.09689266979694366, + 0.12388435751199722, + 0.09581945091485977, + 0.08964087069034576, + 0.13362005352973938, + 0.09075826406478882, + 0.05297842249274254, + 0.07564979046583176 + ], + "credit_map": [ + 0.014595338143408298, + 0.02994999848306179, + 0.012699578888714314, + 0.0206866804510355, + 0.03484179824590683, + 0.023977719247341156, + 0.01762750744819641, + 0.008948715403676033, + 0.008322595618665218, + 0.013924241997301579, + 0.006075429264456034, + 0.009616706520318985, + 0.01834765262901783, + 0.010341684333980083, + 0.007392365951091051, + 0.0034080827608704567, + 0.011777406558394432, + 0.011676586233079433, + 0.011979322880506516, + 0.01372560765594244, + 0.016505872830748558, + 0.01406979002058506, + 0.010855267755687237, + 0.006302819587290287, + 0.013865577057003975, + 0.01635667122900486, + 0.012322155758738518, + 0.017161117866635323, + 0.025215860456228256, + 0.017491141334176064, + 0.013135192915797234, + 0.008336642757058144, + 0.011214038357138634, + 0.01435227319598198, + 0.009568311274051666, + 0.01358955167233944, + 0.016268353909254074, + 0.01378733292222023, + 0.01007874682545662, + 0.006960837636142969, + 0.009146442636847496, + 0.013649548403918743, + 0.005641214549541473, + 0.012649045325815678, + 0.021081190556287766, + 0.015171337872743607, + 0.0077632637694478035, + 0.004538832698017359, + 0.017270762473344803, + 0.015946974977850914, + 0.015700645744800568, + 0.021075120195746422, + 0.01934479922056198, + 0.02008787915110588, + 0.013635624200105667, + 0.010558245703577995, + 0.009606284089386463, + 0.010858616791665554, + 0.00997487735003233, + 0.011755319312214851, + 0.01675848290324211, + 0.012475373223423958, + 0.012485461309552193, + 0.006843851879239082, + 0.0058625019155442715, + 0.008349554613232613, + 0.0037518134340643883, + 0.005825621075928211, + 0.012729860842227936, + 0.008741811849176884, + 0.0052488334476947784, + 0.0024684283416718245, + 0.009076963178813457, + 0.010644746944308281, + 0.00487965764477849, + 0.008925735019147396, + 0.018504692241549492, + 0.013556260615587234, + 0.006177783012390137, + 0.003883946919813752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.014216572046279907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static wide shot with a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are sharp, geometric skyscrapers visible as silhouettes in the lower half of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with a desaturated, black-and-white color grading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moon glow a deep, saturated blood-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky depicted as a deep, almost black canvas?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735123693943024 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735123693943024, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735123693943024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3109619915485382, + 0.08875170350074768, + 0.4204023480415344, + 0.028783883899450302, + 0.027353305369615555, + 0.02622494287788868, + 0.02578626573085785, + 0.028429122641682625, + 0.02262892574071884, + 0.02067745476961136 + ], + "credit_map": [ + 0.03345991298556328, + 0.05456147342920303, + 0.030722757801413536, + 0.04215709865093231, + 0.051262177526950836, + 0.04337674751877785, + 0.0355541966855526, + 0.01986762322485447, + 0.009951216168701649, + 0.011342162266373634, + 0.009739963337779045, + 0.013530310243368149, + 0.015306923538446426, + 0.01220838911831379, + 0.00946554820984602, + 0.007207191549241543, + 0.053980473428964615, + 0.04983724653720856, + 0.050410620868206024, + 0.06926707923412323, + 0.060592178255319595, + 0.05796243995428085, + 0.04822854697704315, + 0.030123794451355934, + 0.0030111621599644423, + 0.0032406141981482506, + 0.002948559820652008, + 0.003493319498375058, + 0.00628748070448637, + 0.004364346154034138, + 0.0036458063405007124, + 0.00179259420838207, + 0.002810728969052434, + 0.0032927750144153833, + 0.0033045602031052113, + 0.004468097817152739, + 0.004336508456617594, + 0.0036849752068519592, + 0.003674582112580538, + 0.0017810791032388806, + 0.002744930097833276, + 0.0031792190857231617, + 0.002534050028771162, + 0.004150899592787027, + 0.004729738924652338, + 0.004469475243240595, + 0.0031713431235402822, + 0.0012452879454940557, + 0.0032249230425804853, + 0.003291553584858775, + 0.0019600847736001015, + 0.003853525733575225, + 0.00425276393070817, + 0.004448980558663607, + 0.003120990702882409, + 0.00163344107568264, + 0.0026994983199983835, + 0.003874069545418024, + 0.003088494762778282, + 0.0031263527926057577, + 0.006475990172475576, + 0.004574461840093136, + 0.0030731039587408304, + 0.001517150434665382, + 0.0027833126951009035, + 0.0028504212386906147, + 0.002488963305950165, + 0.003025870770215988, + 0.004104923456907272, + 0.0031254359055310488, + 0.0028298997785896063, + 0.0014201002195477486, + 0.002370287664234638, + 0.0024951561354100704, + 0.001665327581577003, + 0.0023135263472795486, + 0.004848374053835869, + 0.003662406699731946, + 0.0020455243065953255, + 0.0012768511660397053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.006735123693943024 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_12/index_tensor([6274, 6274])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_12/index_tensor([6274, 6274])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6771d778ce07d8664ee2651a8f36a341d2a1899c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_12/index_tensor([6274, 6274])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In one uninterrupted shot, the camera, at eye level, performs a slow, graceful 360-degree orbit around a serene, white-bearded sage seated in perfect lotus pose beneath a majestic, stylized tree. The sage, his long, flowing white beard resting on his chest, his eyes gently closed, lost in deep meditation, maintaining absolute stillness. His body is clad in a simple off-white silk robe. The tree above him is a fantastical creation, with a gnarled, pearly trunk gracefully winding toward a crown of crystal leaves that shimmer with a soft inner glow. The background is a dreamy, soft-focus, pastel-colored cloudscape. As the camera glides, a point of light appears on the glowing mossy ground next to the sage, sprouting a new life cycle – a magical plant majestically accelerating its life cycle. The stem spirals upward, unfurling its leaves to reveal an intricate, bioluminescent, whimsical pattern. The flowers then blossom into a vibrant spectacle of unfolding fractals and incredible colors, releasing soft pulses of light. Simultaneously, these same intricate patterns begin to flow and echo across the surface of the saint's robes and ripple across the dreamlike background, visually weaving the saint, the plants, and the universe into a single, interconnected tapestry of quiet magic and unfolding wonder.", + "seed": 20847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_12/index_6274_seed_20847_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03779913699254393, + "vlm_logprob": -0.03779913699254393, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.03779913699254393, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038281031884253025 + ], + "opposite_logprob_tokens": [ + -6.628828048706055 + ], + "logprob_reward": -0.0038281031884253025, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0038281031884253025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16559608280658722, + 0.16632811725139618, + 0.0740293487906456, + 0.04248534515500069, + 0.14822274446487427, + 0.024928614497184753, + 0.18208931386470795, + 0.058866214007139206, + 0.06887645274400711, + 0.06857777386903763 + ], + "credit_map": [ + 0.024077609181404114, + 0.025040969252586365, + 0.018894091248512268, + 0.015165170654654503, + 0.02454068697988987, + 0.023562483489513397, + 0.020619681105017662, + 0.013695389032363892, + 0.03837229311466217, + 0.01959933340549469, + 0.010385262779891491, + 0.013293227180838585, + 0.020215312018990517, + 0.036856621503829956, + 0.016334809362888336, + 0.011271258816123009, + 0.013218111358582973, + 0.008244600147008896, + 0.004172662738710642, + 0.00718294782564044, + 0.010480009019374847, + 0.013429357670247555, + 0.010203682817518711, + 0.007097981404513121, + 0.005554531700909138, + 0.007700163405388594, + 0.003916570451110601, + 0.00432272395119071, + 0.006366393528878689, + 0.004706828389316797, + 0.00469377962872386, + 0.005224354099482298, + 0.028920717537403107, + 0.03404346853494644, + 0.011370649561285973, + 0.01326802745461464, + 0.011776619590818882, + 0.017573658376932144, + 0.01817379705607891, + 0.013095802627503872, + 0.0033907308243215084, + 0.0031592247541993856, + 0.0023138292599469423, + 0.0038561266846954823, + 0.003215997712686658, + 0.003185874782502651, + 0.0037698110099881887, + 0.0020370190031826496, + 0.04127238690853119, + 0.056960154324769974, + 0.011646653525531292, + 0.013883302919566631, + 0.018952390179038048, + 0.01853306218981743, + 0.011678694747388363, + 0.009162665344774723, + 0.00747210718691349, + 0.010004239156842232, + 0.006243271753191948, + 0.007238578516989946, + 0.007310481742024422, + 0.005237994249910116, + 0.009944728575646877, + 0.005414816550910473, + 0.007241683546453714, + 0.011380919255316257, + 0.0054540266282856464, + 0.00710200285539031, + 0.008876074105501175, + 0.00864073634147644, + 0.012310752645134926, + 0.007870258763432503, + 0.01047598198056221, + 0.011343355290591717, + 0.005311773624271154, + 0.006993880495429039, + 0.009549619629979134, + 0.006395685952156782, + 0.012750337831676006, + 0.005757137201726437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0038281031884253025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058476023375988 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.02058476023375988, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058476023375988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11366326361894608, + 0.12474357336759567, + 0.13408559560775757, + 0.05763015151023865, + 0.11611492931842804, + 0.0572047159075737, + 0.08109384775161743, + 0.09948252141475677, + 0.13470005989074707, + 0.08128129690885544 + ], + "credit_map": [ + 0.01817522570490837, + 0.018636854365468025, + 0.011190848425030708, + 0.007358161266893148, + 0.02095700614154339, + 0.01975143700838089, + 0.009912004694342613, + 0.007681719493120909, + 0.023814089596271515, + 0.01741648279130459, + 0.010279458947479725, + 0.009858625009655952, + 0.016078200191259384, + 0.0252645593136549, + 0.011717351153492928, + 0.010314813815057278, + 0.02995903044939041, + 0.02046392299234867, + 0.007764206267893314, + 0.009040791541337967, + 0.013089715503156185, + 0.03223400563001633, + 0.01376334112137556, + 0.007770579308271408, + 0.00839676521718502, + 0.011223251931369305, + 0.005410476122051477, + 0.0066859242506325245, + 0.00701082032173872, + 0.007196580525487661, + 0.007225935347378254, + 0.004480401985347271, + 0.01825244538486004, + 0.017718831077218056, + 0.008486105129122734, + 0.0075016384944319725, + 0.022137567400932312, + 0.009193123318254948, + 0.021281298249959946, + 0.011543920263648033, + 0.008796564303338528, + 0.006630919873714447, + 0.008443204686045647, + 0.011345461010932922, + 0.004200706724077463, + 0.004471437074244022, + 0.007000934798270464, + 0.006315488368272781, + 0.012832791544497013, + 0.012287982739508152, + 0.008044365793466568, + 0.013007272966206074, + 0.009405052289366722, + 0.008771299384534359, + 0.009606609120965004, + 0.007138470187783241, + 0.0125977648422122, + 0.01319111417979002, + 0.010546371340751648, + 0.009579936973750591, + 0.01659463532269001, + 0.010675341822206974, + 0.014650982804596424, + 0.011646370403468609, + 0.01639437861740589, + 0.02398035302758217, + 0.01149427704513073, + 0.01569342613220215, + 0.012191045098006725, + 0.01994844526052475, + 0.0246203001588583, + 0.010377831757068634, + 0.012127668596804142, + 0.016614675521850586, + 0.005628985818475485, + 0.007222270127385855, + 0.009309517219662666, + 0.009404338896274567, + 0.01436943095177412, + 0.006604413967579603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.02058476023375988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195299707353115 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195299707353115, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195299707353115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16921685636043549, + 0.14142265915870667, + 0.0718691498041153, + 0.03503746911883354, + 0.04110657796263695, + 0.035855118185281754, + 0.22283324599266052, + 0.07589629292488098, + 0.07861816883087158, + 0.12814447283744812 + ], + "credit_map": [ + 0.018854280933737755, + 0.03457619622349739, + 0.014092402532696724, + 0.0179157517850399, + 0.032555948942899704, + 0.01635323092341423, + 0.02371172606945038, + 0.011157318949699402, + 0.036689747124910355, + 0.021466543897986412, + 0.008321772329509258, + 0.00984827522188425, + 0.009994763880968094, + 0.03933641314506531, + 0.008044001646339893, + 0.0077211433090269566, + 0.009190932847559452, + 0.00875002983957529, + 0.007892238907516003, + 0.006511662621051073, + 0.009901346638798714, + 0.009512844495475292, + 0.013903739862143993, + 0.006206358317285776, + 0.005592772737145424, + 0.004157160874456167, + 0.004059079568833113, + 0.004802104085683823, + 0.004343848209828138, + 0.0034650207962840796, + 0.0052328333258628845, + 0.003384646028280258, + 0.0053754402324557304, + 0.00443004397675395, + 0.00353032979182899, + 0.005290621425956488, + 0.006540282629430294, + 0.003423969028517604, + 0.007440212648361921, + 0.0050756800919771194, + 0.004503753501921892, + 0.005388376768678427, + 0.003332864958792925, + 0.005175616592168808, + 0.0032810152042657137, + 0.0062716747634112835, + 0.004769046325236559, + 0.0031327703036367893, + 0.04802212864160538, + 0.0477658249437809, + 0.02603110298514366, + 0.03519781306385994, + 0.013312146067619324, + 0.014355910941958427, + 0.013111582957208157, + 0.025036733597517014, + 0.009643618948757648, + 0.014548087492585182, + 0.0072022550739347935, + 0.010587604716420174, + 0.009336032904684544, + 0.006706493441015482, + 0.007353633176535368, + 0.010518562979996204, + 0.014502575621008873, + 0.01482035219669342, + 0.005645120050758123, + 0.004258420784026384, + 0.010186022147536278, + 0.011339511722326279, + 0.012281394563615322, + 0.005584773141890764, + 0.017963938415050507, + 0.029663536697626114, + 0.011384759098291397, + 0.011824050918221474, + 0.010687877424061298, + 0.013573463074862957, + 0.016741901636123657, + 0.016304947435855865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.014195299707353115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427636593580246 + ], + "opposite_logprob_tokens": [ + -2.017763614654541 + ], + "logprob_reward": -0.1427636593580246, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427636593580246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17185617983341217, + 0.11140898615121841, + 0.17348447442054749, + 0.05140182375907898, + 0.07030020654201508, + 0.08164273202419281, + 0.09574808180332184, + 0.04324091598391533, + 0.11030653119087219, + 0.09061005711555481 + ], + "credit_map": [ + 0.023780809715390205, + 0.024891909211874008, + 0.018368704244494438, + 0.015563960187137127, + 0.02239682711660862, + 0.027989860624074936, + 0.02052871696650982, + 0.01833540014922619, + 0.03131762892007828, + 0.009595622308552265, + 0.0051207286305725574, + 0.007459074258804321, + 0.008253061212599277, + 0.03315887227654457, + 0.007343597244471312, + 0.009160405956208706, + 0.025865888223052025, + 0.02261253446340561, + 0.009821271523833275, + 0.015078563243150711, + 0.03565790131688118, + 0.01918756030499935, + 0.03265761211514473, + 0.012603148818016052, + 0.006483006291091442, + 0.0077471123076975346, + 0.003966850228607655, + 0.00459479633718729, + 0.009085344150662422, + 0.006715178024023771, + 0.009461358189582825, + 0.003348178695887327, + 0.016853315755724907, + 0.016403310000896454, + 0.004310814663767815, + 0.00479762302711606, + 0.007045007310807705, + 0.007560375612229109, + 0.007788792718201876, + 0.0055409688502550125, + 0.011165250092744827, + 0.010905829258263111, + 0.010441734455525875, + 0.008073375560343266, + 0.009422154165804386, + 0.013752227649092674, + 0.010506454855203629, + 0.00737570459023118, + 0.011218705214560032, + 0.015297427773475647, + 0.009383440017700195, + 0.012208374217152596, + 0.012332383543252945, + 0.010536790825426579, + 0.011985240504145622, + 0.012785724364221096, + 0.005626691039651632, + 0.007391557563096285, + 0.0052107879891991615, + 0.006474749650806189, + 0.003963703289628029, + 0.003234144300222397, + 0.005798760335892439, + 0.005540521815419197, + 0.014932558871805668, + 0.027112876996397972, + 0.011163182556629181, + 0.007577279582619667, + 0.012216292321681976, + 0.01276325061917305, + 0.015900760889053345, + 0.008640331216156483, + 0.008206840604543686, + 0.016411326825618744, + 0.007062187884002924, + 0.01053637359291315, + 0.014499393291771412, + 0.008334694430232048, + 0.015110939741134644, + 0.01044829748570919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.1427636593580246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623862475156784 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623862475156784, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623862475156784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11473498493432999, + 0.2136850655078888, + 0.03786822408437729, + 0.03247610852122307, + 0.05373531952500343, + 0.02296965941786766, + 0.16399747133255005, + 0.069814033806324, + 0.10726677626371384, + 0.18345238268375397 + ], + "credit_map": [ + 0.015172331593930721, + 0.022650357335805893, + 0.009824896231293678, + 0.008907935582101345, + 0.02719051204621792, + 0.009003562852740288, + 0.013829377479851246, + 0.008156011812388897, + 0.070025734603405, + 0.018377477303147316, + 0.008095677942037582, + 0.007212642580270767, + 0.014554213732481003, + 0.07387597113847733, + 0.01115708239376545, + 0.010386248119175434, + 0.00546147720888257, + 0.00554988719522953, + 0.0032882741652429104, + 0.005143360234797001, + 0.004527720622718334, + 0.0052741882391273975, + 0.005484577268362045, + 0.0031387391500175, + 0.005419834051281214, + 0.0052703311666846275, + 0.0025448910892009735, + 0.0038739419542253017, + 0.0030828563030809164, + 0.005319160874933004, + 0.0043121566995978355, + 0.0026529314927756786, + 0.008545929566025734, + 0.007932785898447037, + 0.00609879894182086, + 0.0086261797696352, + 0.006860316265374422, + 0.004117961041629314, + 0.006578482687473297, + 0.0049748653545975685, + 0.002897566184401512, + 0.0032105124555528164, + 0.0025434510316699743, + 0.004302527289837599, + 0.0022163554094731808, + 0.0032405860256403685, + 0.002764906734228134, + 0.0017937547527253628, + 0.03998114541172981, + 0.042651690542697906, + 0.01140214130282402, + 0.014456706121563911, + 0.014911690726876259, + 0.01478070393204689, + 0.01383309904485941, + 0.011980298906564713, + 0.009745570831000805, + 0.01183643564581871, + 0.008014780469238758, + 0.011939417570829391, + 0.006650460883975029, + 0.006568291690200567, + 0.00616689957678318, + 0.008892178535461426, + 0.026380041614174843, + 0.014047947712242603, + 0.007827372290194035, + 0.007033069618046284, + 0.00804795604199171, + 0.028035366907715797, + 0.00936244148761034, + 0.006532575935125351, + 0.0299593023955822, + 0.04948641359806061, + 0.016712557524442673, + 0.016051804646849632, + 0.013328853994607925, + 0.02255118079483509, + 0.020418688654899597, + 0.014943571761250496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.007623862475156784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one uninterrupted shot, the camera, at eye level, performs a slow, graceful 360-degree orbit around a serene, white-bearded sage seated in perfect lotus pose beneath a majestic, stylized tree. The sage, his long, flowing white beard resting on his chest, his eyes gently closed, lost in deep meditation, maintaining absolute stillness. His body is clad in a simple off-white silk robe. The tree above him is a fantastical creation, with a gnarled, pearly trunk gracefully winding toward a crown of crystal leaves that shimmer with a soft inner glow. The background is a dreamy, soft-focus, pastel-colored cloudscape. As the camera glides, a point of light appears on the glowing mossy ground next to the sage, sprouting a new life cycle – a magical plant majestically accelerating its life cycle. The stem spirals upward, unfurling its leaves to reveal an intricate, bioluminescent, whimsical pattern. The flowers then blossom into a vibrant spectacle of unfolding fractals and incredible colors, releasing soft pulses of light. Simultaneously, these same intricate patterns begin to flow and echo across the surface of the saint's robes and ripple across the dreamlike background, visually weaving the saint, the plants, and the universe into a single, interconnected tapestry of quiet magic and unfolding wonder.", + "seed": 20848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_12/index_6274_seed_20848_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016263342276215554, + "vlm_logprob": -0.016263342276215554, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.016263342276215554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038241841830313206 + ], + "opposite_logprob_tokens": [ + -5.753824234008789 + ], + "logprob_reward": -0.0038241841830313206, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0038241841830313206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08703900873661041, + 0.03947719186544418, + 0.06038152426481247, + 0.04282673820853233, + 0.24114759266376495, + 0.3643623888492584, + 0.03787463903427124, + 0.04410586506128311, + 0.04878329113125801, + 0.03400171548128128 + ], + "credit_map": [ + 0.010560651309788227, + 0.017070898786187172, + 0.008030627854168415, + 0.0072869122959673405, + 0.01849265582859516, + 0.010585800744593143, + 0.008009248413145542, + 0.0070022158324718475, + 0.004898007959127426, + 0.007450707256793976, + 0.005642938427627087, + 0.0035714777186512947, + 0.004927558358758688, + 0.005175071768462658, + 0.004194710869342089, + 0.003616721136495471, + 0.009128396399319172, + 0.007961663417518139, + 0.004886300768703222, + 0.004669711459428072, + 0.010213291272521019, + 0.010541536845266819, + 0.008278103545308113, + 0.004702525679022074, + 0.011383702047169209, + 0.00841091014444828, + 0.0037810644134879112, + 0.005918475799262524, + 0.003630231600254774, + 0.00423771608620882, + 0.002692807698622346, + 0.0027718308847397566, + 0.03244612365961075, + 0.05947773903608322, + 0.029350584372878075, + 0.024289634078741074, + 0.027701064944267273, + 0.028623705729842186, + 0.022685430943965912, + 0.016573328524827957, + 0.05643351003527641, + 0.029848188161849976, + 0.05011521279811859, + 0.02736596018075943, + 0.06410147994756699, + 0.04466314986348152, + 0.06258785724639893, + 0.029247049242258072, + 0.0034355849493294954, + 0.00360689265653491, + 0.006194138899445534, + 0.00390238082036376, + 0.005808796733617783, + 0.004135590512305498, + 0.005904983729124069, + 0.0048862723633646965, + 0.004172580782324076, + 0.0032053072936832905, + 0.010151490569114685, + 0.0028611160814762115, + 0.007434952538460493, + 0.0052602048963308334, + 0.005700777750462294, + 0.005319437477737665, + 0.006427946966141462, + 0.005061504431068897, + 0.005242713261395693, + 0.005186463240534067, + 0.008323277346789837, + 0.011196262203156948, + 0.0033608509693294764, + 0.0039842757396399975, + 0.0032806850504130125, + 0.0039191944524645805, + 0.0040410365909338, + 0.003245923202484846, + 0.006284999195486307, + 0.004230897407978773, + 0.0040868306532502174, + 0.004912153817713261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0038241841830313206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979542873799801 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.02979542873799801, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979542873799801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11473319679498672, + 0.043524302542209625, + 0.09575117379426956, + 0.13084813952445984, + 0.12766937911510468, + 0.1785195916891098, + 0.11370429396629333, + 0.08120384067296982, + 0.055056989192962646, + 0.05898907780647278 + ], + "credit_map": [ + 0.01632661744952202, + 0.01846734806895256, + 0.010387216694653034, + 0.007626707199960947, + 0.01981223002076149, + 0.017928125336766243, + 0.013561450876295567, + 0.010623497888445854, + 0.005914059467613697, + 0.012506692670285702, + 0.005106097087264061, + 0.0024986860807985067, + 0.00792025588452816, + 0.005110232159495354, + 0.00237932987511158, + 0.002088950015604496, + 0.02017219364643097, + 0.012330766767263412, + 0.009237566962838173, + 0.005841680336743593, + 0.010330794379115105, + 0.024723289534449577, + 0.008374630473554134, + 0.004740255884826183, + 0.03705083951354027, + 0.045956529676914215, + 0.005975072272121906, + 0.00692058727145195, + 0.009157577529549599, + 0.01618329994380474, + 0.005540949292480946, + 0.004063278902322054, + 0.0160310547798872, + 0.016563022509217262, + 0.012175860814750195, + 0.014940591529011726, + 0.02737618423998356, + 0.012187805958092213, + 0.018855182453989983, + 0.009539680555462837, + 0.026873374357819557, + 0.016469024121761322, + 0.025476032868027687, + 0.013097314164042473, + 0.033691782504320145, + 0.021882081404328346, + 0.02631145529448986, + 0.014718541875481606, + 0.017384426668286324, + 0.010747954249382019, + 0.019915083423256874, + 0.00934569537639618, + 0.017784669995307922, + 0.012085026130080223, + 0.011823012493550777, + 0.014618417248129845, + 0.012357743456959724, + 0.010717959143221378, + 0.01279288437217474, + 0.008898148313164711, + 0.00993396807461977, + 0.008907345123589039, + 0.008664452470839024, + 0.008931341581046581, + 0.00564243970438838, + 0.0044697863049805164, + 0.006662249099463224, + 0.004632842727005482, + 0.012349211610853672, + 0.0073914313688874245, + 0.007600766606628895, + 0.006308262702077627, + 0.006201210431754589, + 0.006699864752590656, + 0.010499387979507446, + 0.005755801219493151, + 0.006782821845263243, + 0.005896942224353552, + 0.006718123331665993, + 0.01043492741882801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.02979542873799801 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3430631458759308, + 0.07348224520683289, + 0.049285225570201874, + 0.07799596339464188, + 0.04808470606803894, + 0.10299621522426605, + 0.05004484951496124, + 0.09189785271883011, + 0.10692549496889114, + 0.05622434243559837 + ], + "credit_map": [ + 0.044270921498537064, + 0.08264991641044617, + 0.02695087157189846, + 0.016659725457429886, + 0.07132333517074585, + 0.046639055013656616, + 0.03859082609415054, + 0.015978489071130753, + 0.010246390476822853, + 0.01693277433514595, + 0.010078256949782372, + 0.00621743593364954, + 0.012414112687110901, + 0.007514013443142176, + 0.005538228899240494, + 0.004541025497019291, + 0.011109685525298119, + 0.007738914340734482, + 0.002803252311423421, + 0.002760775852948427, + 0.005993509199470282, + 0.013464164920151234, + 0.00303929322399199, + 0.0023756250739097595, + 0.011983375996351242, + 0.018420472741127014, + 0.009660265408456326, + 0.010814518667757511, + 0.006222899537533522, + 0.012574238702654839, + 0.0044126734137535095, + 0.003907514736056328, + 0.007252446375787258, + 0.009220467880368233, + 0.004726020153611898, + 0.005799740552902222, + 0.00646144337952137, + 0.004724674392491579, + 0.005575893446803093, + 0.004324018023908138, + 0.01243600994348526, + 0.015341375023126602, + 0.010272908955812454, + 0.011929582804441452, + 0.020466919988393784, + 0.01433328352868557, + 0.011868776753544807, + 0.00634734844788909, + 0.006592511665076017, + 0.0049986462108790874, + 0.007711068727076054, + 0.006428151857107878, + 0.006688500754535198, + 0.0051838611252605915, + 0.007180975750088692, + 0.005261124577373266, + 0.006990231107920408, + 0.006459335330873728, + 0.02216494083404541, + 0.01025471743196249, + 0.014297446236014366, + 0.011355459690093994, + 0.012046866118907928, + 0.00832885131239891, + 0.0128520168364048, + 0.011453085578978062, + 0.006468096747994423, + 0.0076252431608736515, + 0.025981007143855095, + 0.03009025566279888, + 0.005090602673590183, + 0.007365175988525152, + 0.006214400753378868, + 0.006492774933576584, + 0.007706745527684689, + 0.005956399254500866, + 0.0059898304753005505, + 0.007711068727076054, + 0.006608102936297655, + 0.009545018896460533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.00361420470289886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430426187813282 + ], + "opposite_logprob_tokens": [ + -3.1680426597595215 + ], + "logprob_reward": -0.0430426187813282, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430426187813282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12785863876342773, + 0.10742345452308655, + 0.05579974129796028, + 0.09729228913784027, + 0.07931889593601227, + 0.14504699409008026, + 0.09623080492019653, + 0.06976312398910522, + 0.1109975129365921, + 0.11026851832866669 + ], + "credit_map": [ + 0.01468874141573906, + 0.023127174004912376, + 0.013067428022623062, + 0.010021132417023182, + 0.01602516509592533, + 0.022377971559762955, + 0.011817527003586292, + 0.016733486205339432, + 0.01467970386147499, + 0.024542881175875664, + 0.014818414114415646, + 0.010925259441137314, + 0.014223192818462849, + 0.01229450386017561, + 0.008867873810231686, + 0.0070716203190386295, + 0.009678397327661514, + 0.010587054304778576, + 0.005499900318682194, + 0.0032843106891959906, + 0.0074242521077394485, + 0.011394781991839409, + 0.004667594563215971, + 0.0032634532544761896, + 0.022563433274626732, + 0.027320297434926033, + 0.007489289157092571, + 0.010615620762109756, + 0.006923826411366463, + 0.010686810128390789, + 0.006752934772521257, + 0.004940076265484095, + 0.01228471752256155, + 0.019399326294660568, + 0.00931786373257637, + 0.006687245797365904, + 0.010923754423856735, + 0.008826008066534996, + 0.006247911136597395, + 0.005632065236568451, + 0.02222074568271637, + 0.014643524773418903, + 0.017716117203235626, + 0.011125503107905388, + 0.02989477477967739, + 0.018914364278316498, + 0.02086157165467739, + 0.009670390747487545, + 0.013940634205937386, + 0.011696379631757736, + 0.013127563521265984, + 0.008215218782424927, + 0.014788228087127209, + 0.01007895264774561, + 0.012476607225835323, + 0.011907217092812061, + 0.006269785575568676, + 0.005138827953487635, + 0.018915856257081032, + 0.004631279967725277, + 0.010564813390374184, + 0.009104019962251186, + 0.008547978475689888, + 0.006590561475604773, + 0.016076991334557533, + 0.013308838941156864, + 0.009610326029360294, + 0.009428082033991814, + 0.021837009117007256, + 0.027152489870786667, + 0.0055792988277971745, + 0.008004485629498959, + 0.015436604619026184, + 0.01552504301071167, + 0.016914546489715576, + 0.009662563912570477, + 0.013316449709236622, + 0.012777836993336678, + 0.013336303643882275, + 0.013299169950187206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0430426187813282 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010402749758213758 + ], + "opposite_logprob_tokens": [ + -6.876040458679199 + ], + "logprob_reward": -0.0010402749758213758, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010402749758213758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12188887596130371, + 0.05488956719636917, + 0.05433519557118416, + 0.11701875180006027, + 0.06008273363113403, + 0.07998684048652649, + 0.1246856078505516, + 0.06916886568069458, + 0.2523952126502991, + 0.0655483528971672 + ], + "credit_map": [ + 0.01820356212556362, + 0.03292824327945709, + 0.011568611487746239, + 0.007966713048517704, + 0.022297464311122894, + 0.012184029445052147, + 0.009845880791544914, + 0.006894371472299099, + 0.006172935012727976, + 0.010598634369671345, + 0.006453040521591902, + 0.005142356269061565, + 0.008228377439081669, + 0.007277759723365307, + 0.0063536930829286575, + 0.004662771243602037, + 0.007983485236763954, + 0.009244603104889393, + 0.007427316624671221, + 0.00431043840944767, + 0.007005384657531977, + 0.009038028307259083, + 0.0050201681442558765, + 0.004305767826735973, + 0.033101264387369156, + 0.034540776163339615, + 0.008813964203000069, + 0.01396175567060709, + 0.005745474249124527, + 0.01044255867600441, + 0.005254315212368965, + 0.005158647429198027, + 0.00791581254452467, + 0.0107917794957757, + 0.007761870510876179, + 0.00615451205521822, + 0.009008900262415409, + 0.007659904193133116, + 0.006424517370760441, + 0.004365434870123863, + 0.011363118886947632, + 0.007716383319348097, + 0.010899043641984463, + 0.005972504150122404, + 0.017028765752911568, + 0.010804768651723862, + 0.010357527062296867, + 0.0058447327464818954, + 0.023050790652632713, + 0.015507441945374012, + 0.01940040849149227, + 0.009647169150412083, + 0.016455715522170067, + 0.012424956075847149, + 0.012418795377016068, + 0.015780335292220116, + 0.0097993528470397, + 0.0066034672781825066, + 0.010168210603296757, + 0.009160266257822514, + 0.011477210558950901, + 0.008356642909348011, + 0.006700003053992987, + 0.006903714034706354, + 0.03441346436738968, + 0.026625806465744972, + 0.019318073987960815, + 0.02119833417236805, + 0.05284667760133743, + 0.06958569586277008, + 0.010042274370789528, + 0.018364880234003067, + 0.008038345724344254, + 0.009233066812157631, + 0.010587912984192371, + 0.00820947252213955, + 0.007773981429636478, + 0.0074067553505301476, + 0.0067045786418020725, + 0.007594239432364702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.0010402749758213758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one uninterrupted shot, the camera, at eye level, performs a slow, graceful 360-degree orbit around a serene, white-bearded sage seated in perfect lotus pose beneath a majestic, stylized tree. The sage, his long, flowing white beard resting on his chest, his eyes gently closed, lost in deep meditation, maintaining absolute stillness. His body is clad in a simple off-white silk robe. The tree above him is a fantastical creation, with a gnarled, pearly trunk gracefully winding toward a crown of crystal leaves that shimmer with a soft inner glow. The background is a dreamy, soft-focus, pastel-colored cloudscape. As the camera glides, a point of light appears on the glowing mossy ground next to the sage, sprouting a new life cycle – a magical plant majestically accelerating its life cycle. The stem spirals upward, unfurling its leaves to reveal an intricate, bioluminescent, whimsical pattern. The flowers then blossom into a vibrant spectacle of unfolding fractals and incredible colors, releasing soft pulses of light. Simultaneously, these same intricate patterns begin to flow and echo across the surface of the saint's robes and ripple across the dreamlike background, visually weaving the saint, the plants, and the universe into a single, interconnected tapestry of quiet magic and unfolding wonder.", + "seed": 20849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_12/index_6274_seed_20849_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028188627399504184, + "vlm_logprob": -0.028188627399504184, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.028188627399504184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0075566633604466915 + ], + "opposite_logprob_tokens": [ + -5.007556438446045 + ], + "logprob_reward": -0.0075566633604466915, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0075566633604466915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06960249692201614, + 0.018275631591677666, + 0.12433165311813354, + 0.060117416083812714, + 0.04673853889107704, + 0.08931304514408112, + 0.03372245281934738, + 0.487724632024765, + 0.034776169806718826, + 0.035397980362176895 + ], + "credit_map": [ + 0.011110090650618076, + 0.01542132068425417, + 0.006641607731580734, + 0.007054613903164864, + 0.009998923167586327, + 0.005270312074571848, + 0.006392664276063442, + 0.007712968159466982, + 0.0024085156619548798, + 0.0022841105237603188, + 0.0017750681145116687, + 0.0032190007623285055, + 0.0023318170569837093, + 0.002108815824612975, + 0.0020317647140473127, + 0.0021165378857403994, + 0.018488455563783646, + 0.01676333136856556, + 0.020335186272859573, + 0.01459461823105812, + 0.009959875606000423, + 0.015391716733574867, + 0.015162095427513123, + 0.013636376708745956, + 0.010395267978310585, + 0.00869028177112341, + 0.005876943469047546, + 0.008730818517506123, + 0.005799428094178438, + 0.008498979732394218, + 0.0046694837510585785, + 0.007456216961145401, + 0.008150757290422916, + 0.006652384530752897, + 0.004645133391022682, + 0.008466527797281742, + 0.0036494715604931116, + 0.005751758813858032, + 0.005681787617504597, + 0.0037407209165394306, + 0.015616951510310173, + 0.031199974939227104, + 0.004815035965293646, + 0.003129025688394904, + 0.004223572555929422, + 0.023717528209090233, + 0.002491590566933155, + 0.004119364079087973, + 0.004165466874837875, + 0.0035412812139838934, + 0.003953271545469761, + 0.003919582813978195, + 0.004955926910042763, + 0.0052420878782868385, + 0.0036895570810884237, + 0.004255280364304781, + 0.05788816139101982, + 0.015540688298642635, + 0.15758997201919556, + 0.020687734708189964, + 0.022192278876900673, + 0.04606933146715164, + 0.015066205523908138, + 0.15269024670124054, + 0.003329597180709243, + 0.0031936021987348795, + 0.003927184734493494, + 0.0026484664995223284, + 0.0036335308104753494, + 0.009649084880948067, + 0.003699717577546835, + 0.0046949866227805614, + 0.006011555436998606, + 0.003847519401460886, + 0.006256215739995241, + 0.002115782583132386, + 0.0035390970297157764, + 0.004819145426154137, + 0.0030042841099202633, + 0.00580438319593668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.0075566633604466915 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054950349032878876 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.054950349032878876, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054950349032878876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04075615108013153, + 0.08427851647138596, + 0.13903416693210602, + 0.035703353583812714, + 0.05095114931464195, + 0.019549375399947166, + 0.02228575199842453, + 0.49698251485824585, + 0.06955534964799881, + 0.04090368002653122 + ], + "credit_map": [ + 0.00614976417273283, + 0.0046989209949970245, + 0.004991372115910053, + 0.005688142031431198, + 0.0034665223211050034, + 0.003982611931860447, + 0.005910170264542103, + 0.005868648644536734, + 0.013984454795718193, + 0.013823187910020351, + 0.008087166585028172, + 0.014343444257974625, + 0.008514344692230225, + 0.007218736223876476, + 0.009517127647995949, + 0.008790053427219391, + 0.026710858568549156, + 0.01780412159860134, + 0.01928793080151081, + 0.014871791936457157, + 0.01470237597823143, + 0.014491877518594265, + 0.013199912384152412, + 0.017965296283364296, + 0.006635948549956083, + 0.005445733666419983, + 0.0046818433329463005, + 0.004265956114977598, + 0.0032111306209117174, + 0.0038158788811415434, + 0.003205458400771022, + 0.004441403318196535, + 0.007563774939626455, + 0.007217139936983585, + 0.003991293255239725, + 0.005349243991076946, + 0.007647516205906868, + 0.009411660954356194, + 0.004586270544677973, + 0.005184248089790344, + 0.0025589659344404936, + 0.003348183585330844, + 0.001605082186870277, + 0.002034995472058654, + 0.002078820951282978, + 0.00279554002918303, + 0.001840714132413268, + 0.0032870734576135874, + 0.0032084970735013485, + 0.0023329013492912054, + 0.002521221060305834, + 0.002756064059212804, + 0.0030084133613854647, + 0.0024169026874005795, + 0.0026915562339127064, + 0.0033501966390758753, + 0.06031150743365288, + 0.016330622136592865, + 0.16276738047599792, + 0.02091117389500141, + 0.023989621549844742, + 0.04528381675481796, + 0.014576668851077557, + 0.15281173586845398, + 0.008114088326692581, + 0.006020760629326105, + 0.009454323910176754, + 0.0069502596743404865, + 0.007539812009781599, + 0.013516179285943508, + 0.009018134325742722, + 0.008941792882978916, + 0.005903163924813271, + 0.006035827100276947, + 0.008386529050767422, + 0.003583177225664258, + 0.004210787359625101, + 0.004774884320795536, + 0.0036450845655053854, + 0.004364222753793001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.054950349032878876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2063482403755188, + 0.06726576387882233, + 0.06947997957468033, + 0.03730715811252594, + 0.062439318746328354, + 0.12631672620773315, + 0.05033884570002556, + 0.17740105092525482, + 0.0766603946685791, + 0.12644244730472565 + ], + "credit_map": [ + 0.0335170142352581, + 0.03887953236699104, + 0.023398982360959053, + 0.024654744192957878, + 0.026108751073479652, + 0.01944258250296116, + 0.018911344930529594, + 0.021435314789414406, + 0.009653419256210327, + 0.00924866646528244, + 0.00772064970806241, + 0.008982717990875244, + 0.008763388730585575, + 0.006049495656043291, + 0.007981582544744015, + 0.008865847252309322, + 0.014215312898159027, + 0.012088531628251076, + 0.007424550596624613, + 0.005773133132606745, + 0.00540165277197957, + 0.009026117622852325, + 0.005449559073895216, + 0.010101133026182652, + 0.005762750748544931, + 0.005080993287265301, + 0.0054800622165203094, + 0.004793769679963589, + 0.003606203943490982, + 0.003473761724308133, + 0.004856184124946594, + 0.004253435414284468, + 0.007893611676990986, + 0.00858191680163145, + 0.004017168190330267, + 0.008504647761583328, + 0.010588722303509712, + 0.010810699313879013, + 0.005819196812808514, + 0.0062233638018369675, + 0.02098797634243965, + 0.039315659552812576, + 0.013195418752729893, + 0.006936271209269762, + 0.005910538136959076, + 0.03039488010108471, + 0.0036700458731502295, + 0.005905942991375923, + 0.008550968952476978, + 0.005662671755999327, + 0.007584244478493929, + 0.004272123333066702, + 0.004372376948595047, + 0.009296891279518604, + 0.004483423195779324, + 0.006116149015724659, + 0.022487075999379158, + 0.01723632961511612, + 0.0483010895550251, + 0.009803120978176594, + 0.009334924630820751, + 0.018337639048695564, + 0.006611015181988478, + 0.04528987407684326, + 0.013692506588995457, + 0.006084111053496599, + 0.01674322411417961, + 0.0062497640028595924, + 0.006032553501427174, + 0.011060092598199844, + 0.005848363041877747, + 0.010949783027172089, + 0.020641634240746498, + 0.016129640862345695, + 0.02428615093231201, + 0.010181057266891003, + 0.016565147787332535, + 0.014209162443876266, + 0.01288916077464819, + 0.011540503241121769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.005947155877947807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000225782394409 + ], + "opposite_logprob_tokens": [ + -2.695002317428589 + ], + "logprob_reward": -0.07000225782394409, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000225782394409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1589396446943283, + 0.1322164088487625, + 0.13520771265029907, + 0.060270681977272034, + 0.06434277445077896, + 0.2741350829601288, + 0.029397349804639816, + 0.06038830801844597, + 0.028786469250917435, + 0.056315578520298004 + ], + "credit_map": [ + 0.025703812018036842, + 0.03220399469137192, + 0.015501156449317932, + 0.0193808451294899, + 0.021144067868590355, + 0.013552107848227024, + 0.01750677265226841, + 0.013946885243058205, + 0.022363439202308655, + 0.02128748409450054, + 0.013276795856654644, + 0.020850177854299545, + 0.013631287030875683, + 0.008977826684713364, + 0.015119625255465508, + 0.016709759831428528, + 0.026883570477366447, + 0.02338259480893612, + 0.021502790972590446, + 0.011889156885445118, + 0.009810482151806355, + 0.012239723466336727, + 0.0120385205373168, + 0.017460862174630165, + 0.008274560794234276, + 0.008640374056994915, + 0.010270228609442711, + 0.0069725546054542065, + 0.0052514225244522095, + 0.00789554137736559, + 0.004624723456799984, + 0.008341269567608833, + 0.007701958529651165, + 0.008655058220028877, + 0.005641718860715628, + 0.00708378292620182, + 0.009801643900573254, + 0.011393888853490353, + 0.006154962815344334, + 0.007909753359854221, + 0.048138249665498734, + 0.09303142130374908, + 0.025680599734187126, + 0.01414014957845211, + 0.014003254473209381, + 0.058103013783693314, + 0.0075634559616446495, + 0.013474907726049423, + 0.005331403110176325, + 0.003377289976924658, + 0.004639810416847467, + 0.004306779708713293, + 0.003134411061182618, + 0.002979113021865487, + 0.0023342743515968323, + 0.0032942681573331356, + 0.014290404506027699, + 0.007946355268359184, + 0.00714739877730608, + 0.0034495899453759193, + 0.004829914774745703, + 0.011022395454347134, + 0.0034485391806811094, + 0.008253706619143486, + 0.003816444892436266, + 0.0024162218905985355, + 0.0034988783299922943, + 0.0028857672587037086, + 0.003444978967308998, + 0.004894379526376724, + 0.003999915439635515, + 0.0038298810832202435, + 0.007103148847818375, + 0.00894960667937994, + 0.006074097473174334, + 0.004349823575466871, + 0.01002442929893732, + 0.007531794253736734, + 0.006293139886111021, + 0.005989537574350834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.07000225782394409 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04091867431998253, + 0.04583226889371872, + 0.07164780795574188, + 0.04226003214716911, + 0.05578300729393959, + 0.17360395193099976, + 0.015990499407052994, + 0.5032828450202942, + 0.022193951532244682, + 0.02848697640001774 + ], + "credit_map": [ + 0.006365431938320398, + 0.0072728730738162994, + 0.0037047378718852997, + 0.005479220300912857, + 0.0059637390077114105, + 0.004580595530569553, + 0.0034382652956992388, + 0.004113811068236828, + 0.007139586377888918, + 0.005955526605248451, + 0.004981453530490398, + 0.00684344582259655, + 0.00517707085236907, + 0.0038341726176440716, + 0.005890815053135157, + 0.006010197103023529, + 0.016973264515399933, + 0.016561325639486313, + 0.010491584427654743, + 0.0046823215670883656, + 0.004856368061155081, + 0.006004016846418381, + 0.003990397788584232, + 0.008088530041277409, + 0.006973609793931246, + 0.006358683574944735, + 0.008279313333332539, + 0.004094030242413282, + 0.0032187453471124172, + 0.0048058233223855495, + 0.0031396562699228525, + 0.005390169098973274, + 0.011629958637058735, + 0.00791851058602333, + 0.005653079133480787, + 0.008641554042696953, + 0.005682741291821003, + 0.0058513483963906765, + 0.005357592832297087, + 0.005048222374171019, + 0.03028019890189171, + 0.06034514680504799, + 0.012317628599703312, + 0.006587334908545017, + 0.008489838801324368, + 0.04331029951572418, + 0.004555229097604752, + 0.007718280889093876, + 0.0032455555628985167, + 0.002182801254093647, + 0.0021483225282281637, + 0.0015995055437088013, + 0.0014679731102660298, + 0.0018919284921139479, + 0.0015873450320214033, + 0.0018670682329684496, + 0.06316595524549484, + 0.01607181876897812, + 0.16545860469341278, + 0.020665500313043594, + 0.021420836448669434, + 0.04477669671177864, + 0.013185686431825161, + 0.15853773057460785, + 0.004600198473781347, + 0.002230957383289933, + 0.004825318697839975, + 0.0017246284987777472, + 0.0012319302186369896, + 0.0026626153849065304, + 0.0012594180880114436, + 0.0036588869988918304, + 0.0032496917992830276, + 0.0029208471532911062, + 0.002915583783760667, + 0.002253176411613822, + 0.003385223215445876, + 0.003775122808292508, + 0.003682300914078951, + 0.006305030547082424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.0024867109023034573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one uninterrupted shot, the camera, at eye level, performs a slow, graceful 360-degree orbit around a serene, white-bearded sage seated in perfect lotus pose beneath a majestic, stylized tree. The sage, his long, flowing white beard resting on his chest, his eyes gently closed, lost in deep meditation, maintaining absolute stillness. His body is clad in a simple off-white silk robe. The tree above him is a fantastical creation, with a gnarled, pearly trunk gracefully winding toward a crown of crystal leaves that shimmer with a soft inner glow. The background is a dreamy, soft-focus, pastel-colored cloudscape. As the camera glides, a point of light appears on the glowing mossy ground next to the sage, sprouting a new life cycle – a magical plant majestically accelerating its life cycle. The stem spirals upward, unfurling its leaves to reveal an intricate, bioluminescent, whimsical pattern. The flowers then blossom into a vibrant spectacle of unfolding fractals and incredible colors, releasing soft pulses of light. Simultaneously, these same intricate patterns begin to flow and echo across the surface of the saint's robes and ripple across the dreamlike background, visually weaving the saint, the plants, and the universe into a single, interconnected tapestry of quiet magic and unfolding wonder.", + "seed": 20850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_12/index_6274_seed_20850_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015165586885996164, + "vlm_logprob": -0.015165586885996164, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.015165586885996164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0016777737764641643 + ], + "opposite_logprob_tokens": [ + -6.876677989959717 + ], + "logprob_reward": -0.0016777737764641643, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0016777737764641643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18408071994781494, + 0.09726113080978394, + 0.08496177196502686, + 0.11321435123682022, + 0.0596454031765461, + 0.05526527017354965, + 0.05036575347185135, + 0.06329936534166336, + 0.07668013870716095, + 0.21522606909275055 + ], + "credit_map": [ + 0.023930910974740982, + 0.02347775176167488, + 0.016486426815390587, + 0.018728403374552727, + 0.024600619450211525, + 0.01603766344487667, + 0.03169391304254532, + 0.029125049710273743, + 0.017852460965514183, + 0.013237825594842434, + 0.009637612849473953, + 0.0067592039704322815, + 0.013204987160861492, + 0.012867162004113197, + 0.010722687467932701, + 0.012979201041162014, + 0.011395585723221302, + 0.014325115829706192, + 0.00901530496776104, + 0.008484477177262306, + 0.012326390482485294, + 0.00869104266166687, + 0.012177933938801289, + 0.008545923046767712, + 0.02310953103005886, + 0.01888672634959221, + 0.009444170631468296, + 0.00816916674375534, + 0.010056070983409882, + 0.02230091392993927, + 0.011458230204880238, + 0.009789551608264446, + 0.006741370540112257, + 0.0066150822676718235, + 0.007519436068832874, + 0.006490682251751423, + 0.009264107793569565, + 0.005389468278735876, + 0.011033546179533005, + 0.00659171724691987, + 0.006500819232314825, + 0.008108512498438358, + 0.006563345901668072, + 0.004338267724961042, + 0.008774282410740852, + 0.00927055161446333, + 0.007651926018297672, + 0.00405756663531065, + 0.0061898441053926945, + 0.007184762042015791, + 0.007721630856394768, + 0.005272306967526674, + 0.008179827593266964, + 0.005910297390073538, + 0.004884143825620413, + 0.005022944882512093, + 0.01231487188488245, + 0.008392415009438992, + 0.006559951230883598, + 0.006799164693802595, + 0.008776368573307991, + 0.006180856842547655, + 0.006713825277984142, + 0.00756191648542881, + 0.00991388875991106, + 0.010889227502048016, + 0.008656305260956287, + 0.007402552291750908, + 0.007907318882644176, + 0.009483512490987778, + 0.008919753134250641, + 0.01350758783519268, + 0.014107618480920792, + 0.017727447673678398, + 0.016051921993494034, + 0.05886261165142059, + 0.07097012549638748, + 0.020152650773525238, + 0.00957285612821579, + 0.007780860643833876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0016777737764641643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305209591984749 + ], + "opposite_logprob_tokens": [ + -3.1680521965026855 + ], + "logprob_reward": -0.04305209591984749, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305209591984749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2120041698217392, + 0.1544230431318283, + 0.09139756113290787, + 0.14273542165756226, + 0.08738864213228226, + 0.0376826748251915, + 0.1320682317018509, + 0.049653392285108566, + 0.04851517453789711, + 0.04413170367479324 + ], + "credit_map": [ + 0.026335401460528374, + 0.029985299333930016, + 0.019047264009714127, + 0.023916076868772507, + 0.028553569689393044, + 0.0164876077324152, + 0.03807562589645386, + 0.029603315517306328, + 0.02893819659948349, + 0.01817317306995392, + 0.013082809746265411, + 0.01046354416757822, + 0.02032843604683876, + 0.025446215644478798, + 0.0212763249874115, + 0.016714343801140785, + 0.010616007260978222, + 0.008264681324362755, + 0.018602175638079643, + 0.008654160425066948, + 0.009105218574404716, + 0.009810551069676876, + 0.009114738553762436, + 0.01723003387451172, + 0.016193410381674767, + 0.02093123272061348, + 0.012586048804223537, + 0.014552371576428413, + 0.021692629903554916, + 0.01743004098534584, + 0.02760929800570011, + 0.011740387417376041, + 0.01881490647792816, + 0.021174166351556778, + 0.005948440171778202, + 0.005181292071938515, + 0.011094930581748486, + 0.010105034336447716, + 0.008123224601149559, + 0.006946646608412266, + 0.005511011928319931, + 0.004879775922745466, + 0.0035448919516056776, + 0.004065696615725756, + 0.005405651871114969, + 0.005907313898205757, + 0.005567294545471668, + 0.002801039256155491, + 0.012865876778960228, + 0.017179789021611214, + 0.007934276014566422, + 0.017484014853835106, + 0.026045288890600204, + 0.015561413019895554, + 0.02278069406747818, + 0.012216882780194283, + 0.007867302745580673, + 0.006768242456018925, + 0.00382014038041234, + 0.005137986969202757, + 0.009229863062500954, + 0.005963774863630533, + 0.006569067016243935, + 0.00429701991379261, + 0.006697675213217735, + 0.0070329452864825726, + 0.006964711472392082, + 0.006877987179905176, + 0.0063413651660084724, + 0.005479416344314814, + 0.0037664633709937334, + 0.005354607477784157, + 0.00426227692514658, + 0.005608539097011089, + 0.004142153542488813, + 0.004916823003441095, + 0.011502466164529324, + 0.006981971208006144, + 0.0035015258472412825, + 0.003215949749574065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04305209591984749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636015277355909 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004636015277355909, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636015277355909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0688372552394867, + 0.05751057341694832, + 0.09353765100240707, + 0.12493199110031128, + 0.03368445113301277, + 0.055077917873859406, + 0.08801716566085815, + 0.054603323340415955, + 0.04507758840918541, + 0.37872204184532166 + ], + "credit_map": [ + 0.007707254029810429, + 0.00790492445230484, + 0.008309018798172474, + 0.006195097230374813, + 0.008057673461735249, + 0.00709141418337822, + 0.011401936411857605, + 0.01216993760317564, + 0.009388469159603119, + 0.010866577737033367, + 0.004428558517247438, + 0.0049509573727846146, + 0.009467882104218006, + 0.006104145664721727, + 0.00822371058166027, + 0.004080270882695913, + 0.014893073588609695, + 0.022813830524683, + 0.010004344396293163, + 0.00741787301376462, + 0.007008276879787445, + 0.013404018245637417, + 0.005177236162126064, + 0.012819004245102406, + 0.024546684697270393, + 0.030340751633048058, + 0.009784468449652195, + 0.008569550700485706, + 0.011310622096061707, + 0.025516269728541374, + 0.0075537157244980335, + 0.007309931330382824, + 0.00433771125972271, + 0.004302657209336758, + 0.0054330769926309586, + 0.002147634979337454, + 0.004501734860241413, + 0.0055696372874081135, + 0.00330218649469316, + 0.004089809488505125, + 0.007530171889811754, + 0.007949323393404484, + 0.006386719644069672, + 0.004271802958101034, + 0.010339675471186638, + 0.007929950952529907, + 0.006798031274229288, + 0.003872241359204054, + 0.009705479256808758, + 0.009949649684131145, + 0.01207125373184681, + 0.012143167667090893, + 0.015456314198672771, + 0.008006067015230656, + 0.011148774065077305, + 0.009536463767290115, + 0.009690456092357635, + 0.006311166100203991, + 0.006954167038202286, + 0.005270170047879219, + 0.00744109321385622, + 0.005109951365739107, + 0.004945095628499985, + 0.008881224319338799, + 0.006754626985639334, + 0.00665886327624321, + 0.006844763178378344, + 0.005723459646105766, + 0.0047334530390799046, + 0.004418006632477045, + 0.0046639516949653625, + 0.005280462559312582, + 0.025174831971526146, + 0.030484208837151527, + 0.03119797445833683, + 0.10938402265310287, + 0.12126615643501282, + 0.03479304164648056, + 0.015624835155904293, + 0.010797007009387016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.004636015277355909 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327328361570835 + ], + "opposite_logprob_tokens": [ + -3.773273229598999 + ], + "logprob_reward": -0.02327328361570835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327328361570835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12886333465576172, + 0.06531593948602676, + 0.12548194825649261, + 0.054659489542245865, + 0.08144985884428024, + 0.10855299234390259, + 0.14209383726119995, + 0.07391217350959778, + 0.09105512499809265, + 0.12861531972885132 + ], + "credit_map": [ + 0.014842369593679905, + 0.0145723232999444, + 0.01380860898643732, + 0.013043535873293877, + 0.013055805116891861, + 0.011782967485487461, + 0.02073040045797825, + 0.027027325704693794, + 0.008990936912596226, + 0.009059335105121136, + 0.007156981620937586, + 0.007348309271037579, + 0.010271049104630947, + 0.005420435220003128, + 0.010610093362629414, + 0.0064588007517158985, + 0.020380156114697456, + 0.0354243740439415, + 0.009118213318288326, + 0.01031504012644291, + 0.010520887561142445, + 0.018800079822540283, + 0.011813648976385593, + 0.00910954549908638, + 0.00716191902756691, + 0.009362167678773403, + 0.004341410472989082, + 0.006537070497870445, + 0.0076585677452385426, + 0.00597369484603405, + 0.00823832955211401, + 0.005386330187320709, + 0.011137224733829498, + 0.009814614430069923, + 0.0079340236261487, + 0.006846620235592127, + 0.010417704470455647, + 0.013541381806135178, + 0.009458796121180058, + 0.012299495749175549, + 0.01606915146112442, + 0.013835239224135876, + 0.013776532374322414, + 0.009827064350247383, + 0.021108528599143028, + 0.012280110269784927, + 0.010963731445372105, + 0.010692629963159561, + 0.0159052275121212, + 0.01956726238131523, + 0.015465311706066132, + 0.022590549662709236, + 0.017407741397619247, + 0.015331492759287357, + 0.020262930542230606, + 0.015563322231173515, + 0.014564924873411655, + 0.010382076725363731, + 0.007845781743526459, + 0.006272867321968079, + 0.011804385110735893, + 0.008061385713517666, + 0.00686355447396636, + 0.00811719335615635, + 0.010510641150176525, + 0.01313457079231739, + 0.007309409789741039, + 0.011116601526737213, + 0.01716422848403454, + 0.009713275358080864, + 0.012373900040984154, + 0.009732497856020927, + 0.009180048480629921, + 0.011913226917386055, + 0.011186419054865837, + 0.030175350606441498, + 0.03952556848526001, + 0.01291324757039547, + 0.007847820408642292, + 0.005873627029359341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.02327328361570835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a white-bearded sage seated in a lotus pose visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sage wearing a simple off-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fantastical tree with a pearly trunk and crystal leaves glowing above the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbit around the sage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a magical plant sprout from the ground and bloom with bioluminescent fractal patterns that ripple onto the sage's robes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031887658406049013 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031887658406049013, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031887658406049013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04952934756875038, + 0.038318533450365067, + 0.1148269847035408, + 0.022157475352287292, + 0.02566341869533062, + 0.022303037345409393, + 0.0617724284529686, + 0.043572038412094116, + 0.03306034207344055, + 0.5887963175773621 + ], + "credit_map": [ + 0.007102228235453367, + 0.005807096604257822, + 0.006387590896338224, + 0.0036000139079988003, + 0.005971874576061964, + 0.0047237565740942955, + 0.007463875692337751, + 0.008472914807498455, + 0.009208085015416145, + 0.0034057593438774347, + 0.003143120324239135, + 0.002254847204312682, + 0.003815352451056242, + 0.008888043463230133, + 0.0036153022665530443, + 0.003988024313002825, + 0.016819480806589127, + 0.029241859912872314, + 0.013807760551571846, + 0.007729416247457266, + 0.008086991496384144, + 0.018879801034927368, + 0.005664173513650894, + 0.014597499743103981, + 0.0037186513654887676, + 0.0037309846375137568, + 0.002548947697505355, + 0.0026334584690630436, + 0.0024967051576822996, + 0.0026480627711862326, + 0.00206890725530684, + 0.002311757765710354, + 0.0037226341664791107, + 0.003635168308392167, + 0.0023244014009833336, + 0.0022569154389202595, + 0.004211677238345146, + 0.003490869188681245, + 0.004455987364053726, + 0.0015657648909837008, + 0.002746234880760312, + 0.0026421600487083197, + 0.003371619153767824, + 0.0023838584311306477, + 0.002964933402836323, + 0.0027470437344163656, + 0.002488747937604785, + 0.002958441385999322, + 0.007607132662087679, + 0.007563825696706772, + 0.009310366585850716, + 0.00714788818731904, + 0.008792661130428314, + 0.005776097998023033, + 0.00801567267626524, + 0.007558783516287804, + 0.010668179020285606, + 0.005260690115392208, + 0.006052343174815178, + 0.003909138031303883, + 0.0036789707373827696, + 0.004054839722812176, + 0.004115454386919737, + 0.0058324262499809265, + 0.003793673124164343, + 0.004452963825315237, + 0.006383085157722235, + 0.002804219489917159, + 0.003859582357108593, + 0.002865975257009268, + 0.0038417624309659004, + 0.0050590792670845985, + 0.0379631444811821, + 0.049993276596069336, + 0.0479753352701664, + 0.15940937399864197, + 0.19254495203495026, + 0.05587654560804367, + 0.026747703552246094, + 0.01828598603606224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.0031887658406049013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_tensor([9568, 9568])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_tensor([9568, 9568])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7800239a48ee4c31c17cf6a02d0a4ae006ddbedc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_tensor([9568, 9568])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, shot from a slightly upward angle, the camera slowly panning forward, accompanied by a gentle, irregular sway simulated by wind, captures the breathtaking spectacle of an ancient chain bridge swaying violently over a bottomless chasm. The sky is shrouded in thick, swirling clouds, as dark as ink. Pale, forked lightning occasionally rips through the canopy, each flash momentarily illuminating the bridge's thick chains, coated in dark brown grains of rust, and its slippery, dark wood, soaked by the torrential rain and reflecting a cold, high-gloss light. Under the raging wind, the entire bridge groans under the weight, emitting a harsh sound of metal rubbing and wood squeezing. It twists and throbs in agony like a dying python, and through the gaps in the planks, one can glimpse the endless darkness below, swallowed by the thick fog. This scene was built with the extremely high standards of the UE engine, presenting a film-like texture. Whether it is the traces of rain flowing on the iron chain or the wet texture of the wooden board, it has achieved a realistic level of detail. At the same time, a layer of subtle film grain is covered on the picture, which not only enhances the visual depth, but also creates a tense, epic and doomsday atmosphere of despair.", + "seed": 2885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_9568_seed_2885_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20286285942420362, + "vlm_logprob": -0.20286285942420362, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.20286285942420362, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07245365530252457 + ], + "opposite_logprob_tokens": [ + -2.697453737258911 + ], + "logprob_reward": -0.07245365530252457, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07245365530252457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26322874426841736, + 0.026934480294585228, + 0.04093228280544281, + 0.03451916575431824, + 0.08111856132745743, + 0.06563159078359604, + 0.047767024487257004, + 0.28716257214546204, + 0.05261286348104477, + 0.10009272396564484 + ], + "credit_map": [ + 0.023021113127470016, + 0.05770323798060417, + 0.02352474257349968, + 0.024796999990940094, + 0.03781619295477867, + 0.03864411264657974, + 0.030911760404706, + 0.02681058831512928, + 0.0027483045123517513, + 0.0034332633949816227, + 0.0032076393254101276, + 0.003536767093464732, + 0.003547919448465109, + 0.004829582292586565, + 0.00261619221419096, + 0.003014810848981142, + 0.007881882600486279, + 0.008139991201460361, + 0.0023461682721972466, + 0.004252892918884754, + 0.003221163060516119, + 0.006817682180553675, + 0.004275538492947817, + 0.003996965009719133, + 0.0051489039324223995, + 0.009550596587359905, + 0.0023073323536664248, + 0.003911631181836128, + 0.0043871537782251835, + 0.0030240293126553297, + 0.0030394205823540688, + 0.0031501015182584524, + 0.008741309866309166, + 0.013750670477747917, + 0.007646517362445593, + 0.008646087720990181, + 0.015080118551850319, + 0.010552369058132172, + 0.00786775816231966, + 0.008833734318614006, + 0.005146299488842487, + 0.008172943256795406, + 0.007853468880057335, + 0.011677534319460392, + 0.007574478164315224, + 0.006259759422391653, + 0.011166412383317947, + 0.007780701853334904, + 0.0036601400934159756, + 0.005466302391141653, + 0.003967964090406895, + 0.005014484748244286, + 0.007004708517342806, + 0.00788260530680418, + 0.008192644454538822, + 0.006578176282346249, + 0.03995325416326523, + 0.02957357093691826, + 0.02722625620663166, + 0.03075418621301651, + 0.033964429050683975, + 0.051004961133003235, + 0.031068019568920135, + 0.04361788183450699, + 0.003931352868676186, + 0.006337106693536043, + 0.0072655826807022095, + 0.007255468983203173, + 0.007159724831581116, + 0.0071861292235553265, + 0.007616851013153791, + 0.005860650446265936, + 0.008877977728843689, + 0.009769313968718052, + 0.0066504329442977905, + 0.011542685329914093, + 0.017586536705493927, + 0.013921523466706276, + 0.020073076710104942, + 0.011671177111566067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.07245365530252457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983324766159058 + ], + "opposite_logprob_tokens": [ + -0.5233324766159058 + ], + "logprob_reward": -0.8983324766159058, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983324766159058, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0819341242313385, + 0.01923792064189911, + 0.028804654255509377, + 0.040779102593660355, + 0.33445656299591064, + 0.13053196668624878, + 0.04004203528165817, + 0.10280626267194748, + 0.20025387406349182, + 0.021153518930077553 + ], + "credit_map": [ + 0.008287915028631687, + 0.01223413273692131, + 0.008721391670405865, + 0.010165916755795479, + 0.009842307306826115, + 0.015394528396427631, + 0.0074049546383321285, + 0.009882980026304722, + 0.0019056707387790084, + 0.002258844207972288, + 0.0028512857388705015, + 0.002251621102914214, + 0.0023085481952875853, + 0.002905437257140875, + 0.0024170242249965668, + 0.0023394895251840353, + 0.0053701018914580345, + 0.006784446071833372, + 0.0018213634612038732, + 0.0026914409827440977, + 0.0027991551905870438, + 0.004186580423265696, + 0.0026734794955700636, + 0.002478085458278656, + 0.0037091777194291353, + 0.008237428963184357, + 0.0023610780481249094, + 0.00460019800812006, + 0.00978456623852253, + 0.002603191649541259, + 0.00462565990164876, + 0.004857804160565138, + 0.03736526891589165, + 0.06143265217542648, + 0.03514854982495308, + 0.03569384664297104, + 0.05373106524348259, + 0.03952128440141678, + 0.036621954292058945, + 0.034941934049129486, + 0.009556700475513935, + 0.015527709387242794, + 0.011237190105021, + 0.019444540143013, + 0.01806589402258396, + 0.01817682571709156, + 0.02284754067659378, + 0.015675581991672516, + 0.0032569547183811665, + 0.005218373145908117, + 0.0037944300565868616, + 0.004639579448848963, + 0.005094042979180813, + 0.006511713843792677, + 0.006423038896173239, + 0.005103901494294405, + 0.011235801503062248, + 0.011225832626223564, + 0.010436588898301125, + 0.01203997153788805, + 0.012298094108700752, + 0.015536810271441936, + 0.014185888692736626, + 0.01584726758301258, + 0.012053223326802254, + 0.02896462008357048, + 0.016978057101368904, + 0.03971071168780327, + 0.01827482134103775, + 0.01499553956091404, + 0.04096440225839615, + 0.02831251174211502, + 0.0025388244539499283, + 0.002922656713053584, + 0.001941448776051402, + 0.0029193393420428038, + 0.0029052929021418095, + 0.002475182991474867, + 0.003028692677617073, + 0.0024220808409154415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.8983324766159058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08308248221874237, + 0.022170089185237885, + 0.03375089168548584, + 0.06286510080099106, + 0.18280357122421265, + 0.23279203474521637, + 0.036983802914619446, + 0.06869213283061981, + 0.2389042228460312, + 0.0379556268453598 + ], + "credit_map": [ + 0.008096938952803612, + 0.015187554061412811, + 0.013396347872912884, + 0.008824922144412994, + 0.00956213753670454, + 0.011732554994523525, + 0.00969296507537365, + 0.006589072290807962, + 0.0028188840951770544, + 0.002941570710390806, + 0.0017919177189469337, + 0.0027938273269683123, + 0.0026466958224773407, + 0.00347116868942976, + 0.00229640188626945, + 0.003409623634070158, + 0.007611608598381281, + 0.007702235132455826, + 0.0022619166411459446, + 0.0034733726643025875, + 0.0025652721524238586, + 0.004347854759544134, + 0.002886570291593671, + 0.0029020633082836866, + 0.005887775681912899, + 0.011953919194638729, + 0.0034145531244575977, + 0.006732859183102846, + 0.0176960788667202, + 0.004337686579674482, + 0.006190774030983448, + 0.006651459261775017, + 0.01969187706708908, + 0.031933773308992386, + 0.019380006939172745, + 0.01853122189640999, + 0.03237488493323326, + 0.0213980320841074, + 0.019613156095147133, + 0.0198806244879961, + 0.018108703196048737, + 0.030668744817376137, + 0.021478312090039253, + 0.034630488604307175, + 0.037785038352012634, + 0.023256460204720497, + 0.03873841091990471, + 0.028125882148742676, + 0.0023318775929510593, + 0.003853181144222617, + 0.003933418542146683, + 0.003972123842686415, + 0.006094430573284626, + 0.006457608193159103, + 0.005707337986677885, + 0.004633825272321701, + 0.012143552303314209, + 0.00719043193385005, + 0.006199203431606293, + 0.007226155139505863, + 0.008341306820511818, + 0.010201983153820038, + 0.006399792619049549, + 0.010989709757268429, + 0.029198553413152695, + 0.041101180016994476, + 0.022833002731204033, + 0.028094716370105743, + 0.026072323322296143, + 0.023469988256692886, + 0.03597148507833481, + 0.03216299042105675, + 0.004099436104297638, + 0.003884477773681283, + 0.003651741659268737, + 0.003241546917706728, + 0.0034872416872531176, + 0.008965669199824333, + 0.004109126050025225, + 0.006516391411423683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.004639219027012587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347588747739792 + ], + "opposite_logprob_tokens": [ + -3.6513476371765137 + ], + "logprob_reward": -0.026347588747739792, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347588747739792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28417715430259705, + 0.04472418129444122, + 0.029356088489294052, + 0.03045891597867012, + 0.28501394391059875, + 0.09276473522186279, + 0.03056260384619236, + 0.09866629540920258, + 0.07455983012914658, + 0.02971627563238144 + ], + "credit_map": [ + 0.020577358081936836, + 0.048471786081790924, + 0.033024828881025314, + 0.027500253170728683, + 0.037700504064559937, + 0.043540507555007935, + 0.030890386551618576, + 0.04247152805328369, + 0.004569276701658964, + 0.007748155388981104, + 0.003638385096564889, + 0.0053946455009281635, + 0.00686779385432601, + 0.006913135293871164, + 0.00494472635909915, + 0.0046480633318424225, + 0.00308984168805182, + 0.003905939869582653, + 0.002977401716634631, + 0.0032604136504232883, + 0.003971520345658064, + 0.0036463080905377865, + 0.00456886924803257, + 0.0039357938803732395, + 0.0027190176770091057, + 0.005083940457552671, + 0.002096071606501937, + 0.00356978178024292, + 0.005365682765841484, + 0.003034907393157482, + 0.003939606249332428, + 0.0046499064192175865, + 0.03244595229625702, + 0.04276292771100998, + 0.03380905091762543, + 0.029693372547626495, + 0.04955119267106056, + 0.03370283544063568, + 0.031792622059583664, + 0.03125598281621933, + 0.006698395125567913, + 0.008016753941774368, + 0.008064024150371552, + 0.012171715497970581, + 0.008225618861615658, + 0.011855825781822205, + 0.019104428589344025, + 0.018627969548106194, + 0.00237456476315856, + 0.0036516983527690172, + 0.003200364764779806, + 0.003497153287753463, + 0.003720997367054224, + 0.005368756130337715, + 0.004688980523496866, + 0.004060088191181421, + 0.02211984619498253, + 0.010122486390173435, + 0.00829757284373045, + 0.009026866406202316, + 0.011475060135126114, + 0.011836715042591095, + 0.010781042277812958, + 0.01500670425593853, + 0.008491536602377892, + 0.013662803918123245, + 0.007909432984888554, + 0.011514105834066868, + 0.006757933646440506, + 0.006303549278527498, + 0.011112712323665619, + 0.008807752281427383, + 0.003062209114432335, + 0.003494785400107503, + 0.002944014733657241, + 0.0029380035120993853, + 0.0026488068979233503, + 0.0064426627941429615, + 0.003112080041319132, + 0.0050737131386995316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.026347588747739792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541357427835464 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541357427835464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541357427835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09375602006912231, + 0.024867843836545944, + 0.023501792922616005, + 0.06526694446802139, + 0.23708851635456085, + 0.13996559381484985, + 0.03712468594312668, + 0.13217388093471527, + 0.11447316408157349, + 0.13178157806396484 + ], + "credit_map": [ + 0.008819987997412682, + 0.017897771671414375, + 0.011125744320452213, + 0.014022949151694775, + 0.009865013882517815, + 0.013374000787734985, + 0.009208502247929573, + 0.009442048147320747, + 0.0028779858257621527, + 0.003506096312776208, + 0.00268046953715384, + 0.002772877225652337, + 0.0030624698847532272, + 0.003945917822420597, + 0.0032983350101858377, + 0.0027236926835030317, + 0.004983693826943636, + 0.0043092332780361176, + 0.00198769080452621, + 0.0024819965474307537, + 0.002204370917752385, + 0.003171098418533802, + 0.0020652783568948507, + 0.0022984312381595373, + 0.004983603022992611, + 0.009016525000333786, + 0.004899141378700733, + 0.007950503379106522, + 0.01475292257964611, + 0.006170456297695637, + 0.008120739832520485, + 0.009373052045702934, + 0.026258230209350586, + 0.04052475839853287, + 0.025162044912576675, + 0.02540508657693863, + 0.03968320041894913, + 0.028606802225112915, + 0.02499302662909031, + 0.026455357670783997, + 0.010474211536347866, + 0.014991197735071182, + 0.014335754327476025, + 0.019203651696443558, + 0.015257694758474827, + 0.022577116265892982, + 0.0248775202780962, + 0.018248436972498894, + 0.002684939419850707, + 0.005064396187663078, + 0.0030769205186516047, + 0.004239055793732405, + 0.004917866084724665, + 0.005489106755703688, + 0.006607457064092159, + 0.005044944584369659, + 0.023260263726115227, + 0.012756388634443283, + 0.01212314423173666, + 0.013507286086678505, + 0.015872754156589508, + 0.01920214109122753, + 0.014219366014003754, + 0.02123253233730793, + 0.0062050288543105125, + 0.014170548878610134, + 0.01126217283308506, + 0.023740002885460854, + 0.010378608480095863, + 0.00929755438119173, + 0.023957790806889534, + 0.01546145137399435, + 0.01451766025274992, + 0.012865317985415459, + 0.011284473352134228, + 0.015612639486789703, + 0.014060196466743946, + 0.023821638897061348, + 0.02098797634243965, + 0.018631668761372566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.012541357427835464 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a slightly upward angle, the camera slowly panning forward, accompanied by a gentle, irregular sway simulated by wind, captures the breathtaking spectacle of an ancient chain bridge swaying violently over a bottomless chasm. The sky is shrouded in thick, swirling clouds, as dark as ink. Pale, forked lightning occasionally rips through the canopy, each flash momentarily illuminating the bridge's thick chains, coated in dark brown grains of rust, and its slippery, dark wood, soaked by the torrential rain and reflecting a cold, high-gloss light. Under the raging wind, the entire bridge groans under the weight, emitting a harsh sound of metal rubbing and wood squeezing. It twists and throbs in agony like a dying python, and through the gaps in the planks, one can glimpse the endless darkness below, swallowed by the thick fog. This scene was built with the extremely high standards of the UE engine, presenting a film-like texture. Whether it is the traces of rain flowing on the iron chain or the wet texture of the wooden board, it has achieved a realistic level of detail. At the same time, a layer of subtle film grain is covered on the picture, which not only enhances the visual depth, but also creates a tense, epic and doomsday atmosphere of despair.", + "seed": 2886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_9568_seed_2886_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14295761936809867, + "vlm_logprob": -0.14295761936809867, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.14295761936809867, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04762709513306618 + ], + "opposite_logprob_tokens": [ + -3.1726272106170654 + ], + "logprob_reward": -0.04762709513306618, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04762709513306618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20045247673988342, + 0.07691066712141037, + 0.09943367540836334, + 0.10169057548046112, + 0.05435333773493767, + 0.019825348630547523, + 0.06130366772413254, + 0.054242558777332306, + 0.3011215627193451, + 0.030666204169392586 + ], + "credit_map": [ + 0.01935580186545849, + 0.02179322950541973, + 0.014541951939463615, + 0.023937812075018883, + 0.021246885880827904, + 0.027460243552923203, + 0.036555152386426926, + 0.03556138649582863, + 0.01242961548268795, + 0.011054754257202148, + 0.005612267181277275, + 0.007583118509501219, + 0.007908087223768234, + 0.008566182106733322, + 0.011328991502523422, + 0.01242764201015234, + 0.010900590568780899, + 0.01702069863677025, + 0.008358186110854149, + 0.006737131625413895, + 0.010892816819250584, + 0.02169177681207657, + 0.00925716944038868, + 0.014575297944247723, + 0.01502461638301611, + 0.009639468044042587, + 0.021271781995892525, + 0.004379947669804096, + 0.017370443791151047, + 0.013918854296207428, + 0.010598689317703247, + 0.009486761875450611, + 0.009173834696412086, + 0.00911963265389204, + 0.006513877771794796, + 0.0032913200557231903, + 0.0070310188457369804, + 0.006676146294921637, + 0.004836434032768011, + 0.00771106593310833, + 0.003109435783699155, + 0.0019657558295875788, + 0.0026330009568482637, + 0.0012095959391444921, + 0.0028004227206110954, + 0.0025905093643814325, + 0.0014367032563313842, + 0.00407992210239172, + 0.017612800002098083, + 0.012959124520421028, + 0.005924719851464033, + 0.0021550385281443596, + 0.004581183660775423, + 0.007121460046619177, + 0.004280383232980967, + 0.006668954622000456, + 0.014404281042516232, + 0.015106601640582085, + 0.0044472371228039265, + 0.0028181569650769234, + 0.005636678542941809, + 0.005118119530379772, + 0.003930488135665655, + 0.0027809953317046165, + 0.09761135280132294, + 0.11198814958333969, + 0.007650301791727543, + 0.0064943344332277775, + 0.04746969789266586, + 0.020245594903826714, + 0.005048093385994434, + 0.004613985773175955, + 0.0033893107902258635, + 0.004301376175135374, + 0.0032827178947627544, + 0.005122938193380833, + 0.0028833788819611073, + 0.004022042267024517, + 0.004039794206619263, + 0.003624641802161932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.04762709513306618 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328364610671997 + ], + "opposite_logprob_tokens": [ + -0.7578364610671997 + ], + "logprob_reward": -0.6328364610671997, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328364610671997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053806111216545105, + 0.07414139807224274, + 0.28257396817207336, + 0.019726915284991264, + 0.03014606423676014, + 0.01806533709168434, + 0.23467868566513062, + 0.11708441376686096, + 0.1551298350095749, + 0.014647290110588074 + ], + "credit_map": [ + 0.008554789237678051, + 0.0063059222884476185, + 0.003112109610810876, + 0.00580323301255703, + 0.0042546638287603855, + 0.0070581925101578236, + 0.006893507670611143, + 0.011823696084320545, + 0.01262399461120367, + 0.01664183847606182, + 0.007473947945982218, + 0.005804203916341066, + 0.008081840351223946, + 0.009461202658712864, + 0.006808809470385313, + 0.007245561573654413, + 0.034659244120121, + 0.05812109634280205, + 0.026641225442290306, + 0.017048770561814308, + 0.03534218296408653, + 0.04626164957880974, + 0.027593674138188362, + 0.036906108260154724, + 0.002811623504385352, + 0.002051199320703745, + 0.0027797380462288857, + 0.0013365017948672175, + 0.0031804253812879324, + 0.002581157023087144, + 0.0019846840295940638, + 0.0030015879310667515, + 0.0038422492798417807, + 0.0026983218267560005, + 0.0037502071354538202, + 0.002416393021121621, + 0.0037315275985747576, + 0.004661628510802984, + 0.005148238968104124, + 0.003897498594596982, + 0.003758559701964259, + 0.0016000097384676337, + 0.0019451824482530355, + 0.000891083967871964, + 0.0023541992995887995, + 0.00321874744258821, + 0.0014761765487492085, + 0.0028213781770318747, + 0.08005472272634506, + 0.014970017597079277, + 0.02250794693827629, + 0.007910280488431454, + 0.013342452235519886, + 0.07567095756530762, + 0.01105963159352541, + 0.009162670001387596, + 0.041627436876297, + 0.03732271119952202, + 0.005032332614064217, + 0.004048623610287905, + 0.010348428040742874, + 0.010655584745109081, + 0.003730514319613576, + 0.004318786785006523, + 0.05384715273976326, + 0.05049863085150719, + 0.005707654170691967, + 0.005770939402282238, + 0.017984120175242424, + 0.012612339109182358, + 0.0037835417315363884, + 0.004925438668578863, + 0.0021709648426622152, + 0.0021096852142363787, + 0.0017908122390508652, + 0.0014986699679866433, + 0.0019776765257120132, + 0.002117833122611046, + 0.001274058478884399, + 0.0017075897194445133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.6328364610671997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00194110709708184 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.00194110709708184, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00194110709708184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04826115816831589, + 0.08135098963975906, + 0.2151491492986679, + 0.02289370447397232, + 0.05266658216714859, + 0.020709598436951637, + 0.20045481622219086, + 0.122996024787426, + 0.21466292440891266, + 0.02085503190755844 + ], + "credit_map": [ + 0.006848176009953022, + 0.006129893008619547, + 0.0031661188695579767, + 0.003791258204728365, + 0.005064177326858044, + 0.007002776488661766, + 0.006830332800745964, + 0.009428426623344421, + 0.015524840913712978, + 0.01799033395946026, + 0.008510974235832691, + 0.005654538050293922, + 0.010306516662240028, + 0.009450090117752552, + 0.006519187707453966, + 0.007394508924335241, + 0.02543892152607441, + 0.04418961703777313, + 0.0207337848842144, + 0.013525777496397495, + 0.027024853974580765, + 0.03501027077436447, + 0.021066218614578247, + 0.02815971150994301, + 0.0035880592186003923, + 0.0024944585748016834, + 0.003688238561153412, + 0.0011566332541406155, + 0.0037500099278986454, + 0.003491069423034787, + 0.0019575455226004124, + 0.00276769045740366, + 0.007434310857206583, + 0.006475480273365974, + 0.006324890069663525, + 0.004721204284578562, + 0.007141200825572014, + 0.008223162963986397, + 0.006347205489873886, + 0.0059991255402565, + 0.00279924925416708, + 0.0020764770451933146, + 0.002716611372306943, + 0.0010564045514911413, + 0.0026392468716949224, + 0.003709473880007863, + 0.0016453579301014543, + 0.004066778346896172, + 0.07460936158895493, + 0.017558882012963295, + 0.016447121277451515, + 0.0067386808805167675, + 0.009978570975363255, + 0.058686479926109314, + 0.00859992578625679, + 0.007835784927010536, + 0.04169132560491562, + 0.03870339319109917, + 0.007106284145265818, + 0.0035720246378332376, + 0.012199608609080315, + 0.0115351602435112, + 0.004945083055645227, + 0.0032431536819785833, + 0.0775219202041626, + 0.06438246369361877, + 0.008805016987025738, + 0.009244279004633427, + 0.021764276549220085, + 0.020557982847094536, + 0.0058078765869140625, + 0.006579102948307991, + 0.0030696887988597155, + 0.002637004479765892, + 0.002221850212663412, + 0.0021014795638620853, + 0.0032913056202232838, + 0.003813224844634533, + 0.0017378232441842556, + 0.0019826542120426893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.00194110709708184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185900524258614 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018185900524258614, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185900524258614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06233026832342148, + 0.0798923596739769, + 0.11294645071029663, + 0.030964292585849762, + 0.028933215886354446, + 0.011600298807024956, + 0.12355082482099533, + 0.0995091050863266, + 0.4298490881919861, + 0.02042410522699356 + ], + "credit_map": [ + 0.013196910731494427, + 0.007456141524016857, + 0.004466798156499863, + 0.005457103252410889, + 0.00530254514887929, + 0.0080176442861557, + 0.007484850939363241, + 0.010948280803859234, + 0.016686981543898582, + 0.021643150597810745, + 0.006767620798200369, + 0.005526458844542503, + 0.006546422373503447, + 0.008293941617012024, + 0.006910683121532202, + 0.007517108228057623, + 0.01491381786763668, + 0.023944362998008728, + 0.010930458083748817, + 0.007924974896013737, + 0.014297418296337128, + 0.014578090980648994, + 0.012396549805998802, + 0.013960786163806915, + 0.004208116792142391, + 0.0030615641735494137, + 0.003941507078707218, + 0.0018398131942376494, + 0.004631660878658295, + 0.004602143540978432, + 0.0037804751191288233, + 0.004899016115814447, + 0.0034524479415267706, + 0.0034390054643154144, + 0.0031904459465295076, + 0.0021794147323817015, + 0.004458896815776825, + 0.003340452443808317, + 0.004425478633493185, + 0.004447076469659805, + 0.001324604731053114, + 0.001054870430380106, + 0.0015018864069133997, + 0.000818534754216671, + 0.0015664045931771398, + 0.0017014658078551292, + 0.0011862806277349591, + 0.0024462526198476553, + 0.04368635267019272, + 0.00846729427576065, + 0.010463730432093143, + 0.004504394251853228, + 0.00628465972840786, + 0.038756903260946274, + 0.005759117193520069, + 0.005628383718430996, + 0.02998560480773449, + 0.027732256799936295, + 0.008262217044830322, + 0.004631552845239639, + 0.00920084211975336, + 0.010347946546971798, + 0.004359350074082613, + 0.004989341367036104, + 0.12635581195354462, + 0.14446543157100677, + 0.020114796236157417, + 0.014167800545692444, + 0.06304265558719635, + 0.035315051674842834, + 0.01337040588259697, + 0.013017188757658005, + 0.002887116279453039, + 0.003050378989428282, + 0.002880122745409608, + 0.0027699556667357683, + 0.0022011047694832087, + 0.002731849905103445, + 0.001643204246647656, + 0.0022603743709623814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.018185900524258614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197533018887043 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197533018887043, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197533018887043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08877861499786377, + 0.07525555044412613, + 0.20946116745471954, + 0.028975224122405052, + 0.12111411988735199, + 0.022547295317053795, + 0.18418091535568237, + 0.03164079412817955, + 0.18256162106990814, + 0.05548461899161339 + ], + "credit_map": [ + 0.008273993618786335, + 0.009087464772164822, + 0.007174745202064514, + 0.015469186939299107, + 0.008391306735575199, + 0.01107746921479702, + 0.01774493418633938, + 0.011559521779417992, + 0.014399592764675617, + 0.017802760004997253, + 0.006531693506985903, + 0.006089974194765091, + 0.006678517442196608, + 0.00878824945539236, + 0.007388618309050798, + 0.0075761605985462666, + 0.02449275553226471, + 0.04007972031831741, + 0.018383881077170372, + 0.012414863333106041, + 0.024574080482125282, + 0.040529996156692505, + 0.019712235778570175, + 0.029273660853505135, + 0.004215989727526903, + 0.0032645054161548615, + 0.004248208366334438, + 0.0012377387611195445, + 0.004400686360895634, + 0.004602544475346804, + 0.0026017469353973866, + 0.004403809085488319, + 0.01155819557607174, + 0.008741391822695732, + 0.015108665451407433, + 0.013090579770505428, + 0.017108295112848282, + 0.013632571324706078, + 0.02296270988881588, + 0.018911730498075485, + 0.003979358822107315, + 0.002826876239851117, + 0.002401818986982107, + 0.0013227523304522038, + 0.002523131901398301, + 0.004566544201225042, + 0.0020746521186083555, + 0.002852164674550295, + 0.06745420396327972, + 0.01641842909157276, + 0.015322833321988583, + 0.004694968927651644, + 0.008024921640753746, + 0.05344086512923241, + 0.009905432350933552, + 0.008919278159737587, + 0.00513888243585825, + 0.00547156622633338, + 0.004525783937424421, + 0.00228177011013031, + 0.0031978078186511993, + 0.003604419529438019, + 0.003554097143933177, + 0.0038664706517010927, + 0.06612475216388702, + 0.040788449347019196, + 0.009499975480139256, + 0.009503044188022614, + 0.018888765946030617, + 0.022611867636442184, + 0.006984836421906948, + 0.008159946650266647, + 0.007631484419107437, + 0.012736691161990166, + 0.00588366761803627, + 0.004788175690919161, + 0.007948493584990501, + 0.00627053901553154, + 0.004887313582003117, + 0.0053382571786642075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.014197533018887043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a slightly upward angle, the camera slowly panning forward, accompanied by a gentle, irregular sway simulated by wind, captures the breathtaking spectacle of an ancient chain bridge swaying violently over a bottomless chasm. The sky is shrouded in thick, swirling clouds, as dark as ink. Pale, forked lightning occasionally rips through the canopy, each flash momentarily illuminating the bridge's thick chains, coated in dark brown grains of rust, and its slippery, dark wood, soaked by the torrential rain and reflecting a cold, high-gloss light. Under the raging wind, the entire bridge groans under the weight, emitting a harsh sound of metal rubbing and wood squeezing. It twists and throbs in agony like a dying python, and through the gaps in the planks, one can glimpse the endless darkness below, swallowed by the thick fog. This scene was built with the extremely high standards of the UE engine, presenting a film-like texture. Whether it is the traces of rain flowing on the iron chain or the wet texture of the wooden board, it has achieved a realistic level of detail. At the same time, a layer of subtle film grain is covered on the picture, which not only enhances the visual depth, but also creates a tense, epic and doomsday atmosphere of despair.", + "seed": 2887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_9568_seed_2887_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2311162947677076, + "vlm_logprob": -0.2311162947677076, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.2311162947677076, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06332144886255264 + ], + "opposite_logprob_tokens": [ + -2.813321352005005 + ], + "logprob_reward": -0.06332144886255264, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06332144886255264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10375098884105682, + 0.04507442191243172, + 0.020947810262441635, + 0.04101860150694847, + 0.054904766380786896, + 0.47940707206726074, + 0.07747631520032883, + 0.09058079868555069, + 0.0413961298763752, + 0.045443203300237656 + ], + "credit_map": [ + 0.02404196932911873, + 0.013204981572926044, + 0.009435427375137806, + 0.010729925706982613, + 0.006921470630913973, + 0.02867116592824459, + 0.005700366105884314, + 0.005045661237090826, + 0.010301144793629646, + 0.005585745442658663, + 0.003869761014357209, + 0.0037457316648215055, + 0.005011111963540316, + 0.012194213457405567, + 0.001429067924618721, + 0.0029376400634646416, + 0.0030187168158590794, + 0.0032483020331710577, + 0.0020371517166495323, + 0.0014299878384917974, + 0.002077542943879962, + 0.003242970909923315, + 0.0021206866949796677, + 0.0037724480498582125, + 0.004999317694455385, + 0.004643634427338839, + 0.006914257537573576, + 0.002727895276620984, + 0.0043724472634494305, + 0.005791713483631611, + 0.005903976503759623, + 0.005665353499352932, + 0.007489279378205538, + 0.013223443180322647, + 0.00526290712878108, + 0.005315796937793493, + 0.007036092225462198, + 0.004664812237024307, + 0.005456740036606789, + 0.006455689202994108, + 0.0232735313475132, + 0.009228448383510113, + 0.13544891774654388, + 0.03284537419676781, + 0.025191783905029297, + 0.015989623963832855, + 0.1895029991865158, + 0.047926340252161026, + 0.008186395280063152, + 0.006414154544472694, + 0.007856430485844612, + 0.011088171042501926, + 0.009374533779919147, + 0.007094551809132099, + 0.01974485255777836, + 0.0077172210440039635, + 0.007440481334924698, + 0.0098453713580966, + 0.005278874188661575, + 0.006300789769738913, + 0.016541432589292526, + 0.015193410217761993, + 0.015023624524474144, + 0.014956803061068058, + 0.0028987384866923094, + 0.0038029025308787823, + 0.0026745395734906197, + 0.011568070389330387, + 0.010253572836518288, + 0.0045845406129956245, + 0.0024691529106348753, + 0.0031446071807295084, + 0.005033903755247593, + 0.004930154420435429, + 0.00741317356005311, + 0.007411681581288576, + 0.004092399496585131, + 0.0066228448413312435, + 0.00414400827139616, + 0.005795028060674667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.06332144886255264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053863525390625 + ], + "opposite_logprob_tokens": [ + -0.4288635551929474 + ], + "logprob_reward": -1.053863525390625, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053863525390625, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08336391299962997, + 0.14357954263687134, + 0.05383665859699249, + 0.1632431596517563, + 0.04442984238266945, + 0.28195157647132874, + 0.1459648162126541, + 0.031016336753964424, + 0.02738145925104618, + 0.025232702493667603 + ], + "credit_map": [ + 0.013501478359103203, + 0.012201931327581406, + 0.015428653918206692, + 0.009485805407166481, + 0.007204653695225716, + 0.015128870494663715, + 0.006150437518954277, + 0.004262084607034922, + 0.0357353650033474, + 0.01938963681459427, + 0.009394530206918716, + 0.009222045540809631, + 0.013875366188585758, + 0.04164236783981323, + 0.004107404034584761, + 0.01021282933652401, + 0.008125226013362408, + 0.007771092001348734, + 0.006150275003165007, + 0.004121766891330481, + 0.005310983397066593, + 0.006892037112265825, + 0.005325229372829199, + 0.010140049271285534, + 0.023241650313138962, + 0.014109441079199314, + 0.013367638923227787, + 0.010947800241410732, + 0.01543697901070118, + 0.018244711682200432, + 0.030305776745080948, + 0.03758915886282921, + 0.007295453455299139, + 0.004862672183662653, + 0.00449663819745183, + 0.005566392559558153, + 0.005660820286720991, + 0.005738413892686367, + 0.0046553341671824455, + 0.006154119502753019, + 0.015458719804883003, + 0.0090273292735219, + 0.11546365916728973, + 0.0454491563141346, + 0.012542381882667542, + 0.010402793996036053, + 0.0571390800178051, + 0.016468461602926254, + 0.021326348185539246, + 0.010312835685908794, + 0.016215255483984947, + 0.018847951665520668, + 0.021139053627848625, + 0.012647705152630806, + 0.03264680877327919, + 0.012828845530748367, + 0.004042958840727806, + 0.002711200388148427, + 0.004147015977650881, + 0.0033537454437464476, + 0.004566957708448172, + 0.005181515589356422, + 0.0045449151657521725, + 0.0024680267088115215, + 0.0028433252591639757, + 0.0029973359778523445, + 0.0022145055700093508, + 0.0055628204718232155, + 0.0062315273098647594, + 0.0034968494437634945, + 0.002109169727191329, + 0.0019259272376075387, + 0.0023684801999479532, + 0.0023157664109021425, + 0.004749995656311512, + 0.004553311970084906, + 0.0017804063390940428, + 0.002389894798398018, + 0.0036886725574731827, + 0.0033861740957945585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -1.053863525390625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615273628383875 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615273628383875, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615273628383875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12921099364757538, + 0.12367657572031021, + 0.0410149022936821, + 0.07924777269363403, + 0.04065326973795891, + 0.18811406195163727, + 0.22129006683826447, + 0.06622526794672012, + 0.07280778139829636, + 0.03775930404663086 + ], + "credit_map": [ + 0.02343565784394741, + 0.027810068801045418, + 0.017963046208024025, + 0.012185008265078068, + 0.00858671311289072, + 0.025691242888569832, + 0.00750328553840518, + 0.0060359761118888855, + 0.029928917065262794, + 0.01646343618631363, + 0.007774913217872381, + 0.008043203502893448, + 0.013658682815730572, + 0.035280607640743256, + 0.0035133841447532177, + 0.009013434872031212, + 0.006225853692740202, + 0.005946522578597069, + 0.0040282318368554115, + 0.0032578259706497192, + 0.004473112989217043, + 0.005354001186788082, + 0.0035003165248781443, + 0.008229036815464497, + 0.009898139163851738, + 0.006813868414610624, + 0.008069292642176151, + 0.008186106570065022, + 0.0062937247566878796, + 0.007632184773683548, + 0.015308423899114132, + 0.01704602874815464, + 0.006074398290365934, + 0.004238994792103767, + 0.004319236613810062, + 0.0058579882606863976, + 0.005465000402182341, + 0.004562574438750744, + 0.004757527727633715, + 0.005377549212425947, + 0.011770528741180897, + 0.007304915692657232, + 0.07439091056585312, + 0.03754449635744095, + 0.009999703615903854, + 0.008245300501585007, + 0.02576567232608795, + 0.01309253554791212, + 0.02560211345553398, + 0.017783092334866524, + 0.018047161400318146, + 0.029294194653630257, + 0.04991086572408676, + 0.021668318659067154, + 0.04097004979848862, + 0.018014265224337578, + 0.01388619840145111, + 0.006127709522843361, + 0.010354598984122276, + 0.006881884764879942, + 0.006388685200363398, + 0.007893701083958149, + 0.008345399983227253, + 0.006347088608890772, + 0.006470627151429653, + 0.011861657723784447, + 0.005681811831891537, + 0.016484839841723442, + 0.009818708524107933, + 0.010969342663884163, + 0.006231912411749363, + 0.005288877058774233, + 0.002825450384989381, + 0.00486702611669898, + 0.003941746894270182, + 0.00390714593231678, + 0.00403869291767478, + 0.009082498028874397, + 0.0038422434590756893, + 0.005254500545561314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.003615273628383875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195065014064312 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195065014064312, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195065014064312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03884098678827286, + 0.05798567458987236, + 0.016151154413819313, + 0.07841026037931442, + 0.03158712387084961, + 0.564063310623169, + 0.10657346993684769, + 0.05448431149125099, + 0.023867029696702957, + 0.028036722913384438 + ], + "credit_map": [ + 0.0035111941397190094, + 0.004343653097748756, + 0.005757307168096304, + 0.005025854334235191, + 0.004682979080826044, + 0.0033386352006345987, + 0.008375851437449455, + 0.0038055086042732, + 0.0072922431863844395, + 0.007901757024228573, + 0.0069463420659303665, + 0.006562334019690752, + 0.012652477249503136, + 0.009582514874637127, + 0.002595841186121106, + 0.004452159628272057, + 0.0016747008776292205, + 0.002692013280466199, + 0.0019350668881088495, + 0.001227175467647612, + 0.0021771141327917576, + 0.002144220983609557, + 0.0018608004320412874, + 0.002440060954540968, + 0.01052730530500412, + 0.007151165045797825, + 0.007337719667702913, + 0.00562125351279974, + 0.005710960831493139, + 0.006727981846779585, + 0.016008412465453148, + 0.019325455650687218, + 0.004189454950392246, + 0.005383345764130354, + 0.003396199317649007, + 0.004340943414717913, + 0.0037942053750157356, + 0.0032594779040664434, + 0.003384803654626012, + 0.0038386916276067495, + 0.025784604251384735, + 0.00898837111890316, + 0.1491633504629135, + 0.03638143837451935, + 0.029111968353390694, + 0.01710447482764721, + 0.2329491674900055, + 0.06457989662885666, + 0.016238925978541374, + 0.009282583370804787, + 0.01300134975463152, + 0.015581252984702587, + 0.012208384461700916, + 0.009122478775680065, + 0.0202246755361557, + 0.01091381162405014, + 0.010033898986876011, + 0.005119774490594864, + 0.008503034710884094, + 0.005756945349276066, + 0.004319837316870689, + 0.00698862224817276, + 0.00932321883738041, + 0.004438974428921938, + 0.003546632593497634, + 0.0022507428657263517, + 0.0019910698756575584, + 0.004810689948499203, + 0.004018296021968126, + 0.003096245462074876, + 0.0019497023895382881, + 0.002203649142757058, + 0.002833286300301552, + 0.002819482237100601, + 0.004307015333324671, + 0.006259949412196875, + 0.0018369305180385709, + 0.0033691762946546078, + 0.004080054350197315, + 0.0025308257900178432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.014195065014064312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586160942912102 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586160942912102, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586160942912102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10647857189178467, + 0.057521916925907135, + 0.0463942214846611, + 0.10229973495006561, + 0.05828697234392166, + 0.24168160557746887, + 0.08795925974845886, + 0.21523429453372955, + 0.041883960366249084, + 0.04225948825478554 + ], + "credit_map": [ + 0.021503092721104622, + 0.014701787382364273, + 0.010809894651174545, + 0.010225649923086166, + 0.010633580386638641, + 0.025224018841981888, + 0.006134519353508949, + 0.007246033754199743, + 0.013374396599829197, + 0.00877150148153305, + 0.004411688074469566, + 0.0036115918774157763, + 0.006053658667951822, + 0.015204810537397861, + 0.0018875434761866927, + 0.004206728655844927, + 0.0073173996061086655, + 0.006525814533233643, + 0.00385839631780982, + 0.0031461429316550493, + 0.00481666624546051, + 0.0064658187329769135, + 0.004269839264452457, + 0.009994142688810825, + 0.011351103894412518, + 0.009230165742337704, + 0.01270883996039629, + 0.010508033446967602, + 0.01248627994209528, + 0.015061846002936363, + 0.016436578705906868, + 0.014516890048980713, + 0.007608411833643913, + 0.006989148911088705, + 0.0055524324998259544, + 0.007290595211088657, + 0.008453353308141232, + 0.007663876749575138, + 0.006625245325267315, + 0.008103910833597183, + 0.012998988851904869, + 0.010663656517863274, + 0.0716962069272995, + 0.052670687437057495, + 0.011367588303983212, + 0.010544640943408012, + 0.03457476571202278, + 0.037165068089962006, + 0.007682975847274065, + 0.005329733714461327, + 0.010683931410312653, + 0.015484622679650784, + 0.007939727045595646, + 0.0085101667791605, + 0.022265976294875145, + 0.010062123648822308, + 0.04503893107175827, + 0.020030321553349495, + 0.036383140832185745, + 0.022168949246406555, + 0.017600132152438164, + 0.022963035851716995, + 0.028859572485089302, + 0.022190213203430176, + 0.0034997982438653708, + 0.0034681467805057764, + 0.003365847049281001, + 0.010302293114364147, + 0.008369685150682926, + 0.0058337668888270855, + 0.0034637281205505133, + 0.0035806940868496895, + 0.004514205269515514, + 0.004723048768937588, + 0.006664277520030737, + 0.006474316585808992, + 0.004125559702515602, + 0.00610063411295414, + 0.004963863641023636, + 0.004693580791354179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.020586160942912102 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a slightly upward angle, the camera slowly panning forward, accompanied by a gentle, irregular sway simulated by wind, captures the breathtaking spectacle of an ancient chain bridge swaying violently over a bottomless chasm. The sky is shrouded in thick, swirling clouds, as dark as ink. Pale, forked lightning occasionally rips through the canopy, each flash momentarily illuminating the bridge's thick chains, coated in dark brown grains of rust, and its slippery, dark wood, soaked by the torrential rain and reflecting a cold, high-gloss light. Under the raging wind, the entire bridge groans under the weight, emitting a harsh sound of metal rubbing and wood squeezing. It twists and throbs in agony like a dying python, and through the gaps in the planks, one can glimpse the endless darkness below, swallowed by the thick fog. This scene was built with the extremely high standards of the UE engine, presenting a film-like texture. Whether it is the traces of rain flowing on the iron chain or the wet texture of the wooden board, it has achieved a realistic level of detail. At the same time, a layer of subtle film grain is covered on the picture, which not only enhances the visual depth, but also creates a tense, epic and doomsday atmosphere of despair.", + "seed": 2888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_13/index_9568_seed_2888_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17897540174890308, + "vlm_logprob": -0.17897540174890308, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.17897540174890308, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03198203444480896 + ], + "opposite_logprob_tokens": [ + -3.531981945037842 + ], + "logprob_reward": -0.03198203444480896, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03198203444480896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2741597294807434, + 0.06955239176750183, + 0.0732322558760643, + 0.07898420095443726, + 0.039734333753585815, + 0.06547129899263382, + 0.05561292544007301, + 0.056481145322322845, + 0.03190888464450836, + 0.25486284494400024 + ], + "credit_map": [ + 0.054508283734321594, + 0.02864014357328415, + 0.024116694927215576, + 0.031057137995958328, + 0.020605511963367462, + 0.03211269900202751, + 0.021484477445483208, + 0.061634790152311325, + 0.007964903488755226, + 0.009888331405818462, + 0.0057449545711278915, + 0.010817384347319603, + 0.009141349233686924, + 0.00974578782916069, + 0.007571537513285875, + 0.008678152225911617, + 0.017645616084337234, + 0.005174253135919571, + 0.004806350916624069, + 0.006951871328055859, + 0.004740078002214432, + 0.0033124666661024094, + 0.022025642916560173, + 0.008575976826250553, + 0.015615464188158512, + 0.007267578039318323, + 0.003946889657527208, + 0.012419506907463074, + 0.0047663371078670025, + 0.014787795953452587, + 0.010675458237528801, + 0.009505178779363632, + 0.005259128287434578, + 0.005997860338538885, + 0.0032041745726019144, + 0.005950309801846743, + 0.004982995335012674, + 0.005570873152464628, + 0.004699769429862499, + 0.004069220740348101, + 0.007430867291986942, + 0.009468142874538898, + 0.005062844138592482, + 0.006573416758328676, + 0.012406074441969395, + 0.006636906880885363, + 0.011058547534048557, + 0.006834499072283506, + 0.009724353440105915, + 0.012302287854254246, + 0.0037825768813490868, + 0.0033813214395195246, + 0.0073044560849666595, + 0.006695243529975414, + 0.005173896439373493, + 0.007248791866004467, + 0.004494763910770416, + 0.006744283251464367, + 0.005636246409267187, + 0.005737275816500187, + 0.006295522674918175, + 0.006798718124628067, + 0.00931417103856802, + 0.011460166424512863, + 0.0035116446670144796, + 0.003603410441428423, + 0.004310275427997112, + 0.003024755045771599, + 0.00385337439365685, + 0.004507945850491524, + 0.004782444331794977, + 0.004315038211643696, + 0.03426339104771614, + 0.09871632605791092, + 0.014891502447426319, + 0.009868058376014233, + 0.05652375519275665, + 0.015045776963233948, + 0.009522265754640102, + 0.016031788662075996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.03198203444480896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261737823486328 + ], + "opposite_logprob_tokens": [ + -0.5761737823486328 + ], + "logprob_reward": -0.8261737823486328, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261737823486328, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09094386547803879, + 0.057836055755615234, + 0.021725043654441833, + 0.19875873625278473, + 0.09018837660551071, + 0.04496503993868828, + 0.03952262923121452, + 0.03080768883228302, + 0.05078485235571861, + 0.3744677007198334 + ], + "credit_map": [ + 0.016317520290613174, + 0.015107732266187668, + 0.008529243990778923, + 0.01344019640237093, + 0.009220120497047901, + 0.005406284239143133, + 0.008198722265660763, + 0.014724057167768478, + 0.006422123871743679, + 0.007282812613993883, + 0.005675322376191616, + 0.010661628097295761, + 0.00767433550208807, + 0.007097605150192976, + 0.004671007860451937, + 0.008351225405931473, + 0.003608771599829197, + 0.0022317231632769108, + 0.0016051250277087092, + 0.0019001777982339263, + 0.003041487652808428, + 0.0021521758753806353, + 0.004460209980607033, + 0.002725374884903431, + 0.03372152894735336, + 0.01705220341682434, + 0.012884022668004036, + 0.06681210547685623, + 0.012318653054535389, + 0.010707800276577473, + 0.016819749027490616, + 0.02844269573688507, + 0.017967965453863144, + 0.022199232131242752, + 0.004445693921297789, + 0.0067169503308832645, + 0.006682158913463354, + 0.014298727735877037, + 0.009867618791759014, + 0.008010038174688816, + 0.008703278377652168, + 0.006286323070526123, + 0.00324926245957613, + 0.004359561018645763, + 0.004715530201792717, + 0.008295413106679916, + 0.00473626097664237, + 0.004619413986802101, + 0.006573924329131842, + 0.008673788979649544, + 0.0043748486787080765, + 0.002994746435433626, + 0.004126148764044046, + 0.0032725606579333544, + 0.004745907615870237, + 0.0047607081942260265, + 0.004112378228455782, + 0.0041683753952383995, + 0.0030175652354955673, + 0.0035653910599648952, + 0.004124760627746582, + 0.0031773513182997704, + 0.003475298173725605, + 0.005166569259017706, + 0.007352000568062067, + 0.00747192557901144, + 0.005085817072540522, + 0.005447408650070429, + 0.005900673568248749, + 0.006383745931088924, + 0.004591051489114761, + 0.008552231825888157, + 0.047207195311784744, + 0.1463465392589569, + 0.01744084432721138, + 0.012528170831501484, + 0.09122876077890396, + 0.022561335936188698, + 0.012675690464675426, + 0.024479201063513756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.8261737823486328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019417019793763757 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019417019793763757, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019417019793763757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1833392232656479, + 0.06051307171583176, + 0.10476215183734894, + 0.0900416299700737, + 0.08510048687458038, + 0.059938326478004456, + 0.08007612079381943, + 0.04357833415269852, + 0.08749544620513916, + 0.20515519380569458 + ], + "credit_map": [ + 0.04923674091696739, + 0.015142719261348248, + 0.010737226344645023, + 0.017549533396959305, + 0.01001574657857418, + 0.012924999929964542, + 0.011289488524198532, + 0.05644277110695839, + 0.006305984687060118, + 0.006941801402717829, + 0.00653268676251173, + 0.011655091308057308, + 0.007777698803693056, + 0.00783777516335249, + 0.005919798742979765, + 0.007542237173765898, + 0.023609980940818787, + 0.0075020333752036095, + 0.004723063670098782, + 0.007341483142226934, + 0.007711135782301426, + 0.01325830165296793, + 0.02992563135921955, + 0.010690522380173206, + 0.011295895092189312, + 0.011673442088067532, + 0.005869253072887659, + 0.018711013719439507, + 0.006958988960832357, + 0.007718325592577457, + 0.011191261000931263, + 0.016623452305793762, + 0.014604468829929829, + 0.01844421960413456, + 0.006501833442598581, + 0.009667742997407913, + 0.00825985986739397, + 0.011579888872802258, + 0.010807770304381847, + 0.0052347066812217236, + 0.011087825521826744, + 0.009339557960629463, + 0.004660026170313358, + 0.006014371756464243, + 0.005781200248748064, + 0.011762936599552631, + 0.006724096368998289, + 0.004568308126181364, + 0.01549269538372755, + 0.01953979581594467, + 0.00678042508661747, + 0.005156045313924551, + 0.009403365664184093, + 0.006897509563714266, + 0.007735832594335079, + 0.009070455096662045, + 0.005254605785012245, + 0.005015685223042965, + 0.005827402696013451, + 0.005924124736338854, + 0.005553665105253458, + 0.004942371975630522, + 0.004853531252592802, + 0.006206949707120657, + 0.008789824321866035, + 0.01176542043685913, + 0.006439977325499058, + 0.007260976359248161, + 0.014498506672680378, + 0.014096943661570549, + 0.01146573293954134, + 0.013178056105971336, + 0.027841895818710327, + 0.06882558763027191, + 0.017652170732617378, + 0.009764264337718487, + 0.03877900540828705, + 0.013381308875977993, + 0.011364749632775784, + 0.017546214163303375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0019417019793763757 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31219661235809326, + 0.060831982642412186, + 0.028182627633213997, + 0.12652404606342316, + 0.061588387936353683, + 0.11516454815864563, + 0.033017661422491074, + 0.034544412046670914, + 0.03266816958785057, + 0.19528153538703918 + ], + "credit_map": [ + 0.09622202068567276, + 0.02131274715065956, + 0.012563259340822697, + 0.03020838275551796, + 0.01276317983865738, + 0.014187471941113472, + 0.013164051808416843, + 0.11177551001310349, + 0.008023777976632118, + 0.008824987336993217, + 0.00511319562792778, + 0.006902265828102827, + 0.010835183784365654, + 0.0063522676937282085, + 0.006434227339923382, + 0.008346076123416424, + 0.0040211183950304985, + 0.00269958539865911, + 0.0031555567402392626, + 0.004674418829381466, + 0.002945331856608391, + 0.0032740409951657057, + 0.0037898842711001635, + 0.0036226902157068253, + 0.01363342348486185, + 0.01260824128985405, + 0.008101143874228, + 0.031322915107011795, + 0.006525615695863962, + 0.009185489267110825, + 0.014266028068959713, + 0.030881192535161972, + 0.00894852913916111, + 0.010101760737597942, + 0.004832061473280191, + 0.007547479122877121, + 0.005888151470571756, + 0.00804994534701109, + 0.010089336894452572, + 0.006131122354418039, + 0.017609668895602226, + 0.017776450142264366, + 0.008026697672903538, + 0.008936429396271706, + 0.011003728024661541, + 0.024474099278450012, + 0.012164770625531673, + 0.015172700397670269, + 0.00411559222266078, + 0.0047942171804606915, + 0.0035958027001470327, + 0.0027990664821118116, + 0.004942564759403467, + 0.0033501731231808662, + 0.005604567006230354, + 0.0038156788796186447, + 0.004198107868432999, + 0.003868664149194956, + 0.0045095570385456085, + 0.005777443293482065, + 0.00413499167189002, + 0.0036785572301596403, + 0.0037647259887307882, + 0.004612365271896124, + 0.0033145195338875055, + 0.00555507093667984, + 0.0032865689136087894, + 0.0037600048817694187, + 0.0041373055428266525, + 0.004183777142316103, + 0.0038981435354799032, + 0.004532777238637209, + 0.021078556776046753, + 0.07073736935853958, + 0.00889348890632391, + 0.0061862110160291195, + 0.04959391430020332, + 0.015286334790289402, + 0.00737651577219367, + 0.016129150986671448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.020581839606165886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient chain bridge spanning a deep chasm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bridge swaying violently due to strong wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a sky with thick, dark, swirling clouds and occasional pale forked lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bridge chains coated in dark brown rust and the wooden planks wet with a high-gloss reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197650365531445 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197650365531445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197650365531445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22031553089618683, + 0.06997782737016678, + 0.07057507336139679, + 0.14550741016864777, + 0.10884610563516617, + 0.07277167588472366, + 0.10938873142004013, + 0.033897895365953445, + 0.05370267853140831, + 0.1150171309709549 + ], + "credit_map": [ + 0.033922914415597916, + 0.03310046344995499, + 0.02440912462770939, + 0.0293642096221447, + 0.021266894415020943, + 0.02363276667892933, + 0.025885915383696556, + 0.028733231127262115, + 0.007350794039666653, + 0.0077190520241856575, + 0.007315400056540966, + 0.013011512346565723, + 0.006912782322615385, + 0.009484926238656044, + 0.006582163739949465, + 0.011601193808019161, + 0.018719909712672234, + 0.003899645060300827, + 0.00313623808324337, + 0.004827883094549179, + 0.004356786143034697, + 0.0035710015799850225, + 0.024130716919898987, + 0.007932886481285095, + 0.027397315949201584, + 0.014765961095690727, + 0.008541883900761604, + 0.04799612984061241, + 0.007181000895798206, + 0.015511282719671726, + 0.011191661469638348, + 0.012922159396111965, + 0.014446798712015152, + 0.018557358533143997, + 0.006545146461576223, + 0.012223877012729645, + 0.011019783094525337, + 0.015858681872487068, + 0.016368713229894638, + 0.013825736939907074, + 0.009564276784658432, + 0.009825843386352062, + 0.006295500788837671, + 0.005890864413231611, + 0.008496914990246296, + 0.009697477333247662, + 0.011443665251135826, + 0.011557129211723804, + 0.022649236023426056, + 0.030388325452804565, + 0.00630323588848114, + 0.005518777761608362, + 0.011476208455860615, + 0.009139707311987877, + 0.008516493253409863, + 0.01539673563092947, + 0.00421447679400444, + 0.004345904104411602, + 0.0030366331338882446, + 0.003359962720423937, + 0.004207257181406021, + 0.004067832604050636, + 0.004421041812747717, + 0.006244783755391836, + 0.008182828314602375, + 0.00927529577165842, + 0.0059554073959589005, + 0.005704194772988558, + 0.006914535537362099, + 0.006441298872232437, + 0.0037155381869524717, + 0.007513573858886957, + 0.011667588725686073, + 0.037280552089214325, + 0.00796092115342617, + 0.004546674899756908, + 0.028626546263694763, + 0.008853180333971977, + 0.007521344348788261, + 0.00856031570583582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.014197650365531445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_14/index_tensor([161, 161])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_14/index_tensor([161, 161])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..337010c8905d8e7aa177945b2b59269934e7ce49 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_14/index_tensor([161, 161])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static shot depicts the rear of a pearlescent white, wedge-shaped supercar, parked on a rain-slicked city street at night. Its low-slung chassis is all sharp angles and aggressive lines, culminating in a mesmerizing array of vibrant, glowing crimson hexagonal taillights, brimming with potential energy. The surrounding architecture is a symphony of towering Art Deco skyscrapers, their facades adorned with intricate gilded geometric patterns and bold vertical chrome lines, gleaming in the city lights, their reflections shimmering on the dark, wet asphalt. Suddenly, the taillights flash violently, and the supercar launches forward with explosive and decisive acceleration, its wide rear tires viciously cutting through standing water and sending a dramatic plume of spray into the air. As the car speeds down the street, transforming into a streak of white light, the camera smoothly and swiftly pulls up, rising vertically to reveal a breathtaking, sprawling metropolis stretching to the horizon, a vast network of illuminated streets and monumental buildings creating a powerful and awe-inspiring sense of scale and speed.", + "seed": 14033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_14/index_161_seed_14033_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030856810323894025, + "vlm_logprob": -0.030856810323894025, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.030856810323894025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02859410271048546 + ], + "opposite_logprob_tokens": [ + -4.153594017028809 + ], + "logprob_reward": -0.02859410271048546, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02859410271048546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24933572113513947, + 0.05087887868285179, + 0.023025307804346085, + 0.17011813819408417, + 0.017058461904525757, + 0.03977027162909508, + 0.34668129682540894, + 0.034678228199481964, + 0.04147202894091606, + 0.02698172628879547 + ], + "credit_map": [ + 0.01956593245267868, + 0.019749227911233902, + 0.06431308388710022, + 0.03726194426417351, + 0.05095215514302254, + 0.02237003482878208, + 0.018535146489739418, + 0.016588155180215836, + 0.006817740388214588, + 0.005452804267406464, + 0.006641066167503595, + 0.01306646503508091, + 0.004744976293295622, + 0.004967553075402975, + 0.004978884477168322, + 0.004209380596876144, + 0.002265044953674078, + 0.0031492654234170914, + 0.004560350440442562, + 0.002751357387751341, + 0.001984931528568268, + 0.003558603348210454, + 0.0018606525845825672, + 0.002895097713917494, + 0.01974709890782833, + 0.01547899842262268, + 0.022901024669408798, + 0.014288228936493397, + 0.012857613153755665, + 0.05852017179131508, + 0.009151758626103401, + 0.01717320643365383, + 0.0015156042063608766, + 0.002152244560420513, + 0.0015352948103100061, + 0.0028792580123990774, + 0.0024373969063162804, + 0.002404342172667384, + 0.0020759045146405697, + 0.0020584133453667164, + 0.01057377364486456, + 0.0037220832891762257, + 0.0026619199197739363, + 0.003292286768555641, + 0.006994818802922964, + 0.004458890296518803, + 0.003809350309893489, + 0.004257142543792725, + 0.024120429530739784, + 0.05300087481737137, + 0.04220392182469368, + 0.03198227658867836, + 0.1339402049779892, + 0.01902163028717041, + 0.014996905811131, + 0.02741503342986107, + 0.0029244362376630306, + 0.0069058178924024105, + 0.003281820798292756, + 0.004561432637274265, + 0.0048116655088961124, + 0.0031120621133595705, + 0.005579862743616104, + 0.003501125844195485, + 0.006997536402195692, + 0.006130801979452372, + 0.00793523620814085, + 0.006105923093855381, + 0.00423921225592494, + 0.0029200343415141106, + 0.00382603844627738, + 0.0033172410912811756, + 0.0032402132637798786, + 0.004581519868224859, + 0.005195351783186197, + 0.003142506815493107, + 0.0030991043895483017, + 0.0014411407755687833, + 0.003854976734146476, + 0.002426909049972892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.02859410271048546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018205448985099792 + ], + "opposite_logprob_tokens": [ + -4.018205642700195 + ], + "logprob_reward": -0.018205448985099792, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018205448985099792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2862621545791626, + 0.059945642948150635, + 0.12713658809661865, + 0.044349610805511475, + 0.08853978663682938, + 0.03859961777925491, + 0.12451562285423279, + 0.05874895676970482, + 0.12420035153627396, + 0.04770166054368019 + ], + "credit_map": [ + 0.01923344098031521, + 0.027122605592012405, + 0.06576412916183472, + 0.04370441287755966, + 0.0798066183924675, + 0.01813811995089054, + 0.01735619455575943, + 0.015136657282710075, + 0.00773773156106472, + 0.005817838478833437, + 0.006483153905719519, + 0.010042562149465084, + 0.010586087591946125, + 0.0060626184567809105, + 0.005469182506203651, + 0.007746471557766199, + 0.009632155299186707, + 0.01725325733423233, + 0.022792790085077286, + 0.03325539827346802, + 0.009541725739836693, + 0.012105070985853672, + 0.014329531230032444, + 0.00822665635496378, + 0.0070771207101643085, + 0.005189725663512945, + 0.007261519320309162, + 0.004391702823340893, + 0.0048643676564097404, + 0.007601533085107803, + 0.0036764193791896105, + 0.004287221934646368, + 0.008572622202336788, + 0.009596608579158783, + 0.006730611901730299, + 0.018779845908284187, + 0.011916727758944035, + 0.013160726055502892, + 0.013254286721348763, + 0.006528361700475216, + 0.005828031804412603, + 0.00364798866212368, + 0.0037190790753811598, + 0.005573864560574293, + 0.006044901441782713, + 0.00471515953540802, + 0.004867508076131344, + 0.0042030843906104565, + 0.011771067976951599, + 0.0108627425506711, + 0.03103511407971382, + 0.018350187689065933, + 0.013895968906581402, + 0.007112385705113411, + 0.011936298571527004, + 0.019551848992705345, + 0.0042896568775177, + 0.014450428076088428, + 0.006459274794906378, + 0.009191418066620827, + 0.007722628768533468, + 0.005250111687928438, + 0.00709486473351717, + 0.004290571436285973, + 0.011297717690467834, + 0.02444472722709179, + 0.009428944438695908, + 0.011283118277788162, + 0.014355182647705078, + 0.00972873903810978, + 0.02654709853231907, + 0.017114819958806038, + 0.005878234747797251, + 0.005438038147985935, + 0.007393680978566408, + 0.004763138946145773, + 0.005281943362206221, + 0.0034844218753278255, + 0.011986376717686653, + 0.0034758257679641247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.018205448985099792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039812371134757996, + 0.047983162105083466, + 0.021366607397794724, + 0.13034537434577942, + 0.06380841135978699, + 0.025106675922870636, + 0.5091372132301331, + 0.04200340434908867, + 0.09697334468364716, + 0.02346343733370304 + ], + "credit_map": [ + 0.00401305640116334, + 0.003469152608886361, + 0.009873859584331512, + 0.004016504157334566, + 0.006619500927627087, + 0.004136021248996258, + 0.003116573905572295, + 0.004567698109894991, + 0.013322727754712105, + 0.006779469083994627, + 0.007278887089341879, + 0.0036027180030941963, + 0.004985310137271881, + 0.004962493199855089, + 0.0033924507442861795, + 0.003659096546471119, + 0.002008819952607155, + 0.002137446543201804, + 0.0042469557374715805, + 0.002996002323925495, + 0.0016203788109123707, + 0.004597000312060118, + 0.001825710409320891, + 0.001934289583005011, + 0.03249535709619522, + 0.01790941134095192, + 0.016977347433567047, + 0.016493666917085648, + 0.013199382461607456, + 0.011045673862099648, + 0.011549628339707851, + 0.01067490316927433, + 0.00630280002951622, + 0.0065153175964951515, + 0.005238288547843695, + 0.011642388068139553, + 0.008789033629000187, + 0.010612496174871922, + 0.009595735929906368, + 0.005112348590046167, + 0.003602700773626566, + 0.0018544434569776058, + 0.0023597378749400377, + 0.002828393131494522, + 0.003396750893443823, + 0.004266985692083836, + 0.002523883944377303, + 0.004273778758943081, + 0.0380144901573658, + 0.08378849178552628, + 0.05978136882185936, + 0.03532618656754494, + 0.197196364402771, + 0.028924867510795593, + 0.01960451900959015, + 0.04650092124938965, + 0.003657296299934387, + 0.00867532379925251, + 0.005672840401530266, + 0.004703546408563852, + 0.00653474684804678, + 0.0031270007602870464, + 0.005057495553046465, + 0.004575148690491915, + 0.020335759967565536, + 0.016911569982767105, + 0.011531439609825611, + 0.007311895489692688, + 0.010240373201668262, + 0.006868191063404083, + 0.012548860162496567, + 0.011225242167711258, + 0.0029858064372092485, + 0.0026148550678044558, + 0.004904080182313919, + 0.002246732823550701, + 0.003278455464169383, + 0.002004204550758004, + 0.00384939881041646, + 0.0015799012035131454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283647373318672 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283647373318672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283647373318672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06753852218389511, + 0.05192144587635994, + 0.04379242658615112, + 0.02033173479139805, + 0.029479701071977615, + 0.018903061747550964, + 0.5544564127922058, + 0.045217014849185944, + 0.0924578532576561, + 0.07590179890394211 + ], + "credit_map": [ + 0.005365653429180384, + 0.0050965300761163235, + 0.01640862040221691, + 0.012427587062120438, + 0.012487080879509449, + 0.006308228708803654, + 0.004386059939861298, + 0.005058753304183483, + 0.008899561129510403, + 0.005739973857998848, + 0.0081189488992095, + 0.006660377606749535, + 0.008415447548031807, + 0.005168017465621233, + 0.00421694153919816, + 0.004702173639088869, + 0.005053828004747629, + 0.004292876459658146, + 0.009707057848572731, + 0.003949274774640799, + 0.002456014510244131, + 0.010390715673565865, + 0.003243559505790472, + 0.00469909468665719, + 0.002795499050989747, + 0.0021214524749666452, + 0.002500467235222459, + 0.0016613340703770518, + 0.001513717113994062, + 0.005906120873987675, + 0.0015062782913446426, + 0.0023268640507012606, + 0.0031582925003021955, + 0.003003140678629279, + 0.0024005568120628595, + 0.005076379049569368, + 0.0038002687506377697, + 0.005668398924171925, + 0.0037946458905935287, + 0.002578014275059104, + 0.002600747626274824, + 0.001618312788195908, + 0.0019094649469479918, + 0.002135263988748193, + 0.0028635472990572453, + 0.0021789895836263895, + 0.001948136487044394, + 0.0036485965829342604, + 0.03908667340874672, + 0.08999539166688919, + 0.06955030560493469, + 0.047938596457242966, + 0.21424756944179535, + 0.03085280768573284, + 0.02285194955766201, + 0.03993312641978264, + 0.004094386473298073, + 0.008842671290040016, + 0.004565170966088772, + 0.006854352541267872, + 0.005958566442131996, + 0.0036081066355109215, + 0.007713158614933491, + 0.003580598160624504, + 0.01724541373550892, + 0.0175534188747406, + 0.012140571139752865, + 0.013147668913006783, + 0.010825694538652897, + 0.0059259627014398575, + 0.010007991455495358, + 0.005611120723187923, + 0.005909004248678684, + 0.009948395192623138, + 0.01823321171104908, + 0.009736038744449615, + 0.008228874765336514, + 0.0038955947384238243, + 0.015013271942734718, + 0.004937401972711086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.023283647373318672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000825554132462 + ], + "opposite_logprob_tokens": [ + -2.6950082778930664 + ], + "logprob_reward": -0.07000825554132462, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000825554132462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03653363883495331, + 0.03404780477285385, + 0.029681073501706123, + 0.02267899177968502, + 0.03565305843949318, + 0.013888103887438774, + 0.5930577516555786, + 0.07178331166505814, + 0.13240383565425873, + 0.030272413045167923 + ], + "credit_map": [ + 0.0030058780685067177, + 0.003506999695673585, + 0.008268136531114578, + 0.0079456502571702, + 0.004791506566107273, + 0.00376761332154274, + 0.002798880683258176, + 0.002448974410071969, + 0.005081299226731062, + 0.0034123866353183985, + 0.006083700340241194, + 0.003966018091887236, + 0.007225203327834606, + 0.0032600273843854666, + 0.002192612737417221, + 0.0028265570290386677, + 0.00306598749011755, + 0.0020016252528876066, + 0.008073299191892147, + 0.00199687946587801, + 0.0016556886257603765, + 0.008775847963988781, + 0.001649564248509705, + 0.002462181728333235, + 0.0046735238283872604, + 0.0029830327257514, + 0.0022182364482432604, + 0.0022139758802950382, + 0.0016444214852526784, + 0.005686480551958084, + 0.0014821247896179557, + 0.0017771985149011016, + 0.0038761331234127283, + 0.0035226966720074415, + 0.0030045073945075274, + 0.005952473264187574, + 0.004960010759532452, + 0.006256483029574156, + 0.004951979499310255, + 0.0031287744641304016, + 0.001711945398710668, + 0.001241828897036612, + 0.001290858373977244, + 0.001374807907268405, + 0.002059019636362791, + 0.002688723849132657, + 0.001360519090667367, + 0.0021604010835289955, + 0.042006850242614746, + 0.09492112696170807, + 0.07492652535438538, + 0.04941782355308533, + 0.22965605556964874, + 0.03308829665184021, + 0.02420262061059475, + 0.04483843967318535, + 0.006727476138621569, + 0.016453120857477188, + 0.0080686304718256, + 0.009991224855184555, + 0.009253323078155518, + 0.006003294140100479, + 0.008989814668893814, + 0.006296426523476839, + 0.025734219700098038, + 0.028012458235025406, + 0.013481799513101578, + 0.013958982191979885, + 0.01354622095823288, + 0.007939483970403671, + 0.016680872067809105, + 0.013049804605543613, + 0.003698953427374363, + 0.005717283580452204, + 0.00473332405090332, + 0.0038329684175550938, + 0.0029168003238737583, + 0.001705996342934668, + 0.005656281020492315, + 0.0020108052995055914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.07000825554132462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot depicts the rear of a pearlescent white, wedge-shaped supercar, parked on a rain-slicked city street at night. Its low-slung chassis is all sharp angles and aggressive lines, culminating in a mesmerizing array of vibrant, glowing crimson hexagonal taillights, brimming with potential energy. The surrounding architecture is a symphony of towering Art Deco skyscrapers, their facades adorned with intricate gilded geometric patterns and bold vertical chrome lines, gleaming in the city lights, their reflections shimmering on the dark, wet asphalt. Suddenly, the taillights flash violently, and the supercar launches forward with explosive and decisive acceleration, its wide rear tires viciously cutting through standing water and sending a dramatic plume of spray into the air. As the car speeds down the street, transforming into a streak of white light, the camera smoothly and swiftly pulls up, rising vertically to reveal a breathtaking, sprawling metropolis stretching to the horizon, a vast network of illuminated streets and monumental buildings creating a powerful and awe-inspiring sense of scale and speed.", + "seed": 14034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_14/index_161_seed_14034_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04733815677464008, + "vlm_logprob": -0.04733815677464008, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.04733815677464008, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037422552704811096 + ], + "opposite_logprob_tokens": [ + -3.7874226570129395 + ], + "logprob_reward": -0.037422552704811096, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.037422552704811096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33145788311958313, + 0.06569911539554596, + 0.07407106459140778, + 0.09043537080287933, + 0.046236056834459305, + 0.10811242461204529, + 0.03295322507619858, + 0.10146913677453995, + 0.07731630653142929, + 0.0722493976354599 + ], + "credit_map": [ + 0.03459484502673149, + 0.010157102718949318, + 0.10553579032421112, + 0.028643321245908737, + 0.016574550420045853, + 0.01672980561852455, + 0.10294422507286072, + 0.016278255730867386, + 0.010789789259433746, + 0.0034606773406267166, + 0.007560779806226492, + 0.010539940558373928, + 0.003265309613198042, + 0.006745264399796724, + 0.007645246107131243, + 0.01569211110472679, + 0.004695549141615629, + 0.004235109314322472, + 0.01346132904291153, + 0.010270400904119015, + 0.006300893146544695, + 0.009015563875436783, + 0.010803522542119026, + 0.01528869941830635, + 0.008537370711565018, + 0.007401087787002325, + 0.011497658677399158, + 0.009793942794203758, + 0.007798820734024048, + 0.010886008851230145, + 0.022781265899538994, + 0.011739211156964302, + 0.007879109121859074, + 0.00759521359577775, + 0.008367586880922318, + 0.005642900709062815, + 0.0036221579648554325, + 0.005913987290114164, + 0.0037080750335007906, + 0.0035070281010121107, + 0.012576221488416195, + 0.018123267218470573, + 0.013119947165250778, + 0.014896545559167862, + 0.01221210416406393, + 0.011017720215022564, + 0.011688915081322193, + 0.014477713033556938, + 0.004728676285594702, + 0.00550099229440093, + 0.004250003956258297, + 0.0033784827683120966, + 0.0039342776872217655, + 0.003819925244897604, + 0.00399575661867857, + 0.003345112781971693, + 0.0062299747951328754, + 0.01779817044734955, + 0.01598779857158661, + 0.016271239146590233, + 0.010763372294604778, + 0.011060649529099464, + 0.013854123651981354, + 0.009503807872533798, + 0.005228638183325529, + 0.01000630296766758, + 0.009943024255335331, + 0.010095828212797642, + 0.014454175718128681, + 0.009478769265115261, + 0.010968071408569813, + 0.0071414983831346035, + 0.006624230649322271, + 0.009175261482596397, + 0.008004344068467617, + 0.00854931864887476, + 0.0077653001062572, + 0.014957021921873093, + 0.008633151650428772, + 0.008540770970284939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.037422552704811096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822157695889473 + ], + "opposite_logprob_tokens": [ + -3.5298221111297607 + ], + "logprob_reward": -0.029822157695889473, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822157695889473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2052391618490219, + 0.04292317107319832, + 0.09953723102807999, + 0.19505634903907776, + 0.08263057470321655, + 0.04543408006429672, + 0.03490343317389488, + 0.1358911097049713, + 0.07724110782146454, + 0.08114379644393921 + ], + "credit_map": [ + 0.03267497941851616, + 0.008739020675420761, + 0.03844638168811798, + 0.01822385936975479, + 0.012948437593877316, + 0.02812056429684162, + 0.04771553352475166, + 0.018370388075709343, + 0.006608596071600914, + 0.002717820927500725, + 0.00894962903112173, + 0.006604913622140884, + 0.0036702349316328764, + 0.004168897867202759, + 0.005040192976593971, + 0.005162885412573814, + 0.007031172979623079, + 0.004883183632045984, + 0.01960904523730278, + 0.012134011834859848, + 0.007540907245129347, + 0.012067842297255993, + 0.015525729395449162, + 0.02074533887207508, + 0.015184761956334114, + 0.014763182029128075, + 0.022083960473537445, + 0.02253466472029686, + 0.029465923085808754, + 0.021057037636637688, + 0.04753018915653229, + 0.022436631843447685, + 0.01262260228395462, + 0.01925237663090229, + 0.012665721587836742, + 0.008380831219255924, + 0.008372216485440731, + 0.00813466776162386, + 0.006578398402780294, + 0.006623758003115654, + 0.0055727544240653515, + 0.005654299631714821, + 0.008826794102787971, + 0.006507483776658773, + 0.005048409570008516, + 0.003143993206322193, + 0.00552777573466301, + 0.005152571015059948, + 0.0063546448945999146, + 0.006655764766037464, + 0.005266272462904453, + 0.0036024332512170076, + 0.0037801959551870823, + 0.0034595790784806013, + 0.0028512123972177505, + 0.0029333310667425394, + 0.009312246926128864, + 0.022621171548962593, + 0.022639146074652672, + 0.021404584869742393, + 0.012739531695842743, + 0.016377456486225128, + 0.017632601782679558, + 0.013164373114705086, + 0.008858823217451572, + 0.01203268300741911, + 0.007518794387578964, + 0.00945535022765398, + 0.011228854767978191, + 0.010676863603293896, + 0.00899504590779543, + 0.008474692702293396, + 0.01021897979080677, + 0.010706399567425251, + 0.014644083566963673, + 0.008134905248880386, + 0.009716593660414219, + 0.00878873374313116, + 0.007696815766394138, + 0.01123728882521391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.029822157695889473 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343408972024918 + ], + "opposite_logprob_tokens": [ + -3.65134334564209 + ], + "logprob_reward": -0.026343408972024918, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343408972024918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16904102265834808, + 0.08465737849473953, + 0.08784317970275879, + 0.04827013239264488, + 0.11297202110290527, + 0.03373663127422333, + 0.09203029423952103, + 0.09613979607820511, + 0.09221359342336655, + 0.1830960363149643 + ], + "credit_map": [ + 0.020443974062800407, + 0.0045477584935724735, + 0.04927537590265274, + 0.015021906234323978, + 0.00944189541041851, + 0.005738554056733847, + 0.056640759110450745, + 0.007930816151201725, + 0.010749333538115025, + 0.00678256107494235, + 0.014330546371638775, + 0.014254075475037098, + 0.009494652971625328, + 0.009482487104833126, + 0.009901379235088825, + 0.00966233853250742, + 0.01155011635273695, + 0.009124357253313065, + 0.013035288080573082, + 0.009952673688530922, + 0.005859161727130413, + 0.010813153348863125, + 0.009018872864544392, + 0.018489548936486244, + 0.004643809515982866, + 0.003458383958786726, + 0.0049232193268835545, + 0.006352042313665152, + 0.004829383920878172, + 0.006690243259072304, + 0.011200127191841602, + 0.0061729238368570805, + 0.01879962347447872, + 0.02105417102575302, + 0.01816464774310589, + 0.01610824279487133, + 0.007737047038972378, + 0.008668918162584305, + 0.011039515025913715, + 0.011399851180613041, + 0.0034529881086200476, + 0.004424723330885172, + 0.006799028255045414, + 0.004565664567053318, + 0.003589041531085968, + 0.0030423516873270273, + 0.004039086401462555, + 0.003823748091235757, + 0.016407646238803864, + 0.018006183207035065, + 0.014586280100047588, + 0.008260146714746952, + 0.010171839036047459, + 0.009381758980453014, + 0.007354556117206812, + 0.007861885242164135, + 0.007459729444235563, + 0.009924114681780338, + 0.017207859084010124, + 0.01766493171453476, + 0.008789092302322388, + 0.012152046896517277, + 0.011455046012997627, + 0.011486979201436043, + 0.009224389679729939, + 0.01464556623250246, + 0.01048597414046526, + 0.012150866910815239, + 0.012974018231034279, + 0.011672470718622208, + 0.010125831700861454, + 0.010934475809335709, + 0.01464516669511795, + 0.024654170498251915, + 0.028845980763435364, + 0.018045159056782722, + 0.017777184024453163, + 0.04213262349367142, + 0.012566084042191505, + 0.02442966215312481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.026343408972024918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15573084354400635, + 0.07156524807214737, + 0.04963228851556778, + 0.047688379883766174, + 0.08908342570066452, + 0.0671200081706047, + 0.05050131678581238, + 0.22308827936649323, + 0.10358460247516632, + 0.14200559258460999 + ], + "credit_map": [ + 0.022707480937242508, + 0.005141591187566519, + 0.047759197652339935, + 0.01174839586019516, + 0.007272912655025721, + 0.00956546887755394, + 0.043480247259140015, + 0.00805553887039423, + 0.010573512874543667, + 0.005352822132408619, + 0.007711250334978104, + 0.009170769713819027, + 0.008334704674780369, + 0.006258643232285976, + 0.00637483224272728, + 0.017788715660572052, + 0.006765224039554596, + 0.003952649421989918, + 0.005701775662600994, + 0.005853891838341951, + 0.003273608162999153, + 0.005735623184591532, + 0.01000955794006586, + 0.008339955471456051, + 0.0042273374274373055, + 0.004396296571940184, + 0.005388235207647085, + 0.008621233515441418, + 0.005055977031588554, + 0.004492473788559437, + 0.009662237018346786, + 0.005844590254127979, + 0.015558257699012756, + 0.012748206965625286, + 0.012018843553960323, + 0.014043761417269707, + 0.005862442310899496, + 0.009726718068122864, + 0.010550599545240402, + 0.008574591018259525, + 0.008750444278120995, + 0.01403648592531681, + 0.009259951300919056, + 0.006358475424349308, + 0.008121337741613388, + 0.00601949030533433, + 0.005754305515438318, + 0.008819514885544777, + 0.0074601490050554276, + 0.006141109857708216, + 0.006975031457841396, + 0.006545412354171276, + 0.0051050870679318905, + 0.007783501874655485, + 0.00496179424226284, + 0.005529232323169708, + 0.01917969435453415, + 0.04702243208885193, + 0.03446190804243088, + 0.03236757963895798, + 0.02109392173588276, + 0.020245736464858055, + 0.032704908400774, + 0.01601211167871952, + 0.011771482415497303, + 0.014412596821784973, + 0.01227354072034359, + 0.013316066935658455, + 0.015316045843064785, + 0.013483677059412003, + 0.011514411307871342, + 0.011496782302856445, + 0.011330566368997097, + 0.016170216724276543, + 0.01377896685153246, + 0.018781280145049095, + 0.017705000936985016, + 0.02801700308918953, + 0.018669167533516884, + 0.017553390935063362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703734636306763 + ], + "opposite_logprob_tokens": [ + -2.127037286758423 + ], + "logprob_reward": -0.12703734636306763, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703734636306763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18090376257896423, + 0.04345400631427765, + 0.02870548702776432, + 0.03197824954986572, + 0.08147025108337402, + 0.04961685836315155, + 0.05648079887032509, + 0.2095666080713272, + 0.0879739299416542, + 0.22985006868839264 + ], + "credit_map": [ + 0.019510844722390175, + 0.005110498983412981, + 0.05756477639079094, + 0.014935450628399849, + 0.009837508201599121, + 0.005876624491065741, + 0.0600857213139534, + 0.00798233225941658, + 0.0063539226539433, + 0.0025094265583902597, + 0.005469619762152433, + 0.0067390031181275845, + 0.0031990292482078075, + 0.0034383290912956, + 0.004766176454722881, + 0.010978498496115208, + 0.002458156319335103, + 0.002320937579497695, + 0.005039513111114502, + 0.0029819512274116278, + 0.00266123004257679, + 0.003906004363670945, + 0.003646416589617729, + 0.005691277328878641, + 0.0028474859427660704, + 0.002061937004327774, + 0.004446791484951973, + 0.004830298945307732, + 0.0035956064239144325, + 0.003763455431908369, + 0.006844238378107548, + 0.0035884373355656862, + 0.013443340547382832, + 0.020061979070305824, + 0.011234772391617298, + 0.010843129828572273, + 0.008689108304679394, + 0.0066605121828615665, + 0.004876601509749889, + 0.0056608025915920734, + 0.0066405911929905415, + 0.008137490600347519, + 0.008324491791427135, + 0.0062904455699026585, + 0.004888402763754129, + 0.004300043918192387, + 0.005016630981117487, + 0.006018761545419693, + 0.01046027522534132, + 0.00697995163500309, + 0.008609951473772526, + 0.007065949961543083, + 0.0056474353186786175, + 0.006667608395218849, + 0.006084074266254902, + 0.004965552128851414, + 0.014884634874761105, + 0.025173554196953773, + 0.048640746623277664, + 0.03632361441850662, + 0.019206631928682327, + 0.018728170543909073, + 0.02886865846812725, + 0.017740590497851372, + 0.014553631655871868, + 0.013884404674172401, + 0.009413043968379498, + 0.009627575054764748, + 0.011359626427292824, + 0.011262308806180954, + 0.007063715253025293, + 0.01080961711704731, + 0.023573489859700203, + 0.03350184112787247, + 0.029415246099233627, + 0.02316163294017315, + 0.015905290842056274, + 0.06075093522667885, + 0.013482830487191677, + 0.030058803036808968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.12703734636306763 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot depicts the rear of a pearlescent white, wedge-shaped supercar, parked on a rain-slicked city street at night. Its low-slung chassis is all sharp angles and aggressive lines, culminating in a mesmerizing array of vibrant, glowing crimson hexagonal taillights, brimming with potential energy. The surrounding architecture is a symphony of towering Art Deco skyscrapers, their facades adorned with intricate gilded geometric patterns and bold vertical chrome lines, gleaming in the city lights, their reflections shimmering on the dark, wet asphalt. Suddenly, the taillights flash violently, and the supercar launches forward with explosive and decisive acceleration, its wide rear tires viciously cutting through standing water and sending a dramatic plume of spray into the air. As the car speeds down the street, transforming into a streak of white light, the camera smoothly and swiftly pulls up, rising vertically to reveal a breathtaking, sprawling metropolis stretching to the horizon, a vast network of illuminated streets and monumental buildings creating a powerful and awe-inspiring sense of scale and speed.", + "seed": 14035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_14/index_161_seed_14035_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028447646275162698, + "vlm_logprob": -0.028447646275162698, + "vlm_yes_no_margin": 3.8749999523162844, + "vlm_token_credit": -0.028447646275162698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026745300740003586 + ], + "opposite_logprob_tokens": [ + -4.401745319366455 + ], + "logprob_reward": -0.026745300740003586, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.026745300740003586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18782104551792145, + 0.04779606685042381, + 0.08240001648664474, + 0.07554508745670319, + 0.0522601455450058, + 0.27583634853363037, + 0.06110897287726402, + 0.040047720074653625, + 0.14073054492473602, + 0.03645410016179085 + ], + "credit_map": [ + 0.025384824723005295, + 0.014258584938943386, + 0.0218272116035223, + 0.03806624561548233, + 0.01265731267631054, + 0.01203514076769352, + 0.03536853939294815, + 0.0282231904566288, + 0.0046827602200210094, + 0.006203317549079657, + 0.005305001512169838, + 0.008593407459557056, + 0.00698938500136137, + 0.00296017131768167, + 0.007955403067171574, + 0.005106619093567133, + 0.00829871091991663, + 0.009484249167144299, + 0.004835603758692741, + 0.012860095128417015, + 0.018584376201033592, + 0.005650042090564966, + 0.013306096196174622, + 0.009380838833749294, + 0.006493123713880777, + 0.004705784376710653, + 0.006205287761986256, + 0.017934627830982208, + 0.005985267926007509, + 0.004231627099215984, + 0.017442792654037476, + 0.012546579353511333, + 0.004415641073137522, + 0.0046172961592674255, + 0.0069153388030827045, + 0.009398963302373886, + 0.005338038317859173, + 0.0042179343290627, + 0.010379650630056858, + 0.006977282464504242, + 0.021921299397945404, + 0.044524114578962326, + 0.018675988540053368, + 0.019920552149415016, + 0.0591241829097271, + 0.0397525392472744, + 0.04382522776722908, + 0.028092434629797935, + 0.011003408581018448, + 0.005760500207543373, + 0.011104130186140537, + 0.004450424108654261, + 0.005922143347561359, + 0.006762034259736538, + 0.006245513912290335, + 0.009860818274319172, + 0.005146347917616367, + 0.004439848940819502, + 0.005369760096073151, + 0.005656747613102198, + 0.003479225328192115, + 0.0034679416567087173, + 0.006095562130212784, + 0.006392287090420723, + 0.016678299754858017, + 0.020357761532068253, + 0.030000878497958183, + 0.011026379652321339, + 0.008315100334584713, + 0.015011487528681755, + 0.010326085612177849, + 0.029014548286795616, + 0.0056987060233950615, + 0.004352767486125231, + 0.004376849625259638, + 0.004380499944090843, + 0.0033066608011722565, + 0.003761874744668603, + 0.002663511084392667, + 0.007913229987025261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.026745300740003586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969415068626404 + ], + "opposite_logprob_tokens": [ + -2.929969310760498 + ], + "logprob_reward": -0.054969415068626404, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969415068626404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2642841935157776, + 0.07418949156999588, + 0.09381716698408127, + 0.1279057413339615, + 0.09033263474702835, + 0.16784390807151794, + 0.06675475090742111, + 0.04033244401216507, + 0.037175536155700684, + 0.037364039570093155 + ], + "credit_map": [ + 0.02809508703649044, + 0.019728001207113266, + 0.02894863858819008, + 0.05135820806026459, + 0.028141483664512634, + 0.016077548265457153, + 0.048522405326366425, + 0.04341284558176994, + 0.007118621375411749, + 0.005654364824295044, + 0.013886971399188042, + 0.015269142575562, + 0.006421292200684547, + 0.0044162459671497345, + 0.014547314494848251, + 0.00687553733587265, + 0.007165567018091679, + 0.006242674309760332, + 0.008282124064862728, + 0.017084931954741478, + 0.013829746283590794, + 0.007212504744529724, + 0.016574544832110405, + 0.01742507703602314, + 0.010381742380559444, + 0.007447794079780579, + 0.009493771940469742, + 0.029394861310720444, + 0.00931289978325367, + 0.007389985956251621, + 0.03293323889374733, + 0.021551450714468956, + 0.007208989467471838, + 0.01007465086877346, + 0.012512722052633762, + 0.014811068773269653, + 0.01049069594591856, + 0.008041881956160069, + 0.015140095725655556, + 0.01205253042280674, + 0.013664687983691692, + 0.025384141132235527, + 0.01246224157512188, + 0.012097456492483616, + 0.036239806562662125, + 0.024387413635849953, + 0.02640712633728981, + 0.017201030626893044, + 0.01282375305891037, + 0.00832938589155674, + 0.010730466805398464, + 0.00567222572863102, + 0.009846841916441917, + 0.007814951241016388, + 0.00565237645059824, + 0.005884748883545399, + 0.004747845232486725, + 0.0047187162563204765, + 0.006154800299555063, + 0.004024240653961897, + 0.004090188071131706, + 0.005258487071841955, + 0.0048494539223611355, + 0.00648871622979641, + 0.0037883075419813395, + 0.004721999634057283, + 0.0042537180706858635, + 0.0046156044118106365, + 0.0028671969193965197, + 0.005039931740611792, + 0.004927189089357853, + 0.006961590610444546, + 0.004109947942197323, + 0.004470793530344963, + 0.0053360359743237495, + 0.006772997789084911, + 0.00342170731164515, + 0.0033483400475233793, + 0.003539996687322855, + 0.006364220287650824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.054969415068626404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640995249152184 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640995249152184, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640995249152184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20513875782489777, + 0.051761139184236526, + 0.09198186546564102, + 0.04488096013665199, + 0.07248326390981674, + 0.14301058650016785, + 0.05266120284795761, + 0.09459121525287628, + 0.15240386128425598, + 0.09108718484640121 + ], + "credit_map": [ + 0.02907472476363182, + 0.017537882551550865, + 0.022504795342683792, + 0.041864920407533646, + 0.019547220319509506, + 0.012461729347705841, + 0.040030427277088165, + 0.02211706154048443, + 0.00500941788777709, + 0.004372699186205864, + 0.00617508590221405, + 0.007101294584572315, + 0.008467687293887138, + 0.0035877081099897623, + 0.00837393756955862, + 0.008673306554555893, + 0.006738181225955486, + 0.0068711950443685055, + 0.008476314134895802, + 0.01592077501118183, + 0.017547260969877243, + 0.005797194316983223, + 0.014726834371685982, + 0.01590411737561226, + 0.004778084810823202, + 0.003385033691301942, + 0.004653444048017263, + 0.00859942939132452, + 0.005154496058821678, + 0.002903103129938245, + 0.008788504637777805, + 0.006618867162615061, + 0.004923911299556494, + 0.005628175567835569, + 0.013797590509057045, + 0.011773455888032913, + 0.008536162786185741, + 0.005565520841628313, + 0.012001194059848785, + 0.010257246904075146, + 0.010791695676743984, + 0.023366447538137436, + 0.010572018101811409, + 0.011406516656279564, + 0.02915191650390625, + 0.018756307661533356, + 0.022321665659546852, + 0.01664402335882187, + 0.004882090259343386, + 0.007380568888038397, + 0.005337866488844156, + 0.004615284502506256, + 0.011408754624426365, + 0.00715938163921237, + 0.006569646764546633, + 0.005307607352733612, + 0.011914500035345554, + 0.015514717437326908, + 0.01680777408182621, + 0.010785294696688652, + 0.0068356310948729515, + 0.01271726842969656, + 0.007069746498018503, + 0.012946277856826782, + 0.01337797287851572, + 0.01965264044702053, + 0.017659753561019897, + 0.015520735643804073, + 0.015151385217905045, + 0.025538397952914238, + 0.018074294552206993, + 0.027428681030869484, + 0.010139582678675652, + 0.011319493874907494, + 0.013485799543559551, + 0.012661196291446686, + 0.007534592412412167, + 0.00964916218072176, + 0.008097984828054905, + 0.018199373036623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.008640995249152184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187187612056732 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018187187612056732, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018187187612056732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08845709264278412, + 0.05392276123166084, + 0.10579878091812134, + 0.04579811170697212, + 0.08902657777070999, + 0.041464630514383316, + 0.15243329107761383, + 0.049747031182050705, + 0.27512326836586, + 0.09822849184274673 + ], + "credit_map": [ + 0.008164424449205399, + 0.008669190108776093, + 0.008156218566000462, + 0.018983162939548492, + 0.009488021954894066, + 0.005685004405677319, + 0.019096614792943, + 0.010214448906481266, + 0.005205648019909859, + 0.005779891274869442, + 0.006917320191860199, + 0.009823824279010296, + 0.007180086802691221, + 0.0028768531046807766, + 0.009216145612299442, + 0.006922988686710596, + 0.008969848975539207, + 0.01304162759333849, + 0.006025984417647123, + 0.013340278528630733, + 0.031486876308918, + 0.007131481077522039, + 0.015751594677567482, + 0.01005107443779707, + 0.0043251593597233295, + 0.003636915935203433, + 0.0050249784253537655, + 0.007053153123706579, + 0.005946747958660126, + 0.0036506399046629667, + 0.0089087700471282, + 0.007251744624227285, + 0.010761034674942493, + 0.009034192189574242, + 0.014862549491226673, + 0.014672327786684036, + 0.007767753209918737, + 0.005661877803504467, + 0.014364113099873066, + 0.011902722530066967, + 0.003801469225436449, + 0.0072935097850859165, + 0.0035230277571827173, + 0.004499015863984823, + 0.008148635737597942, + 0.003923133481293917, + 0.005881672725081444, + 0.004394163843244314, + 0.044180598109960556, + 0.012479688972234726, + 0.04090230539441109, + 0.013441182672977448, + 0.006150213070213795, + 0.01129728090018034, + 0.013368966989219189, + 0.010613043792545795, + 0.006226743571460247, + 0.006448348052799702, + 0.008402518928050995, + 0.005567347630858421, + 0.00411773007363081, + 0.004328764043748379, + 0.004362456034868956, + 0.01029311865568161, + 0.03139662742614746, + 0.023877743631601334, + 0.06814064085483551, + 0.025003382936120033, + 0.013525121845304966, + 0.019291045144200325, + 0.029766114428639412, + 0.06412257254123688, + 0.009184207767248154, + 0.009288724511861801, + 0.012142091058194637, + 0.012634895741939545, + 0.011699238792061806, + 0.015866870060563087, + 0.0073388488963246346, + 0.020073605701327324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.018187187612056732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369533270597458 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.03369533270597458, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369533270597458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060143593698740005, + 0.03847198188304901, + 0.055616337805986404, + 0.07322250306606293, + 0.11196800321340561, + 0.12972109019756317, + 0.04514405503869057, + 0.0963871031999588, + 0.22776676714420319, + 0.16155852377414703 + ], + "credit_map": [ + 0.006833069026470184, + 0.006283962167799473, + 0.007012928370386362, + 0.012383596040308475, + 0.005830499343574047, + 0.0033160888124257326, + 0.011328566819429398, + 0.00715488288551569, + 0.0033180059399455786, + 0.002797274850308895, + 0.0040808930061757565, + 0.0076271528378129005, + 0.004399448167532682, + 0.0022882474586367607, + 0.007873966358602047, + 0.006086995825171471, + 0.004595608450472355, + 0.006150668952614069, + 0.0027587926015257835, + 0.008957109414041042, + 0.014714877121150494, + 0.0028153012972325087, + 0.008788215927779675, + 0.006835768464952707, + 0.0056166863068938255, + 0.004275057464838028, + 0.006593117956072092, + 0.01831853948533535, + 0.005533332470804453, + 0.004119434393942356, + 0.017986994236707687, + 0.010779338888823986, + 0.010385436937212944, + 0.011912114918231964, + 0.010554932057857513, + 0.018939077854156494, + 0.019889749586582184, + 0.008680643513798714, + 0.018556108698248863, + 0.01304994523525238, + 0.01078333891928196, + 0.022773221135139465, + 0.007919538766145706, + 0.010009587742388248, + 0.03229016810655594, + 0.016182389110326767, + 0.018684785813093185, + 0.011078059673309326, + 0.013147068209946156, + 0.00434437720105052, + 0.010858711786568165, + 0.003099835244938731, + 0.0018741728272289038, + 0.0029497770592570305, + 0.004082221072167158, + 0.0047878907062113285, + 0.011943619698286057, + 0.014421786181628704, + 0.012832860462367535, + 0.01241092849522829, + 0.007168160751461983, + 0.011583949439227581, + 0.00997986551374197, + 0.016045935451984406, + 0.024827636778354645, + 0.023850195109844208, + 0.04503161087632179, + 0.02428206242620945, + 0.012609731405973434, + 0.019118787720799446, + 0.030321810394525528, + 0.04772493243217468, + 0.016959011554718018, + 0.016748422756791115, + 0.020870979875326157, + 0.0196818970143795, + 0.020042190328240395, + 0.026663413271307945, + 0.009635362774133682, + 0.03095725178718567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.03369533270597458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot depicts the rear of a pearlescent white, wedge-shaped supercar, parked on a rain-slicked city street at night. Its low-slung chassis is all sharp angles and aggressive lines, culminating in a mesmerizing array of vibrant, glowing crimson hexagonal taillights, brimming with potential energy. The surrounding architecture is a symphony of towering Art Deco skyscrapers, their facades adorned with intricate gilded geometric patterns and bold vertical chrome lines, gleaming in the city lights, their reflections shimmering on the dark, wet asphalt. Suddenly, the taillights flash violently, and the supercar launches forward with explosive and decisive acceleration, its wide rear tires viciously cutting through standing water and sending a dramatic plume of spray into the air. As the car speeds down the street, transforming into a streak of white light, the camera smoothly and swiftly pulls up, rising vertically to reveal a breathtaking, sprawling metropolis stretching to the horizon, a vast network of illuminated streets and monumental buildings creating a powerful and awe-inspiring sense of scale and speed.", + "seed": 14036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_14/index_161_seed_14036_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03662006594240665, + "vlm_logprob": -0.03662006594240665, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03662006594240665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01537260040640831 + ], + "opposite_logprob_tokens": [ + -4.7653727531433105 + ], + "logprob_reward": -0.01537260040640831, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01537260040640831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17368273437023163, + 0.07951124012470245, + 0.05897403880953789, + 0.0983525961637497, + 0.07818552106618881, + 0.05449073016643524, + 0.3206404447555542, + 0.048493389040231705, + 0.037709228694438934, + 0.04996005445718765 + ], + "credit_map": [ + 0.015602811239659786, + 0.010744291357696056, + 0.01516643725335598, + 0.04405893385410309, + 0.01644902676343918, + 0.01401200145483017, + 0.042310841381549835, + 0.015338404104113579, + 0.0073250471614301205, + 0.005776249337941408, + 0.017590560019016266, + 0.015101433731615543, + 0.007999038323760033, + 0.005107260774821043, + 0.015974679961800575, + 0.004636968951672316, + 0.006320159416645765, + 0.005259491037577391, + 0.007580826990306377, + 0.009662941098213196, + 0.009902037680149078, + 0.00651570362970233, + 0.007842808030545712, + 0.005890072323381901, + 0.011020706966519356, + 0.009948121383786201, + 0.02221551537513733, + 0.012730692513287067, + 0.013826243579387665, + 0.006034202873706818, + 0.015094929374754429, + 0.0074821836315095425, + 0.010147334076464176, + 0.009880575351417065, + 0.008746026083827019, + 0.009797575883567333, + 0.013915564864873886, + 0.0062683396972715855, + 0.01188526302576065, + 0.007544838357716799, + 0.004984606523066759, + 0.005074535962194204, + 0.005417331587523222, + 0.008092178963124752, + 0.007059736642986536, + 0.0045231967233121395, + 0.01091635599732399, + 0.0084227891638875, + 0.040042899549007416, + 0.015535429120063782, + 0.030831770971417427, + 0.0194388497620821, + 0.015755100175738335, + 0.025555314496159554, + 0.06603027135133743, + 0.10745079815387726, + 0.0029041001107543707, + 0.0038819743786007166, + 0.005897660739719868, + 0.0075264740735292435, + 0.009412139654159546, + 0.005348357371985912, + 0.006710363086313009, + 0.006812318228185177, + 0.0032772806007415056, + 0.0028612350579351187, + 0.0040115369483828545, + 0.005269570741802454, + 0.0070101423189044, + 0.0048570288345217705, + 0.005177096463739872, + 0.005245339125394821, + 0.00411206716671586, + 0.008944814093410969, + 0.005256975535303354, + 0.006485537625849247, + 0.007107334677129984, + 0.004753731191158295, + 0.005755074787884951, + 0.007544517982751131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.01537260040640831 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355136185884476 + ], + "opposite_logprob_tokens": [ + -3.651355028152466 + ], + "logprob_reward": -0.026355136185884476, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355136185884476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17251284420490265, + 0.16879388689994812, + 0.0757620558142662, + 0.030292117968201637, + 0.04353073984384537, + 0.04902361333370209, + 0.3521037995815277, + 0.043901361525058746, + 0.029495202004909515, + 0.034584444016218185 + ], + "credit_map": [ + 0.015150132589042187, + 0.012402643449604511, + 0.018245631828904152, + 0.03499028459191322, + 0.017824584618210793, + 0.01850915141403675, + 0.03940267860889435, + 0.015987714752554893, + 0.011944714933633804, + 0.019507506862282753, + 0.016607634723186493, + 0.02287941798567772, + 0.03995364159345627, + 0.015429560095071793, + 0.0301143117249012, + 0.012357089668512344, + 0.009391987696290016, + 0.007191430311650038, + 0.009218314662575722, + 0.01451561227440834, + 0.011073033325374126, + 0.007077791728079319, + 0.012582309544086456, + 0.004711571149528027, + 0.004787931218743324, + 0.0029787886887788773, + 0.00378286256454885, + 0.0045211054384708405, + 0.004268082790076733, + 0.002812181366607547, + 0.003752730321139097, + 0.0033884295262396336, + 0.006991724018007517, + 0.0061746626161038876, + 0.004630751442164183, + 0.004688349552452564, + 0.00638731149956584, + 0.004466972313821316, + 0.006090741138905287, + 0.00410022446885705, + 0.0037960559129714966, + 0.004609269555658102, + 0.004872099030762911, + 0.008269613608717918, + 0.0045671528205275536, + 0.005929769016802311, + 0.01071119774132967, + 0.00626844959333539, + 0.03925064578652382, + 0.016230721026659012, + 0.03312353417277336, + 0.023013297468423843, + 0.02180316112935543, + 0.026999805122613907, + 0.07664117962121964, + 0.11504141241312027, + 0.002671696711331606, + 0.0041227745823562145, + 0.0058889188803732395, + 0.007135096471756697, + 0.007178247440606356, + 0.005626740399748087, + 0.005433185026049614, + 0.00584469735622406, + 0.0016070448327809572, + 0.002036580117419362, + 0.0033795463386923075, + 0.0049353246577084064, + 0.005552145652472973, + 0.0036469579208642244, + 0.004635218530893326, + 0.003702381392940879, + 0.002719281706959009, + 0.0032867109403014183, + 0.0034009588416665792, + 0.007980157621204853, + 0.004197897855192423, + 0.0034965756349265575, + 0.0039042788557708263, + 0.005598581396043301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.026355136185884476 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067326370626688 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067326370626688, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067326370626688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05001688003540039, + 0.04901871085166931, + 0.05229610577225685, + 0.042665354907512665, + 0.04661673307418823, + 0.134125217795372, + 0.2704629898071289, + 0.1147158145904541, + 0.04232984781265259, + 0.19775229692459106 + ], + "credit_map": [ + 0.006316448096185923, + 0.0040913913398981094, + 0.00696738064289093, + 0.00613042339682579, + 0.0064162276685237885, + 0.005905785132199526, + 0.008449502289295197, + 0.0057397265918552876, + 0.004666891880333424, + 0.0021851188503205776, + 0.015220577828586102, + 0.008312346413731575, + 0.002991587622091174, + 0.008203723467886448, + 0.004766765516251326, + 0.0026716983411461115, + 0.007751634810119867, + 0.004109377972781658, + 0.007727228570729494, + 0.007929254323244095, + 0.0050337412394583225, + 0.0051241228356957436, + 0.007511653006076813, + 0.007109090685844421, + 0.009443860501050949, + 0.003649362362921238, + 0.006339221261441708, + 0.0054906983859837055, + 0.004278120119124651, + 0.004597086925059557, + 0.004966249223798513, + 0.003900758223608136, + 0.007199573330581188, + 0.005640136543661356, + 0.005661191884428263, + 0.005490781739354134, + 0.005018147639930248, + 0.006941096391528845, + 0.006101546809077263, + 0.004564265254884958, + 0.014073554426431656, + 0.016262095421552658, + 0.01934468001127243, + 0.024454742670059204, + 0.012708005495369434, + 0.010719063691794872, + 0.016807982698082924, + 0.019755100831389427, + 0.025306228548288345, + 0.012380479834973812, + 0.02401432953774929, + 0.016468891873955727, + 0.019333122298121452, + 0.02383652701973915, + 0.05757833272218704, + 0.09154509007930756, + 0.006820038892328739, + 0.010664266534149647, + 0.014430626295506954, + 0.012654424645006657, + 0.018008604645729065, + 0.02468395233154297, + 0.009675300680100918, + 0.017778607085347176, + 0.0035117201041430235, + 0.003991667181253433, + 0.005677227396517992, + 0.005905416328459978, + 0.00866009946912527, + 0.004558335989713669, + 0.0054310886189341545, + 0.004594296216964722, + 0.012657741084694862, + 0.024348491802811623, + 0.017516465857625008, + 0.04439456760883331, + 0.022248325869441032, + 0.020985515788197517, + 0.027838848531246185, + 0.02776235342025757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.0067326370626688 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624335587024689 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624335587024689, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624335587024689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06436185538768768, + 0.09339828789234161, + 0.05102994292974472, + 0.05600842460989952, + 0.07425080239772797, + 0.09268967062234879, + 0.25139960646629333, + 0.12838521599769592, + 0.060055967420339584, + 0.12842024862766266 + ], + "credit_map": [ + 0.007036377675831318, + 0.004360981285572052, + 0.006526282988488674, + 0.012098098173737526, + 0.006710972171276808, + 0.009512951597571373, + 0.011120826005935669, + 0.006995365954935551, + 0.007837311364710331, + 0.004141080658882856, + 0.021271347999572754, + 0.013267840258777142, + 0.00932928454130888, + 0.014742851257324219, + 0.01663319580256939, + 0.0061753722839057446, + 0.004877039231359959, + 0.003961409442126751, + 0.007755796425044537, + 0.007897776551544666, + 0.009515946730971336, + 0.005202971398830414, + 0.00766139617189765, + 0.004157603718340397, + 0.006996294483542442, + 0.006593777798116207, + 0.008065531961619854, + 0.0074127474799752235, + 0.0086893355473876, + 0.005844130180776119, + 0.007389018312096596, + 0.005017585586756468, + 0.010814961045980453, + 0.007793706376105547, + 0.011732751503586769, + 0.009621073491871357, + 0.008144978433847427, + 0.011575641110539436, + 0.007922638207674026, + 0.006645054556429386, + 0.008695153519511223, + 0.00995028205215931, + 0.011529404670000076, + 0.01501021720468998, + 0.011187837459146976, + 0.010536412708461285, + 0.013790223747491837, + 0.011990140192210674, + 0.01620367355644703, + 0.011596948839724064, + 0.025222983211278915, + 0.02078753337264061, + 0.02183833345770836, + 0.01891869306564331, + 0.061337534338235855, + 0.07549392431974411, + 0.005915192421525717, + 0.010172891430556774, + 0.010791818611323833, + 0.017416682094335556, + 0.025055453181266785, + 0.0145024498924613, + 0.02103513292968273, + 0.023495586588978767, + 0.0037934081628918648, + 0.004479974042624235, + 0.00582383805885911, + 0.008722684346139431, + 0.013616667129099369, + 0.007459873333573341, + 0.009289477951824665, + 0.006870042532682419, + 0.0077665140852332115, + 0.014982855878770351, + 0.009637419134378433, + 0.021602578461170197, + 0.016846906393766403, + 0.013894729316234589, + 0.023858465254306793, + 0.019830772653222084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.007624335587024689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pearlescent white, wedge-shaped supercar with sharp angles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the taillights vibrant, glowing crimson hexagons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-slicked city street at night with Art Deco skyscrapers featuring gilded geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car launch forward with explosive acceleration, cutting through standing water and creating a plume of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pull up vertically to reveal a sprawling metropolis stretching to the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127015620470047 + ], + "opposite_logprob_tokens": [ + -2.1270155906677246 + ], + "logprob_reward": -0.127015620470047, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127015620470047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024549148976802826, + 0.04153001680970192, + 0.028707312420010567, + 0.02874358929693699, + 0.03966734558343887, + 0.036087557673454285, + 0.4671444892883301, + 0.18540120124816895, + 0.049472544342279434, + 0.09869671612977982 + ], + "credit_map": [ + 0.00292806513607502, + 0.0020243616309016943, + 0.0024187862873077393, + 0.0039380439557135105, + 0.0032041005324572325, + 0.004961604718118906, + 0.003614688292145729, + 0.001459500752389431, + 0.003464794484898448, + 0.0019841354805976152, + 0.007302813697606325, + 0.007164109032601118, + 0.0032057815697044134, + 0.006311855744570494, + 0.009741125628352165, + 0.002355405129492283, + 0.0047679259441792965, + 0.002329003531485796, + 0.0033514914102852345, + 0.003638719441369176, + 0.0029743837658315897, + 0.004516213666647673, + 0.004199566785246134, + 0.002930012531578541, + 0.003895396366715431, + 0.0025132361333817244, + 0.0032392009161412716, + 0.005234821699559689, + 0.003164417576044798, + 0.0030591797549277544, + 0.004844023380428553, + 0.0027933164965361357, + 0.004969191271811724, + 0.004262024536728859, + 0.005880929529666901, + 0.005888658110052347, + 0.004257698077708483, + 0.004756644833832979, + 0.006146313156932592, + 0.00350589188747108, + 0.002988487482070923, + 0.0049388715997338295, + 0.004733442794531584, + 0.005509751383215189, + 0.004194315988570452, + 0.0035104125272482634, + 0.005353465676307678, + 0.004858813248574734, + 0.03293684497475624, + 0.019877368584275246, + 0.04263344779610634, + 0.037000298500061035, + 0.04346160218119621, + 0.034161053597927094, + 0.1151881068944931, + 0.14188583195209503, + 0.010436165146529675, + 0.015388378873467445, + 0.02535671927034855, + 0.026384757831692696, + 0.03372560441493988, + 0.025399014353752136, + 0.021297136321663857, + 0.027413446456193924, + 0.002630149945616722, + 0.004322023130953312, + 0.007237775716930628, + 0.005245382897555828, + 0.011415359564125538, + 0.005571875721216202, + 0.007671360857784748, + 0.005378620699048042, + 0.007338692434132099, + 0.010609287768602371, + 0.009666896425187588, + 0.024959906935691833, + 0.00958195235580206, + 0.009540703147649765, + 0.01489406730979681, + 0.012105213478207588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.127015620470047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_15/index_tensor([923, 923])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_15/index_tensor([923, 923])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a76a3fe0311846aa8f8fc8c26f35d6b0ca847a13 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_15/index_tensor([923, 923])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, as if filmed from a deep-sea submersible, captures a giant octopus as it emerges from the inky abyss. The creature floats majestically and deliberately in the murky waters, its body a marvel of ethereal biology. Its translucent, membranous flesh, shades of deep indigo and ghostly white, undulates with a hypnotic slowness, truly resembling a wisp of smoke in zero gravity. A cool internal bioluminescence pulses rhythmically from within its cloak and along the delicate webs between its arms, casting a soft, ever-shifting glow that illuminates the intricate patterns on its skin. Its enormous, intelligent silver eyes gaze forward into the void, while its tentacles trail behind with languid grace. The only light in the devastating darkness comes from the octopus itself, capturing countless grains of marine snow that drift across the camera like a slow-motion blizzard, creating an atmosphere of mystery and awe, witnessing a primordial, otherworldly being in its profoundly silent realm.", + "seed": 21207, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_15/index_923_seed_21207_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1177029269747436, + "vlm_logprob": -0.1177029269747436, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.1177029269747436, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5334846377372742 + ], + "opposite_logprob_tokens": [ + -0.9084846377372742 + ], + "logprob_reward": -0.5334846377372742, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5334846377372742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0811915174126625, + 0.11735323071479797, + 0.27969250082969666, + 0.04796444997191429, + 0.10418936610221863, + 0.1558072566986084, + 0.07559720426797867, + 0.045480258762836456, + 0.041714880615472794, + 0.0510093979537487 + ], + "credit_map": [ + 0.014060859568417072, + 0.008370423689484596, + 0.010422267019748688, + 0.011463246308267117, + 0.011010372079908848, + 0.00645761052146554, + 0.011446023359894753, + 0.007960706017911434, + 0.01980072818696499, + 0.010848509147763252, + 0.026144472882151604, + 0.012530840933322906, + 0.016126729547977448, + 0.009945978410542011, + 0.010670880787074566, + 0.011285086162388325, + 0.05183302238583565, + 0.006343078333884478, + 0.10605452209711075, + 0.0123814782127738, + 0.009315486066043377, + 0.016638675704598427, + 0.01164536364376545, + 0.06548083573579788, + 0.009809883311390877, + 0.0058131590485572815, + 0.009068752638995647, + 0.004340842366218567, + 0.006737834308296442, + 0.0032282997854053974, + 0.0047016148455441, + 0.004264057148247957, + 0.014147672802209854, + 0.015382454730570316, + 0.01485936064273119, + 0.010547615587711334, + 0.014507392421364784, + 0.007665074430406094, + 0.015212488360702991, + 0.011867301538586617, + 0.02050863392651081, + 0.01648777350783348, + 0.017744002863764763, + 0.017399610951542854, + 0.027600670233368874, + 0.02210448868572712, + 0.018395328894257545, + 0.015566740185022354, + 0.010860197246074677, + 0.012701103463768959, + 0.00838412344455719, + 0.007512002717703581, + 0.009362338110804558, + 0.008736337535083294, + 0.0067181396298110485, + 0.011322947219014168, + 0.006396310403943062, + 0.004355457611382008, + 0.0055204420350492, + 0.004118828568607569, + 0.007067212834954262, + 0.006068017333745956, + 0.005974241066724062, + 0.0059797437861561775, + 0.004855388775467873, + 0.005526786204427481, + 0.0028881544712930918, + 0.0036688554100692272, + 0.009298271499574184, + 0.006428288295865059, + 0.004444593098014593, + 0.004604537971317768, + 0.006028789561241865, + 0.008783221244812012, + 0.00463428720831871, + 0.004658326972275972, + 0.006576668005436659, + 0.006867308635264635, + 0.005108876619488001, + 0.008351916447281837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.5334846377372742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036281016655266285 + ], + "opposite_logprob_tokens": [ + -5.628628253936768 + ], + "logprob_reward": -0.0036281016655266285, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036281016655266285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10314325243234634, + 0.08659350872039795, + 0.23315544426441193, + 0.0811302661895752, + 0.12329984456300735, + 0.16302958130836487, + 0.05579724162817001, + 0.04811149090528488, + 0.043589260429143906, + 0.06215009465813637 + ], + "credit_map": [ + 0.018613852560520172, + 0.011972302570939064, + 0.01437798235565424, + 0.012546981684863567, + 0.014758164063096046, + 0.009600094519555569, + 0.009745358489453793, + 0.011528517119586468, + 0.009394727647304535, + 0.011617420241236687, + 0.01180417463183403, + 0.010020888410508633, + 0.013333824463188648, + 0.008688842877745628, + 0.010957533493638039, + 0.0107760950922966, + 0.05173049122095108, + 0.0061512053944170475, + 0.09080921858549118, + 0.009768874384462833, + 0.00723223714157939, + 0.011850259266793728, + 0.009233067743480206, + 0.04638011008501053, + 0.01674591191112995, + 0.007984576746821404, + 0.016758428886532784, + 0.006625543814152479, + 0.009443451650440693, + 0.007004438433796167, + 0.007520594634115696, + 0.009047325700521469, + 0.02619534358382225, + 0.014256964437663555, + 0.023784205317497253, + 0.009144669398665428, + 0.015477186068892479, + 0.00871503259986639, + 0.015003601089119911, + 0.010722844861447811, + 0.020881744101643562, + 0.013321489095687866, + 0.02119368501007557, + 0.019441910088062286, + 0.02804393693804741, + 0.02503550983965397, + 0.019985975697636604, + 0.015125320293009281, + 0.007057067938148975, + 0.008017176762223244, + 0.00716774258762598, + 0.0061525506898760796, + 0.006997351534664631, + 0.006823049392551184, + 0.006454654037952423, + 0.007127649150788784, + 0.006265414413064718, + 0.007386546116322279, + 0.004814751911908388, + 0.0045263818465173244, + 0.006739578675478697, + 0.006008871830999851, + 0.005042303819209337, + 0.00732764508575201, + 0.005114582367241383, + 0.005555756855756044, + 0.00431578466668725, + 0.003968400880694389, + 0.008257903158664703, + 0.003947856370359659, + 0.004245867487043142, + 0.00818310771137476, + 0.006656710058450699, + 0.006778121925890446, + 0.004837976768612862, + 0.004806922283023596, + 0.008340549655258656, + 0.010549883358180523, + 0.006679512094706297, + 0.013500419445335865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.0036281016655266285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539120391011238 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012539120391011238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539120391011238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08917507529258728, + 0.040253009647130966, + 0.27370011806488037, + 0.06405002623796463, + 0.19070245325565338, + 0.1250845044851303, + 0.05352719873189926, + 0.07246587425470352, + 0.049492038786411285, + 0.041549645364284515 + ], + "credit_map": [ + 0.011634796857833862, + 0.011152329854667187, + 0.015097610652446747, + 0.01197640597820282, + 0.012211539782583714, + 0.00741182966157794, + 0.01031939871609211, + 0.009371164254844189, + 0.009594814851880074, + 0.0030706939287483692, + 0.010145974345505238, + 0.003090444952249527, + 0.0037542067002505064, + 0.003727257950231433, + 0.0026458550710231066, + 0.004223762545734644, + 0.06986052542924881, + 0.005202234257012606, + 0.11046584695577621, + 0.0107950484380126, + 0.007496002595871687, + 0.013903066515922546, + 0.009300624951720238, + 0.04667676240205765, + 0.014773738570511341, + 0.007508150767534971, + 0.01415158249437809, + 0.004813537001609802, + 0.006968972738832235, + 0.005267287604510784, + 0.005217922385782003, + 0.0053488402627408504, + 0.05325475335121155, + 0.020236831158399582, + 0.04831387847661972, + 0.011545666493475437, + 0.014962216839194298, + 0.010215943679213524, + 0.01805897429585457, + 0.01411418803036213, + 0.01687602885067463, + 0.014173329807817936, + 0.010794628411531448, + 0.013228447176516056, + 0.022537220269441605, + 0.0166793093085289, + 0.014734945259988308, + 0.01606059819459915, + 0.006747802719473839, + 0.006403444334864616, + 0.007215714547783136, + 0.0057404013350605965, + 0.006677865516394377, + 0.006376100238412619, + 0.005842114333063364, + 0.00852375477552414, + 0.00835247989743948, + 0.007656308822333813, + 0.0071759410202503204, + 0.007469562813639641, + 0.010243994183838367, + 0.009017993696033955, + 0.011094285175204277, + 0.011455308645963669, + 0.004869901109486818, + 0.006176725495606661, + 0.003624314209446311, + 0.004307189490646124, + 0.0075204758904874325, + 0.007475737482309341, + 0.004812108818441629, + 0.01070559024810791, + 0.004886929877102375, + 0.0082371411845088, + 0.003908120561391115, + 0.004194030072540045, + 0.0057570915669202805, + 0.005263824015855789, + 0.0045524523593485355, + 0.004750056192278862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.012539120391011238 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026333190500736237 + ], + "opposite_logprob_tokens": [ + -3.6513330936431885 + ], + "logprob_reward": -0.026333190500736237, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026333190500736237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10639341175556183, + 0.02409324236214161, + 0.2043868899345398, + 0.05836833640933037, + 0.11261508613824844, + 0.11828242987394333, + 0.0488315187394619, + 0.10276774317026138, + 0.0428374707698822, + 0.18142388761043549 + ], + "credit_map": [ + 0.020313149318099022, + 0.014239577576518059, + 0.018606308847665787, + 0.012597858905792236, + 0.012841549701988697, + 0.009243697859346867, + 0.008970435708761215, + 0.009580831043422222, + 0.0031304177828133106, + 0.0025910111144185066, + 0.003903997130692005, + 0.002501110313460231, + 0.003392226994037628, + 0.0023070613387972116, + 0.002930673537775874, + 0.0033367446158081293, + 0.05005675554275513, + 0.005191768519580364, + 0.08159574121236801, + 0.008083329536020756, + 0.005715702194720507, + 0.010284118354320526, + 0.007032697554677725, + 0.03642678260803223, + 0.01541784405708313, + 0.004648174624890089, + 0.015665855258703232, + 0.004275512415915728, + 0.005821213126182556, + 0.0038799592293798923, + 0.004708524793386459, + 0.003951251041144133, + 0.030833646655082703, + 0.01270131766796112, + 0.027278456836938858, + 0.007655437104403973, + 0.009951118379831314, + 0.006803498603403568, + 0.00937290582805872, + 0.008018704131245613, + 0.014117781072854996, + 0.018623748794198036, + 0.01479469146579504, + 0.015087027102708817, + 0.01669897697865963, + 0.014491467736661434, + 0.012425411492586136, + 0.012043330818414688, + 0.005356272682547569, + 0.005861470941454172, + 0.006223233416676521, + 0.005323519930243492, + 0.005912034772336483, + 0.006025814916938543, + 0.00551991630345583, + 0.008609255775809288, + 0.01042280625551939, + 0.022731920704245567, + 0.006941766943782568, + 0.012469355948269367, + 0.014947257936000824, + 0.011946075595915318, + 0.009687582962214947, + 0.013620981015264988, + 0.0057996055111289024, + 0.008683287538588047, + 0.0031350383069366217, + 0.003903395961970091, + 0.006848536431789398, + 0.0052749053575098515, + 0.004114584997296333, + 0.005078114569187164, + 0.02562580816447735, + 0.03007945604622364, + 0.01609502173960209, + 0.015616950578987598, + 0.02375629171729088, + 0.02523946762084961, + 0.01727866940200329, + 0.02773222140967846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.026333190500736237 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01252958457916975 + ], + "opposite_logprob_tokens": [ + -4.387529373168945 + ], + "logprob_reward": -0.01252958457916975, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01252958457916975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09308898448944092, + 0.03432177007198334, + 0.3158852159976959, + 0.04249775409698486, + 0.07427015900611877, + 0.13562160730361938, + 0.05196705088019371, + 0.10852479189634323, + 0.040313705801963806, + 0.10350896418094635 + ], + "credit_map": [ + 0.018996985629200935, + 0.012781122699379921, + 0.015536096878349781, + 0.012029706500470638, + 0.010028593242168427, + 0.0077807046473026276, + 0.007937769405543804, + 0.007998008280992508, + 0.005051793996244669, + 0.003716473700478673, + 0.00579784344881773, + 0.004649731330573559, + 0.005292696412652731, + 0.0030201307963579893, + 0.003301192307844758, + 0.00349190691486001, + 0.08017564564943314, + 0.005683789495378733, + 0.12759412825107574, + 0.012411358766257763, + 0.00839564111083746, + 0.016187023371458054, + 0.010407905094325542, + 0.0550297275185585, + 0.007156586740165949, + 0.005079870577901602, + 0.005787424277514219, + 0.004990567918866873, + 0.0061372146010398865, + 0.0037390044890344143, + 0.0051110549829900265, + 0.004496030043810606, + 0.016002919524908066, + 0.00976326409727335, + 0.014657266438007355, + 0.006500056944787502, + 0.007207197602838278, + 0.004009881056845188, + 0.009332720190286636, + 0.006796852685511112, + 0.017542922869324684, + 0.012320445850491524, + 0.014745600521564484, + 0.013736382126808167, + 0.028489597141742706, + 0.021170303225517273, + 0.016299035400152206, + 0.011317316442728043, + 0.00894335936754942, + 0.0069613978266716, + 0.006182251963764429, + 0.004876452963799238, + 0.0062862178310751915, + 0.006194328889250755, + 0.0054048397578299046, + 0.007118201814591885, + 0.01956741325557232, + 0.008580870926380157, + 0.011036605574190617, + 0.00905343983322382, + 0.017466919496655464, + 0.01606825925409794, + 0.013958715833723545, + 0.012792563997209072, + 0.0037601073272526264, + 0.004535892978310585, + 0.0023399146739393473, + 0.0028254161588847637, + 0.004661324433982372, + 0.00649245223030448, + 0.0038830055855214596, + 0.011815592646598816, + 0.008512914180755615, + 0.015231737866997719, + 0.007214407902210951, + 0.012983073480427265, + 0.01309195626527071, + 0.013230523094534874, + 0.008126683533191681, + 0.02511766366660595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.01252958457916975 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, as if filmed from a deep-sea submersible, captures a giant octopus as it emerges from the inky abyss. The creature floats majestically and deliberately in the murky waters, its body a marvel of ethereal biology. Its translucent, membranous flesh, shades of deep indigo and ghostly white, undulates with a hypnotic slowness, truly resembling a wisp of smoke in zero gravity. A cool internal bioluminescence pulses rhythmically from within its cloak and along the delicate webs between its arms, casting a soft, ever-shifting glow that illuminates the intricate patterns on its skin. Its enormous, intelligent silver eyes gaze forward into the void, while its tentacles trail behind with languid grace. The only light in the devastating darkness comes from the octopus itself, capturing countless grains of marine snow that drift across the camera like a slow-motion blizzard, creating an atmosphere of mystery and awe, witnessing a primordial, otherworldly being in its profoundly silent realm.", + "seed": 21208, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_15/index_923_seed_21208_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.048174799606204036, + "vlm_logprob": -0.048174799606204036, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.048174799606204036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14073678851127625 + ], + "opposite_logprob_tokens": [ + -2.1407368183135986 + ], + "logprob_reward": -0.14073678851127625, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14073678851127625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24360938370227814, + 0.021275928243994713, + 0.07962938398122787, + 0.0483623668551445, + 0.17393246293067932, + 0.06755375117063522, + 0.31222912669181824, + 0.017549565061926842, + 0.018740979954600334, + 0.017116999253630638 + ], + "credit_map": [ + 0.03392346203327179, + 0.03472211956977844, + 0.04868035018444061, + 0.011360266245901585, + 0.009290908463299274, + 0.06860142201185226, + 0.018207235261797905, + 0.01882363110780716, + 0.0020579739939421415, + 0.0023624789901077747, + 0.0030349469743669033, + 0.0030292714945971966, + 0.002516977721825242, + 0.003585460362955928, + 0.0023191371001303196, + 0.0023696848656982183, + 0.005860236939042807, + 0.007591675966978073, + 0.027409354224801064, + 0.004133503418415785, + 0.004223764408379793, + 0.023962918668985367, + 0.0035271577071398497, + 0.002920778701081872, + 0.008818031288683414, + 0.006367562338709831, + 0.0063828411512076855, + 0.004387720953673124, + 0.004471374209970236, + 0.010646470822393894, + 0.0037940056063234806, + 0.0034943653736263514, + 0.030242469161748886, + 0.017142511904239655, + 0.03570566698908806, + 0.007097920402884483, + 0.009347977116703987, + 0.05535874515771866, + 0.005610588937997818, + 0.013426589779555798, + 0.00869262870401144, + 0.008245524019002914, + 0.010401858948171139, + 0.0066547514870762825, + 0.005103815346956253, + 0.01845702901482582, + 0.0037751549389213324, + 0.006222997792065144, + 0.017906298860907555, + 0.022793065756559372, + 0.10999506711959839, + 0.007237737998366356, + 0.008363474160432816, + 0.13267681002616882, + 0.004421598743647337, + 0.008835063315927982, + 0.0015894556418061256, + 0.002223590388894081, + 0.003130655037239194, + 0.0015850861091166735, + 0.0019653758499771357, + 0.003904792247340083, + 0.0016966325929388404, + 0.0014539782423526049, + 0.0022106729447841644, + 0.0025550907012075186, + 0.0030539478175342083, + 0.0015885294415056705, + 0.0021899219136685133, + 0.003698145505040884, + 0.001749073970131576, + 0.001695598941296339, + 0.002769563579931855, + 0.002297590486705303, + 0.002594882156699896, + 0.0011117230169475079, + 0.0016557881608605385, + 0.0037955602165311575, + 0.001281463773921132, + 0.0016104289097711444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.14073678851127625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590832456946373 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.020590832456946373, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590832456946373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2654178738594055, + 0.05652258172631264, + 0.05152342468500137, + 0.024222733452916145, + 0.03955903649330139, + 0.015166951343417168, + 0.1541350781917572, + 0.31250351667404175, + 0.055390819907188416, + 0.025558041408658028 + ], + "credit_map": [ + 0.03173677250742912, + 0.03949388116598129, + 0.05001773685216904, + 0.011640702374279499, + 0.013608911074697971, + 0.08561813086271286, + 0.014568065293133259, + 0.018733670935034752, + 0.004212740808725357, + 0.007359194569289684, + 0.015487795695662498, + 0.002631275448948145, + 0.0020524768624454737, + 0.019884761422872543, + 0.002356375101953745, + 0.00253796367906034, + 0.0025900150649249554, + 0.004804336000233889, + 0.016760332509875298, + 0.002472162712365389, + 0.0021179281175136566, + 0.019293705001473427, + 0.0016204440034925938, + 0.0018644985975697637, + 0.00491709727793932, + 0.002661400707438588, + 0.00343267060816288, + 0.0015947407810017467, + 0.0016248688334599137, + 0.006755273323506117, + 0.0014457132201641798, + 0.001790969748981297, + 0.006493110675364733, + 0.004070246126502752, + 0.007964644581079483, + 0.0016127243870869279, + 0.002222105860710144, + 0.012407101690769196, + 0.0015320798847824335, + 0.0032570241019129753, + 0.0020847280975431204, + 0.001960241235792637, + 0.0023890151642262936, + 0.0017067649168893695, + 0.0012424681335687637, + 0.003350680461153388, + 0.0009363105054944754, + 0.0014967421302571893, + 0.008312584832310677, + 0.011775469407439232, + 0.05282936990261078, + 0.0038683025632053614, + 0.005289894994348288, + 0.0644311010837555, + 0.0032328986562788486, + 0.004395459312945604, + 0.019223060458898544, + 0.01839008927345276, + 0.11562541127204895, + 0.008529559709131718, + 0.007098212372511625, + 0.12825943529605865, + 0.0059810820966959, + 0.009396657347679138, + 0.004217926412820816, + 0.005555673968046904, + 0.015633929520845413, + 0.003865650622174144, + 0.005163936410099268, + 0.013042252510786057, + 0.004293712321668863, + 0.003617737675085664, + 0.002908173715695739, + 0.0033672216814011335, + 0.003666411153972149, + 0.0022299166303128004, + 0.003374305088073015, + 0.0038419298361986876, + 0.003025819780305028, + 0.003144264919683337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.020590832456946373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699020743370056 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.033699020743370056, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699020743370056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10445243865251541, + 0.1502077430486679, + 0.02206389047205448, + 0.042777594178915024, + 0.18197645246982574, + 0.06493069231510162, + 0.11219338327646255, + 0.2676864564418793, + 0.02676607482135296, + 0.02694527804851532 + ], + "credit_map": [ + 0.01350521482527256, + 0.015702415257692337, + 0.014273063279688358, + 0.007514273747801781, + 0.008574404753744602, + 0.029361965134739876, + 0.007641937583684921, + 0.007879157550632954, + 0.010681240819394588, + 0.016648583114147186, + 0.05012790858745575, + 0.005996095947921276, + 0.003719067433848977, + 0.05475581809878349, + 0.003958509303629398, + 0.004320535343140364, + 0.0020004836842417717, + 0.0037009038496762514, + 0.0050314078107476234, + 0.0019101077923551202, + 0.0024301367811858654, + 0.002813570434227586, + 0.0020866678096354008, + 0.0020906112622469664, + 0.008637660183012486, + 0.00474324868991971, + 0.006018295884132385, + 0.002487714169546962, + 0.0025910737458616495, + 0.01301543414592743, + 0.002149284118786454, + 0.0031348825432360172, + 0.03159227594733238, + 0.01818946562707424, + 0.03567209839820862, + 0.007908494211733341, + 0.00973574724048376, + 0.058040205389261246, + 0.0061071342788636684, + 0.014731018804013729, + 0.009932790882885456, + 0.008005082607269287, + 0.01039956510066986, + 0.004803403280675411, + 0.005206183064728975, + 0.017459936439990997, + 0.0037652868777513504, + 0.005358445458114147, + 0.007071651518344879, + 0.007551145739853382, + 0.03979657590389252, + 0.003435444086790085, + 0.003641351591795683, + 0.045023541897535324, + 0.002225260017439723, + 0.0034484167117625475, + 0.01643363945186138, + 0.0156282689422369, + 0.09843546152114868, + 0.007371940184384584, + 0.007150711491703987, + 0.10893304646015167, + 0.005459177307784557, + 0.008274216204881668, + 0.0024365221615880728, + 0.003893716959282756, + 0.0036978113930672407, + 0.0018164487555623055, + 0.0040449704974889755, + 0.006393189076334238, + 0.0026729940436780453, + 0.0018104201881214976, + 0.004676668904721737, + 0.003627984318882227, + 0.003977099433541298, + 0.0015479808207601309, + 0.0027061065193265676, + 0.005949620623141527, + 0.002010691910982132, + 0.002449125749990344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.033699020743370056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978825569152832 + ], + "opposite_logprob_tokens": [ + -3.5297882556915283 + ], + "logprob_reward": -0.02978825569152832, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978825569152832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055601876229047775, + 0.0454644039273262, + 0.10037443041801453, + 0.02444867603480816, + 0.05412023141980171, + 0.29831740260124207, + 0.2976398169994354, + 0.03608477860689163, + 0.04437118023633957, + 0.04357720538973808 + ], + "credit_map": [ + 0.007559026591479778, + 0.008320676162838936, + 0.012490086257457733, + 0.004892277996987104, + 0.00406784238293767, + 0.010595896281301975, + 0.0038147352170199156, + 0.0038613337092101574, + 0.0028710889164358377, + 0.004584809765219688, + 0.01178797148168087, + 0.004051832482218742, + 0.003476958256214857, + 0.013170270249247551, + 0.002858066698536277, + 0.002663404680788517, + 0.004849891643971205, + 0.008834633976221085, + 0.03218957409262657, + 0.00493952352553606, + 0.0055719465017318726, + 0.03526057302951813, + 0.004648622591048479, + 0.004079666454344988, + 0.0034505792427808046, + 0.0025579812936484814, + 0.0036341689992696047, + 0.0019820856396108866, + 0.0029351371340453625, + 0.006045469548553228, + 0.002009605523198843, + 0.0018336507491767406, + 0.008846611715853214, + 0.005265144631266594, + 0.009466860443353653, + 0.003217158140614629, + 0.00476702768355608, + 0.014685858972370625, + 0.002436920301988721, + 0.005434650927782059, + 0.04219016805291176, + 0.03561954200267792, + 0.04716683179140091, + 0.025902172550559044, + 0.024368956685066223, + 0.08127288520336151, + 0.01683802902698517, + 0.024958821013569832, + 0.01726195216178894, + 0.020145662128925323, + 0.0998438224196434, + 0.011186184361577034, + 0.013364146463572979, + 0.11827704310417175, + 0.008452082984149456, + 0.00910891406238079, + 0.0036098703276365995, + 0.0033921615686267614, + 0.008021577261388302, + 0.0022694135550409555, + 0.003186357906088233, + 0.010697501711547375, + 0.0017987183528020978, + 0.0031091803684830666, + 0.005689643789082766, + 0.004139456432312727, + 0.012015298940241337, + 0.00258758757263422, + 0.0048874556086957455, + 0.009222323074936867, + 0.002747325226664543, + 0.0030820874962955713, + 0.003115565050393343, + 0.00524120545014739, + 0.00738746440038085, + 0.0034589481074362993, + 0.006121200975030661, + 0.011054085567593575, + 0.0033728708513081074, + 0.0038258677814155817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.02978825569152832 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605910062789917 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.01605910062789917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605910062789917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02616981416940689, + 0.19297249615192413, + 0.008506729267537594, + 0.02502400241792202, + 0.039492588490247726, + 0.16313619911670685, + 0.05518696457147598, + 0.44621512293815613, + 0.010535579174757004, + 0.03276049718260765 + ], + "credit_map": [ + 0.005323461256921291, + 0.004702861420810223, + 0.004200172144919634, + 0.0018248760607093573, + 0.0018183111678808928, + 0.003988722339272499, + 0.0023471107706427574, + 0.0019642976112663746, + 0.01646146923303604, + 0.023894593119621277, + 0.06099658086895943, + 0.007385893724858761, + 0.004938927944749594, + 0.06887110322713852, + 0.005001105833798647, + 0.005422818008810282, + 0.0006050058291293681, + 0.0012683480745181441, + 0.00243568979203701, + 0.0007016081362962723, + 0.0006215897155925632, + 0.0014116266975179315, + 0.0007066286634653807, + 0.0007562316604889929, + 0.005145602393895388, + 0.003154603997245431, + 0.003371452447026968, + 0.0016548390267416835, + 0.0016686412272974849, + 0.0067655486054718494, + 0.0015356977237388492, + 0.0017276154831051826, + 0.005966559052467346, + 0.004441466182470322, + 0.007205755449831486, + 0.0023212702944874763, + 0.002345114015042782, + 0.011888657696545124, + 0.0016364026814699173, + 0.0036873642820864916, + 0.02424374409019947, + 0.019110502675175667, + 0.0267178975045681, + 0.013368602842092514, + 0.01333429291844368, + 0.04349798336625099, + 0.00903578195720911, + 0.013827393762767315, + 0.003012408036738634, + 0.003956597764045, + 0.019603073596954346, + 0.0014222582103684545, + 0.001427736016921699, + 0.02339557558298111, + 0.000864915200509131, + 0.0015043948078528047, + 0.027746716514229774, + 0.026673898100852966, + 0.16306346654891968, + 0.011891568079590797, + 0.011033655144274235, + 0.18337444961071014, + 0.0088652353733778, + 0.01356610655784607, + 0.0009654271998442709, + 0.0009911054512485862, + 0.003057206282392144, + 0.0006593599100597203, + 0.0007087401463650167, + 0.0028829004149883986, + 0.0006604391965083778, + 0.0006104002241045237, + 0.006260837893933058, + 0.004012401215732098, + 0.00584891764447093, + 0.0019870074465870857, + 0.00236439798027277, + 0.007653702050447464, + 0.0017348474357277155, + 0.002898384351283312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.01605910062789917 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, as if filmed from a deep-sea submersible, captures a giant octopus as it emerges from the inky abyss. The creature floats majestically and deliberately in the murky waters, its body a marvel of ethereal biology. Its translucent, membranous flesh, shades of deep indigo and ghostly white, undulates with a hypnotic slowness, truly resembling a wisp of smoke in zero gravity. A cool internal bioluminescence pulses rhythmically from within its cloak and along the delicate webs between its arms, casting a soft, ever-shifting glow that illuminates the intricate patterns on its skin. Its enormous, intelligent silver eyes gaze forward into the void, while its tentacles trail behind with languid grace. The only light in the devastating darkness comes from the octopus itself, capturing countless grains of marine snow that drift across the camera like a slow-motion blizzard, creating an atmosphere of mystery and awe, witnessing a primordial, otherworldly being in its profoundly silent realm.", + "seed": 21209, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_15/index_923_seed_21209_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2331891294568777, + "vlm_logprob": -0.2331891294568777, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.2331891294568777, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0632476806640625 + ], + "opposite_logprob_tokens": [ + -0.4382476806640625 + ], + "logprob_reward": -1.0632476806640625, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0632476806640625, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.7334963083267212, + 0.021691204980015755, + 0.04060607776045799, + 0.12373035401105881, + 0.02717120572924614, + 0.01061394065618515, + 0.011431703343987465, + 0.00869354885071516, + 0.009700650349259377, + 0.012865016236901283 + ], + "credit_map": [ + 0.05474814772605896, + 0.016611186787486076, + 0.013967478647828102, + 0.013731325045228004, + 0.02685234323143959, + 0.2953128516674042, + 0.015534160658717155, + 0.29673877358436584, + 0.0026313993148505688, + 0.0033955001272261143, + 0.0026971246115863323, + 0.0015891699586063623, + 0.0024571747053414583, + 0.0032298355363309383, + 0.0025168098509311676, + 0.0031741883140057325, + 0.006470346823334694, + 0.006499267648905516, + 0.004440657328814268, + 0.003170221811160445, + 0.004590935539454222, + 0.0059602572582662106, + 0.0033106717746704817, + 0.006163712590932846, + 0.014959028922021389, + 0.007176171522587538, + 0.009561402723193169, + 0.009057643823325634, + 0.009378178045153618, + 0.028030099347233772, + 0.007729640696197748, + 0.0378381684422493, + 0.0027521722950041294, + 0.0025051634293049574, + 0.00202256185002625, + 0.0033233221620321274, + 0.0027114273980259895, + 0.004560434725135565, + 0.0029021943919360638, + 0.0063939266838133335, + 0.0011408425634726882, + 0.0017220658482983708, + 0.0010578178334981203, + 0.0012413081713020802, + 0.0017848435090854764, + 0.0015797988744452596, + 0.0011850740993395448, + 0.0009021891164593399, + 0.0010819132439792156, + 0.001782611245289445, + 0.001369479694403708, + 0.0016957097686827183, + 0.0012807396706193686, + 0.0017204767791554332, + 0.0013691256754100323, + 0.0011316461022943258, + 0.0009656588663347065, + 0.001552178175188601, + 0.0010876348242163658, + 0.0008343226509168744, + 0.0011575892567634583, + 0.0012536832364276052, + 0.0007380250608548522, + 0.001104456139728427, + 0.0007947750273160636, + 0.0007035068119876087, + 0.002197638154029846, + 0.0010598489316180348, + 0.0008096619276329875, + 0.002227689838036895, + 0.0010310775833204389, + 0.0008764515514485538, + 0.0011553473304957151, + 0.000823162030428648, + 0.002854471793398261, + 0.0012627742253243923, + 0.0009825783781707287, + 0.003105925628915429, + 0.0010203560814261436, + 0.0016603992553427815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -1.0632476806640625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023317772895097733 + ], + "opposite_logprob_tokens": [ + -3.773317813873291 + ], + "logprob_reward": -0.023317772895097733, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023317772895097733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9196898341178894, + 0.007373603992164135, + 0.014345505274832249, + 0.010170381516218185, + 0.0035749853122979403, + 0.006829377729445696, + 0.009552078321576118, + 0.009059647098183632, + 0.013509009033441544, + 0.005895583890378475 + ], + "credit_map": [ + 0.06658719480037689, + 0.020828688517212868, + 0.0173807255923748, + 0.017419034615159035, + 0.033352091908454895, + 0.3722737431526184, + 0.0181184783577919, + 0.3737298548221588, + 0.0009938509901985526, + 0.0009826908353716135, + 0.001007511862553656, + 0.0005668457015417516, + 0.0009274255717173219, + 0.001015769550576806, + 0.0009269214933738112, + 0.0009525880450382829, + 0.002492668107151985, + 0.002321691019460559, + 0.0016448815586045384, + 0.001384092727676034, + 0.0018696573097258806, + 0.0021393559873104095, + 0.001204514759592712, + 0.0012886441545560956, + 0.001213120762258768, + 0.0018873108783736825, + 0.0008587631164118648, + 0.0013599920785054564, + 0.001487037749029696, + 0.001395300030708313, + 0.0008850544109009206, + 0.0010838024318218231, + 0.00032961220131255686, + 0.0004906447720713913, + 0.00042255406151525676, + 0.00038274735561572015, + 0.00047696754336357117, + 0.000562652014195919, + 0.0002320691419299692, + 0.0006777381058782339, + 0.000662192702293396, + 0.0010901816422119737, + 0.0007746367482468486, + 0.0007823696359992027, + 0.0009469789220020175, + 0.0010075555182993412, + 0.0006877899868413806, + 0.0008776726317591965, + 0.0009780196705833077, + 0.0011080214753746986, + 0.0013194893253967166, + 0.0008682945626787841, + 0.0012518654111772776, + 0.0017400860087946057, + 0.0012067323550581932, + 0.0010795696871355176, + 0.0008988587651401758, + 0.0013793661491945386, + 0.00153952999971807, + 0.0009712176397442818, + 0.0009323076228611171, + 0.0018728843424469233, + 0.0006774283829145133, + 0.000788053497672081, + 0.001183937769383192, + 0.0010105186374858022, + 0.0028096456080675125, + 0.0015856159152463078, + 0.0011785626411437988, + 0.0031551120337098837, + 0.0014019693480804563, + 0.0011836463818326592, + 0.0007291018264368176, + 0.00047184157301671803, + 0.0008660402963869274, + 0.0008383747772313654, + 0.0008233430562540889, + 0.000930223090108484, + 0.00048569857608526945, + 0.0007509603747166693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.023317772895097733 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254948042333126 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.01254948042333126, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254948042333126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6183573603630066, + 0.012116854079067707, + 0.010819150134921074, + 0.24782100319862366, + 0.011442056857049465, + 0.006209284067153931, + 0.015189885161817074, + 0.006288817152380943, + 0.00695634912699461, + 0.0647992491722107 + ], + "credit_map": [ + 0.04548975080251694, + 0.014198795892298222, + 0.012007977813482285, + 0.011878220364451408, + 0.02267962321639061, + 0.2499828338623047, + 0.01159245241433382, + 0.2505277097225189, + 0.0013264799490571022, + 0.0020301053300499916, + 0.0019405591301620007, + 0.0010969664435833693, + 0.0016363522736355662, + 0.0015725777484476566, + 0.0011345177190378308, + 0.0013792954850941896, + 0.0016667181625962257, + 0.0017420294461771846, + 0.0012440579012036324, + 0.0014117661630734801, + 0.0012086436618119478, + 0.001396241132169962, + 0.0008928586612455547, + 0.001256835414096713, + 0.0331270694732666, + 0.01525136362761259, + 0.01944909058511257, + 0.018106207251548767, + 0.018241196870803833, + 0.05527263507246971, + 0.01448266301304102, + 0.07389078289270401, + 0.0014918313827365637, + 0.0010424137581139803, + 0.0013632416957989335, + 0.001306322985328734, + 0.001274312031455338, + 0.0018222855869680643, + 0.0009627655381336808, + 0.002178882947191596, + 0.0007230468327179551, + 0.0009234142489731312, + 0.0008052149787545204, + 0.0005029714666306973, + 0.0010162928374484181, + 0.0009250434231944382, + 0.0005964929587207735, + 0.0007168074371293187, + 0.002049605129286647, + 0.0021175649017095566, + 0.001617933390662074, + 0.0017539348918944597, + 0.0020739976316690445, + 0.0022452580742537975, + 0.001621605595573783, + 0.0017099868273362517, + 0.0007258225814439356, + 0.0009939665906131268, + 0.0007569736917503178, + 0.0007407989469356835, + 0.0008392651216126978, + 0.000954432412981987, + 0.0005435527418740094, + 0.0007340048323385417, + 0.0006500890594907105, + 0.0007328623905777931, + 0.0009044281323440373, + 0.0010920583736151457, + 0.0010625917930155993, + 0.0010820628376677632, + 0.0008010963210836053, + 0.0006311610341072083, + 0.005366693716496229, + 0.0033832297194749117, + 0.013814005069434643, + 0.0055906279012560844, + 0.006082308012992144, + 0.015501619316637516, + 0.004607588052749634, + 0.010453175753355026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.01254948042333126 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865148663520813 + ], + "opposite_logprob_tokens": [ + -3.0486514568328857 + ], + "logprob_reward": -0.04865148663520813, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865148663520813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6667479872703552, + 0.022255914285779, + 0.024251863360404968, + 0.12335695326328278, + 0.009545352309942245, + 0.017145469784736633, + 0.011120759882032871, + 0.00569088663905859, + 0.019181780517101288, + 0.10070304572582245 + ], + "credit_map": [ + 0.04808606952428818, + 0.015904966741800308, + 0.013592155650258064, + 0.013889303430914879, + 0.025172337889671326, + 0.2672615051269531, + 0.014130815863609314, + 0.26871079206466675, + 0.0027163077611476183, + 0.003229209454730153, + 0.003722545923665166, + 0.0014145106542855501, + 0.003110497957095504, + 0.0029866578988730907, + 0.001933666761033237, + 0.0031425172928720713, + 0.004543511662632227, + 0.003542687278240919, + 0.0020581050775945187, + 0.0019313606899231672, + 0.00245733093470335, + 0.0032596001401543617, + 0.0031386897899210453, + 0.0033205777872353792, + 0.016341889277100563, + 0.010145763866603374, + 0.011109361425042152, + 0.010584214702248573, + 0.010075556114315987, + 0.024878710508346558, + 0.006856105290353298, + 0.03336535394191742, + 0.0009391072671860456, + 0.0012154611758887768, + 0.0009125059004873037, + 0.000928869063500315, + 0.0011653057299554348, + 0.0015760170063003898, + 0.0008193371468223631, + 0.001988748786970973, + 0.0018008589977398515, + 0.002437632530927658, + 0.0019090690184384584, + 0.001588386483490467, + 0.0030766704585403204, + 0.002641364699229598, + 0.001565287820994854, + 0.0021262008231133223, + 0.0013874671421945095, + 0.0016623249975964427, + 0.0014724461361765862, + 0.000878774153534323, + 0.0015814871294423938, + 0.0020396574400365353, + 0.0011379781644791365, + 0.0009606246603652835, + 0.0005427891155704856, + 0.0009080250165425241, + 0.0008123121806420386, + 0.0005418850341811776, + 0.0006130729452706873, + 0.000865391397383064, + 0.000588291441090405, + 0.0008191195665858686, + 0.002042195526883006, + 0.0016899197362363338, + 0.0035111848264932632, + 0.0023682464379817247, + 0.001960649387910962, + 0.004053875803947449, + 0.0016636336222290993, + 0.0018920740112662315, + 0.007820023223757744, + 0.004681738559156656, + 0.022477414458990097, + 0.007901311852037907, + 0.009022271260619164, + 0.025434384122490883, + 0.00751270167529583, + 0.015853192657232285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.04865148663520813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817922666668892 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.01817922666668892, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817922666668892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4361031949520111, + 0.008658853359520435, + 0.011948935687541962, + 0.27075856924057007, + 0.008587471209466457, + 0.005937095265835524, + 0.004076892510056496, + 0.00313595705665648, + 0.011880476027727127, + 0.23891262710094452 + ], + "credit_map": [ + 0.03138114884495735, + 0.009759156964719296, + 0.008298994973301888, + 0.008001088164746761, + 0.015567717142403126, + 0.17735157907009125, + 0.007811847608536482, + 0.17793159186840057, + 0.001045886194333434, + 0.0009015648975037038, + 0.0011916423682123423, + 0.0006093275733292103, + 0.001085975207388401, + 0.0014340432826429605, + 0.0008245432982221246, + 0.0015658698976039886, + 0.0014416860649362206, + 0.0022058489266783, + 0.00127987505402416, + 0.0008108420297503471, + 0.0013702495489269495, + 0.0018636456225067377, + 0.0009243832319043577, + 0.0020524042192846537, + 0.03530818596482277, + 0.016953453421592712, + 0.021754052489995956, + 0.018877940252423286, + 0.019877461716532707, + 0.060525983572006226, + 0.015557484701275826, + 0.08190397173166275, + 0.0011016877833753824, + 0.0008885858696885407, + 0.0009882185840979218, + 0.001133404322899878, + 0.0007956313784234226, + 0.0013360683806240559, + 0.0008215572452172637, + 0.0015223167138174176, + 0.0008905736613087356, + 0.0006360230036079884, + 0.00068115652538836, + 0.000594454410020262, + 0.0007565670530311763, + 0.0009833279764279723, + 0.0005137159605510533, + 0.0008812761516310275, + 0.0006855013780295849, + 0.0005122564034536481, + 0.0006145514198578894, + 0.0005018002120777965, + 0.0004905604291707277, + 0.0006404616287909448, + 0.0003244449326302856, + 0.00030731616425327957, + 0.00033665119553916156, + 0.0004561901732813567, + 0.0005036803195253015, + 0.0003095151041634381, + 0.00033996495767496526, + 0.000536670908331871, + 0.0002596606791485101, + 0.0003936233406420797, + 0.001083661220036447, + 0.0009998772293329239, + 0.002423357218503952, + 0.0011918818345293403, + 0.001170594827271998, + 0.0029384407680481672, + 0.001143981353379786, + 0.0009286798303946853, + 0.01858142390847206, + 0.011838534846901894, + 0.05290007218718529, + 0.01936393417418003, + 0.020195327699184418, + 0.0602603480219841, + 0.017406154423952103, + 0.03836679458618164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.01817922666668892 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, as if filmed from a deep-sea submersible, captures a giant octopus as it emerges from the inky abyss. The creature floats majestically and deliberately in the murky waters, its body a marvel of ethereal biology. Its translucent, membranous flesh, shades of deep indigo and ghostly white, undulates with a hypnotic slowness, truly resembling a wisp of smoke in zero gravity. A cool internal bioluminescence pulses rhythmically from within its cloak and along the delicate webs between its arms, casting a soft, ever-shifting glow that illuminates the intricate patterns on its skin. Its enormous, intelligent silver eyes gaze forward into the void, while its tentacles trail behind with languid grace. The only light in the devastating darkness comes from the octopus itself, capturing countless grains of marine snow that drift across the camera like a slow-motion blizzard, creating an atmosphere of mystery and awe, witnessing a primordial, otherworldly being in its profoundly silent realm.", + "seed": 21210, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_15/index_923_seed_21210_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11769797708839178, + "vlm_logprob": -0.11769797708839178, + "vlm_yes_no_margin": 3.1249999940395354, + "vlm_token_credit": -0.11769797708839178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.48976460099220276 + ], + "opposite_logprob_tokens": [ + -0.9897645711898804 + ], + "logprob_reward": -0.48976460099220276, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.48976460099220276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02355838753283024, + 0.03833414614200592, + 0.022787868976593018, + 0.8744478821754456, + 0.0122530497610569, + 0.009869086556136608, + 0.009035556577146053, + 0.004058713559061289, + 0.0035543260164558887, + 0.0021009023766964674 + ], + "credit_map": [ + 0.002898511476814747, + 0.0022547338157892227, + 0.003669044701382518, + 0.0018031545914709568, + 0.002004517475143075, + 0.003770815674215555, + 0.0028378234710544348, + 0.004319786559790373, + 0.004218601621687412, + 0.003934601787477732, + 0.00589672289788723, + 0.0018281214870512486, + 0.0033144555054605007, + 0.0054108453914523125, + 0.005373224150389433, + 0.008357572369277477, + 0.002184112323448062, + 0.002674151211977005, + 0.003938406240195036, + 0.0020472537726163864, + 0.002108016749843955, + 0.004217713139951229, + 0.0031928387470543385, + 0.0024253767915070057, + 0.027998920530080795, + 0.008117753081023693, + 0.05246438831090927, + 0.007098513189703226, + 0.011998484842479229, + 0.39973583817481995, + 0.01312281284481287, + 0.35391125082969666, + 0.0013922128127887845, + 0.002113990718498826, + 0.0016206738073378801, + 0.0014537908136844635, + 0.001443652086891234, + 0.0016895204316824675, + 0.0012428475311025977, + 0.0012963623739778996, + 0.001183351967483759, + 0.0007181082619354129, + 0.0017363856313750148, + 0.0006169974221847951, + 0.0009831058559939265, + 0.0027670881245285273, + 0.0006714414339512587, + 0.0011926079168915749, + 0.0013773213140666485, + 0.0011568047339096665, + 0.0010793318506330252, + 0.000996827962808311, + 0.000951045542024076, + 0.0017989202169701457, + 0.0008704100036993623, + 0.0008048951858654618, + 0.0006111165857873857, + 0.0005841566016897559, + 0.0005134635139256716, + 0.0005217390134930611, + 0.00039287511026486754, + 0.0004983857506886125, + 0.000565839116461575, + 0.0003711377503350377, + 0.0007720972062088549, + 0.0003848457708954811, + 0.0005506673478521407, + 0.000294979807222262, + 0.00046438685967586935, + 0.0003578297037165612, + 0.00040067563531920314, + 0.0003288437146693468, + 0.00024217132886406034, + 0.0002492313215043396, + 0.00026214690296910703, + 0.00022799741418566555, + 0.00031202854006551206, + 0.0002832495956681669, + 0.00026930245803669095, + 0.0002547749027144164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.48976460099220276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209403656423092 + ], + "opposite_logprob_tokens": [ + -4.264209270477295 + ], + "logprob_reward": -0.014209403656423092, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209403656423092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.214790940284729, + 0.0587754063308239, + 0.027336545288562775, + 0.5965821146965027, + 0.0383438766002655, + 0.009441603906452656, + 0.028368674218654633, + 0.011275319382548332, + 0.008744964376091957, + 0.006340578198432922 + ], + "credit_map": [ + 0.018937312066555023, + 0.008422520942986012, + 0.019785825163125992, + 0.006487212143838406, + 0.00842830166220665, + 0.05473962426185608, + 0.0111874183639884, + 0.08680272102355957, + 0.005494144279509783, + 0.007649495266377926, + 0.009597133845090866, + 0.00393849890679121, + 0.006112500559538603, + 0.00846783071756363, + 0.0068391840904951096, + 0.010676617734134197, + 0.0024946003686636686, + 0.0035333670675754547, + 0.0048987651243805885, + 0.0025557551998645067, + 0.0024598792660981417, + 0.004228207748383284, + 0.004252832382917404, + 0.002913138596341014, + 0.018814049661159515, + 0.009158375672996044, + 0.03503773361444473, + 0.00784210953861475, + 0.01030734647065401, + 0.2652139663696289, + 0.011580605991184711, + 0.23862788081169128, + 0.004079132340848446, + 0.006470039486885071, + 0.004702536389231682, + 0.0050697107799351215, + 0.004219402093440294, + 0.005084410309791565, + 0.004735538270324469, + 0.003983107395470142, + 0.0014392811572179198, + 0.0005388653371483088, + 0.0016868292586877942, + 0.0006394220981746912, + 0.0010262418072670698, + 0.0023383384104818106, + 0.0006002765730954707, + 0.0011723495554178953, + 0.0031700031831860542, + 0.0035698816645890474, + 0.0033974163234233856, + 0.003040215466171503, + 0.00373471318744123, + 0.0054539586417376995, + 0.0034135787282139063, + 0.002588906092569232, + 0.001118067535571754, + 0.0017295337747782469, + 0.0012606927193701267, + 0.0015103446785360575, + 0.0012955614365637302, + 0.0015527090290561318, + 0.0017986268503591418, + 0.0010097833583131433, + 0.001510749920271337, + 0.0013956535840407014, + 0.0012149378890171647, + 0.0008781917858868837, + 0.0009623817168176174, + 0.0009455220424570143, + 0.000957672658842057, + 0.0008798546041361988, + 0.001116120838560164, + 0.0008771967841312289, + 0.0008446653373539448, + 0.0007049484411254525, + 0.0008954743971116841, + 0.0006559541798196733, + 0.0006603033980354667, + 0.0005859149969182909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.014209403656423092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306385666131973 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.04306385666131973, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306385666131973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03945625573396683, + 0.02563394047319889, + 0.07272100448608398, + 0.7958303093910217, + 0.016293786466121674, + 0.013877976685762405, + 0.013679028488695621, + 0.009888555854558945, + 0.00657403701916337, + 0.006045094225555658 + ], + "credit_map": [ + 0.005270051304250956, + 0.0025670649483799934, + 0.004664615262299776, + 0.0027041244320571423, + 0.0031523467041552067, + 0.006878172047436237, + 0.0041120462119579315, + 0.010107839480042458, + 0.002404271624982357, + 0.0028003789484500885, + 0.005237182602286339, + 0.0016094777965918183, + 0.0018117822473868728, + 0.004079864826053381, + 0.0036604739725589752, + 0.0040305121801793575, + 0.006439456716179848, + 0.0036243097856640816, + 0.025368794798851013, + 0.003387377131730318, + 0.0028939354233443737, + 0.016624368727207184, + 0.003383274655789137, + 0.01099949236959219, + 0.026964640244841576, + 0.008277194574475288, + 0.04731076583266258, + 0.00880791712552309, + 0.012507021427154541, + 0.360713928937912, + 0.013696850277483463, + 0.31755200028419495, + 0.0019091240828856826, + 0.0022366016637533903, + 0.002650248119607568, + 0.0015897880075499415, + 0.0019057989120483398, + 0.0025688549503684044, + 0.001527950051240623, + 0.001905421377159655, + 0.0016493325820192695, + 0.0005751296412199736, + 0.002722774865105748, + 0.0007867764215916395, + 0.001045010518282652, + 0.004651147872209549, + 0.000741995929274708, + 0.001705810660496354, + 0.0014399491483345628, + 0.002004284178838134, + 0.0014490196481347084, + 0.0014242971083149314, + 0.0016638675006106496, + 0.0025681289844214916, + 0.0019062196370214224, + 0.001223263912834227, + 0.0018528118962422013, + 0.001115730032324791, + 0.0015148703241720796, + 0.0010846691438928246, + 0.00139804114587605, + 0.0009224063251167536, + 0.001243007485754788, + 0.0007570211892016232, + 0.0015083970502018929, + 0.0007803384214639664, + 0.001018370850943029, + 0.0006128661916591227, + 0.0007417810847982764, + 0.000551470264326781, + 0.0007624522550031543, + 0.000598361250013113, + 0.0006362458225339651, + 0.0011010307352989912, + 0.0006522996700368822, + 0.000718022114597261, + 0.0008031352772377431, + 0.0007029143744148314, + 0.0007702835137024522, + 0.000661163532640785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.04306385666131973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023276543244719505 + ], + "opposite_logprob_tokens": [ + -3.7732765674591064 + ], + "logprob_reward": -0.023276543244719505, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023276543244719505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0026274158153682947, + 0.0022893778514117002, + 0.005422020331025124, + 0.9844197034835815, + 0.001100838533602655, + 0.0007018721080385149, + 0.0018324741395190358, + 0.0003687245771288872, + 0.0008394205942749977, + 0.0003981769841630012 + ], + "credit_map": [ + 0.0003462666063569486, + 0.00021255994215607643, + 0.00035765956272371113, + 0.00018473912496119738, + 0.000244762166403234, + 0.00044586308649741113, + 0.0002765798126347363, + 0.0005589853972196579, + 0.00024144898634403944, + 0.00024953653337433934, + 0.0004801790928468108, + 0.00012214420712552965, + 0.0002356186305405572, + 0.00033538954448886216, + 0.0002891319745685905, + 0.00033592869294807315, + 0.0005099753034301102, + 0.0003091725520789623, + 0.0018248270498588681, + 0.00024535798002034426, + 0.0002177859132643789, + 0.0012244461104273796, + 0.00028221492539159954, + 0.0008082396816462278, + 0.03232467547059059, + 0.007755668833851814, + 0.0588274709880352, + 0.007407994940876961, + 0.01347870659083128, + 0.4517478942871094, + 0.014413131400942802, + 0.3984641432762146, + 0.00011085695587098598, + 0.00014636525884270668, + 0.00014390863361768425, + 0.00010096450569108129, + 0.00016503049118909985, + 0.00017703835328575224, + 0.00014061108231544495, + 0.00011606313637457788, + 7.37390000722371e-05, + 8.539015107089654e-05, + 8.793905726633966e-05, + 6.745947757735848e-05, + 0.00011393591557862237, + 0.00010745257168309763, + 7.217467646114528e-05, + 9.378123650094494e-05, + 0.00031127987313084304, + 0.0002242111077066511, + 0.0002288552059326321, + 0.00019223667914047837, + 0.0001740547886583954, + 0.0003638736379798502, + 0.0001717399136396125, + 0.00016622281691525131, + 3.941185786970891e-05, + 5.6922952353488654e-05, + 4.121170786675066e-05, + 4.233905201544985e-05, + 7.155080675147474e-05, + 4.209557664580643e-05, + 4.587980583892204e-05, + 2.9312794140423648e-05, + 9.163884533336386e-05, + 0.00011379892384866253, + 8.855012129060924e-05, + 8.874337072484195e-05, + 0.00014354524319060147, + 9.087580838240683e-05, + 0.00012008692283416167, + 0.0001021813222905621, + 5.142816007719375e-05, + 5.406419950304553e-05, + 4.407706728670746e-05, + 3.947949880966917e-05, + 5.6447872339049354e-05, + 4.882139910478145e-05, + 4.806617653230205e-05, + 5.5792599596315995e-05 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.023276543244719505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a wide-angle view from a deep-sea submersible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant octopus with translucent, membranous flesh in shades of deep indigo and ghostly white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the octopus exhibit cool internal bioluminescence pulsing from its body and arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous silver eyes visible on the octopus gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a dark deep-sea setting with marine snow drifting across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175480887293816 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.018175480887293816, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018175480887293816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043492142111063004, + 0.03847287595272064, + 0.21901549398899078, + 0.5780366659164429, + 0.018851900473237038, + 0.07019881904125214, + 0.011892437003552914, + 0.006209488492459059, + 0.003560344921424985, + 0.010269762948155403 + ], + "credit_map": [ + 0.00297681731171906, + 0.0022243924904614687, + 0.004648552276194096, + 0.002745950361713767, + 0.0023753847926855087, + 0.00943430233746767, + 0.004761226009577513, + 0.014325521886348724, + 0.003954476676881313, + 0.003790995804592967, + 0.006990901194512844, + 0.0025804764591157436, + 0.0033956102561205626, + 0.005981802474707365, + 0.0035754914861172438, + 0.008203128352761269, + 0.019494537264108658, + 0.008924023248255253, + 0.08208108693361282, + 0.008894302882254124, + 0.007267972454428673, + 0.0500866137444973, + 0.009934678673744202, + 0.03233228251338005, + 0.020106833428144455, + 0.004879647400230169, + 0.03515172749757767, + 0.005612899549305439, + 0.008121995255351067, + 0.264273464679718, + 0.008903596550226212, + 0.23098652064800262, + 0.0017663545440882444, + 0.0028270441107451916, + 0.002909432863816619, + 0.0021206762176007032, + 0.0021837307140231133, + 0.003231757553294301, + 0.002034831326454878, + 0.0017780734924599528, + 0.00981625635176897, + 0.0025179553776979446, + 0.01464164350181818, + 0.0033703232184052467, + 0.003820913378149271, + 0.023846393451094627, + 0.003056638641282916, + 0.009128705598413944, + 0.00294909393414855, + 0.0013058061012998223, + 0.0018636906752362847, + 0.0009525614441372454, + 0.0009229836869053543, + 0.001551540568470955, + 0.001194633194245398, + 0.0011521285632625222, + 0.0007201023981906474, + 0.0008739964687265456, + 0.000848838419187814, + 0.000739681301638484, + 0.0007908818079158664, + 0.0008289086981676519, + 0.000816932471934706, + 0.0005901474505662918, + 0.0007629395113326609, + 0.0003872539382427931, + 0.0005609342479147017, + 0.00031092693097889423, + 0.0003759770770557225, + 0.0003822439757641405, + 0.000379364937543869, + 0.0004007045354228467, + 0.0019896202720701694, + 0.0013251110212877393, + 0.0013252184726297855, + 0.0009957264410331845, + 0.0012067232746630907, + 0.001229779445566237, + 0.0009098812006413937, + 0.001287703518755734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.018175480887293816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_tensor([3382, 3382])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_tensor([3382, 3382])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d4e5b0e378f88783f796aa80f459c2c831097958 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_tensor([3382, 3382])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, tracking shot, close to the ground, captures a discarded white paper takeout container, like a lone boat, drifting steadily in the current that gathers along the curb of a rainy city street. The waterlogged container, its edges softened and its bottom partially submerged in the murky water, maintains its buoyancy thanks to its ingenious folding structure. Propelled by a gentle, steady current, it sails forward steadily and calmly, occasionally giving a gentle spin when encountering small obstacles like fallen leaves. The wet asphalt, like a black mirror, shimmers with the warm yellow glow of distant streetlights and the blue and red reflections of neon store signs. These shimmering shadows are stretched and distorted by the flowing water, paving a dazzling path for this miniature voyage. The entire scene evokes the unique tranquility and poetry of a rainy night, transforming an ordinary discarded object into an unexpected and beautiful visual spectacle.", + "seed": 5764, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_3382_seed_5764_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00951322978362441, + "vlm_logprob": -0.00951322978362441, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.00951322978362441, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027348646894097328 + ], + "opposite_logprob_tokens": [ + -5.527348518371582 + ], + "logprob_reward": -0.027348646894097328, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.027348646894097328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10644669085741043, + 0.06686908006668091, + 0.04948332905769348, + 0.15863345563411713, + 0.059552405029535294, + 0.053650762885808945, + 0.33383384346961975, + 0.05701813846826553, + 0.04271343722939491, + 0.07179883867502213 + ], + "credit_map": [ + 0.014753797091543674, + 0.012184365652501583, + 0.01537469681352377, + 0.012668806128203869, + 0.012376285158097744, + 0.008733520284295082, + 0.020026348531246185, + 0.010328870266675949, + 0.0086607551202178, + 0.01053039450198412, + 0.008659550920128822, + 0.006193117238581181, + 0.007463006768375635, + 0.007057618349790573, + 0.009711693972349167, + 0.008592940866947174, + 0.005662455689162016, + 0.007971299812197685, + 0.006413578521460295, + 0.006121449172496796, + 0.005528469104319811, + 0.007021002937108278, + 0.0054098572582006454, + 0.00535521749407053, + 0.020560171455144882, + 0.006900272332131863, + 0.055900897830724716, + 0.006022065877914429, + 0.0083440151065588, + 0.046136513352394104, + 0.00491295475512743, + 0.009856581687927246, + 0.006824270822107792, + 0.006793252192437649, + 0.013070191256701946, + 0.006180638447403908, + 0.004729865584522486, + 0.006617186591029167, + 0.00631809001788497, + 0.00901891104876995, + 0.005007417872548103, + 0.004788401070982218, + 0.013145199045538902, + 0.003795633325353265, + 0.00656020687893033, + 0.004405211191624403, + 0.007886311039328575, + 0.00806238129734993, + 0.03342609480023384, + 0.030303897336125374, + 0.04538271948695183, + 0.03493955358862877, + 0.029200179502367973, + 0.03461536392569542, + 0.06375229358673096, + 0.06221375986933708, + 0.006375294178724289, + 0.003542468184605241, + 0.009872560389339924, + 0.007139131426811218, + 0.0035886734258383512, + 0.008648380637168884, + 0.007748052477836609, + 0.010103581473231316, + 0.0045472062192857265, + 0.004990004934370518, + 0.0065505714155733585, + 0.005273817107081413, + 0.003772618481889367, + 0.005752402823418379, + 0.006080924533307552, + 0.005745894741266966, + 0.00781160918995738, + 0.0062215905636549, + 0.00923053827136755, + 0.009024891071021557, + 0.005259355995804071, + 0.010585431940853596, + 0.014043250121176243, + 0.009622176177799702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.027348646894097328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643263787031174 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.007643263787031174, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643263787031174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07613871246576309, + 0.04797567054629326, + 0.027836518362164497, + 0.20461836457252502, + 0.061774469912052155, + 0.027392785996198654, + 0.3211151957511902, + 0.1049688309431076, + 0.03317636623978615, + 0.09500303864479065 + ], + "credit_map": [ + 0.01175388228148222, + 0.013452854938805103, + 0.01122373342514038, + 0.006364789791405201, + 0.008276932872831821, + 0.007150718010962009, + 0.011688962578773499, + 0.006226847879588604, + 0.005884495098143816, + 0.006110255140811205, + 0.005935130640864372, + 0.004515575710684061, + 0.007434184197336435, + 0.005901007447391748, + 0.006768275983631611, + 0.005426755640655756, + 0.002994469366967678, + 0.0033851000480353832, + 0.0035289807710796595, + 0.0031283588614314795, + 0.003313923254609108, + 0.0038306911010295153, + 0.004124993924051523, + 0.0035300040617585182, + 0.027665769681334496, + 0.009701685048639774, + 0.07447104156017303, + 0.007377680856734514, + 0.009208670817315578, + 0.057422831654548645, + 0.005955921486020088, + 0.012814783491194248, + 0.0066911871545016766, + 0.007187788374722004, + 0.009311700239777565, + 0.009526973590254784, + 0.005182269494980574, + 0.00832055602222681, + 0.006978333927690983, + 0.008575670421123505, + 0.002513559302315116, + 0.0019041530322283506, + 0.00494155241176486, + 0.005426507443189621, + 0.001983227673918009, + 0.0024585085920989513, + 0.0040002865716814995, + 0.004164995159953833, + 0.031165046617388725, + 0.029303768649697304, + 0.04726754128932953, + 0.03390015661716461, + 0.03045050799846649, + 0.03335726261138916, + 0.06164204701781273, + 0.05402892455458641, + 0.00796707347035408, + 0.015187077224254608, + 0.009377093985676765, + 0.011939721181988716, + 0.012613688595592976, + 0.007540487684309483, + 0.019591717049479485, + 0.020751988515257835, + 0.00279621803201735, + 0.004621320869773626, + 0.004236566834151745, + 0.0040054009296, + 0.0033604956697672606, + 0.004328480456024408, + 0.005332782864570618, + 0.004495102446526289, + 0.008726031519472599, + 0.01076435949653387, + 0.012324146926403046, + 0.011882936581969261, + 0.008087463676929474, + 0.010308055207133293, + 0.01948040910065174, + 0.013429643586277962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.007643263787031174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022047278471291065 + ], + "opposite_logprob_tokens": [ + -6.127204895019531 + ], + "logprob_reward": -0.0022047278471291065, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022047278471291065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18051062524318695, + 0.06361790746450424, + 0.031474240124225616, + 0.24337929487228394, + 0.09479247778654099, + 0.08491936326026917, + 0.1992911398410797, + 0.036070942878723145, + 0.015279839746654034, + 0.05066416785120964 + ], + "credit_map": [ + 0.027142025530338287, + 0.022508125752210617, + 0.025519637390971184, + 0.01727365329861641, + 0.022503729909658432, + 0.018256431445479393, + 0.030725980177521706, + 0.01658104732632637, + 0.008504382334649563, + 0.008181672543287277, + 0.009046790190041065, + 0.007428034208714962, + 0.0083810705691576, + 0.005833484698086977, + 0.007010180037468672, + 0.009232294745743275, + 0.0037513806018978357, + 0.004748445004224777, + 0.003839031560346484, + 0.0031885155476629734, + 0.004255550913512707, + 0.004275619052350521, + 0.004504128824919462, + 0.002911567920818925, + 0.033435218036174774, + 0.009252467192709446, + 0.08841865509748459, + 0.009318895637989044, + 0.013965080492198467, + 0.06692410260438919, + 0.007242064923048019, + 0.014822793193161488, + 0.009785518981516361, + 0.01106047909706831, + 0.015976278111338615, + 0.014334704726934433, + 0.008661803789436817, + 0.01055152714252472, + 0.012358581647276878, + 0.012063580565154552, + 0.007236904930323362, + 0.005951997824013233, + 0.025001103058457375, + 0.009733425453305244, + 0.0059728799387812614, + 0.007503940723836422, + 0.009758996777236462, + 0.013760111294686794, + 0.020720655098557472, + 0.017600921913981438, + 0.026566721498966217, + 0.020217059180140495, + 0.015399257652461529, + 0.022266095504164696, + 0.03762587532401085, + 0.03889455273747444, + 0.003791176015511155, + 0.0030555769335478544, + 0.0032176857348531485, + 0.0037567131221294403, + 0.0020934888161718845, + 0.0023791282437741756, + 0.008081880398094654, + 0.009695294313132763, + 0.0012031239457428455, + 0.0018718072678893805, + 0.0017935875803232193, + 0.0020859341602772474, + 0.001905258628539741, + 0.001978215528652072, + 0.0024318580981343985, + 0.0020100551191717386, + 0.005764190573245287, + 0.005076825153082609, + 0.005702883005142212, + 0.007825152948498726, + 0.003623276250436902, + 0.0066873361356556416, + 0.009609471075236797, + 0.00637503108009696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0022047278471291065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651750162243843 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.008651750162243843, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651750162243843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17779527604579926, + 0.05796384811401367, + 0.09894097596406937, + 0.09572850167751312, + 0.07585760205984116, + 0.05583415552973747, + 0.20545096695423126, + 0.030736466869711876, + 0.07259302586317062, + 0.12909913063049316 + ], + "credit_map": [ + 0.02471400424838066, + 0.025869334116578102, + 0.02565036341547966, + 0.01614473946392536, + 0.022613950073719025, + 0.014422375708818436, + 0.03241819515824318, + 0.015962325036525726, + 0.00811171904206276, + 0.010987216606736183, + 0.007850057445466518, + 0.005381403956562281, + 0.006698880344629288, + 0.006012944038957357, + 0.006677021738141775, + 0.006244609598070383, + 0.016101470217108727, + 0.01896844431757927, + 0.013584138825535774, + 0.0075898305512964725, + 0.010671359486877918, + 0.010972270742058754, + 0.010568891651928425, + 0.010484583675861359, + 0.012084701098501682, + 0.005625056568533182, + 0.031280193477869034, + 0.004224603064358234, + 0.005692748818546534, + 0.02510000392794609, + 0.00480874627828598, + 0.006912461016327143, + 0.009198240004479885, + 0.009438821114599705, + 0.012141544371843338, + 0.012446160428225994, + 0.004734621848911047, + 0.00811048224568367, + 0.009515316225588322, + 0.010272419080138206, + 0.006329250056296587, + 0.005410659592598677, + 0.006333642173558474, + 0.004673419054597616, + 0.008113238029181957, + 0.006745882797986269, + 0.009826045483350754, + 0.008402027189731598, + 0.01985936425626278, + 0.02015780098736286, + 0.027669312432408333, + 0.0217761043459177, + 0.01962091214954853, + 0.022618988528847694, + 0.037780411541461945, + 0.0359681099653244, + 0.0032099902164191008, + 0.0035919160582125187, + 0.003693503560498357, + 0.0038797978777438402, + 0.0026114988140761852, + 0.0027592882979661226, + 0.005408680997788906, + 0.005581796169281006, + 0.006482662633061409, + 0.007322984281927347, + 0.010193636640906334, + 0.00762237748131156, + 0.007030202075839043, + 0.0119302524253726, + 0.012032304890453815, + 0.009978623129427433, + 0.014403671957552433, + 0.012684051878750324, + 0.020748045295476913, + 0.018478889018297195, + 0.009403541684150696, + 0.013116435147821903, + 0.024560537189245224, + 0.015703963115811348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.008651750162243843 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017177602276206017 + ], + "opposite_logprob_tokens": [ + -6.376717567443848 + ], + "logprob_reward": -0.0017177602276206017, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017177602276206017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13692717254161835, + 0.08987587690353394, + 0.059308044612407684, + 0.1434326469898224, + 0.13498972356319427, + 0.20683793723583221, + 0.0974540039896965, + 0.06220227852463722, + 0.02691016159951687, + 0.04206211119890213 + ], + "credit_map": [ + 0.01836773380637169, + 0.017289483919739723, + 0.019239909946918488, + 0.012260782532393932, + 0.016759272664785385, + 0.015991734340786934, + 0.02311781980097294, + 0.013900438323616982, + 0.011420155875384808, + 0.017560360953211784, + 0.01120421290397644, + 0.011677882634103298, + 0.010790267027914524, + 0.007043097633868456, + 0.012696118094027042, + 0.007483774330466986, + 0.0070700873620808125, + 0.009579705074429512, + 0.008190596476197243, + 0.004733325447887182, + 0.007808124180883169, + 0.00956136453896761, + 0.0061067817732691765, + 0.006258057896047831, + 0.0185310710221529, + 0.012998098507523537, + 0.04572197422385216, + 0.008088690228760242, + 0.0076336320489645, + 0.03293254226446152, + 0.006928137969225645, + 0.010598508641123772, + 0.014869194477796555, + 0.014528798870742321, + 0.025356942787766457, + 0.014110246673226357, + 0.013791768811643124, + 0.017494555562734604, + 0.01821175031363964, + 0.016626451164484024, + 0.01899324171245098, + 0.013776621781289577, + 0.0635211318731308, + 0.02274801954627037, + 0.01717258058488369, + 0.011473200283944607, + 0.025380689650774002, + 0.033772457391023636, + 0.008371895179152489, + 0.011261319741606712, + 0.015084370970726013, + 0.011476704850792885, + 0.012193460948765278, + 0.010289239697158337, + 0.018092628568410873, + 0.010684385895729065, + 0.009245811030268669, + 0.005508047994226217, + 0.00801103562116623, + 0.007242967840284109, + 0.0036852816119790077, + 0.00630104960873723, + 0.009616795927286148, + 0.012591292150318623, + 0.0020460723899304867, + 0.0027492179069668055, + 0.002769121201708913, + 0.0034204311668872833, + 0.0031626576092094183, + 0.0031858349684625864, + 0.005618647672235966, + 0.00395817868411541, + 0.004043882247060537, + 0.003839431796222925, + 0.005674923770129681, + 0.005189229268580675, + 0.0035571553744375706, + 0.006867566145956516, + 0.006815143395215273, + 0.006074781529605389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.0017177602276206017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tracking shot, close to the ground, captures a discarded white paper takeout container, like a lone boat, drifting steadily in the current that gathers along the curb of a rainy city street. The waterlogged container, its edges softened and its bottom partially submerged in the murky water, maintains its buoyancy thanks to its ingenious folding structure. Propelled by a gentle, steady current, it sails forward steadily and calmly, occasionally giving a gentle spin when encountering small obstacles like fallen leaves. The wet asphalt, like a black mirror, shimmers with the warm yellow glow of distant streetlights and the blue and red reflections of neon store signs. These shimmering shadows are stretched and distorted by the flowing water, paving a dazzling path for this miniature voyage. The entire scene evokes the unique tranquility and poetry of a rainy night, transforming an ordinary discarded object into an unexpected and beautiful visual spectacle.", + "seed": 5765, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_3382_seed_5765_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009946010378189384, + "vlm_logprob": -0.009946010378189384, + "vlm_yes_no_margin": 5.7, + "vlm_token_credit": -0.009946010378189384, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031120354309678078 + ], + "opposite_logprob_tokens": [ + -6.656120300292969 + ], + "logprob_reward": -0.031120354309678078, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.031120354309678078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25012660026550293, + 0.0982154905796051, + 0.08137083798646927, + 0.09551283717155457, + 0.12705005705356598, + 0.051419906318187714, + 0.06011679768562317, + 0.07586570084095001, + 0.06607536226511002, + 0.09424632787704468 + ], + "credit_map": [ + 0.03477752208709717, + 0.02273706905543804, + 0.04077862948179245, + 0.03705542907118797, + 0.016906362026929855, + 0.025373345240950584, + 0.03416683152318001, + 0.038331422954797745, + 0.010330980643630028, + 0.008007231168448925, + 0.0171047430485487, + 0.010847014375030994, + 0.007224175613373518, + 0.012874334119260311, + 0.014113939367234707, + 0.01771307922899723, + 0.012576659210026264, + 0.0068337637931108475, + 0.011305175721645355, + 0.006404994986951351, + 0.012599868699908257, + 0.009484386071562767, + 0.011348568834364414, + 0.010817424394190311, + 0.01287308894097805, + 0.00789615698158741, + 0.019999166950583458, + 0.007460694760084152, + 0.01007277425378561, + 0.012001830153167248, + 0.01128511130809784, + 0.013924019411206245, + 0.010968056507408619, + 0.010727010667324066, + 0.013697242364287376, + 0.010594576597213745, + 0.01838473230600357, + 0.014753839001059532, + 0.0288896132260561, + 0.019034994766116142, + 0.004678800702095032, + 0.00370301422663033, + 0.005646103993058205, + 0.0035681084264069796, + 0.0025591186713427305, + 0.01344714593142271, + 0.003956986125558615, + 0.013860626146197319, + 0.003944914322346449, + 0.00667354604229331, + 0.005466587375849485, + 0.00849189143627882, + 0.008428514935076237, + 0.009398441761732101, + 0.008037453517317772, + 0.009675448760390282, + 0.00985361635684967, + 0.008371478877961636, + 0.010200152173638344, + 0.010049527511000633, + 0.006088903173804283, + 0.012967205606400967, + 0.0069345636293292046, + 0.0114002525806427, + 0.009785384871065617, + 0.007320354226976633, + 0.00926926638931036, + 0.008647529408335686, + 0.00550612760707736, + 0.007832510396838188, + 0.007799047511070967, + 0.009915148839354515, + 0.025045275688171387, + 0.008115259930491447, + 0.023647431284189224, + 0.007835094816982746, + 0.005199725273996592, + 0.010493665933609009, + 0.006479479838162661, + 0.007430397439748049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.031120354309678078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004654762800782919 + ], + "opposite_logprob_tokens": [ + -5.379654884338379 + ], + "logprob_reward": -0.004654762800782919, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004654762800782919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2895249128341675, + 0.12807689607143402, + 0.04993828013539314, + 0.09718057513237, + 0.08095179498195648, + 0.10238480567932129, + 0.11543601751327515, + 0.03422467038035393, + 0.03647078573703766, + 0.06581128388643265 + ], + "credit_map": [ + 0.06786995381116867, + 0.0349963940680027, + 0.04919695481657982, + 0.02577570453286171, + 0.01967429742217064, + 0.02354711852967739, + 0.019711550325155258, + 0.048752959817647934, + 0.012127067893743515, + 0.011084010824561119, + 0.014851223677396774, + 0.012995479628443718, + 0.014105720445513725, + 0.018144141882658005, + 0.023311328142881393, + 0.02145792543888092, + 0.006053895223885775, + 0.005720891989767551, + 0.007848717272281647, + 0.0037476741708815098, + 0.0074003650806844234, + 0.004625633358955383, + 0.0074945250526070595, + 0.007046579383313656, + 0.010687959380447865, + 0.009045768529176712, + 0.016257042065262794, + 0.007569862529635429, + 0.011647769249975681, + 0.013444305397570133, + 0.015318662859499454, + 0.013209197670221329, + 0.005973908584564924, + 0.007881797850131989, + 0.007793935481458902, + 0.007159806322306395, + 0.015754373744130135, + 0.011049547232687473, + 0.015263264067471027, + 0.010075160302221775, + 0.012412707321345806, + 0.007295910269021988, + 0.01190139725804329, + 0.006773719098418951, + 0.004506708588451147, + 0.02490515261888504, + 0.006401360966265202, + 0.028187846764922142, + 0.010916018858551979, + 0.009656655602157116, + 0.008261538110673428, + 0.02128814533352852, + 0.014289719983935356, + 0.017174122855067253, + 0.01876676455140114, + 0.015083054080605507, + 0.003907875157892704, + 0.0032678295392543077, + 0.005680728703737259, + 0.0038626149762421846, + 0.0027402201667428017, + 0.005272966343909502, + 0.004549640230834484, + 0.004942794796079397, + 0.005263804458081722, + 0.0029903221875429153, + 0.006977483630180359, + 0.0040475158020854, + 0.0033818797674030066, + 0.004797101952135563, + 0.003942761104553938, + 0.005069916136562824, + 0.019749078899621964, + 0.004970083944499493, + 0.012046062387526035, + 0.008356762118637562, + 0.004115734715014696, + 0.00422570900991559, + 0.006192539352923632, + 0.006155315320938826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004654762800782919 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019488405669108033 + ], + "opposite_logprob_tokens": [ + -6.251948833465576 + ], + "logprob_reward": -0.0019488405669108033, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019488405669108033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2637360394001007, + 0.08207438886165619, + 0.19026263058185577, + 0.120435930788517, + 0.10123661160469055, + 0.03783535584807396, + 0.07231943309307098, + 0.040073659271001816, + 0.048210617154836655, + 0.04381532222032547 + ], + "credit_map": [ + 0.029930345714092255, + 0.019258998334407806, + 0.04705357179045677, + 0.04056172072887421, + 0.018869662657380104, + 0.030173948034644127, + 0.03745509684085846, + 0.040432680398225784, + 0.010457256808876991, + 0.010698203928768635, + 0.012364652007818222, + 0.0070088934153318405, + 0.0067910086363554, + 0.011530781164765358, + 0.00973280519247055, + 0.013490783981978893, + 0.02675987407565117, + 0.020282519981265068, + 0.02583540976047516, + 0.012385217472910881, + 0.028990235179662704, + 0.022718684747815132, + 0.02543134056031704, + 0.027859363704919815, + 0.01254464965313673, + 0.00818368699401617, + 0.017649654299020767, + 0.010263892821967602, + 0.010536810383200645, + 0.031900376081466675, + 0.012800591997802258, + 0.01655626855790615, + 0.011256873607635498, + 0.012806659564375877, + 0.014931445941329002, + 0.0126412408426404, + 0.010006174445152283, + 0.015121911652386189, + 0.010697523131966591, + 0.013774779625236988, + 0.006675670854747295, + 0.004162636585533619, + 0.005236291792243719, + 0.0033404382411390543, + 0.0031084332149475813, + 0.005282003898173571, + 0.003583987010642886, + 0.0064458963461220264, + 0.006442595738917589, + 0.005814528092741966, + 0.006016287952661514, + 0.007455732207745314, + 0.006838041357696056, + 0.012693350203335285, + 0.010103789158165455, + 0.016955099999904633, + 0.0062191118486225605, + 0.0024774540215730667, + 0.005098660942167044, + 0.004255553241819143, + 0.0026009141001850367, + 0.00835780706256628, + 0.0035594250075519085, + 0.007504733745008707, + 0.006512404885143042, + 0.003944195806980133, + 0.006548392120748758, + 0.00461812037974596, + 0.0034784688614308834, + 0.008797272108495235, + 0.006226928438991308, + 0.008084836415946484, + 0.006289671640843153, + 0.00603344663977623, + 0.008220449090003967, + 0.005509768612682819, + 0.004576668608933687, + 0.005328076425939798, + 0.004316394217312336, + 0.0035408465191721916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0019488405669108033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747674662619829 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747674662619829, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747674662619829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3022628128528595, + 0.07025513797998428, + 0.07227712869644165, + 0.11252856254577637, + 0.07627353817224503, + 0.05681378021836281, + 0.061458684504032135, + 0.06487345695495605, + 0.08343255519866943, + 0.09982430934906006 + ], + "credit_map": [ + 0.0593341626226902, + 0.0324304960668087, + 0.05789129436016083, + 0.02771521359682083, + 0.022972987964749336, + 0.028929108753800392, + 0.024500679224729538, + 0.04848887398838997, + 0.00948404986411333, + 0.008241541683673859, + 0.009399272501468658, + 0.006728242617100477, + 0.007285865489393473, + 0.009031865745782852, + 0.008107156492769718, + 0.011977143585681915, + 0.009332023561000824, + 0.007235066965222359, + 0.010418043471872807, + 0.004015459679067135, + 0.01191423274576664, + 0.007325358688831329, + 0.009968836791813374, + 0.012068110518157482, + 0.012122624553740025, + 0.01078298594802618, + 0.016741078346967697, + 0.009783675894141197, + 0.01608714647591114, + 0.01147743035107851, + 0.01847018674015999, + 0.017063427716493607, + 0.0058512007817626, + 0.00928568560630083, + 0.006595975253731012, + 0.0069176494143903255, + 0.015121555887162685, + 0.011258442886173725, + 0.011341352015733719, + 0.009901678189635277, + 0.005351205822080374, + 0.006040982902050018, + 0.004444791004061699, + 0.0065070101991295815, + 0.006178864277899265, + 0.010507775470614433, + 0.009161865338683128, + 0.008621285669505596, + 0.008520410396158695, + 0.005751224234700203, + 0.006307694595307112, + 0.005917474161833525, + 0.004279805812984705, + 0.010333237238228321, + 0.006365231238305569, + 0.013983606360852718, + 0.00948952604085207, + 0.004639035556465387, + 0.007586688734591007, + 0.007981343194842339, + 0.00462065264582634, + 0.014475295320153236, + 0.005242910701781511, + 0.01083800196647644, + 0.01102240476757288, + 0.009543421678245068, + 0.012346104718744755, + 0.008787865750491619, + 0.007959305308759212, + 0.010649149306118488, + 0.011784816160798073, + 0.01133948564529419, + 0.01887327991425991, + 0.010725939646363258, + 0.021895110607147217, + 0.008143825456500053, + 0.005105805583298206, + 0.015815556049346924, + 0.008592231199145317, + 0.010672558099031448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.006747674662619829 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052584195509552956 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.0052584195509552956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052584195509552956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2807686924934387, + 0.2702546715736389, + 0.06629876792430878, + 0.05720369517803192, + 0.09755872935056686, + 0.09781856089830399, + 0.034867316484451294, + 0.03498663380742073, + 0.02287297323346138, + 0.03736989572644234 + ], + "credit_map": [ + 0.027484338730573654, + 0.027246840298175812, + 0.038196247071027756, + 0.04669247195124626, + 0.026208680123090744, + 0.030545037239789963, + 0.04158778861165047, + 0.04280730336904526, + 0.04188228398561478, + 0.024824630469083786, + 0.06086193770170212, + 0.01919666677713394, + 0.0146649070084095, + 0.03235325962305069, + 0.01968429610133171, + 0.05678670108318329, + 0.009883171878755093, + 0.007719679735600948, + 0.009554523043334484, + 0.004793435335159302, + 0.009085784666240215, + 0.005977597087621689, + 0.008313777856528759, + 0.010970797389745712, + 0.0056860800832509995, + 0.0057791126891970634, + 0.008515431545674801, + 0.004833430051803589, + 0.0074668703600764275, + 0.006879459600895643, + 0.00936937890946865, + 0.008673938922584057, + 0.00873342715203762, + 0.00949280895292759, + 0.01064812857657671, + 0.009275996126234531, + 0.014005385339260101, + 0.013519302010536194, + 0.018758384510874748, + 0.013125310651957989, + 0.01264897733926773, + 0.006666171830147505, + 0.010319712571799755, + 0.006263039540499449, + 0.005252666771411896, + 0.023599283769726753, + 0.00625389302149415, + 0.026814822107553482, + 0.003613510401919484, + 0.003243525978177786, + 0.0052852020598948, + 0.0035266864579170942, + 0.0032217144034802914, + 0.005733078345656395, + 0.005369645077735186, + 0.004873953759670258, + 0.005300847347825766, + 0.0025997741613537073, + 0.0037764145527035, + 0.004087676294147968, + 0.0017626129556447268, + 0.008000641129910946, + 0.004473815206438303, + 0.004984855651855469, + 0.004182285163551569, + 0.0017778276233002543, + 0.0030431412160396576, + 0.002941038692370057, + 0.0018596878508105874, + 0.0028391971718519926, + 0.002689589047804475, + 0.0035402073990553617, + 0.00934060849249363, + 0.0028562184888869524, + 0.006847891490906477, + 0.00628206180408597, + 0.0016683534486219287, + 0.0032913729082792997, + 0.0037389930803328753, + 0.0033443986903876066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0052584195509552956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tracking shot, close to the ground, captures a discarded white paper takeout container, like a lone boat, drifting steadily in the current that gathers along the curb of a rainy city street. The waterlogged container, its edges softened and its bottom partially submerged in the murky water, maintains its buoyancy thanks to its ingenious folding structure. Propelled by a gentle, steady current, it sails forward steadily and calmly, occasionally giving a gentle spin when encountering small obstacles like fallen leaves. The wet asphalt, like a black mirror, shimmers with the warm yellow glow of distant streetlights and the blue and red reflections of neon store signs. These shimmering shadows are stretched and distorted by the flowing water, paving a dazzling path for this miniature voyage. The entire scene evokes the unique tranquility and poetry of a rainy night, transforming an ordinary discarded object into an unexpected and beautiful visual spectacle.", + "seed": 5766, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_3382_seed_5766_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011087303422391415, + "vlm_logprob": -0.011087303422391415, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.011087303422391415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031303610652685165 + ], + "opposite_logprob_tokens": [ + -6.531303405761719 + ], + "logprob_reward": -0.031303610652685165, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.031303610652685165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09753644466400146, + 0.07101866602897644, + 0.10940194875001907, + 0.12182694673538208, + 0.050759222358465195, + 0.0637788400053978, + 0.1326570212841034, + 0.06394141167402267, + 0.13306397199630737, + 0.15601560473442078 + ], + "credit_map": [ + 0.010730535723268986, + 0.0067457836121320724, + 0.01108547393232584, + 0.01312632393091917, + 0.007949567399919033, + 0.016745550557971, + 0.015121758915483952, + 0.01603144221007824, + 0.008399879559874535, + 0.0071350326761603355, + 0.009800995700061321, + 0.008830361999571323, + 0.004875022452324629, + 0.007340651471167803, + 0.011364037171006203, + 0.013272671028971672, + 0.008468016050755978, + 0.011887261644005775, + 0.008475459180772305, + 0.008369645103812218, + 0.00916068535298109, + 0.008602596819400787, + 0.02527252770960331, + 0.029165739193558693, + 0.02476327121257782, + 0.008848285302519798, + 0.03509368374943733, + 0.007859813049435616, + 0.0074942647479474545, + 0.011058954522013664, + 0.008219833485782146, + 0.018488824367523193, + 0.008112064562737942, + 0.006471880245953798, + 0.0078052859753370285, + 0.005286613013595343, + 0.005140243098139763, + 0.007323003374040127, + 0.004183209966868162, + 0.006436917930841446, + 0.011836414225399494, + 0.006656025070697069, + 0.008101915009319782, + 0.007926935330033302, + 0.005052773281931877, + 0.0088746203109622, + 0.008198658935725689, + 0.007131494581699371, + 0.030736053362488747, + 0.010662711225450039, + 0.01727919839322567, + 0.008793634362518787, + 0.009204826317727566, + 0.029576485976576805, + 0.013591255061328411, + 0.012812848202884197, + 0.01724262349307537, + 0.003705534851178527, + 0.005992390215396881, + 0.018876606598496437, + 0.003864720929414034, + 0.005774273537099361, + 0.0036534443497657776, + 0.004831809084862471, + 0.020367631688714027, + 0.008126496337354183, + 0.020077386870980263, + 0.008153350092470646, + 0.02965252473950386, + 0.013763492926955223, + 0.013907562009990215, + 0.019015518948435783, + 0.026866449043154716, + 0.00977880135178566, + 0.028748832643032074, + 0.006431611720472574, + 0.03418881818652153, + 0.013925381936132908, + 0.012537469156086445, + 0.023538224399089813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.031303610652685165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646458223462105 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646458223462105, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646458223462105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16809584200382233, + 0.1330462247133255, + 0.06361503899097443, + 0.07990903407335281, + 0.042291123420000076, + 0.02302871085703373, + 0.08140378445386887, + 0.02590113878250122, + 0.26507750153541565, + 0.11763160675764084 + ], + "credit_map": [ + 0.0180705264210701, + 0.01565885730087757, + 0.023045487701892853, + 0.023081505671143532, + 0.013868002220988274, + 0.026670534163713455, + 0.02439054101705551, + 0.023310383781790733, + 0.013890009373426437, + 0.02120991423726082, + 0.011930369772017002, + 0.021124672144651413, + 0.013626987114548683, + 0.011158580891788006, + 0.02161584421992302, + 0.01848985068500042, + 0.006747652310878038, + 0.005053202621638775, + 0.009746829979121685, + 0.005924080032855272, + 0.004711539018899202, + 0.006433852482587099, + 0.012611337006092072, + 0.012386546470224857, + 0.018106743693351746, + 0.0049250503070652485, + 0.027870792895555496, + 0.0059755961410701275, + 0.004175374284386635, + 0.004804228898137808, + 0.00526017602533102, + 0.008791069500148296, + 0.005073526408523321, + 0.0036517437547445297, + 0.00904132705181837, + 0.005588822066783905, + 0.0035594089422374964, + 0.008654745295643806, + 0.004005740396678448, + 0.0027158078737556934, + 0.003232145681977272, + 0.0014045299030840397, + 0.003307932987809181, + 0.0028297100216150284, + 0.0013726165052503347, + 0.004317356739193201, + 0.003909803461283445, + 0.002654615556821227, + 0.018580371513962746, + 0.006536614149808884, + 0.0085949981585145, + 0.0035271490924060345, + 0.004307230934500694, + 0.023938696831464767, + 0.0070976014249026775, + 0.008821127004921436, + 0.0035899390932172537, + 0.003397256601601839, + 0.0022408755030483007, + 0.003765481524169445, + 0.0034220193047076464, + 0.00247571081854403, + 0.0037163663655519485, + 0.0032934879418462515, + 0.05855143070220947, + 0.01495678536593914, + 0.0818050280213356, + 0.021250586956739426, + 0.015538115985691547, + 0.023200305178761482, + 0.024198509752750397, + 0.025576749816536903, + 0.014572332613170147, + 0.0045799994841217995, + 0.03575149178504944, + 0.004959027282893658, + 0.012052853591740131, + 0.00784248672425747, + 0.005065353587269783, + 0.032808054238557816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.007646458223462105 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031986285466700792 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031986285466700792, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031986285466700792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2086118757724762, + 0.10504188388586044, + 0.053972095251083374, + 0.08677221834659576, + 0.03554844111204147, + 0.06323999911546707, + 0.08272790163755417, + 0.02982219308614731, + 0.20179089903831482, + 0.13247257471084595 + ], + "credit_map": [ + 0.02132173627614975, + 0.013048558495938778, + 0.0223158597946167, + 0.02944599650800228, + 0.015018858015537262, + 0.03913149982690811, + 0.034040771424770355, + 0.03428858891129494, + 0.008184660226106644, + 0.004166339058429003, + 0.018678711727261543, + 0.008244054391980171, + 0.005117400083690882, + 0.012252647429704666, + 0.01685144752264023, + 0.03154661878943443, + 0.007906951010227203, + 0.006661662366241217, + 0.007470122072845697, + 0.006888396106660366, + 0.005820629186928272, + 0.006307767238467932, + 0.0073212492279708385, + 0.005595307797193527, + 0.020034752786159515, + 0.0035966660361737013, + 0.030167659744620323, + 0.005395317450165749, + 0.002904884284362197, + 0.007600948214530945, + 0.005407503340393305, + 0.011664479970932007, + 0.0068679568357765675, + 0.0024519828148186207, + 0.006666646338999271, + 0.0036323927342891693, + 0.001991387689486146, + 0.005356294102966785, + 0.0034327246248722076, + 0.005149048753082752, + 0.00796278566122055, + 0.004095393233001232, + 0.014146911911666393, + 0.007062170188874006, + 0.0040213726460933685, + 0.012327076867222786, + 0.007941271178424358, + 0.005683011841028929, + 0.02238059975206852, + 0.00614252220839262, + 0.008267632685601711, + 0.0032887612469494343, + 0.003421245841309428, + 0.028349410742521286, + 0.005430740769952536, + 0.00544698117300868, + 0.004124360624700785, + 0.003895163768902421, + 0.0028494878206402063, + 0.004992990754544735, + 0.004152327310293913, + 0.003945070784538984, + 0.002548885764554143, + 0.003313902299851179, + 0.02936699241399765, + 0.012542270123958588, + 0.04270014539361, + 0.01861703395843506, + 0.02029534801840782, + 0.03799445554614067, + 0.018210140988230705, + 0.02206449955701828, + 0.023719996213912964, + 0.0048109074123203754, + 0.05295393243432045, + 0.004627512767910957, + 0.013663834892213345, + 0.00764174060896039, + 0.005033215507864952, + 0.02002141997218132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0031986285466700792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085351929068565 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085351929068565, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085351929068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1156768947839737, + 0.05789085105061531, + 0.07675917446613312, + 0.1275249570608139, + 0.12064836919307709, + 0.045694176107645035, + 0.20978820323944092, + 0.038455791771411896, + 0.11132406443357468, + 0.09623758494853973 + ], + "credit_map": [ + 0.012307850643992424, + 0.007253857795149088, + 0.011972060427069664, + 0.013845576904714108, + 0.008977444842457771, + 0.022991843521595, + 0.01902649737894535, + 0.0193017590790987, + 0.005355537869036198, + 0.003354998305439949, + 0.007854483090341091, + 0.007345111574977636, + 0.004412359558045864, + 0.008504276163876057, + 0.00812364462763071, + 0.012940436601638794, + 0.008311825804412365, + 0.012987217865884304, + 0.006921111140400171, + 0.010148677043616772, + 0.01021045632660389, + 0.008707519620656967, + 0.009265629574656487, + 0.01020672544836998, + 0.01666071079671383, + 0.008125740103423595, + 0.023915722966194153, + 0.008586319163441658, + 0.008191723376512527, + 0.0171556007117033, + 0.020546795800328255, + 0.02434232085943222, + 0.016498198732733727, + 0.013247721828520298, + 0.023373771458864212, + 0.01293782889842987, + 0.00956844910979271, + 0.021367540583014488, + 0.010798466391861439, + 0.012856381013989449, + 0.00603732792660594, + 0.0025373802054673433, + 0.007325396873056889, + 0.004923027474433184, + 0.00310740084387362, + 0.007670806720852852, + 0.007066329475492239, + 0.007026506122201681, + 0.059640128165483475, + 0.014639526605606079, + 0.01774437725543976, + 0.007186167407780886, + 0.008116188459098339, + 0.07623058557510376, + 0.010961291380226612, + 0.015269910916686058, + 0.008642983622848988, + 0.0037392182275652885, + 0.0035589979961514473, + 0.00786997564136982, + 0.0030953255482017994, + 0.004170355852693319, + 0.003460757900029421, + 0.0039181774482131, + 0.017544681206345558, + 0.008534438908100128, + 0.014512324705719948, + 0.006497511174529791, + 0.02036568708717823, + 0.01937059871852398, + 0.010371105745434761, + 0.014127710834145546, + 0.01832517236471176, + 0.006183987949043512, + 0.02215355820953846, + 0.00539750000461936, + 0.010478968732059002, + 0.012235886417329311, + 0.010185428895056248, + 0.011277071200311184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.011085351929068565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022024677600711584 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022024677600711584, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022024677600711584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15326054394245148, + 0.03584893047809601, + 0.13465027511119843, + 0.08161315321922302, + 0.11546947807073593, + 0.07184258103370667, + 0.08024236559867859, + 0.0423508882522583, + 0.1138182282447815, + 0.17090356349945068 + ], + "credit_map": [ + 0.016609059646725655, + 0.014881114475429058, + 0.023693308234214783, + 0.018752215430140495, + 0.0147767448797822, + 0.02416270039975643, + 0.019557451829314232, + 0.02082795649766922, + 0.0032679697033017874, + 0.0036391576286405325, + 0.0048708138056099415, + 0.0036075799725949764, + 0.0030877073295414448, + 0.004516024142503738, + 0.00458541139960289, + 0.008274262771010399, + 0.010755447670817375, + 0.011646338738501072, + 0.007153146434575319, + 0.010054480284452438, + 0.012582728639245033, + 0.010480216704308987, + 0.031837623566389084, + 0.04014028608798981, + 0.010484528727829456, + 0.004341537598520517, + 0.02087434008717537, + 0.004176876973360777, + 0.004176171962171793, + 0.009630449116230011, + 0.006705132778733969, + 0.021224118769168854, + 0.019482988864183426, + 0.006899719592183828, + 0.018797684460878372, + 0.008118283934891224, + 0.008184190839529037, + 0.02356681600213051, + 0.010579417459666729, + 0.01984037086367607, + 0.01016839500516653, + 0.004864577203989029, + 0.011012365110218525, + 0.007665822748094797, + 0.004904556553810835, + 0.012490714900195599, + 0.009851728565990925, + 0.010884424671530724, + 0.020006747916340828, + 0.005372518207877874, + 0.00793443899601698, + 0.004506109748035669, + 0.004037095699459314, + 0.0234893336892128, + 0.007598778698593378, + 0.007297344971448183, + 0.005830463487654924, + 0.003953043837100267, + 0.004204553086310625, + 0.007312628906220198, + 0.00381896598264575, + 0.007342726923525333, + 0.0043623908422887325, + 0.005526118911802769, + 0.016067760065197945, + 0.007256668992340565, + 0.03577807918190956, + 0.006784509401768446, + 0.02601439505815506, + 0.008569885976612568, + 0.006245479919016361, + 0.007101454306393862, + 0.029257897287607193, + 0.004828053526580334, + 0.06985003501176834, + 0.005577804986387491, + 0.01445911917835474, + 0.009500345215201378, + 0.00676749600097537, + 0.030662821605801582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0022024677600711584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tracking shot, close to the ground, captures a discarded white paper takeout container, like a lone boat, drifting steadily in the current that gathers along the curb of a rainy city street. The waterlogged container, its edges softened and its bottom partially submerged in the murky water, maintains its buoyancy thanks to its ingenious folding structure. Propelled by a gentle, steady current, it sails forward steadily and calmly, occasionally giving a gentle spin when encountering small obstacles like fallen leaves. The wet asphalt, like a black mirror, shimmers with the warm yellow glow of distant streetlights and the blue and red reflections of neon store signs. These shimmering shadows are stretched and distorted by the flowing water, paving a dazzling path for this miniature voyage. The entire scene evokes the unique tranquility and poetry of a rainy night, transforming an ordinary discarded object into an unexpected and beautiful visual spectacle.", + "seed": 5767, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_2/index_3382_seed_5767_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010943496599793434, + "vlm_logprob": -0.010943496599793434, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.010943496599793434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03148198127746582 + ], + "opposite_logprob_tokens": [ + -6.406481742858887 + ], + "logprob_reward": -0.03148198127746582, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.03148198127746582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07484608143568039, + 0.032736897468566895, + 0.028317129239439964, + 0.026799561455845833, + 0.029704943299293518, + 0.05362832918763161, + 0.6647278070449829, + 0.04399894177913666, + 0.03273326903581619, + 0.012507065199315548 + ], + "credit_map": [ + 0.007468412630259991, + 0.009072826243937016, + 0.008493268862366676, + 0.006955369841307402, + 0.006833047606050968, + 0.01215433795005083, + 0.009569716639816761, + 0.014299088157713413, + 0.0036114042159169912, + 0.0033763463143259287, + 0.0028025114443153143, + 0.002680797129869461, + 0.0022632237523794174, + 0.007224953267723322, + 0.005686156917363405, + 0.005091501399874687, + 0.0018495620461180806, + 0.0011670448584482074, + 0.004930334165692329, + 0.00303738284856081, + 0.0008726059459149837, + 0.00639285147190094, + 0.0016488861292600632, + 0.008418458513915539, + 0.004346165806055069, + 0.0020816840697079897, + 0.004113224800676107, + 0.003823558334261179, + 0.001998934894800186, + 0.0036930704955011606, + 0.0026046347338706255, + 0.004138285759836435, + 0.002381965285167098, + 0.003484455868601799, + 0.00261422130279243, + 0.005083771422505379, + 0.0028691471088677645, + 0.003622540272772312, + 0.00438939081504941, + 0.005259447731077671, + 0.003931875340640545, + 0.0022769044153392315, + 0.004974766634404659, + 0.015891721472144127, + 0.002308022463694215, + 0.008941451087594032, + 0.009429278783500195, + 0.005874304100871086, + 0.25114700198173523, + 0.012275733053684235, + 0.3010324537754059, + 0.04252507910132408, + 0.016903357580304146, + 0.011188820004463196, + 0.013434361666440964, + 0.01622096449136734, + 0.007504127454012632, + 0.003732474287971854, + 0.009508510120213032, + 0.0070254686288535595, + 0.004719898570328951, + 0.00396386394277215, + 0.0049851262010633945, + 0.0025594683829694986, + 0.004875560291111469, + 0.0019067857647314668, + 0.004697965458035469, + 0.0033152159303426743, + 0.002677540061995387, + 0.0055062235333025455, + 0.004812362603843212, + 0.004941611085087061, + 0.001405219780281186, + 0.001069848658517003, + 0.002043134532868862, + 0.0022361271549016237, + 0.00122320547234267, + 0.001767949666827917, + 0.0017915819771587849, + 0.0009699977235868573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.03148198127746582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006749569438397884 + ], + "opposite_logprob_tokens": [ + -5.006749629974365 + ], + "logprob_reward": -0.006749569438397884, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006749569438397884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14222091436386108, + 0.05348329246044159, + 0.031559478491544724, + 0.03419540449976921, + 0.022445159032940865, + 0.08210422098636627, + 0.4992651343345642, + 0.043298836797475815, + 0.07622867077589035, + 0.015198874287307262 + ], + "credit_map": [ + 0.013360116630792618, + 0.017268098890781403, + 0.014338868670165539, + 0.011974713765084743, + 0.01464754156768322, + 0.02458832412958145, + 0.01785457879304886, + 0.02818867564201355, + 0.004138368181884289, + 0.0062581016682088375, + 0.004548788070678711, + 0.005274394527077675, + 0.00523905735462904, + 0.008766882121562958, + 0.0112812714651227, + 0.007976426742970943, + 0.002623701700940728, + 0.002776965731754899, + 0.0031218414660543203, + 0.003870683955028653, + 0.0027773415204137564, + 0.007082274649292231, + 0.003987923264503479, + 0.005318746902048588, + 0.004389527719467878, + 0.0026608319021761417, + 0.0036625429056584835, + 0.006731721572577953, + 0.002828072290867567, + 0.0029873610474169254, + 0.005233415402472019, + 0.005701932590454817, + 0.0017872069729492068, + 0.002360246842727065, + 0.0017764141084626317, + 0.003842439502477646, + 0.0019529637647792697, + 0.0026189519558101892, + 0.003840327262878418, + 0.004266607575118542, + 0.0057719252072274685, + 0.0031529166735708714, + 0.011512862518429756, + 0.03191262483596802, + 0.0027044396847486496, + 0.015687769278883934, + 0.004989387933164835, + 0.006372292060405016, + 0.19108644127845764, + 0.008610140532255173, + 0.22742459177970886, + 0.030963944271206856, + 0.012684036046266556, + 0.007516127545386553, + 0.009649796411395073, + 0.011330093257129192, + 0.006989982910454273, + 0.004735814407467842, + 0.007639653515070677, + 0.0092732859775424, + 0.004976743832230568, + 0.003083419520407915, + 0.0039263260550796986, + 0.0026736087165772915, + 0.010993387550115585, + 0.00420716218650341, + 0.010778692550957203, + 0.01723325066268444, + 0.0047743008472025394, + 0.009483382105827332, + 0.008918150328099728, + 0.00984035525470972, + 0.002081547398120165, + 0.0011790167773142457, + 0.0023384748492389917, + 0.002404105616733432, + 0.0011987901525571942, + 0.0019982014782726765, + 0.002394340466707945, + 0.0016043981304392219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.006749569438397884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031969649717211723 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.0031969649717211723, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031969649717211723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23028676211833954, + 0.07620115578174591, + 0.05629262328147888, + 0.21286122500896454, + 0.10874965786933899, + 0.03446417674422264, + 0.1261058747768402, + 0.0258930753916502, + 0.06073993071913719, + 0.06840552389621735 + ], + "credit_map": [ + 0.018280671909451485, + 0.03336402773857117, + 0.02579379640519619, + 0.01798834092915058, + 0.015903044492006302, + 0.04594277963042259, + 0.026286037638783455, + 0.04672805592417717, + 0.009496564976871014, + 0.003787008114159107, + 0.008717763237655163, + 0.00591361615806818, + 0.004092923831194639, + 0.01906410977244377, + 0.00823074858635664, + 0.01689843088388443, + 0.0076330313459038734, + 0.004787956830114126, + 0.004605934023857117, + 0.007458374369889498, + 0.003163074143230915, + 0.012018733657896519, + 0.00721455505117774, + 0.009410963393747807, + 0.02692793495953083, + 0.015536529943346977, + 0.03211287409067154, + 0.04596098139882088, + 0.014018401503562927, + 0.020801501348614693, + 0.02392880991101265, + 0.03357420116662979, + 0.00858386978507042, + 0.012357574887573719, + 0.01072895247489214, + 0.01736239343881607, + 0.01061292178928852, + 0.01842200942337513, + 0.016955766826868057, + 0.013726163655519485, + 0.0034363081213086843, + 0.0015519758453592658, + 0.006528499070554972, + 0.010860346257686615, + 0.0014065703144297004, + 0.002709820168092847, + 0.0032955489587038755, + 0.004675107076764107, + 0.047311075031757355, + 0.0034451058600097895, + 0.05279921740293503, + 0.0069645908661186695, + 0.0033432503696531057, + 0.003400353714823723, + 0.003985300660133362, + 0.004856983199715614, + 0.004539316054433584, + 0.002645346336066723, + 0.0038522693794220686, + 0.004918914753943682, + 0.002380449092015624, + 0.002436485607177019, + 0.003344773082062602, + 0.0017755203880369663, + 0.009291808120906353, + 0.003394771134480834, + 0.007694700732827187, + 0.011985685676336288, + 0.003712984500452876, + 0.007303554564714432, + 0.01076833438128233, + 0.006588092539459467, + 0.008001824840903282, + 0.005326718091964722, + 0.011423398740589619, + 0.008967006579041481, + 0.004866103641688824, + 0.014963584020733833, + 0.008801784366369247, + 0.006055100820958614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0031969649717211723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647849783301353 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647849783301353, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647849783301353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2143021672964096, + 0.058674734085798264, + 0.06035071983933449, + 0.03877194970846176, + 0.11645166575908661, + 0.039962656795978546, + 0.3306373059749603, + 0.030669406056404114, + 0.07515013962984085, + 0.03502929210662842 + ], + "credit_map": [ + 0.018373334780335426, + 0.024768490344285965, + 0.023079894483089447, + 0.01681341603398323, + 0.022669857367873192, + 0.03641918674111366, + 0.027323568239808083, + 0.04485441744327545, + 0.0069067105650901794, + 0.004039186984300613, + 0.004896308295428753, + 0.0026701875030994415, + 0.003733657533302903, + 0.014902593567967415, + 0.005636381451040506, + 0.015889711678028107, + 0.004556653555482626, + 0.00260867434553802, + 0.010215559974312782, + 0.004662674851715565, + 0.0015001872088760138, + 0.01418178528547287, + 0.0037926696240901947, + 0.01883251778781414, + 0.005205864552408457, + 0.002702695317566395, + 0.005425355397164822, + 0.00812778901308775, + 0.0020095258951187134, + 0.004332083277404308, + 0.005327705759555101, + 0.005640931893140078, + 0.007185573223978281, + 0.007841280661523342, + 0.009429664351046085, + 0.021458573639392853, + 0.007249362301081419, + 0.01850012131035328, + 0.014935163781046867, + 0.02985193394124508, + 0.003328533610329032, + 0.002232516650110483, + 0.0038844523951411247, + 0.01269446685910225, + 0.0017317755846306682, + 0.005516984965652227, + 0.005300483200699091, + 0.005273443181067705, + 0.12869904935359955, + 0.00646623782813549, + 0.1477147936820984, + 0.019552363082766533, + 0.007644782308489084, + 0.005584631580859423, + 0.006697805132716894, + 0.00827766489237547, + 0.005885684862732887, + 0.0029842788353562355, + 0.006940145511180162, + 0.004734812304377556, + 0.003210039110854268, + 0.0019076758762821555, + 0.0028551644645631313, + 0.002151607535779476, + 0.011404595337808132, + 0.004413031041622162, + 0.010757570154964924, + 0.012972862459719181, + 0.00554928183555603, + 0.0116943484172225, + 0.006620719097554684, + 0.011737726628780365, + 0.003653887426480651, + 0.0030978661961853504, + 0.005518125835806131, + 0.007099431473761797, + 0.003194632241502404, + 0.003852153429761529, + 0.005665752571076155, + 0.0029474433977156878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.008647849783301353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a discarded white paper takeout container?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the container drifting in murky water along the curb of a rainy city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container move steadily forward and occasionally spin when encountering obstacles like fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the wet asphalt reflect warm yellow streetlights and blue/red neon store signs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464111752808094 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.00464111752808094, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464111752808094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06317650526762009, + 0.1138942539691925, + 0.3183397650718689, + 0.052268777042627335, + 0.07550429552793503, + 0.019612520933151245, + 0.27318274974823, + 0.025389019399881363, + 0.024946680292487144, + 0.03368540108203888 + ], + "credit_map": [ + 0.0060082413256168365, + 0.009470658376812935, + 0.009281563572585583, + 0.005774978082627058, + 0.006507248617708683, + 0.00910449493676424, + 0.006323865614831448, + 0.010705459862947464, + 0.010866876691579819, + 0.014826063066720963, + 0.010361512191593647, + 0.013797069899737835, + 0.009622636251151562, + 0.019086752086877823, + 0.018093949183821678, + 0.0172394011169672, + 0.024185094982385635, + 0.013434110209345818, + 0.05197684094309807, + 0.02020517736673355, + 0.006523598916828632, + 0.07896619290113449, + 0.01928533799946308, + 0.10376344621181488, + 0.006124406587332487, + 0.0055793579667806625, + 0.006021248642355204, + 0.008718525059521198, + 0.0034330065827816725, + 0.007541570346802473, + 0.0069672525860369205, + 0.007883413694798946, + 0.004780460149049759, + 0.004706407897174358, + 0.0062716929242014885, + 0.01392755750566721, + 0.0036879894323647022, + 0.012495256960391998, + 0.010179654695093632, + 0.019455276429653168, + 0.0023088292218744755, + 0.0008792065200395882, + 0.0018084635958075523, + 0.005497455131262541, + 0.0009884695755317807, + 0.0027612089179456234, + 0.002615379635244608, + 0.002753509907051921, + 0.10520505905151367, + 0.005082684103399515, + 0.1216578409075737, + 0.015940984711050987, + 0.006260093301534653, + 0.006150226574391127, + 0.005922406911849976, + 0.0069634499959647655, + 0.004595162346959114, + 0.0017858395585790277, + 0.005771829281002283, + 0.00331519334577024, + 0.0020926191937178373, + 0.0030307795386761427, + 0.002803975949063897, + 0.0019936237949877977, + 0.004689989145845175, + 0.0013945790706202388, + 0.0037795742973685265, + 0.0036487241741269827, + 0.0015176786109805107, + 0.003911941312253475, + 0.0027025272138416767, + 0.003301667282357812, + 0.004623151384294033, + 0.002369462512433529, + 0.005829624366015196, + 0.0036809234879910946, + 0.0024293679744005203, + 0.008259822614490986, + 0.0035969868768006563, + 0.0028960630297660828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.00464111752808094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_tensor([1838, 1838])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_tensor([1838, 1838])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ca6ed2d14651863f6f4d6a1d32af602c975d2694 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_tensor([1838, 1838])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly composed long shot, captured with a telephoto lens, holds the camera still, its focus fixed on a mournful aristocratic woman. She stands alone at the edge of a rugged, barren black volcanic cliff, clad in an intricately patterned silk kimono, embroidered with chrysanthemums and cranes in silver thread, symbols of both nobility and the transience of life. Behind her, the boundless, surging sea of white foam tumbles, its waves crashing again and again with thunderous force against the sharp, dark rocks at the cliff's base, sending a cascade of spray skyward. A harsh, low-angled light, like a stage spotlight, slices her profile from the side, casting deep, sharp shadows across her pale, delicate face, accentuating the sorrowful frown between her brows and the hollowness of her eyes. Wisps of wind-blown black hair brush against her cheeks, while the wide sleeves of her kimono rustle in the sharp, gusty wind. The entire scene presents the ultimate 1940s black-and-white film style. The strong dramatic contrast of light and shadow highlights the loneliness and despair of the characters. The picture has obvious graininess and silver salt texture, creating a grand, tragic and fatalistic visual atmosphere.", + "seed": 19272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_1838_seed_19272_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10835478082299232, + "vlm_logprob": -0.10835478082299232, + "vlm_yes_no_margin": 2.775000011920929, + "vlm_token_credit": -0.10835478082299232, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09211865067481995 + ], + "opposite_logprob_tokens": [ + -2.467118740081787 + ], + "logprob_reward": -0.09211865067481995, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09211865067481995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17276430130004883, + 0.09297222644090652, + 0.035258855670690536, + 0.16541136801242828, + 0.1487247794866562, + 0.1825224906206131, + 0.12330544739961624, + 0.020189331844449043, + 0.03274119272828102, + 0.02610999345779419 + ], + "credit_map": [ + 0.025406550616025925, + 0.02093045599758625, + 0.03180265426635742, + 0.015337597578763962, + 0.015298333950340748, + 0.02450375072658062, + 0.010926785878837109, + 0.02855818346142769, + 0.027781998738646507, + 0.020354438573122025, + 0.007615344598889351, + 0.011243230663239956, + 0.0059607247821986675, + 0.004314925521612167, + 0.00653055589646101, + 0.009170996025204659, + 0.004629615228623152, + 0.004567532334476709, + 0.005330503452569246, + 0.006258207838982344, + 0.0042092218063771725, + 0.0035860275384038687, + 0.0034272903576493263, + 0.003250456415116787, + 0.026673706248402596, + 0.01700119487941265, + 0.03320080786943436, + 0.0135391466319561, + 0.011227591894567013, + 0.024025391787290573, + 0.014417851343750954, + 0.02532566897571087, + 0.02660142071545124, + 0.01683913916349411, + 0.031023694202303886, + 0.01664801500737667, + 0.014229129068553448, + 0.015559467487037182, + 0.012481753714382648, + 0.015342168509960175, + 0.049346309155225754, + 0.015599243342876434, + 0.04627581313252449, + 0.01118235383182764, + 0.009653128683567047, + 0.023419950157403946, + 0.010004887357354164, + 0.01704082265496254, + 0.023067384958267212, + 0.011658073402941227, + 0.03342340886592865, + 0.008254359476268291, + 0.011439471505582333, + 0.015906017273664474, + 0.0075096022337675095, + 0.012047134339809418, + 0.002592179924249649, + 0.002888268092647195, + 0.0033475111704319715, + 0.0019908274989575148, + 0.0025701846461743116, + 0.0020594880916178226, + 0.0022154401522129774, + 0.0025254315696656704, + 0.00709399301558733, + 0.003591675776988268, + 0.005278591066598892, + 0.003217846155166626, + 0.00287074688822031, + 0.00312209315598011, + 0.00299273245036602, + 0.004573512822389603, + 0.005241722799837589, + 0.0026986575685441494, + 0.004286359529942274, + 0.0029373893048614264, + 0.002831432968378067, + 0.0020270217210054398, + 0.0023919434752315283, + 0.0036954667884856462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.09211865067481995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34882253408432007 + ], + "opposite_logprob_tokens": [ + -1.2238225936889648 + ], + "logprob_reward": -0.34882253408432007, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34882253408432007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1011979803442955, + 0.2944059669971466, + 0.0680084154009819, + 0.09845525771379471, + 0.10493194311857224, + 0.1437838077545166, + 0.057489704340696335, + 0.0529029443860054, + 0.025263013318181038, + 0.05356095731258392 + ], + "credit_map": [ + 0.014401025138795376, + 0.009746933355927467, + 0.017293864861130714, + 0.013995392248034477, + 0.009754769504070282, + 0.012334320694208145, + 0.007694669999182224, + 0.015977000817656517, + 0.0972307100892067, + 0.06432577967643738, + 0.02848116122186184, + 0.02978089638054371, + 0.01815635897219181, + 0.0204482302069664, + 0.014907661825418472, + 0.021075164899230003, + 0.008604127913713455, + 0.008096171543002129, + 0.008899963460862637, + 0.01522036548703909, + 0.007725139148533344, + 0.004962112754583359, + 0.009819233790040016, + 0.004681302700191736, + 0.014371367171406746, + 0.011755254119634628, + 0.016817986965179443, + 0.015747757628560066, + 0.008758279494941235, + 0.011988883838057518, + 0.009285259060561657, + 0.009730475954711437, + 0.02299368940293789, + 0.011768070049583912, + 0.01938789337873459, + 0.011130377650260925, + 0.008703405037522316, + 0.011647779494524002, + 0.008836203254759312, + 0.010464520193636417, + 0.03599774092435837, + 0.013848919421434402, + 0.03209071606397629, + 0.008954468183219433, + 0.011547132395207882, + 0.018230564892292023, + 0.00941642839461565, + 0.01369782816618681, + 0.010251722298562527, + 0.00461055850610137, + 0.01408883836120367, + 0.005136455409228802, + 0.005391170270740986, + 0.007292157504707575, + 0.0052139428444206715, + 0.005504854954779148, + 0.00813183281570673, + 0.005907313898205757, + 0.013647432439029217, + 0.0036503677256405354, + 0.004447950050234795, + 0.0075177340768277645, + 0.003432108089327812, + 0.006168206222355366, + 0.0033621869515627623, + 0.00231256615370512, + 0.0033067241311073303, + 0.003054487518966198, + 0.0042577781714499, + 0.0032813011202961206, + 0.0025981655344367027, + 0.0030898042023181915, + 0.008359844796359539, + 0.007676541805267334, + 0.008107865229249, + 0.006349949166178703, + 0.006449289619922638, + 0.0054624686017632484, + 0.005343012977391481, + 0.005811988841742277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.34882253408432007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581606775522232 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581606775522232, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581606775522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1005709245800972, + 0.028428731486201286, + 0.01708281598985195, + 0.25206586718559265, + 0.1855272501707077, + 0.33104875683784485, + 0.024854380637407303, + 0.016981730237603188, + 0.025056710466742516, + 0.01838281750679016 + ], + "credit_map": [ + 0.016004450619220734, + 0.01282338984310627, + 0.017728477716445923, + 0.008902143687009811, + 0.009756348095834255, + 0.012685607187449932, + 0.007818621583282948, + 0.014851890504360199, + 0.007053249981254339, + 0.00497840391471982, + 0.0029692589305341244, + 0.003204080043360591, + 0.0024134316481649876, + 0.003356314031407237, + 0.002369185909628868, + 0.0020848067943006754, + 0.002118931384757161, + 0.0018226571846753359, + 0.00226062024012208, + 0.0041952780447900295, + 0.0021111031528562307, + 0.001374543528072536, + 0.0018685912946239114, + 0.0013310923241078854, + 0.04231221228837967, + 0.02989017963409424, + 0.04913768544793129, + 0.018396617844700813, + 0.018635578453540802, + 0.04403820261359215, + 0.016481688246130943, + 0.033173710107803345, + 0.05061782896518707, + 0.01934712752699852, + 0.03473628684878349, + 0.016913186758756638, + 0.012152672745287418, + 0.018408628180623055, + 0.013370753265917301, + 0.019980769604444504, + 0.08612660318613052, + 0.022128460928797722, + 0.09412531554698944, + 0.01999693363904953, + 0.015775350853800774, + 0.05752447620034218, + 0.01253946777433157, + 0.022832157090306282, + 0.0043668258003890514, + 0.0032981240656226873, + 0.004691399168223143, + 0.002361405873671174, + 0.0030083388555794954, + 0.002408841624855995, + 0.0022899163886904716, + 0.002429528161883354, + 0.002371737500652671, + 0.0017785320524126291, + 0.00428615091368556, + 0.0013813559198752046, + 0.0023142355494201183, + 0.001475277473218739, + 0.0015034970128908753, + 0.0018709453288465738, + 0.0051656123250722885, + 0.0026570744812488556, + 0.003831347683444619, + 0.002272050827741623, + 0.002844045404344797, + 0.004267049953341484, + 0.0020746015943586826, + 0.0019449286628514528, + 0.00383205059915781, + 0.0016245105070993304, + 0.002822364680469036, + 0.0015734891640022397, + 0.0021306120324879885, + 0.0021622306667268276, + 0.0018112147226929665, + 0.0024263428058475256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.020581606775522232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062059711664915085 + ], + "opposite_logprob_tokens": [ + -2.8120596408843994 + ], + "logprob_reward": -0.062059711664915085, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062059711664915085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12467660009860992, + 0.07322428375482559, + 0.015738315880298615, + 0.3331065773963928, + 0.21544554829597473, + 0.09736450016498566, + 0.03077484853565693, + 0.0486714169383049, + 0.03015802428126335, + 0.030839968472719193 + ], + "credit_map": [ + 0.024642102420330048, + 0.022733991965651512, + 0.01955329068005085, + 0.010422970168292522, + 0.014361333101987839, + 0.010522443801164627, + 0.011180897243320942, + 0.01125956978648901, + 0.020546820014715195, + 0.013149915263056755, + 0.01179849449545145, + 0.007444498594850302, + 0.006883582565933466, + 0.004957397934049368, + 0.0043718889355659485, + 0.004071688279509544, + 0.0028111301362514496, + 0.0024751792661845684, + 0.0025796054396778345, + 0.0021140500903129578, + 0.0016675418009981513, + 0.0013903261860832572, + 0.0015808342723175883, + 0.0011196479899808764, + 0.09471554309129715, + 0.015881910920143127, + 0.10587739199399948, + 0.022619260475039482, + 0.018199196085333824, + 0.0386059507727623, + 0.01596883311867714, + 0.021238481625914574, + 0.049107614904642105, + 0.02162637747824192, + 0.03883742913603783, + 0.02397943288087845, + 0.015772156417369843, + 0.02507418394088745, + 0.01650644652545452, + 0.024541905149817467, + 0.017230335623025894, + 0.00891517661511898, + 0.02202356979250908, + 0.008250885643064976, + 0.012093517929315567, + 0.009161341935396194, + 0.009547472931444645, + 0.010142196901142597, + 0.005261982325464487, + 0.0025945447850972414, + 0.009031149558722973, + 0.0033457200042903423, + 0.0027991824317723513, + 0.0019666492007672787, + 0.0024267672561109066, + 0.0033488518092781305, + 0.00936805084347725, + 0.004908604547381401, + 0.005497750360518694, + 0.004108895547688007, + 0.00870435405522585, + 0.005123165436089039, + 0.006344520021229982, + 0.004616075661033392, + 0.004882451146841049, + 0.0034345469903200865, + 0.004874078091233969, + 0.0031183252576738596, + 0.00398979103192687, + 0.0046487972140312195, + 0.002697449177503586, + 0.0025125856045633554, + 0.00540110282599926, + 0.002586428774520755, + 0.004362695850431919, + 0.0030544514302164316, + 0.003831766778603196, + 0.004018737003207207, + 0.0029490317683666945, + 0.004635753575712442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.062059711664915085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191400915384293 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.018191400915384293, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191400915384293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16621024906635284, + 0.0572713278234005, + 0.01786707155406475, + 0.3838020861148834, + 0.1541212946176529, + 0.06826017796993256, + 0.047810547053813934, + 0.03288685902953148, + 0.049493931233882904, + 0.022276397794485092 + ], + "credit_map": [ + 0.02800983190536499, + 0.02584386244416237, + 0.02541377767920494, + 0.012302638031542301, + 0.019722741097211838, + 0.023157158866524696, + 0.013879089616239071, + 0.017881153151392937, + 0.02042960748076439, + 0.013501772657036781, + 0.006326234433799982, + 0.002927665365859866, + 0.003160072723403573, + 0.00479846028611064, + 0.0028946984093636274, + 0.003232821822166443, + 0.0018003791337832808, + 0.001984415343031287, + 0.003199739148840308, + 0.0028944527730345726, + 0.0022519370540976524, + 0.0016000006580725312, + 0.001920179696753621, + 0.002215967047959566, + 0.11244156211614609, + 0.02368254028260708, + 0.12218236178159714, + 0.023214491084218025, + 0.0210738442838192, + 0.04523906111717224, + 0.016043206676840782, + 0.01992504857480526, + 0.024366335943341255, + 0.016499429941177368, + 0.03458959609270096, + 0.018256530165672302, + 0.015840429812669754, + 0.01246958039700985, + 0.01331723015755415, + 0.018782174214720726, + 0.014320033602416515, + 0.007874049246311188, + 0.016501516103744507, + 0.004707328509539366, + 0.005818601232022047, + 0.007198796141892672, + 0.004597035236656666, + 0.007242822088301182, + 0.00836748443543911, + 0.004486215300858021, + 0.008696927689015865, + 0.005359967704862356, + 0.0064682927913963795, + 0.004428273532539606, + 0.005454637110233307, + 0.004548748955130577, + 0.0038758383598178625, + 0.0036722461227327585, + 0.006138606462627649, + 0.0033704559318721294, + 0.004953639581799507, + 0.003029119223356247, + 0.003717367770150304, + 0.004129582550376654, + 0.010853441432118416, + 0.004587945528328419, + 0.009684878401458263, + 0.00345550081692636, + 0.005342741962522268, + 0.004293479491025209, + 0.003931637853384018, + 0.007344305515289307, + 0.0039123655296862125, + 0.003070036182180047, + 0.0030114315450191498, + 0.002712635323405266, + 0.0030253552831709385, + 0.002133576665073633, + 0.0023510949686169624, + 0.0020599018316715956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.018191400915384293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly composed long shot, captured with a telephoto lens, holds the camera still, its focus fixed on a mournful aristocratic woman. She stands alone at the edge of a rugged, barren black volcanic cliff, clad in an intricately patterned silk kimono, embroidered with chrysanthemums and cranes in silver thread, symbols of both nobility and the transience of life. Behind her, the boundless, surging sea of white foam tumbles, its waves crashing again and again with thunderous force against the sharp, dark rocks at the cliff's base, sending a cascade of spray skyward. A harsh, low-angled light, like a stage spotlight, slices her profile from the side, casting deep, sharp shadows across her pale, delicate face, accentuating the sorrowful frown between her brows and the hollowness of her eyes. Wisps of wind-blown black hair brush against her cheeks, while the wide sleeves of her kimono rustle in the sharp, gusty wind. The entire scene presents the ultimate 1940s black-and-white film style. The strong dramatic contrast of light and shadow highlights the loneliness and despair of the characters. The picture has obvious graininess and silver salt texture, creating a grand, tragic and fatalistic visual atmosphere.", + "seed": 19273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_1838_seed_19273_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09707638276740908, + "vlm_logprob": -0.09707638276740908, + "vlm_yes_no_margin": 2.924999988079071, + "vlm_token_credit": -0.09707638276740908, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04710569232702255 + ], + "opposite_logprob_tokens": [ + -3.1721057891845703 + ], + "logprob_reward": -0.04710569232702255, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04710569232702255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12781378626823425, + 0.5357713103294373, + 0.03578229248523712, + 0.030596666038036346, + 0.08083564788103104, + 0.07727556675672531, + 0.020944664254784584, + 0.024373168125748634, + 0.02405736595392227, + 0.04254944995045662 + ], + "credit_map": [ + 0.02091883309185505, + 0.015159622766077518, + 0.009976469911634922, + 0.027370065450668335, + 0.015388851054012775, + 0.01954398863017559, + 0.009318520314991474, + 0.010137438774108887, + 0.15517498552799225, + 0.0921400710940361, + 0.040775928646326065, + 0.1342468559741974, + 0.0207278523594141, + 0.0576469860970974, + 0.017115464434027672, + 0.01794317364692688, + 0.003469163319095969, + 0.0030457035172730684, + 0.0038985840510576963, + 0.0066628530621528625, + 0.006227003876119852, + 0.005317069590091705, + 0.0036915128584951162, + 0.0034704022109508514, + 0.0039419555105268955, + 0.002935715951025486, + 0.0035855958703905344, + 0.004425199702382088, + 0.003963634371757507, + 0.005537036806344986, + 0.0025537426117807627, + 0.0036537847481667995, + 0.009028437547385693, + 0.00831515435129404, + 0.009144426323473454, + 0.021827897056937218, + 0.006648902781307697, + 0.013037314638495445, + 0.006042719818651676, + 0.006790796294808388, + 0.009153601713478565, + 0.009676382876932621, + 0.009308052249252796, + 0.018400950357317924, + 0.0059062703512609005, + 0.01277895737439394, + 0.0055832997895777225, + 0.006468056235462427, + 0.0027336080092936754, + 0.003702897811308503, + 0.0022188154980540276, + 0.0020063198171555996, + 0.0020107515156269073, + 0.00225046812556684, + 0.0025675762444734573, + 0.0034542272333055735, + 0.0033839233219623566, + 0.0034054561983793974, + 0.0035525320563465357, + 0.0029772717971354723, + 0.001650070073083043, + 0.002292627701535821, + 0.003093962324783206, + 0.004017324186861515, + 0.004817958455532789, + 0.0064282906241714954, + 0.004492724314332008, + 0.002086282940581441, + 0.0019089061534032226, + 0.0016444617649540305, + 0.001388715230859816, + 0.0012900267029181123, + 0.008233350701630116, + 0.011129318736493587, + 0.009297432377934456, + 0.0036483241710811853, + 0.0027153226546943188, + 0.002651308197528124, + 0.0019467463716864586, + 0.0029276476707309484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.04710569232702255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31362324953079224 + ], + "opposite_logprob_tokens": [ + -1.3136231899261475 + ], + "logprob_reward": -0.31362324953079224, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31362324953079224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20694568753242493, + 0.28621798753738403, + 0.06874196976423264, + 0.046972110867500305, + 0.10740068554878235, + 0.1496330201625824, + 0.01938907988369465, + 0.04545839875936508, + 0.04209541529417038, + 0.027145616710186005 + ], + "credit_map": [ + 0.03945424035191536, + 0.01931142620742321, + 0.022294972091913223, + 0.05120714381337166, + 0.017245925962924957, + 0.027747150510549545, + 0.013908960856497288, + 0.015775876119732857, + 0.07383902370929718, + 0.04836980253458023, + 0.023009181022644043, + 0.05597490072250366, + 0.02110671065747738, + 0.031269099563360214, + 0.015521887689828873, + 0.017127403989434242, + 0.0051436410285532475, + 0.005417002364993095, + 0.00618400052189827, + 0.007591983303427696, + 0.012586312368512154, + 0.009498871862888336, + 0.011599658988416195, + 0.010720503516495228, + 0.006660799961537123, + 0.005296952556818724, + 0.008453592658042908, + 0.005233660340309143, + 0.007135632447898388, + 0.006944663822650909, + 0.004098180681467056, + 0.0031486295629292727, + 0.009585303254425526, + 0.011723496951162815, + 0.008954617194831371, + 0.01182081364095211, + 0.019595570862293243, + 0.011680887080729008, + 0.015814263373613358, + 0.018225738778710365, + 0.01857062801718712, + 0.013377336785197258, + 0.013371429406106472, + 0.013372190296649933, + 0.027882169932127, + 0.023533865809440613, + 0.02114579826593399, + 0.01837959699332714, + 0.002570372074842453, + 0.0033647799864411354, + 0.0024467436596751213, + 0.0018251033034175634, + 0.0016023056814447045, + 0.0016258582472801208, + 0.0025540711358189583, + 0.0033998466096818447, + 0.005799636710435152, + 0.007469530217349529, + 0.007298616226762533, + 0.00577961653470993, + 0.0024683044757694006, + 0.003656073473393917, + 0.006047266535460949, + 0.006939353421330452, + 0.005214266944676638, + 0.008468402549624443, + 0.011833886615931988, + 0.005048913415521383, + 0.002854958176612854, + 0.0028062222991138697, + 0.0023301246110349894, + 0.0035386381205171347, + 0.004293740727007389, + 0.0037370449863374233, + 0.006290150340646505, + 0.002919003600254655, + 0.00231171608902514, + 0.0017873708857223392, + 0.0029731332324445248, + 0.002833455102518201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.31362324953079224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631906773895025 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631906773895025, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631906773895025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06805627048015594, + 0.5003650188446045, + 0.02180500514805317, + 0.06012745574116707, + 0.07240675389766693, + 0.03387707471847534, + 0.03590391203761101, + 0.021967574954032898, + 0.06703174114227295, + 0.11845923215150833 + ], + "credit_map": [ + 0.006860062479972839, + 0.009588652290403843, + 0.012183026410639286, + 0.012718886137008667, + 0.00629356550052762, + 0.008138769306242466, + 0.00544402003288269, + 0.006829284131526947, + 0.14195984601974487, + 0.07650285959243774, + 0.0424296036362648, + 0.13919268548488617, + 0.013133483938872814, + 0.05443254113197327, + 0.01613774336874485, + 0.01657624915242195, + 0.0023029313888400793, + 0.002005583606660366, + 0.003060627030208707, + 0.004626919981092215, + 0.0027977307327091694, + 0.0027049249038100243, + 0.0024361247196793556, + 0.0018701638327911496, + 0.008870706893503666, + 0.006482729222625494, + 0.009576743468642235, + 0.013578608632087708, + 0.00479225255548954, + 0.007575908675789833, + 0.003782651387155056, + 0.005467848386615515, + 0.0078107574954628944, + 0.006736278533935547, + 0.012221109122037888, + 0.01971062645316124, + 0.005085848271846771, + 0.0116449985653162, + 0.00397264352068305, + 0.005224492866545916, + 0.004038605839014053, + 0.0037584281526505947, + 0.007909731939435005, + 0.006430058740079403, + 0.0029054286424070597, + 0.004803183022886515, + 0.0019652366172522306, + 0.0020664026960730553, + 0.007483051624149084, + 0.008265329524874687, + 0.006291674450039864, + 0.0043294187635183334, + 0.002652300288900733, + 0.002578401006758213, + 0.0019747442565858364, + 0.0023289902601391077, + 0.004158299881964922, + 0.0048354677855968475, + 0.004924650304019451, + 0.002685168059542775, + 0.001277257571928203, + 0.001275615650229156, + 0.0013549946015700698, + 0.0014561207499355078, + 0.009031359106302261, + 0.013271046802401543, + 0.019681459292769432, + 0.008356661535799503, + 0.0039044423028826714, + 0.0036894110962748528, + 0.0039396025240421295, + 0.005157753825187683, + 0.03227195888757706, + 0.017373541370034218, + 0.0354265458881855, + 0.008756431750953197, + 0.005082257557660341, + 0.007061441894620657, + 0.005481301341205835, + 0.007005753926932812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.007631906773895025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969754070043564 + ], + "opposite_logprob_tokens": [ + -2.9299697875976562 + ], + "logprob_reward": -0.054969754070043564, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969754070043564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09627164900302887, + 0.6377996206283569, + 0.03126508742570877, + 0.012448581866919994, + 0.07841450721025467, + 0.051731910556554794, + 0.02255425602197647, + 0.03260992839932442, + 0.0178492683917284, + 0.01905510015785694 + ], + "credit_map": [ + 0.02134321630001068, + 0.009892442263662815, + 0.008488819003105164, + 0.03115733154118061, + 0.005318368785083294, + 0.010606886819005013, + 0.00546614034101367, + 0.003998451866209507, + 0.23882700502872467, + 0.23084326088428497, + 0.02832988277077675, + 0.045174408704042435, + 0.03000461682677269, + 0.03628425672650337, + 0.0137095982208848, + 0.014626682735979557, + 0.00413873977959156, + 0.003096366999670863, + 0.0035992662888020277, + 0.007412544917315245, + 0.0039899833500385284, + 0.0035320473834872246, + 0.0022234306670725346, + 0.0032727126963436604, + 0.0019421702018007636, + 0.001480471109971404, + 0.0018176422454416752, + 0.0020852982997894287, + 0.0012924966868013144, + 0.0011850196169689298, + 0.0010305633768439293, + 0.0016149207949638367, + 0.012173191644251347, + 0.009018689393997192, + 0.01322589535266161, + 0.015214657410979271, + 0.0070326426066458225, + 0.008233926258981228, + 0.007050799205899239, + 0.006464706733822823, + 0.007700616959482431, + 0.007673276588320732, + 0.004426146391779184, + 0.010652178898453712, + 0.006911452393978834, + 0.007560119032859802, + 0.00305825169198215, + 0.0037498781457543373, + 0.003262408310547471, + 0.005138039123266935, + 0.0036101555451750755, + 0.002838514046743512, + 0.002134589944034815, + 0.002071542665362358, + 0.0012969707604497671, + 0.002202036790549755, + 0.0042554582469165325, + 0.004692342597991228, + 0.004908635281026363, + 0.0044118789955973625, + 0.0031859129667282104, + 0.0025417052675038576, + 0.005042408127337694, + 0.003571593901142478, + 0.0026310125831514597, + 0.0027279818896204233, + 0.004334317985922098, + 0.001880208496004343, + 0.0017686154460534453, + 0.001699149957858026, + 0.0012837940594181418, + 0.0015241883229464293, + 0.002859455067664385, + 0.004662750288844109, + 0.002149957697838545, + 0.0021698670461773872, + 0.0013596286298707128, + 0.001449208240956068, + 0.0021254310850054026, + 0.0022788038477301598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.054969754070043564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205131113529205 + ], + "opposite_logprob_tokens": [ + -2.812051296234131 + ], + "logprob_reward": -0.06205131113529205, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205131113529205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17192190885543823, + 0.6243790984153748, + 0.04193528741598129, + 0.01580938883125782, + 0.021531101316213608, + 0.053962480276823044, + 0.015702102333307266, + 0.03054114803671837, + 0.013332325033843517, + 0.010885152965784073 + ], + "credit_map": [ + 0.018666833639144897, + 0.032919444143772125, + 0.03579175844788551, + 0.02719326876103878, + 0.014601458795368671, + 0.01780027523636818, + 0.010388222523033619, + 0.014560660347342491, + 0.19613036513328552, + 0.13439054787158966, + 0.043769702315330505, + 0.13537830114364624, + 0.019371503964066505, + 0.059115055948495865, + 0.017789354547858238, + 0.018434299156069756, + 0.00744623551145196, + 0.004490950144827366, + 0.005670231766998768, + 0.010921236127614975, + 0.004416539333760738, + 0.003219630802050233, + 0.0027329381555318832, + 0.003037522081285715, + 0.0024302098900079727, + 0.0016576071502640843, + 0.0022210017777979374, + 0.0023614433594048023, + 0.0018254688475281, + 0.0020748868118971586, + 0.0015729098813608289, + 0.0016658612294122577, + 0.0034307711757719517, + 0.0026030607987195253, + 0.003405975177884102, + 0.0027493988163769245, + 0.002068232512101531, + 0.002511247992515564, + 0.0023087586741894484, + 0.0024536550045013428, + 0.008062897250056267, + 0.009561263956129551, + 0.0062446799129247665, + 0.012128799222409725, + 0.0034165633842349052, + 0.0067437151446938515, + 0.003911888226866722, + 0.0038926759734749794, + 0.0023968403693288565, + 0.0033392743207514286, + 0.002395144896581769, + 0.001953503116965294, + 0.001343763549812138, + 0.0012314971536397934, + 0.0013125065015628934, + 0.0017295735888183117, + 0.004071834497153759, + 0.004271331243216991, + 0.005510349757969379, + 0.004066984169185162, + 0.0025751525536179543, + 0.003004272934049368, + 0.0035156093072146177, + 0.0035256140399724245, + 0.0017009255243465304, + 0.0020459757652133703, + 0.001933250343427062, + 0.0026656887494027615, + 0.0011305981315672398, + 0.0012663675006479025, + 0.0010346168419346213, + 0.0015549027593806386, + 0.0015181185444816947, + 0.0023709821980446577, + 0.0016536416951566935, + 0.0010097523918375373, + 0.0009175732266157866, + 0.0011744450312107801, + 0.0008755060844123363, + 0.0013651333283632994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.06205131113529205 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly composed long shot, captured with a telephoto lens, holds the camera still, its focus fixed on a mournful aristocratic woman. She stands alone at the edge of a rugged, barren black volcanic cliff, clad in an intricately patterned silk kimono, embroidered with chrysanthemums and cranes in silver thread, symbols of both nobility and the transience of life. Behind her, the boundless, surging sea of white foam tumbles, its waves crashing again and again with thunderous force against the sharp, dark rocks at the cliff's base, sending a cascade of spray skyward. A harsh, low-angled light, like a stage spotlight, slices her profile from the side, casting deep, sharp shadows across her pale, delicate face, accentuating the sorrowful frown between her brows and the hollowness of her eyes. Wisps of wind-blown black hair brush against her cheeks, while the wide sleeves of her kimono rustle in the sharp, gusty wind. The entire scene presents the ultimate 1940s black-and-white film style. The strong dramatic contrast of light and shadow highlights the loneliness and despair of the characters. The picture has obvious graininess and silver salt texture, creating a grand, tragic and fatalistic visual atmosphere.", + "seed": 19274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_1838_seed_19274_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1747742708772421, + "vlm_logprob": -0.1747742708772421, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.1747742708772421, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08249976485967636 + ], + "opposite_logprob_tokens": [ + -2.5824997425079346 + ], + "logprob_reward": -0.08249976485967636, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08249976485967636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8176175951957703, + 0.05147286877036095, + 0.020795337855815887, + 0.01560833677649498, + 0.01799771375954151, + 0.02073480747640133, + 0.010583451949059963, + 0.01637987233698368, + 0.009050961583852768, + 0.01975906640291214 + ], + "credit_map": [ + 0.29472610354423523, + 0.05328560993075371, + 0.26024919748306274, + 0.07000511139631271, + 0.03940675035119057, + 0.054351504892110825, + 0.019346315413713455, + 0.026247067376971245, + 0.006332543212920427, + 0.006677067838609219, + 0.005822440609335899, + 0.004276747349649668, + 0.007522277068346739, + 0.005858088377863169, + 0.005096422974020243, + 0.009887286461889744, + 0.003192250384017825, + 0.002726938109844923, + 0.0023911590687930584, + 0.0019934072624891996, + 0.0038735116831958294, + 0.002148692263290286, + 0.0019979842472821474, + 0.002471397863700986, + 0.0018993946723639965, + 0.002032177522778511, + 0.0019609329756349325, + 0.0014725272776558995, + 0.001836653333157301, + 0.00191358698066324, + 0.0018014658708125353, + 0.002691598841920495, + 0.00284240092150867, + 0.002088123932480812, + 0.001991672208532691, + 0.002161602256819606, + 0.0026980999391525984, + 0.0023931534960865974, + 0.001725463313050568, + 0.002097198273986578, + 0.0033448119647800922, + 0.00188051990699023, + 0.002283930778503418, + 0.0023277855943888426, + 0.0028502237983047962, + 0.004051190800964832, + 0.0015393418725579977, + 0.002457005437463522, + 0.001729099778458476, + 0.001718251034617424, + 0.0016559467185288668, + 0.0009077413706108928, + 0.00124543032143265, + 0.001697183819487691, + 0.000662029895465821, + 0.0009677701164036989, + 0.0027282903902232647, + 0.0021975550334900618, + 0.0015904398169368505, + 0.0017034001648426056, + 0.002186596393585205, + 0.0027320890221744776, + 0.0010549171129241586, + 0.0021865854505449533, + 0.0013380766613408923, + 0.0016385044436901808, + 0.0012312415055930614, + 0.0008176574483513832, + 0.001265611033886671, + 0.0010009799152612686, + 0.0006835369858890772, + 0.0010753544047474861, + 0.003601543139666319, + 0.0029067527502775192, + 0.002720365533605218, + 0.002691355999559164, + 0.001773897442035377, + 0.002657249104231596, + 0.0014106754679232836, + 0.001997230341657996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08249976485967636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6935060620307922 + ], + "opposite_logprob_tokens": [ + -0.6935060620307922 + ], + "logprob_reward": -0.6935060620307922, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6935060620307922, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5051791667938232, + 0.07319601625204086, + 0.041284456849098206, + 0.032578472048044205, + 0.020478056743741035, + 0.02067125402390957, + 0.050860773772001266, + 0.06644399464130402, + 0.0750691369175911, + 0.1142386645078659 + ], + "credit_map": [ + 0.16904756426811218, + 0.038661058992147446, + 0.16283728182315826, + 0.045687466859817505, + 0.02228570356965065, + 0.034894201904535294, + 0.01366744376718998, + 0.018098456785082817, + 0.009238543920218945, + 0.009066799655556679, + 0.007667667232453823, + 0.005701021756976843, + 0.007346753496676683, + 0.010121499188244343, + 0.0077218348160386086, + 0.016331899911165237, + 0.004696007817983627, + 0.004976664669811726, + 0.005865038372576237, + 0.0040290867909789085, + 0.0038369789253920317, + 0.004407538566738367, + 0.004416889511048794, + 0.00905624870210886, + 0.004203522112220526, + 0.003648047801107168, + 0.004378981422632933, + 0.0039632562547922134, + 0.0035014464519917965, + 0.00347158988006413, + 0.00306781567633152, + 0.006343812216073275, + 0.002763489494100213, + 0.003008204745128751, + 0.0024211574345827103, + 0.0015839816769585013, + 0.003500832710415125, + 0.003686829237267375, + 0.0014779862249270082, + 0.0020355742890387774, + 0.002783923177048564, + 0.0018735510529950261, + 0.0024020939599722624, + 0.003368291538208723, + 0.003787580644711852, + 0.0029189190827310085, + 0.0020251392852514982, + 0.001511755515821278, + 0.00788117479532957, + 0.009574740193784237, + 0.0059656105004251, + 0.004378358833491802, + 0.007105366326868534, + 0.008193080313503742, + 0.0033024479635059834, + 0.004459993913769722, + 0.009999880567193031, + 0.012314174324274063, + 0.005842159502208233, + 0.005467981565743685, + 0.010947316884994507, + 0.00824123714119196, + 0.005324260331690311, + 0.008306982927024364, + 0.010831054300069809, + 0.01270085759460926, + 0.00650691008195281, + 0.009968082420527935, + 0.008210115134716034, + 0.008214980363845825, + 0.005824171472340822, + 0.01281296368688345, + 0.01893860474228859, + 0.013421637937426567, + 0.02123536914587021, + 0.015518681146204472, + 0.010616227053105831, + 0.009815964847803116, + 0.012086663395166397, + 0.0126055171713233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.6935060620307922 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083701625466347 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083701625466347, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083701625466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8460479378700256, + 0.037191566079854965, + 0.026412473991513252, + 0.014622276648879051, + 0.010137377306818962, + 0.00517019210383296, + 0.009522728621959686, + 0.030984356999397278, + 0.00598907470703125, + 0.013922152109444141 + ], + "credit_map": [ + 0.3035387098789215, + 0.054998427629470825, + 0.2705594003200531, + 0.07296034693717957, + 0.03944599628448486, + 0.05797289311885834, + 0.020105404779314995, + 0.0264666136354208, + 0.004373321775346994, + 0.0064665572717785835, + 0.007362162694334984, + 0.002757256617769599, + 0.004373728763312101, + 0.005971286911517382, + 0.0019391902023926377, + 0.0039480519481003284, + 0.003439820371568203, + 0.004980714991688728, + 0.004291079472750425, + 0.0018111042445525527, + 0.0023952408228069544, + 0.0025952933356165886, + 0.0020308829843997955, + 0.0048683322966098785, + 0.0022616733331233263, + 0.0024398760870099068, + 0.0025401695165783167, + 0.0014023934490978718, + 0.0012275127228349447, + 0.0014276603469625115, + 0.0011376641923561692, + 0.0021853239741176367, + 0.0015592538984492421, + 0.0012553943088278174, + 0.0011625952320173383, + 0.0010826688958331943, + 0.001523709623143077, + 0.0013865121873095632, + 0.0009295387426391244, + 0.0012377030216157436, + 0.0007030543056316674, + 0.00048300527851097286, + 0.0006744649144820869, + 0.0006399133708328009, + 0.0007563011604361236, + 0.0007542272796854377, + 0.0005498115788213909, + 0.0006094129057601094, + 0.0016682044370099902, + 0.0011293726274743676, + 0.0019383778562769294, + 0.000747097481507808, + 0.0010704101296141744, + 0.0017837554914876819, + 0.000491051294375211, + 0.0006944572669453919, + 0.006099895108491182, + 0.0038950557354837656, + 0.003441523527726531, + 0.002063651103526354, + 0.00348048098385334, + 0.007296412717550993, + 0.0013607193250209093, + 0.0033466117456555367, + 0.0010174794588238, + 0.0010620520915836096, + 0.0009961166651919484, + 0.0005033834022469819, + 0.0005750431446358562, + 0.001063886913470924, + 0.0003087080258410424, + 0.00046240357914939523, + 0.003093860810622573, + 0.002429583575576544, + 0.0024433848448097706, + 0.0012967790244147182, + 0.0011870918096974492, + 0.0010938063496723771, + 0.0011089369654655457, + 0.0012687055859714746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011083701625466347 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811853006482124 + ], + "opposite_logprob_tokens": [ + -3.288118600845337 + ], + "logprob_reward": -0.03811853006482124, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811853006482124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8893856406211853, + 0.02297307923436165, + 0.011953875422477722, + 0.01029905490577221, + 0.029746320098638535, + 0.0047933259047567844, + 0.004387326538562775, + 0.007241035811603069, + 0.007322121877223253, + 0.011898299679160118 + ], + "credit_map": [ + 0.32005032896995544, + 0.05774753540754318, + 0.2839570641517639, + 0.07676858454942703, + 0.04145663231611252, + 0.06034581735730171, + 0.020904170349240303, + 0.02815546840429306, + 0.0027804928831756115, + 0.0033219007309526205, + 0.0021836101077497005, + 0.0021143315825611353, + 0.0022467353846877813, + 0.002413222100585699, + 0.0026721726171672344, + 0.005240613128989935, + 0.0015903367893770337, + 0.002191344043239951, + 0.0018260785145685077, + 0.0009200999047607183, + 0.0013115102192386985, + 0.0015737279318273067, + 0.0008354383753612638, + 0.0017053383635357022, + 0.0014070976758375764, + 0.0016108708223327994, + 0.0012908129720017314, + 0.0008628531941212714, + 0.0015240403590723872, + 0.0014978518011048436, + 0.001136908889748156, + 0.0009686183766461909, + 0.0042076692916452885, + 0.003494103904813528, + 0.0027122879400849342, + 0.003664741525426507, + 0.0059803929179906845, + 0.0035683943424373865, + 0.003238166216760874, + 0.0028805630281567574, + 0.0006837064283899963, + 0.0004553495382424444, + 0.0006936537683941424, + 0.0006808792240917683, + 0.0006606818060390651, + 0.0008305125520564616, + 0.00043053628178313375, + 0.0003580055199563503, + 0.0005673474515788257, + 0.0005255580763332546, + 0.0007795538986101747, + 0.00036974321119487286, + 0.000579783518332988, + 0.0006642558728344738, + 0.0005022443365305662, + 0.0003988399403169751, + 0.0011112723732367158, + 0.0006239944486878812, + 0.0006404136656783521, + 0.0010325744515284896, + 0.0015773327322676778, + 0.0004833606362808496, + 0.0008604456088505685, + 0.0009116408764384687, + 0.0011846713023260236, + 0.0012870695209130645, + 0.0008285182411782444, + 0.0005756058380939066, + 0.0011474081547930837, + 0.0007199255051091313, + 0.0005948422476649284, + 0.0009840800194069743, + 0.002260600682348013, + 0.0025018397718667984, + 0.0010107160778716207, + 0.0014844232937321067, + 0.0012389373732730746, + 0.0015373656060546637, + 0.0008162306039594114, + 0.0010481850476935506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03811853006482124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048663295805454254 + ], + "opposite_logprob_tokens": [ + -3.048663377761841 + ], + "logprob_reward": -0.048663295805454254, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048663295805454254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9539278745651245, + 0.008363132365047932, + 0.006301731336861849, + 0.007680266629904509, + 0.011716362088918686, + 0.0011419320944696665, + 0.002154161687940359, + 0.002268889220431447, + 0.0018282397650182247, + 0.0046173869632184505 + ], + "credit_map": [ + 0.343066930770874, + 0.062002506107091904, + 0.30471232533454895, + 0.08201632648706436, + 0.04440685734152794, + 0.06529068201780319, + 0.022418633103370667, + 0.030013615265488625, + 0.001019289018586278, + 0.0013669568579643965, + 0.0011142621515318751, + 0.0007103417301550508, + 0.001174245961010456, + 0.0009941194439306855, + 0.0006410653586499393, + 0.0013428527163341641, + 0.0008249949314631522, + 0.0008779634372331202, + 0.0010517213959246874, + 0.0005671830731444061, + 0.0007512221345677972, + 0.0008960772538557649, + 0.0005740742199122906, + 0.0007584950653836131, + 0.0010388822993263602, + 0.0012652644654735923, + 0.0009206980466842651, + 0.0006420075078494847, + 0.0008770099957473576, + 0.0013426968362182379, + 0.0006514212000183761, + 0.0009422869770787656, + 0.0017069445457309484, + 0.0012058481806889176, + 0.0018189530819654465, + 0.001415139646269381, + 0.0015247089322656393, + 0.0018712348537519574, + 0.000830736244097352, + 0.0013427975354716182, + 0.00015335521311499178, + 0.0001495944889029488, + 0.00010190736065851524, + 0.0001202229323098436, + 0.0002147828199667856, + 0.00018321236711926758, + 0.00013639726967085153, + 8.245972276199609e-05, + 0.0003236789780203253, + 0.0002626360801514238, + 0.00039842454134486616, + 0.00029024347895756364, + 0.00024238909827545285, + 0.0002962763828691095, + 0.00019022337801288813, + 0.0001502900995546952, + 0.00030174441053532064, + 0.0002238182059954852, + 0.0003368394682183862, + 0.0003348156460560858, + 0.00032134499633684754, + 0.00022289797198027372, + 0.0001877694157883525, + 0.0003396593383513391, + 0.0002374745236011222, + 0.00016839314776007086, + 0.00029479380464181304, + 0.0002751780557446182, + 0.0002653533883858472, + 0.00014778613694943488, + 0.00022568597341887653, + 0.00021357483637984842, + 0.0006121255573816597, + 0.0007536867051385343, + 0.0006941136671230197, + 0.0005287962849251926, + 0.0006025850889272988, + 0.0007080505019985139, + 0.00038574516656808555, + 0.0003322843986097723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.048663295805454254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly composed long shot, captured with a telephoto lens, holds the camera still, its focus fixed on a mournful aristocratic woman. She stands alone at the edge of a rugged, barren black volcanic cliff, clad in an intricately patterned silk kimono, embroidered with chrysanthemums and cranes in silver thread, symbols of both nobility and the transience of life. Behind her, the boundless, surging sea of white foam tumbles, its waves crashing again and again with thunderous force against the sharp, dark rocks at the cliff's base, sending a cascade of spray skyward. A harsh, low-angled light, like a stage spotlight, slices her profile from the side, casting deep, sharp shadows across her pale, delicate face, accentuating the sorrowful frown between her brows and the hollowness of her eyes. Wisps of wind-blown black hair brush against her cheeks, while the wide sleeves of her kimono rustle in the sharp, gusty wind. The entire scene presents the ultimate 1940s black-and-white film style. The strong dramatic contrast of light and shadow highlights the loneliness and despair of the characters. The picture has obvious graininess and silver salt texture, creating a grand, tragic and fatalistic visual atmosphere.", + "seed": 19275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_3/index_1838_seed_19275_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12994861444458367, + "vlm_logprob": -0.12994861444458367, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.12994861444458367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032616499811410904 + ], + "opposite_logprob_tokens": [ + -3.53261661529541 + ], + "logprob_reward": -0.032616499811410904, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.032616499811410904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36869630217552185, + 0.07455715537071228, + 0.14558027684688568, + 0.15274186432361603, + 0.04985515773296356, + 0.044263552874326706, + 0.036946527659893036, + 0.05690059810876846, + 0.046485431492328644, + 0.023973172530531883 + ], + "credit_map": [ + 0.034210894256830215, + 0.10306400805711746, + 0.03186658024787903, + 0.051152873784303665, + 0.031116055324673653, + 0.03783509135246277, + 0.027191367000341415, + 0.0522594228386879, + 0.005767177790403366, + 0.012914839200675488, + 0.008783897385001183, + 0.01213041041046381, + 0.00971140619367361, + 0.007038039620965719, + 0.0046250480227172375, + 0.013586332090198994, + 0.0141060221940279, + 0.01242874190211296, + 0.02791191264986992, + 0.015674900263547897, + 0.013569067232310772, + 0.02028821036219597, + 0.021272238343954086, + 0.02032918855547905, + 0.01868671365082264, + 0.013407453894615173, + 0.018583878874778748, + 0.014076357707381248, + 0.02839699387550354, + 0.013849952258169651, + 0.01117402408272028, + 0.034566495567560196, + 0.005103978794068098, + 0.0037898935843259096, + 0.007049592211842537, + 0.0056534442119300365, + 0.00624535558745265, + 0.00974647980183363, + 0.006581799127161503, + 0.005684610456228256, + 0.008550172671675682, + 0.00498930411413312, + 0.004479526542127132, + 0.005042508244514465, + 0.006971071939915419, + 0.004379518795758486, + 0.003409570548683405, + 0.006441881414502859, + 0.004637481644749641, + 0.0053138453513383865, + 0.005372720770537853, + 0.005234540905803442, + 0.004627218935638666, + 0.00412724306806922, + 0.0033065069001168013, + 0.004326972179114819, + 0.007036325056105852, + 0.005464094690978527, + 0.007722529582679272, + 0.008814598433673382, + 0.005941310431808233, + 0.007996209897100925, + 0.0053690639324486256, + 0.008556469343602657, + 0.003984300885349512, + 0.004714256152510643, + 0.01327090710401535, + 0.0061958045698702335, + 0.003665544791147113, + 0.006294783670455217, + 0.002610280644148588, + 0.00574955390766263, + 0.0020147685427218676, + 0.0037102631758898497, + 0.002703002654016018, + 0.00437574740499258, + 0.0027134709525853395, + 0.002254270017147064, + 0.002868122886866331, + 0.003333526896312833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.032616499811410904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576371967792511 + ], + "opposite_logprob_tokens": [ + -0.826371967792511 + ], + "logprob_reward": -0.576371967792511, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576371967792511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22058390080928802, + 0.06676595658063889, + 0.06607075780630112, + 0.20862600207328796, + 0.05709146335721016, + 0.06421905755996704, + 0.0980914756655693, + 0.08997053653001785, + 0.08946298062801361, + 0.03911786153912544 + ], + "credit_map": [ + 0.022259995341300964, + 0.04182470217347145, + 0.02515507861971855, + 0.03904608264565468, + 0.017234548926353455, + 0.021189169958233833, + 0.024749282747507095, + 0.02912503480911255, + 0.007396414410322905, + 0.009232018142938614, + 0.005408651195466518, + 0.009072543121874332, + 0.010568848811089993, + 0.007807176094502211, + 0.005211384501308203, + 0.012068921700119972, + 0.005936332046985626, + 0.006930375937372446, + 0.010884562507271767, + 0.007000243756920099, + 0.006034285761415958, + 0.007867853157222271, + 0.012550337240099907, + 0.008866765536367893, + 0.029486721381545067, + 0.013725238852202892, + 0.03276931867003441, + 0.01653992384672165, + 0.03410112485289574, + 0.02809896692633629, + 0.010497260838747025, + 0.04340745508670807, + 0.005993694998323917, + 0.004549698904156685, + 0.00765990512445569, + 0.009744769893586636, + 0.00755723612383008, + 0.006822923664003611, + 0.00946157518774271, + 0.005301658529788256, + 0.007876659743487835, + 0.008314494043588638, + 0.007826386950910091, + 0.01234485488384962, + 0.008196162059903145, + 0.005951886996626854, + 0.008794267661869526, + 0.004914342891424894, + 0.010279480367898941, + 0.010504497215151787, + 0.012862852774560452, + 0.00964207760989666, + 0.019466429948806763, + 0.014600402675569057, + 0.011330519802868366, + 0.009405211545526981, + 0.01084563136100769, + 0.014211161062121391, + 0.012243527919054031, + 0.015103216283023357, + 0.010885775089263916, + 0.008562194183468819, + 0.007976141758263111, + 0.010142884217202663, + 0.011718370020389557, + 0.012760404497385025, + 0.006841419730335474, + 0.0132988840341568, + 0.012113122269511223, + 0.012093852274119854, + 0.0074744028970599174, + 0.013162528164684772, + 0.003180496161803603, + 0.005133519880473614, + 0.004495637491345406, + 0.004960980266332626, + 0.007868463173508644, + 0.003969585057348013, + 0.0032593279611319304, + 0.006249851081520319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.576371967792511 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3160814642906189, + 0.08656245470046997, + 0.0798729807138443, + 0.24249820411205292, + 0.07234566658735275, + 0.05113795027136803, + 0.020828578621149063, + 0.05503761023283005, + 0.04852575436234474, + 0.027109390124678612 + ], + "credit_map": [ + 0.03311726078391075, + 0.05453791469335556, + 0.027589736506342888, + 0.036983173340559006, + 0.0236265417188406, + 0.04805079102516174, + 0.02436790242791176, + 0.06780809909105301, + 0.008426723070442677, + 0.005594703368842602, + 0.01284817699342966, + 0.0061839912086725235, + 0.016513826325535774, + 0.01113047543913126, + 0.0070093171671032906, + 0.018855232745409012, + 0.005720007233321667, + 0.0076718577183783054, + 0.012115897610783577, + 0.009876823052763939, + 0.006338697858154774, + 0.006907069124281406, + 0.017258089035749435, + 0.01398453302681446, + 0.039890870451927185, + 0.014332722872495651, + 0.0277173463255167, + 0.024446407333016396, + 0.04178912937641144, + 0.036844402551651, + 0.018346231430768967, + 0.039131078869104385, + 0.006154896225780249, + 0.00456625921651721, + 0.008993962779641151, + 0.005645242519676685, + 0.010777856223285198, + 0.021311072632670403, + 0.0047318702563643456, + 0.010164503008127213, + 0.007343834266066551, + 0.004924117121845484, + 0.008466821163892746, + 0.0070700012147426605, + 0.007457430008798838, + 0.007110130041837692, + 0.0037291941698640585, + 0.005036414600908756, + 0.0030255683232098818, + 0.0024915807880461216, + 0.002044315217062831, + 0.0030822677072137594, + 0.002961802529171109, + 0.0034077647142112255, + 0.0020184246823191643, + 0.0017968520987778902, + 0.006034443620592356, + 0.0053610117174685, + 0.0076407683081924915, + 0.009902385994791985, + 0.005239112768322229, + 0.009082908742129803, + 0.004582474008202553, + 0.007194500882178545, + 0.004558823071420193, + 0.003992595709860325, + 0.003598700975999236, + 0.0029756794683635235, + 0.004944693762809038, + 0.013283574022352695, + 0.0038198716938495636, + 0.011351808905601501, + 0.0023379598278552294, + 0.003110633697360754, + 0.003241375321522355, + 0.004644980188459158, + 0.004475525580346584, + 0.0025237419176846743, + 0.0028627256397157907, + 0.003912444692105055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.004096688237041235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589079707860947 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589079707860947, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589079707860947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22237998247146606, + 0.07756254822015762, + 0.05658818781375885, + 0.33009445667266846, + 0.06639046967029572, + 0.06089349836111069, + 0.06520970165729523, + 0.0741691142320633, + 0.025085102766752243, + 0.021626882255077362 + ], + "credit_map": [ + 0.02131330966949463, + 0.0314343236386776, + 0.02519725076854229, + 0.024403132498264313, + 0.01699906587600708, + 0.03797372430562973, + 0.013297188095748425, + 0.051762014627456665, + 0.0056556761264801025, + 0.007721133530139923, + 0.010769631713628769, + 0.008468813262879848, + 0.012475941330194473, + 0.011670985259115696, + 0.0048831976018846035, + 0.015917176380753517, + 0.005462035536766052, + 0.008085020817816257, + 0.0075255404226481915, + 0.013832985423505306, + 0.004464322701096535, + 0.0055386717431247234, + 0.005059598945081234, + 0.006620020605623722, + 0.048701778054237366, + 0.024323349818587303, + 0.047888875007629395, + 0.031625740230083466, + 0.049621108919382095, + 0.03690866008400917, + 0.023017391562461853, + 0.06800760328769684, + 0.007315259426832199, + 0.006916798651218414, + 0.009171922691166401, + 0.008471041917800903, + 0.009456385858356953, + 0.00974980928003788, + 0.007666302844882011, + 0.007642956450581551, + 0.007314891088753939, + 0.00869214627891779, + 0.006913115736097097, + 0.013231677003204823, + 0.004965740256011486, + 0.004821919836103916, + 0.006758079398423433, + 0.008195937611162663, + 0.006060234736651182, + 0.006382387597113848, + 0.012290549464523792, + 0.006155176088213921, + 0.011056548915803432, + 0.008538566529750824, + 0.005762973800301552, + 0.008963268250226974, + 0.012690998613834381, + 0.006283873692154884, + 0.010791145265102386, + 0.009153001010417938, + 0.00866604782640934, + 0.012081858702003956, + 0.005135925021022558, + 0.009366266429424286, + 0.0018742888933047652, + 0.0027350690215826035, + 0.004001126624643803, + 0.0018492053495720029, + 0.0022799125872552395, + 0.005521215498447418, + 0.0020736984442919493, + 0.004750589840114117, + 0.0017181937582790852, + 0.0026511300820857286, + 0.0037924330681562424, + 0.003702314803376794, + 0.0024273947346955538, + 0.002087520668283105, + 0.0022253436036407948, + 0.003022555261850357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.020589079707860947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot featuring a mournful aristocratic woman standing alone at the edge of a rugged, barren black volcanic cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an intricately patterned silk kimono embroidered with chrysanthemums and cranes in silver thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless sea of white foam crashing against dark rocks with spray rising into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and low-angled, creating strong dramatic contrast with deep shadows on the woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a 1940s black-and-white film style with obvious graininess and silver salt texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160688366740942 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160688366740942, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160688366740942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22960907220840454, + 0.05784669890999794, + 0.07394150644540787, + 0.27871885895729065, + 0.044501036405563354, + 0.023840835317969322, + 0.0839378833770752, + 0.14744967222213745, + 0.033423855900764465, + 0.026730617508292198 + ], + "credit_map": [ + 0.023286228999495506, + 0.0388290211558342, + 0.026946287602186203, + 0.02694757655262947, + 0.018340904265642166, + 0.03819034621119499, + 0.016364865005016327, + 0.04070382937788963, + 0.0047044395469129086, + 0.008094443008303642, + 0.0070021734572947025, + 0.00716519495472312, + 0.006481871008872986, + 0.008590523153543472, + 0.004215019755065441, + 0.011593030765652657, + 0.005961420945823193, + 0.006945693865418434, + 0.009987419471144676, + 0.00847889669239521, + 0.006812466774135828, + 0.009035533294081688, + 0.014454666525125504, + 0.012265413999557495, + 0.035216525197029114, + 0.025883864611387253, + 0.0369940809905529, + 0.027247970923781395, + 0.047053590416908264, + 0.041368886828422546, + 0.017013220116496086, + 0.04794072359800339, + 0.0043220315128564835, + 0.0029027971904724836, + 0.006754047237336636, + 0.005235082004219294, + 0.007125599309802055, + 0.008594432845711708, + 0.004651092458516359, + 0.004915950819849968, + 0.0027498495765030384, + 0.0027845054864883423, + 0.002882193773984909, + 0.0032245570328086615, + 0.0031723196152597666, + 0.0036494294181466103, + 0.0028128877747803926, + 0.0025650914758443832, + 0.007586623542010784, + 0.007199650164693594, + 0.017817728221416473, + 0.007321442477405071, + 0.013081919401884079, + 0.009429525583982468, + 0.007726642303168774, + 0.013774351216852665, + 0.023528045043349266, + 0.01371418870985508, + 0.018863780423998833, + 0.021311381831765175, + 0.015647364780306816, + 0.023013796657323837, + 0.010574485175311565, + 0.020796623080968857, + 0.0022731295321136713, + 0.003078373149037361, + 0.006282226648181677, + 0.002856902312487364, + 0.002291184850037098, + 0.006942362058907747, + 0.0025814450345933437, + 0.007118230685591698, + 0.0021856892853975296, + 0.0033367080613970757, + 0.0029516955837607384, + 0.004684169311076403, + 0.0038560167886316776, + 0.0024095610715448856, + 0.0032004278618842363, + 0.004106347449123859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0160688366740942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_4/index_tensor([3970, 3970])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_4/index_tensor([3970, 3970])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6089dde40bf572adc0feee443fa83c800161dac0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_4/index_tensor([3970, 3970])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, close-up shot, rendered with the signature high dynamic range and sharp image quality of Sony cameras, begins with a massive basketball player. Clad in a tight, sleeveless black athletic tank top, he's bathed in dramatic, high-contrast studio lighting. Sweat shimmers on his tanned skin, highlighting every muscle bulging with energy. With his focus completely focused, his eyes fixed forward, he swung a heavy, cloth-wrapped gong mallet with unstoppable force, striking a massive, ancient bronze gong suspended from a dark wooden frame. At the moment of impact, the camera's focus remains fixed on his forearm, which tenses and bulges with veins from the immense force. Then, in a swift and smooth rack focus, the focus shifts from his arm to the struck gong surface. Under strong sidelight, the gong's surface vibrates violently, creating visible concentric ripples that spread like water waves. Its highly polished surface reflects a dazzling specular highlight, causing the camera to produce a brief and dazzling halo, and the entire screen to instantly flash in a dazzling white light, creating an ultimate sense of impact in which power and sound are visualized, and a grand atmosphere like a commercial blockbuster.", + "seed": 27340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_4/index_3970_seed_27340_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03544616997241974, + "vlm_logprob": -0.03544616997241974, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.03544616997241974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015390795655548573 + ], + "opposite_logprob_tokens": [ + -4.265390872955322 + ], + "logprob_reward": -0.015390795655548573, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015390795655548573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14094728231430054, + 0.08983152359724045, + 0.051325004547834396, + 0.03265715762972832, + 0.09873788803815842, + 0.04352875053882599, + 0.025168219581246376, + 0.04261027276515961, + 0.03018822893500328, + 0.4450056552886963 + ], + "credit_map": [ + 0.021500252187252045, + 0.015158718451857567, + 0.02624737285077572, + 0.01477657351642847, + 0.013992054387927055, + 0.011847804300487041, + 0.011600633151829243, + 0.025823866948485374, + 0.014565353281795979, + 0.006003108341246843, + 0.010395866818726063, + 0.007380183786153793, + 0.011460541747510433, + 0.0130113884806633, + 0.016919080168008804, + 0.010096002370119095, + 0.007137112319469452, + 0.003347529796883464, + 0.007987117394804955, + 0.005828143563121557, + 0.006063066888600588, + 0.003164525842294097, + 0.013716678135097027, + 0.004080831538885832, + 0.004427994601428509, + 0.0038444988895207644, + 0.004563621245324612, + 0.004167195409536362, + 0.0039717587642371655, + 0.004974001087248325, + 0.0029569726902991533, + 0.0037511144764721394, + 0.004838153719902039, + 0.007517282385379076, + 0.007246668916195631, + 0.009665209800004959, + 0.007666659075766802, + 0.014342318288981915, + 0.033333733677864075, + 0.014127857983112335, + 0.007063626777380705, + 0.0050909980200231075, + 0.005514130927622318, + 0.007673841435462236, + 0.006122788414359093, + 0.003649898571893573, + 0.005086530465632677, + 0.003326932666823268, + 0.0027987284120172262, + 0.0032637298572808504, + 0.002874246332794428, + 0.002609015442430973, + 0.002529977820813656, + 0.005060963798314333, + 0.0034692957997322083, + 0.002562261186540127, + 0.003899534232914448, + 0.0037010766100138426, + 0.0052147675305604935, + 0.0036294814199209213, + 0.005685954820364714, + 0.009383488446474075, + 0.005654508247971535, + 0.00544146029278636, + 0.0026572998613119125, + 0.002545960247516632, + 0.0035499415826052427, + 0.002237146720290184, + 0.007536702323704958, + 0.0048676892183721066, + 0.0031763412989676, + 0.0036171493120491505, + 0.027079852297902107, + 0.011705192737281322, + 0.016795670613646507, + 0.015129382722079754, + 0.026801569387316704, + 0.03771534934639931, + 0.1874050348997116, + 0.12237360328435898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.015390795655548573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043138645589351654 + ], + "opposite_logprob_tokens": [ + -3.1681387424468994 + ], + "logprob_reward": -0.043138645589351654, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043138645589351654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17435091733932495, + 0.030777698382735252, + 0.058086834847927094, + 0.04094859957695007, + 0.25223225355148315, + 0.034608710557222366, + 0.024737564846873283, + 0.045716386288404465, + 0.03333759307861328, + 0.3052034080028534 + ], + "credit_map": [ + 0.024783434346318245, + 0.024159058928489685, + 0.031950678676366806, + 0.02312764711678028, + 0.017702292650938034, + 0.01492294855415821, + 0.014486931264400482, + 0.023217931389808655, + 0.00522676482796669, + 0.0022423036862164736, + 0.004267923999577761, + 0.002456539310514927, + 0.0035668956115841866, + 0.002689412795007229, + 0.006643951870501041, + 0.0036839088425040245, + 0.007856274954974651, + 0.002871550153940916, + 0.008856458589434624, + 0.008850659243762493, + 0.005513278767466545, + 0.0038128839805722237, + 0.015764623880386353, + 0.004561109002679586, + 0.004726104438304901, + 0.005264153238385916, + 0.004568862728774548, + 0.006600994151085615, + 0.005530982278287411, + 0.006024950183928013, + 0.004052055533975363, + 0.004180497489869595, + 0.007023494224995375, + 0.040276963263750076, + 0.01241385843604803, + 0.021981196478009224, + 0.011603054590523243, + 0.03256790339946747, + 0.10542074590921402, + 0.02094506099820137, + 0.0060602822341024876, + 0.004434135276824236, + 0.004217473324388266, + 0.005311567801982164, + 0.004475836642086506, + 0.0030879892874509096, + 0.003998458385467529, + 0.0030229694675654173, + 0.0034458583686500788, + 0.002606774913147092, + 0.0034351209178566933, + 0.002044636057689786, + 0.002362447325140238, + 0.005072613712400198, + 0.0035178728867322206, + 0.0022522418294101954, + 0.004823159892112017, + 0.006064379122108221, + 0.005440655164420605, + 0.006618892773985863, + 0.005081379786133766, + 0.007561523001641035, + 0.00487983413040638, + 0.005246563348919153, + 0.0023040040396153927, + 0.003253816394135356, + 0.0049292524345219135, + 0.0023927511647343636, + 0.005678082350641489, + 0.0065589104779064655, + 0.0037671446334570646, + 0.004453637637197971, + 0.019534219056367874, + 0.012444850988686085, + 0.023328406736254692, + 0.015499074943363667, + 0.022900808602571487, + 0.027206657454371452, + 0.11006638407707214, + 0.07422304153442383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.043138645589351654 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20468010008335114, + 0.05778384581208229, + 0.06459453701972961, + 0.13725146651268005, + 0.21256199479103088, + 0.05328796058893204, + 0.06868620961904526, + 0.0443863682448864, + 0.022494429722428322, + 0.13427309691905975 + ], + "credit_map": [ + 0.03129654750227928, + 0.021916862577199936, + 0.023207617923617363, + 0.014493575319647789, + 0.025639699772000313, + 0.019730156287550926, + 0.027899224311113358, + 0.040496405214071274, + 0.007655370514839888, + 0.005871412809938192, + 0.008516207337379456, + 0.0067240153439342976, + 0.006274723447859287, + 0.006590021308511496, + 0.0094346534460783, + 0.006717441603541374, + 0.010395522229373455, + 0.005306071601808071, + 0.009565072134137154, + 0.008409502916038036, + 0.006968244444578886, + 0.005575612653046846, + 0.011625252664089203, + 0.006749255117028952, + 0.014671267941594124, + 0.018413959071040154, + 0.015157213434576988, + 0.02359243854880333, + 0.020437361672520638, + 0.018721291795372963, + 0.011575650423765182, + 0.014682289212942123, + 0.006586518604308367, + 0.05047697573900223, + 0.006389778573065996, + 0.016009273007512093, + 0.009943999350070953, + 0.02364123985171318, + 0.08217664808034897, + 0.017337555065751076, + 0.009319060482084751, + 0.00544316042214632, + 0.006781273055821657, + 0.004770263098180294, + 0.006330073811113834, + 0.005591928027570248, + 0.00779855577275157, + 0.007253649178892374, + 0.009325370192527771, + 0.00855040829628706, + 0.01243799738585949, + 0.0057911048643291, + 0.0056685893796384335, + 0.012923916801810265, + 0.007238578982651234, + 0.0067502460442483425, + 0.0038620256818830967, + 0.005115887150168419, + 0.004717691335827112, + 0.006232934072613716, + 0.005221150815486908, + 0.008388306014239788, + 0.005566448904573917, + 0.0052819205448031425, + 0.0018070221412926912, + 0.0017459032824262977, + 0.0037640410009771585, + 0.002010134281590581, + 0.004383932799100876, + 0.003555953036993742, + 0.0026461484376341105, + 0.002581294160336256, + 0.007116931490600109, + 0.004074264783412218, + 0.006328205578029156, + 0.004391441121697426, + 0.007189763244241476, + 0.013406496495008469, + 0.05535938963294029, + 0.03640660643577576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.008639576844871044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778574109077454 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778574109077454, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778574109077454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13446897268295288, + 0.04486648738384247, + 0.0368681326508522, + 0.04897778853774071, + 0.15473751723766327, + 0.05742361396551132, + 0.10964321345090866, + 0.2552104890346527, + 0.0768335834145546, + 0.08097022771835327 + ], + "credit_map": [ + 0.01856267638504505, + 0.01576852612197399, + 0.019211113452911377, + 0.01244316715747118, + 0.017478976398706436, + 0.012055260129272938, + 0.01960807666182518, + 0.019341174513101578, + 0.009170831181108952, + 0.0033940717112272978, + 0.007022843696177006, + 0.003780787345021963, + 0.004298703745007515, + 0.004640358500182629, + 0.007266960572451353, + 0.005291932262480259, + 0.006891441065818071, + 0.002331603318452835, + 0.006693245843052864, + 0.005338484887033701, + 0.0030198434833437204, + 0.0030109852086752653, + 0.006987894419580698, + 0.002594636520370841, + 0.0044962456449866295, + 0.005232525523751974, + 0.0064448462799191475, + 0.009161781519651413, + 0.005419545341283083, + 0.00673160795122385, + 0.005218855105340481, + 0.006272383499890566, + 0.003989982884377241, + 0.023403562605381012, + 0.006613687612116337, + 0.013389255851507187, + 0.006966958288103342, + 0.0199273694306612, + 0.0666198655962944, + 0.013826831243932247, + 0.010155352763831615, + 0.00915944017469883, + 0.008545721881091595, + 0.006824301555752754, + 0.00803239457309246, + 0.004240299109369516, + 0.006467765662819147, + 0.003998336382210255, + 0.015582621097564697, + 0.011484764516353607, + 0.012417317368090153, + 0.01305230613797903, + 0.009829205460846424, + 0.02174714021384716, + 0.012090594507753849, + 0.01343927625566721, + 0.023286931216716766, + 0.026614118367433548, + 0.03662684187293053, + 0.03526393696665764, + 0.03154785931110382, + 0.04140974208712578, + 0.03224183991551399, + 0.02821921743452549, + 0.00439649960026145, + 0.005764266010373831, + 0.009810230694711208, + 0.0071625495329499245, + 0.010912149213254452, + 0.01657799631357193, + 0.01085815578699112, + 0.011351732537150383, + 0.00467635365203023, + 0.00465203495696187, + 0.008889448828995228, + 0.0051881857216358185, + 0.0051009259186685085, + 0.012413201853632927, + 0.02445986121892929, + 0.015590210445225239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.009778574109077454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028325766324997 + ], + "opposite_logprob_tokens": [ + -2.350283145904541 + ], + "logprob_reward": -0.10028325766324997, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028325766324997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1122652217745781, + 0.042225759476423264, + 0.025880936533212662, + 0.02356819435954094, + 0.07308188080787659, + 0.02969837374985218, + 0.062387049198150635, + 0.03512890264391899, + 0.1015881672501564, + 0.49417558312416077 + ], + "credit_map": [ + 0.017031028866767883, + 0.006876384373754263, + 0.017218677327036858, + 0.008629932068288326, + 0.015752049162983894, + 0.012673215009272099, + 0.014137988910079002, + 0.019945945590734482, + 0.008046823553740978, + 0.003395842621102929, + 0.0060942391864955425, + 0.004391624592244625, + 0.003605013946071267, + 0.00470539228990674, + 0.006780580151826143, + 0.00520623754709959, + 0.0037083211354911327, + 0.0016566066769883037, + 0.0039003074634820223, + 0.003612313186749816, + 0.0024645430967211723, + 0.0013506689574569464, + 0.007139751687645912, + 0.002048423746600747, + 0.0030026703607290983, + 0.00311711011454463, + 0.002945178421214223, + 0.004245111718773842, + 0.003044877899810672, + 0.002722984878346324, + 0.002071040915325284, + 0.002419217023998499, + 0.00257256836630404, + 0.009564604610204697, + 0.0041491324082016945, + 0.006930891890078783, + 0.004792405758053064, + 0.009007206186652184, + 0.028462179005146027, + 0.007602884899824858, + 0.003667803481221199, + 0.004378021229058504, + 0.004065603483468294, + 0.0038371849805116653, + 0.004460406955331564, + 0.0027616366278380156, + 0.0036459125112742186, + 0.002881802385672927, + 0.00805892888456583, + 0.010051836259663105, + 0.009775382466614246, + 0.006310752127319574, + 0.0059365094639360905, + 0.008950270712375641, + 0.007171871606260538, + 0.006131493486464024, + 0.003750505857169628, + 0.004466697108000517, + 0.004868111573159695, + 0.004829722922295332, + 0.004394949413836002, + 0.005188449751585722, + 0.004633495118469, + 0.0029969695024192333, + 0.006228026933968067, + 0.011367705650627613, + 0.010702032595872879, + 0.009640122763812542, + 0.019474446773529053, + 0.016046691685914993, + 0.016354043036699295, + 0.011775088496506214, + 0.028293829411268234, + 0.01332995668053627, + 0.017948811873793602, + 0.015079495497047901, + 0.02802382782101631, + 0.04398156702518463, + 0.2099464237689972, + 0.13757166266441345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.10028325766324997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, close-up shot, rendered with the signature high dynamic range and sharp image quality of Sony cameras, begins with a massive basketball player. Clad in a tight, sleeveless black athletic tank top, he's bathed in dramatic, high-contrast studio lighting. Sweat shimmers on his tanned skin, highlighting every muscle bulging with energy. With his focus completely focused, his eyes fixed forward, he swung a heavy, cloth-wrapped gong mallet with unstoppable force, striking a massive, ancient bronze gong suspended from a dark wooden frame. At the moment of impact, the camera's focus remains fixed on his forearm, which tenses and bulges with veins from the immense force. Then, in a swift and smooth rack focus, the focus shifts from his arm to the struck gong surface. Under strong sidelight, the gong's surface vibrates violently, creating visible concentric ripples that spread like water waves. Its highly polished surface reflects a dazzling specular highlight, causing the camera to produce a brief and dazzling halo, and the entire screen to instantly flash in a dazzling white light, creating an ultimate sense of impact in which power and sound are visualized, and a grand atmosphere like a commercial blockbuster.", + "seed": 27341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_4/index_3970_seed_27341_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09222954679280519, + "vlm_logprob": -0.09222954679280519, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.09222954679280519, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3166554868221283 + ], + "opposite_logprob_tokens": [ + -1.3166555166244507 + ], + "logprob_reward": -0.3166554868221283, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3166554868221283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5066345930099487, + 0.0590292252600193, + 0.09796348214149475, + 0.060474950820207596, + 0.08447613567113876, + 0.038856662809848785, + 0.02849487215280533, + 0.02800893969833851, + 0.04027561843395233, + 0.055785536766052246 + ], + "credit_map": [ + 0.04296303912997246, + 0.059029992669820786, + 0.05308074504137039, + 0.06926633417606354, + 0.05856543779373169, + 0.06641692668199539, + 0.0625661313533783, + 0.09474599361419678, + 0.007374212145805359, + 0.005405091214925051, + 0.007120627909898758, + 0.007132223341614008, + 0.006382419262081385, + 0.00983708631247282, + 0.007411181461066008, + 0.008366384543478489, + 0.005018523428589106, + 0.008286590687930584, + 0.015245731920003891, + 0.010346428491175175, + 0.018907511606812477, + 0.013173220679163933, + 0.011887074448168278, + 0.015098413452506065, + 0.00413996446877718, + 0.007507821545004845, + 0.004026444163173437, + 0.005472165998071432, + 0.00863328855484724, + 0.011719126254320145, + 0.008461364544928074, + 0.010514779947698116, + 0.00722834188491106, + 0.010859919711947441, + 0.006754749454557896, + 0.01080580148845911, + 0.011228082701563835, + 0.01723328046500683, + 0.014435289427638054, + 0.005930679850280285, + 0.004559657070785761, + 0.006360587198287249, + 0.005897295195609331, + 0.006816064473241568, + 0.005365490913391113, + 0.003525801468640566, + 0.0037235578056424856, + 0.0026082098484039307, + 0.004567762371152639, + 0.002786959521472454, + 0.0042780665680766106, + 0.00546239921823144, + 0.003134337952360511, + 0.002280161250382662, + 0.0022837838623672724, + 0.0037014016415923834, + 0.004206334240734577, + 0.002188966376706958, + 0.003652171930298209, + 0.0037635231856256723, + 0.0031533504370599985, + 0.0034823124296963215, + 0.004416691605001688, + 0.003145590191707015, + 0.003326459787786007, + 0.006251926068216562, + 0.0028956818860024214, + 0.00867384672164917, + 0.004034154117107391, + 0.004465662874281406, + 0.007310085464268923, + 0.0033178022131323814, + 0.003797458950430155, + 0.01180524192750454, + 0.003042385447770357, + 0.004297442268580198, + 0.01473715528845787, + 0.005954938940703869, + 0.006527381017804146, + 0.005623533856123686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.3166554868221283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029843909665942192 + ], + "opposite_logprob_tokens": [ + -3.529843807220459 + ], + "logprob_reward": -0.029843909665942192, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029843909665942192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3763270378112793, + 0.13941161334514618, + 0.07270949333906174, + 0.05321287736296654, + 0.08270429819822311, + 0.06128079816699028, + 0.037255968898534775, + 0.029353713616728783, + 0.03172977268695831, + 0.11601436138153076 + ], + "credit_map": [ + 0.035139210522174835, + 0.04777706041932106, + 0.04153944179415703, + 0.050293441861867905, + 0.04584576189517975, + 0.05352247133851051, + 0.03168234974145889, + 0.0705273225903511, + 0.012852493673563004, + 0.017931967973709106, + 0.012542808428406715, + 0.01585107482969761, + 0.02385498769581318, + 0.02510012499988079, + 0.01243048720061779, + 0.01884767971932888, + 0.0034989845007658005, + 0.004953595343977213, + 0.012646079063415527, + 0.005459907930344343, + 0.008978861384093761, + 0.010397784411907196, + 0.006201549898833036, + 0.020572742447257042, + 0.0047258962877094746, + 0.007826612330973148, + 0.004399274010211229, + 0.0038951768074184656, + 0.008001578971743584, + 0.006978493649512529, + 0.009060988202691078, + 0.008324865251779556, + 0.00655398890376091, + 0.009031085297465324, + 0.006901639513671398, + 0.012354203499853611, + 0.011779853142797947, + 0.01431548036634922, + 0.013422315940260887, + 0.008345743641257286, + 0.006240124348551035, + 0.010289181023836136, + 0.008348732255399227, + 0.010853063315153122, + 0.009070728905498981, + 0.004745881538838148, + 0.007265529595315456, + 0.004467566031962633, + 0.006173841655254364, + 0.0034074103459715843, + 0.005868174135684967, + 0.006820240058004856, + 0.004377956502139568, + 0.0028881027828902006, + 0.0023058252409100533, + 0.005414422135800123, + 0.004509489517658949, + 0.0024075843393802643, + 0.004308673553168774, + 0.004084346350282431, + 0.0033135267440229654, + 0.003118787659332156, + 0.004679716192185879, + 0.0029315915890038013, + 0.004371290560811758, + 0.004540101159363985, + 0.0034515061415731907, + 0.004896003287285566, + 0.0031588308047503233, + 0.0034661509562283754, + 0.004200139548629522, + 0.003645755583420396, + 0.0085495226085186, + 0.019289478659629822, + 0.008464996702969074, + 0.01467097271233797, + 0.024823278188705444, + 0.014323725365102291, + 0.015012215822935104, + 0.0108801843598485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.029843909665942192 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632262073457241 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632262073457241, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632262073457241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23806427419185638, + 0.1467283070087433, + 0.07628971338272095, + 0.12411624193191528, + 0.18016548454761505, + 0.04634453356266022, + 0.046910662204027176, + 0.02666223607957363, + 0.05895242840051651, + 0.05576614663004875 + ], + "credit_map": [ + 0.0210401751101017, + 0.02857879363000393, + 0.02448982559144497, + 0.03189535066485405, + 0.027348190546035767, + 0.03163862228393555, + 0.026311997324228287, + 0.04676130786538124, + 0.016539862379431725, + 0.02045203372836113, + 0.014312909916043282, + 0.011884942650794983, + 0.02389306016266346, + 0.02361503802239895, + 0.014361010864377022, + 0.021669449284672737, + 0.003578247968107462, + 0.00643489183858037, + 0.011412748135626316, + 0.007523755077272654, + 0.01485505886375904, + 0.008421818725764751, + 0.009994707070291042, + 0.01406848430633545, + 0.009214836172759533, + 0.016378983855247498, + 0.008401354774832726, + 0.007412888575345278, + 0.020157944411039352, + 0.015970727428793907, + 0.02172630839049816, + 0.024853188544511795, + 0.018680786713957787, + 0.03191542252898216, + 0.017224043607711792, + 0.018917521461844444, + 0.027103964239358902, + 0.02556641772389412, + 0.02270740456879139, + 0.018049927428364754, + 0.004887130111455917, + 0.006604276597499847, + 0.007037642877548933, + 0.008502522483468056, + 0.004720795899629593, + 0.005225813016295433, + 0.004859142936766148, + 0.004507207311689854, + 0.010860479436814785, + 0.005550717003643513, + 0.008708193898200989, + 0.007119372952729464, + 0.004261780064553022, + 0.002632793039083481, + 0.0036612253170460463, + 0.004116097930818796, + 0.002151736756786704, + 0.002581496722996235, + 0.0026861296501010656, + 0.002917904406785965, + 0.002520255744457245, + 0.004065367858856916, + 0.006803566589951515, + 0.0029357783496379852, + 0.0050910166464746, + 0.011385675519704819, + 0.003877360373735428, + 0.008231204934418201, + 0.007023809943348169, + 0.008654125034809113, + 0.007066354621201754, + 0.007622879929840565, + 0.005166227463632822, + 0.007624455727636814, + 0.003931153565645218, + 0.007904154248535633, + 0.008227006532251835, + 0.00858709029853344, + 0.00810941681265831, + 0.0062166438437998295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.007632262073457241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733465939760208 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733465939760208, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733465939760208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22073140740394592, + 0.07242801785469055, + 0.0686006247997284, + 0.174712672829628, + 0.04550871625542641, + 0.042464785277843475, + 0.11939474940299988, + 0.07476722449064255, + 0.08372011035680771, + 0.09767168760299683 + ], + "credit_map": [ + 0.02071128413081169, + 0.02367534674704075, + 0.01811905764043331, + 0.03309804946184158, + 0.027691593393683434, + 0.03317590057849884, + 0.02161332033574581, + 0.042646851390600204, + 0.00734699796885252, + 0.00846510287374258, + 0.009181064553558826, + 0.008562308736145496, + 0.012592547573149204, + 0.009476604871451855, + 0.007307206280529499, + 0.00949618685990572, + 0.0030791801400482655, + 0.005789982620626688, + 0.009969092905521393, + 0.006472878623753786, + 0.008640812709927559, + 0.010350887663662434, + 0.011488109827041626, + 0.012809684500098228, + 0.017373764887452126, + 0.025675423443317413, + 0.012099352665245533, + 0.013270635157823563, + 0.03491407260298729, + 0.019561462104320526, + 0.02571316808462143, + 0.026104794815182686, + 0.004406095948070288, + 0.0055554877035319805, + 0.0037023562472313643, + 0.005918631795793772, + 0.006503107491880655, + 0.006501462776213884, + 0.006902660708874464, + 0.006018917541950941, + 0.006091742776334286, + 0.006500147748738527, + 0.005309718195348978, + 0.007181303109973669, + 0.004155944567173719, + 0.005160487722605467, + 0.0042138975113630295, + 0.003851543879136443, + 0.0348086953163147, + 0.01324875932186842, + 0.027123911306262016, + 0.014774594455957413, + 0.008386299945414066, + 0.00549737736582756, + 0.0075637781992554665, + 0.007991333492100239, + 0.0060689435340464115, + 0.004842222668230534, + 0.006596221588551998, + 0.009304637089371681, + 0.00913948193192482, + 0.011115689761936665, + 0.01638936810195446, + 0.011310659348964691, + 0.005825383123010397, + 0.010312221944332123, + 0.006088822148740292, + 0.014279183931648731, + 0.009964664466679096, + 0.011080075986683369, + 0.016887666657567024, + 0.009282095357775688, + 0.008140027523040771, + 0.014491445384919643, + 0.00846794992685318, + 0.01286150049418211, + 0.018644988536834717, + 0.012239281088113785, + 0.014088654890656471, + 0.008737845346331596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.006733465939760208 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028260946273804 + ], + "opposite_logprob_tokens": [ + -2.350282669067383 + ], + "logprob_reward": -0.10028260946273804, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028260946273804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30878522992134094, + 0.07452201843261719, + 0.04374401271343231, + 0.021864276379346848, + 0.02248333767056465, + 0.031281787902116776, + 0.039039187133312225, + 0.028859280049800873, + 0.04412256181240082, + 0.38529831171035767 + ], + "credit_map": [ + 0.02482043206691742, + 0.04121464863419533, + 0.03509773686528206, + 0.04331062361598015, + 0.036876797676086426, + 0.03614647313952446, + 0.03579813614487648, + 0.05552038922905922, + 0.008081763982772827, + 0.006913845427334309, + 0.013078439980745316, + 0.008936085738241673, + 0.007655583322048187, + 0.0122032780200243, + 0.00640126783400774, + 0.011251749470829964, + 0.0028048367239534855, + 0.003249452216550708, + 0.007117106579244137, + 0.0034159801434725523, + 0.006713178474456072, + 0.0049970983527600765, + 0.004476185422390699, + 0.01097017154097557, + 0.0014637891435995698, + 0.0035859134513884783, + 0.001176202087663114, + 0.0017401393270120025, + 0.00537818344309926, + 0.0014366823015734553, + 0.004552277736365795, + 0.00253108749166131, + 0.001779758371412754, + 0.0028783895540982485, + 0.0020837087649852037, + 0.0030531100928783417, + 0.003200788516551256, + 0.004008612595498562, + 0.003312470857053995, + 0.002166497055441141, + 0.0040772827342152596, + 0.00479842396453023, + 0.0060902745462954044, + 0.0050286296755075455, + 0.002833150327205658, + 0.0034153074957430363, + 0.003180037485435605, + 0.0018586814403533936, + 0.006611567456275225, + 0.0033274805173277855, + 0.006018438376486301, + 0.0055046845227479935, + 0.0040297438390553, + 0.004287234507501125, + 0.003180034924298525, + 0.0060800001956522465, + 0.0038039491046220064, + 0.0025794643443077803, + 0.005148108582943678, + 0.0028638949152082205, + 0.0028358250856399536, + 0.00363603001460433, + 0.005488540977239609, + 0.002503465861082077, + 0.004186124540865421, + 0.005370364058762789, + 0.004937327466905117, + 0.005768897943198681, + 0.005288120824843645, + 0.005362717900425196, + 0.0076683019287884235, + 0.00554070295765996, + 0.02812083065509796, + 0.06683377176523209, + 0.021269913762807846, + 0.031419381499290466, + 0.09326233714818954, + 0.04778361693024635, + 0.057733211666345596, + 0.038875237107276917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.10028260946273804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, close-up shot, rendered with the signature high dynamic range and sharp image quality of Sony cameras, begins with a massive basketball player. Clad in a tight, sleeveless black athletic tank top, he's bathed in dramatic, high-contrast studio lighting. Sweat shimmers on his tanned skin, highlighting every muscle bulging with energy. With his focus completely focused, his eyes fixed forward, he swung a heavy, cloth-wrapped gong mallet with unstoppable force, striking a massive, ancient bronze gong suspended from a dark wooden frame. At the moment of impact, the camera's focus remains fixed on his forearm, which tenses and bulges with veins from the immense force. Then, in a swift and smooth rack focus, the focus shifts from his arm to the struck gong surface. Under strong sidelight, the gong's surface vibrates violently, creating visible concentric ripples that spread like water waves. Its highly polished surface reflects a dazzling specular highlight, causing the camera to produce a brief and dazzling halo, and the entire screen to instantly flash in a dazzling white light, creating an ultimate sense of impact in which power and sound are visualized, and a grand atmosphere like a commercial blockbuster.", + "seed": 27342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_4/index_3970_seed_27342_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07063137385994196, + "vlm_logprob": -0.07063137385994196, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.07063137385994196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25594276189804077 + ], + "opposite_logprob_tokens": [ + -1.5059428215026855 + ], + "logprob_reward": -0.25594276189804077, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25594276189804077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14450863003730774, + 0.5228817462921143, + 0.036621347069740295, + 0.03970027342438698, + 0.040078625082969666, + 0.06177043914794922, + 0.043741948902606964, + 0.05348611995577812, + 0.03523265942931175, + 0.021978240460157394 + ], + "credit_map": [ + 0.02152123861014843, + 0.010212705470621586, + 0.017015600576996803, + 0.010376201942563057, + 0.033386699855327606, + 0.02383454702794552, + 0.00974727887660265, + 0.018414366990327835, + 0.11888876557350159, + 0.014093723148107529, + 0.19571936130523682, + 0.019004132598638535, + 0.016479680314660072, + 0.042178478091955185, + 0.01592782884836197, + 0.10058978945016861, + 0.004755080211907625, + 0.004462949000298977, + 0.006826481781899929, + 0.00339944357983768, + 0.004456001799553633, + 0.00607278011739254, + 0.0037054952699691057, + 0.0029431176371872425, + 0.005162583664059639, + 0.0037366088945418596, + 0.004207402467727661, + 0.004360589198768139, + 0.007591115310788155, + 0.004113757982850075, + 0.00452071987092495, + 0.00600749347358942, + 0.004634997341781855, + 0.005673804320394993, + 0.006005177740007639, + 0.007346943020820618, + 0.0027233196888118982, + 0.004315987695008516, + 0.003555108094587922, + 0.005823285318911076, + 0.005488424561917782, + 0.010681858286261559, + 0.010294412262737751, + 0.007222323678433895, + 0.00429278751835227, + 0.00809297151863575, + 0.0071367789059877396, + 0.008560874499380589, + 0.0022716419771313667, + 0.0029089897871017456, + 0.005040994379669428, + 0.004860273562371731, + 0.005231600254774094, + 0.007911430671811104, + 0.008686535991728306, + 0.006830484140664339, + 0.005923054181039333, + 0.00913180224597454, + 0.0041913739405572414, + 0.007010036613792181, + 0.007641671225428581, + 0.009489910677075386, + 0.004966573789715767, + 0.005131700541824102, + 0.0034901308827102184, + 0.003797761397436261, + 0.0038122585974633694, + 0.004986168816685677, + 0.005105518735945225, + 0.005047092214226723, + 0.003411342157050967, + 0.005582381971180439, + 0.0031955307349562645, + 0.0024218298494815826, + 0.0026707083452492952, + 0.0024570708628743887, + 0.0030547932256013155, + 0.0030239084735512733, + 0.0022435453720390797, + 0.002910852897912264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.25594276189804077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03818359225988388 + ], + "opposite_logprob_tokens": [ + -3.2881836891174316 + ], + "logprob_reward": -0.03818359225988388, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03818359225988388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461695700883865, + 0.4659343957901001, + 0.02956533432006836, + 0.040896836668252945, + 0.04823803901672363, + 0.03431568294763565, + 0.019651085138320923, + 0.16333076357841492, + 0.0717165544629097, + 0.031734369695186615 + ], + "credit_map": [ + 0.013930805027484894, + 0.00937994197010994, + 0.013498476706445217, + 0.009652258828282356, + 0.016291070729494095, + 0.011237646453082561, + 0.010719538666307926, + 0.00990721769630909, + 0.10492278635501862, + 0.012454772368073463, + 0.172378271818161, + 0.017161620780825615, + 0.016121286898851395, + 0.03865228220820427, + 0.013825011439621449, + 0.09041836112737656, + 0.0035473203752189875, + 0.003119846573099494, + 0.004030602518469095, + 0.002841582987457514, + 0.004277621395885944, + 0.004951997194439173, + 0.003440228523686528, + 0.003356132423505187, + 0.005351067986339331, + 0.003155970713123679, + 0.005912933498620987, + 0.005423807073384523, + 0.005443504545837641, + 0.003929452504962683, + 0.005135250277817249, + 0.006544848438352346, + 0.003873896785080433, + 0.008486895821988583, + 0.011547977104783058, + 0.003974442835897207, + 0.005656569264829159, + 0.00404683593660593, + 0.0036557235289365053, + 0.006995697505772114, + 0.0051288544200360775, + 0.0042280713096261024, + 0.005776423495262861, + 0.002988925203680992, + 0.0023049002047628164, + 0.004994213581085205, + 0.0036653748247772455, + 0.005228918977081776, + 0.0024791881442070007, + 0.0022444429341703653, + 0.0019663230050355196, + 0.0015639187768101692, + 0.0022136878687888384, + 0.0039652674458920956, + 0.0026925373822450638, + 0.0025257186498492956, + 0.019110027700662613, + 0.015504163689911366, + 0.013890201225876808, + 0.021752173081040382, + 0.024335946887731552, + 0.020660068839788437, + 0.025990158319473267, + 0.022088022902607918, + 0.006775634828954935, + 0.00665779085829854, + 0.007628208491951227, + 0.011735609732568264, + 0.010103869251906872, + 0.012162928469479084, + 0.006457299925386906, + 0.010195218026638031, + 0.003029716433957219, + 0.004448769148439169, + 0.002888167044147849, + 0.00436423160135746, + 0.004884847905486822, + 0.004139877390116453, + 0.003525744192302227, + 0.004453015513718128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.03818359225988388 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203879982233047 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014203879982233047, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203879982233047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1642434298992157, + 0.34048762917518616, + 0.02089790813624859, + 0.0582122877240181, + 0.15043671429157257, + 0.13246622681617737, + 0.03050203248858452, + 0.04151998832821846, + 0.020044740289449692, + 0.041188981384038925 + ], + "credit_map": [ + 0.028057275339961052, + 0.010921522043645382, + 0.026006290689110756, + 0.007712632417678833, + 0.031779006123542786, + 0.025619223713874817, + 0.014038212597370148, + 0.0201092679053545, + 0.07605309784412384, + 0.01007148902863264, + 0.12490388005971909, + 0.01231968030333519, + 0.015037975274026394, + 0.029513319954276085, + 0.009641412645578384, + 0.06294681131839752, + 0.004032718017697334, + 0.002493220381438732, + 0.0023211382795125246, + 0.002167844446375966, + 0.0026242283638566732, + 0.0028581039514392614, + 0.0020648976787924767, + 0.002335759811103344, + 0.009283914230763912, + 0.005751959979534149, + 0.007325679063796997, + 0.007329126819968224, + 0.011264745146036148, + 0.00518640223890543, + 0.005274683237075806, + 0.006795785389840603, + 0.01258836593478918, + 0.022210506722331047, + 0.022386454045772552, + 0.020482879132032394, + 0.019588716328144073, + 0.014712930656969547, + 0.019501997157931328, + 0.018964873626828194, + 0.025230295956134796, + 0.016276855021715164, + 0.02455354854464531, + 0.011720215901732445, + 0.009848115965723991, + 0.014333812519907951, + 0.009686288423836231, + 0.0208171047270298, + 0.0024711648002266884, + 0.003029888030141592, + 0.00444584246724844, + 0.0030602540355175734, + 0.002993096597492695, + 0.005937436129897833, + 0.004344120621681213, + 0.0042202346958220005, + 0.003342937445268035, + 0.006980529986321926, + 0.0037317490205168724, + 0.004275010898709297, + 0.005630946718156338, + 0.007839526049792767, + 0.005097387824207544, + 0.004621905740350485, + 0.0019566253758966923, + 0.0018643946386873722, + 0.002090022200718522, + 0.0033122405875474215, + 0.0025472482666373253, + 0.0019597092177718878, + 0.002654298674315214, + 0.003660201793536544, + 0.004568165633827448, + 0.006187185179442167, + 0.004872934892773628, + 0.006433885544538498, + 0.005200537387281656, + 0.004387690685689449, + 0.004438658244907856, + 0.005099928006529808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.014203879982233047 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737136282026768 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737136282026768, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737136282026768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08701495826244354, + 0.4782707691192627, + 0.021784020587801933, + 0.0641455128788948, + 0.07035268843173981, + 0.028852736577391624, + 0.093852199614048, + 0.07431555539369583, + 0.04024677723646164, + 0.04116472229361534 + ], + "credit_map": [ + 0.01397862657904625, + 0.006555963773280382, + 0.011883893050253391, + 0.004971511662006378, + 0.016739560291171074, + 0.012639961205422878, + 0.00928607676178217, + 0.010959378443658352, + 0.10800422728061676, + 0.012956146150827408, + 0.17885306477546692, + 0.01655564270913601, + 0.01359797827899456, + 0.03892596811056137, + 0.01431494764983654, + 0.09506281465291977, + 0.0035405289381742477, + 0.00270888302475214, + 0.0026676866691559553, + 0.0029974409844726324, + 0.0025196655187755823, + 0.003088522469624877, + 0.001837976393289864, + 0.002423318335786462, + 0.011318124830722809, + 0.008330641314387321, + 0.011968852952122688, + 0.0101710744202137, + 0.006882976274937391, + 0.004252408631145954, + 0.00551931606605649, + 0.005702129565179348, + 0.006156453862786293, + 0.013804872520267963, + 0.01397427637130022, + 0.009129907935857773, + 0.006841597147285938, + 0.006189931184053421, + 0.005721469409763813, + 0.008534186519682407, + 0.0029241011943668127, + 0.004585961811244488, + 0.004953861702233553, + 0.0039271945133805275, + 0.0023049365263432264, + 0.0032766906078904867, + 0.003241844242438674, + 0.0036381487734615803, + 0.007877062074840069, + 0.008156408555805683, + 0.016873516142368317, + 0.012276054359972477, + 0.007864460349082947, + 0.014433387666940689, + 0.01476901676505804, + 0.011602304875850677, + 0.008844088762998581, + 0.010548909194767475, + 0.0063582854345440865, + 0.016416847705841064, + 0.0076073105446994305, + 0.010167030617594719, + 0.008196665905416012, + 0.006176421884447336, + 0.003994776867330074, + 0.003960255533456802, + 0.004448472522199154, + 0.005987060256302357, + 0.005531344097107649, + 0.005207602865993977, + 0.00495579419657588, + 0.006161476951092482, + 0.004440975375473499, + 0.005502365063875914, + 0.00434102164581418, + 0.00719113927334547, + 0.005778178572654724, + 0.004460987634956837, + 0.0045901271514594555, + 0.0048599326983094215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.006737136282026768 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808949887752533 + ], + "opposite_logprob_tokens": [ + -3.2880895137786865 + ], + "logprob_reward": -0.03808949887752533, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808949887752533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06917273253202438, + 0.414803147315979, + 0.022842558100819588, + 0.046628352254629135, + 0.04717721417546272, + 0.03271550312638283, + 0.05194680392742157, + 0.10456636548042297, + 0.10804609954357147, + 0.10210122168064117 + ], + "credit_map": [ + 0.010684682987630367, + 0.007726843468844891, + 0.009492946788668633, + 0.005697679705917835, + 0.012615188024938107, + 0.009158994071185589, + 0.007808160036802292, + 0.005988243967294693, + 0.0911664143204689, + 0.01115332916378975, + 0.15372580289840698, + 0.015320125967264175, + 0.011593952775001526, + 0.03497954085469246, + 0.012838226743042469, + 0.08402577042579651, + 0.0034915395081043243, + 0.003734016790986061, + 0.0023574684746563435, + 0.002903120359405875, + 0.0024762386456131935, + 0.003997298888862133, + 0.0017066781874746084, + 0.002176197012886405, + 0.00948423519730568, + 0.003712346078827977, + 0.00915913749486208, + 0.006417072843760252, + 0.005180597770959139, + 0.003801510203629732, + 0.0034173985477536917, + 0.005456054117530584, + 0.00586451031267643, + 0.0050771539099514484, + 0.00811954215168953, + 0.0068186125718057156, + 0.004943840205669403, + 0.0038675577379763126, + 0.0055914269760251045, + 0.006894565187394619, + 0.005314663518220186, + 0.005535290110856295, + 0.0044937292113900185, + 0.003360189963132143, + 0.002931237919256091, + 0.003970677964389324, + 0.0027134690899401903, + 0.004396244417876005, + 0.004964631516486406, + 0.005144056864082813, + 0.007375772576779127, + 0.006942100822925568, + 0.00501755578443408, + 0.008182298392057419, + 0.008008859120309353, + 0.006311530247330666, + 0.008355644531548023, + 0.022931743413209915, + 0.00825873576104641, + 0.012106570415198803, + 0.010425504297018051, + 0.01895882561802864, + 0.012581717222929, + 0.010947623290121555, + 0.010780960321426392, + 0.010464555583894253, + 0.010182238183915615, + 0.017432380467653275, + 0.011206733994185925, + 0.017131343483924866, + 0.014654655940830708, + 0.016193227842450142, + 0.010275719687342644, + 0.013312547467648983, + 0.01249587070196867, + 0.014314597472548485, + 0.01207813061773777, + 0.012831686064600945, + 0.01158938743174076, + 0.015203275717794895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.03808949887752533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, close-up shot, rendered with the signature high dynamic range and sharp image quality of Sony cameras, begins with a massive basketball player. Clad in a tight, sleeveless black athletic tank top, he's bathed in dramatic, high-contrast studio lighting. Sweat shimmers on his tanned skin, highlighting every muscle bulging with energy. With his focus completely focused, his eyes fixed forward, he swung a heavy, cloth-wrapped gong mallet with unstoppable force, striking a massive, ancient bronze gong suspended from a dark wooden frame. At the moment of impact, the camera's focus remains fixed on his forearm, which tenses and bulges with veins from the immense force. Then, in a swift and smooth rack focus, the focus shifts from his arm to the struck gong surface. Under strong sidelight, the gong's surface vibrates violently, creating visible concentric ripples that spread like water waves. Its highly polished surface reflects a dazzling specular highlight, causing the camera to produce a brief and dazzling halo, and the entire screen to instantly flash in a dazzling white light, creating an ultimate sense of impact in which power and sound are visualized, and a grand atmosphere like a commercial blockbuster.", + "seed": 27343, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_4/index_3970_seed_27343_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04900564551353455, + "vlm_logprob": -0.04900564551353455, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.04900564551353455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06521180272102356 + ], + "opposite_logprob_tokens": [ + -2.815211772918701 + ], + "logprob_reward": -0.06521180272102356, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06521180272102356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23153173923492432, + 0.08906756341457367, + 0.03181493282318115, + 0.18717926740646362, + 0.09758187085390091, + 0.04167550429701805, + 0.038361672312021255, + 0.06672398000955582, + 0.09649906307458878, + 0.11956442892551422 + ], + "credit_map": [ + 0.043336234986782074, + 0.024713046848773956, + 0.0362255796790123, + 0.018884610384702682, + 0.03467367962002754, + 0.02812300994992256, + 0.022416194900870323, + 0.023159386590123177, + 0.01165897585451603, + 0.006142857484519482, + 0.01168413832783699, + 0.007547595072537661, + 0.011490392498672009, + 0.015035542659461498, + 0.011943941004574299, + 0.013564113527536392, + 0.0029907242860645056, + 0.003110613441094756, + 0.00515323830768466, + 0.003043044125661254, + 0.0032159541733562946, + 0.006107814144343138, + 0.003270984860137105, + 0.004922559950500727, + 0.010957172140479088, + 0.022103028371930122, + 0.010500437580049038, + 0.01217955257743597, + 0.029832296073436737, + 0.036626286804676056, + 0.04054044559597969, + 0.024440046399831772, + 0.006546421907842159, + 0.004357637371867895, + 0.014216525480151176, + 0.03401464968919754, + 0.004984833765774965, + 0.014335782267153263, + 0.014658776111900806, + 0.00446724146604538, + 0.0037720641121268272, + 0.004310037475079298, + 0.0061235385946929455, + 0.004646555054932833, + 0.004528909455984831, + 0.0066056321375072, + 0.005877596791833639, + 0.005811171606183052, + 0.005139409564435482, + 0.005136115476489067, + 0.004747922997921705, + 0.008104497566819191, + 0.0031285141594707966, + 0.005039248615503311, + 0.0024966076016426086, + 0.004569356329739094, + 0.012558434158563614, + 0.004555908031761646, + 0.016714131459593773, + 0.00457816943526268, + 0.005698627792298794, + 0.005840491037815809, + 0.010768027976155281, + 0.006010189652442932, + 0.009786100126802921, + 0.0068482886999845505, + 0.011504198424518108, + 0.0127152344211936, + 0.011028524488210678, + 0.010651422664523125, + 0.01966571994125843, + 0.014299574308097363, + 0.005903872195631266, + 0.013137037865817547, + 0.012023796327412128, + 0.015013575553894043, + 0.014419044367969036, + 0.014502691105008125, + 0.024554317817091942, + 0.020010102540254593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.06521180272102356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637951821088791 + ], + "opposite_logprob_tokens": [ + -3.6513795852661133 + ], + "logprob_reward": -0.02637951821088791, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637951821088791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10343246906995773, + 0.04278763383626938, + 0.02957562543451786, + 0.078868068754673, + 0.1711495816707611, + 0.051164206117391586, + 0.07612122595310211, + 0.2526249885559082, + 0.05154196172952652, + 0.14273418486118317 + ], + "credit_map": [ + 0.015127014368772507, + 0.011020716279745102, + 0.017167260870337486, + 0.007880316115915775, + 0.017473388463258743, + 0.011703245341777802, + 0.01018361933529377, + 0.012876922264695168, + 0.003202257677912712, + 0.0024893353693187237, + 0.003979701083153486, + 0.004313996061682701, + 0.006059049163013697, + 0.009146174415946007, + 0.0066983336582779884, + 0.006898791529238224, + 0.002645551459863782, + 0.0031544778030365705, + 0.0030333336908370256, + 0.002611471340060234, + 0.004573816433548927, + 0.005520844366401434, + 0.0030663462821394205, + 0.004969787783920765, + 0.006127193570137024, + 0.008090385235846043, + 0.00568258436396718, + 0.006809079088270664, + 0.010012145154178143, + 0.016173290088772774, + 0.017845463007688522, + 0.008127938024699688, + 0.01290329359471798, + 0.00926466379314661, + 0.034667156636714935, + 0.05084272101521492, + 0.00877933669835329, + 0.021127482876181602, + 0.02267516404390335, + 0.01088979933410883, + 0.004319282714277506, + 0.006005538161844015, + 0.008957474492490292, + 0.006076798774302006, + 0.005969754420220852, + 0.008387946523725986, + 0.006068414077162743, + 0.005379004869610071, + 0.00920936930924654, + 0.009136222302913666, + 0.009500556625425816, + 0.01591954194009304, + 0.006525533739477396, + 0.012172962538897991, + 0.00462751742452383, + 0.009029529988765717, + 0.05284538120031357, + 0.019121987745165825, + 0.048012617975473404, + 0.014317618682980537, + 0.016957061365246773, + 0.02136312425136566, + 0.05869060382246971, + 0.02131662517786026, + 0.004367399495095015, + 0.004670552909374237, + 0.007348170969635248, + 0.007574068382382393, + 0.007134379353374243, + 0.007111701648682356, + 0.007574948947876692, + 0.0057607474736869335, + 0.007871693931519985, + 0.015443037264049053, + 0.020277364179491997, + 0.023700816556811333, + 0.01900215446949005, + 0.01934349723160267, + 0.020278163254261017, + 0.016817478463053703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.02637951821088791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347240433096886 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347240433096886, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347240433096886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18307606875896454, + 0.0594097264111042, + 0.1290290653705597, + 0.07446194440126419, + 0.17404848337173462, + 0.0667475163936615, + 0.02991824597120285, + 0.10962951928377151, + 0.06941082328557968, + 0.10426852852106094 + ], + "credit_map": [ + 0.026670685037970543, + 0.017032699659466743, + 0.02077409066259861, + 0.017380915582180023, + 0.03509931638836861, + 0.0190384890884161, + 0.03094804286956787, + 0.016131848096847534, + 0.006320560816675425, + 0.00428105890750885, + 0.005834661889821291, + 0.0073355333879590034, + 0.007136599626392126, + 0.00949076283723116, + 0.01067812368273735, + 0.008332423865795135, + 0.01173307653516531, + 0.013334161601960659, + 0.016124585643410683, + 0.01396472193300724, + 0.016786210238933563, + 0.023102251812815666, + 0.014717403799295425, + 0.019266655668616295, + 0.006367056164890528, + 0.005948036443442106, + 0.007099759299308062, + 0.004622662905603647, + 0.011252550408244133, + 0.015464380383491516, + 0.01544116996228695, + 0.008266321383416653, + 0.00972073432058096, + 0.014963780529797077, + 0.023487070575356483, + 0.04741152375936508, + 0.013497356325387955, + 0.02765890397131443, + 0.02019493468105793, + 0.017114171758294106, + 0.007310166954994202, + 0.0054986076429486275, + 0.011801893822848797, + 0.0052030086517333984, + 0.00785021111369133, + 0.010322770103812218, + 0.008794636465609074, + 0.009966224431991577, + 0.0033004353754222393, + 0.0037210641894489527, + 0.004047649446874857, + 0.004892620258033276, + 0.004003292880952358, + 0.0043885475024580956, + 0.002350363414734602, + 0.0032142754644155502, + 0.02822493016719818, + 0.008548092097043991, + 0.015825564041733742, + 0.0056388648226857185, + 0.006762938108295202, + 0.009565659798681736, + 0.027967920526862144, + 0.007095547392964363, + 0.006474530324339867, + 0.005946699529886246, + 0.0087681720033288, + 0.007282423321157694, + 0.010281750001013279, + 0.009475779719650745, + 0.010916708037257195, + 0.010264758951961994, + 0.005570839624851942, + 0.010904605500400066, + 0.009237993508577347, + 0.012740600854158401, + 0.015187526121735573, + 0.01641934923827648, + 0.01932574063539505, + 0.014881874434649944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.026347240433096886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188835397362709 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188835397362709, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188835397362709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08781211823225021, + 0.044781990349292755, + 0.057958539575338364, + 0.06314188987016678, + 0.13025014102458954, + 0.13763020932674408, + 0.10811498761177063, + 0.13903354108333588, + 0.06927435845136642, + 0.16200223565101624 + ], + "credit_map": [ + 0.017158949747681618, + 0.010271567851305008, + 0.014027721248567104, + 0.006859293673187494, + 0.010578631423413754, + 0.011467265896499157, + 0.0100327767431736, + 0.00741590978577733, + 0.0034286160953342915, + 0.003418445121496916, + 0.003977649379521608, + 0.004384530242532492, + 0.006010829005390406, + 0.008894502185285091, + 0.006861464586108923, + 0.007805958390235901, + 0.004229572135955095, + 0.004767231177538633, + 0.004505608230829239, + 0.006639707367867231, + 0.011087710037827492, + 0.010492540895938873, + 0.00902478490024805, + 0.007211389020085335, + 0.004693866241723299, + 0.006725891958922148, + 0.004610415082424879, + 0.006024959497153759, + 0.007601672783493996, + 0.013185139745473862, + 0.012995230965316296, + 0.007304714526981115, + 0.007707254495471716, + 0.012246008031070232, + 0.02598874643445015, + 0.02767685428261757, + 0.009138816967606544, + 0.018509525805711746, + 0.014379383064806461, + 0.014603565447032452, + 0.011978477239608765, + 0.012243303470313549, + 0.021501734852790833, + 0.015519017353653908, + 0.01810351572930813, + 0.020071830600500107, + 0.019417092204093933, + 0.018795233219861984, + 0.012506973929703236, + 0.013402906246483326, + 0.013101759366691113, + 0.021298138424754143, + 0.012360635213553905, + 0.01329493522644043, + 0.008769508451223373, + 0.013380120508372784, + 0.017442425712943077, + 0.012575236149132252, + 0.02952878177165985, + 0.015171392820775509, + 0.016906246542930603, + 0.023121709004044533, + 0.009162923321127892, + 0.015124828554689884, + 0.006146957632154226, + 0.005988668650388718, + 0.01198368426412344, + 0.009250497445464134, + 0.010680531151592731, + 0.01077233999967575, + 0.008314186707139015, + 0.006137489341199398, + 0.01432764157652855, + 0.018373731523752213, + 0.018415160477161407, + 0.02467568777501583, + 0.020796239376068115, + 0.01900193840265274, + 0.017518948763608932, + 0.028892885893583298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.014188835397362709 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive basketball player wearing a tight, sleeveless black athletic tank top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the player swing a heavy, cloth-wrapped mallet to strike a massive, ancient bronze gong suspended from a dark wooden frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera initially focus on the player's forearm, showing it tensing and bulging with veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a swift and smooth rack focus to shift attention from the arm to the gong surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gong surface show visible concentric ripples spreading like water waves after being struck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290083080530167 + ], + "opposite_logprob_tokens": [ + -2.237900733947754 + ], + "logprob_reward": -0.11290083080530167, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290083080530167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10389348119497299, + 0.03211985528469086, + 0.03780350461602211, + 0.07228035479784012, + 0.1271393895149231, + 0.07202161103487015, + 0.06879187375307083, + 0.13661234080791473, + 0.07494869083166122, + 0.27438896894454956 + ], + "credit_map": [ + 0.016801752150058746, + 0.011876099742949009, + 0.016707569360733032, + 0.00727203581482172, + 0.013867244124412537, + 0.012433724477887154, + 0.012767815962433815, + 0.012167228385806084, + 0.0051544466987252235, + 0.0025393227115273476, + 0.004110553301870823, + 0.0029920777305960655, + 0.003144708229228854, + 0.004720449913293123, + 0.004514965694397688, + 0.004943328909575939, + 0.0038699270226061344, + 0.004284276161342859, + 0.004202894866466522, + 0.004981388803571463, + 0.004796267487108707, + 0.005199231207370758, + 0.005532875657081604, + 0.004936640616506338, + 0.009096094407141209, + 0.00504999328404665, + 0.012253277003765106, + 0.0060158343985676765, + 0.007867703214287758, + 0.012920335866510868, + 0.010916346684098244, + 0.008160763420164585, + 0.0071012843400239944, + 0.007422464434057474, + 0.015484124422073364, + 0.04285271093249321, + 0.006249079015105963, + 0.01880534179508686, + 0.01930864527821541, + 0.009915739297866821, + 0.007673454470932484, + 0.005587477702647448, + 0.011640533804893494, + 0.00873289629817009, + 0.009131530299782753, + 0.010427916422486305, + 0.010681960731744766, + 0.008145833387970924, + 0.006781119853258133, + 0.005377375986427069, + 0.006181028671562672, + 0.008241215720772743, + 0.006077714264392853, + 0.0170077383518219, + 0.007108656223863363, + 0.012017018161714077, + 0.036452364176511765, + 0.010822280310094357, + 0.015651291236281395, + 0.005784344393759966, + 0.00965102855116129, + 0.012470756657421589, + 0.03570093214511871, + 0.010079341940581799, + 0.008987603709101677, + 0.004879551474004984, + 0.010235829278826714, + 0.008100450038909912, + 0.010800237767398357, + 0.009191407822072506, + 0.011963455937802792, + 0.010790149681270123, + 0.013226214796304703, + 0.027742348611354828, + 0.03998875245451927, + 0.031048931181430817, + 0.034587278962135315, + 0.031411245465278625, + 0.04648846760392189, + 0.049895696341991425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.11290083080530167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_5/index_tensor([3893, 3893])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_5/index_tensor([3893, 3893])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..332fcc15062b6f17f0538f835f634ab68ea68477 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_5/index_tensor([3893, 3893])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a sleek, matte-black sports motorcycle speeding down a rain-soaked, neon-lit city street at midnight. A rider, clad in a tight black leather jersey and a full-face helmet, leans forward, pressed against the motorcycle's body. The dark visor reflects the shimmering light of the massive, multicolored billboards adorning the tall buildings. The damp asphalt acts like a black mirror, reflecting intensely saturated cyan, magenta, and blue hues. A thin layer of airborne vapor creates a hazy, layered light. As the rider decisively lowers the motorcycle around a sharp turn, the metal side of the motorcycle rubs violently against the rough pavement, sending a trail of dazzling orange-red sparks flying. These sparks are like short-lived flames, hissing through the night, not only illuminating the splashes of the high-speed rotating wheels, but also leaving a fiery trail on the slippery ground. The whole scene is full of speed and power, creating a cyberpunk, adrenaline-pumping atmosphere of tension.", + "seed": 17506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_5/index_3893_seed_17506_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05465520787984133, + "vlm_logprob": -0.05465520787984133, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05465520787984133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019933262839913368 + ], + "opposite_logprob_tokens": [ + -4.019933223724365 + ], + "logprob_reward": -0.019933262839913368, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019933262839913368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17086297273635864, + 0.07087867707014084, + 0.0864156186580658, + 0.113348588347435, + 0.059508103877305984, + 0.05939684063196182, + 0.02768746018409729, + 0.05711205676198006, + 0.2336522787809372, + 0.1211373582482338 + ], + "credit_map": [ + 0.033834993839263916, + 0.01864449493587017, + 0.020888738334178925, + 0.018326232209801674, + 0.012051885016262531, + 0.022246215492486954, + 0.01956496201455593, + 0.025305448099970818, + 0.009489166550338268, + 0.006931752897799015, + 0.008830634877085686, + 0.005895405076444149, + 0.007870284840464592, + 0.00837251078337431, + 0.01373753510415554, + 0.009751379489898682, + 0.011058788746595383, + 0.008458146825432777, + 0.010739312507212162, + 0.01006254367530346, + 0.008493908680975437, + 0.021203061565756798, + 0.007462100125849247, + 0.008937763050198555, + 0.012073960155248642, + 0.013285782188177109, + 0.013418843038380146, + 0.012333850376307964, + 0.019103074446320534, + 0.015402966178953648, + 0.018017102032899857, + 0.00971300434321165, + 0.010789060033857822, + 0.009277698583900928, + 0.003917717840522528, + 0.007125035393983126, + 0.007074200548231602, + 0.005234406795352697, + 0.009946780279278755, + 0.00614320021122694, + 0.009789668954908848, + 0.0070182750932872295, + 0.009564027190208435, + 0.006709378212690353, + 0.00581770297139883, + 0.005190880969166756, + 0.008281747810542583, + 0.0070251584984362125, + 0.004548241849988699, + 0.00538503285497427, + 0.003097236854955554, + 0.0030655746813863516, + 0.003637507325038314, + 0.0018133288249373436, + 0.002430677879601717, + 0.0037098610773682594, + 0.006284787785261869, + 0.006178824231028557, + 0.0062901671044528484, + 0.009486601687967777, + 0.006096079014241695, + 0.007344528567045927, + 0.007965808734297752, + 0.007465258706361055, + 0.02285064198076725, + 0.029889393597841263, + 0.01346611324697733, + 0.04562070965766907, + 0.02424788288772106, + 0.03445481136441231, + 0.03907081484794617, + 0.024051928892731667, + 0.015050448477268219, + 0.015474310144782066, + 0.010098296217620373, + 0.015282981097698212, + 0.02117633819580078, + 0.014233293011784554, + 0.018207214772701263, + 0.011614483781158924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.019933262839913368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298100650310516 + ], + "opposite_logprob_tokens": [ + -2.237981081008911 + ], + "logprob_reward": -0.11298100650310516, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298100650310516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22384442389011383, + 0.07517550140619278, + 0.15886937081813812, + 0.07975196093320847, + 0.04790286719799042, + 0.06348675489425659, + 0.025612499564886093, + 0.03201809898018837, + 0.16332563757896423, + 0.1300128847360611 + ], + "credit_map": [ + 0.03545915335416794, + 0.02485404536128044, + 0.029259424656629562, + 0.022271081805229187, + 0.016698502004146576, + 0.04619969427585602, + 0.01749604381620884, + 0.031606487929821014, + 0.011737419292330742, + 0.007982688024640083, + 0.009799906052649021, + 0.007368376478552818, + 0.007773691788315773, + 0.008369723334908485, + 0.01228931825608015, + 0.009854382835328579, + 0.02162904664874077, + 0.016610154882073402, + 0.025093119591474533, + 0.017973147332668304, + 0.015774071216583252, + 0.031935323029756546, + 0.011012007482349873, + 0.018842505291104317, + 0.01049263495951891, + 0.011826510541141033, + 0.010789548046886921, + 0.013533859513700008, + 0.008706127293407917, + 0.007402032148092985, + 0.009411788545548916, + 0.00758946081623435, + 0.00670935632660985, + 0.007059974130243063, + 0.003179417224600911, + 0.007852188311517239, + 0.00482253497466445, + 0.0041034286841750145, + 0.00882039312273264, + 0.005355572793632746, + 0.007909862324595451, + 0.008938128128647804, + 0.011467531323432922, + 0.008074335753917694, + 0.005223254673182964, + 0.005151495337486267, + 0.009464699774980545, + 0.007257452234625816, + 0.003755154088139534, + 0.003294319147244096, + 0.0035797818563878536, + 0.003591533750295639, + 0.003929531667381525, + 0.002062764950096607, + 0.002258811378851533, + 0.0031406020279973745, + 0.0032466789707541466, + 0.004024243447929621, + 0.0039486363530159, + 0.005256755743175745, + 0.0027309744618833065, + 0.0032455455511808395, + 0.0050861104391515255, + 0.0044791544787585735, + 0.01597149670124054, + 0.02011270821094513, + 0.00854492373764515, + 0.03640462085604668, + 0.015297308564186096, + 0.023204797878861427, + 0.026771705597639084, + 0.017018070444464684, + 0.019932348281145096, + 0.017763758078217506, + 0.012544762343168259, + 0.01563357561826706, + 0.01666056551039219, + 0.009891248308122158, + 0.01691061072051525, + 0.020676016807556152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.11298100650310516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653049357235432 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653049357235432, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653049357235432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18772973120212555, + 0.06366128474473953, + 0.12659576535224915, + 0.13176050782203674, + 0.0933578833937645, + 0.07094278931617737, + 0.021553218364715576, + 0.06833989173173904, + 0.05931689962744713, + 0.17674200236797333 + ], + "credit_map": [ + 0.0284088384360075, + 0.024058278650045395, + 0.020355410873889923, + 0.022192206233739853, + 0.013891431502997875, + 0.030216922983527184, + 0.024177461862564087, + 0.024429183453321457, + 0.008243275806307793, + 0.005770159885287285, + 0.00839115958660841, + 0.005093232728540897, + 0.007189278956502676, + 0.008612028323113918, + 0.011900697834789753, + 0.008461452089250088, + 0.01107072364538908, + 0.01018272340297699, + 0.00839367788285017, + 0.014690011739730835, + 0.01574314385652542, + 0.04136678948998451, + 0.011220365762710571, + 0.013928326778113842, + 0.01460447907447815, + 0.014636191539466381, + 0.015100564807653427, + 0.010898539796471596, + 0.02214701846241951, + 0.02028672955930233, + 0.0196932815015316, + 0.014393698424100876, + 0.013587720692157745, + 0.011977296322584152, + 0.007136572618037462, + 0.012089586816728115, + 0.009231355041265488, + 0.01014311145991087, + 0.017680618911981583, + 0.01151161827147007, + 0.010873381048440933, + 0.00665357057005167, + 0.01155707985162735, + 0.008510295301675797, + 0.006837674416601658, + 0.006634073331952095, + 0.012654092162847519, + 0.007222623098641634, + 0.0031030664686113596, + 0.002863583154976368, + 0.0028743103612214327, + 0.002762535819783807, + 0.002691274043172598, + 0.0013672609347850084, + 0.0019331336952745914, + 0.003958054352551699, + 0.006699695717543364, + 0.007425222545862198, + 0.008157203905284405, + 0.01049756072461605, + 0.008033648133277893, + 0.009100904688239098, + 0.009560399688780308, + 0.008865254931151867, + 0.007404448464512825, + 0.006598311942070723, + 0.004728195257484913, + 0.005202118307352066, + 0.005300190299749374, + 0.012269798666238785, + 0.009676456451416016, + 0.008137381635606289, + 0.015692560002207756, + 0.02519748918712139, + 0.014598820358514786, + 0.02971591055393219, + 0.025918196886777878, + 0.0176782738417387, + 0.024937743321061134, + 0.02300301007926464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.008653049357235432 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706926465034485 + ], + "opposite_logprob_tokens": [ + -2.1270692348480225 + ], + "logprob_reward": -0.12706926465034485, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706926465034485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2975381314754486, + 0.13517527282238007, + 0.05303693935275078, + 0.07992773503065109, + 0.11884172260761261, + 0.08021946996450424, + 0.02146335132420063, + 0.04437212646007538, + 0.06342341005802155, + 0.10600178688764572 + ], + "credit_map": [ + 0.05211066082119942, + 0.036855556070804596, + 0.033375754952430725, + 0.029081864282488823, + 0.02021670900285244, + 0.045099951326847076, + 0.045000310987234116, + 0.03579733148217201, + 0.013155215419828892, + 0.016763441264629364, + 0.009395454078912735, + 0.0113674346357584, + 0.015162508003413677, + 0.01316696498543024, + 0.03696543350815773, + 0.019198834896087646, + 0.00535600446164608, + 0.003734477562829852, + 0.004285098519176245, + 0.005020603071898222, + 0.004492538515478373, + 0.008706743828952312, + 0.010003730654716492, + 0.011437742970883846, + 0.008765741251409054, + 0.014059157110750675, + 0.011529665440320969, + 0.011123212054371834, + 0.007661681156605482, + 0.011716065928339958, + 0.00870030652731657, + 0.006371907889842987, + 0.029008714482188225, + 0.020648330450057983, + 0.009358875453472137, + 0.012005005031824112, + 0.010353273712098598, + 0.00970644224435091, + 0.016277020797133446, + 0.011484062299132347, + 0.012642684392631054, + 0.011323966085910797, + 0.01458107028156519, + 0.008973302319645882, + 0.007011196110397577, + 0.007288867607712746, + 0.010123265907168388, + 0.00827512051910162, + 0.002113519934937358, + 0.002993338042870164, + 0.003009706735610962, + 0.0027030231431126595, + 0.0034421663731336594, + 0.0016283802688121796, + 0.002158306771889329, + 0.003414911450818181, + 0.0050580850802361965, + 0.004623204004019499, + 0.0060614184476435184, + 0.006957067176699638, + 0.0047439285553991795, + 0.005072127562016249, + 0.005918172653764486, + 0.005938119720667601, + 0.005869342479854822, + 0.006527725141495466, + 0.0044760750606656075, + 0.012824328616261482, + 0.005730446428060532, + 0.008601879701018333, + 0.01141319889575243, + 0.007980410009622574, + 0.01419841405004263, + 0.014399023726582527, + 0.00951477326452732, + 0.01710662804543972, + 0.013402664102613926, + 0.009611569344997406, + 0.015629136934876442, + 0.012139574624598026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.12706926465034485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639456048607826 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639456048607826, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639456048607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1639784574508667, + 0.07223481684923172, + 0.06600818037986755, + 0.10326981544494629, + 0.18616396188735962, + 0.056099746376276016, + 0.039163295179605484, + 0.06163576990365982, + 0.16370224952697754, + 0.08774372935295105 + ], + "credit_map": [ + 0.02491704747080803, + 0.02150692790746689, + 0.01854945719242096, + 0.019397549331188202, + 0.010697823017835617, + 0.020562605932354927, + 0.02367391437292099, + 0.024673113599419594, + 0.008325054310262203, + 0.007530921138823032, + 0.006338515784591436, + 0.008522750809788704, + 0.008023195900022984, + 0.007032268214970827, + 0.01582285389304161, + 0.010639244690537453, + 0.007999179884791374, + 0.0067164101637899876, + 0.008687240071594715, + 0.006304736714810133, + 0.005585423205047846, + 0.006527016870677471, + 0.009500347077846527, + 0.014687824994325638, + 0.010796032845973969, + 0.01358666643500328, + 0.01707194559276104, + 0.015108861029148102, + 0.01089806854724884, + 0.011981024406850338, + 0.012640230357646942, + 0.011186981573700905, + 0.024061480537056923, + 0.018620509654283524, + 0.016064681112766266, + 0.02952618896961212, + 0.02010052651166916, + 0.02455616183578968, + 0.03299286961555481, + 0.020241523161530495, + 0.006983985658735037, + 0.008378631435334682, + 0.00906042754650116, + 0.006225736811757088, + 0.005204200278967619, + 0.005402851849794388, + 0.007831350900232792, + 0.007012559100985527, + 0.0074402159079909325, + 0.006871374789625406, + 0.004099283833056688, + 0.00487116863951087, + 0.004282817244529724, + 0.0026146965101361275, + 0.0033183011692017317, + 0.005665432196110487, + 0.006767715327441692, + 0.005726810544729233, + 0.007038821466267109, + 0.009857933968305588, + 0.007655265741050243, + 0.005874678026884794, + 0.009596358053386211, + 0.009118182584643364, + 0.01848060078918934, + 0.019271276891231537, + 0.010944195091724396, + 0.025954069569706917, + 0.0167755838483572, + 0.028659341856837273, + 0.025992505252361298, + 0.017624666914343834, + 0.0123064573854208, + 0.012048307806253433, + 0.00775587884709239, + 0.012722741812467575, + 0.012265169993042946, + 0.010150031186640263, + 0.008896742016077042, + 0.011598384939134121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.004639456048607826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a sleek, matte-black sports motorcycle speeding down a rain-soaked, neon-lit city street at midnight. A rider, clad in a tight black leather jersey and a full-face helmet, leans forward, pressed against the motorcycle's body. The dark visor reflects the shimmering light of the massive, multicolored billboards adorning the tall buildings. The damp asphalt acts like a black mirror, reflecting intensely saturated cyan, magenta, and blue hues. A thin layer of airborne vapor creates a hazy, layered light. As the rider decisively lowers the motorcycle around a sharp turn, the metal side of the motorcycle rubs violently against the rough pavement, sending a trail of dazzling orange-red sparks flying. These sparks are like short-lived flames, hissing through the night, not only illuminating the splashes of the high-speed rotating wheels, but also leaving a fiery trail on the slippery ground. The whole scene is full of speed and power, creating a cyberpunk, adrenaline-pumping atmosphere of tension.", + "seed": 17507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_5/index_3893_seed_17507_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02843079622834921, + "vlm_logprob": -0.02843079622834921, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02843079622834921, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014700782485306263 + ], + "opposite_logprob_tokens": [ + -4.264700889587402 + ], + "logprob_reward": -0.014700782485306263, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014700782485306263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14265897870063782, + 0.09402021765708923, + 0.06013602763414383, + 0.08510816842317581, + 0.0695396363735199, + 0.10072853416204453, + 0.10667042434215546, + 0.13553979992866516, + 0.1455916315317154, + 0.06000663712620735 + ], + "credit_map": [ + 0.018740255385637283, + 0.016679642722010612, + 0.020321005955338478, + 0.029465416446328163, + 0.011089053004980087, + 0.022767042741179466, + 0.01142856478691101, + 0.012167977169156075, + 0.013186219148337841, + 0.011288991197943687, + 0.014523888006806374, + 0.010732048191130161, + 0.015281382948160172, + 0.0062822382897138596, + 0.009637037292122841, + 0.013088393025100231, + 0.009463689289987087, + 0.007273979485034943, + 0.007476625498384237, + 0.005434955004602671, + 0.0116944694891572, + 0.006619102321565151, + 0.006447514519095421, + 0.0057256827130913734, + 0.009550036862492561, + 0.008044430986046791, + 0.01897607371211052, + 0.02020401693880558, + 0.006275930907577276, + 0.010311156511306763, + 0.00695760827511549, + 0.00478890398517251, + 0.008596856147050858, + 0.006081173196434975, + 0.014383235946297646, + 0.010038946755230427, + 0.0057887425646185875, + 0.009086132980883121, + 0.006953511852771044, + 0.00861102994531393, + 0.009861699305474758, + 0.011983811855316162, + 0.012810438871383667, + 0.011361071839928627, + 0.011741427704691887, + 0.010400069877505302, + 0.01200896967202425, + 0.02056102640926838, + 0.024188295006752014, + 0.010632884688675404, + 0.014622061513364315, + 0.007962310686707497, + 0.00852151494473219, + 0.013088862411677837, + 0.012283391319215298, + 0.01537108514457941, + 0.01386574562638998, + 0.02111176587641239, + 0.01661977358162403, + 0.015655338764190674, + 0.014360198751091957, + 0.015349062159657478, + 0.013282513245940208, + 0.025295384228229523, + 0.023039426654577255, + 0.01727139763534069, + 0.015064500272274017, + 0.019143899902701378, + 0.02104441449046135, + 0.01592593640089035, + 0.01270739734172821, + 0.0213946420699358, + 0.006240055430680513, + 0.00789314042776823, + 0.0044130319729447365, + 0.005913420580327511, + 0.015163672156631947, + 0.006694805342704058, + 0.005744581110775471, + 0.007943923585116863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.014700782485306263 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899660617113113 + ], + "opposite_logprob_tokens": [ + -2.5789966583251953 + ], + "logprob_reward": -0.07899660617113113, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899660617113113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15118113160133362, + 0.1207587718963623, + 0.04687204211950302, + 0.14852607250213623, + 0.1426466703414917, + 0.08556322753429413, + 0.05447670444846153, + 0.05344854295253754, + 0.11162162572145462, + 0.08490519225597382 + ], + "credit_map": [ + 0.02200184017419815, + 0.01600993052124977, + 0.026078762486577034, + 0.024567050859332085, + 0.01397429034113884, + 0.02659536339342594, + 0.010008259676396847, + 0.011945636011660099, + 0.015123169869184494, + 0.012092918157577515, + 0.017398787662386894, + 0.015160255134105682, + 0.021236546337604523, + 0.014224889688193798, + 0.012725313194096088, + 0.01279687974601984, + 0.008193306624889374, + 0.0036064847372472286, + 0.008288715034723282, + 0.006981805898249149, + 0.005122231785207987, + 0.007920735515654087, + 0.0033984770998358727, + 0.003360282862558961, + 0.018151039257645607, + 0.014146610163152218, + 0.032440897077322006, + 0.037137191742658615, + 0.010893117636442184, + 0.015453060157597065, + 0.011890115216374397, + 0.008414031006395817, + 0.019157379865646362, + 0.01406810525804758, + 0.02850150875747204, + 0.017377469688653946, + 0.013564600609242916, + 0.022029295563697815, + 0.008661291562020779, + 0.019287006929516792, + 0.008429663255810738, + 0.0073353610932827, + 0.007314434740692377, + 0.007132441271096468, + 0.012245718389749527, + 0.00970003567636013, + 0.010186743922531605, + 0.023218825459480286, + 0.00872842874377966, + 0.005493344739079475, + 0.007129644975066185, + 0.0036804855335503817, + 0.005774987395852804, + 0.007829050533473492, + 0.006861628964543343, + 0.008979137055575848, + 0.003370706457644701, + 0.009254654869437218, + 0.004479996860027313, + 0.008855147287249565, + 0.005726816598325968, + 0.0044931937009096146, + 0.008201869204640388, + 0.009066160768270493, + 0.01591763086616993, + 0.014312813989818096, + 0.009232690557837486, + 0.015881286934018135, + 0.014824076555669308, + 0.0107894167304039, + 0.0171627476811409, + 0.013500966131687164, + 0.01751507632434368, + 0.008389048278331757, + 0.008998971432447433, + 0.008788738399744034, + 0.010055013000965118, + 0.008127555251121521, + 0.01240294985473156, + 0.010627838782966137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.07899660617113113 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742820143699646 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742820143699646, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742820143699646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.193141371011734, + 0.20653966069221497, + 0.08254742622375488, + 0.07267387211322784, + 0.07361346483230591, + 0.09393090754747391, + 0.1133938878774643, + 0.0731700211763382, + 0.05237407609820366, + 0.03861531242728233 + ], + "credit_map": [ + 0.019718436524271965, + 0.02285459265112877, + 0.0349644273519516, + 0.037849023938179016, + 0.018847208470106125, + 0.021220611408352852, + 0.013705319724977016, + 0.023981750011444092, + 0.01954607293009758, + 0.024372151121497154, + 0.020314255729317665, + 0.024683700874447823, + 0.05593132600188255, + 0.018019163981080055, + 0.02420452982187271, + 0.019468456506729126, + 0.010107830166816711, + 0.008195996284484863, + 0.009135616011917591, + 0.011080091819167137, + 0.007443500682711601, + 0.024201802909374237, + 0.004164545796811581, + 0.008218041621148586, + 0.008783457800745964, + 0.007109510712325573, + 0.014736788347363472, + 0.01684020832180977, + 0.004708486143499613, + 0.008317185565829277, + 0.0065247309394180775, + 0.0056535108014941216, + 0.00920669361948967, + 0.009381323121488094, + 0.011859672144055367, + 0.00920356810092926, + 0.007044895086437464, + 0.007910823449492455, + 0.008934080600738525, + 0.010072409175336361, + 0.009932232089340687, + 0.009379788301885128, + 0.010904182679951191, + 0.01023348793387413, + 0.012338737957179546, + 0.009997527115046978, + 0.011222471483051777, + 0.01992248184978962, + 0.0184489656239748, + 0.00999872013926506, + 0.014583863317966461, + 0.011453675106167793, + 0.01024069543927908, + 0.015800487250089645, + 0.013777031563222408, + 0.01909044198691845, + 0.006827039644122124, + 0.00953663606196642, + 0.008336616680026054, + 0.010901485569775105, + 0.011533075012266636, + 0.008150502108037472, + 0.008467141538858414, + 0.009417527355253696, + 0.007448056247085333, + 0.006754585076123476, + 0.004529977682977915, + 0.008143842220306396, + 0.009353718720376492, + 0.0050449371337890625, + 0.005426987074315548, + 0.005671975668519735, + 0.006845253054052591, + 0.0053568133153021336, + 0.0032274778932332993, + 0.004627511836588383, + 0.005213066469877958, + 0.003228906076401472, + 0.005143941380083561, + 0.004972340073436499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.006742820143699646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810039907693863 + ], + "opposite_logprob_tokens": [ + -3.288100481033325 + ], + "logprob_reward": -0.03810039907693863, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810039907693863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0439545102417469, + 0.03903881832957268, + 0.06552205979824066, + 0.05031394213438034, + 0.07101448625326157, + 0.09142281115055084, + 0.055069610476493835, + 0.4678911864757538, + 0.0954332947731018, + 0.02033928409218788 + ], + "credit_map": [ + 0.005007977597415447, + 0.005416756495833397, + 0.0060575539246201515, + 0.006632884033024311, + 0.005572475492954254, + 0.0063529470935463905, + 0.0033327164128422737, + 0.005581197794526815, + 0.0053572519682347775, + 0.0038988881278783083, + 0.004710420034825802, + 0.004214724991470575, + 0.008509042672812939, + 0.00502681452780962, + 0.0036324604880064726, + 0.0036892155185341835, + 0.012737750075757504, + 0.006199404131621122, + 0.008179671131074429, + 0.007168378680944443, + 0.01146444771438837, + 0.008318133652210236, + 0.005969609599560499, + 0.00548466295003891, + 0.004778714384883642, + 0.0044680265709757805, + 0.009768465533852577, + 0.007964552380144596, + 0.003641755785793066, + 0.011550351046025753, + 0.003834974952042103, + 0.004307105205953121, + 0.009678758680820465, + 0.009184845723211765, + 0.008951116353273392, + 0.01158334780484438, + 0.006688740104436874, + 0.008981234394013882, + 0.008329019881784916, + 0.007617420516908169, + 0.011083368211984634, + 0.010171390138566494, + 0.007785469759255648, + 0.006054268684238195, + 0.013432233594357967, + 0.010499146766960621, + 0.009294702671468258, + 0.023102235049009323, + 0.010942029766738415, + 0.005356912035495043, + 0.007153472863137722, + 0.004088100977241993, + 0.005121921189129353, + 0.006692550145089626, + 0.00661453977227211, + 0.00910008605569601, + 0.04540276899933815, + 0.0728013664484024, + 0.05005843564867973, + 0.04914829879999161, + 0.050390273332595825, + 0.054837435483932495, + 0.05348844826221466, + 0.0917641818523407, + 0.014613321982324123, + 0.011114080436527729, + 0.010387172922492027, + 0.013100333511829376, + 0.010907568968832493, + 0.010800719261169434, + 0.010291504673659801, + 0.014218597672879696, + 0.0023844989482313395, + 0.0022642449475824833, + 0.001614149659872055, + 0.0029371262062340975, + 0.0028234082274138927, + 0.0025702365674078465, + 0.003100987523794174, + 0.00264463247731328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.03810039907693863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613373264670372 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613373264670372, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613373264670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1761086881160736, + 0.056543342769145966, + 0.07782510668039322, + 0.08446627855300903, + 0.12730787694454193, + 0.08332815766334534, + 0.08584565669298172, + 0.06315923482179642, + 0.17950570583343506, + 0.06590993702411652 + ], + "credit_map": [ + 0.015975728631019592, + 0.01755380816757679, + 0.02364322356879711, + 0.0323360450565815, + 0.020067930221557617, + 0.023431722074747086, + 0.01941956952214241, + 0.023680657148361206, + 0.009714453481137753, + 0.005630169995129108, + 0.009865635074675083, + 0.005277764983475208, + 0.008019255474209785, + 0.00505554024130106, + 0.005073636304587126, + 0.007906890474259853, + 0.010365920141339302, + 0.008549856022000313, + 0.017411667853593826, + 0.010848498903214931, + 0.01121053658425808, + 0.007992813363671303, + 0.005821052007377148, + 0.0056247650645673275, + 0.008963849395513535, + 0.008924594148993492, + 0.014971432276070118, + 0.01940973289310932, + 0.00645325006917119, + 0.010172327049076557, + 0.009057310409843922, + 0.006513787899166346, + 0.017672695219516754, + 0.01302280928939581, + 0.026775317266583443, + 0.01949845813214779, + 0.009255996905267239, + 0.010723479092121124, + 0.013776636682450771, + 0.016582494601607323, + 0.011535927653312683, + 0.0082246707752347, + 0.012709235772490501, + 0.008198595605790615, + 0.011460869573056698, + 0.007984884083271027, + 0.011404448188841343, + 0.01180952787399292, + 0.009896504692733288, + 0.007038963492959738, + 0.011269963346421719, + 0.010363026522099972, + 0.00964669231325388, + 0.013922100886702538, + 0.010910598561167717, + 0.012797812931239605, + 0.0071823871694505215, + 0.008712164126336575, + 0.007109564263373613, + 0.006617579143494368, + 0.010284991934895515, + 0.009614828042685986, + 0.007317381910979748, + 0.006320346146821976, + 0.024955373257398605, + 0.0237237848341465, + 0.016878021880984306, + 0.025521734729409218, + 0.027598878368735313, + 0.0194962527602911, + 0.014242267236113548, + 0.02708941325545311, + 0.010900206863880157, + 0.006031768396496773, + 0.005564089398831129, + 0.00816834345459938, + 0.01005363930016756, + 0.0064956978894770145, + 0.009368879720568657, + 0.009327314794063568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.003613373264670372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a sleek, matte-black sports motorcycle speeding down a rain-soaked, neon-lit city street at midnight. A rider, clad in a tight black leather jersey and a full-face helmet, leans forward, pressed against the motorcycle's body. The dark visor reflects the shimmering light of the massive, multicolored billboards adorning the tall buildings. The damp asphalt acts like a black mirror, reflecting intensely saturated cyan, magenta, and blue hues. A thin layer of airborne vapor creates a hazy, layered light. As the rider decisively lowers the motorcycle around a sharp turn, the metal side of the motorcycle rubs violently against the rough pavement, sending a trail of dazzling orange-red sparks flying. These sparks are like short-lived flames, hissing through the night, not only illuminating the splashes of the high-speed rotating wheels, but also leaving a fiery trail on the slippery ground. The whole scene is full of speed and power, creating a cyberpunk, adrenaline-pumping atmosphere of tension.", + "seed": 17508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_5/index_3893_seed_17508_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025686694495379924, + "vlm_logprob": -0.025686694495379924, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.025686694495379924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008785760030150414 + ], + "opposite_logprob_tokens": [ + -5.133785724639893 + ], + "logprob_reward": -0.008785760030150414, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008785760030150414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.078123077750206, + 0.01826375722885132, + 0.033604711294174194, + 0.04340618848800659, + 0.02525428682565689, + 0.6369806528091431, + 0.025253044441342354, + 0.06827935576438904, + 0.03399471566081047, + 0.03684024512767792 + ], + "credit_map": [ + 0.008196347393095493, + 0.011032771319150925, + 0.01020707469433546, + 0.01712765172123909, + 0.0049945455975830555, + 0.009301945567131042, + 0.00937624555081129, + 0.007886495441198349, + 0.001778566394932568, + 0.0023537487722933292, + 0.002657649340108037, + 0.0031182831153273582, + 0.0017154153902083635, + 0.002451783511787653, + 0.0024136530701071024, + 0.0017746586818248034, + 0.004896848928183317, + 0.005756967235356569, + 0.00411229208111763, + 0.003078317968174815, + 0.004935590550303459, + 0.0048949094489216805, + 0.002734568202868104, + 0.0031952145509421825, + 0.004428725223988295, + 0.006285651586949825, + 0.0051222313195466995, + 0.006325190421193838, + 0.005225398577749729, + 0.005379063542932272, + 0.006070692557841539, + 0.004569237120449543, + 0.002292484976351261, + 0.0019027726957574487, + 0.004545122385025024, + 0.0035521413665264845, + 0.0024695885367691517, + 0.003198706079274416, + 0.003147104289382696, + 0.004146366845816374, + 0.09635043889284134, + 0.030462706461548805, + 0.2691902220249176, + 0.03760068863630295, + 0.02269797958433628, + 0.04633159562945366, + 0.08277005702257156, + 0.051576972007751465, + 0.004546924959868193, + 0.002725069411098957, + 0.0033469591289758682, + 0.0022229314781725407, + 0.0029092784970998764, + 0.0025063049979507923, + 0.002634730190038681, + 0.004360845312476158, + 0.011141673661768436, + 0.006733182817697525, + 0.016523150727152824, + 0.005317612085491419, + 0.005989909637719393, + 0.006665022112429142, + 0.006337261758744717, + 0.009571543894708157, + 0.005468207877129316, + 0.005656079389154911, + 0.004392871633172035, + 0.0033801631070673466, + 0.0038556677754968405, + 0.0030421805568039417, + 0.004874897189438343, + 0.0033246458042412996, + 0.005760729778558016, + 0.004119028802961111, + 0.0054432787001132965, + 0.004673020914196968, + 0.003336863359436393, + 0.004817330278456211, + 0.005125898867845535, + 0.003564096987247467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.008785760030150414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866158962249756 + ], + "opposite_logprob_tokens": [ + -3.048661708831787 + ], + "logprob_reward": -0.04866158962249756, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866158962249756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22006778419017792, + 0.10366050899028778, + 0.061952099204063416, + 0.0386250801384449, + 0.026418747380375862, + 0.19086867570877075, + 0.06699530780315399, + 0.06113266572356224, + 0.10825923085212708, + 0.12201989442110062 + ], + "credit_map": [ + 0.026930613443255424, + 0.035652998834848404, + 0.02551492676138878, + 0.037130359560251236, + 0.017977986484766006, + 0.03732648864388466, + 0.02399175614118576, + 0.015542648732662201, + 0.008652540855109692, + 0.006837937515228987, + 0.016317853704094887, + 0.036179330199956894, + 0.004728579893708229, + 0.007726944983005524, + 0.013425123877823353, + 0.0097922058776021, + 0.006632804870605469, + 0.0053660674020648, + 0.008757811971008778, + 0.006543729919940233, + 0.006586749572306871, + 0.013529127463698387, + 0.007528567686676979, + 0.007007238920778036, + 0.00393675547093153, + 0.003906399477273226, + 0.0038450369611382484, + 0.004232481587678194, + 0.004729865118861198, + 0.007232932839542627, + 0.005435583181679249, + 0.00530602689832449, + 0.0030224903021007776, + 0.0030584915075451136, + 0.005154200829565525, + 0.002423820085823536, + 0.0022634530905634165, + 0.0034249979071319103, + 0.0028027219232171774, + 0.0042685712687671185, + 0.024010971188545227, + 0.008151710964739323, + 0.06825309246778488, + 0.013340235687792301, + 0.00984492152929306, + 0.009624608792364597, + 0.03948451578617096, + 0.018158627673983574, + 0.018412858247756958, + 0.00759326433762908, + 0.00872128363698721, + 0.005502797197550535, + 0.005105926189571619, + 0.005906127858906984, + 0.007710449397563934, + 0.008042600005865097, + 0.009655182249844074, + 0.00686185946688056, + 0.011757840402424335, + 0.0049965763464570045, + 0.005996404215693474, + 0.005039918702095747, + 0.007285221014171839, + 0.009539661929011345, + 0.022090168669819832, + 0.02175847254693508, + 0.013070098124444485, + 0.008437996730208397, + 0.010458760894834995, + 0.007910655811429024, + 0.012740334495902061, + 0.011792744509875774, + 0.018429674208164215, + 0.0164132472127676, + 0.0182346161454916, + 0.012070581316947937, + 0.012391801923513412, + 0.016711346805095673, + 0.015444295480847359, + 0.012324335984885693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.04866158962249756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674506975337863 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.00674506975337863, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674506975337863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10614822804927826, + 0.05358127877116203, + 0.09895388036966324, + 0.1248244196176529, + 0.03566373139619827, + 0.1476956307888031, + 0.03753542900085449, + 0.2911164164543152, + 0.03738728538155556, + 0.06709371507167816 + ], + "credit_map": [ + 0.009157069958746433, + 0.02076273411512375, + 0.013506889343261719, + 0.019078614190220833, + 0.008178546093404293, + 0.011266950517892838, + 0.013938885182142258, + 0.010258529335260391, + 0.006283110473304987, + 0.005884106736630201, + 0.00687722722068429, + 0.007459748070687056, + 0.0062789819203317165, + 0.0063933697529137135, + 0.007999734953045845, + 0.006404999177902937, + 0.009081829339265823, + 0.016988161951303482, + 0.010744458995759487, + 0.012932186014950275, + 0.009837638586759567, + 0.014087587594985962, + 0.016481874510645866, + 0.008800143375992775, + 0.012554721906781197, + 0.01565292477607727, + 0.01636294461786747, + 0.017977634444832802, + 0.014990607276558876, + 0.01552150584757328, + 0.017983337864279747, + 0.013780741021037102, + 0.0030015769880264997, + 0.0038602780550718307, + 0.006220194511115551, + 0.003778889775276184, + 0.0033879743423312902, + 0.005157287232577801, + 0.0035136069636791945, + 0.006743918638676405, + 0.02066091075539589, + 0.011793551035225391, + 0.03179463371634483, + 0.008893397636711597, + 0.007772910408675671, + 0.014133183285593987, + 0.024471284821629524, + 0.02817574329674244, + 0.005516025237739086, + 0.004361925181001425, + 0.0043205684050917625, + 0.0036909086629748344, + 0.004232834558933973, + 0.004290822893381119, + 0.00570994196459651, + 0.0054123965092003345, + 0.04285275563597679, + 0.03382911533117294, + 0.054032571613788605, + 0.03137635439634323, + 0.029836883768439293, + 0.030055323615670204, + 0.030048087239265442, + 0.039085302501916885, + 0.004451493266969919, + 0.004107219632714987, + 0.003967165015637875, + 0.002860691864043474, + 0.006868482101708651, + 0.004593638237565756, + 0.005525391548871994, + 0.005013202782720327, + 0.00898379273712635, + 0.006180575583130121, + 0.011652356944978237, + 0.008285637944936752, + 0.004331224597990513, + 0.00914896372705698, + 0.009550081565976143, + 0.008961078710854053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.00674506975337863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062043581157922745 + ], + "opposite_logprob_tokens": [ + -2.8120436668395996 + ], + "logprob_reward": -0.062043581157922745, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062043581157922745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06127805635333061, + 0.03137274086475372, + 0.08232248574495316, + 0.027154840528964996, + 0.03257494047284126, + 0.5901032090187073, + 0.019008399918675423, + 0.0905776247382164, + 0.018955888226628304, + 0.046651776880025864 + ], + "credit_map": [ + 0.005915556102991104, + 0.00904062483459711, + 0.006786207668483257, + 0.01341971755027771, + 0.005084062926471233, + 0.009649267420172691, + 0.006433135364204645, + 0.0049494793638587, + 0.0034871120005846024, + 0.0032601316925138235, + 0.003937887027859688, + 0.0078473761677742, + 0.0023404352832585573, + 0.003879643976688385, + 0.003880066331475973, + 0.002740084193646908, + 0.0104998042806983, + 0.0124601935967803, + 0.008720474317669868, + 0.009942940436303616, + 0.011889344081282616, + 0.010500099509954453, + 0.010972682386636734, + 0.007336935494095087, + 0.003286184975877404, + 0.003669929224997759, + 0.003036455949768424, + 0.003264873754233122, + 0.003798461752012372, + 0.003384504234418273, + 0.0035148856695741415, + 0.0031995417084544897, + 0.004429802298545837, + 0.0033749316353350878, + 0.004629714880138636, + 0.002996434224769473, + 0.003048623213544488, + 0.004946043249219656, + 0.003319533308967948, + 0.005829854402691126, + 0.07562319189310074, + 0.02381761744618416, + 0.24219729006290436, + 0.037013497203588486, + 0.024256953969597816, + 0.03528614714741707, + 0.09187497943639755, + 0.060033511370420456, + 0.004205878358334303, + 0.0026056382339447737, + 0.002440469339489937, + 0.001650982303544879, + 0.0018220209749415517, + 0.0016040520276874304, + 0.0022312086075544357, + 0.002448149025440216, + 0.01380087062716484, + 0.008650046773254871, + 0.02035396173596382, + 0.00782867707312107, + 0.008493009954690933, + 0.009031973779201508, + 0.008637947961688042, + 0.013781134970486164, + 0.002435940084978938, + 0.0015038895653560758, + 0.002228344324976206, + 0.0015073047252371907, + 0.004095441196113825, + 0.0026783605571836233, + 0.0023309653624892235, + 0.0021756391506642103, + 0.005256348289549351, + 0.005795045290142298, + 0.00585496611893177, + 0.006426598876714706, + 0.0032559537794440985, + 0.006853036116808653, + 0.006895299535244703, + 0.006314527243375778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.062043581157922745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021974719129502773 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021974719129502773, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021974719129502773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1397721916437149, + 0.04334287345409393, + 0.06062255799770355, + 0.04485521465539932, + 0.04287230595946312, + 0.43607911467552185, + 0.02557419054210186, + 0.07515338808298111, + 0.06602793186903, + 0.0657002329826355 + ], + "credit_map": [ + 0.01669110171496868, + 0.030811218544840813, + 0.015875672921538353, + 0.028270915150642395, + 0.014732361771166325, + 0.010837099514901638, + 0.013241845183074474, + 0.009311984293162823, + 0.0037997919134795666, + 0.003408892313018441, + 0.007737497333437204, + 0.014552591368556023, + 0.0028309649787843227, + 0.002814887324348092, + 0.00503009557723999, + 0.003168154042214155, + 0.008231745101511478, + 0.007739132735878229, + 0.0070260087959468365, + 0.006762357894331217, + 0.009110546670854092, + 0.008739311248064041, + 0.006651518866419792, + 0.006361937616020441, + 0.006217984948307276, + 0.0071968454867601395, + 0.005124193616211414, + 0.0045926934108138084, + 0.0040392340160906315, + 0.005458880215883255, + 0.005382949486374855, + 0.006842431146651506, + 0.005881869234144688, + 0.0052126627415418625, + 0.005867809522897005, + 0.004329119808971882, + 0.004277276806533337, + 0.0052519552409648895, + 0.004362327046692371, + 0.0076892864890396595, + 0.06010459363460541, + 0.017170850187540054, + 0.1264614760875702, + 0.024872614070773125, + 0.016525771468877792, + 0.026013432070612907, + 0.09305959194898605, + 0.07187075912952423, + 0.003878468181937933, + 0.003453496377915144, + 0.004325077403336763, + 0.002744918456301093, + 0.0031314806547015905, + 0.002640388673171401, + 0.002799251116812229, + 0.0026011085137724876, + 0.01599900797009468, + 0.01264841016381979, + 0.014856886118650436, + 0.0055302842520177364, + 0.005689830053597689, + 0.006231515668332577, + 0.006564599461853504, + 0.007632847875356674, + 0.009767943061888218, + 0.007878797128796577, + 0.008707284927368164, + 0.005155721213668585, + 0.010683652944862843, + 0.008021296001970768, + 0.0077468534000217915, + 0.008066384121775627, + 0.007913394831120968, + 0.008534444496035576, + 0.012005333788692951, + 0.008683237247169018, + 0.004602321423590183, + 0.010176816023886204, + 0.007864857092499733, + 0.0059198299422860146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0021974719129502773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a sleek, matte-black sports motorcycle speeding down a rain-soaked, neon-lit city street at midnight. A rider, clad in a tight black leather jersey and a full-face helmet, leans forward, pressed against the motorcycle's body. The dark visor reflects the shimmering light of the massive, multicolored billboards adorning the tall buildings. The damp asphalt acts like a black mirror, reflecting intensely saturated cyan, magenta, and blue hues. A thin layer of airborne vapor creates a hazy, layered light. As the rider decisively lowers the motorcycle around a sharp turn, the metal side of the motorcycle rubs violently against the rough pavement, sending a trail of dazzling orange-red sparks flying. These sparks are like short-lived flames, hissing through the night, not only illuminating the splashes of the high-speed rotating wheels, but also leaving a fiery trail on the slippery ground. The whole scene is full of speed and power, creating a cyberpunk, adrenaline-pumping atmosphere of tension.", + "seed": 17509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_5/index_3893_seed_17509_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.034351046895608306, + "vlm_logprob": -0.034351046895608306, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.034351046895608306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02423076517879963 + ], + "opposite_logprob_tokens": [ + -3.774230718612671 + ], + "logprob_reward": -0.02423076517879963, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02423076517879963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07025132328271866, + 0.23422472178936005, + 0.027758093550801277, + 0.04851723462343216, + 0.038877882063388824, + 0.07277961820363998, + 0.028260612860322, + 0.057020146399736404, + 0.06828543543815613, + 0.35402488708496094 + ], + "credit_map": [ + 0.01199184637516737, + 0.007500885985791683, + 0.00781773030757904, + 0.010683895088732243, + 0.008598657324910164, + 0.00816792156547308, + 0.008983699604868889, + 0.006506696343421936, + 0.023966066539287567, + 0.00889602955430746, + 0.06635809689760208, + 0.08392363786697388, + 0.007597395684570074, + 0.00671892985701561, + 0.02022198773920536, + 0.016542596742510796, + 0.005129717290401459, + 0.0023672098759561777, + 0.0031633193138986826, + 0.0038274708203971386, + 0.003465840592980385, + 0.0027189527172595263, + 0.003635817440226674, + 0.003449768526479602, + 0.0041458033956587315, + 0.005464881658554077, + 0.005082421936094761, + 0.007109909318387508, + 0.0063946908339858055, + 0.004978164564818144, + 0.006426359061151743, + 0.00891500897705555, + 0.005788648035377264, + 0.004080893937498331, + 0.007151723839342594, + 0.004155178088694811, + 0.003628604579716921, + 0.004891611635684967, + 0.0038554146885871887, + 0.0053258114494383335, + 0.006361559499055147, + 0.007468818686902523, + 0.018035273998975754, + 0.007330033928155899, + 0.006080173421651125, + 0.012145930901169777, + 0.00553850969299674, + 0.009819325990974903, + 0.0029736419674009085, + 0.003244235645979643, + 0.0033005611039698124, + 0.0029658519197255373, + 0.0035579833202064037, + 0.005273152608424425, + 0.0030313690658658743, + 0.003913820721209049, + 0.007375005166977644, + 0.012026423588395119, + 0.0072943065315485, + 0.007138303481042385, + 0.009727035649120808, + 0.004231809172779322, + 0.005385186057537794, + 0.0038420851342380047, + 0.007371380925178528, + 0.0047934153117239475, + 0.00878383219242096, + 0.006073107477277517, + 0.01429901085793972, + 0.010774233378469944, + 0.008279533125460148, + 0.00791092962026596, + 0.022589776664972305, + 0.02195262722671032, + 0.046847421675920486, + 0.026641439646482468, + 0.01874469965696335, + 0.09414901584386826, + 0.030130591243505478, + 0.0929693728685379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.02423076517879963 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908166736364365 + ], + "opposite_logprob_tokens": [ + -2.4640817642211914 + ], + "logprob_reward": -0.08908166736364365, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908166736364365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10853392630815506, + 0.06325910240411758, + 0.042079366743564606, + 0.07640951871871948, + 0.09089899063110352, + 0.11798769980669022, + 0.0833195298910141, + 0.028743932023644447, + 0.07445695251226425, + 0.3143109679222107 + ], + "credit_map": [ + 0.015158494934439659, + 0.01240003015846014, + 0.015068721026182175, + 0.014537982642650604, + 0.009105058386921883, + 0.012841574847698212, + 0.018505463376641273, + 0.010916600935161114, + 0.007021178491413593, + 0.003403390059247613, + 0.014804472215473652, + 0.0215054452419281, + 0.00408248882740736, + 0.0022871282417327166, + 0.005593735724687576, + 0.004561265464872122, + 0.00698568532243371, + 0.004773289430886507, + 0.005193219054490328, + 0.0062875146977603436, + 0.003651648061349988, + 0.005001521669328213, + 0.0053759231232106686, + 0.004810566082596779, + 0.008384219370782375, + 0.00884346105158329, + 0.007759867701679468, + 0.011070040985941887, + 0.009539267048239708, + 0.007165658287703991, + 0.009230914525687695, + 0.014416088350117207, + 0.009165706112980843, + 0.008689068257808685, + 0.018572699278593063, + 0.00704324571415782, + 0.005248623434454203, + 0.020085347816348076, + 0.00758106354624033, + 0.014513234607875347, + 0.010441824793815613, + 0.014027179218828678, + 0.01709207519888878, + 0.01377484854310751, + 0.012897152453660965, + 0.021896136924624443, + 0.013040146790444851, + 0.014818336814641953, + 0.008201172575354576, + 0.011645927093923092, + 0.01367341447621584, + 0.0120043084025383, + 0.006748343352228403, + 0.009615868330001831, + 0.0071894023567438126, + 0.014241092838346958, + 0.0043322560377418995, + 0.005535407457500696, + 0.0038502193056046963, + 0.0031402253080159426, + 0.0035483455285429955, + 0.0023745454382151365, + 0.003496590070426464, + 0.002466343343257904, + 0.008295486681163311, + 0.005529109388589859, + 0.011249073781073093, + 0.010250498540699482, + 0.008703428320586681, + 0.011228262446820736, + 0.008440425619482994, + 0.010760670527815819, + 0.020251581445336342, + 0.02022555284202099, + 0.03820401057600975, + 0.02330484800040722, + 0.017316458746790886, + 0.08586180210113525, + 0.027006391435861588, + 0.08214030414819717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.08908166736364365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255324762314558 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.01255324762314558, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255324762314558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04699741676449776, + 0.48953375220298767, + 0.03655708581209183, + 0.05595717579126358, + 0.033820416778326035, + 0.036685261875391006, + 0.028224462643265724, + 0.06530437618494034, + 0.09118176996707916, + 0.11573828756809235 + ], + "credit_map": [ + 0.006869876757264137, + 0.0049018338322639465, + 0.006757938303053379, + 0.005857565905898809, + 0.005922853481024504, + 0.00614122161641717, + 0.006125069223344326, + 0.004421056713908911, + 0.049150500446558, + 0.018250256776809692, + 0.14201846718788147, + 0.17836815118789673, + 0.010838588699698448, + 0.012359918095171452, + 0.04461805894970894, + 0.03392980992794037, + 0.008268686011433601, + 0.004180029965937138, + 0.004909702576696873, + 0.005390009842813015, + 0.002705165883526206, + 0.003296701470389962, + 0.004121475853025913, + 0.0036853135097771883, + 0.004832540638744831, + 0.006361223757266998, + 0.005684177856892347, + 0.008554832078516483, + 0.007489715237170458, + 0.0070309992879629135, + 0.007499294355511665, + 0.008504386991262436, + 0.0045367139391601086, + 0.0034471009857952595, + 0.007111730985343456, + 0.003676760010421276, + 0.0028481846675276756, + 0.0035050418227910995, + 0.003631538711488247, + 0.005063344724476337, + 0.003206690074875951, + 0.004491137340664864, + 0.008585038594901562, + 0.0037123083602637053, + 0.0031990318093448877, + 0.005694437772035599, + 0.0029860895592719316, + 0.0048105246387422085, + 0.002232382306829095, + 0.003011387540027499, + 0.004331695847213268, + 0.002386606065556407, + 0.003673193510621786, + 0.0059689925983548164, + 0.0031407931819558144, + 0.0034794106613844633, + 0.008038324303925037, + 0.01249400433152914, + 0.008215772919356823, + 0.0066432831808924675, + 0.013789008371531963, + 0.0046046581119298935, + 0.00747759360820055, + 0.004041722975671291, + 0.011681759729981422, + 0.007606129162013531, + 0.013099800795316696, + 0.010685480199754238, + 0.01296616904437542, + 0.013009384274482727, + 0.011644124053418636, + 0.010488915257155895, + 0.006312299519777298, + 0.007208025548607111, + 0.013106913305819035, + 0.009299278259277344, + 0.0074493419378995895, + 0.029734337702393532, + 0.0134147759526968, + 0.029213301837444305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.01255324762314558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043071962893009186 + ], + "opposite_logprob_tokens": [ + -3.168071985244751 + ], + "logprob_reward": -0.043071962893009186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043071962893009186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10001782327890396, + 0.17777161300182343, + 0.02984173409640789, + 0.03148612752556801, + 0.04927306994795799, + 0.04609375447034836, + 0.0672542080283165, + 0.03226388990879059, + 0.23520208895206451, + 0.2307957112789154 + ], + "credit_map": [ + 0.01248852163553238, + 0.006546489894390106, + 0.01749100722372532, + 0.0124203534796834, + 0.010534582659602165, + 0.014807897619903088, + 0.01600067690014839, + 0.009728288277983665, + 0.019769128412008286, + 0.006907688919454813, + 0.05060061067342758, + 0.057234060019254684, + 0.007561260834336281, + 0.00618758937343955, + 0.01667504943907261, + 0.012836235575377941, + 0.009804142639040947, + 0.0033320877701044083, + 0.005113882012665272, + 0.002601241460070014, + 0.0020014529582113028, + 0.002081620041280985, + 0.0019622896797955036, + 0.002945018233731389, + 0.003797226119786501, + 0.0044682882726192474, + 0.005119736306369305, + 0.0038936312776058912, + 0.002564008580520749, + 0.0037744403816759586, + 0.002971962559968233, + 0.004896834958344698, + 0.004553311970084906, + 0.004171376116573811, + 0.008318016305565834, + 0.00456669507548213, + 0.0033930365461856127, + 0.01197617407888174, + 0.0038048892747610807, + 0.008489571511745453, + 0.004208501894026995, + 0.00496114045381546, + 0.006778532639145851, + 0.006355804856866598, + 0.0049929325468838215, + 0.009130003862082958, + 0.004444088786840439, + 0.005222748499363661, + 0.008266367018222809, + 0.009076098911464214, + 0.009106609970331192, + 0.008275214582681656, + 0.00504026934504509, + 0.00908276904374361, + 0.00625618128105998, + 0.012150696478784084, + 0.0042249602265655994, + 0.004794815089553595, + 0.0059745898470282555, + 0.003344717901200056, + 0.004997915122658014, + 0.0025218736845999956, + 0.003346842946484685, + 0.003058173693716526, + 0.022161444649100304, + 0.01510176993906498, + 0.046695392578840256, + 0.025761038064956665, + 0.02660985477268696, + 0.033374205231666565, + 0.02495027333498001, + 0.040548112243413925, + 0.01432317215949297, + 0.014034183695912361, + 0.029981082305312157, + 0.014470865949988365, + 0.011485799215734005, + 0.06875849515199661, + 0.014917049556970596, + 0.062825046479702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.043071962893009186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, matte-black sports motorcycle visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a rain-soaked, neon-lit city street at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the rider wear a tight black leather jersey and a full-face helmet with a dark visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are orange-red sparks flying from the motorcycle rubbing against the pavement during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the motorcycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817591419443488 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.002817591419443488, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817591419443488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3100453317165375, + 0.13944201171398163, + 0.02873588539659977, + 0.030434375628829002, + 0.06019295006990433, + 0.05265744403004646, + 0.033050402998924255, + 0.025309721007943153, + 0.07848241925239563, + 0.24164943397045135 + ], + "credit_map": [ + 0.035035111010074615, + 0.023695148527622223, + 0.06791234016418457, + 0.04207751154899597, + 0.025441981852054596, + 0.041411787271499634, + 0.043055180460214615, + 0.03141627460718155, + 0.013544304296374321, + 0.00638031680136919, + 0.03671078383922577, + 0.05150436982512474, + 0.0035516005009412766, + 0.004045682959258556, + 0.01315144170075655, + 0.010553522035479546, + 0.0058431425131857395, + 0.0027277220506221056, + 0.003796530421823263, + 0.004069717600941658, + 0.002524896524846554, + 0.003141275141388178, + 0.003244660096243024, + 0.0033879419788718224, + 0.003229151712730527, + 0.0025083112996071577, + 0.0033833675552159548, + 0.0044775307178497314, + 0.004294942133128643, + 0.0033881620038300753, + 0.004085752181708813, + 0.0050671580247581005, + 0.005451642442494631, + 0.004954786505550146, + 0.009205118753015995, + 0.004685975611209869, + 0.003579372074455023, + 0.018889745697379112, + 0.0043631792068481445, + 0.009063131175935268, + 0.0044114976190030575, + 0.005456357263028622, + 0.01115095242857933, + 0.00609171949326992, + 0.004523996263742447, + 0.009526568464934826, + 0.004628845024853945, + 0.006867504213005304, + 0.0036924288142472506, + 0.004649891052395105, + 0.00539692398160696, + 0.00374017097055912, + 0.0035263437312096357, + 0.004353368189185858, + 0.0035737832076847553, + 0.0041174935176968575, + 0.0035373340360820293, + 0.0038758590817451477, + 0.005406952928751707, + 0.0027892650105059147, + 0.002197634894400835, + 0.0021782999392598867, + 0.0027716911863535643, + 0.0025526832323521376, + 0.007473244797438383, + 0.005447735544294119, + 0.01181725226342678, + 0.008252717554569244, + 0.008630364201962948, + 0.012095602229237556, + 0.008899503387510777, + 0.0158659927546978, + 0.01636621356010437, + 0.014819481410086155, + 0.03339427337050438, + 0.016322670504450798, + 0.013867235742509365, + 0.06664378196001053, + 0.018543878570199013, + 0.06169189140200615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.002817591419443488 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_tensor([5656, 5656])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_tensor([5656, 5656])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..52c201d8ad3545d3e5f871f3605be714b674bf1a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_tensor([5656, 5656])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins from an overhead angle, then gradually tilts upward and zooms out as the subject ascends, revealing a scene imbued with the riches of Eastern mythology. At the center of the frame, a majestic Chinese dragon, constructed entirely of clay, bears the subtle fingerprints and tool marks of hand-molding. Each scale appears individually molded and affixed, its edges slightly roughened, lending it a unique earthy texture. The dragon's head is majestic, its horns and flowing whiskers formed from golden clay, while its eyes, two smooth black glass beads, stand out against the matte clay skin. It soars slowly through the mist-shrouded mountains. Its movements are not fluid but rather rhythmic, with subtle pauses, each frame ascending, creating a unique frame-by-frame animation effect. Its serpentine body rises and falls rhythmically, its four claws making subtle strokes in the air. The backdrop is a rolling, inky-green mountain range, its peaks enveloped and obscured by a thick, cotton-like cloud, also sculpted from white clay. A few ancient, sturdy pines dot the landscape, evoking the texture of clay sculptures. Soft, diffused light evenly illuminates the scene, highlighting the matte texture of the clay, devoid of harsh reflections. The overall palette is simple and elegant, creating an artistic atmosphere of tranquility, solemnity, and the warmth of craftsmanship.", + "seed": 17853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_5656_seed_17853_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03443221924826503, + "vlm_logprob": -0.03443221924826503, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03443221924826503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08412861824035645 + ], + "opposite_logprob_tokens": [ + -2.5841286182403564 + ], + "logprob_reward": -0.08412861824035645, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08412861824035645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1879260092973709, + 0.07154183834791183, + 0.03485482931137085, + 0.034233104437589645, + 0.0616881437599659, + 0.08407782763242722, + 0.03486225754022598, + 0.36725252866744995, + 0.07400960475206375, + 0.049553874880075455 + ], + "credit_map": [ + 0.048367690294981, + 0.01575714536011219, + 0.04899480193853378, + 0.01111957710236311, + 0.018991196528077126, + 0.011810519732534885, + 0.017997197806835175, + 0.014887874014675617, + 0.020766036584973335, + 0.004500776994973421, + 0.015617692843079567, + 0.005733459256589413, + 0.00611447636038065, + 0.005956947337836027, + 0.006645804736763239, + 0.006206647492945194, + 0.007175181061029434, + 0.00337474443949759, + 0.006105728913098574, + 0.002722892677411437, + 0.004796603228896856, + 0.003582604927942157, + 0.0037270148750394583, + 0.003370058024302125, + 0.005424736998975277, + 0.0026819442864507437, + 0.005309181287884712, + 0.0037952889688313007, + 0.0028485783841460943, + 0.005883260164409876, + 0.00452085118740797, + 0.0037692624609917402, + 0.010471036657691002, + 0.007468936499208212, + 0.012896674685180187, + 0.005978972185403109, + 0.005215808283537626, + 0.005847389809787273, + 0.007576873525977135, + 0.0062324535101652145, + 0.01216536108404398, + 0.007792649790644646, + 0.014154922217130661, + 0.0094459755346179, + 0.007894866168498993, + 0.009748794138431549, + 0.012912617065012455, + 0.009962641634047031, + 0.007358701433986425, + 0.002484692493453622, + 0.003920228220522404, + 0.004494166001677513, + 0.0030685828533023596, + 0.0048366631381213665, + 0.003990170080214739, + 0.004709053318947554, + 0.02087821252644062, + 0.010901234112679958, + 0.13165000081062317, + 0.018315158784389496, + 0.008720282465219498, + 0.09552296996116638, + 0.010042286477982998, + 0.07122237235307693, + 0.010519014671444893, + 0.008558269590139389, + 0.012552112340927124, + 0.010097037069499493, + 0.006366502959281206, + 0.008201475255191326, + 0.008781480602920055, + 0.00893370434641838, + 0.009439962916076183, + 0.00774411903694272, + 0.007274704519659281, + 0.004715531598776579, + 0.002997785806655884, + 0.0067131491377949715, + 0.004074813798069954, + 0.006593808066099882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.08412861824035645 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004664611537009478 + ], + "opposite_logprob_tokens": [ + -5.379664421081543 + ], + "logprob_reward": -0.004664611537009478, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004664611537009478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10884901136159897, + 0.19297266006469727, + 0.05973847210407257, + 0.04090746119618416, + 0.05673922225832939, + 0.04223232343792915, + 0.05031788721680641, + 0.34002959728240967, + 0.050966501235961914, + 0.05724688991904259 + ], + "credit_map": [ + 0.02786395326256752, + 0.010450457222759724, + 0.02648850902915001, + 0.0067649660632014275, + 0.010027173906564713, + 0.008727549575269222, + 0.011089912615716457, + 0.00743648549541831, + 0.06240706518292427, + 0.008741538040339947, + 0.059108566492795944, + 0.014599988237023354, + 0.00887836329638958, + 0.011842053383588791, + 0.010123936459422112, + 0.017271149903535843, + 0.008074041455984116, + 0.005827341228723526, + 0.008814576081931591, + 0.006336734630167484, + 0.007132888305932283, + 0.006890437565743923, + 0.007424999959766865, + 0.009237448684871197, + 0.010279737412929535, + 0.002931758062914014, + 0.00782491359859705, + 0.0034650424495339394, + 0.0037065318319946527, + 0.004740005359053612, + 0.003648146288469434, + 0.004311321768909693, + 0.010938267223536968, + 0.008189529180526733, + 0.009708539582788944, + 0.005587435327470303, + 0.00475664809346199, + 0.005792357958853245, + 0.006647488567978144, + 0.005118954926729202, + 0.0044458904303610325, + 0.004621716681867838, + 0.00648544542491436, + 0.005104771815240383, + 0.003562274621799588, + 0.004643191117793322, + 0.008038251660764217, + 0.005330774001777172, + 0.010334976948797703, + 0.0038542754482477903, + 0.006174386478960514, + 0.006330868694931269, + 0.004489625804126263, + 0.006930478382855654, + 0.0055708857253193855, + 0.006632388569414616, + 0.018598441034555435, + 0.009336280636489391, + 0.12316489964723587, + 0.019082313403487206, + 0.008664599619805813, + 0.09077931195497513, + 0.009436260908842087, + 0.06096746772527695, + 0.0065148803405463696, + 0.00801511388272047, + 0.006091548129916191, + 0.007618709001690149, + 0.004473733250051737, + 0.0061238789930939674, + 0.007070364896208048, + 0.005058271810412407, + 0.009028158150613308, + 0.005892412271350622, + 0.00812991987913847, + 0.0062436372973024845, + 0.00593189150094986, + 0.00740335276350379, + 0.007084425073117018, + 0.007533090654760599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.004664611537009478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038129545748233795 + ], + "opposite_logprob_tokens": [ + -3.2881295680999756 + ], + "logprob_reward": -0.038129545748233795, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038129545748233795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07959231734275818, + 0.375740110874176, + 0.03695671632885933, + 0.08206753432750702, + 0.031507458537817, + 0.044062599539756775, + 0.03299086168408394, + 0.22136320173740387, + 0.04384912922978401, + 0.051870085299015045 + ], + "credit_map": [ + 0.017792467027902603, + 0.008183539845049381, + 0.013204352930188179, + 0.007149731274694204, + 0.008714282885193825, + 0.00788433663547039, + 0.007978657260537148, + 0.008684946224093437, + 0.13322342932224274, + 0.0150247011333704, + 0.13002294301986694, + 0.021604342386126518, + 0.019020535051822662, + 0.017705151811242104, + 0.017234614118933678, + 0.02190438099205494, + 0.006554540246725082, + 0.004024007823318243, + 0.005962136201560497, + 0.003743999172002077, + 0.0038789124228060246, + 0.004939929582178593, + 0.004490265157073736, + 0.0033629240933805704, + 0.014524804428219795, + 0.007095293141901493, + 0.013628673739731312, + 0.006787709891796112, + 0.006568362005054951, + 0.01956861838698387, + 0.007860910147428513, + 0.006033163517713547, + 0.008104025386273861, + 0.003803401719778776, + 0.005547100678086281, + 0.0028035377617925406, + 0.0019313122611492872, + 0.003121373476460576, + 0.003213519463315606, + 0.002983188722282648, + 0.0060225301422178745, + 0.0033129097428172827, + 0.007874425500631332, + 0.005867920350283384, + 0.003763888729736209, + 0.004520363640040159, + 0.0073937601409852505, + 0.005306802224367857, + 0.00537376431748271, + 0.0028833174146711826, + 0.004083007574081421, + 0.0051817456260323524, + 0.002705465769395232, + 0.003746303031221032, + 0.004036228638142347, + 0.004981030244380236, + 0.012202542275190353, + 0.007243945728987455, + 0.07637359946966171, + 0.014803637750446796, + 0.007343742530792952, + 0.05635996535420418, + 0.009736988693475723, + 0.0372987799346447, + 0.005502261687070131, + 0.005465369671583176, + 0.005449909251183271, + 0.006766290869563818, + 0.0032407427206635475, + 0.0050038620829582214, + 0.005916003603488207, + 0.006504688877612352, + 0.009042050689458847, + 0.006624213885515928, + 0.01096657756716013, + 0.004479211755096912, + 0.003512119874358177, + 0.0069967289455235004, + 0.004055920522660017, + 0.0061932639218866825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.038129545748233795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196044661104679 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002196044661104679, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196044661104679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2433197945356369, + 0.09178824722766876, + 0.025514986366033554, + 0.08321846276521683, + 0.06207584962248802, + 0.044352758675813675, + 0.10666998475790024, + 0.14100004732608795, + 0.04977088421583176, + 0.15228907763957977 + ], + "credit_map": [ + 0.06729776412248611, + 0.02157445251941681, + 0.07346563041210175, + 0.015617432072758675, + 0.019738025963306427, + 0.010816570371389389, + 0.019858818501234055, + 0.014951095916330814, + 0.02876541018486023, + 0.004850433673709631, + 0.027823707088828087, + 0.005651828367263079, + 0.005476110614836216, + 0.006077133119106293, + 0.005544672254472971, + 0.007598953787237406, + 0.0036899573169648647, + 0.002644505351781845, + 0.003640404436737299, + 0.0023062771651893854, + 0.003642122494056821, + 0.003760740626603365, + 0.0035520887468010187, + 0.0022788895294070244, + 0.01298533659428358, + 0.005607155151665211, + 0.013360747136175632, + 0.0070505933836102486, + 0.006073173601180315, + 0.021597320213913918, + 0.009882509708404541, + 0.006661629304289818, + 0.009240929037332535, + 0.00580549705773592, + 0.00930069386959076, + 0.006439147517085075, + 0.005836178082972765, + 0.009885234758257866, + 0.008190217427909374, + 0.007377949543297291, + 0.007128079421818256, + 0.004751968197524548, + 0.006708608940243721, + 0.004347437992691994, + 0.003694225801154971, + 0.005664679687470198, + 0.006591319106519222, + 0.005466438829898834, + 0.022981815040111542, + 0.009976888075470924, + 0.013877487741410732, + 0.01119846012443304, + 0.006657761055976152, + 0.015037464909255505, + 0.0069708251394331455, + 0.019969284534454346, + 0.007712082006037235, + 0.0046752928756177425, + 0.049838319420814514, + 0.007904063910245895, + 0.003689468139782548, + 0.0366082601249218, + 0.004591380711644888, + 0.025981171056628227, + 0.004892835393548012, + 0.0072449627332389355, + 0.005296691320836544, + 0.008329059928655624, + 0.005273111164569855, + 0.006006828974932432, + 0.007107443641871214, + 0.005619949195533991, + 0.029718119651079178, + 0.02338522858917713, + 0.020336296409368515, + 0.015113044530153275, + 0.008318670094013214, + 0.02055654674768448, + 0.01149674504995346, + 0.02336442843079567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.002196044661104679 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304227605462074 + ], + "opposite_logprob_tokens": [ + -3.1680421829223633 + ], + "logprob_reward": -0.04304227605462074, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304227605462074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07979445904493332, + 0.10139143466949463, + 0.06423870474100113, + 0.042934391647577286, + 0.04158323258161545, + 0.03978969529271126, + 0.06340668350458145, + 0.44651588797569275, + 0.06127341836690903, + 0.05907212570309639 + ], + "credit_map": [ + 0.019021177664399147, + 0.0069559537805616856, + 0.02030729316174984, + 0.006869048345834017, + 0.006645427085459232, + 0.004525059834122658, + 0.009589157067239285, + 0.005881340242922306, + 0.038201265037059784, + 0.003692759433761239, + 0.036550238728523254, + 0.005008404143154621, + 0.004282544832676649, + 0.004772807005792856, + 0.00377889652736485, + 0.005104511976242065, + 0.00958876684308052, + 0.006010629236698151, + 0.008989877067506313, + 0.004803815856575966, + 0.008185334503650665, + 0.01386105827987194, + 0.006290782708674669, + 0.0065084416419267654, + 0.00811251811683178, + 0.0030955267138779163, + 0.00808288250118494, + 0.003680561203509569, + 0.004611673299223185, + 0.00718105211853981, + 0.00453902455046773, + 0.0036311529111117125, + 0.00851026177406311, + 0.0037016437854617834, + 0.005607558414340019, + 0.004750879947096109, + 0.004505663178861141, + 0.004401891492307186, + 0.00601887796074152, + 0.004086455795913935, + 0.005090815015137196, + 0.0045433687046170235, + 0.0077686551958322525, + 0.004684321582317352, + 0.003474925644695759, + 0.00549194635823369, + 0.004414888098835945, + 0.004320776555687189, + 0.010261693969368935, + 0.00516004441305995, + 0.007556798402220011, + 0.007402437273412943, + 0.00540503952652216, + 0.00975846592336893, + 0.00826998334378004, + 0.00959222111850977, + 0.024554939940571785, + 0.011654636822640896, + 0.16284194588661194, + 0.024332240223884583, + 0.011230500414967537, + 0.11849921196699142, + 0.012140141800045967, + 0.08126227557659149, + 0.006774642039090395, + 0.008857103995978832, + 0.010540427640080452, + 0.008095324970781803, + 0.004732480272650719, + 0.006763305980712175, + 0.007864083163440228, + 0.007646048907190561, + 0.008250445127487183, + 0.005908884573727846, + 0.009914073161780834, + 0.007058234419673681, + 0.0056377314031124115, + 0.007266474887728691, + 0.007513710763305426, + 0.007522569037973881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.04304227605462074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from an overhead angle, then gradually tilts upward and zooms out as the subject ascends, revealing a scene imbued with the riches of Eastern mythology. At the center of the frame, a majestic Chinese dragon, constructed entirely of clay, bears the subtle fingerprints and tool marks of hand-molding. Each scale appears individually molded and affixed, its edges slightly roughened, lending it a unique earthy texture. The dragon's head is majestic, its horns and flowing whiskers formed from golden clay, while its eyes, two smooth black glass beads, stand out against the matte clay skin. It soars slowly through the mist-shrouded mountains. Its movements are not fluid but rather rhythmic, with subtle pauses, each frame ascending, creating a unique frame-by-frame animation effect. Its serpentine body rises and falls rhythmically, its four claws making subtle strokes in the air. The backdrop is a rolling, inky-green mountain range, its peaks enveloped and obscured by a thick, cotton-like cloud, also sculpted from white clay. A few ancient, sturdy pines dot the landscape, evoking the texture of clay sculptures. Soft, diffused light evenly illuminates the scene, highlighting the matte texture of the clay, devoid of harsh reflections. The overall palette is simple and elegant, creating an artistic atmosphere of tranquility, solemnity, and the warmth of craftsmanship.", + "seed": 17854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_5656_seed_17854_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04988030651584267, + "vlm_logprob": -0.04988030651584267, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.04988030651584267, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08731929957866669 + ], + "opposite_logprob_tokens": [ + -2.5873193740844727 + ], + "logprob_reward": -0.08731929957866669, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08731929957866669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1577971875667572, + 0.1608976423740387, + 0.025788437575101852, + 0.04298219829797745, + 0.06172994524240494, + 0.15268124639987946, + 0.12839661538600922, + 0.03187025710940361, + 0.13257957994937897, + 0.1052769124507904 + ], + "credit_map": [ + 0.02113720402121544, + 0.01755242794752121, + 0.019937338307499886, + 0.02219635806977749, + 0.013271375559270382, + 0.017057286575436592, + 0.025686418637633324, + 0.02095877379179001, + 0.01553089078515768, + 0.028960922732949257, + 0.015316366218030453, + 0.03391800448298454, + 0.010659951716661453, + 0.011275838129222393, + 0.01913866400718689, + 0.026096992194652557, + 0.004680568818002939, + 0.00290355971083045, + 0.0032130705658346415, + 0.003650544909760356, + 0.0017878743819892406, + 0.003027161583304405, + 0.002611069241538644, + 0.003914590459316969, + 0.005058135837316513, + 0.0049602435901761055, + 0.004141110926866531, + 0.007742499001324177, + 0.004143475089222193, + 0.0060724192298948765, + 0.005258942488580942, + 0.005605374462902546, + 0.006863733287900686, + 0.005831678863614798, + 0.010568992234766483, + 0.0069202096201479435, + 0.004545886069536209, + 0.010850139893591404, + 0.005182049702852964, + 0.01096725556999445, + 0.017830347642302513, + 0.01458838488906622, + 0.019797777757048607, + 0.012141665443778038, + 0.019835302606225014, + 0.029087357223033905, + 0.012658883817493916, + 0.02674153633415699, + 0.011930272914469242, + 0.013263680972158909, + 0.021407868713140488, + 0.017496025189757347, + 0.01166708767414093, + 0.02569231577217579, + 0.012093636207282543, + 0.014845723286271095, + 0.0032290343660861254, + 0.0037700203247368336, + 0.00842281710356474, + 0.0036032856442034245, + 0.0033274898305535316, + 0.002980373566970229, + 0.0030944838654249907, + 0.003442753804847598, + 0.03012634627521038, + 0.009534268639981747, + 0.050069548189640045, + 0.007724533323198557, + 0.006248163990676403, + 0.015803713351488113, + 0.006963237654417753, + 0.006109763868153095, + 0.017610326409339905, + 0.01562827080488205, + 0.017531007528305054, + 0.009779630228877068, + 0.011477534659206867, + 0.015171403996646404, + 0.009544303640723228, + 0.008534431457519531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.08731929957866669 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004129929933696985 + ], + "opposite_logprob_tokens": [ + -5.504129886627197 + ], + "logprob_reward": -0.004129929933696985, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004129929933696985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12357736378908157, + 0.08042213320732117, + 0.04688645526766777, + 0.045789461582899094, + 0.06342707574367523, + 0.14493922889232635, + 0.047895681113004684, + 0.04255349189043045, + 0.30624738335609436, + 0.09826172888278961 + ], + "credit_map": [ + 0.013261446729302406, + 0.014120780862867832, + 0.014083693735301495, + 0.019458448514342308, + 0.010292311199009418, + 0.011609056033194065, + 0.02350260131061077, + 0.01724901981651783, + 0.006931978743523359, + 0.013907373882830143, + 0.007037506438791752, + 0.018604610115289688, + 0.00607956200838089, + 0.008417762815952301, + 0.010014750063419342, + 0.00942858587950468, + 0.005383695010095835, + 0.00579856289550662, + 0.006114217918366194, + 0.007095881272107363, + 0.0038402637001127005, + 0.005751499906182289, + 0.0061256797052919865, + 0.006776655558496714, + 0.006337932776659727, + 0.0049158260226249695, + 0.00496482802554965, + 0.007465267553925514, + 0.0056279427371919155, + 0.006222815252840519, + 0.004920612089335918, + 0.005334238056093454, + 0.007070457097142935, + 0.005160444416105747, + 0.011828232556581497, + 0.0041978005319833755, + 0.0038169478066265583, + 0.014034024439752102, + 0.004236311186105013, + 0.013082853518426418, + 0.012681704014539719, + 0.012215596623718739, + 0.020493164658546448, + 0.010951191186904907, + 0.01609286479651928, + 0.043476030230522156, + 0.011282289400696754, + 0.01774638332426548, + 0.004977497272193432, + 0.0041972436010837555, + 0.008998577482998371, + 0.005617177113890648, + 0.0036604406777769327, + 0.007069649174809456, + 0.005359613802284002, + 0.008015479892492294, + 0.005006211809813976, + 0.004119995050132275, + 0.01300759520381689, + 0.0045888787135481834, + 0.0032514638733118773, + 0.004601465072482824, + 0.0033988833893090487, + 0.0045789978466928005, + 0.08696652948856354, + 0.01144517119973898, + 0.13186240196228027, + 0.011898776516318321, + 0.008369333110749722, + 0.033609453588724136, + 0.011455035768449306, + 0.010640697553753853, + 0.009970488026738167, + 0.01586225815117359, + 0.021211255341768265, + 0.00906516332179308, + 0.01045902818441391, + 0.016827702522277832, + 0.007696190848946571, + 0.007169646676629782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.004129929933696985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331264689564705 + ], + "opposite_logprob_tokens": [ + -3.773312568664551 + ], + "logprob_reward": -0.02331264689564705, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331264689564705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057721443474292755, + 0.07584143429994583, + 0.04662488400936127, + 0.03843259811401367, + 0.03951901197433472, + 0.09760235995054245, + 0.12156930565834045, + 0.04201717674732208, + 0.3986184895038605, + 0.08205325901508331 + ], + "credit_map": [ + 0.007303665857762098, + 0.006105078849941492, + 0.008637594990432262, + 0.007142374757677317, + 0.004876808729022741, + 0.007215152960270643, + 0.007496955804526806, + 0.00894380733370781, + 0.0072878217324614525, + 0.013792142271995544, + 0.007741055451333523, + 0.015363839454948902, + 0.005529444199055433, + 0.007231061812490225, + 0.009048698469996452, + 0.009847376495599747, + 0.009181804023683071, + 0.0038965202402323484, + 0.0046018147841095924, + 0.004670472349971533, + 0.007466593757271767, + 0.005173565354198217, + 0.0032842345535755157, + 0.008349879644811153, + 0.005995128303766251, + 0.004428043961524963, + 0.005225031636655331, + 0.0038297332357615232, + 0.005159430205821991, + 0.004061471670866013, + 0.0043363627046346664, + 0.005397396627813578, + 0.0030552891548722982, + 0.005834641400724649, + 0.005986648146063089, + 0.0029955862555652857, + 0.005081749055534601, + 0.006897824816405773, + 0.0042350622825324535, + 0.005432212259620428, + 0.009084113873541355, + 0.012541231699287891, + 0.01765117421746254, + 0.008655580691993237, + 0.00935499370098114, + 0.022150959819555283, + 0.007191967684775591, + 0.010972336865961552, + 0.010100604966282845, + 0.00677759712561965, + 0.02769959531724453, + 0.011509601958096027, + 0.007393151056021452, + 0.030303506180644035, + 0.008867578580975533, + 0.018917663022875786, + 0.004092758521437645, + 0.004399504512548447, + 0.006746228318661451, + 0.003991318866610527, + 0.004476986825466156, + 0.005178782623261213, + 0.004207375925034285, + 0.008924221619963646, + 0.1129787340760231, + 0.01715874671936035, + 0.1690666228532791, + 0.01752944104373455, + 0.009888489730656147, + 0.043498389422893524, + 0.015146265737712383, + 0.013351831585168839, + 0.010354279540479183, + 0.013425500132143497, + 0.01716052182018757, + 0.007144805043935776, + 0.008114713244140148, + 0.011135762557387352, + 0.006794597487896681, + 0.007923080585896969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.02331264689564705 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624572142958641 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624572142958641, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624572142958641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322096735239029, + 0.043255385011434555, + 0.050618644803762436, + 0.035319000482559204, + 0.09575987607240677, + 0.06014345958828926, + 0.07982388883829117, + 0.03551460802555084, + 0.3539852201938629, + 0.11337029188871384 + ], + "credit_map": [ + 0.017583368346095085, + 0.012516291812062263, + 0.01874307543039322, + 0.015781564638018608, + 0.008767379447817802, + 0.019941214472055435, + 0.016601024195551872, + 0.022275762632489204, + 0.0046763657592237, + 0.004513029009103775, + 0.005423251539468765, + 0.005718356464058161, + 0.003580030519515276, + 0.009367333725094795, + 0.003909818362444639, + 0.006067198701202869, + 0.008784660138189793, + 0.004091726150363684, + 0.005086044780910015, + 0.004653938114643097, + 0.008205090649425983, + 0.006711150519549847, + 0.005733320955187082, + 0.007352715823799372, + 0.005192420445382595, + 0.003330220002681017, + 0.005195198580622673, + 0.003302582073956728, + 0.003636474721133709, + 0.00503545580431819, + 0.003984121605753899, + 0.005642528645694256, + 0.012383082881569862, + 0.008139975368976593, + 0.016003623604774475, + 0.008245747536420822, + 0.008314142003655434, + 0.01545945554971695, + 0.010312088765203953, + 0.016901759430766106, + 0.006305697839707136, + 0.005675424821674824, + 0.007607420906424522, + 0.00498161930590868, + 0.006499429699033499, + 0.014550278894603252, + 0.0050053950399160385, + 0.009518190287053585, + 0.007640262134373188, + 0.008707521483302116, + 0.010530390776693821, + 0.0073537882417440414, + 0.008401730097830296, + 0.01618976518511772, + 0.00869046151638031, + 0.012309974059462547, + 0.004520025569945574, + 0.005218620877712965, + 0.005855113733559847, + 0.0032844373490661383, + 0.0035725198686122894, + 0.004092987626791, + 0.003227313281968236, + 0.005743588786572218, + 0.09702096879482269, + 0.012997271493077278, + 0.14899933338165283, + 0.014392688870429993, + 0.010897721163928509, + 0.039675261825323105, + 0.014501935802400112, + 0.015500045381486416, + 0.013395535759627819, + 0.01453388761729002, + 0.017313601449131966, + 0.008765588514506817, + 0.012557068839669228, + 0.018949495628476143, + 0.011709877289831638, + 0.01614522933959961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.007624572142958641 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701508402824402 + ], + "opposite_logprob_tokens": [ + -2.1270151138305664 + ], + "logprob_reward": -0.12701508402824402, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701508402824402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07809370011091232, + 0.08858814090490341, + 0.07148172706365585, + 0.02554214932024479, + 0.04457724094390869, + 0.14900454878807068, + 0.0969042032957077, + 0.040713921189308167, + 0.37108585238456726, + 0.034008514136075974 + ], + "credit_map": [ + 0.008117669261991978, + 0.009881684556603432, + 0.008723626844584942, + 0.012896982952952385, + 0.006391485687345266, + 0.007861603982746601, + 0.014803406782448292, + 0.009417243301868439, + 0.00885507371276617, + 0.01640590839087963, + 0.008468040265142918, + 0.01697765290737152, + 0.005477845668792725, + 0.006668878719210625, + 0.013211287558078766, + 0.012523452751338482, + 0.019336514174938202, + 0.0062127625569701195, + 0.006997046992182732, + 0.007048997096717358, + 0.00485446210950613, + 0.00503029627725482, + 0.006653439719229937, + 0.015348211862146854, + 0.0030233124271035194, + 0.0026393933221697807, + 0.0037287792656570673, + 0.002824899973347783, + 0.0028377908747643232, + 0.0027246237732470036, + 0.003391451667994261, + 0.00437189731746912, + 0.0038047160487622023, + 0.003758518723770976, + 0.008235633373260498, + 0.004411613568663597, + 0.00408103596419096, + 0.006058007013052702, + 0.004764118697494268, + 0.00946359895169735, + 0.011005107313394547, + 0.014360599219799042, + 0.023455286398530006, + 0.008763125166296959, + 0.014519904740154743, + 0.047301825135946274, + 0.012647578492760658, + 0.016951119527220726, + 0.011433384381234646, + 0.013841213658452034, + 0.013369770720601082, + 0.013891594484448433, + 0.013108842074871063, + 0.010486460290849209, + 0.012081470340490341, + 0.008691470138728619, + 0.004139958880841732, + 0.003625415964052081, + 0.011726722121238708, + 0.004912710748612881, + 0.0034237520303577185, + 0.0054502934217453, + 0.003535578027367592, + 0.0038994918577373028, + 0.08597736805677414, + 0.027500439435243607, + 0.14138561487197876, + 0.019767804071307182, + 0.017858460545539856, + 0.04389640688896179, + 0.018951982259750366, + 0.015747781842947006, + 0.0038410888519138098, + 0.0056465426459908485, + 0.0071363309398293495, + 0.0024678159970790148, + 0.0030029648914933205, + 0.006442487705498934, + 0.002509992104023695, + 0.0029612891376018524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.12701508402824402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from an overhead angle, then gradually tilts upward and zooms out as the subject ascends, revealing a scene imbued with the riches of Eastern mythology. At the center of the frame, a majestic Chinese dragon, constructed entirely of clay, bears the subtle fingerprints and tool marks of hand-molding. Each scale appears individually molded and affixed, its edges slightly roughened, lending it a unique earthy texture. The dragon's head is majestic, its horns and flowing whiskers formed from golden clay, while its eyes, two smooth black glass beads, stand out against the matte clay skin. It soars slowly through the mist-shrouded mountains. Its movements are not fluid but rather rhythmic, with subtle pauses, each frame ascending, creating a unique frame-by-frame animation effect. Its serpentine body rises and falls rhythmically, its four claws making subtle strokes in the air. The backdrop is a rolling, inky-green mountain range, its peaks enveloped and obscured by a thick, cotton-like cloud, also sculpted from white clay. A few ancient, sturdy pines dot the landscape, evoking the texture of clay sculptures. Soft, diffused light evenly illuminates the scene, highlighting the matte texture of the clay, devoid of harsh reflections. The overall palette is simple and elegant, creating an artistic atmosphere of tranquility, solemnity, and the warmth of craftsmanship.", + "seed": 17855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_5656_seed_17855_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03603299991227686, + "vlm_logprob": -0.03603299991227686, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03603299991227686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09740511327981949 + ], + "opposite_logprob_tokens": [ + -2.472405195236206 + ], + "logprob_reward": -0.09740511327981949, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09740511327981949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13152238726615906, + 0.27225881814956665, + 0.06591102480888367, + 0.02409491501748562, + 0.05466293916106224, + 0.16564548015594482, + 0.04839591681957245, + 0.03203281760215759, + 0.1012180894613266, + 0.10425756126642227 + ], + "credit_map": [ + 0.019970616325736046, + 0.01302951667457819, + 0.01785394921898842, + 0.01818721555173397, + 0.013650549575686455, + 0.022986218333244324, + 0.011755716055631638, + 0.01408860832452774, + 0.04370836913585663, + 0.01935080625116825, + 0.03627666085958481, + 0.022761031985282898, + 0.026601741090416908, + 0.03982942923903465, + 0.026314599439501762, + 0.05741621181368828, + 0.008880837820470333, + 0.006485728081315756, + 0.00763959065079689, + 0.008201544173061848, + 0.006492066662758589, + 0.009700557217001915, + 0.011001955717802048, + 0.007508750539273024, + 0.003091102931648493, + 0.002515167463570833, + 0.003929153084754944, + 0.002517721150070429, + 0.0027992352843284607, + 0.003709841752424836, + 0.0031271055340766907, + 0.002405589446425438, + 0.005280871409922838, + 0.00801642145961523, + 0.006341783795505762, + 0.007988370954990387, + 0.00722949905321002, + 0.006116336211562157, + 0.008113139308989048, + 0.005576527211815119, + 0.029040217399597168, + 0.012005774304270744, + 0.0273244958370924, + 0.012427554465830326, + 0.008927302435040474, + 0.031622111797332764, + 0.020359905436635017, + 0.023938138037919998, + 0.0070348395965993404, + 0.005688227713108063, + 0.005393002647906542, + 0.003579395590350032, + 0.00441515538841486, + 0.011012603528797626, + 0.003969418350607157, + 0.007303277961909771, + 0.003892898093909025, + 0.0051347180269658566, + 0.003526975866407156, + 0.0036532641388475895, + 0.004945097025483847, + 0.003961831331253052, + 0.0027452418580651283, + 0.004172796383500099, + 0.028254898265004158, + 0.009370218962430954, + 0.033314503729343414, + 0.005956660490483046, + 0.005896936170756817, + 0.006891773547977209, + 0.004895533435046673, + 0.006637571845203638, + 0.018300406634807587, + 0.006939630024135113, + 0.03899935632944107, + 0.010045572184026241, + 0.006905927788466215, + 0.0061067757196724415, + 0.009213192388415337, + 0.007746710442006588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.09740511327981949 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005990052595734596 + ], + "opposite_logprob_tokens": [ + -5.130990028381348 + ], + "logprob_reward": -0.005990052595734596, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005990052595734596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24944345653057098, + 0.08062571287155151, + 0.06678661704063416, + 0.0426226481795311, + 0.07575790584087372, + 0.05344530940055847, + 0.0569126270711422, + 0.10035406053066254, + 0.10776011645793915, + 0.16629156470298767 + ], + "credit_map": [ + 0.039059653878211975, + 0.021936068311333656, + 0.043881140649318695, + 0.028407493606209755, + 0.03023393265902996, + 0.026782264932990074, + 0.02656196802854538, + 0.032580938190221786, + 0.010081997141242027, + 0.008121170103549957, + 0.012610417790710926, + 0.010266343131661415, + 0.0071054077707231045, + 0.011910000815987587, + 0.008872313424944878, + 0.011658066883683205, + 0.009491403587162495, + 0.004852317739278078, + 0.00837680697441101, + 0.007721053436398506, + 0.005122978240251541, + 0.01207754760980606, + 0.009713783860206604, + 0.009430722333490849, + 0.006262450013309717, + 0.0031349011696875095, + 0.008145267143845558, + 0.00506529351696372, + 0.0025538275949656963, + 0.008517412468791008, + 0.0053763664327561855, + 0.0035671317018568516, + 0.007333631627261639, + 0.009921682998538017, + 0.010134597308933735, + 0.011383835226297379, + 0.006516534835100174, + 0.009625200182199478, + 0.011920532211661339, + 0.008921892382204533, + 0.006043938919901848, + 0.0052935220301151276, + 0.011213370598852634, + 0.00444505549967289, + 0.0025928241666406393, + 0.006593857891857624, + 0.005379620939493179, + 0.011883116327226162, + 0.00837793294340372, + 0.010428308509290218, + 0.006451246328651905, + 0.007299500517547131, + 0.006714168004691601, + 0.005222274921834469, + 0.006040093954652548, + 0.006379102822393179, + 0.01671665720641613, + 0.014187057502567768, + 0.010839051567018032, + 0.012204388156533241, + 0.012883136980235577, + 0.011394215747714043, + 0.008011963218450546, + 0.014117586426436901, + 0.024026865139603615, + 0.011158233508467674, + 0.02874855510890484, + 0.010054984129965305, + 0.008386321365833282, + 0.009360562078654766, + 0.00748442392796278, + 0.008540171198546886, + 0.03299911320209503, + 0.012106341309845448, + 0.07384473085403442, + 0.01096972543746233, + 0.006389733403921127, + 0.010655375197529793, + 0.008727574720978737, + 0.010598969645798206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.005990052595734596 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018219729885458946 + ], + "opposite_logprob_tokens": [ + -4.018219947814941 + ], + "logprob_reward": -0.018219729885458946, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018219729885458946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2427506297826767, + 0.21643505990505219, + 0.11719891428947449, + 0.04430951178073883, + 0.045540399849414825, + 0.07415669411420822, + 0.028076468035578728, + 0.05483226478099823, + 0.07077790796756744, + 0.10592204332351685 + ], + "credit_map": [ + 0.04215717315673828, + 0.022714003920555115, + 0.035108502954244614, + 0.028840944170951843, + 0.03008892387151718, + 0.028094615787267685, + 0.02622598223388195, + 0.029520533978939056, + 0.03290701285004616, + 0.014459132216870785, + 0.027807367965579033, + 0.018705660477280617, + 0.01558070257306099, + 0.05037832260131836, + 0.01626461371779442, + 0.04033226892352104, + 0.02160978876054287, + 0.00766646396368742, + 0.016296254470944405, + 0.011242400854825974, + 0.010422326624393463, + 0.02303001470863819, + 0.015278861857950687, + 0.011652808636426926, + 0.008563550189137459, + 0.0043160212226212025, + 0.0077814701944589615, + 0.00490250438451767, + 0.0038898794446140528, + 0.005680304020643234, + 0.005086422897875309, + 0.0040893591940402985, + 0.0045852577313780785, + 0.0032190545462071896, + 0.009103246033191681, + 0.003974276129156351, + 0.005014478694647551, + 0.008797497488558292, + 0.0047703031450510025, + 0.006076289806514978, + 0.013072005473077297, + 0.006456997711211443, + 0.010467493906617165, + 0.005365788005292416, + 0.006421325728297234, + 0.016033995896577835, + 0.007856873795390129, + 0.008482222445309162, + 0.0026972710620611906, + 0.003754611825570464, + 0.0032323463819921017, + 0.0036589186638593674, + 0.0037622551899403334, + 0.003195134224370122, + 0.004038207698613405, + 0.0037377269472926855, + 0.007280707824975252, + 0.009071161970496178, + 0.0057914904318749905, + 0.0048894151113927364, + 0.008439244702458382, + 0.0069369059056043625, + 0.004558507818728685, + 0.007864835672080517, + 0.00856047309935093, + 0.01103896927088499, + 0.008985604159533978, + 0.012062877416610718, + 0.008869687095284462, + 0.005774512887001038, + 0.007785724475979805, + 0.007700070273131132, + 0.019097844138741493, + 0.00853317603468895, + 0.039815232157707214, + 0.006746046245098114, + 0.007653831969946623, + 0.009047499857842922, + 0.005577873904258013, + 0.009450547397136688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018219729885458946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036140859592705965 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036140859592705965, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036140859592705965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1443452090024948, + 0.07961374521255493, + 0.06831176578998566, + 0.051928646862506866, + 0.05455324426293373, + 0.14447827637195587, + 0.13844171166419983, + 0.058706991374492645, + 0.029315665364265442, + 0.23030471801757812 + ], + "credit_map": [ + 0.025795919820666313, + 0.011826924048364162, + 0.01580796390771866, + 0.01822390966117382, + 0.01619967631995678, + 0.02128720097243786, + 0.015764659270644188, + 0.01943894661962986, + 0.01117171160876751, + 0.006372157949954271, + 0.008612426929175854, + 0.008332950994372368, + 0.007450135424733162, + 0.012669151648879051, + 0.008816041052341461, + 0.016189172863960266, + 0.010436232201755047, + 0.004924645181745291, + 0.011836147867143154, + 0.005930724088102579, + 0.0046433391980826855, + 0.016416678205132484, + 0.006858449429273605, + 0.007265547290444374, + 0.010171255096793175, + 0.0025784422177821398, + 0.012366454117000103, + 0.00406691525131464, + 0.00303631485439837, + 0.011597986333072186, + 0.003105932381004095, + 0.005005346145480871, + 0.007228467613458633, + 0.005042641889303923, + 0.010338605381548405, + 0.00637398287653923, + 0.005785379558801651, + 0.007829694077372551, + 0.0051440466195344925, + 0.006810428109019995, + 0.025274796411395073, + 0.011300666257739067, + 0.02228654734790325, + 0.010661656036973, + 0.008308631367981434, + 0.03128694370388985, + 0.01646142452955246, + 0.018897611647844315, + 0.021860403940081596, + 0.01169196330010891, + 0.015339134261012077, + 0.011248491704463959, + 0.012353121303021908, + 0.030234012752771378, + 0.013758333399891853, + 0.021956240758299828, + 0.0070048184134066105, + 0.010294481180608273, + 0.007048099301755428, + 0.007039487361907959, + 0.006487617269158363, + 0.00802784226834774, + 0.004908435046672821, + 0.00789620727300644, + 0.0031838580034673214, + 0.004514927510172129, + 0.0034279185347259045, + 0.004541369620710611, + 0.003586194245144725, + 0.0037291045300662518, + 0.003308771410956979, + 0.0030235201120376587, + 0.042348943650722504, + 0.015377960167825222, + 0.09813622385263443, + 0.018871335312724113, + 0.013824879191815853, + 0.012690586969256401, + 0.013279846869409084, + 0.015774931758642197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0036140859592705965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493601784110069 + ], + "opposite_logprob_tokens": [ + -2.929935932159424 + ], + "logprob_reward": -0.05493601784110069, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493601784110069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671406626701355, + 0.08270809799432755, + 0.03932402655482292, + 0.04492342472076416, + 0.04736558720469475, + 0.214144766330719, + 0.1960238218307495, + 0.07445995509624481, + 0.059703923761844635, + 0.07420578598976135 + ], + "credit_map": [ + 0.03159957379102707, + 0.015777302905917168, + 0.023785028606653214, + 0.02111646719276905, + 0.01599694788455963, + 0.021414048969745636, + 0.013795704580843449, + 0.023655565455555916, + 0.010937143117189407, + 0.007368914317339659, + 0.009466635063290596, + 0.01184456143528223, + 0.007934841327369213, + 0.015692558139562607, + 0.009443172253668308, + 0.010020263493061066, + 0.007063291501253843, + 0.00216060527600348, + 0.006456498987972736, + 0.002825477160513401, + 0.0021913822274655104, + 0.008612074889242649, + 0.0035852647852152586, + 0.006429426837712526, + 0.006063256878405809, + 0.0017309436807408929, + 0.011728297919034958, + 0.002690007211640477, + 0.0024337389040738344, + 0.012947501614689827, + 0.0032566539011895657, + 0.00407301913946867, + 0.007833796553313732, + 0.0031637202482670546, + 0.009828946553170681, + 0.004123596474528313, + 0.003086941549554467, + 0.009527970105409622, + 0.0034509969409555197, + 0.0063496180810034275, + 0.037600159645080566, + 0.014227200299501419, + 0.03312774375081062, + 0.016507992520928383, + 0.010556631721556187, + 0.04492225497961044, + 0.028232013806700706, + 0.028970740735530853, + 0.028363995254039764, + 0.019739840179681778, + 0.02438148483633995, + 0.009973195381462574, + 0.021485231816768646, + 0.050346821546554565, + 0.010290181264281273, + 0.0314430333673954, + 0.01062942948192358, + 0.011553632095456123, + 0.009296469390392303, + 0.00760422321036458, + 0.008314649574458599, + 0.009866522625088692, + 0.005298352334648371, + 0.01189667358994484, + 0.011447676457464695, + 0.00611353712156415, + 0.013496137224137783, + 0.005617784336209297, + 0.005635423585772514, + 0.005441307555884123, + 0.0056852372363209724, + 0.00626681512221694, + 0.011649210005998611, + 0.005288435146212578, + 0.023257406428456306, + 0.006670105271041393, + 0.0064285146072506905, + 0.00819343514740467, + 0.004442439414560795, + 0.00827623438090086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.05493601784110069 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from an overhead angle, then gradually tilts upward and zooms out as the subject ascends, revealing a scene imbued with the riches of Eastern mythology. At the center of the frame, a majestic Chinese dragon, constructed entirely of clay, bears the subtle fingerprints and tool marks of hand-molding. Each scale appears individually molded and affixed, its edges slightly roughened, lending it a unique earthy texture. The dragon's head is majestic, its horns and flowing whiskers formed from golden clay, while its eyes, two smooth black glass beads, stand out against the matte clay skin. It soars slowly through the mist-shrouded mountains. Its movements are not fluid but rather rhythmic, with subtle pauses, each frame ascending, creating a unique frame-by-frame animation effect. Its serpentine body rises and falls rhythmically, its four claws making subtle strokes in the air. The backdrop is a rolling, inky-green mountain range, its peaks enveloped and obscured by a thick, cotton-like cloud, also sculpted from white clay. A few ancient, sturdy pines dot the landscape, evoking the texture of clay sculptures. Soft, diffused light evenly illuminates the scene, highlighting the matte texture of the clay, devoid of harsh reflections. The overall palette is simple and elegant, creating an artistic atmosphere of tranquility, solemnity, and the warmth of craftsmanship.", + "seed": 17856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_6/index_5656_seed_17856_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01734100147150457, + "vlm_logprob": -0.01734100147150457, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.01734100147150457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0373983234167099 + ], + "opposite_logprob_tokens": [ + -3.537398338317871 + ], + "logprob_reward": -0.0373983234167099, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0373983234167099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.159966841340065, + 0.048327017575502396, + 0.14432097971439362, + 0.1294284164905548, + 0.046067312359809875, + 0.17827603220939636, + 0.04105202853679657, + 0.03648040071129799, + 0.13544417917728424, + 0.08063683658838272 + ], + "credit_map": [ + 0.013436528854072094, + 0.014449993148446083, + 0.013714184053242207, + 0.014537421986460686, + 0.015696881338953972, + 0.020837493240833282, + 0.027789361774921417, + 0.039504967629909515, + 0.007906898856163025, + 0.005761642009019852, + 0.007604263722896576, + 0.005824596621096134, + 0.006781966425478458, + 0.004216909408569336, + 0.004127101507037878, + 0.0061036390252411366, + 0.018432874232530594, + 0.0180575642734766, + 0.015043485909700394, + 0.01691969484090805, + 0.012400094419717789, + 0.018328970298171043, + 0.01514976006001234, + 0.029988545924425125, + 0.014266053214669228, + 0.007259222213178873, + 0.006021382287144661, + 0.006156759802252054, + 0.0067529198713600636, + 0.04532632604241371, + 0.005428648553788662, + 0.03821711242198944, + 0.006774481851607561, + 0.0046620797365903854, + 0.0070790741592645645, + 0.0059064580127596855, + 0.004156903829425573, + 0.006908543407917023, + 0.004745222628116608, + 0.005834547337144613, + 0.01933509297668934, + 0.012338067404925823, + 0.027542822062969208, + 0.02039201743900776, + 0.01558306161314249, + 0.032981906086206436, + 0.011645005084574223, + 0.03845806419849396, + 0.00561658525839448, + 0.0043863593600690365, + 0.0069313435815274715, + 0.004619356710463762, + 0.004225289914757013, + 0.003455782076343894, + 0.0032484973780810833, + 0.00856881681829691, + 0.004774256143718958, + 0.004119225777685642, + 0.006484714802354574, + 0.004216288682073355, + 0.005345105659216642, + 0.003966139629483223, + 0.003823696868494153, + 0.0037509698886424303, + 0.0199272520840168, + 0.011021151207387447, + 0.020569652318954468, + 0.01403425820171833, + 0.014085677452385426, + 0.018701326102018356, + 0.010902339592576027, + 0.02620251290500164, + 0.010453472845256329, + 0.00817798636853695, + 0.013129647821187973, + 0.007161600515246391, + 0.008754624985158443, + 0.011080204509198666, + 0.007910958491265774, + 0.013968336395919323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0373983234167099 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003222037572413683 + ], + "opposite_logprob_tokens": [ + -5.7532219886779785 + ], + "logprob_reward": -0.003222037572413683, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003222037572413683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12244464457035065, + 0.033328939229249954, + 0.04167231172323227, + 0.5208251476287842, + 0.02406633459031582, + 0.09176862239837646, + 0.02637014165520668, + 0.023241251707077026, + 0.07653942704200745, + 0.039743196219205856 + ], + "credit_map": [ + 0.012171471491456032, + 0.01494405698031187, + 0.008946088142693043, + 0.014018746092915535, + 0.01622040942311287, + 0.015552016906440258, + 0.026916708797216415, + 0.0136751439422369, + 0.004711981862783432, + 0.0031294471118599176, + 0.0036223633214831352, + 0.004960115998983383, + 0.0034419961739331484, + 0.00358246429823339, + 0.0029946009162813425, + 0.006885966751724482, + 0.005666859447956085, + 0.006174693815410137, + 0.00425383634865284, + 0.005162269342690706, + 0.0045097521506249905, + 0.005112985614687204, + 0.004302425310015678, + 0.0064894892275333405, + 0.05579185485839844, + 0.021586956456303596, + 0.014443359337747097, + 0.01774611324071884, + 0.024148160591721535, + 0.2007395178079605, + 0.014234572649002075, + 0.1721346378326416, + 0.0032359003089368343, + 0.0025098565965890884, + 0.005413923412561417, + 0.0022922419011592865, + 0.001722145127132535, + 0.0036002376582473516, + 0.001953655621036887, + 0.0033383751288056374, + 0.009411437436938286, + 0.005092304199934006, + 0.014798873104155064, + 0.006960506550967693, + 0.006814184132963419, + 0.01932222954928875, + 0.0046508340165019035, + 0.024718252941966057, + 0.0034853036049753428, + 0.002045759465545416, + 0.0050711664371192455, + 0.0028145869728177786, + 0.002997308038175106, + 0.0023118590470403433, + 0.0022262358106672764, + 0.005417921580374241, + 0.003257456235587597, + 0.0017191126244142652, + 0.006019723601639271, + 0.0014152256771922112, + 0.0019810800440609455, + 0.003498959122225642, + 0.001759655773639679, + 0.0035900368820875883, + 0.01228413823992014, + 0.005852847825735807, + 0.011245721019804478, + 0.006085061002522707, + 0.009034891612827778, + 0.010956459678709507, + 0.006364939268678427, + 0.014715366996824741, + 0.0047027235850691795, + 0.0033941396977752447, + 0.007777398452162743, + 0.0032522878609597683, + 0.004437986295670271, + 0.005835505668073893, + 0.003617342561483383, + 0.006725810468196869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.003222037572413683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020613020285964012 + ], + "opposite_logprob_tokens": [ + -3.895612955093384 + ], + "logprob_reward": -0.020613020285964012, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020613020285964012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10549546033143997, + 0.058537960052490234, + 0.17260994017124176, + 0.14365482330322266, + 0.054098572582006454, + 0.10621076822280884, + 0.04910369589924812, + 0.03586240112781525, + 0.21941621601581573, + 0.0550101064145565 + ], + "credit_map": [ + 0.007141162641346455, + 0.009816343896090984, + 0.012313778512179852, + 0.012855199165642262, + 0.008821639232337475, + 0.011847427114844322, + 0.028209635987877846, + 0.014490275643765926, + 0.0062799472361803055, + 0.007701079361140728, + 0.006170176435261965, + 0.008351746015250683, + 0.007246947847306728, + 0.006436815485358238, + 0.00759838055819273, + 0.008752870373427868, + 0.01979885995388031, + 0.023842742666602135, + 0.015582315623760223, + 0.01679043099284172, + 0.016431378200650215, + 0.024150123819708824, + 0.015011850744485855, + 0.041002243757247925, + 0.020200753584504128, + 0.014441176317632198, + 0.008619027212262154, + 0.011543597094714642, + 0.01085939072072506, + 0.04177747294306755, + 0.008629691787064075, + 0.02758372761309147, + 0.008389578200876713, + 0.005500773899257183, + 0.01138234231621027, + 0.005014614202082157, + 0.00458097318187356, + 0.00820946879684925, + 0.0033914169762283564, + 0.00762940663844347, + 0.010596493259072304, + 0.008494814857840538, + 0.01596636138856411, + 0.007863318547606468, + 0.010036874562501907, + 0.020688241347670555, + 0.00569978728890419, + 0.026864873245358467, + 0.004995172377675772, + 0.004511524923145771, + 0.007341472897678614, + 0.006078606005758047, + 0.005987375508993864, + 0.003884918289259076, + 0.006630221381783485, + 0.009674401953816414, + 0.00493647949770093, + 0.0019227344309911132, + 0.010363744571805, + 0.0030442418064922094, + 0.003341704374179244, + 0.004792228806763887, + 0.0025450605899095535, + 0.0049162073992192745, + 0.03222109377384186, + 0.020016027614474297, + 0.02815459854900837, + 0.022500628605484962, + 0.03462885692715645, + 0.031105950474739075, + 0.02169840782880783, + 0.02909066341817379, + 0.006894448306411505, + 0.0036927429027855396, + 0.008976726792752743, + 0.003383248345926404, + 0.004624216351658106, + 0.009464249014854431, + 0.0050909677520394325, + 0.012883505783975124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.020613020285964012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021946171764284372 + ], + "opposite_logprob_tokens": [ + -6.127194404602051 + ], + "logprob_reward": -0.0021946171764284372, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021946171764284372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1515769511461258, + 0.04601137712597847, + 0.06413604319095612, + 0.11704393476247787, + 0.08862540870904922, + 0.13113881647586823, + 0.0722699910402298, + 0.09190505743026733, + 0.1503630429506302, + 0.08692940324544907 + ], + "credit_map": [ + 0.015092368237674236, + 0.010246221907436848, + 0.014670567587018013, + 0.01419879961758852, + 0.012527892366051674, + 0.022668777033686638, + 0.028010357171297073, + 0.03416196256875992, + 0.005593497771769762, + 0.005895424634218216, + 0.004791199695318937, + 0.005180070176720619, + 0.005768485367298126, + 0.004937297664582729, + 0.004640933591872454, + 0.009204464964568615, + 0.008090074174106121, + 0.0073822420090436935, + 0.006149229593575001, + 0.006148867774754763, + 0.0058921948075294495, + 0.00880917813628912, + 0.0062497626058757305, + 0.015414494089782238, + 0.01358611136674881, + 0.007686800789088011, + 0.0063685099594295025, + 0.006601003464311361, + 0.00504011707380414, + 0.03934931755065918, + 0.004937026184052229, + 0.033475056290626526, + 0.013328077271580696, + 0.007677485700696707, + 0.02183530665934086, + 0.006129743997007608, + 0.005614744499325752, + 0.015143798664212227, + 0.0062035671435296535, + 0.012692688032984734, + 0.014724927954375744, + 0.0064623053185641766, + 0.022901562973856926, + 0.00872128363698721, + 0.007900693453848362, + 0.029041728004813194, + 0.0044205887243151665, + 0.03696572780609131, + 0.011228104121983051, + 0.00496381102129817, + 0.013730248436331749, + 0.0076049561612308025, + 0.007299823220819235, + 0.005574841983616352, + 0.006052810233086348, + 0.01581539586186409, + 0.013581212610006332, + 0.005073562264442444, + 0.024764567613601685, + 0.007916317321360111, + 0.005185153800994158, + 0.014121252112090588, + 0.007812711410224438, + 0.01345028541982174, + 0.02325153723359108, + 0.011281760409474373, + 0.023322435095906258, + 0.011581006459891796, + 0.01910654455423355, + 0.022451166063547134, + 0.013797884806990623, + 0.025570707395672798, + 0.01021096296608448, + 0.006433676462620497, + 0.014150208793580532, + 0.006059128791093826, + 0.00917227752506733, + 0.01373233087360859, + 0.00876645091921091, + 0.018404366448521614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0021946171764284372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Chinese dragon constructed entirely of clay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the dragon's body exhibit visible hand-molding marks, roughened edges, and individually molded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon's horns and whiskers made of golden clay while its eyes are smooth black glass beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rolling, inky-green mountain range with white clay clouds and clay-textured pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from an overhead angle, then tilt upward and zoom out as the dragon ascends?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023277008906006813 + ], + "opposite_logprob_tokens": [ + -3.7732770442962646 + ], + "logprob_reward": -0.023277008906006813, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023277008906006813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1074092909693718, + 0.0667678639292717, + 0.21653200685977936, + 0.17706410586833954, + 0.03682399541139603, + 0.07088647782802582, + 0.05560528486967087, + 0.06375473737716675, + 0.07589016854763031, + 0.12926606833934784 + ], + "credit_map": [ + 0.008142566308379173, + 0.010197051800787449, + 0.007039438001811504, + 0.012469510547816753, + 0.01361146755516529, + 0.013487608171999454, + 0.025756558403372765, + 0.016705088317394257, + 0.009281677193939686, + 0.006043472792953253, + 0.007097805384546518, + 0.007420902140438557, + 0.0076346551068127155, + 0.0072974213398993015, + 0.006305177230387926, + 0.015686746686697006, + 0.027105005457997322, + 0.030164601281285286, + 0.022033479064702988, + 0.018354177474975586, + 0.019156912341713905, + 0.030168566852808, + 0.01706084981560707, + 0.05248841643333435, + 0.019305424764752388, + 0.010182764381170273, + 0.007484302390366793, + 0.009378510527312756, + 0.009352597407996655, + 0.062428656965494156, + 0.00678126560524106, + 0.05215059965848923, + 0.005480038467794657, + 0.0036296197213232517, + 0.008904626592993736, + 0.004046603571623564, + 0.0028625368140637875, + 0.0037644945550709963, + 0.003083039540797472, + 0.005053035914897919, + 0.009151214733719826, + 0.003955536987632513, + 0.011102060787379742, + 0.00495967548340559, + 0.004553474485874176, + 0.014627442695200443, + 0.002541160676628351, + 0.019995911046862602, + 0.008012259379029274, + 0.00792518351227045, + 0.0085346270352602, + 0.0065370346419513226, + 0.006242079194635153, + 0.004808313213288784, + 0.005206119269132614, + 0.008339671418070793, + 0.00706676347181201, + 0.006212722044438124, + 0.011442295275628567, + 0.00862707570195198, + 0.007739964872598648, + 0.007694350089877844, + 0.006823127623647451, + 0.008148439228534698, + 0.012976949103176594, + 0.006783671677112579, + 0.01052074134349823, + 0.00856187753379345, + 0.010655537247657776, + 0.010123592801392078, + 0.0056617846712470055, + 0.010606016963720322, + 0.014236307702958584, + 0.00745553569868207, + 0.026730788871645927, + 0.008465197868645191, + 0.00805251020938158, + 0.02512473240494728, + 0.00770160648971796, + 0.03149939328432083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.023277008906006813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_tensor([5912, 5912])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_tensor([5912, 5912])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f669e17d79a437954bba41f3e164c9829df0d5ea --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_tensor([5912, 5912])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, extreme close-up, rendered in HR Giger's signature dark, surreal style, captures a biomechanical pangolin as it slowly unfurls from a defensive ball. The creature's form is a nightmarish fusion of polished chrome and dark, damp chitin, the tight framing focused on the terrifyingly complex movements of its interlocking plates. With deliberate mechanical slowness, the chrome scales, reflecting a distorted, cold light, begin to separate from the smooth, oily chitin plates, revealing the gleaming, muscular flesh beneath. This underlying tissue is a grotesque tapestry of translucent muscle fibers interwoven with black, pulsating conduits and thin metallic wiring, all slick with a viscous fluid. The scene is illuminated by a harsh keylight from above, carving the creature out of the oppressive, ambiguous darkness and creating a stark interplay of metallic shimmer and deep, cavernous shadows. The color palette is a cool, desaturated monochrome of gunmetal and obsidian, emphasizing the alien textures and creating a claustrophobic, deeply unsettling atmosphere of biomechanical horror.", + "seed": 4947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4947_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2035910298116505, + "vlm_logprob": -0.2035910298116505, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.2035910298116505, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020708072930574417 + ], + "opposite_logprob_tokens": [ + -4.145708084106445 + ], + "logprob_reward": -0.020708072930574417, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020708072930574417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18735459446907043, + 0.04651787132024765, + 0.06569889932870865, + 0.06924441456794739, + 0.12972113490104675, + 0.08464065939188004, + 0.054325785487890244, + 0.20308104157447815, + 0.07129545509815216, + 0.08812010288238525 + ], + "credit_map": [ + 0.03998073562979698, + 0.013697976246476173, + 0.012758615426719189, + 0.014581928960978985, + 0.014516932889819145, + 0.01594879850745201, + 0.020680192857980728, + 0.055189408361911774, + 0.005179505329579115, + 0.007979434914886951, + 0.00564822880551219, + 0.005435449071228504, + 0.004790994338691235, + 0.0037871503736823797, + 0.005682420451194048, + 0.008014686405658722, + 0.008639084175229073, + 0.006750059314072132, + 0.006357330828905106, + 0.006650702562183142, + 0.009113889187574387, + 0.008699329569935799, + 0.010349265299737453, + 0.009139235131442547, + 0.007842098362743855, + 0.009276032447814941, + 0.006544596049934626, + 0.005031656939536333, + 0.015966452658176422, + 0.007308880798518658, + 0.008444026112556458, + 0.008830673061311245, + 0.03280191868543625, + 0.010047618299722672, + 0.01110024843364954, + 0.011154066771268845, + 0.012403439730405807, + 0.011946545913815498, + 0.015678981319069862, + 0.024588322266936302, + 0.007212755270302296, + 0.0073761846870183945, + 0.008743618614971638, + 0.01295869704335928, + 0.01073292177170515, + 0.01017748098820448, + 0.008098257705569267, + 0.019340746104717255, + 0.0048125311732292175, + 0.00853884220123291, + 0.0044816709123551846, + 0.0066006858833134174, + 0.008215151727199554, + 0.00614410312846303, + 0.004632431548088789, + 0.010900367982685566, + 0.014963897876441479, + 0.0210981797426939, + 0.053178414702415466, + 0.019095489755272865, + 0.012581015937030315, + 0.015953952446579933, + 0.015069619752466679, + 0.05114047974348068, + 0.007816425524652004, + 0.01052939798682928, + 0.009561982937157154, + 0.008211174979805946, + 0.008558082394301891, + 0.007565249223262072, + 0.00788606982678175, + 0.011167071759700775, + 0.010731391608715057, + 0.009437515400350094, + 0.009925369173288345, + 0.01014045998454094, + 0.010921324603259563, + 0.011773483827710152, + 0.01372503861784935, + 0.011465522460639477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.020708072930574417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033731408417224884 + ], + "opposite_logprob_tokens": [ + -3.408731460571289 + ], + "logprob_reward": -0.033731408417224884, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033731408417224884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.262776255607605, + 0.0613669790327549, + 0.025324679911136627, + 0.05989721044898033, + 0.06825044751167297, + 0.06113073229789734, + 0.09685946255922318, + 0.24962376058101654, + 0.06025295332074165, + 0.05451754108071327 + ], + "credit_map": [ + 0.043860383331775665, + 0.016216231510043144, + 0.0157671719789505, + 0.0212406013160944, + 0.015351871028542519, + 0.029815934598445892, + 0.02552083134651184, + 0.09500323981046677, + 0.00971115380525589, + 0.008553058840334415, + 0.00741218775510788, + 0.006155950948596001, + 0.004772265441715717, + 0.005823406390845776, + 0.007565243169665337, + 0.011373712681233883, + 0.003191764932125807, + 0.0026535089127719402, + 0.002704823622480035, + 0.002322608372196555, + 0.003012838074937463, + 0.003056036774069071, + 0.005334693472832441, + 0.003048405284062028, + 0.007263640873134136, + 0.009309579618275166, + 0.0051040551625192165, + 0.004167483188211918, + 0.013110365718603134, + 0.005065279547125101, + 0.007408055942505598, + 0.0084687490016222, + 0.01149486843496561, + 0.006665864959359169, + 0.00830019824206829, + 0.007230258546769619, + 0.008018816821277142, + 0.009150231257081032, + 0.007235073950141668, + 0.010155132040381432, + 0.005422177258878946, + 0.008363188244402409, + 0.008656715042889118, + 0.009927185252308846, + 0.0071157305501401424, + 0.0058136917650699615, + 0.007142786867916584, + 0.008689254522323608, + 0.013130756095051765, + 0.011285651475191116, + 0.009934647008776665, + 0.010632447898387909, + 0.011498795822262764, + 0.012349423952400684, + 0.014377085492014885, + 0.01365065947175026, + 0.015147579833865166, + 0.026256300508975983, + 0.06575776636600494, + 0.02598695643246174, + 0.015766805037856102, + 0.017211049795150757, + 0.01807883381843567, + 0.06541847437620163, + 0.007208407390862703, + 0.005429078359156847, + 0.008836527355015278, + 0.007301441393792629, + 0.0077079893089830875, + 0.007648412138223648, + 0.009877374395728111, + 0.006243722978979349, + 0.0059744687750935555, + 0.007829392328858376, + 0.006277158390730619, + 0.0055394601076841354, + 0.005907917860895395, + 0.008506170473992825, + 0.007987654767930508, + 0.006495315115898848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.033731408417224884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104642663151026 + ], + "opposite_logprob_tokens": [ + -5.5041046142578125 + ], + "logprob_reward": -0.004104642663151026, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104642663151026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13445384800434113, + 0.0587906613945961, + 0.02797674387693405, + 0.08903999626636505, + 0.18215103447437286, + 0.0639854297041893, + 0.09124778211116791, + 0.12784358859062195, + 0.08381069451570511, + 0.14070014655590057 + ], + "credit_map": [ + 0.03305314853787422, + 0.011702446267008781, + 0.008584997616708279, + 0.00979454442858696, + 0.009283162653446198, + 0.013226591050624847, + 0.012099594809114933, + 0.036709368228912354, + 0.007670469116419554, + 0.008530923165380955, + 0.0060987165197730064, + 0.0064448644407093525, + 0.005932742729783058, + 0.00571334408596158, + 0.008633935824036598, + 0.009765665978193283, + 0.00398501381278038, + 0.0028993599116802216, + 0.0035204188898205757, + 0.002687206258997321, + 0.003517200704663992, + 0.0029286800418049097, + 0.004692391492426395, + 0.0037464734632521868, + 0.01012382935732603, + 0.014860937371850014, + 0.00691459933295846, + 0.006394701078534126, + 0.020598560571670532, + 0.008007981814444065, + 0.009853219613432884, + 0.01228616014122963, + 0.0424155592918396, + 0.013749145902693272, + 0.017255017533898354, + 0.015593593008816242, + 0.020808542147278786, + 0.016253039240837097, + 0.01955295354127884, + 0.03652319684624672, + 0.005589847452938557, + 0.01312288548797369, + 0.008063765242695808, + 0.00510826800018549, + 0.010091695934534073, + 0.008123821578919888, + 0.006351647432893515, + 0.007533503230661154, + 0.01649063639342785, + 0.014508380554616451, + 0.010715213604271412, + 0.010636040940880775, + 0.011593999341130257, + 0.011516732163727283, + 0.00813446007668972, + 0.007652321830391884, + 0.00946517288684845, + 0.021929334849119186, + 0.017495326697826385, + 0.014255341142416, + 0.018679693341255188, + 0.01253652386367321, + 0.018822308629751205, + 0.014659899286925793, + 0.010070029646158218, + 0.009340579621493816, + 0.012647598050534725, + 0.011253438889980316, + 0.012067865580320358, + 0.009292351081967354, + 0.01133904978632927, + 0.007799783255904913, + 0.018760623410344124, + 0.01682845689356327, + 0.013248858973383904, + 0.014877629466354847, + 0.01837037317454815, + 0.020008014515042305, + 0.0257332194596529, + 0.012872990220785141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.004104642663151026 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157110691070557 + ], + "opposite_logprob_tokens": [ + -1.7015711069107056 + ], + "logprob_reward": -0.20157110691070557, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157110691070557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15387575328350067, + 0.12217717617750168, + 0.03751186653971672, + 0.07795847207307816, + 0.08075092732906342, + 0.19844792783260345, + 0.11541847884654999, + 0.11650190502405167, + 0.033644601702690125, + 0.06371280550956726 + ], + "credit_map": [ + 0.034231528639793396, + 0.013255958445370197, + 0.011384539306163788, + 0.01278766430914402, + 0.010750602930784225, + 0.009974820539355278, + 0.02066745050251484, + 0.04082319512963295, + 0.01918264478445053, + 0.015804562717676163, + 0.012881357222795486, + 0.011293948628008366, + 0.0093733761459589, + 0.012538855895400047, + 0.016718149185180664, + 0.024384282529354095, + 0.004272034391760826, + 0.0037459228187799454, + 0.0036853663623332977, + 0.005018165335059166, + 0.005992607679218054, + 0.0034297064412385225, + 0.006370030343532562, + 0.004998038988560438, + 0.008409657515585423, + 0.009720928966999054, + 0.0062543293461203575, + 0.005669405683875084, + 0.019457409158349037, + 0.006639162544161081, + 0.009889364242553711, + 0.01191821601241827, + 0.012807713821530342, + 0.008954695425927639, + 0.014083468355238438, + 0.00873317290097475, + 0.009532123804092407, + 0.006990061607211828, + 0.008577565662562847, + 0.011072139255702496, + 0.015387337654829025, + 0.03226080164313316, + 0.0242361668497324, + 0.029925301671028137, + 0.02538822591304779, + 0.019867874681949615, + 0.020293375477194786, + 0.03108886070549488, + 0.014409733936190605, + 0.02058245800435543, + 0.014831339940428734, + 0.013156182132661343, + 0.017003756016492844, + 0.013207817450165749, + 0.009917482733726501, + 0.012309720739722252, + 0.010133389383554459, + 0.016725877299904823, + 0.022828923538327217, + 0.013566605746746063, + 0.011687560006976128, + 0.011663444340229034, + 0.009389371611177921, + 0.02050674520432949, + 0.004207389894872904, + 0.003809772664681077, + 0.00398494815453887, + 0.00530104897916317, + 0.0033836299553513527, + 0.004489901941269636, + 0.004507114645093679, + 0.003960799891501665, + 0.011342578567564487, + 0.0044967057183384895, + 0.009086096659302711, + 0.006618624087423086, + 0.006301499903202057, + 0.010126566514372826, + 0.007996505126357079, + 0.007744240574538708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.20157110691070557 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578399181365967 + ], + "opposite_logprob_tokens": [ + -0.6328399181365967 + ], + "logprob_reward": -0.7578399181365967, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578399181365967, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09361281245946884, + 0.0291278176009655, + 0.01208131480962038, + 0.025905033573508263, + 0.20415693521499634, + 0.09970176219940186, + 0.03923339769244194, + 0.2810315191745758, + 0.1133478507399559, + 0.1018015444278717 + ], + "credit_map": [ + 0.014759735204279423, + 0.005235093180090189, + 0.0072616273537278175, + 0.009204035624861717, + 0.0066942814737558365, + 0.011022582650184631, + 0.009495630860328674, + 0.02993982471525669, + 0.003988876007497311, + 0.0037810439243912697, + 0.0024652888532727957, + 0.0037236500065773726, + 0.0027928301133215427, + 0.002365612890571356, + 0.005175656173378229, + 0.004834860097616911, + 0.0013294132659211755, + 0.0011663851328194141, + 0.0010187861043959856, + 0.001469729351811111, + 0.0017273453995585442, + 0.001227742643095553, + 0.0020286154467612505, + 0.002113296650350094, + 0.0031113093718886375, + 0.003249336965382099, + 0.0021065091714262962, + 0.0027184709906578064, + 0.004159938544034958, + 0.0034222418908029795, + 0.0032008057460188866, + 0.003936422523111105, + 0.0443757139146328, + 0.021938640624284744, + 0.025901157408952713, + 0.016669543460011482, + 0.013454530388116837, + 0.01853056252002716, + 0.02233681082725525, + 0.0409499816596508, + 0.00515694310888648, + 0.007324874866753817, + 0.007600910030305386, + 0.02710229903459549, + 0.006871972698718309, + 0.008112814277410507, + 0.00800907053053379, + 0.029522879049181938, + 0.005165846087038517, + 0.006453787442296743, + 0.0038935435004532337, + 0.004731787368655205, + 0.005211586598306894, + 0.004011064302176237, + 0.005434524267911911, + 0.00433125626295805, + 0.02689391002058983, + 0.06106855720281601, + 0.03894006460905075, + 0.03128129988908768, + 0.036850590258836746, + 0.02544531039893627, + 0.03373066335916519, + 0.026821129024028778, + 0.012733595445752144, + 0.01364713441580534, + 0.012636618688702583, + 0.016227001324295998, + 0.015021797269582748, + 0.01633964478969574, + 0.013211008161306381, + 0.013531050644814968, + 0.01208494696766138, + 0.011347542516887188, + 0.010594774037599564, + 0.010566544719040394, + 0.014404159039258957, + 0.014553370885550976, + 0.01766613870859146, + 0.010584063827991486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.7578399181365967 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, static, extreme close-up, rendered in HR Giger's signature dark, surreal style, captures a biomechanical pangolin as it slowly unfurls from a defensive ball. The creature's form is a nightmarish fusion of polished chrome and dark, damp chitin, the tight framing focused on the terrifyingly complex movements of its interlocking plates. With deliberate mechanical slowness, the chrome scales, reflecting a distorted, cold light, begin to separate from the smooth, oily chitin plates, revealing the gleaming, muscular flesh beneath. This underlying tissue is a grotesque tapestry of translucent muscle fibers interwoven with black, pulsating conduits and thin metallic wiring, all slick with a viscous fluid. The scene is illuminated by a harsh keylight from above, carving the creature out of the oppressive, ambiguous darkness and creating a stark interplay of metallic shimmer and deep, cavernous shadows. The color palette is a cool, desaturated monochrome of gunmetal and obsidian, emphasizing the alien textures and creating a claustrophobic, deeply unsettling atmosphere of biomechanical horror.", + "seed": 4948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4948_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04234171523712575, + "vlm_logprob": -0.04234171523712575, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04234171523712575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016356905922293663 + ], + "opposite_logprob_tokens": [ + -4.516356945037842 + ], + "logprob_reward": -0.016356905922293663, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.016356905922293663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11264137923717499, + 0.05015046149492264, + 0.043338652700185776, + 0.05538112670183182, + 0.05293998122215271, + 0.031129835173487663, + 0.044810134917497635, + 0.037279367446899414, + 0.450932115316391, + 0.12139694392681122 + ], + "credit_map": [ + 0.009361965581774712, + 0.010124343447387218, + 0.018465055152773857, + 0.019423937425017357, + 0.012984027154743671, + 0.01550904382020235, + 0.014343391172587872, + 0.012429609894752502, + 0.005455708596855402, + 0.006462200544774532, + 0.005163270048797131, + 0.005032609216868877, + 0.0055419206619262695, + 0.008431225083768368, + 0.00569729320704937, + 0.008366228081285954, + 0.002845540875568986, + 0.003292752429842949, + 0.007736385799944401, + 0.006151182111352682, + 0.005143287125974894, + 0.006977280136197805, + 0.006790736224502325, + 0.0044014835730195045, + 0.004556719679385424, + 0.004003171809017658, + 0.00568962749093771, + 0.005243274383246899, + 0.005935182329267263, + 0.009337115101516247, + 0.006342941429466009, + 0.014273090288043022, + 0.0023543047718703747, + 0.00993192009627819, + 0.005868792999535799, + 0.007591500878334045, + 0.004416831769049168, + 0.004373681731522083, + 0.011215862818062305, + 0.007187078706920147, + 0.0042838468216359615, + 0.003614597488194704, + 0.0040787686593830585, + 0.003954370506107807, + 0.003222976578399539, + 0.0033878523390740156, + 0.005598791874945164, + 0.0029886283446103334, + 0.004167821258306503, + 0.004814865067601204, + 0.004622204229235649, + 0.005570065230131149, + 0.0076223742216825485, + 0.005957060493528843, + 0.006960301660001278, + 0.005095437169075012, + 0.0031035705469548702, + 0.005274055991321802, + 0.006388506386429071, + 0.004574282560497522, + 0.004972801078110933, + 0.0042595588602125645, + 0.005034598056226969, + 0.0036719883792102337, + 0.03186492249369621, + 0.020390016958117485, + 0.010057829320430756, + 0.01843567006289959, + 0.16408400237560272, + 0.1632712185382843, + 0.02364327199757099, + 0.019185174256563187, + 0.016439121216535568, + 0.0129022728651762, + 0.013505294919013977, + 0.01285603642463684, + 0.01470308005809784, + 0.012511643581092358, + 0.016395583748817444, + 0.022083895280957222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.016356905922293663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373682498931885 + ], + "opposite_logprob_tokens": [ + -3.4087367057800293 + ], + "logprob_reward": -0.03373682498931885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373682498931885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053256116807460785, + 0.04450330138206482, + 0.029690934345126152, + 0.06387770175933838, + 0.08150613307952881, + 0.046778835356235504, + 0.03459493815898895, + 0.030128199607133865, + 0.505539059638977, + 0.11012472212314606 + ], + "credit_map": [ + 0.003865134669467807, + 0.004965933971107006, + 0.011384868994355202, + 0.0058937459252774715, + 0.005401623900979757, + 0.010028719902038574, + 0.0072754425927996635, + 0.004440657794475555, + 0.00433050561696291, + 0.005447369068861008, + 0.0044730231165885925, + 0.004725505597889423, + 0.003855984890833497, + 0.007439319044351578, + 0.004109066445380449, + 0.010122531093657017, + 0.0031040585599839687, + 0.0035340366885066032, + 0.00473895063623786, + 0.003862660378217697, + 0.0024402535054832697, + 0.004203511867672205, + 0.005041454918682575, + 0.0027660089544951916, + 0.006249413825571537, + 0.006539680063724518, + 0.007956113666296005, + 0.0055738696828484535, + 0.006422233302146196, + 0.008137334138154984, + 0.00685981335118413, + 0.016139252111315727, + 0.004392314236611128, + 0.008654252626001835, + 0.01068070624023676, + 0.010482802987098694, + 0.0104692243039608, + 0.010044313035905361, + 0.016511449590325356, + 0.010271080769598484, + 0.004000057466328144, + 0.004649593029171228, + 0.005861662793904543, + 0.010526072233915329, + 0.004842453636229038, + 0.0038714876864105463, + 0.008391023613512516, + 0.0046364925801754, + 0.0029110382311046124, + 0.0037280917167663574, + 0.006052989047020674, + 0.003638783236965537, + 0.004658782854676247, + 0.0033491384238004684, + 0.006857048254460096, + 0.003399068024009466, + 0.0022356812842190266, + 0.005897775758057833, + 0.002964311046525836, + 0.003547362983226776, + 0.0046895937994122505, + 0.003524692729115486, + 0.004523586481809616, + 0.002745195524767041, + 0.034799639135599136, + 0.023499932140111923, + 0.009944847784936428, + 0.016770057380199432, + 0.18627557158470154, + 0.18761979043483734, + 0.02560068666934967, + 0.02102859690785408, + 0.013469808734953403, + 0.010851438157260418, + 0.010380142368376255, + 0.01345840934664011, + 0.015918053686618805, + 0.010808281600475311, + 0.01638924703001976, + 0.018849339336156845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.03373682498931885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003628576872870326 + ], + "opposite_logprob_tokens": [ + -5.628628730773926 + ], + "logprob_reward": -0.003628576872870326, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003628576872870326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08608084917068481, + 0.047956958413124084, + 0.08948284387588501, + 0.10535595566034317, + 0.05589660629630089, + 0.1902969777584076, + 0.06040014326572418, + 0.04599433392286301, + 0.2243228256702423, + 0.09421258419752121 + ], + "credit_map": [ + 0.007257174700498581, + 0.00921577587723732, + 0.015259861014783382, + 0.011524495668709278, + 0.010146087035536766, + 0.010726547800004482, + 0.011747381649911404, + 0.010203513316810131, + 0.0030088119674474, + 0.00804146844893694, + 0.005842880345880985, + 0.005954778753221035, + 0.005003155674785376, + 0.008121115155518055, + 0.006460337899625301, + 0.005524406675249338, + 0.005912850610911846, + 0.0066572995856404305, + 0.013022758066654205, + 0.01207155454903841, + 0.013283491134643555, + 0.01726975105702877, + 0.013158464804291725, + 0.008106664754450321, + 0.01084413193166256, + 0.010594061575829983, + 0.020720740780234337, + 0.01130717247724533, + 0.01051319669932127, + 0.015704935416579247, + 0.013140537776052952, + 0.012531180866062641, + 0.0036490652710199356, + 0.00765260448679328, + 0.008968924172222614, + 0.00822694692760706, + 0.005376105662435293, + 0.006355029996484518, + 0.010509009473025799, + 0.005158916115760803, + 0.01649335026741028, + 0.018463637679815292, + 0.02662000246345997, + 0.04122070223093033, + 0.022381484508514404, + 0.018636072054505348, + 0.0336458794772625, + 0.012835821136832237, + 0.005845536943525076, + 0.005227856803685427, + 0.008765819482505322, + 0.008745737373828888, + 0.010215268470346928, + 0.008403069339692593, + 0.008055143058300018, + 0.0051417043432593346, + 0.004261615686118603, + 0.009804422967135906, + 0.0053390031680464745, + 0.006158731412142515, + 0.0063281008042395115, + 0.004805892705917358, + 0.005814299453049898, + 0.0034822607412934303, + 0.013971745036542416, + 0.009663120843470097, + 0.005836266092956066, + 0.01230346318334341, + 0.08269869536161423, + 0.08000492304563522, + 0.010676326230168343, + 0.009168269112706184, + 0.009003461338579655, + 0.009287302382290363, + 0.008077705278992653, + 0.013430175371468067, + 0.01346349436789751, + 0.011203712783753872, + 0.014974087476730347, + 0.014772631227970123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.003628576872870326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899539172649384 + ], + "opposite_logprob_tokens": [ + -2.5789954662323 + ], + "logprob_reward": -0.07899539172649384, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899539172649384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11861538141965866, + 0.0957593321800232, + 0.09773072600364685, + 0.05445628985762596, + 0.05160621926188469, + 0.06981214880943298, + 0.12838850915431976, + 0.07413570582866669, + 0.11765298992395401, + 0.1918426752090454 + ], + "credit_map": [ + 0.010609609074890614, + 0.014963828958570957, + 0.018488775938749313, + 0.021498627960681915, + 0.010896142572164536, + 0.016470296308398247, + 0.013301585800945759, + 0.01238651666790247, + 0.009892167523503304, + 0.014606391079723835, + 0.013983730226755142, + 0.010904209688305855, + 0.00937200989574194, + 0.010619854554533958, + 0.010374032892286777, + 0.016006939113140106, + 0.005976867862045765, + 0.01124902069568634, + 0.021369025111198425, + 0.014527443796396255, + 0.010328039526939392, + 0.009962322190403938, + 0.013516456820070744, + 0.010801544412970543, + 0.005212732125073671, + 0.0057763829827308655, + 0.007299970835447311, + 0.0055852062068879604, + 0.007783570792526007, + 0.007294365204870701, + 0.005879603326320648, + 0.009624461643397808, + 0.00326918950304389, + 0.008428315632045269, + 0.006943810265511274, + 0.00724535109475255, + 0.006993754766881466, + 0.0036763979587703943, + 0.011649928987026215, + 0.0033994708210229874, + 0.007285765837877989, + 0.007360351737588644, + 0.009556143544614315, + 0.01384895108640194, + 0.0073328749276697636, + 0.004123941995203495, + 0.012949281372129917, + 0.007354840636253357, + 0.010062901303172112, + 0.010283485054969788, + 0.01658765599131584, + 0.023857152089476585, + 0.020888997241854668, + 0.009736263193190098, + 0.027043871581554413, + 0.009928181767463684, + 0.006771563086658716, + 0.015403364785015583, + 0.006759290583431721, + 0.008362115360796452, + 0.010544593445956707, + 0.008402332663536072, + 0.01253752876073122, + 0.005354922264814377, + 0.01177904848009348, + 0.015124538913369179, + 0.009170755743980408, + 0.016965361312031746, + 0.017548177391290665, + 0.015756262466311455, + 0.016026441007852554, + 0.0152824055403471, + 0.024913154542446136, + 0.019906478002667427, + 0.019803430885076523, + 0.023211924359202385, + 0.02390965446829796, + 0.02407955192029476, + 0.025220884010195732, + 0.030797600746154785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.07899539172649384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789908766746521 + ], + "opposite_logprob_tokens": [ + -2.578990936279297 + ], + "logprob_reward": -0.0789908766746521, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789908766746521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08759349584579468, + 0.07160186767578125, + 0.11441191285848618, + 0.097148597240448, + 0.14748777449131012, + 0.06163153424859047, + 0.08215196430683136, + 0.024828718975186348, + 0.15294475853443146, + 0.160199373960495 + ], + "credit_map": [ + 0.007806430570781231, + 0.010727076791226864, + 0.011692141182720661, + 0.013557789847254753, + 0.011599147692322731, + 0.011759240180253983, + 0.011237804777920246, + 0.00921386107802391, + 0.006676150485873222, + 0.011568286456167698, + 0.011027460917830467, + 0.007878079079091549, + 0.008836292661726475, + 0.010939247906208038, + 0.006647635251283646, + 0.008028705604374409, + 0.008776907809078693, + 0.010631299577653408, + 0.015817983075976372, + 0.016044866293668747, + 0.017987217754125595, + 0.01311879139393568, + 0.020029904320836067, + 0.012004940770566463, + 0.007886750623583794, + 0.007246027700603008, + 0.011316363699734211, + 0.009198077954351902, + 0.009363834746181965, + 0.01214904710650444, + 0.01070302352309227, + 0.02928546443581581, + 0.008409153670072556, + 0.030222224071621895, + 0.01572386547923088, + 0.02447008714079857, + 0.014367274008691311, + 0.00857013650238514, + 0.03465445339679718, + 0.011070579290390015, + 0.0075698005966842175, + 0.006939708720892668, + 0.00815916433930397, + 0.00927573349326849, + 0.0059088184498250484, + 0.005991609767079353, + 0.01112439390271902, + 0.006662303116172552, + 0.006885242182761431, + 0.0063512856140732765, + 0.012684543617069721, + 0.0164292324334383, + 0.006599895190447569, + 0.006983594968914986, + 0.020203372463583946, + 0.0060147992335259914, + 0.0020552119240164757, + 0.0035467089619487524, + 0.0032790161203593016, + 0.0028852156829088926, + 0.0029526634607464075, + 0.003260253230109811, + 0.004021876025944948, + 0.002827773801982403, + 0.013486324809491634, + 0.013042464852333069, + 0.009572818875312805, + 0.019981440156698227, + 0.03550123795866966, + 0.031424250453710556, + 0.013741358183324337, + 0.016194863244891167, + 0.016935063526034355, + 0.016475433483719826, + 0.01445132214576006, + 0.01949349045753479, + 0.01925635151565075, + 0.021505216136574745, + 0.029858127236366272, + 0.02222435362637043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.0789908766746521 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up, rendered in HR Giger's signature dark, surreal style, captures a biomechanical pangolin as it slowly unfurls from a defensive ball. The creature's form is a nightmarish fusion of polished chrome and dark, damp chitin, the tight framing focused on the terrifyingly complex movements of its interlocking plates. With deliberate mechanical slowness, the chrome scales, reflecting a distorted, cold light, begin to separate from the smooth, oily chitin plates, revealing the gleaming, muscular flesh beneath. This underlying tissue is a grotesque tapestry of translucent muscle fibers interwoven with black, pulsating conduits and thin metallic wiring, all slick with a viscous fluid. The scene is illuminated by a harsh keylight from above, carving the creature out of the oppressive, ambiguous darkness and creating a stark interplay of metallic shimmer and deep, cavernous shadows. The color palette is a cool, desaturated monochrome of gunmetal and obsidian, emphasizing the alien textures and creating a claustrophobic, deeply unsettling atmosphere of biomechanical horror.", + "seed": 4949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4949_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07335316096432506, + "vlm_logprob": -0.07335316096432506, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.07335316096432506, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06349308043718338 + ], + "opposite_logprob_tokens": [ + -2.938493013381958 + ], + "logprob_reward": -0.06349308043718338, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06349308043718338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11240153759717941, + 0.0702521950006485, + 0.059101421386003494, + 0.06706337630748749, + 0.07544029504060745, + 0.3688430190086365, + 0.06872335076332092, + 0.05180305615067482, + 0.06652017682790756, + 0.05985154211521149 + ], + "credit_map": [ + 0.016024624928832054, + 0.01056207437068224, + 0.007580884732306004, + 0.011652383022010326, + 0.017660679295659065, + 0.01845986396074295, + 0.011457739397883415, + 0.01900329440832138, + 0.008635356090962887, + 0.007377552334219217, + 0.006402306724339724, + 0.006259368732571602, + 0.006445589940994978, + 0.012068388983607292, + 0.013679979369044304, + 0.009383659809827805, + 0.014276763424277306, + 0.003662113333120942, + 0.0031841809395700693, + 0.0034545385278761387, + 0.004489116370677948, + 0.007145713549107313, + 0.005794810596853495, + 0.017094189301133156, + 0.007669161539524794, + 0.006711794063448906, + 0.007326735649257898, + 0.008050000295042992, + 0.009086044505238533, + 0.0081276660785079, + 0.011926816776394844, + 0.00816516112536192, + 0.009550017304718494, + 0.006503300275653601, + 0.012622704729437828, + 0.007595671806484461, + 0.007764960639178753, + 0.006814439315348864, + 0.01576242782175541, + 0.008826779201626778, + 0.10305081307888031, + 0.021585889160633087, + 0.025480564683675766, + 0.019901346415281296, + 0.02118621952831745, + 0.01820245012640953, + 0.13241924345493317, + 0.027016522362828255, + 0.008310611359775066, + 0.007200249936431646, + 0.006844212766736746, + 0.007295204792171717, + 0.01078403927385807, + 0.00814454723149538, + 0.010167294181883335, + 0.009977200999855995, + 0.006751114036887884, + 0.006821250077337027, + 0.005750049836933613, + 0.007095971144735813, + 0.006973985116928816, + 0.004944344516843557, + 0.005511809140443802, + 0.007954533211886883, + 0.007030211389064789, + 0.009660185314714909, + 0.008646492846310139, + 0.005718550179153681, + 0.006280102301388979, + 0.008759693242609501, + 0.012401707470417023, + 0.008023236878216267, + 0.0072928378358483315, + 0.009965837933123112, + 0.00833839364349842, + 0.006791430991142988, + 0.007800507824867964, + 0.007220088504254818, + 0.005787279922515154, + 0.006655171513557434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.06349308043718338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161525905132294 + ], + "opposite_logprob_tokens": [ + -1.7016152143478394 + ], + "logprob_reward": -0.20161525905132294, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161525905132294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1241970807313919, + 0.04881802201271057, + 0.1893526166677475, + 0.07389889657497406, + 0.07185821235179901, + 0.14933760464191437, + 0.08371362090110779, + 0.05238879472017288, + 0.11401814967393875, + 0.09241697192192078 + ], + "credit_map": [ + 0.022584913298487663, + 0.01333176251500845, + 0.008322938345372677, + 0.010646755807101727, + 0.013773821294307709, + 0.018180597573518753, + 0.01610914245247841, + 0.021247150376439095, + 0.009989951737225056, + 0.004400311037898064, + 0.003714354243129492, + 0.003278518794104457, + 0.004424855578690767, + 0.006536159664392471, + 0.005884705111384392, + 0.01058916375041008, + 0.04946973919868469, + 0.011077008210122585, + 0.010882962495088577, + 0.01302131824195385, + 0.012092758901417255, + 0.015419957228004932, + 0.018074586987495422, + 0.059314291924238205, + 0.008520900271832943, + 0.009709272533655167, + 0.005727578420192003, + 0.008441121317446232, + 0.0068938639014959335, + 0.010349107906222343, + 0.014413365162909031, + 0.009843687526881695, + 0.008776046335697174, + 0.008299189619719982, + 0.01770761050283909, + 0.007052036002278328, + 0.005914070177823305, + 0.007563336752355099, + 0.008500552736222744, + 0.008045370690524578, + 0.042500972747802734, + 0.008305242285132408, + 0.008791727013885975, + 0.006369779352098703, + 0.007895628921687603, + 0.0069594779051840305, + 0.056671857833862305, + 0.011842912063002586, + 0.014639374800026417, + 0.006072319578379393, + 0.007820035330951214, + 0.007785221096128225, + 0.0071110837161540985, + 0.0104356175288558, + 0.02321138046681881, + 0.006638588849455118, + 0.006105124019086361, + 0.0035404337104409933, + 0.007117181085050106, + 0.007168282754719257, + 0.006144959479570389, + 0.007970805279910564, + 0.005327526479959488, + 0.009014484472572803, + 0.017862999811768532, + 0.012202968820929527, + 0.014282329007983208, + 0.015344205312430859, + 0.009926033206284046, + 0.016949769109487534, + 0.01532241702079773, + 0.012127424590289593, + 0.008594213053584099, + 0.011362574994564056, + 0.008399262093007565, + 0.012441092170774937, + 0.016622215509414673, + 0.010891737416386604, + 0.013385227881371975, + 0.01072064507752657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.20161525905132294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003630239749327302 + ], + "opposite_logprob_tokens": [ + -5.6286301612854 + ], + "logprob_reward": -0.003630239749327302, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003630239749327302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08233463019132614, + 0.04198281094431877, + 0.05217127501964569, + 0.09800422936677933, + 0.09934446215629578, + 0.14100217819213867, + 0.06148611009120941, + 0.05501140281558037, + 0.12324792891740799, + 0.24541500210762024 + ], + "credit_map": [ + 0.015100209973752499, + 0.0056955646723508835, + 0.005378397181630135, + 0.007453379686921835, + 0.0118497209623456, + 0.011458644643425941, + 0.010776936076581478, + 0.014621774666011333, + 0.008359122090041637, + 0.0033413430210202932, + 0.004186544567346573, + 0.0031550144776701927, + 0.0036424885038286448, + 0.0056471447460353374, + 0.006620188243687153, + 0.007030960638076067, + 0.010269977152347565, + 0.003540605306625366, + 0.0037539107725024223, + 0.0038945309352129698, + 0.004119002725929022, + 0.00750700943171978, + 0.006842892616987228, + 0.012243339791893959, + 0.008607670664787292, + 0.011117669753730297, + 0.010242206044495106, + 0.01182505302131176, + 0.013839050196111202, + 0.01165717002004385, + 0.018922168761491776, + 0.011793233454227448, + 0.02481844648718834, + 0.005629748106002808, + 0.009192033670842648, + 0.006671789102256298, + 0.006055481266230345, + 0.00912073627114296, + 0.011734960600733757, + 0.02612125873565674, + 0.03860572353005409, + 0.008884863927960396, + 0.008378161117434502, + 0.00980397593230009, + 0.009203889407217503, + 0.008044919930398464, + 0.04742800444364548, + 0.010652635246515274, + 0.010579144582152367, + 0.004380817525088787, + 0.005240251775830984, + 0.005196504294872284, + 0.008756261318922043, + 0.006065658759325743, + 0.015927758067846298, + 0.005339706316590309, + 0.006337261293083429, + 0.005704930052161217, + 0.004390236921608448, + 0.007341873366385698, + 0.008340572007000446, + 0.007793960627168417, + 0.0076469965279102325, + 0.007455566897988319, + 0.009659424424171448, + 0.016751816496253014, + 0.013067188672721386, + 0.012661519460380077, + 0.022943662479519844, + 0.022102635353803635, + 0.018030934035778046, + 0.008030741475522518, + 0.021793629974126816, + 0.039509281516075134, + 0.02040395326912403, + 0.04713562875986099, + 0.028840316459536552, + 0.02555251121520996, + 0.028741398826241493, + 0.03343825042247772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.003630239749327302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495959892868996 + ], + "opposite_logprob_tokens": [ + -2.929959535598755 + ], + "logprob_reward": -0.05495959892868996, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495959892868996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05777495354413986, + 0.14956924319267273, + 0.08605354279279709, + 0.0426141619682312, + 0.3163665533065796, + 0.038078341633081436, + 0.10153226554393768, + 0.13768018782138824, + 0.031023439019918442, + 0.03930733725428581 + ], + "credit_map": [ + 0.008615373633801937, + 0.005533162038773298, + 0.004829030949622393, + 0.00625851983204484, + 0.007242200430482626, + 0.00908836629241705, + 0.006269353441894054, + 0.009938949719071388, + 0.02887655608355999, + 0.013201815076172352, + 0.011863045394420624, + 0.011581290513277054, + 0.010927997529506683, + 0.019993700087070465, + 0.013577993027865887, + 0.03954685106873512, + 0.02236500196158886, + 0.00412764260545373, + 0.005352155305445194, + 0.005162087734788656, + 0.00530818710103631, + 0.006147446110844612, + 0.008747403509914875, + 0.028843622654676437, + 0.004556851927191019, + 0.005377411376684904, + 0.0040681082755327225, + 0.0051026721484959126, + 0.0056891716085374355, + 0.005986413918435574, + 0.0069497572258114815, + 0.004883775487542152, + 0.09258023649454117, + 0.016242563724517822, + 0.02343369834125042, + 0.016843708232045174, + 0.014470462687313557, + 0.02922622486948967, + 0.029929107055068016, + 0.09364055097103119, + 0.0035206901375204325, + 0.0027020585257560015, + 0.004351957235485315, + 0.0051561142317950726, + 0.005057771690189838, + 0.005094493273645639, + 0.006461774464696646, + 0.00573347881436348, + 0.023650195449590683, + 0.006190186832100153, + 0.009665735997259617, + 0.007330696564167738, + 0.005835584364831448, + 0.0090012326836586, + 0.02937709540128708, + 0.010481537319719791, + 0.016326749697327614, + 0.014375103637576103, + 0.015427515842020512, + 0.019675128161907196, + 0.020725980401039124, + 0.01632416434586048, + 0.014028005301952362, + 0.02079753391444683, + 0.004836736712604761, + 0.0032084081321954727, + 0.0036505667958408594, + 0.002658256096765399, + 0.0038164381403476, + 0.004890866111963987, + 0.005047227721661329, + 0.002914938610047102, + 0.005699999630451202, + 0.004233796149492264, + 0.0044745588675141335, + 0.006922633852809668, + 0.005291757173836231, + 0.004243890289217234, + 0.0047199660912156105, + 0.0037207372952252626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.05495959892868996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067626655101776 + ], + "opposite_logprob_tokens": [ + -3.168067693710327 + ], + "logprob_reward": -0.043067626655101776, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067626655101776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08401218056678772, + 0.06134778633713722, + 0.09462542086839676, + 0.06321478635072708, + 0.3480331301689148, + 0.04278434440493584, + 0.039624135941267014, + 0.09686394035816193, + 0.05883388593792915, + 0.1106603816151619 + ], + "credit_map": [ + 0.012280580587685108, + 0.010058765299618244, + 0.006982275750488043, + 0.0064494130201637745, + 0.007023852784186602, + 0.015979662537574768, + 0.009485384449362755, + 0.015752242878079414, + 0.011739054694771767, + 0.0051299030892550945, + 0.004988210275769234, + 0.005949023179709911, + 0.00466211186721921, + 0.007245289161801338, + 0.007960259914398193, + 0.013673932291567326, + 0.017924174666404724, + 0.006087905261665583, + 0.011670257896184921, + 0.009672073647379875, + 0.005808582995086908, + 0.008543957024812698, + 0.011361933313310146, + 0.023556532338261604, + 0.008568883873522282, + 0.008855391293764114, + 0.006640092935413122, + 0.007287910208106041, + 0.007940758019685745, + 0.007732003927230835, + 0.008671028539538383, + 0.007518715225160122, + 0.042358893901109695, + 0.03173520788550377, + 0.07369571179151535, + 0.03225162252783775, + 0.027934107929468155, + 0.04051033407449722, + 0.05739700794219971, + 0.042150240391492844, + 0.007932701148092747, + 0.0029950225725769997, + 0.004679892212152481, + 0.004334887955337763, + 0.002742190146818757, + 0.0031967952381819487, + 0.012018590234220028, + 0.004884264897555113, + 0.007106466684490442, + 0.0027249502018094063, + 0.0034128096885979176, + 0.0034816728439182043, + 0.0033391411416232586, + 0.003206513822078705, + 0.011200995184481144, + 0.0051515898667275906, + 0.009653160348534584, + 0.008283709175884724, + 0.009355973452329636, + 0.012916469015181065, + 0.014861341565847397, + 0.012727344408631325, + 0.012934403494000435, + 0.016131537035107613, + 0.00835795234888792, + 0.00528964027762413, + 0.00590731343254447, + 0.004758098628371954, + 0.006034859921783209, + 0.01163014117628336, + 0.011767548508942127, + 0.005088334903120995, + 0.010684817098081112, + 0.015015878714621067, + 0.009048797190189362, + 0.020380664616823196, + 0.013858490623533726, + 0.012577405199408531, + 0.018357964232563972, + 0.010736364871263504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.043067626655101776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up, rendered in HR Giger's signature dark, surreal style, captures a biomechanical pangolin as it slowly unfurls from a defensive ball. The creature's form is a nightmarish fusion of polished chrome and dark, damp chitin, the tight framing focused on the terrifyingly complex movements of its interlocking plates. With deliberate mechanical slowness, the chrome scales, reflecting a distorted, cold light, begin to separate from the smooth, oily chitin plates, revealing the gleaming, muscular flesh beneath. This underlying tissue is a grotesque tapestry of translucent muscle fibers interwoven with black, pulsating conduits and thin metallic wiring, all slick with a viscous fluid. The scene is illuminated by a harsh keylight from above, carving the creature out of the oppressive, ambiguous darkness and creating a stark interplay of metallic shimmer and deep, cavernous shadows. The color palette is a cool, desaturated monochrome of gunmetal and obsidian, emphasizing the alien textures and creating a claustrophobic, deeply unsettling atmosphere of biomechanical horror.", + "seed": 4950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_7/index_5912_seed_4950_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06377654066309332, + "vlm_logprob": -0.06377654066309332, + "vlm_yes_no_margin": 3.2249999761581423, + "vlm_token_credit": -0.06377654066309332, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023448433727025986 + ], + "opposite_logprob_tokens": [ + -4.023448467254639 + ], + "logprob_reward": -0.023448433727025986, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.023448433727025986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0886947438120842, + 0.06922087073326111, + 0.039411019533872604, + 0.04297729581594467, + 0.038038015365600586, + 0.03173723816871643, + 0.12313622236251831, + 0.032406650483608246, + 0.2655239403247833, + 0.268854022026062 + ], + "credit_map": [ + 0.0139089310541749, + 0.007006167434155941, + 0.015172220766544342, + 0.007391196675598621, + 0.014331666752696037, + 0.012915250845253468, + 0.010205615311861038, + 0.007763697300106287, + 0.016319239512085915, + 0.003737517399713397, + 0.015028017573058605, + 0.00482042646035552, + 0.0067604463547468185, + 0.008205345831811428, + 0.0061535765416920185, + 0.008196297101676464, + 0.004251659847795963, + 0.0023137007374316454, + 0.0034607581328600645, + 0.004128993488848209, + 0.004233343061059713, + 0.004532601684331894, + 0.01114911213517189, + 0.005340848583728075, + 0.003901047632098198, + 0.0029568939935415983, + 0.0045388732105493546, + 0.005238990765064955, + 0.005597090348601341, + 0.006830400321632624, + 0.00491164019331336, + 0.00900236051529646, + 0.003278665244579315, + 0.004876017104834318, + 0.003077226458117366, + 0.002894297242164612, + 0.0073443143628537655, + 0.006140838842839003, + 0.005435001105070114, + 0.004991654306650162, + 0.0031688848976045847, + 0.0045442660339176655, + 0.003918222617357969, + 0.003780847182497382, + 0.004167138133198023, + 0.004803469870239496, + 0.0037837892305105925, + 0.003570619272068143, + 0.009825153276324272, + 0.020639142021536827, + 0.015289274975657463, + 0.01286393217742443, + 0.016841238364577293, + 0.02423972822725773, + 0.01222776249051094, + 0.01120998989790678, + 0.00263720634393394, + 0.002654933137819171, + 0.003340719733387232, + 0.005556497722864151, + 0.0051619913429021835, + 0.004682356026023626, + 0.005266851279884577, + 0.003106094431132078, + 0.012919639237225056, + 0.015089577995240688, + 0.016344573348760605, + 0.09450092911720276, + 0.014143791049718857, + 0.016544297337532043, + 0.07701142877340317, + 0.018969712778925896, + 0.07211141288280487, + 0.03318515792489052, + 0.02164236083626747, + 0.018834833055734634, + 0.016897577792406082, + 0.033990561962127686, + 0.018623296171426773, + 0.053568821400403976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.023448433727025986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270962506532669 + ], + "opposite_logprob_tokens": [ + -2.127096176147461 + ], + "logprob_reward": -0.1270962506532669, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270962506532669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046344272792339325, + 0.04795132204890251, + 0.03480206057429314, + 0.059027258306741714, + 0.031907688826322556, + 0.0336993932723999, + 0.050985801964998245, + 0.08440208435058594, + 0.3127947747707367, + 0.29808536171913147 + ], + "credit_map": [ + 0.004610608797520399, + 0.004083401523530483, + 0.004586323164403439, + 0.0056509654968976974, + 0.010227523744106293, + 0.007351312320679426, + 0.005188195966184139, + 0.00464594317600131, + 0.012590020895004272, + 0.0027961500454694033, + 0.012039368972182274, + 0.0032242992892861366, + 0.0035326345823705196, + 0.0039164102636277676, + 0.005657676141709089, + 0.004194761626422405, + 0.00374300847761333, + 0.0028265246655792, + 0.003817614633589983, + 0.0038207366596907377, + 0.00562158552929759, + 0.005965492222458124, + 0.00484911585226655, + 0.0041579799726605415, + 0.0038985502906143665, + 0.0042404718697071075, + 0.005789753049612045, + 0.006113890092819929, + 0.012258127331733704, + 0.008531155064702034, + 0.01133694127202034, + 0.0068583725951612, + 0.002891409443691373, + 0.0049025933258235455, + 0.0032621652353554964, + 0.0023834689054638147, + 0.006631945725530386, + 0.004852119367569685, + 0.0033976139966398478, + 0.0035863728262484074, + 0.0029148668982088566, + 0.004273738246411085, + 0.003004735568538308, + 0.0039115226827561855, + 0.003428457072004676, + 0.006097440607845783, + 0.004843960981816053, + 0.005224668886512518, + 0.0033087055198848248, + 0.007544702850282192, + 0.005612388253211975, + 0.010885292664170265, + 0.0059125605039298534, + 0.005638891831040382, + 0.006508451420813799, + 0.00557481124997139, + 0.006145854014903307, + 0.007292137015610933, + 0.008684972301125526, + 0.01516637485474348, + 0.009620330296456814, + 0.009265458211302757, + 0.020123936235904694, + 0.008103027008473873, + 0.0111905038356781, + 0.020543769001960754, + 0.023023031651973724, + 0.10240129381418228, + 0.015318590216338634, + 0.01652730442583561, + 0.09080824255943298, + 0.032982029020786285, + 0.1068079024553299, + 0.02854408323764801, + 0.014998159371316433, + 0.010198689997196198, + 0.01210612803697586, + 0.031216423958539963, + 0.016409317031502724, + 0.07780466228723526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.1270962506532669 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963153671473265 + ], + "opposite_logprob_tokens": [ + -5.130963325500488 + ], + "logprob_reward": -0.005963153671473265, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963153671473265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09166304767131805, + 0.040115322917699814, + 0.03937675431370735, + 0.07531967014074326, + 0.025339512154459953, + 0.09465324133634567, + 0.14809121191501617, + 0.050803277641534805, + 0.18437032401561737, + 0.2502676248550415 + ], + "credit_map": [ + 0.01122133620083332, + 0.007783149369060993, + 0.01163925789296627, + 0.007814031094312668, + 0.022496970370411873, + 0.013261710293591022, + 0.008825492113828659, + 0.00862109661102295, + 0.0068313851952552795, + 0.002330582356080413, + 0.0076981461606919765, + 0.0034028489608317614, + 0.0043929098173975945, + 0.0041539715602993965, + 0.00668939296156168, + 0.004616085439920425, + 0.0033982915338128805, + 0.002816365100443363, + 0.0033348840661346912, + 0.0036721087526530027, + 0.005296470131725073, + 0.0059453449212014675, + 0.009836282581090927, + 0.005077005829662085, + 0.006090928800404072, + 0.005881749093532562, + 0.00482726376503706, + 0.005784034729003906, + 0.015730252489447594, + 0.012777869589626789, + 0.014950393699109554, + 0.00927718821913004, + 0.0026076706126332283, + 0.002931850729510188, + 0.0024279430508613586, + 0.0014895829372107983, + 0.0051629929803311825, + 0.0038369237445294857, + 0.0033091322984546423, + 0.003573415568098426, + 0.009442419745028019, + 0.012892356142401695, + 0.00867689959704876, + 0.009221279993653297, + 0.010202552191913128, + 0.01795763336122036, + 0.012918688356876373, + 0.01334141381084919, + 0.01097888220101595, + 0.024686051532626152, + 0.015353943221271038, + 0.019249754026532173, + 0.018410831689834595, + 0.028655940666794777, + 0.016928955912590027, + 0.013826847076416016, + 0.004723373334854841, + 0.005884113255888224, + 0.005337852984666824, + 0.008628403767943382, + 0.0061413259245455265, + 0.008125892840325832, + 0.006581498775631189, + 0.005380816757678986, + 0.011063319630920887, + 0.012947018258273602, + 0.013877362944185734, + 0.06089329719543457, + 0.012463425286114216, + 0.011883694678544998, + 0.04577314853668213, + 0.01546905655413866, + 0.08525385707616806, + 0.02547452226281166, + 0.011755062267184258, + 0.011202291585505009, + 0.01372545212507248, + 0.025772931054234505, + 0.013988646678626537, + 0.0630948394536972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.005963153671473265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032142698764801 + ], + "opposite_logprob_tokens": [ + -2.3503215312957764 + ], + "logprob_reward": -0.10032142698764801, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032142698764801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05438836291432381, + 0.08985849469900131, + 0.03970242291688919, + 0.09802157431840897, + 0.029546869918704033, + 0.03475295752286911, + 0.03914770111441612, + 0.0933818370103836, + 0.14615881443023682, + 0.37504100799560547 + ], + "credit_map": [ + 0.007676796056330204, + 0.00419175811111927, + 0.009342171251773834, + 0.005129353608936071, + 0.00948482844978571, + 0.006220141425728798, + 0.005508872214704752, + 0.0068344431929290295, + 0.028581984341144562, + 0.004676568787544966, + 0.02626214548945427, + 0.004532115999609232, + 0.004956615623086691, + 0.00640224851667881, + 0.006977846845984459, + 0.0074689630419015884, + 0.006003447808325291, + 0.0028194966726005077, + 0.005044350866228342, + 0.005055722314864397, + 0.005671738646924496, + 0.005391211248934269, + 0.005223395302891731, + 0.004493062384426594, + 0.007416425738483667, + 0.007377401925623417, + 0.010915440507233143, + 0.009539184160530567, + 0.016372503712773323, + 0.015797583386301994, + 0.015643469989299774, + 0.014959569089114666, + 0.005457806400954723, + 0.0031780328135937452, + 0.0042164940387010574, + 0.0024298967327922583, + 0.0029785519000142813, + 0.0037018521688878536, + 0.00306553952395916, + 0.00451869610697031, + 0.0030083628371357918, + 0.003492653137072921, + 0.0033015673980116844, + 0.0021811537444591522, + 0.00390882883220911, + 0.007497752550989389, + 0.004425572231411934, + 0.006937065627425909, + 0.0033467316534370184, + 0.005779210012406111, + 0.003998695407062769, + 0.00394267775118351, + 0.005077177193015814, + 0.007104512304067612, + 0.005437610205262899, + 0.0044610886834561825, + 0.005905818659812212, + 0.006590799894183874, + 0.009751713834702969, + 0.017959440127015114, + 0.010693513788282871, + 0.009967018850147724, + 0.019623544067144394, + 0.012889985926449299, + 0.0050422982312738895, + 0.007338909897953272, + 0.006618643179535866, + 0.05614747107028961, + 0.006039443425834179, + 0.008808127604424953, + 0.04826326668262482, + 0.007900652475655079, + 0.13872109353542328, + 0.034033406525850296, + 0.013403930701315403, + 0.011993352323770523, + 0.02032018080353737, + 0.03682196885347366, + 0.018687361851334572, + 0.1010596975684166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.10032142698764801 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a biomechanical pangolin composed of polished chrome and dark chitin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pangolin shown slowly unfurling from a defensive ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature reveal translucent muscle fibers, black pulsating conduits, and metallic wiring beneath its plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a harsh keylight from above creating stark metallic shimmer and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette a cool, desaturated monochrome of gunmetal and obsidian?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062053438276052475 + ], + "opposite_logprob_tokens": [ + -2.8120534420013428 + ], + "logprob_reward": -0.062053438276052475, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062053438276052475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05394658073782921, + 0.08132998645305634, + 0.05048873648047447, + 0.02559957094490528, + 0.03261470049619675, + 0.03198060765862465, + 0.027099553495645523, + 0.1386594921350479, + 0.09111908078193665, + 0.46716171503067017 + ], + "credit_map": [ + 0.0075211916118860245, + 0.003801629412919283, + 0.006340912543237209, + 0.005079830531030893, + 0.008762241341173649, + 0.006940904539078474, + 0.00915940385311842, + 0.006340467371046543, + 0.022940125316381454, + 0.004311548080295324, + 0.02031361311674118, + 0.005209547467529774, + 0.00460905022919178, + 0.005597494076937437, + 0.010281291790306568, + 0.008067317306995392, + 0.0083383833989501, + 0.003785699373111129, + 0.005676476750522852, + 0.005375951528549194, + 0.005787887144833803, + 0.007867787033319473, + 0.007769515737891197, + 0.005887033883482218, + 0.003030011896044016, + 0.0019221138209104538, + 0.0027648117393255234, + 0.002585164038464427, + 0.0047839307226240635, + 0.0035223187878727913, + 0.003344683675095439, + 0.003646536497399211, + 0.0052719274535775185, + 0.004598899278789759, + 0.003631783649325371, + 0.002200524788349867, + 0.005279305391013622, + 0.004130077548325062, + 0.003771137213334441, + 0.003731043776497245, + 0.002729246160015464, + 0.0034459400922060013, + 0.0035941344685852528, + 0.003392444923520088, + 0.003232982475310564, + 0.006225128658115864, + 0.0037781691644340754, + 0.00558256171643734, + 0.0018483753083273768, + 0.004213403910398483, + 0.004438605159521103, + 0.0027418143581598997, + 0.0031314780935645103, + 0.0055737970396876335, + 0.0025598653592169285, + 0.0025922150816768408, + 0.00646958639845252, + 0.009797302074730396, + 0.008496013469994068, + 0.03873928263783455, + 0.012217224575579166, + 0.01690658926963806, + 0.025369973853230476, + 0.02066352218389511, + 0.0038907185662537813, + 0.007992768660187721, + 0.005964940879493952, + 0.027466852217912674, + 0.0069512478075921535, + 0.006329876370728016, + 0.02449517883360386, + 0.00802749302238226, + 0.16353052854537964, + 0.04294821247458458, + 0.018817922100424767, + 0.016803402453660965, + 0.03279871121048927, + 0.05010456591844559, + 0.023547157645225525, + 0.11861118674278259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.062053438276052475 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_8/index_tensor([7245, 7245])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_8/index_tensor([7245, 7245])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fe43f170b7f4bd07cfea71468e4b633a88fd6962 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_8/index_tensor([7245, 7245])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot focuses precisely on a pair of hands with vibrant skin and distinct knuckles, carefully cradling a square, layered French mille-feuille pastry. The pastry is evenly coated with a thick, delicate layer of snow-white icing. Warm afternoon sunlight streams through the window on the left, casting soft golden outlines on the pastry and the backs of the hands. In the background, a wooden kitchen counter and a partially blurred white tiled wall are softly blurred. As another pair of equally warm hands slowly reaches in from the right, ready to gently receive this sweet gift, the gentle touch and exchange of hands slightly disturbs the top layer of powdered sugar on the pastry. At this moment, the scene enters an extreme slow motion effect, with countless tiny particles of powdered sugar rising up and dancing like shimmering stardust in the golden light penetrating the kitchen, each particle clearly visible before slowly dispersing. The main colors of the entire picture are warm beige and wood color, with soft light and shadow, full of tranquility, healing comfort, creating a pure and warm atmosphere.", + "seed": 13534, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_8/index_7245_seed_13534_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17697645714506507, + "vlm_logprob": -0.17697645714506507, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.17697645714506507, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.760953426361084 + ], + "opposite_logprob_tokens": [ + -0.635953426361084 + ], + "logprob_reward": -0.760953426361084, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.760953426361084, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09678096324205399, + 0.03246092423796654, + 0.020001068711280823, + 0.030573803931474686, + 0.016642579808831215, + 0.04318176582455635, + 0.10271477699279785, + 0.13947968184947968, + 0.38246992230415344, + 0.13569457828998566 + ], + "credit_map": [ + 0.0070931497029960155, + 0.012904349714517593, + 0.00512847350910306, + 0.008977306075394154, + 0.02102000080049038, + 0.020330606028437614, + 0.010281119495630264, + 0.011045954190194607, + 0.002632140414789319, + 0.0036356658674776554, + 0.00294996565207839, + 0.002759018214419484, + 0.007417996879667044, + 0.006565961986780167, + 0.0028964332304894924, + 0.0036037417594343424, + 0.0029914926271885633, + 0.001774461125023663, + 0.0016983079258352518, + 0.0035045838449150324, + 0.002689753659069538, + 0.0026360522024333477, + 0.0020149494521319866, + 0.0026914672926068306, + 0.0027228309772908688, + 0.003445300506427884, + 0.0026325518265366554, + 0.005663592368364334, + 0.004680803511291742, + 0.004012868274003267, + 0.0035554436035454273, + 0.0038604135625064373, + 0.0015035083051770926, + 0.002268560230731964, + 0.0017575620440766215, + 0.0031416709534823895, + 0.0018610650440678, + 0.0030000924598425627, + 0.0017147073522210121, + 0.001395413069985807, + 0.003133634105324745, + 0.005493987817317247, + 0.003575971582904458, + 0.006044973619282246, + 0.007893803529441357, + 0.005962473340332508, + 0.003860343713313341, + 0.0072165788151323795, + 0.010708345100283623, + 0.014695370569825172, + 0.005828670226037502, + 0.01113977376371622, + 0.023773374035954475, + 0.022350642830133438, + 0.008530491963028908, + 0.005688108038157225, + 0.010862365365028381, + 0.020858632400631905, + 0.009754455648362637, + 0.013471201993525028, + 0.0330403596162796, + 0.021323854103684425, + 0.014801768586039543, + 0.015367056243121624, + 0.043868109583854675, + 0.07788818329572678, + 0.01886013336479664, + 0.03552927449345589, + 0.03131633996963501, + 0.04302568733692169, + 0.04949316382408142, + 0.0824890211224556, + 0.015331528149545193, + 0.02120872028172016, + 0.014637666754424572, + 0.022679410874843597, + 0.01298435963690281, + 0.012390948832035065, + 0.017868995666503906, + 0.018592944368720055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.760953426361084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754779256880283 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.006754779256880283, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754779256880283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16554291546344757, + 0.030098650604486465, + 0.020665090531110764, + 0.02605195716023445, + 0.05742360278964043, + 0.023130057379603386, + 0.03384731337428093, + 0.14450180530548096, + 0.13767032325267792, + 0.3610682785511017 + ], + "credit_map": [ + 0.013754372484982014, + 0.020649917423725128, + 0.008645941503345966, + 0.01792754977941513, + 0.03315094858407974, + 0.040644533932209015, + 0.016148721799254417, + 0.014620935544371605, + 0.00237961090169847, + 0.003827338106930256, + 0.003064516931772232, + 0.0056556700728833675, + 0.003475168952718377, + 0.004982262384146452, + 0.00353655475191772, + 0.0031775289680808783, + 0.0035566759761422873, + 0.002636753721162677, + 0.0019501085625961423, + 0.0033983425237238407, + 0.001875267713330686, + 0.003105126554146409, + 0.0017942137783393264, + 0.0023485999554395676, + 0.0027692699804902077, + 0.0028288026805967093, + 0.002358583267778158, + 0.0047935983166098595, + 0.0035926317796111107, + 0.0034711929038167, + 0.00299831829033792, + 0.003239558544009924, + 0.00400565704330802, + 0.00749213295057416, + 0.0033349411096423864, + 0.009995169937610626, + 0.009700529277324677, + 0.012462951242923737, + 0.0051245070062577724, + 0.005307713523507118, + 0.0016176962526515126, + 0.003961904440075159, + 0.001989812357351184, + 0.0026836248580366373, + 0.004956488963216543, + 0.0029419211205095053, + 0.0024678464978933334, + 0.0025107611436396837, + 0.002480764640495181, + 0.004404392093420029, + 0.002715233014896512, + 0.005580027122050524, + 0.006392586510628462, + 0.004350872244685888, + 0.004541794303804636, + 0.0033816432114690542, + 0.012478942982852459, + 0.024315120652318, + 0.013389826752245426, + 0.015681149438023567, + 0.02336946316063404, + 0.03510778397321701, + 0.011020557023584843, + 0.009138967841863632, + 0.012128854170441628, + 0.016527174040675163, + 0.014154311269521713, + 0.0240586306899786, + 0.017982231453061104, + 0.015069960616528988, + 0.020566042512655258, + 0.017183126881718636, + 0.07628646492958069, + 0.13927465677261353, + 0.014003384858369827, + 0.01690603420138359, + 0.024423573166131973, + 0.0481850765645504, + 0.022434117272496223, + 0.019554972648620605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.006754779256880283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096450749784708 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096450749784708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096450749784708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13029196858406067, + 0.10663994401693344, + 0.043052319437265396, + 0.0650290697813034, + 0.03105819970369339, + 0.031073369085788727, + 0.12780946493148804, + 0.0788751170039177, + 0.1984468549489975, + 0.18772368133068085 + ], + "credit_map": [ + 0.01758929155766964, + 0.013074756599962711, + 0.009083636105060577, + 0.018366582691669464, + 0.018559258431196213, + 0.023462112993001938, + 0.014023860916495323, + 0.016132473945617676, + 0.016236472874879837, + 0.01127470564097166, + 0.007339920848608017, + 0.008529291488230228, + 0.01427864097058773, + 0.03162594139575958, + 0.007942680269479752, + 0.009412290528416634, + 0.006682432722300291, + 0.003626501653343439, + 0.003210168331861496, + 0.0042135464027523994, + 0.009275363758206367, + 0.006840150337666273, + 0.004161375109106302, + 0.005042780190706253, + 0.006840739399194717, + 0.0074637942016124725, + 0.0037481295876204967, + 0.006894520949572325, + 0.014153745025396347, + 0.014339255169034004, + 0.005185421090573072, + 0.006403465289622545, + 0.002889356343075633, + 0.0029962880071252584, + 0.002296177903190255, + 0.004324004054069519, + 0.0053730309009552, + 0.006639671046286821, + 0.0036687932442873716, + 0.002870876109227538, + 0.0033094945829361677, + 0.00391788687556982, + 0.002283457200974226, + 0.0023506684228777885, + 0.007673661224544048, + 0.005458150990307331, + 0.0023128404282033443, + 0.003767212387174368, + 0.015344101935625076, + 0.017685696482658386, + 0.006913990713655949, + 0.008924025110900402, + 0.01707638055086136, + 0.03174901753664017, + 0.01594569906592369, + 0.014170563779771328, + 0.0072195944376289845, + 0.011174461804330349, + 0.006321716122329235, + 0.009873890317976475, + 0.011437318287789822, + 0.014449402689933777, + 0.008870340883731842, + 0.009528395719826221, + 0.02099703811109066, + 0.036917541176080704, + 0.018516870215535164, + 0.027023011818528175, + 0.023798905313014984, + 0.01860804297029972, + 0.025581205263733864, + 0.027004241943359375, + 0.0354958213865757, + 0.0628729909658432, + 0.009258209727704525, + 0.014502673409879208, + 0.014210106804966927, + 0.024012073874473572, + 0.014217901974916458, + 0.01315389946103096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.004096450749784708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001537084579468 + ], + "opposite_logprob_tokens": [ + -2.6950154304504395 + ], + "logprob_reward": -0.07001537084579468, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001537084579468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06279443949460983, + 0.023898474872112274, + 0.021166354417800903, + 0.019594883546233177, + 0.02531140111386776, + 0.03728363290429115, + 0.24152004718780518, + 0.11405041068792343, + 0.30746039748191833, + 0.14691995084285736 + ], + "credit_map": [ + 0.004364013206213713, + 0.007498546969145536, + 0.0035217406693845987, + 0.006016320548951626, + 0.013085179962217808, + 0.015292055904865265, + 0.005576142575591803, + 0.007440438028424978, + 0.0022705302108079195, + 0.0024196403101086617, + 0.00268819322809577, + 0.002687769941985607, + 0.0035412872675806284, + 0.005516657140105963, + 0.0024478293489664793, + 0.0023265653289854527, + 0.0031490428373217583, + 0.002110124099999666, + 0.0017655210103839636, + 0.0026786960661411285, + 0.0029446063563227654, + 0.004050016403198242, + 0.0021978402510285378, + 0.0022705073934048414, + 0.002216234803199768, + 0.001558201853185892, + 0.0019026568625122309, + 0.004035153891891241, + 0.0025054432917386293, + 0.002535315230488777, + 0.0025858632288873196, + 0.0022560141514986753, + 0.001944471849128604, + 0.004068712703883648, + 0.002647948917001486, + 0.004415202420204878, + 0.003377961926162243, + 0.0032545628491789103, + 0.0037727539893239737, + 0.001829788088798523, + 0.003485928988084197, + 0.0055172862485051155, + 0.0028537290636450052, + 0.0034425861667841673, + 0.004988388624042273, + 0.01006824430078268, + 0.003411969868466258, + 0.0035155017394572496, + 0.026617716997861862, + 0.033949848264455795, + 0.014143290929496288, + 0.01837484911084175, + 0.028475472703576088, + 0.05732036381959915, + 0.034275881946086884, + 0.02836262807250023, + 0.008718256838619709, + 0.017058752477169037, + 0.010261177085340023, + 0.014615446329116821, + 0.017486991360783577, + 0.02387412264943123, + 0.01126122660934925, + 0.010774449445307255, + 0.03375578299164772, + 0.05605977773666382, + 0.018318209797143936, + 0.030563978478312492, + 0.023188846185803413, + 0.03595879673957825, + 0.03985923156142235, + 0.06975577026605606, + 0.02380209229886532, + 0.040888961404561996, + 0.009159539826214314, + 0.014749963767826557, + 0.012619543820619583, + 0.016523126512765884, + 0.015781838446855545, + 0.013394884765148163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.07001537084579468 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306225851178169 + ], + "opposite_logprob_tokens": [ + -3.168062210083008 + ], + "logprob_reward": -0.04306225851178169, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306225851178169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05302444100379944, + 0.05080472677946091, + 0.0252035241574049, + 0.03030623309314251, + 0.027592815458774567, + 0.05018848925828934, + 0.14984562993049622, + 0.19769364595413208, + 0.14201819896697998, + 0.2733222544193268 + ], + "credit_map": [ + 0.004707625135779381, + 0.005175356287509203, + 0.00557083822786808, + 0.008914798498153687, + 0.007730159908533096, + 0.008604764938354492, + 0.005543772131204605, + 0.006777129601687193, + 0.0056461733765900135, + 0.00578866433352232, + 0.003955396823585033, + 0.003337186062708497, + 0.009416811168193817, + 0.01124928891658783, + 0.004792448598891497, + 0.006618760991841555, + 0.002672802656888962, + 0.002676971023902297, + 0.001941372756846249, + 0.0031722928397357464, + 0.0034602752421051264, + 0.004713142290711403, + 0.0027255110908299685, + 0.003841157304123044, + 0.004043778870254755, + 0.002714066533371806, + 0.0024389999452978373, + 0.0028119415510445833, + 0.005868406035006046, + 0.0056577506475150585, + 0.003362666117027402, + 0.0034086250234395266, + 0.001671782461926341, + 0.00312059186398983, + 0.0036216643638908863, + 0.004997546318918467, + 0.004022551700472832, + 0.004228154197335243, + 0.0027303192764520645, + 0.003200206672772765, + 0.00495904590934515, + 0.006528276950120926, + 0.003127780742943287, + 0.004201748874038458, + 0.009823145344853401, + 0.012288175523281097, + 0.0038430867716670036, + 0.005417235195636749, + 0.015430808998644352, + 0.020855259150266647, + 0.007677210960537195, + 0.014338788576424122, + 0.016207654029130936, + 0.03394334390759468, + 0.023088691756129265, + 0.018303876742720604, + 0.020216310396790504, + 0.03099513240158558, + 0.0163886621594429, + 0.014597168192267418, + 0.04720940440893173, + 0.038426656275987625, + 0.014822331257164478, + 0.015038002282381058, + 0.011241813190281391, + 0.017342546954751015, + 0.014618366956710815, + 0.026369206607341766, + 0.018792500719428062, + 0.011663099750876427, + 0.019381290301680565, + 0.022609379142522812, + 0.0588667169213295, + 0.10585305839776993, + 0.010616320185363293, + 0.016142023727297783, + 0.017984561622142792, + 0.03840072825551033, + 0.01292566116899252, + 0.01253318227827549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.04306225851178169 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot focuses precisely on a pair of hands with vibrant skin and distinct knuckles, carefully cradling a square, layered French mille-feuille pastry. The pastry is evenly coated with a thick, delicate layer of snow-white icing. Warm afternoon sunlight streams through the window on the left, casting soft golden outlines on the pastry and the backs of the hands. In the background, a wooden kitchen counter and a partially blurred white tiled wall are softly blurred. As another pair of equally warm hands slowly reaches in from the right, ready to gently receive this sweet gift, the gentle touch and exchange of hands slightly disturbs the top layer of powdered sugar on the pastry. At this moment, the scene enters an extreme slow motion effect, with countless tiny particles of powdered sugar rising up and dancing like shimmering stardust in the golden light penetrating the kitchen, each particle clearly visible before slowly dispersing. The main colors of the entire picture are warm beige and wood color, with soft light and shadow, full of tranquility, healing comfort, creating a pure and warm atmosphere.", + "seed": 13535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_8/index_7245_seed_13535_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04487700667232275, + "vlm_logprob": -0.04487700667232275, + "vlm_yes_no_margin": 3.3749999523162844, + "vlm_token_credit": -0.04487700667232275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04445001110434532 + ], + "opposite_logprob_tokens": [ + -4.044449806213379 + ], + "logprob_reward": -0.04445001110434532, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.04445001110434532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12522336840629578, + 0.02105838432908058, + 0.011410645209252834, + 0.020997360348701477, + 0.026762772351503372, + 0.02339509315788746, + 0.3459162414073944, + 0.31507962942123413, + 0.03728238865733147, + 0.07287411391735077 + ], + "credit_map": [ + 0.009228846989572048, + 0.013759348541498184, + 0.0210152305662632, + 0.02163483388721943, + 0.012457424774765968, + 0.01782272197306156, + 0.018716147169470787, + 0.010588811710476875, + 0.002787089440971613, + 0.0036034206859767437, + 0.002034547505900264, + 0.0020515266805887222, + 0.0022605638951063156, + 0.0030006698798388243, + 0.003169847186654806, + 0.002150719752535224, + 0.0015949252992868423, + 0.0018766751745715737, + 0.0010089683346450329, + 0.0010078820632770658, + 0.0017064684070646763, + 0.001596471294760704, + 0.0012303919065743685, + 0.0013888622634112835, + 0.005867357831448317, + 0.0026602528523653746, + 0.0013486872194334865, + 0.0017882183892652392, + 0.0016381453024223447, + 0.0032631629146635532, + 0.00199749693274498, + 0.0024340390227735043, + 0.006103273946791887, + 0.0040124766528606415, + 0.0021628523245453835, + 0.0035598932299762964, + 0.002638459438458085, + 0.003922417759895325, + 0.002358590252697468, + 0.002004808746278286, + 0.0034892328549176455, + 0.003046610625460744, + 0.002679743804037571, + 0.0029785409569740295, + 0.002095324918627739, + 0.003317652503028512, + 0.0025564844254404306, + 0.0032315044663846493, + 0.021139629185199738, + 0.04424212500452995, + 0.01001691073179245, + 0.008365952409803867, + 0.020979607477784157, + 0.1519778072834015, + 0.012578885070979595, + 0.0766153335571289, + 0.07826574146747589, + 0.13309746980667114, + 0.00821534264832735, + 0.010712414048612118, + 0.026023849844932556, + 0.033279478549957275, + 0.009211678057909012, + 0.016273627057671547, + 0.004146359395235777, + 0.003867735154926777, + 0.003845903789624572, + 0.0059111109003424644, + 0.004716069437563419, + 0.006059205159544945, + 0.0036842250265181065, + 0.005051779095083475, + 0.014537760987877846, + 0.007085824850946665, + 0.005114729516208172, + 0.017884843051433563, + 0.005246864631772041, + 0.0056269061751663685, + 0.007520337589085102, + 0.00985685084015131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.04445001110434532 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209477782249451 + ], + "opposite_logprob_tokens": [ + -2.8120946884155273 + ], + "logprob_reward": -0.06209477782249451, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209477782249451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11276040971279144, + 0.014813638292253017, + 0.03601767122745514, + 0.02042418159544468, + 0.0076672849245369434, + 0.014925485476851463, + 0.39282888174057007, + 0.3120742440223694, + 0.036643121391534805, + 0.051845040172338486 + ], + "credit_map": [ + 0.007763168308883905, + 0.013164308853447437, + 0.017310112714767456, + 0.01837521232664585, + 0.011376678012311459, + 0.016203584149479866, + 0.018171383067965508, + 0.010395974852144718, + 0.0013395067071542144, + 0.0026779125910252333, + 0.0018233557930216193, + 0.0017225835472345352, + 0.001348858349956572, + 0.0017691318644210696, + 0.002666657790541649, + 0.001465632813051343, + 0.00754280062392354, + 0.004542108625173569, + 0.0036119320429861546, + 0.002839642111212015, + 0.00440341979265213, + 0.005452836863696575, + 0.003624193137511611, + 0.0040007405914366245, + 0.005863696336746216, + 0.0029616174288094044, + 0.0013921037316322327, + 0.0013947748811915517, + 0.0016274758381769061, + 0.0037082056514918804, + 0.0014765913365408778, + 0.001999719301238656, + 0.0011050630128011107, + 0.0009762162808328867, + 0.0006659370847046375, + 0.0012544173514470458, + 0.0007377215661108494, + 0.0009222976514138281, + 0.0011365768732503057, + 0.0008690556860528886, + 0.0014294234570115805, + 0.002207099227234721, + 0.0012924748007208109, + 0.00230409181676805, + 0.001676372718065977, + 0.0020047740545123816, + 0.0024252133443951607, + 0.0015860367566347122, + 0.02347104623913765, + 0.050280939787626266, + 0.01089667621999979, + 0.010445713996887207, + 0.024003397673368454, + 0.1713927835226059, + 0.014849090948700905, + 0.0874892845749855, + 0.07254498451948166, + 0.13147591054439545, + 0.007981548085808754, + 0.01352379098534584, + 0.02561194822192192, + 0.033736452460289, + 0.010139767080545425, + 0.01705985702574253, + 0.0037640309892594814, + 0.0038734753616154194, + 0.0037818029522895813, + 0.0074393837712705135, + 0.00315056461840868, + 0.005806971341371536, + 0.0037276712246239185, + 0.005099225323647261, + 0.0072138127870857716, + 0.005350599065423012, + 0.006407096050679684, + 0.007257458288222551, + 0.005764645058661699, + 0.006051542703062296, + 0.0061353472992777824, + 0.0076645431108772755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.06209477782249451 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22956600785255432, + 0.029731901362538338, + 0.044561561197042465, + 0.04065261408686638, + 0.038265787065029144, + 0.06727366149425507, + 0.1910005807876587, + 0.23105134069919586, + 0.05257618799805641, + 0.07532042264938354 + ], + "credit_map": [ + 0.029625194147229195, + 0.04140602797269821, + 0.03165512531995773, + 0.043008953332901, + 0.011262425221502781, + 0.03512314707040787, + 0.020292475819587708, + 0.017192644998431206, + 0.0055847386829555035, + 0.005711181554943323, + 0.0022481547202914953, + 0.003053348045796156, + 0.002859017811715603, + 0.00427648238837719, + 0.0035220677964389324, + 0.0024769101291894913, + 0.007026481907814741, + 0.010230747982859612, + 0.003398801200091839, + 0.0038311961106956005, + 0.003914887085556984, + 0.006154296454042196, + 0.004316691774874926, + 0.005688460078090429, + 0.00576030882075429, + 0.006658698432147503, + 0.002935438882559538, + 0.007417409215122461, + 0.004202141426503658, + 0.005950942635536194, + 0.0038537157233804464, + 0.0038739570882171392, + 0.008236429654061794, + 0.0055625238455832005, + 0.003766290843486786, + 0.003196666482836008, + 0.0037093844730407, + 0.006559197790920734, + 0.003386617172509432, + 0.0038486747071146965, + 0.008874303661286831, + 0.010911387391388416, + 0.004393009468913078, + 0.00852400902658701, + 0.007662651129066944, + 0.008658990263938904, + 0.010530713945627213, + 0.007718597538769245, + 0.013146760873496532, + 0.023012038320302963, + 0.007641934789717197, + 0.011999999172985554, + 0.01274827029556036, + 0.07271108031272888, + 0.011747200973331928, + 0.03799330070614815, + 0.0570882223546505, + 0.09480918198823929, + 0.006763764191418886, + 0.007856886833906174, + 0.017430834472179413, + 0.024779103696346283, + 0.00888359360396862, + 0.013439755886793137, + 0.006398509256541729, + 0.007300641853362322, + 0.003837724681943655, + 0.005364913027733564, + 0.014927289448678493, + 0.006593597587198019, + 0.003832013811916113, + 0.004321498330682516, + 0.010130558162927628, + 0.00968096125870943, + 0.003203359665349126, + 0.017447911202907562, + 0.003290027379989624, + 0.004836308304220438, + 0.005440505221486092, + 0.02129078470170498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01420364435762167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867158457636833 + ], + "opposite_logprob_tokens": [ + -3.0486714839935303 + ], + "logprob_reward": -0.04867158457636833, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867158457636833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0691896602511406, + 0.01964564621448517, + 0.019760290160775185, + 0.030001454055309296, + 0.015371808782219887, + 0.033841442316770554, + 0.2340727150440216, + 0.2861253619194031, + 0.11665501445531845, + 0.17533662915229797 + ], + "credit_map": [ + 0.005945492535829544, + 0.008486572653055191, + 0.009833840653300285, + 0.010629037395119667, + 0.010139771737158298, + 0.009752456098794937, + 0.008575450628995895, + 0.005827036220580339, + 0.0024763550609350204, + 0.002885193796828389, + 0.00189546262845397, + 0.0029041252564638853, + 0.0024056173861026764, + 0.002300747437402606, + 0.002949375193566084, + 0.0018287681741639972, + 0.0047559356316924095, + 0.0029381467029452324, + 0.0014354620361700654, + 0.001707854913547635, + 0.002118361182510853, + 0.003096275730058551, + 0.0014130553463473916, + 0.0022951981518417597, + 0.00820823572576046, + 0.004221382550895214, + 0.002162097254768014, + 0.0030243764631450176, + 0.002016473561525345, + 0.004858557600528002, + 0.002501616021618247, + 0.003008713247254491, + 0.0015511688543483615, + 0.001836408511735499, + 0.0010042544454336166, + 0.0026601378340274096, + 0.001868615159764886, + 0.0023000515066087246, + 0.0022950239945203066, + 0.0018561480101197958, + 0.0043298834934830666, + 0.0054840329103171825, + 0.002400952158495784, + 0.004180059302598238, + 0.00303953536786139, + 0.004945181310176849, + 0.004174881149083376, + 0.005286916624754667, + 0.015120117925107479, + 0.02943548373878002, + 0.0072996485978364944, + 0.008284516632556915, + 0.015626179054379463, + 0.09907311946153641, + 0.009474650956690311, + 0.04975901171565056, + 0.06706273555755615, + 0.11554034054279327, + 0.008760083466768265, + 0.012145351618528366, + 0.021707428619265556, + 0.030391067266464233, + 0.009739711880683899, + 0.020778629928827286, + 0.013020634651184082, + 0.013065940700471401, + 0.012510542757809162, + 0.017332706600427628, + 0.0168985016644001, + 0.017997117713093758, + 0.011778646148741245, + 0.014050926081836224, + 0.01796541176736355, + 0.015438801608979702, + 0.021991316229104996, + 0.04385460540652275, + 0.014987476170063019, + 0.017435863614082336, + 0.017674343660473824, + 0.025988800451159477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.04867158457636833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496501550078392 + ], + "opposite_logprob_tokens": [ + -2.929965019226074 + ], + "logprob_reward": -0.05496501550078392, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496501550078392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07339520752429962, + 0.022142130881547928, + 0.0493200458586216, + 0.04761580750346184, + 0.027792643755674362, + 0.040290795266628265, + 0.4133917987346649, + 0.07583871483802795, + 0.19176125526428223, + 0.05845162272453308 + ], + "credit_map": [ + 0.006968371570110321, + 0.008505369536578655, + 0.008615422062575817, + 0.009723372757434845, + 0.009665437042713165, + 0.009837922640144825, + 0.011389817111194134, + 0.008689496666193008, + 0.002929236274212599, + 0.004798101261258125, + 0.0021215963643044233, + 0.0032736919820308685, + 0.002079492434859276, + 0.0029084477573633194, + 0.002287812065333128, + 0.001743751927278936, + 0.013359125703573227, + 0.0071325525641441345, + 0.003737601451575756, + 0.003977352753281593, + 0.0035498884972184896, + 0.007822819985449314, + 0.004041551612317562, + 0.00569915259256959, + 0.013752066530287266, + 0.0063856071792542934, + 0.002626595553010702, + 0.003841114230453968, + 0.004190817475318909, + 0.008076882921159267, + 0.004332089796662331, + 0.004410634748637676, + 0.0039467038586735725, + 0.003920051269233227, + 0.0024949528742581606, + 0.00326681905426085, + 0.0031124919187277555, + 0.0038726909551769495, + 0.004179251845926046, + 0.002999682677909732, + 0.0070613957941532135, + 0.0068265460431575775, + 0.00320110353641212, + 0.0035914054606109858, + 0.004886300303041935, + 0.005179496016353369, + 0.004636341240257025, + 0.004908206406980753, + 0.025528330355882645, + 0.054069943726062775, + 0.01196771115064621, + 0.010011812672019005, + 0.024518582969903946, + 0.18125879764556885, + 0.014719880186021328, + 0.09131670743227005, + 0.01602441817522049, + 0.0221506766974926, + 0.0036619880702346563, + 0.0055968849919736385, + 0.007580567616969347, + 0.009823705069720745, + 0.004813747946172953, + 0.006186732091009617, + 0.012941590510308743, + 0.015807420015335083, + 0.006186106242239475, + 0.06545495241880417, + 0.05517927557229996, + 0.008424580097198486, + 0.008341898210346699, + 0.01942543126642704, + 0.013730446808040142, + 0.0068656462244689465, + 0.0035124067217111588, + 0.010471208952367306, + 0.0036492382641881704, + 0.0045968834310770035, + 0.006407261360436678, + 0.009218533523380756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.05496501550078392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot focuses precisely on a pair of hands with vibrant skin and distinct knuckles, carefully cradling a square, layered French mille-feuille pastry. The pastry is evenly coated with a thick, delicate layer of snow-white icing. Warm afternoon sunlight streams through the window on the left, casting soft golden outlines on the pastry and the backs of the hands. In the background, a wooden kitchen counter and a partially blurred white tiled wall are softly blurred. As another pair of equally warm hands slowly reaches in from the right, ready to gently receive this sweet gift, the gentle touch and exchange of hands slightly disturbs the top layer of powdered sugar on the pastry. At this moment, the scene enters an extreme slow motion effect, with countless tiny particles of powdered sugar rising up and dancing like shimmering stardust in the golden light penetrating the kitchen, each particle clearly visible before slowly dispersing. The main colors of the entire picture are warm beige and wood color, with soft light and shadow, full of tranquility, healing comfort, creating a pure and warm atmosphere.", + "seed": 13536, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_8/index_7245_seed_13536_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04308515433222056, + "vlm_logprob": -0.04308515433222056, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04308515433222056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10940882563591003 + ], + "opposite_logprob_tokens": [ + -2.4844088554382324 + ], + "logprob_reward": -0.10940882563591003, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10940882563591003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08893083781003952, + 0.159652441740036, + 0.07282734662294388, + 0.08480725437402725, + 0.05846317484974861, + 0.22460868954658508, + 0.06692638993263245, + 0.08996962755918503, + 0.03373706713318825, + 0.1200772225856781 + ], + "credit_map": [ + 0.00817506667226553, + 0.015549687668681145, + 0.007266080006957054, + 0.011805971153080463, + 0.012512250803411007, + 0.01231971476227045, + 0.010671508498489857, + 0.01063055731356144, + 0.009589357301592827, + 0.02109600603580475, + 0.01083252765238285, + 0.01873401366174221, + 0.03148598223924637, + 0.03971491754055023, + 0.012660695239901543, + 0.015538923442363739, + 0.006211092695593834, + 0.01209903322160244, + 0.006165453232824802, + 0.009150543250143528, + 0.010799364186823368, + 0.009248899295926094, + 0.012283828109502792, + 0.006869122385978699, + 0.0070466031320393085, + 0.011218656785786152, + 0.007526871282607317, + 0.010653333738446236, + 0.011993518099188805, + 0.009972143918275833, + 0.01767553947865963, + 0.008720580488443375, + 0.006256415043026209, + 0.006967340130358934, + 0.006864440627396107, + 0.007113120052963495, + 0.008285416290163994, + 0.009463712573051453, + 0.007050695829093456, + 0.0064620268531143665, + 0.017314404249191284, + 0.02958172932267189, + 0.017346758395433426, + 0.02521822601556778, + 0.03481336683034897, + 0.04788980260491371, + 0.02511933073401451, + 0.027325045317411423, + 0.006018619053065777, + 0.007503821048885584, + 0.0055097006261348724, + 0.00959731638431549, + 0.008994621224701405, + 0.009784596040844917, + 0.010975505225360394, + 0.008542202413082123, + 0.0059092347510159016, + 0.011439514346420765, + 0.009401463903486729, + 0.013934636488556862, + 0.012968294322490692, + 0.01252814382314682, + 0.012152413837611675, + 0.01163591630756855, + 0.003557262010872364, + 0.005040425807237625, + 0.00265026674605906, + 0.0032636099494993687, + 0.004863926209509373, + 0.003732156241312623, + 0.006003230810165405, + 0.0046261874958872795, + 0.015252060256898403, + 0.012269461527466774, + 0.012155733071267605, + 0.01394672878086567, + 0.01694732904434204, + 0.01340926717966795, + 0.01671823114156723, + 0.019378406926989555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.10940882563591003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609283819794655 + ], + "opposite_logprob_tokens": [ + -3.8956093788146973 + ], + "logprob_reward": -0.020609283819794655, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609283819794655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19075310230255127, + 0.10172471404075623, + 0.049351293593645096, + 0.05205495283007622, + 0.07028617709875107, + 0.10800818353891373, + 0.09003163874149323, + 0.08829116821289062, + 0.06322041898965836, + 0.1862783282995224 + ], + "credit_map": [ + 0.01575026474893093, + 0.028186704963445663, + 0.017221910879015923, + 0.026883874088525772, + 0.026176517829298973, + 0.023158233612775803, + 0.030663730576634407, + 0.022711871191859245, + 0.007265279069542885, + 0.0126488683745265, + 0.011082927696406841, + 0.013594095595180988, + 0.015569359995424747, + 0.014922045171260834, + 0.016411790624260902, + 0.010230348445475101, + 0.00656888447701931, + 0.006053147837519646, + 0.004358629696071148, + 0.005825018044561148, + 0.008014841005206108, + 0.007496395148336887, + 0.004963906016200781, + 0.006070474628359079, + 0.0038316473364830017, + 0.008235647343099117, + 0.007063461001962423, + 0.007041954435408115, + 0.0065931882709264755, + 0.006101679522544146, + 0.006574421655386686, + 0.006612951867282391, + 0.00789882242679596, + 0.007869303226470947, + 0.012095394544303417, + 0.006998037453740835, + 0.00886981375515461, + 0.009814834222197533, + 0.00702459504827857, + 0.009715369902551174, + 0.007318808697164059, + 0.0111925657838583, + 0.01048439834266901, + 0.01070377230644226, + 0.015604936517775059, + 0.023463303223252296, + 0.012206644751131535, + 0.017033761367201805, + 0.006839417852461338, + 0.008879086002707481, + 0.011496171355247498, + 0.01495556440204382, + 0.012377328239381313, + 0.011219494976103306, + 0.011810201220214367, + 0.012454373762011528, + 0.006634393706917763, + 0.009338595904409885, + 0.008178622461855412, + 0.012992601841688156, + 0.013137570582330227, + 0.013809198513627052, + 0.011753331869840622, + 0.012446852400898933, + 0.005272200331091881, + 0.006733722984790802, + 0.005211178213357925, + 0.007418384775519371, + 0.0069300034083426, + 0.011304909363389015, + 0.010384262539446354, + 0.009965754114091396, + 0.018863363191485405, + 0.014618981629610062, + 0.02169162407517433, + 0.01838270202279091, + 0.026166241616010666, + 0.022717386484146118, + 0.038143571466207504, + 0.025694454088807106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.020609283819794655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544064782559872 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.012544064782559872, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544064782559872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035280562937259674, + 0.772088348865509, + 0.023324323818087578, + 0.042262520641088486, + 0.024534093216061592, + 0.02069230191409588, + 0.012508642859756947, + 0.0276900976896286, + 0.01321207731962204, + 0.02840711548924446 + ], + "credit_map": [ + 0.0036607205402106047, + 0.005207891575992107, + 0.0025909319519996643, + 0.0023700890596956015, + 0.007243286818265915, + 0.008309721946716309, + 0.0026138038374483585, + 0.003284112084656954, + 0.04698896035552025, + 0.10051098465919495, + 0.05397140607237816, + 0.08677057921886444, + 0.15788237750530243, + 0.1934690922498703, + 0.057236429303884506, + 0.07525843381881714, + 0.002368609420955181, + 0.0036587994545698166, + 0.001421588473021984, + 0.003248178632929921, + 0.0038177964743226767, + 0.0032828873954713345, + 0.00324048544280231, + 0.0022859766613692045, + 0.003775515593588352, + 0.007623034063726664, + 0.003324020653963089, + 0.004047205671668053, + 0.007958345115184784, + 0.007906400598585606, + 0.0038868526462465525, + 0.0037411414086818695, + 0.00246168440207839, + 0.00424626749008894, + 0.0035411762073636055, + 0.0026384727098047733, + 0.0031997920013964176, + 0.0032650171779096127, + 0.0026824534870684147, + 0.0024992264807224274, + 0.0015958643052726984, + 0.0026522709522396326, + 0.002117329044267535, + 0.0019655649084597826, + 0.002948424778878689, + 0.004183609504252672, + 0.0022253338247537613, + 0.003003902966156602, + 0.001206312677823007, + 0.0021677871700376272, + 0.0016689632320776582, + 0.0015841901767998934, + 0.001805274747312069, + 0.0010048975236713886, + 0.0018707193667069077, + 0.0012004970340058208, + 0.0035385945811867714, + 0.006366230547428131, + 0.004028056748211384, + 0.003035652218386531, + 0.0038660704158246517, + 0.0024075855035334826, + 0.002584580099210143, + 0.0018633275758475065, + 0.0011166357435286045, + 0.0025498804170638323, + 0.001400474924594164, + 0.0017359225312247872, + 0.0017406214028596878, + 0.0018836132949218154, + 0.0015527044888585806, + 0.0012322237016633153, + 0.0027078280691057444, + 0.003676172811537981, + 0.0023906033020466566, + 0.0025154314935207367, + 0.003749934723600745, + 0.0050045340321958065, + 0.0039021330885589123, + 0.004460476338863373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.012544064782559872 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802024364471436 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802024364471436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802024364471436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05507633090019226, + 0.22166582942008972, + 0.06432634592056274, + 0.052213530987501144, + 0.03671523183584213, + 0.10715501010417938, + 0.08102158457040787, + 0.0571405291557312, + 0.05225076526403427, + 0.27243486046791077 + ], + "credit_map": [ + 0.004833702463656664, + 0.008152022957801819, + 0.006426671985536814, + 0.006916465703397989, + 0.00853398721665144, + 0.008005240000784397, + 0.006544810254126787, + 0.005663429386913776, + 0.013477888889610767, + 0.029078559949994087, + 0.016897717490792274, + 0.027597077190876007, + 0.04367399588227272, + 0.05390733480453491, + 0.01590038277208805, + 0.021132854744791985, + 0.004064437001943588, + 0.0074477847665548325, + 0.0046464973129332066, + 0.004071615170687437, + 0.01702873595058918, + 0.016482528299093246, + 0.0053930324502289295, + 0.005191713105887175, + 0.0038774602580815554, + 0.0034608840942382812, + 0.005554260220378637, + 0.005766707938164473, + 0.010791073553264141, + 0.008820065297186375, + 0.007041811477392912, + 0.006901272572577, + 0.0036127292551100254, + 0.006544589530676603, + 0.00392339238896966, + 0.0034003856126219034, + 0.005231546238064766, + 0.006262280512601137, + 0.0033514576498419046, + 0.004388853441923857, + 0.006297798827290535, + 0.010940239764750004, + 0.012904576025903225, + 0.017083410173654556, + 0.013642502017319202, + 0.020675988867878914, + 0.011266962625086308, + 0.014343529008328915, + 0.006914183031767607, + 0.011565901339054108, + 0.010227270424365997, + 0.014120607636868954, + 0.010400035418570042, + 0.00913951825350523, + 0.010299619287252426, + 0.008354444988071918, + 0.0042976150289177895, + 0.006798847112804651, + 0.0053017581813037395, + 0.007866975851356983, + 0.008133330382406712, + 0.008322206325829029, + 0.008238241076469421, + 0.008181555196642876, + 0.004868249408900738, + 0.008727881126105785, + 0.0028889712411910295, + 0.00508664408698678, + 0.007285133935511112, + 0.0069182319566607475, + 0.010274868458509445, + 0.006200787145644426, + 0.02572987601161003, + 0.03177136182785034, + 0.028907151892781258, + 0.0241978969424963, + 0.035599321126937866, + 0.04793412983417511, + 0.03396037593483925, + 0.04433473199605942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.029802024364471436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043061573058366776 + ], + "opposite_logprob_tokens": [ + -3.1680614948272705 + ], + "logprob_reward": -0.043061573058366776, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043061573058366776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10243720561265945, + 0.08286135643720627, + 0.0998208299279213, + 0.12448269873857498, + 0.09262637048959732, + 0.2298118770122528, + 0.07810333371162415, + 0.01933969184756279, + 0.09288334101438522, + 0.0776333212852478 + ], + "credit_map": [ + 0.00863938219845295, + 0.01249574963003397, + 0.00919182226061821, + 0.014211723580956459, + 0.01659196987748146, + 0.017982730641961098, + 0.014746399596333504, + 0.008577420376241207, + 0.005985169671475887, + 0.010459409095346928, + 0.006747947074472904, + 0.01218090858310461, + 0.01446918211877346, + 0.017812874168157578, + 0.006636261474341154, + 0.008569608442485332, + 0.0079144686460495, + 0.01526940893381834, + 0.007184971123933792, + 0.007314417511224747, + 0.02297966368496418, + 0.020374024286866188, + 0.011041902005672455, + 0.007741979323327541, + 0.011286169290542603, + 0.0221287589520216, + 0.014651552774012089, + 0.011194348335266113, + 0.020972903817892075, + 0.018751289695501328, + 0.010790031403303146, + 0.014707651920616627, + 0.008325794711709023, + 0.012613538652658463, + 0.010662594810128212, + 0.011425672098994255, + 0.011052180081605911, + 0.013439804315567017, + 0.011710677295923233, + 0.013396102003753185, + 0.016301138326525688, + 0.02908988483250141, + 0.022935498505830765, + 0.027469418942928314, + 0.03328312188386917, + 0.045197173953056335, + 0.028657760471105576, + 0.02687789313495159, + 0.007147501222789288, + 0.011854573152959347, + 0.009998000226914883, + 0.011435475200414658, + 0.011315272189676762, + 0.008815482258796692, + 0.01061646081507206, + 0.006920568645000458, + 0.0016752308001741767, + 0.0024670991115272045, + 0.002387912478297949, + 0.0024379759561270475, + 0.002259865403175354, + 0.002243399852886796, + 0.003632688196375966, + 0.0022355190012604, + 0.007556895259767771, + 0.011241883970797062, + 0.008420856669545174, + 0.01215420849621296, + 0.012750486843287945, + 0.014965731650590897, + 0.014488083310425282, + 0.011305189691483974, + 0.008872314356267452, + 0.009148512035608292, + 0.00674403365701437, + 0.0076216706074774265, + 0.013515839353203773, + 0.011037162505090237, + 0.012526816688477993, + 0.008166976273059845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.043061573058366776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot focuses precisely on a pair of hands with vibrant skin and distinct knuckles, carefully cradling a square, layered French mille-feuille pastry. The pastry is evenly coated with a thick, delicate layer of snow-white icing. Warm afternoon sunlight streams through the window on the left, casting soft golden outlines on the pastry and the backs of the hands. In the background, a wooden kitchen counter and a partially blurred white tiled wall are softly blurred. As another pair of equally warm hands slowly reaches in from the right, ready to gently receive this sweet gift, the gentle touch and exchange of hands slightly disturbs the top layer of powdered sugar on the pastry. At this moment, the scene enters an extreme slow motion effect, with countless tiny particles of powdered sugar rising up and dancing like shimmering stardust in the golden light penetrating the kitchen, each particle clearly visible before slowly dispersing. The main colors of the entire picture are warm beige and wood color, with soft light and shadow, full of tranquility, healing comfort, creating a pure and warm atmosphere.", + "seed": 13537, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_8/index_7245_seed_13537_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.040769101865589616, + "vlm_logprob": -0.040769101865589616, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.040769101865589616, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04242695868015289 + ], + "opposite_logprob_tokens": [ + -3.6674270629882812 + ], + "logprob_reward": -0.04242695868015289, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04242695868015289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25315696001052856, + 0.029233630746603012, + 0.013327536173164845, + 0.012999627739191055, + 0.004920642822980881, + 0.6029860377311707, + 0.015233278274536133, + 0.015122450888156891, + 0.007821389473974705, + 0.04519848898053169 + ], + "credit_map": [ + 0.017876094207167625, + 0.046772100031375885, + 0.048783037811517715, + 0.023302098736166954, + 0.020597174763679504, + 0.01967940293252468, + 0.02249740995466709, + 0.0536496639251709, + 0.0035232510417699814, + 0.003786316607147455, + 0.0038698946591466665, + 0.004494726657867432, + 0.0026428212877362967, + 0.0033588288351893425, + 0.00306481821462512, + 0.004492974374443293, + 0.0012875217944383621, + 0.0024479255080223083, + 0.0021017869003117085, + 0.001534862443804741, + 0.0011419294169172645, + 0.0010820745956152678, + 0.001543916412629187, + 0.0021875181701034307, + 0.0008640949381515384, + 0.002347497269511223, + 0.0022515899036079645, + 0.0016284830635413527, + 0.001307228347286582, + 0.0009128407691605389, + 0.0010089341085404158, + 0.0026789598632603884, + 0.0005672815605066717, + 0.0006920784944668412, + 0.0009045199258252978, + 0.000439780589658767, + 0.0004756726266350597, + 0.0008196458220481873, + 0.00044305642950348556, + 0.0005786077817901969, + 0.05756273493170738, + 0.1465451866388321, + 0.1225171908736229, + 0.02509370446205139, + 0.06373409181833267, + 0.12417379766702652, + 0.0245344378054142, + 0.038824860006570816, + 0.0016631881007924676, + 0.002022483851760626, + 0.002459886483848095, + 0.0014682576293125749, + 0.0015621326165273786, + 0.002963111735880375, + 0.0014353018486872315, + 0.00165891635697335, + 0.0014624135801568627, + 0.0026763773057609797, + 0.002722366014495492, + 0.0017943333368748426, + 0.0013884155778214335, + 0.001450223382562399, + 0.0019985029939562082, + 0.0016298178816214204, + 0.0006202366785146296, + 0.001340663293376565, + 0.0011554951779544353, + 0.0008790029678493738, + 0.001038078567944467, + 0.0007665182347409427, + 0.0010023282375186682, + 0.001019067014567554, + 0.003182864049449563, + 0.013538520783185959, + 0.015511590987443924, + 0.003321316558867693, + 0.0021983899641782045, + 0.002769794315099716, + 0.0027512998785823584, + 0.0019247106974944472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.04242695868015289 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007656513713300228 + ], + "opposite_logprob_tokens": [ + -4.882656574249268 + ], + "logprob_reward": -0.007656513713300228, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007656513713300228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2850511968135834, + 0.24460534751415253, + 0.06961619853973389, + 0.08365830779075623, + 0.04340999200940132, + 0.09712562710046768, + 0.03389860689640045, + 0.028167279437184334, + 0.01993156224489212, + 0.09453591704368591 + ], + "credit_map": [ + 0.01970178447663784, + 0.048717230558395386, + 0.05463444069027901, + 0.024184267967939377, + 0.021123092621564865, + 0.02866690419614315, + 0.022126082330942154, + 0.06589742004871368, + 0.02902268059551716, + 0.03304172307252884, + 0.03219391405582428, + 0.03414241969585419, + 0.02277839183807373, + 0.02922648750245571, + 0.023932723328471184, + 0.040267009288072586, + 0.0085751973092556, + 0.013086879625916481, + 0.012042501010000706, + 0.005538361147046089, + 0.007035824004560709, + 0.008796356618404388, + 0.0058199861086905, + 0.008721085265278816, + 0.006073391065001488, + 0.015060018748044968, + 0.014291222207248211, + 0.009162526577711105, + 0.007540971506386995, + 0.006070632953196764, + 0.0065360404551029205, + 0.018923502415418625, + 0.0028943491633981466, + 0.005538654047995806, + 0.006737021263688803, + 0.0037655034102499485, + 0.005098555237054825, + 0.003940691705793142, + 0.004354167263954878, + 0.011081050150096416, + 0.010893261060118675, + 0.011806207709014416, + 0.013722156174480915, + 0.011053625494241714, + 0.010885548777878284, + 0.01699029840528965, + 0.007334496825933456, + 0.014440033584833145, + 0.006051226053386927, + 0.004090332891792059, + 0.00554530369117856, + 0.003211531089618802, + 0.003417754778638482, + 0.004926052875816822, + 0.003610155312344432, + 0.0030462497379630804, + 0.0036948593333363533, + 0.004209584556519985, + 0.0043876864947378635, + 0.0019849101081490517, + 0.0037057045847177505, + 0.003353666514158249, + 0.003953350707888603, + 0.0028775169048458338, + 0.001927600591443479, + 0.003726644441485405, + 0.0029257566202431917, + 0.002558114705607295, + 0.002152839209884405, + 0.0024587863590568304, + 0.0023867974523454905, + 0.0017950221663340926, + 0.0062401192262768745, + 0.02904958836734295, + 0.033149391412734985, + 0.006852095481008291, + 0.005276115145534277, + 0.005771125201135874, + 0.004544011317193508, + 0.003653476247563958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.007656513713300228 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073880717158318 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073880717158318, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073880717158318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18777433037757874, + 0.11545882374048233, + 0.08206116408109665, + 0.00871692132204771, + 0.01423877663910389, + 0.4628508687019348, + 0.028766801580786705, + 0.04613630473613739, + 0.011505050584673882, + 0.04249093309044838 + ], + "credit_map": [ + 0.015107895247638226, + 0.0329909585416317, + 0.032329775393009186, + 0.023499276489019394, + 0.019844159483909607, + 0.013472400605678558, + 0.01849566213786602, + 0.03203418478369713, + 0.013220255263149738, + 0.017059089615941048, + 0.01610262133181095, + 0.017124418169260025, + 0.009547919034957886, + 0.011914998292922974, + 0.011479312554001808, + 0.019010210409760475, + 0.007038780953735113, + 0.01510821096599102, + 0.016532093286514282, + 0.0066064391285181046, + 0.007812673225998878, + 0.007866373285651207, + 0.006973844487220049, + 0.014122744090855122, + 0.0010486261453479528, + 0.0011471955804154277, + 0.001335518667474389, + 0.0009634066373109818, + 0.001079427427612245, + 0.0013834815472364426, + 0.000512635160703212, + 0.001246630446985364, + 0.0014148663030937314, + 0.0017923690611496568, + 0.002516576088964939, + 0.001568336971104145, + 0.001720812520943582, + 0.0018558211158961058, + 0.0011629421496763825, + 0.002207052195444703, + 0.04014316201210022, + 0.10559341311454773, + 0.09243331849575043, + 0.02347657084465027, + 0.04719776287674904, + 0.08923768997192383, + 0.021483371034264565, + 0.043285585939884186, + 0.0029641222208738327, + 0.0038626212626695633, + 0.0051353126764297485, + 0.0029589568730443716, + 0.004046328365802765, + 0.0039176722057163715, + 0.0029361783526837826, + 0.0029456098563969135, + 0.0062358807772397995, + 0.007543662562966347, + 0.008115879260003567, + 0.004369179252535105, + 0.0038177315145730972, + 0.005038261413574219, + 0.0050943768583238125, + 0.0059213340282440186, + 0.0015995235880836844, + 0.0019199354574084282, + 0.0016615670174360275, + 0.0011086539598181844, + 0.0017197694396600127, + 0.0010925154201686382, + 0.0010272867511957884, + 0.0013757998822256923, + 0.0031900506000965834, + 0.012605293653905392, + 0.013899369165301323, + 0.0030041737481951714, + 0.0024478561244904995, + 0.0033556141424924135, + 0.001863372977823019, + 0.00212520407512784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.016073880717158318 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864989593625069 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864989593625069, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864989593625069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40762704610824585, + 0.20452353358268738, + 0.08562511205673218, + 0.035702191293239594, + 0.040802668780088425, + 0.09989292919635773, + 0.04188926890492439, + 0.04130389541387558, + 0.02207985147833824, + 0.020553532987833023 + ], + "credit_map": [ + 0.028916263952851295, + 0.07684647291898727, + 0.07888839393854141, + 0.04299222677946091, + 0.03415453061461449, + 0.02788618952035904, + 0.0369153656065464, + 0.08102761954069138, + 0.0258425772190094, + 0.026434842497110367, + 0.027146777138113976, + 0.0320158451795578, + 0.015167501755058765, + 0.02082589641213417, + 0.023454030975699425, + 0.033636052161455154, + 0.007589097134768963, + 0.01571488566696644, + 0.015508887358009815, + 0.0084683233872056, + 0.008891199715435505, + 0.007839088328182697, + 0.006001060362905264, + 0.01561256404966116, + 0.0025161802768707275, + 0.005377770867198706, + 0.00635672640055418, + 0.004136250354349613, + 0.003638245165348053, + 0.002598195569589734, + 0.002906970912590623, + 0.008171849884092808, + 0.0021583004854619503, + 0.004414435476064682, + 0.0076168738305568695, + 0.004958948586136103, + 0.004201038740575314, + 0.0026039108633995056, + 0.003585440805181861, + 0.011263720691204071, + 0.010691974312067032, + 0.026663128286600113, + 0.02058645710349083, + 0.0038892312441021204, + 0.009646521881222725, + 0.019544662907719612, + 0.004029615316540003, + 0.004841337911784649, + 0.003264722414314747, + 0.0057072420604527, + 0.0061305006965994835, + 0.00799240916967392, + 0.0055181849747896194, + 0.003360518952831626, + 0.005535943899303675, + 0.004379746969789267, + 0.003903063712641597, + 0.005389143247157335, + 0.006313296500593424, + 0.003806634107604623, + 0.0036726200487464666, + 0.007734703365713358, + 0.003822578117251396, + 0.006661855615675449, + 0.001804826082661748, + 0.0037818418350070715, + 0.003257764969021082, + 0.0026682845782488585, + 0.0023658890277147293, + 0.0018290493171662092, + 0.003294646507129073, + 0.003077548462897539, + 0.001911831903271377, + 0.004910875111818314, + 0.005670284852385521, + 0.0016812999965623021, + 0.0011819718638435006, + 0.0017206821357831359, + 0.001953235361725092, + 0.0015233511803671718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04864989593625069 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a pair of hands with vibrant skin and distinct knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square, layered French mille-feuille pastry coated with thick snow-white icing visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include warm afternoon sunlight streaming from the left, casting golden outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do two pairs of hands exchange the pastry, causing powdered sugar particles to rise in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene with a blurred background of a wooden counter and white tiled wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903826028108597 + ], + "opposite_logprob_tokens": [ + -2.464038372039795 + ], + "logprob_reward": -0.08903826028108597, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903826028108597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2757699191570282, + 0.12024061381816864, + 0.12745751440525055, + 0.03462444245815277, + 0.032483913004398346, + 0.16197186708450317, + 0.14020012319087982, + 0.05375571548938751, + 0.02086314558982849, + 0.0326327383518219 + ], + "credit_map": [ + 0.021064680069684982, + 0.05045275017619133, + 0.050136446952819824, + 0.03441993519663811, + 0.028511788696050644, + 0.018926287069916725, + 0.02601579949259758, + 0.04624227061867714, + 0.013724349439144135, + 0.019064869731664658, + 0.017810329794883728, + 0.016552595421671867, + 0.01019206177443266, + 0.011665744706988335, + 0.010764011181890965, + 0.02046666294336319, + 0.010411106050014496, + 0.023376623168587685, + 0.023706521838903427, + 0.014079397544264793, + 0.01291877031326294, + 0.009753212332725525, + 0.009534227661788464, + 0.023677654564380646, + 0.0030784481205046177, + 0.0047003403306007385, + 0.005131290294229984, + 0.0041390592232346535, + 0.003989208023995161, + 0.003741943510249257, + 0.0025188373401761055, + 0.00732531538233161, + 0.003017893759533763, + 0.004071445669978857, + 0.005262632388621569, + 0.0030380338430404663, + 0.004413785878568888, + 0.0051519074477255344, + 0.0023796812165528536, + 0.005148532800376415, + 0.010302312672138214, + 0.041295431554317474, + 0.038504041731357574, + 0.007993658073246479, + 0.014830841682851315, + 0.025644974783062935, + 0.008726098574697971, + 0.014674510806798935, + 0.016121823340654373, + 0.02443336881697178, + 0.020672054961323738, + 0.023318680003285408, + 0.014581138268113136, + 0.011370333842933178, + 0.01480837445706129, + 0.01489434577524662, + 0.0059356666170060635, + 0.010373007506132126, + 0.008137671276926994, + 0.004184048157185316, + 0.004852887708693743, + 0.00840853713452816, + 0.004669611342251301, + 0.007194290868937969, + 0.0021499355789273977, + 0.002993151778355241, + 0.0030027704779058695, + 0.0025186603888869286, + 0.0026157007087022066, + 0.002271756064146757, + 0.0024568275548517704, + 0.0028543442022055387, + 0.0031900564208626747, + 0.008685658685863018, + 0.009603898972272873, + 0.003111443715170026, + 0.0018522364553064108, + 0.002042410196736455, + 0.0025107157416641712, + 0.0016363193280994892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.08903826028108597 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_9/index_tensor([7040, 7040])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_9/index_tensor([7040, 7040])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f910fdc7c71c8c39ecc97af57a53590c95110f2d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_9/index_tensor([7040, 7040])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide, eye-level tracking shot follows a lone traveler as they glide smoothly through a vast, modern airport terminal bathed in the golden glow of the afternoon sun. Light streams through vast floor-to-ceiling windows, casting long, dramatic shadows across the highly polished concrete floor and creating cinematic lens flares as the camera moves. The traveler, a quiet figure with medium-length dark brown hair, wears a practical olive green canvas jacket paired with a simple white T-shirt and dark denim jeans for a long journey. They walk with a steady, deliberate pace, one hand firmly grasping the extended handle of a sleek, dark gray hardshell rolling suitcase, its wheels gliding almost silently across the floor. In the background, the vast expanse of the concourse stretches outward, its high, vaulted ceiling supported by minimalist steel beams, while other passengers become blurred, anonymous figures in the distance. Through the expansive glass, the silhouettes of airplanes are visible on the sun-drenched tarmac, imbuing the scene with a hopeful and contemplative atmosphere, capturing a moment of transition and the purposeful beginning of an important journey to a distant gate.", + "seed": 29000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_9/index_7040_seed_29000_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010451881773769855, + "vlm_logprob": -0.010451881773769855, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.010451881773769855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005821061320602894 + ], + "opposite_logprob_tokens": [ + -6.755821228027344 + ], + "logprob_reward": -0.005821061320602894, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005821061320602894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1315937638282776, + 0.06660568714141846, + 0.09286811202764511, + 0.07990039139986038, + 0.07349071651697159, + 0.06979820132255554, + 0.11558623611927032, + 0.10534197837114334, + 0.13112564384937286, + 0.133689284324646 + ], + "credit_map": [ + 0.014165369793772697, + 0.016030384227633476, + 0.019258348271250725, + 0.01425300445407629, + 0.017358141019940376, + 0.018863549456000328, + 0.015147306025028229, + 0.01651766151189804, + 0.011943436227738857, + 0.007302949205040932, + 0.008441450074315071, + 0.00482319388538599, + 0.007116171531379223, + 0.008322962559759617, + 0.007677304092794657, + 0.0109782163053751, + 0.014596524648368359, + 0.00854278914630413, + 0.01058996096253395, + 0.0070212651044130325, + 0.0069381180219352245, + 0.011814861558377743, + 0.005009869579225779, + 0.028354724869132042, + 0.008158762939274311, + 0.010383822955191135, + 0.01178672630339861, + 0.008144873194396496, + 0.01215976569801569, + 0.010858766734600067, + 0.008742228150367737, + 0.009665446355938911, + 0.009783055633306503, + 0.008526140823960304, + 0.00715928990393877, + 0.006829138845205307, + 0.010937086306512356, + 0.012980855070054531, + 0.009325248189270496, + 0.007949906401336193, + 0.008682871237397194, + 0.010200249962508678, + 0.009900146164000034, + 0.007129695266485214, + 0.010933585464954376, + 0.006161043886095285, + 0.008579308167099953, + 0.008211296051740646, + 0.014962357468903065, + 0.015147028490900993, + 0.015792563557624817, + 0.0122286481782794, + 0.01285503525286913, + 0.01544574461877346, + 0.014503979124128819, + 0.014650876633822918, + 0.016355890780687332, + 0.01033381000161171, + 0.013879683800041676, + 0.01810752972960472, + 0.006807738449424505, + 0.015756050124764442, + 0.010385947301983833, + 0.01371532492339611, + 0.029634809121489525, + 0.02053293026983738, + 0.01237315684556961, + 0.011351116932928562, + 0.012040683999657631, + 0.012325338087975979, + 0.01608516275882721, + 0.016782449558377266, + 0.016190528869628906, + 0.009836175478994846, + 0.021005850285291672, + 0.01718599908053875, + 0.014516712166368961, + 0.022373629733920097, + 0.015965009108185768, + 0.01661538891494274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005821061320602894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651915282011032 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004651915282011032, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651915282011032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08176678419113159, + 0.0395875908434391, + 0.039028599858284, + 0.058172114193439484, + 0.4284142255783081, + 0.04005609452724457, + 0.1038525402545929, + 0.06480541825294495, + 0.030541397631168365, + 0.11377526074647903 + ], + "credit_map": [ + 0.00878628995269537, + 0.00833928119391203, + 0.014214270748198032, + 0.011373027227818966, + 0.00557396886870265, + 0.010777872055768967, + 0.008854089304804802, + 0.013847985304892063, + 0.0059403833001852036, + 0.004710002802312374, + 0.007139981724321842, + 0.0042640529572963715, + 0.0038601665291935205, + 0.003183069871738553, + 0.005524900741875172, + 0.004965032916516066, + 0.0064393202774226665, + 0.004217239562422037, + 0.004350706469267607, + 0.004295006860047579, + 0.0043115150183439255, + 0.004415316041558981, + 0.004409317392855883, + 0.006590178702026606, + 0.005731372162699699, + 0.006775661837309599, + 0.009474007412791252, + 0.006138850003480911, + 0.007337923627346754, + 0.008509430103003979, + 0.007535307202488184, + 0.006669559981673956, + 0.05318952724337578, + 0.07592330127954483, + 0.034724995493888855, + 0.04821131378412247, + 0.049661844968795776, + 0.08377007395029068, + 0.03176678717136383, + 0.05116639658808708, + 0.005202700849622488, + 0.006930182222276926, + 0.004835869185626507, + 0.0036381736863404512, + 0.005516091361641884, + 0.005225410684943199, + 0.003736837999895215, + 0.004970828536897898, + 0.015350878238677979, + 0.01613175868988037, + 0.013271980918943882, + 0.012303391471505165, + 0.008271080441772938, + 0.014307646080851555, + 0.011032425798475742, + 0.013183373026549816, + 0.009312017820775509, + 0.012104753404855728, + 0.006664579268544912, + 0.006580944173038006, + 0.0038278272841125727, + 0.011117947287857533, + 0.006975443102419376, + 0.008221910335123539, + 0.004137816838920116, + 0.003378345863893628, + 0.003519082674756646, + 0.003540733829140663, + 0.002942524151876569, + 0.00409570150077343, + 0.00473928265273571, + 0.0041879089549183846, + 0.01279471442103386, + 0.006006081122905016, + 0.020146135240793228, + 0.019448455423116684, + 0.010825169272720814, + 0.013146130368113518, + 0.016843242570757866, + 0.01456532720476389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.004651915282011032 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188944086432457 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.018188944086432457, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188944086432457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10203714668750763, + 0.055232468992471695, + 0.06819034367799759, + 0.12712818384170532, + 0.07380361109972, + 0.07506439089775085, + 0.14773273468017578, + 0.08167891204357147, + 0.07697626948356628, + 0.19215598702430725 + ], + "credit_map": [ + 0.010846158489584923, + 0.012247051112353802, + 0.014477242715656757, + 0.013634027913212776, + 0.011396089568734169, + 0.013142529875040054, + 0.01412709429860115, + 0.012166946195065975, + 0.0070131197571754456, + 0.005986418575048447, + 0.008119861595332623, + 0.007066558115184307, + 0.006345696747303009, + 0.005293601658195257, + 0.007035463582724333, + 0.008371751755475998, + 0.006173369474709034, + 0.008884811773896217, + 0.007166213355958462, + 0.006387737113982439, + 0.006658751051872969, + 0.007849249057471752, + 0.00643988186493516, + 0.01863032765686512, + 0.012584013864398003, + 0.016872690990567207, + 0.016420599073171616, + 0.016075054183602333, + 0.017001833766698837, + 0.015338039956986904, + 0.018930457532405853, + 0.013905499130487442, + 0.008961870335042477, + 0.00743458466604352, + 0.011012572795152664, + 0.01171767245978117, + 0.007895673625171185, + 0.009815658442676067, + 0.008936241269111633, + 0.008029336109757423, + 0.007099006325006485, + 0.011186869814991951, + 0.007464176043868065, + 0.007370918523520231, + 0.011929430067539215, + 0.008329140022397041, + 0.010613445192575455, + 0.011071408167481422, + 0.012580911628901958, + 0.019829684868454933, + 0.012624015100300312, + 0.016391471028327942, + 0.02918918989598751, + 0.013907311484217644, + 0.023873072117567062, + 0.019337082281708717, + 0.007809831295162439, + 0.01384210865944624, + 0.0070729684084653854, + 0.011860688216984272, + 0.007295264396816492, + 0.012388838455080986, + 0.01106966845691204, + 0.010339547879993916, + 0.011227725073695183, + 0.011251581832766533, + 0.005352024920284748, + 0.007499973755329847, + 0.015319131314754486, + 0.007493141107261181, + 0.009747433476150036, + 0.009085253812372684, + 0.0168603602796793, + 0.01438615657389164, + 0.02356639876961708, + 0.032913174480199814, + 0.019085431471467018, + 0.03260117396712303, + 0.026577696204185486, + 0.026165585964918137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.018188944086432457 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067785322666168 + ], + "opposite_logprob_tokens": [ + -4.511067867279053 + ], + "logprob_reward": -0.011067785322666168, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067785322666168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18902850151062012, + 0.1524515599012375, + 0.036730196326971054, + 0.07490513473749161, + 0.11680686473846436, + 0.03586580231785774, + 0.058563295751810074, + 0.1093006506562233, + 0.06145879998803139, + 0.16488920152187347 + ], + "credit_map": [ + 0.018157213926315308, + 0.020177310332655907, + 0.024547766894102097, + 0.028280213475227356, + 0.026177920401096344, + 0.02311318926513195, + 0.021740229800343513, + 0.026834649965167046, + 0.021752841770648956, + 0.016760313883423805, + 0.02344794198870659, + 0.018569163978099823, + 0.01602005399763584, + 0.011980373412370682, + 0.022111844271421432, + 0.02180902473628521, + 0.003919142298400402, + 0.004045519977807999, + 0.004749682731926441, + 0.004751159343868494, + 0.004133738577365875, + 0.0050718821585178375, + 0.004261928144842386, + 0.005797142628580332, + 0.007682696916162968, + 0.009716801345348358, + 0.009675600565969944, + 0.00884127989411354, + 0.010423346422612667, + 0.009742015041410923, + 0.009888519532978535, + 0.008934873156249523, + 0.015175477601587772, + 0.01817692257463932, + 0.014346346259117126, + 0.0122977951541543, + 0.016342569142580032, + 0.015676358714699745, + 0.009709016419947147, + 0.015082383528351784, + 0.003631786908954382, + 0.00535851065069437, + 0.004882531706243753, + 0.0031008596997708082, + 0.004547886550426483, + 0.005322301760315895, + 0.004542374052107334, + 0.004479550290852785, + 0.006825090851634741, + 0.008258932270109653, + 0.006154543720185757, + 0.006480913609266281, + 0.008439500816166401, + 0.0069963568821549416, + 0.007960041053593159, + 0.007447915617376566, + 0.013394816778600216, + 0.0212055966258049, + 0.010004463605582714, + 0.009853126481175423, + 0.006732661742717028, + 0.021734921261668205, + 0.012580805458128452, + 0.013794262893497944, + 0.01151255238801241, + 0.0059151556342840195, + 0.008667625486850739, + 0.006409384775906801, + 0.004099955316632986, + 0.00793028436601162, + 0.007595979608595371, + 0.009327864274382591, + 0.013152694329619408, + 0.013883206993341446, + 0.02323310635983944, + 0.022491371259093285, + 0.016923412680625916, + 0.02311295084655285, + 0.02769005112349987, + 0.024402404204010963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.011067785322666168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012529702857136726 + ], + "opposite_logprob_tokens": [ + -4.3875298500061035 + ], + "logprob_reward": -0.012529702857136726, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012529702857136726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20905651152133942, + 0.11229147762060165, + 0.04838116094470024, + 0.04924002289772034, + 0.09562254697084427, + 0.08480604737997055, + 0.0439012236893177, + 0.057815950363874435, + 0.19320324063301086, + 0.10568179190158844 + ], + "credit_map": [ + 0.022715797647833824, + 0.02256610244512558, + 0.030739936977624893, + 0.033115725964307785, + 0.027039116248488426, + 0.024152809754014015, + 0.022557483986020088, + 0.026169534772634506, + 0.016388509422540665, + 0.010406030341982841, + 0.014208531007170677, + 0.01372657623142004, + 0.008618812076747417, + 0.013212798163294792, + 0.011963125318288803, + 0.02376709319651127, + 0.008557155728340149, + 0.00562082277610898, + 0.006381654180586338, + 0.005102533847093582, + 0.005529378540813923, + 0.005112707149237394, + 0.005366010125726461, + 0.006710901390761137, + 0.004061520099639893, + 0.005564106162637472, + 0.004259693901985884, + 0.008489721454679966, + 0.00739072123542428, + 0.006531150080263615, + 0.006966856773942709, + 0.005976254120469093, + 0.008561555296182632, + 0.012807019986212254, + 0.006911692675203085, + 0.012431505136191845, + 0.015130535699427128, + 0.015570051036775112, + 0.011407442390918732, + 0.012802747078239918, + 0.010903495363891125, + 0.011138485744595528, + 0.01164302323013544, + 0.009279042482376099, + 0.011365707963705063, + 0.008728309534490108, + 0.011295591481029987, + 0.0104523915797472, + 0.0048146434128284454, + 0.005711310543119907, + 0.005479589104652405, + 0.006623167544603348, + 0.004837217275053263, + 0.0049294885247945786, + 0.006034858990460634, + 0.005470942705869675, + 0.008168207481503487, + 0.008419758640229702, + 0.005959332920610905, + 0.00671754963696003, + 0.004310172982513905, + 0.009418747387826443, + 0.007244231645017862, + 0.00757795013487339, + 0.017915578559041023, + 0.021545803174376488, + 0.01834961585700512, + 0.02585633657872677, + 0.030808981508016586, + 0.021496787667274475, + 0.03514277562499046, + 0.022087357938289642, + 0.008499826304614544, + 0.007303758058696985, + 0.01597657799720764, + 0.012167824432253838, + 0.008763732388615608, + 0.01915886066854, + 0.0172067079693079, + 0.016604503616690636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.012529702857136726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, eye-level tracking shot follows a lone traveler as they glide smoothly through a vast, modern airport terminal bathed in the golden glow of the afternoon sun. Light streams through vast floor-to-ceiling windows, casting long, dramatic shadows across the highly polished concrete floor and creating cinematic lens flares as the camera moves. The traveler, a quiet figure with medium-length dark brown hair, wears a practical olive green canvas jacket paired with a simple white T-shirt and dark denim jeans for a long journey. They walk with a steady, deliberate pace, one hand firmly grasping the extended handle of a sleek, dark gray hardshell rolling suitcase, its wheels gliding almost silently across the floor. In the background, the vast expanse of the concourse stretches outward, its high, vaulted ceiling supported by minimalist steel beams, while other passengers become blurred, anonymous figures in the distance. Through the expansive glass, the silhouettes of airplanes are visible on the sun-drenched tarmac, imbuing the scene with a hopeful and contemplative atmosphere, capturing a moment of transition and the purposeful beginning of an important journey to a distant gate.", + "seed": 29001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_9/index_7040_seed_29001_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010949767380952834, + "vlm_logprob": -0.010949767380952834, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.010949767380952834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01225842721760273 + ], + "opposite_logprob_tokens": [ + -6.762258529663086 + ], + "logprob_reward": -0.01225842721760273, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.01225842721760273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08183655142784119, + 0.07645920664072037, + 0.09316589683294296, + 0.08251208066940308, + 0.08172162622213364, + 0.08106566965579987, + 0.06307439506053925, + 0.2252325713634491, + 0.13032904267311096, + 0.08460298180580139 + ], + "credit_map": [ + 0.008585792034864426, + 0.010441415943205357, + 0.00979748647660017, + 0.011601991020143032, + 0.007905689999461174, + 0.012083731591701508, + 0.009756362065672874, + 0.011664089746773243, + 0.006332810968160629, + 0.011629102751612663, + 0.007509942166507244, + 0.01012114156037569, + 0.00872884877026081, + 0.00784242618829012, + 0.012761085294187069, + 0.011533849872648716, + 0.013243625871837139, + 0.011974611319601536, + 0.01339976117014885, + 0.010487030260264874, + 0.007943650707602501, + 0.013531469739973545, + 0.0085737444460392, + 0.014012014493346214, + 0.010866768658161163, + 0.009479016065597534, + 0.015590638853609562, + 0.010809381492435932, + 0.01001090556383133, + 0.008203321136534214, + 0.007699393667280674, + 0.009852662682533264, + 0.010103083215653896, + 0.008237462490797043, + 0.012251936830580235, + 0.012764617800712585, + 0.008465161547064781, + 0.007835901342332363, + 0.009587588720023632, + 0.012475881725549698, + 0.007013027556240559, + 0.009273272007703781, + 0.011813109740614891, + 0.011526140384376049, + 0.011592124588787556, + 0.005657977424561977, + 0.012611279264092445, + 0.011578744277358055, + 0.009878568351268768, + 0.008216463960707188, + 0.0072337063029408455, + 0.008397537283599377, + 0.007350863888859749, + 0.006670859642326832, + 0.007470941171050072, + 0.007855452597141266, + 0.017789475619792938, + 0.021281151100993156, + 0.033427491784095764, + 0.03510792925953865, + 0.013325942680239677, + 0.025892147794365883, + 0.02629975602030754, + 0.052108701318502426, + 0.015230252407491207, + 0.017938323318958282, + 0.019760306924581528, + 0.020284002646803856, + 0.010808628052473068, + 0.014223223552107811, + 0.015026946552097797, + 0.017057368531823158, + 0.013147050514817238, + 0.008584341034293175, + 0.01613144762814045, + 0.009126733988523483, + 0.0079265795648098, + 0.008894849568605423, + 0.008856288157403469, + 0.011935697868466377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.01225842721760273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004112834110856056 + ], + "opposite_logprob_tokens": [ + -5.504112720489502 + ], + "logprob_reward": -0.004112834110856056, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004112834110856056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15100648999214172, + 0.058335013687610626, + 0.10185348242521286, + 0.07808265835046768, + 0.058313656598329544, + 0.05515958368778229, + 0.05701703205704689, + 0.0872611328959465, + 0.07843956351280212, + 0.2745313346385956 + ], + "credit_map": [ + 0.018355298787355423, + 0.017319433391094208, + 0.01971081644296646, + 0.02391645684838295, + 0.012846879661083221, + 0.022742915898561478, + 0.014734035357832909, + 0.021380653604865074, + 0.005910890642553568, + 0.007251202128827572, + 0.007682240102440119, + 0.007272061426192522, + 0.0055414168164134026, + 0.0076566231437027454, + 0.008847644552588463, + 0.008172934874892235, + 0.013043141923844814, + 0.014975498430430889, + 0.013489115051925182, + 0.01603822410106659, + 0.006756564602255821, + 0.016843609511852264, + 0.009755904786288738, + 0.010951423086225986, + 0.008718419820070267, + 0.005906193051487207, + 0.020344315096735954, + 0.01250402256846428, + 0.007248821668326855, + 0.0060862209647893906, + 0.006567292381078005, + 0.010707375593483448, + 0.007599493023008108, + 0.0047330092638731, + 0.010655384510755539, + 0.008968939073383808, + 0.004674743395298719, + 0.00709562050178647, + 0.006575177423655987, + 0.008011288940906525, + 0.008217818103730679, + 0.005836159456521273, + 0.007439813110977411, + 0.009283945895731449, + 0.007535767275840044, + 0.003954703453928232, + 0.006198022980242968, + 0.006693349219858646, + 0.006136162206530571, + 0.00791961420327425, + 0.006022998131811619, + 0.008876380510628223, + 0.006885071750730276, + 0.0076356991194188595, + 0.006765391677618027, + 0.006775715854018927, + 0.008160010911524296, + 0.009457660838961601, + 0.011983577162027359, + 0.014894789084792137, + 0.005234370473772287, + 0.008928710594773293, + 0.009356055408716202, + 0.01924595795571804, + 0.009681770578026772, + 0.0072328923270106316, + 0.015339907258749008, + 0.015331842005252838, + 0.006429821252822876, + 0.0061735063791275024, + 0.009673837572336197, + 0.008575988002121449, + 0.038983315229415894, + 0.009241249412298203, + 0.07805077731609344, + 0.029424086213111877, + 0.017206981778144836, + 0.028599930927157402, + 0.02294079400599003, + 0.050084199756383896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.004112834110856056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068367287516594 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068367287516594, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068367287516594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1255655139684677, + 0.03213541954755783, + 0.048167768865823746, + 0.17563822865486145, + 0.1565835326910019, + 0.08197968453168869, + 0.05319778993725777, + 0.10171684622764587, + 0.07013477385044098, + 0.15488047897815704 + ], + "credit_map": [ + 0.012629781849682331, + 0.015663234516978264, + 0.014033116400241852, + 0.02050090953707695, + 0.01675432175397873, + 0.014689693227410316, + 0.013059708289802074, + 0.01823473535478115, + 0.0039955987595021725, + 0.004429152235388756, + 0.003224360989406705, + 0.003512716619297862, + 0.005299105774611235, + 0.003412913763895631, + 0.004444151651114225, + 0.0038174150977283716, + 0.00461241789162159, + 0.00638474291190505, + 0.008713511750102043, + 0.006558351218700409, + 0.004170956555753946, + 0.005098791792988777, + 0.004560441244393587, + 0.008068552240729332, + 0.012085613794624805, + 0.017963120713829994, + 0.01688026823103428, + 0.01963200606405735, + 0.051370300352573395, + 0.012868307530879974, + 0.024719588458538055, + 0.02011900581419468, + 0.009272798895835876, + 0.01260469388216734, + 0.013589104637503624, + 0.01752929762005806, + 0.036415111273527145, + 0.015774661675095558, + 0.03377661481499672, + 0.0176212377846241, + 0.008647824637591839, + 0.00971054658293724, + 0.011102546006441116, + 0.015241267159581184, + 0.010824231430888176, + 0.006112154573202133, + 0.012472559697926044, + 0.007868544198572636, + 0.004855677485466003, + 0.005798736121505499, + 0.006057393737137318, + 0.006820752285420895, + 0.005926464684307575, + 0.006138582248240709, + 0.008873187005519867, + 0.008726995438337326, + 0.009090447798371315, + 0.01400015503168106, + 0.013741517439484596, + 0.016738232225179672, + 0.007088596001267433, + 0.010598104447126389, + 0.012102044187486172, + 0.01835775002837181, + 0.011984353885054588, + 0.007032237946987152, + 0.010073201730847359, + 0.014036763459444046, + 0.003935557324439287, + 0.0051091513596475124, + 0.011189908720552921, + 0.006773591507226229, + 0.020304011180996895, + 0.009119924157857895, + 0.03865780308842659, + 0.017051218077540398, + 0.01265027653425932, + 0.016624826937913895, + 0.013541472144424915, + 0.026930946856737137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.016068367287516594 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534411624073982 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534411624073982, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534411624073982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14128166437149048, + 0.04334678873419762, + 0.08930255472660065, + 0.0634591206908226, + 0.11721735447645187, + 0.10580527037382126, + 0.06369821727275848, + 0.1319400817155838, + 0.11075835675001144, + 0.13319067656993866 + ], + "credit_map": [ + 0.016067376360297203, + 0.014848562888801098, + 0.016241183504462242, + 0.02339107356965542, + 0.01341786328703165, + 0.01838623359799385, + 0.01862773671746254, + 0.020301630720496178, + 0.0046801320277154446, + 0.005078773479908705, + 0.004998197313398123, + 0.0047839791513979435, + 0.003823615610599518, + 0.006582271307706833, + 0.00689620990306139, + 0.0065036010928452015, + 0.007958962582051754, + 0.01146068423986435, + 0.020063363015651703, + 0.007341600023210049, + 0.0076032038778066635, + 0.009280047379434109, + 0.009767048992216587, + 0.015827640891075134, + 0.0050365193746984005, + 0.005009130109101534, + 0.01091307308524847, + 0.013317985460162163, + 0.007282697129994631, + 0.00488507142290473, + 0.009309019893407822, + 0.007705621421337128, + 0.009766512550413609, + 0.010903660207986832, + 0.013275125995278358, + 0.01551375724375248, + 0.023884598165750504, + 0.011329052969813347, + 0.014706436544656754, + 0.01783820055425167, + 0.013948642648756504, + 0.01121607981622219, + 0.01526104286313057, + 0.021029703319072723, + 0.013658449985086918, + 0.006743852514773607, + 0.012226653285324574, + 0.011720840819180012, + 0.006572120822966099, + 0.0050447918474674225, + 0.006850986275821924, + 0.009992537088692188, + 0.006754780653864145, + 0.0060753850266337395, + 0.01094655692577362, + 0.011461053043603897, + 0.012049966491758823, + 0.020273050293326378, + 0.018239129334688187, + 0.02123289927840233, + 0.007172204088419676, + 0.010975164361298084, + 0.014459334313869476, + 0.02753831073641777, + 0.011262491345405579, + 0.009478718973696232, + 0.015854910016059875, + 0.02963590994477272, + 0.007897287607192993, + 0.006360209546983242, + 0.02001052163541317, + 0.010258292779326439, + 0.01729605160653591, + 0.010808292776346207, + 0.030812939628958702, + 0.015739930793642998, + 0.011360811069607735, + 0.014140239916741848, + 0.011671680957078934, + 0.021360713988542557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.012534411624073982 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774796664714813 + ], + "opposite_logprob_tokens": [ + -4.634774684906006 + ], + "logprob_reward": -0.009774796664714813, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774796664714813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1191665455698967, + 0.06938683241605759, + 0.1475730836391449, + 0.09046117216348648, + 0.0410064272582531, + 0.10778908431529999, + 0.05553441867232323, + 0.1687915027141571, + 0.09678374230861664, + 0.10350722819566727 + ], + "credit_map": [ + 0.009751294739544392, + 0.011096215806901455, + 0.020483503118157387, + 0.019988253712654114, + 0.010410900227725506, + 0.01983260177075863, + 0.011261538602411747, + 0.01634223759174347, + 0.005602490156888962, + 0.008364011533558369, + 0.0059729465283453465, + 0.010225988924503326, + 0.007546859793365002, + 0.0062121073715388775, + 0.013667731545865536, + 0.011794702149927616, + 0.013417602516710758, + 0.02138620987534523, + 0.03743560239672661, + 0.010339498519897461, + 0.006487999577075243, + 0.01226803008466959, + 0.01317552663385868, + 0.03306262567639351, + 0.013400179333984852, + 0.01546396128833294, + 0.01393633522093296, + 0.01050030067563057, + 0.007735158782452345, + 0.007540733553469181, + 0.010454094968736172, + 0.011430412530899048, + 0.002626292174682021, + 0.003597670467570424, + 0.005119701381772757, + 0.004282086621969938, + 0.009062917903065681, + 0.0038955819327384233, + 0.006212498992681503, + 0.006209674756973982, + 0.014241619035601616, + 0.008896333165466785, + 0.01441211812198162, + 0.024000028148293495, + 0.012247118167579174, + 0.00784913171082735, + 0.016314810141921043, + 0.00982793141156435, + 0.00609866576269269, + 0.00519516970962286, + 0.00715069891884923, + 0.008152312599122524, + 0.006194234825670719, + 0.004662503022700548, + 0.010385655798017979, + 0.007695178035646677, + 0.01720448024570942, + 0.025929732248187065, + 0.022555062547326088, + 0.024512365460395813, + 0.012116613797843456, + 0.0161884855479002, + 0.017701303586363792, + 0.03258345276117325, + 0.01796659640967846, + 0.009182332083582878, + 0.01573142409324646, + 0.01474007498472929, + 0.007486824411898851, + 0.009484331123530865, + 0.010218634270131588, + 0.011973521672189236, + 0.011537043377757072, + 0.01437960285693407, + 0.014272527769207954, + 0.02114633470773697, + 0.008518325164914131, + 0.006756703369319439, + 0.013963928446173668, + 0.012932763434946537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.009774796664714813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, eye-level tracking shot follows a lone traveler as they glide smoothly through a vast, modern airport terminal bathed in the golden glow of the afternoon sun. Light streams through vast floor-to-ceiling windows, casting long, dramatic shadows across the highly polished concrete floor and creating cinematic lens flares as the camera moves. The traveler, a quiet figure with medium-length dark brown hair, wears a practical olive green canvas jacket paired with a simple white T-shirt and dark denim jeans for a long journey. They walk with a steady, deliberate pace, one hand firmly grasping the extended handle of a sleek, dark gray hardshell rolling suitcase, its wheels gliding almost silently across the floor. In the background, the vast expanse of the concourse stretches outward, its high, vaulted ceiling supported by minimalist steel beams, while other passengers become blurred, anonymous figures in the distance. Through the expansive glass, the silhouettes of airplanes are visible on the sun-drenched tarmac, imbuing the scene with a hopeful and contemplative atmosphere, capturing a moment of transition and the purposeful beginning of an important journey to a distant gate.", + "seed": 29002, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_9/index_7040_seed_29002_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013895916193723679, + "vlm_logprob": -0.013895916193723679, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.013895916193723679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014359474182128906 + ], + "opposite_logprob_tokens": [ + -5.014359474182129 + ], + "logprob_reward": -0.014359474182128906, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.014359474182128906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029044199734926224, + 0.02974517270922661, + 0.09159813076257706, + 0.021500200033187866, + 0.04366319999098778, + 0.024159464985132217, + 0.06868953257799149, + 0.6698715686798096, + 0.011104893870651722, + 0.010623574256896973 + ], + "credit_map": [ + 0.0028434842824935913, + 0.004344503860920668, + 0.0031808577477931976, + 0.004076443146914244, + 0.0033810127060860395, + 0.003995719365775585, + 0.0035200435668230057, + 0.003702137153595686, + 0.003659867448732257, + 0.003365499898791313, + 0.00296945427544415, + 0.004717798437923193, + 0.004361047875136137, + 0.0035215618554502726, + 0.00407792441546917, + 0.003072020597755909, + 0.007905805483460426, + 0.01154131069779396, + 0.01080715749412775, + 0.014122677966952324, + 0.01503161434084177, + 0.007400491740554571, + 0.013310450129210949, + 0.01147862896323204, + 0.0019109881250187755, + 0.0013847560621798038, + 0.00481807067990303, + 0.0027954471297562122, + 0.0011224912013858557, + 0.0019926708191633224, + 0.004806859418749809, + 0.0026689174119383097, + 0.003983247559517622, + 0.005901210475713015, + 0.006447266787290573, + 0.005989382974803448, + 0.00556974159553647, + 0.004337766207754612, + 0.006029433570802212, + 0.005405152682214975, + 0.001765213324688375, + 0.002230773912742734, + 0.004596227314323187, + 0.003655637614428997, + 0.003206406719982624, + 0.0030152983963489532, + 0.0026455207262188196, + 0.0030443882569670677, + 0.005507708992809057, + 0.005713098216801882, + 0.014585476368665695, + 0.0074948593974113464, + 0.0062408773228526115, + 0.01241213083267212, + 0.006387423723936081, + 0.010347968898713589, + 0.027148203924298286, + 0.008127818815410137, + 0.2501184642314911, + 0.022181566804647446, + 0.015747031196951866, + 0.006278419401496649, + 0.28003233671188354, + 0.06023778021335602, + 0.0011836511548608541, + 0.0010806213831529021, + 0.002262407448142767, + 0.001288585364818573, + 0.0007406339864246547, + 0.0015831892378628254, + 0.0012782406993210316, + 0.0016875648871064186, + 0.0008010258898139, + 0.000972271547652781, + 0.0018206755630671978, + 0.0019499873742461205, + 0.0011270404793322086, + 0.0007833654526621103, + 0.0015990500105544925, + 0.0015701580559834838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.014359474182128906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553129345178604 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553129345178604, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553129345178604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07844778895378113, + 0.05394238606095314, + 0.04438232257962227, + 0.053973015397787094, + 0.06624723970890045, + 0.047792699187994, + 0.025593752041459084, + 0.5927261114120483, + 0.014838065952062607, + 0.022056536749005318 + ], + "credit_map": [ + 0.009621931239962578, + 0.008493746630847454, + 0.009297632612287998, + 0.01296515017747879, + 0.007128774654120207, + 0.006122560240328312, + 0.013629022985696793, + 0.01118898019194603, + 0.007789373863488436, + 0.005494051147252321, + 0.006820240989327431, + 0.012261036783456802, + 0.005641370080411434, + 0.0038206803146749735, + 0.00599882286041975, + 0.006116810720413923, + 0.005226669833064079, + 0.005163080990314484, + 0.005503165535628796, + 0.006044267676770687, + 0.007367799989879131, + 0.003947145771235228, + 0.005537667777389288, + 0.005592527333647013, + 0.003801139537245035, + 0.0029499472584575415, + 0.01276188064366579, + 0.005073133390396833, + 0.004581414628773928, + 0.002665730193257332, + 0.015663353726267815, + 0.00647641671821475, + 0.003422216046601534, + 0.005886271130293608, + 0.008307360112667084, + 0.006129754241555929, + 0.005807606503367424, + 0.003298531984910369, + 0.021112972870469093, + 0.012282531708478928, + 0.00356918852776289, + 0.005931810941547155, + 0.008884016424417496, + 0.006504873279482126, + 0.007198592647910118, + 0.004904572851955891, + 0.004068631678819656, + 0.006731011904776096, + 0.002407762687653303, + 0.002968371845781803, + 0.004338564816862345, + 0.004923458211123943, + 0.0023663502652198076, + 0.002479251939803362, + 0.0025944614317268133, + 0.0035155327059328556, + 0.02408611588180065, + 0.009931973181664944, + 0.22261002659797668, + 0.019351158291101456, + 0.01397654041647911, + 0.005807941313832998, + 0.24544154107570648, + 0.05152084305882454, + 0.0011679816525429487, + 0.001530995243228972, + 0.0028937829192727804, + 0.0023657793644815683, + 0.0009858830599114299, + 0.0016118035418912768, + 0.002594235586002469, + 0.001687605050392449, + 0.002440202981233597, + 0.0024268082343041897, + 0.0035529460292309523, + 0.0025671045295894146, + 0.002535836538299918, + 0.0020104653667658567, + 0.0035007742699235678, + 0.003022399265319109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.012553129345178604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010585789568722248, + 0.005039803683757782, + 0.015597952529788017, + 0.012622350826859474, + 0.022470748052001, + 0.0046431515365839005, + 0.007800907827913761, + 0.9143784642219543, + 0.003057723632082343, + 0.003803121391683817 + ], + "credit_map": [ + 0.0011250306852161884, + 0.0012642493238672614, + 0.0011577238328754902, + 0.0017615548567846417, + 0.0012239770730957389, + 0.000843906425870955, + 0.002307476010173559, + 0.0009018704295158386, + 0.000708435254637152, + 0.0005454552010633051, + 0.0005201725289225578, + 0.0007152842590585351, + 0.0006184287485666573, + 0.00044681827421300113, + 0.000743142212741077, + 0.0007420667097903788, + 0.0018947829958051443, + 0.001662860973738134, + 0.0017695620190352201, + 0.0022935920860618353, + 0.0026221268344670534, + 0.0012357825180515647, + 0.0023916084319353104, + 0.0017276352737098932, + 0.00081522116670385, + 0.0010781217133626342, + 0.0022825717460364103, + 0.0012448416091501713, + 0.0017671588575467467, + 0.0007104665855877101, + 0.0032421087380498648, + 0.0014818597119301558, + 0.001038716989569366, + 0.0010656045051291585, + 0.0038599579129368067, + 0.0019704841542989016, + 0.0014312827261164784, + 0.000827849202323705, + 0.008219900541007519, + 0.004056950099766254, + 0.0004854707221966237, + 0.0004447243991307914, + 0.0007286809850484133, + 0.0008961225976236165, + 0.00033877641544677317, + 0.00040094886207953095, + 0.0006423103041015565, + 0.0007061170181259513, + 0.0010546722915023565, + 0.0011205447372049093, + 0.0013350368244573474, + 0.0009523344924673438, + 0.0007467776304110885, + 0.0008432302856817842, + 0.0008610700606368482, + 0.0008872409234754741, + 0.0367155559360981, + 0.011834758333861828, + 0.3399244546890259, + 0.03035609796643257, + 0.021605374291539192, + 0.007893659174442291, + 0.38309723138809204, + 0.08295124024152756, + 0.0004047835827805102, + 0.0003206910041626543, + 0.0006367901805788279, + 0.0004183253622613847, + 0.0002198667061747983, + 0.00023506463912781328, + 0.0004400266916491091, + 0.00038217531982809305, + 0.00019603238615673035, + 0.00027904065791517496, + 0.0004533030733000487, + 0.0012879152782261372, + 0.00018157920567318797, + 0.0002634676930028945, + 0.0006118802702985704, + 0.0005299022886902094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.014192949049174786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418930571526289 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.01418930571526289, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418930571526289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03698016703128815, + 0.015856049954891205, + 0.06447037309408188, + 0.03874460980296135, + 0.1855206936597824, + 0.04154326021671295, + 0.05050944909453392, + 0.541227400302887, + 0.011189673095941544, + 0.013958415016531944 + ], + "credit_map": [ + 0.003850498702377081, + 0.006040117237716913, + 0.0031359612476080656, + 0.006008140277117491, + 0.0041151526384055614, + 0.004748037084937096, + 0.0046716248616576195, + 0.004410628695040941, + 0.0016987987328320742, + 0.0018153932178393006, + 0.001943483599461615, + 0.0024816023651510477, + 0.001831972156651318, + 0.001517527736723423, + 0.0024084337055683136, + 0.002158836228772998, + 0.0053549702279269695, + 0.005886727944016457, + 0.008684922941029072, + 0.011026585474610329, + 0.010473421774804592, + 0.005064722616225481, + 0.011896229349076748, + 0.006082786712795496, + 0.0025219193194061518, + 0.002457191003486514, + 0.007567720487713814, + 0.004350978881120682, + 0.003444902366027236, + 0.0024847351014614105, + 0.010370329953730106, + 0.00554682407528162, + 0.006864557042717934, + 0.007548946421593428, + 0.019280368462204933, + 0.012941374443471432, + 0.01525266282260418, + 0.008464416489005089, + 0.07043492794036865, + 0.04473341628909111, + 0.0026460858061909676, + 0.005116631276905537, + 0.005045171361416578, + 0.012386022135615349, + 0.0029319904278963804, + 0.0027764872647821903, + 0.00546117452904582, + 0.0051796892657876015, + 0.0038789266254752874, + 0.005352438427507877, + 0.007036362774670124, + 0.008992891758680344, + 0.004712813068181276, + 0.008591950871050358, + 0.0059847659431397915, + 0.00595929054543376, + 0.02194044180214405, + 0.0067812735214829445, + 0.20327302813529968, + 0.017668042331933975, + 0.01280770730227232, + 0.005170208401978016, + 0.22551915049552917, + 0.04806746542453766, + 0.0010155224008485675, + 0.0011203745380043983, + 0.0022937762551009655, + 0.0017855647020041943, + 0.0007875719456933439, + 0.0008362950757145882, + 0.0020514768548309803, + 0.0012990901013836265, + 0.00118045206181705, + 0.0011615273542702198, + 0.0021870131604373455, + 0.0016649471363052726, + 0.002452879911288619, + 0.001569915795698762, + 0.0020355393644422293, + 0.0017061387188732624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.01418930571526289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014184722676873207 + ], + "opposite_logprob_tokens": [ + -4.264184951782227 + ], + "logprob_reward": -0.014184722676873207, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014184722676873207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0692804828286171, + 0.027072926983237267, + 0.05601664260029793, + 0.02728055790066719, + 0.043588317930698395, + 0.10392860323190689, + 0.09382542222738266, + 0.5254332423210144, + 0.022685309872031212, + 0.030888507142663002 + ], + "credit_map": [ + 0.00837729312479496, + 0.007378728128969669, + 0.005789424758404493, + 0.009526839479804039, + 0.008121001534163952, + 0.009514573961496353, + 0.012291635386645794, + 0.00828099250793457, + 0.0024684167001396418, + 0.00277339993044734, + 0.0025994882453233004, + 0.005734511651098728, + 0.0036665466614067554, + 0.004125590901821852, + 0.0030967730563133955, + 0.0026082005351781845, + 0.005742622539401054, + 0.006586354225873947, + 0.0077059329487383366, + 0.008343469351530075, + 0.009056066162884235, + 0.004607101436704397, + 0.008802810683846474, + 0.00517228152602911, + 0.001828133244998753, + 0.0026810767594724894, + 0.0037685935385525227, + 0.0033829829189926386, + 0.0048012700863182545, + 0.0019478260073810816, + 0.00612025149166584, + 0.00275042443536222, + 0.0024334865156561136, + 0.001971794757992029, + 0.004786917939782143, + 0.003556192386895418, + 0.0035666772164404392, + 0.0019711144268512726, + 0.015659315511584282, + 0.00964281801134348, + 0.006434221286326647, + 0.012994367629289627, + 0.01468981709331274, + 0.026411209255456924, + 0.01029759831726551, + 0.007854283787310123, + 0.011398141272366047, + 0.01384896133095026, + 0.009200853295624256, + 0.00942568015307188, + 0.019172513857483864, + 0.010946610011160374, + 0.007690786384046078, + 0.014023213647305965, + 0.010231074877083302, + 0.013134687207639217, + 0.022056296467781067, + 0.009795689024031162, + 0.19483040273189545, + 0.0223587267100811, + 0.013752278871834278, + 0.007995205000042915, + 0.20945695042610168, + 0.04518771544098854, + 0.00228080409578979, + 0.0029132848139852285, + 0.0033967609051615, + 0.00342738744802773, + 0.0015950388042256236, + 0.0022698824759572744, + 0.00318712554872036, + 0.0036150244995951653, + 0.0016872353153303266, + 0.0021916022524237633, + 0.005888466257601976, + 0.007723723538219929, + 0.00148570176679641, + 0.0014507215237244964, + 0.004600753542035818, + 0.005860301665961742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.014184722676873207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, eye-level tracking shot follows a lone traveler as they glide smoothly through a vast, modern airport terminal bathed in the golden glow of the afternoon sun. Light streams through vast floor-to-ceiling windows, casting long, dramatic shadows across the highly polished concrete floor and creating cinematic lens flares as the camera moves. The traveler, a quiet figure with medium-length dark brown hair, wears a practical olive green canvas jacket paired with a simple white T-shirt and dark denim jeans for a long journey. They walk with a steady, deliberate pace, one hand firmly grasping the extended handle of a sleek, dark gray hardshell rolling suitcase, its wheels gliding almost silently across the floor. In the background, the vast expanse of the concourse stretches outward, its high, vaulted ceiling supported by minimalist steel beams, while other passengers become blurred, anonymous figures in the distance. Through the expansive glass, the silhouettes of airplanes are visible on the sun-drenched tarmac, imbuing the scene with a hopeful and contemplative atmosphere, capturing a moment of transition and the purposeful beginning of an important journey to a distant gate.", + "seed": 29003, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000162/group_9/index_7040_seed_29003_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010045909322798252, + "vlm_logprob": -0.010045909322798252, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.010045909322798252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008667821995913982 + ], + "opposite_logprob_tokens": [ + -6.883667945861816 + ], + "logprob_reward": -0.008667821995913982, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.008667821995913982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.245767280459404, + 0.10655443370342255, + 0.15505459904670715, + 0.10256195068359375, + 0.07991093397140503, + 0.04363875836133957, + 0.021386971697211266, + 0.09303544461727142, + 0.09987325966358185, + 0.05221632868051529 + ], + "credit_map": [ + 0.02320259064435959, + 0.02954636514186859, + 0.031575821340084076, + 0.041987545788288116, + 0.037541136145591736, + 0.013167408294975758, + 0.041781771928071976, + 0.026964642107486725, + 0.01442919485270977, + 0.013869172893464565, + 0.014559818431735039, + 0.016010412946343422, + 0.009498625993728638, + 0.006937566678971052, + 0.016969185322523117, + 0.014280456118285656, + 0.02995273470878601, + 0.018526844680309296, + 0.024419274181127548, + 0.028427034616470337, + 0.01344351563602686, + 0.012218138203024864, + 0.01050752867013216, + 0.017559530213475227, + 0.017653433606028557, + 0.010044693946838379, + 0.020127078518271446, + 0.009837707504630089, + 0.014953123405575752, + 0.008299175649881363, + 0.008598068729043007, + 0.013048669323325157, + 0.010598769411444664, + 0.00810984242707491, + 0.011555559001863003, + 0.015059416182339191, + 0.008982419967651367, + 0.009408989921212196, + 0.008700554259121418, + 0.007495380938053131, + 0.004712781868875027, + 0.0030078995041549206, + 0.012023783288896084, + 0.003823887789621949, + 0.003382432274520397, + 0.004081787075847387, + 0.0033257470931857824, + 0.009280440397560596, + 0.0028396781999617815, + 0.001809759414754808, + 0.0023169887717813253, + 0.0022094622254371643, + 0.0020803422667086124, + 0.0022540672216564417, + 0.0041555133648216724, + 0.003721159417182207, + 0.006145757157355547, + 0.005242826882749796, + 0.030967863276600838, + 0.005503268912434578, + 0.004916775971651077, + 0.007322864141315222, + 0.005188200157135725, + 0.027747884392738342, + 0.005396914668381214, + 0.004621672909706831, + 0.006507945712655783, + 0.021632282063364983, + 0.004248259589076042, + 0.014391075819730759, + 0.007720255292952061, + 0.0353548526763916, + 0.004365865606814623, + 0.004608983173966408, + 0.005194979719817638, + 0.004651591181755066, + 0.00414088973775506, + 0.013212759979069233, + 0.0038823538925498724, + 0.012158908881247044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008667821995913982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655712284147739 + ], + "opposite_logprob_tokens": [ + -5.379655838012695 + ], + "logprob_reward": -0.004655712284147739, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655712284147739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11541646718978882, + 0.04615532606840134, + 0.1304057389497757, + 0.025916287675499916, + 0.17644892632961273, + 0.05432794988155365, + 0.028453126549720764, + 0.3598540127277374, + 0.04040435329079628, + 0.022617805749177933 + ], + "credit_map": [ + 0.012254907749593258, + 0.018675409257411957, + 0.016577133908867836, + 0.010533730499446392, + 0.026969200000166893, + 0.009337968192994595, + 0.012016106396913528, + 0.009052015841007233, + 0.007999984547495842, + 0.00623661233112216, + 0.005627502221614122, + 0.00561886141076684, + 0.006437117233872414, + 0.003909952472895384, + 0.0036237090826034546, + 0.0067015900276601315, + 0.03195933625102043, + 0.014867004938423634, + 0.018231473863124847, + 0.007876814343035221, + 0.02361379563808441, + 0.00852030236274004, + 0.011804105713963509, + 0.013532925397157669, + 0.003432891797274351, + 0.0034466260112822056, + 0.006073688622564077, + 0.0028343135491013527, + 0.00343190785497427, + 0.0025437138974666595, + 0.001894479850307107, + 0.0022586665581911802, + 0.020305756479501724, + 0.01944516785442829, + 0.029937081038951874, + 0.02920694835484028, + 0.018888365477323532, + 0.021367371082305908, + 0.021104473620653152, + 0.016193781048059464, + 0.0061283898539841175, + 0.0036824815906584263, + 0.015112501569092274, + 0.004065980669111013, + 0.0038527061697095633, + 0.004319436382502317, + 0.003132185898721218, + 0.01403427217155695, + 0.0025426193606108427, + 0.002743818098679185, + 0.0025521162897348404, + 0.0035561274271458387, + 0.0038231739308685064, + 0.002359742298722267, + 0.006189057603478432, + 0.004686474334448576, + 0.021606124937534332, + 0.014832724817097187, + 0.1319466084241867, + 0.02033953368663788, + 0.010041123256087303, + 0.01897130347788334, + 0.016999104991555214, + 0.1251174807548523, + 0.004656502511352301, + 0.004558265674859285, + 0.005168640520423651, + 0.0043553742580115795, + 0.003014110727235675, + 0.005278652533888817, + 0.005386576987802982, + 0.007986231707036495, + 0.0015970021486282349, + 0.002929392969235778, + 0.0022414522245526314, + 0.0020135845988988876, + 0.002316711936146021, + 0.0051774620078504086, + 0.0017158406553789973, + 0.00462636211887002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.004655712284147739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606578752398491 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.01606578752398491, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606578752398491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.189442977309227, + 0.09013620018959045, + 0.21893957257270813, + 0.05054434388875961, + 0.04010242968797684, + 0.07119253277778625, + 0.017096523195505142, + 0.19197173416614532, + 0.06724724173545837, + 0.06332645565271378 + ], + "credit_map": [ + 0.014580932445824146, + 0.026461239904165268, + 0.026209047064185143, + 0.027652451768517494, + 0.037149589508771896, + 0.014919961802661419, + 0.027574824169278145, + 0.014894930645823479, + 0.011855010874569416, + 0.011355666443705559, + 0.011742815375328064, + 0.010111398063600063, + 0.010838213376700878, + 0.006301600951701403, + 0.011133861728012562, + 0.01679762825369835, + 0.0362757183611393, + 0.02951674349606037, + 0.0362822487950325, + 0.03661121055483818, + 0.02911040559411049, + 0.013429144397377968, + 0.018374161794781685, + 0.019339943304657936, + 0.006996667478233576, + 0.004927529022097588, + 0.009580563753843307, + 0.004676728043705225, + 0.014022083953022957, + 0.0028294865041971207, + 0.004466651007533073, + 0.0030446373857557774, + 0.003915220033377409, + 0.003627395024523139, + 0.004938794765621424, + 0.006747949402779341, + 0.006162824109196663, + 0.0043588606640696526, + 0.006681014783680439, + 0.0036703674122691154, + 0.008215253241360188, + 0.0046686166897416115, + 0.023272907361388206, + 0.004764004610478878, + 0.0035085075069218874, + 0.004876295570284128, + 0.003918421920388937, + 0.01796852797269821, + 0.00157838873565197, + 0.0013824172783643007, + 0.0016704535810276866, + 0.001952750259079039, + 0.0024058958515524864, + 0.0017200285801663995, + 0.0038404464721679688, + 0.0025461420882493258, + 0.01118492428213358, + 0.007606442552059889, + 0.06947075575590134, + 0.010573710314929485, + 0.005629891064018011, + 0.010191618464887142, + 0.008492507971823215, + 0.06882188469171524, + 0.003606516635045409, + 0.005702056922018528, + 0.0034151801373809576, + 0.013644115999341011, + 0.004181436263024807, + 0.008778003044426441, + 0.005911523476243019, + 0.022008415311574936, + 0.0038778542075306177, + 0.004209213890135288, + 0.005333803128451109, + 0.004872786346822977, + 0.005873607937246561, + 0.019444994628429413, + 0.004452879074960947, + 0.015261318534612656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.01606578752398491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774561040103436 + ], + "opposite_logprob_tokens": [ + -4.634774684906006 + ], + "logprob_reward": -0.009774561040103436, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774561040103436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16207177937030792, + 0.07487110048532486, + 0.08862922340631485, + 0.04391501843929291, + 0.0511908084154129, + 0.05101035162806511, + 0.0315370038151741, + 0.29433828592300415, + 0.04309503361582756, + 0.15934136509895325 + ], + "credit_map": [ + 0.008794333785772324, + 0.023569775745272636, + 0.02303396351635456, + 0.02852909453213215, + 0.03403579071164131, + 0.01008386630564928, + 0.023338399827480316, + 0.010686559602618217, + 0.00995594635605812, + 0.010125590488314629, + 0.011516477912664413, + 0.010901988483965397, + 0.00950678065419197, + 0.0056391567923128605, + 0.007250621914863586, + 0.009974542073905468, + 0.01674114167690277, + 0.011404067277908325, + 0.013547804206609726, + 0.013254423625767231, + 0.011795971542596817, + 0.007339599076658487, + 0.006456870585680008, + 0.008089352399110794, + 0.006528841331601143, + 0.0037201568484306335, + 0.008787021040916443, + 0.005312757100909948, + 0.006763567682355642, + 0.0033859123941510916, + 0.005245862994343042, + 0.004170906264334917, + 0.005770498421043158, + 0.004477981477975845, + 0.008265688084065914, + 0.008949533104896545, + 0.004949039779603481, + 0.007297900039702654, + 0.005752825643867254, + 0.005727343261241913, + 0.008527147583663464, + 0.003993109799921513, + 0.011686926707625389, + 0.004948579240590334, + 0.004165247548371553, + 0.00542040541768074, + 0.00395484734326601, + 0.008314091712236404, + 0.002632297109812498, + 0.0031138944905251265, + 0.0034419160801917315, + 0.00494832918047905, + 0.0040505737997591496, + 0.0030331804882735014, + 0.006007245276123285, + 0.004309569485485554, + 0.017488328740000725, + 0.011309749446809292, + 0.10579525679349899, + 0.016760505735874176, + 0.009942183271050453, + 0.015829043462872505, + 0.012397120706737041, + 0.10481616109609604, + 0.004818654619157314, + 0.006034534890204668, + 0.004157504066824913, + 0.003604111960157752, + 0.002188835758715868, + 0.007638109382241964, + 0.00409812992438674, + 0.010555156506597996, + 0.006443378049880266, + 0.0053044455125927925, + 0.012560289353132248, + 0.006880633998662233, + 0.009275522083044052, + 0.05770710110664368, + 0.008381668478250504, + 0.05278833955526352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.009774561040103436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, modern airport terminal with high vaulted ceilings and steel beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by golden afternoon sun streaming through floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the traveler wearing an olive green canvas jacket, a white T-shirt, and dark denim jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the traveler pulling a sleek, dark gray hardshell rolling suitcase with one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a wide, eye-level tracking shot moving smoothly alongside the traveler?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011065663769841194 + ], + "opposite_logprob_tokens": [ + -4.511065483093262 + ], + "logprob_reward": -0.011065663769841194, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011065663769841194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07516124099493027, + 0.0798039436340332, + 0.2533246874809265, + 0.02295610122382641, + 0.062475502490997314, + 0.028592916205525398, + 0.06551318615674973, + 0.3251003921031952, + 0.06643643230199814, + 0.020635630935430527 + ], + "credit_map": [ + 0.0065927328541874886, + 0.009534542448818684, + 0.010437033139169216, + 0.012349606491625309, + 0.008443490602076054, + 0.0098803099244833, + 0.00826207920908928, + 0.009661443531513214, + 0.01109849102795124, + 0.010550322942435741, + 0.011057475581765175, + 0.007739135064184666, + 0.011959991417825222, + 0.007823582738637924, + 0.009958806447684765, + 0.00961613655090332, + 0.05132712423801422, + 0.03171512857079506, + 0.04165422543883324, + 0.03739048168063164, + 0.02719326876103878, + 0.01713704504072666, + 0.018790746107697487, + 0.02811664342880249, + 0.003506774315610528, + 0.0027418392710387707, + 0.004621441476047039, + 0.002141831675544381, + 0.00299969594925642, + 0.002302591223269701, + 0.002530528698116541, + 0.002111397683620453, + 0.005828583613038063, + 0.007368342485278845, + 0.007772226817905903, + 0.008653935976326466, + 0.009954327717423439, + 0.008073005825281143, + 0.010081703774631023, + 0.004743377212435007, + 0.0040848092176020145, + 0.002578994259238243, + 0.0058296541683375835, + 0.0032934648916125298, + 0.0035405505914241076, + 0.00340548949316144, + 0.0027809753082692623, + 0.0030789796728640795, + 0.008842569775879383, + 0.005205721594393253, + 0.007684573996812105, + 0.005858046934008598, + 0.008765009231865406, + 0.005468135699629784, + 0.013984746299684048, + 0.00970438588410616, + 0.01902160793542862, + 0.012828771024942398, + 0.11295638233423233, + 0.01908382773399353, + 0.012471737340092659, + 0.018081240355968475, + 0.016790291294455528, + 0.11386653035879135, + 0.0028551011346280575, + 0.003452856559306383, + 0.0035976003855466843, + 0.01650511473417282, + 0.004537184257060289, + 0.007386512588709593, + 0.007875822484493256, + 0.020226238295435905, + 0.001670687459409237, + 0.0030741416849195957, + 0.0029918747022747993, + 0.002580734435468912, + 0.0025446268264204264, + 0.002824801718816161, + 0.00249386765062809, + 0.0024548950605094433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.011065663769841194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file